title
stringlengths 14
154
| paper_url
stringlengths 42
42
| authors
sequencelengths 1
21
| type
stringclasses 3
values | abstract
stringlengths 413
2.52k
| keywords
stringlengths 4
397
| TL;DR
stringlengths 5
250
⌀ | submission_number
int64 2
14.3k
| arxiv_id
stringlengths 10
10
⌀ | embedding
sequencelengths 384
384
| github_url
stringlengths 0
126
| github_stars
int64 0
55k
| num_models
int64 0
82
| num_datasets
int64 0
7
| num_spaces
int64 0
100
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Linear Multistep Solver Distillation for Fast Sampling of Diffusion Models | https://openreview.net/forum?id=vkOFOUDLTn | [
"Yuchen Liang",
"Xiangzhong Fang",
"Hanting Chen",
"Yunhe Wang"
] | Poster | Sampling from diffusion models can be seen as solving the corresponding
probability flow ordinary differential equation (ODE).
The solving process requires a significant number of function
evaluations (NFE), making it time-consuming.
Recently, several solver search frameworks have attempted to find
better-performing model-specific solvers. However, predicting the impact of
intermediate solving strategies on final sample quality remains challenging,
rendering the search process inefficient.
In this paper, we propose a novel method for designing
solving strategies. We first introduce a unified prediction formula
for linear multistep solvers. Subsequently, we present a solver distillation
framework, which enables a student solver to mimic the sampling trajectory
generated by a teacher solver with more steps. We utilize the mean Euclidean
distance between the student and teacher sampling trajectories as a metric,
facilitating rapid adjustment and optimization of intermediate solving strategies.
The design space of our framework encompasses multiple aspects,
including prediction coefficients, time step schedules, and time scaling
factors.
Our framework has the ability to complete a solver search
for Stable-Diffusion in under 12 total GPU hours.
Compared to previous reinforcement learning-based
search frameworks,
our approach achieves over a 10$\times$ increase in search efficiency.
With just 5 NFE, we achieve FID scores of 3.23 on CIFAR10, 7.16 on ImageNet-64,
5.44 on LSUN-Bedroom, and 12.52 on MS-COCO, resulting in a 2$\times$ sampling acceleration ratio
compared to handcrafted solvers. | Diffusion Probabilistic Model, Diffusion Sampler, Solver Schedule | We provide a solver distillation framework for diffusion models and search for solver schedules based on it. | 165 | null | [
-0.11243943870067596,
-0.0410580150783062,
0.0426013357937336,
0.010704639367759228,
0.019438503310084343,
-0.12209674715995789,
-0.09960082918405533,
0.012071793898940086,
0.05512786656618118,
-0.010126853361725807,
-0.07445307821035385,
-0.03335864841938019,
-0.03172942250967026,
0.07803778350353241,
-0.03359553590416908,
-0.033845577389001846,
-0.017713584005832672,
0.08378889411687851,
-0.04712960124015808,
-0.07162240147590637,
-0.004274621605873108,
-0.036992333829402924,
-0.009949455969035625,
-0.012664224952459335,
0.02366991713643074,
0.0049910880625247955,
0.08172407746315002,
-0.10217641294002533,
-0.0021143658086657524,
-0.022029222920536995,
0.09734436869621277,
-0.002071630209684372,
-0.012025121599435806,
-0.021843325346708298,
-0.08431196212768555,
-0.004021703265607357,
-0.0867052674293518,
0.017344309017062187,
-0.03842048719525337,
-0.00337775400839746,
-0.019706184044480324,
0.031905028969049454,
0.008374355733394623,
0.0786847397685051,
0.09548389911651611,
-0.016461264342069626,
0.00005779555067420006,
-0.0717139020562172,
0.0790502205491066,
0.021407855674624443,
-0.04532564803957939,
-0.05163038522005081,
-0.058660611510276794,
-0.035743579268455505,
0.01950409822165966,
-0.02643599361181259,
0.086492158472538,
-0.03480808436870575,
0.033669013530015945,
-0.04645105078816414,
-0.08393460512161255,
-0.06115516647696495,
-0.041286855936050415,
-0.004455970134586096,
-0.07851918041706085,
-0.01677826978266239,
0.0412108451128006,
0.09320700913667679,
0.04639463126659393,
0.09810248017311096,
-0.07315929979085922,
0.05133671313524246,
-0.029497642070055008,
-0.015681181102991104,
0.056460633873939514,
-0.005977306514978409,
0.009719897992908955,
0.003651693696156144,
-0.014120999723672867,
-0.03354644030332565,
0.04903916269540787,
0.004789027851074934,
0.006184765603393316,
-0.05004554241895676,
0.06129033491015434,
-0.05380827561020851,
0.023045893758535385,
0.04546245187520981,
0.14856405556201935,
-0.08053523302078247,
0.009077348746359348,
0.10109180957078934,
-0.08518577367067337,
-0.0752519965171814,
0.019713709130883217,
0.011946318671107292,
0.0009051855304278433,
0.01155521348118782,
0.019372988492250443,
0.03961056098341942,
-0.045320283621549606,
0.08632896095514297,
0.0665004625916481,
-0.0845583975315094,
-0.054738014936447144,
-0.03934619203209877,
0.054943110793828964,
0.08523237705230713,
0.043182920664548874,
-0.014719976112246513,
0.03427991271018982,
-0.0019415264250710607,
0.06652684509754181,
0.10587606579065323,
0.08091790229082108,
-0.053939204663038254,
0.055854734033346176,
-0.10381261259317398,
0.008621471002697945,
0.0887756273150444,
-0.012709097936749458,
0.006465636193752289,
-0.035861145704984665,
-0.06737594306468964,
-0.010144604369997978,
-0.02581515535712242,
-0.02662409096956253,
1.8389707294042417e-33,
0.002744884230196476,
0.027198005467653275,
0.04614663124084473,
0.03345036879181862,
0.043685875833034515,
0.021262992173433304,
0.0814184695482254,
0.050080858170986176,
0.015209258534014225,
0.05246540158987045,
-0.009783823043107986,
-0.026690902188420296,
-0.12026380747556686,
0.07359291613101959,
0.02955676056444645,
-0.03445712849497795,
-0.030694058164954185,
0.0346248559653759,
-0.03729424253106117,
0.016655396670103073,
0.04811009392142296,
-0.09984329342842102,
-0.0364810936152935,
-0.055783141404390335,
-0.021535059437155724,
0.05458301678299904,
-0.008932695724070072,
-0.0035722567699849606,
0.057912491261959076,
-0.018521616235375404,
-0.07050495594739914,
-0.009993672370910645,
-0.13356620073318481,
0.055200885981321335,
0.04085231572389603,
0.025793395936489105,
0.007111004088073969,
-0.016469784080982208,
0.005169850308448076,
-0.024703627452254295,
-0.04700060933828354,
-0.014607596211135387,
0.07705231010913849,
0.02624925971031189,
-0.10621137917041779,
-0.025712789967656136,
-0.026482511311769485,
-0.031383614987134933,
0.021870315074920654,
-0.044007208198308945,
-0.025384332984685898,
-0.007776562124490738,
-0.0538502000272274,
-0.09454738348722458,
-0.009843987412750721,
0.07739589363336563,
0.038804586976766586,
0.002447127364575863,
-0.02672399766743183,
0.037104032933712006,
-0.013954682275652885,
0.04085855931043625,
-0.0733497142791748,
0.0032684006728231907,
0.05335574969649315,
0.04535176604986191,
-0.05122058838605881,
0.07252869755029678,
0.11829861998558044,
0.03585787117481232,
-0.015209531411528587,
-0.04964426904916763,
0.00901687704026699,
-0.07146744430065155,
0.07784443348646164,
-0.04038221016526222,
0.07871279120445251,
-0.010847087018191814,
0.0011532724602147937,
-0.025607608258724213,
-0.025412889197468758,
-0.06905337423086166,
-0.07514693588018417,
-0.0616275928914547,
0.004318065941333771,
-0.07378564029932022,
0.027898332104086876,
-0.06093925982713699,
-0.0601501390337944,
-0.0871129184961319,
-0.0696941390633583,
0.02787856012582779,
-0.015927331522107124,
0.0655340626835823,
0.022341741248965263,
-2.2357540720785046e-33,
0.07021761685609818,
-0.02926774136722088,
0.06246104836463928,
0.11913082003593445,
0.07433614134788513,
-0.05405564233660698,
-0.011231972835958004,
-0.051039665937423706,
0.1529415249824524,
-0.0982866957783699,
-0.0476546548306942,
-0.041446294635534286,
0.027488360181450844,
0.013272058218717575,
0.05508585274219513,
0.04156140610575676,
0.010410331189632416,
0.022141804918646812,
-0.022100400179624557,
-0.00030616039293818176,
-0.061829887330532074,
0.0619681216776371,
-0.06629002094268799,
-0.03353376314043999,
-0.0505361407995224,
-0.030615149065852165,
0.027492385357618332,
0.038134071975946426,
-0.0008864455739967525,
0.023945217952132225,
-0.05094274505972862,
-0.04369622841477394,
0.04103633761405945,
-0.029322586953639984,
-0.056355349719524384,
0.09565013647079468,
0.014680224470794201,
0.021176328882575035,
-0.01897265389561653,
0.08163745701313019,
0.06914286315441132,
-0.0790579691529274,
0.0020857544150203466,
0.014446236193180084,
0.030552128329873085,
0.09195110946893692,
-0.05134425312280655,
0.018809104338288307,
0.029171587899327278,
0.04564845189452171,
0.010302764363586903,
0.04809468984603882,
-0.05258263647556305,
0.019856562837958336,
0.0039244345389306545,
0.018204132094979286,
-0.05587540566921234,
-0.0424831360578537,
0.061083871871232986,
0.028125924989581108,
-0.010187789797782898,
0.007760840468108654,
0.05429346486926079,
-0.01201455295085907,
-0.04348495602607727,
-0.027462361380457878,
-0.047435466200113297,
0.038533493876457214,
-0.015317302197217941,
0.08619832247495651,
-0.08574821054935455,
-0.009664373472332954,
0.08529845625162125,
-0.011089966632425785,
0.003384264884516597,
-0.10023419559001923,
0.005587887018918991,
-0.08344527333974838,
0.020544176921248436,
-0.013995267450809479,
0.004839614499360323,
0.05769367888569832,
0.009505445137619972,
-0.01580040529370308,
-0.013629394583404064,
-0.06060121953487396,
0.054980941116809845,
0.08671068400144577,
0.05196527764201164,
-0.024937642738223076,
-0.007274650037288666,
0.00879796501249075,
0.06686322391033173,
0.02914516255259514,
-0.0565742552280426,
-4.4253205544464436e-8,
-0.06330548971891403,
0.05958335101604462,
0.020518779754638672,
0.018413178622722626,
0.06628849357366562,
0.0320650190114975,
-0.0036280592903494835,
0.017864709720015526,
-0.010069294832646847,
0.025985579937696457,
0.13518381118774414,
-0.02468704804778099,
0.018991662189364433,
-0.036796994507312775,
0.01341153308749199,
0.10126239061355591,
0.017862897366285324,
-0.025277700275182724,
-0.029034728184342384,
-0.026776662096381187,
0.020770259201526642,
-0.0024322045501321554,
-0.018624361604452133,
-0.012152639217674732,
0.021191006526350975,
0.022522814571857452,
-0.003957625478506088,
0.02982516586780548,
0.09821437299251556,
0.019146427512168884,
0.022359589114785194,
0.019374698400497437,
0.02832200564444065,
0.017875615507364273,
0.03899897262454033,
0.009279889985918999,
-0.0000818757907836698,
-0.012463628314435482,
0.0299617238342762,
-0.02437000162899494,
0.018829232081770897,
0.0407695434987545,
-0.050267934799194336,
0.0030009800102561712,
0.009817923419177532,
0.025941133499145508,
-0.03483156859874725,
-0.017181579023599625,
0.048668622970581055,
0.08895892649888992,
0.009143389761447906,
0.00011784940579673275,
-0.024622904136776924,
0.021154126152396202,
0.11633896827697754,
-0.04623367264866829,
-0.06489773839712143,
0.004757300019264221,
0.04994245991110802,
-0.03608700633049011,
0.008912929333746433,
0.0062558893114328384,
-0.015801670029759407,
0.05930573120713234
] | 0 | 0 | 0 | 0 |
|
ImagineNav: Prompting Vision-Language Models as Embodied Navigator through Scene Imagination | https://openreview.net/forum?id=vQFw9ryKyK | [
"Xinxin Zhao",
"Wenzhe Cai",
"Likun Tang",
"Teng Wang"
] | Poster | Visual navigation is an essential skill for home-assistance robots, providing the object-searching ability to accomplish long-horizon daily tasks. Many recent approaches use Large Language Models (LLMs) for commonsense inference to improve exploration efficiency. However, the planning process of LLMs is limited within texts and it is difficult to represent the spatial occupancy and geometry layout only by texts. Both are important for making rational navigation decisions. In this work, we seek to unleash the spatial perception and planning ability of Vision-Language Models (VLMs), and explore whether the VLM, with only on-board camera captured RGB/RGB-D stream inputs, can efficiently finish the visual navigation tasks in a mapless manner. We achieve this by developing the imagination-powered navigation framework ImagineNav, which imagines the future observation images at valuable robot views and translates the complex navigation planning process into a rather simple best-view image selection problem for VLM. To generate appropriate candidate robot views for imagination, we introduce the Where2Imagine module, which is distilled to align with human navigation habits. Finally, to reach the VLM preferred views, an off-the-shelf point-goal navigation policy is utilized. Empirical experiments on the challenging open-vocabulary object navigation benchmarks demonstrates the superiority of our proposed system. | Robotics, Visual Navigation, Vision-Language Model, Scene Imagination | We propose a mapless visual navigation system by proposing a imagination-based visual prompting for pre-trained large vision-language models. | 159 | 2410.09874 | [
0.06055102124810219,
-0.08710850030183792,
0.042327895760536194,
-0.03522607311606407,
0.10461201518774033,
0.004041256383061409,
-0.011834794655442238,
0.004298200830817223,
0.017470763996243477,
0.009470040909945965,
-0.05296311154961586,
-0.06308770924806595,
0.046148039400577545,
0.07627267390489578,
0.0836159735918045,
0.02198314107954502,
0.0848735123872757,
0.05333930253982544,
-0.0534193217754364,
0.009579874575138092,
0.03723405674099922,
0.03549426421523094,
0.0631398931145668,
-0.028576664626598358,
0.00020928270532749593,
0.05180205777287483,
0.08916754275560379,
-0.05447645112872124,
0.007122632581740618,
-0.03469476103782654,
0.014365541748702526,
0.0010093499440699816,
-0.0025114051532000303,
0.01578674279153347,
-0.011227584443986416,
0.029690533876419067,
-0.0790112316608429,
-0.03049139492213726,
-0.04384636878967285,
-0.09589458256959915,
-0.032823849469423294,
0.02045399136841297,
0.05142553150653839,
-0.007389783393591642,
0.06930268555879593,
0.08356941491365433,
0.04645300284028053,
-0.041659869253635406,
0.06644388288259506,
0.013481676578521729,
-0.15742284059524536,
-0.08381116390228271,
-0.0024434244260191917,
-0.05344371870160103,
-0.04741213098168373,
0.046583667397499084,
-0.014238272793591022,
-0.07060687988996506,
0.09280665218830109,
-0.0379246324300766,
-0.00875244103372097,
-0.033921219408512115,
0.04325828701257706,
0.00449138879776001,
-0.06362265348434448,
-0.050241876393556595,
0.011314762756228447,
0.016360605135560036,
0.017038198187947273,
-0.018436532467603683,
-0.01747548021376133,
0.03904510289430618,
0.046731047332286835,
-0.014845401048660278,
0.013960149139165878,
-0.01105944998562336,
0.03313633054494858,
-0.09115270525217056,
0.07402557879686356,
-0.07384029030799866,
0.036606986075639725,
0.017266836017370224,
-0.019766759127378464,
0.07754170894622803,
0.04849104583263397,
0.04802565649151802,
0.01125423889607191,
0.03673003241419792,
0.06514977663755417,
-0.013394389301538467,
-0.009992220439016819,
-0.13395078480243683,
0.013331265188753605,
-0.022971460595726967,
-0.007517190650105476,
0.002021028893068433,
-0.028167923912405968,
-0.12002202123403549,
-0.0073488205671310425,
0.006760671269148588,
0.01272277720272541,
0.018804725259542465,
0.06394311040639877,
-0.03503071516752243,
0.003921132534742355,
-0.027881529182195663,
0.07400580495595932,
0.06494184583425522,
0.01874181069433689,
-0.04782570153474808,
0.052229270339012146,
-0.0546351782977581,
0.023078076541423798,
0.08361094444990158,
-0.02236807346343994,
0.0116298021748662,
0.052596744149923325,
0.02651946432888508,
0.041120391339063644,
-0.0038713538087904453,
-0.016942497342824936,
0.025482453405857086,
0.06635288149118423,
-0.020035719498991966,
0.05456577613949776,
-0.06495894491672516,
0.013485334813594818,
-3.472113683676447e-34,
0.004964117426425219,
0.009019659832119942,
0.0581149123609066,
0.031473640352487564,
0.09145420789718628,
0.007886159233748913,
0.026172438636422157,
0.04923761263489723,
0.002960427198559046,
0.01867079734802246,
-0.008674495853483677,
-0.04499367997050285,
-0.03966682404279709,
0.12883591651916504,
0.05158254876732826,
-0.021670736372470856,
0.002921260427683592,
0.031163033097982407,
-0.0801241397857666,
0.04527938365936279,
0.03850176930427551,
0.014155064709484577,
0.044554367661476135,
-0.053426846861839294,
0.012509592808783054,
0.024325868114829063,
0.0576201006770134,
-0.04392480477690697,
-0.00035571487387642264,
-0.017423855140805244,
-0.08053592592477798,
0.03341543301939964,
-0.0793747752904892,
0.049586690962314606,
-0.01713535375893116,
-0.005720486398786306,
-0.06580032408237457,
-0.01933322474360466,
0.014578512869775295,
0.01568906381726265,
-0.01272636279463768,
0.07666050642728806,
-0.04941016435623169,
-0.04330946505069733,
-0.1059497520327568,
-0.004743269179016352,
0.010955325327813625,
0.10148554295301437,
-0.0990447923541069,
0.021609969437122345,
0.06338612735271454,
0.029309751465916634,
-0.05359756946563721,
-0.05626332387328148,
-0.018961433321237564,
-0.07793576270341873,
0.032120659947395325,
0.06265345215797424,
0.0429391972720623,
-0.028540510684251785,
0.018404057249426842,
0.0014639039291068912,
0.04880440607666969,
0.08063971996307373,
0.058710966259241104,
0.020298650488257408,
0.019165746867656708,
0.0038033355958759785,
0.04953478276729584,
-0.036485087126493454,
-0.00973974447697401,
-0.017435695976018906,
0.039378512650728226,
-0.06900952011346817,
-0.01196274533867836,
-0.010582146234810352,
-0.014613726176321507,
-0.08934775739908218,
0.10927300155162811,
0.06656482070684433,
-0.04398104175925255,
-0.01356535591185093,
-0.0175851471722126,
0.009525172412395477,
0.04182705655694008,
-0.020458180457353592,
0.03930282220244408,
-0.07514125853776932,
-0.002438692143186927,
-0.07782302051782608,
-0.04846365004777908,
-0.09239059686660767,
-0.02733660861849785,
-0.0576825775206089,
-0.027497833594679832,
1.2567326046954975e-34,
0.022688960656523705,
-0.04957878589630127,
0.04213658347725868,
0.04438513517379761,
-0.06462991237640381,
-0.13146594166755676,
0.0666518434882164,
-0.006410634610801935,
-0.006332089193165302,
-0.11895903944969177,
-0.07326977699995041,
0.04889431595802307,
0.038873281329870224,
0.09462966024875641,
0.05773652344942093,
-0.04835839197039604,
0.05129333585500717,
-0.045640893280506134,
-0.02580011636018753,
0.025183768942952156,
-0.050070539116859436,
0.054177142679691315,
-0.04767243191599846,
-0.03989996388554573,
-0.022028520703315735,
0.08392328023910522,
0.05109049379825592,
0.034818921238183975,
-0.053574446588754654,
-0.004116257186979055,
-0.0478828139603138,
0.011807339265942574,
0.019546756520867348,
-0.03881736472249031,
-0.011801229789853096,
0.06266774237155914,
0.017121989279985428,
-0.032783087342977524,
-0.0749642625451088,
0.11511756479740143,
0.046608030796051025,
-0.07569153606891632,
-0.030238661915063858,
-0.005879046395421028,
-0.06555866450071335,
0.014100194908678532,
-0.038363147526979446,
0.05002664774656296,
-0.006138315889984369,
0.051519088447093964,
-0.042741965502500534,
0.06016138568520546,
-0.07397421449422836,
-0.06563419103622437,
-0.04083525761961937,
-0.024406161159276962,
-0.06185591220855713,
0.004583415575325489,
0.11144521832466125,
-0.03234298527240753,
-0.03912996128201485,
-0.006351036950945854,
-0.027299605309963226,
0.02165093459188938,
-0.01768854819238186,
0.01196278352290392,
-0.07548722624778748,
-0.028584983199834824,
0.03073425218462944,
-0.051200393587350845,
0.05466364324092865,
-0.07919947057962418,
0.023773353546857834,
0.03226521983742714,
0.016378670930862427,
0.021839100867509842,
0.0584416463971138,
0.021895626559853554,
0.11128684133291245,
-0.06069105863571167,
0.008261368609964848,
-0.04012457653880119,
0.017503485083580017,
0.09886520355939865,
0.05974576622247696,
0.04261448234319687,
-0.049092117697000504,
-0.017426960170269012,
0.03777981176972389,
0.03090924583375454,
-0.02815532311797142,
0.004381329286843538,
0.013569407165050507,
0.13015000522136688,
-0.05853598192334175,
-3.872745324429161e-8,
-0.09112674742937088,
0.022996066138148308,
0.05319847911596298,
0.02039915695786476,
0.02798951044678688,
-0.050826460123062134,
-0.01858895644545555,
0.06049978360533714,
-0.014372065663337708,
0.021139945834875107,
0.06362858414649963,
-0.019613170996308327,
-0.030517304316163063,
-0.00523733114823699,
0.03351601958274841,
0.03570028394460678,
0.058669574558734894,
-0.013371673412621021,
-0.022463448345661163,
0.006614152807742357,
0.027589593082666397,
-0.03438584879040718,
-0.024542467668652534,
0.02120133675634861,
-0.015605189837515354,
0.007001680321991444,
-0.09147071093320847,
0.021060306578874588,
0.06645839661359787,
-0.0739629939198494,
0.07311256229877472,
0.0023440751247107983,
0.008487872779369354,
0.050932787358760834,
0.00964940432459116,
-0.03920657932758331,
-0.06463555246591568,
-0.038174938410520554,
0.02479511871933937,
-0.048033323138952255,
0.03443199396133423,
0.06198393553495407,
0.000767551944591105,
0.00439969077706337,
-0.02063649334013462,
0.00805734097957611,
0.007248847745358944,
-0.1403811275959015,
-0.04079345241189003,
-0.02041475661098957,
-0.13438662886619568,
0.00571019109338522,
0.006238812115043402,
0.09967149049043655,
0.07736348360776901,
0.00018761822138912976,
0.026841599494218826,
-0.0707004964351654,
-0.035972271114587784,
0.11501910537481308,
0.009196719154715538,
0.058535970747470856,
-0.11073258519172668,
-0.05609028413891792
] | 0 | 0 | 0 | 0 |
|
Navigation-Guided Sparse Scene Representation for End-to-End Autonomous Driving | https://openreview.net/forum?id=Vv76fCYffN | [
"Peidong Li",
"Dixiao Cui"
] | Poster | End-to-End Autonomous Driving (E2EAD) methods typically rely on supervised perception tasks to extract explicit scene information (e.g., objects, maps). This reliance necessitates expensive annotations and constrains deployment and data scalability in real-time applications. In this paper, we introduce SSR, a novel framework that utilizes only 16 navigation-guided tokens as Sparse Scene Representation, efficiently extracting crucial scene information for E2EAD. Our method eliminates the need for human-designed supervised sub-tasks, allowing computational resources to concentrate on essential elements directly related to navigation intent. We further introduce a temporal enhancement module, aligning predicted future scenes with actual future scenes through self-supervision. SSR achieves a 27.2\% relative reduction in L2 error and a 51.6\% decrease in collision rate to UniAD in nuScenes, with a 10.9× faster inference speed and 13× faster training time. Moreover, SSR outperforms VAD-Base with a 48.6-point improvement on driving score in CARLA's Town05 Long benchmark. This framework represents a significant leap in real-time autonomous driving systems and paves the way for future scalable deployment. Code is available at https://github.com/PeidongLi/SSR. | End-to-End, Autonomous Driving, Sparse Scene Representation | null | 152 | 2409.18341 | [
0.03712772578001022,
-0.06805137544870377,
0.02890012226998806,
0.009264457039535046,
0.08459903299808502,
-0.009293039329349995,
-0.010430509224534035,
0.012511178851127625,
-0.037030383944511414,
-0.0422922782599926,
-0.05378983914852142,
-0.057781368494033813,
-0.018533343449234962,
0.08828778564929962,
-0.020107608288526535,
-0.0050098467618227005,
0.15068994462490082,
0.059764545410871506,
-0.10449286550283432,
-0.04359887167811394,
-0.011489702388644218,
0.03493325784802437,
0.016242779791355133,
0.03440336138010025,
0.05592162162065506,
-0.026560185477137566,
0.015786128118634224,
-0.08657220005989075,
-0.0016799794975668192,
-0.04315229505300522,
0.06355073302984238,
-0.014780599623918533,
-0.048558685928583145,
-0.036537133157253265,
-0.03986835852265358,
-0.024866556748747826,
0.0011291747214272618,
0.023256992921233177,
-0.055078089237213135,
-0.10699764639139175,
0.012721486389636993,
0.014847733080387115,
-0.011823141947388649,
0.07718945294618607,
0.07580118626356125,
0.07738012075424194,
0.058849040418863297,
-0.06262923032045364,
0.08628686517477036,
-0.0323624312877655,
-0.08914606273174286,
0.006892967503517866,
0.02087065763771534,
-0.03976071998476982,
-0.07599000632762909,
0.08291635662317276,
0.020248932763934135,
-0.020432190969586372,
-0.02153860032558441,
0.0538802333176136,
0.01117137260735035,
-0.08881966769695282,
0.052158500999212265,
-0.022484462708234787,
-0.02525218203663826,
-0.03532654047012329,
0.07195905596017838,
-0.010191948153078556,
0.017735503613948822,
0.04847743734717369,
0.0479041263461113,
0.04498559609055519,
0.06677208095788956,
0.03529387339949608,
0.06667003780603409,
0.004608186427503824,
-0.005260175094008446,
-0.03787703067064285,
0.06037183478474617,
-0.09297304600477219,
0.006020217668265104,
-0.013951594941318035,
0.009256651625037193,
0.037879738956689835,
0.07362805306911469,
-0.03760547935962677,
0.023011906072497368,
0.08543407171964645,
-0.005469393916428089,
0.040004316717386246,
-0.043437808752059937,
-0.12022357434034348,
0.017290296033024788,
-0.03323493152856827,
-0.07990588247776031,
-0.044981446117162704,
-0.0035155981313437223,
-0.09373724460601807,
-0.007498553488403559,
0.06895484775304794,
-0.02575879916548729,
0.0515742190182209,
-0.0294779185205698,
-0.04955705255270004,
-0.006040952634066343,
0.05080389603972435,
0.0997193306684494,
0.04031708836555481,
-0.032573238015174866,
-0.0012934838887304068,
0.08982973545789719,
-0.04785959795117378,
-0.01208515651524067,
-0.054572224617004395,
-0.0007135104387998581,
0.022169625386595726,
0.009947462007403374,
0.020505765452980995,
0.0533318929374218,
0.05838879197835922,
-0.04589313268661499,
-0.08433037251234055,
0.013886187225580215,
0.05350305140018463,
0.001084032584913075,
-0.08172415941953659,
-0.017573129385709763,
4.0099790318169354e-34,
-0.02467331849038601,
0.06630873680114746,
0.036185432225465775,
0.002691216068342328,
0.048821497708559036,
0.020520595833659172,
-0.029986707493662834,
0.022165890783071518,
-0.07677070051431656,
0.018492795526981354,
-0.001045142998918891,
-0.08713293075561523,
-0.0019506977405399084,
0.059588704258203506,
0.03497811034321785,
-0.020043451339006424,
-0.026719022542238235,
0.04973750561475754,
-0.08128633350133896,
0.004111382178962231,
0.010107593610882759,
0.019529519602656364,
-0.06297793239355087,
-0.07161227613687515,
0.06993043422698975,
-0.03629371151328087,
0.03108801320195198,
0.002039693295955658,
0.025402303785085678,
0.0035571397747844458,
-0.11644633114337921,
0.05463618412613869,
-0.013346608728170395,
0.08579511195421219,
0.03907434642314911,
0.018595295026898384,
-0.0817241370677948,
-0.033816859126091,
0.0568171925842762,
0.031726136803627014,
0.005019159056246281,
0.06208384409546852,
-0.01317231822758913,
-0.04147743061184883,
-0.13494384288787842,
-0.07937315851449966,
0.05169132351875305,
0.12080015987157822,
-0.04198893904685974,
0.05407123267650604,
0.04244290664792061,
0.014250196516513824,
-0.027540283277630806,
-0.09152057021856308,
-0.08452805876731873,
-0.06544199585914612,
0.049298785626888275,
0.0007829610258340836,
0.01409939955919981,
-0.0036829521413892508,
0.04132974520325661,
-0.02177436277270317,
-0.009953544475138187,
0.02895570732653141,
0.0006479075527749956,
0.021300554275512695,
0.04505985602736473,
0.08502174913883209,
0.07378238439559937,
-0.01533055491745472,
-0.06356961280107498,
0.01892395131289959,
0.07055766880512238,
-0.07040616869926453,
0.12170956283807755,
0.005800400394946337,
0.06921876221895218,
-0.06233043596148491,
0.010801147669553757,
0.05050007253885269,
-0.04305444657802582,
-0.004208947066217661,
-0.02344217523932457,
0.018496815115213394,
0.03379816561937332,
-0.0034091209527105093,
0.021475711837410927,
0.0029728782828897238,
-0.024513836950063705,
-0.016552217304706573,
-0.01830470748245716,
-0.08692748099565506,
-0.047249432653188705,
-0.008981633931398392,
-0.039368968456983566,
2.843133518434067e-34,
0.09504964202642441,
0.057164691388607025,
-0.011242971755564213,
-0.02411567233502865,
-0.03380527347326279,
-0.06996098160743713,
-0.04319453239440918,
-0.016390230506658554,
0.016955772414803505,
-0.03411591798067093,
-0.03481283038854599,
-0.0400230698287487,
0.08925764262676239,
0.0034014214761555195,
-0.03333355858922005,
0.02373327501118183,
0.0007981662056408823,
-0.04014202579855919,
-0.04430507868528366,
-0.03503449261188507,
-0.05214505270123482,
0.07884394377470016,
-0.0278354212641716,
-0.03408980369567871,
-0.07363646477460861,
0.05293840169906616,
0.037764161825180054,
0.08473663032054901,
-0.009435592219233513,
-0.03760648891329765,
-0.012072675861418247,
-0.012021293863654137,
-0.010135059244930744,
-0.12516775727272034,
-0.03257330134510994,
0.04644596576690674,
0.06971801072359085,
-0.03838576748967171,
-0.06162317097187042,
0.037177253514528275,
0.034676238894462585,
-0.02696523256599903,
-0.019374456256628036,
0.0029926744755357504,
-0.0323101170361042,
0.001970787299796939,
-0.04616660997271538,
0.040696125477552414,
-0.0168125182390213,
0.10609068721532822,
-0.07196298241615295,
0.08001888543367386,
-0.12881217896938324,
0.017877863720059395,
-0.022649552673101425,
-0.0841488316655159,
-0.04704516381025314,
0.03061666712164879,
0.05746614560484886,
-0.008128338493406773,
-0.0026736066211014986,
0.03770605847239494,
-0.082334965467453,
0.011506515555083752,
0.019282078370451927,
-0.08688973635435104,
-0.08757548034191132,
-0.014373766258358955,
-0.004559273365885019,
0.02365671657025814,
0.006041505374014378,
0.008493366651237011,
0.005838421173393726,
0.0006017148261889815,
-0.02919621765613556,
0.00725511321797967,
0.10915481299161911,
-0.005532752722501755,
0.0776117816567421,
-0.03951757028698921,
0.03596583381295204,
-0.03445000946521759,
0.05803227797150612,
0.06335349380970001,
0.030026856809854507,
0.12631171941757202,
-0.015194067731499672,
0.03097428008913994,
0.06541473418474197,
0.05099902302026749,
-0.031536441296339035,
0.0008118512923829257,
-0.013724506832659245,
0.09725409001111984,
-0.06274078786373138,
-4.124436614461047e-8,
-0.05288514122366905,
0.08613774180412292,
-0.008973340503871441,
-0.018969517201185226,
-0.044989798218011856,
-0.052953362464904785,
0.02216179296374321,
0.14362646639347076,
-0.026022378355264664,
0.0336155965924263,
0.07350631058216095,
-0.02140979841351509,
-0.03786393254995346,
-0.037394773215055466,
0.007056776899844408,
0.051211342215538025,
0.029087740927934647,
-0.014247252605855465,
-0.03004375472664833,
0.029454153031110764,
-0.026517534628510475,
0.01457615103572607,
0.006729361601173878,
-0.036737438291311264,
0.03600344806909561,
-0.0021810324396938086,
0.006106889806687832,
0.09293994307518005,
0.0697375163435936,
-0.1025007888674736,
-0.017843017354607582,
0.016881922259926796,
0.035990502685308456,
0.0042371139861643314,
0.0026877562049776316,
0.032987356185913086,
0.022652216255664825,
0.04997309297323227,
-0.03895794227719307,
-0.054634448140859604,
0.02424372173845768,
0.07969687879085541,
-0.03575446084141731,
0.04901838302612305,
-0.017554979771375656,
0.019148850813508034,
0.03290474787354469,
-0.06959589570760727,
-0.026724373921751976,
-0.08437985926866531,
-0.03531826287508011,
-0.003113359212875366,
-0.003171948017552495,
0.10516471415758133,
0.04911116883158684,
0.07447529584169388,
0.03320854902267456,
-0.05368852615356445,
0.03252727538347244,
0.08409047871828079,
-0.02794649824500084,
0.02077219821512699,
-0.05900031328201294,
-0.0072057549841701984
] | https://github.com/peidongli/ssr | 151 | 0 | 0 | 0 |
An Evolved Universal Transformer Memory | https://openreview.net/forum?id=s1kyHkdTmi | [
"Edoardo Cetin",
"Qi Sun",
"Tianyu Zhao",
"Yujin Tang"
] | Poster | Prior methods propose to offset the escalating costs of modern foundation models by dropping specific parts of their contexts with hand-designed rules, while attempting to preserve their original performance. We overcome this trade-off with Neural Attention Memory Models (NAMMs), introducing a learned network for memory management that improves both the performance and efficiency of transformers. We evolve NAMMs atop pre-trained transformers to provide different latent contexts focusing on the most relevant information for individual layers and attention heads. NAMMs are universally applicable to any model using self-attention as they condition exclusively on the values in the produced attention matrices. Learning NAMMs on a small set of problems, we achieve substantial performance improvements across multiple long-context benchmarks while cutting the model's input contexts up to a fraction of the original sizes. We show the generality of our conditioning enables zero-shot transfer of NAMMs trained only on language to entirely new transformer architectures even across input modalities, with their benefits carrying over to vision and reinforcement learning. | Transformers, Evolution, Memory, KV cache, attention | We propose a new memory system designed for transformers able to achieve performance improvements in language tasks, and zero-shot transferring to different architectures and input modalities. | 135 | 2410.13166 | [
-0.016621509566903114,
-0.04375999793410301,
0.005955132190138102,
0.05790427699685097,
0.048210904002189636,
0.03043806552886963,
-0.0135655477643013,
-0.004718804731965065,
0.030041730031371117,
-0.018334027379751205,
0.013369821943342686,
-0.011979381553828716,
0.04136578366160393,
0.03452284261584282,
0.029508236795663834,
0.031166408210992813,
0.08346763998270035,
0.08991064876317978,
-0.10736845433712006,
-0.028946807608008385,
-0.003593944478780031,
0.010741189122200012,
0.054379653185606,
-0.023494506254792213,
0.02616082690656185,
0.011899827048182487,
-0.025786934420466423,
-0.07081536948680878,
0.08285056054592133,
-0.014064081944525242,
0.021013885736465454,
-0.007601887919008732,
-0.04994041472673416,
0.056419044733047485,
-0.07406661659479141,
0.05517430603504181,
-0.08071105182170868,
-0.006959030870348215,
-0.033117856830358505,
-0.05856846272945404,
0.04878056049346924,
0.010840361937880516,
-0.017574308440089226,
0.047043099999427795,
0.053268373012542725,
0.060668643563985825,
0.08210358023643494,
-0.04698857665061951,
0.015493059530854225,
-0.06860276311635971,
-0.031176546588540077,
-0.04919160529971123,
-0.03352908045053482,
0.055410515516996384,
-0.011645480990409851,
0.0576738640666008,
0.07353273779153824,
0.02700984477996826,
-0.05880054831504822,
-0.01748579740524292,
-0.09386138617992401,
-0.10398751497268677,
-0.020312417298555374,
-0.0492260679602623,
-0.05647283047437668,
-0.005408384837210178,
0.0655546560883522,
0.05682309716939926,
-0.043075308203697205,
-0.010217375122010708,
-0.050353799015283585,
0.09219863265752792,
-0.0008320473716594279,
0.04734304919838905,
0.0745861679315567,
0.030021412298083305,
0.018697528168559074,
-0.07340822368860245,
0.03158176690340042,
-0.04612923413515091,
0.03847050666809082,
-0.01587587408721447,
0.003963636700063944,
0.004038470331579447,
0.11909832805395126,
0.049830976873636246,
0.00831521674990654,
0.039493314921855927,
0.016851916909217834,
-0.028283048421144485,
-0.0713924765586853,
-0.1045125275850296,
0.03774445876479149,
-0.01716144196689129,
0.0319380983710289,
-0.005918325390666723,
0.007119455374777317,
-0.05666637048125267,
-0.0711747482419014,
0.061546988785266876,
0.03701796010136604,
0.01690538041293621,
0.003978312015533447,
0.046088095754384995,
0.014654233120381832,
-0.005221362691372633,
0.0027376143261790276,
0.09106267988681793,
-0.04216773808002472,
-0.047897882759571075,
0.06629011780023575,
0.05440281704068184,
-0.00033175694989040494,
-0.047774530947208405,
0.07315501570701599,
-0.09014344960451126,
0.013384169898927212,
-0.06982352584600449,
-0.03153089061379433,
0.06529279798269272,
-0.013697528280317783,
0.015060328878462315,
-0.04732154682278633,
0.053590524941682816,
-0.011064033955335617,
-0.007735094986855984,
-0.062037743628025055,
1.5958632529618304e-33,
0.04253853112459183,
0.06237592175602913,
0.026451561599969864,
0.01715742237865925,
0.030701490119099617,
0.04005611315369606,
0.05710873007774353,
0.053886570036411285,
0.03165121003985405,
0.011951602064073086,
0.01981440745294094,
0.055063385516405106,
-0.07373640686273575,
0.05155937746167183,
0.01924772933125496,
-0.014426333829760551,
-0.03873320296406746,
0.029797859489917755,
0.0030626454390585423,
0.012363544665277004,
-0.018518123775720596,
0.033795978873968124,
0.01746167056262493,
-0.04536610469222069,
0.008363647386431694,
-0.005050349980592728,
0.049908123910427094,
-0.05822659656405449,
-0.035989079624414444,
-0.029919259250164032,
-0.12450949847698212,
0.06704656779766083,
-0.008422967046499252,
0.030219729989767075,
0.02896728366613388,
-0.08489872515201569,
0.007061013486236334,
-0.08647480607032776,
0.03455384820699692,
-0.060043785721063614,
-0.00777484942227602,
0.08932610601186752,
-0.022808009758591652,
0.017970209941267967,
-0.09095053374767303,
-0.11992521584033966,
0.003241470782086253,
0.02179158851504326,
-0.024944735690951347,
-0.040248047560453415,
0.04884682223200798,
0.014366497285664082,
-0.07632192224264145,
-0.03965621069073677,
0.014315137639641762,
-0.040475357323884964,
0.01480898354202509,
0.05319938063621521,
0.037726402282714844,
0.04360968619585037,
0.017695168033242226,
0.04230376332998276,
-0.02644338458776474,
0.14474254846572876,
0.03576783835887909,
0.042468175292015076,
-0.004276207182556391,
0.11055131256580353,
0.02248232252895832,
-0.06899932026863098,
-0.043587516993284225,
-0.01808360405266285,
-0.09395937621593475,
-0.018193749710917473,
0.08936040848493576,
-0.028691696003079414,
0.06900645792484283,
-0.06960310786962509,
0.005042810924351215,
-0.036228328943252563,
0.050158143043518066,
0.040264323353767395,
-0.03531135991215706,
-0.0017100016120821238,
-0.004214782267808914,
-0.00586103368550539,
0.06391005218029022,
-0.09590775519609451,
-0.03726203367114067,
-0.03361842781305313,
-0.04729048162698746,
-0.06721829622983932,
-0.018044985830783844,
-0.011732983402907848,
-0.05299927294254303,
-1.5480412040345027e-33,
0.02939419634640217,
-0.0010775153059512377,
-0.08969324082136154,
0.04894080013036728,
-0.027465349063277245,
-0.06816660612821579,
0.03128833696246147,
0.02652919478714466,
-0.04519815742969513,
-0.04207427427172661,
-0.025784894824028015,
-0.018010390922427177,
0.05197590962052345,
-0.03414534404873848,
0.016270313411951065,
-0.052714426070451736,
-0.028289012610912323,
-0.020068932324647903,
-0.010470686480402946,
-0.033460889011621475,
0.02136131562292576,
0.10005846619606018,
-0.06775849312543869,
0.026788368821144104,
-0.12420659512281418,
0.01833530329167843,
-0.10857165604829788,
0.06329409778118134,
0.019118689000606537,
-0.02023753523826599,
-0.0791749581694603,
0.007772911339998245,
0.07567429542541504,
0.01063915342092514,
-0.09452318400144577,
0.05565587431192398,
0.034008946269750595,
0.011422669515013695,
-0.03051215410232544,
0.03795619681477547,
0.03791240230202675,
-0.07291380316019058,
-0.029666507616639137,
0.04859081283211708,
-0.08070387691259384,
-0.038886670023202896,
-0.08388110995292664,
-0.019528446719050407,
-0.014737560413777828,
0.0571775920689106,
-0.04661191627383232,
0.006538074463605881,
-0.09708057343959808,
-0.0747595801949501,
-0.07979553192853928,
-0.08688525855541229,
0.028614191338419914,
-0.06590556353330612,
0.13904546201229095,
-0.03487413004040718,
-0.006049700081348419,
-0.009283673018217087,
-0.02675679326057434,
-0.1178784891963005,
-0.005339848343282938,
-0.004543462302535772,
0.03217317536473274,
0.0023700627498328686,
0.061310019344091415,
0.010292681865394115,
0.026895930990576744,
0.02887582592666149,
0.01050417311489582,
-0.0655684694647789,
0.00159934489056468,
0.02092091366648674,
0.0378451943397522,
-0.03313238546252251,
0.023213939741253853,
-0.021026594564318657,
-0.0464676134288311,
-0.022430939599871635,
0.000792471575550735,
0.027941659092903137,
0.08272504061460495,
0.11769640445709229,
0.07023083418607712,
0.016683563590049744,
0.055368732661008835,
-0.005479168612509966,
-0.05695096403360367,
0.06518785655498505,
0.03534003719687462,
0.04550469294190407,
-0.060711417347192764,
-4.2618662376980865e-8,
-0.11586011201143265,
0.08795526623725891,
0.022488653659820557,
0.07333705574274063,
0.026165327057242393,
-0.0627080425620079,
0.004095250740647316,
0.14155659079551697,
-0.035063158720731735,
0.02043984644114971,
0.06491570174694061,
0.028188396245241165,
0.011004995554685593,
-0.06427673250436783,
0.006441265810281038,
0.0699881911277771,
0.042295344173908234,
0.027644500136375427,
-0.006544800475239754,
-0.042452599853277206,
0.0030487352050840855,
0.04362601786851883,
0.02716090902686119,
-0.04495944082736969,
0.04932675510644913,
-0.0854812040925026,
-0.004367867484688759,
0.08479241281747818,
0.07002485543489456,
0.024889802560210228,
-0.015311665832996368,
-0.01280330028384924,
0.013353602960705757,
-0.06525017321109772,
-0.011724323034286499,
0.08802858740091324,
0.07148554921150208,
-0.005041520576924086,
0.0512072890996933,
-0.05217508599162102,
0.024928253144025803,
0.029733818024396896,
-0.06839363276958466,
0.07232742756605148,
-0.030953146517276764,
0.04101802036166191,
-0.008191165514290333,
-0.09822457283735275,
0.007749213837087154,
-0.02700980193912983,
0.0712348222732544,
0.004456134047359228,
-0.03449700027704239,
0.05936805531382561,
0.05332354083657265,
-0.013021026737987995,
0.05310963839292526,
-0.05289503559470177,
-0.014375390484929085,
0.08298949152231216,
-0.0008542448049411178,
0.021485019475221634,
-0.0562068335711956,
0.044627971947193146
] | https://github.com/sakanaai/evo-memory | 305 | 0 | 0 | 0 |
Interpretable Bilingual Multimodal Large Language Model for Diverse Biomedical Tasks | https://openreview.net/forum?id=YuHQTo6G9S | [
"Lehan Wang",
"Haonan Wang",
"Honglong Yang",
"Jiaji Mao",
"Zehong Yang",
"Jun Shen",
"Xiaomeng Li"
] | Poster | Several medical Multimodal Large Languange Models (MLLMs) have been developed to address tasks involving visual images with textual instructions across various medical modalities, achieving impressive results.
Most current medical generalist models are region-agnostic, treating the entire image as a holistic representation. However, they struggle to identify which specific regions they are focusing on when generating a sentence.
To mimic the behavior of doctors, who typically begin by reviewing the entire image before concentrating on specific regions for a thorough evaluation, we aim to enhance the capability of medical MLLMs in understanding anatomical regions within entire medical scans.
To achieve it, we first formulate \textbf{Region-Centric tasks} and construct a \textbf{large-scale dataset, MedRegInstruct,} to incorporate regional information into training. Combining our collected dataset with other medical multimodal corpora for training, we propose a \textbf{Region-Aware medical MLLM, MedRegA}, which is the first bilingual generalist medical AI system to simultaneously handle image-level and region-level medical vision-language tasks across a broad range of modalities. Our MedRegA not only enables three region-centric tasks, but also achieves the best performance for visual question answering, report generation and medical image classification over 8 modalities, showcasing significant versatility. Experiments demonstrate that our model can not only accomplish powerful performance across various medical vision-language tasks in bilingual settings, but also recognize and detect structures in multimodal medical scans, boosting the interpretability and user interactivity of medical MLLMs. The codes and model will be made publicly available. | Multimodal Large Language Model, Biomedicine, Region-Text | We present an interpretable, region-aware bilingual multimodal large language model for biomedicine with a large-scale region-centric dataset. | 132 | 2410.18387 | [
0.06939306855201721,
-0.043411895632743835,
0.008395426906645298,
-0.07220842689275742,
0.03559399023652077,
-0.026054512709379196,
0.004969924688339233,
-0.03986124321818352,
0.016549265012145042,
-0.08188391476869583,
-0.035675786435604095,
-0.09284631162881851,
-0.009460448287427425,
0.09348131716251373,
-0.031220799311995506,
-0.022164911031723022,
-0.013512127101421356,
0.03205607831478119,
-0.09684520959854126,
0.047421883791685104,
0.05060785263776779,
0.06234379857778549,
0.07769018411636353,
0.00552487513050437,
0.02164464071393013,
-0.004312820732593536,
0.017686311155557632,
-0.0850134864449501,
0.009298902004957199,
-0.007716653868556023,
0.0452193059027195,
0.003233686788007617,
-0.009284306317567825,
0.08464649319648743,
-0.019625576213002205,
0.08838985115289688,
-0.07237625122070312,
0.050388868898153305,
-0.036043595522642136,
0.007402624469250441,
-0.03692963719367981,
0.061083052307367325,
0.03390415757894516,
0.05415712296962738,
0.14739485085010529,
0.02442668378353119,
-0.05179239809513092,
-0.029137935489416122,
0.06962445378303528,
0.005494109354913235,
-0.12230975925922394,
-0.07476722449064255,
0.0037090477999299765,
0.10679935663938522,
-0.029996782541275024,
0.0036492934450507164,
0.061206791549921036,
-0.06957145035266876,
-0.011872806586325169,
-0.03317246958613396,
-0.06962130963802338,
-0.011771749705076218,
0.016230236738920212,
-0.018994469195604324,
0.004839892033487558,
-0.05529540032148361,
-0.010670818388462067,
-0.04585014656186104,
-0.06953389942646027,
0.03373447433114052,
-0.017590412870049477,
0.023056259378790855,
0.037363145500421524,
0.05729833245277405,
-0.046510785818099976,
0.04831588640809059,
0.008357653394341469,
-0.03949621319770813,
0.03577627241611481,
-0.1150120347738266,
0.0538797453045845,
0.04018140584230423,
0.06460857391357422,
-0.03722860664129257,
0.09199125319719315,
-0.023246852681040764,
-0.005806155037134886,
0.015319091267883778,
0.007862591184675694,
0.008898625150322914,
-0.019016548991203308,
0.0018587958766147494,
0.04258958250284195,
-0.02302468754351139,
0.023928998038172722,
0.008804824203252792,
0.06872323155403137,
-0.06490769982337952,
0.030046705156564713,
0.06668511033058167,
0.03755686432123184,
-0.007906204089522362,
0.05168456584215164,
-0.03007330186665058,
-0.030170351266860962,
-0.04348669573664665,
0.04014381393790245,
0.005727273877710104,
0.06859132647514343,
-0.12442592531442642,
0.04453277960419655,
0.04927652329206467,
-0.030822014436125755,
0.015885764732956886,
0.008243315853178501,
-0.015441255643963814,
0.037416599690914154,
-0.028963681310415268,
0.055729616433382034,
0.06257714331150055,
-0.09642016887664795,
-0.0034737831447273493,
-0.04478209465742111,
-0.02771572582423687,
0.05155036225914955,
0.007376571651548147,
-0.03549819439649582,
8.451100922511677e-34,
0.09119795262813568,
-0.002751715015619993,
0.058208659291267395,
0.07271162420511246,
0.02284615859389305,
-0.0010805983329191804,
-0.05947162210941315,
-0.0187922865152359,
-0.04871474578976631,
-0.046312279999256134,
-0.06609408557415009,
0.059474799782037735,
-0.03587251529097557,
0.1316688507795334,
0.006645749788731337,
0.017771467566490173,
-0.017231227830052376,
0.028194891288876534,
-0.02364397794008255,
0.025767985731363297,
0.036420788615942,
-0.010772865265607834,
-0.014520530588924885,
0.013003967702388763,
-0.025045285001397133,
-0.013255941681563854,
0.05875613912940025,
-0.09349137544631958,
-0.03100917674601078,
0.034747008234262466,
-0.20138351619243622,
0.04224400222301483,
0.03518441691994667,
0.04003685340285301,
0.025462402030825615,
-0.03459113463759422,
-0.0023734713904559612,
-0.03806634992361069,
0.011955269612371922,
0.05479050055146217,
0.052104443311691284,
0.012407572939991951,
0.003197432728484273,
0.0012786701554432511,
-0.011333885602653027,
0.0024075645487755537,
-0.015018422156572342,
0.07281583547592163,
-0.0698421522974968,
-0.028545208275318146,
0.004739316646009684,
-0.008054950274527073,
0.009242416359484196,
-0.11329183727502823,
0.0350487194955349,
0.06805875897407532,
0.008350005373358727,
0.04005127027630806,
0.07593145221471786,
-0.06574973464012146,
0.05071160942316055,
0.06292364746332169,
0.03315608948469162,
0.10383523255586624,
0.003939381800591946,
-0.031210778281092644,
-0.0033927804324775934,
0.015490884892642498,
0.0445357970893383,
-0.011690243147313595,
-0.11361689865589142,
0.041645873337984085,
0.02968095801770687,
0.0022324316669255495,
-0.010460085235536098,
0.024892818182706833,
0.007356002926826477,
-0.07102293521165848,
0.021282631903886795,
0.037537917494773865,
-0.027441207319498062,
-0.034822821617126465,
-0.03808462619781494,
-0.008336267434060574,
0.02120661921799183,
-0.0019563171081244946,
0.05918635427951813,
-0.05234091728925705,
-0.04178333282470703,
-0.03949686139822006,
-0.07181128114461899,
0.02317851595580578,
-0.01212243176996708,
-0.069104865193367,
0.006199454423040152,
-1.5284579273272841e-33,
0.048860061913728714,
-0.013119987212121487,
-0.016178544610738754,
0.012446465902030468,
-0.010298652574419975,
-0.07512209564447403,
-0.0023376585450023413,
0.04575682058930397,
-0.027234725654125214,
-0.06262408196926117,
-0.005777968559414148,
-0.0810670331120491,
0.03268733248114586,
-0.031997933983802795,
-0.0406937301158905,
-0.008813715539872646,
-0.0038217490073293447,
0.07751525938510895,
-0.014519141055643559,
0.06977133452892303,
-0.0001084217510651797,
0.06783781945705414,
-0.06677798926830292,
0.052756208926439285,
-0.09033305943012238,
0.048066869378089905,
-0.047165412455797195,
0.030193688347935677,
-0.0817437395453453,
-0.059152957051992416,
-0.049957726150751114,
-0.017321331426501274,
0.03391565382480621,
-0.055912356823682785,
-0.09790961444377899,
0.0509912371635437,
0.0016949052223935723,
-0.10648711770772934,
-0.011279364116489887,
0.08634092658758163,
0.07697688043117523,
-0.041598137468099594,
-0.05910179018974304,
-0.0014441931853070855,
-0.016167595982551575,
-0.11756142973899841,
-0.05713152512907982,
-0.02312394790351391,
-0.034355439245700836,
0.03677469119429588,
-0.09614808112382889,
-0.014855555258691311,
-0.09967007488012314,
0.022683260962367058,
-0.020019222050905228,
-0.09932778775691986,
-0.06547492742538452,
-0.07873464375734329,
-0.017962301149964333,
-0.05872640758752823,
-0.019165081903338432,
-0.07801754027605057,
-0.019027767702937126,
-0.012948359362781048,
0.020613493397831917,
-0.001543241087347269,
0.056241318583488464,
-0.02192899025976658,
0.07279519736766815,
0.016160961240530014,
0.038205116987228394,
-0.030998481437563896,
0.0040384503081440926,
0.016465654596686363,
-0.008795736357569695,
-0.0014384608948603272,
0.018031762912869453,
-0.04836052283644676,
0.007687854114919901,
-0.059055596590042114,
0.07547688484191895,
-0.11043079197406769,
0.04575483873486519,
0.063261479139328,
0.07507241517305374,
0.05899443477392197,
-0.012089776806533337,
-0.017132343724370003,
0.045372772961854935,
-0.00023763546778354794,
-0.04217671602964401,
0.06721001863479614,
-0.025674019008874893,
0.024233859032392502,
-0.016566768288612366,
-4.118789931339961e-8,
-0.07156148552894592,
0.05756527557969093,
-0.02323463186621666,
0.007108546327799559,
-0.009449330158531666,
-0.0850682184100151,
-0.08862584829330444,
0.07337164133787155,
-0.008375728502869606,
0.010140875354409218,
-0.00041173346107825637,
-0.02633267268538475,
-0.015275055542588234,
-0.1203363910317421,
0.04047268256545067,
0.09314388781785965,
0.006636491511017084,
0.09586159139871597,
-0.010380735620856285,
0.03070138767361641,
0.02657230757176876,
0.005810727830976248,
0.05726249888539314,
-0.02562432363629341,
0.044042736291885376,
-0.03458819538354874,
-0.05555994063615799,
0.0514533706009388,
0.009191506542265415,
-0.034333162009716034,
0.006551464553922415,
0.00019504512601997703,
-0.03401489555835724,
0.044301602989435196,
-0.0021598045714199543,
-0.016671113669872284,
0.010900121182203293,
-0.03133292496204376,
0.04303765669465065,
-0.060648392885923386,
0.12674455344676971,
0.035587556660175323,
-0.023472722619771957,
0.03811793401837349,
0.02981732226908207,
-0.0387895330786705,
0.007271105889230967,
-0.03016088902950287,
0.02150331437587738,
-0.009853105060756207,
0.007157726678997278,
-0.012876010499894619,
0.05490540340542793,
0.13499897718429565,
0.06299862265586853,
0.03754105418920517,
0.07175363600254059,
0.014969242736697197,
0.04571477696299553,
0.12001130729913712,
0.022160604596138,
0.06153729185461998,
-0.09585636109113693,
-0.061108823865652084
] | 0 | 0 | 0 | 0 |
|
MindSearch: Mimicking Human Minds Elicits Deep AI Searcher | https://openreview.net/forum?id=xgtXkyqw1f | [
"Zehui Chen",
"Kuikun Liu",
"Qiuchen Wang",
"Jiangning Liu",
"Wenwei Zhang",
"Kai Chen",
"Feng Zhao"
] | Poster | Information seeking and integration is a complex cognitive task that consumes enormous time and effort. Inspired by the remarkable progress of Large Language Models, recent works attempt to solve this task by combining LLMs and search engines. However, these methods still obtain unsatisfying performance due to three challenges: (1) complex requests often cannot be accurately and completely retrieved by the search engine once (2) corresponding information to be integrated is spread over multiple web pages along with massive noise, and (3) a large number of web pages with long contents may quickly exceed the maximum context length of LLMs. Inspired by the cognitive process when humans solve these problems, we introduce MindSearch to mimic the human minds in web information seeking and integration, which can be instantiated by a simple yet effective LLM-based multi-agent framework. The WebPlanner models the human mind of multi-step information seeking as a dynamic graph construction process: it decomposes the user query into atomic sub-questions as nodes in the graph and progressively extends the graph based on the search result from WebSearcher. Tasked with each sub-question, WebSearcher performs hierarchical information retrieval with search engines and collects valuable information for WebPlanner. The multi-agent design of MindSearch enables the whole framework to seek and integrate information parallelly from larger-scale (e.g., more than 300) web pages in 3 minutes, which is worth 3 hours of human effort. MindSearch demonstrates significant improvement in the response quality in terms of depth and breadth, on both close-set and open-set QA problems. Besides, responses from MindSearch based on InternLM2.5-7B are preferable by humans to ChatGPT-Web and Perplexity.ai applications, which implies that MindSearch can already deliver a competitive solution to the proprietary AI search engine. | language model, search engine, multi-agent system | null | 129 | 2407.20183 | [
-0.06609879434108734,
-0.04005308449268341,
0.010973416268825531,
-0.013964030891656876,
0.022728271782398224,
-0.061854053288698196,
-0.017479728907346725,
0.02095063589513302,
0.0032008904963731766,
-0.08436498045921326,
-0.014848309569060802,
-0.0811857134103775,
0.09194926172494888,
0.02810635417699814,
0.028384285047650337,
0.017634157091379166,
0.07599879801273346,
0.035520583391189575,
-0.05341225117444992,
-0.042319174855947495,
0.05876829847693443,
0.0357222743332386,
0.057986289262771606,
-0.0746385008096695,
-0.055701401084661484,
0.003115232102572918,
-0.06480458378791809,
-0.10507384687662125,
0.044630859047174454,
0.02459157072007656,
0.10086974501609802,
0.001484955777414143,
-0.024849100038409233,
0.1398826241493225,
-0.04723887890577316,
-0.03454355522990227,
-0.09335588663816452,
0.04070521891117096,
0.02164716273546219,
-0.03138434886932373,
-0.03353177011013031,
-0.05141971632838249,
-0.03394908830523491,
0.02550794556736946,
0.03002292849123478,
0.0562858022749424,
-0.09242215752601624,
-0.0370006337761879,
0.06265021115541458,
0.012280265800654888,
-0.16402164101600647,
-0.042010996490716934,
0.01572629250586033,
0.017275499179959297,
0.01302400790154934,
-0.012374714016914368,
0.006083663087338209,
0.020282041281461716,
-0.04163799062371254,
-0.04070374369621277,
0.004892340861260891,
-0.04933961480855942,
0.06717909872531891,
-0.0018276561750099063,
0.01495268102735281,
-0.01145923975855112,
0.004557822365313768,
-0.003900606418028474,
0.04879467934370041,
-0.026940524578094482,
0.03801429644227028,
0.0191055815666914,
-0.06480338424444199,
0.04821142181754112,
-0.01763179711997509,
-0.023170465603470802,
-0.047802750021219254,
-0.03766744211316109,
0.0167778879404068,
-0.014428666792809963,
-0.05335644632577896,
-0.008679913356900215,
0.016467316076159477,
0.04229765757918358,
0.1083349883556366,
-0.028098585084080696,
0.027407566085457802,
-0.025951005518436432,
0.0341411717236042,
0.013370601460337639,
-0.06879273802042007,
-0.13080190122127533,
0.007311749272048473,
-0.022429293021559715,
0.09807826578617096,
0.06772492825984955,
-0.03313606604933739,
-0.03764164820313454,
-0.03392220288515091,
0.015360480174422264,
0.005059312097728252,
0.04556114971637726,
0.029825478792190552,
-0.09827319532632828,
-0.09722214192152023,
0.03293761983513832,
0.033386945724487305,
0.12713660299777985,
0.01869223453104496,
-0.044433463364839554,
-0.05831199511885643,
0.040433187037706375,
0.0786106288433075,
-0.01801482029259205,
0.04993625357747078,
-0.04677613824605942,
0.08388769626617432,
0.05396999791264534,
0.07610471546649933,
0.05785926803946495,
0.007680633570998907,
0.10706435889005661,
0.004111480433493853,
0.005194791592657566,
0.0827396959066391,
-0.03624509647488594,
-0.0327911414206028,
2.4169025269860193e-33,
0.08530577272176743,
0.034774184226989746,
0.03700379282236099,
0.05990985408425331,
0.028727807104587555,
0.01344528328627348,
0.032911114394664764,
0.11497385054826736,
-0.06335647404193878,
0.029565241187810898,
-0.050806429237127304,
0.07409592717885971,
-0.032599352300167084,
0.07152791321277618,
-0.017605887725949287,
-0.0328947938978672,
-0.03396199271082878,
0.01690267026424408,
-0.01579679176211357,
-0.06900420784950256,
0.04564420133829117,
0.013718998059630394,
0.1043965220451355,
-0.04097357392311096,
0.07614847272634506,
0.023109164088964462,
-0.010196270421147346,
-0.06896550208330154,
0.056341663002967834,
-0.010683882050216198,
-0.11155867576599121,
-0.004599167034029961,
-0.08322428166866302,
-0.0027657151222229004,
-0.016738582402467728,
-0.025062983855605125,
-0.031721893697977066,
-0.0013368724612519145,
0.005967112258076668,
-0.029370451346039772,
-0.03530376777052879,
0.009589528664946556,
0.05741078406572342,
-0.029070712625980377,
-0.09443732351064682,
-0.04344426468014717,
-0.07277694344520569,
0.037088096141815186,
0.033258579671382904,
-0.03464757651090622,
0.060504425317049026,
-0.00952917244285345,
-0.025319095700979233,
0.038675710558891296,
0.05003933236002922,
0.005510649643838406,
0.01110601332038641,
0.04367682710289955,
-0.015201704576611519,
0.038069069385528564,
0.0037341571878641844,
-0.026876268908381462,
0.009515221230685711,
0.08967414498329163,
0.08073064684867859,
-0.014419554732739925,
-0.010660805739462376,
0.08448362350463867,
0.0553782619535923,
0.022692125290632248,
0.025020647794008255,
0.0042413887567818165,
0.04660903289914131,
-0.06473541259765625,
-0.013501228764653206,
-0.04128003492951393,
-0.0001334048865828663,
-0.08646070957183838,
0.01907447725534439,
-0.02043125405907631,
0.03809084743261337,
-0.07975433021783829,
0.010295751504600048,
-0.02277093008160591,
0.0173439122736454,
-0.008270594291388988,
0.021587833762168884,
-0.09206787496805191,
0.02786475606262684,
-0.04170148819684982,
-0.01788632944226265,
-0.002657948760315776,
-0.03257559984922409,
-0.020747199654579163,
-0.013741377741098404,
-1.9003574321270816e-33,
-0.009699435904622078,
-0.10329126566648483,
-0.018233738839626312,
0.05470624193549156,
0.03787509351968765,
-0.048441510647535324,
-0.03914690017700195,
-0.051548149436712265,
-0.038248855620622635,
0.02739284187555313,
-0.06703527271747589,
-0.05480988323688507,
0.05169733241200447,
-0.046792108565568924,
-0.04544799402356148,
0.04030943661928177,
0.07759096473455429,
0.0030552062671631575,
-0.007052871864289045,
0.024741746485233307,
-0.024973547086119652,
0.08389303833246231,
-0.1343873143196106,
-0.024834316223859787,
0.005397457629442215,
0.02038511447608471,
0.019141340628266335,
0.05719931796193123,
-0.021273471415042877,
0.0366300530731678,
0.00346551276743412,
-0.04404391348361969,
-0.04608374834060669,
-0.044298745691776276,
-0.0478675402700901,
0.11493248492479324,
0.000997701776213944,
0.0018831577617675066,
-0.055284999310970306,
0.060400739312171936,
0.02047642692923546,
-0.013328070752322674,
-0.0036899535916745663,
-0.010782630182802677,
-0.039592668414115906,
0.02493133209645748,
-0.1510167419910431,
0.07376382499933243,
-0.038474954664707184,
0.0412069708108902,
-0.00493094464763999,
-0.028166957199573517,
-0.040674906224012375,
-0.06638098508119583,
-0.08497112989425659,
-0.015690401196479797,
-0.07153838872909546,
0.04475576430559158,
-0.04006902873516083,
-0.01840585097670555,
-0.047160983085632324,
0.019799841567873955,
0.015874087810516357,
-0.006165653001517057,
-0.0012010206701233983,
-0.04522309824824333,
-0.005865566898137331,
0.03622833639383316,
-0.07615707814693451,
-0.05736057460308075,
0.05619901418685913,
-0.07210667431354523,
0.02111946977674961,
-0.00032521371031180024,
0.08298631757497787,
0.041800983250141144,
0.0600597970187664,
-0.00428652111440897,
0.016026146709918976,
-0.06358293443918228,
-0.056820958852767944,
0.02206357941031456,
0.013197015970945358,
0.035623498260974884,
0.006659312639385462,
0.010541420429944992,
-0.03429439291357994,
0.14365656673908234,
0.0292632095515728,
-0.03528950363397598,
0.02449413388967514,
0.036374010145664215,
-0.02127605862915516,
0.06819238513708115,
-0.04215012490749359,
-3.9943294893873826e-8,
-0.06919451057910919,
0.04943561553955078,
-0.03994285315275192,
0.047668032348155975,
0.06800497323274612,
0.0090462826192379,
0.012614666484296322,
0.10449713468551636,
-0.0587504543364048,
-0.009929860010743141,
0.10790612548589706,
-0.039699140936136246,
0.001299519557505846,
0.015661943703889847,
0.03179154917597771,
0.009655379690229893,
-0.011410052888095379,
-0.051271162927150726,
0.012411016039550304,
-0.08643623441457748,
0.05786505714058876,
0.02060239017009735,
-0.06508799642324448,
0.07500769942998886,
0.03834790363907814,
0.03775724768638611,
-0.05973893031477928,
0.05302351713180542,
-0.05811610072851181,
0.01764495112001896,
-0.024035103619098663,
0.039950545877218246,
-0.04654516279697418,
0.03272898867726326,
0.05573791638016701,
0.03008812852203846,
-0.01516942959278822,
-0.04300222173333168,
-0.08808261901140213,
-0.047704264521598816,
0.07079395651817322,
0.11241637915372849,
0.04596918448805809,
0.03335252404212952,
0.0652105063199997,
-0.01624712534248829,
-0.037603847682476044,
-0.06571079790592194,
0.09906505793333054,
-0.01716500148177147,
-0.00006039752042852342,
-0.032353900372982025,
0.019482076168060303,
0.053755875676870346,
0.06887336075305939,
-0.016096320003271103,
-0.02764616347849369,
-0.01369418390095234,
0.08518008142709732,
0.07685451209545135,
0.07906875014305115,
-0.021536003798246384,
-0.04112986847758293,
0.027344107627868652
] | https://github.com/internlm/mindsearch | 6,326 | 0 | 0 | 0 |
PseDet: Revisiting the Power of Pseudo Label in Incremental Object Detection | https://openreview.net/forum?id=Iu8FVcUmVp | [
"Qiuchen Wang",
"Zehui Chen",
"Chenhongyi Yang",
"Jiaming Liu",
"Zhenyu Li",
"Feng Zhao"
] | Poster | Incremental Objection Detection (IOD) facilitates the expansion of the usage scope of object detectors without forgetting previously acquired knowledge. Current approaches mostly adopt response-level knowledge distillation to overcome forgetting issues, by conducting implicit memory replay from the teacher model on new training data. However, this indirect learning paradigm does not fully leverage the knowledge generated by the teacher model. In this paper, we dive deeper into the mechanism of pseudo-labeling in incremental object detection by investigating three critical problems: (a) the upper bound quality of the pseudo labels is greatly limited by the previous model, (b) fixed score thresholds for label filtering, without considering the distribution across categories, and (c) the confidence score generated by the model does not well reflect the quality of the localization. Based on these observations, we propose a simple yet effective pseudo-labeling continual object detection framework, namely PseDet. Specifically, we introduce the spatio-temporal enhancement module to alleviate the negative effects when learning noisy data from the previous model. Considering the score distribution divergence across different classes, we propose the Categorical Adaptive Label Selector with a simple mathematical prior and fast K-Means pre-computation to dynamically determine the class-wise filtering threshold. In order to align the label score with the localization quality of the pseudo labels, we project the score through non-linear mapping to calibrate the distribution and integrate it into the new-step supervision. Extensive experiments on the competitive COCO benchmarks demonstrate the effectiveness and generalization of PseDet. Notably, it achieves 43.5+/41.2+ mAP under the 1/4-step incremental settings, achieving new state-of-the-art performance. | Incremental Object Detection, Catastrophic Forgetting, Pseudo Labeling | null | 126 | null | [
-0.0023209406062960625,
-0.041660141199827194,
-0.030624790117144585,
0.07617364823818207,
0.08156781643629074,
0.06343299895524979,
0.06400272995233536,
-0.041187018156051636,
0.061444710940122604,
-0.0791003555059433,
0.00728339608758688,
0.012607039883732796,
-0.00464525306597352,
0.05025259777903557,
-0.04741019755601883,
-0.021323150023818016,
0.1251630038022995,
0.09182470291852951,
-0.022163476794958115,
-0.01836962066590786,
-0.011170324869453907,
0.021241547539830208,
0.05328097194433212,
0.05334647744894028,
-0.007071354892104864,
0.03862515091896057,
-0.051066718995571136,
-0.04888681694865227,
0.05828366428613663,
-0.02563820406794548,
-0.0238136388361454,
0.04351849853992462,
-0.004246291238814592,
0.08444675803184509,
-0.07527164369821548,
-0.051164411008358,
-0.021073196083307266,
0.10314790159463882,
0.03459155187010765,
0.022588029503822327,
0.025600934401154518,
0.025903116911649704,
-0.009673701599240303,
-0.03816460072994232,
-0.016750458627939224,
0.027781501412391663,
0.0061728740110993385,
-0.06269636005163193,
0.03400108590722084,
0.018766023218631744,
0.0007575093186460435,
-0.021656449884176254,
-0.007763493340462446,
0.020584799349308014,
-0.024904608726501465,
0.02851187065243721,
0.04087309166789055,
0.009749144315719604,
-0.013075518421828747,
0.07579098641872406,
-0.05565066263079643,
-0.11188247054815292,
0.00045943379518575966,
-0.007845074869692326,
-0.0688524842262268,
0.03515223413705826,
0.014505323022603989,
0.002106758998706937,
0.04218368977308273,
0.09787939488887787,
-0.004500032402575016,
0.16481758654117584,
0.046602845191955566,
0.051572736352682114,
0.017173321917653084,
-0.018166586756706238,
-0.043562743812799454,
-0.04635291546583176,
0.07633090764284134,
-0.0677557960152626,
-0.01194482110440731,
-0.021803855895996094,
0.05807820335030556,
-0.06922577321529388,
0.12586282193660736,
-0.009200983680784702,
0.0770053043961525,
0.019420698285102844,
-0.03237153962254524,
-0.011153293773531914,
-0.04220994561910629,
-0.04893629997968674,
0.01321995910257101,
0.016684262081980705,
0.00864106323570013,
-0.038180142641067505,
-0.010484644211828709,
-0.023542677983641624,
0.01684405654668808,
0.04288235306739807,
-0.028607208281755447,
0.04057058319449425,
-0.0630960613489151,
-0.04151064157485962,
0.010418986901640892,
-0.060914039611816406,
0.0010368401417508721,
0.054311469197273254,
0.07126113027334213,
0.0036143388133496046,
0.006414735224097967,
0.012965304777026176,
-0.01705818623304367,
-0.01526479609310627,
-0.004571176134049892,
0.009238348342478275,
0.07644718885421753,
-0.04560422524809837,
-0.024475689977407455,
-0.0430435948073864,
0.057287026196718216,
-0.002618728671222925,
-0.029015760868787766,
0.01058194413781166,
0.0022594572510570288,
-0.03536726534366608,
-0.042903605848550797,
2.825191938973917e-33,
0.024863574653863907,
0.011987943202257156,
0.02263493463397026,
-0.049723461270332336,
-0.006109078414738178,
-0.007916291244328022,
-0.023434199392795563,
0.04963015392422676,
0.04354391619563103,
-0.0036533742677420378,
0.021733559668064117,
0.05787859857082367,
-0.057343918830156326,
0.02150045521557331,
0.06000232324004173,
-0.05227336287498474,
-0.003467310918495059,
0.04375721514225006,
-0.045838646590709686,
0.012946642935276031,
0.00563031155616045,
-0.06270485371351242,
0.00597417214885354,
-0.04980596899986267,
0.07636676728725433,
0.04957938566803932,
0.02831992693245411,
0.048963941633701324,
0.04031597450375557,
0.015462704002857208,
-0.049466993659734726,
-0.025711150839924812,
-0.005840080790221691,
0.04095645621418953,
0.022521624341607094,
-0.11335282772779465,
0.017433393746614456,
0.008767603896558285,
0.018054239451885223,
-0.06979218870401382,
0.007884393446147442,
0.036630965769290924,
-0.053900014609098434,
-0.06113731116056442,
-0.03633236885070801,
-0.08893295377492905,
0.05349560081958771,
0.05960813909769058,
-0.0885591134428978,
0.04694008827209473,
0.014920942485332489,
0.018495453521609306,
-0.06608428806066513,
-0.010509773157536983,
-0.13025718927383423,
0.01241656206548214,
-0.02151433750987053,
0.02300366945564747,
-0.006858004257082939,
-0.03312801569700241,
0.059987716376781464,
0.07558993995189667,
0.0073282960802316666,
0.11651673913002014,
0.048189494758844376,
0.0572957843542099,
-0.017214760184288025,
0.05235375836491585,
0.06100776419043541,
-0.02006733976304531,
0.0017733066342771053,
0.03761495649814606,
-0.07042769342660904,
-0.01945357769727707,
0.03198621794581413,
-0.047209158539772034,
0.00886981375515461,
-0.047545526176691055,
0.010597296059131622,
-0.011245990172028542,
0.0025784948375076056,
-0.1249418705701828,
-0.01398183312267065,
-0.04752657562494278,
-0.04609757289290428,
0.020092876628041267,
0.09177762269973755,
-0.10875603556632996,
-0.027081971988081932,
0.02369476854801178,
-0.04916132614016533,
-0.029163328930735588,
-0.07374177873134613,
0.04137249290943146,
-0.005698623135685921,
-2.070709155093286e-33,
0.07959181070327759,
-0.029451515525579453,
-0.02403535693883896,
0.05065353587269783,
-0.021334078162908554,
-0.03986182063817978,
-0.0036726784892380238,
0.1041622906923294,
0.007380343973636627,
-0.15767845511436462,
-0.04811892658472061,
-0.1013910248875618,
-0.030463742092251778,
0.04324641823768616,
0.0035170281771570444,
0.005158310290426016,
-0.08685795217752457,
0.055269692093133926,
-0.046196479350328445,
-0.006950322072952986,
0.007567476946860552,
0.06869398057460785,
-0.04148620739579201,
-0.012760070152580738,
-0.07219044119119644,
0.05050414428114891,
-0.023104164749383926,
0.062394991517066956,
0.0288394745439291,
-0.05594100058078766,
-0.016031840816140175,
-0.010045926086604595,
0.005864901468157768,
-0.07671885192394257,
-0.07607923448085785,
0.02202954888343811,
0.10028047859668732,
-0.0450490340590477,
-0.03450838476419449,
0.08952300995588303,
0.09596633911132812,
-0.02729775756597519,
-0.06931692361831665,
0.010367744602262974,
-0.026836175471544266,
-0.10536917299032211,
-0.028124447911977768,
0.07272998243570328,
-0.0046445089392364025,
0.01525272335857153,
-0.07986833900213242,
-0.013366155326366425,
-0.10379605740308762,
0.043325502425432205,
-0.07110858708620071,
-0.07897157222032547,
-0.06672070920467377,
-0.020025525242090225,
0.04584716260433197,
0.016471631824970245,
-0.030268428847193718,
-0.036966580897569656,
-0.02453990839421749,
-0.025122707709670067,
0.030361687764525414,
-0.0027634145226329565,
-0.01737898215651512,
0.009789952076971531,
0.0077787768095731735,
0.04703701287508011,
0.047490231692790985,
0.08969263732433319,
-0.10858003050088882,
-0.0032779222819954157,
-0.03589337319135666,
0.04262175038456917,
-0.02133111283183098,
-0.011477584019303322,
0.004697870463132858,
-0.00775251816958189,
0.009949476458132267,
-0.07935724407434464,
-0.02642669342458248,
0.11399401724338531,
0.06068415567278862,
0.05241336300969124,
0.045365847647190094,
0.055751342326402664,
-0.0033567885402590036,
0.03382530063390732,
-0.045369502156972885,
0.011873863637447357,
0.016450632363557816,
0.06206488236784935,
-0.07096937298774719,
-4.7567738192810793e-8,
-0.07817557454109192,
0.030028775334358215,
-0.052900563925504684,
0.024307677522301674,
0.08241880685091019,
-0.014557581394910812,
-0.027435488998889923,
0.09526333957910538,
-0.07655981183052063,
-0.07819550484418869,
0.06507251411676407,
0.004726404789835215,
-0.012885736301541328,
-0.07427088916301727,
-0.033841874450445175,
0.07388428598642349,
0.07915716618299484,
0.04317416250705719,
0.02722565457224846,
0.05967142432928085,
0.04345046356320381,
-0.052426017820835114,
0.0562850646674633,
-0.044684071093797684,
-0.026487452909350395,
-0.003279738128185272,
0.08368286490440369,
0.10913488268852234,
0.053060855716466904,
-0.024087045341730118,
0.004223929718136787,
0.025090940296649933,
0.014230367727577686,
0.01709992066025734,
0.042970478534698486,
0.060879118740558624,
0.016623707488179207,
-0.06710093468427658,
-0.002105415565893054,
-0.04656904563307762,
0.004701555240899324,
-0.07571002840995789,
-0.03429507464170456,
0.029831474646925926,
-0.008141690865159035,
0.043102845549583435,
0.025723442435264587,
-0.14322592318058014,
0.02882504090666771,
-0.0013255755184218287,
-0.034521497786045074,
-0.0116927120834589,
-0.040291495621204376,
0.11324635148048401,
0.082097627222538,
0.014691454358398914,
0.04372921586036682,
-0.03219735622406006,
-0.028404762968420982,
0.13001517951488495,
0.03639144450426102,
0.03524884954094887,
-0.052067629992961884,
0.046051375567913055
] | 0 | 0 | 0 | 0 |
|
Weak-to-Strong Generalization Through the Data-Centric Lens | https://openreview.net/forum?id=uogG8BfLs2 | [
"Changho Shin",
"John Cooper",
"Frederic Sala"
] | Poster | The weak-to-strong generalization phenomenon is the driver for important machine learning applications including highly data-efficient learning and, most recently, performing superalignment. While decades of research have resulted in numerous algorithms that produce strong empirical performance, understanding what aspects of data enable weak-to-strong generalization has been understudied. We propose a simple data-centric mechanism that characterizes weak-to-strong generalization: the overlap density. Intuitively, generalization tracks the number of points that contain overlaps, i.e., both easy patterns (learnable by a weak model) and challenging patterns (only learnable by a stronger model), as with such points, weak predictions can be used to learn challenging patterns by stronger models. And, we provide a practical overlap detection algorithm to find overlap density from data. Finally, we provide an algorithm to learn, among multiple sources of data, which to query when seeking to maximize overlap density and thereby enhance weak-to-strong generalization. We provide a theoretical result showing that the generalization benefit is a function of the overlap density and a regret bound of our data selection algorithm. Empirically, we validate the mechanism and the overlap detection algorithm on a wide array of settings. | weak to strong generalization, data-centric AI | We characterize data property that induces weak-to-strong generalization. | 124 | 2412.03881 | [
-0.05792729929089546,
-0.095936119556427,
0.029224807396531105,
0.05124553665518761,
0.11024993658065796,
-0.024255333468317986,
0.035732660442590714,
-0.10953269153833389,
-0.04634508863091469,
-0.05589594319462776,
-0.04092669114470482,
-0.016153691336512566,
0.046837836503982544,
0.01530675869435072,
0.005707993637770414,
0.014036211185157299,
0.07123654335737228,
-0.002647599671036005,
-0.11822652816772461,
-0.03820312023162842,
0.02254997380077839,
0.034160152077674866,
0.003557349555194378,
0.03535420075058937,
0.057356733828783035,
0.030539704486727715,
-0.0023832309525460005,
0.02630370296537876,
0.0041624377481639385,
0.020662419497966766,
0.027499757707118988,
-0.02777627669274807,
-0.005397188477218151,
-0.006602813024073839,
-0.08122970908880234,
-0.02739912085235119,
-0.08075878769159317,
0.12128439545631409,
0.02960599958896637,
-0.004724038764834404,
0.029588237404823303,
0.021592186763882637,
0.05891238898038864,
-0.013608395121991634,
0.02704901061952114,
0.09652842581272125,
-0.041421931236982346,
-0.05607151985168457,
0.025026755407452583,
-0.027766220271587372,
-0.04948669672012329,
-0.06380557268857956,
-0.026655660942196846,
0.06478052586317062,
0.029504794627428055,
0.008998066186904907,
0.014185154810547829,
-0.00443872669711709,
-0.037602923810482025,
0.02530476823449135,
-0.035508666187524796,
-0.08943238854408264,
-0.06436172872781754,
-0.02762085385620594,
0.018427547067403793,
-0.024025535210967064,
0.03166656941175461,
0.09996704757213593,
-0.019440803676843643,
0.10292909294366837,
-0.010021687485277653,
0.16069965064525604,
-0.07242514193058014,
0.09060214459896088,
0.07155144214630127,
0.10720299929380417,
-0.0116850221529603,
-0.028638483956456184,
-0.016245273873209953,
-0.048349618911743164,
-0.05282195657491684,
0.02530616894364357,
0.04376565292477608,
0.03927549347281456,
0.10813187062740326,
-0.03582153096795082,
0.03623133525252342,
-0.014784511178731918,
0.021049071103334427,
-0.010776805691421032,
-0.03735790401697159,
-0.044068753719329834,
-0.01275718491524458,
-0.033863384276628494,
0.01887994259595871,
0.10317660868167877,
-0.016603339463472366,
-0.09448552876710892,
0.002830607583746314,
0.05876988545060158,
-0.03573103994131088,
0.05253945291042328,
0.04155731201171875,
-0.0006502760807052255,
0.0664689689874649,
-0.005125262774527073,
-0.048317499458789825,
0.01273177657276392,
0.031299933791160583,
-0.11475740373134613,
0.02010118030011654,
0.04489636421203613,
-0.009215899743139744,
-0.0007815314456820488,
-0.03269939124584198,
0.05399201810359955,
0.05437833443284035,
0.014389277435839176,
-0.009286540560424328,
0.03595539554953575,
0.005965039599686861,
-0.011297679506242275,
0.012398503720760345,
0.03084060549736023,
0.019974371418356895,
-0.025969678536057472,
-0.08612169325351715,
1.859253126784837e-34,
0.04462890699505806,
-0.06332957744598389,
0.05480458587408066,
0.04768885299563408,
0.06432333588600159,
-0.05347022786736488,
-0.10008829832077026,
-0.02101825550198555,
0.020775964483618736,
0.10751549154520035,
-0.026740316301584244,
0.07376378029584885,
-0.009814443066716194,
0.016248181462287903,
0.031097223982214928,
0.07281046360731125,
0.0043184226378798485,
0.054299622774124146,
-0.056755419820547104,
-0.06686904281377792,
0.05581275373697281,
-0.07839065790176392,
-0.014119457453489304,
-0.019386867061257362,
-0.012813001871109009,
0.015209942124783993,
0.01189698651432991,
0.00047623662976548076,
0.012317204847931862,
0.016446975991129875,
-0.0708521232008934,
0.02599998563528061,
0.09807699918746948,
0.09642531722784042,
0.03336549922823906,
0.02543729543685913,
-0.024462025612592697,
-0.0748702883720398,
0.021184442564845085,
0.08169655501842499,
-0.0001692707446636632,
0.014022193849086761,
0.033260028809309006,
-0.04065874591469765,
0.017704036086797714,
-0.030743088573217392,
-0.07299532741308212,
-0.018311792984604836,
-0.02513272315263748,
-0.011873683892190456,
0.06349055469036102,
-0.01487345527857542,
-0.032697293907403946,
-0.05742499232292175,
-0.07342007756233215,
0.03375576436519623,
0.0009712769533507526,
0.018749436363577843,
0.02428933046758175,
0.051188286393880844,
-0.05257188156247139,
0.03499292582273483,
0.03539936617016792,
0.07000521570444107,
-0.030829036608338356,
0.024893758818507195,
-0.028989141806960106,
0.061028946191072464,
0.05936544016003609,
-0.00831556599587202,
0.018375271931290627,
-0.006562208291143179,
-0.025125853717327118,
-0.07552874833345413,
0.06520210951566696,
0.026037070900201797,
0.06600220501422882,
-0.05561460182070732,
0.05816231667995453,
0.026381731033325195,
-0.01969756744801998,
-0.02471703104674816,
0.06571167707443237,
-0.07073462754487991,
-0.07126245647668839,
-0.020576288923621178,
0.003306998871266842,
-0.05642328038811684,
-0.06962504982948303,
-0.06800171732902527,
-0.12808583676815033,
0.003397085005417466,
-0.04752740263938904,
-0.09468462318181992,
-0.021533485502004623,
-2.009927464619135e-33,
0.0038109682500362396,
-0.029241884127259254,
0.02238457277417183,
0.01574135571718216,
0.012549425475299358,
-0.0827595666050911,
-0.05051899328827858,
-0.06792499870061874,
-0.05346855893731117,
-0.06228594854474068,
-0.018119513988494873,
0.02603270672261715,
0.10803774744272232,
-0.03598466143012047,
-0.03115103393793106,
0.015553797595202923,
0.02648952417075634,
0.05284016206860542,
-0.021282624453306198,
0.021864047273993492,
-0.0020171189680695534,
0.010201041586697102,
-0.0642145499587059,
0.0418546125292778,
-0.01786472275853157,
-0.018865840509533882,
-0.12251574546098709,
0.04870462417602539,
0.008876333013176918,
0.022377435117959976,
-0.1103500947356224,
-0.03437121585011482,
0.02822992019355297,
-0.11167288571596146,
-0.032877836376428604,
0.07891637831926346,
-0.028758730739355087,
0.026011448353528976,
-0.034646835178136826,
0.05362912639975548,
0.0011790001299232244,
0.005524429492652416,
-0.09123031795024872,
-0.008479084819555283,
0.024806132540106773,
-0.05892814323306084,
-0.043314315378665924,
0.0070136007852852345,
0.04660765826702118,
-0.041826698929071426,
-0.07459364831447601,
0.05765688791871071,
0.01739255152642727,
0.043919630348682404,
-0.020578792318701744,
-0.0010632288176566362,
0.03633676841855049,
0.0198811162263155,
0.0440104678273201,
0.022875022143125534,
-0.025821208953857422,
-0.01671980880200863,
-0.009340828284621239,
0.002203567186370492,
0.05415821447968483,
-0.023287348449230194,
-0.0073178852908313274,
-0.0414285883307457,
0.012633653357625008,
-0.010845107957720757,
0.012206445448100567,
0.016841288655996323,
-0.016575824469327927,
0.026211922988295555,
-0.09323769807815552,
0.06221906095743179,
0.09898161888122559,
-0.06769724935293198,
-0.02467811480164528,
0.04030679166316986,
0.032049115747213364,
-0.030967924743890762,
-0.00859676580876112,
0.0635843500494957,
0.06424257904291153,
0.08651603758335114,
-0.03381979465484619,
0.05303828418254852,
0.04113916680216789,
-0.08319088816642761,
-0.0735076367855072,
0.04279845207929611,
-0.1269608587026596,
-0.014314359053969383,
-0.050722770392894745,
-4.145285004142352e-8,
-0.03352409973740578,
0.02837621606886387,
-0.021551741287112236,
-0.04510147497057915,
0.10946393758058548,
0.0033069488126784563,
-0.06412773579359055,
0.142048180103302,
-0.04293941333889961,
-0.06479817628860474,
0.04613310843706131,
-0.0063083600252866745,
-0.03322267904877663,
-0.027551718056201935,
0.04685552418231964,
0.04527415335178375,
-0.026052795350551605,
0.027426229789853096,
-0.07074928283691406,
0.04563945531845093,
-0.03582712635397911,
-0.026354333385825157,
0.04957641288638115,
-0.04431721568107605,
0.06456883996725082,
-0.06683813780546188,
-0.060285378247499466,
0.054240886121988297,
0.03312534838914871,
-0.0273085106164217,
0.005064932629466057,
-0.034606195986270905,
0.03420525789260864,
0.0295934546738863,
0.09768635779619217,
0.15325593948364258,
0.006348200608044863,
0.028415841981768608,
-0.08875986188650131,
-0.06404206901788712,
0.005030243191868067,
0.03381343558430672,
0.037894587963819504,
0.03849852457642555,
0.0720272809267044,
0.03293384611606598,
0.0459076426923275,
-0.045029353350400925,
0.0052198925986886024,
0.00017572562501300126,
0.017033027485013008,
-0.02461150288581848,
0.08394129574298859,
0.03377868980169296,
-0.005861841142177582,
0.02145862765610218,
-0.06694880872964859,
-0.008520727977156639,
0.0027874631341546774,
0.037863072007894516,
0.028273863717913628,
-0.058347608894109726,
-0.07447528094053268,
0.06442756950855255
] | 0 | 0 | 0 | 0 |
|
Distilling Structural Representations into Protein Sequence Models | https://openreview.net/forum?id=KXrgDM3mVD | [
"Jeffrey Ouyang-Zhang",
"Chengyue Gong",
"Yue Zhao",
"Philipp Kraehenbuehl",
"Adam Klivans",
"Daniel Jesus Diaz"
] | Poster | Protein language (or sequence) models, like the popular ESM2, are now widely used tools for extracting evolution-based protein representations and have achieved significant success on core downstream biological tasks.
A major open problem is how to obtain representations that best capture both the sequence evolutionary history and the atomic structural properties of proteins in general.
We introduce **I**mplicit **S**equence **M**odel, a sequence-only input model with structurally-enriched representations that outperforms state-of-the-art sequence models on several well-studied benchmarks including mutation stability assessment and structure prediction.
Our key innovations are a microenvironment-based Autoencoder for generating structure tokens and a self-supervised training objective that distills these tokens into ESM2's pre-trained model.
Notably, we make ISM's structure-enriched weights easily accessible for any application using the ESM2 framework. | biology, proteins, sequence, structure, autoencoder, esm | we finetune esm to improve structural representations | 121 | null | [
-0.0795896127820015,
-0.0932774618268013,
0.022277016192674637,
0.002364162355661392,
0.06155841797590256,
0.03774118050932884,
-0.04764885827898979,
0.018878232687711716,
0.03791840001940727,
-0.1067722737789154,
-0.010317827574908733,
-0.12280962616205215,
-0.009999834932386875,
0.08719708770513535,
-0.030383579432964325,
0.015406785532832146,
-0.004103883169591427,
0.04248813912272453,
-0.03933758661150932,
-0.06450005620718002,
0.018832944333553314,
0.06059189885854721,
0.0017348480178043246,
0.005453344434499741,
0.01455579325556755,
0.025463348254561424,
0.05296933278441429,
0.01316265482455492,
0.04942941293120384,
-0.04217882454395294,
0.08851846307516098,
-0.06839016824960709,
0.0337650328874588,
0.01835048198699951,
0.01048933807760477,
0.046273987740278244,
-0.09168834984302521,
-0.0939566045999527,
-0.03445475175976753,
-0.1416965276002884,
0.06674637645483017,
0.03335171192884445,
0.053775716572999954,
0.019300295040011406,
0.049871429800987244,
-0.03773271664977074,
0.021834151819348335,
-0.04046805575489998,
-0.0016564883990213275,
0.031189551576972008,
-0.04110126197338104,
-0.023916075006127357,
-0.053293194621801376,
0.061115726828575134,
0.014901283197104931,
0.007178093772381544,
0.010104600340127945,
-0.009770067408680916,
-0.03282399848103523,
-0.052850499749183655,
0.059033628553152084,
-0.09190523624420166,
0.03207199275493622,
0.000627579225692898,
-0.0169376228004694,
0.01460506021976471,
0.04285982623696327,
0.13875208795070648,
0.059096306562423706,
0.001144655980169773,
-0.01975727081298828,
0.032241154462099075,
-0.020395442843437195,
0.0621507428586483,
0.04499226436018944,
0.1303291916847229,
-0.024480057880282402,
0.06722237914800644,
0.07478886097669601,
-0.04554428160190582,
-0.02422069013118744,
0.013427370227873325,
0.07490385323762894,
-0.017842324450612068,
0.07048558443784714,
0.03489431366324425,
0.03762667998671532,
0.07278109341859818,
0.0507938414812088,
0.04794443026185036,
-0.006651619449257851,
-0.0359128899872303,
0.08602451533079147,
-0.04698744788765907,
0.021362783387303352,
0.05406521260738373,
0.024875441566109657,
-0.04476487636566162,
-0.012031635269522667,
0.051858555525541306,
0.017986707389354706,
0.08111554384231567,
0.05783272534608841,
0.0038112413603812456,
0.06095433235168457,
-0.036197613924741745,
0.016806893050670624,
0.07492998242378235,
0.006800685077905655,
0.004560338798910379,
0.03281250596046448,
0.05299952253699303,
0.06792277842760086,
0.03074093721807003,
0.05198191478848457,
-0.005331198684871197,
0.008381026796996593,
-0.06442608684301376,
0.056645289063453674,
0.10198643058538437,
0.008459831587970257,
-0.03361489623785019,
-0.0806518942117691,
0.0675007626414299,
-0.03875621035695076,
-0.01406142208725214,
-0.1280345767736435,
1.1185888443603951e-35,
-0.02045230194926262,
0.05066109448671341,
0.02141188457608223,
0.030387308448553085,
0.003808459034189582,
-0.11143390834331512,
0.002270109485834837,
0.04282994940876961,
-0.0735604539513588,
-0.0441066175699234,
-0.042709898203611374,
-0.03956165537238121,
0.012173542752861977,
0.07369471341371536,
0.0199830811470747,
-0.07682851701974869,
-0.047691840678453445,
0.05338359251618385,
-0.00015339507081080228,
-0.015689373016357422,
-0.055754225701093674,
-0.007455313112586737,
-0.031515005975961685,
-0.05133642256259918,
0.0776403397321701,
0.01277050282806158,
0.06214110925793648,
-0.04017053544521332,
-0.07477828115224838,
0.00016534788301214576,
-0.0934884250164032,
-0.04782343655824661,
-0.025097884237766266,
0.04501647129654884,
0.03875517472624779,
-0.041753482073545456,
0.030271079391241074,
-0.04745907336473465,
-0.008929620496928692,
-0.05729071423411369,
0.03240194171667099,
0.04532269388437271,
-0.024300467222929,
0.011142618022859097,
-0.03537015989422798,
-0.12274474650621414,
-0.01860976219177246,
-0.03720603138208389,
-0.07848366349935532,
-0.12491153180599213,
0.037827786058187485,
0.006246274337172508,
0.008139691315591335,
-0.01342005841434002,
-0.038685284554958344,
0.03693859279155731,
-0.011448523961007595,
0.08383574336767197,
0.04030090197920799,
0.017677072435617447,
-0.004586735740303993,
0.04788966476917267,
0.021511683240532875,
0.04463203251361847,
0.05247822403907776,
-0.0001563225669087842,
-0.05317461118102074,
0.004836907144635916,
-0.00831264816224575,
0.11523830890655518,
0.03611793741583824,
-0.010993191972374916,
-0.022178271785378456,
-0.02168608084321022,
0.031015628948807716,
-0.004486788995563984,
0.0074940212070941925,
-0.09840662032365799,
-0.03238803520798683,
0.11276710778474808,
-0.006296793930232525,
0.06779021769762039,
-0.07268846780061722,
0.030083710327744484,
-0.059012819081544876,
-0.1332150250673294,
0.0784442275762558,
-0.08820705115795135,
-0.00875107292085886,
-0.01365527231246233,
0.039865802973508835,
-0.09567548334598541,
0.043313849717378616,
-0.013637497089803219,
-0.06158727779984474,
-1.06918502945682e-33,
-0.00422350550070405,
0.043874308466911316,
-0.01414340641349554,
0.0014636649284511805,
-0.05442295968532562,
-0.07790455222129822,
-0.0461956150829792,
0.08428938686847687,
-0.02910114824771881,
-0.025612086057662964,
0.032836053520441055,
-0.0048967767506837845,
0.0551832914352417,
-0.03853438422083855,
0.051890697330236435,
0.056109145283699036,
-0.052061744034290314,
-0.00926644541323185,
0.09249872714281082,
0.04458237811923027,
0.02723808027803898,
0.04052421450614929,
-0.05485184118151665,
0.05926084518432617,
-0.04950990527868271,
0.047283872961997986,
0.012171863578259945,
0.017627276480197906,
0.00899121817201376,
0.011544267646968365,
-0.1404070109128952,
0.0346582755446434,
-0.0222021471709013,
-0.03005516529083252,
-0.11023742705583572,
0.060049254447221756,
-0.014145568013191223,
-0.033293094485998154,
0.045998092740774155,
0.00697824452072382,
0.034825295209884644,
0.01331649161875248,
-0.08210238814353943,
0.05064740777015686,
-0.004524432122707367,
-0.006953899282962084,
-0.013527504168450832,
0.02182036265730858,
0.0439024493098259,
-0.027873720973730087,
-0.017491783946752548,
0.04032217711210251,
-0.08763818442821503,
-0.020587697625160217,
-0.06207456439733505,
-0.08983764797449112,
0.005964663811028004,
-0.013937328942120075,
-0.016449272632598877,
0.012475813739001751,
-0.07290739566087723,
-0.04279756918549538,
0.03957575187087059,
-0.0448547787964344,
-0.022374730557203293,
-0.09776449203491211,
-0.04064106568694115,
-0.021093936637043953,
-0.06189621239900589,
-0.0641210526227951,
0.06070490553975105,
-0.004747456405311823,
0.039397355169057846,
0.08806868642568588,
0.04787673428654671,
-0.10885006934404373,
-0.02085229568183422,
-0.013367650099098682,
-0.0016672994242981076,
-0.07345101982355118,
-0.04823799431324005,
0.06777054071426392,
0.021745754405856133,
0.07160153239965439,
0.07786578685045242,
0.023075120523571968,
0.04493086040019989,
0.06676875799894333,
0.004893663804978132,
0.029811933636665344,
-0.029766414314508438,
-0.005112233106046915,
-0.035674817860126495,
0.06715887039899826,
-0.01242064218968153,
-3.9861550504838306e-8,
-0.018967939540743828,
-0.012977144680917263,
0.06731472164392471,
0.03624589741230011,
-0.011752830818295479,
0.029161496087908745,
-0.03937341272830963,
0.06370772421360016,
0.032912272959947586,
-0.002392907626926899,
0.04607085511088371,
-0.016706185415387154,
-0.07959432154893875,
-0.03665570169687271,
-0.03814465180039406,
0.1323775202035904,
0.011583833023905754,
0.02369583398103714,
-0.03222363069653511,
-0.012565295211970806,
0.02682887390255928,
0.10188554972410202,
-0.0008233094704337418,
-0.005412871949374676,
0.05834698677062988,
-0.06586074829101562,
-0.013842238113284111,
0.03699444234371185,
0.06240850314497948,
-0.04380762577056885,
-0.08135015517473221,
0.03948047012090683,
-0.016436440870165825,
-0.013305900618433952,
-0.07672609388828278,
0.06524799019098282,
-0.02380281873047352,
-0.03611604496836662,
-0.025856446474790573,
0.01594829373061657,
0.0191172044724226,
0.093690887093544,
-0.10260363668203354,
-0.03873898461461067,
-0.014493398368358612,
0.01767723448574543,
-0.01184010598808527,
0.017987340688705444,
-0.03643233701586723,
-0.05531288683414459,
0.029406143352389336,
-0.050295520573854446,
-0.00972443725913763,
-0.019941426813602448,
-0.03267848491668701,
-0.005113608203828335,
0.00048333662562072277,
-0.03368633612990379,
0.08284205943346024,
0.012560678645968437,
0.006442584563046694,
-0.01178828440606594,
0.013474112376570702,
-0.023101279512047768
] | 0 | 0 | 0 | 0 |
|
Probabilistic Language-Image Pre-Training | https://openreview.net/forum?id=D5X6nPGFUY | [
"Sanghyuk Chun",
"Wonjae Kim",
"Song Park",
"Sangdoo Yun"
] | Poster | Vision-language models (VLMs) embed aligned image-text pairs into a joint space but often rely on deterministic embeddings, assuming a one-to-one correspondence between images and texts. This oversimplifies real-world relationships, which are inherently many-to-many, with multiple captions describing a single image and vice versa. We introduce Probabilistic Language-Image Pre-training (ProLIP), the first probabilistic VLM pre-trained on a billion-scale image-text dataset using only probabilistic objectives, achieving a strong zero-shot capability (e.g., 74.6% ImageNet zero-shot accuracy with ViT-B/16). ProLIP efficiently estimates uncertainty by an ``uncertainty token'' without extra parameters. We also introduce a novel inclusion loss that enforces distributional inclusion relationships between image-text pairs and between original and masked inputs. Experiments demonstrate that, by leveraging uncertainty estimates, ProLIP benefits downstream tasks and aligns with intuitive notions of uncertainty, e.g., shorter texts being more uncertain and more general inputs including specific ones. Utilizing text uncertainties, we further improve ImageNet accuracy from 74.6% to 75.8% (under a few-shot setting), supporting the practical advantages of our probabilistic approach. The code is available at https://github.com/naver-ai/prolip | vision-langauge-pretraining, vision-language-model, probabilistic-embeddings | We introduce Probabilistic Language-Image Pre-training (ProLIP), the first probabilistic VLM pre-trained on a billion-scale image-text dataset using only probabilistic objectives, achieving a strong zero-shot capability | 115 | 2410.18857 | [
0.025236861780285835,
-0.04066736251115799,
0.00769431097432971,
0.017998216673731804,
0.0936751514673233,
0.027371404692530632,
0.012817158363759518,
0.002370770089328289,
0.11094509065151215,
-0.04739588499069214,
0.037489522248506546,
-0.06550557166337967,
0.11894526332616806,
0.07116373628377914,
-0.03660202771425247,
-0.011545867659151554,
0.0793059915304184,
0.053158801048994064,
-0.1083592027425766,
-0.03743918240070343,
-0.016656529158353806,
0.0039032481145113707,
0.1039765253663063,
-0.021352574229240417,
0.06533476710319519,
-0.04343757778406143,
-0.04007336497306824,
0.01962985284626484,
0.007815990597009659,
-0.044437747448682785,
0.011441352777183056,
-0.025863170623779297,
-0.022807754576206207,
0.0368158221244812,
0.004684231244027615,
0.0871562510728836,
-0.016692468896508217,
0.03920835629105568,
0.05196963995695114,
-0.025986840948462486,
0.015586580149829388,
-0.002571233082562685,
0.04558219015598297,
0.00891059823334217,
0.08179344236850739,
0.0376710519194603,
0.00212112907320261,
-0.07086267322301865,
0.037515804171562195,
-0.016086390241980553,
-0.05429161339998245,
-0.031237127259373665,
0.007065792102366686,
-0.013278789818286896,
-0.06298443675041199,
0.047002118080854416,
0.01718747615814209,
-0.0242303479462862,
0.011860406957566738,
-0.03120991587638855,
-0.013398639857769012,
-0.09316807240247726,
-0.008070003241300583,
0.035102199763059616,
-0.00022194618941284716,
-0.02380204014480114,
0.028447357937693596,
0.05599255859851837,
-0.037098854780197144,
0.04524335265159607,
-0.005138577427715063,
0.08383390307426453,
-0.021213747560977936,
0.0914924368262291,
-0.04666954651474953,
0.06517927348613739,
0.056852374225854874,
-0.07074747234582901,
0.03634348511695862,
-0.01688995026051998,
0.0753181129693985,
-0.00010379683953942731,
0.08163803070783615,
0.015320167876780033,
0.10141703486442566,
0.02349981479346752,
-0.0023607746697962284,
0.035258375108242035,
-0.02322101965546608,
-0.025873253121972084,
-0.14254218339920044,
-0.11395584046840668,
0.07595989853143692,
0.03972006216645241,
0.004036589991301298,
0.05279672518372536,
-0.022033117711544037,
-0.0909770131111145,
-0.03626702353358269,
0.10435720533132553,
0.030719652771949768,
0.012775286100804806,
0.04749884828925133,
-0.02197122946381569,
0.04074607789516449,
-0.01362097542732954,
0.015559489838778973,
0.027070386335253716,
0.017406225204467773,
-0.08475666493177414,
0.06582767516374588,
-0.046668365597724915,
-0.0521276593208313,
0.019614826887845993,
-0.014339347369968891,
0.002789237769320607,
0.0010703494772315025,
-0.006178140640258789,
0.0099110696464777,
-0.031249141320586205,
-0.08580745756626129,
0.04014718905091286,
0.01588672772049904,
-0.032010793685913086,
-0.026728658005595207,
-0.05668484419584274,
0.045951761305332184,
-3.138953100303227e-34,
0.06776763498783112,
0.006812937557697296,
0.07451766729354858,
0.03958703577518463,
0.07093705981969833,
-0.02458268217742443,
-0.04504586011171341,
0.025955062359571457,
-0.06366097927093506,
-0.022006485611200333,
-0.008585580624639988,
-0.040041666477918625,
-0.009859586134552956,
0.07913610339164734,
0.0006417628028430045,
0.00867580994963646,
-0.024044543504714966,
0.003415787359699607,
0.03170905262231827,
0.04799869656562805,
-0.05813542753458023,
-0.03971460089087486,
-0.04643523320555687,
-0.01930118538439274,
0.006386867258697748,
0.024849796667695045,
0.058463890105485916,
-0.08519498258829117,
-0.056026890873909,
-0.009324390441179276,
-0.15841446816921234,
0.05789155513048172,
0.0748966708779335,
-0.024003634229302406,
0.08823230862617493,
-0.09597570449113846,
-0.04899002984166145,
-0.05645567923784256,
0.00036387366708368063,
-0.02909034676849842,
-0.029096737504005432,
0.07458700239658356,
-0.0033689311239868402,
-0.03365163877606392,
-0.05320427939295769,
-0.06139422953128815,
-0.01310588326305151,
0.077414371073246,
-0.062013834714889526,
-0.010500573553144932,
0.035283397883176804,
0.05774042010307312,
-0.06342577934265137,
0.024066532030701637,
-0.03182878717780113,
-0.0531533844769001,
0.02362876385450363,
-0.0023989654146134853,
0.012636815197765827,
-0.05350368469953537,
-0.00601219991222024,
0.009073624387383461,
-0.036296434700489044,
0.04451458901166916,
0.08792552351951599,
0.09908124804496765,
0.001542530837468803,
-0.0004142277757637203,
0.05969489738345146,
-0.010105262510478497,
-0.049498867243528366,
-0.04293106496334076,
-0.039217524230480194,
-0.08083730190992355,
0.05299586430191994,
0.08192786574363708,
0.03211984783411026,
-0.08812794089317322,
0.06252222508192062,
0.10171454399824142,
-0.01827600784599781,
-0.0519060418009758,
0.02834184467792511,
-0.0730600506067276,
-0.08341298252344131,
-0.06665817648172379,
0.04016026481986046,
-0.09192534536123276,
0.03283865004777908,
-0.06989025324583054,
-0.017485488206148148,
-0.03138184919953346,
-0.050846558064222336,
0.019310640171170235,
0.033856023102998734,
9.017151961955495e-34,
0.014956476166844368,
0.09713329374790192,
-0.005940000992268324,
0.015442821197211742,
-0.009725782088935375,
-0.03513139858841896,
0.07653308659791946,
0.02227490395307541,
0.05814732611179352,
-0.0679217129945755,
-0.07125601172447205,
-0.03386756032705307,
0.02063564397394657,
-0.06965401023626328,
-0.011254457756876945,
0.011853402480483055,
-0.011199005879461765,
0.038507670164108276,
-0.022905956953763962,
0.008549320511519909,
0.02851724438369274,
0.04780338332056999,
-0.01929762400686741,
0.06265146285295486,
0.04639579728245735,
0.1153726577758789,
0.01354797463864088,
0.010179208591580391,
-0.07550736516714096,
-0.02982400916516781,
0.0188777893781662,
-0.013332808390259743,
0.009138301014900208,
-0.038529232144355774,
-0.06031601130962372,
0.04223675653338432,
0.01229493971914053,
0.0017810020362958312,
0.019272545352578163,
0.0337667390704155,
0.07583890110254288,
-0.0706508457660675,
-0.08857589960098267,
-0.042506713420152664,
-0.07204496115446091,
-0.041310638189315796,
-0.047876764088869095,
-0.07417193055152893,
0.054720617830753326,
0.06671588867902756,
-0.08407679945230484,
0.08008645474910736,
-0.07559124380350113,
0.04187857359647751,
-0.022021174430847168,
-0.14931736886501312,
-0.0951916053891182,
-0.021771106868982315,
0.08042217791080475,
-0.010790206491947174,
-0.010562222450971603,
-0.009465094655752182,
-0.01012294739484787,
0.031620774418115616,
-0.019726473838090897,
-0.03813227638602257,
-0.06675226241350174,
-0.022452320903539658,
-0.0021234876476228237,
0.01168062724173069,
0.0490587055683136,
-0.06067671254277229,
-0.00856614951044321,
0.011349587701261044,
-0.05237337201833725,
0.02214011363685131,
0.0490763895213604,
-0.012531820684671402,
0.006999281235039234,
-0.049452584236860275,
0.056221798062324524,
-0.03442666307091713,
0.036354873329401016,
0.0640549287199974,
0.1061677485704422,
0.1295708417892456,
0.016915740445256233,
-0.04576843976974487,
0.06599953770637512,
0.04494898021221161,
-0.06539702415466309,
0.013076829724013805,
-0.010219496674835682,
0.0541786327958107,
-0.019638026133179665,
-3.7863589597009195e-8,
-0.026199564337730408,
0.06916147470474243,
-0.02715528942644596,
0.05204987898468971,
-0.01416091714054346,
-0.04308965802192688,
-0.056594185531139374,
0.03576431795954704,
-0.027599547058343887,
-0.06056690961122513,
0.033978454768657684,
0.06660949438810349,
-0.05387263372540474,
-0.12078867107629776,
-0.07574300467967987,
0.0892254114151001,
-0.0007812284748069942,
0.04272765666246414,
0.005698648281395435,
0.03819757699966431,
0.01598464883863926,
0.002784023294225335,
-0.004641328472644091,
-0.07014497369527817,
-0.009193887934088707,
-0.06238631159067154,
-0.036371856927871704,
0.10599015653133392,
0.08070903271436691,
-0.07054539769887924,
0.025665175169706345,
-0.019525781273841858,
0.01985054649412632,
-0.006575062405318022,
0.033059317618608475,
0.07045810669660568,
0.02617720328271389,
-0.04380325973033905,
-0.031309377402067184,
-0.04929368570446968,
0.012897050939500332,
0.04254607856273651,
-0.0003023988101631403,
0.017935745418071747,
0.03376622870564461,
0.03419160097837448,
0.027698712423443794,
-0.07845818251371384,
0.031419120728969574,
-0.006460906006395817,
-0.009824574925005436,
0.0023445875849574804,
-0.013155413791537285,
0.11921079456806183,
0.014861736446619034,
-0.005609571933746338,
0.04728024825453758,
-0.020990805700421333,
0.016848113387823105,
0.09372696280479431,
0.0013919747434556484,
0.028007792308926582,
-0.018178462982177734,
-0.03815346956253052
] | https://github.com/naver-ai/prolip | 50 | 0 | 0 | 0 |
Find A Winning Sign: Sign Is All We Need to Win the Lottery | https://openreview.net/forum?id=cLtE4qoPlD | [
"Junghun Oh",
"Sungyong Baik",
"Kyoung Mu Lee"
] | Poster | The Lottery Ticket Hypothesis (LTH) posits the existence of a sparse subnetwork (a.k.a. winning ticket) that can generalize comparably to its over-parameterized counterpart when trained from scratch.
The common approach to finding a winning ticket is to preserve the original strong generalization through Iterative Pruning (IP) and transfer information useful for achieving the learned generalization by applying the resulting sparse mask to an untrained network.
However, existing IP methods still struggle to generalize their observations beyond ad-hoc initialization and small-scale architectures or datasets, or they bypass these challenges by applying their mask to trained weights instead of initialized ones.
In this paper, we demonstrate that the parameter sign configuration plays a crucial role in conveying useful information for generalization to any randomly initialized network.
Through linear mode connectivity analysis, we observe that a sparse network trained by an existing IP method can retain its basin of attraction if its parameter signs and normalization layer parameters are preserved.
To take a step closer to finding a winning ticket, we alleviate the reliance on normalization layer parameters by preventing high error barriers along the linear path between the sparse network trained by our method and its counterpart with initialized normalization layer parameters.
Interestingly, across various architectures and datasets, we observe that any randomly initialized network can be optimized to exhibit low error barriers along the linear path to the sparse network trained by our method by inheriting its sparsity and parameter sign information, potentially achieving performance comparable to the original.
The code is available at https://github.com/JungHunOh/AWS_ICLR2025.git. | lottery ticket hypothesis, network pruning, linear mode connectivity | We demonstrate that an effective signed mask can allow any randomly initialized network to win the lottery. | 109 | null | [
-0.01696937158703804,
-0.022060714662075043,
0.05458107590675354,
0.031496793031692505,
0.023097598925232887,
0.04348728060722351,
0.03397117927670479,
-0.07093514502048492,
-0.06269939988851547,
-0.07835137844085693,
-0.012588932178914547,
0.015467979945242405,
0.05976221337914467,
-0.016434241086244583,
0.0022987222764641047,
-0.003334518289193511,
0.09862656146287918,
0.04797887057065964,
-0.04632585868239403,
-0.04337182268500328,
-0.0322895273566246,
-0.010423527099192142,
-0.065415158867836,
-0.022170644253492355,
0.08596876263618469,
-0.02646886743605137,
0.017799701541662216,
-0.00533585250377655,
0.01917308010160923,
0.0034555168822407722,
0.03998008370399475,
0.0011668875813484192,
-0.013213557191193104,
0.032347772270441055,
-0.04655992239713669,
0.0038567115552723408,
-0.03747570887207985,
0.07092917710542679,
-0.016755485907197,
-0.03673871234059334,
0.06562021374702454,
-0.029007596895098686,
-0.026288332417607307,
0.1022518202662468,
0.01951841451227665,
0.05767533555626869,
0.014133709482848644,
-0.001329046324826777,
0.05955333635210991,
-0.04807163029909134,
0.020112795755267143,
-0.01067379955202341,
0.012525667436420918,
0.011410807259380817,
0.05601048842072487,
-0.020023910328745842,
-0.02517852932214737,
0.0470266118645668,
-0.09820276498794556,
0.013334172777831554,
-0.03121279366314411,
-0.07638625055551529,
-0.04667367786169052,
-0.0071473815478384495,
-0.03332448750734329,
-0.0012089849915355444,
0.05959183722734451,
0.06014174968004227,
0.027232825756072998,
0.050094444304704666,
0.06791116297245026,
0.07868845015764236,
-0.068522147834301,
0.10322657227516174,
0.09567146748304367,
0.10707012563943863,
-0.006682638078927994,
-0.0029493095353245735,
0.002598848659545183,
-0.01385578978806734,
0.0020349170081317425,
0.0022690000478178263,
-0.046969179064035416,
0.024530651047825813,
0.12072934955358505,
-0.009015639312565327,
-0.05112959071993828,
0.016759954392910004,
0.04187985509634018,
-0.048628807067871094,
-0.07900087535381317,
0.029071619734168053,
-0.027838312089443207,
-0.016981642693281174,
0.012892666272819042,
0.07027379423379898,
-0.0060471901670098305,
-0.07875706255435944,
-0.022076377645134926,
0.14567382633686066,
-0.04316846281290054,
-0.02650781348347664,
0.03201661258935928,
0.00834160391241312,
0.1219991147518158,
0.054452355951070786,
0.01854165643453598,
-0.004876340739428997,
0.0789831280708313,
-0.08891651779413223,
-0.06636527925729752,
-0.0037448410876095295,
0.024828940629959106,
0.013344254344701767,
0.018371665850281715,
-0.005979460664093494,
-0.019671164453029633,
0.029183469712734222,
-0.03775345906615257,
0.04373021051287651,
-0.05244717746973038,
-0.0385054349899292,
0.02044707164168358,
0.032674215734004974,
-0.024951213970780373,
0.06942056119441986,
-0.06378636509180069,
2.0370119564880342e-33,
0.0779353678226471,
0.07186026871204376,
0.0025385902263224125,
-0.006772385910153389,
0.002072693081572652,
0.009075941517949104,
0.054376356303691864,
0.01216342207044363,
0.02298634499311447,
0.03907577693462372,
-0.09100445359945297,
0.007130044046789408,
0.003974733408540487,
0.07423126697540283,
0.021323125809431076,
0.01765989139676094,
-0.010300100781023502,
0.035801421850919724,
0.04573292285203934,
-0.10773404687643051,
0.018878858536481857,
-0.050348129123449326,
-0.006718765012919903,
-0.0996282622218132,
0.021990757435560226,
-0.08034162223339081,
0.006956881377846003,
-0.0789814218878746,
0.006162668112665415,
0.006237916648387909,
-0.05499656870961189,
0.013506080023944378,
0.03998244181275368,
0.015699805691838264,
0.003049368504434824,
-0.00392204849049449,
0.029547149315476418,
-0.053681474179029465,
0.016422735527157784,
-0.007157653570175171,
-0.051466554403305054,
-0.004333459306508303,
-0.05665702745318413,
-0.018360968679189682,
-0.09870926290750504,
-0.017962411046028137,
-0.03083508089184761,
0.0721077173948288,
-0.04656527563929558,
0.012367462739348412,
0.0230819471180439,
-0.02373439632356167,
-0.02575964294373989,
-0.0678471103310585,
-0.046342529356479645,
-0.0640740916132927,
0.052469562739133835,
0.023386605083942413,
-0.005236487369984388,
0.06573135405778885,
-0.031630102545022964,
-0.01880396157503128,
-0.02416517399251461,
0.11233770102262497,
-0.002163828117772937,
0.018328048288822174,
-0.05881374701857567,
-0.03591359406709671,
0.07232602685689926,
-0.06962503492832184,
-0.0027740923687815666,
0.04211197420954704,
-0.06760307401418686,
-0.0632772296667099,
0.04209361597895622,
-0.02253207005560398,
0.04517055302858353,
-0.026689857244491577,
-0.02697722241282463,
0.035551805049180984,
-0.010157575830817223,
0.08964050561189651,
-0.034546010196208954,
-0.016975902020931244,
-0.021802550181746483,
0.0240553580224514,
0.04541249945759773,
-0.07732582837343216,
-0.01575538143515587,
-0.07896246016025543,
-0.03815842047333717,
-0.0010731462389230728,
-0.05461709201335907,
-0.02702188864350319,
0.01708207279443741,
-1.1381322730718588e-33,
-0.02637064829468727,
0.10359089821577072,
-0.06480146199464798,
0.0009161397465504706,
-0.03423909470438957,
-0.048089783638715744,
0.014127179980278015,
0.020056717097759247,
-0.06608216464519501,
-0.020927928388118744,
0.011244121007621288,
0.0028097808826714754,
0.04473535716533661,
-0.07150162756443024,
0.01673566736280918,
-0.06608347594738007,
-0.05977495759725571,
0.043299417942762375,
0.06836932897567749,
0.011509903706610203,
-0.030653946101665497,
0.052314236760139465,
-0.1371547281742096,
0.0003430218785069883,
0.05345906689763069,
-0.02175540290772915,
-0.012173877097666264,
0.11243149638175964,
-0.0023131701163947582,
-0.02465309016406536,
-0.007325317244976759,
0.0237599927932024,
-0.0439852699637413,
-0.07231035083532333,
0.019762510433793068,
0.10085730999708176,
0.054218702018260956,
0.05669896677136421,
-0.0003861749137286097,
0.06645918637514114,
-0.030689090490341187,
-0.04585113748908043,
-0.08177585899829865,
0.0007147329743020236,
0.02173932082951069,
-0.038442060351371765,
-0.0374513678252697,
0.0010381614556536078,
-0.030000073835253716,
0.0201144777238369,
-0.001633331528864801,
-0.007091388572007418,
-0.03158450499176979,
0.019390027970075607,
-0.02439754456281662,
0.07373490929603577,
-0.04697868973016739,
0.06592389941215515,
0.13544227182865143,
-0.0013083674712106586,
-0.018320145085453987,
-0.08911636471748352,
-0.058810677379369736,
0.04607314243912697,
-0.026573410257697105,
-0.0834130272269249,
0.044787850230932236,
-0.016216212883591652,
-0.05753370374441147,
0.014853906817734241,
-0.018472446128726006,
0.041403498500585556,
0.034715764224529266,
0.04404481500387192,
-0.02621450088918209,
-0.07756280153989792,
0.03514857590198517,
0.05937568098306656,
-0.03439530357718468,
0.008206319995224476,
0.008026715368032455,
0.07015946507453918,
-0.02131902612745762,
-0.0481097511947155,
0.11215025186538696,
0.1415662169456482,
0.0012706407578662038,
0.036921426653862,
0.04407229274511337,
-0.026881011202931404,
-0.05944310128688812,
0.030085226520895958,
-0.011001533828675747,
0.06455129384994507,
-0.020336877554655075,
-4.693915656162062e-8,
-0.10805845260620117,
0.08782663196325302,
-0.03456013649702072,
-0.017429381608963013,
0.08588466793298721,
0.013280884362757206,
0.026912227272987366,
0.04617249220609665,
0.02262735925614834,
-0.04533911123871803,
-0.0029963271226733923,
0.009682347066700459,
-0.12267002463340759,
-0.03960523381829262,
-0.07370118796825409,
0.0730445384979248,
-0.015068520791828632,
-0.08746086061000824,
-0.03978924825787544,
0.021519608795642853,
-0.0019311661599203944,
0.012636967934668064,
-0.09463687241077423,
-0.024160919710993767,
0.06605328619480133,
-0.09520522505044937,
0.011254421435296535,
0.1206786185503006,
0.06166452541947365,
0.05490245297551155,
-0.058957818895578384,
-0.010653704404830933,
0.12360738217830658,
-0.01695765182375908,
0.016603168100118637,
0.16560128331184387,
0.022604642435908318,
0.030519725754857063,
-0.05055597424507141,
-0.05427734926342964,
-0.039484549313783646,
0.029620517045259476,
0.002524352865293622,
-0.06094849482178688,
0.02651609666645527,
-0.013810002245008945,
-0.00016287622565869242,
0.008320804685354233,
0.06914655864238739,
-0.08852426707744598,
0.05781228095293045,
-0.00302856951020658,
0.0128416633233428,
0.040556035935878754,
0.05770231783390045,
-0.025966200977563858,
0.005462903529405594,
-0.06691645085811615,
0.0697278156876564,
0.0505584292113781,
-0.017497975379228592,
-0.006428265944123268,
-0.11884120106697083,
0.031497228890657425
] | 0 | 0 | 0 | 0 |
|
IPDreamer: Appearance-Controllable 3D Object Generation with Complex Image Prompts | https://openreview.net/forum?id=3PguviI7Uf | [
"Bohan Zeng",
"Shanglin Li",
"Yutang Feng",
"Ling Yang",
"Juan Zhang",
"Hong Li",
"Jiaming Liu",
"Conghui He",
"Wentao Zhang",
"Jianzhuang Liu",
"Baochang Zhang",
"Shuicheng YAN"
] | Poster | Recent advances in 3D generation have been remarkable, with methods such as DreamFusion leveraging large-scale text-to-image diffusion-based models to guide 3D object generation. These methods enable the synthesis of detailed and photorealistic textured objects. However, the appearance of 3D objects produced by such text-to-3D models is often unpredictable, and it is hard for single-image-to-3D methods to deal with images lacking a clear subject, complicating the generation of appearance-controllable 3D objects from complex images. To address these challenges, we present IPDreamer, a novel method that captures intricate appearance features from complex **I**mage **P**rompts and aligns the synthesized 3D object with these extracted features, enabling high-fidelity, appearance-controllable 3D object generation. Our experiments demonstrate that IPDreamer consistently generates high-quality 3D objects that align with both the textual and complex image prompts, highlighting its promising capability in appearance-controlled, complex 3D object generation. | 3D generation, Diffusion model | null | 108 | 2310.05375 | [
-0.10940971225500107,
0.011506324633955956,
0.03471267223358154,
0.022687751799821854,
0.023883458226919174,
-0.06056280806660652,
-0.07949979603290558,
-0.0800030380487442,
0.035752058029174805,
-0.05960967391729355,
-0.06023332104086876,
-0.07545992732048035,
0.01753847487270832,
-0.004014344420284033,
0.030040092766284943,
-0.011540672741830349,
0.02208474464714527,
0.09086957573890686,
-0.0428568534553051,
0.07451090961694717,
-0.013398319482803345,
-0.07642791420221329,
0.023753322660923004,
-0.030045881867408752,
0.017006954178214073,
0.06784319132566452,
0.041141197085380554,
-0.0542941689491272,
0.021858658641576767,
-0.05756111443042755,
0.05132833495736122,
-0.018309492617845535,
-0.0030565536580979824,
-0.013302292674779892,
-0.05726457014679909,
0.05673587694764137,
-0.03457299992442131,
0.03395960107445717,
-0.04260292649269104,
-0.07040536403656006,
0.05303823575377464,
0.008920490741729736,
-0.001724040717817843,
0.029785536229610443,
0.03753819316625595,
0.02158927172422409,
0.006298454944044352,
-0.01609726808965206,
0.050088562071323395,
-0.046158403158187866,
-0.059943787753582,
-0.1280633807182312,
0.010427656583487988,
-0.019021041691303253,
0.010322021320462227,
0.07949692010879517,
0.05315433442592621,
-0.08188850432634354,
0.03921229764819145,
-0.010037321597337723,
-0.02019617147743702,
0.034103311598300934,
0.05102526769042015,
-0.012628311291337013,
-0.030496617779135704,
-0.034753330051898956,
0.055837083607912064,
-0.029143474996089935,
-0.007870030589401722,
-0.04514685273170471,
-0.05893871188163757,
-0.0006240941002033651,
-0.055525802075862885,
0.022250240668654442,
0.04345647245645523,
0.07012137025594711,
-0.024122050032019615,
-0.10719496011734009,
-0.052097152918577194,
-0.020170530304312706,
0.032273050397634506,
-0.009207863360643387,
0.0405273362994194,
0.05750363692641258,
-0.07697030156850815,
0.049874141812324524,
0.011887798085808754,
0.040856413543224335,
-0.012205606326460838,
0.06310287863016129,
-0.07827960699796677,
-0.03224135562777519,
-0.06257288157939911,
-0.04036048799753189,
-0.03860313072800636,
0.0014662573812529445,
0.019660143181681633,
-0.11859316378831863,
0.02293921634554863,
0.04097738862037659,
0.02800087071955204,
-0.042433906346559525,
0.07230338454246521,
-0.02677980810403824,
0.048079054802656174,
-0.03044815920293331,
0.06102301925420761,
-0.0788334533572197,
0.006043802015483379,
0.03414810821413994,
0.05076059326529503,
-0.04114849120378494,
-0.03259163349866867,
-0.04705389589071274,
-0.005758124403655529,
0.03518027439713478,
-0.043234195560216904,
0.04023154452443123,
0.04437624663114548,
-0.0364266000688076,
0.009977971203625202,
0.012855798937380314,
-0.03987045958638191,
0.02643643133342266,
0.0020278876181691885,
-0.023878442123532295,
0.0010016437154263258,
-1.3617799915493841e-33,
-0.0064746420830488205,
0.05036156624555588,
0.022058239206671715,
0.12213713675737381,
0.09124387800693512,
-0.008217550814151764,
-0.01725340262055397,
-0.0060910121537745,
-0.0765048936009407,
0.01098540797829628,
-0.007487141061574221,
-0.11338341236114502,
-0.02653888426721096,
0.17168766260147095,
0.05197956785559654,
-0.06985073536634445,
-0.03447748348116875,
0.06528475880622864,
-0.04328896477818489,
0.09623009711503983,
0.0009050266235135496,
-0.003503252286463976,
0.010042735375463963,
-0.013691701926290989,
-0.0377783365547657,
0.05221563205122948,
0.04104245454072952,
0.023167897015810013,
0.03361830115318298,
-0.03817922621965408,
0.010718190111219883,
-0.01787715218961239,
0.05935077369213104,
-0.0447332002222538,
-0.004021531902253628,
0.010755381546914577,
-0.05556309223175049,
-0.08692813664674759,
-0.005722458939999342,
0.047591302543878555,
-0.008331780321896076,
0.029058970510959625,
-0.06516905128955841,
0.01744481734931469,
-0.0642264261841774,
0.09812406450510025,
0.06151885539293289,
0.0882856622338295,
-0.07945895195007324,
-0.023572849109768867,
0.056672655045986176,
0.02521209791302681,
-0.016491472721099854,
-0.07136260718107224,
-0.1178436279296875,
-0.09394398331642151,
0.008048429153859615,
-0.07340610772371292,
0.06151626259088516,
-0.003713975427672267,
0.009144533425569534,
0.05295082926750183,
-0.057104140520095825,
0.03909936919808388,
0.029663313180208206,
-0.01025936845690012,
0.06972888857126236,
0.05745476484298706,
0.02376951277256012,
0.04548538476228714,
-0.025117652490735054,
-0.010130158625543118,
-0.02481474168598652,
-0.03499755635857582,
0.05821233242750168,
0.041874196380376816,
0.005017105955630541,
-0.03989759460091591,
0.014080743305385113,
0.05329166725277901,
-0.10446596890687943,
0.08024067431688309,
0.008110634982585907,
-0.06549585610628128,
-0.0697840228676796,
-0.08645468950271606,
0.004567285068333149,
-0.042343057692050934,
-0.017492013052105904,
-0.03022681176662445,
-0.003577840281650424,
-0.07292652875185013,
0.014701188541948795,
-0.011879947036504745,
-0.022449610754847527,
5.260476350706416e-34,
0.05727497488260269,
-0.026013318449258804,
-0.058625757694244385,
-0.01797918789088726,
-0.006678031757473946,
-0.03152642771601677,
0.00219777156598866,
0.04445656016469002,
0.02272893488407135,
-0.021901099011301994,
0.019332801923155785,
0.07734186202287674,
0.002461763098835945,
-0.09074471890926361,
-0.019567085430026054,
-0.05779058113694191,
0.036416590213775635,
0.002302939770743251,
-0.07573217153549194,
0.0008377190679311752,
0.03220926225185394,
0.015638498589396477,
-0.044824209064245224,
-0.008156967349350452,
-0.12039730697870255,
0.1054723784327507,
0.026968521997332573,
0.03677375987172127,
0.03226202726364136,
0.06860765814781189,
-0.008425683714449406,
-0.013111808337271214,
0.06421288102865219,
-0.0046777790412306786,
-0.06673763692378998,
0.04305306822061539,
0.04871697351336479,
0.016053088009357452,
0.011022488586604595,
0.03352323919534683,
0.004123640712350607,
-0.04802663251757622,
-0.07032676041126251,
0.07640910893678665,
-0.0486658476293087,
0.010279224254190922,
-0.027392691001296043,
-0.010573790408670902,
0.05169663578271866,
0.108673095703125,
-0.055289801210165024,
0.06805568933486938,
-0.027870066463947296,
-0.023275062441825867,
0.0108846640214324,
-0.10204529017210007,
-0.15951800346374512,
-0.0584220215678215,
0.09802447259426117,
0.03491315245628357,
-0.06898147612810135,
-0.02888474613428116,
0.021071694791316986,
-0.030746491625905037,
-0.026876473799347878,
0.07137955725193024,
0.026723774150013924,
-0.035774581134319305,
0.04513269290328026,
0.03980937972664833,
0.07481975853443146,
-0.007064805831760168,
0.0359545461833477,
0.07565557211637497,
0.04119284078478813,
-0.07973285764455795,
0.05263479799032211,
0.04051071032881737,
0.016991455107927322,
-0.04459880664944649,
0.07604380697011948,
-0.051368631422519684,
0.01716463826596737,
0.10708685219287872,
0.09945093095302582,
-0.01116042397916317,
-0.034026943147182465,
0.039901264011859894,
-0.0068070306442677975,
0.09216829389333725,
-0.021343249827623367,
0.11759711056947708,
0.05534563213586807,
0.10770022869110107,
-0.0056129200384020805,
-3.636509404714161e-8,
-0.009084037505090237,
-0.02258152887225151,
0.05193678289651871,
-0.011371800675988197,
-0.07640355080366135,
-0.026169653981924057,
0.033574484288692474,
-0.043573059141635895,
-0.030859194695949554,
-0.05338592454791069,
0.02329445257782936,
-0.01748722791671753,
-0.02117667905986309,
-0.031090741977095604,
0.09337052702903748,
0.04625537618994713,
-0.08272185176610947,
0.04884343221783638,
-0.019164254888892174,
-0.030497124418616295,
-0.08096186816692352,
0.04373864829540253,
-0.0005491416086442769,
-0.06785503774881363,
0.04927390441298485,
0.023780152201652527,
-0.048910386860370636,
0.05326792970299721,
0.03891964256763458,
-0.0038792446721345186,
0.057675790041685104,
0.02945866622030735,
0.03485347330570221,
0.00633928133174777,
-0.02499064803123474,
0.010042853653430939,
-0.0818273052573204,
-0.030744142830371857,
0.024454018101096153,
-0.10723518580198288,
0.04359874129295349,
0.05737144500017166,
0.03133687749505043,
0.0070800986140966415,
0.02665429748594761,
-0.007165545132011175,
0.06626275926828384,
-0.09569956362247467,
0.01646888256072998,
-0.013157368637621403,
-0.04904000833630562,
-0.027476344257593155,
0.04903300851583481,
0.0803578644990921,
0.042368389666080475,
0.011296036653220654,
-0.02718162350356579,
0.02855118364095688,
0.06456725299358368,
0.07302238792181015,
0.0360945425927639,
-0.018140139058232307,
0.027237864211201668,
-0.013200410641729832
] | https://github.com/zengbohan0217/ipdreamer | 53 | 0 | 0 | 0 |
Gated Delta Networks: Improving Mamba2 with Delta Rule | https://openreview.net/forum?id=r8H7xhYPwz | [
"Songlin Yang",
"Jan Kautz",
"Ali Hatamizadeh"
] | Poster | Linear Transformers have gained attention as efficient alternatives to standard Transformers, but their performance in retrieval and long-context tasks has been limited. To address these limitations, recent work has explored two distinct mechanisms: gating for adaptive memory control and the delta update rule for precise memory modifications. We observe that these mechanisms are complementary—gating enables rapid memory erasure while the delta rule facilitates targeted updates. Building on this insight, we introduce the gated delta rule and develop a parallel training algorithm optimized for modern hardware. Our proposed architecture, Gated DeltaNet, consistently surpasses existing models like Mamba2 and DeltaNet across multiple benchmarks, including language modeling, common-sense reasoning, in-context retrieval, length extrapolation, and long-context understanding. We further enhance performance by developing hybrid architectures that combine Gated DeltaNet layers with sliding window attention or Mamba2 layers, achieving both improved training efficiency and superior task performance. | linear RNN, state-space model, linear transformer, subquadractic model, linear attention, delta rule, mamba | We introduce Gated DeltaNet, which combines the gating mechanism from Mamba2 with the delta rule from DeltaNet, achieving superior performance compared to both models individually. | 90 | 2412.06464 | [
-0.008918526582419872,
-0.05851667746901512,
-0.0010449144756421447,
0.04088224098086357,
0.002797139575704932,
0.016835661605000496,
-0.0011319746263325214,
-0.016640791669487953,
0.001394469290971756,
-0.06892209500074387,
0.032784413546323776,
-0.02628740668296814,
-0.00047773291589692235,
0.0012679906794801354,
0.06301847845315933,
-0.002957361750304699,
0.050683241337537766,
0.0842227190732956,
-0.08600810170173645,
-0.09944359958171844,
0.06594343483448029,
0.01509971171617508,
0.06385358422994614,
-0.028757786378264427,
-0.040024783462285995,
0.050583336502313614,
-0.06435972452163696,
-0.05287644639611244,
0.07966736704111099,
-0.03908358886837959,
0.06812556087970734,
0.02806277386844158,
-0.04557032510638237,
0.052495792508125305,
-0.07610747218132019,
-0.006022003013640642,
-0.11399750411510468,
0.061764754354953766,
-0.003000345779582858,
-0.11754776537418365,
0.0014302098425105214,
-0.0364835225045681,
0.055854201316833496,
0.053084596991539,
0.04516986384987831,
-0.0020841199439018965,
0.013372482731938362,
-0.05576160550117493,
0.08493468165397644,
0.012176070362329483,
-0.04688270017504692,
-0.02637845277786255,
0.030626192688941956,
0.05034271627664566,
-0.0003064821648877114,
0.09009073674678802,
0.01830963045358658,
0.08611123263835907,
-0.03882119432091713,
-0.0012700518127530813,
-0.057917047291994095,
-0.06610038876533508,
-0.029569948092103004,
0.005562172271311283,
-0.07287471741437912,
-0.043605152517557144,
0.08932201564311981,
0.00839594379067421,
-0.006393529940396547,
-0.00961928442120552,
-0.02466418407857418,
0.0670267641544342,
0.01878643035888672,
0.06859651207923889,
-0.00555421132594347,
0.06122235953807831,
0.08899574726819992,
-0.04822832718491554,
0.07113979756832123,
-0.04718732461333275,
0.039498887956142426,
-0.06148802489042282,
0.037082280963659286,
-0.004641386680305004,
0.09944786876440048,
-0.0036626134533435106,
-0.04803088679909706,
0.053089439868927,
0.05350669473409653,
-0.01543370634317398,
0.03493793308734894,
-0.13328832387924194,
0.05617561936378479,
-0.020048731938004494,
0.01013185828924179,
0.05516434833407402,
-0.02150532230734825,
-0.037666477262973785,
-0.10837827622890472,
0.08027694374322891,
0.05355890095233917,
0.056555118411779404,
-0.02398415468633175,
0.006522130686789751,
0.00943529698997736,
0.00836872961372137,
0.03525441139936447,
0.09596472233533859,
0.027536598965525627,
-0.06368561834096909,
0.008914466947317123,
0.05729665607213974,
0.01784912310540676,
-0.004663340747356415,
0.027686821296811104,
-0.052852168679237366,
0.005644470453262329,
-0.058665547519922256,
-0.004848243668675423,
0.05335403233766556,
-0.0637856274843216,
-0.02773900330066681,
-0.06057336553931236,
-0.01520534884184599,
-0.012829088605940342,
-0.02824569307267666,
-0.07225006818771362,
2.207530085373277e-33,
0.054861508309841156,
-0.020418377593159676,
0.0010005001677200198,
-0.014141988940536976,
0.0721588209271431,
-0.028686635196208954,
0.04581896588206291,
0.011179340071976185,
-0.07317441701889038,
0.0014526317827403545,
-0.004523547366261482,
-0.010049697011709213,
-0.10484005510807037,
0.08274353295564651,
0.04712004214525223,
-0.04075288400053978,
-0.07718943804502487,
-0.0023544859141111374,
0.025978943333029747,
-0.048072680830955505,
0.02880457416176796,
0.00611521303653717,
0.04395308345556259,
-0.019482608884572983,
0.07060170918703079,
-0.0166445542126894,
0.059937749058008194,
-0.02811453491449356,
-0.031030597165226936,
0.021693158894777298,
-0.14240582287311554,
0.007320199627429247,
-0.007265177089720964,
-0.006721526850014925,
0.055460356175899506,
-0.05075932666659355,
0.015936441719532013,
-0.054819364100694656,
0.03848366439342499,
-0.05654594302177429,
-0.01265355758368969,
0.005704063922166824,
0.04209010675549507,
-0.02311810664832592,
-0.1400858610868454,
-0.08938274532556534,
-0.028371747583150864,
0.018137464299798012,
-0.011602116748690605,
0.05054376646876335,
0.042458970099687576,
-0.003509319154545665,
-0.020246196538209915,
-0.013999252580106258,
0.041796423494815826,
-0.03068222478032112,
0.03295658528804779,
0.13067178428173065,
0.06696867197751999,
0.1219511553645134,
0.01047475729137659,
0.049614958465099335,
-0.0254793893545866,
0.16034352779388428,
-0.006498346105217934,
0.04474741220474243,
0.017215488478541374,
0.1147652342915535,
0.02425505965948105,
-0.061919182538986206,
-0.0012737123761326075,
-0.04629454389214516,
0.03439091145992279,
0.04807214438915253,
0.08305524289608002,
-0.017560750246047974,
0.0007416012231260538,
-0.09915253520011902,
-0.02148428186774254,
0.02114224247634411,
0.04936633259057999,
0.008095811121165752,
-0.026256483048200607,
0.04459267109632492,
-0.009321637451648712,
-0.03257297724485397,
0.027986062690615654,
-0.06305976212024689,
0.007443703711032867,
-0.058715663850307465,
-0.05708913877606392,
-0.04692737013101578,
0.031204113736748695,
0.0031324790325015783,
-0.07348859310150146,
-1.6400757995481644e-33,
-0.014088449999690056,
0.013076741248369217,
-0.030374251306056976,
0.029448596760630608,
-0.021505530923604965,
-0.05202031508088112,
0.05579233542084694,
0.03904392942786217,
-0.10423216223716736,
-0.10875242203474045,
-0.06964769959449768,
0.012437903322279453,
0.045278385281562805,
-0.04873077571392059,
0.04368709772825241,
-0.01567378267645836,
-0.05585961416363716,
0.0016238734824582934,
-0.050966162234544754,
0.028804177418351173,
0.03547772392630577,
0.09696751832962036,
-0.09640487283468246,
0.030802197754383087,
-0.05805910751223564,
-0.010241219773888588,
-0.05361441895365715,
0.08301174640655518,
-0.015388919040560722,
-0.041861023753881454,
-0.04989178851246834,
0.0454765148460865,
0.018798794597387314,
-0.062438081949949265,
-0.0364401675760746,
-0.029365474358201027,
0.0583677738904953,
-0.031325992196798325,
-0.06265255063772202,
0.038493696600198746,
0.09486957639455795,
-0.025470523163676262,
-0.019105980172753334,
0.011841299943625927,
-0.009803892113268375,
0.07420118898153305,
-0.11898567527532578,
-0.04275749623775482,
-0.02852497063577175,
0.038275692611932755,
-0.0883571058511734,
-0.005433614365756512,
-0.0487004891037941,
0.02073262631893158,
-0.010982517153024673,
-0.08609455823898315,
-0.020663606002926826,
-0.003807953791692853,
0.07285454869270325,
-0.015413200482726097,
-0.011099297553300858,
-0.04301838204264641,
-0.03433579206466675,
-0.07366465032100677,
0.04899461939930916,
-0.0030284279491752386,
0.07580764591693878,
0.0034539697226136923,
-0.021180376410484314,
0.009543132036924362,
-0.02647908218204975,
0.016464829444885254,
-0.061322543770074844,
0.014946323819458485,
-0.014764937572181225,
-0.021275872364640236,
-0.036129195243120193,
-0.09793305397033691,
0.013715806417167187,
-0.0037558069452643394,
-0.03189282864332199,
0.028530815616250038,
-0.001498183119110763,
0.028671737760305405,
0.03264228627085686,
0.09728923439979553,
0.010455602779984474,
0.061324574053287506,
0.06052341312170029,
-0.023851320147514343,
-0.019399607554078102,
0.017275378108024597,
0.01725502498447895,
0.05580548197031021,
-0.05685074254870415,
-3.949688220927783e-8,
-0.0729871392250061,
0.06971067935228348,
-0.027001582086086273,
0.028278490528464317,
0.027374088764190674,
-0.05242802947759628,
-0.05656419321894646,
0.10043418407440186,
0.027616193518042564,
-0.020175140351057053,
0.08193432539701462,
-0.04889420419931412,
0.012831438332796097,
-0.01601545140147209,
0.013210119679570198,
0.10350518673658371,
0.07441705465316772,
0.017517847940325737,
0.016645506024360657,
-0.08832041174173355,
0.022669045254588127,
0.041110411286354065,
0.05376233160495758,
0.04906371608376503,
0.018526237457990646,
-0.05810531601309776,
-0.044267550110816956,
0.13366302847862244,
0.07014531642198563,
-0.07017498463392258,
-0.055610787123441696,
0.0036163891199976206,
0.031171929091215134,
-0.039257124066352844,
0.012238804250955582,
0.03561703488230705,
0.05910512059926987,
-0.03408459201455116,
0.03698592260479927,
0.035131268203258514,
0.031157350167632103,
0.003887397237122059,
0.012745668180286884,
0.06728418916463852,
-0.021384717896580696,
0.032828062772750854,
-0.02687140926718712,
-0.03916233777999878,
-0.03321511670947075,
-0.06280959397554398,
0.07596959918737411,
0.011336589232087135,
-0.038860104978084564,
0.011797750368714333,
0.040439024567604065,
-0.012069630436599255,
0.008426814340054989,
0.009982078336179256,
0.041133128106594086,
0.0675034373998642,
0.01562713459134102,
0.03301446884870529,
-0.10032974183559418,
-0.005269445013254881
] | https://github.com/NVlabs/GatedDeltaNet | 157 | 0 | 0 | 0 |
Rethinking Classifier Re-Training in Long-Tailed Recognition: Label Over-Smooth Can Balance | https://openreview.net/forum?id=OeKp3AdiVO | [
"Siyu Sun",
"Han Lu",
"Jiangtong Li",
"Yichen Xie",
"Tianjiao Li",
"Xiaokang Yang",
"Liqing Zhang",
"Junchi Yan"
] | Poster | In the field of long-tailed recognition, the Decoupled Training paradigm has shown exceptional promise by dividing training into two stages: representation learning and classifier re-training. While previous work has tried to improve both stages simultaneously, this complicates isolating the effect of classifier re-training. Recent studies reveal that simple regularization can produce strong feature representations, highlighting the need to reassess classifier re-training methods. In this study, we revisit classifier re-training methods based on a unified feature representation and re-evaluate their performances.
We propose two new metrics, Logits Magnitude and Regularized Standard Deviation, to compare the differences and similarities between various methods.
Using these two newly proposed metrics, we demonstrate that when the Logits Magnitude across classes is nearly balanced, further reducing its overall value can effectively decrease errors and disturbances during training, leading to better model performance.
Based on our analysis using these metrics, we observe that adjusting the logits could improve model performance, leading us to develop a simple label over-smoothing approach to adjust the logits without requiring prior knowledge of class distribution.
This method softens the original one-hot labels by assigning a probability slightly higher than $\frac{1}{K}$ to the true class and slightly lower than $\frac{1}{K}$ to the other classes, where $K$ is the number of classes.
Our method achieves state-of-the-art performance on various imbalanced datasets, including CIFAR100-LT, ImageNet-LT, and iNaturalist2018. | Long-Tailed Recognition and Decoupled Training | In this study, we propose a new logits-based metric for classifier re-training in long-tailed recognition, introducing a simple retargeting method "Label Over-Smooth" that achieves state-of-the-art performance on imbalanced datasets. | 76 | null | [
-0.05228743329644203,
-0.04701472818851471,
0.03499578312039375,
-0.0012320706155151129,
0.07957324385643005,
0.05883415788412094,
-0.030719934031367302,
0.002783567178994417,
-0.05997413769364357,
-0.08627451211214066,
-0.07445785403251648,
-0.004049777053296566,
0.009659632109105587,
0.009184246882796288,
-0.059029340744018555,
-0.07245099544525146,
0.09677086770534515,
0.1067163422703743,
-0.1021616980433464,
0.020360685884952545,
0.03796087205410004,
0.032139282673597336,
0.06695718318223953,
0.11149781197309494,
0.03028811886906624,
-0.023736821487545967,
0.0008911007316783071,
-0.021591894328594208,
-0.026799079030752182,
-0.0776609554886818,
-0.021252958104014397,
0.02287001721560955,
-0.003981754649430513,
-0.014368888922035694,
-0.11923844367265701,
0.02218818850815296,
0.022586040198802948,
0.10487144440412521,
0.007141892798244953,
-0.04259751737117767,
0.0012164871441200376,
-0.028901709243655205,
-0.027927586808800697,
0.05573049187660217,
0.021301545202732086,
0.015199500136077404,
0.016592632979154587,
-0.052158694714307785,
0.0535413958132267,
0.05682719126343727,
0.009624293074011803,
0.00009584771760273725,
-0.035280242562294006,
0.08732644468545914,
-0.06212545931339264,
0.09260009229183197,
0.0633927434682846,
0.04112549498677254,
0.014164279215037823,
0.022274350747466087,
-0.030293185263872147,
-0.06983678042888641,
0.0794006958603859,
-0.003169001778587699,
-0.035842906683683395,
-0.07292920351028442,
-0.0012175824958831072,
0.03146734833717346,
0.001471583149395883,
0.03188634291291237,
-0.01671619340777397,
0.1137552410364151,
-0.07616934925317764,
0.05862963944673538,
0.011189376935362816,
0.0341792106628418,
-0.043546367436647415,
0.10101749002933502,
0.048142027109861374,
-0.04418645426630974,
-0.0569511353969574,
0.002496228087693453,
0.087003193795681,
-0.01540206465870142,
0.16530515253543854,
0.0013709616614505649,
-0.03901689499616623,
-0.017226744443178177,
-0.044616520404815674,
-0.030171196907758713,
-0.015158100984990597,
0.007047905586659908,
0.07490532845258713,
-0.06108979880809784,
-0.036861639469861984,
0.03776898607611656,
-0.01570148579776287,
-0.005572586786001921,
0.015073168091475964,
0.09492853283882141,
-0.10338908433914185,
0.04852508753538132,
-0.0011295543517917395,
-0.06049162149429321,
0.07180887460708618,
-0.021402746438980103,
0.015236878767609596,
-0.031762223690748215,
0.05855454504489899,
-0.08352062851190567,
0.04542647674679756,
-0.010402745567262173,
-0.08539923280477524,
0.0326736643910408,
0.055654678493738174,
0.05406109243631363,
-0.03627548739314079,
-0.01462002843618393,
-0.08954999595880508,
0.015021231025457382,
-0.04191930964589119,
0.0030180271714925766,
-0.0153787387534976,
-0.0278947576880455,
0.051133960485458374,
0.07362326979637146,
-0.015939125791192055,
5.124527949475422e-33,
-0.005785764195024967,
0.08928961306810379,
0.0015227808617055416,
-0.05532876029610634,
-0.022875696420669556,
-0.007307976018637419,
-0.03675737977027893,
0.05792278051376343,
0.0008146155159920454,
0.043925922363996506,
0.04617879167199135,
0.06459212303161621,
0.03225164860486984,
0.06864555180072784,
0.03218446671962738,
0.031059198081493378,
-0.05619928985834122,
0.0434793196618557,
-0.02043028362095356,
-0.10056176036596298,
-0.020129084587097168,
-0.03457646444439888,
0.06369069963693619,
-0.09885136038064957,
0.019890962168574333,
-0.005059387069195509,
0.03260494023561478,
0.007866937667131424,
-0.03534184768795967,
0.0162179134786129,
-0.010649007745087147,
-0.0062177712097764015,
0.0920235738158226,
-0.0527445487678051,
0.03747954219579697,
-0.11632896959781647,
0.016497205942869186,
-0.015353504568338394,
0.04498637840151787,
-0.01718774065375328,
-0.0439266562461853,
0.0005841598613187671,
-0.08554670214653015,
-0.05482245609164238,
-0.030232617631554604,
-0.008175432682037354,
0.016101596876978874,
-0.024630369618535042,
-0.07981909811496735,
0.006571529433131218,
0.001318231225013733,
-0.03014424815773964,
-0.024860626086592674,
-0.05047011002898216,
-0.07053328305482864,
0.0007761269225738943,
0.06347046792507172,
0.06902918219566345,
-0.0770583525300026,
0.019340114668011665,
-0.033885300159454346,
0.08125100284814835,
-0.0018830799963325262,
0.0487341582775116,
-0.003569386899471283,
-0.001776835648342967,
0.048097409307956696,
0.14534325897693634,
0.024072660133242607,
-0.031943049281835556,
0.04942438006401062,
0.02187136560678482,
-0.05082424357533455,
-0.04241891950368881,
0.057559434324502945,
-0.05011782795190811,
0.07250244915485382,
-0.02393665723502636,
-0.03255636990070343,
0.013052215799689293,
0.007586558815091848,
0.0017153703374788165,
0.004560945555567741,
-0.07224338501691818,
-0.023777572438120842,
-0.038563694804906845,
0.012591327540576458,
-0.0658607929944992,
-0.02693977579474449,
0.050840895622968674,
-0.05436919257044792,
0.01826227456331253,
-0.07446792721748352,
0.042975105345249176,
-0.02153121866285801,
-3.805520799439103e-33,
0.006275620777159929,
0.093102365732193,
-0.06955263763666153,
0.05464428663253784,
-0.0029480077791959047,
-0.0029100680258125067,
0.014385450631380081,
0.07778651267290115,
-0.07874327152967453,
-0.09566967934370041,
0.05797720327973366,
-0.06128481402993202,
-0.034130699932575226,
-0.021498488262295723,
0.018066171556711197,
0.025364430621266365,
-0.06108922138810158,
0.07483596354722977,
-0.011081397533416748,
-0.017225056886672974,
0.0722498670220375,
0.043194081634283066,
-0.08706601709127426,
0.06079860404133797,
-0.04363549128174782,
0.019193412736058235,
-0.019893554970622063,
0.05308821052312851,
0.062146734446287155,
-0.05073617771267891,
-0.04448457807302475,
-0.03893846273422241,
-0.014930682256817818,
-0.07387525588274002,
-0.042161308228969574,
-0.028107475489377975,
0.012637271545827389,
-0.018803758546710014,
0.014439924620091915,
0.12810924649238586,
0.05952128767967224,
-0.03655405715107918,
-0.07903105020523071,
-0.04874993488192558,
-0.007676200475543737,
-0.1097344383597374,
0.0036311827134341,
0.033988043665885925,
0.014639109373092651,
0.020534953102469444,
0.011872738599777222,
-0.025055421516299248,
-0.0908961147069931,
0.1012471541762352,
-0.0026169591583311558,
-0.04903059080243111,
-0.0017356566386297345,
-0.0023541825357824564,
-0.04733206331729889,
0.026766696944832802,
-0.08342485874891281,
0.023294756188988686,
0.024264005944132805,
0.028425641357898712,
0.017928767949342728,
-0.06799248605966568,
0.021567899733781815,
0.03906216099858284,
0.02045021392405033,
0.01715383492410183,
0.012038126587867737,
0.011828752234578133,
0.006286354269832373,
0.030331440269947052,
-0.0720769539475441,
0.009599101729691029,
0.038325272500514984,
-0.035512540489435196,
-0.018221141770482063,
0.016419576480984688,
-0.03769432380795479,
-0.022626249119639397,
-0.015453264117240906,
0.08196555823087692,
0.10435112565755844,
0.052788473665714264,
-0.0327153243124485,
0.046655286103487015,
0.04544179141521454,
0.00684405816718936,
0.01399452518671751,
0.05739869922399521,
-0.012660648673772812,
0.09437458217144012,
0.01656978763639927,
-4.6322405466980854e-8,
-0.07037485390901566,
0.03744147717952728,
0.024990644305944443,
0.04482365399599075,
0.09620417654514313,
0.03995110094547272,
-0.05246104300022125,
0.12877428531646729,
-0.10291123390197754,
-0.03994695097208023,
0.03539382666349411,
-0.03375919908285141,
-0.03275615721940994,
-0.031537845730781555,
-0.03159511834383011,
0.06695112586021423,
0.001389243989251554,
0.08807740360498428,
-0.014514639042317867,
0.027314364910125732,
0.01880633644759655,
-0.04757285118103027,
0.08004669100046158,
-0.0352281890809536,
-0.0388650968670845,
-0.09360799193382263,
-0.005064068827778101,
0.06338676810264587,
0.055723126977682114,
-0.04987592622637749,
-0.06874985247850418,
0.06943439692258835,
-0.020456437021493912,
-0.03021441586315632,
0.07063204795122147,
0.058133676648139954,
-0.009279447607696056,
-0.002739816438406706,
-0.021963197737932205,
0.07714688777923584,
-0.02283143624663353,
0.05400549992918968,
-0.03821616247296333,
-0.00662511819973588,
0.032613061368465424,
-0.010196720249950886,
0.061782509088516235,
-0.04206814616918564,
0.011022316291928291,
-0.04365471377968788,
-0.02101069502532482,
-0.06273075938224792,
-0.005894174333661795,
0.03880367800593376,
-0.00030185794457793236,
0.0638485923409462,
-0.022784482687711716,
-0.07403058558702469,
-0.02024112083017826,
0.0651656910777092,
0.027078574523329735,
-0.07030431926250458,
-0.06404039263725281,
0.031701818108558655
] | 0 | 0 | 0 | 0 |
|
STBLLM: Breaking the 1-Bit Barrier with Structured Binary LLMs | https://openreview.net/forum?id=6XUSDvBFkV | [
"Peijie Dong",
"Lujun Li",
"Yuedong Zhong",
"DaYou Du",
"Ruibo FAN",
"Yuhan Chen",
"Zhenheng Tang",
"Qiang Wang",
"Wei Xue",
"Yike Guo",
"Xiaowen Chu"
] | Poster | In this paper, we present the first structural binarization method for LLM compression to less than 1-bit precision. Although LLMs have achieved remarkable performance, their memory-bound nature during the inference stage hinders the adoption of resource-constrained devices. Reducing weights to 1-bit precision through binarization substantially enhances computational efficiency. We observe that randomly flipping some weights in binarized LLMs does not significantly degrade the model's performance, suggesting the potential for further compression. To exploit this, our STBLLM employs an N:M sparsity technique to achieve structural binarization of the weights. Specifically, we introduce a novel Standardized Importance (SI) metric, which considers weight magnitude and input feature norm to more accurately assess weight significance. Then, we propose a layer-wise approach, allowing different layers of the LLM to be sparsified with varying N:M ratios, thereby balancing compression and accuracy. Furthermore, we implement a fine-grained grouping strategy for less important weights, applying distinct quantization schemes to sparse, intermediate, and dense regions. Finally, we design a specialized CUDA kernel to support structural binarization. We conduct extensive experiments on LLaMA, OPT, and Mistral family. STBLLM achieves a perplexity of 11.07 at 0.55 bits per weight, outperforming the BiLLM by 3×. The results demonstrate that our approach performs better than other compressed binarization LLM methods while significantly reducing memory requirements. Code is released at https://github.com/pprp/STBLLM. | structured sparsification, language model, model compression, binary neural networks, computational efficiency | We introduces STBLLM, a novel approach that breaks the 1-bit barrier in language models by leveraging Structured Binary LLMs. | 66 | null | [
0.047347236424684525,
-0.011056693270802498,
-0.03095739148557186,
0.017119744792580605,
0.06383581459522247,
-0.04419979453086853,
-0.02622379921376705,
-0.06319459527730942,
-0.022598786279559135,
-0.06897193193435669,
-0.07416290044784546,
-0.0607232004404068,
-0.01212849747389555,
0.004420770797878504,
-0.01798170432448387,
-0.01576227881014347,
0.04943947494029999,
0.09570220857858658,
-0.08319211006164551,
-0.02654053084552288,
0.06349057704210281,
-0.011686938814818859,
-0.0023305374197661877,
-0.01337343454360962,
0.11569653451442719,
-0.031623825430870056,
-0.016152212396264076,
0.030605964362621307,
0.04351042956113815,
-0.10018032789230347,
0.05183005332946777,
0.023489931598305702,
0.0648481547832489,
-0.007678699214011431,
-0.01900039240717888,
0.010146599262952805,
0.025605136528611183,
-0.03133505582809448,
-0.04967525973916054,
-0.07645445317029953,
-0.04415019974112511,
-0.00453029153868556,
0.011480896733701229,
0.038136739283800125,
0.018159594386816025,
0.0907328650355339,
0.08214099705219269,
0.05700242519378662,
0.05844685807824135,
-0.022359594702720642,
0.06807776540517807,
0.07766719907522202,
-0.011601877398788929,
0.10352585464715958,
-0.05627044662833214,
-0.019839243963360786,
0.04515450820326805,
-0.012790086679160595,
-0.037316858768463135,
-0.0011746914824470878,
-0.05344799906015396,
-0.07362338900566101,
0.041783407330513,
-0.0012123386841267347,
0.01976088434457779,
0.012332881800830364,
0.07244240492582321,
-0.010935471393167973,
0.03533737733960152,
-0.0032687196508049965,
0.0160459466278553,
-0.03720196336507797,
-0.046318233013153076,
0.10165759176015854,
-0.013434134423732758,
0.08205042779445648,
0.008634510450065136,
0.021858686581254005,
0.0572432242333889,
-0.015545805916190147,
-0.025942476466298103,
0.00034321201383136213,
0.07226181030273438,
-0.02372194454073906,
0.07485118508338928,
0.003268864704295993,
0.003238801611587405,
0.056177444756031036,
-0.00272998190484941,
0.04645109176635742,
-0.02520827203989029,
0.0185165423899889,
-0.02746826596558094,
0.04569314420223236,
0.06812059879302979,
-0.06120045855641365,
-0.016459261998534203,
-0.045590583235025406,
-0.03764422982931137,
0.09811989963054657,
0.020518604665994644,
0.008969035930931568,
-0.006924846209585667,
0.046741727739572525,
0.05084969103336334,
-0.03337368369102478,
0.04290317744016647,
0.06659691780805588,
0.00352286733686924,
-0.05967705696821213,
0.05861020088195801,
0.06576269119977951,
0.016373077407479286,
-0.02374471351504326,
0.03419817239046097,
-0.051312584429979324,
-0.006209825165569782,
0.025313328951597214,
0.03376556187868118,
-0.0006248938152566552,
-0.04908556118607521,
-0.02565823867917061,
-0.137593075633049,
-0.02501748315989971,
0.018010757863521576,
-0.04917493090033531,
-0.08627515286207199,
1.482879878748577e-33,
0.06989721953868866,
0.041846293956041336,
-0.03730616718530655,
-0.021810319274663925,
0.03999266028404236,
0.03706052526831627,
0.002030441304668784,
0.0701131671667099,
-0.009739680215716362,
0.0389842726290226,
-0.11509451270103455,
-0.019244613125920296,
-0.021584052592515945,
0.03953428193926811,
0.0833335891366005,
0.006887956988066435,
-0.04357226938009262,
0.003635349916294217,
0.010458716191351414,
-0.12044551968574524,
-0.020611528307199478,
-0.016700411215424538,
0.0028163916431367397,
-0.027047444134950638,
0.04242904484272003,
-0.011980701237916946,
0.05019408091902733,
-0.06771348416805267,
-0.016884062439203262,
0.033163681626319885,
-0.12705181539058685,
-0.043461672961711884,
0.06928735971450806,
-0.049297064542770386,
0.0786423459649086,
-0.026107236742973328,
-0.03645235672593117,
-0.01954067125916481,
0.039266228675842285,
-0.037838369607925415,
0.0008455029455944896,
0.01253841258585453,
-0.015279805287718773,
-0.08046990633010864,
-0.1395430713891983,
-0.0359494574368,
0.018027672544121742,
0.013143624179065228,
-0.06999234110116959,
-0.034002695232629776,
0.08193831145763397,
-0.011846124194562435,
-0.07386394590139389,
0.011008696630597115,
0.0018035826506093144,
-0.08045898377895355,
0.06839095056056976,
0.0005023474805057049,
0.03866354376077652,
0.09528108686208725,
-0.052157942205667496,
0.08701305836439133,
0.02074507810175419,
0.05907648429274559,
0.010134777054190636,
0.01750592701137066,
-0.0019028817769140005,
0.09524199366569519,
-0.059153731912374496,
-0.012290969491004944,
-0.1237197294831276,
0.02579319290816784,
0.006147271487861872,
-0.042407531291246414,
0.0006700551020912826,
0.006166765931993723,
0.08945590257644653,
-0.08859437704086304,
0.028045546263456345,
0.009640350006520748,
0.01594221591949463,
0.04318323731422424,
-0.01529536210000515,
-0.0679924264550209,
-0.12557922303676605,
0.00021590878895949572,
0.016715815290808678,
-0.0711318626999855,
-0.07529862225055695,
-0.04267424717545509,
-0.029512714594602585,
-0.052551425993442535,
0.024182813242077827,
-0.01669071801006794,
-0.06642252951860428,
-1.2855043671764337e-33,
-0.07968369126319885,
0.047288235276937485,
0.0629560649394989,
0.09742725640535355,
-0.06327943503856659,
-0.078253373503685,
0.004649595357477665,
-0.061618901789188385,
0.004954375792294741,
-0.02503855898976326,
-0.00337256770581007,
-0.0032017610501497984,
0.009146008640527725,
-0.01709224097430706,
-0.004695475101470947,
0.007490306161344051,
-0.11337429285049438,
0.06890273839235306,
-0.04301498457789421,
0.03874834254384041,
-0.018627526238560677,
0.08086506277322769,
-0.044916946440935135,
0.0002611999516375363,
-0.029518265277147293,
0.009651959873735905,
-0.037361517548561096,
0.016840962693095207,
0.07920358330011368,
-0.017308371141552925,
-0.03368159011006355,
-0.0009375683730468154,
-0.03142629191279411,
-0.0344989076256752,
-0.01995905302464962,
0.019747912883758545,
0.04355726018548012,
0.052161116153001785,
0.034010931849479675,
0.011431346647441387,
0.07341834902763367,
0.03940416872501373,
-0.09643224626779556,
-0.014382176101207733,
0.029457954689860344,
0.003657371737062931,
-0.06988184154033661,
0.02575155906379223,
0.03323204070329666,
-0.029638957232236862,
0.03758029639720917,
0.12186387181282043,
0.04068446904420853,
0.08696410059928894,
0.05347765237092972,
-0.048868514597415924,
-0.0976073294878006,
-0.004452066496014595,
0.050069451332092285,
-0.0045472015626728535,
-0.0693068727850914,
-0.04938751086592674,
0.010188333690166473,
-0.08752909302711487,
0.01655559055507183,
-0.06985487043857574,
0.017721863463521004,
0.0012720766244456172,
-0.06185118108987808,
0.04235655814409256,
0.012306639924645424,
-0.032003868371248245,
0.04383724182844162,
0.021365683525800705,
-0.0553167462348938,
-0.038110360503196716,
-0.0073822736740112305,
-0.02171877957880497,
-0.023794757202267647,
0.07270562648773193,
0.03024938330054283,
0.055043965578079224,
0.02772449515759945,
0.01632962003350258,
0.1050676628947258,
0.09789590537548065,
0.051696255803108215,
0.005552572663873434,
0.03047798201441765,
0.03246186673641205,
-0.09637901186943054,
0.10399485379457474,
0.011053272522985935,
0.0370461530983448,
-0.03735491633415222,
-4.2910695441378266e-8,
0.01907912828028202,
-0.04166463389992714,
-0.0452863983809948,
-0.06398189812898636,
0.0188657995313406,
-0.010998563840985298,
-0.007402283139526844,
0.08616484701633453,
0.009840997867286205,
-0.0450560636818409,
0.13035549223423004,
-0.040285564959049225,
-0.13376255333423615,
-0.062295567244291306,
-0.04119189456105232,
0.07863864302635193,
0.010119476355612278,
-0.04909975081682205,
0.03746020793914795,
-0.020401280373334885,
-0.02663733996450901,
0.023087069392204285,
0.028000349178910255,
-0.004517436493188143,
0.02531534992158413,
-0.06755433231592178,
0.025651762261986732,
0.11473734676837921,
0.11338578909635544,
0.022648204118013382,
0.02340894564986229,
-0.007313310168683529,
0.07687082141637802,
-0.016219059005379677,
0.019168371334671974,
0.06147563084959984,
-0.0073186494410037994,
0.0818546861410141,
0.038355376571416855,
-0.01378586608916521,
0.0023341497872024775,
-0.0017738762544468045,
-0.016502656042575836,
0.029352238401770592,
-0.012699795886874199,
-0.013077698647975922,
0.0069952793419361115,
0.033058054745197296,
-0.08494820445775986,
-0.022251714020967484,
0.07253677397966385,
0.04419824853539467,
-0.04265584424138069,
0.055004458874464035,
-0.026893112808465958,
0.043919242918491364,
-0.04524984210729599,
-0.05454856529831886,
0.02372737042605877,
0.0447092242538929,
0.01802181638777256,
-0.07942362129688263,
-0.07298509776592255,
-0.023683741688728333
] | 0 | 0 | 0 | 0 |
|
3DTrajMaster: Mastering 3D Trajectory for Multi-Entity Motion in Video Generation | https://openreview.net/forum?id=Gx04TnVjee | [
"Xiao FU",
"Xian Liu",
"Xintao Wang",
"Sida Peng",
"Menghan Xia",
"Xiaoyu Shi",
"Ziyang Yuan",
"Pengfei Wan",
"Di ZHANG",
"Dahua Lin"
] | Poster | This paper aims to manipulate multi-entity 3D motions in video generation. Previous methods on controllable video generation primarily leverage 2D control signals to manipulate object motions and have achieved remarkable synthesis results. However, 2D control signals are inherently limited in expressing the 3D nature of object motions. To overcome this problem, we introduce 3DTrajMaster, a robust controller that regulates multi-entity dynamics in 3D space, given user-desired 6DoF pose (location and rotation) sequences of entities. At the core of our approach is a plug-and-play 3D-motion grounded object injector that fuses multiple input entities with their respective 3D trajectories through a gated self-attention mechanism. In addition, we exploit an injector architecture to preserve the video diffusion prior, which is crucial for generalization ability. To mitigate video quality degradation, we introduce a domain adaptor during training and employ an annealed sampling strategy during inference. To address the lack of suitable training data, we construct a 360-Motion Dataset, which first correlates collected 3D human and animal assets with GPT-generated trajectory and then captures their motion with 12 evenly-surround cameras on diverse 3D UE platforms. Extensive experiments show that 3DTrajMaster sets a new state-of-the-art in both accuracy and generalization for controlling multi-entity 3D motions. Project page: http://fuxiao0719.github.io/projects/3dtrajmaster | Controllable Video Generation, 3D Motion Control, Multi-Entity Motion | 3DTrajMaster masters multiple entity motions in 3D space for text-to-video (T2V) generation by leveraging entity-specific 6DoF pose sequences as additional inputs. | 47 | 2412.07759 | [
-0.010611212812364101,
-0.1673278659582138,
0.00938943400979042,
-0.06433499604463577,
0.0474700927734375,
-0.011560274288058281,
-0.016004841774702072,
-0.06713496893644333,
0.09716586023569107,
0.015639008954167366,
-0.03437883406877518,
-0.07278979569673538,
-0.04420504346489906,
0.016086479648947716,
0.0350089892745018,
0.04580400884151459,
0.06870830059051514,
0.13839396834373474,
-0.07816386967897415,
0.036017272621393204,
0.044127386063337326,
-0.04867933318018913,
0.08058879524469376,
0.017428608611226082,
-0.057516444474458694,
-0.017604805529117584,
0.059395961463451385,
-0.12779337167739868,
-0.01692773588001728,
-0.019439389929175377,
0.05946589633822441,
-0.014624408446252346,
-0.03949339687824249,
0.0018968690419569612,
-0.06349179893732071,
0.061237286776304245,
-0.028405245393514633,
-0.030320972204208374,
-0.06339102238416672,
-0.02526254951953888,
0.0416281521320343,
0.042256005108356476,
0.08858156949281693,
-0.033724427223205566,
0.03207669407129288,
0.11645442992448807,
-0.02040543407201767,
-0.08497012406587601,
0.07002653181552887,
0.0329849012196064,
-0.06301630288362503,
-0.07868099957704544,
0.01543225534260273,
0.0007390711107291281,
-0.008194888941943645,
0.03620672971010208,
0.011271893046796322,
-0.010378528386354446,
0.06488248705863953,
-0.025268668308854103,
0.05069386959075928,
-0.019173650071024895,
0.04628608003258705,
-0.024437930434942245,
-0.022661805152893066,
-0.07009942084550858,
0.06026134267449379,
-0.027691353112459183,
-0.09111529588699341,
0.023678593337535858,
-0.09946298599243164,
-0.025083431974053383,
0.009859164245426655,
-0.03156112879514694,
0.027441514655947685,
0.07580867409706116,
-0.022601014003157616,
-0.03440885990858078,
-0.00666264072060585,
-0.07895935326814651,
0.1120610311627388,
0.014339038170874119,
0.06341050565242767,
0.024789676070213318,
0.024355880916118622,
0.061271388083696365,
0.011392924934625626,
0.07387383282184601,
-0.07451213896274567,
0.05176778510212898,
-0.09638003259897232,
-0.012010140344500542,
0.049745649099349976,
-0.060962289571762085,
0.002053305273875594,
-0.04667746275663376,
-0.014693591743707657,
-0.12292743474245071,
0.026688871905207634,
0.031645942479372025,
-0.006857473403215408,
0.033772874623537064,
0.026566725224256516,
0.0709870457649231,
0.07674700021743774,
-0.017718931660056114,
-0.0019512205617502332,
0.04924675449728966,
-0.022892998531460762,
0.00007696224201936275,
0.0083865812048316,
0.017999427393078804,
-0.009827635250985622,
0.008915881626307964,
0.05850478634238243,
0.06558102369308472,
-0.05373046547174454,
0.015006202273070812,
-0.005718572996556759,
-0.0025971659924834967,
0.047079019248485565,
-0.03656373918056488,
-0.03110118769109249,
0.0030251413118094206,
0.036976464092731476,
-0.06400279700756073,
0.04087759926915169,
-1.3048749528745503e-33,
0.06996358931064606,
0.07299172133207321,
0.07966631650924683,
0.06751332432031631,
-0.014321879483759403,
0.03513231873512268,
0.019940992817282677,
0.03422914445400238,
-0.03662776201963425,
-0.016544485464692116,
-0.0012545466888695955,
-0.0706007331609726,
-0.04240240901708603,
0.0857679545879364,
0.05584314092993736,
-0.015811564400792122,
-0.02942768856883049,
0.007367256563156843,
0.04068046808242798,
0.07752195000648499,
0.02358231320977211,
0.014613069593906403,
-0.08819454163312912,
-0.024964047595858574,
-0.015983736142516136,
0.034113846719264984,
0.024556055665016174,
-0.005633944645524025,
-0.019429290667176247,
-0.028100404888391495,
-0.024034246802330017,
-0.03591713681817055,
-0.05327693372964859,
0.032638005912303925,
0.03686930239200592,
-0.017528578639030457,
-0.06037772446870804,
-0.06518837064504623,
-0.04010534659028053,
0.02396554872393608,
0.04705340042710304,
0.01338364277034998,
-0.052271921187639236,
-0.00909288227558136,
-0.12857653200626373,
-0.027605576440691948,
0.10946141183376312,
0.06622046232223511,
-0.08331092447042465,
0.037683695554733276,
0.07174983620643616,
0.05218516290187836,
0.01909957453608513,
-0.10446470230817795,
-0.08304906636476517,
-0.026947159320116043,
0.04721424728631973,
-0.03417963534593582,
0.025166021659970284,
0.023152228444814682,
-0.03764239326119423,
0.002210531383752823,
-0.02442089654505253,
0.019716650247573853,
0.00419508246704936,
0.008010435849428177,
-0.008583592250943184,
0.056042540818452835,
0.09275918453931808,
0.04278197139501572,
0.01668435148894787,
0.026329128071665764,
-0.0028967391699552536,
-0.08157174289226532,
0.053743526339530945,
0.03559379652142525,
0.0053581902757287025,
-0.059994377195835114,
-0.020164906978607178,
0.036916330456733704,
-0.027927953749895096,
-0.022790497168898582,
-0.012644326314330101,
0.0178946815431118,
-0.02503635361790657,
-0.026993749663233757,
0.018658051267266273,
-0.054679155349731445,
-0.036235958337783813,
-0.04434410482645035,
-0.023801878094673157,
-0.08437339216470718,
-0.07995597273111343,
0.004655996337532997,
0.02297007106244564,
1.2696142549472088e-33,
0.0844895988702774,
0.03678707033395767,
-0.04987740144133568,
0.02546115033328533,
-0.0006759737152606249,
-0.02827592007815838,
0.03692138195037842,
0.0163115207105875,
-0.01203097403049469,
-0.08937195688486099,
-0.05838966369628906,
-0.05105620250105858,
0.0342419371008873,
-0.010302070528268814,
0.019491003826260567,
-0.020770316943526268,
-0.03628731891512871,
-0.06317716836929321,
-0.03541184961795807,
-0.0013488198164850473,
0.036922845989465714,
0.05110494792461395,
0.02146945521235466,
-0.0015200028428807855,
-0.0550973005592823,
0.10216861963272095,
-0.0031116348691284657,
0.056318722665309906,
0.04160117357969284,
0.017275676131248474,
0.008330564945936203,
0.022109119221568108,
0.03718461096286774,
-0.0197482630610466,
-0.09749878197908401,
0.020890628919005394,
0.03299112245440483,
0.013481011614203453,
-0.07131677120923996,
0.019136715680360794,
0.024226821959018707,
-0.055221278220415115,
-0.02129889652132988,
0.12039488554000854,
-0.007255579810589552,
0.04068220779299736,
-0.025781605392694473,
0.06179054081439972,
0.03446527570486069,
0.11563204228878021,
-0.13920049369335175,
0.059726350009441376,
-0.05328615754842758,
-0.01940551958978176,
0.013028166256844997,
-0.08215604722499847,
-0.03136549890041351,
-0.06737938523292542,
0.1287544071674347,
-0.0296589657664299,
-0.0550205335021019,
0.03333068639039993,
-0.03661976382136345,
0.023045703768730164,
0.007072471082210541,
0.03661327809095383,
0.006293893791735172,
-0.012659081257879734,
0.03527899459004402,
-0.005833479110151529,
0.07373721897602081,
0.01065234374254942,
0.05460666865110397,
-0.024652162566781044,
-0.000618798949290067,
-0.07174914330244064,
0.037931837141513824,
-0.060430508106946945,
0.09398003667593002,
-0.020722072571516037,
0.03138219192624092,
-0.06701208651065826,
0.055866215378046036,
0.09159941226243973,
0.01145879551768303,
0.06776586920022964,
0.05421605333685875,
0.03768869861960411,
-0.07184155285358429,
0.07162296026945114,
-0.028743380680680275,
0.09229128062725067,
-0.010560925118625164,
0.030414940789341927,
-0.01334130484610796,
-3.53490960947056e-8,
-0.11597646772861481,
0.02970794215798378,
0.03573770076036453,
0.022091468796133995,
-0.06260370463132858,
-0.019311249256134033,
0.017924685031175613,
0.07113376259803772,
0.06534017622470856,
-0.03898723050951958,
0.009323571808636189,
-0.016119487583637238,
-0.00808227900415659,
-0.02954079955816269,
0.01868913322687149,
0.043453529477119446,
-0.038746628910303116,
-0.01836027204990387,
-0.011163510382175446,
0.09757713228464127,
-0.05459380894899368,
0.020505258813500404,
0.04493776336312294,
-0.09713462740182877,
0.023497838526964188,
0.033413153141736984,
0.003961675334721804,
0.006702220533043146,
0.08142629265785217,
0.004437616094946861,
-0.017792463302612305,
0.01449726615101099,
0.013747901655733585,
-0.019744819030165672,
-0.09778713434934616,
0.016341611742973328,
-0.009043354541063309,
-0.012430368922650814,
0.03241587057709694,
-0.0678989365696907,
-0.031501006335020065,
0.10718902200460434,
-0.05336882174015045,
0.014046179130673409,
-0.029231151565909386,
0.044080737978219986,
0.0030164276249706745,
-0.08549125492572784,
0.0471423976123333,
-0.016854681074619293,
-0.11335370689630508,
-0.04150199517607689,
0.02313917502760887,
0.10273271799087524,
0.0542655885219574,
0.025590427219867706,
0.02191087231040001,
-0.03536760061979294,
0.04130645468831062,
0.07018084079027176,
-0.058185916393995285,
-0.046154558658599854,
-0.019101213663816452,
0.004154291469603777
] | 0 | 0 | 0 | 0 |
|
Animate-X: Universal Character Image Animation with Enhanced Motion Representation | https://openreview.net/forum?id=1IuwdOI4Zb | [
"Shuai Tan",
"Biao Gong",
"Xiang Wang",
"Shiwei Zhang",
"DanDan Zheng",
"Ruobing Zheng",
"Kecheng Zheng",
"Jingdong Chen",
"Ming Yang"
] | Poster | Character image animation, which generates high-quality videos from a reference image and target pose sequence, has seen significant progress in recent years. However, most existing methods only apply to human figures, which usually do not generalize well on anthropomorphic characters commonly used in industries like gaming and entertainment. Our in-depth analysis suggests to attribute this limitation to their insufficient modeling of motion, which is unable to comprehend the movement pattern of the driving video, thus imposing a pose sequence rigidly onto the target character. To this end, this paper proposes $\texttt{Animate-X}$, a universal animation framework based on LDM for various character types (collectively named $\texttt{X}$), including anthropomorphic characters. To enhance motion representation, we introduce the Pose Indicator, which captures comprehensive motion pattern from the driving video through both implicit and explicit manner. The former leverages CLIP visual features of a driving video to extract its gist of motion, like the overall movement pattern and temporal relations among motions, while the latter strengthens the generalization of LDM by simulating possible inputs in advance that may arise during inference. Moreover, we introduce a new Animated Anthropomorphic Benchmark ($\texttt{$A^2$Bench}$) to evaluate the performance of $\texttt{Animate-X}$ on universal and widely applicable animation images. Extensive experiments demonstrate the superiority and effectiveness of $\texttt{Animate-X}$ compared to state-of-the-art methods. | Animation, Anthropomorphic, Video Generation, Pose | A universal animation framework based on LDM for various character types (collectively named X), including anthropomorphic characters. | 37 | null | [
-0.016034254804253578,
-0.08777117729187012,
0.0507277175784111,
-0.10697950422763824,
0.03873179852962494,
0.06436669826507568,
0.053195830434560776,
-0.037563640624284744,
0.051684193313121796,
-0.07052908092737198,
0.03750377520918846,
-0.04451950266957283,
-0.028018930926918983,
0.05316083878278732,
0.004677697084844112,
-0.005511918105185032,
0.10660512745380402,
0.10384007543325424,
-0.014746565371751785,
0.023995155468583107,
0.06797460466623306,
-0.0882905125617981,
-0.019978148862719536,
-0.038821037858724594,
-0.02248765528202057,
0.003154950914904475,
-0.009382475167512894,
-0.02237134985625744,
0.06279084086418152,
-0.009651859290897846,
0.0006767433951608837,
-0.0001707981718936935,
0.078703373670578,
0.020946120843291283,
-0.09906523674726486,
0.0990549698472023,
-0.03960287943482399,
-0.010007462464272976,
-0.07024495303630829,
-0.06123904138803482,
0.022973179817199707,
0.060921892523765564,
0.07054606825113297,
0.01096279639750719,
0.08256812393665314,
0.07703147083520889,
0.012423854321241379,
-0.07191749662160873,
-0.010248729027807713,
0.06885494291782379,
-0.08457452803850174,
-0.010168892331421375,
0.06720855087041855,
-0.003362566465511918,
-0.004308312200009823,
-0.0360691100358963,
0.04664595425128937,
-0.06407107412815094,
0.02910098060965538,
-0.031127016991376877,
0.024275489151477814,
0.01494321133941412,
0.06573976576328278,
0.03155739977955818,
-0.022751949727535248,
-0.07774478942155838,
0.06334066390991211,
-0.03349696844816208,
-0.024446474388241768,
0.06837159395217896,
-0.08059099316596985,
-0.10098465532064438,
0.024289151653647423,
-0.053067442029714584,
-0.04497932270169258,
0.022888999432325363,
0.029765641316771507,
-0.06556407362222672,
0.010600327514111996,
-0.030220409855246544,
0.0323554091155529,
-0.04474011808633804,
0.039967022836208344,
0.010954296216368675,
0.051726195961236954,
0.05820154771208763,
-0.048124708235263824,
0.042891960591077805,
-0.066020168364048,
0.11634906381368637,
-0.13674496114253998,
-0.0183245986700058,
0.10846026241779327,
-0.07403992116451263,
0.07947208732366562,
0.0068340725265443325,
0.020044252276420593,
-0.07565397024154663,
0.03794527053833008,
0.035535603761672974,
-0.03326497599482536,
0.027178660035133362,
0.06707876175642014,
-0.007012567017227411,
0.04059981182217598,
0.04277342930436134,
0.01656527817249298,
-0.018940512090921402,
0.039651889353990555,
-0.0073606581427156925,
0.022218050435185432,
-0.07896671444177628,
-0.021975941956043243,
-0.005653871688991785,
0.05832517147064209,
-0.019351642578840256,
-0.034609440714120865,
-0.0343112088739872,
0.006678166799247265,
-0.009595070034265518,
0.05668971687555313,
-0.01859249919652939,
-0.048071060329675674,
0.04816094785928726,
0.05604216083884239,
-0.06651619076728821,
-0.027413077652454376,
-4.830487404837835e-34,
0.028368528932332993,
0.07010958343744278,
0.019220521673560143,
0.02564748004078865,
-0.0013468661345541477,
-0.030558886006474495,
-0.04537016525864601,
0.0202666986733675,
-0.002460663439705968,
0.03151538223028183,
0.05026491358876228,
-0.06347845494747162,
-0.03561359643936157,
0.11100844293832779,
0.018391337245702744,
-0.01121226791292429,
-0.03455434739589691,
-0.018361836671829224,
0.018488110974431038,
0.08241676539182663,
0.008659696206450462,
0.06578146666288376,
-0.04191943630576134,
-0.025731191039085388,
-0.053011924028396606,
0.05505271255970001,
-0.0018436081008985639,
-0.055348146706819534,
0.005041550379246473,
-0.027122393250465393,
-0.04785081371665001,
-0.016671644523739815,
-0.0733976662158966,
-0.03783215582370758,
0.06401629000902176,
-0.0898466557264328,
-0.0669640451669693,
-0.07298672199249268,
0.024488480761647224,
0.08299065381288528,
-0.00811819825321436,
-0.005273283924907446,
-0.09408622235059738,
-0.047233715653419495,
-0.06877158582210541,
-0.007899579592049122,
0.10269659012556076,
0.10529451817274094,
-0.0922570452094078,
0.059883274137973785,
0.02507053129374981,
-0.0020996187813580036,
0.05237337574362755,
-0.08501338213682175,
-0.06536159664392471,
-0.0048104021698236465,
0.011909696273505688,
-0.020187530666589737,
-0.041461773216724396,
0.030084965750575066,
-0.07019497454166412,
0.009576026350259781,
0.04687444120645523,
-0.03796841576695442,
-0.024946488440036774,
-0.002824584022164345,
0.00448836712166667,
0.013309415429830551,
0.049526866525411606,
0.06834869831800461,
0.03300849348306656,
-0.0284538883715868,
0.009749880991876125,
-0.03529996797442436,
0.03708326444029808,
0.04643241688609123,
-0.03478462994098663,
-0.06595676392316818,
-0.07290968298912048,
0.07098828256130219,
-0.06970488280057907,
0.001401238376274705,
0.04140901565551758,
-0.029279451817274094,
0.017837096005678177,
-0.05465880408883095,
0.057020340114831924,
-0.027079127728939056,
-0.0384649820625782,
-0.10846512019634247,
-0.028400316834449768,
-0.03521467745304108,
-0.04987296462059021,
0.03914760798215866,
0.07564301788806915,
-4.933500658410696e-34,
0.076908178627491,
0.05198301747441292,
-0.019352341070771217,
-0.0006969301612116396,
0.011853094212710857,
-0.04633716866374016,
0.029148543253540993,
0.0037809270434081554,
-0.011927017010748386,
-0.013139158487319946,
-0.08318819105625153,
-0.016270941123366356,
-0.0016915228916332126,
-0.03815394267439842,
0.023666149005293846,
0.04128311574459076,
0.020252497866749763,
0.003318936564028263,
-0.11045088618993759,
0.012650948017835617,
0.03247717395424843,
0.0337701179087162,
-0.01779787242412567,
0.008282840251922607,
-0.041875310242176056,
0.11632762104272842,
0.04990293085575104,
0.06049547716975212,
0.03847009316086769,
0.00648430036380887,
-0.08603838086128235,
0.03179038316011429,
0.005975186824798584,
0.0016940913628786802,
-0.09156011044979095,
0.005179804749786854,
-0.02676885575056076,
0.015153130516409874,
-0.0005364878452382982,
0.03634008392691612,
0.07456667721271515,
-0.0676204040646553,
-0.021583762019872665,
0.03265541419386864,
0.023191146552562714,
0.011944188736379147,
-0.020477814599871635,
0.07004613429307938,
-0.003406449221074581,
0.059145569801330566,
-0.07499753683805466,
0.041915442794561386,
-0.020528405904769897,
-0.04292500391602516,
-0.058778803795576096,
-0.021435802802443504,
-0.04442090913653374,
0.0009837577817961574,
0.0495072677731514,
-0.0067560505121946335,
0.013376166112720966,
0.0058022672310471535,
-0.003396248212084174,
0.013356784358620644,
-0.07770366966724396,
-0.005098535213619471,
-0.048276033252477646,
-0.08650897443294525,
-0.06519538909196854,
-0.06736597418785095,
0.03651721775531769,
0.007881302386522293,
0.019005144014954567,
0.012113934382796288,
-0.057549312710762024,
-0.05582209676504135,
0.0711558610200882,
0.024476515129208565,
0.09000594913959503,
-0.07217547297477722,
0.017502937465906143,
-0.04909716174006462,
0.0941685289144516,
0.08206088840961456,
0.0422929972410202,
0.1605842113494873,
-0.08267749100923538,
0.04441201686859131,
-0.022805223241448402,
0.024916375055909157,
-0.044554416090250015,
0.14190328121185303,
-0.03637220710515976,
0.06944605708122253,
-0.03360014036297798,
-4.377418250101073e-8,
-0.13956588506698608,
0.014162290841341019,
0.016535261645913124,
-0.048810429871082306,
-0.06464169174432755,
-0.02440824545919895,
0.04042389988899231,
-0.057564184069633484,
0.010556917637586594,
-0.04990622401237488,
0.09627240896224976,
-0.008636592887341976,
0.017142385244369507,
0.017589149996638298,
-0.01294784527271986,
0.05391037464141846,
0.01122921984642744,
0.01206190139055252,
-0.072774738073349,
0.06921054422855377,
-0.026152808219194412,
-0.000628494075499475,
-0.04462531954050064,
-0.04301639646291733,
-0.0412648506462574,
-0.06274883449077606,
-0.04971735551953316,
0.03484372794628143,
0.05810210853815079,
0.0332358218729496,
-0.02400839887559414,
0.07403219491243362,
-0.0014393676538020372,
-0.07477809488773346,
0.009275771677494049,
-0.00288598146289587,
0.008919347077608109,
-0.01561054214835167,
0.015041732229292393,
-0.03381805866956711,
0.015336815267801285,
0.05610725283622742,
0.05124497786164284,
0.035141512751579285,
0.008363889530301094,
0.06737099587917328,
0.0381939634680748,
-0.10173244774341583,
0.010241837240755558,
0.011851854622364044,
-0.0035325183998793364,
-0.059023596346378326,
-0.05607622489333153,
0.07137532532215118,
0.040303852409124374,
0.007937480695545673,
0.028293075039982796,
0.03447922319173813,
0.09798578917980194,
0.026332585141062737,
-0.006809265818446875,
-0.036769766360521317,
0.014165429398417473,
-0.0350603312253952
] | 0 | 0 | 0 | 0 |
|
Inner Information Analysis Algorithm for Deep Neural Network based on Community | https://openreview.net/forum?id=awz1JPyXNK | [
"Guipeng Lan",
"Shuai Xiao",
"Meng Xi",
"Jiabao Wen",
"Jiachen Yang"
] | Poster | Deep learning has achieved advancements across a variety of forefront fields. However, its inherent 'black box' characteristic poses challenges to the comprehension and trustworthiness of the decision-making processes within neural networks. To mitigate these challenges, we introduce InnerSightNet, an inner information analysis algorithm designed to illuminate the inner workings of deep neural networks through the perspectives of community. This approach is aimed at deciphering the intricate patterns of neurons within deep neural networks, thereby shedding light on the networks' information processing and decision-making pathways. InnerSightNet operates in three primary phases, 'neuronization-aggregation-evaluation'. Initially, it transforms learnable units into a structured network of neurons. Subsequently, these neurons are aggregated into distinct communities according to representation attributes. The final phase involves the evaluation of these communities' roles and functionalities, to unpick the information flow and decision-making. By transcending focus on single-layer or individual neuron, InnerSightNet broadens the horizon for deep neural network interpretation. InnerSightNet offers a unique vantage point, enabling insights into the collective behavior of communities within the overarching architecture, thereby enhancing transparency and trust in deep learning systems. | inner information analysis, transparency, knowledge mining | null | 36 | null | [
-0.05237323045730591,
-0.07908467948436737,
-0.008611793629825115,
0.05361417308449745,
0.07176671177148819,
-0.05926494672894478,
0.018360203132033348,
-0.0826345682144165,
0.02073604427278042,
-0.06548447161912918,
-0.014924164861440659,
-0.0509936697781086,
-0.06064649671316147,
0.044556695967912674,
-0.1090531274676323,
-0.004363605752587318,
0.019856099039316177,
0.03199444338679314,
-0.10723267495632172,
-0.05174693465232849,
0.0664539560675621,
-0.023620355874300003,
-0.04049914702773094,
-0.0007861152989789844,
-0.025252381339669228,
0.0629807859659195,
-0.06676466017961502,
-0.042798105627298355,
0.010000137612223625,
-0.05921364575624466,
0.112525075674057,
0.03228670358657837,
0.012865344062447548,
0.05700678005814552,
-0.04762760549783707,
0.06547985970973969,
-0.06934859603643417,
0.0882144421339035,
0.025179438292980194,
-0.03600766882300377,
0.022735819220542908,
-0.009235164150595665,
-0.0008335376624017954,
0.04275934398174286,
0.05736953020095825,
0.06488543748855591,
-0.021461747586727142,
-0.08061172068119049,
0.04589783027768135,
-0.019288672134280205,
-0.07388806343078613,
-0.026348978281021118,
-0.01642250269651413,
0.0141508299857378,
0.04717589169740677,
0.06101243942975998,
-0.06284236907958984,
0.021361207589507103,
-0.06578204780817032,
-0.030696097761392593,
0.06240687519311905,
-0.027477115392684937,
-0.012463880702853203,
0.01320527121424675,
0.026451999321579933,
0.028717990964651108,
0.021372312679886818,
0.06130657717585564,
-0.0065190354362130165,
-0.05421233922243118,
0.12958799302577972,
0.09012483060359955,
-0.01886570081114769,
-0.04487008601427078,
0.03081846982240677,
0.06040211766958237,
0.047002796083688736,
-0.05120694637298584,
0.05141235515475273,
-0.06280893832445145,
0.0786449983716011,
0.10693637281656265,
-0.0031744446605443954,
0.029510345309972763,
0.10757249593734741,
-0.03501695767045021,
0.01099767442792654,
-0.04521019384264946,
0.01740274950861931,
0.008027189411222935,
-0.09569625556468964,
-0.05149365961551666,
0.006510306149721146,
-0.05151653662323952,
0.10650206357240677,
0.013480054214596748,
-0.08027170598506927,
-0.06623201817274094,
-0.0674097016453743,
0.06031414493918419,
-0.02607727237045765,
-0.03994232416152954,
0.009102018550038338,
-0.07340757548809052,
0.04471003636717796,
0.0023671381641179323,
0.06760517507791519,
0.10330867767333984,
0.08835625648498535,
0.007170728873461485,
-0.041440125554800034,
0.001951879938133061,
0.06575798243284225,
0.01858704909682274,
0.03663303703069687,
0.007465803995728493,
0.0568215548992157,
0.09564375877380371,
0.08039644360542297,
0.08572030067443848,
-0.05121474713087082,
0.03173614293336868,
0.02128509245812893,
-0.014566129073500633,
-0.004588726907968521,
0.05583176389336586,
-0.09216557443141937,
6.495692243035207e-34,
-0.02260194718837738,
-0.02166990377008915,
-0.03957926109433174,
-0.010616306215524673,
0.024309098720550537,
-0.06973925232887268,
-0.026402702555060387,
0.021420283243060112,
-0.03889032080769539,
0.044430866837501526,
-0.013210540637373924,
0.053352341055870056,
-0.03432418406009674,
0.09318757802248001,
0.06551963835954666,
0.018939485773444176,
-0.09302321076393127,
0.06523836404085159,
0.042098212987184525,
-0.07841889560222626,
0.021337542682886124,
-0.04837688058614731,
-0.012166742235422134,
-0.06460706144571304,
-0.011839988641440868,
-0.043735820800065994,
0.04404332488775253,
0.021043967455625534,
-0.02461880072951317,
0.009642048738896847,
-0.043491464108228683,
0.05047106742858887,
-0.010824687778949738,
-0.0543355718255043,
0.07330788671970367,
0.011384221725165844,
0.029732128605246544,
-0.01667601428925991,
0.05900229513645172,
0.019131623208522797,
-0.02409791201353073,
-0.012205129489302635,
-0.06618057191371918,
-0.04089905321598053,
-0.04283882677555084,
0.03697021305561066,
-0.04186633229255676,
0.055940110236406326,
-0.06132669746875763,
-0.05860232561826706,
0.0354146845638752,
-0.04147757589817047,
-0.0008355200407095253,
-0.010272744111716747,
-0.06866340339183807,
0.013417345471680164,
0.05114307999610901,
0.047315482050180435,
0.03285655751824379,
-0.01727680116891861,
0.02325783297419548,
0.02265501208603382,
-0.05434033274650574,
0.06462861597537994,
-0.024674225598573685,
0.0031212426256388426,
-0.014267979189753532,
0.11486584693193436,
0.08268589526414871,
-0.04232371225953102,
-0.08725269138813019,
0.058125630021095276,
0.05154414102435112,
-0.09207256883382797,
0.00833389163017273,
0.014044133946299553,
-0.00783628411591053,
-0.05662722885608673,
0.00649831211194396,
0.014909550547599792,
-0.001537592732347548,
0.01235166285187006,
-0.05077015608549118,
-0.021005410701036453,
-0.004932165611535311,
0.025074757635593414,
0.03017517551779747,
-0.008009796030819416,
0.0344659648835659,
0.013088383711874485,
-0.02729187160730362,
-0.03255912661552429,
0.016005221754312515,
-0.026033295318484306,
-0.03724164515733719,
-1.6808072298171336e-33,
-0.07859038561582565,
0.036538682878017426,
-0.0021821933332830667,
-0.01866184175014496,
0.023072324693202972,
-0.02926059626042843,
-0.08742082118988037,
-0.05593652278184891,
-0.04012769088149071,
0.010291983373463154,
-0.01577315293252468,
-0.013952846638858318,
0.06786775588989258,
0.0041798185557127,
0.05436444655060768,
-0.03722601756453514,
-0.006912905722856522,
-0.01098260935395956,
0.054195985198020935,
-0.028276147320866585,
0.02447528950870037,
0.04112609475851059,
-0.10238472372293472,
0.07206333428621292,
0.03685002774000168,
0.01420596707612276,
-0.011346649378538132,
0.01778486557304859,
0.011467378586530685,
-0.014370789751410484,
-0.019769109785556793,
-0.026090798899531364,
-0.05447251349687576,
-0.06604225188493729,
0.05147441476583481,
0.12584266066551208,
0.026778949424624443,
-0.1543324738740921,
-0.02539302036166191,
-0.03959539905190468,
0.031115001067519188,
-0.056606657803058624,
-0.04880121722817421,
-0.017348412424325943,
-0.028470801189541817,
-0.033323220908641815,
0.0220396239310503,
0.03533826768398285,
-0.02801559306681156,
0.04369666054844856,
0.05672279745340347,
-0.040438227355480194,
-0.008901533670723438,
0.03142356872558594,
-0.06956227123737335,
0.02002355456352234,
0.06864476948976517,
0.052701327949762344,
0.11599701642990112,
0.031601373106241226,
-0.02309522219002247,
-0.09116113185882568,
-0.08372526615858078,
0.017818627879023552,
-0.05522189661860466,
0.0010250286431983113,
0.01976374164223671,
0.04077308252453804,
0.009443238377571106,
0.03370538726449013,
0.006120097823441029,
-0.009876496158540249,
-0.008427775464951992,
0.011626515537500381,
-0.04118020087480545,
-0.036387886852025986,
0.08961483091115952,
-0.01379469782114029,
-0.012259005568921566,
-0.011739568784832954,
0.03573102876543999,
-0.05023664981126785,
0.049247317016124725,
0.0017226593336090446,
0.12739133834838867,
0.023970389738678932,
0.029325388371944427,
0.10807055979967117,
0.007259243633598089,
0.02766464278101921,
-0.06085728108882904,
-0.02834526263177395,
-0.10745154321193695,
-0.05345737561583519,
-0.08351900428533554,
-4.0515555355113975e-8,
-0.05397938936948776,
0.014730086550116539,
-0.008710860274732113,
0.05605728551745415,
0.03707636147737503,
-0.06847172975540161,
-0.008419879712164402,
0.05247417837381363,
-0.032406285405159,
0.0023766059894114733,
0.10503292828798294,
0.033529095351696014,
-0.09515701979398727,
-0.011851298622786999,
0.07092878967523575,
0.06257512420415878,
0.0032619640696793795,
-0.034047722816467285,
0.031074505299329758,
-0.04567603021860123,
0.036805395036935806,
-0.03157896548509598,
-0.015927299857139587,
0.025240397080779076,
0.037451907992362976,
-0.0832202285528183,
-0.10873881727457047,
0.04906643182039261,
0.017863279208540916,
-0.01244180928915739,
0.05681169033050537,
0.08542418479919434,
-0.017588064074516296,
-0.011326872743666172,
0.0018443536246195436,
0.1321159303188324,
0.03829629346728325,
0.006932820193469524,
0.05284750461578369,
0.004307860974222422,
-0.02264147438108921,
0.07219720631837845,
0.020508507266640663,
0.01660243608057499,
-0.0594857893884182,
-0.016196224838495255,
-0.017931250855326653,
-0.012427778914570808,
-0.030142787843942642,
-0.058740030974149704,
0.00910292100161314,
-0.035992320626974106,
0.030263779684901237,
0.128785640001297,
-0.008889870718121529,
-0.06184431165456772,
-0.06188981235027313,
-0.03548853099346161,
0.027085551992058754,
0.090067557990551,
0.029203491285443306,
0.03283081576228142,
-0.10587380081415176,
-0.023333854973316193
] | 0 | 0 | 0 | 0 |
|
AniSDF: Fused-Granularity Neural Surfaces with Anisotropic Encoding for High-Fidelity 3D Reconstruction | https://openreview.net/forum?id=v1f6c7wVBm | [
"Jingnan Gao",
"Zhuo Chen",
"Xiaokang Yang",
"Yichao Yan"
] | Poster | Neural radiance fields have recently revolutionized novel-view synthesis and achieved high-fidelity renderings.
However, these methods sacrifice the geometry for the rendering quality, limiting their further applications including relighting and deformation.
How to synthesize photo-realistic rendering while reconstructing accurate geometry remains an unsolved problem. In this work, we present AniSDF, a novel approach that learns fused-granularity neural surfaces with physics-based encoding for high-fidelity 3D reconstruction. Different from previous neural surfaces, our fused-granularity geometry structure balances the overall structures and fine geometric details, producing accurate geometry reconstruction.
To disambiguate geometry from reflective appearance, we introduce blended radiance fields to model diffuse and specularity following the anisotropic spherical Gaussian encoding, a physics-based rendering pipeline. With these designs, AniSDF can reconstruct objects with complex structures and produce high-quality renderings.
Furthermore, our method is a unified model that does not require complex hyperparameter tuning for specific objects.
Extensive experiments demonstrate that our method boosts the quality of SDF-based methods by a great scale in both geometry reconstruction and novel-view synthesis. | Surface Reconstruction, Neural Radiance Field | null | 22 | 2410.01202 | [
-0.07472830265760422,
-0.10499492287635803,
0.08706904202699661,
0.026608511805534363,
-0.013324353843927383,
-0.058364447206258774,
-0.032081007957458496,
-0.049139320850372314,
0.037272997200489044,
-0.02930922619998455,
-0.06107690930366516,
-0.07133207470178604,
-0.027995161712169647,
0.04397330433130264,
0.0046997456811368465,
-0.0052451156079769135,
0.043456632643938065,
0.09686030447483063,
-0.0968475192785263,
0.025798168033361435,
0.0030860065016895533,
-0.05150899291038513,
-0.017656097188591957,
-0.03581516444683075,
0.09929255396127701,
0.022160395979881287,
0.06450505554676056,
-0.036502063274383545,
0.00589323416352272,
-0.013162759132683277,
0.07140214741230011,
-0.012157206423580647,
-0.09649937599897385,
0.020742623135447502,
-0.0045052398927509785,
0.025726567953824997,
-0.025417914614081383,
0.010651507414877415,
-0.011019296944141388,
-0.018116716295480728,
0.008264577947556973,
0.018040185794234276,
-0.028944149613380432,
0.04088455066084862,
0.051313366740942,
0.06700795888900757,
0.014782574027776718,
-0.134978786110878,
0.07851822674274445,
-0.011269448325037956,
-0.0210659671574831,
-0.03267180919647217,
-0.00031121610663831234,
-0.023986969143152237,
0.012982169166207314,
0.13961295783519745,
-0.011336013674736023,
-0.08578457683324814,
0.003352222964167595,
0.0015182048082351685,
-0.020888367667794228,
0.053696341812610626,
0.038380201905965805,
0.005863250233232975,
0.010698108933866024,
-0.006299832370132208,
0.010702826082706451,
-0.061494793742895126,
0.0014080413384363055,
-0.006913663819432259,
-0.04413957521319389,
0.010297717526555061,
0.0162130668759346,
-0.07713436335325241,
0.016079459339380264,
0.04160167649388313,
0.002102398546412587,
-0.015574153512716293,
0.01676677167415619,
-0.07466047257184982,
0.04709812253713608,
0.08716702461242676,
0.02688564732670784,
0.02004886046051979,
-0.011949192732572556,
0.0445481538772583,
0.02513633482158184,
0.13472239673137665,
-0.06460313498973846,
0.014274547807872295,
-0.04434700310230255,
0.030622249469161034,
-0.037078097462654114,
-0.07114547491073608,
-0.013963690958917141,
0.0032262157183140516,
0.029577432200312614,
-0.1306479573249817,
0.04821097478270531,
0.03592359647154808,
0.005014716647565365,
-0.038796693086624146,
0.07162643224000931,
-0.011723301373422146,
0.005838207434862852,
-0.052060432732105255,
0.016533343121409416,
0.07297846674919128,
0.004383376333862543,
0.023001277819275856,
0.0876188799738884,
0.02783476933836937,
-0.02763131819665432,
0.009123357944190502,
-0.012764648534357548,
0.04175547510385513,
-0.024167044088244438,
0.02637539431452751,
0.03659617155790329,
0.029566271230578423,
-0.0006073940312489867,
0.021913060918450356,
-0.07967927306890488,
0.08724909275770187,
-0.0004978969809599221,
-0.08450967818498611,
-0.011358542367815971,
1.0605245748271367e-33,
0.06317351758480072,
0.10604655742645264,
-0.0021997787989676,
0.023719800636172295,
0.05925894156098366,
0.01901303045451641,
0.058668430894613266,
0.0001044239616021514,
0.03783514350652695,
0.019795896485447884,
-0.10959834605455399,
0.010660221800208092,
-0.0332694947719574,
0.09296412765979767,
-0.03491899371147156,
-0.008714726194739342,
0.0026416475884616375,
0.012579027563333511,
-0.06798768043518066,
0.02836729772388935,
0.003387056291103363,
0.021664947271347046,
0.06493940204381943,
-0.03128478676080704,
-0.08924965560436249,
0.0736381933093071,
-0.01767432689666748,
0.03736144304275513,
-0.054448455572128296,
-0.002835770370438695,
-0.05052905157208443,
-0.014297956600785255,
0.002287417184561491,
-0.014867779798805714,
0.00029612472280859947,
-0.0197946485131979,
0.03403143957257271,
-0.07316745072603226,
-0.08130888640880585,
-0.02956940047442913,
-0.027575399726629257,
0.0928162932395935,
0.028546350076794624,
0.04250461980700493,
-0.05375437065958977,
0.025545038282871246,
0.053265608847141266,
0.020279226824641228,
-0.04209044948220253,
0.02663194015622139,
0.006544575560837984,
-0.03851153701543808,
0.034519825130701065,
-0.10938636213541031,
-0.04548848792910576,
-0.049861181527376175,
-0.002609112998470664,
-0.031762201339006424,
0.06395729631185532,
0.002807226497679949,
0.05822451412677765,
0.0038347486406564713,
-0.0828123465180397,
0.011984073556959629,
-0.05162188038229942,
0.03181922808289528,
0.015587937086820602,
0.06986981630325317,
0.020451176911592484,
0.025982815772294998,
-0.023149600252509117,
0.049588125199079514,
-0.010859157890081406,
0.003973456099629402,
0.043187033385038376,
0.006730765104293823,
0.0907023623585701,
-0.058882538229227066,
0.0472986102104187,
0.0381198413670063,
-0.0885697603225708,
0.07567872107028961,
-0.03410768508911133,
-0.07518836110830307,
-0.08712378889322281,
-0.006449567619711161,
0.029467426240444183,
0.03920404240489006,
-0.02621774934232235,
-0.08679404109716415,
-0.06951092928647995,
-0.10799115896224976,
0.039121516048908234,
-0.11137916892766953,
-0.043705087155103683,
-1.1473970971017578e-33,
0.027596833184361458,
-0.027726953849196434,
-0.14647312462329865,
0.04109392315149307,
-0.0009393673972226679,
-0.023542998358607292,
-0.02214937098324299,
0.036704398691654205,
-0.10045336931943893,
-0.041972722858190536,
0.0008296582964248955,
0.020761763677001,
0.012823694385588169,
-0.07832605391740799,
0.002704954007640481,
-0.008322345092892647,
-0.015376823954284191,
-0.04762644320726395,
-0.07157653570175171,
-0.04247138276696205,
0.009586875326931477,
0.03901075944304466,
-0.03643043711781502,
0.010881906375288963,
-0.07009868323802948,
0.09871887415647507,
-0.03770484775304794,
-0.029206635430455208,
0.011726511642336845,
0.07198141515254974,
0.008410116657614708,
-0.028286969289183617,
0.03101847693324089,
-0.053585853427648544,
-0.04829413443803787,
0.04245346412062645,
0.06487166881561279,
0.001824354869313538,
-0.08103538304567337,
-0.06139951944351196,
-0.003506610868498683,
-0.08562725782394409,
0.02048826590180397,
0.09016362577676773,
-0.021629521623253822,
0.028165267780423164,
-0.01852337457239628,
-0.014638548716902733,
0.014221702702343464,
0.05090589076280594,
-0.04305373504757881,
-0.001544137136079371,
-0.024538200348615646,
-0.050787754356861115,
-0.030640624463558197,
-0.06184391677379608,
-0.08151903003454208,
-0.06346911936998367,
0.08427996933460236,
-0.031226912513375282,
-0.0552898533642292,
-0.05307089537382126,
-0.00502517307177186,
0.0622427836060524,
-0.01449501607567072,
0.037422019988298416,
-0.07072701305150986,
-0.029290547594428062,
0.019896438345313072,
0.044932473450899124,
-0.006531663704663515,
0.009625984355807304,
0.030053531751036644,
0.013576927594840527,
0.03620636463165283,
-0.051735881716012955,
0.06889638304710388,
0.009653313085436821,
0.05481948330998421,
0.02371375635266304,
0.09745080024003983,
-0.022198084741830826,
-0.0036850811447948217,
0.12136319279670715,
0.14476794004440308,
0.04581807553768158,
-0.036858342587947845,
0.007017952390015125,
0.025155046954751015,
0.0014646168565377593,
0.041413988918066025,
0.049403246492147446,
-0.0268330629914999,
0.04156259074807167,
0.03150905296206474,
-3.767718936842357e-8,
-0.0007575950585305691,
0.08126035332679749,
0.055774983018636703,
-0.04783681035041809,
-0.03496577590703964,
-0.035821136087179184,
0.046991314738988876,
0.02406240813434124,
-0.04410611838102341,
-0.1018882766366005,
0.0028113205917179585,
-0.10922995209693909,
-0.02691151574254036,
-0.008802460506558418,
0.08759325742721558,
0.1545485258102417,
0.028804628178477287,
0.056929539889097214,
0.008229821920394897,
0.08360809832811356,
-0.11149869859218597,
0.0205655749887228,
-0.007712342776358128,
-0.046183839440345764,
0.04885252192616463,
0.040443167090415955,
-0.08378664404153824,
0.062064945697784424,
0.1061854362487793,
-0.026628859341144562,
0.0384814590215683,
0.008657826110720634,
0.052008602768182755,
-0.001513482304289937,
-0.07502453774213791,
0.012372501194477081,
0.028729962185025215,
-0.020497659221291542,
-0.053130291402339935,
-0.05366014689207077,
-0.025074388831853867,
0.08934145420789719,
0.04916093870997429,
0.025205129757523537,
0.04778100550174713,
0.02912125550210476,
0.059131965041160583,
-0.06641379743814468,
-0.0060433391481637955,
0.06606072932481766,
-0.035764291882514954,
-0.03722946345806122,
-0.029444757848978043,
0.0782223790884018,
-0.0003049500519409776,
-0.04464446008205414,
0.0035344762727618217,
-0.03392577916383743,
0.050143539905548096,
0.02202252298593521,
0.04497816786170006,
0.01476510800421238,
-0.03642457723617554,
0.028154978528618813
] | 0 | 0 | 0 | 0 |
|
Scaling Large Language Model-based Multi-Agent Collaboration | https://openreview.net/forum?id=K3n5jPkrU6 | [
"Chen Qian",
"Zihao Xie",
"YiFei Wang",
"Wei Liu",
"Kunlun Zhu",
"Hanchen Xia",
"Yufan Dang",
"Zhuoyun Du",
"Weize Chen",
"Cheng Yang",
"Zhiyuan Liu",
"Maosong Sun"
] | Poster | Recent breakthroughs in large language model-driven autonomous agents have revealed that multi-agent collaboration often surpasses each individual through collective reasoning. Inspired by the neural scaling law—increasing neurons enhances performance, this study explores whether the continuous addition of collaborative agents can yield similar benefits. Technically, we utilize directed acyclic graphs to organize agents into a multi-agent collaboration network (MacNet), upon which their interactive reasoning is topologically orchestrated for autonomous task solving. Extensive evaluations reveal that it effectively supports collaboration among over a thousand agents, with irregular topologies outperforming regular ones. We also identify a collaborative scaling law—the overall performance follows a logistic growth pattern as agents scale, with collaborative emergence occurring earlier than traditional neural emergence. We speculate this may be because scaling agents catalyzes their multidimensional considerations during interactive reflection and refinement, thereby producing more comprehensive artifacts. The code is available at https://github.com/OpenBMB/ChatDev/tree/macnet. | Large Language Model, Autonomous Agent, Multi-Agent Collaboration, Interactive Reasoning | We examine the impact of scaling LLM agents in multi-agent task solving, extending traditional scaling from training (neuron collaboration) to inference (agent collaboration) & circumventing resource-intensive retraining via inference-time thinking. | 18 | 2406.07155 | [
0.022416207939386368,
-0.138315811753273,
-0.02197199873626232,
0.011339872144162655,
0.05939061567187309,
-0.02035641111433506,
-0.04995759576559067,
-0.0271598007529974,
0.06188514083623886,
-0.020012840628623962,
-0.03865066170692444,
-0.02358873188495636,
0.02860191836953163,
0.08857321739196777,
0.004032230470329523,
0.024952614679932594,
0.0236109159886837,
-0.02245965413749218,
-0.07015404850244522,
-0.08649321645498276,
-0.036233190447092056,
-0.01755056530237198,
0.010320090688765049,
-0.06692706793546677,
0.016652828082442284,
-0.00635921536013484,
-0.03205440193414688,
-0.11269967257976532,
0.027826977893710136,
0.037190355360507965,
0.06098209694027901,
0.0799766331911087,
-0.012841947376728058,
0.028814369812607765,
0.00949941761791706,
0.0802314504981041,
-0.03402263671159744,
0.05541115626692772,
0.0649096667766571,
-0.00730386795476079,
0.005715216044336557,
0.04070163518190384,
0.0298137329518795,
0.008207345381379128,
0.04043295606970787,
0.0011322987265884876,
-0.08758309483528137,
0.031785085797309875,
0.023390525951981544,
0.014368611387908459,
-0.10761874914169312,
-0.07039128988981247,
0.03658324480056763,
0.047553740441799164,
0.03982176259160042,
-0.011968342587351799,
-0.013077069073915482,
-0.002208967227488756,
0.006365305744111538,
-0.10227133333683014,
-0.02793373540043831,
-0.04296377673745155,
-0.006718939170241356,
0.004663025960326195,
-0.022361768409609795,
0.04724808782339096,
0.04179748520255089,
0.020765608176589012,
-0.021516917273402214,
0.02905089780688286,
0.04666169360280037,
0.022259017452597618,
0.0170972291380167,
-0.006590703967958689,
0.09260585159063339,
0.032565388828516006,
-0.03176753595471382,
-0.0627017691731453,
0.05532598868012428,
-0.0967845767736435,
-0.060302890837192535,
0.040041614323854446,
0.00588280288502574,
0.027959447354078293,
0.06732931733131409,
-0.009057164192199707,
0.019602302461862564,
0.015145689249038696,
0.009768380783498287,
0.013598531484603882,
-0.05021383613348007,
0.014001824893057346,
0.025241881608963013,
0.008416993543505669,
-0.000562639266718179,
0.07660983502864838,
-0.02591579034924507,
-0.06664552539587021,
-0.02585810050368309,
0.10115903615951538,
-0.04378323256969452,
0.022792942821979523,
0.028217144310474396,
-0.01413219515234232,
0.008876753970980644,
0.0015310103772208095,
0.07484791427850723,
0.032906338572502136,
0.047483623027801514,
-0.05243506655097008,
-0.03373958170413971,
0.04278326779603958,
0.014378420077264309,
0.05403176322579384,
0.07016629725694656,
-0.021531235426664352,
0.04425247013568878,
0.04140881448984146,
0.037042420357465744,
0.03633444756269455,
0.06534627825021744,
0.010667342692613602,
-0.023552821949124336,
0.011246603913605213,
0.0730811282992363,
0.06855793297290802,
-0.0006621578359045088,
1.0054491747422586e-33,
0.08915101736783981,
0.04858594760298729,
0.05164782702922821,
0.04850470647215843,
0.09801927953958511,
-0.03666375204920769,
0.0037722119595855474,
0.029057715088129044,
-0.04914845898747444,
-0.0506511852145195,
-0.06585879623889923,
0.10645005106925964,
-0.012007045559585094,
0.0914354994893074,
-0.01054519321769476,
-0.039407648146152496,
-0.02377963997423649,
0.06617660820484161,
-0.04627905786037445,
-0.08718565106391907,
0.02214319445192814,
-0.0014866848941892385,
0.03340356796979904,
0.016107207164168358,
-0.03967902064323425,
-0.023884152993559837,
0.011826645582914352,
-0.06433282792568207,
0.029186462983489037,
0.008037375286221504,
-0.08196058869361877,
0.020666001364588737,
-0.04401280730962753,
0.09000468254089355,
-0.07351338118314743,
0.009514975361526012,
0.0017283326014876366,
-0.09564752876758575,
0.07439751923084259,
0.04637538269162178,
-0.015103685669600964,
0.01622912473976612,
0.04433361068367958,
-0.05793480947613716,
-0.03843313083052635,
-0.03599239140748978,
-0.07452762126922607,
0.014031056314706802,
-0.061133868992328644,
-0.111829474568367,
0.01384420320391655,
0.04638412594795227,
0.05180644616484642,
0.019149545580148697,
0.03160626441240311,
0.0012597362510859966,
0.01162679772824049,
0.017107976600527763,
0.07423391193151474,
-0.004025632049888372,
0.033899690955877304,
0.0071520376950502396,
-0.010661900043487549,
0.12548954784870148,
0.04226218909025192,
0.04654974862933159,
-0.024527186527848244,
0.0712616965174675,
0.1468602567911148,
-0.019616330042481422,
-0.014983120374381542,
0.0770459696650505,
-0.04157962650060654,
-0.04925745353102684,
0.00960589200258255,
-0.017027603462338448,
0.0159149207174778,
-0.12848739326000214,
-0.0007419848698191345,
0.07325379550457001,
-0.0794723853468895,
-0.03910686448216438,
-0.06599894911050797,
-0.020029699429869652,
-0.024518385529518127,
-0.04844767600297928,
0.006790585815906525,
-0.00946834683418274,
-0.01704780012369156,
0.028267664834856987,
-0.05445657670497894,
-0.02560797519981861,
0.010742110200226307,
-0.016472702845931053,
-0.022190310060977936,
-1.9086013209461922e-33,
-0.029745668172836304,
-0.056515760719776154,
-0.019200915470719337,
0.02457331120967865,
-0.01682271622121334,
-0.09770575165748596,
-0.05120999738574028,
-0.07212948054075241,
-0.05939849093556404,
-0.0056030903942883015,
-0.02606855146586895,
-0.03642481192946434,
0.056712038815021515,
-0.047226037830114365,
0.0339028537273407,
-0.09389698505401611,
0.04564462974667549,
0.032217659056186676,
0.07162394374608994,
-0.007940571755170822,
0.07433748245239258,
0.03482314944267273,
-0.045537907630205154,
-0.015158021822571754,
0.04635155200958252,
-0.008593949489295483,
-0.04171086102724075,
-0.000042150924855377525,
0.031678456813097,
0.07378114759922028,
-0.04834342375397682,
-0.061790402978658676,
-0.004264077637344599,
-0.06715968251228333,
-0.014166350476443768,
0.1376982480287552,
-0.03145173192024231,
-0.03771728649735451,
-0.08226565271615982,
0.04075796902179718,
0.027056369930505753,
-0.11697731167078018,
-0.10745780169963837,
-0.017331307753920555,
-0.003499693935737014,
0.05593038350343704,
-0.0699828565120697,
0.03377870097756386,
-0.06511324644088745,
-0.0038604638539254665,
0.04255690798163414,
0.016399875283241272,
0.01877680979669094,
-0.10949508100748062,
-0.058325450867414474,
0.03731352835893631,
0.053171735256910324,
0.010986346751451492,
0.0007837301818653941,
-0.02850828319787979,
-0.0669940859079361,
-0.008355735801160336,
0.021160518750548363,
-0.008144368417561054,
-0.04496828839182854,
-0.016524624079465866,
-0.037073247134685516,
0.0473390594124794,
0.005918844603002071,
0.006529828533530235,
0.029489193111658096,
-0.009750846773386002,
-0.0565924271941185,
0.013112579472362995,
0.030844002962112427,
-0.009676752611994743,
-0.02053959295153618,
-0.05305255576968193,
-0.006732056848704815,
-0.043387580662965775,
-0.039964593946933746,
0.06726107746362686,
0.030853411182761192,
0.08586811274290085,
0.022315165027976036,
0.00820457935333252,
-0.015325148589909077,
0.15576031804084778,
-0.0009588098619133234,
0.12328589707612991,
0.019008835777640343,
-0.05617307499051094,
-0.026315344497561455,
0.006135902367532253,
-0.14018435776233673,
-4.23887804856804e-8,
-0.11946797370910645,
0.13068193197250366,
0.06259716302156448,
0.03748512640595436,
0.042672447860240936,
-0.06958481669425964,
0.02335526794195175,
0.03159090876579285,
-0.009741449728608131,
0.04560050368309021,
0.07656006515026093,
-0.018319150432944298,
0.03679857403039932,
-0.006123958621174097,
0.032522328197956085,
0.029324369505047798,
-0.01864543929696083,
-0.07844264805316925,
-0.03282386064529419,
-0.08711618185043335,
0.007572794333100319,
0.0020917656365782022,
-0.06227680295705795,
0.014062168076634407,
-0.006002555135637522,
-0.030446255579590797,
-0.08798371255397797,
0.07477260380983353,
-0.03669806197285652,
-0.01036901492625475,
-0.023525435477495193,
0.020478904247283936,
-0.00211421144194901,
-0.0007728885975666344,
0.058946963399648666,
0.06389676779508591,
-0.06287164986133575,
-0.018819158896803856,
-0.003094422398135066,
-0.11341637372970581,
-0.00010656182712409645,
0.07292480766773224,
0.03012162446975708,
0.01082503143697977,
0.04276350885629654,
-0.014552936889231205,
-0.07205744832754135,
-0.053169600665569305,
0.03640686348080635,
-0.11204834282398224,
-0.04641370847821236,
-0.034159693866968155,
-0.03385493904352188,
0.06337317079305649,
0.07855331897735596,
0.023895110934972763,
-0.012256801128387451,
-0.07056364417076111,
0.09022960811853409,
0.013787663541734219,
0.013624142855405807,
0.05381226912140846,
-0.03844699263572693,
-0.04292530566453934
] | https://github.com/openbmb/chatdev | 26,752 | 0 | 0 | 0 |
Lotus: Diffusion-based Visual Foundation Model for High-quality Dense Prediction | https://openreview.net/forum?id=stK7iOPH9Q | [
"Jing He",
"Haodong LI",
"Wei Yin",
"Yixun Liang",
"Leheng Li",
"Kaiqiang Zhou",
"Hongbo Zhang",
"Bingbing Liu",
"Ying-Cong Chen"
] | Poster | Leveraging the visual priors of pre-trained text-to-image diffusion models offers a promising solution to enhance zero-shot generalization in dense prediction tasks. However, existing methods often uncritically use the original diffusion formulation, which may not be optimal due to the fundamental differences between dense prediction and image generation. In this paper, we provide a systemic analysis of the diffusion formulation for the dense prediction, focusing on both quality and efficiency. And we find that the original parameterization type for image generation, which learns to predict noise, is harmful for dense prediction; the multi-step noising/denoising diffusion process is also unnecessary and challenging to optimize. Based on these insights, we introduce $\textbf{Lotus}$, a diffusion-based visual foundation model with a simple yet effective adaptation protocol for dense prediction. Specifically, Lotus is trained to directly predict annotations instead of noise, thereby avoiding harmful variance. We also reformulate the diffusion process into a single-step procedure, simplifying optimization and significantly boosting inference speed. Additionally, we introduce a novel tuning strategy called detail preserver, which achieves more accurate and fine-grained predictions. Without scaling up the training data or model capacity, Lotus achieves SoTA performance in zero-shot depth and normal estimation across various datasets. It also enhances efficiency, being significantly faster than most existing diffusion-based methods. Lotus' superior quality and efficiency also enable a wide range of practical applications, such as joint estimation, single/multi-view 3D reconstruction, etc. | Diffusion Models; Dense Prediction; Monocular Depth Estimation; Monocular Normal Estimation | Based on pre-trained Stable Diffusion, Lotus delivers SoTA performance on monocular depth & normal estimation with simple yet effective fine-tuning protocol. | 13 | 2409.18124 | [
0.034689873456954956,
-0.1136474758386612,
0.04812193661928177,
0.06434965133666992,
0.12781476974487305,
-0.015041869133710861,
-0.02057083323597908,
-0.06905803084373474,
0.022503435611724854,
-0.07557657361030579,
-0.038023535162210464,
-0.05463312938809395,
-0.0021822659764438868,
0.04899267852306366,
-0.05847499519586563,
0.012817595154047012,
0.0857890248298645,
0.06482052057981491,
-0.10131355375051498,
-0.07079645246267319,
-0.0539998859167099,
-0.04828091338276863,
0.010932385921478271,
-0.009049241431057453,
0.10184983909130096,
0.025523092597723007,
-0.002849364187568426,
-0.031107814982533455,
0.017492011189460754,
-0.0497421957552433,
0.0029466883279383183,
-0.05174080282449722,
-0.03898691013455391,
0.0007874754373915493,
-0.05647331103682518,
0.02197588048875332,
0.0014408602146431804,
0.030909763649106026,
-0.010415023192763329,
-0.013206289149820805,
0.0259095449000597,
0.02886064536869526,
0.028429750353097916,
0.06316569447517395,
0.07128164917230606,
0.05347491800785065,
0.0009141335031017661,
-0.07427963614463806,
0.07490314543247223,
-0.0009527323418296874,
-0.07757674902677536,
-0.06454413384199142,
-0.06610806286334991,
0.005779670085757971,
-0.03995436802506447,
0.0441981665790081,
0.07795960456132889,
0.03856092691421509,
0.02774880826473236,
-0.018598072230815887,
-0.04609716683626175,
-0.061321817338466644,
0.04195624217391014,
0.008916225284337997,
-0.01818048022687435,
0.005161325912922621,
0.0565083883702755,
0.01106902863830328,
-0.033130548894405365,
0.020444946363568306,
-0.04782019183039665,
0.08263349533081055,
-0.028239820152521133,
-0.01388890203088522,
-0.025484662503004074,
0.05273938179016113,
-0.0010256101377308369,
-0.03428763896226883,
0.02893158607184887,
-0.037982191890478134,
0.05903248488903046,
0.04198914393782616,
0.06068343296647072,
-0.05227452889084816,
0.060795798897743225,
0.0201996061950922,
0.03880111500620842,
0.014325197786092758,
-0.03462294489145279,
-0.009210389107465744,
-0.07074106484651566,
0.012657906860113144,
-0.06757058203220367,
0.026271477341651917,
-0.02359410934150219,
0.06318547576665878,
-0.03652503713965416,
-0.09752769768238068,
-0.02312505804002285,
0.07266125082969666,
0.00023845241230446845,
0.019726991653442383,
0.02451305463910103,
-0.029022863134741783,
0.06425375491380692,
-0.017337121069431305,
0.091990165412426,
0.04982917383313179,
0.007371103391051292,
0.008796009235084057,
0.06872048228979111,
0.05153495445847511,
0.011412248946726322,
-0.01017732359468937,
0.01957150734961033,
-0.051124293357133865,
-0.004400440491735935,
-0.018656035885214806,
-0.005735331680625677,
0.03755291923880577,
0.04271624982357025,
-0.036611735820770264,
-0.12238417565822601,
-0.03788583353161812,
-0.030619584023952484,
-0.046650443226099014,
-0.010859673842787743,
1.820841933240666e-36,
-0.009437034837901592,
-0.020138580352067947,
0.139881893992424,
0.04956359043717384,
0.04543885961174965,
-0.016154339537024498,
-0.04334660619497299,
-0.06874218583106995,
-0.015531951561570168,
-0.062103331089019775,
-0.024048905819654465,
-0.09151479601860046,
-0.10170828551054001,
0.15481212735176086,
0.0517861545085907,
-0.07103050500154495,
-0.026639118790626526,
0.04862624406814575,
0.0037639420479536057,
0.007892249152064323,
0.0014719977043569088,
0.020606650039553642,
-0.08702613413333893,
-0.06706536561250687,
0.017209850251674652,
0.016875294968485832,
0.12095078825950623,
-0.07252334803342819,
-0.006595146842300892,
-0.0005506762536242604,
-0.16863514482975006,
-0.006489673163741827,
0.04169490560889244,
0.011053274385631084,
0.020038150250911713,
-0.03677746653556824,
-0.018772458657622337,
-0.007333043031394482,
-0.0007515049073845148,
-0.014834919944405556,
-0.015084672719240189,
0.06830614060163498,
-0.01288554910570383,
-0.02360084466636181,
-0.09353563189506531,
0.009172067977488041,
-0.05862288549542427,
0.06000358238816261,
-0.04570912569761276,
-0.03848044201731682,
0.030086083337664604,
-0.06050918251276016,
-0.004659689497202635,
-0.041727371513843536,
-0.022794334217905998,
-0.0504777617752552,
0.07510879635810852,
-0.05886676162481308,
-0.022638902068138123,
-0.04681456834077835,
0.02550816349685192,
0.022999675944447517,
-0.038203828036785126,
0.08106838911771774,
0.002959019271656871,
-0.0028646057471632957,
0.018313035368919373,
0.0181419737637043,
-0.006278018932789564,
-0.05309418588876724,
-0.02938139997422695,
0.04241214320063591,
0.00906450767070055,
-0.07893005758523941,
0.036761146038770676,
-0.003990798722952604,
0.05989496037364006,
0.0026380312629044056,
0.0710723027586937,
0.03312468156218529,
-0.035752613097429276,
-0.022919533774256706,
-0.007435251958668232,
-0.10057680308818817,
-0.00968505535274744,
0.032259490340948105,
0.03449954837560654,
-0.023562878370285034,
-0.06035042926669121,
-0.06872568279504776,
-0.04574665054678917,
-0.008747407235205173,
0.014606871642172337,
-0.013014007359743118,
-0.011303809471428394,
3.412225001566936e-34,
0.025348175317049026,
0.11260690540075302,
0.013204898685216904,
0.03798820450901985,
-0.03068746067583561,
-0.01362390723079443,
-0.018924962729215622,
0.034792833030223846,
0.0872267559170723,
-0.07147150486707687,
-0.017111605033278465,
-0.024998433887958527,
-0.042435936629772186,
-0.029741575941443443,
0.0010182560654357076,
-0.022906340658664703,
0.027771346271038055,
0.0045602587051689625,
-0.13028275966644287,
0.031359706073999405,
0.001953383209183812,
0.03268478065729141,
-0.050844959914684296,
0.0693555474281311,
-0.02788582630455494,
0.04369297996163368,
-0.02581794001162052,
0.1484665870666504,
-0.0026295455172657967,
-0.04582491144537926,
-0.010736016556620598,
-0.01665802113711834,
0.05432441458106041,
-0.06119277700781822,
-0.07356461882591248,
0.058436471968889236,
0.06834246963262558,
-0.033521559089422226,
-0.048011414706707,
0.10412916541099548,
0.04336058720946312,
-0.03131682053208351,
-0.08043462038040161,
0.013553240336477757,
-0.025184910744428635,
-0.028770269826054573,
-0.03599189966917038,
-0.035260699689388275,
0.04058824107050896,
0.06348711252212524,
-0.02875504456460476,
0.09712004661560059,
-0.0792464166879654,
0.08460021764039993,
-0.05258721858263016,
-0.09507555514574051,
-0.08228342235088348,
0.0009679486392997205,
0.10104969143867493,
-0.028491871431469917,
-0.018749196082353592,
0.020264355465769768,
-0.011174316518008709,
-0.056455932557582855,
-0.028648750856518745,
-0.016658978536725044,
-0.08189427107572556,
-0.023000899702310562,
-0.0010270384373143315,
0.03189825639128685,
-0.033772438764572144,
0.006434329319745302,
0.0283112321048975,
0.027586203068494797,
0.01690247468650341,
-0.06969621777534485,
0.06822190433740616,
-0.00018617267778608948,
0.02678181417286396,
0.027484102174639702,
0.06103364750742912,
0.006122278980910778,
0.04951055720448494,
0.015394133515655994,
0.10038968920707703,
0.08492746949195862,
0.018939385190606117,
0.0003442915331106633,
0.09515220671892166,
0.025687623769044876,
-0.04855890944600105,
-0.010234048590064049,
-0.011790475808084011,
0.03754225745797157,
-0.08415056020021439,
-4.126038888330186e-8,
-0.052996814250946045,
0.03439189866185188,
-0.00048561024595983326,
-0.010043589398264885,
-0.007976186461746693,
-0.06403714418411255,
0.06683871150016785,
0.11997600644826889,
0.034903306514024734,
-0.04339209571480751,
0.07607050240039825,
-0.015210025943815708,
-0.07883918285369873,
-0.07756432145833969,
0.024390162900090218,
0.08946198225021362,
0.035972412675619125,
0.034235630184412,
0.030183790251612663,
0.015676027163863182,
-0.03889930993318558,
0.022659819573163986,
-0.0022880700416862965,
-0.03781402111053467,
0.08367052674293518,
-0.04219232499599457,
0.001680721645243466,
0.07237137109041214,
0.08527836948633194,
0.006519664544612169,
0.04223961383104324,
0.04324078559875488,
0.03277600556612015,
-0.034951094537973404,
-0.04497867822647095,
0.08568613976240158,
0.0874122753739357,
0.030754614621400833,
-0.027989346534013748,
-0.04380648955702782,
-0.01771574281156063,
0.07264436781406403,
0.01014974620193243,
-0.02492697536945343,
-0.024397818371653557,
-0.007371844723820686,
0.07277395576238632,
-0.06883527338504791,
0.03944456949830055,
0.009006824344396591,
-0.013979516923427582,
-0.024422699585556984,
0.014157976023852825,
0.07120651006698608,
0.035848941653966904,
-0.05331984534859657,
0.04318246990442276,
0.058410029858350754,
0.044227082282304764,
0.10090601444244385,
0.011052128858864307,
-0.007954435423016548,
0.06234326586127281,
0.0898110643029213
] | 0 | 0 | 0 | 0 |
|
ARB-LLM: Alternating Refined Binarizations for Large Language Models | https://openreview.net/forum?id=ZU8OdDLTts | [
"Zhiteng Li",
"Xianglong Yan",
"Tianao Zhang",
"Haotong Qin",
"Dong Xie",
"Jiang Tian",
"zhongchao shi",
"Linghe Kong",
"Yulun Zhang",
"Xiaokang Yang"
] | Poster | Large Language Models (LLMs) have greatly pushed forward advancements in natural language processing, yet their high memory and computational demands hinder practical deployment. Binarization, as an effective compression technique, can shrink model weights to just 1 bit, significantly reducing the high demands on computation and memory. However, current binarization methods struggle to narrow the distribution gap between binarized and full-precision weights, while also overlooking the column deviation in LLM weight distribution. To tackle these issues, we propose ARB-LLM, a novel 1-bit post-training quantization (PTQ) technique tailored for LLMs. To narrow the distribution shift between binarized and full-precision weights, we first design an alternating refined binarization (ARB) algorithm to progressively update the binarization parameters, which significantly reduces the quantization error. Moreover, considering the pivot role of calibration data and the column deviation in LLM weights, we further extend ARB to ARB-X and ARB-RC. In addition, we refine the weight partition strategy with column-group bitmap (CGB), which further enhance performance. Equipping ARB-X and ARB-RC with CGB, we obtain ARB-LLM$_{\text{X}}$ and ARB-LLM$ _{\text{RC}} $ respectively, which significantly outperform state-of-the-art (SOTA) binarization methods for LLMs.
As a binary PTQ method, our ARB-LLM$ _{\text{RC}} $ is the first to surpass FP16 models of the same size. Code: https://github.com/ZHITENGLI/ARB-LLM. | Binarization, LLM | We propose ARB-LLM, which alternately updates binarization parameters to reduce quantization error during the binarization of large language models (LLMs). | 12 | null | [
0.01377187855541706,
-0.03836969658732414,
-0.042877402156591415,
-0.010964175686240196,
-0.014514478854835033,
0.017991824075579643,
-0.02123463712632656,
-0.03642008453607559,
0.03206954151391983,
-0.04094961658120155,
-0.020545311272144318,
-0.011334403418004513,
-0.015125260688364506,
0.0016156429192051291,
-0.010930569842457771,
0.031198877841234207,
0.013397249393165112,
0.10466212779283524,
-0.08578427135944366,
-0.07567290961742401,
0.0235279630869627,
-0.012309950776398182,
-0.042488958686590195,
-0.007913310080766678,
0.10462191700935364,
0.01732327975332737,
-0.023397114127874374,
0.01495425496250391,
0.02099190279841423,
-0.061361998319625854,
0.037265926599502563,
0.02779700793325901,
0.07914125919342041,
0.01775822788476944,
-0.060153305530548096,
0.0473807118833065,
-0.0023373370058834553,
0.008183225058019161,
0.06048683449625969,
-0.07069657742977142,
-0.04879377782344818,
-0.031330786645412445,
0.0496617928147316,
-0.013809112831950188,
0.08699186891317368,
0.043855562806129456,
0.05431685224175453,
0.05923464894294739,
0.018617628142237663,
-0.009288634173572063,
0.05689820647239685,
0.08145461976528168,
-0.03399767726659775,
0.09468089789152145,
-0.11319281905889511,
-0.09333968907594681,
0.0330171138048172,
0.022266659885644913,
-0.04658912494778633,
-0.04349914565682411,
-0.10429280996322632,
-0.026095852255821228,
0.021297911182045937,
0.030494675040245056,
-0.03749221935868263,
-0.034782689064741135,
0.024456234648823738,
-0.009633002802729607,
0.01926334947347641,
0.008410821668803692,
-0.01584801822900772,
-0.027287399396300316,
-0.08553434163331985,
0.10479826480150223,
-0.08366389572620392,
0.09505154192447662,
0.049392055720090866,
0.01266450434923172,
0.11752548068761826,
-0.0014668010408058763,
-0.01637386344373226,
0.010188438929617405,
0.14371395111083984,
-0.026463588699698448,
0.07216481119394302,
-0.01537417434155941,
-0.025115039199590683,
0.047071997076272964,
-0.02718663215637207,
0.03962313011288643,
0.0008076280355453491,
0.027272319421172142,
0.060775235295295715,
0.0459442064166069,
0.00011910582543350756,
-0.024738645181059837,
0.011514678597450256,
0.01801941730082035,
-0.04751897603273392,
0.11540259420871735,
0.05484074354171753,
0.06827212125062943,
-0.0190581027418375,
-0.020766358822584152,
-0.010996640659868717,
-0.044243138283491135,
0.028324874117970467,
0.04491782560944557,
0.06362903118133545,
-0.06537759304046631,
0.03281855955719948,
0.05337582901120186,
-0.043923553079366684,
-0.031055236235260963,
0.02727961726486683,
0.018119579181075096,
-0.020841538906097412,
-0.0015965739730745554,
-0.005955622997134924,
0.016328996047377586,
-0.10595015436410904,
0.021691234782338142,
-0.0946078822016716,
0.016783032566308975,
0.004029151983559132,
-0.052521660923957825,
-0.09842979162931442,
3.120414039105073e-33,
0.054545771330595016,
0.054173968732357025,
0.010393660515546799,
-0.03347349539399147,
0.06396768987178802,
0.051068395376205444,
0.04010677710175514,
0.0340053029358387,
-0.020795907825231552,
0.014931678771972656,
0.016625868156552315,
-0.01176100317388773,
-0.02127372846007347,
0.023919690400362015,
0.039516154676675797,
-0.008851864375174046,
-0.06098957359790802,
-0.0065891132690012455,
-0.05571100115776062,
-0.09432672709226608,
0.014213998802006245,
-0.033289507031440735,
0.024597181007266045,
0.03943532705307007,
0.018409578129649162,
0.008512841537594795,
0.11419829726219177,
-0.02666146121919155,
-0.06811528652906418,
0.05337121710181236,
-0.10865852981805801,
-0.034944791346788406,
0.01134840864688158,
-0.08440377563238144,
0.04154156893491745,
-0.09672529250383377,
-0.01938180811703205,
-0.0056603169068694115,
0.026180773973464966,
-0.025618983432650566,
0.017995450645685196,
0.03343285620212555,
0.043562181293964386,
-0.07915792614221573,
-0.09658975154161453,
-0.0026846746914088726,
0.0017679862212389708,
0.05083014443516731,
-0.05548799782991409,
-0.05704478174448013,
0.07464385032653809,
-0.00923349428921938,
-0.06308026611804962,
0.0738534927368164,
0.06020301952958107,
0.00014872541942168027,
0.0908803939819336,
0.017557863146066666,
0.009329080581665039,
0.023735247552394867,
0.001941713155247271,
0.06663165986537933,
0.026938263326883316,
0.10314331948757172,
0.021972322836518288,
0.005899196490645409,
-0.009720641188323498,
0.055984653532505035,
-0.033247023820877075,
0.016751930117607117,
-0.07643213868141174,
-0.010434141382575035,
0.051322199404239655,
-0.006099850870668888,
-0.00008292723214253783,
-0.03724977746605873,
0.10680391639471054,
-0.05229053646326065,
-0.007069704122841358,
-0.00862403679639101,
0.020670538768172264,
0.03629831224679947,
0.05894826352596283,
-0.06471244245767593,
-0.06202257052063942,
-0.10483752191066742,
0.035833779722452164,
-0.005938208196312189,
-0.04944531247019768,
-0.046631090342998505,
-0.024471497163176537,
-0.013936594128608704,
-0.032047759741544724,
-0.025221584364771843,
-0.04687994346022606,
-2.147824523243105e-33,
-0.05778305605053902,
0.0036431760527193546,
-0.053191497921943665,
0.10899645835161209,
-0.07556673884391785,
-0.0936913937330246,
0.05113577842712402,
-0.018667951226234436,
0.03477367386221886,
-0.032231654971838,
-0.03431423380970955,
-0.04265806823968887,
0.017260486260056496,
0.04284209758043289,
0.03190551698207855,
-0.02895703725516796,
-0.03992817923426628,
0.11736290901899338,
-0.03341394662857056,
0.10437910258769989,
-0.06452315300703049,
0.08670277893543243,
-0.050874125212430954,
-0.00990564376115799,
-0.018329087644815445,
-0.00869995541870594,
-0.05351565405726433,
0.05379148945212364,
0.02413414791226387,
0.011056311428546906,
-0.05294368788599968,
0.007658598944544792,
-0.03464651480317116,
-0.0764770433306694,
-0.09034773707389832,
-0.00034906432847492397,
0.06083178520202637,
0.012785366736352444,
0.06021775305271149,
0.051260512322187424,
0.06582610309123993,
-0.06472567468881607,
-0.11401308327913284,
-0.006337967701256275,
0.02110208384692669,
-0.007879572920501232,
-0.11547736823558807,
0.005756190977990627,
-0.01570768468081951,
-0.02917044796049595,
-0.00357182789593935,
0.04954129084944725,
0.024461187422275543,
0.10394129902124405,
0.009918343275785446,
-0.09528212249279022,
-0.09709721058607101,
-0.01722637005150318,
0.008442648686468601,
-0.036395955830812454,
-0.09011675417423248,
-0.010274877771735191,
0.07893651723861694,
-0.1147390753030777,
0.027469512075185776,
-0.047121044248342514,
-0.02889108657836914,
-0.0006869153003208339,
-0.007867291569709778,
-0.002855663187801838,
0.06300276517868042,
-0.01973991096019745,
0.025127042084932327,
0.07062084227800369,
-0.010792955756187439,
-0.02998870424926281,
0.024544350802898407,
-0.026451939716935158,
-0.014515958726406097,
-0.013101580552756786,
-0.011929870583117008,
0.0421626903116703,
0.011678122915327549,
0.04767170175909996,
0.0318000502884388,
0.06825683265924454,
0.006864729803055525,
0.050246890634298325,
-0.0013527923729270697,
0.06567411124706268,
-0.05661330372095108,
0.09502658993005753,
0.08886300772428513,
0.019066864624619484,
-0.07973560690879822,
-4.258417263258707e-8,
-0.036081522703170776,
0.02446806989610195,
-0.04478057473897934,
-0.011656663380563259,
0.058129653334617615,
-0.0019112128065899014,
-0.04685252159833908,
0.07617124170064926,
0.04162847250699997,
-0.09262709319591522,
0.08804517984390259,
-0.00043747469317168,
-0.07028757035732269,
-0.06010841578245163,
-0.059550706297159195,
0.10531303286552429,
0.0030966538470238447,
-0.026795143261551857,
0.03197086602449417,
-0.013814181089401245,
-0.02386298216879368,
0.0754896029829979,
0.035399872809648514,
-0.008590775541961193,
0.027406061068177223,
-0.05584435537457466,
-0.028587639331817627,
0.09239786863327026,
0.10128853470087051,
0.013988783583045006,
0.00016392098041251302,
0.005670185666531324,
0.012389762327075005,
-0.0294124074280262,
0.08001085370779037,
0.042245130985975266,
0.008078082464635372,
0.019577862694859505,
-0.015281436033546925,
-0.019796373322606087,
0.03796916455030441,
-0.028746377676725388,
-0.07640343904495239,
0.04148418456315994,
0.0001343272306257859,
-0.05780300498008728,
-0.006250468082726002,
-0.012672564014792442,
-0.029356975108385086,
-0.0952596664428711,
0.06370139867067337,
0.05717615410685539,
-0.042449887841939926,
-0.0015270733274519444,
0.008420991711318493,
0.06329351663589478,
-0.025794005021452904,
-0.03929326310753822,
0.04378950595855713,
0.04435862600803375,
-0.008430150337517262,
-0.05646849423646927,
-0.09346126019954681,
-0.027235379442572594
] | 0 | 0 | 0 | 0 |
|
Understanding and Enhancing the Transferability of Jailbreaking Attacks | https://openreview.net/forum?id=asR9FVd4eL | [
"Runqi Lin",
"Bo Han",
"Fengwang Li",
"Tongliang Liu"
] | Poster | Jailbreaking attacks can effectively manipulate open-source large language models (LLMs) to produce harmful responses. However, these attacks exhibit limited transferability, failing to disrupt proprietary LLMs consistently. To reliably identify vulnerabilities in proprietary LLMs, this work investigates the transferability of jailbreaking attacks by analysing their impact on the model's intent perception. By incorporating adversarial sequences, these attacks can redirect the source LLM's focus away from malicious-intent tokens in the original input, thereby obstructing the model's intent recognition and eliciting harmful responses. Nevertheless, these adversarial sequences fail to mislead the target LLM's intent perception, allowing the target LLM to refocus on malicious-intent tokens and abstain from responding. Our analysis further reveals the inherent $\textit{distributional dependency}$ within the generated adversarial sequences, whose effectiveness stems from overfitting the source LLM's parameters, resulting in limited transferability to target LLMs. To this end, we propose the Perceived-importance Flatten (PiF) method, which uniformly disperses the model's focus across neutral-intent tokens in the original input, thus obscuring malicious-intent tokens without relying on overfitted adversarial sequences. Extensive experiments demonstrate that PiF provides an effective and efficient red-teaming evaluation for proprietary LLMs. | Jailbreaking Attack, Black-box Transferable Attack, Large Language Model, Red-teaming Evaluation | null | 10 | 2502.03052 | [
-0.0706317350268364,
-0.09190788865089417,
0.0147081995382905,
-0.029161591082811356,
0.09996513277292252,
0.04832601547241211,
0.020331084728240967,
-0.031076103448867798,
0.04462905600667,
-0.06502930819988251,
-0.04387656971812248,
-0.0367627814412117,
0.013082238845527172,
-0.024702195078134537,
-0.009657037444412708,
0.019342312589287758,
0.10327130556106567,
0.12720271944999695,
-0.08047422021627426,
0.004733419977128506,
0.08922747522592545,
0.00671918410807848,
-0.0003669393190648407,
-0.035285834223032,
-0.0054341708309948444,
0.01108659990131855,
-0.0759085938334465,
-0.03061068058013916,
0.011179756373167038,
-0.024185042828321457,
0.047977425158023834,
0.05003936588764191,
-0.011247623711824417,
0.004886076785624027,
-0.08805350959300995,
0.002572072669863701,
-0.020587991923093796,
0.008361145853996277,
0.05372603237628937,
-0.09728682786226273,
0.0043908581137657166,
0.0614314042031765,
-0.0036678574979305267,
0.052078548818826675,
0.0005145111354067922,
-0.07159765064716339,
0.021037407219409943,
-0.0032295426353812218,
0.008611819706857204,
-0.02594749815762043,
-0.028133146464824677,
0.018829112872481346,
-0.0163103137165308,
0.10847503691911697,
-0.06745816767215729,
-0.08340653777122498,
0.0017528310418128967,
0.10920486599206924,
0.034965284168720245,
0.0441826656460762,
-0.035269029438495636,
0.0014119510306045413,
-0.011707770638167858,
0.018391795456409454,
-0.0037658363580703735,
-0.05177721753716469,
0.04226907342672348,
0.06783898174762726,
-0.07870978116989136,
0.07023803144693375,
0.008839672431349754,
0.03369705006480217,
-0.07565223425626755,
-0.026889996603131294,
0.004937945865094662,
0.08102589100599289,
-0.01800396665930748,
0.07495096325874329,
-0.014484312385320663,
-0.09311183542013168,
0.0522066093981266,
-0.009930793195962906,
0.06500310450792313,
0.013414676301181316,
0.004230928607285023,
0.07150162756443024,
-0.052596550434827805,
0.03430195152759552,
0.12631821632385254,
0.054895587265491486,
-0.036458224058151245,
0.013500948436558247,
0.05906621739268303,
0.019675597548484802,
0.08939646184444427,
-0.008304130285978317,
-0.04363277181982994,
-0.084093376994133,
-0.06330178678035736,
0.09334589540958405,
0.009408928453922272,
0.029905179515480995,
0.010996345430612564,
-0.05816854536533356,
0.07546861469745636,
0.0008478273521177471,
0.0006513951229862869,
-0.022358957678079605,
-0.00042106289765797555,
-0.005245030391961336,
0.051215991377830505,
0.06401219218969345,
0.06540986150503159,
-0.02375326119363308,
0.008614357560873032,
-0.028684698045253754,
-0.05631816387176514,
0.0044633192010223866,
0.0045742313377559185,
0.01678205095231533,
0.0077525293454527855,
0.06515011936426163,
0.02340935356914997,
0.026552099734544754,
-0.06264115869998932,
-0.06386862695217133,
-0.09791494160890579,
1.2522922436618884e-33,
-0.0014962995192036033,
0.00768644642084837,
-0.05297986790537834,
0.021914198994636536,
0.025421645492315292,
-0.05503324791789055,
0.003925824072211981,
0.09365707635879517,
0.0038407070096582174,
-0.0028774875681847334,
-0.01520626712590456,
0.010593850165605545,
-0.02011844515800476,
0.0783727839589119,
0.057633038610219955,
0.07637188583612442,
-0.024199923500418663,
0.0026107218582183123,
0.02314993180334568,
-0.01804766058921814,
-0.005742594599723816,
-0.00795581005513668,
0.012830127961933613,
-0.02268611639738083,
0.01671707071363926,
-0.014382916502654552,
0.016254935413599014,
-0.03870859369635582,
-0.04802609980106354,
0.03669658675789833,
-0.09788266569375992,
-0.03087303787469864,
0.026197822764515877,
0.00477428175508976,
0.028468040749430656,
-0.03528624773025513,
-0.024796275421977043,
-0.05627466365695,
0.023333633318543434,
-0.036832623183727264,
-0.024214349687099457,
0.02741849236190319,
-0.02411569096148014,
0.0716976597905159,
-0.00828291941434145,
-0.02778758481144905,
-0.07884545624256134,
0.004559260327368975,
-0.007823606953024864,
-0.06259570270776749,
-0.026978757232427597,
0.02329181134700775,
0.026530016213655472,
-0.03183549642562866,
-0.15013183653354645,
-0.013942623510956764,
-0.055089447647333145,
0.014034871011972427,
-0.01988544501364231,
0.09026459604501724,
-0.04519053176045418,
0.05964554846286774,
-0.006913045886904001,
0.053395744413137436,
0.020683206617832184,
-0.05928237363696098,
-0.07451507449150085,
0.0292322039604187,
-0.01114736869931221,
0.028596417978405952,
-0.06491842120885849,
0.0077807605266571045,
-0.05749257653951645,
0.016915643587708473,
-0.05322416126728058,
-0.04376358538866043,
0.08284042775630951,
0.003014837857335806,
0.04513856768608093,
-0.01710539683699608,
0.046251051127910614,
-0.016292808577418327,
0.10441767424345016,
-0.017407089471817017,
-0.11019746959209442,
-0.033376481384038925,
0.004538927227258682,
-0.10508005321025848,
-0.02308226376771927,
0.04036416485905647,
-0.013583945110440254,
-0.03770655021071434,
-0.03200308978557587,
0.03407253324985504,
-0.010241839103400707,
-2.665274197739199e-33,
-0.11786411702632904,
0.0020005994010716677,
-0.044801775366067886,
0.05998170003294945,
-0.08532115817070007,
-0.018757672980427742,
-0.04979081451892853,
0.06887402385473251,
-0.047530919313430786,
0.027011197060346603,
-0.13225583732128143,
-0.008636931888759136,
0.0913843959569931,
0.007610350381582975,
0.02155345119535923,
-0.0697927251458168,
0.03244907408952713,
0.0686473399400711,
-0.07770780473947525,
0.04816741868853569,
0.029827890917658806,
0.07416924089193344,
-0.06284089386463165,
0.012170626781880856,
-0.026636222377419472,
-0.010196039453148842,
-0.05698973685503006,
0.04962804540991783,
0.008369820192456245,
0.017255187034606934,
-0.015838410705327988,
0.055027302354574203,
-0.0557473823428154,
-0.0050806524232029915,
-0.06963830441236496,
0.06875894218683243,
0.06732968986034393,
0.03207334503531456,
-0.03847002238035202,
0.027685031294822693,
0.08247124403715134,
-0.021723421290516853,
-0.08758172392845154,
-0.02936282753944397,
0.004329513292759657,
-0.05152731388807297,
-0.031680069863796234,
0.0083867646753788,
0.08650828897953033,
-0.08690910786390305,
0.0548153892159462,
0.04297627881169319,
0.02525719255208969,
0.05677009001374245,
-0.00324838375672698,
-0.06672888994216919,
0.05786101892590523,
-0.02740650624036789,
0.07263008505105972,
-0.035276710987091064,
0.00012794444046448916,
-0.016615428030490875,
0.008658302947878838,
-0.1038266271352768,
0.07051125913858414,
-0.05132574588060379,
-0.003979051951318979,
0.018021317198872566,
0.03308435156941414,
-0.04583648964762688,
0.09443502873182297,
-0.031363990157842636,
-0.0024860387202352285,
0.03336512669920921,
-0.04267919808626175,
-0.03464045003056526,
0.00016023953503463417,
-0.11156231164932251,
-0.0947817713022232,
-0.007993439212441444,
0.09186064451932907,
-0.0035646571777760983,
0.029425637796521187,
0.0038835646118968725,
0.06246575340628624,
0.017815662547945976,
0.09746517241001129,
0.03514210879802704,
-0.006292496342211962,
0.03422367572784424,
0.059166692197322845,
0.03714318573474884,
-0.10330687463283539,
0.07693227380514145,
-0.0031215392518788576,
-4.196728298211383e-8,
0.014350282028317451,
-0.03530336543917656,
-0.03400126099586487,
0.042029812932014465,
0.02731756493449211,
0.0358944870531559,
-0.09949444979429245,
0.03193947672843933,
-0.0009762760018929839,
-0.0994296744465828,
-0.0011323225917294621,
0.015271523967385292,
-0.04682593420147896,
-0.05236067995429039,
-0.04304305091500282,
0.09671462327241898,
-0.045801542699337006,
-0.008084430359303951,
-0.03357522189617157,
0.02352984994649887,
-0.02051473967730999,
0.02876306138932705,
-0.007779624313116074,
-0.03902614861726761,
0.027862589806318283,
-0.0907711610198021,
-0.00352925481274724,
0.028172513470053673,
0.07913704961538315,
0.062148984521627426,
0.031216217204928398,
-0.018235761672258377,
-0.09377492964267731,
-0.04801769182085991,
-0.07321104407310486,
0.13730403780937195,
-0.015893621370196342,
-0.06172588840126991,
0.05502210557460785,
0.015970123931765556,
-0.001349488040432334,
0.046579357236623764,
-0.02294858731329441,
0.023019518703222275,
-0.0030753070022910833,
-0.05258499085903168,
-0.006210291292518377,
-0.13099829852581024,
0.10730228573083878,
0.015303990803658962,
0.07146476209163666,
-0.04687554016709328,
0.01053746696561575,
0.02966476045548916,
0.025012196972966194,
-0.006994554307311773,
-0.03280150517821312,
-0.0066897813230752945,
0.014578769914805889,
0.0774451270699501,
0.07711226493120193,
0.00732412887737155,
0.04963850975036621,
-0.015209364704787731
] | https://github.com/tmllab/2025_ICLR_PiF | 23 | 0 | 0 | 0 |
FLOPS: Forward Learning with OPtimal Sampling | https://openreview.net/forum?id=z1nSpA2dAW | [
"Tao Ren",
"Zishi Zhang",
"Jinyang Jiang",
"Guanghao Li",
"Zeliang Zhang",
"Mingqian Feng",
"Yijie Peng"
] | Poster | Given the limitations of backpropagation, perturbation-based gradient computation methods have recently gained focus for learning with only forward passes, also referred to as queries. Conventional forward learning consumes enormous queries on each data point for accurate gradient estimation through Monte Carlo sampling, which hinders the scalability of those algorithms. However, not all data points deserve equal queries for gradient estimation. In this paper, we study the problem of improving the forward learning efficiency from a novel perspective: how to reduce the gradient estimation variance with minimum cost? For this, we allocate the optimal number of queries within a set budget during training to balance estimation accuracy and computational efficiency. Specifically, with a simplified proxy objective and a reparameterization technique, we derive a novel plug-and-play query allocator with minimal parameters. Theoretical results are carried out to verify its optimality. We conduct extensive experiments for fine-tuning Vision Transformers on various datasets and further deploy the allocator to two black-box applications: prompt tuning and multimodal alignment for foundation models. All findings demonstrate that our proposed allocator significantly enhances the scalability of forward-learning algorithms, paving the way for real-world applications. The implementation is available at https://github.com/RTkenny/FLOPS-Forward-Learning-with-OPtimal-Sampling. | stochastic optimization, zeroth-order optimization, bp-free training | Design an optimal query allocator for forward learning | 8 | 2410.05966 | [
-0.04596570134162903,
-0.11229798197746277,
-0.0308977197855711,
0.07836222648620605,
0.022895988076925278,
0.02182024158537388,
-0.009066248312592506,
-0.02548227086663246,
-0.001478097285144031,
-0.022879809141159058,
-0.10958167165517807,
-0.03339719399809837,
0.016845159232616425,
-0.04730096086859703,
-0.05172232165932655,
0.046512141823768616,
0.06703249365091324,
0.16152824461460114,
-0.045975785702466965,
-0.055403538048267365,
-0.043846432119607925,
-0.005248670000582933,
0.05895015969872475,
-0.013074839487671852,
-0.006721390876919031,
0.04156256094574928,
-0.038170430809259415,
-0.07244621217250824,
-0.03655260056257248,
-0.023539135232567787,
-0.028239168226718903,
-0.07180602103471756,
-0.056836433708667755,
0.012071090750396252,
-0.08351241797208786,
0.009553611278533936,
-0.09384553134441376,
0.039362240582704544,
-0.04274518042802811,
0.014420131221413612,
0.017726769670844078,
0.01801343634724617,
-0.001979321241378784,
0.030760882422327995,
0.0874854326248169,
0.05982537195086479,
0.03908968344330788,
-0.01947603188455105,
0.02247445657849312,
0.0022076419554650784,
-0.021824650466442108,
-0.05145905166864395,
-0.08114715665578842,
0.03837380185723305,
-0.038538336753845215,
0.06171889975667,
0.03560568392276764,
-0.005134094972163439,
-0.0038399905897676945,
0.06768066436052322,
-0.01849334128201008,
-0.13449817895889282,
0.027089953422546387,
-0.00966720748692751,
-0.04282359033823013,
-0.058266349136829376,
0.06155826896429062,
0.015658942982554436,
0.03708745539188385,
-0.009492747485637665,
-0.07502350211143494,
0.14106185734272003,
-0.003399953478947282,
-0.0075223068706691265,
0.0032715427223592997,
0.07519803941249847,
0.08495744317770004,
-0.022683432325720787,
0.03155011311173439,
-0.10243505984544754,
0.04646940529346466,
0.012922082096338272,
0.029624879360198975,
0.002301894361153245,
0.16369201242923737,
-0.03978235274553299,
-0.02593143656849861,
0.009195702150464058,
0.058465585112571716,
-0.04270434379577637,
-0.018240483477711678,
0.00011195730621693656,
-0.05545476824045181,
-0.002772144740447402,
-0.007146028336137533,
0.022769825533032417,
-0.014743694104254246,
-0.11177804321050644,
-0.012718371115624905,
0.07148289680480957,
-0.0006361124687828124,
0.06908723711967468,
0.11662404239177704,
0.029972350224852562,
-0.005142522975802422,
-0.0016250833868980408,
0.04451625794172287,
0.018372271209955215,
0.02175355888903141,
-0.10037677735090256,
0.052314285188913345,
0.034143272787332535,
0.02061762660741806,
-0.016738975420594215,
-0.0012634831946343184,
0.00438797939568758,
-0.015698304399847984,
-0.08748199790716171,
-0.012893137522041798,
0.0519600585103035,
-0.06993595510721207,
0.05563894659280777,
-0.003805737243965268,
0.028776368126273155,
0.02649758569896221,
-0.07755432277917862,
-0.06899896264076233,
3.6728927105717016e-33,
0.024805491790175438,
0.00990938302129507,
0.07501441985368729,
-0.010224373079836369,
0.026093740016222,
0.040649838745594025,
0.10548491775989532,
0.055299051105976105,
0.00569947762414813,
0.06904628872871399,
0.002017734572291374,
-0.007693000603467226,
-0.0927312821149826,
0.08595272153615952,
0.03900241479277611,
-0.034085460007190704,
-0.025961343199014664,
0.05179303139448166,
-0.01789691112935543,
-0.03975740820169449,
0.03066948428750038,
-0.06004687026143074,
-0.04492699354887009,
-0.03248944506049156,
0.0357741117477417,
0.04218779876828194,
0.035528093576431274,
0.031506333500146866,
-0.028809094801545143,
0.01095921266824007,
-0.05996014550328255,
-0.02998470328748226,
-0.041848231106996536,
-0.01452072523534298,
0.06835312396287918,
-0.051608603447675705,
-0.057777225971221924,
0.029308568686246872,
0.02699991501867771,
-0.019847530871629715,
-0.028068404644727707,
0.03765098750591278,
0.04566485807299614,
-0.017986252903938293,
-0.14547061920166016,
-0.13312354683876038,
-0.007666385732591152,
-0.0033418131060898304,
0.0018092746613547206,
-0.03564970940351486,
0.01771111600100994,
-0.04314877465367317,
-0.05656099691987038,
-0.06212668493390083,
-0.009221288375556469,
-0.029447225853800774,
0.04248971864581108,
0.09268441796302795,
0.00020725508511532098,
0.07734015583992004,
-0.04367242380976677,
0.007780418265610933,
-0.06429778039455414,
0.091380774974823,
-0.037085432559251785,
0.08753855526447296,
0.037570953369140625,
0.05211098492145538,
0.03530525788664818,
-0.007882459089159966,
0.025730952620506287,
-0.0046639335341751575,
-0.007309095002710819,
-0.06009795516729355,
0.05321906879544258,
0.04096176475286484,
0.021690450608730316,
0.012358399108052254,
0.03237570822238922,
0.024716340005397797,
-0.03791186958551407,
0.019428109750151634,
-0.006170423701405525,
-0.0023637861013412476,
-0.013907771557569504,
-0.0267346054315567,
0.07239518314599991,
-0.04903841391205788,
-0.007760254200547934,
0.012303054332733154,
-0.08656277507543564,
-0.017024097964167595,
-0.044647689908742905,
-0.018382450565695763,
-0.05505456030368805,
-3.8556071443597024e-33,
-0.006347546819597483,
0.04983719438314438,
-0.020776668563485146,
0.07643221318721771,
0.02687068097293377,
-0.035378385335206985,
-0.006781923118978739,
0.018211696296930313,
0.0479709655046463,
-0.06481163203716278,
-0.08811943978071213,
-0.03146642819046974,
0.026584986597299576,
-0.018047813326120377,
0.03652598336338997,
0.08927887678146362,
-0.05716070905327797,
0.022369084879755974,
0.002761882496997714,
-0.005841441452503204,
-0.04069376364350319,
0.09417527914047241,
-0.0657941997051239,
0.02368267811834812,
-0.06704436242580414,
-0.007694021333009005,
-0.007348231039941311,
0.06510885059833527,
-0.014899029396474361,
0.011202933266758919,
-0.03394963964819908,
-0.10124893486499786,
-0.022597871720790863,
-0.010987738147377968,
-0.0933474451303482,
0.04129345715045929,
-0.03366347774863243,
0.027817586436867714,
0.0169749166816473,
0.10018564015626907,
0.09639810025691986,
-0.04502628371119499,
-0.052461057901382446,
0.024868348613381386,
-0.013018561527132988,
-0.04449023678898811,
-0.0914306715130806,
-0.029966164380311966,
0.026004312559962273,
0.01482031773775816,
-0.05514845624566078,
0.04611874371767044,
-0.046684686094522476,
0.006524938158690929,
-0.08389081060886383,
-0.013724724762141705,
-0.022939864546060562,
0.07201524078845978,
0.1074545755982399,
0.0059340340085327625,
-0.012691189534962177,
-0.02216176688671112,
-0.008774928748607635,
-0.07215564697980881,
0.007464420981705189,
0.019878357648849487,
-0.015779253095388412,
-0.04897763580083847,
0.038350801914930344,
0.0356624610722065,
-0.01595322974026203,
0.009484133683145046,
0.026968585327267647,
0.0013219773536548018,
-0.0306827612221241,
0.04982072487473488,
0.024996982887387276,
-0.055503927171230316,
0.059109270572662354,
0.040001124143600464,
0.04232501611113548,
-0.014037234708666801,
0.06602603942155838,
0.048278484493494034,
0.0791681706905365,
0.04816410690546036,
0.1027282252907753,
0.061188746243715286,
0.05065642297267914,
-0.05055883154273033,
0.01146117877215147,
-0.022853491827845573,
0.0034983300138264894,
-0.025473766028881073,
-0.030845211818814278,
-4.7058119179155256e-8,
-0.08209801465272903,
0.12801335752010345,
0.048340946435928345,
0.03860624507069588,
0.07957743108272552,
-0.02967172861099243,
0.03471364453434944,
0.12730737030506134,
-0.0797296091914177,
-0.0682232677936554,
0.09883756935596466,
-0.0022920812480151653,
0.05035688728094101,
-0.027277588844299316,
0.005734659731388092,
0.07936716079711914,
-0.013248798437416553,
-0.016943322494626045,
-0.03342305123806,
0.03018372319638729,
-0.06627865880727768,
0.05972254276275635,
0.011597682721912861,
-0.04087517410516739,
0.01950313337147236,
-0.03973142057657242,
-0.003436405211687088,
0.019576873630285263,
0.11256151646375656,
-0.06312108039855957,
0.02843681536614895,
-0.008957683108747005,
0.052872247993946075,
-0.04114724323153496,
0.035663705319166183,
0.05997544899582863,
-0.006532120518386364,
-0.03830467164516449,
-0.040560200810432434,
-0.02946281246840954,
-0.028340119868516922,
0.03847094625234604,
-0.006901497952640057,
0.01806548796594143,
-0.032288357615470886,
0.009793128818273544,
-0.04425753280520439,
-0.0805623009800911,
-0.01211907435208559,
-0.02656560018658638,
0.09338083863258362,
-0.056121356785297394,
-0.022866113111376762,
0.037087757140398026,
0.0841955915093422,
-0.04515508562326431,
0.014631879515945911,
-0.007832101546227932,
-0.05329642817378044,
0.103905089199543,
-0.03389216214418411,
-0.05195557326078415,
-0.07324102520942688,
0.004644064698368311
] | 0 | 0 | 0 | 0 |
|
Storybooth: Training-Free Multi-Subject Consistency for Improved Visual Storytelling | https://openreview.net/forum?id=JZLon6cvx8 | [
"Jaskirat Singh",
"Junshen K Chen",
"Jonas K Kohler",
"Michael F Cohen"
] | Poster | Consistent text-to-image generation depicting the *same* subjects across different images has gained significant recent attention due to its widespread applications in the fields of visual-storytelling and multiple-shot video generation. While remarkable, existing methods often require costly finetuning for each subject and struggle to maintain consistency across multiple characters. In this work, we first analyse the reason for these limitations. Our exploration reveals that the primary-issue stems from *self-attention leakage*, which is exacerbated when trying to ensure consistency across multiple-characters. Motivated by these findings, we next propose a simple yet effective *training and optimization-free approach* for improving multiple-character consistency. In particular, we first leverage multi-modal *chain-of-thought* reasoning in order to *apriori* localize the different subjects across the storyboard frames. The final storyboard images are then generated using a modified diffusion model which includes *1) a bounded cross-attention layer* for ensuring adherence to the initially predicted layout, and *2) a bounded cross-frame self-attention layer* for reducing inter-character attention leakage. Furthermore, we also propose a novel *cross-frame token-merging layer* which allows for improved fine-grain consistency for the storyboard characters.
Experimental analysis reveals that proposed approach is not only $\times 30$ faster than prior training-based methods (*eg, textual inversion, dreambooth-lora*) but also surpasses the prior *state-of-the-art*, exhibiting improved multi-character consistency and text-to-image alignment performance. | consistent text-to-image generation, visual storytelling, story generation | null | 4 | null | [
0.04595397040247917,
-0.10815487802028656,
0.038179729133844376,
0.005441921763122082,
0.11944001168012619,
0.029405834153294563,
-0.03396425396203995,
-0.03580089658498764,
0.08704342693090439,
-0.07469720393419266,
-0.001755211502313614,
-0.05056097358465195,
0.007679593283683062,
0.048629917204380035,
-0.00785066932439804,
-0.036054566502571106,
0.0516679622232914,
0.06533898413181305,
-0.08381989598274231,
-0.04030868411064148,
0.043302394449710846,
-0.030016286298632622,
0.10444000363349915,
0.008971654810011387,
-0.019987305626273155,
-0.005175062920898199,
-0.03242993727326393,
-0.07245216518640518,
0.04200337827205658,
-0.06034056469798088,
0.043310124427080154,
-0.02310074120759964,
-0.057979922741651535,
0.029706405475735664,
-0.02767655812203884,
0.06873249262571335,
-0.07555733621120453,
0.08079049736261368,
-0.021230414509773254,
-0.012003039941191673,
0.016920670866966248,
0.03868122398853302,
0.011670286767184734,
0.05579213425517082,
0.08066999912261963,
-0.006000151392072439,
-0.03756660223007202,
0.02262122556567192,
0.024876190349459648,
-0.00538539607077837,
-0.053288064897060394,
-0.02029603347182274,
-0.006878461688756943,
0.007569837849587202,
-0.0030121158342808485,
0.046402476727962494,
0.06423403322696686,
0.043199993669986725,
0.03831718489527702,
-0.015157358720898628,
-0.04309887811541557,
-0.03872194141149521,
-0.011747339740395546,
-0.021356595680117607,
0.06614210456609726,
-0.04624522477388382,
0.05586287006735802,
0.02431778609752655,
-0.058613672852516174,
0.013429766520857811,
-0.003601107047870755,
0.005079944152384996,
-0.002974356757476926,
-0.021118858829140663,
0.029930774122476578,
0.11827889829874039,
0.009996732696890831,
-0.10752375423908234,
0.031190451234579086,
-0.028133675456047058,
0.008315738290548325,
-0.02176605351269245,
0.06087304651737213,
-0.03378651291131973,
0.10300798714160919,
-0.028546802699565887,
-0.006713045295327902,
0.018252696841955185,
-0.048743050545454025,
0.02056490257382393,
-0.0557975172996521,
-0.0821627750992775,
0.09864243119955063,
-0.009419416077435017,
0.05138217657804489,
0.05234243720769882,
0.007190958596765995,
-0.0732722133398056,
-0.025793954730033875,
0.03406304121017456,
0.013054617680609226,
-0.016742520034313202,
0.04737890884280205,
0.03114495985209942,
0.06747914850711823,
-0.0304513368755579,
0.08305597305297852,
0.03438263386487961,
-0.008280298672616482,
-0.06106331944465637,
0.07739534974098206,
0.018369629979133606,
-0.049285463988780975,
0.03282913193106651,
0.08212681859731674,
-0.00029804956284351647,
0.01590762473642826,
-0.03862354904413223,
0.04218940809369087,
-0.004540538880974054,
0.010974563658237457,
-0.01496302243322134,
-0.013664914295077324,
0.0064015514217317104,
0.07114875316619873,
-0.09362585842609406,
0.077072374522686,
-1.0097692083170266e-33,
0.03245711326599121,
0.03437986597418785,
0.01833866536617279,
0.07349184155464172,
0.029314160346984863,
-0.010402003303170204,
-0.04939095675945282,
0.011306763626635075,
-0.05346677452325821,
0.005552910268306732,
0.029932962730526924,
-0.04730949550867081,
-0.07279078662395477,
0.11363562941551208,
0.025584308430552483,
-0.04196389392018318,
-0.0562620535492897,
0.010995118878781796,
0.019861577078700066,
0.08746741712093353,
0.03256145119667053,
-0.0274768378585577,
-0.00342868291772902,
-0.025905432179570198,
-0.0652013048529625,
-0.00582152558490634,
0.12251557409763336,
-0.06778473407030106,
-0.0172211192548275,
0.0006591823766939342,
-0.13502095639705658,
0.0464913435280323,
0.0409877672791481,
0.0349050834774971,
-0.0010804206831380725,
-0.08427302539348602,
-0.04990677163004875,
-0.04075950011610985,
0.039531685411930084,
0.06226564198732376,
-0.013731180690228939,
0.057688724249601364,
0.008937494829297066,
-0.024418190121650696,
-0.03479021042585373,
-0.0028291307389736176,
0.002993451664224267,
0.08985771238803864,
-0.11013063788414001,
-0.05218590423464775,
0.024236226454377174,
0.04463741555809975,
-0.016554050147533417,
-0.09885130822658539,
0.00987331848591566,
-0.06416556239128113,
0.05226204916834831,
-0.029509995132684708,
0.08994799107313156,
-0.006572614423930645,
0.08059633523225784,
0.0435493141412735,
-0.04670777544379234,
0.118848517537117,
0.038094960153102875,
0.08393885940313339,
0.02574303187429905,
0.0319468192756176,
0.03788158297538757,
-0.03889781981706619,
-0.06034719571471214,
-0.05083026364445686,
0.023456549271941185,
-0.05533139407634735,
-0.042847245931625366,
0.022319039329886436,
0.0050414069555699825,
-0.05961136147379875,
-0.012276099063456059,
0.056946784257888794,
0.01249634101986885,
-0.022290309891104698,
-0.08665778487920761,
-0.05180199444293976,
-0.03848661109805107,
-0.08039231598377228,
0.03268124535679817,
-0.05782485380768776,
-0.04989172890782356,
-0.0477774403989315,
0.04531028866767883,
-0.007050734478980303,
0.029199786484241486,
-0.04128098115324974,
0.015417298302054405,
2.31009343685046e-33,
0.04875817522406578,
0.03435466066002846,
-0.043753236532211304,
0.0018905616598203778,
0.07355372607707977,
-0.07693290710449219,
-0.008447418920695782,
0.008940807543694973,
0.04442672058939934,
-0.0889982357621193,
-0.10865076631307602,
-0.02851111814379692,
-0.05275535210967064,
-0.06675872951745987,
-0.000971725326962769,
0.015473389066755772,
0.060104478150606155,
0.046724580228328705,
-0.09806255251169205,
0.012976397760212421,
0.04372480884194374,
0.025571582838892937,
-0.0876021534204483,
0.02798670344054699,
0.02993343397974968,
0.07962773740291595,
-0.029795823618769646,
0.06026868149638176,
-0.0008092545904219151,
-0.016192689538002014,
-0.03120914287865162,
-0.012898333370685577,
0.07312723994255066,
-0.04388125240802765,
-0.07971473783254623,
0.07431919872760773,
0.042692456394433975,
-0.017690442502498627,
-0.019587062299251556,
0.10377415269613266,
0.07807675004005432,
-0.08160781860351562,
-0.060070157051086426,
0.03617563471198082,
-0.05032210424542427,
0.0025489996187388897,
-0.10676567256450653,
-0.048790231347084045,
-0.027080345898866653,
0.11645149439573288,
-0.11649106442928314,
0.011466444469988346,
-0.046599093824625015,
-0.04640616849064827,
-0.06810939311981201,
-0.07418778538703918,
-0.08752050995826721,
0.04762296751141548,
0.06960228830575943,
-0.04886433854699135,
-0.012431721203029156,
0.019060717895627022,
0.017238903790712357,
-0.08217145502567291,
0.029232874512672424,
-0.06528858840465546,
-0.015537059865891933,
-0.06130087003111839,
0.013729847967624664,
0.039420802146196365,
0.00590862799435854,
-0.008715400472283363,
0.04633010923862457,
0.02501811645925045,
0.057362738996744156,
0.02695825882256031,
0.006851245649158955,
-0.05293136462569237,
0.011485776863992214,
-0.07572311908006668,
-0.036131955683231354,
-0.02581629529595375,
0.010538073256611824,
0.08563121408224106,
0.040154214948415756,
0.0755147635936737,
0.06368403136730194,
0.08837290108203888,
0.05258416384458542,
-0.01702357642352581,
-0.0491601899266243,
0.02874140441417694,
0.06500224024057388,
0.056377723813056946,
-0.021570121869444847,
-4.0036564286083376e-8,
-0.06595903635025024,
-0.05247725173830986,
-0.011643834412097931,
0.0695154219865799,
-0.029346244409680367,
-0.08925490081310272,
-0.004786042030900717,
0.012251192703843117,
0.016897791996598244,
0.0028297461103647947,
0.01672653667628765,
0.01965482346713543,
-0.016054226085543633,
-0.04139527678489685,
-0.03356615826487541,
0.007264540996402502,
-0.017552662640810013,
0.04149811714887619,
-0.028287414461374283,
-0.04355204105377197,
0.059038229286670685,
0.05060382932424545,
0.03469785302877426,
-0.06188635155558586,
-0.05353916436433792,
0.004868577234447002,
-0.027712024748325348,
0.026087936013936996,
0.04421501234173775,
-0.06210843473672867,
0.0700211226940155,
-0.006291007157415152,
0.02344951219856739,
-0.049373749643564224,
-0.02792404405772686,
0.03400915116071701,
0.04230659827589989,
0.047165922820568085,
0.020278997719287872,
-0.03398279473185539,
0.048295121639966965,
0.007499590050429106,
0.057175301015377045,
0.07885640114545822,
0.04110848903656006,
0.012837626971304417,
-0.020875973626971245,
-0.06618545204401016,
0.05973386391997337,
-0.06540267914533615,
-0.04694424942135811,
-0.05281501263380051,
-0.0331205390393734,
0.0841754674911499,
0.0887005627155304,
-0.03345494344830513,
0.08666488528251648,
0.035384319722652435,
0.1116022914648056,
0.030859345570206642,
-0.00515633262693882,
0.04443960264325142,
-0.03179101645946503,
-0.024742290377616882
] | 0 | 0 | 0 | 0 |
|
How new data permeates LLM knowledge and how to dilute it | https://openreview.net/forum?id=NGKQoaqLpo | [
"Chen Sun",
"Renat Aksitov",
"Andrey Zhmoginov",
"Nolan Andrew Miller",
"Max Vladymyrov",
"Ulrich Rueckert",
"Been Kim",
"Mark Sandler"
] | Spotlight | Large language models continually learn through the accumulation of gradient-based updates, but how individual pieces of new information affect existing knowledge, leading to both beneficial generalization and problematic hallucination, remains poorly understood. We demonstrate that when learning new information, LLMs exhibit a "priming" effect: learning a new fact can cause the model to inappropriately apply that knowledge in unrelated contexts.
To systematically study this phenomenon, we introduce "Outlandish," a carefully curated dataset of 1320 diverse text samples designed to probe how new knowledge permeates through an LLM's existing knowledge base. Using this dataset, we show that the degree of priming after learning new information can be predicted by measuring the token probability of key words before training. This relationship holds robustly across different model architectures (PALM-2, Gemma, Llama), sizes, and training stages.
Finally, we develop two novel techniques to modulate how new knowledge affects existing model behavior: (1) a ``stepping-stone'' text augmentation strategy and (2) an ``ignore-k'' update pruning method. These approaches reduce undesirable priming effects by 50-95% while preserving the model's ability to learn new information. Our findings provide both empirical insights into how LLMs learn and practical tools for improving the specificity of knowledge insertion in language models. Further materials: https://sunchipsster1.github.io/projects/outlandish/ | fine-tuning, hallucinations, knowledge injection, memory, LLMs | null | 13,888 | null | [
0.010717143304646015,
-0.08185147494077682,
0.0461600124835968,
0.07783441245555878,
0.07353334873914719,
-0.00283587328158319,
-0.02553175576031208,
-0.026421280577778816,
0.04900741204619408,
0.01624351367354393,
0.030265264213085175,
0.028935078531503677,
0.03614821285009384,
-0.035160087049007416,
0.021459946408867836,
0.016528096050024033,
0.03537170588970184,
0.08126001060009003,
-0.14030107855796814,
-0.07321435213088989,
0.023027408868074417,
0.01663060486316681,
0.05590604245662689,
0.009195000864565372,
-0.014017236419022083,
0.0350685641169548,
-0.017659993842244148,
0.010208009742200375,
0.049735505133867264,
-0.06552117317914963,
0.015746135264635086,
0.09885767847299576,
0.02240488864481449,
0.05341003090143204,
-0.04831703379750252,
0.015392065048217773,
-0.03158416226506233,
0.10651649534702301,
0.026942215859889984,
-0.05741789564490318,
0.011900200508534908,
-0.01679772511124611,
0.0031784488819539547,
0.062023598700761795,
0.10395114868879318,
0.0038148663006722927,
0.010986408218741417,
0.024278463795781136,
-0.034288447350263596,
-0.0023074839264154434,
-0.058862462639808655,
-0.06776547431945801,
-0.004694810602813959,
0.012086630798876286,
-0.03038700856268406,
0.032329630106687546,
0.006920613814145327,
0.07506558299064636,
-0.0538778230547905,
-0.034812018275260925,
-0.015878669917583466,
-0.11528486758470535,
-0.07507862150669098,
-0.02709522470831871,
-0.0430193655192852,
-0.057027556002140045,
-0.010486808605492115,
0.08559959381818771,
-0.021914290264248848,
0.0467103011906147,
0.0719885528087616,
0.08509090542793274,
-0.010680139996111393,
0.08718036115169525,
0.04023413732647896,
0.048335347324609756,
0.007409754209220409,
0.0281947273761034,
0.03973643109202385,
-0.023218030110001564,
0.009732902981340885,
0.05943905934691429,
0.06333201378583908,
0.004833166021853685,
0.08443576097488403,
-0.029116759076714516,
0.043709591031074524,
0.01744321547448635,
-0.03641160577535629,
-0.0184265598654747,
-0.045842524617910385,
-0.09062142670154572,
0.013875920325517654,
-0.0075804805383086205,
0.010417616926133633,
0.04685346782207489,
-0.02219030074775219,
-0.08373325318098068,
-0.03985007107257843,
0.08045274764299393,
-0.007815707474946976,
0.14313365519046783,
-0.00022794299002271146,
-0.003227448556572199,
0.00158514897339046,
-0.019655976444482803,
0.01007845252752304,
0.05096178501844406,
0.036323368549346924,
-0.04035884886980057,
0.02546190470457077,
0.10653326660394669,
-0.0076319691725075245,
-0.023887570947408676,
-0.005633885972201824,
-0.007650172803550959,
0.024695873260498047,
-0.02460823580622673,
-0.04822156950831413,
0.0952257513999939,
-0.022318586707115173,
0.022738810628652573,
0.029847722500562668,
-0.008083774708211422,
-0.07073277980089188,
-0.02275862917304039,
-0.07091164588928223,
1.224728186832052e-33,
0.06495334953069687,
0.061166875064373016,
-0.008165953680872917,
0.08491821587085724,
0.05362085625529289,
0.0003984845825470984,
0.002423277124762535,
0.011432196013629436,
0.01091486494988203,
-0.033246006816625595,
-0.003447069553658366,
0.001360702677629888,
-0.040193501859903336,
0.02908216044306755,
0.03494129329919815,
0.002389239612966776,
-0.06718195229768753,
0.044219594448804855,
-0.03331001102924347,
-0.04165565222501755,
0.04187522456049919,
0.018261874094605446,
0.015610630623996258,
-0.09840359538793564,
-0.03105899877846241,
0.06481246650218964,
0.0875452533364296,
-0.027562301605939865,
-0.08502065390348434,
0.01862172782421112,
-0.07306777685880661,
-0.008902185596525669,
0.011502182111144066,
0.04682087153196335,
0.0029087925795465708,
-0.04358184337615967,
0.020531482994556427,
-0.11931372433900833,
0.0757775604724884,
-0.05688314884901047,
-0.022224998101592064,
0.047672171145677567,
0.026440860703587532,
-0.018853522837162018,
-0.07325610518455505,
-0.03681575506925583,
0.005461425520479679,
-0.043939489871263504,
-0.03569963946938515,
-0.039904139935970306,
0.04974199831485748,
0.03303823247551918,
-0.021658770740032196,
0.044648051261901855,
-0.0930560752749443,
0.047876082360744476,
-0.03590203821659088,
-0.04206198826432228,
0.03621998429298401,
-0.010147451423108578,
0.03639225661754608,
0.06002539396286011,
-0.007635351736098528,
0.16317687928676605,
0.05043884366750717,
0.024490978568792343,
-0.06456942111253738,
0.028635099530220032,
0.031062515452504158,
-0.05591553449630737,
-0.011114024557173252,
-0.0291855838149786,
-0.06443493068218231,
-0.024482296779751778,
-0.01232830062508583,
-0.036369726061820984,
0.021640043705701828,
-0.04009260609745979,
0.010025355964899063,
0.0020915830973535776,
0.07844927906990051,
-0.06205618381500244,
0.009689483791589737,
-0.08054909855127335,
-0.0677332878112793,
-0.015087663196027279,
0.055599842220544815,
-0.09486697614192963,
0.014605027623474598,
-0.008833852596580982,
-0.007653688546270132,
-0.0022499943152070045,
-0.04202635586261749,
0.005124260671436787,
0.02860180102288723,
-1.1506902261584872e-33,
-0.05101761594414711,
-0.05588725581765175,
-0.03259407728910446,
0.0374283529818058,
-0.055414170026779175,
-0.02722279727458954,
0.003377315355464816,
0.06335338950157166,
-0.06747332960367203,
-0.12192975729703903,
-0.04581121355295181,
-0.0238684993237257,
0.043481215834617615,
-0.010744179598987103,
0.008032477460801601,
0.011649583466351032,
0.017845939844846725,
0.07224810123443604,
0.031846169382333755,
0.07018055766820908,
-0.00599667290225625,
0.031275298446416855,
-0.12230835109949112,
0.039539679884910583,
-0.00172906881198287,
-0.0002999439020641148,
-0.028574150055646896,
0.10657226294279099,
-0.0356159508228302,
-0.006550672464072704,
0.00039759432547725737,
-0.0005400136578828096,
-0.02108755148947239,
0.035719387233257294,
-0.08328466862440109,
-0.009924634359776974,
-0.0024042848963290453,
-0.029858602210879326,
-0.02009911835193634,
0.08634355664253235,
0.0018515682313591242,
-0.028775108978152275,
-0.0677594467997551,
-0.030176986008882523,
-0.036382220685482025,
-0.08669529855251312,
-0.08578606694936752,
-0.02102484740316868,
0.054548244923353195,
-0.03021669015288353,
0.03854847326874733,
-0.0008541218703612685,
-0.034965917468070984,
-0.011437155306339264,
-0.01122679840773344,
-0.05157284066081047,
-0.020788785070180893,
-0.01647414267063141,
0.027720972895622253,
0.022085053846240044,
-0.03966878727078438,
0.023432772606611252,
-0.01745861954987049,
-0.04711654409766197,
-0.001887906459160149,
-0.10276689380407333,
-0.07457828521728516,
0.02824554219841957,
0.006934817414730787,
-0.014897696673870087,
0.021588997915387154,
0.034942127764225006,
-0.06827185302972794,
-0.019033988937735558,
-0.013171239756047726,
-0.007837063632905483,
0.003682386828586459,
-0.08917029201984406,
-0.07261094450950623,
-0.0914241224527359,
-0.01880558393895626,
-0.015110626816749573,
0.005633044522255659,
0.056318655610084534,
0.06908125430345535,
0.12501311302185059,
0.04694291949272156,
0.10220427066087723,
-0.008360177278518677,
0.018267439678311348,
-0.037644773721694946,
-0.026086382567882538,
-0.1137421652674675,
0.0618089959025383,
-0.05983225628733635,
-4.344869708461374e-8,
-0.08115866780281067,
0.02949182316660881,
-0.05962316691875458,
0.12325215339660645,
0.041880201548337936,
-0.07609228044748306,
-0.07820053398609161,
0.127120241522789,
-0.02016521617770195,
-0.07850862294435501,
0.07608013600111008,
0.03689724579453468,
-0.03698918968439102,
-0.03157154843211174,
0.030948108062148094,
0.09624616801738739,
0.08132673054933548,
0.02887074649333954,
-0.0348680205643177,
-0.034116048365831375,
0.07664638012647629,
0.0376814603805542,
0.04777289927005768,
-0.021805651485919952,
0.05554358288645744,
-0.06709074229001999,
-0.02460353821516037,
0.08060488849878311,
0.012694882228970528,
-0.013606112450361252,
0.023906204849481583,
-0.048406362533569336,
-0.03280468285083771,
0.002785697113722563,
0.0657460168004036,
0.129008948802948,
0.059205446392297745,
-0.06805649399757385,
0.014626838266849518,
-0.02126719057559967,
0.032716456800699234,
0.039098046720027924,
-0.0005993582890368998,
0.08719683438539505,
-0.05077272653579712,
-0.026133572682738304,
-0.03723446652293205,
-0.05061584338545799,
0.014528282918035984,
-0.06608554720878601,
0.06001191958785057,
0.016871877014636993,
0.04120643064379692,
0.018960827961564064,
0.09026620537042618,
0.07992783188819885,
-0.023694613948464394,
0.044074300676584244,
0.01930898427963257,
0.03499511629343033,
0.07078149914741516,
-0.0564020574092865,
-0.06354138255119324,
0.032240498811006546
] | 0 | 0 | 0 | 0 |
|
UniCBE: An Uniformity-driven Comparing Based Evaluation Framework with Unified Multi-Objective Optimization | https://openreview.net/forum?id=rpwGUtTeA5 | [
"Peiwen Yuan",
"Shaoxiong Feng",
"Yiwei Li",
"Xinglin Wang",
"Yueqi Zhang",
"Jiayi Shi",
"Chuyi Tan",
"Boyuan Pan",
"Yao Hu",
"Kan Li"
] | Spotlight | Human preference plays a significant role in measuring large language models and guiding them to align with human values. Unfortunately, current comparing-based evaluation (CBE) methods typically focus on a single optimization objective, failing to effectively utilize scarce yet valuable preference signals. To address this, we delve into key factors that can enhance the accuracy, convergence, and scalability of CBE: suppressing sampling bias, balancing descending process of uncertainty, and mitigating updating uncertainty.
Following the derived guidelines, we propose UniCBE, a unified uniformity-driven CBE framework which simultaneously optimize these core objectives by constructing and integrating three decoupled sampling probability matrices, each designed to ensure uniformity in specific aspects. We further ablate the optimal tuple sampling and preference aggregation strategies to achieve efficient CBE.
On the AlpacaEval benchmark, UniCBE saves over 17% of evaluation budgets while achieving a Pearson correlation with ground truth exceeding 0.995, demonstrating excellent accuracy and convergence. In scenarios where new models are continuously introduced, UniCBE can even save over 50% of evaluation costs, highlighting its improved scalability. | evaluation, efficient, scalability, accuracy, convergence | We propose UniCBE, a comparing-based evaluation framework with better scalability, accuracy and convergence. | 13,837 | 2502.11454 | [
-0.022420838475227356,
-0.08136887848377228,
-0.07460621744394302,
0.002961109858006239,
0.02771775983273983,
0.0013224761933088303,
-0.019044535234570503,
-0.019628584384918213,
0.03382909670472145,
-0.0208330936729908,
-0.09686801582574844,
-0.11434504389762878,
0.028704164549708366,
0.01412348635494709,
0.05375071242451668,
-0.03069210611283779,
0.07388154417276382,
0.042282022535800934,
-0.11111528426408768,
-0.04015004262328148,
0.01148646604269743,
0.0749746784567833,
0.04362662881612778,
-0.07922421395778656,
0.040064968168735504,
-0.06224534660577774,
-0.02110804058611393,
-0.019303882494568825,
-0.01827087253332138,
0.008165636099874973,
0.04768640175461769,
0.03211532160639763,
0.08719462156295776,
-0.042152661830186844,
-0.02527010254561901,
-0.008774071000516415,
-0.10069023072719574,
-0.052235737442970276,
0.018390433862805367,
0.026585234329104424,
-0.08044068515300751,
-0.015228562988340855,
-0.01415333617478609,
0.05252235382795334,
0.023989899083971977,
0.031130146235227585,
-0.04865977540612221,
-0.011506893672049046,
0.043773550540208817,
0.05546336993575096,
-0.04507480934262276,
0.030616970732808113,
0.02789878658950329,
-0.0026960368268191814,
-0.02753114514052868,
0.02461814135313034,
-0.01265591848641634,
-0.005426970310509205,
0.03737783804535866,
-0.13169896602630615,
-0.13079291582107544,
-0.1332194060087204,
-0.02250208891928196,
0.07166043668985367,
0.03634258359670639,
-0.053822483867406845,
-0.004575349856168032,
0.022088726982474327,
-0.00009616922761779279,
0.10451089590787888,
-0.050430938601493835,
0.10317566245794296,
0.032937925308942795,
0.10117907077074051,
0.015958309173583984,
0.10693853348493576,
-0.010083422064781189,
-0.056334372609853745,
0.0462724044919014,
-0.051042068749666214,
-0.025429725646972656,
-0.0005937748355790973,
0.06732833385467529,
0.020116262137889862,
0.15323737263679504,
0.012735736556351185,
0.018834546208381653,
0.03403763845562935,
0.001063415315002203,
0.027268335223197937,
-0.038939837366342545,
0.021749911829829216,
0.07674509286880493,
0.02104094624519348,
-0.014569477178156376,
0.03373957425355911,
-0.00965762697160244,
-0.05085097253322601,
-0.04616689682006836,
0.09137294441461563,
0.013676067814230919,
0.05900523439049721,
0.008806823752820492,
-0.06419568508863449,
-0.03643564507365227,
-0.09169483184814453,
0.00659495173022151,
0.03837010636925697,
0.012449636124074459,
-0.06323177367448807,
-0.024731669574975967,
0.005197138991206884,
0.006830128375440836,
0.024973737075924873,
0.015399581752717495,
0.0015933471731841564,
0.04186674952507019,
-0.021319566294550896,
0.05526300147175789,
0.04268322139978409,
-0.05119498446583748,
0.0029490331653505564,
0.03992419317364693,
0.03171301633119583,
-0.02484830468893051,
0.04017110541462898,
0.023548994213342667,
2.9967364213642213e-33,
-0.043115511536598206,
0.01929568313062191,
0.09810581058263779,
-0.009209859184920788,
-0.02512693777680397,
-0.021786943078041077,
-0.009330732747912407,
0.022270988672971725,
-0.07810463011264801,
-0.007202227599918842,
0.011529293842613697,
0.13464216887950897,
-0.03912446275353432,
0.07158651947975159,
0.06268326193094254,
0.028422115370631218,
-0.007051211781799793,
0.11773458123207092,
-0.06201944127678871,
-0.01264667883515358,
0.08384936302900314,
0.02354830503463745,
0.008138335309922695,
-0.0333278551697731,
0.003661296097561717,
-0.014332501217722893,
0.03536384180188179,
-0.005930073093622923,
-0.040461476892232895,
0.0038594373036175966,
-0.09040248394012451,
-0.031141547486186028,
0.021938418969511986,
-0.011205376125872135,
0.04466928169131279,
-0.04734577611088753,
0.014585498720407486,
0.0009057358838617802,
-0.0763971135020256,
0.009668497368693352,
-0.034923333674669266,
0.022572703659534454,
0.11175337433815002,
-0.04640251770615578,
-0.045868661254644394,
-0.033486220985651016,
0.005204599816352129,
-0.005816231016069651,
0.059091582894325256,
-0.09528017044067383,
0.05517066642642021,
0.022480661049485207,
0.00025204732082784176,
0.04448787495493889,
-0.01352309063076973,
-0.04031069204211235,
0.05254222825169563,
0.06644385308027267,
-0.009904952719807625,
-0.01279002707451582,
-0.03532036021351814,
0.07238365709781647,
0.006393577437847853,
-0.03998322784900665,
0.049666907638311386,
0.07530055195093155,
-0.025960354134440422,
0.009222110733389854,
0.042233098298311234,
0.024229254573583603,
-0.003737831022590399,
0.04142041876912117,
-0.02103094756603241,
-0.043010372668504715,
-0.022777264937758446,
0.07646546512842178,
0.03289899230003357,
0.00574883259832859,
0.0799076110124588,
-0.003264723578467965,
0.04381650313735008,
0.006249688100069761,
-0.022710008546710014,
-0.12906354665756226,
-0.02196715958416462,
-0.09800849109888077,
0.04045712947845459,
-0.029915524646639824,
-0.036882102489471436,
-0.034450698643922806,
0.009728004224598408,
0.028480317443609238,
-0.07384771853685379,
0.024415358901023865,
0.025976281613111496,
-3.608409131247422e-33,
-0.03811582177877426,
-0.019016079604625702,
0.04870900511741638,
0.09451105445623398,
-0.029265452176332474,
-0.026667771860957146,
0.04245763272047043,
-0.01691070757806301,
0.015402527526021004,
-0.09889901429414749,
-0.05279620364308357,
-0.07090812176465988,
0.11423502117395401,
-0.050533004105091095,
0.059876613318920135,
0.035145606845617294,
-0.07779288291931152,
0.048700857907533646,
0.01878325827419758,
0.06079552322626114,
0.049834225326776505,
0.08721546083688736,
-0.09711067378520966,
0.022535279393196106,
-0.05532491207122803,
0.021983465179800987,
-0.01797318458557129,
-0.03390868753194809,
0.009859895333647728,
-0.06035012751817703,
-0.05015159770846367,
-0.0033567468635737896,
-0.09635062515735626,
0.00019364574109204113,
-0.06876450031995773,
-0.012146856635808945,
0.016056589782238007,
0.016724662855267525,
-0.014474744908511639,
0.16079184412956238,
0.0034913402050733566,
0.01827651634812355,
-0.06854956597089767,
-0.043079257011413574,
-0.000354477931978181,
0.0032348858658224344,
-0.07903245091438293,
-0.000046283625124488026,
-0.02494082599878311,
-0.041891586035490036,
-0.018110590055584908,
0.04239948093891144,
-0.07476159185171127,
0.08308444172143936,
-0.010397573001682758,
-0.07473500072956085,
-0.05183476582169533,
-0.07964283227920532,
0.072239451110363,
0.05157697573304176,
-0.06567569822072983,
0.07462725043296814,
-0.008829301223158836,
-0.032639265060424805,
-0.008456527255475521,
-0.01405265275388956,
0.009764982387423515,
0.010148502886295319,
-0.033522818237543106,
-0.031039904803037643,
-0.02221418172121048,
-0.07854995131492615,
-0.021770911291241646,
-0.04791458696126938,
-0.047818101942539215,
0.04814542084932327,
0.0931355357170105,
-0.040035609155893326,
0.006688036024570465,
-0.007529271300882101,
0.03248367831110954,
0.04899316653609276,
0.029932191595435143,
0.06183109059929848,
0.032963961362838745,
0.08683480322360992,
0.0534357912838459,
0.056889791041612625,
0.03213028237223625,
0.044945430010557175,
0.0009219817584380507,
-0.014223855920135975,
0.06642284244298935,
0.06446846574544907,
-0.07179700583219528,
-5.0742759327704334e-8,
-0.0835983008146286,
0.03820708766579628,
-0.007622567471116781,
0.09168313443660736,
-0.016711439937353134,
-0.06766097992658615,
-0.06870285421609879,
0.0012749609304592013,
-0.029965940862894058,
-0.07461743801832199,
0.044418271631002426,
-0.0007327936473302543,
-0.0626659095287323,
-0.024244921281933784,
-0.0173163004219532,
0.07669788599014282,
-0.010515832342207432,
0.03965700417757034,
0.024702511727809906,
-0.02076692320406437,
0.012025650590658188,
0.056787874549627304,
0.02295738458633423,
-0.02420889213681221,
0.03486417606472969,
-0.0035002930089831352,
-0.055586714297533035,
0.018647432327270508,
0.07724776864051819,
-0.0162222720682621,
-0.03468047082424164,
-0.009180772118270397,
-0.04907410219311714,
-0.03135617822408676,
0.05917274206876755,
0.05790851637721062,
-0.04591843858361244,
-0.03457944467663765,
0.01174592413008213,
-0.05496394634246826,
0.06496911495923996,
0.07827417552471161,
-0.11376701295375824,
0.0290006622672081,
0.059089239686727524,
-0.017098139971494675,
-0.03285522013902664,
-0.04306218773126602,
0.007645787671208382,
-0.07718001306056976,
-0.015263592824339867,
-0.08397921174764633,
-0.03786924481391907,
0.027523640543222427,
0.032594192773103714,
0.013816519640386105,
-0.0448044054210186,
-0.017133602872490883,
0.024657050147652626,
0.071947380900383,
0.09136505424976349,
-0.04487672075629234,
-0.08109351992607117,
-0.014816002920269966
] | 0 | 0 | 0 | 0 |
|
NetMoE: Accelerating MoE Training through Dynamic Sample Placement | https://openreview.net/forum?id=1qP3lsatCR | [
"Xinyi Liu",
"Yujie Wang",
"Fangcheng Fu",
"Xupeng Miao",
"Shenhan Zhu",
"Xiaonan Nie",
"Bin CUI"
] | Spotlight | Mixture of Experts (MoE) is a widely used technique to expand model sizes for better model quality while maintaining the computation cost constant. In a nutshell, an MoE model consists of multiple experts in each model layer and routes the training tokens to only a fixed number of experts rather than all. In distributed training, as experts are distributed among different GPUs, All-to-All communication is necessary to exchange the training tokens among the GPUs after each time of expert routing. Due to the frequent and voluminous data exchanges, All-to-All communication has become a notable challenge to training efficiency.
In this paper, we manage to accelerate All-to-All communication in MoE models from the training sample perspective, which is unexplored so far. In particular, we put forward the observation that tokens in the same training sample have certain levels of locality in expert routing. Motivated by this, we develop NetMoE, which takes such locality into account and dynamically rearranges the placement of training samples to minimize All-to-All communication costs. Specifically, we model the All-to-All communication given the sample placement and formulate an integer programming problem to deduce the optimal placement in polynomial time. Experiments with 32 GPUs show that NetMoE achieves a maximum efficiency improvement of $1.67 \times$ compared with current MoE training frameworks. | Mixture of Experts, All-to-All communication, Distributed training | null | 13,820 | null | [
-0.018349871039390564,
-0.05554937943816185,
0.023789793252944946,
0.07570154219865799,
0.057451412081718445,
-0.0571279413998127,
0.0056574526242911816,
0.014541939832270145,
-0.07083068042993546,
-0.02132176049053669,
-0.09394297003746033,
-0.010161792859435081,
0.004582414869219065,
-0.018418235704302788,
-0.0304307509213686,
0.010499643161892891,
0.16043546795845032,
0.07175538688898087,
-0.11343555897474289,
-0.037635110318660736,
-0.019871311262249947,
-0.0033927266485989094,
0.04700285196304321,
0.02119368687272072,
-0.001521528698503971,
0.020925357937812805,
0.05542069301009178,
-0.06975220888853073,
0.04148567467927933,
-0.02107188291847706,
-0.005079067777842283,
-0.03304918855428696,
0.03459179773926735,
0.06108042597770691,
-0.006678987760096788,
0.046059370040893555,
-0.08277017623186111,
0.02475857362151146,
0.01767982915043831,
-0.004424666985869408,
0.06662943214178085,
0.06120183691382408,
-0.039051029831171036,
-0.020107798278331757,
0.03618212789297104,
-0.016875071451067924,
0.04877348616719246,
0.016808466985821724,
0.01358860358595848,
0.006809940095990896,
-0.0328497588634491,
-0.08968997746706009,
-0.032837241888046265,
0.0408446229994297,
-0.03663269430398941,
0.017676018178462982,
0.007381713017821312,
-0.05315939709544182,
-0.009298187680542469,
0.07723212987184525,
-0.04763492941856384,
-0.07064545899629593,
-0.073845773935318,
-0.019191453233361244,
-0.016565781086683273,
-0.02563878707587719,
0.04340196028351784,
0.08029217272996902,
0.0010678501566872,
-0.02513556368649006,
0.005844359751790762,
0.09164892882108688,
-0.09353069216012955,
0.05708635225892067,
-0.03805840387940407,
0.03827666491270065,
0.06659375876188278,
0.04212701693177223,
0.045784227550029755,
-0.007195932324975729,
0.022441787645220757,
0.047257713973522186,
0.04550090804696083,
-0.08540958911180496,
0.11941687762737274,
-0.06288320571184158,
0.015118985436856747,
0.018967950716614723,
0.08286537230014801,
0.0027002571150660515,
-0.09427326917648315,
0.0722680613398552,
-0.00032216127146966755,
0.007095640990883112,
0.03951770067214966,
0.03300299867987633,
0.025361446663737297,
-0.0701112449169159,
0.004157721064984798,
0.12618570029735565,
-0.041620176285505295,
0.03563060238957405,
0.013284852728247643,
0.010560579597949982,
-0.03284931182861328,
0.04390993341803551,
-0.02988535910844803,
0.020727386698126793,
0.020534638315439224,
-0.008841744624078274,
0.01610325649380684,
0.04784892126917839,
-0.005937016103416681,
0.0066222455352544785,
-0.034694503992795944,
0.05610113590955734,
-0.03419966995716095,
0.011827691458165646,
-0.03650452569127083,
0.09272519499063492,
-0.12961345911026,
-0.03625034540891647,
0.03226904198527336,
-0.01653524488210678,
0.009076548740267754,
0.0005327254184521735,
-0.05452436953783035,
2.8602903635795473e-33,
0.02732996456325054,
0.07075393199920654,
0.003729619551450014,
-0.02529054880142212,
0.06324328482151031,
0.03881649300456047,
0.08618401736021042,
0.03013700805604458,
-0.001041608746163547,
-0.0625574141740799,
-0.0559609979391098,
0.015435861423611641,
0.014109605923295021,
0.015866266563534737,
0.05443748086690903,
-0.006440524011850357,
-0.012488149106502533,
0.09150738269090652,
0.03378577530384064,
-0.023812185972929,
0.06346520036458969,
-0.0449744313955307,
-0.09419415146112442,
0.0006186818354763091,
-0.014195559546351433,
-0.001392386038787663,
0.03821883350610733,
-0.0332517996430397,
0.0011840984225273132,
0.021326057612895966,
-0.08260039985179901,
-0.053689274936914444,
0.03928028419613838,
0.05929284170269966,
0.10903354734182358,
-0.013364600017666817,
-0.008375300094485283,
-0.042141299694776535,
0.005135160870850086,
-0.009022836573421955,
-0.026349147781729698,
0.005910313222557306,
0.014185146428644657,
-0.08749115467071533,
-0.13526694476604462,
-0.046016644686460495,
-0.03984551504254341,
0.00020609261991921812,
-0.023168453946709633,
-0.027451593428850174,
0.02800593711435795,
-0.007797377184033394,
-0.004359201993793249,
-0.03417983651161194,
-0.025027170777320862,
-0.03355640172958374,
0.008812589570879936,
0.056981369853019714,
0.11664800345897675,
0.1042608916759491,
-0.04515615850687027,
0.005461984779685736,
-0.01267575565725565,
0.08013094216585159,
0.05699041113257408,
-0.05428864061832428,
-0.09204733371734619,
0.0179611723870039,
0.1266663521528244,
-0.030483795329928398,
-0.0526297353208065,
0.008886483497917652,
-0.016001125797629356,
-0.006741788703948259,
0.0382048562169075,
-0.039178475737571716,
0.004771355539560318,
-0.03469597175717354,
-0.03260545805096626,
0.04394238069653511,
-0.005262005142867565,
0.10023485124111176,
-0.04660724103450775,
-0.04723609611392021,
0.024590447545051575,
-0.04375222697854042,
0.003418676322326064,
-0.04900999367237091,
-0.04291212931275368,
-0.026883218437433243,
-0.05987542122602463,
-0.030117759481072426,
-0.029527155682444572,
-0.0004986498970538378,
0.013515152037143707,
-2.0168154941729688e-33,
0.015363974496722221,
0.059831403195858,
-0.013092979788780212,
0.10077023506164551,
-0.031671177595853806,
-0.08333829790353775,
0.005403083749115467,
-0.012629839591681957,
-0.022759493440389633,
-0.013230489566922188,
-0.11454887688159943,
-0.049789175391197205,
0.08263231813907623,
-0.021116439253091812,
0.0445963516831398,
0.03721374273300171,
-0.08386629819869995,
-0.02771049179136753,
0.04307182878255844,
0.013803331181406975,
-0.01388805452734232,
0.06642387062311172,
-0.0851181223988533,
-0.014638915657997131,
0.024829652160406113,
0.01997586153447628,
-0.010186419822275639,
0.11931076645851135,
0.031052857637405396,
0.04646509885787964,
-0.025565920397639275,
-0.026566121727228165,
-0.060810819268226624,
-0.06651201844215393,
-0.036921609193086624,
0.10209102183580399,
0.0735991969704628,
0.08662206679582596,
-0.004969396162778139,
0.15426413714885712,
0.07181676477193832,
-0.1256328672170639,
-0.051566850394010544,
0.05114302411675453,
0.03303070738911629,
0.010522166267037392,
-0.04798677936196327,
-0.08191919326782227,
-0.0763518214225769,
-0.00782544631510973,
-0.04191666096448898,
0.03220338001847267,
-0.07940881699323654,
0.030766692012548447,
-0.018064776435494423,
-0.01561636570841074,
-0.003753932425752282,
0.043422602117061615,
0.0037309974431991577,
-0.029122021049261093,
0.045206934213638306,
-0.06582704186439514,
0.05708326771855354,
0.035453423857688904,
0.024287059903144836,
-0.0056067099794745445,
0.04302597418427467,
-0.023439759388566017,
-0.030107514932751656,
0.04951905086636543,
-0.030599376186728477,
-0.030988939106464386,
-0.03500935807824135,
0.025033535435795784,
0.00014798935444559902,
0.08910281956195831,
0.040518179535865784,
0.006207008846104145,
-0.01657637394964695,
0.018684370443224907,
-0.05567440763115883,
0.05468292161822319,
-0.007670215796679258,
-0.01190749928355217,
0.07933232933282852,
-0.006807521916925907,
0.1255199909210205,
0.08263494074344635,
-0.03976419195532799,
-0.00672526378184557,
-0.03645537048578262,
-0.010501005686819553,
0.08249109238386154,
0.01371341198682785,
-0.05149071663618088,
-4.6445730816913056e-8,
-0.06860093772411346,
0.053065478801727295,
0.0030088182538747787,
0.060705769807100296,
0.07861839234828949,
-0.0015784656861796975,
-0.02008766122162342,
0.0356961227953434,
-0.0022759526036679745,
0.014163387008011341,
0.045621324330568314,
-0.012948204763233662,
-0.0419670045375824,
-0.021605513989925385,
0.004804013762623072,
0.09234529733657837,
-0.062204062938690186,
-0.07906349003314972,
-0.057536184787750244,
0.01068050041794777,
0.019869763404130936,
0.009868985041975975,
0.04951862245798111,
0.0007752895471639931,
0.05045488849282265,
-0.07993567734956741,
0.048585522919893265,
0.07221560925245285,
0.03427409380674362,
0.016815392300486565,
-0.10787869244813919,
-0.0644836276769638,
0.07128767669200897,
-0.014538943767547607,
0.033471763134002686,
0.02523265965282917,
-0.03339184448122978,
-0.06419234722852707,
0.0026468734722584486,
-0.09026651829481125,
-0.014135854318737984,
-0.036990370601415634,
0.0018573260167613626,
0.0033088026102632284,
0.11521721631288528,
-0.004239468835294247,
-0.03174079209566116,
-0.03268546611070633,
0.016175327822566032,
0.008848840370774269,
-0.060683831572532654,
0.009926632978022099,
-0.04522356763482094,
0.06557755917310715,
0.07241398096084595,
-0.017623811960220337,
-0.03961096704006195,
-0.08994998782873154,
0.036132365465164185,
0.031070342287421227,
-0.0592450350522995,
-0.02369220368564129,
-0.10533551126718521,
-0.022810528054833412
] | 0 | 0 | 0 | 0 |
|
TabReD: Analyzing Pitfalls and Filling the Gaps in Tabular Deep Learning Benchmarks | https://openreview.net/forum?id=L14sqcrUC3 | [
"Ivan Rubachev",
"Nikolay Kartashev",
"Yury Gorishniy",
"Artem Babenko"
] | Spotlight | Advances in machine learning research drive progress in real-world applications.
To ensure this progress, it is important to understand the potential pitfalls on the way from a novel method's success on academic benchmarks to its practical deployment. In this work, we analyze existing tabular deep learning benchmarks and find two common characteristics of tabular data in typical industrial applications that are underrepresented in the datasets usually used for evaluation in the literature.
First, in real-world deployment scenarios, distribution of data often changes over time. To account for this distribution drift, time-based train/test splits should be used in evaluation. However, existing academic tabular datasets often lack timestamp metadata to enable such evaluation.
Second, a considerable portion of datasets in production settings stem from extensive data acquisition and feature engineering pipelines. This can have an impact on the absolute and relative number of predictive, uninformative, and correlated features compared to academic datasets.
In this work, we aim to understand how recent research advances in tabular deep learning transfer to these underrepresented conditions.
To this end, we introduce TabReD -- a collection of eight industry-grade tabular datasets.
We reassess a large number of tabular ML models and techniques on TabReD. We demonstrate that evaluation on both time-based data splits and richer feature sets leads to different methods ranking, compared to evaluation on random splits and smaller number of features, which are common in academic benchmarks. Furthermore, simple MLP-like architectures and GBDT show the best results on the TabReD datasets, while other methods are less effective in the new setting. | Tabular Data, Benchmarks, Reality Check, Tabular Deep Learning, Applications | We introduce TabReD, a collection of industry-grade tabular datasets, filling the gaps in academic benchmarks. Our evaluation reveals performance differences for various models and techniques in a new setting. | 13,784 | 2406.19380 | [
-0.0709138736128807,
-0.07633477449417114,
0.04901348054409027,
0.049543824046850204,
0.07262536138296127,
0.023013221099972725,
-0.07132302969694138,
-0.09199965745210648,
-0.08174216002225876,
-0.06048471853137016,
-0.05487588047981262,
-0.0515829361975193,
-0.05649540200829506,
0.006219205912202597,
-0.10261208564043045,
-0.041980598121881485,
0.07100917398929596,
0.035708218812942505,
-0.12363516539335251,
0.012839826755225658,
0.02132921852171421,
0.00907549075782299,
0.03986093029379845,
-0.010895208455622196,
0.08401593565940857,
0.03984905406832695,
0.004750634543597698,
-0.08619692176580429,
-0.05451096221804619,
-0.039252910763025284,
-0.04767954722046852,
0.01915297843515873,
-0.09303689002990723,
0.05303911864757538,
-0.026448803022503853,
0.003337040077894926,
-0.01813271827995777,
0.06351009756326675,
0.06757984310388565,
-0.0026833838783204556,
-0.01027852576225996,
-0.1141190454363823,
0.0179073978215456,
0.03884303569793701,
0.01778666488826275,
-0.001982256304472685,
0.05020364001393318,
-0.08784482628107071,
-0.007795525249093771,
0.06079567223787308,
-0.11717905849218369,
0.03409954905509949,
0.0030040726996958256,
0.06994806975126266,
-0.013880329206585884,
0.08592899143695831,
0.044663604348897934,
0.0007207918097265065,
-0.025874385610222816,
0.0690310001373291,
-0.05153527110815048,
-0.06264868378639221,
0.006002993322908878,
-0.03782547265291214,
-0.05938754603266716,
-0.04464287683367729,
-0.015757987275719643,
0.04421384260058403,
-0.00422202842310071,
-0.02295927330851555,
0.016248242929577827,
0.1375425010919571,
-0.06159285828471184,
-0.0077155414037406445,
-0.006383407395333052,
0.031524110585451126,
0.04303218796849251,
0.011831646785140038,
0.06188772991299629,
-0.13519802689552307,
0.04126325994729996,
0.02125825360417366,
-0.00040999220800586045,
0.0009071163949556649,
0.11355286091566086,
-0.02032218500971794,
0.07142504304647446,
0.07343290001153946,
-0.04485237970948219,
-0.02575649693608284,
-0.0054216752760112286,
-0.004322137217968702,
0.028468292206525803,
-0.007608615793287754,
-0.010808160528540611,
0.03476403281092644,
0.07073906809091568,
-0.03566797822713852,
0.015507551841437817,
0.0664423406124115,
-0.029525233432650566,
0.056048180907964706,
0.0558629035949707,
0.0183646809309721,
-0.014500902034342289,
-0.051691941916942596,
0.07466589659452438,
-0.0125072430819273,
0.04441145062446594,
-0.06435342133045197,
0.12643049657344818,
0.01905231550335884,
-0.013334943912923336,
-0.04757533594965935,
0.03749698027968407,
-0.01487512607127428,
-0.08990002423524857,
-0.07097869366407394,
0.013012039475142956,
0.10783357918262482,
-0.06231227144598961,
0.013862619176506996,
-0.00622629513964057,
0.061861664056777954,
-0.05367663875222206,
0.07180847972631454,
-0.047295574098825455,
4.337348239213692e-33,
0.01716090552508831,
0.0018074061954393983,
0.004480482079088688,
-0.002011877251788974,
0.07258395850658417,
-0.08112192153930664,
-0.026795586571097374,
0.07355387508869171,
-0.006046060472726822,
0.04497021809220314,
-0.1159973293542862,
-0.0007522177184000611,
-0.0851554423570633,
0.07380544394254684,
0.041307806968688965,
0.047148607671260834,
-0.07823128253221512,
0.04888032376766205,
-0.006220247130841017,
-0.009460216388106346,
0.017995504662394524,
-0.05563714727759361,
-0.04753918945789337,
0.0014281031908467412,
0.05374691262841225,
-0.031120698899030685,
0.011688871309161186,
0.035719383507966995,
0.0026134310755878687,
0.0319199338555336,
-0.01513300184160471,
0.022822272032499313,
-0.013162624090909958,
0.0046228449791669846,
0.0431164875626564,
-0.03314855694770813,
-0.04303107410669327,
-0.010319185443222523,
0.047236040234565735,
0.03690777346491814,
0.009883498772978783,
0.04415220767259598,
-0.048472195863723755,
-0.016299225389957428,
-0.023166542872786522,
-0.09920449554920197,
0.0390603132545948,
-0.008459876291453838,
-0.005768387112766504,
-0.018875140696763992,
0.05200876295566559,
0.0009929608786478639,
0.017326977103948593,
-0.043668173253536224,
-0.013869908638298512,
-0.014577290043234825,
0.026550747454166412,
0.010733606293797493,
0.055786434561014175,
0.07743139564990997,
-0.05505015701055527,
0.05081789940595627,
-0.04038939252495766,
0.02505761943757534,
-0.10514660179615021,
0.09219079464673996,
0.0377141498029232,
0.0855892226099968,
0.05671568587422371,
-0.046911366283893585,
-0.029988564550876617,
0.010036202147603035,
-0.03519825637340546,
-0.09076764434576035,
0.09323595464229584,
0.0028691578190773726,
0.060931798070669174,
-0.04419538751244545,
0.04502546787261963,
-0.01145726628601551,
0.09090139716863632,
0.10818042606115341,
-0.027900012210011482,
-0.05846491456031799,
-0.0377676784992218,
0.007428427692502737,
0.01800175942480564,
-0.10011893510818481,
-0.0003054903936572373,
0.012015284970402718,
-0.02274388074874878,
0.016909020021557808,
-0.058591198176145554,
-0.03746050223708153,
0.04056158661842346,
-3.726582313019556e-33,
-0.018572259694337845,
0.005159176420420408,
-0.06126847490668297,
0.06544054299592972,
0.013290810398757458,
0.001105594914406538,
-0.015101910568773746,
0.026526961475610733,
0.04156701639294624,
-0.028585515916347504,
0.032826054841279984,
-0.08374225348234177,
-0.00008284794603241608,
-0.04087606072425842,
-0.013970188796520233,
0.014186686836183071,
-0.05564112961292267,
-0.045669540762901306,
-0.03034469485282898,
0.0008293981663882732,
0.05282292142510414,
0.0362786203622818,
-0.050326552242040634,
0.026860959827899933,
-0.050656404346227646,
-0.04822387918829918,
-0.021780401468276978,
0.049562081694602966,
-0.030442306771874428,
-0.04758133366703987,
0.0008036819635890424,
-0.0415242500603199,
-0.0857313722372055,
-0.030316883698105812,
-0.02801409177482128,
0.04962017387151718,
0.044516805559396744,
-0.0671921968460083,
-0.04828048497438431,
0.0360652357339859,
0.09580060094594955,
0.037509217858314514,
-0.06011151522397995,
0.055534183979034424,
-0.037934809923172,
0.0323532372713089,
-0.014859382063150406,
-0.02659286931157112,
0.025276973843574524,
0.035473983734846115,
-0.00293279648758471,
-0.02871677838265896,
-0.0553494468331337,
-0.01226517092436552,
0.0075176735408604145,
-0.026001153513789177,
-0.007262190338224173,
0.011164996773004532,
-0.020621150732040405,
0.06408553570508957,
-0.10256290435791016,
-0.0029632726218551397,
0.029173800721764565,
-0.02433030866086483,
0.005744160618633032,
-0.010071846656501293,
0.014368983916938305,
0.009267098270356655,
-0.005780564155429602,
0.03350558131933212,
0.021016515791416168,
-0.026836439967155457,
-0.0048037683591246605,
-0.003309406340122223,
-0.05294313654303551,
0.017557214945554733,
0.04509427398443222,
-0.026937900111079216,
-0.08707553148269653,
-0.021027816459536552,
0.0792093575000763,
-0.025838052853941917,
-0.004819655325263739,
-0.012129785493016243,
0.02627950720489025,
0.124168761074543,
0.05920940265059471,
0.0739082619547844,
-0.010825044475495815,
-0.007873794063925743,
-0.04573439434170723,
-0.03719598427414894,
-0.12057865411043167,
0.042055729776620865,
-0.008423100225627422,
-4.172492040765974e-8,
-0.04319775849580765,
0.12927088141441345,
0.020216723904013634,
0.04999673739075661,
0.049856822937726974,
-0.06459652632474899,
0.03787531331181526,
0.1581880897283554,
0.006352864671498537,
-0.03693489730358124,
0.10679656267166138,
0.014077923260629177,
-0.033014412969350815,
0.0006402398576028645,
0.012132604606449604,
0.1205078586935997,
0.02609804831445217,
-0.02793642319738865,
-0.005756499711424112,
-0.011699513532221317,
0.07565807551145554,
0.06253364682197571,
0.10823763161897659,
0.0054098633117973804,
-0.024394337087869644,
-0.09213588386774063,
-0.06419585645198822,
0.06901976466178894,
0.054172974079847336,
-0.05958452820777893,
0.02632797136902809,
0.021274955943226814,
0.03398365527391434,
-0.08753761649131775,
0.05947129428386688,
0.05063247308135033,
0.048627737909555435,
0.012001017108559608,
0.011859630234539509,
0.039564285427331924,
-0.023103808984160423,
0.02614717185497284,
-0.07086006551980972,
-0.011330794543027878,
-0.028257982805371284,
-0.027162333950400352,
-0.06957037001848221,
0.0116583202034235,
0.014009465463459492,
-0.01753801852464676,
0.06433581560850143,
-0.02506970800459385,
0.03740771487355232,
0.0672781765460968,
0.064728282392025,
0.04204869642853737,
-0.014780319295823574,
-0.06246298551559448,
0.018925495445728302,
0.03694707155227661,
0.024311339482665062,
-0.05355334281921387,
-0.07715987414121628,
0.06845519691705704
] | https://github.com/yandex-research/tabred | 69 | 0 | 0 | 0 |
Improving the Sparse Structure Learning of Spiking Neural Networks from the View of Compression Efficiency | https://openreview.net/forum?id=gcouwCx7dG | [
"Jiangrong Shen",
"Qi Xu",
"Gang Pan",
"Badong Chen"
] | Spotlight | The human brain utilizes spikes for information transmission and dynamically reorganizes its network structure to boost energy efficiency and cognitive capabilities throughout its lifespan. Drawing inspiration from this spike-based computation, Spiking Neural Networks (SNNs) have been developed to construct event-driven models that emulate this efficiency. Despite these advances, deep SNNs continue to suffer from over-parameterization during training and inference, a stark contrast to the brain’s ability to self-organize. Furthermore, existing sparse SNNs are challenged by maintaining optimal pruning levels due to a static pruning ratio, resulting in either under or over-pruning.
In this paper, we propose a novel two-stage dynamic structure learning approach for deep SNNs, aimed at maintaining effective sparse training from scratch while optimizing compression efficiency.
The first stage evaluates the compressibility of existing sparse subnetworks within SNNs using the PQ index, which facilitates an adaptive determination of the rewiring ratio for synaptic connections based on data compression insights. In the second stage, this rewiring ratio critically informs the dynamic synaptic connection rewiring process, including both pruning and regrowth. This approach significantly improves the exploration of sparse structures training in deep SNNs, adapting sparsity dynamically from the point view of compression efficiency.
Our experiments demonstrate that this sparse training approach not only aligns with the performance of current deep SNNs models but also significantly improves the efficiency of compressing sparse SNNs. Crucially, it preserves the advantages of initiating training with sparse models and offers a promising solution for implementing Edge AI on neuromorphic hardware. | spiking neural networks | null | 13,685 | 2502.13572 | [
-0.05661286041140556,
-0.03195523843169212,
0.02472963184118271,
0.0546659380197525,
-0.03618103265762329,
0.01350307371467352,
-0.0033240285702049732,
-0.06280139833688736,
0.04377887770533562,
-0.06404995173215866,
-0.006080621853470802,
-0.007861383259296417,
-0.02270558848977089,
0.0794944018125534,
-0.034455958753824234,
-0.008183605968952179,
-0.004962574224919081,
0.06495293229818344,
-0.07583685964345932,
-0.04141739755868912,
0.046699147671461105,
-0.1068260595202446,
-0.013126862235367298,
0.0759514793753624,
0.053856950253248215,
0.062026478350162506,
-0.018022147938609123,
-0.060693591833114624,
0.03085467778146267,
0.005166388116776943,
0.05810188874602318,
-0.07522417604923248,
-0.002658743178471923,
-0.04407874494791031,
0.02344452030956745,
0.008856913074851036,
-0.03605971485376358,
0.023404676467180252,
0.010706959292292595,
0.012516239657998085,
0.05134292319417,
-0.002515623811632395,
0.0033828718587756157,
0.07208860665559769,
0.03955591842532158,
0.05488713085651398,
0.005959379952400923,
-0.0659150555729866,
-0.04020394757390022,
-0.01740824244916439,
0.01457968819886446,
0.0513540655374527,
0.02373139001429081,
0.04649334400892258,
0.0686265230178833,
0.08676718175411224,
-0.06194286420941353,
0.03034476935863495,
-0.04606273025274277,
0.009795492514967918,
0.005896188784390688,
-0.06027545779943466,
0.036796022206544876,
-0.0037009548395872116,
-0.034148991107940674,
0.012158427387475967,
0.0781145989894867,
0.09748199582099915,
0.031129002571105957,
0.04597558453679085,
0.025750160217285156,
0.0779886543750763,
-0.014458175748586655,
0.03761420398950577,
0.07340694218873978,
0.003901680698618293,
-0.000755344342906028,
-0.012839186936616898,
0.03972204029560089,
-0.06074192747473717,
0.040459517389535904,
0.033266350626945496,
-0.031009985134005547,
-0.025872519239783287,
0.12958140671253204,
0.008477517403662205,
-0.021595323458313942,
0.05619198456406593,
-0.041992951184511185,
-0.02973032556474209,
-0.06421440094709396,
-0.02523694559931755,
-0.007704886142164469,
-0.04072340950369835,
0.02172415517270565,
0.016730012372136116,
-0.012642988935112953,
-0.006064965855330229,
-0.0729295015335083,
0.06414590775966644,
-0.021752027794718742,
-0.009604010730981827,
0.006761110387742519,
0.006986089516431093,
0.0335795059800148,
-0.02954631857573986,
0.016065126284956932,
0.06703799217939377,
0.013400036841630936,
0.01961306296288967,
0.07852260768413544,
0.056493841111660004,
0.012276673689484596,
0.04581204429268837,
0.03292294591665268,
-0.029575040563941002,
-0.04226967319846153,
0.05608240142464638,
0.08129590004682541,
0.059533946216106415,
0.04418566823005676,
-0.02538934163749218,
-0.0806986466050148,
-0.015078315511345863,
0.00127853371668607,
-0.02836475521326065,
-0.11147120594978333,
3.3524114408932826e-33,
0.008307641372084618,
0.07051178812980652,
0.04612990841269493,
-0.039497021585702896,
0.039630234241485596,
-0.02277807518839836,
0.004577340092509985,
0.03361964970827103,
0.0310208760201931,
0.07982886582612991,
-0.16508139669895172,
-0.015592323616147041,
-0.023868661373853683,
0.08877192437648773,
-0.01790779083967209,
-0.047317080199718475,
-0.043772485107183456,
0.041099805384874344,
-0.010345078073441982,
-0.16767939925193787,
0.019008304923772812,
0.01680777408182621,
-0.018049608916044235,
-0.049380455166101456,
0.007775363977998495,
-0.07755450904369354,
0.03828508034348488,
-0.04369046911597252,
-0.05738169327378273,
0.006341450847685337,
-0.061915840953588486,
0.049448851495981216,
-0.04720091447234154,
-0.05313127487897873,
0.060143254697322845,
-0.058100033551454544,
0.04195129871368408,
-0.016167497262358665,
0.015701716765761375,
0.0019544686656445265,
0.021972140297293663,
0.00955521035939455,
-0.003954593557864428,
-0.06644844263792038,
-0.09177996218204498,
-0.006112800911068916,
0.08202345669269562,
0.015259488485753536,
-0.0576292909681797,
-0.0930965319275856,
0.025655442848801613,
-0.05191468819975853,
0.03212929889559746,
0.02898593805730343,
-0.01970498450100422,
-0.07280511409044266,
0.004791208542883396,
-0.003553517162799835,
0.0918571799993515,
0.09073839336633682,
-0.05485633760690689,
0.00560274813324213,
-0.047012634575366974,
0.09888613969087601,
0.09350727498531342,
0.051789991557598114,
0.018896400928497314,
0.04339182749390602,
0.04507707431912422,
-0.08675514161586761,
0.06787053495645523,
0.043605413287878036,
-0.028615398332476616,
-0.10137304663658142,
0.01381123811006546,
0.016326092183589935,
0.08819206804037094,
0.02087431587278843,
-0.12753626704216003,
0.008490215055644512,
-0.06966060400009155,
-0.050350431352853775,
0.00606419425457716,
0.07109788805246353,
-0.020649177953600883,
0.020292978733778,
0.08072759956121445,
-0.033154502511024475,
-0.01607605442404747,
0.024027591571211815,
-0.010628112591803074,
-0.0720449909567833,
0.058660946786403656,
-0.11542075127363205,
-0.027148369699716568,
-1.751607805580137e-33,
-0.038173943758010864,
0.04572900012135506,
-0.035316962748765945,
0.02588726580142975,
0.011366992257535458,
-0.02371620573103428,
0.02319582924246788,
0.011092747561633587,
-0.08504467457532883,
-0.0608532689511776,
-0.016926893964409828,
-0.040830012410879135,
0.005590210668742657,
-0.040500201284885406,
0.049506016075611115,
0.032870516180992126,
-0.12712378799915314,
-0.005186385940760374,
0.04027875140309334,
-0.012985613197088242,
-0.005946527235209942,
0.05885954201221466,
-0.09114252775907516,
-0.032606810331344604,
-0.0076148370280861855,
0.04290683567523956,
-0.023097531870007515,
0.04900859668850899,
0.0519036166369915,
-0.008487740531563759,
-0.11084092408418655,
-0.027624892070889473,
-0.1269097775220871,
-0.0759173110127449,
0.0054159946739673615,
0.06940090656280518,
0.025735341012477875,
-0.02521340548992157,
0.017344776540994644,
0.00034822075394913554,
0.04824460670351982,
0.024051833897829056,
-0.03325103968381882,
0.07961694151163101,
0.09894157201051712,
-0.014424018561840057,
-0.07462918013334274,
-0.0013870218535885215,
-0.023483749479055405,
0.08881715685129166,
-0.004045497160404921,
0.018474377691745758,
-0.10081585496664047,
-0.04422752559185028,
-0.034926287829875946,
-0.08137158304452896,
-0.030151009559631348,
-0.01554634515196085,
0.04013063758611679,
-0.026703471317887306,
-0.0356605239212513,
-0.11317688226699829,
-0.045891329646110535,
0.052191730588674545,
-0.023862458765506744,
-0.04402047395706177,
0.039241813123226166,
0.007505467627197504,
0.0004255165986251086,
-0.015997083857655525,
-0.04788509011268616,
0.006765919737517834,
0.02469516173005104,
-0.0011793371522799134,
-0.0366290844976902,
-0.011973408982157707,
-0.004148097708821297,
-0.019542738795280457,
0.017977844923734665,
0.0001755928242346272,
0.012690444476902485,
-0.028609437867999077,
-0.05156753212213516,
-0.025990208610892296,
0.04794654622673988,
0.14649848639965057,
0.0465298630297184,
0.08489450812339783,
0.06116935610771179,
0.02901807799935341,
-0.014172463677823544,
-0.009208740666508675,
0.03959769383072853,
0.0656377375125885,
-0.0039021256379783154,
-4.1018296315087355e-8,
-0.02993464283645153,
0.06840178370475769,
0.0010214932262897491,
0.026386866346001625,
0.08221302181482315,
-0.04162435978651047,
0.03209461271762848,
0.08830142021179199,
0.01503175962716341,
-0.07223770022392273,
0.08442305028438568,
-0.0951559916138649,
-0.008887331932783127,
0.020423710346221924,
0.01164667122066021,
0.10746976733207703,
0.034920234233140945,
-0.006844361778348684,
-0.01614510454237461,
-0.030410094186663628,
0.012341995723545551,
0.05714686959981918,
-0.0445093959569931,
0.055946968495845795,
0.0047819409519433975,
-0.06327097117900848,
-0.013158638030290604,
0.08237268030643463,
0.02087409421801567,
-0.1272345781326294,
0.0064224423840641975,
0.039682790637016296,
0.042962126433849335,
0.016901668161153793,
-0.05114219710230827,
0.08324141800403595,
0.04323316365480423,
0.08416881412267685,
-0.0033079672139137983,
-0.07687772065401077,
-0.09081222862005234,
-0.017539771273732185,
0.03690284490585327,
0.004444062244147062,
-0.07893816381692886,
-0.0667080506682396,
-0.02960476092994213,
-0.018268702551722527,
0.027576018124818802,
-0.0030961381271481514,
0.00029045104747638106,
0.057569608092308044,
-0.005981349386274815,
0.004794703330844641,
0.044761426746845245,
-0.0007804377819411457,
0.0308254212141037,
0.038356706500053406,
-0.039111535996198654,
0.03163887932896614,
0.06323384493589401,
0.038901783525943756,
-0.10232260823249817,
0.0038063295651227236
] | 0 | 0 | 0 | 0 |
|
JudgeLM: Fine-tuned Large Language Models are Scalable Judges | https://openreview.net/forum?id=xsELpEPn4A | [
"Lianghui Zhu",
"Xinggang Wang",
"Xinlong Wang"
] | Spotlight | Evaluating Large Language Models (LLMs) in open-ended scenarios is challenging because existing benchmarks and metrics can not measure them comprehensively. To address this problem, we propose to fine-tune LLMs as scalable judges (JudgeLM) to evaluate LLMs efficiently and effectively in open-ended benchmarks. We first propose a comprehensive, large-scale, high-quality dataset containing task seeds, LLMs-generated answers, and GPT-4-generated judgments for fine-tuning high-performance judges, as well as a new benchmark for evaluating the judges. We train JudgeLM at different scales from 7B, 13B, to 33B parameters, and conduct a systematic analysis of its capabilities and behaviors. We then analyze the key biases in fine-tuning LLM as a judge and consider them as position bias, knowledge bias, and format bias. To address these issues, JudgeLM introduces a bag of techniques including swap augmentation, reference support, and reference drop, which clearly enhance the judge's performance. JudgeLM obtains the state-of-the-art judge performance on both the existing PandaLM benchmark and our proposed new benchmark. Our JudgeLM is efficient and the JudgeLM-7B only needs 3 minutes to judge 5K samples with 8 A100 GPUs. JudgeLM obtains high agreement with the teacher judge, achieving an agreement exceeding 90% that even surpasses human-to-human agreement. JudgeLM also demonstrates extended capabilities in being judges of the single answer, multimodal models, multiple answers, multi-turn chat, etc. | LLM Judging | We propose a high-quality LLM judge dataset and a series of strong LLM judges to address the biases in LLM judging. | 13,655 | 2310.17631 | [
-0.03363100066781044,
-0.049636583775281906,
-0.05255745351314545,
0.008088720962405205,
0.039707500487565994,
0.024856679141521454,
0.0018334555206820369,
0.021873030811548233,
-0.02616797387599945,
-0.037093911319971085,
-0.08732421696186066,
-0.046657856553792953,
0.002358492696657777,
-0.02119232527911663,
-0.006923082750290632,
0.05243156850337982,
0.10597100853919983,
0.05901525542140007,
-0.07082536071538925,
-0.08498386293649673,
-0.015250063501298428,
0.058255884796381,
0.02526792325079441,
-0.026604000478982925,
0.00525025324895978,
-0.07066347450017929,
-0.06113024801015854,
-0.045331161469221115,
-0.01847672648727894,
-0.07985010743141174,
-0.011348818428814411,
0.09815753251314163,
-0.03400671109557152,
0.03136749193072319,
-0.04035532847046852,
0.04491886869072914,
-0.03128260374069214,
-0.009016153402626514,
0.0050238086842000484,
-0.04471207410097122,
0.05926290154457092,
0.05619435012340546,
0.03226287662982941,
0.021366942673921585,
0.05517836660146713,
-0.017129337415099144,
0.007988754659891129,
-0.014784636907279491,
-0.024550052359700203,
0.011999815702438354,
-0.08187997341156006,
-0.015552196651697159,
0.014297365210950375,
0.05792444571852684,
-0.028732256963849068,
0.034482598304748535,
0.03735971823334694,
0.0252431258559227,
0.004935003351420164,
-0.07932648807764053,
-0.04462840035557747,
-0.11417938768863678,
-0.06899082660675049,
-0.019276181235909462,
-0.061707645654678345,
-0.049547791481018066,
0.02184244990348816,
0.012061625719070435,
-0.066607765853405,
0.01631726138293743,
-0.036547593772411346,
0.10406908392906189,
0.026362476870417595,
0.09736432880163193,
-0.008405419066548347,
0.07320541888475418,
0.02323569916188717,
-0.0363953560590744,
0.1104109138250351,
-0.06312583386898041,
-0.023266565054655075,
-0.055055830627679825,
0.02543003484606743,
-0.0054467832669615746,
0.12656423449516296,
-0.043416980654001236,
0.033149171620607376,
0.003055483801290393,
0.042489610612392426,
0.06271038949489594,
0.033759620040655136,
-0.003636633977293968,
0.07142532616853714,
-0.02409532479941845,
0.016187312081456184,
0.019723767414689064,
0.023830363526940346,
-0.08647643029689789,
0.016725847497582436,
0.11514604091644287,
0.017920566722750664,
0.15604476630687714,
0.010810717940330505,
-0.04969911649823189,
-0.05037129297852516,
-0.03169027715921402,
0.0062753427773714066,
0.052168961614370346,
-0.02594859153032303,
-0.11324749886989594,
0.028193477541208267,
0.05777395889163017,
0.07002181559801102,
-0.007759676780551672,
0.03012000024318695,
-0.005204546265304089,
-0.030380677431821823,
-0.02449999935925007,
-0.041330937296152115,
0.041301921010017395,
-0.08270289748907089,
0.06244804337620735,
0.005451645236462355,
-0.018776873126626015,
-0.025218959897756577,
0.0426391139626503,
-0.02267240732908249,
9.95802682390327e-34,
0.06297865509986877,
0.05055389180779457,
0.0006163361831568182,
-0.016681451350450516,
0.013726537115871906,
-0.005456928163766861,
0.017438054084777832,
0.05500313267111778,
-0.12274560332298279,
-0.036450646817684174,
-0.02047557942569256,
0.04539231210947037,
-0.03333752974867821,
0.02772430330514908,
0.03913063183426857,
0.029499882832169533,
-0.050447408109903336,
0.0989774763584137,
-0.054065097123384476,
-0.0036884143482893705,
0.05210265889763832,
0.01325406413525343,
-0.0015855884412303567,
-0.042093999683856964,
0.02475987747311592,
-0.0029213009402155876,
0.04970082268118858,
-0.020968303084373474,
-0.09207841008901596,
0.007769342511892319,
-0.11411623656749725,
-0.0633312314748764,
0.052525512874126434,
0.07430130243301392,
0.061156146228313446,
-0.049161940813064575,
-0.04876026138663292,
-0.0288569163531065,
0.06779124587774277,
-0.019703587517142296,
-0.05738751217722893,
0.042841870337724686,
0.052201054990291595,
-0.003602738957852125,
-0.09428330510854721,
-0.021410075947642326,
-0.0016748573398217559,
-0.027045033872127533,
0.020101165398955345,
-0.0222341138869524,
0.05573931336402893,
-0.00515366904437542,
0.04681485891342163,
0.07016659528017044,
-0.016679465770721436,
0.029429586604237556,
0.032628387212753296,
0.04220554232597351,
0.024654891341924667,
0.10309586673974991,
0.014425577595829964,
0.10121230036020279,
-0.03252139687538147,
0.065597765147686,
0.03532131761312485,
0.0009909102227538824,
-0.0525956004858017,
0.06294668465852737,
0.06628362834453583,
-0.049707766622304916,
-0.003004227066412568,
-0.04160380735993385,
0.01671069860458374,
0.024706512689590454,
0.009558620862662792,
-0.02512611821293831,
0.04253561049699783,
-0.046297721564769745,
0.004649153910577297,
-0.01993538811802864,
0.04856549948453903,
0.08364500850439072,
0.03850376978516579,
-0.05718166381120682,
-0.04225220903754234,
-0.12089309841394424,
-0.006635117810219526,
-0.060037173330783844,
0.0042204540222883224,
-0.004204579163342714,
-0.05973697453737259,
-0.03286825865507126,
-0.02709933929145336,
-0.024005362764000893,
0.04239721596240997,
-4.490636375980765e-34,
-0.004802511539310217,
-0.05810936912894249,
-0.004663544241338968,
0.14017577469348907,
-0.06764442473649979,
-0.06950541585683823,
-0.015954220667481422,
0.03373502939939499,
-0.015738146379590034,
-0.06758090108633041,
-0.046783700585365295,
-0.0695161446928978,
0.06967392563819885,
0.0038547597359865904,
0.031205810606479645,
-0.04819284379482269,
-0.06705030053853989,
0.02516569197177887,
-0.008745276369154453,
0.05605480074882507,
0.05390280485153198,
0.08798045665025711,
-0.07193153351545334,
0.07331083714962006,
-0.034126367419958115,
-0.004706340376287699,
-0.02551550418138504,
-0.03841891139745712,
0.0037681912072002888,
-0.004675768315792084,
-0.013394416309893131,
-0.023870263248682022,
-0.02200157381594181,
0.020125966519117355,
0.005310805514454842,
-0.015134825371205807,
0.14201976358890533,
-0.03378070145845413,
-0.006256321910768747,
0.13237616419792175,
0.045925598591566086,
-0.00026144026196561754,
-0.05189971253275871,
-0.01948932558298111,
0.024210534989833832,
0.04759446904063225,
-0.09496861696243286,
0.007721145637333393,
0.012258998118340969,
-0.008564069867134094,
-0.07099983841180801,
-0.0025255107320845127,
-0.030845660716295242,
0.05391721799969673,
0.014920715242624283,
-0.14848178625106812,
-0.04324223846197128,
-0.05554557964205742,
0.005142930429428816,
0.04090753197669983,
-0.013980750925838947,
-0.011385506018996239,
0.016357123851776123,
-0.09494353085756302,
0.04520079120993614,
-0.03132430464029312,
-0.029139239341020584,
-0.04427220672369003,
0.001015305519104004,
-0.003047528676688671,
-0.0009364046272821724,
-0.05502558872103691,
-0.01908920891582966,
0.018389010801911354,
-0.024837879464030266,
0.008756929077208042,
0.04415256157517433,
-0.008522827178239822,
-0.008433098904788494,
-0.03259142115712166,
0.03250695392489433,
-0.009690966457128525,
0.047548968344926834,
0.019608575850725174,
0.01592555269598961,
0.055785514414310455,
0.057269711047410965,
0.07058292627334595,
0.0007376022986136377,
0.06427686661481857,
-0.06425383687019348,
0.0013288211775943637,
0.020542534068226814,
0.04393579438328743,
-0.05270356312394142,
-4.1185341359550875e-8,
-0.026396000757813454,
0.005532876122742891,
-0.021942123770713806,
0.13224297761917114,
-0.0601535402238369,
-0.03294597938656807,
-0.0778636559844017,
0.0945315882563591,
0.011852730996906757,
-0.0012754028430208564,
0.06677959114313126,
-0.0008484659483656287,
-0.05844252556562424,
-0.04930149391293526,
0.02470759116113186,
0.0827900767326355,
-0.0071785347536206245,
0.04776379466056824,
-0.05304916203022003,
0.004543013870716095,
0.00022616850037593395,
0.07663460820913315,
0.03127708286046982,
-0.02350856363773346,
0.04564569890499115,
-0.08102545887231827,
-0.02474002167582512,
0.09664634615182877,
0.08663414418697357,
-0.01960744708776474,
0.04584565386176109,
0.004212810657918453,
-0.018804090097546577,
-0.07281612604856491,
0.08113488554954529,
0.04210891202092171,
-0.06271806359291077,
0.05316044017672539,
0.09540030360221863,
0.06191962957382202,
0.02910173125565052,
0.08693643659353256,
-0.0948442816734314,
0.056415457278490067,
0.0017495423089712858,
-0.09052899479866028,
-0.09550676494836807,
0.0023985346779227257,
-0.012959256768226624,
-0.06388555467128754,
0.003360806731507182,
-0.09016044437885284,
-0.0025065129157155752,
0.02056995779275894,
0.02715166099369526,
0.08802931010723114,
-0.019137894734740257,
-0.001079382374882698,
0.025560691952705383,
0.036484185606241226,
0.011950845830142498,
-0.06867356598377228,
-0.01868320256471634,
0.013956363312900066
] | https://github.com/baaivision/judgelm | 360 | 0 | 0 | 0 |
Direct Post-Training Preference Alignment for Multi-Agent Motion Generation Model Using Implicit Feedback from Pre-training Demonstrations | https://openreview.net/forum?id=8UFG9D8xeU | [
"Thomas Tian",
"Kratarth Goel"
] | Spotlight | Recent advancements in Large Language Models (LLMs) have revolutionized motion generation models in embodied applications such as autonomous driving and robotic manipulation. While LLM-type auto-regressive motion generation models benefit from training scalability, there remains a discrepancy between their token prediction objectives and human preferences. As a result, models pre-trained solely with token-prediction objectives often generate behaviors that deviate from what humans would prefer, making post-training preference alignment crucial for producing human-preferred motions. Unfortunately, post-training alignment requires extensive preference rankings of motions generated by the pre-trained model, which are costly and time-consuming to annotate, especially in multi-agent motion generation settings. Recently, there has been growing interest in leveraging expert demonstrations previously used during pre-training to scalably generate preference data for post-training alignment. However, these methods often adopt an adversarial assumption, treating all pre-trained model-generated samples as unpreferred examples and relying solely on pre-training expert demonstrations to construct preferred examples. This adversarial approach overlooks the valuable signal provided by preference rankings among the model's own generations, ultimately reducing alignment effectiveness and potentially leading to misaligned behaviors. In this work, instead of treating all generated samples as equally bad, we propose a principled approach that leverages implicit preferences encoded in pre-training expert demonstrations to construct preference rankings among the pre-trained model's generations, offering more nuanced preference alignment guidance with zero human cost. We apply our approach to large-scale traffic simulation (more than 100 agents) and demonstrate its effectiveness in improving the realism of pre-trained model's generated behaviors, making a lightweight 1M motion generation model comparable to state-of-the-art large imitation-based models by relying solely on implicit feedback from pre-training demonstrations, without requiring additional post-training human preference annotations or incurring high computational costs. Furthermore, we provide an in-depth analysis of preference data scaling laws and their effects on over-optimization, offering valuable insights for future studies. | Efficient Post-training Preference Alignment, Alignment from demonstrations, Multi-agent Motion Generation | We propose an efficient post-training alignment approach that significantly improves the pre-trained motion generation model’s quality without requiring additional post-training human preference annotation or expansive compute. | 13,515 | 2503.20105 | [
-0.042067378759384155,
-0.11026836186647415,
0.06044331192970276,
0.029018376022577286,
0.025598876178264618,
0.08526059985160828,
-0.02922169677913189,
0.0013365483609959483,
0.05983851104974747,
0.0008642533794045448,
-0.017936520278453827,
-0.05718313530087471,
0.005409069824963808,
0.01900157332420349,
0.006707802414894104,
0.0032834550365805626,
0.15234492719173431,
0.07966224104166031,
-0.08103863149881363,
-0.012369589880108833,
0.031624533236026764,
-0.020480550825595856,
0.06661991775035858,
0.03638192638754845,
-0.003030474530532956,
-0.05656138435006142,
0.014485113322734833,
-0.008733615279197693,
0.004408818203955889,
-0.03932420536875725,
0.018668094649910927,
-0.012839230708777905,
0.019126243889331818,
-0.037039078772068024,
-0.06562251597642899,
0.04148288071155548,
-0.1122778058052063,
-0.0026870141737163067,
0.041990555822849274,
-0.05034412443637848,
-0.01486513763666153,
-0.05375240370631218,
0.05145694315433502,
-0.033436160534620285,
0.037592701613903046,
0.07851610332727432,
0.012655750848352909,
-0.029635809361934662,
0.00030501032597385347,
0.022136889398097992,
-0.05544636398553848,
-0.04040353745222092,
0.05415937677025795,
-0.0416344553232193,
-0.007756163831800222,
0.012016226537525654,
0.037333160638809204,
0.05336213856935501,
0.0812162384390831,
-0.04632949084043503,
-0.01302121952176094,
-0.07215015590190887,
-0.013758717104792595,
-0.01868094503879547,
-0.04176788404583931,
-0.07254497706890106,
0.04472719132900238,
0.004327747505158186,
-0.021398508921265602,
0.10329742729663849,
-0.0007231882191263139,
0.03857863321900368,
0.05796746909618378,
0.014924761839210987,
-0.011184634640812874,
0.07352741062641144,
0.014042890630662441,
-0.0040087709203362465,
-0.02831551991403103,
-0.08033790439367294,
0.024996669963002205,
-0.015447800979018211,
0.01655985228717327,
0.05203121155500412,
0.021971777081489563,
0.0445873960852623,
-0.027075735852122307,
0.05753292143344879,
0.09681004285812378,
-0.009540186263620853,
-0.14001527428627014,
-0.06623782217502594,
-0.01036817580461502,
-0.019888652488589287,
0.017492657527327538,
0.062235862016677856,
-0.05934933200478554,
-0.03966231271624565,
-0.051546093076467514,
0.0768895298242569,
0.05141587182879448,
0.07764733582735062,
0.0015449139755219221,
0.08638489246368408,
0.048664551228284836,
-0.059660904109478,
0.01647925190627575,
0.039167456328868866,
0.12245747447013855,
-0.012132976204156876,
0.06773041188716888,
-0.020150255411863327,
0.014483204111456871,
0.05089262127876282,
-0.00017563137225806713,
-0.03905784711241722,
-0.012693622149527073,
-0.010096991434693336,
-0.0030967509374022484,
0.03040730208158493,
0.002514938823878765,
-0.055518824607133865,
0.04281015694141388,
0.0236428901553154,
0.06940242648124695,
-0.05613734573125839,
-0.04255909100174904,
1.938924299751607e-33,
0.030809251591563225,
0.04793698340654373,
0.13177625834941864,
-0.009269289672374725,
0.04016921669244766,
-0.05057910084724426,
-0.002816007472574711,
-0.014259420335292816,
0.0456966869533062,
-0.03945085406303406,
-0.013285933062434196,
-0.004122368060052395,
-0.0476309172809124,
0.11091192066669464,
-0.049860358238220215,
0.027342040091753006,
-0.013108484447002411,
0.023362042382359505,
-0.051937878131866455,
0.04315660521388054,
0.07024648040533066,
0.05771557241678238,
-0.07404836267232895,
-0.04301625117659569,
-0.02521343342959881,
0.012878789566457272,
-0.012264501303434372,
-0.03881186619400978,
-0.019086744636297226,
-0.00005981143476674333,
-0.10373389720916748,
-0.041903164237737656,
-0.002954539842903614,
0.010180728510022163,
0.04103594645857811,
-0.0052378177642822266,
-0.03508560359477997,
-0.06708156317472458,
-0.004297239705920219,
0.03956669941544533,
0.02677498385310173,
0.0013596357312053442,
-0.002036820398643613,
-0.027760900557041168,
-0.09122861921787262,
-0.09201151877641678,
0.017724143341183662,
0.06991128623485565,
-0.08830967545509338,
-0.06287360191345215,
0.05535658448934555,
0.06729170680046082,
-0.012434879317879677,
-0.04464368149638176,
-0.04922376945614815,
-0.03444785997271538,
0.06467939913272858,
0.004887090530246496,
0.012487443163990974,
-0.0017370515270158648,
-0.0985182598233223,
0.02344246581196785,
0.011371517553925514,
0.017503313720226288,
-0.0013850360410287976,
-0.008426403626799583,
0.0021485062316060066,
0.05366675183176994,
0.07917965948581696,
0.014040290378034115,
-0.014354786835610867,
-0.004756271839141846,
-0.11064241826534271,
-0.04852944239974022,
0.043774284422397614,
0.010897700674831867,
0.023539826273918152,
-0.062328703701496124,
0.07170208543539047,
-0.0005011402536183596,
-0.00003491355892037973,
-0.02829624153673649,
-0.03592223674058914,
-0.057970546185970306,
0.02388741821050644,
-0.06985212862491608,
0.039758168160915375,
-0.10409165173768997,
0.027868201956152916,
-0.02282973937690258,
-0.042292140424251556,
-0.0397077240049839,
-0.08576540648937225,
0.10574659705162048,
-0.009221749380230904,
-2.6167058935700914e-33,
-0.0052543384954333305,
0.0232409555464983,
0.0009685847326181829,
0.0856204628944397,
-0.048323169350624084,
-0.00867246463894844,
0.003096181433647871,
-0.007361250463873148,
-0.02646593190729618,
-0.04655134677886963,
-0.0757635161280632,
-0.05168452858924866,
0.032959043979644775,
-0.03647236526012421,
0.04995382949709892,
-0.020756958052515984,
-0.025324350222945213,
0.00439823605120182,
0.02975345030426979,
-0.015748975798487663,
0.013691643252968788,
0.07839364558458328,
-0.0640483945608139,
0.08144460618495941,
-0.019503796473145485,
0.018697066232562065,
0.07011159509420395,
0.08422042429447174,
-0.016278916969895363,
0.02063359133899212,
-0.04505159333348274,
-0.029043689370155334,
0.024777419865131378,
-0.00843797717243433,
-0.057236865162849426,
0.11889450252056122,
0.03046984039247036,
0.02923201024532318,
-0.04729965701699257,
0.1495828926563263,
0.02291693538427353,
-0.055494341999292374,
-0.0328981988132,
0.033909328281879425,
-0.07326473295688629,
-0.020694460719823837,
-0.057713910937309265,
0.009054625406861305,
0.046130768954753876,
0.051905274391174316,
-0.016268694773316383,
0.079936683177948,
-0.06591680645942688,
-0.04882552847266197,
-0.08237617462873459,
-0.10293973982334137,
0.012183775193989277,
0.014512992464005947,
0.10930873453617096,
-0.02116384357213974,
0.009342547506093979,
0.00950284767895937,
0.023817554116249084,
-0.03365008905529976,
-0.00749660050496459,
-0.01675083301961422,
-0.03363659232854843,
0.01868206262588501,
0.008881320245563984,
-0.008864734321832657,
0.008871810510754585,
-0.033181071281433105,
-0.022828232496976852,
-0.013639198616147041,
0.02156905271112919,
-0.007096422836184502,
0.05510764569044113,
-0.02480740286409855,
0.029890425503253937,
-0.09624247252941132,
0.013744316063821316,
-0.0016710085328668356,
0.07403812557458878,
0.07053632289171219,
-0.0034564873203635216,
0.13338857889175415,
0.014385567046701908,
0.08100979775190353,
-0.03423740342259407,
0.05118810757994652,
0.021435415372252464,
0.06497414410114288,
0.0027610117103904486,
0.03326372802257538,
-0.09663040935993195,
-4.1023685781738095e-8,
-0.15378570556640625,
0.04674885421991348,
0.1328124850988388,
0.08103343844413757,
-0.03942536935210228,
-0.017771724611520767,
-0.006892607547342777,
0.046035125851631165,
-0.03569250926375389,
-0.07605349272489548,
0.03395707160234451,
0.020562462508678436,
-0.001994698541238904,
-0.05699377879500389,
-0.028359487652778625,
0.02917824499309063,
-0.04155030846595764,
0.01761624962091446,
-0.05555713176727295,
0.017532870173454285,
0.004813818261027336,
0.01726705953478813,
0.011678609997034073,
-0.08343863487243652,
0.07516485452651978,
-0.060250524431467056,
-0.04998139292001724,
0.09982404857873917,
0.005207514390349388,
0.019978199154138565,
-0.028716029599308968,
-0.02641548402607441,
0.02943778969347477,
-0.049143318086862564,
0.0006682792445644736,
0.07773734629154205,
-0.006021060049533844,
-0.06750496476888657,
0.02465713769197464,
-0.09006097167730331,
-0.008669431321322918,
0.08000212907791138,
-0.03301788121461868,
-0.007553504314273596,
-0.02604680508375168,
-0.00629838090389967,
0.028703903779387474,
-0.11804676800966263,
0.02087455242872238,
-0.06329512596130371,
-0.04055861756205559,
-0.07017511129379272,
-0.01978473737835884,
0.1030106246471405,
0.07833147794008255,
0.053450148552656174,
-0.03747846931219101,
-0.0033397511579096317,
0.04319023713469505,
0.09723338484764099,
0.05141653120517731,
0.011157282628118992,
-0.050958894193172455,
-0.019651630893349648
] | 0 | 0 | 0 | 0 |
|
Online Preference Alignment for Language Models via Count-based Exploration | https://openreview.net/forum?id=cfKZ5VrhXt | [
"Chenjia Bai",
"Yang Zhang",
"Shuang Qiu",
"Qiaosheng Zhang",
"Kang Xu",
"Xuelong Li"
] | Spotlight | Reinforcement Learning from Human Feedback (RLHF) has shown great potential in fine-tuning Large Language Models (LLMs) to align with human preferences. Existing methods perform preference alignment from a fixed dataset, which can be limited in data coverage and the resulting reward model is hard to generalize in out-of-distribution responses. Thus, online RLHF is more desirable to empower the LLM to explore outside the support of the initial dataset by iteratively collecting the prompt-response pairs. In this paper, we study the fundamental problem in online RLHF, i.e., how to explore for LLM. We give a theoretical motivation in linear reward assumption to show that an optimistic reward with an upper confidence bound (UCB) term leads to a provably efficient RLHF policy. Then, we reformulate our objective to direct preference optimization with an exploration term, where the UCB-term can be converted to a count-based exploration bonus. We further propose a practical algorithm, named Count-based Online Preference Optimization (COPO), which leverages a simple coin-flip counting module to estimate the pseudo-count of a prompt-response pair in previously collected data. COPO encourages LLMs to balance exploration and preference optimization in an iterative manner, which enlarges the exploration space and the entire data coverage of iterative LLM policies. We conduct online RLHF experiments on Zephyr and Llama-3 models. The results on instruction-following and standard academic benchmarks show that COPO significantly increases performance. | Reinforcement Learning from Human Feedback, RLHF, Preference Alignment, Exploration, LLMs | We propose count-based online preference optimization for LLM alignment that leverages coin-flip counting to encourage exploration in online RLHF. | 13,482 | 2501.12735 | [
-0.0640619769692421,
-0.09233816713094711,
-0.006883716210722923,
0.020084965974092484,
-0.017764177173376083,
0.08543065935373306,
0.0388631671667099,
0.058960724622011185,
0.0699266791343689,
0.014036508277058601,
-0.02031645178794861,
-0.07485969364643097,
0.07122191041707993,
0.009893307462334633,
0.05408412590622902,
-0.012082226574420929,
0.11882808059453964,
0.010523511096835136,
-0.03292778506875038,
-0.027083907276391983,
-0.013196472078561783,
0.030560409650206566,
0.060073014348745346,
-0.04026694968342781,
-0.01686505414545536,
-0.021140702068805695,
0.01491633988916874,
-0.01797465980052948,
0.02139361947774887,
0.005619600415229797,
0.008560116402804852,
0.027794867753982544,
0.04272731766104698,
-0.05409112200140953,
-0.0553494393825531,
0.010624065063893795,
-0.18754567205905914,
-0.08866238594055176,
-0.017732949927449226,
-0.01746240258216858,
-0.07965190708637238,
0.017770325765013695,
0.03344704210758209,
0.0037769158370792866,
0.023282790556550026,
0.034737564623355865,
-0.004047274589538574,
0.021661508828401566,
0.028614632785320282,
0.0545133501291275,
-0.015440686605870724,
0.015232633799314499,
-0.0005506961606442928,
0.03355111554265022,
-0.014430548064410686,
0.043805114924907684,
0.00041304357000626624,
0.01896229013800621,
0.035404860973358154,
-0.03933819755911827,
-0.04591601714491844,
-0.16895975172519684,
-0.010322880931198597,
0.010734200477600098,
-0.049577414989471436,
-0.06827861070632935,
-0.0154038704931736,
0.06866741925477982,
-0.037040553987026215,
0.07929721474647522,
0.00394632900133729,
0.07428057491779327,
0.06054463982582092,
0.04791458323597908,
0.014657524414360523,
0.07097883522510529,
0.006103612016886473,
-0.06387752294540405,
0.03409312292933464,
0.027492348104715347,
-0.042590364813804626,
0.004589549265801907,
0.04220355674624443,
0.0038537366781383753,
0.062165722250938416,
-0.018470095470547676,
0.0289910975843668,
0.010021110065281391,
0.10113190114498138,
-0.01786939613521099,
-0.03506425768136978,
-0.019529420882463455,
0.0027406751178205013,
0.011024446226656437,
-0.03430395573377609,
0.08898332715034485,
0.0007663536234758794,
-0.05639296770095825,
-0.05480683594942093,
0.06521393358707428,
0.036103349179029465,
0.09246017783880234,
-0.0547620989382267,
-0.08479805290699005,
-0.07323391735553741,
-0.03135934844613075,
-0.05274035409092903,
0.08953358232975006,
0.0545821413397789,
-0.1264382302761078,
0.003469131188467145,
0.032175637781620026,
0.053604401648044586,
0.07119344919919968,
-0.03209326043725014,
-0.04563216492533684,
0.05348663777112961,
-0.04705929756164551,
0.05988301336765289,
-0.0003121830231975764,
0.040446944534778595,
-0.040595363825559616,
0.06274842470884323,
-0.0330100879073143,
0.021980399265885353,
-0.0014852453023195267,
0.025304429233074188,
1.9180473593688849e-34,
0.03750436753034592,
0.04950202256441116,
0.06522514671087265,
-0.0028192829340696335,
0.057382982224226,
-0.020703265443444252,
0.09236713498830795,
0.056725356727838516,
-0.0914536863565445,
-0.008921816945075989,
0.04860103502869606,
0.04201700538396835,
0.0001542928075650707,
0.07273411005735397,
-0.011396205984055996,
-0.03782786801457405,
-0.010914076119661331,
0.04434879496693611,
-0.05755048245191574,
-0.055653806775808334,
0.06901621073484421,
0.017825057730078697,
0.04900125786662102,
-0.05354775860905647,
0.047632619738578796,
-0.019255168735980988,
-0.008947661146521568,
-0.05497187003493309,
-0.05673229321837425,
-0.013203172013163567,
-0.0491151288151741,
-0.031829118728637695,
0.012712709605693817,
-0.008493702858686447,
-0.015292319469153881,
-0.02413027547299862,
-0.07678870111703873,
-0.023442376405000687,
-0.03638415411114693,
-0.015482230111956596,
0.027995554730296135,
0.03897717967629433,
0.0174283217638731,
0.05878503620624542,
-0.11397020518779755,
-0.0631098598241806,
0.0259241946041584,
-0.007814921438694,
-0.05903985723853111,
-0.04296286776661873,
0.04204850643873215,
0.017357483506202698,
-0.06052311137318611,
0.03425802290439606,
-0.03505988046526909,
-0.009506055153906345,
-0.04087506979703903,
0.06471424549818039,
-0.03798853978514671,
0.0022874625865370035,
0.0028544068336486816,
0.04465210810303688,
0.03050742857158184,
0.013634822331368923,
0.02926591783761978,
0.06706074625253677,
-0.03479773551225662,
-0.020536713302135468,
0.04054667428135872,
0.017068564891815186,
0.03797677531838417,
0.022426599636673927,
-0.02339908480644226,
-0.06906060129404068,
0.04304250702261925,
0.042006634175777435,
0.059309471398591995,
-0.05149879679083824,
0.07452847808599472,
-0.03734209015965462,
0.07150585949420929,
-0.038525596261024475,
0.015836214646697044,
-0.04247937723994255,
-0.010286597535014153,
-0.09644615650177002,
0.03634177893400192,
-0.05201276019215584,
-0.023264549672603607,
-0.06176469847559929,
0.0018136051949113607,
0.01139125507324934,
-0.028538929298520088,
0.017974304035305977,
0.05458059161901474,
-2.7226283579079935e-34,
-0.030840005725622177,
-0.09909718483686447,
0.09862128645181656,
0.06771302223205566,
-0.02293124794960022,
-0.018483364954590797,
-0.018557187169790268,
-0.039863817393779755,
-0.009034614078700542,
-0.0513400062918663,
-0.09364823997020721,
0.0545482262969017,
0.03645046800374985,
0.006539423484355211,
0.04065363109111786,
-0.033159930258989334,
-0.01618199236690998,
0.023261111229658127,
0.014353384263813496,
0.04417921602725983,
0.0016681362176313996,
0.0897999256849289,
-0.16473256051540375,
0.0657706931233406,
-0.007378057111054659,
0.03479418903589249,
0.010226287879049778,
0.02565794438123703,
0.012205176055431366,
0.05073920264840126,
-0.03711484745144844,
-0.050308793783187866,
-0.11374565958976746,
-0.02522904984652996,
-0.03704861178994179,
0.06611748039722443,
0.04031535983085632,
0.04143765568733215,
-0.0708855614066124,
0.1958165168762207,
0.025484902784228325,
-0.06982914358377457,
-0.024891583248972893,
-0.04026424139738083,
-0.04403055086731911,
-0.01545008085668087,
-0.0805264338850975,
0.012633861042559147,
-0.016571037471294403,
-0.0003908198850695044,
-0.0037212122697383165,
0.11517419666051865,
-0.047903601080179214,
0.06557769328355789,
-0.023874841630458832,
-0.09073946624994278,
-0.09632528573274612,
-0.06001262366771698,
0.07891248911619186,
-0.016815602779388428,
-0.04451064020395279,
0.0395590104162693,
0.026395495980978012,
-0.002968212589621544,
-0.01367754116654396,
-0.05111455172300339,
0.0038847378455102444,
0.012794200330972672,
-0.04053666815161705,
-0.028902528807520866,
0.004601435270160437,
-0.020767483860254288,
-0.00022199128579813987,
-0.00769569119438529,
0.028498278930783272,
0.02687801420688629,
0.018646778538823128,
0.00022763591550756246,
0.01462879590690136,
0.008517846465110779,
0.0012484207982197404,
0.06572192907333374,
0.04152832552790642,
0.01922232285141945,
0.08382192254066467,
0.03402058035135269,
0.05816102400422096,
0.12328051030635834,
0.00013065016537439078,
-0.01844923011958599,
0.004999864846467972,
-0.06436426192522049,
0.04725097119808197,
0.04978293552994728,
-0.03571756184101105,
-4.4927709552666784e-8,
-0.11084248870611191,
-0.04831084981560707,
0.06274241954088211,
0.07422587275505066,
0.04348248988389969,
-0.004664401989430189,
0.016798818483948708,
0.03386075049638748,
-0.04342819005250931,
-0.04670695215463638,
0.020130200311541557,
0.09323956072330475,
0.01828196458518505,
-0.042234260588884354,
-0.08706536889076233,
0.06765516847372055,
-0.012421863153576851,
0.0010095542529597878,
0.009384174831211567,
0.024019761011004448,
0.01824893243610859,
0.006200628820806742,
0.05445064976811409,
-0.06730681657791138,
0.05548090860247612,
-0.01616230607032776,
0.04119007661938667,
0.09350240230560303,
0.036655910313129425,
-0.009168293327093124,
-0.007391930557787418,
-0.08267658203840256,
-0.015625251457095146,
-0.020153479650616646,
0.0637425035238266,
0.04550258815288544,
0.00285359681583941,
-0.051692090928554535,
-0.022987980395555496,
-0.07178863137960434,
0.015728609636425972,
0.06157476454973221,
-0.014856056310236454,
0.020035570487380028,
0.014305992983281612,
0.007602890487760305,
-0.04430606588721275,
-0.12775155901908875,
0.02595381997525692,
-0.13717825710773468,
-0.0052228220738470554,
-0.06939242035150528,
-0.03601319342851639,
0.031790461391210556,
0.043357495218515396,
0.033365823328495026,
-0.0671701580286026,
-0.025354832410812378,
0.05408576503396034,
0.09539928287267685,
0.022899063304066658,
-0.032467689365148544,
-0.07385291159152985,
-0.005637510679662228
] | https://github.com/baichenjia/copo | 14 | 0 | 0 | 0 |
Internet of Agents: Weaving a Web of Heterogeneous Agents for Collaborative Intelligence | https://openreview.net/forum?id=o1Et3MogPw | [
"Weize Chen",
"Ziming You",
"Ran Li",
"yitong guan",
"Chen Qian",
"Chenyang Zhao",
"Cheng Yang",
"Ruobing Xie",
"Zhiyuan Liu",
"Maosong Sun"
] | Spotlight | The rapid advancement of large language models (LLMs) has paved the way for the development of highly capable autonomous agents. However, existing multi-agent frameworks often struggle with integrating diverse capable third-party agents due to reliance on agents defined within their own ecosystems. They also face challenges in simulating distributed environments, as most frameworks are limited to single-device setups. Furthermore, these frameworks often rely on hard-coded communication pipelines, limiting their adaptability to dynamic task requirements. Inspired by the concept of the Internet, we propose the Internet of Agents (IoA), a novel framework that addresses these limitations by providing a flexible and scalable platform for LLM-based multi-agent collaboration. IoA introduces an agent integration protocol, an instant-messaging-like architecture design, and dynamic mechanisms for agent teaming and conversation flow control. Through extensive experiments on general assistant tasks, embodied AI tasks, and retrieval-augmented generation benchmarks, we demonstrate that IoA consistently outperforms state-of-the-art baselines, showcasing its ability to facilitate effective collaboration among heterogeneous agents. IoA represents a step towards linking diverse agents in an Internet-like environment, where agents can seamlessly collaborate to achieve greater intelligence and capabilities. We will release our code to facilitate further research. | llm agent, multi-agent | We propose IoA, a novel framework inspired by the Internet for effective collaboration among diverse LLM agents. IoA enables autonomous conversation flow, integration of heterogeneous agents, etc. It outperforms SoTA baselines in various tasks. | 13,450 | 2407.07061 | [
-0.0277743898332119,
-0.10784541070461273,
-0.007971874438226223,
-0.0005608632927760482,
0.04373760148882866,
-0.03957424685359001,
-0.04498998075723648,
-0.05466241389513016,
0.021827848628163338,
-0.05187911540269852,
-0.037278614938259125,
-0.09361826628446579,
-0.0020135759841650724,
0.02552291564643383,
0.10358377546072006,
0.020337717607617378,
0.1433471441268921,
-0.01726771704852581,
-0.10095580667257309,
-0.03317279741168022,
-0.002234847517684102,
0.05128082260489464,
0.033832598477602005,
-0.012691481970250607,
-0.023406585678458214,
0.012363266199827194,
-0.019676487892866135,
-0.0792088732123375,
-0.007439238019287586,
0.00959524605423212,
0.0810416042804718,
0.0028878808952867985,
-0.03662601113319397,
0.06991243362426758,
-0.03775486350059509,
0.06355301290750504,
-0.031727712601423264,
0.020880531519651413,
-0.022335590794682503,
-0.04406994208693504,
-0.046969249844551086,
-0.06941135972738266,
-0.04061759263277054,
0.016154484823346138,
0.012064133770763874,
0.0006384075968526304,
-0.06739548593759537,
-0.010674994438886642,
-0.015874572098255157,
0.05505247786641121,
-0.060976967215538025,
-0.0928344652056694,
0.030981173738837242,
0.08689939975738525,
-0.036257367581129074,
0.02937612310051918,
0.009850451722741127,
-0.0002629603259265423,
0.027908755466341972,
0.0008702451596036553,
0.008819817565381527,
-0.04665306955575943,
0.09550944715738297,
0.03521725907921791,
-0.05556348338723183,
-0.007351120468229055,
0.033835794776678085,
0.049229398369789124,
-0.05077226087450981,
-0.07184675335884094,
-0.04636445268988609,
0.005782701540738344,
0.01876901648938656,
0.008672134019434452,
-0.0013850133400410414,
-0.009041300974786282,
-0.021418292075395584,
-0.01914231851696968,
0.05857093259692192,
0.004935435485094786,
-0.06960737705230713,
0.04557948186993599,
-0.018371084704995155,
0.04806811735033989,
0.0350518636405468,
0.031710535287857056,
-0.006415503565222025,
0.05504865199327469,
0.015368393622338772,
0.017516953870654106,
-0.05607091262936592,
-0.009885555133223534,
0.11939652264118195,
-0.011767600663006306,
0.09773237258195877,
0.04632330685853958,
-0.01943168230354786,
-0.09905610978603363,
-0.0446135513484478,
0.07721143960952759,
-0.040461406111717224,
0.038054030388593674,
-0.011887801811099052,
-0.012750071473419666,
-0.03824181482195854,
0.014462443999946117,
0.0447474867105484,
0.031729474663734436,
0.028455529361963272,
-0.025410879403352737,
-0.037562672048807144,
-0.027944358065724373,
0.07662589102983475,
0.0013502441579475999,
0.048877209424972534,
-0.008335024118423462,
0.02392939291894436,
0.07426528632640839,
0.0599249042570591,
0.039988573640584946,
0.00517355976626277,
-0.04825635626912117,
-0.0051732915453612804,
0.024114377796649933,
0.12738049030303955,
0.1137295514345169,
0.025266008451581,
-1.3245746769908971e-33,
0.13454122841358185,
0.05498099699616432,
0.0351908802986145,
0.09032659977674484,
0.07526795566082001,
-0.030203156173229218,
0.06623905897140503,
0.03386946767568588,
-0.1239025667309761,
-0.045022301375865936,
-0.0000039880255826574285,
0.05475299805402756,
0.00921692606061697,
0.05586490407586098,
0.011815874837338924,
-0.02662665955722332,
-0.0739232674241066,
0.07832761853933334,
0.051473066210746765,
-0.015039699152112007,
0.013907535932958126,
-0.026936843991279602,
0.04230956360697746,
0.03594709560275078,
0.022983800619840622,
0.0639435276389122,
0.07791441679000854,
-0.03759355843067169,
0.10714652389287949,
-0.013349120505154133,
-0.07618238031864166,
0.0036140086594969034,
0.02909798175096512,
0.0539422482252121,
-0.02859918773174286,
-0.04770280793309212,
-0.08105869591236115,
-0.07548192888498306,
-0.0013827410293743014,
0.05547558143734932,
0.012608813121914864,
0.013925157487392426,
0.0007832869305275381,
-0.09150377660989761,
-0.10839425027370453,
-0.10792706906795502,
-0.05909419059753418,
0.04722430184483528,
0.00349070830270648,
-0.074202761054039,
0.019484268501400948,
0.024266140535473824,
0.012865974567830563,
0.024979285895824432,
0.015961244702339172,
-0.03647206723690033,
-0.03831205144524574,
0.02887045592069626,
0.012418712489306927,
-0.018147185444831848,
0.005763006396591663,
0.04003484174609184,
-0.021076299250125885,
0.08858706057071686,
0.11556577682495117,
-0.07316891103982925,
0.0038712911773473024,
0.12065324932336807,
0.056089144200086594,
-0.03941044956445694,
-0.034526512026786804,
-0.006604560650885105,
0.04119078069925308,
0.0170363150537014,
-0.018315045163035393,
-0.0020023200195282698,
0.0227340217679739,
-0.08356364071369171,
0.039721060544252396,
0.008184756152331829,
-0.03390559181571007,
0.01245300192385912,
-0.056195735931396484,
-0.04181006923317909,
-0.032977502793073654,
-0.09180799126625061,
-0.053550027310848236,
-0.05177677050232887,
-0.04977498948574066,
0.06522756814956665,
-0.051228757947683334,
-0.01953859254717827,
-0.013011020608246326,
0.029150618240237236,
0.01252730842679739,
-4.9486278013378405e-34,
-0.03465614840388298,
-0.04719800874590874,
-0.09771815687417984,
0.051490288227796555,
0.06892905384302139,
-0.07594326138496399,
-0.010348405689001083,
-0.06134362891316414,
-0.04900968819856644,
0.009070811793208122,
-0.07316768169403076,
0.008226473815739155,
0.037678517401218414,
-0.055952705442905426,
0.039587412029504776,
-0.04650837928056717,
0.02771228738129139,
-0.06738974153995514,
0.0026144192088395357,
0.013587268069386482,
0.0016816930146887898,
0.04391874745488167,
-0.028531139716506004,
-0.007592152338474989,
-0.028285931795835495,
0.022792603820562363,
0.006102492567151785,
-0.016725171357393265,
-0.003210392314940691,
-0.011844201013445854,
0.018718939274549484,
-0.011015040799975395,
-0.031211193650960922,
-0.08879196643829346,
-0.013409758917987347,
0.07947541773319244,
0.000704832433257252,
-0.0016775255789980292,
-0.04895130544900894,
0.0212644562125206,
0.07473950833082199,
-0.11884210258722305,
-0.0706789642572403,
0.00187678390648216,
-0.014868194237351418,
0.07294059544801712,
-0.14239078760147095,
0.016832975670695305,
-0.04671918600797653,
-0.000050046986871166155,
-0.0003769306349568069,
0.035208698362112045,
-0.05999620258808136,
-0.07039030641317368,
-0.06458976119756699,
-0.03248158097267151,
0.03759869188070297,
-0.0014856215566396713,
-0.04134639725089073,
-0.0056260921992361546,
-0.02864113822579384,
-0.04294246435165405,
0.05332876369357109,
0.03692065551877022,
0.005167039111256599,
-0.02158057875931263,
-0.033601004630327225,
0.021756263449788094,
-0.02128884568810463,
0.015762055292725563,
0.05759222060441971,
-0.06446649134159088,
-0.029506515711545944,
0.008627159520983696,
0.07564329355955124,
-0.02025824971497059,
-0.017457468435168266,
-0.05723254382610321,
0.03191540390253067,
-0.04633643478155136,
-0.016438933089375496,
0.09866216033697128,
0.06633012741804123,
0.042613785713911057,
0.06182383373379707,
0.011300468817353249,
-0.02435191161930561,
0.14653833210468292,
-0.003006715327501297,
0.030687779188156128,
0.023433031514286995,
0.0036555237602442503,
-0.04590579494833946,
0.046151190996170044,
-0.06776788085699081,
-4.1266293493436024e-8,
-0.0598759762942791,
0.03614402934908867,
-0.02570682391524315,
0.014039169065654278,
-0.06842013448476791,
-0.003035247325897217,
-0.0057334015145897865,
0.010083723813295364,
0.09761371463537216,
0.01413465104997158,
0.019262002781033516,
-0.0582568533718586,
0.0211598239839077,
-0.006470148451626301,
0.0844898670911789,
0.029598478227853775,
-0.07469776272773743,
-0.028798051178455353,
0.013384511694312096,
-0.08632059395313263,
0.06627284735441208,
0.010505998507142067,
-0.03062460757791996,
-0.008567138575017452,
0.014718955382704735,
-0.010573416948318481,
0.0247630774974823,
0.05553664639592171,
-0.03143385425209999,
0.015843503177165985,
-0.02993328869342804,
-0.035942669957876205,
0.0015647243708372116,
-0.01931757852435112,
0.03198125958442688,
0.03309033066034317,
-0.04482423886656761,
-0.05386030301451683,
0.029553676024079323,
-0.07308880239725113,
0.020592913031578064,
0.08472129702568054,
-0.055353470146656036,
-0.05241934210062027,
0.051540400832891464,
0.01244939211755991,
-0.09280180186033249,
-0.10591617971658707,
0.015123781748116016,
-0.031210122630000114,
-0.08424058556556702,
-0.04674742370843887,
0.037098877131938934,
0.11480358242988586,
0.06771702319383621,
-0.03001708909869194,
0.03577958419919014,
-0.06739005446434021,
0.15220892429351807,
0.05080663785338402,
0.045857202261686325,
0.0370933823287487,
-0.04032869637012482,
-0.0219358392059803
] | https://github.com/openbmb/ioa | 708 | 0 | 0 | 0 |
Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning | https://openreview.net/forum?id=A6Y7AqlzLW | [
"Amrith Setlur",
"Chirag Nagpal",
"Adam Fisch",
"Xinyang Geng",
"Jacob Eisenstein",
"Rishabh Agarwal",
"Alekh Agarwal",
"Jonathan Berant",
"Aviral Kumar"
] | Spotlight | A promising approach for improving reasoning in large language models is to use process reward models (PRMs). PRMs provide feedback at each step of a multi-step reasoning trace, improving credit assignment over outcome reward models (ORMs) that only provide feedback at the final step. However, collecting dense, per-step human labels is not scalable, and training PRMs from automatically-labeled data has thus far led to limited gains. With the goal of using PRMs to improve a *base* policy via test-time search and reinforcement learning (RL), we ask: ``How should we design process rewards?'' Our key insight is that, to be effective, the process reward for a step should measure
*progress*: a change in the likelihood of producing a correct response in the future, before and after taking the step, as measured under a *prover* policy distinct from the base policy. Such progress values can {distinguish} good and bad steps generated by the base policy, even though the base policy itself cannot. Theoretically, we show that even weaker provers can improve the base policy, as long as they distinguish steps without being too misaligned with the base policy. Our results show that process rewards defined as progress under such provers improve the efficiency of exploration during test-time search and online RL. We empirically validate our claims by training **process advantage verifiers (PAVs)** to measure progress under such provers and show that compared to ORM, they are >8% more accurate, and 1.5-5x more compute-efficient. Equipped with these insights, our PAVs enable **one of the first results** showing a 6x gain in sample efficiency for a policy trained using online RL with PRMs vs. ORMs. | LLM, Math Reasoning, Process Supervision, Reward Models, RL, Search | null | 13,389 | 2410.08146 | [
-0.12491114437580109,
-0.009988374076783657,
-0.02050800435245037,
-0.03350024297833443,
0.04042520746588707,
0.043297477066516876,
0.017289070412516594,
0.07202646136283875,
0.01853165775537491,
0.04867430031299591,
-0.04700401425361633,
-0.05875076726078987,
0.052071694284677505,
0.11056089401245117,
0.021478282287716866,
0.007526901550590992,
0.04420539736747742,
0.005903720855712891,
-0.01477536279708147,
-0.07762563228607178,
-0.03653407469391823,
-0.03140786662697792,
0.10088659077882767,
0.05281798914074898,
-0.05584275349974632,
0.002662826096639037,
-0.012967025861144066,
-0.02014932408928871,
0.06914485991001129,
-0.03707237169146538,
0.0776493027806282,
-0.013996086083352566,
0.07805493474006653,
-0.011759639717638493,
-0.06194550171494484,
0.05788161978125572,
-0.05440625920891762,
-0.03207746893167496,
-0.03503815084695816,
-0.04997580125927925,
-0.007705824449658394,
-0.08055313676595688,
-0.010200297459959984,
0.016013843938708305,
0.05727684870362282,
-0.003458996070548892,
0.08454298973083496,
-0.014719958417117596,
-0.10290665179491043,
-0.02279515378177166,
-0.04534880444407463,
-0.04942581057548523,
0.01435826811939478,
-0.01694456674158573,
-0.086111880838871,
0.02159045822918415,
0.13776463270187378,
-0.026274196803569794,
-0.03644135221838951,
-0.061741627752780914,
-0.06991035491228104,
-0.13019151985645294,
-0.07691751420497894,
-0.009038127027451992,
-0.03592841327190399,
0.02304486557841301,
0.011443892493844032,
0.046492744237184525,
-0.019842838868498802,
0.0478975884616375,
-0.008512688800692558,
0.0377662368118763,
-0.025670794770121574,
-0.009868256747722626,
-0.018535571172833443,
0.024506879970431328,
-0.052855052053928375,
-0.007891426794230938,
0.014714580029249191,
-0.03769451007246971,
-0.07205470651388168,
-0.01872285269200802,
-0.0006122341146692634,
0.008324543945491314,
0.03677082434296608,
-0.04438629001379013,
0.021303726360201836,
-0.014821126125752926,
0.08452065289020538,
0.009034514427185059,
0.04562375694513321,
0.0469735823571682,
-0.035286422818899155,
-0.022330591455101967,
0.015019354410469532,
0.05876817926764488,
-0.08810126036405563,
-0.10062213242053986,
-0.010908522643148899,
0.042139675468206406,
0.05110044777393341,
0.17664578557014465,
0.03561504930257797,
-0.08749199658632278,
0.03265673294663429,
-0.006860361434519291,
-0.007947545498609543,
0.048579148948192596,
0.08744210004806519,
-0.13574858009815216,
0.010313755832612514,
0.011064855381846428,
0.07395409792661667,
0.08670909702777863,
0.02509363181889057,
-0.004593776073306799,
-0.020144877955317497,
-0.05444415286183357,
0.013188144192099571,
0.037087224423885345,
0.0008321916102431715,
-0.019686613231897354,
-0.028011050075292587,
0.0214910339564085,
0.0400373674929142,
-0.0003946842916775495,
-0.003703133901581168,
4.097767796836725e-33,
0.07676710933446884,
0.04859888553619385,
0.04589412361383438,
0.01365609373897314,
0.0497860312461853,
0.03315874934196472,
0.057975441217422485,
0.021851735189557076,
-0.005152628757059574,
-0.049545712769031525,
-0.003686182666569948,
0.018767690286040306,
-0.03930364176630974,
0.06310955435037613,
0.0182057935744524,
-0.03281453624367714,
-0.0327768549323082,
0.04982798174023628,
-0.02403370663523674,
-0.00761415297165513,
0.06402809917926788,
-0.08551912754774094,
0.00542425224557519,
-0.06588799506425858,
0.08436629176139832,
0.0008077852544374764,
0.01977871172130108,
0.004346833098679781,
-0.036113765090703964,
-0.03960490599274635,
0.021457383409142494,
-0.022382458671927452,
0.016017666086554527,
0.062078870832920074,
-0.004154783673584461,
-0.015071731992065907,
-0.029830530285835266,
0.007105085067451,
0.08252719044685364,
-0.06944786757230759,
0.003063978860154748,
-0.019553152844309807,
0.022646479308605194,
0.062397558242082596,
-0.05005015432834625,
-0.024517584592103958,
-0.058328207582235336,
-0.03480727970600128,
-0.09410251677036285,
0.01013830304145813,
0.1210957020521164,
0.035018742084503174,
-0.022611483931541443,
-0.11674370616674423,
-0.0249743964523077,
0.02845788560807705,
-0.028012318536639214,
0.05990977585315704,
-0.05653410404920578,
-0.0066728428937494755,
0.03730153664946556,
0.0008746915264055133,
-0.019006261602044106,
0.08839672058820724,
-0.0590280145406723,
0.07221141457557678,
-0.06431131809949875,
0.00815894640982151,
0.07139701396226883,
0.03032003901898861,
0.020978737622499466,
0.047399286180734634,
0.03917858749628067,
-0.06154841557145119,
0.10701900720596313,
-0.06997082382440567,
0.017397750169038773,
-0.1015491858124733,
0.025601407513022423,
0.001343553769402206,
-0.028842099010944366,
0.006546672899276018,
-0.014968370087444782,
-0.00026994006475433707,
0.08165813982486725,
-0.050885532051324844,
-0.012454676441848278,
-0.08073686808347702,
-0.0008743212674744427,
-0.005877203308045864,
-0.03199084848165512,
-0.041831254959106445,
-0.004200637806206942,
0.08334788680076599,
-0.0010685488814488053,
-3.9069611864692664e-33,
0.009735818020999432,
-0.04752451926469803,
-0.0018750784220173955,
0.06972388178110123,
-0.011423951014876366,
-0.014075825922191143,
-0.031185591593384743,
-0.007608251180499792,
0.040497153997421265,
-0.011228866875171661,
-0.06703433394432068,
0.007408919744193554,
0.013912145979702473,
0.0439189076423645,
-0.028319409117102623,
-0.07416623085737228,
0.0466342493891716,
0.040439117699861526,
0.08466033637523651,
0.019008561968803406,
0.045635923743247986,
0.1538945883512497,
-0.13092947006225586,
0.08479847013950348,
-0.007455410435795784,
0.0016137317288666964,
-0.0019064318621531129,
0.02584541030228138,
0.03918775916099548,
0.06589978188276291,
-0.04194681718945503,
-0.09673424065113068,
-0.029805384576320648,
-0.04259324073791504,
-0.030927935615181923,
-0.004445226397365332,
0.07380922138690948,
0.011307697743177414,
-0.04351525008678436,
0.11309532076120377,
0.036349985748529434,
-0.05293990671634674,
-0.09621761739253998,
-0.0031501888297498226,
-0.002605772577226162,
-0.02922111190855503,
-0.02174503728747368,
-0.050866469740867615,
-0.09385289251804352,
-0.013039187528192997,
-0.009131924249231815,
0.028709273785352707,
-0.028887247666716576,
0.013820174150168896,
-0.03509986028075218,
-0.04379066452383995,
-0.026091018691658974,
-0.082664854824543,
-0.02985868602991104,
-0.010828669182956219,
-0.10145672410726547,
0.06415250897407532,
0.023952318355441093,
0.025783343240618706,
0.03143475204706192,
0.003798715304583311,
0.026471542194485664,
0.022274868562817574,
0.003570176661014557,
0.01860397681593895,
0.09343264997005463,
0.026388196274638176,
-0.03354192525148392,
0.03768323361873627,
0.09856964647769928,
0.027559999376535416,
-0.04547644406557083,
-0.06267381459474564,
-0.04676804691553116,
-0.02756948210299015,
-0.06595081835985184,
-0.004534151870757341,
0.04357120767235756,
0.044460248202085495,
-0.042810067534446716,
0.012096169404685497,
0.07304537296295166,
0.08040551841259003,
-0.0010388721711933613,
0.00465577095746994,
-0.0054308264516294,
-0.014490032568573952,
0.006522873416543007,
0.04684198647737503,
-0.0007821053732186556,
-5.173492567678295e-8,
-0.0524287223815918,
0.050254520028829575,
0.014107265509665012,
0.07428283989429474,
0.07194780558347702,
-0.013182939030230045,
-0.04932881519198418,
0.06524980813264847,
-0.0397176556289196,
-0.037577927112579346,
0.062018923461437225,
0.006016966421157122,
-0.029909860342741013,
-0.04700331762433052,
0.03056122548878193,
0.06798527389764786,
0.02613246440887451,
-0.0017628002678975463,
-0.048058148473501205,
-0.025838224217295647,
0.051193058490753174,
0.03178122267127037,
-0.01294289156794548,
-0.039869558066129684,
-0.012043152004480362,
-0.0756206288933754,
0.0471644401550293,
0.10173339396715164,
0.06561727821826935,
-0.048777684569358826,
0.025928104296326637,
-0.037063781172037125,
-0.0016545206308364868,
0.06011733412742615,
0.1423300802707672,
0.009725846350193024,
0.02558200992643833,
-0.007171193603426218,
0.0105902049690485,
0.01410675048828125,
-0.032827310264110565,
0.05709993094205856,
-0.011586561799049377,
-0.0068917712196707726,
-0.009684748947620392,
0.0034251180477440357,
-0.15969811379909515,
-0.07950520515441895,
-0.020588159561157227,
-0.08288589119911194,
0.04093312472105026,
-0.019400376826524734,
-0.036523472517728806,
0.054724596440792084,
0.05767235904932022,
0.010277277790009975,
0.04144403338432312,
-0.04328945279121399,
-0.053640834987163544,
0.04316957667469978,
0.08136888593435287,
0.02172616496682167,
-0.055731967091560364,
-0.0032349287066608667
] | 0 | 0 | 0 | 0 |
|
MQuAKE-Remastered: Multi-Hop Knowledge Editing Can Only Be Advanced with Reliable Evaluations | https://openreview.net/forum?id=m9wG6ai2Xk | [
"Shaochen Zhong",
"Yifan Lu",
"Lize Shao",
"Bhargav Bhushanam",
"Xiaocong Du",
"Yixin Wan",
"Yucheng Shi",
"Daochen Zha",
"Yiwei Wang",
"Ninghao Liu",
"Kaixiong Zhou",
"Shuai Xu",
"Kai-Wei Chang",
"Louis Feng",
"Vipin Chaudhary",
"Xia Hu"
] | Spotlight | Large language models (LLMs) can give out erroneous answers to factually rooted questions either as a result of undesired training outcomes or simply because the world has moved on after a certain knowledge cutoff date. Under such scenarios, *knowledge editing* often comes to the rescue by delivering efficient patches for such erroneous answers without significantly altering the rest, where many editing methods have seen reasonable success when the editing targets are simple and direct (e.g., *``what club does Lionel Messi currently play for?''*).
However, knowledge fragments like this are often deeply intertwined in the real world, making effectively propagating the editing effect to non-directly related questions a practical challenge (to entertain an extreme example: [*"What car did the wife of the owner of the club that Messi currently plays for used to get to school in the 80s?"*](youtube.com/watch?v=DbwiHC1Fu-E\&t=132s)). Prior arts have coined this task as *multi-hop knowledge editing* with the most popular dataset being MQuAKE, serving as the sole evaluation benchmark for many later proposed editing methods due to the expensive nature of constructing knowledge editing datasets at scale.
In this work, we reveal that **up to 33\% or 76\% of \mquake{}'s questions and ground truth labels are, in fact, corrupted in various fashions due to some unintentional clerical or procedural oversights**. Our work provides a detailed audit of MQuAKE's error pattern and a comprehensive fix without sacrificing its dataset capacity. Additionally, we benchmarked almost all proposed MQuAKE-evaluated editing methods on our post-fix dataset, **MQuAKE-Remastered**. We observe that many methods try to overfit the original MQuAKE by exploiting some dataset idiosyncrasies of MQuAKE. We provide a guideline on how to approach such datasets faithfully and show that a simple, minimally invasive approach — **GWalk** — can offer beyond SOTA editing performance without such exploitation. The MQuAKE-Remastered datasets and utilities are available at [huggingface.co/datasets/henryzhongsc/MQuAKE-Remastered](https://huggingface.co/datasets/henryzhongsc/MQuAKE-Remastered) and [github.com/henryzhongsc/MQuAKE-Remastered](https://github.com/henryzhongsc/MQuAKE-Remastered), respectively. | knowledge edit, model edit, multi-hop, question answering, natural language processing, dataset audit | Updating one knowledge fact will produce a ripple effect, making multi-hop knowledge editing (MHKE) a desired capability for reliable LLMs. We reveal many unknown errors of MQuAKE — the most popular MHKE dataset — though an audit and fix everything. | 13,349 | null | [
0.020513102412223816,
0.036475975066423416,
0.030204176902770996,
0.018693357706069946,
-0.003472816664725542,
0.05216093361377716,
-0.06292528659105301,
0.015668945387005806,
-0.014613715931773186,
0.053508460521698,
-0.04746336117386818,
-0.0012795131187886,
0.0745711475610733,
0.020897151902318,
-0.019291236996650696,
0.06162480264902115,
0.014701379463076591,
0.05964302271604538,
-0.07278343290090561,
-0.06410699337720871,
-0.03120696358382702,
-0.011992184445261955,
0.0353664867579937,
0.028118841350078583,
-0.003570509608834982,
-0.06182963028550148,
-0.028732992708683014,
0.014407495968043804,
-0.001038066460750997,
-0.04843658208847046,
0.015734612941741943,
0.14774546027183533,
0.001649949001148343,
0.08452238887548447,
-0.06897514313459396,
0.05739741027355194,
0.0039820801466703415,
0.1405244916677475,
-0.004741525277495384,
-0.08203930407762527,
-0.032668475061655045,
-0.04480576515197754,
-0.011825640685856342,
-0.03280450403690338,
0.0663650780916214,
0.0539003349840641,
-0.011106881313025951,
-0.011738544330000877,
-0.03598518669605255,
0.05660633370280266,
-0.09929698705673218,
-0.1274629831314087,
-0.021891333162784576,
0.00883448962122202,
0.026703525334596634,
0.019815858453512192,
-0.03508233278989792,
0.09212269634008408,
0.019285066053271294,
-0.03960549831390381,
-0.01147541869431734,
-0.0723828449845314,
-0.03135836869478226,
0.012857845984399319,
0.006095751654356718,
-0.06472054868936539,
-0.03578709065914154,
0.08045646548271179,
-0.037215907126665115,
0.013520310632884502,
0.05548525229096413,
-0.009769544005393982,
-0.05577720329165459,
-0.0015962349716573954,
-0.025022607296705246,
0.05737820640206337,
-0.04086406156420708,
0.027557531371712685,
-0.03461240604519844,
-0.07150930166244507,
0.09323474019765854,
-0.02610364370048046,
-0.0005475422949530184,
-0.0221402607858181,
0.07170764356851578,
0.006642573047429323,
-0.021436018869280815,
0.0027414162177592516,
0.06056346371769905,
-0.0001041264840750955,
-0.1041417196393013,
-0.07339063286781311,
0.0839574858546257,
-0.03470320627093315,
0.0900317057967186,
0.026999864727258682,
0.028712382540106773,
-0.03518020361661911,
-0.0321320965886116,
0.08931857347488403,
0.055177751928567886,
0.13990645110607147,
-0.021258976310491562,
-0.0921044796705246,
0.03611946105957031,
0.005113878287374973,
0.07702247053384781,
0.0693632960319519,
0.07178844511508942,
-0.12036173790693283,
-0.05583234876394272,
0.02776084654033184,
0.03653952106833458,
-0.072756327688694,
-0.0067219361662864685,
0.023093167692422867,
0.05806028097867966,
0.08572680503129959,
-0.0686367079615593,
0.03122428059577942,
-0.03847270831465721,
0.0340622216463089,
0.05488131567835808,
0.04228919744491577,
-0.0012497863499447703,
-0.027224764227867126,
-0.05857924371957779,
7.483367484015552e-33,
0.11706133931875229,
0.07934022694826126,
0.04792318120598793,
0.0471891313791275,
0.035529974848032,
0.02047421596944332,
-0.04275912046432495,
-0.033601079136133194,
0.007603620644658804,
-0.0035919693764299154,
0.04114710912108421,
-0.0025688966270536184,
-0.05943894386291504,
0.014329084195196629,
0.05507519096136093,
0.038748692721128464,
-0.10345660895109177,
0.00407444080337882,
-0.020480353385210037,
-0.0016767688794061542,
0.05033750459551811,
-0.027911048382520676,
-0.0402677096426487,
-0.08778972923755646,
-0.03769363835453987,
0.07234249264001846,
0.08366096764802933,
0.00797183346003294,
0.09782931953668594,
-0.013923917897045612,
-0.08668483793735504,
0.010459048673510551,
-0.05650655925273895,
0.04202452301979065,
-0.018142765387892723,
-0.04872192069888115,
-0.0361689031124115,
-0.11837410926818848,
0.033225711435079575,
-0.0013191484613344073,
-0.06127333268523216,
-0.03896212577819824,
-0.016066431999206543,
-0.048155661672353745,
-0.0583827942609787,
0.001047560479491949,
-0.0010484665399417281,
-0.03687618672847748,
-0.012541668489575386,
0.002584384521469474,
0.06441882997751236,
0.045666515827178955,
0.003788059577345848,
-0.0659511610865593,
0.0010847779922187328,
0.0040291547775268555,
0.008964638225734234,
-0.042517516762018204,
0.025558751076459885,
0.056517839431762695,
0.018466288223862648,
0.02609560638666153,
-0.050828538835048676,
0.10161685198545456,
0.028718777000904083,
0.10869446396827698,
-0.013314959593117237,
0.052590399980545044,
0.09085459262132645,
-0.025083232671022415,
-0.0624191053211689,
-0.03496583178639412,
-0.0812888965010643,
-0.07006799429655075,
0.015365800820291042,
-0.04007631167769432,
-0.07274851948022842,
-0.001174007193185389,
0.0702824518084526,
-0.052979838103055954,
0.0170454028993845,
-0.04112793505191803,
-0.07190898805856705,
-0.05419956520199776,
0.03532743081450462,
-0.011775564402341843,
0.06348585337400436,
-0.08692843466997147,
0.011869722045958042,
0.02211363986134529,
-0.02994590252637863,
-0.03659738600254059,
-0.04959529638290405,
0.06330320984125137,
0.025985851883888245,
-6.221115037946856e-33,
-0.027127714827656746,
-0.039811406284570694,
-0.04439248889684677,
0.10618803650140762,
0.029901742935180664,
-0.05160561949014664,
-0.022344835102558136,
0.00773300975561142,
0.00914001651108265,
-0.1286087930202484,
-0.04911825433373451,
-0.09155314415693283,
0.004096987657248974,
-0.01804564893245697,
-0.04422096908092499,
-0.00569195207208395,
0.018997497856616974,
-0.02224956452846527,
-0.022799549624323845,
0.034412600100040436,
-0.008829913102090359,
-0.03241905942559242,
-0.07338067889213562,
0.07450651377439499,
-0.027753179892897606,
0.08208002895116806,
0.0008618349675089121,
0.031057611107826233,
-0.03082018718123436,
0.07283348590135574,
-0.02963707223534584,
-0.08787336945533752,
-0.009833859279751778,
-0.04654250293970108,
-0.04978765547275543,
0.03495560958981514,
0.06279610097408295,
-0.03386315703392029,
-0.04754401370882988,
0.02300986461341381,
0.0528741255402565,
-0.0010383735643699765,
-0.07132703065872192,
0.001772832009010017,
-0.01959831453859806,
0.06034226343035698,
-0.1506420224905014,
0.0019561978988349438,
0.03100505657494068,
0.00512118823826313,
0.0021176692098379135,
-0.07081598043441772,
-0.07869258522987366,
-0.03075955994427204,
0.042436499148607254,
-0.026170535013079643,
-0.0050725797191262245,
-0.01683078519999981,
-0.027822114527225494,
-0.016020679846405983,
-0.044833675026893616,
0.0016065748641267419,
-0.03177254647016525,
-0.011346674524247646,
0.07251381129026413,
0.009529462084174156,
-0.039908889681100845,
0.057252660393714905,
-0.0774877592921257,
-0.03019186109304428,
-0.019177837297320366,
-0.01639314740896225,
0.010984708555042744,
-0.08417615294456482,
0.07929437607526779,
-0.023496858775615692,
0.06859789043664932,
-0.0348966084420681,
0.005565689876675606,
-0.10621348768472672,
-0.03427913039922714,
-0.02814985066652298,
0.06572160869836807,
0.047906726598739624,
0.009023027494549751,
0.09732834994792938,
-0.0125743318349123,
0.09078919142484665,
0.08609364181756973,
0.030745919793844223,
-0.010033111087977886,
0.033663056790828705,
-0.02524455450475216,
0.03891769424080849,
-0.055926207453012466,
-6.077950587268788e-8,
-0.03571753203868866,
0.03056713566184044,
-0.07263793051242828,
0.05061740800738335,
0.058800384402275085,
-0.009729636833071709,
-0.05331335589289665,
0.09307347983121872,
0.023112040013074875,
0.018303966149687767,
0.004353641998022795,
0.04049601033329964,
-0.009402751922607422,
-0.010412100702524185,
-0.010699541307985783,
0.05120798572897911,
-0.00886434968560934,
0.059884048998355865,
-0.050961658358573914,
0.006522543728351593,
0.0858597531914711,
0.010676941834390163,
0.023809684440493584,
0.048025865107774734,
0.04758976027369499,
-0.06277332454919815,
-0.07679478079080582,
0.051709502935409546,
-0.012201961129903793,
0.0387926883995533,
-0.032250482589006424,
-0.051718708127737045,
0.01989947445690632,
0.06102477014064789,
0.025586167350411415,
-0.033267706632614136,
0.017269760370254517,
0.032665375620126724,
0.019933423027396202,
-0.011966673657298088,
0.00893832091242075,
0.02903653308749199,
-0.028329934924840927,
0.045088622719049454,
-0.009897454641759396,
-0.07399175316095352,
-0.02843494340777397,
-0.01377320010215044,
-0.03995563089847565,
-0.005191626958549023,
-0.0542367659509182,
0.03903314098715782,
0.060584891587495804,
-0.0018501348095014691,
0.10067716985940933,
0.032924678176641464,
-0.02962348237633705,
0.06278114020824432,
0.016291720792651176,
-0.04803002253174782,
0.09716738760471344,
-0.07065488398075104,
0.029952209442853928,
0.02267548069357872
] | 0 | 0 | 0 | 0 |
|
Competition Dynamics Shape Algorithmic Phases of In-Context Learning | https://openreview.net/forum?id=XgH1wfHSX8 | [
"Core Francisco Park",
"Ekdeep Singh Lubana",
"Hidenori Tanaka"
] | Spotlight | In-Context Learning (ICL) has significantly expanded the general-purpose nature of large language models, allowing them to adapt to novel tasks using merely the inputted context. This has motivated a series of papers that analyze tractable synthetic domains and postulate precise mechanisms that may underlie ICL. However, the use of relatively distinct setups that often lack a sequence modeling nature to them makes it unclear how general the reported insights from such studies are. Motivated by this, we propose a synthetic sequence modeling task that involves learning to simulate a finite mixture of Markov chains. As we show, models trained on this task reproduce most well-known results on ICL, hence offering a unified setting for studying the concept. Building on this setup, we demonstrate we can explain a model’s behavior by decomposing it into four broad algorithms that combine a fuzzy retrieval vs. inference approach with either unigram or bigram statistics of the context. These algorithms engage in a competitive dynamics to dominate model behavior, with the precise experimental conditions dictating which algorithm ends up superseding others: e.g., we find merely varying context size or amount of training yields (at times sharp) transitions between which algorithm dictates the model behavior, revealing a mechanism that explains the transient nature of ICL. In this sense, we argue ICL is best thought of as a mixture of different algorithms, each with its own peculiarities, instead of a monolithic capability. This also implies that making general claims about ICL that hold universally across all settings may be infeasible. | In-Context Learning, Circuit Competition, Markov Chains, Training Dynamics, Generalization | In-context learning consists of phases of multiple algorithmic solutions, many phenomena are explained by this decomposition. | 13,272 | 2412.01003 | [
-0.06783401966094971,
-0.10330625623464584,
0.03839113935828209,
-0.0012864888412877917,
0.07852370291948318,
0.012148678302764893,
0.026250455528497696,
-0.0450117401778698,
0.09581033140420914,
-0.032126106321811676,
-0.0692504346370697,
-0.06544885039329529,
0.05074656382203102,
0.05872988700866699,
0.006287181284278631,
-0.04446299374103546,
0.08366646617650986,
-0.019225267693400383,
-0.06698264926671982,
-0.03265489637851715,
0.0740990862250328,
0.046409036964178085,
0.06508848816156387,
-0.02345859259366989,
0.017169931903481483,
0.03392104059457779,
0.013817137107253075,
-0.02719935029745102,
0.03804462403059006,
0.00009115356078837067,
0.037187766283750534,
0.06389732658863068,
0.024178070947527885,
-0.02853536605834961,
0.016911080107092857,
0.06939815729856491,
-0.10252109915018082,
0.018095839768648148,
0.029773006215691566,
-0.017003007233142853,
-0.01116835419088602,
0.0033972691744565964,
0.013092955574393272,
0.06628481298685074,
0.10121416300535202,
-0.0010390224633738399,
-0.007569545414298773,
-0.03603655472397804,
-0.002622137311846018,
-0.0055942595936357975,
-0.15174512565135956,
-0.0023012820165604353,
-0.003051258623600006,
0.050966452807188034,
-0.01783985085785389,
0.041445355862379074,
0.05928120017051697,
0.035870376974344254,
-0.001214070594869554,
-0.06926356256008148,
-0.049063391983509064,
-0.13908201456069946,
-0.031228553503751755,
0.0006360130500979722,
-0.07233467698097229,
0.018503859639167786,
0.004004315007477999,
0.10960375517606735,
-0.03079136833548546,
0.0037199039943516254,
0.011155132204294205,
0.10127057880163193,
-0.011562438681721687,
0.061430372297763824,
0.07619926333427429,
0.08108925819396973,
-0.00026545129367150366,
-0.03280283510684967,
0.021578913554549217,
-0.07048618048429489,
0.008526040241122246,
0.0013897990575060248,
0.08725611865520477,
-0.045327600091695786,
0.06763827800750732,
0.031943824142217636,
0.045561742037534714,
-0.0031456290744245052,
0.06085561215877533,
-0.009356380440294743,
-0.04857034608721733,
-0.04539426788687706,
0.044822148978710175,
0.010125860571861267,
0.06736161559820175,
0.06148839369416237,
0.047928337007761,
-0.004292400553822517,
0.09024614095687866,
0.131851464509964,
0.06999579817056656,
-0.010943434201180935,
-0.013874230906367302,
-0.056436676532030106,
0.004566724877804518,
0.026048583909869194,
-0.028006434440612793,
0.008035249076783657,
0.03715050593018532,
-0.08624868839979172,
0.03907948359847069,
0.02417224831879139,
0.01940501667559147,
0.05482613667845726,
-0.029807014390826225,
-0.0469193160533905,
0.052106525748968124,
-0.07956820726394653,
-0.006685665808618069,
0.07435861974954605,
-0.02287832647562027,
0.005785734858363867,
-0.06908518821001053,
0.026460355147719383,
-0.05363842099905014,
-0.0641268864274025,
-0.0951138511300087,
1.3734912213614358e-33,
0.0487634539604187,
-0.0602998286485672,
0.016662459820508957,
0.05213804170489311,
-0.009387414902448654,
0.014224841259419918,
-0.044338300824165344,
-0.051941972225904465,
0.0141108063980937,
-0.0013579867081716657,
-0.023943975567817688,
0.0502343475818634,
-0.06114352494478226,
0.010454787872731686,
0.06073435768485069,
-0.02368554100394249,
-0.06538688391447067,
0.06338585913181305,
-0.007857904769480228,
-0.03142615035176277,
0.006200739182531834,
0.01312743965536356,
-0.021439330652356148,
-0.025793686509132385,
-0.016448717564344406,
0.08002876490354538,
0.06700563430786133,
-0.06328251957893372,
-0.05455351248383522,
0.0018016843823716044,
-0.06677626818418503,
-0.03501640632748604,
-0.07962246984243393,
0.10642185062170029,
0.03280835226178169,
-0.011835861951112747,
0.024560170248150826,
-0.0547906756401062,
-0.023273196071386337,
-0.012604805640876293,
-0.032683953642845154,
-0.03194084390997887,
0.0455436073243618,
0.05928365886211395,
-0.11298821866512299,
-0.034412261098623276,
-0.06204766035079956,
0.02839232049882412,
-0.09757532179355621,
-0.013559299521148205,
0.09632541239261627,
0.023515930399298668,
0.06055570766329765,
-0.08640232682228088,
0.021579265594482422,
0.0186076108366251,
-0.062453534454107285,
-0.028892485424876213,
0.028874659910798073,
0.04666213318705559,
0.031438253819942474,
-0.04351126775145531,
-0.01033898163586855,
0.08488929271697998,
-0.012560641393065453,
0.03397061303257942,
-0.08787748217582703,
0.027431629598140717,
0.041830386966466904,
-0.008969884365797043,
0.026274146512150764,
0.05483930930495262,
-0.059902042150497437,
-0.02502441592514515,
0.09102494269609451,
-0.030522966757416725,
0.03342979773879051,
-0.0810413733124733,
-0.060912758111953735,
0.06938698142766953,
0.02390121854841709,
-0.0062065282836556435,
-0.032847531139850616,
-0.08486898988485336,
-0.09407990425825119,
-0.02296311967074871,
0.020217979326844215,
-0.17419637739658356,
-0.02712172083556652,
-0.04673369973897934,
-0.09311000257730484,
-0.07875877618789673,
-0.047626711428165436,
0.06869441270828247,
0.04289363697171211,
-3.660669038691088e-33,
-0.026135586202144623,
-0.02147078514099121,
0.012441328726708889,
0.0020834023598581553,
-0.03231726959347725,
-0.025683818385004997,
0.037220656871795654,
0.024785077199339867,
-0.04567786306142807,
-0.09036226570606232,
-0.027768131345510483,
-0.014394877478480339,
0.08916809409856796,
-0.01531304232776165,
0.010587225668132305,
-0.02066541090607643,
0.03588985279202461,
-0.008980881422758102,
0.01569635234773159,
0.04215602949261665,
0.06380520015954971,
-0.006401377264410257,
-0.1251644790172577,
-0.012206999585032463,
-0.06412917375564575,
0.05161208659410477,
-0.09672600030899048,
0.05728430673480034,
-0.014964248985052109,
0.08043354749679565,
-0.060976020991802216,
-0.04229442775249481,
-0.055212266743183136,
0.024369580671191216,
-0.1432679295539856,
0.06556230038404465,
0.054233912378549576,
-0.049860693514347076,
-0.05293426662683487,
0.02717096358537674,
0.0035129101015627384,
-0.04270452260971069,
-0.0038470153231173754,
0.03248174116015434,
-0.04122990369796753,
-0.05776885524392128,
-0.05760044977068901,
-0.02254670113325119,
0.05239493027329445,
-0.05058646202087402,
0.003421130357310176,
0.0641123354434967,
-0.049119558185338974,
-0.00025608838768675923,
-0.0545119009912014,
-0.023677537217736244,
-0.07368884980678558,
0.002845871727913618,
0.02498530223965645,
-0.024916958063840866,
-0.06253096461296082,
-0.05190671607851982,
-0.011834035627543926,
0.011099088005721569,
0.018414204940199852,
-0.035071562975645065,
-0.04783015698194504,
-0.07095855474472046,
0.02906518243253231,
-0.05140301585197449,
0.052267588675022125,
0.03587532788515091,
-0.03590799868106842,
0.018579132854938507,
0.03139331191778183,
-0.025967497378587723,
0.017867187038064003,
-0.03383948653936386,
-0.014302961528301239,
-0.023965584114193916,
0.021801795810461044,
-0.01815260760486126,
0.056102924048900604,
0.0727851539850235,
0.014360466040670872,
0.07113903015851974,
0.05020656809210777,
0.029588274657726288,
0.046354930847883224,
0.055017877370119095,
0.025631897151470184,
0.02417011559009552,
0.006171749904751778,
0.08190469443798065,
-0.0313577838242054,
-5.198425867547485e-8,
-0.06858179718255997,
0.06347130239009857,
0.01115583535283804,
0.011006571352481842,
0.03603516146540642,
0.07699976861476898,
-0.021182920783758163,
0.028565386310219765,
-0.017021674662828445,
-0.009737853892147541,
0.03674780949950218,
0.0008835974149405956,
-0.015062033198773861,
-0.023606592789292336,
-0.05216218903660774,
0.1231808066368103,
0.026826610788702965,
0.005538014695048332,
0.01714598387479782,
-0.01452705729752779,
-0.009714251384139061,
0.020671065896749496,
0.02033061906695366,
0.063785620033741,
0.022186294198036194,
-0.07476448267698288,
-0.09248639643192291,
0.07151208817958832,
0.08394057303667068,
0.0478084459900856,
-0.06221628189086914,
0.002142166020348668,
-0.03321387246251106,
-0.07139574736356735,
-0.04704403504729271,
0.06521591544151306,
0.0016076373867690563,
-0.06715308874845505,
0.00586424395442009,
-0.0878085121512413,
-0.004238816909492016,
0.09430671483278275,
-0.015164284966886044,
0.026153750717639923,
0.077417753636837,
-0.00014660332817584276,
0.010997720062732697,
-0.07451771199703217,
0.07532556354999542,
0.015738707035779953,
0.03944879770278931,
-0.03723747283220291,
-0.03321906179189682,
0.013149833306670189,
0.029912441968917847,
0.02976744994521141,
-0.028870059177279472,
0.007006744854152203,
0.01251714676618576,
0.08855398744344711,
-0.03169640153646469,
0.00950415525585413,
-0.03012797236442566,
-0.0005595990805886686
] | https://github.com/cfpark00/markov-mixtures | 4 | 0 | 0 | 0 |
In vivo cell-type and brain region classification via multimodal contrastive learning | https://openreview.net/forum?id=10JOlFIPjt | [
"Han Yu",
"Hanrui Lyu",
"YiXun Xu",
"Charlie Windolf",
"Eric Kenji Lee",
"Fan Yang",
"Andrew M Shelton",
"Olivier Winter",
"International Brain Laboratory",
"Eva L Dyer",
"Chandramouli Chandrasekaran",
"Nicholas A. Steinmetz",
"Liam Paninski",
"Cole Lincoln Hurwitz"
] | Spotlight | Current electrophysiological approaches can track the activity of many neurons, yet it is usually unknown which cell-types or brain areas are being recorded without further molecular or histological analysis. Developing accurate and scalable algorithms for identifying the cell-type and brain region of recorded neurons is thus crucial for improving our understanding of neural computation. In this work, we develop a multimodal contrastive learning approach for neural data that can be fine-tuned for different downstream tasks, including inference of cell-type and brain location. We utilize multimodal contrastive learning to jointly embed the activity autocorrelations and extracellular waveforms of individual neurons. We demonstrate that our embedding approach, Neuronal Embeddings via MultimOdal Contrastive Learning (NEMO), paired with supervised fine-tuning, achieves state-of-the-art cell-type classification for two opto-tagged datasets and brain region classification for the public International Brain Laboratory Brain-wide Map dataset. Our method represents a promising step towards accurate cell-type and brain region classification from electrophysiological recordings. | contrastive learning, electrophysiology, extracellular, multimodal, neuroscience, cell type, brain region, Neuropixels, deep learning | null | 13,258 | null | [
0.01999383606016636,
-0.12731842696666718,
0.03603392466902733,
-0.018322201445698738,
0.007013739086687565,
-0.03384207561612129,
0.0017971224151551723,
-0.030734552070498466,
-0.04821966588497162,
-0.010376328602433205,
-0.040952809154987335,
-0.11164697259664536,
-0.07659032940864563,
0.08844519406557083,
-0.06059623509645462,
0.002821688074618578,
0.02761070616543293,
0.11101339757442474,
-0.052002862095832825,
-0.012527617625892162,
0.0501352958381176,
-0.011651496402919292,
0.02847522310912609,
-0.03988935798406601,
0.09293843060731888,
0.03704531490802765,
0.0200963094830513,
0.028430232778191566,
-0.03023645654320717,
-0.035390522330999374,
0.036775555461645126,
0.00048466355656273663,
-0.046174295246601105,
0.0958528220653534,
0.014746762812137604,
0.05189046263694763,
-0.06903602182865143,
-0.005270887166261673,
0.015528863295912743,
-0.008200726471841335,
0.06426917016506195,
0.0328218974173069,
-0.018540002405643463,
0.06732802093029022,
0.04760313779115677,
0.058095358312129974,
0.02757302112877369,
-0.11185703426599503,
0.05938338860869408,
-0.0321563221514225,
0.017933737486600876,
-0.05841394513845444,
0.009363480843603611,
0.10875801742076874,
-0.03822481632232666,
0.04644618183374405,
-0.04793325811624527,
-0.005992083810269833,
0.009669226594269276,
0.018932051956653595,
0.0127261308953166,
-0.01870543137192726,
0.05378789082169533,
-0.08692557364702225,
-0.00014715097495354712,
0.006262895651161671,
0.022120250388979912,
0.028740420937538147,
-0.12223498523235321,
0.027580363675951958,
0.011405786499381065,
0.028515199199318886,
-0.02706790156662464,
-0.01811016909778118,
-0.018960796296596527,
-0.014225715771317482,
0.04449288547039032,
-0.028980785980820656,
0.092178575694561,
-0.1363343447446823,
0.04567793384194374,
0.0026652971282601357,
-0.039973825216293335,
-0.0075431205332279205,
0.1739213764667511,
0.055250395089387894,
-0.021163426339626312,
0.051311343908309937,
-0.010369809344410896,
0.036259837448596954,
-0.03449683636426926,
-0.0201687291264534,
0.03365271911025047,
-0.03497380018234253,
0.0010259191039949656,
-0.04241757467389107,
0.08096092939376831,
0.04039999842643738,
0.04049772396683693,
0.0779515877366066,
0.01571466773748398,
0.025190751999616623,
0.00872189924120903,
0.01105667743831873,
0.029977871105074883,
-0.05946885049343109,
0.06012309342622757,
0.03514307737350464,
0.03914138302206993,
-0.042198605835437775,
0.015138719230890274,
0.04585369676351547,
-0.030727392062544823,
0.028214208781719208,
0.04372572526335716,
-0.0008259105961769819,
0.009038039483129978,
0.051610950380563736,
0.119633249938488,
0.03689984604716301,
-0.03737073019146919,
-0.006361340172588825,
-0.04812297224998474,
0.011635092087090015,
0.06582071632146835,
-0.039748724550008774,
-0.11923467367887497,
3.1316078677317703e-34,
0.010045642033219337,
-0.05801762640476227,
0.007581816054880619,
0.048038240522146225,
-0.026409611105918884,
0.004801255185157061,
-0.050562288612127304,
0.016290906816720963,
-0.050628263503313065,
0.0595134012401104,
0.005916720721870661,
0.09458893537521362,
-0.028685428202152252,
0.15174120664596558,
-0.0043395026586949825,
-0.003345804288983345,
-0.1127815768122673,
-0.07376841455698013,
0.026171907782554626,
-0.04579510912299156,
-0.003755162702873349,
-0.04063071310520172,
-0.043457068502902985,
0.021081531420350075,
-0.02910578064620495,
-0.0641309842467308,
0.012816309928894043,
-0.028541600331664085,
0.01263225544244051,
0.02029488980770111,
-0.15326593816280365,
0.006500838790088892,
-0.0026612067595124245,
-0.014279046095907688,
0.054453711956739426,
-0.056391552090644836,
0.006879796739667654,
0.02277102880179882,
0.027745144441723824,
-0.017076367512345314,
0.010904747061431408,
0.006164273247122765,
-0.03980953246355057,
-0.08104684948921204,
-0.060028307139873505,
0.006074934732168913,
0.001589422463439405,
0.05286325514316559,
0.013916700147092342,
-0.07723506540060043,
0.005276271607726812,
-0.0356566458940506,
0.08211691677570343,
-0.10468405485153198,
-0.0025592446327209473,
0.024800525978207588,
-0.015472757630050182,
0.0634160116314888,
0.08657921105623245,
0.02008252590894699,
-0.08377328515052795,
0.024746572598814964,
0.035444702953100204,
0.06435548514127731,
-0.01592560112476349,
0.09708626568317413,
-0.10318060219287872,
0.04999152943491936,
0.05355076864361763,
-0.024638663977384567,
0.01868126168847084,
0.039796896278858185,
0.04350639134645462,
-0.09512155503034592,
0.02610667608678341,
0.058785635977983475,
0.02228669449687004,
0.03488409146666527,
-0.04621627926826477,
-0.021104954183101654,
0.019685985520482063,
-0.01538886222988367,
-0.07133767008781433,
0.002772618317976594,
-0.0408560112118721,
-0.02479744330048561,
0.013163506984710693,
-0.07553606480360031,
-0.05359921231865883,
0.0033030505292117596,
-0.08232872188091278,
-0.04783010482788086,
0.02043985202908516,
-0.05817512422800064,
-0.015293681062757969,
-2.2588200263065226e-33,
-0.0063215200789272785,
0.035882093012332916,
-0.04447971656918526,
-0.008533335290849209,
0.037842411547899246,
0.03409508615732193,
0.053159765899181366,
0.03904164955019951,
0.015275614336133003,
0.04804141819477081,
0.0018584411591291428,
0.023539045825600624,
0.020279739052057266,
0.00006464360194513574,
0.01945876143872738,
0.028436122462153435,
-0.08198775351047516,
0.11004529893398285,
0.015421736985445023,
0.004589463584125042,
-0.07591141760349274,
0.07063304632902145,
0.025417698547244072,
0.10128619521856308,
-0.05907708406448364,
0.07689305394887924,
-0.09219444543123245,
-0.005147708114236593,
-0.09398522973060608,
-0.09503588825464249,
-0.088025763630867,
-0.023137841373682022,
-0.025801358744502068,
-0.05802905187010765,
-0.0314350388944149,
0.017929119989275932,
0.023938124999403954,
-0.06764169782400131,
-0.04103054478764534,
0.007681558839976788,
0.07309513539075851,
-0.011737791821360588,
-0.030047889798879623,
-0.02565104514360428,
0.006646239198744297,
0.027427522465586662,
-0.025207746773958206,
0.01987733133137226,
-0.03967178612947464,
0.05443274602293968,
-0.02567480504512787,
-0.014425016939640045,
-0.1281442940235138,
-0.059964414685964584,
-0.014181935228407383,
-0.029756363481283188,
0.05046892166137695,
-0.037341900169849396,
0.032813020050525665,
0.01812339760363102,
-0.09048508107662201,
-0.09997657686471939,
0.013507455587387085,
0.02526690438389778,
-0.023058781400322914,
-0.03429979085922241,
0.021899376064538956,
-0.054329052567481995,
0.06777562201023102,
0.029104195535182953,
-0.010243234224617481,
0.06691781431436539,
-0.007749099750071764,
-0.035058170557022095,
0.007064299192279577,
0.0025029222015291452,
0.03510699048638344,
-0.026502244174480438,
0.03312348946928978,
-0.028094826266169548,
0.10205867886543274,
-0.13286881148815155,
-0.04995577409863472,
0.055448878556489944,
0.05716490000486374,
0.055797476321458817,
0.05235956236720085,
0.028638748452067375,
0.06281564384698868,
-0.04975935071706772,
0.020845137536525726,
0.005819507408887148,
-0.0652746930718422,
-0.004822514485567808,
-0.02987099252641201,
-3.779879875764891e-8,
-0.0038909113500267267,
0.011679012328386307,
0.04731963202357292,
-0.055865757167339325,
0.013375398702919483,
-0.11511358618736267,
0.016359753906726837,
0.09387820959091187,
0.022703764960169792,
0.029102949425578117,
0.10751847922801971,
-0.06952900439500809,
0.0029017545748502016,
-0.060817789286375046,
0.024851199239492416,
0.0382659025490284,
-0.023800624534487724,
-0.004040462896227837,
0.0004533515893854201,
-0.0014698769664391875,
0.0025486985687166452,
0.025725804269313812,
0.0845290869474411,
0.012630187906324863,
0.10162155330181122,
-0.08019629120826721,
-0.049554966390132904,
0.019515076652169228,
0.01469845324754715,
-0.0808662474155426,
0.04171108454465866,
0.06269890815019608,
0.014203709550201893,
-0.05425289273262024,
-0.05295506492257118,
0.0042917151004076,
0.03775930404663086,
0.011805970221757889,
-0.07955197244882584,
-0.015951007604599,
-0.0660112202167511,
0.011092109605669975,
-0.07264923304319382,
0.04077526926994324,
0.01734967529773712,
0.0013323307503014803,
0.050875771790742874,
-0.021789146587252617,
0.013891558162868023,
-0.008185779675841331,
-0.010834354907274246,
0.0284915491938591,
0.06376548856496811,
0.030182307586073875,
0.025078168138861656,
-0.008697881363332272,
-0.0656948909163475,
-0.019870445132255554,
0.00018684800306800753,
0.0008799983188509941,
0.04022115841507912,
0.054315242916345596,
-0.09833430498838425,
-0.005894440691918135
] | 0 | 0 | 0 | 0 |
|
Physics of Language Models: Part 3.3, Knowledge Capacity Scaling Laws | https://openreview.net/forum?id=FxNNiUgtfa | [
"Zeyuan Allen-Zhu",
"Yuanzhi Li"
] | Spotlight | Scaling laws describe the relationship between the size of language models and their capabilities. Unlike prior studies that evaluate a model's capability via loss or benchmarks, we estimate information-theoretically the number of knowledge \emph{bits} a model stores. We focus on factual knowledge represented as tuples, such as (USA, capital, Washington D.C.) from a Wikipedia page. Through multiple controlled datasets, we establish that language models can and only can store \emph{2 bits of knowledge per parameter, even when quantized to int8}, and such knowledge can be flexibly extracted for downstream applications.
More broadly, we present 12 results on how (1) training duration, (2) model architecture, (3) quantization, (4) sparsity constraints such as MoE, and (5) data signal-to-noise ratio affect a model's knowledge storage capacity. | scaling laws, knowledge capacity, language models | null | 13,207 | 2404.05405 | [
0.013546437956392765,
-0.0739705041050911,
-0.008058713749051094,
0.05617754906415939,
0.029345432296395302,
0.028443291783332825,
-0.022525958716869354,
0.003215441945940256,
0.03498510643839836,
-0.049942318350076675,
-0.02202429063618183,
0.010354331694543362,
0.043999698013067245,
0.022035837173461914,
0.02675107680261135,
0.010512253269553185,
0.06020035967230797,
0.043627217411994934,
-0.19928696751594543,
-0.06331373006105423,
0.07295946776866913,
-0.0322813056409359,
-0.007888210006058216,
0.021427912637591362,
0.06643705815076828,
-0.04329884424805641,
-0.05955246835947037,
-0.03758278116583824,
0.0422821044921875,
-0.04606470838189125,
0.04199449345469475,
-0.01611228659749031,
0.09277860820293427,
0.040082961320877075,
0.037421513348817825,
0.04311925545334816,
0.034832075238227844,
-0.020276330411434174,
0.03085847944021225,
0.010832743719220161,
-0.0006262263050302863,
-0.010065129958093166,
0.02722720243036747,
0.050527412444353104,
0.07808515429496765,
0.06112394481897354,
0.03225470334291458,
0.05081089213490486,
-0.07232017815113068,
0.006266779266297817,
-0.05067681148648262,
-0.0152052603662014,
0.0061676693148911,
0.047799933701753616,
-0.024450697004795074,
-0.040666110813617706,
-0.01624768227338791,
0.012747744098305702,
-0.10790431499481201,
0.03594992682337761,
-0.09231752902269363,
-0.1225496158003807,
-0.057251524180173874,
-0.020492151379585266,
-0.05666838586330414,
-0.028608091175556183,
-0.02486586757004261,
-0.011319784447550774,
-0.03808455541729927,
0.04044632613658905,
-0.017322782427072525,
0.010830597952008247,
-0.07644736021757126,
0.03319329023361206,
0.009506582282483578,
-0.008131218142807484,
0.09195883572101593,
0.030372485518455505,
0.050644584000110626,
-0.07103601098060608,
0.00016861430776771158,
0.01064439956098795,
0.029226360842585564,
-0.029018310829997063,
0.04955517128109932,
-0.06160570681095123,
0.05445167049765587,
0.03640531376004219,
0.03635680675506592,
0.00997229665517807,
-0.045680418610572815,
-0.025075186043977737,
0.03827124461531639,
0.029752839356660843,
0.08530144393444061,
0.04031024128198624,
0.003950241953134537,
-0.026916075497865677,
-0.01398063637316227,
0.04971633478999138,
0.03359917551279068,
0.10729954391717911,
0.035482753068208694,
-0.016560876742005348,
-0.0014603513991460204,
-0.06310198456048965,
-0.008605696260929108,
0.048549991101026535,
0.011651236563920975,
-0.03565891832113266,
0.05212186649441719,
0.08249181509017944,
0.028173018246889114,
0.0055434429086744785,
0.07070895284414291,
-0.015719037503004074,
-0.06237870454788208,
-0.00882240291684866,
-0.00409015454351902,
0.013659161515533924,
-0.05454469099640846,
0.0037361131981015205,
0.03845363110303879,
-0.009412186220288277,
-0.00009169832628685981,
0.003996605519205332,
-0.06566411256790161,
3.612856402902117e-34,
0.02839343436062336,
0.07838113605976105,
-0.04260866343975067,
0.03039107285439968,
0.026920156553387642,
0.020474329590797424,
0.010819204151630402,
-0.0023003153037279844,
-0.04895220324397087,
-0.016657745465636253,
-0.0467388741672039,
0.06260867416858673,
-0.04343288764357567,
0.001293353270739317,
0.10234608501195908,
0.09215088188648224,
-0.012991195544600487,
0.09996846318244934,
-0.005865596700459719,
-0.10144463926553726,
0.12340123951435089,
-0.059007782489061356,
-0.06292932480573654,
0.012160621583461761,
0.005342200398445129,
-0.03592078387737274,
0.020993975922465324,
0.004076319746673107,
-0.10321266204118729,
-0.00399712473154068,
-0.0989934429526329,
-0.005936430301517248,
0.04978900030255318,
-0.005339578725397587,
0.06121058762073517,
-0.05423693731427193,
0.02895423024892807,
-0.04729095846414566,
0.03282708302140236,
-0.05005209520459175,
0.011210815981030464,
0.04591301456093788,
0.05006486177444458,
-0.031737588346004486,
-0.09797371923923492,
-0.06659483164548874,
0.03249676153063774,
-0.05272909253835678,
-0.10075556486845016,
-0.07610107213258743,
0.07857708632946014,
-0.005594492424279451,
-0.03454824164509773,
-0.006578275468200445,
0.11875269562005997,
0.017735954374074936,
-0.00612733606249094,
0.013649002648890018,
0.02539650723338127,
0.09631343930959702,
-0.06512012332677841,
0.0049887108616530895,
0.10546961426734924,
0.058914393186569214,
0.08423451334238052,
0.04055376723408699,
-0.08666715770959854,
0.020829729735851288,
0.05521947145462036,
-0.016125313937664032,
-0.027695780619978905,
-0.06090604141354561,
-0.03431561589241028,
-0.022135887295007706,
-0.002196050249040127,
-0.07197471708059311,
0.05010254681110382,
-0.13919967412948608,
0.010556158609688282,
0.083967424929142,
-0.03205648437142372,
-0.0763922706246376,
0.046725526452064514,
-0.019873185083270073,
-0.028440775349736214,
-0.07270131260156631,
0.03158263489603996,
-0.08170411735773087,
-0.04002774879336357,
-0.07624644041061401,
-0.07699400931596756,
-0.04546055942773819,
-0.04173414781689644,
-0.05886963754892349,
-0.03211265429854393,
-1.3511227577267377e-33,
-0.042187564074993134,
-0.011149273253977299,
-0.07614956796169281,
0.08804038166999817,
-0.06084267422556877,
-0.08037728816270828,
-0.011422567069530487,
0.12403451651334763,
-0.057513441890478134,
-0.12090075761079788,
-0.056840457022190094,
-0.08282364159822464,
0.0835290253162384,
-0.009494805708527565,
-0.04943319037556648,
0.017360901460051537,
-0.0007361340103670955,
-0.030868947505950928,
0.07155704498291016,
0.08084998279809952,
-0.03178948163986206,
-0.08916773647069931,
-0.046168431639671326,
0.08116643875837326,
0.041123632341623306,
-0.010299541056156158,
-0.09769802540540695,
-0.009624117985367775,
-0.009242995642125607,
0.03835703805088997,
-0.03772129490971565,
-0.10148395597934723,
-0.021692123264074326,
-0.023484056815505028,
-0.05604834854602814,
-0.05426882579922676,
0.06054491177201271,
0.008763521909713745,
0.03381212428212166,
0.09508110582828522,
0.07305828481912613,
-0.02908020280301571,
-0.02720693126320839,
-0.014627953059971333,
0.021225305274128914,
-0.001986138988286257,
-0.05079542472958565,
-0.020383872091770172,
0.016730625182390213,
-0.05251963064074516,
0.07183913141489029,
0.030275588855147362,
0.06511476635932922,
0.0527908056974411,
-0.020341219380497932,
-0.051900722086429596,
-0.006751715671271086,
-0.00370719819329679,
0.038071099668741226,
0.021243689581751823,
0.0024303183890879154,
-0.04292374849319458,
-0.036969512701034546,
-0.014718822203576565,
-0.02185736782848835,
-0.01399831660091877,
-0.03618915006518364,
0.013741103932261467,
-0.0283723883330822,
-0.03406130522489548,
0.023140009492635727,
-0.0371989868581295,
-0.01731995679438114,
0.04424348101019859,
-0.07070182263851166,
-0.010167242959141731,
0.011555715464055538,
-0.03748380020260811,
-0.013246056623756886,
0.01856493391096592,
-0.048555515706539154,
0.08868125826120377,
0.057640671730041504,
-0.018300481140613556,
0.058361511677503586,
0.07936879247426987,
0.07353057712316513,
0.04272366315126419,
0.024416711181402206,
0.007595180068165064,
-0.013721272349357605,
0.03277035057544708,
-0.05257478728890419,
0.07225767523050308,
-0.08406173437833786,
-4.436802214513591e-8,
-0.043027110397815704,
0.05605759099125862,
0.031787410378456116,
-0.022299576550722122,
0.06055601313710213,
0.005234965588897467,
-0.0420706570148468,
0.06310877203941345,
0.03893132507801056,
0.032389525324106216,
0.06677732616662979,
0.07027681171894073,
-0.09394209086894989,
0.016259878873825073,
0.024621790274977684,
0.10943622887134552,
0.0396069772541523,
-0.014949670061469078,
0.03171117231249809,
-0.015225283801555634,
0.09676428884267807,
0.014150181785225868,
-0.00718197925016284,
0.01515947561711073,
0.04320146515965462,
-0.03239147737622261,
-0.002921365899965167,
0.1345929354429245,
0.025553688406944275,
0.0167210903018713,
0.010592257604002953,
-0.03987406566739082,
0.05078108608722687,
0.0004399029421620071,
0.07579177618026733,
0.02183278277516365,
-0.0360269658267498,
-0.049452997744083405,
0.003474203171208501,
-0.03366288170218468,
0.04108597710728645,
0.0028543362859636545,
-0.052190616726875305,
0.03155235946178436,
0.08687926083803177,
0.019299061968922615,
-0.06672343611717224,
-0.028728483244776726,
-0.0012644194066524506,
0.043979983776807785,
0.059259314090013504,
0.04269019141793251,
-0.0014312495477497578,
0.04132132604718208,
0.04938759282231331,
0.044473569840192795,
-0.008462068624794483,
-0.019085951149463654,
-0.009080839343369007,
-0.0440431572496891,
0.01846592128276825,
-0.009264183230698109,
-0.03507918864488602,
0.022672876715660095
] | 0 | 1 | 0 | 0 |
|
Scalable Decision-Making in Stochastic Environments through Learned Temporal Abstraction | https://openreview.net/forum?id=pQsllTesiE | [
"Baiting Luo",
"Ava Pettet",
"Aron Laszka",
"Abhishek Dubey",
"Ayan Mukhopadhyay"
] | Spotlight | Sequential decision-making in high-dimensional continuous action spaces, particularly in stochastic environments, faces significant computational challenges. We explore this challenge in the traditional offline RL setting, where an agent must learn how to make decisions based on data collected through a stochastic behavior policy. We present \textit{Latent Macro Action Planner} (L-MAP), which addresses this challenge by learning a set of temporally extended macro-actions through a state-conditional Vector Quantized Variational Autoencoder (VQ-VAE), effectively reducing action dimensionality. L-MAP employs a (separate) learned prior model that acts as a latent transition model and allows efficient sampling of plausible actions. During planning, our approach accounts for stochasticity in both the environment and the behavior policy by using Monte Carlo tree search (MCTS). In offline RL settings, including stochastic continuous control tasks, L-MAP efficiently searches over discrete latent actions to yield high expected returns.
Empirical results demonstrate that L-MAP maintains low decision latency despite increased action dimensionality. Notably, across tasks ranging from continuous control with inherently stochastic dynamics to high-dimensional robotic hand manipulation, L-MAP significantly outperforms existing model-based methods and performs on par with strong model-free actor-critic baselines, highlighting the effectiveness of the proposed approach in planning in complex and stochastic environments with high-dimensional action spaces. | Sequential Decision-Making, Monte Carlo Tree Search, Temporal Abstraction, Planning, Model-based Reinforcement Learning, Offline Reinforcement Learning | A scalable approach for sequential decision-making in high-dimensional continuous action spaces by learning macro actions and using MCTS. | 13,144 | 2502.21186 | [
-0.023428797721862793,
-0.1065014973282814,
0.04091968759894371,
0.0669555515050888,
0.05776842683553696,
0.033226627856492996,
-0.048679351806640625,
-0.031280841678380966,
0.0074038696475327015,
0.019389955326914787,
-0.008464573882520199,
-0.03132881596684456,
-0.05155710503458977,
0.04262128472328186,
-0.010840815491974354,
0.010835861787199974,
0.0640840083360672,
0.05988454073667526,
-0.06921981275081635,
-0.018048807978630066,
0.01681126467883587,
0.06509072333574295,
0.07601795345544815,
-0.01721947081387043,
0.035253845155239105,
0.009341464377939701,
0.07095044106245041,
-0.0735883116722107,
-0.02904462441802025,
-0.02227885276079178,
0.10258179157972336,
-0.0015129860257729888,
0.02279037982225418,
-0.07600821554660797,
-0.05333281308412552,
0.09782157093286514,
-0.1429152488708496,
-0.008533097803592682,
-0.07431318610906601,
-0.07489634305238724,
-0.015589050948619843,
0.01539415679872036,
-0.0070032463409006596,
0.03483477979898453,
0.022624047473073006,
-0.02670706994831562,
-0.004963320214301348,
-0.07950972020626068,
0.05011076107621193,
-0.009655652567744255,
-0.0009564162464812398,
-0.008156705647706985,
0.018768364563584328,
-0.047033946961164474,
-0.014583717100322247,
0.020237306132912636,
0.03690629079937935,
-0.004374136682599783,
0.06108876317739487,
-0.08280973136425018,
0.012712668627500534,
-0.10620544105768204,
0.0032148021273314953,
-0.09959656000137329,
-0.08537847548723221,
-0.029766488820314407,
-0.0070607163943350315,
0.0388687439262867,
0.053679294884204865,
-0.038937367498874664,
-0.015228023752570152,
0.05050346255302429,
-0.08079046756029129,
0.004660648759454489,
0.0318940132856369,
-0.008932740427553654,
0.012017144821584225,
-0.07972131669521332,
0.07063660025596619,
-0.006843076087534428,
-0.011811494827270508,
-0.012210839428007603,
-0.02868829295039177,
0.06996449083089828,
0.09742618352174759,
-0.000815459992736578,
0.014562220312654972,
0.1256055384874344,
0.09460189938545227,
-0.01773729920387268,
-0.001784131396561861,
0.017827944830060005,
-0.004605497699230909,
-0.03878936544060707,
-0.04336590692400932,
-0.03396647050976753,
0.012498797848820686,
-0.07057713717222214,
0.03283582255244255,
0.036869607865810394,
0.02312292903661728,
0.06658943742513657,
0.025156857445836067,
0.04166476055979729,
0.05349276587367058,
-0.08094922453165054,
0.03524136170744896,
0.023537879809737206,
0.020639896392822266,
-0.11446387320756912,
-0.021096935495734215,
-0.03868027403950691,
0.043605636805295944,
0.006666886154562235,
0.015381721779704094,
0.031437575817108154,
0.033244501799345016,
-0.02182447537779808,
-0.03970230743288994,
0.048787862062454224,
0.010852688923478127,
-0.03104846365749836,
-0.02215464599430561,
-0.00836354773491621,
0.014953572303056717,
-0.05022395774722099,
-0.04040658101439476,
1.633543908048431e-33,
0.05211795121431351,
-0.04179462417960167,
0.07428676635026932,
0.055887505412101746,
0.07027638703584671,
-0.030385028570890427,
0.0397118404507637,
-0.0023308179806917906,
0.013078171759843826,
-0.009030865505337715,
0.02891625091433525,
-0.013504128903150558,
-0.08501733094453812,
0.09302292764186859,
0.08293192088603973,
0.000034605454857228324,
-0.091031514108181,
0.0010005473159253597,
-0.016369521617889404,
-0.037533171474933624,
0.04322439432144165,
-0.03778538852930069,
-0.04085801914334297,
-0.016394460573792458,
0.022127971053123474,
-0.031155703589320183,
0.06686850637197495,
0.025580590590834618,
0.01592095196247101,
-0.005208395421504974,
-0.08419948071241379,
-0.021946413442492485,
-0.012590458616614342,
0.055701564997434616,
0.0350135937333107,
0.05710156261920929,
-0.09341629594564438,
-0.037048231810331345,
-0.048005811870098114,
-0.048677507787942886,
0.012718061916530132,
-0.04027487337589264,
0.02389652654528618,
0.019828325137495995,
-0.08107800781726837,
-0.12106369435787201,
0.029826350510120392,
-0.014227760024368763,
-0.05404813587665558,
0.011215034872293472,
0.10382407903671265,
0.023829512298107147,
0.013394729234278202,
-0.0694594457745552,
-0.01326905470341444,
0.02290501818060875,
0.035424042493104935,
0.045564133673906326,
0.0005339807248674333,
0.07922974228858948,
-0.0042354390025138855,
0.027801990509033203,
0.018442004919052124,
0.08453378081321716,
0.006308405194431543,
0.057205237448215485,
-0.03909027576446533,
0.1126444935798645,
0.05829113349318504,
0.007161509245634079,
0.006312144920229912,
0.056928496807813644,
0.05721665173768997,
-0.12609021365642548,
0.07210034132003784,
0.011796348728239536,
0.03543759882450104,
-0.1344502866268158,
-0.006534187588840723,
0.026223493739962578,
-0.039276912808418274,
-0.022495655342936516,
-0.05973440781235695,
0.02242887206375599,
0.03116939775645733,
-0.09460612386465073,
-0.006661694031208754,
-0.07949327677488327,
-0.017157940194010735,
0.04082845151424408,
-0.09106901288032532,
-0.05706178769469261,
-0.006374478805810213,
0.04241415485739708,
0.00953592173755169,
-2.7049160916817498e-33,
-0.01589982584118843,
-0.05260244011878967,
0.047324590384960175,
0.02767036110162735,
0.0010096868500113487,
-0.03591320291161537,
-0.02553480491042137,
-0.05103764683008194,
0.06775893270969391,
-0.0912315621972084,
-0.11588069051504135,
-0.023160550743341446,
0.05026286467909813,
-0.0002266445371787995,
0.05743683502078056,
0.003243186743929982,
0.0023922480177134275,
0.02034660428762436,
-0.0026570495683699846,
0.061959151178598404,
-0.008427918888628483,
0.003815034171566367,
-0.05469023436307907,
0.006485708523541689,
-0.031812649220228195,
0.025109004229307175,
0.0374298021197319,
0.04499904438853264,
-0.007022042293101549,
0.015964940190315247,
-0.03318190947175026,
-0.07277572154998779,
0.003903678385540843,
0.015528247691690922,
-0.0626644641160965,
0.04237806797027588,
0.07143362611532211,
0.043396856635808945,
-0.01968930847942829,
0.06227360665798187,
0.03072931617498398,
-0.1310349702835083,
-0.04764535278081894,
0.006919075734913349,
0.06140923127532005,
0.040313493460416794,
-0.014884453266859055,
-0.0062149022705852985,
-0.009806622751057148,
0.049700066447257996,
0.004950548522174358,
0.09949608892202377,
-0.07400563359260559,
0.026517916470766068,
-0.0313236229121685,
-0.007581150159239769,
-0.02375892922282219,
-0.011029627174139023,
0.09528066962957382,
0.048017337918281555,
-0.007520364597439766,
-0.010429725982248783,
0.05858732387423515,
-0.023882664740085602,
-0.050350386649370193,
-0.09410539269447327,
-0.028165806084871292,
0.011732913553714752,
0.0411868654191494,
-0.037460967898368835,
0.030588719993829727,
-0.028767818585038185,
-0.026623783633112907,
-0.01166294515132904,
0.04473089426755905,
-0.10619884729385376,
0.0019723731093108654,
-0.02279832772910595,
0.016444645822048187,
-0.028427129611372948,
0.05505596846342087,
0.05138557776808739,
0.03449258580803871,
-0.002480957191437483,
0.007345287594944239,
0.07929862290620804,
-0.05095946043729782,
0.03739536553621292,
0.04701413959264755,
0.022809552028775215,
-0.018222572281956673,
0.014792794361710548,
-0.014160770922899246,
0.02796660177409649,
-0.09969601035118103,
-3.8195935303519946e-8,
-0.03626976162195206,
0.10685253143310547,
0.07405095547437668,
-0.04191932454705238,
0.12483429163694382,
0.004444415215402842,
0.02791653946042061,
0.05159204453229904,
0.039374105632305145,
0.03609488904476166,
0.1237783208489418,
0.007018541917204857,
0.01117285992950201,
-0.13454779982566833,
-0.0028059124015271664,
-0.006480008829385042,
0.012181124649941921,
-0.06711319088935852,
0.021084705367684364,
0.0267266184091568,
0.04712185263633728,
0.030772756785154343,
0.002337409183382988,
-0.016386505216360092,
0.005236038472503424,
-0.05645706504583359,
-0.010753864422440529,
0.12518903613090515,
0.06537574529647827,
0.021792525425553322,
0.06255532801151276,
0.026783673092722893,
0.011863384395837784,
0.022617653012275696,
-0.053853895515203476,
-0.022967081516981125,
0.0747881680727005,
-0.07771573960781097,
0.008991328068077564,
-0.041600048542022705,
-0.018435578793287277,
0.03522448241710663,
-0.02401730976998806,
0.0058389343321323395,
-0.06663908809423447,
0.05809571593999863,
-0.06955000013113022,
-0.0784279927611351,
0.08251491189002991,
-0.010785363614559174,
-0.015657395124435425,
-0.04437016695737839,
-0.006182078272104263,
0.07528352737426758,
0.10680409520864487,
0.04979518800973892,
0.058288026601076126,
-0.06491599977016449,
0.019386054947972298,
0.005134832579642534,
-0.039221152663230896,
-0.06940588355064392,
-0.12361202389001846,
0.03817104175686836
] | https://github.com/baitingluo/l-map | 0 | 0 | 0 | 0 |
Modeling Complex System Dynamics with Flow Matching Across Time and Conditions | https://openreview.net/forum?id=hwnObmOTrV | [
"Martin Rohbeck",
"Edward De Brouwer",
"Charlotte Bunne",
"Jan-Christian Huetter",
"Anne Biton",
"Kelvin Y. Chen",
"Aviv Regev",
"Romain Lopez"
] | Spotlight | Modeling the dynamics of complex real-world systems from temporal snapshot data is crucial for understanding phenomena such as gene regulation, climate change, and financial market fluctuations. Researchers have recently proposed a few methods based either on the Schroedinger Bridge or Flow Matching to tackle this problem, but these approaches remain limited in their ability to effectively combine data from multiple time points and different experimental settings. This integration is essential in real-world scenarios where observations from certain combinations of time points and experimental conditions are missing, either because of experimental costs or sensory failure. To address this challenge, we propose a novel method named Multi-Marginal Flow Matching (MMFM). MMFM first constructs a flow using smooth spline-based interpolation across time points and conditions and regresses it with a neural network using the classifier-free guided Flow Matching framework. This framework allows for the sharing of contextual information about the dynamics across multiple trajectories. We demonstrate the effectiveness of our method on both synthetic and real-world datasets, including a recent single-cell genomics data set with around a hundred chemical perturbations across time points. Our results show that MMFM significantly outperforms existing methods at imputing data at missing time points. | Flow Matching, dynamical systems | null | 12,998 | null | [
-0.11732758581638336,
-0.09664812684059143,
0.040452741086483,
0.029130496084690094,
0.047552451491355896,
0.031189516186714172,
-0.06585350632667542,
-0.04159657284617424,
0.09176567196846008,
-0.06509120017290115,
-0.06473902612924576,
-0.09399430453777313,
-0.011932993307709694,
0.025814462453126907,
-0.06926926225423813,
0.030722394585609436,
-0.02988593466579914,
0.0530211441218853,
-0.02533704601228237,
-0.0142307598143816,
-0.0013533314922824502,
0.03203311190009117,
-0.10728440433740616,
0.014461084268987179,
0.05820140615105629,
0.06750927865505219,
-0.008398744277656078,
0.010012952610850334,
-0.08578941226005554,
0.06147104129195213,
0.0706155076622963,
0.06100011616945267,
-0.09575587511062622,
0.03593974933028221,
-0.014919093810021877,
0.007836012169718742,
-0.056550540030002594,
0.0867026075720787,
0.031144198030233383,
-0.01157945953309536,
0.056143105030059814,
0.03600245714187622,
0.030352845788002014,
0.03132876008749008,
0.08759814500808716,
0.07693464308977127,
0.0016852633561939,
-0.021744634956121445,
-0.013598362915217876,
0.047074250876903534,
-0.08902920037508011,
0.02236281894147396,
0.0027130702510476112,
0.06870570778846741,
0.041846342384815216,
0.06377945840358734,
0.026025598868727684,
-0.02485731616616249,
0.028240852057933807,
0.045521821826696396,
0.04742007702589035,
-0.1129164844751358,
-0.03739066794514656,
-0.0009883403545245528,
0.02580089494585991,
-0.03166940063238144,
-0.011969809420406818,
0.10260649025440216,
0.029206808656454086,
-0.0035026459954679012,
-0.034539639949798584,
0.054780200123786926,
-0.025433821603655815,
-0.007320535369217396,
0.009623279795050621,
0.04188620671629906,
0.05155373364686966,
0.025860700756311417,
0.05013516545295715,
-0.07125711441040039,
-0.012288899160921574,
-0.00995886605232954,
0.02889157459139824,
-0.017151586711406708,
0.06870605796575546,
0.0313226543366909,
0.02603309601545334,
0.09112904965877533,
0.06015939638018608,
-0.011566178873181343,
-0.038673873990774155,
0.012890025973320007,
0.06826435774564743,
-0.06585391610860825,
0.05354418233036995,
0.061150070279836655,
0.004902144894003868,
-0.015376322902739048,
0.08571986854076385,
0.037496693432331085,
-0.013389530591666698,
0.024780895560979843,
0.040340881794691086,
0.06195961683988571,
0.08255012333393097,
-0.02607424184679985,
0.029898880049586296,
0.038871943950653076,
0.03248830512166023,
0.005469274707138538,
-0.022099582478404045,
-0.04308374598622322,
0.0530603751540184,
0.015807492658495903,
0.011645342223346233,
-0.06022370606660843,
-0.022185945883393288,
-0.05905037745833397,
0.033910784870386124,
0.11960447579622269,
-0.11757105588912964,
-0.02888190746307373,
-0.048097964376211166,
0.020700471475720406,
-0.017297843471169472,
0.08899538964033127,
-0.045244134962558746,
1.0898190879110782e-33,
-0.01954067312180996,
-0.012183455750346184,
0.04165506362915039,
0.01925375871360302,
0.02224235236644745,
-0.033530209213495255,
-0.08131281286478043,
-0.03251232951879501,
-0.05233046039938927,
0.05285544693470001,
-0.07290703058242798,
-0.019055170938372612,
-0.08018054068088531,
0.03655821830034256,
-0.0904599204659462,
-0.016375476494431496,
-0.08744063228368759,
0.02097783237695694,
0.057641495019197464,
0.03610475733876228,
-0.034495845437049866,
-0.08613841980695724,
-0.025837693363428116,
-0.05055180937051773,
0.01351851411163807,
-0.02184442989528179,
0.05817485600709915,
0.004234968218952417,
0.0288444384932518,
-0.011181621812283993,
-0.04729553684592247,
0.005390554200857878,
0.002161964075639844,
-0.035621143877506256,
0.04030948877334595,
-0.049782074987888336,
0.06550391763448715,
-0.027508076280355453,
-0.019501125440001488,
0.01801641844213009,
-0.0790926069021225,
0.013994747772812843,
-0.07666183263063431,
-0.009276618249714375,
-0.05711166188120842,
-0.08029656857252121,
-0.05218977853655815,
0.04000967741012573,
-0.039648015052080154,
0.0435565784573555,
0.000023486631107516587,
0.04247729107737541,
-0.0010958603816106915,
-0.12356825917959213,
-0.10437095910310745,
0.013500144705176353,
-0.024455880746245384,
-0.007377780973911285,
-0.011200827546417713,
0.07320813834667206,
-0.1261993646621704,
0.04082544147968292,
-0.022291075438261032,
-0.039441950619220734,
0.07271192222833633,
0.0597502738237381,
-0.04844838008284569,
0.07055170834064484,
0.06287949532270432,
0.014576463028788567,
0.023029906675219536,
-0.004248506389558315,
-0.0031916212756186724,
-0.01938437484204769,
0.05833393335342407,
-0.030139770358800888,
0.051460906863212585,
0.04471663013100624,
-0.07364820688962936,
0.09413309395313263,
-0.06895165890455246,
-0.008419143036007881,
0.0409216433763504,
0.02274993434548378,
0.007112535648047924,
-0.08324316143989563,
-0.012625188566744328,
-0.09723255783319473,
-0.06563485413789749,
-0.10984916985034943,
-0.03365471586585045,
-0.03695708140730858,
-0.016889790073037148,
0.020689966157078743,
0.027308443561196327,
-2.9132799958756852e-33,
-0.03929390385746956,
0.0059265512973070145,
0.0334581695497036,
0.03554436191916466,
0.047056715935468674,
0.020000731572508812,
-0.059401337057352066,
-0.03376185894012451,
0.10479950159788132,
0.04223478212952614,
0.05632268637418747,
0.006358967162668705,
0.03099989704787731,
-0.06716031581163406,
0.03584425896406174,
0.014735438860952854,
-0.0003012904489878565,
-0.08275992423295975,
0.032283999025821686,
0.04143602401018143,
0.04557017236948013,
0.00906321406364441,
-0.10516896843910217,
0.018809521570801735,
-0.10885629802942276,
0.05232223495841026,
0.008975240401923656,
0.01087866723537445,
-0.06531230360269547,
0.009299292229115963,
-0.0543450303375721,
-0.039044614881277084,
-0.05930478125810623,
-0.0284841638058424,
-0.03829290345311165,
0.08038601279258728,
-0.006318359635770321,
-0.0013357118004933,
-0.023226773366332054,
0.010997657664120197,
0.03849164396524429,
0.013812296092510223,
-0.08813879638910294,
0.021973742172122,
-0.01587725430727005,
0.048646245151758194,
-0.05276622250676155,
0.07413174211978912,
0.043599940836429596,
0.04550880193710327,
-0.011060363613069057,
0.11494094878435135,
-0.10266637802124023,
-0.020378515124320984,
-0.018231341615319252,
-0.0897943526506424,
0.025128858163952827,
-0.005650379229336977,
-0.06123203784227371,
-0.027169806882739067,
-0.05515744537115097,
-0.0311297494918108,
0.0013273971853777766,
0.02436390146613121,
0.012512545101344585,
-0.04411447048187256,
0.02782217226922512,
-0.07746542245149612,
0.011680865660309792,
-0.02304205857217312,
0.03833593800663948,
0.019638666883111,
0.0009049547952599823,
-0.03531208261847496,
0.0788302794098854,
-0.011522410437464714,
0.00824996829032898,
-0.020863618701696396,
-0.00916969869285822,
0.08069627732038498,
0.026606576517224312,
0.052482422441244125,
0.1058669239282608,
0.08032470941543579,
0.10947694629430771,
0.059163130819797516,
-0.035948578268289566,
-0.0017668182263150811,
0.05869220569729805,
-0.021957149729132652,
-0.01666106842458248,
0.004199265036731958,
-0.04979689419269562,
-0.017089728266000748,
-0.09409957379102707,
-5.0459373568401134e-8,
0.012408601120114326,
0.028798235580325127,
0.05007883161306381,
-0.008729216642677784,
-0.03338368609547615,
-0.006973111070692539,
-0.019340911880135536,
0.07557997107505798,
0.04806813970208168,
-0.016145924106240273,
0.09592226147651672,
0.008023925125598907,
-0.06982401013374329,
0.009046788327395916,
-0.01105183269828558,
-0.03257586061954498,
-0.006942146457731724,
-0.03729455918073654,
-0.02066466026008129,
0.03617362678050995,
-0.05459079518914223,
0.05926607549190521,
-0.0061500780284404755,
0.05846963822841644,
0.12301566451787949,
-0.05914818495512009,
-0.001752415206283331,
0.044069804251194,
0.10178026556968689,
-0.02656153216958046,
-0.0011477317893877625,
-0.010152054019272327,
-0.0113332225009799,
-0.013903562910854816,
-0.006416554097086191,
-0.03754067420959473,
0.03620945289731026,
0.04933229461312294,
-0.09273471683263779,
0.04155563563108444,
0.0016187059227377176,
0.09164654463529587,
-0.11539915949106216,
-0.014126869849860668,
0.04503665864467621,
0.044520843774080276,
0.026571476832032204,
-0.06584298610687256,
0.06731811910867691,
0.027490053325891495,
-0.01590512879192829,
0.0291343554854393,
0.012453104369342327,
0.03569343313574791,
0.02859146147966385,
-0.017503244802355766,
-0.08377664536237717,
-0.06430578231811523,
0.05622434616088867,
0.01336885429918766,
-0.05108661949634552,
-0.02100639045238495,
-0.09007355570793152,
-0.038385942578315735
] | 0 | 0 | 0 | 0 |
|
Joint Reward and Policy Learning with Demonstrations and Human Feedback Improves Alignment | https://openreview.net/forum?id=VCbqXtS5YY | [
"Chenliang Li",
"Siliang Zeng",
"Zeyi Liao",
"Jiaxiang Li",
"Dongyeop Kang",
"Alfredo Garcia",
"Mingyi Hong"
] | Spotlight | Aligning to human preferences and/or intentions is an important requirement for contemporary foundation models. To ensure alignment, popular approaches such as reinforcement learning with human feedback (RLHF) break down the task into three stages: (i) a model is computed with supervised fine-tuning (SFT) based upon large demonstrations data, (ii) a reward model (RM) is estimated based upon human feedback data, and (iii) reinforcement learning (RL) is used to further refine the SFT model by optimizing the estimated reward model. Demonstrations and human feedback data reflect human user preferences in different ways. As a result, the reward model estimate obtained from only human feedback data is likely not as accurate as a reward model estimate obtained from both demonstration and human feedback data. A policy model that optimizes the reward model estimate obtained from both demonstration and human feedback data will likely exhibit better alignment performance. We introduce a tractable algorithm for finding the reward and policy models and provide a finite-time performance guarantee. Additionally, we demonstrate the efficiency of the proposed solution with extensive experiments including alignment problems in LLMs and robotic control problems in MuJoCo. We observe that the proposed solutions outperform the existing alignment algorithm by large margins, especially when the amounts of demonstration and preference data are unbalanced. | Alignment, Inverse Reinforcement Learning, Reinforment Learning from Human Feedback | null | 12,997 | null | [
-0.04413554444909096,
-0.08691760152578354,
0.030182315036654472,
-0.048447027802467346,
0.02830958366394043,
0.04207223653793335,
-0.009276514872908592,
0.02737458609044552,
0.039986319839954376,
0.06083489954471588,
-0.00836652796715498,
-0.07612507045269012,
0.07171092927455902,
0.07498268038034439,
0.022028394043445587,
0.00830246414989233,
0.10257157683372498,
0.0268369372934103,
-0.04726671427488327,
-0.01614668220281601,
0.037167470902204514,
0.039588674902915955,
0.09933052211999893,
0.005136598367244005,
-0.09868577122688293,
-0.00808230135589838,
0.030764222145080566,
-0.019718751311302185,
-0.010203972458839417,
-0.039224665611982346,
0.027618039399385452,
-0.04919227957725525,
0.05845934897661209,
-0.08837757259607315,
-0.10976770520210266,
0.07992193102836609,
-0.15786121785640717,
-0.03830275312066078,
-0.04931626096367836,
-0.047886643558740616,
-0.0735880583524704,
-0.025428246706724167,
0.03336923569440842,
0.009830180555582047,
0.05241475626826286,
0.06265178322792053,
-0.003089416539296508,
-0.07415637373924255,
0.054308291524648666,
0.02741706743836403,
-0.02102375403046608,
0.008917583152651787,
0.06394189596176147,
0.0008846779819577932,
0.04583714157342911,
0.0375790074467659,
0.03593568131327629,
-0.006607310380786657,
0.06371080875396729,
-0.04057570919394493,
0.005704991519451141,
-0.11015372723340988,
-0.05622497946023941,
-0.04213601350784302,
-0.022019604220986366,
-0.06583674252033234,
-0.040200646966695786,
0.010111391544342041,
-0.009740937501192093,
-0.0014485879801213741,
-0.043279845267534256,
0.025302503257989883,
0.038855984807014465,
-0.016227979212999344,
-0.011361476965248585,
0.04981931671500206,
0.013146874494850636,
-0.007679312024265528,
0.007690885569900274,
-0.04082499444484711,
-0.028416180983185768,
-0.008857179433107376,
0.03040062077343464,
0.018012065440416336,
0.05362466350197792,
-0.013454982079565525,
0.008535892702639103,
0.10595910251140594,
0.109170101583004,
-0.002777061890810728,
-0.04392779618501663,
-0.0017793882871046662,
-0.04459327459335327,
-0.03375173732638359,
-0.046279098838567734,
0.05465975031256676,
-0.02061045542359352,
-0.0385650172829628,
-0.08637844026088715,
0.09058671444654465,
-0.019533108919858932,
0.12668584287166595,
-0.02126302197575569,
-0.004120601341128349,
0.002201644005253911,
-0.05556202679872513,
-0.020389750599861145,
0.05031251162290573,
0.07898200303316116,
-0.13276338577270508,
0.005628087557852268,
-0.06547269225120544,
-0.024414459243416786,
0.1513816863298416,
-0.005292047280818224,
-0.05342784896492958,
-0.00289780180901289,
-0.07388001680374146,
-0.0711798369884491,
0.01711207628250122,
0.03404625132679939,
-0.055936526507139206,
0.04515891894698143,
0.023535149171948433,
0.008376946672797203,
-0.043191924691200256,
-0.019193565472960472,
2.7378854021444532e-33,
0.038882240653038025,
0.05797833949327469,
0.06119199097156525,
0.02553817443549633,
0.04071465879678726,
0.014598317444324493,
0.04602036252617836,
0.03973881155252457,
-0.0480806864798069,
-0.0013996694469824433,
-0.0021495220717042685,
-0.040556035935878754,
-0.05426212027668953,
0.10671356320381165,
-0.016011886298656464,
-0.024220911785960197,
-0.02819686383008957,
0.03488266095519066,
-0.06148713082075119,
0.006702940445393324,
0.07872340828180313,
0.029376544058322906,
-0.0031484998762607574,
-0.04864294454455376,
0.05514652281999588,
-0.04679972678422928,
0.0063012344762682915,
0.01775551214814186,
-0.034297212958335876,
-0.03968704491853714,
-0.032074179500341415,
-0.0028703303541988134,
-0.01569700613617897,
0.010419814847409725,
0.020308909937739372,
-0.016297265887260437,
-0.01821576990187168,
-0.051750652492046356,
-0.016903147101402283,
-0.028000863268971443,
0.014946495182812214,
0.01207443792372942,
0.03706997632980347,
0.014901336282491684,
-0.10435047745704651,
-0.07035756856203079,
0.030693357810378075,
-0.03667011111974716,
0.01114693284034729,
0.009179587475955486,
0.01560710184276104,
0.0661875382065773,
0.03100915439426899,
-0.09648311883211136,
-0.06859750300645828,
0.017658192664384842,
0.006311800330877304,
0.05462639406323433,
-0.06553144007921219,
0.026944199576973915,
-0.015575106255710125,
-0.00824044644832611,
0.0013851859839633107,
0.03034489043056965,
-0.03260597959160805,
0.07653077691793442,
-0.0179695263504982,
0.050108544528484344,
0.056638773530721664,
0.050058167427778244,
0.007060955744236708,
0.027074433863162994,
0.035822752863168716,
-0.04410543292760849,
0.061292506754398346,
-0.01752307638525963,
0.02293332852423191,
-0.09440968185663223,
0.03527860715985298,
-0.007204913534224033,
-0.03681924566626549,
-0.00037496857112273574,
-0.04320703074336052,
-0.04323482885956764,
-0.0064359400421381,
-0.09038837254047394,
0.02591267041862011,
-0.06090759113430977,
0.015040024183690548,
0.013940691947937012,
-0.0644238144159317,
0.03528902679681778,
-0.02929220348596573,
0.12329050898551941,
0.03913315013051033,
-3.581071540751111e-33,
0.061334650963544846,
-0.02578127384185791,
0.09221938997507095,
0.009458946995437145,
-0.03402506560087204,
0.029389888048171997,
-0.01877288520336151,
-0.004786541219800711,
-0.032415155321359634,
-0.035281952470541,
-0.04873929172754288,
-0.024930093437433243,
-0.0005914910580031574,
-7.825599368516123e-7,
0.005407894961535931,
0.014930113218724728,
-0.010938544757664204,
0.030095906928181648,
0.056958600878715515,
-0.010027023032307625,
0.03789331391453743,
0.13114485144615173,
-0.11904291063547134,
0.03622724115848541,
-0.01350585650652647,
0.017112743109464645,
0.04657678306102753,
0.022272782400250435,
-0.0032319233287125826,
0.07126855850219727,
-0.09472887217998505,
-0.06457895785570145,
-0.08738883584737778,
0.030305100604891777,
-0.06982146203517914,
0.029933951795101166,
0.01876947470009327,
0.002523108385503292,
-0.01790178008377552,
0.08640974014997482,
-0.016197804361581802,
-0.018498992547392845,
-0.06282826513051987,
0.01715240813791752,
0.007625488098710775,
-0.06276361644268036,
-0.07737705856561661,
-0.015848221257328987,
-0.0325499102473259,
-0.03188853710889816,
-0.0442720502614975,
0.0539521761238575,
-0.06418837606906891,
-0.07208319753408432,
-0.04727795720100403,
-0.035909503698349,
-0.03690877556800842,
-0.00175268214661628,
0.12012418359518051,
-0.04146525263786316,
-0.0331730842590332,
0.04236552119255066,
0.00745740532875061,
0.07992978394031525,
0.0038292494136840105,
0.07694130390882492,
0.03128896281123161,
0.022346338257193565,
0.02195795252919197,
0.03441100940108299,
0.013027689419686794,
0.005318942479789257,
0.03887251019477844,
-0.02115567773580551,
0.09189386665821075,
-0.05910623446106911,
0.007162787020206451,
0.0013050216948613524,
0.0034974957816302776,
-0.0895611122250557,
0.05026647821068764,
0.03231627121567726,
0.05620187148451805,
0.04066222533583641,
-0.04239688068628311,
0.09448191523551941,
0.058225978165864944,
0.11646022647619247,
-0.01768677309155464,
-0.009249214082956314,
-0.039555441588163376,
0.03241315856575966,
0.07909490168094635,
0.031914036720991135,
-0.01994701288640499,
-4.775413842139642e-8,
-0.13749496638774872,
0.027147933840751648,
0.06350452452898026,
0.03843345865607262,
-0.029391922056674957,
0.032256126403808594,
0.0046011460945010185,
-0.006305369548499584,
-0.008094157092273235,
-0.08175920695066452,
0.04114275425672531,
0.08355363458395004,
-0.0007030537817627192,
-0.05439215153455734,
-0.0755695104598999,
0.05812729895114899,
-0.02159249596297741,
0.04513632133603096,
0.008221209049224854,
0.08876384794712067,
0.03548649325966835,
0.028880435973405838,
0.019556967541575432,
-0.06075191870331764,
0.0026945050340145826,
-0.08094670623540878,
-0.05714032053947449,
0.16233763098716736,
0.005558284465223551,
0.007329284213483334,
0.021192461252212524,
-0.07451346516609192,
0.03887409344315529,
0.00913612823933363,
0.11534006148576736,
-0.0032089389860630035,
0.009363760240375996,
-0.06987816840410233,
0.004268386401236057,
-0.017881965264678,
0.014805048704147339,
0.14234301447868347,
-0.01730627566576004,
0.007459789980202913,
0.027822788804769516,
0.013123387470841408,
-0.0621110275387764,
-0.08506715297698975,
0.01422516256570816,
-0.08142876625061035,
0.05313779041171074,
-0.043575722724199295,
-0.04797636717557907,
0.024604549631476402,
0.03917178139090538,
0.014586484991014004,
-0.013821727596223354,
-0.07873209565877914,
0.05936665087938309,
0.05120258778333664,
0.010938743129372597,
-0.01156316976994276,
-0.0723477154970169,
-0.03130975738167763
] | 0 | 0 | 0 | 0 |
|
Counterfactual Realizability | https://openreview.net/forum?id=uuriavczkL | [
"Arvind Raghavan",
"Elias Bareinboim"
] | Spotlight | It is commonly believed that, in a real-world environment, samples can only be drawn from observational and interventional distributions, corresponding to Layers 1 and 2 of the *Pearl Causal Hierarchy*. Layer 3, representing counterfactual distributions, is believed to be inaccessible by definition. However, Bareinboim, Forney, and Pearl (2015) introduced a procedure that allows an agent to sample directly from a counterfactual distribution, leaving open the question of what other counterfactual quantities can be estimated directly via physical experimentation. We resolve this by introducing a formal definition of realizability, the ability to draw samples from a distribution, and then developing a complete algorithm to determine whether an arbitrary counterfactual distribution is realizable given fundamental physical constraints, such as the inability to go back in time and subject the same unit to a different experimental condition. We illustrate the implications of this new framework for counterfactual data collection using motivating examples from causal fairness and causal reinforcement learning. While the baseline approach in these motivating settings typically follows an interventional or observational strategy, we show that a counterfactual strategy provably dominates both. | causal inference, experiment design, causal reinforcement learning, counterfactual reasoning | A complete algorithm for which counterfactual (Layer 3) distributions can be experimentally realized; and its implications for optimal decision-making | 12,927 | 2503.11870 | [
-0.10810037702322006,
-0.03585352376103401,
0.03485383838415146,
0.08162577450275421,
0.04532885551452637,
0.02516629733145237,
-0.035044290125370026,
-0.046592436730861664,
0.027082698419690132,
0.03964103385806084,
-0.016019722446799278,
-0.10057047754526138,
0.07175067067146301,
0.0653698518872261,
-0.0693427324295044,
-0.07780716568231583,
0.10745655000209808,
0.06399554759263992,
-0.0536239892244339,
-0.0017371941357851028,
0.0321684256196022,
-0.013506014831364155,
0.06369441747665405,
0.004736238159239292,
-0.052457913756370544,
-0.026035837829113007,
-0.0345802865922451,
-0.03965597227215767,
0.007124884519726038,
0.0034352182410657406,
0.051391955465078354,
0.05301549285650253,
-0.09916595369577408,
-0.06992407888174057,
-0.011288325302302837,
0.009705211967229843,
-0.020058738067746162,
0.02614799328148365,
0.011890844441950321,
0.04330790787935257,
-0.02256772853434086,
-0.010809269733726978,
-0.06260126829147339,
-0.0065987491980195045,
0.047295160591602325,
-0.021407995373010635,
0.05319651588797569,
0.04363393783569336,
-0.08763515949249268,
0.016509737819433212,
-0.07195240259170532,
0.044943924993276596,
0.012568921782076359,
0.004118788056075573,
0.08827801048755646,
0.012157458811998367,
0.11247634142637253,
0.006888336036354303,
-0.006632099859416485,
-0.019680412486195564,
-0.060621265321969986,
-0.08949974924325943,
-0.11042143404483795,
-0.02073420211672783,
-0.03566930070519447,
0.06362613290548325,
-0.06649203598499298,
0.08705387264490128,
0.03794213384389877,
0.06321363151073456,
0.032973166555166245,
0.06908511370420456,
-0.030640624463558197,
0.04253370687365532,
0.049096982926130295,
-0.00014002711395733058,
-0.04598983749747276,
0.001694267732091248,
-0.01032630167901516,
-0.05129664018750191,
0.02238580211997032,
-0.052948445081710815,
0.012984270229935646,
0.05225517973303795,
0.06458935886621475,
-0.06738219410181046,
0.010903491638600826,
0.06985723972320557,
0.04885439947247505,
0.06450431793928146,
0.00986047089099884,
-0.0024565565399825573,
0.051746636629104614,
0.0397116057574749,
0.04568920657038689,
0.05818437784910202,
0.007896021008491516,
-0.04820086434483528,
0.06435736268758774,
0.05859585106372833,
0.054188162088394165,
-0.002237957436591387,
0.002214189851656556,
0.034736793488264084,
0.00789634883403778,
-0.09209004044532776,
-0.026643306016921997,
-0.005400152876973152,
-0.014650837518274784,
-0.0686056911945343,
0.026396630331873894,
-0.04707258567214012,
0.043426722288131714,
-0.03765949606895447,
0.0006955849239602685,
-0.05421784892678261,
0.01606808230280876,
0.010238570161163807,
-0.01055212039500475,
-0.10276353359222412,
0.0876893699169159,
-0.03712236136198044,
0.09532859921455383,
-0.014253053814172745,
0.08647304028272629,
0.012442002072930336,
0.00903240405023098,
3.6508334570525675e-33,
0.053168877959251404,
-0.01941061019897461,
0.011031638830900192,
-0.04823632165789604,
0.028557367622852325,
-0.042236633598804474,
-0.07528916001319885,
0.012600904330611229,
0.11417051404714584,
0.011091097258031368,
0.06965745240449905,
0.10188068449497223,
-0.08853790909051895,
0.018184348940849304,
0.05493561923503876,
0.021036555990576744,
-0.07005959749221802,
0.11370126903057098,
0.010201863944530487,
-0.05321794003248215,
0.03424336761236191,
-0.06799101829528809,
-0.016992369666695595,
-0.07720012962818146,
-0.03278138488531113,
0.009718435816466808,
-0.009963671676814556,
0.027028841897845268,
-0.027605384588241577,
0.024422423914074898,
-0.009057427756488323,
0.006715117022395134,
0.04056686908006668,
0.037197090685367584,
-0.039994336664676666,
-0.04905230924487114,
-0.06807919591665268,
-0.08148325234651566,
-0.06983385235071182,
-0.011099297553300858,
0.000719882023986429,
-0.009428410790860653,
0.004640444181859493,
0.05826558545231819,
-0.008807496167719364,
-0.1590900421142578,
0.005437406711280346,
-0.019374972209334373,
-0.05191297456622124,
-0.0733644962310791,
0.11636769026517868,
0.07978291809558868,
0.0790175348520279,
-0.10106999427080154,
-0.018005896359682083,
0.05025652050971985,
-0.028856666758656502,
-0.014256127178668976,
-0.03661194071173668,
0.05951021611690521,
0.018468376249074936,
0.0006202423828653991,
-0.040828872472047806,
0.09099163860082626,
-0.048901285976171494,
0.07582934200763702,
-0.050885170698165894,
-0.02350042387843132,
-0.014796511270105839,
0.046057816594839096,
0.020228596404194832,
0.05634593218564987,
-0.060418155044317245,
-0.028750812634825706,
0.04052088037133217,
-0.02332952246069908,
0.03290114551782608,
0.032370325177907944,
-0.00255004083737731,
-0.0310487300157547,
-0.03948310390114784,
-0.07238275557756424,
-0.040644899010658264,
0.03962643817067146,
-0.060861025005578995,
-0.026826202869415283,
0.002288062823936343,
0.005889496766030788,
-0.0042478907853364944,
-0.06640758365392685,
-0.05178903415799141,
-0.03026067465543747,
0.015122427605092525,
0.06619954854249954,
0.017016522586345673,
-4.60340338576507e-33,
-0.06589330732822418,
-0.0018915956607088447,
0.01583433337509632,
0.00716871814802289,
0.023880071938037872,
-0.014440189115703106,
-0.033698685467243195,
-0.0905219092965126,
0.08284355700016022,
-0.07372108101844788,
-0.09100837260484695,
-0.0034702750854194164,
0.06647203117609024,
0.00278560072183609,
-0.04698306322097778,
-0.03508647531270981,
0.06152394786477089,
0.023183569312095642,
-0.04111483693122864,
-0.009128288365900517,
0.041795846074819565,
0.005718901287764311,
-0.05823669582605362,
-0.038300398737192154,
-0.007039784919470549,
0.021756883710622787,
-0.026390785351395607,
0.028750691562891006,
0.04964914172887802,
0.026077907532453537,
-0.02218277007341385,
-0.014794263988733292,
0.009664712473750114,
-0.03539612144231796,
-0.04685956612229347,
0.065302774310112,
0.09055270999670029,
0.010899011045694351,
-0.08294955641031265,
0.0341402031481266,
0.08237072825431824,
-0.0073545342311263084,
-0.11369215697050095,
-0.028101667761802673,
0.02924589067697525,
-0.017164498567581177,
-0.05742473155260086,
0.043208785355091095,
0.056459005922079086,
-0.07149173319339752,
-0.0705542340874672,
0.08983273059129715,
0.036449410021305084,
0.0704997181892395,
-0.033907338976860046,
-0.048834118992090225,
0.02061728946864605,
-0.03818556293845177,
0.10025036334991455,
0.06637104600667953,
-0.049375422298908234,
-0.043000299483537674,
-0.01603819988667965,
0.062290992587804794,
0.024888385087251663,
0.021666036918759346,
-0.034206490963697433,
-0.0017788517288863659,
-0.006605137139558792,
0.002828158438205719,
0.06084926798939705,
0.055069372057914734,
-0.06986981630325317,
-0.04706136882305145,
0.06706554442644119,
-0.002528268378227949,
-0.07128497958183289,
0.004213846754282713,
-0.04576585814356804,
-0.0076197353191673756,
0.04948253929615021,
0.00039698451291769743,
0.056450005620718,
-0.004328547976911068,
-0.06271838396787643,
0.033888015896081924,
-0.008169617503881454,
0.023530786857008934,
0.06568848341703415,
-0.003701678942888975,
-0.04620052129030228,
-0.04871877655386925,
-0.09347223490476608,
0.05971954017877579,
-0.06321346014738083,
-5.1487830887708697e-8,
-0.008626151829957962,
0.005135236773639917,
0.09943278133869171,
0.0005265917861834168,
0.08216497302055359,
0.0050761159509420395,
-0.06826168298721313,
-0.012132888659834862,
-0.028185587376356125,
-0.07806975394487381,
0.06715570390224457,
-0.008716159500181675,
0.0443149209022522,
-0.05722244456410408,
-0.011192897334694862,
0.0541868582367897,
0.003314414294436574,
-0.05915680527687073,
-0.025487205013632774,
-0.0015702237142249942,
0.07783592492341995,
0.050069659948349,
-0.033422838896512985,
-0.05669303610920906,
0.027371926233172417,
-0.0173635371029377,
-0.002809236291795969,
0.053899265825748444,
0.05854080989956856,
0.008903488516807556,
0.027389638125896454,
-0.031843315809965134,
0.005773134063929319,
-0.03228452801704407,
-0.014993020333349705,
0.04901568591594696,
0.017580898478627205,
-0.02673248015344143,
-0.01778494566679001,
-0.0763336569070816,
-0.017963135614991188,
0.0694807916879654,
-0.09272453188896179,
0.07125546038150787,
0.11414753645658493,
-0.025118883699178696,
-0.10250452160835266,
-0.03586428984999657,
0.006765384692698717,
0.018070099875330925,
-0.052317310124635696,
-0.04794323816895485,
0.014370839111506939,
0.033439040184020996,
0.06327473372220993,
0.06646782159805298,
0.009805586189031601,
-0.03022294119000435,
-0.05903305858373642,
0.07008332014083862,
0.13405388593673706,
0.05604766309261322,
-0.04015570878982544,
-0.010637945495545864
] | 0 | 0 | 0 | 0 |
|
Robustness Reprogramming for Representation Learning | https://openreview.net/forum?id=SuH5SdOXpe | [
"Zhichao Hou",
"MohamadAli Torkamani",
"Hamid Krim",
"Xiaorui Liu"
] | Spotlight | This work tackles an intriguing and fundamental open challenge in representation learning: Given a well-trained deep learning model, can it be reprogrammed to enhance its robustness against adversarial or noisy input perturbations without altering its parameters?
To explore this, we revisit the core feature transformation mechanism in representation learning and propose a novel non-linear robust pattern matching technique as a robust alternative. Furthermore, we introduce three model reprogramming paradigms to offer flexible control of robustness under different efficiency requirements. Comprehensive experiments and ablation studies across diverse learning models ranging from basic linear model and MLPs to shallow and modern deep ConvNets demonstrate the effectiveness
of our approaches.
This work not only opens a promising and orthogonal direction for improving adversarial defenses in deep learning beyond existing methods but also provides new insights into designing more resilient AI systems with robust statistics.
Our implementation is available at https://github.com/chris-hzc/Robustness-Reprogramming. | Adversarial Robustness, Robustness Reprogramming, Robust Representation Learning | null | 12,910 | 2410.04577 | [
-0.08285338431596756,
-0.10268264263868332,
0.02154623530805111,
0.06456300616264343,
-0.03221408277750015,
0.07049129903316498,
-0.04000496119260788,
-0.10203071683645248,
-0.06156647950410843,
-0.10526593029499054,
-0.07281279563903809,
-0.026930328458547592,
0.043027717620134354,
0.025889208540320396,
-0.11001285910606384,
0.03269362822175026,
0.13774126768112183,
0.1313733607530594,
-0.10858950763940811,
-0.0021532652899622917,
0.01021642331033945,
0.02722952328622341,
-0.04211731255054474,
0.03450818359851837,
0.07694298774003983,
-0.024887772276997566,
-0.04040612280368805,
-0.056615572422742844,
-0.005460872780531645,
-0.08416762202978134,
0.042375411838293076,
-0.06999780982732773,
-0.041405558586120605,
-0.025792602449655533,
-0.03744152560830116,
0.06034651771187782,
-0.04974013566970825,
0.006936612538993359,
0.04802606999874115,
-0.09290529787540436,
-0.0065785981714725494,
0.007007341366261244,
0.011690592393279076,
0.030171124264597893,
0.04398185387253761,
0.0049914587289094925,
0.044017039239406586,
-0.06820639967918396,
0.02852690778672695,
-0.0028795809485018253,
-0.023885006085038185,
0.002948951441794634,
-0.008062247186899185,
0.09195071458816528,
0.014818049035966396,
-0.025629354640841484,
0.006909723859280348,
0.05935562774538994,
-0.012709390372037888,
0.021712694317102432,
-0.020655103027820587,
-0.052653174847364426,
0.03139592334628105,
0.01859615370631218,
-0.030499067157506943,
-0.04960430786013603,
0.012354684993624687,
0.041338011622428894,
0.004163045436143875,
0.05287961661815643,
0.003088295226916671,
0.11897975206375122,
-0.056170251220464706,
-0.008455662056803703,
0.059060171246528625,
0.10359703749418259,
-0.004533743020147085,
-0.059749554842710495,
0.10324929654598236,
-0.0453467071056366,
0.04748421534895897,
-0.0032799551263451576,
0.06865561008453369,
0.02215145342051983,
0.16281184554100037,
-0.0015825993614271283,
-0.025713618844747543,
0.01673700474202633,
0.027385598048567772,
-0.023483315482735634,
-0.05798175930976868,
-0.03641873970627785,
0.07300161570310593,
-0.01649322174489498,
-0.009089401923120022,
0.012055979110300541,
0.023943530395627022,
-0.07424603402614594,
-0.062011756002902985,
0.09320556372404099,
-0.06486578285694122,
0.04536215588450432,
0.056438643485307693,
0.002385841216892004,
0.04494093731045723,
-0.043056733906269073,
0.07875814288854599,
-0.011715717613697052,
-0.007333129644393921,
-0.07933156192302704,
0.07771418988704681,
-0.00992344506084919,
0.03432203084230423,
0.06200197711586952,
0.02834603376686573,
-0.011624176055192947,
-0.041839297860860825,
-0.012415125034749508,
-0.024034908041357994,
0.04251335188746452,
-0.058236218988895416,
0.033673837780952454,
-0.04176297038793564,
0.05576154962182045,
0.006741667632013559,
0.05137118697166443,
-0.025321971625089645,
3.0686112887612808e-33,
0.02274962328374386,
0.08646006882190704,
0.02465014159679413,
-0.005917551461607218,
0.05581808090209961,
-0.09179243445396423,
-0.07306650280952454,
0.03559189289808273,
-0.05177309364080429,
0.029243703931570053,
-0.004398955497890711,
-0.04501927271485329,
-0.06455894559621811,
0.1186441034078598,
-0.00883680209517479,
0.06056465953588486,
-0.004113911185413599,
0.012068009003996849,
-0.017795398831367493,
-0.07729299366474152,
0.010526304133236408,
-0.013648291118443012,
-0.05900082364678383,
-0.0548991858959198,
0.005591633263975382,
-0.003885493380948901,
0.05067215487360954,
-0.015528487041592598,
0.0038865713868290186,
0.025241151452064514,
-0.0946434736251831,
0.05762387812137604,
0.01330568641424179,
0.03574245423078537,
0.02362685464322567,
-0.07045704871416092,
0.011003127321600914,
-0.020973538979887962,
-0.015828032046556473,
0.03235118091106415,
0.041752737015485764,
0.05157243087887764,
-0.014767962507903576,
-0.026539357379078865,
0.012178828939795494,
-0.00576110789552331,
0.03999730572104454,
-0.014565672725439072,
-0.10898847877979279,
-0.08901942521333694,
-0.006194645538926125,
0.05695538595318794,
-0.054118990898132324,
-0.05906284600496292,
-0.02759656123816967,
-0.06593594700098038,
-0.012064083479344845,
0.028303442522883415,
-0.00730884587392211,
0.10649807006120682,
-0.01697005331516266,
0.0529383085668087,
0.025529388338327408,
0.027912026271224022,
-0.022955287247896194,
-0.014095439575612545,
0.0011745915981009603,
0.06266570091247559,
0.031979452818632126,
0.025470295920968056,
-0.021785888820886612,
0.025978319346904755,
-0.044956643134355545,
-0.06782163679599762,
-0.012092385441064835,
0.003781002713367343,
0.09025395661592484,
-0.0689905658364296,
-0.04372316598892212,
0.005768323317170143,
-0.04067236930131912,
0.015099981799721718,
-0.016444068402051926,
0.02974780648946762,
-0.03584704548120499,
-0.021269654855132103,
-0.0010086183901876211,
-0.044561758637428284,
-0.014547565020620823,
0.014383569359779358,
0.0224617850035429,
-0.06726235151290894,
-0.03913464769721031,
-0.04199731722474098,
-0.02380651980638504,
-2.518872253816037e-33,
-0.007664126344025135,
0.0924801230430603,
-0.04973648488521576,
0.005649566650390625,
-0.08503571897745132,
-0.02872920222580433,
-0.005885954946279526,
0.061036862432956696,
-0.0074709258042275906,
-0.05241193622350693,
-0.07102172821760178,
-0.029940389096736908,
0.037381161004304886,
-0.07187192142009735,
0.011690468527376652,
-0.07200763374567032,
-0.08251124620437622,
0.05832565575838089,
-0.02004133351147175,
0.0082332082092762,
0.06272098422050476,
0.08996928483247757,
-0.08954855054616928,
0.0502893403172493,
-0.02009790576994419,
0.030978234484791756,
-0.09440983086824417,
-0.0061882552690804005,
0.11026708036661148,
0.012485393323004246,
-0.056940775364637375,
0.019132887944579124,
-0.02251816913485527,
-0.005571871530264616,
-0.06812764704227448,
0.07272837311029434,
-0.001933684921823442,
0.009889773093163967,
-0.03059772402048111,
0.04399145394563675,
0.0775076374411583,
0.012703849002718925,
-0.07831075042486191,
0.01555237639695406,
0.03657278046011925,
-0.002369658788666129,
0.031508952379226685,
-0.0451408252120018,
0.03430642932653427,
-0.02460874803364277,
-0.023439232259988785,
-0.03268667683005333,
-0.086464524269104,
0.06380040943622589,
-0.06555840373039246,
-0.02377157285809517,
0.007221458945423365,
-0.01376278791576624,
0.06765427440404892,
0.06974110752344131,
-0.07678022980690002,
-0.049557022750377655,
-0.005051746033132076,
-0.03548034280538559,
-0.004009629599750042,
0.007822683081030846,
-0.022452330216765404,
0.052961550652980804,
0.05092528089880943,
-0.005907938349992037,
0.06982596218585968,
-0.0433991365134716,
0.047677572816610336,
0.03516267240047455,
-0.027892587706446648,
-0.05020217224955559,
0.03726913779973984,
-0.08732592314481735,
-0.0632636621594429,
0.025823581963777542,
0.061452608555555344,
-0.02499917522072792,
0.04138868674635887,
0.06664322316646576,
0.10388390719890594,
0.07452398538589478,
0.008151823654770851,
0.0014433561591431499,
0.01707250438630581,
0.0386483259499073,
-0.035475704818964005,
0.09392893314361572,
-0.025568677112460136,
0.015558800660073757,
-0.00698187155649066,
-4.2437424241370536e-8,
-0.1060512438416481,
0.059018876403570175,
0.01246180385351181,
0.034054841846227646,
0.04064752534031868,
-0.007623672019690275,
-0.029008906334638596,
0.05818619951605797,
-0.0191512368619442,
-0.005472642369568348,
-0.00967202428728342,
-0.053199008107185364,
0.014502379111945629,
-0.1059761643409729,
-0.02756558544933796,
0.1454283744096756,
-0.004025932401418686,
0.057501211762428284,
0.011845607310533524,
-0.007734702434390783,
-0.015837999060750008,
0.007052448112517595,
0.0009984195930883288,
-0.010150015354156494,
-0.014410357922315598,
-0.0872264876961708,
-0.06676200032234192,
0.051446810364723206,
0.045384082943201065,
0.061454955488443375,
-0.03184852376580238,
0.030695710331201553,
0.09013210237026215,
-0.01896197535097599,
-0.016546448692679405,
0.07404498010873795,
0.01391131617128849,
-0.01254009548574686,
-0.018442556262016296,
-0.05825464800000191,
0.02037919871509075,
0.0960187241435051,
-0.02150334045290947,
-0.012299034744501114,
-0.03225856274366379,
0.04062149301171303,
0.06585506349802017,
-0.09534395486116409,
0.009070780128240585,
0.03066587820649147,
0.07133711874485016,
-0.025639887899160385,
0.07863885909318924,
0.08194166421890259,
0.04874918609857559,
0.016094541177153587,
-0.00844974722713232,
-0.06403593719005585,
0.009652023203670979,
0.13279709219932556,
0.021434320136904716,
0.03810221701860428,
-0.0001022461146931164,
-0.043425142765045166
] | 0 | 0 | 0 | 0 |
|
LayerDAG: A Layerwise Autoregressive Diffusion Model for Directed Acyclic Graph Generation | https://openreview.net/forum?id=kam84eEmub | [
"Mufei Li",
"Viraj Shitole",
"Eli Chien",
"Changhai Man",
"Zhaodong Wang",
"Srinivas",
"Ying Zhang",
"Tushar Krishna",
"Pan Li"
] | Spotlight | Directed acyclic graphs (DAGs) serve as crucial data representations in domains such as hardware synthesis and compiler/program optimization for computing systems. DAG generative models facilitate the creation of synthetic DAGs, which can be used for benchmarking computing systems while preserving intellectual property. However, generating realistic DAGs is challenging due to their inherent directional and logical dependencies. This paper introduces LayerDAG, an autoregressive diffusion model, to address these challenges. LayerDAG decouples the strong node dependencies into manageable units that can be processed sequentially. By interpreting the partial order of nodes as a sequence of bipartite graphs, LayerDAG leverages autoregressive generation to model directional dependencies and employs diffusion models to capture logical dependencies within each bipartite graph. Comparative analyses demonstrate that LayerDAG outperforms existing DAG generative models in both expressiveness and generalization, particularly for generating large-scale DAGs with up to 400 nodes—a critical scenario for system benchmarking. Extensive experiments on both synthetic and real-world flow graphs from various computing platforms show that LayerDAG generates valid DAGs with superior statistical properties and benchmarking performance. The synthetic DAGs generated by LayerDAG enhance the training of ML-based surrogate models, resulting in improved accuracy in predicting performance metrics of real-world DAGs across diverse computing platforms. | directed acyclic graphs, graph generation, discrete diffusion, autoregressive model | null | 12,908 | 2411.02322 | [
-0.07021880149841309,
-0.05100439488887787,
0.014199699275195599,
0.05054610222578049,
0.03970032557845116,
-0.11667343229055405,
-0.08831540495157242,
-0.018572572618722916,
0.011157887056469917,
-0.044763579964637756,
-0.07086203992366791,
-0.07928800582885742,
-0.011867199093103409,
-0.01983572728931904,
-0.00126753153745085,
0.004044420551508665,
0.061637718230485916,
-0.018989553675055504,
-0.0148467430844903,
-0.11255845427513123,
0.013705039396882057,
0.019356239587068558,
-0.11516744643449783,
-0.022134847939014435,
0.07420115917921066,
0.019659994170069695,
-0.0003669812867883593,
-0.03670359030365944,
0.045794527977705,
-0.03287770599126816,
0.049824152141809464,
0.032399795949459076,
-0.09526833891868591,
0.07104586064815521,
-0.007689308375120163,
0.0882406160235405,
-0.024088887497782707,
-0.04735507816076279,
0.001337080611847341,
-0.03782697767019272,
0.00676583731546998,
0.0682331919670105,
0.011767568066716194,
0.05564482510089874,
0.04284447431564331,
-0.04905439540743828,
-0.1042955070734024,
0.040332842618227005,
-0.06300559639930725,
0.03475973382592201,
-0.06567111611366272,
-0.005753399804234505,
-0.07293256372213364,
0.026405658572912216,
0.033771879971027374,
0.016163956373929977,
0.10170452296733856,
0.0036985245533287525,
-0.03050384484231472,
0.03253905475139618,
-0.08413679152727127,
-0.08105310797691345,
-0.010176110081374645,
-0.06383628398180008,
-0.061826568096876144,
-0.0055361660197377205,
0.08671995252370834,
0.07483892142772675,
0.06908012181520462,
0.05343242362141609,
-0.012684540823101997,
0.039018452167510986,
-0.07196490466594696,
0.04575176164507866,
-0.010198375210165977,
0.070885568857193,
0.0684858113527298,
-0.05202902853488922,
0.002268259646371007,
-0.14555491507053375,
-0.012407967820763588,
0.09294408559799194,
0.008336960338056087,
0.0696583017706871,
0.01961357332766056,
0.0005830558366142213,
0.041288409382104874,
0.01840570941567421,
0.03571205958724022,
0.016617048531770706,
-0.08008579164743423,
0.035664938390254974,
0.04380771517753601,
0.0010367579525336623,
0.10349252074956894,
0.05999714881181717,
0.042372189462184906,
-0.04434267804026604,
-0.015092872083187103,
0.0583786740899086,
0.028189053758978844,
0.06683915108442307,
0.04988779127597809,
-0.08268193155527115,
0.03322354331612587,
-0.01633419096469879,
-0.0011921277036890388,
0.056501854211091995,
-0.08543340116739273,
-0.03758193925023079,
-0.003864279482513666,
0.08901790529489517,
0.05138421058654785,
-0.07071700692176819,
-0.004414470866322517,
-0.07436686009168625,
-0.03697647154331207,
-0.10997883230447769,
-0.0021730870939791203,
0.02152697555720806,
-0.03512179106473923,
0.04880852997303009,
-0.0484037883579731,
0.05282618850469589,
0.01458014827221632,
-0.04157597944140434,
0.007398282177746296,
4.8463934964433914e-33,
-0.004752010107040405,
0.011095063760876656,
0.036790236830711365,
0.0010260703274980187,
0.040562473237514496,
-0.0473623126745224,
0.07103173434734344,
0.02092844247817993,
-0.05820157751441002,
-0.026925653219223022,
-0.06734615564346313,
0.0315970703959465,
-0.09898477792739868,
0.10892164707183838,
0.09424135088920593,
-0.03868147358298302,
-0.004397986922413111,
0.02551465854048729,
-0.02148282527923584,
-0.056335579603910446,
0.00798777211457491,
-0.00403290381655097,
-0.021650133654475212,
0.03809133172035217,
0.03955760598182678,
0.03294522687792778,
-0.0036519314162433147,
0.0205350611358881,
-0.05965336412191391,
0.009305432438850403,
-0.0981065034866333,
-0.03975685313344002,
0.038586586713790894,
0.030037904158234596,
-0.032060518860816956,
-0.013110702857375145,
-0.05031895264983177,
-0.08780275285243988,
0.016048843041062355,
0.004146140068769455,
0.025221526622772217,
0.0497576966881752,
-0.024081680923700333,
0.013279152102768421,
-0.05324491485953331,
-0.027104778215289116,
-0.06763456761837006,
0.00540371285751462,
-0.02264348231256008,
-0.05560353398323059,
0.07264771312475204,
0.01645977608859539,
0.07872244715690613,
0.037971608340740204,
-0.01691831834614277,
-0.021449627354741096,
0.028229964897036552,
0.001960017019882798,
0.10296144336462021,
0.11385682970285416,
0.021443411707878113,
0.03625041991472244,
-0.017109151929616928,
0.07269175350666046,
0.020939595997333527,
0.04048331454396248,
0.001152946031652391,
0.05009685456752777,
0.05989938601851463,
0.010499559342861176,
-0.045193761587142944,
-0.012673890218138695,
-0.02070377767086029,
-0.014366371557116508,
0.0519678071141243,
-0.014878642745316029,
-0.030319008976221085,
-0.0902448445558548,
-0.041556842625141144,
0.009241190738976002,
-0.10701882839202881,
0.08839524537324905,
0.024244459345936775,
-0.016172949224710464,
0.039832953363657,
-0.0997815653681755,
0.009252810850739479,
0.04795236140489578,
-0.0672491043806076,
-0.003890046151354909,
-0.008109226822853088,
-0.029393548145890236,
0.04805904999375343,
-0.005403321702033281,
0.031545329838991165,
-6.438056193810955e-33,
-0.057865310460329056,
0.06704799830913544,
0.04778279364109039,
0.08537135273218155,
-0.016656165942549706,
-0.12012510001659393,
-0.06146794930100441,
-0.02084333822131157,
0.0009038406424224377,
0.010485764592885971,
-0.06162344291806221,
-0.05151617154479027,
0.04871247336268425,
-0.03702813386917114,
0.043927643448114395,
-0.023259282112121582,
-0.042117342352867126,
-0.09485580772161484,
0.009277934208512306,
0.07858641445636749,
-0.03821030631661415,
0.08427629619836807,
-0.025789989158511162,
-0.04302230849862099,
-0.008821276016533375,
-0.00014447675494011492,
-0.04268795996904373,
-0.024059781804680824,
0.028411220759153366,
0.01242232695221901,
0.01811154931783676,
0.06280473619699478,
-0.05281675606966019,
-0.022789085283875465,
0.030942095443606377,
0.029412098228931427,
0.027716441079974174,
-0.011185304261744022,
-0.05545211583375931,
0.04339700564742088,
0.007161297369748354,
-0.05182486027479172,
-0.06397391855716705,
0.025961441919207573,
0.047489069402217865,
0.09872184693813324,
-0.05014258995652199,
0.014524593949317932,
-0.021917374804615974,
-0.06466735154390335,
-0.022246478125452995,
0.11725213378667831,
0.007783064153045416,
0.007382150739431381,
0.011610555462539196,
-0.036125298589468,
-0.04584372416138649,
0.032276324927806854,
0.06333182007074356,
-0.07615697383880615,
-0.03747780621051788,
-0.09972403943538666,
0.06523289531469345,
-0.122869111597538,
-0.056723687797784805,
-0.04148019850254059,
-0.039797425270080566,
-0.03852144256234169,
-0.05370260775089264,
0.057621777057647705,
0.02167837880551815,
0.06846408545970917,
-0.030341316014528275,
-0.0165571216493845,
-0.026417991146445274,
-0.029396628960967064,
0.022790750488638878,
0.0684611052274704,
0.03712612763047218,
0.029212607070803642,
-0.07299064844846725,
0.0912427082657814,
0.019168447703123093,
0.03162867948412895,
-0.06967350840568542,
0.013218511827290058,
-0.00898407120257616,
0.009489004500210285,
-0.029093684628605843,
0.08088338375091553,
-0.02879934385418892,
-0.030016085132956505,
-0.036382827907800674,
0.05304544046521187,
-0.050274789333343506,
-4.44434036239727e-8,
-0.009714479558169842,
0.059725355356931686,
0.09194213151931763,
-0.003105028299614787,
0.033057767897844315,
-0.015029278583824635,
0.03308914229273796,
0.09950302541255951,
0.02290506660938263,
0.06836028397083282,
0.03795706108212471,
-0.028900649398565292,
0.006346364039927721,
-0.07667392492294312,
-0.027438674122095108,
0.06096261739730835,
0.026400910690426826,
-0.04952078312635422,
0.012753923423588276,
-0.07096943259239197,
-0.02215484343469143,
0.03494372218847275,
-0.04351579397916794,
0.10192358493804932,
0.03863340988755226,
-0.08202528208494186,
0.04280482977628708,
-0.015117978677153587,
0.06962130963802338,
0.08925352245569229,
0.017659906297922134,
0.009915389120578766,
0.008720869198441505,
0.004541090689599514,
0.0639301985502243,
0.028441738337278366,
0.03847875818610191,
0.020213432610034943,
0.02846529707312584,
-0.05057220533490181,
-0.009580524638295174,
0.03966517373919487,
-0.028793010860681534,
0.0333709754049778,
0.00312428898178041,
-0.04196399822831154,
-0.017242901027202606,
-0.047588661313056946,
-0.013161243870854378,
-0.050409626215696335,
-0.041358910501003265,
-0.0734647735953331,
-0.06610137969255447,
-0.003970193210989237,
0.015977241098880768,
0.04955014958977699,
-0.04759332165122032,
-0.1258930116891861,
0.0433528758585453,
0.005587208084762096,
-0.07479887455701828,
0.018881361931562424,
0.041120365262031555,
-0.019279535859823227
] | https://github.com/graph-com/layerdag | 16 | 0 | 0 | 0 |
DailyDilemmas: Revealing Value Preferences of LLMs with Quandaries of Daily Life | https://openreview.net/forum?id=PGhiPGBf47 | [
"Yu Ying Chiu",
"Liwei Jiang",
"Yejin Choi"
] | Spotlight | As users increasingly seek guidance from LLMs for decision-making in daily life, many of these decisions are not clear-cut and depend significantly on the personal values and ethical standards of people. We present DailyDilemmas, a dataset of 1,360 moral dilemmas encountered in everyday life. Each dilemma presents two possible actions, along with affected parties and relevant human values for each action. Based on these dilemmas, we gather a repository of human values covering diverse everyday topics, such as interpersonal relationships, workplace, and environmental issues. With DailyDilemmas, we evaluate LLMs on these dilemmas to determine what action they will choose and the values represented by these action choices. Then, we analyze values through the lens of five theoretical frameworks inspired by sociology, psychology, and philosophy, including the World Values Survey, Moral Foundations Theory, Maslow's Hierarchy of Needs, Aristotle's Virtues, and Plutchik's Wheel of Emotions. For instance, we find LLMs are most aligned with self-expression over survival in World Values Survey and care over loyalty in Moral Foundations Theory. Interestingly, we find substantial preference differences in models for some core values. For example, for truthfulness, Mixtral-8x7B neglects it by 9.7% while GPT-4-turbo selects it by 9.4%. We also study the recent guidance released by OpenAI (ModelSpec), and Anthropic (Constitutional AI) to understand how their designated principles reflect their models' actual value prioritization when facing nuanced moral reasoning in daily-life settings. Finally, we find that end users cannot effectively steer such prioritization using system prompts. | language model, moral dilemma, model alignment, machine ethics, value alignment | null | 12,857 | 2410.02683 | [
-0.05334113538265228,
0.044221870601177216,
-0.07353256642818451,
-0.0390559583902359,
0.0346234105527401,
-0.009983944706618786,
0.08726073801517487,
0.006621744018048048,
0.07154709845781326,
-0.0274835005402565,
0.004474939312785864,
-0.03216593712568283,
-0.01878536492586136,
0.08205427974462509,
0.08328847587108612,
0.029055776074528694,
-0.026560073718428612,
-0.025240013375878334,
-0.09963276982307434,
0.06348731368780136,
-0.07499539107084274,
0.013967187143862247,
0.0361541248857975,
0.031945303082466125,
-0.08236882835626602,
0.01869177259504795,
0.09798930585384369,
-0.007424202747642994,
-0.003346315585076809,
-0.0032432943116873503,
0.019803345203399658,
0.08466216176748276,
0.047567445784807205,
-0.0008564611780457199,
-0.019551346078515053,
0.03021758422255516,
0.01502932794392109,
-0.015852633863687515,
-0.047753117978572845,
0.03492147848010063,
-0.01160376239567995,
-0.06883813440799713,
0.036284059286117554,
0.011362064629793167,
-0.05264697223901749,
-0.029119055718183517,
0.029224637895822525,
0.011168163269758224,
-0.07071307301521301,
-0.015208455733954906,
-0.03449377417564392,
-0.007683242205530405,
-0.08003149926662445,
0.03124103508889675,
0.07636986672878265,
0.01353420503437519,
0.053963057696819305,
0.007176676299422979,
0.06388570368289948,
-0.02172374166548252,
0.026345599442720413,
-0.04630628228187561,
-0.01249279361218214,
0.04064466059207916,
0.039882902055978775,
-0.002157695824280381,
-0.057904377579689026,
0.06189141422510147,
-0.0889110416173935,
0.013996989466249943,
-0.05921635031700134,
-0.02863415516912937,
-0.06524699926376343,
0.02522987499833107,
-0.036793291568756104,
-0.011839413084089756,
-0.0021813109051436186,
-0.05490203946828842,
0.031141892075538635,
0.003426380455493927,
-0.01301902811974287,
0.02179992012679577,
-0.05726930871605873,
0.041573334485292435,
-0.006026166491210461,
-0.03087947890162468,
-0.021307818591594696,
0.07099105417728424,
0.07918135076761246,
0.06086844205856323,
-0.06018735468387604,
-0.022898457944393158,
-0.018290583044290543,
0.010440579615533352,
-0.008205081336200237,
0.03830889239907265,
-0.08626006543636322,
0.0004826165677513927,
-0.04107650741934776,
0.039367929100990295,
0.012397348880767822,
0.03465371951460838,
-0.04428517818450928,
-0.0034386939369142056,
0.05695921927690506,
-0.008776510134339333,
-0.10880088061094284,
0.04282475635409355,
0.004077181685715914,
0.037965450435876846,
-0.11492261290550232,
0.026886845007538795,
0.07057332992553711,
-0.0398922823369503,
0.1579483598470688,
-0.039217934012413025,
0.0461743138730526,
-0.012963763438165188,
0.04492994397878647,
-0.007426823489367962,
0.041926153004169464,
-0.023903504014015198,
0.07091158628463745,
0.009632544592022896,
0.038746368139982224,
0.0017045271815732121,
-0.047676388174295425,
1.8089697258483568e-33,
0.02318834327161312,
-0.04197642579674721,
0.03670881316065788,
0.02610916644334793,
-0.030215032398700714,
-0.017164338380098343,
-0.05508802458643913,
0.04293154180049896,
0.015212749131023884,
0.052294619381427765,
0.014204055070877075,
0.11650555580854416,
-0.02653166651725769,
0.06829167157411575,
0.013358408585190773,
0.002096267184242606,
-0.03869418054819107,
-0.07178202271461487,
-0.026933787390589714,
0.03346814587712288,
0.05232846736907959,
0.026623833924531937,
0.051453959196805954,
-0.019334876909852028,
0.0037254297640174627,
-0.03636586666107178,
0.08004816621541977,
-0.04944460093975067,
0.06662403792142868,
0.0003606148820836097,
0.0275808647274971,
0.036244649440050125,
0.08062005788087845,
-0.036897629499435425,
-0.014757636934518814,
0.07421620190143585,
-0.10233547538518906,
-0.026992451399564743,
0.018784385174512863,
-0.09562420845031738,
-0.07877925783395767,
0.012637416832149029,
0.016051506623625755,
0.019256193190813065,
0.010054855607450008,
-0.013027823530137539,
0.11518406122922897,
-0.00896602775901556,
-0.021790387108922005,
0.02707582339644432,
0.038725972175598145,
0.006436204072088003,
0.038603443652391434,
-0.06780709326267242,
-0.1486128866672516,
-0.019227417185902596,
-0.03599463403224945,
-0.03546423092484474,
-0.11415483802556992,
-0.07796849310398102,
-0.04594928398728371,
-0.022279027849435806,
0.01802782528102398,
-0.03877180442214012,
0.01251931395381689,
0.056035131216049194,
-0.05052433907985687,
-0.05515472590923309,
-0.036230992525815964,
-0.017298538237810135,
0.05762273818254471,
0.045489996671676636,
-0.021698491647839546,
-0.07581690698862076,
-0.06571581959724426,
0.03388221189379692,
0.0473923534154892,
-0.03612564504146576,
-0.023322638124227524,
0.001999620348215103,
-0.00424133101478219,
0.050894416868686676,
0.022407393902540207,
-0.07610300928354263,
0.10864472389221191,
-0.025789903476834297,
0.06753061711788177,
-0.08450373262166977,
0.031628746539354324,
0.04887780174612999,
-0.06239261478185654,
0.012316890060901642,
-0.003110784338787198,
0.02804519422352314,
-0.01323925331234932,
-4.023025457907628e-33,
-0.021946623921394348,
-0.10295557975769043,
-0.05342091992497444,
0.07122883945703506,
0.06758938729763031,
-0.006852149032056332,
-0.0567656047642231,
0.05324707552790642,
0.01419195719063282,
-0.0010136135388165712,
-0.07977930456399918,
-0.005532113369554281,
0.05877089872956276,
0.036073025315999985,
0.0057772439904510975,
-0.0702410414814949,
-0.02017849311232567,
-0.07295123487710953,
0.0029431525617837906,
-0.006198982708156109,
-0.04148527607321739,
0.13407984375953674,
-0.15153847634792328,
0.056158650666475296,
-0.00801760982722044,
0.09157730638980865,
0.03849368542432785,
-0.04757067561149597,
0.042491745203733444,
-0.0070913685485720634,
-0.02856067195534706,
-0.015466987155377865,
-0.0338907390832901,
-0.049256470054388046,
0.00995665043592453,
-0.014660272747278214,
0.010578125715255737,
-0.041612621396780014,
-0.02738492377102375,
0.03265140578150749,
-0.07940162718296051,
0.02497064135968685,
-0.03352316841483116,
-0.004734022077172995,
-0.08701801300048828,
-0.026000553742051125,
-0.019587231799960136,
-0.0233957227319479,
0.02622259594500065,
-0.04805908352136612,
0.05671986564993858,
-0.020908690989017487,
-0.003491171868517995,
0.08591832965612411,
0.05491353198885918,
-0.031166018918156624,
-0.00826442800462246,
-0.028474336490035057,
0.07013820856809616,
0.022678682580590248,
0.004840155132114887,
0.06865867972373962,
-0.030354749411344528,
0.008837435394525528,
-0.0758766382932663,
-0.007771727629005909,
-0.020106187090277672,
0.03796214982867241,
-0.15655231475830078,
0.016641665250062943,
0.004736034665256739,
-0.07815173268318176,
-0.011844581924378872,
-0.12513159215450287,
0.007091901730746031,
-0.014271787367761135,
-0.03703722730278969,
0.0028539623599499464,
-0.0556977204978466,
0.05504193529486656,
-0.009573854506015778,
-0.03638675808906555,
0.045738767832517624,
0.05679183453321457,
-0.08403923362493515,
-0.05753489211201668,
0.03347964584827423,
0.14576849341392517,
-0.005115351174026728,
0.026457922533154488,
-0.009236249141395092,
0.026042602956295013,
-0.062398385256528854,
-0.03936728835105896,
-0.04790273308753967,
-4.939845155149669e-8,
0.01006345171481371,
-0.07665946334600449,
-0.023249834775924683,
0.03425801172852516,
-0.04533678665757179,
-0.041569773107767105,
0.08539675921201706,
-0.040868163108825684,
-0.03559350594878197,
0.1320425122976303,
0.1031489148736,
0.05278247594833374,
-0.0705735981464386,
0.060457419604063034,
-0.01716405712068081,
0.018044009804725647,
0.057593658566474915,
-0.07198353856801987,
0.033908866345882416,
0.06514004617929459,
0.12056179344654083,
-0.00822747964411974,
-0.0024293179158121347,
-0.04391169175505638,
0.018613679334521294,
0.015218328684568405,
-0.046823497861623764,
0.030614815652370453,
-0.0005174991674721241,
0.09812804311513901,
0.034316446632146835,
-0.004944856744259596,
-0.03596360608935356,
-0.004224743694067001,
-0.01777092181146145,
-0.05255863443017006,
-0.06437814235687256,
-0.021104346960783005,
0.04261583089828491,
0.06221015751361847,
0.031362779438495636,
0.006074795499444008,
-0.016259251162409782,
0.0397537499666214,
-0.0015799140091985464,
-0.021035870537161827,
-0.04988166689872742,
0.03289115056395531,
-0.08961377292871475,
0.03167534992098808,
-0.062324728816747665,
-0.016443530097603798,
0.04979697987437248,
0.0008489257306791842,
0.03864148259162903,
-0.039551325142383575,
0.050915829837322235,
0.05720945820212364,
-0.005798978731036186,
-0.037212904542684555,
0.08695055544376373,
-0.08731025457382202,
-0.08066387474536896,
-0.006298060994595289
] | 0 | 0 | 1 | 0 |
|
Learning-Augmented Frequent Directions | https://openreview.net/forum?id=WcZLG8XxhD | [
"Anders Aamand",
"Justin Y. Chen",
"Siddharth Gollapudi",
"Sandeep Silwal",
"Hao WU"
] | Spotlight | An influential paper of Hsu et al. (ICLR'19) introduced the study of learning-augmented streaming algorithms in the context of frequency estimation. A fundamental problem in the streaming literature, the goal of frequency estimation is to approximate the number of occurrences of items appearing in a long stream of data using only a small amount of memory. Hsu et al. develop a natural framework to combine the worst-case guarantees of popular solutions such as CountMin and CountSketch with learned predictions of high frequency elements. They demonstrate that learning the underlying structure of data can be used to yield better streaming algorithms, both in theory and practice.
We simplify and generalize past work on learning-augmented frequency estimation. Our first contribution is a learning-augmented variant of the Misra-Gries algorithm which improves upon the error of learned CountMin and learned CountSketch and achieves the state-of-the-art performance of randomized algorithms (Aamand et al., NeurIPS'23) with a simpler, deterministic algorithm. Our second contribution is to adapt learning-augmentation to a high-dimensional generalization of frequency estimation corresponding to finding important directions (top singular vectors) of a matrix given its rows one-by-one in a stream. We analyze a learning-augmented variant of the Frequent Directions algorithm, extending the theoretical and empirical understanding of learned predictions to matrix streaming. | learning-augmented algorithms, algorithms with predictions, data streams, streaming algorithms, frequency estimation, heavy hitters, frequent directions, low-rank approximation | We simplify and generalize existing work on frequency estimation with learned predictions. | 12,778 | 2503.00937 | [
-0.05444219335913658,
-0.09812323749065399,
-0.06184941530227661,
0.03695618361234665,
0.03263600170612335,
0.05354524031281471,
0.03673214465379715,
-0.07180455327033997,
-0.025380009785294533,
0.004410764202475548,
-0.011275370605289936,
0.019020596519112587,
0.05070076882839203,
0.00019174609042238444,
-0.036459632217884064,
-0.025844072923064232,
0.0339525006711483,
0.04814312234520912,
0.011094963178038597,
-0.09050809592008591,
-0.007414801511913538,
0.018217291682958603,
-0.005755375139415264,
-0.0344143882393837,
0.007213082630187273,
0.04339470714330673,
0.004630723502486944,
-0.0038079959340393543,
0.01275106891989708,
-0.007706907112151384,
0.005212054587900639,
0.060335300862789154,
0.00969456136226654,
-0.002016040962189436,
-0.09623165428638458,
-0.045399926602840424,
-0.02735934779047966,
0.10298564285039902,
-0.020066000521183014,
0.05884820595383644,
0.09850678592920303,
0.0731712058186531,
0.02667677029967308,
0.008555612526834011,
0.026409683749079704,
0.06256366521120071,
0.0011148054618388414,
-0.06543242931365967,
-0.05202232301235199,
0.06944958865642548,
-0.001219102181494236,
0.029258739203214645,
-0.027958042919635773,
0.11168038845062256,
0.0004399385361466557,
-0.04805444926023483,
0.007330779917538166,
-0.06462152302265167,
0.018038716167211533,
0.0626874789595604,
-0.007205438334494829,
-0.16969306766986847,
-0.040402453392744064,
-0.1072978749871254,
-0.08703437447547913,
-0.05049683153629303,
0.028307365253567696,
0.11358394473791122,
-0.03944757208228111,
0.09846099466085434,
-0.02468251995742321,
0.08984223753213882,
-0.04408549889922142,
0.10167046636343002,
0.04893990233540535,
0.010534759610891342,
-0.010990431532263756,
-0.04189102351665497,
0.009925159625709057,
0.04057395085692406,
-0.039444342255592346,
0.017010673880577087,
0.0343346931040287,
-0.01498191338032484,
0.14102678000926971,
-0.020579829812049866,
0.02953464537858963,
0.03243207186460495,
0.023491770029067993,
-0.008529840037226677,
-0.04683161526918411,
0.015420107170939445,
0.0283681508153677,
-0.02672285959124565,
-0.011043719947338104,
0.07664629071950912,
0.04426685348153114,
-0.04333006963133812,
0.012858767062425613,
0.07836467772722244,
0.03753865882754326,
-0.0022933497093617916,
-0.024815451353788376,
0.04604946821928024,
-0.04511319100856781,
-0.02492521144449711,
0.027770135551691055,
0.12086115032434464,
0.05054929852485657,
-0.07007616013288498,
-0.0013339690631255507,
0.015506498515605927,
0.04681629687547684,
-0.02251715026795864,
0.03723747283220291,
0.028422383591532707,
0.03020116686820984,
-0.04560362175107002,
0.0378088653087616,
0.0010477685136720538,
-0.007742875721305609,
-0.01566455140709877,
0.07736786454916,
0.050955019891262054,
0.02602577954530716,
0.06761712580919266,
-0.017911409959197044,
2.0694236418180664e-33,
0.02299472689628601,
0.03911568596959114,
0.06966327875852585,
-0.02933468110859394,
0.052285730838775635,
-0.010427357628941536,
0.01860419102013111,
0.06778723746538162,
0.10530035942792892,
0.0520867258310318,
-0.030039722099900246,
0.05738109350204468,
-0.014521026983857155,
0.08779653906822205,
-0.012471686117351055,
-0.04997611790895462,
0.017645694315433502,
0.014576464891433716,
-0.014279746450483799,
-0.10014484822750092,
0.05000942200422287,
-0.12732766568660736,
0.032602932304143906,
-0.0813683420419693,
0.0028891877736896276,
0.0002913977368734777,
0.03738764673471451,
-0.049567811191082,
0.061445631086826324,
0.008115134201943874,
-0.02722020074725151,
0.017997920513153076,
0.044250939041376114,
0.017856566235423088,
0.0451485700905323,
-0.09636186808347702,
0.024093415588140488,
0.011476488783955574,
-0.02184881642460823,
-0.04714391008019447,
0.022860392928123474,
0.016337715089321136,
-0.008232297375798225,
-0.11273019760847092,
-0.10971839725971222,
-0.0026705372147262096,
0.024977387860417366,
0.0017549669137224555,
-0.015432578511536121,
-0.01568610593676567,
0.078013576567173,
-0.07508476078510284,
-0.06479498744010925,
0.05437900871038437,
0.02857343852519989,
0.03694707900285721,
-0.033475808799266815,
0.0283454991877079,
0.05974366515874863,
0.06828204542398453,
-0.010008655488491058,
0.05190230906009674,
0.07463749498128891,
0.009187906049191952,
0.012579016387462616,
0.057468216866254807,
0.03318469598889351,
0.055333368480205536,
0.05494587495923042,
-0.01243614125996828,
0.026570575311779976,
0.016518499702215195,
-0.010480077005922794,
-0.03507279232144356,
0.0054758405312895775,
0.018114592880010605,
0.04017474874854088,
-0.0694715604186058,
-0.03746664524078369,
0.025730878114700317,
0.04580751806497574,
-0.009082138538360596,
0.002105858875438571,
-0.06271477043628693,
-0.0784156396985054,
-0.020910238847136497,
0.04207178205251694,
-0.1070418655872345,
-0.04423294961452484,
-0.09850817173719406,
-0.045912183821201324,
0.025940053164958954,
0.0022115062456578016,
-0.027605537325143814,
-0.05619807168841362,
-1.6837810468537217e-33,
-0.010159130208194256,
0.016863372176885605,
0.02744239754974842,
0.012232636101543903,
0.04488105699419975,
-0.059879809617996216,
-0.04996282607316971,
-0.00866701453924179,
0.043817006051540375,
-0.10713794827461243,
-0.029760021716356277,
-0.06309707462787628,
-0.022120831534266472,
0.03984389826655388,
0.05419327691197395,
0.009554555639624596,
-0.08233103156089783,
0.037585485726594925,
0.006277780048549175,
-0.03240644559264183,
-0.08405088633298874,
-0.04565253481268883,
-0.0034381269942969084,
0.015895260497927666,
-0.041609976440668106,
-0.029403623193502426,
0.022376373410224915,
0.048253364861011505,
0.012024658732116222,
-0.008446280844509602,
-0.036636833101511,
-0.03726140037178993,
-0.0348166897892952,
-0.031687650829553604,
-0.02516874298453331,
0.07663290947675705,
0.07492052763700485,
0.0025834457483142614,
-0.09263518452644348,
0.05244196951389313,
0.05814841762185097,
-0.01719657890498638,
-0.015905676409602165,
-0.025963932275772095,
0.02072419784963131,
-0.015015196986496449,
-0.07193129509687424,
0.07520569860935211,
0.020973222330212593,
-0.024361293762922287,
-0.01618080399930477,
0.007173840422183275,
-0.04186336696147919,
0.0269716028124094,
-0.004632354713976383,
-0.04766194149851799,
0.015021911822259426,
0.026729051023721695,
0.023837117478251457,
0.009842434898018837,
-0.06816864758729935,
-0.0431693010032177,
-0.09506314247846603,
-0.05308009311556816,
0.057430725544691086,
0.019535211846232414,
-0.0000829844138934277,
-0.009490273892879486,
-0.059636328369379044,
0.04244157299399376,
-0.05368858575820923,
-0.037259768694639206,
-0.042848654091358185,
0.002566791605204344,
-0.022468142211437225,
0.00376196950674057,
0.04163273796439171,
-0.04821525886654854,
-0.09730789810419083,
0.00887138582766056,
-0.031157078221440315,
-0.03095940500497818,
-0.05295338109135628,
0.052178990095853806,
-0.023192012682557106,
0.07481426745653152,
0.08324019610881805,
-0.05084212124347687,
0.030207699164748192,
-0.06374482810497284,
-0.06726136058568954,
0.02985342964529991,
-0.08272265642881393,
0.05766717344522476,
0.006224868819117546,
-4.301208988977123e-8,
-0.1486026644706726,
-0.032227616757154465,
-0.01865159533917904,
-0.06002100184559822,
0.1281043291091919,
0.013531733304262161,
-0.012623107060790062,
0.07842136174440384,
-0.005917930044233799,
-0.1240614652633667,
0.07160705327987671,
-0.007212813012301922,
0.01417495310306549,
0.027072003111243248,
-0.033491771668195724,
0.05980768799781799,
0.022857965901494026,
-0.0033237633761018515,
0.03997877985239029,
0.014856972731649876,
-0.0054418290965259075,
0.06893203407526016,
0.10238644480705261,
-0.0726892352104187,
-0.014692380093038082,
-0.05795884504914284,
0.06894320994615555,
0.08957961946725845,
0.06360837072134018,
0.022894438356161118,
-0.035435114055871964,
-0.023167908191680908,
0.08359159529209137,
-0.023498445749282837,
0.06061582639813423,
0.08012878894805908,
-0.02362215146422386,
-0.05062461271882057,
-0.07770611345767975,
-0.08372452110052109,
-0.025448687374591827,
0.08142693340778351,
-0.01407045777887106,
-0.005421476438641548,
0.009780352003872395,
0.03225812315940857,
0.04530690982937813,
-0.03252989798784256,
0.001437563099898398,
-0.05645100772380829,
-0.008607223629951477,
-0.024577172473073006,
0.10712700337171555,
0.016781436279416084,
0.05690058693289757,
0.011227975599467754,
-0.04977552592754364,
-0.11183089762926102,
0.028642840683460236,
0.04102927818894386,
0.0019305255264043808,
0.05453481152653694,
-0.07212516665458679,
0.079892098903656
] | 0 | 0 | 0 | 0 |
|
No Need to Talk: Asynchronous Mixture of Language Models | https://openreview.net/forum?id=pHOH8FVrTp | [
"Anastasiia Filippova",
"Angelos Katharopoulos",
"David Grangier",
"Ronan Collobert"
] | Spotlight | We introduce SMALLTALK LM, an innovative method for training a mixture of language models in an almost asynchronous manner. Each
model of the mixture specializes in distinct parts of the data distribution, without the need of high-bandwidth communication between the nodes training each model. At inference, a lightweight router directs a given sequence to a single expert, according to a short prefix. This inference scheme naturally uses a fraction of the parameters from the overall mixture model. Unlike prior works on asynchronous LLM training, our routing method does not rely on full corpus clustering or access to metadata, making it more suitable for real-world applications. Our experiments on language modeling demonstrate that SMALLTALK LM achieves significantly lower perplexity than dense model baselines for the same total training FLOPs and an almost identical inference cost. Finally, in our downstream evaluations we outperform the dense baseline on 75% of the tasks. | language models, distributed learning, divide and conquer, efficient inference | null | 12,727 | 2410.03529 | [
-0.006305719260126352,
-0.11090155690908432,
0.015393409878015518,
0.07533189654350281,
0.09846437722444534,
0.012446928769350052,
-0.06701210141181946,
-0.01955467090010643,
0.017095837742090225,
-0.09853237122297287,
-0.02949901856482029,
-0.03835073113441467,
-0.027839845046401024,
-0.00884360820055008,
-0.024509310722351074,
0.011287904344499111,
0.10956618934869766,
0.050526123493909836,
-0.06734445691108704,
-0.1000000610947609,
-0.02167847752571106,
0.06378207355737686,
0.07048697769641876,
0.018618233501911163,
0.05353540927171707,
0.024728339165449142,
-0.07394073158502579,
-0.030059954151511192,
0.05199296772480011,
-0.009032437577843666,
0.0632404088973999,
0.10076523572206497,
-0.011950268410146236,
0.049534767866134644,
-0.026162954047322273,
0.04188955947756767,
-0.03328642621636391,
0.026101943105459213,
0.011013561859726906,
-0.017921438440680504,
-0.023616936057806015,
-0.009273563511669636,
0.007799305487424135,
0.03949223831295967,
0.025292472913861275,
-0.05970638245344162,
-0.016092736274003983,
-0.004852248355746269,
-0.010696135461330414,
0.05597468838095665,
-0.06669563055038452,
-0.03495444357395172,
0.00623894389718771,
0.09320122003555298,
-0.051507048308849335,
0.030106380581855774,
0.030711045488715172,
0.07044380158185959,
-0.031553104519844055,
0.009951918385922909,
-0.07420645654201508,
-0.14357849955558777,
-0.041717905551195145,
0.0033355725463479757,
-0.03329847753047943,
-0.00843796320259571,
0.05206606909632683,
0.06655742228031158,
0.00011755045125028118,
0.004993097856640816,
0.01641451008617878,
0.06395532190799713,
-0.015170837752521038,
0.12439360469579697,
-0.041857749223709106,
-0.0003860292781610042,
0.0332784578204155,
-0.05085848271846771,
0.032415032386779785,
-0.03930473327636719,
-0.026851605623960495,
-0.010364696383476257,
0.025042656809091568,
-0.04027935862541199,
0.11280475556850433,
-0.022865192964673042,
0.015856590121984482,
0.02445228397846222,
-0.03413588926196098,
0.005507167428731918,
-0.07681521028280258,
-0.05505479872226715,
0.0639176294207573,
-0.023854684084653854,
-0.03049592301249504,
0.06627467274665833,
0.04045150801539421,
-0.06997506320476532,
0.010519548319280148,
0.115340456366539,
-0.03539817035198212,
0.07247915863990784,
0.021811766549944878,
-0.07554198801517487,
-0.0025425793137401342,
-0.007983813062310219,
0.023144036531448364,
0.028635229915380478,
0.03558503836393356,
-0.010458982549607754,
0.013268394395709038,
0.09266095608472824,
0.002347454894334078,
-0.014037290588021278,
0.019218290224671364,
-0.013463840819895267,
0.04060455039143562,
-0.07245608419179916,
0.008585838600993156,
0.0736413523554802,
-0.060410648584365845,
0.0322139710187912,
0.02858296036720276,
0.0015395678346976638,
0.04596502333879471,
-0.03453224152326584,
-0.028193438425660133,
3.954685889770667e-33,
0.13892881572246552,
0.025217227637767792,
0.01927764154970646,
0.0104259317740798,
0.03644188120961189,
-0.011971701867878437,
0.08175846189260483,
0.030830103904008865,
-0.04713329300284386,
-0.03690901771187782,
0.0011344460071995854,
-0.011540429666638374,
-0.01872638240456581,
-0.011224518530070782,
0.013287125155329704,
-0.025330016389489174,
-0.05826304107904434,
0.07638872414827347,
-0.041077714413404465,
-0.006987663451582193,
0.051888249814510345,
0.008388938382267952,
-0.003061441471800208,
-0.004162423778325319,
0.02703404240310192,
-0.012063262984156609,
0.09418316930532455,
-0.0803276002407074,
0.022122526541352272,
0.020904436707496643,
-0.09652641415596008,
-0.003988107666373253,
0.026441609486937523,
0.02241147868335247,
0.03598373010754585,
-0.018803298473358154,
-0.04045752063393593,
-0.0697348415851593,
0.02760913595557213,
-0.07462877780199051,
-0.026157699525356293,
0.010896822437644005,
0.0341145358979702,
-0.07212768495082855,
-0.1380310207605362,
-0.09077317267656326,
-0.06091177091002464,
0.020210085436701775,
0.0037792399525642395,
-0.028040075674653053,
0.06036979705095291,
-0.021956346929073334,
-0.013315126299858093,
-0.0034252554178237915,
0.029204167425632477,
-0.017968272790312767,
-0.011177384294569492,
0.03884704038500786,
0.0455993115901947,
0.11051011085510254,
0.02387065440416336,
-0.006569236516952515,
-0.02133859135210514,
0.07986902445554733,
0.0757741704583168,
-0.003893798217177391,
-0.05426400527358055,
0.1034674271941185,
0.07123898714780807,
-0.06586389988660812,
-0.0267514456063509,
-0.0277505274862051,
0.005263498518615961,
-0.003039285773411393,
0.07487530261278152,
0.05709806829690933,
-0.007333578076213598,
-0.11172139644622803,
-0.0478944331407547,
0.02026064693927765,
0.018076401203870773,
-0.026139771565794945,
-0.0058653089217841625,
-0.004657309968024492,
-0.01858835108578205,
0.03299582749605179,
0.02791726402938366,
-0.05330189689993858,
0.011560233309864998,
0.0328097939491272,
-0.07260561734437943,
-0.04211709648370743,
-0.007107778452336788,
-0.00025035208091139793,
0.005643632262945175,
-2.2493872351541507e-33,
-0.051423829048871994,
0.030377887189388275,
-0.027600519359111786,
0.12186522036790848,
-0.026955557987093925,
-0.05807291716337204,
0.0552733913064003,
0.03450392931699753,
-0.06561063975095749,
0.010854477994143963,
-0.047171447426080704,
-0.04370913654565811,
0.00954689085483551,
-0.04121657460927963,
0.08553282171487808,
-0.042193442583084106,
0.030852697789669037,
0.014982790686190128,
0.09469295293092728,
0.12070262432098389,
-0.04265259951353073,
0.016756240278482437,
-0.1405511349439621,
0.0001408539537806064,
0.005008287262171507,
0.004640053492039442,
-0.029278788715600967,
0.05958053469657898,
-0.0562177412211895,
-0.005150376353412867,
0.009109042584896088,
-0.02779644913971424,
-0.03661530837416649,
-0.10776925832033157,
-0.08208198845386505,
0.05757562443614006,
0.007957855239510536,
0.0024124342016875744,
-0.000753298751078546,
0.027042828500270844,
0.08341023325920105,
-0.05378741770982742,
-0.06365212798118591,
-0.012203888967633247,
-0.018889380618929863,
-0.08133815228939056,
-0.1573263704776764,
-0.005150505341589451,
-0.04815142974257469,
-0.04566684737801552,
-0.025335805490612984,
0.043561507016420364,
-0.010590186342597008,
-0.01283586211502552,
-0.03280484676361084,
-0.07053320109844208,
-0.06911704689264297,
-0.03950692340731621,
0.019122561439871788,
-0.019253768026828766,
-0.029577184468507767,
-0.04819849133491516,
0.03964556008577347,
-0.0557556189596653,
0.06216088682413101,
-0.09946796298027039,
-0.011550438590347767,
-0.03755722567439079,
-0.04493200033903122,
0.018940569832921028,
0.08428497612476349,
0.01202273927628994,
0.020361831411719322,
0.08962611109018326,
0.06023433804512024,
-0.007741718087345362,
0.013253685086965561,
-0.059751708060503006,
-0.040451180189847946,
0.0020378264598548412,
0.023114528506994247,
0.053076401352882385,
-0.008588948287069798,
-0.01616729609668255,
0.10835648328065872,
0.1018647626042366,
0.023640237748622894,
0.0729241892695427,
-0.0071894340217113495,
0.048098959028720856,
0.0035175506491214037,
0.03398629277944565,
0.003758437465876341,
0.01848127506673336,
-0.08805745095014572,
-4.6519215146645365e-8,
-0.08615642786026001,
0.043500978499650955,
-0.018360499292612076,
0.06688807904720306,
0.0542205274105072,
-0.018304139375686646,
0.004666890483349562,
0.06578577309846878,
0.028686901554465294,
0.005247490014880896,
0.08071626722812653,
-0.012261994183063507,
-0.04308990016579628,
-0.04469560086727142,
-0.012990361079573631,
0.0664009228348732,
0.015524506568908691,
-0.02395263873040676,
-0.020145233720541,
0.005840224679559469,
0.0728001818060875,
0.05059737712144852,
-0.02858029678463936,
0.03234872967004776,
0.0754132941365242,
-0.043363023549318314,
0.04178368300199509,
0.13263869285583496,
0.04224260151386261,
-0.06045833230018616,
-0.07941663265228271,
0.007064175326377153,
-0.04686547443270683,
-0.02870393730700016,
0.0020307491067796946,
0.08632782101631165,
0.03620481118559837,
-0.03939148783683777,
0.0021143308840692043,
-0.013826444745063782,
0.06586981564760208,
-0.015304273925721645,
-0.03539836034178734,
0.04593563824892044,
0.057968251407146454,
0.0124095743522048,
-0.06263658404350281,
-0.13915404677391052,
0.03466656804084778,
-0.05073167011141777,
0.021318379789590836,
-0.07230958342552185,
0.0011208999203518033,
0.05268922448158264,
0.08164328336715698,
0.033108677715063095,
-0.029175488278269768,
-0.04867374897003174,
0.009898616932332516,
0.047901857644319534,
0.010285002179443836,
0.014740796759724617,
-0.03720450773835182,
0.019886011257767677
] | 0 | 0 | 0 | 0 |
|
Estimating the Probabilities of Rare Outputs in Language Models | https://openreview.net/forum?id=DC8bsa9bzY | [
"Gabriel Wu",
"Jacob Hilton"
] | Spotlight | We consider the problem of *low probability estimation*: given a machine learning model and a formally-specified input distribution, how can we estimate the probability of a binary property of the model's output, even when that probability is too small to estimate by random sampling? This problem is motivated by the need to improve worst-case performance, which distribution shift can make much more likely. We study low probability estimation in the context of argmax sampling from small transformer language models. We compare two types of methods: importance sampling, which involves searching for inputs giving rise to the rare output, and activation extrapolation, which involves extrapolating a probability distribution fit to the model's logits. We find that importance sampling outperforms activation extrapolation, but both outperform naive sampling. Finally, we explain how minimizing the probability estimate of an undesirable behavior generalizes adversarial training, and argue that new methods for low probability estimation are needed to provide stronger guarantees about worst-case performance. | low probabilities, adversarial training, importance sampling | We present methods for estimating the probability that a language model outputs a rare token on a given input distribution. | 12,679 | 2410.13211 | [
-0.1080813780426979,
-0.06966905295848846,
-0.00403509009629488,
0.04017093405127525,
0.08866368234157562,
0.005609826650470495,
0.040136050432920456,
0.03661799803376198,
0.061456240713596344,
-0.03690146654844284,
-0.055579379200935364,
-0.05455799773335457,
0.05696579068899155,
-0.03827428072690964,
-0.01957215927541256,
-0.0014692210825160146,
0.04962756112217903,
0.12163839489221573,
-0.12381209433078766,
-0.07858054339885712,
0.009634464047849178,
0.024417951703071594,
0.02619924768805504,
-0.041172150522470474,
0.07435587793588638,
-0.09003979712724686,
-0.019341325387358665,
-0.032366808503866196,
-0.032105449587106705,
-0.010029880329966545,
-0.03602452948689461,
0.008232302032411098,
0.015400838106870651,
-0.02754291333258152,
0.004123901482671499,
0.02852451056241989,
-0.07492578029632568,
-0.013309843838214874,
0.05777820572257042,
-0.0632440522313118,
0.026021800935268402,
-0.024982914328575134,
0.016481757164001465,
-0.008356488309800625,
0.00839947909116745,
-0.03596556559205055,
0.05707535520195961,
-0.019714806228876114,
-0.015070497989654541,
0.030609335750341415,
-0.026837270706892014,
-0.006116957403719425,
0.028195681050419807,
0.07892389595508575,
-0.08897329121828079,
-0.05143536627292633,
0.032597124576568604,
0.03070261888206005,
0.0019880803301930428,
-0.019159426912665367,
-0.08131348341703415,
-0.08829505741596222,
-0.021519701927900314,
-0.003508408786728978,
-0.05014290660619736,
-0.06485424190759659,
0.04420015215873718,
0.08579037338495255,
-0.0602867528796196,
0.09734035283327103,
-0.005758258979767561,
0.07469844818115234,
-0.06214871257543564,
0.09491249918937683,
0.08580826222896576,
-0.04352904111146927,
0.025559352710843086,
-0.045634567737579346,
0.017016233876347542,
-0.03351663053035736,
0.04000120237469673,
-0.01946546509861946,
0.011169741861522198,
0.02855023927986622,
0.12582409381866455,
0.013524224981665611,
-0.030859466642141342,
0.037672676146030426,
0.07905331254005432,
-0.02705548144876957,
-0.11761768162250519,
-0.024007372558116913,
-0.004034191370010376,
0.0344543382525444,
0.04567057639360428,
0.06614357233047485,
-0.006828194484114647,
-0.06971371173858643,
-0.032862965017557144,
0.13682231307029724,
-0.05776652693748474,
0.047114674001932144,
0.09592767059803009,
-0.053978145122528076,
0.05752938613295555,
-0.06868483871221542,
-0.016849400475621223,
0.005800491198897362,
0.0034978315234184265,
-0.04782671481370926,
0.021418439224362373,
0.05565810948610306,
0.024096118286252022,
0.012963857501745224,
0.05161561444401741,
0.009198391810059547,
0.023707907646894455,
-0.03476858139038086,
-0.15604561567306519,
0.027202092111110687,
-0.1026594266295433,
0.08733735233545303,
0.007867120206356049,
0.03798775002360344,
0.013018753379583359,
0.039878204464912415,
0.01567511260509491,
1.0857230411114395e-33,
0.060410402715206146,
0.03531515598297119,
0.030482234433293343,
-0.028895007446408272,
0.05086571350693703,
0.028254766017198563,
0.04944079741835594,
0.0006736695067957044,
0.06185481697320938,
0.004912491887807846,
-0.06254258751869202,
-0.002644066931679845,
0.03520749881863594,
0.03861265629529953,
0.008283250965178013,
0.08393089473247528,
-0.007526361849159002,
0.06322140246629715,
-0.010079748928546906,
0.02211907133460045,
0.01900084689259529,
0.004167867824435234,
-0.016420060768723488,
-0.05204232782125473,
-0.04498981684446335,
0.06951369345188141,
0.043776266276836395,
-0.04076838120818138,
-0.033058859407901764,
0.013391475193202496,
-0.08484568446874619,
0.02936137281358242,
0.03416816517710686,
-0.05909617990255356,
0.0586770661175251,
-0.028406383469700813,
0.013090543448925018,
-0.031183922663331032,
-0.04792476445436478,
0.013500003144145012,
-0.03882031515240669,
0.04550381749868393,
0.047054022550582886,
0.016542542725801468,
-0.033877842128276825,
-0.04067330062389374,
0.014818449504673481,
-0.04974961280822754,
-0.04162583500146866,
-0.10465496778488159,
0.06166413426399231,
0.008242136798799038,
-0.02359040081501007,
0.02189079485833645,
0.004671817645430565,
-0.0022620137315243483,
0.03614388033747673,
0.018335599452257156,
0.03890569135546684,
0.029928799718618393,
-0.048524655401706696,
0.03382021188735962,
0.022952305153012276,
0.07646070420742035,
0.047404851764440536,
0.004215385764837265,
-0.007746489718556404,
0.013401242904365063,
0.025603847578167915,
-0.015677357092499733,
0.03898444026708603,
-0.03818598389625549,
-0.07027820497751236,
-0.06010253354907036,
0.035641398280858994,
0.050665248185396194,
0.09207388013601303,
-0.061065468937158585,
0.059016335755586624,
0.047845155000686646,
0.008910461328923702,
-0.08914995938539505,
0.0468238927423954,
-0.06724943220615387,
-0.04625677317380905,
0.03245057165622711,
0.04586373642086983,
-0.010241945274174213,
0.009291701018810272,
0.018965043127536774,
-0.09456592798233032,
0.00786826852709055,
-0.1240866407752037,
-0.03599931299686432,
0.015894193202257156,
-1.4995785106859768e-33,
-0.06861608475446701,
0.03155570849776268,
-0.0017444993136450648,
0.032978348433971405,
-0.0674290880560875,
-0.03612053766846657,
0.012669738382101059,
0.008519869297742844,
-0.002600142266601324,
-0.05278371647000313,
-0.09888521581888199,
-0.020491503179073334,
0.11954130232334137,
-0.03686334565281868,
0.027635589241981506,
-0.018136102706193924,
-0.01317155733704567,
0.07831543684005737,
0.03844243288040161,
0.04496780410408974,
0.10655289888381958,
0.07188031077384949,
-0.11109331995248795,
0.06435044854879379,
-0.03154027834534645,
-0.03287694975733757,
-0.08413808047771454,
0.074517622590065,
-0.03447763994336128,
-0.07049063593149185,
-0.06640730798244476,
-0.03791217505931854,
-0.03459487855434418,
-0.019812623038887978,
-0.04671052098274231,
0.021784130483865738,
0.04577018320560455,
0.03215789422392845,
-0.05469147115945816,
0.1056203693151474,
0.01625719480216503,
-0.031238503754138947,
-0.08065232634544373,
-0.010178462602198124,
-0.04737504944205284,
-0.062174249440431595,
-0.05127077177166939,
-0.016212372109293938,
0.013730129227042198,
-0.020252855494618416,
0.0045996676199138165,
0.018843889236450195,
-0.03842370957136154,
0.07761473208665848,
-0.024751048535108566,
-0.06219460070133209,
-0.06000107154250145,
-0.02454555407166481,
0.07027240842580795,
0.0812506228685379,
-0.05769842490553856,
0.03461727872490883,
-0.0312667153775692,
-0.05646824464201927,
-0.009602162055671215,
-0.021886268630623817,
-0.040308889001607895,
-0.0018069614889100194,
0.0540149100124836,
0.010819545015692711,
0.09907212853431702,
-0.017380213364958763,
0.026887455955147743,
0.04842238128185272,
-0.039071619510650635,
0.03238055855035782,
0.02185574546456337,
-0.03503363952040672,
-0.023763271048665047,
-0.043546851724386215,
0.040770359337329865,
0.0018161751795560122,
0.0366184301674366,
0.002173397224396467,
0.03621778264641762,
0.04008728638291359,
0.06977786123752594,
-0.0337405726313591,
0.055416855961084366,
-0.026205826550722122,
-0.06122734025120735,
0.07497648894786835,
-0.025203410536050797,
0.050366420298814774,
-0.07467561960220337,
-4.273451637004655e-8,
-0.058536626398563385,
0.07403832674026489,
0.002220629248768091,
0.06236191838979721,
0.10177204012870789,
0.028780324384570122,
-0.08453365415334702,
0.0525771901011467,
-0.024139391258358955,
-0.07457729429006577,
0.07546458393335342,
-0.001225596060976386,
-0.006230862811207771,
-0.05489007383584976,
0.0113941989839077,
0.08362485468387604,
0.02369876019656658,
0.03903226554393768,
-0.01692088507115841,
0.009568684734404087,
0.026948196813464165,
0.07861039787530899,
0.013483268208801746,
-0.08648238331079483,
-0.0030269634444266558,
-0.07847046107053757,
0.03155885636806488,
0.01063801720738411,
0.031354546546936035,
-0.000528307689819485,
-0.07295654714107513,
-0.02008829638361931,
-0.01118879672139883,
-0.009346610866487026,
0.03172785043716431,
0.1489715576171875,
-0.0227969940751791,
-0.08111698925495148,
-0.029326852411031723,
-0.06551393866539001,
0.03856010362505913,
0.04234572499990463,
-0.037468113005161285,
-0.011969486251473427,
-0.023952113464474678,
-0.04284016415476799,
-0.0012462587328627706,
-0.1125217154622078,
0.06418541073799133,
-0.025538591668009758,
0.055958379060029984,
-0.05027599632740021,
0.018168257549405098,
0.05562310665845871,
0.006583470851182938,
0.02998635172843933,
-0.005145188421010971,
-0.05134889855980873,
-0.03718692436814308,
0.08364851027727127,
0.03633905574679375,
0.020588209852576256,
-0.053778085857629776,
0.016582820564508438
] | https://github.com/alignment-research-center/low-probability-estimation | 7 | 0 | 0 | 0 |
Quality Measures for Dynamic Graph Generative Models | https://openreview.net/forum?id=8bjspmAMBk | [
"Ryien Hosseini",
"Filippo Simini",
"Venkatram Vishwanath",
"Rebecca Willett",
"Henry Hoffmann"
] | Spotlight | Deep generative models have recently achieved significant success in modeling graph data, including dynamic graphs, where topology and features evolve over time. However, unlike in vision and natural language domains, evaluating generative models for dynamic graphs is challenging due to the difficulty of visualizing their output, making quantitative metrics essential. In this work, we develop a new quality metric for evaluating generative models of dynamic graphs. Current metrics for dynamic graphs typically involve discretizing the continuous-evolution of graphs into static snapshots and then applying conventional graph similarity measures. This approach has several limitations: (a) it models temporally related events as i.i.d. samples, failing to capture the non-uniform evolution of dynamic graphs; (b) it lacks a unified measure that is sensitive to both features and topology; (c) it fails to provide a scalar metric, requiring multiple metrics without clear superiority; and (d) it requires explicitly instantiating each static snapshot, leading to impractical runtime demands that hinder evaluation at scale. We propose a novel metric based on the Johnson-Lindenstrauss lemma, applying random projections directly to dynamic graph data. This results in an expressive, scalar, and application-agnostic measure of dynamic graph similarity that overcomes the limitations of traditional methods. We also provide a comprehensive empirical evaluation of metrics for continuous-time dynamic graphs, demonstrating the effectiveness of our approach compared to existing methods. Our implementation is available at https://github.com/ryienh/jl-metric. | generative models, dynamic graphs, evaluation metrics | We introduce a novel metric leveraging random projections to evaluate generative models for dynamic graphs. | 12,628 | 2503.01720 | [
-0.04038408026099205,
-0.07282998412847519,
0.030536942183971405,
0.036890752613544464,
0.02250429429113865,
-0.029155870899558067,
-0.06784460693597794,
-0.021604027599096298,
0.04074804484844208,
-0.029702581465244293,
-0.03506798297166824,
-0.07474088668823242,
0.050820913165807724,
0.07628150284290314,
-0.041122402995824814,
0.0057593039236962795,
0.12397423386573792,
-0.0016405396163463593,
-0.04294336587190628,
-0.04369790479540825,
0.03142689913511276,
0.04769878834486008,
0.023128364235162735,
-0.01167973130941391,
0.06433761119842529,
-0.011427950114011765,
0.04204927384853363,
0.010481733828783035,
0.09306414425373077,
0.006326243281364441,
0.010857466608285904,
0.007454945705831051,
-0.07474054396152496,
-0.0008058552048169076,
-0.004189938306808472,
0.0629635751247406,
0.002230711281299591,
0.01835918240249157,
0.06622277945280075,
-0.019993621855974197,
0.016481677070260048,
0.05321109667420387,
0.022837748751044273,
0.02193080447614193,
-0.021407069638371468,
0.00036749380524270236,
-0.10395687818527222,
-0.021352162584662437,
-0.07313254475593567,
0.060903046280145645,
-0.09747471660375595,
-0.027057478204369545,
-0.030238013714551926,
-0.025336937978863716,
0.07223746180534363,
0.046879105269908905,
0.0009815847733989358,
0.000011962077223870438,
0.039379071444272995,
0.02998078614473343,
0.008774037472903728,
-0.073446124792099,
0.014593388885259628,
-0.01577458716928959,
-0.023731473833322525,
0.032376304268836975,
0.0173005573451519,
0.06555833667516708,
0.02325022779405117,
0.09664829820394516,
-0.04937882348895073,
0.08410754799842834,
-0.09776684641838074,
0.0211102943867445,
0.003756533144041896,
0.07176557928323746,
0.020073436200618744,
-0.010340016335248947,
-0.05955987423658371,
-0.10454519838094711,
0.041039302945137024,
0.01641874760389328,
0.024280834943056107,
0.011480730958282948,
-0.003999638371169567,
0.051222365349531174,
0.02838536724448204,
-0.00019156065536662936,
0.02015180140733719,
0.014956573955714703,
-0.12400619685649872,
0.05202769115567207,
-0.03502818942070007,
-0.04592886567115784,
0.021705947816371918,
-0.004649811424314976,
0.0725584402680397,
-0.07123047858476639,
0.039703864604234695,
0.053162019699811935,
-0.018253255635499954,
0.011576377786695957,
0.012803818099200726,
-0.054332200437784195,
0.035473957657814026,
0.036322060972452164,
-0.01950708217918873,
0.03146900609135628,
0.0033252518624067307,
-0.06005113199353218,
0.04242493584752083,
0.03729896992444992,
0.003016670700162649,
0.03387635201215744,
0.0023338429164141417,
-0.0926472544670105,
-0.0042138732969760895,
-0.07611706107854843,
-0.014897779561579227,
0.059995073825120926,
0.024482926353812218,
0.054087795317173004,
-0.021513596177101135,
-0.03735777363181114,
-0.02608637511730194,
-0.04240413010120392,
0.03419933468103409,
1.1556777201193315e-33,
0.11264846473932266,
-0.016791749745607376,
0.031139807775616646,
0.06514493376016617,
0.09657014906406403,
0.0420529767870903,
-0.05651620775461197,
-0.0008782990626059473,
-0.003580884076654911,
-0.024529840797185898,
-0.07936132699251175,
0.0681493878364563,
-0.05647130683064461,
0.07486839592456818,
0.06652239710092545,
-0.02198999747633934,
-0.007659582886844873,
0.043423499912023544,
-0.017023572698235512,
-0.06078764423727989,
0.020300930365920067,
0.013875013217329979,
0.016961222514510155,
0.03797202557325363,
0.040492646396160126,
0.04705096036195755,
0.014211687259376049,
0.04753937944769859,
-0.041198376566171646,
-0.02542829141020775,
-0.10271143168210983,
-0.0077913678251206875,
0.03139892593026161,
0.04401228204369545,
-0.03813733533024788,
-0.03626535087823868,
-0.07076385617256165,
-0.036228131502866745,
-0.01016730535775423,
-0.10627862066030502,
-0.004538819193840027,
0.010248216800391674,
-0.0437890999019146,
-0.03138372674584389,
-0.046633195132017136,
0.004510944243520498,
0.030579471960663795,
-0.06556679308414459,
0.018683137372136116,
-0.10886625945568085,
0.08205287158489227,
-0.02998887188732624,
0.05400000512599945,
0.04421607404947281,
-0.08578972518444061,
-0.05095149949193001,
0.013734559528529644,
0.032543063163757324,
0.04433232173323631,
0.10651899874210358,
0.008334433659911156,
0.02487986907362938,
-0.013379630632698536,
-0.00971556082367897,
0.057917531579732895,
0.032431069761514664,
0.010861381888389587,
0.008898425847291946,
0.05954154580831528,
-0.01986529305577278,
0.04911348596215248,
0.022646095603704453,
-0.07984702289104462,
-0.07729901373386383,
0.04076549783349037,
0.004924280568957329,
-0.022179383784532547,
-0.058910563588142395,
-0.05691956728696823,
0.08012422174215317,
-0.07046832889318466,
-0.07602804899215698,
0.020962797105312347,
-0.07065685838460922,
0.02755984291434288,
-0.02231515385210514,
0.09303483366966248,
-0.014785414561629295,
-0.011974048800766468,
-0.010102886706590652,
-0.022064710035920143,
-0.05977027490735054,
0.02841341309249401,
0.013062383979558945,
0.06421858072280884,
-2.670629492862156e-33,
-0.06087988242506981,
0.06764411181211472,
0.050221484154462814,
0.14699888229370117,
-0.02313150465488434,
-0.07597009092569351,
-0.08062489330768585,
0.03137877956032753,
-0.060319676995277405,
-0.05152462422847748,
-0.08088403195142746,
-0.08348440378904343,
-0.002494393615052104,
-0.04034663736820221,
0.014549854211509228,
0.01582496240735054,
-0.015300532802939415,
-0.06463898718357086,
-0.07665282487869263,
0.03920239582657814,
0.07904097437858582,
0.05782649293541908,
-0.09114926308393478,
-0.01756552793085575,
0.07288981974124908,
0.10517711192369461,
0.01200693380087614,
0.014725986868143082,
-0.01188034750521183,
-0.024938680231571198,
-0.0411427840590477,
-0.03397649899125099,
-0.06155187636613846,
-0.06322488933801651,
-0.018414471298456192,
0.017608724534511566,
0.0853644609451294,
-0.019445741549134254,
0.0016295767854899168,
0.06953185051679611,
-0.01828719489276409,
-0.04629693925380707,
-0.10112307220697403,
0.06094417721033096,
0.012026166543364525,
0.049759309738874435,
-0.0044464184902608395,
0.02224564366042614,
-0.03442494571208954,
-0.0286127757281065,
-0.0066289231181144714,
0.11383955180644989,
-0.03679176792502403,
0.0363331213593483,
-0.029738791286945343,
-0.13874903321266174,
-0.028127087280154228,
0.028576891869306564,
0.006205495912581682,
0.02835647016763687,
-0.05411631986498833,
-0.07762779295444489,
-0.013387004844844341,
0.01058644987642765,
-0.0339127741754055,
-0.027100568637251854,
-0.0076877521350979805,
-0.11462043225765228,
-0.017850184813141823,
0.0667494684457779,
0.03863389045000076,
-0.04340841993689537,
-0.09762337058782578,
-0.023364154621958733,
0.045684657990932465,
-0.011973842047154903,
0.06400609016418457,
0.029294747859239578,
0.008120057173073292,
-0.03478972241282463,
-0.019952761009335518,
0.010466543026268482,
0.024251025170087814,
-0.019030431285500526,
-0.010236036032438278,
-0.0026085996069014072,
0.021692190319299698,
0.04938403144478798,
0.06386975198984146,
0.11427450180053711,
0.010163292288780212,
-0.06847508251667023,
-0.12471555918455124,
0.05522352457046509,
-0.012388588860630989,
-4.1451873045161847e-8,
-0.0950067788362503,
0.10763922333717346,
0.04581921547651291,
0.003431516932323575,
0.015173060819506645,
0.024558333680033684,
0.1068601906299591,
0.13597372174263,
-0.011074955575168133,
0.006443855818361044,
0.0676666647195816,
-0.058811720460653305,
-0.0181706752628088,
0.017263632267713547,
-0.019143003970384598,
-0.0034475347492843866,
-0.04927660897374153,
0.009105054661631584,
0.008223691955208778,
0.04287485405802727,
-0.045810189098119736,
-0.013002627529203892,
-0.04644711688160896,
0.03337991237640381,
-0.009981540031731129,
-0.10067860782146454,
0.00021976785501465201,
0.01195436529815197,
0.03323826938867569,
-0.010954974219202995,
-0.010313841514289379,
0.05861227959394455,
0.08076522499322891,
-0.04866764694452286,
-0.0876941829919815,
0.052778612822294235,
-0.017512964084744453,
0.019188301637768745,
-0.019275521859526634,
0.01586163602769375,
-0.0003627117839641869,
0.10960356146097183,
-0.037172358483076096,
0.013459745794534683,
-0.018428055569529533,
0.013272952288389206,
0.04872484505176544,
0.03780137002468109,
0.04639627784490585,
-0.03142630308866501,
-0.0687170997262001,
-0.0952613577246666,
-0.042993057519197464,
-0.06310618668794632,
0.020858244970440865,
-0.011571384035050869,
-0.02399292215704918,
-0.003314235247671604,
0.12321584671735764,
0.013966876082122326,
-0.013294555246829987,
-0.07282950729131699,
-0.019346680492162704,
-0.08135950565338135
] | https://github.com/ryienh/jl-metric | 1 | 0 | 0 | 0 |
Correlated Proxies: A New Definition and Improved Mitigation for Reward Hacking | https://openreview.net/forum?id=msEr27EejF | [
"Cassidy Laidlaw",
"Shivam Singhal",
"Anca Dragan"
] | Spotlight | Because it is difficult to precisely specify complex objectives, reinforcement learning policies are often optimized using proxy reward functions that only approximate the true goal. However, optimizing proxy rewards frequently leads to reward hacking: the optimized reward function ceases to be a good proxy and the resulting policy performs poorly with respect to the unspecified true reward. Principled solutions to reward hacking have been impeded by the lack of a good definition for the problem. To address this gap, we introduce a definition of reward hacking based on the correlation between proxy and true rewards for states and actions seen by a “reference policy” that breaks down under optimization. We show that this definition captures reward hacking behavior across several realistic settings, including in reinforcement learning from human feedback (RLHF). Using our formulation, we show theoretically that regularization to the reference policy can effectively prevent reward hacking. While the current practice in RLHF applies a KL penalty between action distributions for this purpose, our theory suggests regularizing the χ2 divergence between the policies’ occupancy measures can be more effective. We intuitively show the benefits of this type of regularization and demonstrate that it better mitigates reward hacking in practice across four realistic settings, including RLHF. Our code is available at https://github.com/cassidylaidlaw/orpo. | reward hacking, reward gaming, overoptimization, occupancy measures | We introduce a new definition of reward hacking, which leads to better regularization strategies for preventing it. | 12,616 | 2403.03185 | [
-0.09923408180475235,
-0.048315975815057755,
0.018158044666051865,
0.0005987075855955482,
-0.01355880219489336,
0.09433520585298538,
0.07781368494033813,
-0.08470811694860458,
0.04487292841076851,
0.07608591765165329,
-0.005175102036446333,
0.008724621497094631,
0.09805093705654144,
0.05898740142583847,
0.04470453038811684,
0.03483366593718529,
0.08899515122175217,
0.04578956589102745,
-0.07771290093660355,
-0.027939317747950554,
0.017532240599393845,
-0.10968635231256485,
0.061606064438819885,
-0.01684931106865406,
-0.05277608335018158,
-0.03494947403669357,
-0.03810814395546913,
-0.032996077090501785,
-0.022931328043341637,
0.011012149974703789,
0.03575487807393074,
-0.0004284910100977868,
-0.03223014995455742,
-0.09222472459077835,
-0.07511841505765915,
0.017121553421020508,
-0.11260142922401428,
-0.029455948621034622,
-0.019851844757795334,
-0.03715148940682411,
-0.04440336674451828,
-0.000649989175144583,
0.04210592061281204,
0.03597281873226166,
-0.02453099563717842,
0.016453148797154427,
0.0788206160068512,
-0.009400447830557823,
0.013728403486311436,
-0.0658167228102684,
0.0033652116544544697,
0.01339690200984478,
0.014630861580371857,
0.029774611815810204,
0.0020430476870387793,
-0.028389986604452133,
0.06870991736650467,
0.07303496450185776,
0.0492030568420887,
-0.008076938800513744,
-0.010744050145149231,
-0.11214873939752579,
-0.0176522359251976,
0.01849443092942238,
-0.05658730864524841,
-0.026236310601234436,
-0.047354329377412796,
0.04356377199292183,
0.006799759343266487,
0.06192636862397194,
-0.09147140383720398,
0.032404378056526184,
-0.03411364555358887,
-0.015354844741523266,
0.13036492466926575,
0.03911074250936508,
-0.09121115505695343,
-0.03195984661579132,
0.039410728961229324,
-0.11478625982999802,
0.030668359249830246,
0.03573213145136833,
-0.015433000400662422,
0.047293901443481445,
0.09945913404226303,
0.0025228599552065134,
0.01028311438858509,
-0.030376814305782318,
0.16777053475379944,
0.010458084754645824,
-0.004557449836283922,
-0.0013600733364000916,
0.03612460941076279,
0.0003912757383659482,
-0.0610480010509491,
-0.020427528768777847,
0.0069085052236914635,
-0.05805734172463417,
-0.10903167724609375,
0.07213752716779709,
0.03377377241849899,
0.07406960427761078,
-0.03169895336031914,
-0.050531916320323944,
0.09571745246648788,
-0.004193936940282583,
0.02385525405406952,
0.04277310147881508,
0.06460952758789062,
-0.036562178283929825,
-0.028555337339639664,
0.06963817030191422,
0.021987665444612503,
0.025708885863423347,
-0.017994677647948265,
0.061335861682891846,
0.0507851243019104,
-0.02748706378042698,
-0.008661167696118355,
0.0542234368622303,
0.06024210900068283,
-0.05690276622772217,
0.10141444206237793,
0.0005522142164409161,
0.008615353144705296,
0.010841316543519497,
-0.032262515276670456,
3.3320419605034555e-33,
0.07654180377721786,
0.008158044889569283,
-0.009361430071294308,
-0.030336009338498116,
0.019765637814998627,
-0.014726052060723305,
0.03590117767453194,
0.024267680943012238,
-0.019301949068903923,
-0.027284882962703705,
0.03262493014335632,
0.00860266387462616,
0.007430154364556074,
0.14287489652633667,
0.08463327586650848,
0.02877235971391201,
-0.08372408896684647,
0.0559433214366436,
0.05065787211060524,
-0.06682027876377106,
0.0422343835234642,
-0.09442934393882751,
0.010119197890162468,
-0.05793491005897522,
0.07278160750865936,
-0.028965016826987267,
-0.01589575596153736,
-0.004550989251583815,
0.021784013137221336,
0.0008622338646091521,
0.02707739919424057,
0.007366507314145565,
0.033128540962934494,
0.05119634419679642,
-0.008054152131080627,
0.0034984215162694454,
0.05694327875971794,
-0.0750904306769371,
-0.007841434329748154,
0.02666868455708027,
-0.06437688320875168,
-0.017905356362462044,
0.0075561897829174995,
0.013798132538795471,
0.004294434562325478,
-0.12428399175405502,
0.006803581956773996,
-0.04822295904159546,
-0.015310856513679028,
-0.03134237974882126,
0.028857896104454994,
0.06705410778522491,
0.04764886945486069,
-0.07453040033578873,
-0.0981307104229927,
0.0042425403371453285,
-0.009026754647493362,
0.01637018844485283,
-0.018987789750099182,
0.06694769114255905,
0.03165789321064949,
0.0058454605750739574,
-0.02615213394165039,
0.08711466193199158,
-0.05151907727122307,
0.023251768201589584,
0.025504017248749733,
0.021446239203214645,
0.015638580545783043,
0.0464838445186615,
0.0012355106882750988,
0.039589084684848785,
-0.03873245418071747,
-0.09842535108327866,
0.030723443254828453,
0.0030854246579110622,
0.045128919184207916,
-0.004964339546859264,
0.02564464509487152,
-0.054016996175050735,
0.004536549560725689,
-0.05525222048163414,
-0.026588836684823036,
0.0007664330769330263,
-0.06714919954538345,
-0.05153822898864746,
0.009494244121015072,
-0.14177775382995605,
-0.029747482389211655,
0.0053123533725738525,
-0.053336478769779205,
-0.007649463135749102,
-0.020876768976449966,
0.07271743565797806,
0.010354874655604362,
-4.0316697494899875e-33,
-0.039696622639894485,
-0.0352848544716835,
0.0031743068248033524,
-0.010915498249232769,
-0.00021750497398898005,
0.013707688078284264,
-0.011595652438700199,
-0.05760909616947174,
-0.010037880390882492,
-0.01825786381959915,
-0.05174986645579338,
0.014534012414515018,
0.03777457773685455,
0.009577206335961819,
0.007404040079563856,
-0.09389130026102066,
-0.015221690759062767,
0.06535804271697998,
-0.04123575612902641,
0.018097391352057457,
0.05127020180225372,
0.07227276265621185,
-0.06875797361135483,
0.0392170175909996,
0.030337275937199593,
0.007242904044687748,
0.013533341698348522,
0.07617087662220001,
0.02721271477639675,
0.08869761973619461,
-0.013678600080311298,
0.02759668603539467,
-0.07781369984149933,
0.0010286468314006925,
-0.05069911479949951,
0.019725916907191277,
0.026004072278738022,
0.041731081902980804,
-0.13381485641002655,
0.0303557850420475,
0.046413786709308624,
-0.0659933015704155,
-0.062356967478990555,
-0.0032529672607779503,
0.01904837228357792,
-0.03914039209485054,
-0.06564967334270477,
-0.001078232889994979,
-0.021292919293045998,
-0.018036721274256706,
-0.003856623312458396,
-0.02892949990928173,
-0.039014894515275955,
0.036919232457876205,
-0.12445603311061859,
-0.007221440318971872,
-0.05579804256558418,
0.005474640056490898,
0.1335701048374176,
0.020019838586449623,
-0.0715160220861435,
0.023654179647564888,
-0.06243675947189331,
0.1065678820014,
-0.028446728363633156,
-0.03755439817905426,
-0.0076259104534983635,
0.048308901488780975,
0.019526684656739235,
-0.0033842993434518576,
0.08238744735717773,
0.024520784616470337,
0.032280776649713516,
-0.03223547711968422,
0.011107726022601128,
0.036641210317611694,
0.012300029397010803,
-0.042375143617391586,
-0.07315759360790253,
0.05526844784617424,
0.0694880485534668,
0.015273243188858032,
0.07320252805948257,
-0.0246447641402483,
-0.061755381524562836,
0.04635069519281387,
0.0024076588451862335,
0.10237618535757065,
0.007937022484838963,
-0.06466268002986908,
-0.024774884805083275,
-0.05051817372441292,
-0.06762208789587021,
-0.01317452359944582,
0.005232378374785185,
-4.901386674305286e-8,
-0.014490979723632336,
0.06523330509662628,
0.048120953142642975,
0.03508319333195686,
0.035380810499191284,
0.02482943795621395,
-0.04806698113679886,
-0.0043532694689929485,
-0.044527728110551834,
-0.04487015679478645,
0.11813469976186752,
0.04036654904484749,
0.0032046623528003693,
-0.0878973975777626,
-0.07167164236307144,
0.02984701097011566,
-0.03447451442480087,
-0.012748733162879944,
-0.020517511293292046,
0.060414671897888184,
0.027816545218229294,
0.018715908750891685,
-0.006118481513112783,
-0.07865870743989944,
-0.03941365331411362,
-0.0520160086452961,
0.012156259268522263,
0.05749291554093361,
0.06697481125593185,
0.0233792532235384,
-0.002321838168427348,
-0.050851400941610336,
0.05037294328212738,
0.037930816411972046,
-0.02661103941500187,
0.0986848920583725,
-0.028534119948744774,
-0.10238613933324814,
-0.010650331154465675,
0.01756282150745392,
-0.013706273399293423,
-0.005643472541123629,
0.04028948023915291,
-0.03003779612481594,
-0.05877544358372688,
0.029398975893855095,
-0.0794091522693634,
-0.04459536448121071,
0.08905133605003357,
-0.10680381208658218,
0.057733792811632156,
-0.05833935737609863,
-0.028335345908999443,
0.046090323477983475,
0.06526000797748566,
-0.0047407448291778564,
-0.00843511987477541,
-0.08499555289745331,
-0.01606210321187973,
0.06862406432628632,
0.06742047518491745,
0.022630007937550545,
-0.07622136920690536,
-0.029826268553733826
] | https://github.com/cassidylaidlaw/orpo | 15 | 0 | 0 | 0 |
Holistically Evaluating the Environmental Impact of Creating Language Models | https://openreview.net/forum?id=04qx93Viwj | [
"Jacob Morrison",
"Clara Na",
"Jared Fernandez",
"Tim Dettmers",
"Emma Strubell",
"Jesse Dodge"
] | Spotlight | As the performance of artificial intelligence systems has dramatically increased, so too has the environmental impact of creating these systems. While many model developers release estimates of the power consumption and carbon emissions from the final training runs for their latest models, there is comparatively little transparency into the impact of model development, hardware manufacturing, and total water usage throughout. In this work, we estimate the real-world environmental impact of developing a series of language models, ranging from 20 million to 13 billion active parameters, trained on up to 5.6 trillion tokens each. When accounting for hardware manufacturing, model development, and our final training runs, we find that our series of models released **493 metric tons** of carbon emissions, equivalent to powering about 98 homes in the United States for one year, and consumed **2.769 million liters of water**, equivalent to about 24.5 years of water usage by a person in the United States, even though our data center is extremely water-efficient. We measure and report the environmental impact of our model development; to the best of our knowledge we are the first to do so for LLMs, and we find that model development, the impact of which is generally not disclosed by most model developers, amounted to **~50%** of that of training. By looking at detailed time series data for power consumption, we also find that power usage throughout training is not consistent, fluctuating between ~15% and ~85% of our hardware's maximum power draw, with negative implications for grid-scale planning as demand continues to grow. We close with a discussion on the continued difficulty of estimating the environmental impact of AI systems, and key takeaways for model developers and the public at large. | machine learning, artificial intelligence, language model, large language models, environmental impact, carbon emissions, water usage | Measuring the environmental impact, including carbon emissions and water usage, from training a series of language models. | 12,568 | 2503.05804 | [
0.012427812442183495,
-0.037363212555646896,
0.055720239877700806,
0.07187322527170181,
0.06552722305059433,
-0.05045237019658089,
-0.029276492074131966,
0.020603422075510025,
0.03865455463528633,
0.020510300993919373,
-0.08168358355760574,
-0.0939147025346756,
0.060098737478256226,
0.0030515959952026606,
-0.026418717578053474,
0.03518866002559662,
0.008547552861273289,
-0.051448654383420944,
-0.1005348414182663,
-0.05948704481124878,
0.06428354978561401,
0.031199797987937927,
-0.00038059570943005383,
0.006277464330196381,
0.04945247247815132,
0.03546913340687752,
0.026058457791805267,
-0.01258750632405281,
0.014357229694724083,
0.023548444733023643,
-0.017336638644337654,
0.036100760102272034,
0.04480032995343208,
-0.005168020259588957,
0.05995795875787735,
0.0026272437535226345,
0.05378729850053787,
-0.055456191301345825,
0.044720038771629333,
-0.045533813536167145,
-0.058784108608961105,
-0.08738714456558228,
0.04552662372589111,
-0.00516603933647275,
0.036854907870292664,
0.0118624372407794,
0.043128568679094315,
-0.053702160716056824,
-0.09538109600543976,
-0.009692270308732986,
-0.037814557552337646,
-0.06982661783695221,
-0.022563345730304718,
-0.05164671689271927,
-0.0354006253182888,
-0.0808725506067276,
0.06813715398311615,
-0.06525012105703354,
0.009223981760442257,
-0.0209749024361372,
-0.03875499591231346,
-0.10384932905435562,
-0.09522393345832825,
0.02078208327293396,
0.04265958070755005,
0.016200486570596695,
-0.04049692675471306,
0.05694101005792618,
-0.005456285085529089,
0.004176119342446327,
-0.011567083187401295,
0.05900908634066582,
-0.022055361419916153,
-0.06306125968694687,
-0.04318876191973686,
-0.007969479076564312,
0.0761590376496315,
0.020398413762450218,
0.15536701679229736,
-0.09448370337486267,
0.07207529991865158,
-0.034368496388196945,
-0.013027983717620373,
0.02634524740278721,
0.03810727968811989,
0.03247006610035896,
0.029135866090655327,
0.014522573910653591,
0.042488228529691696,
-0.004303105175495148,
-0.05633432790637016,
-0.05936003103852272,
0.09203716367483139,
0.02512974850833416,
-0.006064542103558779,
0.12176170945167542,
0.01229940913617611,
-0.0818723738193512,
0.02002626657485962,
0.03640935197472572,
-0.03842915967106819,
0.09478544443845749,
0.06394501030445099,
-0.05473216250538826,
0.037908025085926056,
-0.03651793673634529,
0.020336152985692024,
0.03720443323254585,
0.03612634912133217,
-0.021819373592734337,
0.013250858522951603,
0.03121674247086048,
-0.08357124775648117,
0.06097656860947609,
0.08394268155097961,
-0.005197583697736263,
0.01664629392325878,
-0.05563049763441086,
0.019489185884594917,
0.058132756501436234,
-0.0354633554816246,
0.039882026612758636,
-0.019593577831983566,
0.0386933796107769,
0.03243555501103401,
-0.008264211006462574,
-0.057201024144887924,
1.0323229959829715e-33,
0.03252477943897247,
0.041843533515930176,
0.020996056497097015,
0.03433959558606148,
0.05276865139603615,
-0.006654942873865366,
-0.010191441513597965,
0.06743457168340683,
0.025159131735563278,
0.011608944274485111,
-0.05008159577846527,
0.04450087994337082,
-0.0560903400182724,
0.050535380840301514,
0.15079303085803986,
-0.0369437150657177,
-0.019262362271547318,
0.03838770091533661,
-0.03893755003809929,
0.003182303858920932,
0.036676377058029175,
-0.09594415128231049,
0.02100210264325142,
-0.0174192413687706,
0.08759339898824692,
-0.0008326231618411839,
0.07317380607128143,
0.015874410048127174,
-0.06027206778526306,
0.0035287688951939344,
0.015219145454466343,
0.025242870673537254,
-0.0029871000442653894,
0.056562259793281555,
0.017907751724123955,
-0.026065943762660027,
0.017623022198677063,
0.036834150552749634,
0.046769287437200546,
0.02156778797507286,
-0.0416245311498642,
0.09140347689390182,
0.017002694308757782,
-0.08107650279998779,
-0.009089316241443157,
-0.02475484274327755,
0.06493882834911346,
-0.05052000656723976,
0.024926230311393738,
-0.05146007984876633,
0.01074178982526064,
0.04573812708258629,
-0.03293353319168091,
-0.0006168110994622111,
-0.007902407087385654,
-0.05799403041601181,
0.03590472415089607,
-0.10774950683116913,
0.022007092833518982,
0.037264615297317505,
-0.10844657570123672,
0.042406052350997925,
0.01745210401713848,
0.048118919134140015,
0.08520389348268509,
0.03722738102078438,
0.05228468030691147,
0.05937160924077034,
0.07132691144943237,
-0.037104252725839615,
-0.03234444931149483,
-0.06818979978561401,
-0.024034487083554268,
-0.023174550384283066,
0.00446351757273078,
-0.029527757316827774,
0.05218562111258507,
-0.11766297370195389,
-0.05561406910419464,
0.08227882534265518,
-0.03542925417423248,
0.00236141518689692,
-0.011481883004307747,
-0.06891775876283646,
0.05333620309829712,
-0.02443690225481987,
-0.011926045641303062,
-0.03435207158327103,
0.042402565479278564,
0.03361603617668152,
-0.03185654431581497,
-0.06535638123750687,
-0.0207674503326416,
-0.035909950733184814,
-0.060090504586696625,
-2.614640880603483e-33,
-0.08200033754110336,
0.024311726912856102,
-0.04828626662492752,
0.05926171690225601,
-0.024066142737865448,
-0.12295924872159958,
-0.02794467657804489,
0.02557089738547802,
0.010813518427312374,
-0.02156948298215866,
-0.04512602835893631,
0.02678629942238331,
0.015093090012669563,
0.013667904771864414,
0.055848825722932816,
-0.04617340862751007,
-0.04700260981917381,
-0.07450321316719055,
0.058891359716653824,
-0.00495964614674449,
0.003495992161333561,
0.034766245633363724,
-0.05326450243592262,
-0.008387244306504726,
-0.05737987160682678,
0.04405499994754791,
-0.0877465233206749,
-0.015793027356266975,
0.06526844203472137,
-0.000020943385607097298,
-0.07480001449584961,
0.03065013699233532,
0.01060554850846529,
0.009864277206361294,
-0.05420014634728432,
-0.017105165868997574,
0.052580345422029495,
-0.01110111828893423,
0.030445266515016556,
0.04497108981013298,
0.07828019559383392,
-0.06537572294473648,
-0.10005293041467667,
-0.01335070375353098,
-0.020815178751945496,
0.001539557590149343,
-0.040718916803598404,
-0.043653201311826706,
0.023983223363757133,
-0.0644594132900238,
0.10197900980710983,
-0.004915949888527393,
-0.10824450850486755,
-0.03866918012499809,
-0.04880697652697563,
-0.07544756680727005,
0.10811164975166321,
-0.049104176461696625,
-0.028454100713133812,
0.009207075461745262,
-0.026561379432678223,
0.042580947279930115,
0.04493798315525055,
-0.046247448772192,
-0.04601050913333893,
-0.006371107883751392,
-0.03659431263804436,
0.06982570886611938,
-0.009073661640286446,
-0.0683898851275444,
-0.004644744098186493,
-0.022130141034722328,
-0.07846067100763321,
0.017300115898251534,
-0.08751758933067322,
-0.014134908095002174,
-0.04256002604961395,
-0.06769002228975296,
-0.016565201804041862,
0.01089164987206459,
-0.03217916190624237,
0.11749031394720078,
0.009471765719354153,
-0.0045348964631557465,
0.054343510419130325,
-0.04434669762849808,
-0.001613063970580697,
-0.025935089215636253,
-0.043673135340213776,
0.0821802169084549,
-0.10972407460212708,
0.0313897468149662,
-0.10483580082654953,
0.07762696593999863,
-0.0663798525929451,
-4.829390221061658e-8,
-0.05456925183534622,
0.04136526957154274,
0.09207868576049805,
0.06810548901557922,
0.021296212449669838,
-0.03527490049600601,
0.013114575296640396,
0.08449003100395203,
0.04390542954206467,
-0.006666169036179781,
0.04259806498885155,
-0.03501991927623749,
0.0061456505209207535,
0.04825607314705849,
0.01573621854186058,
0.08397729694843292,
0.003158760257065296,
-0.023941822350025177,
-0.08832736313343048,
-0.028017189353704453,
0.07982134819030762,
0.04400882497429848,
-0.0009666080004535615,
0.02665250562131405,
0.08234335482120514,
-0.0779232606291771,
-0.032829031348228455,
0.07900863140821457,
0.0019889851100742817,
0.03347340598702431,
0.022665711119771004,
0.010509784333407879,
-0.037039998918771744,
-0.0004457776958588511,
0.03386394679546356,
-0.03537680581212044,
-0.015180793590843678,
-0.08210791647434235,
0.029376864433288574,
0.04500091075897217,
-0.0031825401820242405,
0.06963114440441132,
-0.073068767786026,
0.044647689908742905,
0.05936037749052048,
-0.08395261317491531,
-0.10235486179590225,
-0.08830893784761429,
0.03313158079981804,
0.014114477671682835,
-0.015188547782599926,
-0.03840586915612221,
0.006935985758900642,
0.05483948439359665,
0.06821881979703903,
0.07472553104162216,
-0.05637619271874428,
-0.032533448189496994,
-0.036526475101709366,
0.03504433482885361,
0.06982900202274323,
-0.050718169659376144,
-0.03845031559467316,
0.007301649544388056
] | 0 | 0 | 0 | 0 |
|
Mixture-of-Agents Enhances Large Language Model Capabilities | https://openreview.net/forum?id=h0ZfDIrj7T | [
"Junlin Wang",
"Jue WANG",
"Ben Athiwaratkun",
"Ce Zhang",
"James Zou"
] | Spotlight | Recent advances in large language models (LLMs) demonstrate substantial capabilities in natural language understanding and generation tasks. With the growing number of LLMs, how to harness the collective expertise of multiple LLMs is an exciting open direction. Toward this goal, we propose a new approach that leverages the collective strengths of multiple LLMs through a Mixture-of-Agents (MoA) methodology. In our approach, we construct a layered MoA architecture wherein each layer comprises multiple LLM agents. Each agent takes all the outputs from agents in the previous layer as auxiliary information in generating its response. MoA models achieves state-of-art performance on AlpacaEval 2.0, Arena-Hard, MT-Bench, and FLASK, surpassing GPT-4 Omni. For example, our MoA using only open-source LLMs achieves a score of 65.1% on AlpacaEval 2.0 compared to 57.5% by GPT-4 Omni. | Multi-Agent Inference, Large Language Model | This paper presents a method that synergistically leverage multiple LLMs to significantly improve their performance. | 12,544 | 2406.04692 | [
-0.012137015350162983,
-0.07308120280504227,
0.007192264311015606,
0.03808698058128357,
0.060496460646390915,
-0.06447654217481613,
-0.0787184089422226,
-0.04949672892689705,
0.050739552825689316,
-0.10231509804725647,
-0.06603047996759415,
-0.1063876822590828,
0.04829150810837746,
-0.0005771434516645968,
0.023150505498051643,
0.05098200589418411,
0.10331741720438004,
0.053048089146614075,
-0.055024534463882446,
-0.052779197692871094,
0.04374944418668747,
0.04054410010576248,
0.06348104029893875,
-0.025540558621287346,
0.020179033279418945,
0.04324381798505783,
-0.09200002253055573,
-0.02803756110370159,
0.08698048442602158,
-0.02963724173605442,
0.0770527645945549,
0.07497238367795944,
-0.007474695332348347,
0.06962568312883377,
-0.05906197428703308,
0.10719160735607147,
-0.08443159610033035,
0.052756063640117645,
0.020402971655130386,
0.0008948687463998795,
-0.04469449818134308,
0.06895428150892258,
0.02202838845551014,
0.0028918967582285404,
0.06063641607761383,
-0.07890791445970535,
-0.027858205139636993,
-0.0018505941843613982,
-0.005858606193214655,
0.07824184000492096,
-0.07578808069229126,
-0.09133970737457275,
-0.003941210452467203,
0.0620836466550827,
-0.060267604887485504,
0.012303542345762253,
-0.018106909468770027,
0.0142462607473135,
-0.01288579497486353,
-0.07230748981237411,
-0.06257998198270798,
-0.08723903447389603,
-0.026341712102293968,
0.010958004742860794,
0.03973730280995369,
-0.0367538146674633,
0.006433720234781504,
0.07404210418462753,
-0.022169269621372223,
-0.06811965256929398,
-0.015462576411664486,
0.005268862936645746,
0.034288130700588226,
0.040704600512981415,
0.023716488853096962,
0.06700669974088669,
0.018367120996117592,
0.009681493043899536,
0.10423146933317184,
0.00048411564785055816,
0.001480569364503026,
0.016467571258544922,
0.00834562350064516,
-0.016363825649023056,
0.09363624453544617,
0.016747936606407166,
0.02600245550274849,
0.09613930433988571,
0.00526126753538847,
0.025244401767849922,
-0.059557266533374786,
-0.008268059231340885,
0.10507820546627045,
0.04377283155918121,
0.03333398327231407,
0.07786885648965836,
0.025966210290789604,
-0.11013634502887726,
-0.002372763818129897,
0.10529468953609467,
-0.01595270447432995,
0.0006501651951111853,
0.012779127806425095,
-0.03796767815947533,
-0.054209038615226746,
-0.05600598827004433,
0.040879882872104645,
0.0635782927274704,
-0.03688449040055275,
-0.058493729680776596,
-0.006312204524874687,
0.08431682735681534,
0.07133445143699646,
-0.020639168098568916,
0.0417281873524189,
-0.007863402366638184,
0.0013390134554356337,
-0.0037896365392953157,
-0.018649108707904816,
0.03177817538380623,
-0.06387748569250107,
0.012861154042184353,
0.022061379626393318,
0.026309337466955185,
0.0701286643743515,
0.03380459547042847,
-0.0012820811243727803,
8.016441682485981e-34,
0.12662087380886078,
0.015891076996922493,
0.02640374004840851,
0.0836581140756607,
0.0927208960056305,
0.006393107119947672,
0.04524282366037369,
0.015291839838027954,
-0.05226346477866173,
-0.032918043434619904,
0.007879220880568027,
0.04944869503378868,
-0.06817959994077682,
0.02266623266041279,
0.0966615304350853,
-0.028665095567703247,
-0.030469819903373718,
0.020809324458241463,
-0.04300854355096817,
-0.039769530296325684,
-0.01610647328197956,
0.0446004681289196,
0.06422341614961624,
0.022182080894708633,
0.03160997852683067,
0.06776829063892365,
0.0896330177783966,
-0.08518640697002411,
-0.038440853357315063,
0.01324333343654871,
-0.0861688107252121,
-0.035863522440195084,
-0.03468611091375351,
0.03560055419802666,
-0.040025386959314346,
-0.009914102032780647,
-0.09075217694044113,
-0.08204572647809982,
0.045503661036491394,
-0.020985960960388184,
-0.09234199672937393,
0.039409339427948,
0.05343968793749809,
-0.06423849612474442,
-0.11917231231927872,
-0.023853862658143044,
-0.04414704814553261,
0.06738872826099396,
0.027224566787481308,
-0.04074115306138992,
0.04058624804019928,
0.005825051572173834,
-0.03543669730424881,
0.03485990688204765,
-0.03534460440278053,
0.013694281689822674,
-0.04453936591744423,
0.05575244501233101,
0.03937628120183945,
0.04973309114575386,
0.008042794652283192,
0.06618558615446091,
0.008546385914087296,
0.08376234024763107,
0.06333591789007187,
-0.019798606634140015,
-0.014957040548324585,
0.12304820120334625,
0.06377621740102768,
0.021742017939686775,
0.0016141448868438601,
-0.008247950114309788,
-0.001817859592847526,
-0.009596370160579681,
0.017701080068945885,
-0.041014160960912704,
0.05704815685749054,
-0.1545659452676773,
0.046428244560956955,
0.010599985718727112,
0.02039492130279541,
0.04344174638390541,
-0.013195560313761234,
-0.06969539821147919,
-0.04822370409965515,
0.012058276683092117,
-0.02263072319328785,
-0.04876885190606117,
-0.012334000319242477,
0.04102832451462746,
-0.1193993091583252,
-0.04444415122270584,
-0.020810607820749283,
-0.018777403980493546,
-0.025455569848418236,
-1.980637267803513e-34,
-0.03065381571650505,
-0.0315890796482563,
-0.006452719680964947,
0.05302702635526657,
-0.009761810302734375,
-0.08655378222465515,
-0.054142292588949203,
-0.0035817271564155817,
-0.0488334596157074,
0.03734380751848221,
-0.03499985858798027,
-0.02819894254207611,
0.05885874107480049,
-0.009253507480025291,
0.03123517334461212,
-0.0724726989865303,
0.010958009399473667,
-0.014263163320720196,
0.08316211402416229,
0.016947345808148384,
-0.015417568385601044,
0.06687089055776596,
-0.0428396537899971,
0.018718475475907326,
-0.02820182405412197,
0.008299586363136768,
-0.060039397329092026,
0.03925108164548874,
0.029731785878539085,
0.019178396090865135,
0.013730589300394058,
0.014190305024385452,
-0.04115324467420578,
-0.01474814023822546,
-0.05757131800055504,
0.04711342602968216,
0.05638531595468521,
0.044670283794403076,
-0.015093176625669003,
0.04685273766517639,
0.057950638234615326,
-0.0851765051484108,
-0.06724220514297485,
0.00376121629960835,
-0.028734559193253517,
0.045671116560697556,
-0.08913575112819672,
0.028554661199450493,
0.005226483568549156,
-0.010988677851855755,
-0.03344687074422836,
-0.011096803471446037,
-0.05062449350953102,
-0.052586618810892105,
-0.009840755723416805,
-0.08778456598520279,
-0.03136331960558891,
-0.08010061830282211,
-0.026355180889368057,
-0.06425917148590088,
-0.04737156629562378,
-0.0030185054056346416,
0.01981147937476635,
-0.09505210816860199,
-0.009851394221186638,
-0.02671119011938572,
-0.06374819576740265,
0.02126547507941723,
-0.011733479797840118,
0.019979456439614296,
0.05892554670572281,
-0.0259443037211895,
-0.038620125502347946,
0.0035016725305467844,
0.04332440346479416,
0.03212594985961914,
-0.0493607372045517,
-0.051214784383773804,
0.02350756525993347,
-0.09390262514352798,
-0.005967414006590843,
0.08808105438947678,
0.047350745648145676,
0.0292779840528965,
0.09024199843406677,
0.046120431274175644,
0.044173337519168854,
0.11603334546089172,
0.035226620733737946,
0.06534208357334137,
-0.007126488257199526,
0.02597266249358654,
0.03742855042219162,
0.07917863130569458,
-0.07809572666883469,
-3.936607484433807e-8,
-0.02046469785273075,
0.004643091931939125,
-0.018184000626206398,
0.03560037910938263,
-0.035966504365205765,
-0.0069960071705281734,
-0.026836024597287178,
0.0455225445330143,
0.02882958948612213,
-0.006832457147538662,
0.0828714668750763,
-0.031441912055015564,
-0.052492253482341766,
-0.019606085494160652,
0.03189466893672943,
0.06213989108800888,
-0.03849908336997032,
-0.00638345442712307,
-0.03131929785013199,
-0.10486564040184021,
0.08162230998277664,
0.01339062862098217,
0.018051696941256523,
0.0713096484541893,
0.04970093443989754,
-0.059603530913591385,
-0.06395243853330612,
0.04826739802956581,
-0.006499841809272766,
0.014825915917754173,
-0.0025653603952378035,
-0.028168056160211563,
-0.03162789344787598,
-0.039622869342565536,
0.05744923651218414,
0.044140782207250595,
-0.02635345049202442,
-0.023555127903819084,
0.049094077199697495,
-0.059441447257995605,
0.017953947186470032,
0.03687458485364914,
0.003789691487327218,
0.0392192080616951,
0.02960503287613392,
-0.03207102790474892,
-0.09448569267988205,
-0.11937334388494492,
0.08017213642597198,
-0.07171749323606491,
-0.0343342199921608,
-0.07955505698919296,
-0.02169371396303177,
0.0688142254948616,
0.01862446591258049,
-0.021263567730784416,
-0.011563941836357117,
-0.06445740908384323,
0.08887351304292679,
-0.05214408412575722,
0.02210407704114914,
-0.03152516111731529,
-0.003010852262377739,
-0.021623393520712852
] | 0 | 0 | 0 | 7 |
|
Answer, Assemble, Ace: Understanding How LMs Answer Multiple Choice Questions | https://openreview.net/forum?id=6NNA0MxhCH | [
"Sarah Wiegreffe",
"Oyvind Tafjord",
"Yonatan Belinkov",
"Hannaneh Hajishirzi",
"Ashish Sabharwal"
] | Spotlight | Multiple-choice question answering (MCQA) is a key competence of performant transformer language models that is tested by mainstream benchmarks. However, recent evidence shows that models can have quite a range of performance, particularly when the task format is diversified slightly (such as by shuffling answer choice order). In this work we ask: how do successful models perform formatted MCQA? We employ vocabulary projection and activation patching methods to localize key hidden states that encode relevant information for predicting the correct answer. We find that prediction of a specific answer symbol is causally attributed to a few middle layers, and specifically their multi-head self-attention mechanisms. We show that subsequent layers increase the probability of the predicted answer symbol in vocabulary space, and that this probability increase is associated with a sparse set of attention heads with unique roles. We additionally uncover differences in how different models adjust to alternative symbols. Finally, we demonstrate that a synthetic task can disentangle sources of model error to pinpoint when a model has learned formatted MCQA, and show that logit differences between answer choice tokens continue to grow over the course of training. | interpretability; multiple-choice question answering | null | 12,533 | 2407.15018 | [
0.0007995844935066998,
-0.10965869575738907,
0.04713261500000954,
0.06526907533407211,
0.07021603733301163,
0.013337214477360249,
0.026812465861439705,
0.04173387959599495,
0.07782554626464844,
0.021665966138243675,
-0.04035406559705734,
-0.08665552735328674,
0.03942028805613518,
0.04106881842017174,
-0.0029499810189008713,
0.019639350473880768,
0.0365213118493557,
0.0360625721514225,
-0.07542219012975693,
-0.04350734129548073,
0.053531695157289505,
-0.03411080688238144,
0.06784454733133316,
0.01521441899240017,
0.041972965002059937,
-0.0004503587551880628,
-0.04936730116605759,
-0.058309268206357956,
-0.05797310173511505,
-0.04618757963180542,
0.053387586027383804,
-0.009212026372551918,
-0.022733576595783234,
0.08280173689126968,
-0.09119953215122223,
0.07117937505245209,
-0.07683253288269043,
-0.01803010143339634,
0.017459748312830925,
-0.09317392855882645,
-0.01191474124789238,
-0.036335840821266174,
0.044386327266693115,
0.03910559415817261,
0.09992243349552155,
-0.014116991311311722,
0.003076310735195875,
-0.01669345237314701,
-0.0009390095947310328,
-0.028102871030569077,
-0.03757118433713913,
-0.04238810017704964,
-0.00911734625697136,
-0.008121555671095848,
-0.03509296104311943,
0.02551455982029438,
0.032980188727378845,
0.07974410057067871,
-0.03260171040892601,
-0.04329594969749451,
-0.09258271008729935,
-0.0470244325697422,
-0.08907751739025116,
0.0058779167011380196,
-0.0037166208494454622,
-0.05245814844965935,
-0.026689348742365837,
0.01712622307240963,
-0.10982073098421097,
0.029715316370129585,
-0.022557292133569717,
0.044994063675403595,
0.03695875033736229,
-0.024357248097658157,
0.13876758515834808,
0.015286353416740894,
0.06160806491971016,
-0.027582217007875443,
0.046412065625190735,
-0.03465283289551735,
0.012563897296786308,
0.021177399903535843,
0.04417983442544937,
0.024551715701818466,
0.15331298112869263,
0.039233118295669556,
-0.024658797308802605,
0.06005612388253212,
-0.021505892276763916,
-0.06361563503742218,
-0.06822721660137177,
-0.11580866575241089,
0.022189415991306305,
-0.03971322253346443,
0.0732160434126854,
0.04022189974784851,
0.021021587774157524,
-0.02850302681326866,
-0.01517824549227953,
0.06113317981362343,
0.0507199726998806,
0.04050601273775101,
0.007139141671359539,
-0.05395665392279625,
-0.04155189171433449,
-0.043219320476055145,
-0.000741704017855227,
0.06879720836877823,
0.016903268173336983,
-0.0794193223118782,
0.018974486738443375,
0.045758605003356934,
0.026789238676428795,
-0.0033074889797717333,
0.015671532601118088,
-0.019992653280496597,
-0.0076773385517299175,
0.0005320695927366614,
0.007813844829797745,
0.034500934183597565,
-0.03397752344608307,
0.04530412703752518,
0.05786785110831261,
-0.055241066962480545,
0.05636297166347504,
-0.03090563789010048,
-0.02739134058356285,
1.3515731189998965e-33,
0.06720339506864548,
0.009067391976714134,
0.044381216168403625,
0.029811382293701172,
0.027997516095638275,
-0.007679513189941645,
-0.011178681626915932,
0.031490523368120193,
0.014946067705750465,
-0.007679032627493143,
0.01999894715845585,
0.00825184304267168,
-0.029219936579465866,
0.07346697151660919,
0.004898297134786844,
0.058098502457141876,
-0.08356736600399017,
0.06846003234386444,
-0.11647160351276398,
-0.011512973345816135,
0.01802954077720642,
0.03309999778866768,
0.035407185554504395,
-0.034123897552490234,
-0.008412332274019718,
-0.02143603004515171,
0.0681576207280159,
-0.10118590295314789,
-0.08943501859903336,
0.019887614995241165,
-0.12835726141929626,
-0.02848486602306366,
-0.023530954495072365,
-0.013512185774743557,
0.048905149102211,
-0.043490007519721985,
-0.013831852003932,
-0.07371699064970016,
0.02305482141673565,
0.0015440966235473752,
-0.01166441012173891,
0.032262224704027176,
0.055427681654691696,
0.05343911051750183,
-0.07618376612663269,
-0.06028749793767929,
-0.043521661311388016,
0.00966684054583311,
-0.018535058945417404,
-0.011443701572716236,
0.0744854286313057,
0.014815187081694603,
-0.04584939405322075,
-0.053437452763319016,
0.060310691595077515,
0.02105087600648403,
0.005484013818204403,
0.044619638472795486,
0.000895132077857852,
0.020556997507810593,
0.025204619392752647,
0.05300624668598175,
-0.055248767137527466,
0.1084219366312027,
0.041058920323848724,
0.020308731123805046,
0.009835334494709969,
0.0775614082813263,
0.08895605057477951,
0.06962724775075912,
0.000544751004781574,
-0.022524354979395866,
-0.04085348919034004,
-0.053197070956230164,
0.04109730198979378,
0.019635289907455444,
0.055005766451358795,
-0.09266648441553116,
0.02504517324268818,
-0.03253953158855438,
0.05942046269774437,
-0.04304501414299011,
-0.04234858229756355,
-0.07613810151815414,
-0.0754474624991417,
-0.03856772184371948,
0.09123116731643677,
-0.10662104934453964,
-0.005480631720274687,
-0.0066874148324131966,
0.0012604722287505865,
-0.052643779665231705,
0.005601738579571247,
-0.055053289979696274,
-0.02119406685233116,
-1.164734169408448e-33,
-0.0011195966508239508,
-0.06329245865345001,
-0.021974891424179077,
0.022440830245614052,
-0.054614629596471786,
-0.031051157042384148,
0.03856853023171425,
-0.006215415429323912,
-0.05303054675459862,
-0.041629329323768616,
-0.05142897740006447,
-0.001871069660410285,
0.06517297029495239,
0.05331520363688469,
0.034755364060401917,
-0.020015770569443703,
0.020840227603912354,
0.05808357894420624,
0.000791157886851579,
0.04904049262404442,
-0.02888050116598606,
0.10209730267524719,
-0.0407399944961071,
0.05318994075059891,
-0.0556655116379261,
0.027730422094464302,
-0.08754362910985947,
0.09051885455846786,
0.013043004088103771,
-0.04465194419026375,
-0.04339583218097687,
0.03287921100854874,
0.030296454206109047,
-0.05288666486740112,
-0.028239067643880844,
0.0913156270980835,
0.012666203081607819,
-0.043257154524326324,
-0.03320806473493576,
0.10548173636198044,
0.007898897863924503,
-0.08067218959331512,
-0.041778311133384705,
0.02800595387816429,
-0.08710843324661255,
-0.028319934383034706,
-0.09835647791624069,
-0.043693140149116516,
-0.009159645065665245,
0.0667167380452156,
-0.035927969962358475,
-0.03288379684090614,
-0.07249104976654053,
-0.014752158895134926,
-0.07964587211608887,
-0.013009228743612766,
-0.08996972441673279,
-0.03912271186709404,
0.0863254964351654,
0.00663553923368454,
-0.02344733476638794,
0.009668934158980846,
0.10172048211097717,
-0.04079589992761612,
0.04296981543302536,
-0.04218041151762009,
-0.04190004989504814,
-0.07227291166782379,
0.04507959634065628,
-0.047741711139678955,
0.043697603046894073,
-0.03340279310941696,
0.057676609605550766,
-0.008711413480341434,
0.037036363035440445,
-0.008821227587759495,
0.006980113219469786,
-0.03417925909161568,
-0.08015633374452591,
-0.04472882300615311,
-0.016614781692624092,
-0.001784941996447742,
-0.017225315794348717,
0.058457378298044205,
-0.004467916674911976,
0.05442756414413452,
0.06727855652570724,
0.09620816260576248,
0.026508310809731483,
-0.024325966835021973,
0.024911941960453987,
0.007099522277712822,
-0.013458248227834702,
0.07576017826795578,
-0.07425959408283234,
-4.663458597065073e-8,
-0.05251971259713173,
0.06745477020740509,
0.01698499359190464,
0.05490504950284958,
-0.05159110575914383,
-0.02287614718079567,
-0.03663884103298187,
0.0632244423031807,
-0.009176049381494522,
0.03230378031730652,
0.06638485938310623,
0.018657159060239792,
-0.04351431876420975,
-0.055110298097133636,
0.0828239694237709,
0.10496455430984497,
0.04172811657190323,
0.06396668404340744,
-0.04382110759615898,
-0.10420295596122742,
0.07324869930744171,
0.0595356747508049,
0.011831635609269142,
0.012918202206492424,
-0.02289574407041073,
0.020376885309815407,
-0.07063287496566772,
0.13973557949066162,
0.03178425505757332,
0.014971813187003136,
0.051599275320768356,
-0.026293087750673294,
-0.07063513249158859,
-0.05295988917350769,
-0.011616927571594715,
0.07540354877710342,
0.08017036318778992,
-0.008665673434734344,
0.06103510037064552,
0.027075881138443947,
0.011444968171417713,
0.031004566699266434,
-0.04349095746874809,
0.10190637409687042,
-0.009792722761631012,
-0.009237638674676418,
-0.05605313926935196,
-0.06395427882671356,
0.03470524027943611,
-0.10652618855237961,
0.020382951945066452,
-0.03243120014667511,
-0.06993304193019867,
0.07562330365180969,
0.035116180777549744,
0.0342252142727375,
0.03310596942901611,
-0.01889825239777565,
-0.011480830609798431,
0.04551911726593971,
0.04518018662929535,
0.008076593279838562,
-0.021704107522964478,
0.03223758563399315
] | 0 | 0 | 0 | 0 |
|
Provable Uncertainty Decomposition via Higher-Order Calibration | https://openreview.net/forum?id=TId1SHe8JG | [
"Gustaf Ahdritz",
"Aravind Gollakota",
"Parikshit Gopalan",
"Charlotte Peale",
"Udi Wieder"
] | Spotlight | We give a principled method for decomposing the predictive uncertainty of a model into aleatoric and epistemic components with explicit semantics relating them to the real-world data distribution. While many works in the literature have proposed such decompositions, they lack the type of formal guarantees we provide. Our method is based on the new notion of higher-order calibration, which generalizes ordinary calibration to the setting of higher-order predictors that predict _mixtures_ over label distributions at every point. We show how to measure as well as achieve higher-order calibration using access to $k$-snapshots, namely examples where each point has $k$ independent conditional labels. Under higher-order calibration, the estimated aleatoric uncertainty at a point is guaranteed to match the real-world aleatoric uncertainty averaged over all points where the prediction is made. To our knowledge, this is the first formal guarantee of this type that places no assumptions whatsoever on the real-world data distribution. Importantly, higher-order calibration is also applicable to existing higher-order predictors such as Bayesian and ensemble models and provides a natural evaluation metric for such models. We demonstrate through experiments that our method produces meaningful uncertainty decompositions in tasks such as image classification. | uncertainty quantification, calibration, trustworthy ML, mixture learning | We provide provable guarantees for uncertainty decomposition using the new notion of higher-order calibration. | 12,521 | 2412.18808 | [
-0.08355963975191116,
-0.03685460612177849,
-0.028970103710889816,
0.027181874960660934,
0.09379154443740845,
-0.010251418687403202,
0.04993445426225662,
-0.033267129212617874,
0.0625305250287056,
-0.026068398728966713,
-0.003717592917382717,
-0.09043607115745544,
0.12841790914535522,
0.04059317708015442,
-0.01743893325328827,
-0.0781828910112381,
0.04141337797045708,
0.12506210803985596,
-0.08893153816461563,
-0.0201973058283329,
-0.008995878510177135,
0.05003870651125908,
-0.05968547612428665,
0.04823810234665871,
0.10986808687448502,
-0.015164042823016644,
-0.03516450524330139,
0.01197352446615696,
-0.03535974770784378,
0.0006283199181780219,
0.028628600761294365,
-0.010020866058766842,
0.02466949261724949,
-0.05092795938253403,
0.032244063913822174,
-0.03675982728600502,
-0.057506341487169266,
0.01183183491230011,
0.04280906915664673,
0.030003104358911514,
0.006627941969782114,
0.007917561568319798,
0.016626358032226562,
0.04792146012187004,
-0.047729041427373886,
0.0785105749964714,
0.07201920449733734,
-0.04161021485924721,
-0.03198607638478279,
0.06103513017296791,
-0.01907353289425373,
0.007160663604736328,
-0.03857050836086273,
-0.0014962995192036033,
-0.051526013761758804,
0.04788326844573021,
0.01284337043762207,
-0.06476827710866928,
-0.010423123836517334,
0.01405114121735096,
-0.07839392125606537,
-0.04395502060651779,
-0.01874380372464657,
-0.00846280250698328,
0.02986106090247631,
0.012999973259866238,
0.050176575779914856,
0.03439556807279587,
0.04433150216937065,
0.1340324431657791,
-0.03733876720070839,
0.09178712964057922,
-0.014090901240706444,
-0.00484516192227602,
-0.014743437990546227,
-0.010984900407493114,
0.0907788872718811,
-0.015647077932953835,
-0.015917889773845673,
0.01570008136332035,
-0.03963668644428253,
0.06462924927473068,
-0.01013513095676899,
-0.031402695924043655,
0.07336228340864182,
-0.0009767848532646894,
0.012407657690346241,
0.03769677132368088,
0.06068579480051994,
0.012768919579684734,
-0.1238546073436737,
-0.04022445157170296,
0.014350298792123795,
0.02866010181605816,
0.023530568927526474,
0.03548937663435936,
0.05591528117656708,
-0.09564044326543808,
0.11369014531373978,
0.06204888969659805,
-0.08657358586788177,
-0.024534694850444794,
0.09217791259288788,
-0.11430689692497253,
0.03926187381148338,
-0.1090955138206482,
-0.05020776018500328,
-0.012701605446636677,
-0.026053957641124725,
-0.025488438084721565,
0.028835497796535492,
-0.03925754874944687,
0.02407148852944374,
-0.03197960928082466,
0.03461478278040886,
0.022221248596906662,
0.003945742268115282,
-0.015582136809825897,
0.025623708963394165,
-0.08251429349184036,
-0.08311021327972412,
-0.009640371426939964,
0.09377149492502213,
0.06931605935096741,
0.04458325356245041,
0.02466624416410923,
0.030487118288874626,
4.6854999112264986e-33,
0.030555689707398415,
0.01424678135663271,
0.06310723721981049,
0.002594067482277751,
0.03360689431428909,
0.0516088493168354,
-0.03508073836565018,
-0.04476839676499367,
0.0533490888774395,
0.00769532285630703,
0.06714769452810287,
0.04111015051603317,
-0.04466474428772926,
0.008775533176958561,
0.10214118659496307,
0.0561731681227684,
-0.08045800030231476,
0.06515426188707352,
-0.07443641126155853,
0.026694294065237045,
0.06890571862459183,
-0.07940803468227386,
-0.06954348087310791,
-0.037744276225566864,
-0.06610768288373947,
0.032839249819517136,
0.028501396998763084,
0.0881100594997406,
-0.05883817747235298,
-0.02497600018978119,
-0.0644318014383316,
-0.023877916857600212,
0.04736294597387314,
-0.033521030098199844,
0.006393016315996647,
-0.06169271469116211,
-0.04650092124938965,
-0.011038865894079208,
-0.07388965785503387,
0.0030358440708369017,
0.025759628042578697,
0.07344114035367966,
-0.02041742019355297,
0.009366949088871479,
-0.01459361519664526,
-0.026143064722418785,
-0.04851285368204117,
0.06268081068992615,
0.030497686937451363,
-0.08634303510189056,
0.010004879906773567,
-0.027150020003318787,
0.015513596124947071,
-0.033629655838012695,
-0.041254278272390366,
-0.010762542486190796,
-0.014484535902738571,
-0.0274905227124691,
-0.00881884153932333,
-0.015794623643159866,
-0.0705452635884285,
0.00753090251237154,
0.09171445667743683,
0.04230564087629318,
0.010182447731494904,
0.09878654778003693,
-0.023543840274214745,
0.02052524872124195,
0.07723474502563477,
-0.038936808705329895,
-0.07633347809314728,
-0.0022510490380227566,
0.01470175851136446,
-0.06530356407165527,
0.06260278820991516,
0.03640828654170036,
-0.035084158182144165,
-0.044662922620773315,
0.042170971632003784,
-0.011702744290232658,
-0.023101335391402245,
0.03754313662648201,
0.03146633133292198,
0.026812555268406868,
0.03697160258889198,
-0.03616835176944733,
-0.025391899049282074,
0.04355306923389435,
-0.1071881428360939,
-0.027969839051365852,
-0.13682106137275696,
-0.009503636509180069,
-0.006286452524363995,
-0.014640271663665771,
-0.013658070005476475,
-4.94530840732921e-33,
-0.028158707544207573,
-0.017102226614952087,
0.0571412667632103,
0.09489230066537857,
-0.03786509856581688,
-0.06255730241537094,
-0.04953274875879288,
0.08179871737957001,
-0.007750667165964842,
-0.08194982260465622,
-0.09201284497976303,
-0.010706202127039433,
0.06247282773256302,
0.024927951395511627,
0.02102462388575077,
0.028762172907590866,
-0.02689269371330738,
-0.0033133686520159245,
-0.05596635118126869,
-0.004099877085536718,
0.0344778336584568,
-0.046593230217695236,
-0.04719671234488487,
0.04704504832625389,
0.000585482397582382,
0.00554853118956089,
-0.03597377613186836,
0.07857675105333328,
-0.01787620410323143,
-0.03948233649134636,
-0.03395650163292885,
-0.056441884487867355,
0.04705837368965149,
-0.03240535780787468,
-0.09906432032585144,
-0.024920757859945297,
-0.004353481810539961,
0.04415690898895264,
-0.0980430319905281,
0.04963454231619835,
-0.007585579063743353,
-0.018993815407156944,
-0.1282138228416443,
-0.06958352774381638,
0.021962536498904228,
-0.01233627274632454,
0.0703795850276947,
-0.06131931394338608,
-0.00040807612822391093,
0.01141088642179966,
-0.02703470177948475,
0.06528564542531967,
-0.09418412297964096,
0.09178924560546875,
-0.03421255946159363,
-0.018077148124575615,
-0.13178236782550812,
-0.013794180005788803,
0.052001383155584335,
0.005765121430158615,
-0.03373163938522339,
-0.024581192061305046,
-0.005967363715171814,
0.07047093659639359,
0.027100248262286186,
-0.026637814939022064,
-0.034529320895671844,
-0.05704803019762039,
0.025791997089982033,
0.011132451705634594,
0.08394567668437958,
-0.056020162999629974,
0.02270614169538021,
0.04784485697746277,
0.04116731509566307,
-0.022505836561322212,
0.11303442716598511,
0.06003224849700928,
0.03942270576953888,
0.037212103605270386,
0.027922537177801132,
0.0054235574789345264,
0.07927709817886353,
0.09134943783283234,
0.06494681537151337,
0.008396538905799389,
0.03874807804822922,
-0.03379553556442261,
0.07003454864025116,
-0.01280173659324646,
-0.030240526422858238,
0.03789335489273071,
-0.05153568089008331,
0.08483488857746124,
-0.01511032972484827,
-4.7727350960258264e-8,
-0.03736598417162895,
0.05727046728134155,
-0.03144055977463722,
-0.009047377854585648,
0.012468220666050911,
0.032732877880334854,
-0.02857975661754608,
-0.0031603833194822073,
0.005223400890827179,
-0.06467688828706741,
0.12331397086381912,
-0.012869643978774548,
-0.031216643750667572,
-0.021304680034518242,
-0.027243558317422867,
0.05333179607987404,
0.013775957748293877,
0.029830316081643105,
-0.03315731883049011,
0.0034359891433268785,
-0.004713649395853281,
-0.006510505452752113,
0.0406319759786129,
-0.05999809503555298,
0.05591404438018799,
-0.04163585230708122,
0.030505992472171783,
0.11230337619781494,
0.03314876928925514,
0.05731616169214249,
0.0026727744843810797,
-0.02876562811434269,
0.03554953634738922,
0.022489814087748528,
0.007764651905745268,
0.1287037879228592,
-0.05029937997460365,
0.015397588722407818,
0.025685781612992287,
-0.03480597585439682,
-0.04961738735437393,
0.0015248296549543738,
-0.13536515831947327,
0.014853359200060368,
0.10576880723237991,
-0.012812767177820206,
-0.007722888141870499,
-0.04127548635005951,
0.005100382026284933,
-0.012255847454071045,
-0.051081810146570206,
0.05503249913454056,
-0.014614822342991829,
0.028192317113280296,
0.010744485072791576,
0.019700178876519203,
-0.05696287006139755,
-0.07817040383815765,
-0.01381088700145483,
0.017240559682250023,
0.04472614452242851,
-0.03181175887584686,
-0.05461699888110161,
-0.01868532970547676
] | 0 | 0 | 0 | 0 |
|
Sparse components distinguish visual pathways & their alignment to neural networks | https://openreview.net/forum?id=IqHeDe2lbl | [
"Ammar I Marvi",
"Nancy Kanwisher",
"Meenakshi Khosla"
] | Spotlight | The ventral, dorsal, and lateral streams in high-level human visual cortex are implicated in distinct functional processes. Yet, deep neural networks (DNNs) trained on a single task model the entire visual system surprisingly well, hinting at common computational principles across these pathways. To explore this inconsistency, we applied a novel sparse decomposition approach to identify the dominant components of visual representations within each stream. Consistent with traditional neuroscience research, we find a clear difference in component response profiles across the three visual streams—identifying components selective for faces, places, bodies, text, and food in the ventral stream; social interactions, implied motion, and hand actions in the lateral stream; and some less interpretable components in the dorsal stream. Building on this, we introduce Sparse Component Alignment (SCA), a new method for measuring representational alignment between brains and machines that better captures the latent neural tuning of these two visual systems. We find that standard visual DNNs are more aligned with ventral than either dorsal or lateral representations. SCA reveals these distinctions with greater resolution than conventional population-level geometry, offering a measure of representational alignment that is sensitive to a system’s underlying axes of neural tuning. | visual representations, alignment, sparse decomposition, neural pathways, brain and machine vision | null | 12,453 | null | [
0.021653763949871063,
-0.12681451439857483,
0.04419200122356415,
0.03287399560213089,
0.0189824141561985,
0.026841355487704277,
-0.028528224676847458,
-0.03476070985198021,
0.08020776510238647,
-0.056197118014097214,
-0.0530971996486187,
-0.06902667880058289,
-0.029138989746570587,
0.06610006839036942,
-0.05263454467058182,
-0.1063854768872261,
0.06619058549404144,
0.021438010036945343,
-0.009866626933217049,
0.04180598631501198,
0.004916056524962187,
-0.03638362139463425,
0.03517827391624451,
0.024101467803120613,
0.023038893938064575,
0.05264541506767273,
0.021610775962471962,
-0.011763805523514748,
0.0028990970458835363,
-0.06199661269783974,
0.06614761799573898,
0.03763723745942116,
-0.0074221170507371426,
-0.00696956692263484,
-0.041731446981430054,
0.11499907076358795,
0.026786791160702705,
0.027943842113018036,
0.004887630697339773,
-0.07220546156167984,
0.041327375918626785,
0.05349091440439224,
0.003907612059265375,
0.10102012753486633,
0.06032905727624893,
0.07798450440168381,
-0.018425796180963516,
-0.057253722101449966,
-0.02946673333644867,
-0.06580915302038193,
-0.029697490856051445,
0.021426230669021606,
0.0289043877273798,
0.11148487031459808,
-0.00777203356847167,
0.05266972631216049,
0.007455869112163782,
0.004992627073079348,
-0.026355650275945663,
0.044200461357831955,
0.02028971165418625,
-0.027515467256307602,
0.02258291095495224,
-0.0028866457287222147,
0.06559724360704422,
0.031505830585956573,
0.023216858506202698,
-0.026311179623007774,
-0.08186795562505722,
0.012492171488702297,
-0.018324291333556175,
0.05358557775616646,
-0.03994838148355484,
0.009261463768780231,
0.026725932955741882,
-0.019328277558088303,
0.007645764388144016,
-0.03862909600138664,
0.05082850530743599,
-0.03207870200276375,
0.02573424205183983,
0.07594484090805054,
-0.04472624137997627,
0.028025837615132332,
0.05784182623028755,
-0.0131472023203969,
-0.03346283361315727,
-0.006970119196921587,
-0.028718536719679832,
-0.0022213647607713938,
-0.050960347056388855,
-0.021383406594395638,
0.01731325127184391,
-0.08820006996393204,
-0.015491313301026821,
0.07665520161390305,
0.05495225265622139,
0.0190635547041893,
-0.0033654291182756424,
0.05390830338001251,
-0.02453068643808365,
-0.040242522954940796,
0.01916438341140747,
0.060104403644800186,
0.020586051046848297,
0.0011329128174111247,
0.04332519695162773,
0.03225922957062721,
0.009809688664972782,
-0.10408779978752136,
-0.016089443117380142,
-0.028872769325971603,
-0.03190291300415993,
0.12579643726348877,
0.014233666472136974,
-0.11321191489696503,
0.006077569909393787,
-0.06891059875488281,
0.09186014533042908,
0.010003888048231602,
-0.0376967117190361,
-0.03166301175951958,
0.0022132014855742455,
-0.016751598566770554,
0.0754011794924736,
0.0041086310520768166,
-0.07972905784845352,
2.0309952621217473e-33,
0.0025439278688281775,
0.04661901667714119,
0.08418258279561996,
0.029688511043787003,
0.028281765058636665,
-0.027464885264635086,
0.005565928295254707,
0.03280381113290787,
0.026379423215985298,
0.09397124499082565,
-0.12451140582561493,
0.001170547679066658,
-0.008738573640584946,
0.13813695311546326,
-0.03265252336859703,
-0.02965419553220272,
-0.03663129359483719,
0.046231143176555634,
-0.06277304887771606,
-0.04354792833328247,
-0.013410599902272224,
-0.008760704658925533,
-0.023643501102924347,
-0.0688663050532341,
-0.03033408708870411,
-0.09993806481361389,
-0.05852249264717102,
-0.08366668224334717,
-0.09427453577518463,
-0.017657632008194923,
-0.0842406377196312,
0.07304863631725311,
0.029927680268883705,
0.001236838405020535,
0.018720604479312897,
-0.04977438971400261,
0.058405645191669464,
-0.04187498986721039,
0.03126705065369606,
0.02446369268000126,
0.022679593414068222,
0.055044326931238174,
-0.06390336900949478,
-0.07181793451309204,
-0.05338546261191368,
0.019910363480448723,
0.05764865502715111,
0.08166813850402832,
-0.03497673198580742,
-0.029271958395838737,
0.05085122957825661,
0.06482381373643875,
0.08904556185007095,
-0.05142562836408615,
-0.0021491728257387877,
0.002369316527619958,
0.003615549299865961,
-0.04593905061483383,
0.031094014644622803,
-0.029805976897478104,
0.008211473934352398,
0.004198603332042694,
-0.03230820596218109,
0.047885701060295105,
0.08038663119077682,
0.05113328620791435,
-0.02118781767785549,
0.04146046191453934,
0.05406608432531357,
-0.010279856622219086,
-0.027094902470707893,
0.07509694248437881,
-0.027148231863975525,
-0.07519655674695969,
0.049815405160188675,
-0.011434128507971764,
0.09859592467546463,
-0.05238833278417587,
-0.045815352350473404,
0.05244174972176552,
-0.032976988703012466,
0.00872642733156681,
-0.04152941703796387,
-0.030614279210567474,
-0.0677393302321434,
0.018201977014541626,
0.09443158656358719,
-0.0833524763584137,
-0.04427485913038254,
-0.022558074444532394,
-0.02304830588400364,
-0.08417556434869766,
0.03473931923508644,
-0.09932249784469604,
-0.014297920279204845,
-1.1034089964680339e-33,
-0.05420269817113876,
0.00766001595184207,
0.02303166873753071,
-0.05850188434123993,
-0.027151472866535187,
-0.01679214835166931,
0.01935865357518196,
-0.015485654585063457,
0.026420557871460915,
-0.014676270075142384,
0.029714951291680336,
-0.030670607462525368,
-0.027068832889199257,
0.011672289110720158,
0.044136010110378265,
0.05325430631637573,
-0.04792772978544235,
0.032410379499197006,
0.07494629919528961,
-0.09555396437644958,
0.03829566761851311,
0.07112663239240646,
-0.04673287272453308,
-0.016000857576727867,
-0.016314692795276642,
0.03216107562184334,
0.07312514632940292,
0.07178516685962677,
-0.02148248627781868,
-0.025704555213451385,
-0.019196517765522003,
0.006694663781672716,
-0.06373375654220581,
-0.06382451206445694,
-0.01775338128209114,
0.08282017707824707,
-0.05365679785609245,
-0.0031682513654232025,
-0.013486338779330254,
0.0109693119302392,
0.04211116209626198,
0.006945458240807056,
-0.05314624682068825,
0.06469504535198212,
0.010942598804831505,
-0.05658261105418205,
-0.05832314491271973,
-0.022028479725122452,
-0.06823135912418365,
0.1459549367427826,
0.0021724628750234842,
0.07599254697561264,
-0.09708873182535172,
-0.01617533341050148,
-0.010377254337072372,
-0.07229392975568771,
0.054748114198446274,
0.01740284450352192,
0.09057992696762085,
-0.03476667404174805,
0.005012383684515953,
0.015355902723968029,
-0.10716063529253006,
-0.015116467140614986,
0.02355770952999592,
0.02435281313955784,
0.005233307834714651,
0.02264692820608616,
0.059001702815294266,
-0.01700582541525364,
0.029220687225461006,
-0.08641994744539261,
0.01974884420633316,
0.004788040183484554,
0.04132528230547905,
-0.021687550470232964,
0.01111784391105175,
0.03838686645030975,
-0.021770304068922997,
0.020156143233180046,
0.02716638520359993,
-0.02475447952747345,
-0.057192616164684296,
0.02494102530181408,
0.024073507636785507,
0.10811507701873779,
0.0026314756833016872,
-0.029466688632965088,
0.02439919486641884,
0.03441883996129036,
-0.047356151044368744,
-0.013230429962277412,
-0.020902233198285103,
-0.00578932324424386,
-0.06756245344877243,
-4.220262184162493e-8,
-0.1157892718911171,
0.009314012713730335,
-0.012204269878566265,
-0.01831188052892685,
0.00804024189710617,
0.0002404344268143177,
0.038608238101005554,
0.061422597616910934,
-0.042618487030267715,
-0.033432431519031525,
0.09206864982843399,
-0.0468582846224308,
-0.004435029812157154,
-0.06675569713115692,
0.08711344003677368,
0.05595903843641281,
-0.035214994102716446,
0.020919442176818848,
-0.0748053640127182,
0.05917494371533394,
0.0461973138153553,
0.04071148484945297,
-0.010748378001153469,
-0.014512582682073116,
0.07088813185691833,
-0.10027477145195007,
-0.09061381965875626,
0.1544799655675888,
0.025010284036397934,
-0.05058087781071663,
0.04746323078870773,
0.07251035422086716,
0.04965676739811897,
-0.03734942525625229,
0.026529641821980476,
0.08069625496864319,
0.015343205071985722,
0.027056392282247543,
-0.0808117687702179,
0.0005553997470997274,
-0.09195073693990707,
0.024299608543515205,
0.022916821762919426,
0.025132624432444572,
-0.0021950015798211098,
0.02423791028559208,
-0.025839222595095634,
-0.03661486506462097,
0.1042763888835907,
-0.029558101668953896,
-0.06742795556783676,
0.018300678580999374,
0.01146368682384491,
0.1442473828792572,
-0.02109343372285366,
-0.022149894386529922,
0.014635404571890831,
-0.011867308989167213,
0.016868410632014275,
0.015799667686223984,
0.06498780846595764,
0.11476943641901016,
-0.0750790610909462,
-0.023736774921417236
] | 0 | 0 | 0 | 0 |
|
Reducing Hallucinations in Large Vision-Language Models via Latent Space Steering | https://openreview.net/forum?id=LBl7Hez0fF | [
"Sheng Liu",
"Haotian Ye",
"James Zou"
] | Spotlight | Hallucination poses a challenge to the deployment of large vision-language models (LVLMs) in applications. Unlike in large language models (LLMs), hallucination in LVLMs often arises from misalignments between visual inputs and textual outputs. This paper investigates the underlying mechanisms of hallucination, focusing on the unique structure of LVLMs that distinguishes them from LLMs. We identify that hallucinations often arise from the sensitivity of text decoders to vision inputs, a natural phenomenon when image encoders and text decoders are pre-trained separately. Inspired by this, we introduce Visual and Textual Intervention (VTI), a novel technique designed to reduce hallucinations by steering latent space representations during inference to enhance the stability of vision features. As a task-agnostic test-time intervention, VTI can be easily applied to any problem without additional training costs. Extensive experiments demonstrate that it can effectively reduce hallucinations and outperform baseline methods across multiple metrics, highlighting the critical role of vision feature stability in LVLMs. | Large Vision-Language Models, Multimodal large language model, Hallucination | null | 12,379 | null | [
0.03146817907691002,
-0.06212790310382843,
0.03824356570839882,
0.00409964332357049,
0.06582603603601456,
0.058127060532569885,
-0.0791563019156456,
-0.048144664615392685,
0.07931016385555267,
-0.07610389590263367,
-0.04166651517152786,
-0.03349217772483826,
0.030713316053152084,
0.042049020528793335,
-0.01563897356390953,
-0.008223719894886017,
0.038080550730228424,
0.09036336839199066,
-0.07821008563041687,
-0.04618595167994499,
0.07543168216943741,
-0.024444999173283577,
0.026531167328357697,
-0.004337576683610678,
0.07674413919448853,
0.048380427062511444,
-0.04027664661407471,
-0.04432469978928566,
-0.012609959580004215,
-0.029796887189149857,
-0.002457366557791829,
0.06735806167125702,
-0.014388425275683403,
-0.033645376563072205,
-0.031155060976743698,
0.053404223173856735,
-0.06886526942253113,
0.04022900015115738,
-0.021523142233490944,
-0.051837071776390076,
-0.03631060943007469,
0.014734392054378986,
0.040597230195999146,
0.04578346014022827,
0.005005212035030127,
-0.044057995080947876,
-0.006646275520324707,
-0.006912234239280224,
-0.018604209646582603,
-0.056206926703453064,
-0.03696175664663315,
0.009945878759026527,
0.02737489901483059,
0.015524955466389656,
-0.05626002699136734,
0.028658172115683556,
0.09507209062576294,
0.07134249061346054,
0.024737561121582985,
-0.008818338625133038,
-0.014013177715241909,
-0.03992685675621033,
0.011145737953484058,
-0.015914414077997208,
-0.04425937682390213,
0.030991947278380394,
0.024278519675135612,
0.027286425232887268,
-0.03248198702931404,
0.010676693171262741,
-0.029984839260578156,
0.046580832451581955,
-0.05224744975566864,
0.06582880765199661,
0.04854576662182808,
0.036855146288871765,
-0.0251607708632946,
-0.1243610754609108,
0.09271848201751709,
-0.04346771910786629,
0.028458986431360245,
0.08054701238870621,
0.03464903309941292,
0.022584279999136925,
0.09837178885936737,
0.0448271818459034,
-0.003228104906156659,
0.042538948357105255,
-0.012514608912169933,
0.027144111692905426,
-0.05278382822871208,
-0.18008926510810852,
-0.0109469685703516,
0.013454957865178585,
-0.01918896660208702,
-0.029059631749987602,
-0.0366082638502121,
-0.05495794489979744,
-0.01986290141940117,
0.015798352658748627,
0.04203403368592262,
-0.048127975314855576,
-0.0011532569769769907,
-0.055890537798404694,
0.03103797324001789,
-0.04098550230264664,
0.16646713018417358,
0.051560234278440475,
0.01797506958246231,
-0.012169904075562954,
0.03653965890407562,
0.03047291375696659,
-0.030086496844887733,
0.05420827120542526,
-0.013483456335961819,
-0.005651154089719057,
0.019675185903906822,
-0.027278432622551918,
0.04686744511127472,
0.050892964005470276,
0.04433296248316765,
0.012664216570556164,
0.022632109001278877,
-0.04943966493010521,
0.044874049723148346,
-0.07539466768503189,
-0.071052685379982,
2.468804459665662e-34,
0.005542182829231024,
-0.004955641459673643,
0.025729266926646233,
0.028761552646756172,
0.08240992575883865,
-0.023483391851186752,
-0.07643217593431473,
0.052830491214990616,
0.02853536233305931,
-0.015445316210389137,
-0.03402513265609741,
-0.057011254131793976,
-0.01838492415845394,
0.06452147662639618,
0.014057429507374763,
0.023648297414183617,
-0.08072299510240555,
0.05346919596195221,
-0.07371270656585693,
-0.009054209105670452,
-0.018825802952051163,
-0.05053909868001938,
-0.012998169288039207,
0.016786156222224236,
0.0001382655173074454,
0.035208653658628464,
0.020776236429810524,
-0.08303636312484741,
-0.012588695622980595,
0.01035163551568985,
-0.15010838210582733,
0.07396702468395233,
0.04015018418431282,
0.015783023089170456,
0.017409080639481544,
0.0040747723542153835,
0.040153078734874725,
-0.04791518673300743,
0.020938120782375336,
-0.039926979690790176,
-0.025088422000408173,
0.08187021315097809,
-0.011884830892086029,
-0.059326931834220886,
-0.03940378502011299,
-0.007700870279222727,
0.049374036490917206,
0.006521799135953188,
-0.1121504157781601,
-0.007261637132614851,
0.04488058015704155,
0.08310984075069427,
-0.035892292857170105,
-0.03307303041219711,
-0.028573455289006233,
-0.02413961850106716,
0.04337603598833084,
-0.03201421722769737,
0.017391161993145943,
-0.015533548779785633,
0.04489033296704292,
0.079579658806324,
0.03532349690794945,
0.09250207990407944,
0.010395468212664127,
-0.07167388498783112,
0.04545632377266884,
-0.02824396826326847,
-0.03381137549877167,
-0.0853266716003418,
-0.08606630563735962,
-0.025123750790953636,
-0.061378270387649536,
0.012805953621864319,
-0.022639110684394836,
-0.031987257301807404,
0.059170693159103394,
-0.05627472326159477,
-0.008160465396940708,
0.014569326303899288,
0.02934594266116619,
-0.09343471378087997,
0.007584436796605587,
-0.07376154512166977,
-0.02130828984081745,
-0.06367585808038712,
-0.011624748818576336,
-0.12342449277639389,
-0.06829571723937988,
-0.07450493425130844,
-0.021700045093894005,
-0.04701414704322815,
0.029496094211935997,
-0.015053551644086838,
-0.0196013692766428,
-1.5995467747127274e-33,
0.0031067535746842623,
-0.017306823283433914,
0.09576449543237686,
0.003167468588799238,
-0.008529708720743656,
-0.03484487906098366,
0.06557176262140274,
0.0022987443953752518,
0.0543757826089859,
-0.06404709815979004,
-0.03523695468902588,
-0.007060293573886156,
-0.005859497468918562,
-0.0077478401362895966,
0.05302475020289421,
-0.10189516097307205,
0.06089463829994202,
0.009744197130203247,
-0.03630058467388153,
0.02699868008494377,
0.0479300282895565,
0.07089004665613174,
-0.0685315951704979,
-0.06218826025724411,
-0.03278886154294014,
0.04688606411218643,
0.010060913860797882,
0.06747395545244217,
0.019445106387138367,
-0.01899305358529091,
0.004630169365555048,
0.03875596821308136,
-0.01530985813587904,
-0.018858496099710464,
-0.019009705632925034,
0.07664085924625397,
0.022591890767216682,
-0.018671825528144836,
-0.10408973693847656,
0.06676734238862991,
0.09032265841960907,
-0.010304625146090984,
-0.009701931849122047,
0.06158694252371788,
-0.058836884796619415,
-0.031213922426104546,
-0.02435903437435627,
-0.03951234370470047,
0.01968667097389698,
0.0580834224820137,
-0.038086216896772385,
0.0017945080762729049,
-0.06296675652265549,
-0.014774968847632408,
-0.1146889328956604,
-0.07447753846645355,
-0.09582540392875671,
0.03519917652010918,
0.09060779213905334,
0.005282609257847071,
-0.03782501071691513,
0.029084067791700363,
-0.021945487707853317,
-0.03742897883057594,
-0.0246212650090456,
-0.04056159034371376,
-0.0631852000951767,
0.05395200476050377,
0.06486871093511581,
0.03892526030540466,
0.06701505184173584,
-0.021519089117646217,
-0.01921072043478489,
0.10577960312366486,
0.0175867211073637,
-0.019995246082544327,
0.010207680985331535,
-0.03898816183209419,
0.007007272448390722,
-0.09727111458778381,
0.011455311439931393,
-0.04199257493019104,
0.031992796808481216,
0.11152835935354233,
0.06156699359416962,
0.10208313912153244,
-0.0008221553871408105,
0.04751608520746231,
0.0037505056243389845,
0.032577771693468094,
-0.04629233852028847,
0.03962422534823418,
0.016716163605451584,
0.06819047033786774,
-0.03546034172177315,
-3.5641924966967053e-8,
-0.10496817529201508,
-0.01901114173233509,
-0.04564221575856209,
0.017338894307613373,
-0.023455150425434113,
-0.09436371177434921,
-0.029408590868115425,
0.11707363277673721,
-0.035443998873233795,
-0.0694701224565506,
0.06099971383810043,
-0.02611960843205452,
-0.013899361714720726,
-0.10704459249973297,
0.00526403309777379,
0.05469866842031479,
-0.0017234610859304667,
0.09711454063653946,
-0.01620286889374256,
-0.021674908697605133,
0.05958205461502075,
0.047369908541440964,
-0.007965587079524994,
-0.06976248323917389,
-0.009651746600866318,
0.028281625360250473,
-0.0023173305671662092,
0.0948975458741188,
0.033975835889577866,
-0.06269115209579468,
0.14231836795806885,
0.088779017329216,
0.01110970787703991,
0.012168643064796925,
-0.044079434126615524,
0.045353882014751434,
-0.05694638565182686,
-0.006903049070388079,
0.04361915588378906,
-0.018444174900650978,
0.04284252971410751,
0.02613821066915989,
-0.011193581856787205,
0.019496183842420578,
-0.04754781350493431,
-0.011253368109464645,
0.07762370258569717,
-0.08976812660694122,
0.02407960779964924,
-0.04637500271201134,
-0.040385060012340546,
0.04205621778964996,
0.01109706424176693,
0.12039116024971008,
0.07778403162956238,
0.012540452182292938,
0.10801657289266586,
-0.004237103741616011,
0.014707440510392189,
0.024754947051405907,
0.013803323730826378,
0.009785417467355728,
-0.054706595838069916,
-0.017270402982831
] | 0 | 0 | 0 | 0 |
|
Mitigating Memorization in Language Models | https://openreview.net/forum?id=MGKDBuyv4p | [
"Mansi Sakarvadia",
"Aswathy Ajith",
"Arham Mushtaq Khan",
"Nathaniel C Hudson",
"Caleb Geniesse",
"Kyle Chard",
"Yaoqing Yang",
"Ian Foster",
"Michael W. Mahoney"
] | Spotlight | Language models (LMs) can “memorize” information, i.e., encode training data in their weights in such a way that inference-time queries can lead to verbatim regurgitation of that data. This ability to extract training data can be problematic, for example, when data are private or sensitive. In this work, we investigate methods to mitigate memorization: three regularizer-based, three fine-tuning-based, and eleven machine unlearning-based methods, with five of the latter being new methods that we introduce. We also introduce TinyMem, a suite of small, computationally-efficient LMs for the rapid development and evaluation of memorization-mitigation methods. We demonstrate that the mitigation methods that we develop using TinyMem can successfully be applied to production-grade LMs, and we determine via experiment that: regularizer-based mitigation methods are slow and ineffective at curbing memorization; fine-tuning-based methods
are effective at curbing memorization, but overly expensive, especially for retaining higher accuracies; and unlearning-based methods are faster and more effective, allowing for the precise localization and removal of memorized information from LM weights prior to inference. We show, in particular, that our proposed unlearning method BalancedSubnet outperforms other mitigation methods at removing
memorized information while preserving performance on target tasks. | language models, memorization, machine unlearning, regularization, fine-tuning, natural language processing | We study memorization mitigation methods in LMs, introduce 5 new methods, and demonstrate that unlearning methods, particularly our method BalancedSubnet, are more effective than regularizer and fine-tuning approaches in removing memorization. | 12,369 | 2410.02159 | [
0.010992765426635742,
-0.02036140114068985,
0.011605501174926758,
0.07094498723745346,
0.05285363644361496,
0.046658311039209366,
0.0030539962463080883,
0.00019183187396265566,
-0.03701894357800484,
-0.02031511627137661,
0.007367035839706659,
0.0004841158224735409,
0.020412612706422806,
0.002026353031396866,
0.01446561049669981,
0.020756591111421585,
0.02567356824874878,
0.10632502287626266,
-0.12277762591838837,
-0.10958565771579742,
0.08554548025131226,
0.028384696692228317,
0.026555469259619713,
0.054989043623209,
0.06590985506772995,
-0.024324744939804077,
-0.005838827230036259,
-0.013515200465917587,
0.04468759894371033,
-0.027129314839839935,
-0.01596987433731556,
0.02677452191710472,
-0.029661599546670914,
0.03182610496878624,
-0.02157818153500557,
0.019912704825401306,
-0.07377542555332184,
0.006798295304179192,
-0.005428887438029051,
-0.05358993634581566,
-0.05970176309347153,
0.012634700164198875,
0.03835039585828781,
-0.0034831541124731302,
0.041710738092660904,
-0.006427074782550335,
0.012902664951980114,
-0.009304005652666092,
-0.04468701779842377,
0.0006176178576424718,
-0.04072590172290802,
0.0006951421382836998,
-0.02867603860795498,
0.025072596967220306,
-0.06733091920614243,
-0.013283555395901203,
0.019742820411920547,
0.056274719536304474,
-0.04749923199415207,
0.01831594668328762,
-0.0978277251124382,
-0.12016916275024414,
-0.07517176866531372,
0.0025045524816960096,
-0.10339450091123581,
-0.03136160224676132,
0.07234223932027817,
0.06442374736070633,
-0.009935559704899788,
0.08436867594718933,
-0.007504217326641083,
0.06578055769205093,
-0.06714029610157013,
0.05758345499634743,
0.03462442383170128,
0.04653574526309967,
-0.03899233788251877,
-0.04632100835442543,
0.06455381214618683,
-0.03655438870191574,
-0.03221717104315758,
-0.03828189894556999,
0.02613748423755169,
-0.007603735662996769,
0.11175515502691269,
0.024026663973927498,
0.06678369641304016,
0.007800733670592308,
0.06009936332702637,
-0.01526621263474226,
0.01952829211950302,
-0.053468745201826096,
0.005117303691804409,
0.009942635893821716,
-0.03628256916999817,
0.040407679975032806,
0.023087631911039352,
-0.016761740669608116,
-0.03937103971838951,
0.08526348322629929,
0.05580109730362892,
0.12932446599006653,
-0.020213276147842407,
0.004217938520014286,
-0.01434938982129097,
0.017303336411714554,
0.060898806899785995,
0.05864706635475159,
0.03321613743901253,
-0.03331322968006134,
0.015458501875400543,
0.11454036086797714,
-0.000400366960093379,
0.026779407635331154,
0.015263769775629044,
-0.03934257850050926,
0.06441150605678558,
-0.11648418754339218,
0.0007591289468109608,
0.09590386599302292,
-0.010150864720344543,
0.051793988794088364,
-0.009834488853812218,
0.014922672882676125,
0.018146490678191185,
-0.003422340378165245,
-0.06422705948352814,
1.4516452489728018e-33,
0.09112802147865295,
0.06989122927188873,
-0.06011316925287247,
-0.00795137882232666,
0.00833820179104805,
0.03509306162595749,
0.052452415227890015,
0.05065137520432472,
-0.0424591526389122,
0.00008649569645058364,
0.04716004431247711,
-0.022072117775678635,
-0.005482932552695274,
0.03504218906164169,
0.06428764015436172,
-0.013556722551584244,
-0.06147077679634094,
0.0751856192946434,
-0.01879408024251461,
-0.04102097079157829,
0.0264569278806448,
-0.01643437147140503,
0.05702124536037445,
-0.05145327374339104,
-0.0005701920599676669,
-0.02834007330238819,
0.0867193415760994,
-0.02890012226998806,
0.00017914906493388116,
0.03454837575554848,
-0.0630771815776825,
-0.07531565427780151,
-0.003187604248523712,
-0.009117349050939083,
0.03247340768575668,
-0.05306385084986687,
0.002233293605968356,
-0.06450661271810532,
0.051268745213747025,
-0.0433666966855526,
-0.036825042217969894,
0.0128089664503932,
0.04441341757774353,
-0.00629042461514473,
-0.08569738268852234,
-0.07783560454845428,
-0.022908920422196388,
-0.013694941066205502,
0.006601574830710888,
-0.08782734721899033,
0.07767558097839355,
0.02166380174458027,
-0.054591163992881775,
-0.053531333804130554,
0.010890068486332893,
0.023816194385290146,
0.06251701712608337,
-0.011464790441095829,
-0.02273447997868061,
0.018926506862044334,
0.00021952265524305403,
0.06406739354133606,
0.03963882476091385,
0.09546531736850739,
0.05381179228425026,
-0.007765706162899733,
-0.0320160947740078,
0.0668596550822258,
0.0491030253469944,
-0.06149207800626755,
-0.026338329538702965,
-0.0766419917345047,
0.0024502191226929426,
-0.05118144676089287,
0.06270407140254974,
-0.03137301281094551,
0.12334276735782623,
-0.07748774439096451,
-0.009985808283090591,
-0.03448133543133736,
0.10874061286449432,
-0.033174172043800354,
0.0197227131575346,
-0.029049448668956757,
-0.058905117213726044,
-0.11250487715005875,
0.05808158963918686,
-0.055389653891325,
0.02684382162988186,
0.004532377235591412,
-0.040486861020326614,
-0.013881340622901917,
-0.013758242130279541,
-0.04348514601588249,
-0.07122576981782913,
-1.7972403126080658e-33,
-0.016721537336707115,
0.023256950080394745,
0.006917175836861134,
0.10247476398944855,
-0.03031899593770504,
-0.09763312339782715,
0.008994674310088158,
0.05002550408244133,
-0.04209651052951813,
-0.11415906995534897,
-0.03655783087015152,
-0.03505581617355347,
0.023233631625771523,
-0.013169370591640472,
0.052010904997587204,
-0.018962418660521507,
0.05402909591794014,
0.051378779113292694,
-0.0004025690141133964,
0.08922318369150162,
-0.07188525795936584,
0.07298936694860458,
-0.07931765168905258,
0.06162717565894127,
-0.012141847051680088,
-0.02325158379971981,
-0.09843695163726807,
0.03538293018937111,
0.060510627925395966,
-0.031236208975315094,
-0.06210503727197647,
-0.007989895530045033,
-0.03366692364215851,
-0.08520781248807907,
-0.06722607463598251,
0.03192739561200142,
0.01733468659222126,
0.0077794757671654224,
0.015164460986852646,
0.06585905700922012,
0.09641950577497482,
-0.0015680119395256042,
-0.09439998120069504,
-0.0141191640868783,
-0.015753010287880898,
-0.07106293737888336,
-0.13813072443008423,
-0.005031452979892492,
-0.005660133436322212,
-0.006930720992386341,
-0.03800395503640175,
-0.01773463748395443,
-0.04122563451528549,
0.044752754271030426,
-0.002773141721263528,
-0.07703884690999985,
-0.007412903010845184,
-0.09834835678339005,
0.0009578748140484095,
0.0009201659122481942,
-0.11912696808576584,
0.004839691799134016,
0.04616006463766098,
-0.0869784951210022,
0.022233521565794945,
-0.04047434404492378,
0.0046656630001962185,
0.01923891343176365,
-0.04032306745648384,
0.016547556966543198,
0.10581448674201965,
0.015345456078648567,
-0.024532679468393326,
0.04852450266480446,
-0.03261345252394676,
-0.016014862805604935,
-0.014238697476685047,
-0.05193017050623894,
-0.07306809723377228,
-0.03256221488118172,
0.07006658613681793,
0.003303623991087079,
-0.004693890456110239,
0.013438532128930092,
0.02168327569961548,
0.020499812439084053,
0.03198708966374397,
0.08977402746677399,
0.010331748984754086,
-0.0005080324481241405,
-0.020532578229904175,
-0.009790890850126743,
0.05428113043308258,
0.07697076350450516,
-0.06341429799795151,
-4.1385803228877194e-8,
-0.05963248386979103,
-0.038127601146698,
0.014251693151891232,
0.09180400520563126,
0.008176865987479687,
-0.12418744713068008,
-0.06026333570480347,
0.10692226886749268,
0.017527423799037933,
-0.05187688022851944,
0.07780282944440842,
-0.0018019871786236763,
-0.08935008198022842,
-0.03261962905526161,
0.014019385911524296,
0.12284360080957413,
0.01951547898352146,
0.06505736708641052,
-0.023015569895505905,
-0.016356367617845535,
-0.04561619088053703,
0.009084898047149181,
-0.0113484812900424,
0.00460099708288908,
0.03495060279965401,
-0.0471465028822422,
0.054458893835544586,
0.1299234926700592,
0.09548166394233704,
0.016255352646112442,
-0.041162919253110886,
0.020620590075850487,
-0.018695246428251266,
-0.02982380986213684,
0.016098693013191223,
0.07536055892705917,
0.015015066601336002,
-0.03267911076545715,
0.0008013226906768978,
0.0027335560880601406,
0.0543668270111084,
0.03735916689038277,
-0.04326004162430763,
0.06612398475408554,
0.0360080860555172,
-0.03590428829193115,
-0.0436648353934288,
-0.13128654658794403,
0.017421603202819824,
-0.09176807850599289,
0.012145962566137314,
0.025355789810419083,
0.011396580375730991,
0.04593007266521454,
0.011533534154295921,
0.07376690208911896,
0.005266308784484863,
-0.008455624803900719,
-0.017957303673028946,
0.0036387646105140448,
-0.004340295679867268,
0.03852565214037895,
-0.04151338338851929,
0.03924715891480446
] | https://github.com/msakarvadia/memorization | 14 | 0 | 0 | 0 |
Effective post-training embedding compression via temperature control in contrastive training | https://openreview.net/forum?id=szRmEM8Kx5 | [
"Georgiana Dinu",
"Corey D Barrett",
"Yi Xiang",
"Miguel Romero Calvo",
"Anna Currey",
"Xing Niu"
] | Spotlight | Fixed-size learned representations (dense representations, or embeddings) are widely used in many machine learning applications across language, vision or speech modalities. This paper investigates the role of the temperature parameter in contrastive training for text embeddings. We shed light on the impact this parameter has on the intrinsic dimensionality of the embedding spaces obtained, and show that lower intrinsic dimensionality is further correlated with effective compression of embeddings. We still observe a trade-off between absolute performance and effective compression and we propose temperature aggregation methods which reduce embedding size by an order of magnitude with minimal impact on quality. | representation learning, embeddings, text retrieval, nlp | null | 12,360 | null | [
-0.002216445980593562,
-0.10343662649393082,
-0.011477216146886349,
0.09357873350381851,
0.023569103330373764,
0.05739464983344078,
-0.04147094115614891,
-0.03185252845287323,
0.08255290240049362,
-0.04693716764450073,
-0.0181210245937109,
0.06301844120025635,
0.07859821617603302,
0.028031770139932632,
-0.10250382870435715,
0.04640158265829086,
0.10925561934709549,
0.0813576802611351,
-0.12771491706371307,
-0.07852227985858917,
-0.03818439319729805,
-0.00048826317652128637,
0.01935710944235325,
-0.0363689549267292,
0.10941438376903534,
-0.007261648774147034,
-0.05192875489592552,
-0.0003831361245829612,
0.055154234170913696,
0.03074944019317627,
0.008670899085700512,
-0.027532294392585754,
0.06833191215991974,
0.05082850158214569,
-0.03709441423416138,
0.08131200075149536,
0.006187974940985441,
-0.009934680536389351,
-0.0665968731045723,
0.019226210191845894,
0.041293468326330185,
0.06021534651517868,
0.04169561713933945,
0.07508732378482819,
0.05454877018928528,
0.004221568815410137,
-0.036198940128088,
-0.03624829277396202,
-0.06471426784992218,
-0.02511059306561947,
-0.014231071807444096,
-0.012685609981417656,
-0.03882299363613129,
0.07599412649869919,
-0.13660594820976257,
-0.0008750590495765209,
-0.04170950874686241,
0.06429395079612732,
-0.013434264808893204,
-0.012192404828965664,
-0.0054868306033313274,
-0.02696436457335949,
0.003949446137994528,
0.00006903057510498911,
0.0567639134824276,
-0.025772232562303543,
0.07054916024208069,
0.10151491314172745,
-0.07413498312234879,
0.0520307756960392,
-0.03051455318927765,
0.08483802527189255,
-0.06806155294179916,
0.030701391398906708,
-0.009515340439975262,
0.05859362706542015,
0.05489034205675125,
-0.08944547921419144,
0.08030837774276733,
-0.03284880518913269,
0.031352996826171875,
-0.07482896000146866,
0.08867845684289932,
-0.02368483692407608,
0.05984441190958023,
0.0183766707777977,
0.10409007221460342,
-0.011320080608129501,
-0.0668739303946495,
0.022169597446918488,
-0.03416953980922699,
-0.017650863155722618,
0.02398797869682312,
-0.0262876246124506,
-0.015269913710653782,
0.013531297445297241,
0.01669629104435444,
0.039303917437791824,
-0.02309272438287735,
0.03477241098880768,
0.017619909718632698,
-0.008177286013960838,
0.012674949131906033,
0.06304365396499634,
0.021052071824669838,
-0.08350686728954315,
0.03995237126946449,
-0.037117160856723785,
0.03202938288450241,
-0.0747438371181488,
0.011575669050216675,
-0.011754839681088924,
-0.03992438316345215,
0.0038928473368287086,
0.04007764533162117,
-0.054604046046733856,
-0.045425895601511,
-0.027055099606513977,
0.04976082220673561,
-0.006276343949139118,
-0.028012968599796295,
0.10493367165327072,
-0.032526757568120956,
0.058560438454151154,
-0.03298800066113472,
-0.0385495088994503,
-0.00016585882985964417,
1.3834057815266525e-33,
0.05049867555499077,
0.04428696259856224,
-0.02237771265208721,
0.03979397192597389,
0.07693132013082504,
0.0007414041319862008,
-0.06499187648296356,
0.06013571470975876,
0.006887031719088554,
-0.023256603628396988,
-0.016134563833475113,
0.05225393921136856,
0.028850866481661797,
0.10991824418306351,
0.058176711201667786,
-0.0047206818126142025,
-0.003214094787836075,
0.017777414992451668,
-0.057874489575624466,
-0.021686777472496033,
-0.017230797559022903,
0.030776528641581535,
0.03011695295572281,
-0.011102750897407532,
-0.05893610045313835,
-0.04139633849263191,
0.08010562509298325,
0.0002020039682975039,
-0.09008500725030899,
-0.015384306199848652,
-0.1596612185239792,
-0.06556178629398346,
-0.02917703241109848,
0.001858276897110045,
0.012529505416750908,
-0.03867702931165695,
-0.029620608314871788,
0.007112202700227499,
0.028284527361392975,
-0.03646146506071091,
0.023531539365649223,
-0.0028620094526559114,
0.036416634917259216,
-0.04202434420585632,
-0.0663522407412529,
-0.040655385702848434,
0.015905732288956642,
0.0095163369551301,
-0.12056130915880203,
-0.04982424154877663,
0.01955007016658783,
-0.021729132160544395,
0.019485648721456528,
-0.036373041570186615,
0.07015152275562286,
-0.04937456548213959,
0.03349080681800842,
0.0152248190715909,
0.061807066202163696,
-0.0073464964516460896,
-0.08665800839662552,
-0.028049876913428307,
0.0554933175444603,
0.05891935154795647,
0.02397185191512108,
0.06366552412509918,
-0.02635095827281475,
0.0321110300719738,
0.02040712535381317,
-0.04726925492286682,
0.0007581519312225282,
0.06453146040439606,
-0.04715739190578461,
-0.06241105869412422,
-0.00223514623939991,
0.005560154095292091,
0.051253896206617355,
-0.06839658319950104,
-0.05103159323334694,
0.07795097678899765,
-0.06792504340410233,
-0.06678707897663116,
0.044989727437496185,
-0.14363521337509155,
-0.08060047030448914,
-0.1264428347349167,
0.002186279511079192,
-0.060590751469135284,
0.05698181688785553,
0.02262219600379467,
-0.10368369519710541,
-0.07626482844352722,
-0.00912504643201828,
0.011860616505146027,
0.0867435559630394,
-1.811188471435534e-33,
-0.018463555723428726,
0.02058342844247818,
-0.06389040499925613,
0.08710477501153946,
-0.05330011621117592,
0.016551697626709938,
0.06351783126592636,
0.09432023018598557,
-0.036104071885347366,
-0.021546870470046997,
0.0022657730150967836,
-0.033592965453863144,
0.058868855237960815,
-0.10184834897518158,
0.01282708253711462,
0.0014333969447761774,
-0.06371773034334183,
0.057147424668073654,
0.01631447859108448,
0.0169054064899683,
-0.047237664461135864,
0.02377416379749775,
0.03303153067827225,
0.11734297126531601,
0.06456388533115387,
0.03502552583813667,
-0.022156983613967896,
-0.019930511713027954,
-0.04769214615225792,
-0.028127938508987427,
-0.02074294351041317,
-0.04806462675333023,
-0.06362110376358032,
0.006943494081497192,
-0.10945161432027817,
0.02541421912610531,
0.0778835266828537,
-0.007032915949821472,
0.006977583281695843,
0.04634702205657959,
0.06402254104614258,
-0.03093019872903824,
-0.055553097277879715,
0.01788531057536602,
0.030506908893585205,
-0.023293733596801758,
-0.06226145476102829,
-0.10014661401510239,
0.11851198971271515,
0.04314625263214111,
0.009403528645634651,
-0.01858479157090187,
-0.04015115648508072,
0.0801902487874031,
-0.012392111122608185,
-0.12324268370866776,
-0.042387306690216064,
0.0030030745547264814,
0.056956447660923004,
-0.0349607914686203,
-0.01527146901935339,
-0.09497805684804916,
0.018615808337926865,
-0.07860095798969269,
-0.04575212672352791,
-0.08871317654848099,
-0.022659920156002045,
0.014646380208432674,
0.028468754142522812,
0.0421721525490284,
0.051031410694122314,
-0.005147208459675312,
0.02170984447002411,
0.04396006464958191,
-0.005501252599060535,
-0.005833796691149473,
0.09484265744686127,
0.028677156195044518,
-0.07507560402154922,
-0.020862005650997162,
0.034423455595970154,
0.04437754303216934,
0.03043409436941147,
0.04870069399476051,
0.0856095403432846,
0.08297344297170639,
0.03485623002052307,
-0.0036743064410984516,
-0.00697294482961297,
0.009178933687508106,
0.002556700725108385,
0.02232258766889572,
0.021630335599184036,
0.04612797498703003,
-0.028459588065743446,
-3.4378228264131394e-8,
-0.0832659974694252,
0.0004860554763581604,
0.035807304084300995,
-0.003173162927851081,
0.010570494458079338,
-0.09482178092002869,
-0.021146515384316444,
0.06254605948925018,
-0.0061506908386945724,
0.00817343033850193,
0.05954146012663841,
0.0016084227245301008,
-0.01391949225217104,
-0.08407063782215118,
-0.014212923124432564,
-0.01416739821434021,
-0.01321276742964983,
-0.027448300272226334,
0.04699011892080307,
-0.02665320225059986,
0.011264799162745476,
0.04709149897098541,
0.06360828876495361,
-0.007324951700866222,
0.0026631858199834824,
-0.000498369219712913,
-0.017700856551527977,
0.028740519657731056,
0.0783061757683754,
-0.03519688546657562,
-0.0354803092777729,
0.07102420181035995,
-0.009200147353112698,
-0.09806901216506958,
-0.00935904961079359,
0.06161437928676605,
0.07409702986478806,
0.013101653195917606,
-0.04525503143668175,
-0.03717254102230072,
0.016422629356384277,
0.004966425243765116,
0.001146618160419166,
-0.03215521201491356,
0.052518825978040695,
-0.004157037939876318,
0.027526838704943657,
0.022508397698402405,
-0.00892573781311512,
0.04631855711340904,
0.04040306806564331,
-0.024117454886436462,
-0.037949007004499435,
0.06411387771368027,
0.023596474900841713,
-0.048205289989709854,
0.0038995530921965837,
0.03228602558374405,
-0.03119301237165928,
0.048577114939689636,
-0.025734761729836464,
0.005604292266070843,
-0.05855617672204971,
-0.04396374151110649
] | 0 | 0 | 0 | 0 |
|
TopoNets: High performing vision and language models with brain-like topography | https://openreview.net/forum?id=THqWPzL00e | [
"Mayukh Deb",
"Mainak Deb",
"Apurva Ratan Murty"
] | Spotlight | Neurons in the brain are organized such that nearby cells tend to share similar functions. AI models lack this organization, and past efforts to introduce topography have often led to trade-offs between topography and task performance. In this work, we present *TopoLoss*, a new loss function that promotes spatially organized topographic representations in AI models without significantly sacrificing task performance. TopoLoss is highly adaptable and can be seamlessly integrated into the training of leading model architectures. We validate our method on both vision (ResNet-18, ResNet-50, ViT) and language models (GPT-Neo-125M, NanoGPT), collectively *TopoNets*. TopoNets are the highest performing supervised topographic models to date, exhibiting brain-like properties such as localized feature processing, lower dimensionality, and increased efficiency. TopoNets also predict responses in the brain and replicate the key topographic signatures observed in the brain’s visual and language cortices, further bridging the gap between biological and artificial systems. This work establishes a robust and generalizable framework for integrating topography into AI, advancing the development of high performing models that more closely emulate the computational strategies of the human brain. Our project page: https://toponets.github.io | topography, neuro-inspired, convolutional neural networks, Transformers, visual cortex, neuroscience | A generalizable framework for inducing brain-like topography in neural networks without compromising task performance | 12,347 | 2501.16396 | [
-0.003644985146820545,
-0.11229217797517776,
0.058446504175662994,
-0.01264047808945179,
0.06998679786920547,
0.03845877945423126,
-0.009131455793976784,
0.05411705747246742,
0.01948324590921402,
-0.02441546507179737,
-0.10009103268384933,
-0.0664231926202774,
0.019482240080833435,
0.0637926459312439,
-0.028854049742221832,
-0.03038846328854561,
-0.015941770747303963,
0.09858371317386627,
-0.03006720542907715,
-0.07449184358119965,
0.07359617948532104,
0.030807405710220337,
0.011631881818175316,
0.0064156921580433846,
0.010097244754433632,
0.06346502900123596,
0.027315979823470116,
-0.010704032145440578,
-0.06075270101428032,
-0.01900077611207962,
0.0400407612323761,
-0.017285676673054695,
-0.05073283612728119,
0.06940034031867981,
-0.09011712670326233,
0.09050052613019943,
-0.09229686111211777,
0.01849975436925888,
-0.01055874489247799,
-0.10270803421735764,
-0.009955765679478645,
0.017236197367310524,
-0.01358312088996172,
0.06890685111284256,
0.10898144543170929,
0.04973387345671654,
0.024792656302452087,
-0.049192044883966446,
-0.028227761387825012,
-0.04126487299799919,
-0.07239382714033127,
-0.052788712084293365,
-0.019291149452328682,
0.06712809205055237,
0.003512782510370016,
0.1726524978876114,
0.036871880292892456,
-0.028041645884513855,
0.019202105700969696,
0.021852806210517883,
0.022179121151566505,
-0.0494123212993145,
-0.01069698017090559,
-0.05904104560613632,
0.04823208600282669,
-0.004415728617459536,
0.04686983302235603,
0.030958320945501328,
-0.06467486917972565,
0.012851995415985584,
0.06870844960212708,
0.08351780474185944,
-0.02921510674059391,
-0.029244055971503258,
0.000589831848628819,
0.050019361078739166,
0.08464862406253815,
0.020644960924983025,
0.02760469540953636,
-0.04384331405162811,
0.01225136686116457,
0.11611133068799973,
0.07522939890623093,
0.01210013497620821,
0.12617945671081543,
0.062224939465522766,
-0.02534504048526287,
0.04277839511632919,
-0.014482482336461544,
0.0125115429982543,
-0.06154889985918999,
-0.09340909123420715,
-0.046004511415958405,
-0.05907387658953667,
-0.005891742184758186,
-0.04526215046644211,
-0.009486749768257141,
-0.11160366237163544,
-0.021954119205474854,
0.017973115667700768,
0.03332386538386345,
-0.058162398636341095,
0.018766051158308983,
-0.0044664740562438965,
0.03800438344478607,
-0.030698370188474655,
0.13675151765346527,
0.031120391562581062,
-0.0026192914228886366,
0.00924176536500454,
0.0002073003852274269,
-0.023232324048876762,
-0.012255175039172173,
0.07791309058666229,
-0.014683992601931095,
0.05670725554227829,
-0.03692861646413803,
0.025719765573740005,
0.06432492285966873,
0.00990009494125843,
-0.04771970584988594,
0.019729331135749817,
-0.0014892503386363387,
-0.023600954562425613,
0.00486009893938899,
-0.0042012473568320274,
-0.0590195469558239,
8.020251937221576e-34,
0.02665936015546322,
0.02994385175406933,
0.010315635241568089,
0.047779202461242676,
0.07244833558797836,
-0.018633756786584854,
0.005078009329736233,
0.016534145921468735,
-0.034860532730817795,
0.041279930621385574,
-0.1036529541015625,
0.011424737051129341,
-0.07697078585624695,
0.1467045694589615,
0.011937237344682217,
-0.03082006424665451,
-0.0038651328068226576,
0.013925501145422459,
-0.06907886266708374,
-0.0699581578373909,
-0.03892681747674942,
0.029283691197633743,
0.028114549815654755,
-0.06467677652835846,
-0.05597498267889023,
0.005324335303157568,
-0.029368722811341286,
-0.06467195600271225,
-0.04228286072611809,
0.012791674584150314,
-0.11011625081300735,
0.07136081159114838,
0.0017719537718221545,
0.0034933933056890965,
0.017904451116919518,
-0.034450072795152664,
0.03415636718273163,
-0.05739673227071762,
0.049495358020067215,
-0.00977666862308979,
0.04300355911254883,
0.12472745031118393,
0.04595550149679184,
-0.05127771571278572,
-0.0697031170129776,
0.03015146590769291,
0.05183882266283035,
0.04207328334450722,
-0.017572809010744095,
-0.007423473056405783,
-0.00714173074811697,
-0.020306488499045372,
0.007616687566041946,
-0.07825765758752823,
-0.03405600041151047,
0.044366639107465744,
-0.009536197409033775,
0.041103295981884,
0.07462552189826965,
0.010046353563666344,
0.018610136583447456,
-0.0335831344127655,
-0.008066258393228054,
0.13169744610786438,
0.02451564371585846,
0.007499208673834801,
-0.029270565137267113,
0.06762157380580902,
0.03637675568461418,
0.07192234694957733,
0.012478562071919441,
0.06559247523546219,
0.005991602316498756,
-0.03360104188323021,
0.0047582476399838924,
-0.029791831970214844,
0.0846550241112709,
-0.09112716466188431,
0.001523256883956492,
-0.014437352307140827,
-0.005854816175997257,
-0.024783484637737274,
-0.08952433615922928,
-0.08765501528978348,
0.00482838274911046,
-0.007336121518164873,
0.059227026998996735,
-0.06299993395805359,
0.0063647120259702206,
0.009871771559119225,
-0.06496161967515945,
-0.09926522523164749,
0.057230863720178604,
-0.05463091656565666,
-0.09782890230417252,
-9.171878242592402e-34,
-0.08859355747699738,
0.03062576986849308,
-0.004325082525610924,
-0.00641533313319087,
-0.09353271126747131,
-0.03772041201591492,
0.06322097033262253,
0.03991633653640747,
-0.044345200061798096,
-0.029249759390950203,
-0.002107340609654784,
-0.016604138538241386,
0.030312104150652885,
0.0033213619608432055,
0.046112168580293655,
-0.04200196638703346,
-0.04966851696372032,
-0.03553837165236473,
0.0484967976808548,
-0.048144012689590454,
-0.002014400204643607,
0.09162037819623947,
-0.07264808565378189,
-0.009938573464751244,
-0.05131053179502487,
0.03523172438144684,
-0.07123453170061111,
0.07573720812797546,
-0.0026719626039266586,
-0.016442859545350075,
-0.047588665038347244,
0.049213092774152756,
-0.07838771492242813,
-0.010313736274838448,
-0.05776679515838623,
0.11152024567127228,
0.057054754346609116,
-0.11158369481563568,
-0.0775609090924263,
0.05266910418868065,
0.022301780059933662,
-0.022308560088276863,
0.014759117737412453,
0.027708429843187332,
-0.016294533386826515,
0.028253091499209404,
-0.09751731902360916,
0.013013163581490517,
-0.07820623368024826,
0.014182482846081257,
-0.07161293178796768,
-0.012824825942516327,
-0.11343562602996826,
0.011478124186396599,
-0.05078325420618057,
-0.03961077332496643,
0.0005138055421411991,
0.013161787763237953,
0.011525332927703857,
-0.05942046269774437,
-0.014546677470207214,
-0.07554451376199722,
-0.02358921989798546,
-0.028168555349111557,
0.012079186737537384,
0.001180914812721312,
-0.03432775288820267,
0.008094824850559235,
0.0744367465376854,
0.034681059420108795,
-0.026866566389799118,
-0.01799527183175087,
0.054691605269908905,
0.025596795603632927,
-0.01639038696885109,
-0.04426172375679016,
0.02661365456879139,
0.0000206210752367042,
-0.0019886598456650972,
-0.06752094626426697,
0.024166204035282135,
0.014578220434486866,
-0.06484132260084152,
0.0682305172085762,
0.10892277956008911,
0.06860149651765823,
0.03617630526423454,
0.07222836464643478,
0.020219719037413597,
-0.0015481708105653524,
-0.04001310095191002,
0.002081269398331642,
-0.012814036570489407,
0.06581372767686844,
-0.08393117785453796,
-3.9886909775077584e-8,
-0.06546325981616974,
-0.01928715966641903,
0.01851346157491207,
0.008139574900269508,
-0.02048793062567711,
-0.08804218471050262,
0.05545254424214363,
0.1230054646730423,
0.010339517146348953,
-0.026276033371686935,
0.003294228110462427,
-0.09150032699108124,
0.00013344960461836308,
-0.04006000608205795,
0.04161734879016876,
0.07414953410625458,
0.0011042124824598432,
0.06730873137712479,
-0.016688477247953415,
-0.011336096562445164,
-0.0406494215130806,
0.06341811269521713,
0.02544940635561943,
0.024015408009290695,
0.035459425300359726,
-0.04016142338514328,
-0.056235022842884064,
0.07337933778762817,
-0.04116992652416229,
-0.01651999168097973,
0.027144338935613632,
0.04277052730321884,
0.06285084038972855,
0.027657153084874153,
-0.0222671739757061,
0.0751379206776619,
0.0376288928091526,
0.0587642602622509,
-0.020898625254631042,
-0.027863824740052223,
0.013717515394091606,
0.046130333095788956,
-0.006253968924283981,
0.030515605583786964,
0.039711736142635345,
0.0023897592909634113,
0.04265788942575455,
-0.03796554356813431,
0.08288323134183884,
-0.03875889256596565,
-0.054628994315862656,
0.06664400547742844,
0.0000881111336639151,
0.07574784755706787,
0.03767701983451843,
-0.005918610841035843,
-0.020556747913360596,
-0.026876773685216904,
0.00895135197788477,
0.03928650915622711,
0.034247543662786484,
0.02538624405860901,
-0.04269171878695488,
-0.039773128926754
] | https://github.com/toponets/topoloss | 56 | 0 | 0 | 0 |
INCLUDE: Evaluating Multilingual Language Understanding with Regional Knowledge | https://openreview.net/forum?id=k3gCieTXeY | [
"Angelika Romanou",
"Negar Foroutan",
"Anna Sotnikova",
"Sree Harsha Nelaturu",
"Shivalika Singh",
"Rishabh Maheshwary",
"Micol Altomare",
"Zeming Chen",
"Mohamed A. Haggag",
"Snegha A",
"Alfonso Amayuelas",
"Azril Hafizi Amirudin",
"Danylo Boiko",
"Michael Chang",
"Jenny Chim",
"Gal Cohen",
"Aditya Kumar Dalmia",
"Abraham Diress",
"Sharad Duwal",
"Daniil Dzenhaliou",
"et al. (37 additional authors not shown)"
] | Spotlight | The performance differential of large language models (LLM) between languages hinders their effective deployment in many regions, inhibiting the potential economic and societal value of generative AI tools in many communities. However, the development of functional LLMs in many languages (i.e., multilingual LLMs) is bottlenecked by the lack of high-quality evaluation resources in languages other than English. Moreover, current practices in multilingual benchmark construction often translate English resources, ignoring the regional and cultural knowledge of the environments in which multilingual systems would be used. In this work, we construct an evaluation suite of 197,243 QA pairs from local exam sources to measure the capabilities of multilingual LLMs in a variety of regional contexts.
Our novel resource, INCLUDE, is a comprehensive knowledge- and reasoning-centric benchmark across 44 written languages that evaluates multilingual LLMs for performance in the actual language environments where they would be deployed. | evaluation, multilinguality, large language models | null | 12,318 | 2411.19799 | [
0.006888067349791527,
-0.052401915192604065,
0.011915650218725204,
-0.004887950140982866,
-0.005756088066846132,
-0.02224772982299328,
-0.02024153061211109,
0.01235619280487299,
0.04330706596374512,
-0.03071882389485836,
-0.08921413123607635,
-0.11306694149971008,
0.03600448742508888,
0.07694593816995621,
0.029373332858085632,
0.0008647639187984169,
0.025054454803466797,
0.00721115805208683,
-0.024836162105202675,
-0.09020226448774338,
0.023897532373666763,
0.05609504505991936,
0.0631289854645729,
-0.03522232919931412,
0.11394376307725906,
-0.07704457640647888,
-0.03926398605108261,
-0.02916841208934784,
0.03849128633737564,
-0.01203994546085596,
0.007008071057498455,
0.08747263252735138,
-0.05212947726249695,
0.07652550935745239,
0.003243526676669717,
0.062275439500808716,
-0.037517137825489044,
-0.07735840231180191,
0.054091908037662506,
0.050505101680755615,
-0.038797542452812195,
-0.0027106101624667645,
0.04442537948489189,
-0.08086913079023361,
0.11029303073883057,
0.04884423688054085,
-0.027893351390957832,
0.02832895703613758,
-0.049541693180799484,
0.01822444424033165,
-0.08534416556358337,
-0.006879510823637247,
0.0001862058270489797,
0.03864472359418869,
-0.019297687336802483,
0.020872177556157112,
0.06691375374794006,
0.01295365672558546,
-0.01889261230826378,
-0.017626836895942688,
-0.09470108896493912,
-0.12451430410146713,
0.009610085748136044,
0.045946408063173294,
-0.01291545107960701,
-0.0543447881937027,
-0.023652156814932823,
0.07189023494720459,
-0.09964460879564285,
0.01965809054672718,
-0.035912808030843735,
0.058417897671461105,
-0.02151627466082573,
0.05891792103648186,
0.013441928662359715,
0.05390412360429764,
-0.026298556476831436,
-0.03353032469749451,
0.029412640258669853,
-0.08736521750688553,
0.04779810830950737,
0.004525924567133188,
-0.006161608267575502,
-0.022795211523771286,
0.1332533359527588,
-0.021739093586802483,
0.0073441253043711185,
0.013023720122873783,
0.033649709075689316,
0.04713032394647598,
-0.0038389151450246572,
-0.015561094507575035,
0.05956817418336868,
0.05027071014046669,
0.11254806816577911,
0.043640416115522385,
0.04639256000518799,
-0.06460237503051758,
-0.025388896465301514,
0.07002781331539154,
0.027118150144815445,
0.08356047421693802,
0.03690719231963158,
-0.0767357274889946,
-0.10100989043712616,
0.0007042245706543326,
0.027912115678191185,
0.01881430856883526,
0.04945770278573036,
-0.09618845582008362,
0.015084381215274334,
0.08506828546524048,
0.06706400215625763,
0.006559388712048531,
0.019705701619386673,
0.002147669205442071,
0.0905052199959755,
0.0029214646201580763,
0.08632547408342361,
0.06327104568481445,
-0.09704884141683578,
0.0018390112090855837,
0.04779747501015663,
0.003048342652618885,
0.0682922899723053,
0.037388794124126434,
-0.03075134940445423,
-9.155427864959991e-35,
0.08786656707525253,
0.045099787414073944,
-0.028055405244231224,
0.05377882719039917,
0.011618140153586864,
-0.03267228230834007,
0.03565121442079544,
0.028451574966311455,
-0.09383860975503922,
-0.07174979895353317,
-0.03341294452548027,
0.11910267174243927,
-0.03362976759672165,
0.03536257892847061,
0.038678430020809174,
0.0696110799908638,
0.055656034499406815,
0.0416809506714344,
-0.037592124193906784,
0.02620038017630577,
0.08591120690107346,
-0.0494157038629055,
0.027805348858237267,
-0.008252691477537155,
0.010520460084080696,
0.020309828221797943,
0.09288948774337769,
-0.045040395110845566,
-0.018401095643639565,
0.0011309231631457806,
-0.11281994730234146,
-0.07511764019727707,
0.0017558617983013391,
0.08504131436347961,
0.02058124728500843,
-0.05784808471798897,
-0.06608578562736511,
-0.011627164669334888,
-0.018099647015333176,
0.02054854854941368,
-0.06952795386314392,
0.02990911528468132,
0.015048316679894924,
0.010537260212004185,
-0.03849254176020622,
-0.050948381423950195,
-0.03624981641769409,
-0.017203642055392265,
0.03461093083024025,
-0.07499435544013977,
-0.00278492015786469,
-0.02906915917992592,
-0.002343089086934924,
-0.006563379894942045,
0.04234902560710907,
0.02479635365307331,
0.015352262184023857,
0.02701769582927227,
0.0036850092001259327,
0.03205357864499092,
-0.02269027568399906,
0.0634189248085022,
0.01964041404426098,
0.03485879302024841,
0.03175027668476105,
0.023417195305228233,
-0.04931119456887245,
0.07708831131458282,
0.07014551758766174,
-0.010065005160868168,
-0.02339206449687481,
-0.03535390645265579,
0.06483697891235352,
-0.03810177743434906,
0.025225425139069557,
0.042184874415397644,
-0.01874636672437191,
-0.12163276225328445,
0.10470091551542282,
-0.017290744930505753,
0.037358902394771576,
-0.015839992091059685,
-0.05531982704997063,
-0.06537219882011414,
0.03281183913350105,
-0.0842675194144249,
0.00832162331789732,
-0.09507019072771072,
0.05965767428278923,
-0.00015020497085060924,
-0.026931900531053543,
-0.04589509591460228,
-0.04720624163746834,
-0.05390182510018349,
0.018051840364933014,
-1.9235547340734796e-34,
-0.04442417249083519,
-0.05533133074641228,
-0.030785974115133286,
0.08601085841655731,
-0.06499689817428589,
-0.0790315791964531,
-0.036468569189310074,
0.03916221484541893,
-0.03319917991757393,
-0.05205010995268822,
-0.0590614415705204,
-0.04032813012599945,
0.06569495797157288,
0.059745002537965775,
0.010497973300516605,
-0.03550820052623749,
0.01769711636006832,
0.007420863024890423,
0.03134748712182045,
0.11370038241147995,
-0.013251040130853653,
0.04993809014558792,
-0.04908638447523117,
0.06290649622678757,
-0.05704595893621445,
0.0046674031764268875,
-0.08915703743696213,
-0.04147312790155411,
-0.09756540507078171,
0.0035207881592214108,
0.007014904171228409,
-0.06629040837287903,
-0.04133304953575134,
0.008026684634387493,
-0.05670858919620514,
0.014578022994101048,
0.0760076716542244,
-0.0615127831697464,
-0.00010734129318734631,
0.04660830646753311,
0.04369548335671425,
-0.04011103883385658,
-0.01830309070646763,
0.004418461117893457,
-0.04097320884466171,
0.0299486443400383,
-0.1098252609372139,
0.005998533684760332,
-0.015866050496697426,
-0.007991759106516838,
0.011750859208405018,
-0.02696608006954193,
-0.04095792770385742,
-0.00799897126853466,
0.013418310321867466,
-0.11118756979703903,
-0.04591945931315422,
-0.07867063581943512,
-0.07248615473508835,
-0.026489362120628357,
-0.0364224947988987,
0.01638280600309372,
0.034439440816640854,
-0.031241396442055702,
0.011204233393073082,
-0.022060127928853035,
-0.014864570461213589,
0.03630474582314491,
-0.01248291414231062,
-0.00791259203106165,
0.016478443518280983,
-0.048966243863105774,
-0.07477761059999466,
-0.0023300473112612963,
-0.015582939609885216,
0.047152724117040634,
0.007612751796841621,
-0.03854081779718399,
0.028706412762403488,
-0.0620860792696476,
0.009664226323366165,
-0.03064815141260624,
-0.019319970160722733,
-0.039891183376312256,
0.037247948348522186,
0.022895725443959236,
0.030735887587070465,
0.10265113413333893,
0.012492071837186813,
0.024275748059153557,
0.019462598487734795,
0.08321361988782883,
-0.027738915756344795,
-0.006132778711616993,
-0.04483662545681,
-3.788012392647033e-8,
-0.04557333141565323,
0.009951410815119743,
-0.042393751442432404,
0.04658517986536026,
-0.017694653943181038,
-0.07808881998062134,
-0.0596582293510437,
0.08883441984653473,
-0.003996782470494509,
0.02552170492708683,
0.02869119495153427,
0.004280791152268648,
-0.026834595948457718,
-0.07481356710195541,
0.028803875669836998,
0.0881139412522316,
-0.00552463298663497,
0.09003566950559616,
-0.032982245087623596,
-0.032432906329631805,
0.1271907240152359,
0.06166178733110428,
0.0028382129967212677,
0.002391338814049959,
0.007620964199304581,
-0.07208508998155594,
-0.051539015024900436,
0.033734530210494995,
0.05627187713980675,
-0.07479660958051682,
0.012093051336705685,
-0.028162511065602303,
-0.03810287266969681,
-0.026035599410533905,
0.04800979793071747,
0.04491976276040077,
-0.05352850630879402,
0.042553238570690155,
0.0687279999256134,
0.02883620746433735,
0.06632371246814728,
-0.006250828038901091,
-0.04778466746211052,
0.07237087935209274,
0.0026998950634151697,
-0.05575764924287796,
-0.1364988386631012,
-0.0234522707760334,
-0.0036582713946700096,
-0.0437663272023201,
-0.017425065860152245,
-0.10344486683607101,
0.0017953923670575023,
0.07773348689079285,
0.06532391905784607,
0.030244337394833565,
-0.06470457464456558,
-0.045969102531671524,
0.07968970388174057,
0.0623650886118412,
0.04045761004090309,
-0.0032911058515310287,
-0.012499368749558926,
0.024405617266893387
] | 0 | 0 | 7 | 0 |
|
Deep Learning Alternatives Of The Kolmogorov Superposition Theorem | https://openreview.net/forum?id=SyVPiehSbg | [
"Leonardo Ferreira Guilhoto",
"Paris Perdikaris"
] | Spotlight | This paper explores alternative formulations of the Kolmogorov Superposition Theorem (KST) as a foundation for neural network design. The original KST formulation, while mathematically elegant, presents practical challenges due to its limited insight into the structure of inner and outer functions and the large number of unknown variables it introduces. Kolmogorov-Arnold Networks (KANs) leverage KST for function approximation, but they have faced scrutiny due to mixed results compared to traditional multilayer perceptrons (MLPs) and practical limitations imposed by the original KST formulation. To address these issues, we introduce ActNet, a scalable deep learning model that builds on the KST and overcomes some of the drawbacks of Kolmogorov's original formulation. We evaluate ActNet in the context of Physics-Informed Neural Networks (PINNs), a framework well-suited for leveraging KST's strengths in low-dimensional function approximation, particularly for simulating partial differential equations (PDEs). In this challenging setting, where models must learn latent functions without direct measurements, ActNet consistently outperforms KANs across multiple benchmarks and is competitive against the current best MLP-based approaches. These results present ActNet as a promising new direction for KST-based deep learning applications, particularly in scientific computing and PDE simulation tasks. | Kolmogorov-Arnold Representation Theorem, Function Approximation, Physics Informed Neural Networks, AI4Science | This paper introduces ActNet, a neural network architecture based on alternative KST formulations, overcoming limitations of KANs and showing improved performance in PINNs for simulating partial differential equations. | 12,259 | 2410.01990 | [
-0.13516977429389954,
-0.0909438505768776,
0.0960012823343277,
0.10099576413631439,
-0.022470710799098015,
0.07252292335033417,
-0.020610079169273376,
-0.009788895957171917,
0.04930446669459343,
0.05012018233537674,
0.007280878722667694,
-0.03768879547715187,
-0.0536155104637146,
0.021958880126476288,
-0.05030345171689987,
-0.011759818531572819,
-0.0072290110401809216,
0.05611829832196236,
-0.1325681358575821,
0.0020650236401706934,
0.013997121714055538,
0.04698411002755165,
0.0021418030373752117,
-0.009628447704017162,
0.019748514518141747,
-0.02649685926735401,
-0.00739901140332222,
0.017701487988233566,
-0.05619809031486511,
0.03479000926017761,
0.09643367677927017,
0.0024855274241417646,
-0.032753124833106995,
-0.024433664977550507,
-0.06599677354097366,
0.1039985716342926,
-0.08102263510227203,
0.057667359709739685,
-0.012477314099669456,
-0.06547858566045761,
0.012364121153950691,
0.021830067038536072,
-0.028154544532299042,
0.025712033733725548,
0.11404263228178024,
0.005374194588512182,
0.036876387894153595,
-0.06805886328220367,
0.058331914246082306,
0.06215960532426834,
-0.011552534066140652,
-0.002162832533940673,
0.0706537663936615,
0.022735480219125748,
0.09824021905660629,
-0.03177040070295334,
-0.036004163324832916,
0.0006384976441040635,
-0.03137395903468132,
0.005547940265387297,
0.06929731369018555,
-0.07997504621744156,
-0.02499016933143139,
-0.021540194749832153,
-0.0011826963163912296,
0.01659920997917652,
-0.005874143447726965,
0.08523029834032059,
0.013116251677274704,
0.03415325656533241,
-0.012298131361603737,
0.07271894812583923,
-0.08888734132051468,
-0.018949078395962715,
0.07688815146684647,
-0.02369527518749237,
0.05997322127223015,
0.01065355446189642,
0.07854016125202179,
-0.05258256569504738,
0.04127580299973488,
0.028600936755537987,
-0.05638676509261131,
-0.04316761717200279,
0.10810806602239609,
0.030742865055799484,
0.012096531689167023,
0.06524144858121872,
-0.017653103917837143,
-0.04035226255655289,
-0.02793434076011181,
-0.019732672721147537,
0.01511581800878048,
-0.09105779230594635,
0.028399281203746796,
-0.011277880519628525,
-0.01854783669114113,
-0.08745396882295609,
-0.042580779641866684,
0.1479421705007553,
0.01626356691122055,
0.04861375316977501,
0.017307789996266365,
0.009626105427742004,
0.06700533628463745,
-0.017273765057325363,
-0.02076849341392517,
-0.004843630827963352,
0.030808255076408386,
-0.09924972802400589,
0.026822630316019058,
0.01082916371524334,
0.005630986765027046,
0.0027866889722645283,
0.04365627467632294,
-0.05586815997958183,
0.0436638705432415,
-0.03421751782298088,
0.01643589325249195,
0.028200464323163033,
-0.10462338477373123,
0.015446161851286888,
-0.03262544795870781,
0.0497121699154377,
0.00919475220143795,
0.001743778819218278,
-0.10551972687244415,
3.768597052511837e-33,
-0.027013830840587616,
0.009163395501673222,
0.08094217628240585,
-0.004699496552348137,
0.04772675409913063,
-0.06939646601676941,
0.04835895448923111,
0.02847711183130741,
0.012614944949746132,
0.0025066579692065716,
-0.09742681682109833,
0.025388645008206367,
-0.04666167125105858,
0.021459052339196205,
0.01221993938088417,
0.043832845985889435,
-0.08897709101438522,
-0.012323228642344475,
0.05113288387656212,
-0.03518667072057724,
0.04863539710640907,
-0.1266515702009201,
-0.02539459802210331,
-0.06344271451234818,
0.010296940803527832,
-0.012587257660925388,
0.033563364297151566,
0.009707501158118248,
-0.06432180106639862,
-0.003941514063626528,
-0.05881546437740326,
0.0686940997838974,
-0.04440316557884216,
0.03590284660458565,
0.010999303311109543,
-0.016158903017640114,
-0.029777085408568382,
-0.07050695270299911,
0.019157517701387405,
-0.00592126464471221,
-0.0019592554308474064,
0.0032246997579932213,
-0.006929939612746239,
0.04822924733161926,
-0.06509716063737869,
0.004858974367380142,
0.05718675255775452,
0.0018841656856238842,
0.016861235722899437,
-0.010523411445319653,
-0.019595010206103325,
-0.007993729785084724,
0.02687665820121765,
-0.08791404217481613,
0.03286158666014671,
0.02316950634121895,
0.03328405320644379,
0.01306168269366026,
0.028196090832352638,
0.059256669133901596,
-0.010763692669570446,
0.028466593474149704,
0.02465151436626911,
0.0807824656367302,
-0.0007182805566117167,
0.0349438339471817,
-0.08588504791259766,
0.08306951075792313,
0.048183832317590714,
-0.05094060301780701,
0.023493587970733643,
0.07366164028644562,
0.047053638845682144,
-0.09497620910406113,
0.0909905657172203,
0.03541666269302368,
-0.021381985396146774,
-0.09564998745918274,
-0.030688222497701645,
0.06231771782040596,
0.01610645279288292,
0.026654504239559174,
-0.05260980874300003,
0.03613889217376709,
-0.07600975781679153,
-0.027341973036527634,
0.034717217087745667,
-0.05668661743402481,
0.036285314708948135,
-0.05775774270296097,
-0.1112419068813324,
-0.09592126309871674,
-0.006808154284954071,
-0.04581337794661522,
0.001043385942466557,
-3.8918277987283834e-33,
-0.06594320386648178,
0.0647183209657669,
-0.04707237705588341,
0.005002036225050688,
-0.010314013808965683,
-0.01859857514500618,
-0.040512554347515106,
-0.030992241576313972,
-0.023591501638293266,
-0.0035485236439853907,
0.03972684219479561,
-0.030390501022338867,
-0.024613112211227417,
-0.06884095072746277,
0.07165303826332092,
0.0409700870513916,
-0.060023870319128036,
-0.014655317179858685,
0.06937570124864578,
-0.00023153658548835665,
0.010028611868619919,
0.07676346600055695,
-0.051148977130651474,
0.06235286593437195,
-0.04102342575788498,
-0.05098727345466614,
-0.008341476321220398,
0.07715686410665512,
0.015606090426445007,
-0.015472263097763062,
-0.03762150555849075,
-0.09422500431537628,
-0.06402730941772461,
0.00864674523472786,
-0.021059852093458176,
0.02997869811952114,
0.0862186923623085,
0.02127978764474392,
-0.021337412297725677,
-0.0713113471865654,
0.05153152346611023,
-0.02687539905309677,
-0.045014169067144394,
0.03606313094496727,
-0.004248976241797209,
0.0025941929779946804,
0.015719879418611526,
-0.015289939939975739,
0.011722719296813011,
0.05839339271187782,
0.014082618057727814,
-0.03416832163929939,
-0.06210494041442871,
0.0004442734643816948,
-0.10937439650297165,
0.012763255275785923,
-0.018869761377573013,
0.0540827214717865,
0.06262568384408951,
0.030837101861834526,
-0.10313327610492706,
-0.11811213940382004,
0.043488599359989166,
0.022535253316164017,
-0.020174531266093254,
-0.031402017921209335,
-0.07144857943058014,
0.13027487695217133,
0.052272554486989975,
-0.011165604926645756,
-0.0686446949839592,
-0.06132951378822327,
0.043718934059143066,
0.024402206763625145,
-0.009809487499296665,
-0.00839269906282425,
-0.0029818022157996893,
-0.10794075578451157,
0.01025490090250969,
0.04092692211270332,
0.03884095326066017,
-0.048110391944646835,
0.04678412526845932,
0.07796624302864075,
0.06623818725347519,
0.05184725672006607,
0.029495196416974068,
-0.017384368926286697,
0.03805205971002579,
-0.03158191964030266,
-0.010258960537612438,
0.09905265271663666,
-0.026053983718156815,
0.0016482226783409715,
-0.04051974415779114,
-4.6817444143698594e-8,
0.03595622628927231,
0.08061221241950989,
0.04764799028635025,
-0.05034402012825012,
0.08122855424880981,
-0.04853446036577225,
-0.028166910633444786,
-0.002769971964880824,
-0.025225568562746048,
-0.002297190483659506,
0.04799234867095947,
-0.02653580717742443,
0.013519752770662308,
-0.12777461111545563,
-0.03900127112865448,
0.04877278208732605,
0.018910815939307213,
0.022848889231681824,
-0.004560074768960476,
-0.027484090998768806,
-0.04524366185069084,
0.0880046933889389,
0.03840886056423187,
-0.042083147913217545,
0.03426027670502663,
-0.03238765522837639,
-0.012084973976016045,
0.011084732599556446,
0.06425956636667252,
0.016938146203756332,
-0.03501090407371521,
0.005054861307144165,
-0.019265683367848396,
0.03896518796682358,
0.020107701420783997,
0.07303754985332489,
0.017106343060731888,
-0.0012019176501780748,
-0.025227652862668037,
0.013505849987268448,
-0.08584639430046082,
0.031181689351797104,
-0.020235862582921982,
0.031427592039108276,
-0.0016985612455755472,
0.001591501641087234,
-0.038675110787153244,
-0.10265432298183441,
0.1020544171333313,
0.06888190656900406,
0.07395043224096298,
0.06259758025407791,
-0.008476910181343555,
0.05815427377820015,
0.08087024837732315,
0.05701614171266556,
-0.027379736304283142,
-0.08604495972394943,
-0.01908843219280243,
0.11926378309726715,
-0.029591629281640053,
0.04358138144016266,
-0.06618652492761612,
0.04634396731853485
] | https://github.com/predictiveintelligencelab/actnet | 15 | 0 | 0 | 0 |
Vision Language Models are In-Context Value Learners | https://openreview.net/forum?id=friHAl5ofG | [
"Yecheng Jason Ma",
"Joey Hejna",
"Chuyuan Fu",
"Dhruv Shah",
"Jacky Liang",
"Zhuo Xu",
"Sean Kirmani",
"Peng Xu",
"Danny Driess",
"Ted Xiao",
"Osbert Bastani",
"Dinesh Jayaraman",
"Wenhao Yu",
"Tingnan Zhang",
"Dorsa Sadigh",
"Fei Xia"
] | Spotlight | Predicting temporal progress from visual trajectories is important for intelligent robots that can learn, adapt, and improve. However, learning such progress estimator, or temporal value function, across different tasks and domains requires both a large amount of diverse data and methods which can scale and generalize. To address these challenges, we present Generative Value Learning (GVL), a universal value function estimator that leverages the world knowledge embedded in vision-language models (VLMs) to predict task progress. Naively asking a VLM to predict values for a video sequence performs poorly due to the strong temporal correlation between successive frames. Instead, GVL poses value estimation as a temporal ordering problem over shuffled video frames; this seemingly more challenging task encourages VLMs to more fully exploit their underlying semantic and temporal grounding capabilities to differentiate frames based on their perceived task progress, consequently producing significantly better value predictions. Without any robot or task specific training, GVL can in-context zero-shot and few-shot predict effective values for more than 300 distinct real-world tasks across diverse robot platforms, including challenging bimanual manipulation tasks. Furthermore, we demonstrate that GVL permits flexible multi-modal in-context learning via examples from heterogeneous tasks and embodiments, such as human videos. The generality of GVL enables various downstream applications pertinent to visuomotor policy learning, including dataset filtering, success detection, and value-weighted regression -- all without any model training or finetuning. | robot learning, vision-language model, value estimation, manipulation | null | 12,258 | 2411.04549 | [
-0.004998279269784689,
-0.09126976877450943,
0.03765994310379028,
-0.016490692272782326,
0.07454117387533188,
0.04253368824720383,
-0.01226718444377184,
-0.0680987760424614,
0.022634558379650116,
-0.04396010935306549,
-0.011371081694960594,
-0.09706614166498184,
-0.043558236211538315,
0.11235001683235168,
0.004159859381616116,
0.018154997378587723,
0.08562573790550232,
0.06509821861982346,
-0.08031807094812393,
-0.05221644416451454,
0.022236522287130356,
0.03930582106113434,
0.08897815644741058,
0.027179773896932602,
0.0270864125341177,
0.0642160102725029,
0.000052947463700547814,
-0.08652214705944061,
0.07193493843078613,
-0.048469386994838715,
0.0442364439368248,
-0.009214654564857483,
-0.031008463352918625,
-0.017963992431759834,
-0.056771352887153625,
0.08943787217140198,
-0.08044344931840897,
0.024881836026906967,
0.023708360269665718,
-0.05743906646966934,
-0.007036478258669376,
-0.010527187027037144,
0.0797567144036293,
-0.015850622206926346,
0.021657556295394897,
0.07973090559244156,
0.03368750959634781,
-0.1096528023481369,
0.009289246052503586,
0.009328296408057213,
-0.06236863508820534,
-0.027386115863919258,
0.033974483609199524,
0.023377403616905212,
-0.02342492900788784,
0.059443291276693344,
0.04856320098042488,
-0.010568183846771717,
0.08095458149909973,
-0.01069081760942936,
-0.023674551397562027,
-0.10231949388980865,
0.03768126294016838,
-0.010868165642023087,
-0.060992851853370667,
-0.0645265281200409,
0.029283765703439713,
-0.02531951479613781,
-0.05671625956892967,
-0.008661212399601936,
0.0010363941546529531,
0.06597806513309479,
0.03826484456658363,
0.031236784532666206,
0.003571196924895048,
0.03154979646205902,
0.05382009595632553,
-0.07354585081338882,
0.05658598244190216,
-0.0497266985476017,
0.0438443161547184,
-0.01632097363471985,
-0.014589758589863777,
0.04563722759485245,
0.1568576842546463,
0.019877731800079346,
0.018369752913713455,
0.127375066280365,
-0.027206435799598694,
0.02825368009507656,
-0.10051286220550537,
-0.059589728713035583,
0.03463909402489662,
-0.03170061483979225,
0.03967662155628204,
-0.0024718004278838634,
-0.0071509648114442825,
-0.11741407215595245,
-0.0056890943087637424,
0.015394645743072033,
0.002949365181848407,
0.02710092067718506,
-0.004297060891985893,
-0.00017619515710975975,
0.04945505037903786,
-0.010298343375325203,
0.03459090739488602,
0.04363592714071274,
-0.009574498981237411,
-0.05894127115607262,
0.00889262743294239,
-0.015876688063144684,
0.09134708344936371,
0.014349912293255329,
0.0030364198610186577,
0.0019126915140077472,
-0.015307240188121796,
0.002766059013083577,
-0.005764431785792112,
0.03687266260385513,
0.00025661112158559263,
-0.008430711925029755,
0.004694168455898762,
0.027213873341679573,
0.003977652173489332,
-0.06880653649568558,
0.06248706206679344,
1.3014565602010614e-33,
0.05596921965479851,
-0.018275178968906403,
0.1273297369480133,
0.03169558197259903,
0.07240468263626099,
-0.0034538498148322105,
-0.016028763726353645,
0.019096041098237038,
-0.02015990950167179,
-0.05240572988986969,
-0.005433319602161646,
-0.017898529767990112,
-0.07341168820858002,
0.10968944430351257,
-0.01841195486485958,
0.024713074788451195,
-0.0030865429434925318,
0.015436590649187565,
-0.012590181082487106,
0.07706893235445023,
0.08343547582626343,
-0.04386221989989281,
-0.0299757681787014,
-0.013511392287909985,
0.03551109880208969,
-0.01281070988625288,
0.06327864527702332,
-0.049943190068006516,
0.013783436268568039,
-0.021716952323913574,
-0.14064262807369232,
0.047957003116607666,
-0.0030573697295039892,
0.06684493273496628,
0.019552232697606087,
-0.033789195120334625,
-0.06612113118171692,
-0.05205653980374336,
0.013720683753490448,
0.01600820943713188,
-0.00422888295724988,
0.046718403697013855,
-0.024673396721482277,
-0.027472125366330147,
-0.06789739429950714,
-0.0854247659444809,
0.06781134009361267,
0.10256906598806381,
-0.11659228056669235,
0.02940019592642784,
0.04057478904724121,
0.06684321910142899,
-0.045497827231884,
-0.07131607085466385,
-0.04388154298067093,
-0.003723959904164076,
0.07100211083889008,
0.030244022607803345,
-0.04626844450831413,
-0.020088864490389824,
0.028135310858488083,
0.05681530758738518,
-0.05232042074203491,
0.07968340069055557,
-0.017712421715259552,
0.017826305702328682,
0.04714163392782211,
0.06679899245500565,
0.07071460783481598,
0.009326259605586529,
0.01949704997241497,
0.03506558761000633,
0.004819790367037058,
-0.0863688737154007,
0.039616622030735016,
0.047899361699819565,
0.020311808213591576,
-0.1409105509519577,
0.0100602637976408,
0.0862627625465393,
-0.05119673162698746,
-0.007918653078377247,
0.005340718664228916,
-0.012057811953127384,
0.058323245495557785,
-0.03708145022392273,
0.04274551942944527,
-0.0680922195315361,
-0.023838108405470848,
-0.07376199215650558,
-0.07982659339904785,
-0.08631113916635513,
-0.03826165571808815,
0.027294637635350227,
-0.012357761152088642,
-1.1698104682940771e-33,
0.0365605466067791,
-0.0048074666410684586,
0.0035149489995092154,
0.0625571608543396,
0.013803083449602127,
-0.059341754764318466,
-0.026709700003266335,
0.038876038044691086,
0.03201889619231224,
-0.0647394061088562,
-0.05613922327756882,
-0.04331541433930397,
-0.025975743308663368,
0.025706421583890915,
0.030833443626761436,
-0.0263200830668211,
0.008953607641160488,
-0.0033354824408888817,
-0.009520860388875008,
-0.005221148021519184,
0.0482882522046566,
0.12343084067106247,
-0.061098869889974594,
0.016653725877404213,
-0.0803019255399704,
0.03114442713558674,
0.013348383828997612,
0.016855396330356598,
-0.011382379569113255,
0.057726167142391205,
0.013613137416541576,
-0.023303352296352386,
0.03236251696944237,
-0.024855833500623703,
-0.06977582722902298,
0.021715868264436722,
0.02398647367954254,
-0.0498090460896492,
-0.04042496532201767,
0.12921689450740814,
0.07123146951198578,
-0.07208702713251114,
-0.0792650580406189,
0.006103167776018381,
-0.03384563699364662,
-0.03387484326958656,
-0.017797691747546196,
-0.021489234641194344,
0.008200859650969505,
0.06784015148878098,
-0.04893367737531662,
0.07001078873872757,
-0.1414673626422882,
-0.05104506015777588,
-0.031249210238456726,
-0.09434852749109268,
-0.06745348870754242,
-0.03903329372406006,
0.0716104507446289,
-0.01245348434895277,
-0.03552445396780968,
0.02106340229511261,
0.012955904938280582,
0.035925935953855515,
-0.00474359979853034,
0.05093776062130928,
-0.0540722981095314,
-0.03693830221891403,
0.020971577614545822,
-0.020921578630805016,
0.03483108431100845,
-0.02395312674343586,
0.03542272746562958,
0.0036712184082716703,
0.008176267147064209,
0.0005426732823252678,
0.00009211908763973042,
-0.041469354182481766,
0.060778528451919556,
-0.05069809779524803,
0.024396851658821106,
-0.04470758140087128,
0.03665396571159363,
0.056104350835084915,
-0.02630503661930561,
0.10594140738248825,
0.009854682721197605,
0.054901257157325745,
0.07718898355960846,
-0.002615689765661955,
-0.06536184251308441,
-0.004113906994462013,
-0.056677721440792084,
0.037502266466617584,
-0.06669171899557114,
-3.7503671279637274e-8,
-0.0797034502029419,
0.05865948647260666,
-0.01917707920074463,
0.046692099422216415,
-0.017605461180210114,
0.005120472051203251,
-0.020737597718834877,
0.01830650493502617,
0.04875701665878296,
0.005604041274636984,
0.12235704064369202,
0.004059856291860342,
0.010830861516296864,
-0.057296935468912125,
-0.03318723663687706,
0.05021410807967186,
0.058386653661727905,
-0.02233853191137314,
0.03891441226005554,
0.03158017247915268,
0.03389585390686989,
-0.041803281754255295,
0.017411213368177414,
-0.06458692252635956,
-0.04665794223546982,
-0.06558214873075485,
-0.06555238366127014,
0.07754379510879517,
0.09373340755701065,
-0.05379441753029823,
0.011437490582466125,
-0.03240305930376053,
0.026886632665991783,
-0.02494792826473713,
0.00817146711051464,
0.013015683740377426,
-0.02677004225552082,
-0.04552371799945831,
0.04929053783416748,
-0.009234530851244926,
0.02675926685333252,
0.11163980513811111,
0.010093123652040958,
0.014880592003464699,
-0.06509795784950256,
0.03335936367511749,
-0.06084060296416283,
-0.09019552916288376,
0.002444891957566142,
0.004058670252561569,
-0.03605633229017258,
0.004049781709909439,
0.0207675788551569,
0.10312645137310028,
0.12260912358760834,
-0.017113719135522842,
0.042237572371959686,
-0.09423351287841797,
0.02742724120616913,
0.06433283537626266,
-0.012574061751365662,
-0.03413723036646843,
-0.042836397886276245,
-0.015342064201831818
] | 0 | 0 | 0 | 0 |
|
Approaching Rate-Distortion Limits in Neural Compression with Lattice Transform Coding | https://openreview.net/forum?id=Tv36j85SqR | [
"Eric Lei",
"Hamed Hassani",
"Shirin Saeedi Bidokhti"
] | Spotlight | Neural compression has brought tremendous progress in designing lossy compressors with good rate-distortion (RD) performance at low complexity. Thus far, neural compression design involves transforming the source to a latent vector, which is then rounded to integers and entropy coded. While this approach has been shown to be optimal on a few specific sources, we show that it can be highly sub-optimal on synthetic sources whose intrinsic dimensionality is greater than one. With integer rounding in the latent space, the quantization regions induced by neural transformations, remain square-like and fail to match those of optimal vector quantization. We demonstrate that this phenomenon is due to the choice of scalar quantization in the latent space, and not the transform design. By employing lattice quantization instead, we propose Lattice Transform Coding (LTC) and show that it approximately recovers optimal vector quantization at reasonable complexity. On real-world sources, LTC improves upon standard neural compressors. LTC also provides a framework that can integrate structurally (near) optimal information-theoretic designs into lossy compression; examples include block coding, which yields coding gain over optimal one-shot coding and approaches the asymptotically-achievable rate-distortion function, as well as nested lattice quantization for low complexity fixed-rate coding. | Neural compression, vector quantization, lattice quantization, nonlinear transform coding | null | 12,196 | 2403.07320 | [
-0.06984714418649673,
0.02583077736198902,
0.028408972546458244,
-0.024671537801623344,
0.008188145235180855,
0.03521231934428215,
-0.019765522330999374,
-0.09488873928785324,
0.09100355207920074,
-0.026289617642760277,
-0.02348555065691471,
-0.012807277962565422,
-0.01688138023018837,
0.03435654193162918,
-0.048003293573856354,
-0.018665127456188202,
0.018368609249591827,
0.194170281291008,
-0.06977109611034393,
-0.026153337210416794,
0.054796282202005386,
-0.030039438977837563,
0.016962850466370583,
-0.05317951738834381,
0.11857791990041733,
0.020163889974355698,
-0.0037536530289798975,
-0.05084729194641113,
0.08787354081869125,
-0.017706332728266716,
0.01605323515832424,
-0.00535019813105464,
0.008629665710031986,
0.01831403747200966,
-0.09429508447647095,
0.06134065240621567,
0.013471291400492191,
-0.0022998289205133915,
-0.006322816014289856,
0.03559387847781181,
0.01850791461765766,
0.10416824370622635,
0.016541995108127594,
0.08431412279605865,
-0.031506601721048355,
0.023304466158151627,
0.06756164133548737,
-0.05907285213470459,
-0.02435782179236412,
-0.06546949595212936,
0.009308419190347195,
0.06616564840078354,
-0.024663278833031654,
0.052540041506290436,
-0.07176193594932556,
-0.03643130138516426,
-0.017002390697598457,
0.04463982582092285,
-0.0365116223692894,
-0.015481451526284218,
-0.03650609776377678,
-0.0016657476080581546,
0.07396085560321808,
-0.040611304342746735,
-0.0008918268722482026,
-0.009253785014152527,
0.03950313851237297,
0.03984050080180168,
-0.02182845026254654,
0.029514556750655174,
-0.03261101618409157,
-0.04368167743086815,
-0.11363039910793304,
0.08837173879146576,
0.034610409289598465,
0.035920701920986176,
0.07084599137306213,
0.01738356240093708,
0.02930532954633236,
0.03531114384531975,
0.009724129922688007,
0.00035292989923618734,
0.009831080213189125,
0.018718484789133072,
0.08418095856904984,
-0.002005308400839567,
0.015438607893884182,
0.10173573344945908,
0.029804930090904236,
-0.006946024019271135,
-0.029616419225931168,
0.05561430752277374,
0.04047182947397232,
-0.004779322538524866,
0.10248243808746338,
-0.025472505018115044,
0.025671498849987984,
0.029827555641531944,
0.03220372647047043,
0.03666733577847481,
0.06989546120166779,
-0.01470625214278698,
-0.01100032590329647,
0.0214876476675272,
0.006306762341409922,
-0.08580829948186874,
0.06478307396173477,
0.09726537019014359,
0.02164580672979355,
-0.02543368749320507,
0.04722577705979347,
0.0155385946854949,
0.010249188169836998,
-0.07179708033800125,
0.0793854370713234,
-0.044139742851257324,
-0.05337430536746979,
0.04710150137543678,
0.008988089859485626,
-0.0036328439600765705,
-0.05383238196372986,
-0.010089104063808918,
-0.13665355741977692,
-0.011483214795589447,
-0.020715495571494102,
-0.00966156367212534,
-0.1369822919368744,
2.891933385806737e-33,
-0.004298197105526924,
0.0986969843506813,
0.009377229027450085,
0.005463294219225645,
0.08277805149555206,
0.034753043204545975,
0.021226828917860985,
-0.01685963198542595,
0.06287364661693573,
0.08375932276248932,
-0.06357316672801971,
-0.03549777716398239,
-0.06806038320064545,
0.030586276203393936,
0.03266517445445061,
0.009472963400185108,
-0.05544161796569824,
0.025407329201698303,
-0.00028853188268840313,
-0.07582683861255646,
0.021619373932480812,
-0.052752651274204254,
-0.010520519688725471,
0.03281060978770256,
-0.03601697087287903,
-0.05286407098174095,
0.03920572251081467,
-0.05445365607738495,
-0.07843261957168579,
-0.030745042487978935,
-0.14096200466156006,
-0.005127643700689077,
0.013507428579032421,
-0.057829175144433975,
0.026255333796143532,
-0.04137678071856499,
0.06521525979042053,
-0.021596897393465042,
-0.003370958846062422,
-0.021117977797985077,
0.002446359721943736,
0.046349894255399704,
0.012316727079451084,
-0.005365447141230106,
-0.028873538598418236,
-0.02679555118083954,
0.09453821927309036,
0.027774784713983536,
-0.06349316239356995,
-0.11456920206546783,
0.055306654423475266,
0.01210783515125513,
-0.009283614344894886,
0.013596441596746445,
0.0888359472155571,
-0.021273594349622726,
0.07856469601392746,
-0.03559551388025284,
0.029584871605038643,
0.1264737844467163,
-0.06530101597309113,
0.018904654309153557,
0.0400422103703022,
-0.00009855580719886348,
-0.033619340509176254,
-0.04666730388998985,
-0.025418393313884735,
-0.035643309354782104,
0.003469964489340782,
0.03838766738772392,
-0.038087885826826096,
0.04924885928630829,
-0.06356050819158554,
-0.05274425074458122,
0.06715568900108337,
-0.029328666627407074,
0.05066267028450966,
-0.0895310789346695,
-0.025591647252440453,
-0.006167866755276918,
-0.05737932398915291,
0.020784450694918633,
-0.007727721240371466,
0.00503157265484333,
-0.006433498580008745,
-0.009615601040422916,
0.029529457911849022,
-0.04861018806695938,
-0.044748686254024506,
0.05804630368947983,
-0.09939856827259064,
-0.09955764561891556,
0.04527220502495766,
-0.07322775572538376,
-0.04394655302166939,
-2.63775752802538e-33,
-0.10762961208820343,
0.036462292075157166,
-0.07072044909000397,
0.035394374281167984,
-0.038334235548973083,
-0.07565204054117203,
0.018621614202857018,
-0.04597654193639755,
0.06618764251470566,
-0.03657640144228935,
0.004845238756388426,
-0.060234326869249344,
0.0029104172717779875,
0.013435399159789085,
-0.003105951938778162,
0.06346802413463593,
-0.009767349809408188,
0.006915771868079901,
-0.034064892679452896,
-0.040095675736665726,
0.04293233901262283,
0.08470059931278229,
-0.016372600570321083,
0.02283504605293274,
0.024816129356622696,
0.03978157788515091,
-0.05461630970239639,
0.12300249189138412,
0.06217991188168526,
0.01612197421491146,
0.0018276391783729196,
-0.060399726033210754,
-0.03804437816143036,
-0.04908757284283638,
-0.04363233223557472,
-0.024175390601158142,
0.008511265739798546,
0.03681117296218872,
-0.0316871702671051,
0.006624900735914707,
0.047048479318618774,
-0.0908617302775383,
-0.036377426236867905,
0.08138827979564667,
0.04090169444680214,
-0.04422256350517273,
-0.015336706303060055,
-0.05504684895277023,
-0.00028549355920404196,
0.09760691970586777,
0.018687333911657333,
0.049608007073402405,
-0.04718911647796631,
0.08013980090618134,
-0.025438692420721054,
-0.06368426978588104,
-0.11368739604949951,
0.02107623778283596,
0.1307990998029709,
0.0463724322617054,
-0.049575816839933395,
-0.057173799723386765,
0.033728815615177155,
-0.11182630807161331,
-0.032311975955963135,
-0.06354279816150665,
0.0212996955960989,
0.0530712753534317,
0.02793286181986332,
0.03728867322206497,
0.03665504604578018,
0.025889497250318527,
0.05209629237651825,
0.07342153042554855,
-0.03350000083446503,
-0.037761151790618896,
0.04815170541405678,
0.06136900186538696,
0.0026641651056706905,
0.06863069534301758,
0.03347126394510269,
0.08422794938087463,
0.021924126893281937,
-0.020328635349869728,
0.026053037494421005,
0.05490994453430176,
0.07307661324739456,
0.06238611787557602,
0.05107611417770386,
0.011351990513503551,
-0.0036237966269254684,
0.01355279702693224,
-0.03555141016840935,
0.013013855554163456,
-0.039629001170396805,
-4.076650483852973e-8,
-0.07984918355941772,
0.047302376478910446,
-0.007842677645385265,
-0.048268891870975494,
0.020780570805072784,
-0.04351428523659706,
0.011552268639206886,
0.047916900366544724,
0.021848758682608604,
-0.07727128267288208,
0.14388567209243774,
-0.02766759879887104,
-0.025322824716567993,
-0.03466763719916344,
0.009499904699623585,
0.007501560263335705,
-0.0423906147480011,
-0.0043713427148759365,
0.03758351504802704,
-0.021253235638141632,
-0.08150289952754974,
0.07960464805364609,
-0.02865692786872387,
-0.019941430538892746,
-0.011117279529571533,
-0.09441397339105606,
-0.03068007342517376,
0.0022822467144578695,
0.07466686517000198,
-0.0014744416112080216,
-0.05216247960925102,
0.03195728361606598,
0.10071122646331787,
0.004807406570762396,
-0.03821367397904396,
-0.04315343499183655,
0.010359617881476879,
-0.042666394263505936,
0.01378779485821724,
-0.07187268137931824,
-0.033591724932193756,
-0.021340729668736458,
-0.02656492218375206,
-0.0027685672976076603,
-0.031188547611236572,
-0.027083424851298332,
0.04847490042448044,
-0.008026570081710815,
-0.03988143056631088,
0.04923273250460625,
0.06888075917959213,
0.02639688178896904,
-0.00002811042759276461,
-0.011173989623785019,
0.08580244332551956,
-0.041877053678035736,
-0.04421648755669594,
-0.039072051644325256,
-0.004657404497265816,
0.02600274421274662,
-0.07740363478660583,
0.01433571521192789,
-0.05145493522286415,
-0.016465172171592712
] | 0 | 0 | 0 | 0 |
|
Wasserstein Distances, Neuronal Entanglement, and Sparsity | https://openreview.net/forum?id=cnKhHxN3xj | [
"Shashata Sawmya",
"Linghao Kong",
"Ilia Markov",
"Dan Alistarh",
"Nir N Shavit"
] | Spotlight | Disentangling polysemantic neurons is at the core of many current approaches to interpretability of large language models. Here we attempt to study how disentanglement can be used to understand performance, particularly under weight sparsity, a leading post-training optimization technique. We suggest a novel measure for estimating neuronal entanglement: the Wasserstein distance of a neuron's output distribution to a Gaussian. Moreover, we show the existence of a small number of highly entangled "Wasserstein Neurons" in each linear layer of an LLM, characterized by their highly non-Gaussian output distributions, their role in mapping similar inputs to dissimilar outputs, and their significant impact on model accuracy. To study these phenomena, we propose a new experimental framework for disentangling polysemantic neurons. Our framework separates each layer's inputs to create a mixture of experts where each neuron's output is computed by a mixture of neurons of lower Wasserstein distance, each better at maintaining accuracy when sparsified without retraining. We provide strong evidence that this is because the mixture of sparse experts is effectively disentangling the input-output relationship of individual neurons, in particular the difficult Wasserstein neurons. | Polysemanticity, Disentanglement, Wasserstein Distance, Sparsity, Large Language Models | We show that the Wasserstein distance of a neuron's output distribution to a Gaussian is a pertinent indicator for its degree of entanglement, and propose a new framework for better investigating disentaglement under sparsity. | 12,168 | 2405.15756 | [
-0.016371311619877815,
-0.12515975534915924,
0.017893295735120773,
0.02754191868007183,
0.018677080050110817,
0.04563787579536438,
0.020552320405840874,
-0.045210063457489014,
0.039528317749500275,
-0.1079670637845993,
-0.03699371963739395,
-0.04461653158068657,
-0.004750108812004328,
0.0441366583108902,
-0.14945179224014282,
0.03262782096862793,
0.17346687614917755,
0.08526083827018738,
-0.10826948285102844,
0.0036616288125514984,
0.05215511843562126,
-0.021273652091622353,
-0.012212869711220264,
0.011860010214149952,
0.06144542619585991,
-0.008596888743340969,
0.01707223430275917,
-0.04812539368867874,
-0.03325798735022545,
0.018508637323975563,
0.032100655138492584,
-0.06993057578802109,
-0.01516338624060154,
0.059597257524728775,
-0.042092908173799515,
0.10500257462263107,
-0.0525641143321991,
0.05741060897707939,
0.05910046398639679,
-0.029514634981751442,
-0.008610780350863934,
0.11385619640350342,
0.05797464773058891,
0.060513392090797424,
0.03915071487426758,
0.03773803636431694,
-0.00007990514859557152,
-0.06833766400814056,
0.006037626415491104,
-0.008237901143729687,
-0.01716873236000538,
0.019934313371777534,
-0.000655206385999918,
0.025507446378469467,
0.004955063574016094,
-0.019899435341358185,
0.04516202211380005,
0.07250290364027023,
-0.06816242635250092,
0.029608730226755142,
-0.01861819252371788,
-0.10636740922927856,
0.011705095879733562,
-0.02461249753832817,
0.024771293625235558,
0.012408650480210781,
0.04407171905040741,
0.055301643908023834,
-0.06722629070281982,
0.02290765568614006,
0.0208489540964365,
0.12565472722053528,
-0.131475031375885,
0.00035411957651376724,
0.031575269997119904,
-0.017204293981194496,
0.035922661423683167,
0.01854102872312069,
0.07406090945005417,
-0.04777652025222778,
0.01772836595773697,
0.046962372958660126,
0.06321640312671661,
0.0036428594030439854,
0.15193293988704681,
0.023762574419379234,
0.005761970765888691,
0.08600756525993347,
0.00810574647039175,
-0.01011175848543644,
-0.0811195969581604,
-0.04996844008564949,
-0.08221685141324997,
-0.0033040756825357676,
0.006383747328072786,
-0.0061530242674052715,
0.01371176727116108,
0.02883812226355076,
-0.03327963501214981,
0.044615570455789566,
0.012243103235960007,
0.031000591814517975,
-0.0035567847080528736,
-0.0809859186410904,
-0.0027733196038752794,
-0.037473201751708984,
0.05076386407017708,
0.02745235152542591,
0.023584436625242233,
-0.025993365794420242,
0.029819926247000694,
0.010529606603085995,
0.051062002778053284,
0.09094226360321045,
0.005009844433516264,
0.04626774042844772,
0.0911923423409462,
0.0026117407251149416,
0.023338166996836662,
0.06670328974723816,
-0.049419794231653214,
0.030881522223353386,
-0.008529654704034328,
0.042222872376441956,
-0.0010308589553460479,
-0.026531299576163292,
-0.026652691885828972,
2.081531233631536e-33,
0.018535954877734184,
0.005483962129801512,
-0.013885816559195518,
0.06570372730493546,
0.025606954470276833,
0.029370427131652832,
0.053905922919511795,
-0.018098685890436172,
0.02447778917849064,
0.005146494135260582,
-0.0719894990324974,
0.013291284441947937,
-0.05077153816819191,
0.04694873094558716,
0.06158263981342316,
-0.01384708657860756,
-0.054988641291856766,
-0.020721185952425003,
-0.0013010866241529584,
-0.06251120567321777,
0.03687315434217453,
-0.03186056390404701,
-0.07173073291778564,
-0.03654387593269348,
-0.07269700616598129,
-0.07705314457416534,
0.04144641011953354,
-0.04801502078771591,
-0.08249429613351822,
-0.017563682049512863,
-0.14070245623588562,
-0.006375832948833704,
0.019713420420885086,
0.013027515262365341,
0.045995041728019714,
-0.07982195913791656,
0.03492646664381027,
-0.013337492011487484,
0.017990464344620705,
-0.0845377966761589,
-0.02049194648861885,
0.03995642066001892,
0.02439666911959648,
-0.12725184857845306,
-0.06360944360494614,
-0.03905403986573219,
0.0324564054608345,
-0.06178602576255798,
-0.019012415781617165,
-0.15610510110855103,
0.07208694517612457,
0.0260788444429636,
0.02364828996360302,
-0.02779434435069561,
-0.01780579797923565,
-0.053599193692207336,
0.013826281763613224,
0.031423646956682205,
0.04572657123208046,
0.11275146156549454,
-0.039072345942258835,
0.07324106246232986,
0.02431909367442131,
0.061973631381988525,
0.03280934691429138,
0.031216196715831757,
-0.05132989212870598,
0.04575544595718384,
0.04890502989292145,
-0.007637254428118467,
-0.05430928245186806,
-0.030043579638004303,
-0.0007586174760945141,
-0.08068302273750305,
0.0010443634819239378,
-0.009984256699681282,
-0.010201253928244114,
-0.049397166818380356,
-0.029352547600865364,
0.002179564908146858,
0.023273438215255737,
-0.013643513433635235,
-0.08015946298837662,
-0.050412800163030624,
-0.10963979363441467,
-0.030765827745199203,
0.07688241451978683,
-0.039925117045640945,
-0.06196358799934387,
0.02724132314324379,
-0.0035949095617979765,
-0.03440673649311066,
0.03726082667708397,
-0.059135351330041885,
0.023957327008247375,
-3.0503488819953027e-33,
-0.13002946972846985,
0.04968344792723656,
0.0038539813831448555,
0.02828330732882023,
-0.03777695819735527,
0.020267168059945107,
-0.007066361606121063,
0.06753493845462799,
-0.039593156427145004,
-0.011716886423528194,
-0.020735567435622215,
-0.04032043367624283,
0.05648628994822502,
0.03795161098241806,
0.00824996829032898,
0.026015814393758774,
-0.04838626831769943,
0.058270398527383804,
0.06566042453050613,
-0.05315834656357765,
0.05222565308213234,
0.07405289262533188,
-0.0808883011341095,
0.043293297290802,
-0.03431946784257889,
0.07437731325626373,
-0.015741217881441116,
0.09892091900110245,
-0.01990557089447975,
0.03095618635416031,
-0.044485654681921005,
-0.0111424895003438,
-0.08424419909715652,
0.01993137039244175,
-0.08110784739255905,
0.07477361708879471,
0.028169361874461174,
0.03791610896587372,
-0.034217823296785355,
0.012731594033539295,
0.022963806986808777,
-0.014039109461009502,
-0.034544795751571655,
0.05729250609874725,
0.006771543994545937,
-0.029966851696372032,
-0.10412167012691498,
-0.07159678637981415,
-0.018033158034086227,
0.0012659078929573298,
-0.031926851719617844,
0.06208394095301628,
-0.10748697072267532,
-0.00021049455972388387,
-0.11234767735004425,
-0.04773515462875366,
-0.0510491319000721,
-0.033874668180942535,
0.04616745561361313,
-0.05105435848236084,
-0.08936844766139984,
-0.04104723781347275,
0.019596455618739128,
0.014812729321420193,
0.0057401773519814014,
0.001969932811334729,
-0.0121610714122653,
0.03574944660067558,
0.02825953997671604,
0.02305261790752411,
0.019513653591275215,
-0.02668091654777527,
0.05402476340532303,
0.08180983364582062,
0.006148101761937141,
0.010269684717059135,
0.013601506128907204,
-0.022752195596694946,
-0.021165451034903526,
0.03667745739221573,
0.04865908622741699,
-0.013671998865902424,
-0.01423840131610632,
0.025770502164959908,
0.1294623166322708,
0.04324488714337349,
0.02011093497276306,
0.02077558822929859,
0.02478446252644062,
-0.019061889499425888,
0.024177715182304382,
0.005030313041061163,
-0.002440025797113776,
-0.007643718738108873,
-0.03495163843035698,
-4.136700937351634e-8,
-0.03599880263209343,
-0.060922201722860336,
-0.009863374754786491,
-0.042104579508304596,
-0.04947924241423607,
-0.08529321104288101,
0.021120764315128326,
0.06596032530069351,
-0.07805238664150238,
-0.018608739599585533,
0.12388669699430466,
-0.029828175902366638,
-0.1311652958393097,
-0.0027802004478871822,
0.015021897852420807,
0.036142654716968536,
-0.0008414723561145365,
0.04555020481348038,
0.03218045085668564,
-0.030705243349075317,
0.0573042593896389,
0.0026542814448475838,
-0.01226862519979477,
0.0017376320902258158,
-0.004222309682518244,
-0.06986773759126663,
0.007069576997309923,
0.088534876704216,
0.03223756328225136,
0.042868923395872116,
0.01387039851397276,
0.009598162956535816,
0.026295840740203857,
-0.011803449131548405,
-0.06610609591007233,
0.11558825522661209,
-0.042104821652173996,
-0.002238902961835265,
0.0026455800980329514,
0.02821243554353714,
-0.022988051176071167,
0.01791706122457981,
0.021750997751951218,
0.04143929108977318,
0.02631944604218006,
-0.03865910694003105,
0.03398946672677994,
-0.07081633061170578,
0.055700499564409256,
-0.04064830020070076,
0.038142796605825424,
0.002027906011790037,
0.019263602793216705,
0.032572053372859955,
0.027697952464222908,
-0.0004018763720523566,
-0.07650099694728851,
-0.053998760879039764,
-0.02833467908203602,
0.0369705967605114,
-0.019703270867466927,
0.07285445928573608,
-0.05508864298462868,
0.022972099483013153
] | https://github.com/shavit-lab/sparse-expansion | 6 | 0 | 0 | 0 |
Walk the Talk? Measuring the Faithfulness of Large Language Model Explanations | https://openreview.net/forum?id=4ub9gpx9xw | [
"Katie Matton",
"Robert Ness",
"John Guttag",
"Emre Kiciman"
] | Spotlight | Large language models (LLMs) are capable of generating *plausible* explanations of how they arrived at an answer to a question. However, these explanations can misrepresent the model's "reasoning" process, i.e., they can be *unfaithful*. This, in turn, can lead to over-trust and misuse. We introduce a new approach for measuring the faithfulness of LLM explanations. First, we provide a rigorous definition of faithfulness. Since LLM explanations mimic human explanations, they often reference high-level *concepts* in the input question that purportedly influenced the model. We define faithfulness in terms of the difference between the set of concepts that the LLM's *explanations imply* are influential and the set that *truly* are. Second, we present a novel method for estimating faithfulness that is based on: (1) using an auxiliary LLM to modify the values of concepts within model inputs to create realistic counterfactuals, and (2) using a hierarchical Bayesian model to quantify the causal effects of concepts at both the example- and dataset-level. Our experiments show that our method can be used to quantify and discover interpretable patterns of unfaithfulness. On a social bias task, we uncover cases where LLM explanations hide the influence of social bias. On a medical question answering task, we uncover cases where LLM explanations provide misleading claims about which pieces of evidence influenced the model's decisions. | large language models, faithful explanations, explainability, safety, counterfactual reasoning | We introduce a novel method for measuring the faithfulness of explanations given by LLMs. | 12,125 | null | [
-0.08044278621673584,
-0.10459448397159576,
0.04485197365283966,
0.038708802312612534,
0.08558410406112671,
-0.03470901772379875,
0.015385672450065613,
0.02522071823477745,
0.06843479722738266,
0.00011725824879249558,
-0.06865672022104263,
-0.07085546851158142,
0.07014018297195435,
-0.02494415082037449,
-0.027958719059824944,
-0.020781259983778,
0.13057082891464233,
0.01911383494734764,
-0.10925599932670593,
-0.09781424701213837,
0.06662292033433914,
-0.036181651055812836,
0.05114467442035675,
0.04132666438817978,
-0.018586285412311554,
-0.04805333912372589,
0.028386415913701057,
0.014019669033586979,
-0.04375802353024483,
0.03445924445986748,
-0.029537325724959373,
0.05013909935951233,
-0.055286768823862076,
0.030437860637903214,
-0.06792055070400238,
0.03627597168087959,
-0.01792113669216633,
0.036769360303878784,
-0.007097027264535427,
-0.03391660377383232,
-0.07026125490665436,
0.039464715868234634,
0.08577080816030502,
-0.011448309756815434,
0.05319366976618767,
-0.06498955190181732,
-0.006830455735325813,
0.03167925029993057,
-0.1002303957939148,
0.02134544774889946,
-0.11886821687221527,
-0.025233400985598564,
-0.015591803006827831,
0.015177411958575249,
-0.0189137514680624,
-0.08182168751955032,
0.07315577566623688,
0.0707017183303833,
-0.016106214374303818,
-0.03313388302922249,
-0.030681023374199867,
-0.07870351523160934,
-0.06977656483650208,
0.015508141368627548,
-0.02350715361535549,
0.06393381208181381,
-0.0231262668967247,
0.05338707193732262,
-0.08465714007616043,
0.07261776179075241,
0.026437101885676384,
0.05337780341506004,
-0.013880988582968712,
-0.01945548504590988,
-0.05297812819480896,
-6.712065214742324e-8,
0.022445229813456535,
0.053053341805934906,
0.012833693996071815,
-0.07124185562133789,
-0.006556080654263496,
0.011721239425241947,
0.03247543424367905,
0.007894852198660374,
0.05069417506456375,
-0.031902462244033813,
0.018582476302981377,
0.03309446573257446,
0.006537866313010454,
0.04963982477784157,
-0.0378517284989357,
-0.06127063184976578,
-0.009795348159968853,
-0.011833487078547478,
0.08333931863307953,
0.0695270448923111,
-0.00280989077873528,
-0.04915481060743332,
-0.009878907352685928,
0.022295087575912476,
-0.007753623649477959,
0.13398244976997375,
0.00951781589537859,
-0.12423723936080933,
0.04883251339197159,
-0.007626368198543787,
-0.05988164618611336,
-0.018686829134821892,
0.016167229041457176,
-0.04123510792851448,
-0.03157317265868187,
0.05085005983710289,
-0.003932460211217403,
0.06914213299751282,
0.08048537373542786,
-0.05838926509022713,
0.11772260069847107,
-0.05310296267271042,
-0.059782445430755615,
-0.006507426034659147,
0.040026742964982986,
0.06724458187818527,
0.08840261399745941,
0.028390949591994286,
-0.000138430084916763,
-0.05239636451005936,
-0.05475896969437599,
-9.924970637061201e-34,
0.06136117875576019,
0.024589696899056435,
0.03737029433250427,
0.1387142837047577,
0.03637591004371643,
-0.01224907860159874,
-0.0410156324505806,
0.014520282857120037,
0.0671592578291893,
0.015293795615434647,
-0.03940879926085472,
0.06968653202056885,
-0.0732906237244606,
-0.02113429084420204,
0.02042049542069435,
0.11360899358987808,
-0.08400719612836838,
0.03624575212597847,
-0.05569223314523697,
-0.04509284719824791,
0.034330181777477264,
-0.020257407799363136,
0.010564419440925121,
-0.07876206934452057,
0.018411679193377495,
-0.04426224157214165,
0.018002372235059738,
-0.015426555648446083,
-0.04510076344013214,
0.010938260704278946,
-0.07329228520393372,
-0.07876421511173248,
0.040456660091876984,
0.01082136482000351,
-0.01897095888853073,
-0.09635154157876968,
0.004079992882907391,
-0.07068269699811935,
0.07233338803052902,
-0.06274870783090591,
0.014771905727684498,
-0.062224067747592926,
0.016874542459845543,
0.04573040083050728,
0.02148621156811714,
-0.033979933708906174,
-0.0240806695073843,
-0.061596695333719254,
-0.015805698931217194,
-0.035725608468055725,
0.09361313283443451,
0.05069120228290558,
0.02561444789171219,
-0.033653922379016876,
-0.02634129300713539,
0.009423800744116306,
-0.010511546395719051,
0.006882102694362402,
-0.027090132236480713,
0.042605843394994736,
0.05154353007674217,
0.03023041971027851,
-0.01715543307363987,
0.062484998255968094,
0.0011909210588783026,
0.004234326537698507,
-0.038404159247875214,
0.04152687266469002,
-0.02220051735639572,
0.023581815883517265,
-0.023027585819363594,
-0.003944482654333115,
-0.08195748925209045,
-0.01626870408654213,
-0.001145040849223733,
0.011553797870874405,
-0.04043331369757652,
-0.01046786829829216,
0.0794702097773552,
0.014396727085113525,
0.01192724984139204,
-0.07602427899837494,
0.0015374138019979,
-0.010786225087940693,
-0.08299954235553741,
-0.018852708861231804,
0.03088252991437912,
-0.038697827607393265,
-0.0042244684882462025,
-0.0356874018907547,
-0.001557711511850357,
-0.04900334030389786,
0.009411129169166088,
-0.005486628971993923,
-0.027075715363025665,
-1.8069269370718184e-33,
-0.07717124372720718,
-0.026027308776974678,
0.024796849116683006,
0.07794849574565887,
-0.04409807175397873,
-0.031207187101244926,
-0.04941066727042198,
0.04602727293968201,
0.02358766831457615,
-0.03847944363951683,
-0.060891833156347275,
-0.021144328638911247,
0.06840734928846359,
0.026722857728600502,
0.04858386144042015,
-0.0807139202952385,
0.08535972237586975,
0.03708530589938164,
-0.002118469914421439,
0.05195518583059311,
0.05759314447641373,
0.017939968034625053,
-0.0729510486125946,
-0.005372639279812574,
-0.020165473222732544,
-0.004188092425465584,
-0.0016747090267017484,
0.036475349217653275,
0.031697750091552734,
0.03769010305404663,
-0.007415966130793095,
0.04101310297846794,
-0.012716684490442276,
-0.029462773352861404,
-0.0821850523352623,
0.05298225209116936,
0.06240909546613693,
0.030379049479961395,
-0.05975855886936188,
0.021654192358255386,
0.06986648589372635,
-0.04154065251350403,
-0.138815775513649,
-0.0779070109128952,
-0.08804885298013687,
-0.07952960580587387,
-0.03384366258978844,
-0.000660842633806169,
0.0074478150345385075,
-0.06195788457989693,
-0.0017515795771032572,
-0.0005973510560579598,
0.03122149407863617,
0.09420118480920792,
-0.03143046796321869,
-0.008954828605055809,
-0.03446321189403534,
-0.044428758323192596,
0.07993172109127045,
0.013861842453479767,
-0.05510862171649933,
0.04279693216085434,
-0.003973710350692272,
-0.058657050132751465,
0.01652725227177143,
0.006205846555531025,
-0.04170089215040207,
-0.004848617594689131,
0.045677393674850464,
0.01874884031713009,
0.08733654022216797,
-0.02275925874710083,
-0.06567735970020294,
0.015543533489108086,
0.018218448385596275,
0.03137364611029625,
-0.02893879823386669,
-0.11317763477563858,
-0.02237090840935707,
-0.07992008328437805,
0.06943777948617935,
-0.048304226249456406,
0.07086692005395889,
0.014307009987533092,
0.025222815573215485,
-0.05404289439320564,
0.025408245623111725,
0.06898427754640579,
0.011357469484210014,
0.04392058402299881,
-0.005206070840358734,
-0.022856328636407852,
-0.10804753005504608,
0.03404718264937401,
-0.07626569271087646,
-4.6550848509241405e-8,
0.010888966731727123,
0.046035945415496826,
0.019999057054519653,
0.06774041056632996,
0.07285982370376587,
-0.03738585487008095,
-0.08027710020542145,
0.03716351464390755,
-0.029679354280233383,
-0.011381329968571663,
0.07099085301160812,
0.005618261639028788,
-0.022869782522320747,
0.002362776780501008,
-0.026722978800535202,
0.10846994817256927,
-0.02952129766345024,
-0.004277655854821205,
0.023712866008281708,
0.015865541994571686,
0.08968239277601242,
0.019455688074231148,
-0.020134100690484047,
0.02373644895851612,
0.06285399943590164,
-0.010358286090195179,
-0.06297269463539124,
0.13833723962306976,
0.00808950886130333,
0.0597851537168026,
0.007812945172190666,
-0.03877793624997139,
-0.11143463104963303,
0.017974581569433212,
-0.007006281055510044,
0.06999272853136063,
0.0018059257417917252,
-0.052382007241249084,
0.05078593268990517,
-0.0409773588180542,
0.04160277545452118,
0.07740694284439087,
-0.06689032912254333,
0.0630822479724884,
0.11479569971561432,
0.0003897397837135941,
-0.05708559975028038,
-0.0637398511171341,
-0.025424139574170113,
-0.004405620042234659,
-0.07502569258213043,
-0.025836901739239693,
-0.022739192470908165,
0.06552916765213013,
0.00551685132086277,
0.0029291431419551373,
-0.025827743113040924,
0.038351770490407944,
-0.035203978419303894,
0.048379041254520416,
0.08206810057163239,
0.10079821944236755,
0.043209511786699295,
0.007416488137096167
] | 0 | 0 | 0 | 0 |
|
Implicit Bias of Mirror Flow for Shallow Neural Networks in Univariate Regression | https://openreview.net/forum?id=IF0Q9KY3p2 | [
"Shuang Liang",
"Guido Montufar"
] | Spotlight | We examine the implicit bias of mirror flow in least squares error regression with wide and shallow neural networks. For a broad class of potential functions, we show that mirror flow exhibits lazy training and has the same implicit bias as ordinary gradient flow when the network width tends to infinity. For univariate ReLU networks, we characterize this bias through a variational problem in function space. Our analysis includes prior results for ordinary gradient flow as a special case and lifts limitations which required either an intractable adjustment of the training data or networks with skip connections. We further introduce \emph{scaled potentials} and show that for these, mirror flow still exhibits lazy training but is not in the kernel regime. For univariate networks with absolute value activations, we show that mirror flow with scaled potentials induces a rich class of biases, which generally cannot be captured by an RKHS norm. A takeaway is that whereas the parameter initialization determines how strongly the curvature of the learned function is penalized at different locations of the input space, the scaled potential determines how the different magnitudes of the curvature are penalized. | implicit bias, overparametrized neural network, mirror descent, univariate regression, lazy training | Univariate wide shallow networks trained with mirror flow are biased towards smooth interpolants and scaled mirror potentials induce biases affecting the different magnitudes of the curvature. | 12,117 | 2410.03988 | [
-0.059086039662361145,
-0.11168239265680313,
0.1044759526848793,
0.04971418157219887,
0.035830359905958176,
0.021190837025642395,
-0.05167538672685623,
-0.02801271714270115,
0.021437158808112144,
-0.06843055039644241,
-0.01820746250450611,
0.044085364788770676,
-0.022705387324094772,
-0.030233286321163177,
-0.03879927098751068,
-0.02321796305477619,
0.08644299954175949,
0.1086127981543541,
-0.0755748599767685,
-0.04599878937005997,
0.040228139609098434,
-0.030862420797348022,
-0.06176867336034775,
0.021077079698443413,
0.061732660979032516,
-0.05197382718324661,
0.00022561363584827632,
0.030757257714867592,
-0.04721646383404732,
-0.06381671130657196,
0.04995874688029289,
0.002252239966765046,
-0.047411393374204636,
-0.0437878780066967,
-0.10079822689294815,
0.05031654238700867,
-0.020578131079673767,
0.040621332824230194,
0.021587450057268143,
-0.014593291096389294,
0.01058198232203722,
-0.03202958405017853,
0.025368234142661095,
0.04508338123559952,
0.0635821744799614,
-0.009913601912558079,
0.04390677064657211,
-0.06911095231771469,
-0.02724926918745041,
-0.006236768327653408,
0.031470928341150284,
-0.024181948974728584,
-0.05271702632308006,
0.023407042026519775,
0.049034666270017624,
-0.04912072420120239,
-0.007368167862296104,
0.10541584342718124,
-0.006586928386241198,
0.07124994695186615,
0.014755085110664368,
-0.16812659800052643,
0.016296476125717163,
0.01714288257062435,
-0.04168158024549484,
0.010732037015259266,
0.015357955358922482,
0.010811206884682178,
0.007065195124596357,
0.04548465833067894,
0.024179792031645775,
0.06492015719413757,
-0.11220785975456238,
-0.03438024967908859,
0.05871643126010895,
-0.004662093706429005,
0.09404724836349487,
-0.0015820697881281376,
-0.00590856745839119,
-0.033166900277137756,
0.03844200819730759,
0.017001817002892494,
0.07790584862232208,
-0.009485145099461079,
0.04994508996605873,
0.01000978983938694,
-0.05439482629299164,
-0.018241198733448982,
0.033579885959625244,
0.026592329144477844,
0.012587489560246468,
0.011372104287147522,
-0.07250633090734482,
-0.0446617566049099,
0.09290558844804764,
-0.0675300657749176,
0.027617178857326508,
-0.04957490414381027,
-0.08082376420497894,
0.1067744642496109,
-0.01581364870071411,
0.058934446424245834,
0.01556057296693325,
0.09214752912521362,
0.05441424250602722,
-0.013090918771922588,
0.0639987587928772,
0.03400110453367233,
0.0212783794850111,
-0.09385009109973907,
-0.023266103118658066,
0.018252380192279816,
0.032098911702632904,
0.004724621307104826,
-0.01864018477499485,
-0.11447477340698242,
0.018699798732995987,
-0.08897443860769272,
-0.06644915789365768,
0.062246065586805344,
-0.06989507377147675,
-0.004246431868523359,
-0.07385387271642685,
0.03185408189892769,
-0.014824734069406986,
0.024184735491871834,
-0.07568056881427765,
3.76206093657928e-33,
-0.00462538143619895,
0.06611226499080658,
0.03860311210155487,
-0.09386846423149109,
0.01878488063812256,
-0.04120919108390808,
0.07500994950532913,
0.0013211107579991221,
0.013068417087197304,
-0.022014345973730087,
-0.0806434154510498,
0.009132441133260727,
-0.07636339217424393,
0.05184273421764374,
-0.022242886945605278,
0.05236515402793884,
0.00024409643083345145,
0.06514029204845428,
0.016374817118048668,
0.003194360760971904,
0.013024430721998215,
-0.06802991777658463,
-0.01126447506248951,
-0.0672103688120842,
-0.07352962344884872,
-0.05261003226041794,
0.0344742126762867,
0.009031254798173904,
-0.0992802083492279,
0.023030411452054977,
-0.0733380988240242,
-0.04683966562151909,
0.037073440849781036,
0.018650105223059654,
0.027152130380272865,
-0.045874036848545074,
0.0426953099668026,
0.026215583086013794,
0.06135934218764305,
0.022255484014749527,
-0.003032483160495758,
0.06710163503885269,
0.02445414662361145,
-0.0427066907286644,
-0.01752602681517601,
-0.09520658850669861,
0.033323004841804504,
0.07308974117040634,
-0.004007464274764061,
-0.031069034710526466,
-0.0012472773669287562,
0.038440946489572525,
-0.03026244230568409,
-0.05987292528152466,
-0.016785025596618652,
-0.0022362752351909876,
0.005475526209920645,
0.04766114428639412,
-0.03023064136505127,
0.0707743689417839,
-0.011585881933569908,
0.05422133579850197,
-0.04018986597657204,
0.0030330789741128683,
0.003636367851868272,
0.040876537561416626,
-0.045625776052474976,
-0.01588558964431286,
0.06994370371103287,
0.029118137434124947,
0.0034190949518233538,
0.08782470226287842,
-0.0794917494058609,
-0.09871279448270798,
0.08202537149190903,
0.04235062003135681,
-0.008297417312860489,
-0.09774958342313766,
0.05527864024043083,
0.09664198756217957,
-0.029422428458929062,
0.040224045515060425,
-0.031143970787525177,
-0.05620342120528221,
-0.035921577364206314,
-0.03129342198371887,
0.021287890151143074,
-0.15747219324111938,
-0.02569187618792057,
-0.0020844663958996534,
-0.03541174158453941,
-0.04500246420502663,
-0.04298568144440651,
-0.03505440801382065,
0.0020081126131117344,
-3.784734018554734e-33,
0.04011274129152298,
0.06221342459321022,
-0.011450633406639099,
0.020566359162330627,
-0.017002636566758156,
0.08094393461942673,
0.02014699950814247,
0.11060964316129684,
-0.021243363618850708,
-0.010749305598437786,
0.007268455345183611,
-0.025769559666514397,
-0.01414395309984684,
0.01952221989631653,
0.045228488743305206,
-0.034993644803762436,
-0.039313312619924545,
-0.03133956715464592,
-0.0005314664449542761,
-0.03481082618236542,
0.05763021111488342,
0.08069377392530441,
0.020073285326361656,
0.08075456321239471,
-0.048295799642801285,
-0.007693430408835411,
-0.058990564197301865,
0.0786733329296112,
-0.03352632001042366,
0.025416629388928413,
-0.03268247842788696,
-0.016009604558348656,
-0.06963891535997391,
0.010254881344735622,
-0.018234556540846825,
0.13186068832874298,
-0.011997010558843613,
0.05507921427488327,
-0.06166662275791168,
0.011602927930653095,
0.04573517292737961,
-0.00250246818177402,
-0.06295378506183624,
-0.03238888457417488,
0.0040907650254666805,
-0.056710440665483475,
-0.01504562422633171,
0.05407366901636124,
-0.03061254322528839,
0.05085344985127449,
-0.0417264886200428,
0.014067655429244041,
-0.0968286544084549,
0.04095933213829994,
-0.12316645681858063,
-0.02372012287378311,
-0.0030839992687106133,
0.03368508443236351,
0.10820356756448746,
0.03034779243171215,
-0.03274596855044365,
-0.060416098684072495,
-0.05807444453239441,
-0.061361901462078094,
-0.0552971176803112,
-0.001917488407343626,
0.037046924233436584,
-0.004465787205845118,
0.1287231743335724,
0.042912330478429794,
0.03782958909869194,
0.0041904631070792675,
0.04128390550613403,
-0.007221419829875231,
-0.028440099209547043,
-0.0037851757369935513,
0.01545552909374237,
-0.02389160916209221,
-0.048464562743902206,
0.0026568190660327673,
0.06582516431808472,
0.04928655922412872,
0.09660850465297699,
0.039489999413490295,
0.10392827540636063,
0.00729434285312891,
0.008853020146489143,
0.01575646549463272,
0.021374382078647614,
0.004585964605212212,
-0.00899609923362732,
-0.01812783256173134,
0.0019661514088511467,
0.027767501771450043,
-0.018540283665060997,
-4.4001456700470953e-8,
-0.024124661460518837,
0.03711742162704468,
0.04834114760160446,
0.07421108335256577,
-0.011546642519533634,
-0.057145148515701294,
0.07234200090169907,
0.04874207451939583,
-0.08982284367084503,
0.038700852543115616,
0.02815867029130459,
0.013138839974999428,
-0.007970284670591354,
-0.09325771033763885,
0.05313863605260849,
0.05941379442811012,
0.02629772573709488,
0.06773798167705536,
-0.028800485655665398,
-0.024678930640220642,
0.030675087124109268,
-0.02949841320514679,
-0.010167543776333332,
-0.02908220700919628,
0.09182502329349518,
-0.11813583970069885,
-0.09336718916893005,
0.01869538612663746,
0.05777589604258537,
-0.009644126519560814,
-0.004206379875540733,
-0.007681406568735838,
0.0052365027368068695,
-0.007489981595426798,
0.029092736542224884,
0.0666499063372612,
0.06415904313325882,
0.02059323899447918,
-0.048906274139881134,
-0.01047185342758894,
-0.05674709007143974,
0.03656844422221184,
0.06848794221878052,
-0.024785347282886505,
-0.0015935314586386085,
0.002047223737463355,
0.0033056128304451704,
-0.006189250387251377,
0.017981482669711113,
-0.08166960626840591,
0.08820337057113647,
-0.02803901396691799,
-0.0030479298438876867,
0.04901004210114479,
0.029056284576654434,
0.0007688520709052682,
0.0014018173096701503,
-0.04354925453662872,
-0.08295244723558426,
0.07218635827302933,
0.03153296187520027,
0.12032284587621689,
-0.0038554647471755743,
-0.09123031795024872
] | 0 | 0 | 0 | 0 |
|
WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild | https://openreview.net/forum?id=MKEHCx25xp | [
"Bill Yuchen Lin",
"Yuntian Deng",
"Khyathi Chandu",
"Abhilasha Ravichander",
"Valentina Pyatkin",
"Nouha Dziri",
"Ronan Le Bras",
"Yejin Choi"
] | Spotlight | We introduce WildBench, an automated evaluation framework designed to benchmark large language models (LLMs) using challenging, real-world user queries. WildBench consists of 1,024 tasks carefully selected from over one million human-chatbot conversation logs. For automated evaluation with WildBench, we have developed two metrics, WB-Reward and WB-Score, which are computable using advanced LLMs such as GPT-4-turbo. WildBench evaluation uses task-specific checklists to evaluate model outputs systematically and provides structured explanations that justify the scores and comparisons, resulting in more reliable and interpretable automatic judgments. WB-Reward employs fine-grained pairwise comparisons between model responses, generating five potential outcomes: much better, slightly better, slightly worse, much worse, or a tie. Unlike previous evaluations that employed a single baseline model, we selected three baseline models at varying performance levels to ensure a comprehensive pairwise evaluation. Additionally, we propose a simple method to mitigate length bias, by converting outcomes of “slightly better/worse” to “tie” if the winner response exceeds the loser one by more than K characters. WB-Score evaluates the quality of model outputs individually, making it a fast and cost-efficient evaluation metric. WildBench results demonstrate a strong correlation with the human-voted Elo ratings from Chatbot Arena on hard tasks. Specifically, WB-Reward achieves a Pearson correlation of 0.98 with top-ranking models. Additionally, WB-Score reaches 0.95, surpassing both ArenaHard’s 0.91 and AlpacaEval2.0’s 0.89 for length-controlled win rates, as well as the 0.87 for regular win rates. | LLM, Evaluation, Benchmarking | WildBench evaluates LLMs with hard and real tasks from users with metrics that are highly correlated with human-voted Elo. | 12,093 | 2406.04770 | [
-0.032383937388658524,
-0.04548662155866623,
-0.017453288659453392,
0.043991655111312866,
0.05826376751065254,
0.0017843835521489382,
-0.001009283703751862,
0.03438350558280945,
0.003950995858758688,
-0.03203883394598961,
-0.11870736628770828,
-0.09125509113073349,
0.03011307679116726,
0.0399276427924633,
-0.009542479179799557,
0.004538201726973057,
0.12060528248548508,
-0.020960871130228043,
-0.05830658599734306,
-0.012689366936683655,
-0.0012625694507732987,
0.06915324926376343,
0.11153513193130493,
-0.00598633149638772,
0.03760789707303047,
-0.07991597056388855,
-0.048376668244600296,
-0.02014349400997162,
-0.018825167790055275,
0.0023263823240995407,
-0.009390383958816528,
-0.017018329352140427,
0.02607988938689232,
0.097722128033638,
-0.06912187486886978,
-0.015571723692119122,
-0.08448424190282822,
-0.06848742067813873,
-0.037081241607666016,
-0.05669058486819267,
-0.060479361563920975,
-0.04327201843261719,
0.024877991527318954,
-0.000026144882212975062,
0.08768341690301895,
-0.029184184968471527,
-0.0527111254632473,
-0.040682755410671234,
0.02314106933772564,
0.03634054213762283,
-0.03865605965256691,
-0.018037648871541023,
0.05853660777211189,
0.07992613315582275,
-0.09630395472049713,
0.10821042209863663,
0.0063137090764939785,
0.032262858003377914,
0.04055337607860565,
-0.04201919957995415,
-0.03345340117812157,
-0.13955241441726685,
-0.07968583703041077,
-0.012226048856973648,
-0.02194424904882908,
-0.055855121463537216,
-0.018845263868570328,
0.039421964436769485,
-0.010838833637535572,
0.027460403740406036,
0.006671717390418053,
0.026490557938814163,
-0.01289243996143341,
0.07046818733215332,
-0.009986753575503826,
0.04932089522480965,
-0.021196847781538963,
-0.03636312857270241,
0.08449889719486237,
0.002973235910758376,
-0.06516721844673157,
-0.0671977624297142,
0.09326083213090897,
0.016916053369641304,
0.09224995970726013,
-0.017695782706141472,
0.07314170151948929,
0.04296750947833061,
0.0008444962440989912,
0.05157598853111267,
-0.024634482339024544,
-0.07114337384700775,
0.08582177013158798,
0.04402991756796837,
0.008471814915537834,
0.026202233508229256,
0.02760200947523117,
-0.010749599896371365,
-0.030315985903143883,
0.06367349624633789,
0.007344533689320087,
0.07218408584594727,
-0.03478318080306053,
-0.1296054869890213,
0.018424971029162407,
0.04019360989332199,
0.037859540432691574,
0.042243942618370056,
0.01023064460605383,
-0.10950381308794022,
-0.00475004268810153,
0.054430391639471054,
0.09653452783823013,
0.02832072041928768,
0.07359770685434341,
-0.008303768932819366,
0.045910488814115524,
-0.0374574214220047,
-0.010519523173570633,
0.051615819334983826,
0.04597073793411255,
0.05170024186372757,
0.053509827703237534,
-0.04684646427631378,
0.041839055716991425,
0.04436565935611725,
0.03162955120205879,
3.2244030448246126e-33,
0.11943469196557999,
0.04362111538648605,
0.02416081912815571,
0.03815092518925667,
0.05078989639878273,
0.039482466876506805,
-0.03272156044840813,
0.07671300321817398,
-0.12327399104833603,
0.008518628776073456,
-0.06481069326400757,
0.04611925408244133,
-0.019427083432674408,
0.03819192200899124,
0.04850277677178383,
0.0392492339015007,
-0.061271507292985916,
0.016859285533428192,
-0.002883264096453786,
0.05263126641511917,
0.020709959790110588,
-0.006999408360570669,
0.0324489064514637,
-0.012518815696239471,
0.03396815061569214,
0.05365283414721489,
0.04520860314369202,
-0.07284364104270935,
-0.04344313219189644,
0.019680336117744446,
-0.13605523109436035,
-0.06907529383897781,
0.005820718128234148,
0.017006412148475647,
0.04926977679133415,
-0.011153297498822212,
-0.06145217642188072,
0.003983835689723492,
-0.011489180848002434,
-0.005782940424978733,
-0.041560519486665726,
-0.033830784261226654,
-0.011651512235403061,
-0.03506724163889885,
-0.05687195062637329,
-0.04458262026309967,
-0.03705893084406853,
0.0073960269801318645,
-0.01697329618036747,
0.05076294764876366,
0.043686509132385254,
0.05068346485495567,
-0.005920378491282463,
0.06188201904296875,
-0.022748855873942375,
-0.015962710604071617,
0.0499575138092041,
0.030979132279753685,
-0.016795383766293526,
0.003940687049180269,
0.04620879143476486,
0.0346401073038578,
-0.02523116208612919,
0.03607802465558052,
0.046207770705223083,
-0.018322084099054337,
-0.04740606248378754,
0.08963645994663239,
-0.008516494184732437,
-0.0015570662217214704,
-0.011448826640844345,
-0.005278122145682573,
0.0597909539937973,
-0.014108462259173393,
0.021994391456246376,
0.01905042678117752,
0.0855800211429596,
-0.06381994485855103,
0.04829292371869087,
-0.02943071909248829,
0.028801564127206802,
-0.04931197687983513,
-0.01552544441074133,
-0.12048791348934174,
-0.061622485518455505,
-0.06808214634656906,
0.015203775838017464,
-0.09995070844888687,
-0.020073488354682922,
0.0035569521132856607,
-0.07304394990205765,
-0.02061128430068493,
-0.02722807601094246,
0.008466963656246662,
0.0058088526129722595,
-1.4875451219534029e-33,
-0.002113313414156437,
-0.010982275009155273,
-0.026634536683559418,
0.08055030554533005,
-0.017124298959970474,
-0.05652710795402527,
0.014815831556916237,
-0.018060851842164993,
-0.01723650097846985,
-0.0118741849437356,
-0.00948949996381998,
-0.03297881782054901,
0.0649942010641098,
-0.015342886559665203,
0.1027509942650795,
-0.011914723552763462,
-0.06894256919622421,
-0.04773768037557602,
-0.013370099477469921,
0.018725130707025528,
0.06132372096180916,
0.1354433000087738,
-0.12184813618659973,
0.033185411244630814,
-0.04754257574677467,
0.008248802274465561,
-0.01490382943302393,
-0.030603131279349327,
0.002904694527387619,
-0.009447659365832806,
0.03503460809588432,
0.012641342356801033,
-0.05523783713579178,
-0.013373215682804585,
0.005223321262747049,
0.03374452888965607,
0.09765054285526276,
-0.020685823634266853,
-0.022228052839636803,
0.08871998637914658,
0.05462484806776047,
-0.030145803466439247,
-0.09061972796916962,
0.009552177041769028,
0.0009448874043300748,
0.042336784303188324,
-0.14779828488826752,
-0.027977582067251205,
-0.06179239600896835,
0.016109073534607887,
-0.026007182896137238,
-0.00794905237853527,
-0.032269950956106186,
0.026310615241527557,
-0.05946822836995125,
-0.18698695302009583,
-0.07222382724285126,
-0.06673356890678406,
0.021076127886772156,
-0.0062482221983373165,
-0.08763442933559418,
-0.013343795202672482,
0.0727907195687294,
-0.017834460362792015,
0.03813368082046509,
-0.056493233889341354,
0.0048135812394320965,
0.02271888218820095,
-0.013226018287241459,
-0.018608123064041138,
0.03552970662713051,
-0.04233965650200844,
-0.008761722594499588,
0.03378041088581085,
0.06084033101797104,
0.05534476786851883,
-0.011464870534837246,
-0.037751730531454086,
-0.010212389752268791,
-0.031235819682478905,
0.015465550124645233,
0.013983598910272121,
0.1020088940858841,
0.013013266026973724,
0.04011145606637001,
0.08071806281805038,
0.029856782406568527,
0.10333192348480225,
-0.008880821987986565,
0.05534126237034798,
-0.041621945798397064,
0.021219687536358833,
0.01644851267337799,
0.027658291161060333,
-0.03354169428348541,
-4.068296277637273e-8,
-0.06454313546419144,
0.033722616732120514,
-0.06042705103754997,
0.08628866821527481,
-0.02377244643867016,
-0.02897663600742817,
-0.092411108314991,
0.11255823820829391,
-0.02121734619140625,
0.00946219451725483,
0.07127554714679718,
0.013492336496710777,
-0.0397685207426548,
-0.016834363341331482,
0.012540055438876152,
0.039054617285728455,
-0.04475058242678642,
0.05337104573845863,
-0.017950531095266342,
-0.0003219031495973468,
0.08406860381364822,
0.08958900719881058,
0.005433578509837389,
0.0029685311019420624,
-0.039824072271585464,
-0.031838372349739075,
-0.01473488099873066,
0.12301132827997208,
0.03409450873732567,
0.0018802274717018008,
0.028820548206567764,
0.027192292734980583,
-0.040701575577259064,
-0.02091227099299431,
0.052352990955114365,
0.01338750310242176,
-0.04743773862719536,
-0.03081286884844303,
0.024773044511675835,
0.06429518759250641,
0.05049077421426773,
0.043234240263700485,
-0.03557579964399338,
0.009375447407364845,
-0.009679815731942654,
-0.03266020491719246,
-0.07235852628946304,
-0.10223232209682465,
0.0616007000207901,
-0.075158029794693,
0.01856980100274086,
-0.08412761986255646,
-0.033330321311950684,
0.06257859617471695,
0.05444948375225067,
0.026178626343607903,
-0.015069002285599709,
-0.029821079224348068,
0.04906517267227173,
0.021740658208727837,
0.0019085335079580545,
0.021557902917265892,
-0.04861791059374809,
0.006298101507127285
] | https://github.com/allenai/wildbench | 221 | 0 | 2 | 6 |
Revisiting Zeroth-Order Optimization: Minimum-Variance Two-Point Estimators and Directionally Aligned Perturbations | https://openreview.net/forum?id=ywFOSIT9ik | [
"Shaocong Ma",
"Heng Huang"
] | Spotlight | In this paper, we explore the two-point zeroth-order gradient estimator and identify the distribution of random perturbations that minimizes the estimator's asymptotic variance as the perturbation stepsize tends to zero. We formulate it as a constrained functional optimization problem over the space of perturbation distributions. Our findings reveal that such desired perturbations can align directionally with the true gradient, instead of maintaining a fixed length. While existing research has largely focused on fixed-length perturbations, the potential advantages of directional alignment have been overlooked. To address this gap, we delve into the theoretical and empirical properties of the directionally aligned perturbation (DAP) scheme, which adaptively offers higher accuracy along critical directions. Additionally, we provide a convergence analysis for stochastic gradient descent using $\delta$-unbiased random perturbations, extending existing complexity bounds to a wider range of perturbations. Through empirical evaluations on both synthetic problems and practical tasks, we demonstrate that DAPs outperform traditional methods under specific conditions. | zeroth-order optimization, SGD, convergence analysis | This paper discusses the minimum-variance condition for two-point zeroth-order gradient estimators and proposes a new random perturbation. | 12,003 | null | [
-0.05819118395447731,
-0.0764893889427185,
0.004038337618112564,
0.015238383784890175,
-0.01291574165225029,
0.014276528730988503,
-0.05183333903551102,
-0.03462415933609009,
0.0820067822933197,
-0.05657569319009781,
-0.005311399232596159,
0.042538534849882126,
0.016338594257831573,
0.0025973685551434755,
-0.012922776862978935,
0.037042438983917236,
0.03116706572473049,
0.09500116109848022,
0.009012619033455849,
-0.003903444856405258,
0.0037267219740897417,
0.08162970095872879,
-0.01523457933217287,
-0.0038764402270317078,
0.07648896425962448,
-0.043222445994615555,
-0.033381517976522446,
0.0029666218906641006,
-0.04874342307448387,
-0.05687038227915764,
0.059730060398578644,
-0.03313430771231651,
-0.09693815559148788,
-0.03112705610692501,
-0.0017755628796294332,
0.024104498326778412,
-0.014209063723683357,
0.038132041692733765,
0.018675394356250763,
0.014227064326405525,
0.00969421211630106,
0.0387752465903759,
0.019805697724223137,
-0.016661357134580612,
0.07377933710813522,
-0.004842026624828577,
0.06020541861653328,
-0.08887182921171188,
0.034988537430763245,
-0.013304460793733597,
-0.06990519165992737,
-0.004407297819852829,
-0.0627489909529686,
0.000011083840945502743,
-0.013491710647940636,
-0.0007562179234810174,
0.02833062782883644,
0.026670588180422783,
0.018656020984053612,
0.07264076918363571,
0.08445193618535995,
-0.11625736951828003,
0.03528992459177971,
-0.006969298701733351,
-0.0038248838391155005,
-0.08441254496574402,
0.10594700276851654,
-0.030369356274604797,
-0.010626398958265781,
0.059261977672576904,
-0.0663767084479332,
0.08799909800291061,
-0.042270347476005554,
0.06986520439386368,
0.024222444742918015,
0.04834984615445137,
0.06365465372800827,
0.017375193536281586,
0.06129048764705658,
-0.10037738084793091,
0.058768317103385925,
0.09114174544811249,
0.06768396496772766,
-0.003738455008715391,
0.02272278442978859,
-0.004283060319721699,
-0.03353250026702881,
-0.01892082951962948,
0.10504844039678574,
-0.015724314376711845,
-0.06643523275852203,
-0.014981894753873348,
-0.15503741800785065,
0.009563425555825233,
0.062804214656353,
-0.002322367625311017,
0.0230596661567688,
-0.04500719904899597,
-0.0264957956969738,
0.08340959250926971,
0.02675996907055378,
0.05108741298317909,
-0.04644493758678436,
-0.03948039934039116,
0.039996784180402756,
-0.07067840546369553,
0.002787864301353693,
0.02054859884083271,
-0.004970773123204708,
-0.05595375597476959,
0.06164255738258362,
-0.005475576967000961,
0.00819204468280077,
0.025407439097762108,
-0.06677444279193878,
-0.0655021071434021,
0.009064769372344017,
-0.03698182851076126,
0.03395475447177887,
0.027545791119337082,
-0.10225678235292435,
-0.055628784000873566,
-0.029071977362036705,
0.0994931235909462,
-0.015920177102088928,
0.06957677006721497,
-0.023248912766575813,
4.167816607228179e-33,
-0.0031800761353224516,
0.09370507299900055,
0.13303343951702118,
-0.03267199546098709,
0.03190353140234947,
-0.032724201679229736,
-0.011098839342594147,
0.004862256348133087,
-0.05372146517038345,
0.06880185753107071,
0.01644531451165676,
-0.06794572621583939,
-0.07394344359636307,
-0.01825130730867386,
0.042462997138500214,
-0.0911165252327919,
-0.005327711813151836,
0.06961002945899963,
-0.021035974845290184,
-0.03984520584344864,
0.06935940682888031,
-0.018970051780343056,
-0.06831423938274384,
-0.015676992014050484,
0.01858855038881302,
-0.040699124336242676,
0.021330559626221657,
-0.011897561140358448,
-0.10654742270708084,
0.006184838246554136,
-0.10681914538145065,
-0.028886504471302032,
0.015150565654039383,
-0.04466121271252632,
0.045643243938684464,
-0.07861991971731186,
-0.051391128450632095,
-0.05030704662203789,
-0.024832848459482193,
0.018184185028076172,
-0.03246819227933884,
0.11021773517131805,
-0.017081642523407936,
-0.028534337878227234,
-0.05571291223168373,
-0.054664891213178635,
-0.0266328863799572,
0.07297360897064209,
0.05982895940542221,
-0.08404570817947388,
-0.0943283811211586,
0.07895006984472275,
-0.03653860464692116,
0.015644658356904984,
-0.016964251175522804,
-0.02206387184560299,
0.021451354026794434,
-0.03314599394798279,
-0.0318976528942585,
0.031157920137047768,
-0.07660216093063354,
-0.009657849557697773,
-0.033457279205322266,
-0.020569918677210808,
0.03164241835474968,
0.007198249455541372,
0.02532579004764557,
0.018577396869659424,
0.06610541045665741,
-0.010608842596411705,
0.04318644106388092,
0.013928571715950966,
0.012644733302295208,
-0.06922654807567596,
0.038233280181884766,
-0.012686643749475479,
0.024731095880270004,
-0.06604685634374619,
0.032612260431051254,
-0.009720000438392162,
-0.011462746188044548,
0.08875259757041931,
-0.0022951033897697926,
-0.008457941003143787,
-0.05483393743634224,
-0.05444619432091713,
0.02937646582722664,
-0.02687210589647293,
-0.0305746179074049,
-0.014665119349956512,
-0.09553219377994537,
0.0220554880797863,
-0.02794499695301056,
0.024238411337137222,
0.016394570469856262,
-4.5302479642095835e-33,
-0.10405479371547699,
-0.012905065901577473,
0.04991719499230385,
0.028590548783540726,
0.060062311589717865,
-0.017416460439562798,
-0.02121804840862751,
0.016193440183997154,
0.001506536966189742,
-0.018503330647945404,
-0.015126209706068039,
-0.02914372831583023,
-0.038946427404880524,
0.03879094123840332,
0.028390096500515938,
0.06502073258161545,
-0.012631947174668312,
0.08484816551208496,
0.038921792060136795,
0.04721841961145401,
0.016202788800001144,
0.06058398634195328,
-0.018420463427901268,
0.07036414742469788,
0.060775432735681534,
0.00938490591943264,
0.03074650838971138,
0.09176289290189743,
-0.06243313103914261,
-0.02617832086980343,
-0.02290324680507183,
0.097641721367836,
-0.011009404435753822,
0.008945389650762081,
-0.05566922575235367,
0.05087917670607567,
-0.031200824305415154,
0.03488893434405327,
0.07325522601604462,
0.10487045347690582,
0.10805032402276993,
0.016123060137033463,
-0.02883121185004711,
-0.044330745935440063,
0.07412045449018478,
-0.053286947309970856,
-0.014243465848267078,
0.006975216791033745,
0.006166826467961073,
-0.014093996025621891,
-0.031120169907808304,
0.04668770357966423,
-0.10964539647102356,
0.06022948399186134,
-0.04294075071811676,
-0.03552571311593056,
-0.08363648504018784,
0.002896167803555727,
0.09573468565940857,
-0.018010534346103668,
-0.04427550733089447,
0.025680620223283768,
-0.02024238184094429,
0.030263781547546387,
-0.021876292303204536,
0.043171778321266174,
0.03941410407423973,
-0.036623019725084305,
0.02387169562280178,
0.03524927422404289,
-0.014623919501900673,
-0.0026310018729418516,
-0.006471558474004269,
0.055507320910692215,
-0.07054464519023895,
-0.03663313388824463,
0.07574447244405746,
-0.05979828163981438,
-0.028502896428108215,
-0.03432152047753334,
0.10607266426086426,
0.06119757890701294,
0.03834644332528114,
-0.004251193720847368,
0.032111551612615585,
0.08303354680538177,
0.0775199681520462,
0.09313064068555832,
0.01298118568956852,
-0.019610870629549026,
-0.021846801042556763,
-0.002847997471690178,
0.011269998736679554,
-0.07197024673223495,
-0.047397613525390625,
-4.9553705139260273e-8,
-0.07576185464859009,
-0.002108038403093815,
0.07406803220510483,
0.039994847029447556,
0.01673048734664917,
0.022727463394403458,
0.02471991255879402,
-0.008981537073850632,
0.03431609645485878,
-0.0965089350938797,
0.024090072140097618,
-0.0070944251492619514,
-0.032961174845695496,
-0.06996084749698639,
-0.04588685929775238,
0.0962958112359047,
-0.01751146838068962,
-0.05182605981826782,
-0.024813059717416763,
0.030819637700915337,
-0.1109093427658081,
0.12339267134666443,
0.014573088847100735,
-0.05758409947156906,
0.01734725944697857,
-0.10078438371419907,
0.003804099513217807,
0.0826396644115448,
0.16520202159881592,
-0.055828630924224854,
-0.06102711707353592,
0.003415395738556981,
0.011394063010811806,
0.004402666352689266,
-0.021283764392137527,
0.03917916119098663,
0.05065825954079628,
0.024525372311472893,
-0.055911120027303696,
-0.0741383284330368,
-0.04497798532247543,
0.05897056311368942,
0.004880120977759361,
0.005964819807559252,
0.0033594234846532345,
0.022277185693383217,
-0.03204796835780144,
-0.06124364957213402,
-0.015689345076680183,
-0.0827096477150917,
0.04129533842206001,
0.01836663857102394,
-0.003013595473021269,
0.057675864547491074,
0.01740824431180954,
0.02369890734553337,
-0.08077114075422287,
-0.06353309750556946,
-0.027835873886942863,
0.0708698108792305,
0.0017195886466652155,
-0.03188224509358406,
-0.05470258370041847,
-0.03902171179652214
] | 0 | 0 | 0 | 0 |
|
Adaptive Batch Size for Privately Finding Second-Order Stationary Points | https://openreview.net/forum?id=ikkvC1UnnE | [
"Daogao Liu",
"Kunal Talwar"
] | Spotlight | There is a gap between finding a first-order stationary point (FOSP) and a second-order stationary point (SOSP) under differential privacy constraints, and it remains unclear whether privately finding an SOSP is more challenging than finding an FOSP. Specifically, Ganesh et al. (2023) claimed that an $\alpha$-SOSP can be found with $\alpha=\Tilde{O}(\frac{1}{n^{1/3}}+(\frac{\sqrt{d}}{n\epsilon})^{3/7})$, where $n$ is the dataset size, $d$ is the dimension, and $\epsilon$ is the differential privacy parameter.
However, a recent analysis revealed an issue in their saddle point escape procedure, leading to weaker guarantees.
Building on the SpiderBoost algorithm framework, we propose a new approach that uses adaptive batch sizes and incorporates the binary tree mechanism.
Our method not only corrects this issue but also improves the results for privately finding an SOSP, achieving $\alpha=\Tilde{O}(\frac{1}{n^{1/3}}+(\frac{\sqrt{d}}{n\epsilon})^{1/2})$.
This improved bound matches the state-of-the-art for finding a FOSP, suggesting that privately finding an SOSP may be achievable at no additional cost. | Differential privacy, non-convex optimization, adaptive batch size | We improved the previous results in finding second-order stationary points privately by utlizing adaptive batch size and the tree mechanism. | 11,995 | 2410.07502 | [
-0.08794960379600525,
-0.04159093275666237,
-0.01650426909327507,
0.0206191036850214,
0.05214986950159073,
-0.050522565841674805,
-0.01626136526465416,
-0.015861675143241882,
0.03877926990389824,
-0.019439345225691795,
0.02714599296450615,
0.06429429352283478,
-0.030325470492243767,
-0.07575216144323349,
0.022852404043078423,
-0.010971760377287865,
0.0429709255695343,
0.055908821523189545,
-0.003013594076037407,
0.040721528232097626,
-0.022565169259905815,
-0.029120037332177162,
-0.021974511444568634,
0.04202038049697876,
0.032843004912137985,
-0.005445356946438551,
0.04872481897473335,
-0.0016335907857865095,
0.038341354578733444,
-0.08555560559034348,
0.06165340915322304,
0.017653336748480797,
0.04870857298374176,
-0.017247088253498077,
0.03428169712424278,
-0.014034388586878777,
-0.010828021913766861,
0.02477097511291504,
-0.005506480578333139,
0.06322037428617477,
0.028670674189925194,
0.06978955864906311,
0.024203486740589142,
0.007454123347997665,
0.0038785748183727264,
0.01596558466553688,
0.0019207327859476209,
0.020181555300951004,
-0.013174056075513363,
-0.10991138219833374,
0.005034805741161108,
0.02272222749888897,
-0.06288151443004608,
-0.0065714712254703045,
-0.07752469182014465,
-0.05419033393263817,
-0.00813019648194313,
-0.0670572817325592,
-0.04128429666161537,
0.09535102546215057,
0.01969018392264843,
-0.06909259408712387,
0.04129916429519653,
-0.01793564483523369,
0.01655321754515171,
0.025293836370110512,
0.07418741285800934,
-0.07226171344518661,
0.010709608905017376,
0.05236918106675148,
-0.09283202886581421,
0.08704747259616852,
-0.065526083111763,
0.04938855767250061,
0.027235381305217743,
-0.0055399248376488686,
0.059268850833177567,
0.048913970589637756,
-0.006971118506044149,
-0.037140391767024994,
0.0068200742825865746,
0.015975037589669228,
0.015338639728724957,
-0.03521919623017311,
-0.003000409109517932,
-0.0510638989508152,
-0.04131712764501572,
0.02576712891459465,
0.11844570934772491,
0.02795092575252056,
-0.006747231353074312,
0.01618197001516819,
-0.12168125808238983,
-0.02042490430176258,
0.03158065304160118,
0.021169213578104973,
-0.003712646896019578,
0.08013609796762466,
0.019384846091270447,
0.049329619854688644,
0.021425537765026093,
0.05753680691123009,
-0.02167019248008728,
0.024223674088716507,
0.035635389387607574,
0.026819387450814247,
0.019093632698059082,
0.03833945095539093,
0.05527424439787865,
0.001703912392258644,
0.04774441570043564,
0.011097813956439495,
-0.0050608622841537,
0.018865181133151054,
-0.010212860070168972,
-0.10069325566291809,
-0.025285780429840088,
-0.03737272322177887,
0.02975555695593357,
0.005744826514273882,
0.01460858155041933,
-0.02882317267358303,
0.029850218445062637,
0.03645425662398338,
-0.010511647909879684,
0.040339160710573196,
-0.05063266307115555,
1.4492673524707829e-33,
0.046237844973802567,
0.09268903732299805,
0.06557168811559677,
-0.1505114734172821,
-0.03105657733976841,
0.08850818127393723,
0.01807759329676628,
-0.015420440584421158,
-0.035949986428022385,
0.07724352926015854,
-0.03786914423108101,
-0.0806017816066742,
0.020590290427207947,
-0.07336772978305817,
0.09533768147230148,
-0.015168353915214539,
0.03718050941824913,
0.08934742957353592,
0.013939036056399345,
-0.0937100276350975,
0.19514116644859314,
-0.03393958508968353,
-0.033524010330438614,
-0.06145312637090683,
0.017941081896424294,
-0.025980539619922638,
-0.010548567399382591,
-0.07363230735063553,
0.01502805482596159,
0.03431988134980202,
-0.0390426404774189,
0.012309681624174118,
-0.027455246075987816,
0.06616732478141785,
0.09465517848730087,
-0.06099998578429222,
-0.03825598955154419,
-0.08003349602222443,
0.04531893506646156,
-0.0045817396603524685,
-0.008469986729323864,
0.028013162314891815,
0.015365404076874256,
-0.030769314616918564,
-0.12169548869132996,
-0.046957239508628845,
0.0005927444435656071,
0.10504899173974991,
-0.005292727146297693,
-0.029336273670196533,
-0.03672933578491211,
0.04398087412118912,
-0.09832913428544998,
0.014491042122244835,
0.008477838709950447,
-0.0004680272832047194,
0.009372470900416374,
-0.0021787125151604414,
0.019433708861470222,
0.14815206825733185,
0.004218278452754021,
0.03208967670798302,
0.024215172976255417,
-0.038428690284490585,
-0.0006313286139629781,
0.002843414666131139,
0.00387434265576303,
-0.05018358305096626,
0.07498505711555481,
-0.045257534831762314,
0.046419620513916016,
0.028665531426668167,
0.0029703981708735228,
-0.0505581758916378,
0.005790199153125286,
-0.08879179507493973,
0.12199261784553528,
-0.023391425609588623,
-0.0021703168749809265,
-0.06594239175319672,
-0.030967814847826958,
-0.024219611659646034,
0.000972158566582948,
0.005043401848524809,
-0.03795918822288513,
-0.08366798609495163,
0.02610623836517334,
0.027999475598335266,
-0.09353481233119965,
-0.07970453053712845,
-0.05139369145035744,
-0.08403515815734863,
-0.018355127424001694,
-0.0011040031677111983,
-0.05446478724479675,
-6.5650358486516635e-34,
-0.036197904497385025,
-0.025278359651565552,
0.015082696452736855,
0.006016988772898912,
0.027837902307510376,
-0.05854564905166626,
-0.08424790948629379,
0.01871422491967678,
0.008777867071330547,
-0.012201585806906223,
-0.12956614792346954,
0.061645474284887314,
0.04653236269950867,
-0.05508435145020485,
0.06057549640536308,
0.05836037918925285,
-0.014256403781473637,
0.023837804794311523,
0.006945847999304533,
0.05858580768108368,
0.0045280964113771915,
0.08108661323785782,
-0.017654957249760628,
0.11308819055557251,
0.07424791157245636,
0.004436693154275417,
-0.0696292519569397,
0.11099271476268768,
-0.010779641568660736,
0.023429427295923233,
-0.06274528056383133,
0.030710464343428612,
-0.05552966147661209,
-0.03680245950818062,
-0.043152812868356705,
0.010272407904267311,
-0.051386866718530655,
0.040589798241853714,
0.01919647492468357,
0.1374744027853012,
0.00070792599581182,
0.011379094794392586,
-0.05267266556620598,
-0.0009838254190981388,
0.06330222636461258,
0.06742426753044128,
-0.07480819523334503,
0.06378212571144104,
0.03621978312730789,
0.07526616007089615,
-0.038623061031103134,
0.037536244839429855,
-0.004101456142961979,
0.10729499906301498,
-0.020178338512778282,
-0.02506420947611332,
-0.06760100275278091,
0.05281122028827667,
-0.018644804134964943,
-0.052880678325891495,
0.029425209388136864,
0.030719410628080368,
-0.04970722645521164,
0.03281998634338379,
0.002336671808734536,
-0.033044639974832535,
-0.04776670038700104,
-0.07672354578971863,
-0.026237236335873604,
0.010364490561187267,
-0.03363368287682533,
-0.06580527126789093,
-0.020359158515930176,
0.04673929512500763,
-0.05270755663514137,
0.002684421371668577,
0.07694243639707565,
-0.03130622208118439,
0.009665981866419315,
0.09463685750961304,
-0.04980434477329254,
0.13625766336917877,
0.08822609484195709,
0.0015671384753659368,
0.04863160476088524,
0.03225715085864067,
0.10070547461509705,
0.023380734026432037,
-0.010606448166072369,
-0.03241884335875511,
-0.029215948656201363,
-0.006448054686188698,
-0.019203511998057365,
0.018135536462068558,
-0.012868411839008331,
-4.857029267668622e-8,
-0.047850657254457474,
-0.04410845413804054,
0.00957921426743269,
0.06002571061253548,
0.13129326701164246,
-0.03805847838521004,
-0.008123764768242836,
-0.026592053472995758,
-0.04692843183875084,
-0.036614689975976944,
0.08813588321208954,
-0.03543449938297272,
-0.019022420048713684,
0.003935184795409441,
0.018521877005696297,
0.03146461024880409,
0.008644462563097477,
-0.116030752658844,
-0.05066147819161415,
-0.003166338661685586,
-0.09351528435945511,
0.03330620005726814,
-0.02192593552172184,
-0.038345493376255035,
0.05347110703587532,
-0.04122588038444519,
0.022211993113160133,
-0.025976968929171562,
0.062235262244939804,
0.06160115450620651,
-0.009203270077705383,
-0.023515023291110992,
0.06308037787675858,
0.0535779744386673,
-0.025058040395379066,
0.07256986200809479,
-0.027394678443670273,
0.05207544192671776,
-0.04739657789468765,
0.023885603994131088,
-0.020313220098614693,
0.04118780419230461,
0.026201477274298668,
-0.00948280468583107,
0.02614910528063774,
0.01755405031144619,
-0.00908999890089035,
-0.0751161277294159,
-0.0016145356930792332,
-0.04765375331044197,
0.031560018658638,
0.01363224908709526,
-0.0607902929186821,
0.05958237126469612,
0.06293149292469025,
0.005326541606336832,
-0.09603837132453918,
-0.04991277679800987,
0.035624463111162186,
0.05064307898283005,
-0.033037398010492325,
-0.049062978476285934,
-0.054013725370168686,
-0.023587798699736595
] | 0 | 0 | 0 | 0 |
|
Adversarial Perturbations Cannot Reliably Protect Artists From Generative AI | https://openreview.net/forum?id=yfW1x7uBS5 | [
"Robert Hönig",
"Javier Rando",
"Nicholas Carlini",
"Florian Tramèr"
] | Spotlight | Artists are increasingly concerned about advancements in image generation models that can closely replicate their unique artistic styles.
In response, several protection tools against style mimicry have been developed that incorporate small adversarial perturbations into artworks published online. In this work, we evaluate the effectiveness of popular protections---with millions of downloads---and show they only provide a false sense of security. We find that low-effort and "off-the-shelf" techniques, such as image upscaling, are sufficient to create robust mimicry methods that significantly degrade existing protections. Through a user study, we demonstrate that **all existing protections can be easily bypassed**, leaving artists vulnerable to style mimicry. We caution that tools based on adversarial perturbations cannot reliably protect artists from the misuse of generative AI, and urge the development of alternative protective solutions. | security, adversarial, style mimicry, generative ai | We show that adversarial perturbations are not a reliable strategy to protect artists' images from being used to train generative models. | 11,941 | null | [
-0.14693306386470795,
0.004119046498090029,
0.06024959683418274,
0.04654814675450325,
0.0437418632209301,
0.02233046293258667,
-0.052657198160886765,
-0.11163860559463501,
-0.013663461431860924,
-0.062367524951696396,
-0.06961504369974136,
-0.03159284219145775,
0.09694495797157288,
-0.07872969657182693,
-0.02872871235013008,
0.041263699531555176,
0.07185974717140198,
0.13315819203853607,
-0.034327466040849686,
0.025014832615852356,
-0.029850229620933533,
-0.03391548991203308,
-0.016507407650351524,
-0.03857272118330002,
0.03980134427547455,
-0.024014241993427277,
-0.02798219956457615,
-0.023435480892658234,
0.05174810066819191,
-0.07492848485708237,
0.00538759957998991,
-0.03615419566631317,
-0.04442986845970154,
0.01303890161216259,
-0.08658912777900696,
0.07644394785165787,
-0.050415847450494766,
0.07844188064336777,
0.030250443145632744,
-0.07364398241043091,
-0.04731244221329689,
0.017633317038416862,
-0.07270461320877075,
0.00019475119188427925,
0.00014865932462271303,
-0.07086515426635742,
0.01674656942486763,
-0.018136706203222275,
-0.00911178719252348,
-0.01894402876496315,
-0.04642114415764809,
-0.06339891999959946,
-0.0136953666806221,
-0.030285291373729706,
0.010189667344093323,
-0.05333087965846062,
0.06780113279819489,
0.03184729069471359,
0.007890868000686169,
0.07396432012319565,
-0.008874130435287952,
-0.013765132986009121,
-0.01170234102755785,
-0.012742326594889164,
0.02438853308558464,
0.011210838332772255,
0.002673159120604396,
0.0757594183087349,
0.0016791396774351597,
0.04729744419455528,
0.07786720991134644,
-0.006739851087331772,
-0.05832381173968315,
0.04515509307384491,
0.07250077277421951,
0.10896617919206619,
-0.07970201969146729,
-0.08056797087192535,
-0.01549459621310234,
-0.0989057719707489,
0.0680733174085617,
-0.06085791438817978,
0.0412934236228466,
0.04779481887817383,
0.051807865500450134,
0.04878192022442818,
0.00047983694821596146,
-0.020786309614777565,
0.08921674638986588,
0.008736134506762028,
-0.03799380734562874,
0.0025861207395792007,
0.0501207634806633,
0.022747652605175972,
0.0861634910106659,
-0.0318659283220768,
-0.005673109088093042,
-0.09124945849180222,
-0.03727070987224579,
0.0734887346625328,
-0.019935455173254013,
0.00740802800282836,
0.05664364993572235,
-0.0011331901187077165,
0.09851585328578949,
-0.05258592590689659,
0.11931080371141434,
-0.008832346647977829,
-0.004007717594504356,
0.01816205307841301,
0.08516936749219894,
0.06995003670454025,
-0.018076376989483833,
-0.029369760304689407,
0.047162655740976334,
-0.02252895012497902,
-0.05035281181335449,
0.025432683527469635,
0.007645073812454939,
0.03034345805644989,
0.03576904162764549,
0.09079279750585556,
0.0200587697327137,
0.01641719974577427,
0.007019467186182737,
-0.06170262396335602,
-0.057962432503700256,
9.30182546965596e-34,
0.053867243230342865,
0.031683262437582016,
0.04063374921679497,
0.013663571327924728,
0.07407378405332565,
-0.04481544345617294,
-0.04432985931634903,
0.00809398852288723,
-0.027921391651034355,
-0.03783801943063736,
0.024862848222255707,
-0.04529256001114845,
-0.06253331899642944,
0.14279602468013763,
0.02763027511537075,
0.058226328343153,
0.023098161444067955,
-0.004420040640980005,
0.017298297956585884,
0.00006825332093285397,
0.015934711322188377,
-0.08362958580255508,
-0.00956558808684349,
0.02442031167447567,
-0.07559856027364731,
0.00930267944931984,
0.010406799614429474,
-0.03822522237896919,
0.06850191950798035,
0.03130626678466797,
-0.14783257246017456,
-0.014534004032611847,
0.09040927141904831,
0.03552931174635887,
-0.03301380202174187,
0.03499279543757439,
-0.008338543586432934,
-0.03897763788700104,
0.0193894375115633,
0.06125393137335777,
-0.04021010175347328,
0.028619129210710526,
-0.010981477797031403,
-0.008925956673920155,
-0.03462151437997818,
0.022897370159626007,
-0.012269506230950356,
0.04075220972299576,
-0.06292645633220673,
0.008854859508574009,
0.0043482002802193165,
0.09409113973379135,
0.014861539006233215,
-0.006958727724850178,
-0.08412992209196091,
-0.08924947679042816,
-0.01568034663796425,
-0.037586186081171036,
0.039132259786129,
-0.004614923615008593,
0.035590652376413345,
0.048350509256124496,
0.011084482073783875,
0.04160375893115997,
-0.05058718100190163,
-0.012978428043425083,
0.03017113357782364,
-0.06862187385559082,
0.004945593420416117,
0.049547892063856125,
-0.09454826265573502,
-0.0025209321174770594,
-0.08697289228439331,
-0.057884231209754944,
-0.05249569192528725,
-0.10113056749105453,
0.05075930431485176,
-0.0028974006418138742,
0.06120502948760986,
-0.008621645160019398,
-0.08825476467609406,
0.012117304839193821,
-0.007470286451280117,
-0.011161348782479763,
-0.06324304640293121,
-0.05107360705733299,
0.020915163680911064,
0.00792111735790968,
-0.01803024671971798,
0.007342766970396042,
-0.0023343826178461313,
-0.05118362233042717,
-0.01488830428570509,
0.006268143653869629,
-0.04336579889059067,
-1.6768329568854004e-33,
-0.06623225659132004,
0.0190484169870615,
-0.005742669105529785,
0.052469402551651,
-0.07219281792640686,
-0.03909587860107422,
-0.05137665569782257,
0.07712078839540482,
-0.015430294908583164,
-0.05183933302760124,
-0.029908835887908936,
0.04212850332260132,
-0.00004560537126963027,
-0.09655993431806564,
-0.011743325740098953,
-0.12256333231925964,
-0.0001266064791707322,
0.04494258016347885,
-0.08541826158761978,
-0.02754206582903862,
0.07226957380771637,
0.05949285626411438,
-0.0021163327619433403,
0.047127965837717056,
-0.048877451568841934,
0.0415385439991951,
-0.07752750813961029,
0.006000255234539509,
0.05558884143829346,
0.021248571574687958,
0.03178078308701515,
-0.01535804569721222,
0.013756061904132366,
-0.0036683459766209126,
-0.046616263687610626,
0.029114218428730965,
0.04084327071905136,
0.11021170765161514,
-0.02522851526737213,
0.08413524180650711,
0.025367071852087975,
-0.00526596512645483,
-0.05325356498360634,
0.027179617434740067,
-0.029268810525536537,
-0.06593754142522812,
-0.023605896160006523,
-0.003411021316424012,
0.04306284710764885,
-0.01661357842385769,
0.05377435311675072,
-0.02724449522793293,
-0.00014287562225945294,
-0.028243450447916985,
-0.041405368596315384,
-0.09324052929878235,
-0.000959983270149678,
0.0222671777009964,
0.02934681810438633,
0.10197591036558151,
-0.05261605232954025,
-0.07437125593423843,
-0.07155117392539978,
-0.04098542034626007,
-0.041638754308223724,
0.05491640791296959,
-0.04145766794681549,
0.07456369698047638,
0.007107212208211422,
-0.00017682058387435973,
0.0602455697953701,
0.038670096546411514,
-0.019583724439144135,
0.07860711216926575,
-0.01979733072221279,
-0.04255180433392525,
0.04640811309218407,
0.01625922881066799,
0.03208959847688675,
-0.023712603375315666,
0.03137156739830971,
-0.048395294696092606,
0.041765280067920685,
0.00043467938667163253,
0.06270696222782135,
0.057919953018426895,
-0.02367520146071911,
-0.02004474401473999,
-0.013105448335409164,
0.06545956432819366,
0.01744408905506134,
0.06693832576274872,
-0.06703273952007294,
0.03434650972485542,
-0.004817149136215448,
-3.322183061982287e-8,
-0.033137913793325424,
0.02852199599146843,
0.06576001644134521,
0.0537874661386013,
0.02477903664112091,
0.028732914477586746,
-0.0371248759329319,
-0.05129978060722351,
-0.002633035648614168,
-0.03405012935400009,
0.05497721955180168,
-0.04669859632849693,
0.04482171684503555,
-0.05375990644097328,
-0.02848002500832081,
-0.0020596308168023825,
-0.01910097524523735,
0.09224525094032288,
-0.06800447404384613,
-0.08301867544651031,
-0.0660526305437088,
0.04139348119497299,
-0.05238818749785423,
-0.11202376335859299,
0.04245263710618019,
-0.04574957117438316,
-0.05107304826378822,
-0.03399709612131119,
-0.022734181955456734,
0.120635025203228,
-0.034596994519233704,
0.00413620937615633,
0.031792230904102325,
0.027544183656573296,
-0.0016265264712274075,
0.08870188891887665,
-0.03370092436671257,
-0.03065497800707817,
-0.0668560117483139,
-0.02951824851334095,
0.004307892639189959,
0.003780838567763567,
0.07688520848751068,
-0.017071649432182312,
0.013411757536232471,
-0.03910517692565918,
0.1025209128856659,
-0.07642166316509247,
0.05244182050228119,
0.046476855874061584,
0.00015696106129325926,
-0.041998181492090225,
0.10347814857959747,
0.09785988181829453,
0.0009434907115064561,
-0.008967595174908638,
-0.018355274572968483,
0.07803986966609955,
0.04841378703713417,
0.1035507395863533,
0.0567462258040905,
-0.03036533109843731,
0.050991591066122055,
0.006572321057319641
] | 0 | 0 | 0 | 0 |
|
Better autoregressive regression with LLMs via regression-aware fine-tuning | https://openreview.net/forum?id=xGs7Ch3Vyo | [
"Michal Lukasik",
"Zhao Meng",
"Harikrishna Narasimhan",
"Yin-Wen Chang",
"Aditya Krishna Menon",
"Felix Yu",
"Sanjiv Kumar"
] | Spotlight | Decoder-based large language models (LLMs) have proven highly versatile, with remarkable successes even on problems ostensibly removed from traditional language generation. One such example is solving regression problems, where the targets are real numbers rather than textual tokens. A common approach to use LLMs on such problems is to perform fine-tuning based on the cross-entropy loss, and use autoregressive sampling at inference time. Another approach relies on fine-tuning a separate predictive head with a suitable loss such as squared error. While each approach has had success, there has been limited study on principled ways of using decoder LLMs for regression. In this work, we compare different prior works under a unified view, and introduce regression-aware fine-tuning(RAFT), a novel approach based on the Bayes-optimal decision rule. We demonstrate how RAFT improves over established baselines on several benchmarks and model families. | regression, LLMs | null | 11,905 | null | [
-0.06585971266031265,
-0.10763540863990784,
0.05405563488602638,
0.048228174448013306,
0.07661376148462296,
0.06765922904014587,
0.019240934401750565,
-0.0026255392003804445,
0.036253053694963455,
-0.021252507343888283,
-0.07665295898914337,
-0.03364315629005432,
0.010646780952811241,
0.01278680469840765,
-0.021819578483700752,
0.013131781481206417,
0.06966373324394226,
0.09085974097251892,
-0.036748290061950684,
-0.08046805113554001,
0.004506788682192564,
0.05826286971569061,
0.0037234441842883825,
-0.009687033481895924,
0.11421534419059753,
-0.028625428676605225,
-0.01605837233364582,
-0.0024486256297677755,
0.0013363693142309785,
0.023542292416095734,
-0.0009715634514577687,
0.029919693246483803,
-0.041751157492399216,
0.017552893608808517,
-0.042682841420173645,
0.00989854708313942,
-0.11063047498464584,
0.005068253725767136,
-0.021664131432771683,
-0.0018811490153893828,
-0.014464793726801872,
-0.03698371350765228,
0.05129488185048103,
0.040889810770750046,
0.08355896174907684,
-0.02721807174384594,
0.0075524915009737015,
0.0030479049310088158,
0.0012496205745264888,
0.009296092204749584,
-0.05077149346470833,
-0.022660383954644203,
0.010740033350884914,
0.0011213356629014015,
-0.04930470511317253,
-0.009913760237395763,
0.02227581851184368,
0.02130819670855999,
0.006613204255700111,
-0.026243174448609352,
-0.048211801797151566,
-0.15337403118610382,
-0.03243180736899376,
-0.029541287571191788,
-0.023232311010360718,
0.02111990936100483,
0.007293100468814373,
0.12589848041534424,
-0.03868534415960312,
0.12349812686443329,
-0.04783383756875992,
0.05089198797941208,
-0.01963716931641102,
0.06677518784999847,
0.034906357526779175,
0.027895350009202957,
0.020741792395710945,
-0.06110874190926552,
0.10639046132564545,
-0.046652261167764664,
-0.017465682700276375,
-0.0076339952647686005,
0.045911677181720734,
-0.013678609393537045,
0.13019663095474243,
0.033020295202732086,
0.05899772793054581,
0.0347219742834568,
0.06494839489459991,
0.07026161253452301,
-0.07017456740140915,
-0.12936937808990479,
0.01869252882897854,
0.02884743921458721,
-0.05095018818974495,
0.00023191387299448252,
0.018503768369555473,
-0.06409632414579391,
0.019973741844296455,
0.06753014028072357,
0.0023049386218190193,
0.14336642622947693,
0.03377711772918701,
-0.08818195760250092,
-0.037861328572034836,
-0.05077685788273811,
0.060609880834817886,
0.10620315372943878,
0.039840858429670334,
-0.0591651126742363,
0.003684992203488946,
0.0836327001452446,
0.05400008708238602,
0.048663120716810226,
0.040196437388658524,
-0.05734655261039734,
0.0543559268116951,
-0.09657640755176544,
-0.0373094342648983,
0.04403473064303398,
-0.054722901433706284,
0.06377299875020981,
-0.0046502226032316685,
0.02274242229759693,
0.017781250178813934,
0.0256390068680048,
0.001365541247650981,
3.226822359135399e-33,
0.014509642496705055,
0.07706508040428162,
0.023217877373099327,
0.020692501217126846,
0.016434891149401665,
0.01384593453258276,
-0.03698861971497536,
0.051982611417770386,
-0.03786760941147804,
0.07211559265851974,
-0.01736615039408207,
-0.05468916520476341,
-0.07095836848020554,
0.061359964311122894,
0.029217449948191643,
0.04081575199961662,
-0.06020664796233177,
0.04510116949677467,
-0.006305215880274773,
-0.006943192798644304,
-0.007113531231880188,
-0.00981308426707983,
-0.03234032914042473,
-0.12339466065168381,
0.00104386149905622,
0.043876346200704575,
0.08469981700181961,
-0.033609066158533096,
-0.041300125420093536,
0.012358086183667183,
-0.13730677962303162,
-0.0776115357875824,
0.056653689593076706,
0.008080486208200455,
0.045923057943582535,
-0.04068218544125557,
-0.0627908855676651,
0.00004994318442186341,
-0.04471726343035698,
0.02037762477993965,
-0.01135244220495224,
0.05590888112783432,
0.04288481920957565,
-0.02899322658777237,
-0.036252882331609726,
-0.10156815499067307,
-0.03101487085223198,
-0.003208072856068611,
0.004714535549283028,
-0.035619135946035385,
0.04668693244457245,
-0.010183889418840408,
-0.06843624264001846,
0.00003346233643242158,
0.0007536183693446219,
0.011998890899121761,
-0.03126385435461998,
0.019211767241358757,
-0.003224861342459917,
0.03892407566308975,
-0.0037302153650671244,
0.034843914210796356,
0.019561264663934708,
-0.022249344736337662,
0.054535236209630966,
0.04804982990026474,
-0.010534949600696564,
0.05736285820603371,
0.015023132786154747,
-0.04077673330903053,
-0.030378302559256554,
-0.06458023190498352,
-0.004206039942800999,
-0.029904186725616455,
0.07285207509994507,
-0.012143826112151146,
0.0788501426577568,
-0.06149900704622269,
0.0156116196885705,
0.012322695925831795,
0.06379393488168716,
-0.00503120431676507,
0.032212357968091965,
-0.04367614910006523,
0.005679624620825052,
-0.07104863226413727,
0.05273131653666496,
-0.10487516969442368,
-0.049387313425540924,
-0.01019271556288004,
-0.0723547711968422,
-0.04435774311423302,
-0.03636854141950607,
-0.02755037695169449,
0.02941073291003704,
-2.9263083303736348e-33,
-0.04092179611325264,
-0.008575071580708027,
0.039542097598314285,
0.04851856455206871,
-0.05679002404212952,
-0.1114031970500946,
-0.03407372534275055,
0.03346697986125946,
-0.0016271431231871247,
-0.12915602326393127,
-0.08207424730062485,
-0.08344686776399612,
0.0404813252389431,
0.04195443168282509,
0.06101936101913452,
-0.017106568440794945,
-0.012763937935233116,
0.00008477736992063001,
0.004362180829048157,
0.055180758237838745,
-0.01310062501579523,
0.09453261643648148,
-0.03676041588187218,
0.049721647053956985,
-0.03801949322223663,
-0.0038615933153778315,
-0.017294490709900856,
0.09788993000984192,
-0.04155771806836128,
-0.04843704774975777,
0.0018721430096775293,
0.009591584093868732,
0.03783116117119789,
-0.00473674526438117,
-0.1076963022351265,
0.030898908153176308,
0.03792865201830864,
-0.020069248974323273,
-0.02902791276574135,
0.09597287327051163,
0.08164980262517929,
-0.015753943473100662,
-0.07864321023225784,
-0.018882237374782562,
-0.030699245631694794,
0.0007431913982145488,
-0.09997119009494781,
0.00007946079131215811,
0.05376812815666199,
0.01264884788542986,
-0.07600778341293335,
0.011092981323599815,
-0.07222642004489899,
0.05995197966694832,
-0.04265972226858139,
-0.1243387907743454,
-0.07630164921283722,
-0.06009417772293091,
0.044478680938482285,
0.0005359398783184588,
-0.04419462010264397,
-0.017868509516119957,
0.09015339612960815,
-0.08282537758350372,
0.04988078027963638,
-0.05959799513220787,
-0.012173967435956001,
0.014287262223660946,
0.018000151962041855,
-0.026468748226761818,
0.018385980278253555,
-0.0636918693780899,
-0.04150442034006119,
0.0753837525844574,
0.020583240315318108,
-0.03598184511065483,
0.011513132601976395,
-0.012712392956018448,
-0.005557835102081299,
-0.010687017813324928,
0.029980771243572235,
0.014153169468045235,
0.02724057249724865,
0.08794084936380386,
0.030984491109848022,
0.028334731236100197,
0.05786251276731491,
0.04846864193677902,
0.019751479849219322,
-0.027548255398869514,
-0.03706920146942139,
0.035663407295942307,
0.037764325737953186,
0.06311562657356262,
-0.035689517855644226,
-4.50329338264055e-8,
-0.06541874259710312,
0.058664362877607346,
0.003112690057605505,
0.06186428666114807,
-0.02722381055355072,
-0.058637700974941254,
-0.07871518284082413,
0.062164321541786194,
-0.05873379111289978,
-0.02012401446700096,
0.034509506076574326,
0.03420629724860191,
-0.05816604197025299,
-0.030587349086999893,
-0.05049659684300423,
0.04509173333644867,
-0.00961585808545351,
0.07425817847251892,
-0.040910713374614716,
-0.010872180573642254,
0.017440924420952797,
0.0878826305270195,
0.00427722604945302,
-0.02230355702340603,
0.07258380204439163,
-0.042175840586423874,
0.023499557748436928,
0.06530505418777466,
0.09163794666528702,
0.020346201956272125,
0.004019640851765871,
0.04205639287829399,
-0.051208678632974625,
-0.03754201531410217,
0.004854694940149784,
0.038867510855197906,
0.021962707862257957,
-0.0520448237657547,
0.06345245987176895,
-0.06790277361869812,
0.06442400813102722,
0.044487908482551575,
-0.07792576402425766,
0.05298168584704399,
-0.0001781201281119138,
-0.004598052706569433,
0.04029933363199234,
-0.09952440857887268,
0.05429396778345108,
-0.08336298912763596,
0.09268521517515182,
-0.03748076781630516,
0.04365992918610573,
0.038777757436037064,
0.07633738964796066,
0.0513320155441761,
-0.01850242353975773,
-0.018350275233387947,
0.03615047037601471,
0.029694456607103348,
0.007642379496246576,
0.05095218867063522,
-0.01986168511211872,
0.00539240101352334
] | 0 | 0 | 0 | 0 |
|
Analyzing Neural Scaling Laws in Two-Layer Networks with Power-Law Data Spectra | https://openreview.net/forum?id=wFD16gwpze | [
"Roman Worschech",
"Bernd Rosenow"
] | Spotlight | Neural scaling laws describe how the performance of deep neural networks scales with key factors such as training data size, model complexity, and training time, often following power-law behaviors over multiple orders of magnitude. Despite their empirical observation, the theoretical understanding of these scaling laws remains limited. In this work, we employ techniques from statistical mechanics to analyze one-pass stochastic gradient descent within a student-teacher framework, where both the student and teacher are two-layer neural networks. Our study primarily focuses on the generalization error and its behavior in response to data covariance matrices that exhibit power-law spectra.
For linear activation functions, we derive analytical expressions for the generalization error, exploring different learning regimes and identifying conditions under which power-law scaling emerges. Additionally, we extend our analysis to non-linear activation functions in the feature learning regime, investigating how power-law spectra in the data covariance matrix impact learning dynamics. Importantly, we find that the length of the symmetric plateau depends on the number of distinct eigenvalues of the data covariance matrix and the number of hidden units, demonstrating how these plateaus behave under various configurations. In addition, our results reveal a transition from exponential to power-law convergence in the specialized phase when the data covariance matrix possesses a power-law spectrum. This work contributes to the theoretical understanding of neural scaling laws and provides insights into optimizing learning performance in practical scenarios involving complex data structures. | Statistical mechanics, neural scaling laws | null | 11,895 | 2410.09005 | [
-0.01374397985637188,
-0.0852392390370369,
0.04431258887052536,
0.07158727943897247,
0.04273277521133423,
0.0002511962375137955,
0.013933172449469566,
-0.03394336998462677,
0.04953800514340401,
-0.03060564026236534,
0.05877893045544624,
0.008348890580236912,
-0.02397531270980835,
0.051925934851169586,
-0.004225284326821566,
0.01553304772824049,
0.0235296543687582,
0.1288144290447235,
-0.1520877480506897,
0.004427660722285509,
0.0601615346968174,
0.048715539276599884,
-0.015561056323349476,
0.030509358271956444,
0.06161341071128845,
-0.010278016328811646,
-0.032166991382837296,
-0.016613498330116272,
-0.05016081780195236,
-0.03628602996468544,
0.04461398720741272,
-0.06935248523950577,
-0.017783068120479584,
0.06669843196868896,
-0.01794813573360443,
-0.018213476985692978,
0.09640359878540039,
0.08064956963062286,
0.01748540811240673,
0.07218953967094421,
0.042221855372190475,
0.0210120789706707,
0.04402219131588936,
0.0291108600795269,
0.03589479252696037,
-0.052626147866249084,
0.03991264849901199,
-0.08442500233650208,
-0.035913970321416855,
-0.030796043574810028,
0.0005370954750105739,
-0.00541719002649188,
-0.06977704167366028,
0.01635567843914032,
0.025586847215890884,
-0.03209713101387024,
0.08844376355409622,
0.0793723464012146,
0.006209124810993671,
0.028418613597750664,
-0.026663120836019516,
-0.12252330780029297,
-0.03996262326836586,
-0.004369967617094517,
0.0030247799586504698,
0.012179899029433727,
0.07162873446941376,
-0.006419590674340725,
-0.0202658548951149,
0.025399912148714066,
-0.02578541822731495,
0.03184165805578232,
-0.07982362061738968,
-0.012298311106860638,
0.07544125616550446,
-0.01196006778627634,
-0.011644338257610798,
0.0024042753502726555,
0.0517403818666935,
-0.07387725263834,
0.007669283542782068,
0.026217065751552582,
0.012530946172773838,
-0.023431790992617607,
0.10303578525781631,
-0.02967745251953602,
0.04198094457387924,
0.05125647783279419,
0.04689671844244003,
-0.038533370941877365,
-0.020358970388770103,
0.04848337545990944,
-0.034155845642089844,
0.028697336092591286,
0.06569322943687439,
-0.06369968503713608,
-0.006943803280591965,
-0.12946562469005585,
0.003777035977691412,
0.03298522159457207,
-0.07475768774747849,
0.0045638843439519405,
-0.009110304526984692,
0.0799170508980751,
0.02217087708413601,
0.003968140110373497,
0.04834407940506935,
-0.00022302927391137928,
0.07959680259227753,
-0.018624888733029366,
0.05962083861231804,
0.029947124421596527,
-0.05908219888806343,
0.015589676797389984,
0.02367805317044258,
0.01688886433839798,
0.08467277139425278,
0.016711296513676643,
-0.03742149844765663,
0.038053322583436966,
-0.04847293719649315,
-0.002283307258039713,
-0.03566920757293701,
-0.04357217997312546,
-0.014555375091731548,
-0.002654962008818984,
-0.055162545293569565,
1.002524489696115e-33,
-0.023725779727101326,
0.010104025714099407,
-0.002905711065977812,
0.04273455962538719,
0.04554055631160736,
-0.07702948153018951,
-0.013484803959727287,
-0.044141922146081924,
0.023984530940651894,
0.06694388389587402,
-0.07818011194467545,
0.08087071031332016,
0.02130187302827835,
0.06134714186191559,
0.04605857655405998,
0.07136780023574829,
-0.0022670396137982607,
-0.0722181424498558,
0.043382249772548676,
-0.07253766059875488,
0.09898395836353302,
-0.018357574939727783,
0.02295853942632675,
0.013203105889260769,
-0.018272975459694862,
-0.03739291429519653,
0.012512371875345707,
0.015625379979610443,
-0.06772640347480774,
0.010169276967644691,
-0.023310426622629166,
-0.03145608305931091,
-0.0034496174193918705,
0.014237064868211746,
0.0065319957211613655,
0.009196123108267784,
0.06696458905935287,
-0.013028405606746674,
0.08260288089513779,
-0.02039523422718048,
-0.034549955278635025,
0.0023950752802193165,
-0.012452444061636925,
-0.024637850001454353,
-0.1049424409866333,
-0.1257917582988739,
0.07098998874425888,
-0.015179937705397606,
-0.018426641821861267,
-0.1421779990196228,
-0.051052093505859375,
0.02379426918923855,
-0.0036168063525110483,
-0.030602235347032547,
0.010392718948423862,
0.0687483698129654,
0.0073173982091248035,
0.04556916654109955,
0.061307910829782486,
0.030536405742168427,
-0.019602002575993538,
-0.0005119657726027071,
0.013113591820001602,
0.04905937984585762,
0.03688779100775719,
0.0015365411527454853,
-0.06279626488685608,
0.0666334480047226,
0.09418846666812897,
-0.04400458559393883,
0.041498083621263504,
0.0011566224275156856,
-0.035736940801143646,
-0.10797269642353058,
0.03728505223989487,
0.006669727154076099,
0.03147398307919502,
-0.1281784027814865,
-0.028454292565584183,
0.10143295675516129,
-0.07559608668088913,
-0.023994144052267075,
0.02149694599211216,
-0.047036148607730865,
-0.10552989691495895,
-0.03838982433080673,
-0.013544746674597263,
-0.09394282847642899,
-0.00046095348079688847,
0.011595009826123714,
-0.059706561267375946,
-0.02925075590610504,
0.031635548919439316,
-0.04332327842712402,
0.01412414200603962,
-2.2053071153532765e-33,
-0.12614284455776215,
0.013745958916842937,
-0.02133864536881447,
0.08112349361181259,
0.06299016624689102,
-0.03178111091256142,
-0.0613589845597744,
0.038828134536743164,
-0.021097736433148384,
-0.04171338677406311,
0.042397789657115936,
-0.07388486713171005,
-0.08613406866788864,
-0.06816872209310532,
0.13508452475070953,
0.009423836134374142,
-0.05428241193294525,
0.06444855034351349,
0.03116726689040661,
-0.059566088020801544,
0.045750606805086136,
0.013240876607596874,
-0.013405237346887589,
0.035575129091739655,
0.03431680426001549,
-0.05115702375769615,
-0.08921736478805542,
0.010927421040832996,
0.06691844016313553,
0.016727304086089134,
-0.10193164646625519,
-0.023196492344141006,
-0.023621322587132454,
0.06632735580205917,
-0.02790415659546852,
0.08226330578327179,
-0.030573464930057526,
-0.03638743981719017,
-0.04022097587585449,
0.06725382804870605,
0.015091175213456154,
-0.04655890539288521,
-0.0032094637863337994,
-0.05556103214621544,
0.06611739099025726,
-0.032712485641241074,
0.05862264707684517,
0.07169277220964432,
-0.023438485339283943,
-0.04574599862098694,
0.07036080956459045,
0.06435877829790115,
-0.0688813328742981,
0.05224515497684479,
-0.04773116484284401,
0.03445759043097496,
0.07020840793848038,
-0.0017503872513771057,
0.06747055053710938,
0.03882979229092598,
-0.030375922098755836,
-0.06821834295988083,
-0.04431015998125076,
0.02904362976551056,
-0.0992896556854248,
-0.02815992757678032,
-0.0035166002344340086,
0.014097120612859726,
0.03117532841861248,
0.029718007892370224,
-0.01064351387321949,
-0.007126777432858944,
-0.0025088072288781404,
-0.0426994189620018,
-0.0960010215640068,
-0.07029663771390915,
-0.028798284009099007,
-0.0651206448674202,
-0.06267831474542618,
-0.022185496985912323,
0.08490922302007675,
0.015473349019885063,
0.0014341931091621518,
-0.0143196452409029,
0.09490123391151428,
0.0749029815196991,
0.03985446318984032,
0.0364169180393219,
0.03124874271452427,
-0.012326075695455074,
-0.013300394639372826,
-0.0015769212041050196,
-0.060523148626089096,
-0.007488100789487362,
-0.04480120167136192,
-4.462658509396533e-8,
-0.04670418053865433,
0.05378522723913193,
0.0818040743470192,
0.015077891759574413,
0.06824193894863129,
0.02057109959423542,
0.0314151905477047,
0.02714960277080536,
-0.011533796787261963,
0.02716962620615959,
0.1201426237821579,
-0.024001771584153175,
0.044163983315229416,
-0.059588056057691574,
-0.03533494845032692,
0.05160269886255264,
-0.0017344107618555427,
0.010227881371974945,
0.029618090018630028,
-0.0565311424434185,
-0.01756681129336357,
-0.03706803545355797,
0.06326473504304886,
-0.02147666923701763,
-0.0015780167886987329,
-0.037682708352804184,
-0.06849158555269241,
0.1172274574637413,
0.052846331149339676,
-0.04182706028223038,
-0.003991507925093174,
-0.006882904097437859,
0.047062020748853683,
-0.034274835139513016,
0.04201067239046097,
0.045480456203222275,
0.0867241621017456,
0.026413841173052788,
-0.010479902848601341,
-0.0009008330525830388,
-0.023027630522847176,
-0.010041764006018639,
0.0970739871263504,
0.04778512567281723,
0.008568419143557549,
0.055447809398174286,
-0.06753053516149521,
-0.0516672283411026,
0.06929349154233932,
-0.007407955825328827,
0.09262087941169739,
0.007872809655964375,
-0.012088587507605553,
0.06544338911771774,
0.017829975113272667,
0.02094031684100628,
-0.04302659630775452,
-0.05956880748271942,
-0.04072660952806473,
0.00634953286498785,
0.024328596889972687,
0.003700033063068986,
-0.11700467020273209,
-0.0584612600505352
] | 0 | 0 | 0 | 0 |
|
Differential learning kinetics govern the transition from memorization to generalization during in-context learning | https://openreview.net/forum?id=INyi7qUdjZ | [
"Alex Nguyen",
"Gautam Reddy"
] | Spotlight | Transformers exhibit in-context learning (ICL): the ability to use novel information presented in the context without additional weight updates. Recent work shows that ICL emerges when models are trained on a sufficiently diverse set of tasks and the transition from memorization to generalization is sharp with increasing task diversity. One interpretation is that a network's limited capacity to memorize favors generalization. Here, we examine the mechanistic underpinnings of this transition using a small transformer applied to a synthetic ICL task. Using theory and experiment, we show that the sub-circuits that memorize and generalize can be viewed as largely independent. The relative *rates* at which these sub-circuits learn explains the transition from memorization to generalization, rather than capacity constraints. We uncover a memorization scaling law, which determines the task diversity threshold at which the network generalizes. The theory quantitatively explains a variety of other ICL-related phenomena, including the long-tailed distribution of when ICL is acquired, the bimodal behavior of solutions close to the task diversity threshold, the influence of contextual and data distributional statistics on ICL, and the transient nature of ICL. | in-context learning, mechanistic interpretability, small transformers, memorization | null | 11,880 | 2412.00104 | [
-0.018513403832912445,
-0.09904149919748306,
0.025550121441483498,
0.025285467505455017,
0.08864076435565948,
-0.06494167447090149,
0.042846500873565674,
0.01114710234105587,
0.11297374218702316,
-0.007905177772045135,
-0.03711608052253723,
-0.0022692782804369926,
0.01314825564622879,
0.06221075356006622,
0.04461653158068657,
-0.039645906537771225,
0.03562014922499657,
0.08112616091966629,
-0.1674538552761078,
-0.015640251338481903,
0.07040230184793472,
-0.00990187656134367,
0.07830782234668732,
0.006598707754164934,
-0.002542755100876093,
0.055583152920007706,
0.01894409768283367,
-0.046579036861658096,
0.021433476358652115,
-0.056959349662065506,
0.036814481019973755,
0.05587972700595856,
-0.014250703155994415,
0.013845758512616158,
-0.025357669219374657,
0.057138241827487946,
-0.1025288999080658,
0.06683080643415451,
0.04733440279960632,
0.02871563471853733,
0.01867852918803692,
0.06531525403261185,
0.0045706406235694885,
0.04735502228140831,
0.07587350904941559,
0.01786775141954422,
-0.005349520593881607,
-0.11117503046989441,
-0.05176481604576111,
-0.01595333218574524,
-0.01843029074370861,
0.012698186561465263,
-0.06021776795387268,
0.07815942168235779,
0.008980858139693737,
0.04038833826780319,
0.07058754563331604,
0.04676302522420883,
-0.046117473393678665,
0.004993576556444168,
-0.054881107062101364,
-0.14247934520244598,
-0.052623771131038666,
-0.05247274786233902,
-0.053386300802230835,
-0.01496878731995821,
0.03974798321723938,
0.036681175231933594,
-0.016799071803689003,
-0.033845335245132446,
-0.03602781519293785,
0.05192277953028679,
-0.08027474582195282,
0.02375330775976181,
0.14483468234539032,
-0.046469394117593765,
0.04473089054226875,
-0.030622312799096107,
0.05065993592143059,
-0.08856494724750519,
0.03349241614341736,
0.036743275821208954,
0.012757032178342342,
-0.030801819637417793,
0.11450154334306717,
0.024935239925980568,
0.01970202475786209,
0.036981746554374695,
0.010954310186207294,
-0.013633626513183117,
-0.040454067289829254,
-0.050850436091423035,
-0.03786563500761986,
-0.02788172848522663,
0.040360040962696075,
0.009447471238672733,
0.036194272339344025,
-0.03389286994934082,
0.019524894654750824,
0.07837027311325073,
0.04282167926430702,
-0.0064644659869372845,
-0.026469172909855843,
0.06538142263889313,
0.005920994095504284,
-0.025532392784953117,
0.021729115396738052,
0.0647471696138382,
0.050400059670209885,
-0.026321865618228912,
-0.00007915434252936393,
0.05626744031906128,
0.04941931366920471,
0.05242978408932686,
-0.016533026471734047,
-0.027112819254398346,
0.04909291863441467,
-0.030639061704277992,
0.02452176623046398,
0.1076323613524437,
-0.028298629447817802,
-0.021254003047943115,
-0.07525311410427094,
0.008084756322205067,
-0.07578308880329132,
-0.015830256044864655,
-0.048971813172101974,
1.7988332911244223e-33,
0.049107626080513,
-0.002102096565067768,
0.014764184132218361,
0.050668954849243164,
0.06774474680423737,
-0.03748916834592819,
-0.011975646950304508,
0.02730533294379711,
-0.0028175278566777706,
0.0003257663920521736,
0.025931628420948982,
0.0981067344546318,
-0.04615818336606026,
0.0903075784444809,
0.025867151096463203,
0.027186524122953415,
-0.10985099524259567,
0.028131883591413498,
0.020310353487730026,
-0.0879630595445633,
0.04367542639374733,
-0.019726812839508057,
0.025565538555383682,
-0.019565576687455177,
-0.043801937252283096,
-0.022882472723722458,
0.00923020951449871,
0.015384328551590443,
-0.011291594244539738,
-0.02021792344748974,
-0.10744797438383102,
0.05564265325665474,
-0.04887540265917778,
0.009065408259630203,
0.02952355518937111,
0.016977230086922646,
0.03162689134478569,
-0.051590364426374435,
0.008592535741627216,
-0.0384351871907711,
-0.025721952319145203,
0.002358394442126155,
-0.002277553314343095,
-0.020377466455101967,
-0.10567568242549896,
-0.055545687675476074,
0.04711858928203583,
0.0025682621635496616,
-0.06560467928647995,
-0.03568284213542938,
-0.00815717875957489,
0.0030234954319894314,
-0.0003388472832739353,
-0.08376234024763107,
0.06809131056070328,
0.07539089024066925,
0.00048513978254050016,
0.014909405261278152,
0.04644173011183739,
0.0669623389840126,
-0.0056821126490831375,
0.08307211101055145,
-0.018237337470054626,
0.166856586933136,
0.05955883488059044,
0.13404013216495514,
-0.07178016006946564,
0.0659925565123558,
0.07731284946203232,
-0.04099269583821297,
0.02511642314493656,
0.04957205057144165,
-0.09566786885261536,
-0.07154514640569687,
0.05935809016227722,
-0.027283675968647003,
0.054780084639787674,
-0.07564806193113327,
-0.03962159529328346,
0.02527572028338909,
-0.01229297835379839,
-0.0591660737991333,
-0.02340034581720829,
-0.014424595050513744,
-0.09169196337461472,
-0.007419969886541367,
0.027376478537917137,
-0.10987628996372223,
-0.04064279794692993,
-0.015328386798501015,
-0.039266061037778854,
-0.03875463083386421,
-0.04636760801076889,
-0.009361039847135544,
-0.0021615761797875166,
-3.819044494603329e-33,
-0.045260436832904816,
-0.00736731244251132,
-0.046259574592113495,
0.02619830146431923,
0.004588763229548931,
0.000727941922377795,
-0.0035186205059289932,
-0.001058835070580244,
-0.08639836311340332,
-0.08686275780200958,
-0.0064377631060779095,
-0.000974026508629322,
-0.02635294757783413,
0.013380581513047218,
-0.03587905690073967,
-0.0413692332804203,
-0.010130519978702068,
0.0835544541478157,
0.02242392487823963,
0.01620817929506302,
0.041996244341135025,
0.0727534145116806,
-0.09816030412912369,
0.006171052344143391,
-0.04045958071947098,
0.020109685137867928,
-0.13308928906917572,
0.07941029220819473,
0.04766973853111267,
0.006507920101284981,
-0.02879122644662857,
0.013810142874717712,
-0.037806786596775055,
-0.010705887340009212,
-0.025150764733552933,
0.09376253187656403,
0.000396352814277634,
0.06140199303627014,
-0.05996561050415039,
0.044229090213775635,
0.006171881686896086,
-0.044934678822755814,
-0.026364121586084366,
0.02013351395726204,
0.05132606625556946,
-0.06232254207134247,
-0.06156664714217186,
-0.05181241035461426,
0.025192076340317726,
-0.006394552998244762,
0.011808100156486034,
0.006524263881146908,
-0.04752366617321968,
-0.062074918299913406,
-0.05721941962838173,
-0.02676454372704029,
0.015278893522918224,
-0.04601221904158592,
0.0846971943974495,
-0.053279902786016464,
-0.030807318165898323,
-0.0439615398645401,
-0.01797366328537464,
-0.0700056329369545,
-0.03179231286048889,
-0.022744478657841682,
-0.02315320447087288,
-0.0688532143831253,
0.05397425964474678,
0.050838299095630646,
0.02776019461452961,
0.039133187383413315,
0.006933493539690971,
-0.08297514170408249,
-0.018978362902998924,
-0.0547596774995327,
-0.0024483983870595694,
-0.026283880695700645,
-0.09449845552444458,
-0.01977650076150894,
-0.013658780604600906,
-0.012964152731001377,
0.008331176824867725,
-0.03134208545088768,
-0.04218935966491699,
0.06928639113903046,
0.028078196570277214,
0.08194468170404434,
0.04442887380719185,
-0.05618758499622345,
-0.02147780731320381,
-0.012644987553358078,
-0.05250495672225952,
0.03908475860953331,
-0.0793798491358757,
-4.798738473255071e-8,
-0.029577983543276787,
0.06786016374826431,
-0.012439731508493423,
0.01744547113776207,
0.07428067177534103,
0.040649399161338806,
-0.025836074724793434,
0.05852404609322548,
-0.009501042775809765,
0.009836127050220966,
0.07027605175971985,
-0.004920711740851402,
0.021239008754491806,
0.017825022339820862,
0.06514274328947067,
0.14445839822292328,
0.05111139640212059,
-0.01882532238960266,
0.03682396933436394,
-0.040966425091028214,
-0.020758843049407005,
-0.019162487238645554,
0.007472645025700331,
-0.031658463180065155,
-0.00044874308514408767,
-0.10889261960983276,
0.005078034941107035,
0.07808488607406616,
0.009186167269945145,
-0.004261309746652842,
-0.02646474353969097,
0.022620392963290215,
-0.0363628976047039,
-0.032820865511894226,
0.020879415795207024,
0.03730103373527527,
0.042524825781583786,
-0.04993337765336037,
0.018774062395095825,
-0.022608673200011253,
-0.03183887526392937,
0.011303630657494068,
-0.03684880957007408,
0.09439080953598022,
-0.013217407278716564,
-0.00009940536983776838,
-0.061351507902145386,
-0.08376144617795944,
0.02818402647972107,
-0.039394501596689224,
0.10330216586589813,
0.03168151155114174,
-0.017404276877641678,
0.007986712269484997,
0.09765565395355225,
0.0654371902346611,
-0.00010836014553206041,
-0.0748787522315979,
-0.04274170473217964,
0.049195609986782074,
0.012687000446021557,
0.03413531184196472,
-0.11168351769447327,
-0.018889648839831352
] | 0 | 0 | 0 | 0 |
|
Multi-session, multi-task neural decoding from distinct cell-types and brain regions | https://openreview.net/forum?id=IuU0wcO0mo | [
"Mehdi Azabou",
"Krystal Xuejing Pan",
"Vinam Arora",
"Ian Jarratt Knight",
"Eva L Dyer",
"Blake Aaron Richards"
] | Spotlight | Recent work has shown that scale is important for improved brain decoding, with more data leading to greater decoding accuracy. However, large-scale decoding across many different datasets is challenging because neural circuits are heterogeneous---each brain region contains a unique mix of cellular sub-types, and the responses to different stimuli are diverse across regions and sub-types. It is unknown whether it is possible to pre-train and transfer brain decoding models between distinct tasks, cellular sub-types, and brain regions. To address these questions, we developed a multi-task transformer architecture and trained it on the entirety of the Allen Institute's Brain Observatory dataset. This dataset contains responses from over 100,000 neurons in 6 areas of the brains of mice, observed with two-photon calcium imaging, recorded while the mice observed different types of visual stimuli. Our results demonstrate that transfer is indeed possible -combining data from different sources is beneficial for a number of downstream decoding tasks. As well, we can transfer the model between regions and sub-types, demonstrating that there is in fact common information in diverse circuits that can be extracted by an appropriately designed model. Interestingly, we found that the model's latent representations showed clear distinctions between different brain regions and cellular sub-types, even though it was never given any information about these distinctions. Altogether, our work demonstrates that training a large-scale neural decoding model on diverse data is possible, and this provides a means of studying the differences and similarities between heterogeneous neural circuits. | neural population, multi-task, transformer, tokenization, two-photon calcium imaging, visual stimuli, cell types | This paper introduces a multi-session, multi-task framework for building large-scale, interpretable models of calcium activity recorded across visual areas and distinct cell-types. | 11,864 | null | [
-0.016955342143774033,
-0.08954102545976639,
0.09178239107131958,
0.009321631863713264,
0.040756259113550186,
-0.027489617466926575,
-0.0037682780530303717,
-0.07523079961538315,
0.03473413363099098,
-0.01905878819525242,
-0.10157661139965057,
-0.0961613804101944,
-0.05509846284985542,
0.08195527642965317,
0.015812207013368607,
-0.0066192676313221455,
0.0022528839763253927,
0.06670708954334259,
-0.09050175547599792,
-0.04318258911371231,
0.021745339035987854,
-0.0483856201171875,
0.004567859694361687,
0.005863449536263943,
0.03785815089941025,
0.05235585197806358,
-0.01070400234311819,
-0.06476332247257233,
-0.021649137139320374,
-0.01093677431344986,
0.08536000549793243,
0.00190010666847229,
-0.05776558071374893,
0.06136523559689522,
-0.05426228418946266,
0.07165969163179398,
-0.0886484757065773,
-0.0521707646548748,
0.027091681957244873,
-0.05552402138710022,
0.04197360947728157,
0.04501769319176674,
0.024205127730965614,
0.05150346830487251,
-0.021352674812078476,
0.029540924355387688,
-0.01100099366158247,
-0.09730725735425949,
0.015748819336295128,
-0.07812591642141342,
0.04869553819298744,
0.006128569599241018,
0.019455015659332275,
0.17893797159194946,
-0.019169898703694344,
0.12126338481903076,
-0.001895994646474719,
0.027105331420898438,
0.00588492676615715,
0.017173223197460175,
-0.07546447962522507,
-0.014886626042425632,
-0.012507817707955837,
-0.04945755377411842,
0.007325476035475731,
0.06282015889883041,
-0.0073810457251966,
0.00991287361830473,
-0.0847722515463829,
-0.034622982144355774,
-0.02436175011098385,
0.026584895327687263,
-0.024466311559081078,
-0.042022351175546646,
0.0717155858874321,
-0.016032135114073753,
0.036185428500175476,
0.015681765973567963,
0.08363664895296097,
-0.08465293794870377,
0.022260814905166626,
0.024049222469329834,
-0.028970075771212578,
0.0077299270778894424,
0.18219627439975739,
0.001635649474337697,
-0.06056533753871918,
0.09066951274871826,
-0.028233354911208153,
-0.011553874239325523,
-0.1124270036816597,
-0.051096007227897644,
0.07796823978424072,
-0.021569114178419113,
-0.03235230967402458,
-0.002818215172737837,
0.01734011247754097,
0.026665519922971725,
0.07033590227365494,
0.03372464329004288,
0.014852373860776424,
0.033220645040273666,
0.06459833681583405,
0.017367510125041008,
-0.0009613598813302815,
-0.054237786680459976,
0.04482262209057808,
0.08919080346822739,
0.0022143188398331404,
-0.0317358672618866,
0.034577250480651855,
0.10558497160673141,
-0.014327578246593475,
0.028722595423460007,
0.0825972855091095,
-0.0011315037263557315,
0.007969561964273453,
0.08069611340761185,
0.024642447009682655,
0.042879898101091385,
-0.03104180470108986,
-0.04727689176797867,
-0.027399640530347824,
0.014152140356600285,
0.05199255421757698,
-0.007728362921625376,
-0.097066730260849,
2.56791681686822e-33,
0.001138988183811307,
0.0157624501734972,
0.015787532553076744,
0.051030296832323074,
0.02580423466861248,
0.0082011166960001,
0.018704717978835106,
0.014920883812010288,
0.012217235751450062,
0.05449959263205528,
-0.023774493485689163,
0.0442487932741642,
-0.04900531843304634,
0.12778154015541077,
0.024804672226309776,
0.02502928115427494,
-0.07104405015707016,
-0.00018967091455124319,
0.04540513828396797,
-0.07790286093950272,
0.0729450210928917,
-0.01569916121661663,
0.008775528520345688,
-0.0014588349731639028,
-0.0059540411457419395,
-0.05324356630444527,
-0.01346040703356266,
-0.05578005686402321,
-0.04019508510828018,
-0.011016609147191048,
-0.15564730763435364,
0.036782339215278625,
0.021430714055895805,
0.001156875747255981,
-0.002514960477128625,
-0.07243821769952774,
0.08959237486124039,
-0.03000275231897831,
0.013696825131773949,
0.01642674207687378,
0.07853960245847702,
0.04180091619491577,
0.026999402791261673,
-0.03663627430796623,
-0.027921389788389206,
-0.007675846107304096,
-0.030190180987119675,
-0.007944654673337936,
-0.03251531720161438,
-0.017544541507959366,
-0.0029597238171845675,
-0.04163609817624092,
-0.012159159407019615,
-0.10686015337705612,
0.041232477873563766,
-0.014729805290699005,
-0.012251303531229496,
0.029488788917660713,
0.11095046997070312,
0.007297770120203495,
0.017185714095830917,
-0.012513687834143639,
-0.03309445083141327,
0.08388654887676239,
0.07228240370750427,
0.060119230300188065,
-0.05166646093130112,
0.048456426709890366,
-0.023710496723651886,
0.005098683759570122,
-0.050599172711372375,
0.03219979256391525,
-0.02747614122927189,
-0.10599583387374878,
0.0954139456152916,
0.007200835272669792,
0.018840940669178963,
-0.007324338890612125,
-0.06606391817331314,
0.03621932491660118,
-0.00659968052059412,
-0.040492985397577286,
-0.08078885078430176,
0.03753071278333664,
0.008823218755424023,
0.08227401226758957,
0.0027958920691162348,
-0.1008114218711853,
-0.09482602030038834,
-0.009991198778152466,
0.019532235339283943,
-0.018207339569926262,
-0.001780316000804305,
-0.12327463179826736,
-0.06684429198503494,
-2.373748843670387e-33,
-0.05001817271113396,
0.013857703655958176,
-0.057115379720926285,
-0.004649519454687834,
0.012353165075182915,
0.008756941184401512,
0.052929360419511795,
-0.055760495364665985,
-0.009858669713139534,
0.006240005139261484,
-0.051464907824993134,
-0.003689225995913148,
0.008024711161851883,
-0.03815887123346329,
-0.014793582260608673,
-0.010565361008048058,
-0.0171213336288929,
0.04951481148600578,
0.042417656630277634,
-0.0470510832965374,
-0.026535669341683388,
0.14028280973434448,
-0.05925901234149933,
0.03780561313033104,
-0.014714915305376053,
0.02501247636973858,
-0.058869749307632446,
0.07829612493515015,
0.061374325305223465,
-0.010473039932549,
-0.08979254961013794,
-0.000509477686136961,
0.005474421661347151,
-0.02125490829348564,
-0.036239318549633026,
0.037605103105306625,
0.04499007761478424,
-0.0331709124147892,
-0.0565849207341671,
0.05466790869832039,
0.04840952157974243,
-0.06447001546621323,
-0.04722251370549202,
0.08216195553541183,
-0.025333428755402565,
-0.004959568381309509,
-0.09630950540304184,
-0.0011032314505428076,
-0.07339280843734741,
0.10337033122777939,
-0.007407044526189566,
0.019920380786061287,
-0.148157998919487,
-0.06268514692783356,
-0.01057839673012495,
-0.05953195318579674,
0.02042844519019127,
0.002186001744121313,
0.03578741103410721,
-0.062007248401641846,
-0.016171282157301903,
-0.02638142928481102,
-0.0047631883062422276,
-0.029939955100417137,
0.05387158691883087,
0.03213515132665634,
0.013351853005588055,
-0.015457582660019398,
0.05738658830523491,
0.04101133719086647,
-0.003496319055557251,
-0.04281355440616608,
0.056940652430057526,
-0.0525163970887661,
0.06777576357126236,
-0.009898442775011063,
0.002093242947012186,
-0.02785061113536358,
0.003517642617225647,
0.039052993059158325,
-0.06843835860490799,
-0.029121331870555878,
-0.008771750144660473,
0.0658520981669426,
0.09882581233978271,
0.03625890240073204,
-0.002793373540043831,
0.06283935159444809,
0.025080841034650803,
-0.06369949877262115,
-0.036011163145303726,
0.029890790581703186,
0.027670877054333687,
-0.014971135184168816,
0.020596228539943695,
-4.6184176483166084e-8,
-0.04135272279381752,
0.023377414792776108,
-0.06748457998037338,
-0.012335645966231823,
-0.019831834360957146,
-0.07531366497278214,
-0.05123846232891083,
0.06390193849802017,
-0.008476123213768005,
0.02838163822889328,
0.078859344124794,
-0.06342623382806778,
-0.004567709751427174,
0.05188675597310066,
0.060398299247026443,
0.0559900663793087,
0.06164650246500969,
0.02121870033442974,
-0.018849758431315422,
-0.03823763504624367,
-0.011814132332801819,
0.05361902341246605,
0.031483303755521774,
-0.023128196597099304,
0.07698013633489609,
-0.058735840022563934,
-0.0695190504193306,
0.09459687769412994,
-0.014493875205516815,
-0.0370294414460659,
0.01342739351093769,
0.019262755289673805,
0.004472706466913223,
-0.029745018109679222,
0.010502209886908531,
-0.02329123206436634,
0.033054251223802567,
0.01444578543305397,
0.01717231795191765,
-0.012379913590848446,
-0.05593293905258179,
-0.028184382244944572,
-0.13445061445236206,
0.06655239313840866,
0.009660759009420872,
-0.029269881546497345,
-0.010807154700160027,
-0.014317456632852554,
0.06155465170741081,
0.05619719997048378,
0.008508431725203991,
0.05682041868567467,
0.04576209932565689,
0.04427023604512215,
0.04370551556348801,
-0.0005193929537199438,
0.004561924375593662,
-0.07807833701372147,
0.0493462048470974,
0.0024194002617150545,
0.051133543252944946,
0.031585998833179474,
-0.0898372009396553,
-0.05755150690674782
] | 0 | 0 | 0 | 0 |
|
Layer Swapping for Zero-Shot Cross-Lingual Transfer in Large Language Models | https://openreview.net/forum?id=vQhn4wrQ6j | [
"Lucas Bandarkar",
"Benjamin Muller",
"Pritish Yuvraj",
"Rui Hou",
"Nayan Singhal",
"Hongjiang Lv",
"Bing Liu"
] | Spotlight | Model merging, such as model souping, is the practice of combining different models with the same architecture together without further training. In this work, we present a model merging methodology that addresses the difficulty of fine-tuning Large Language Models (LLMs) for target tasks in non-English languages, where task-specific data is often unavailable. We focus on mathematical reasoning and without in-language math data, facilitate cross-lingual transfer by composing language and math capabilities. Starting from the same pretrained model, we fine-tune separate "experts" on math instruction data in English and on generic instruction data in the target language. We then replace the top and bottom transformer layers of the math expert directly with layers from the language expert, which consequently enhances math performance in the target language. The resulting merged models outperform the individual experts and other merging methods on the math benchmark, MGSM, by 10% across four major languages where math instruction data is scarce. In addition, this layer swapping is simple, inexpensive, and intuitive, as it is based on an interpretative analysis of the most important parameter changes during the fine-tuning of each expert. The ability to successfully re-compose LLMs for cross-lingual transfer in this manner opens up future possibilities to combine model expertise, create modular solutions, and transfer reasoning capabilities across languages all post hoc. | model souping, model merging, cross-lingual transfer, multilingual, math, mathematical reasoning, LLM, SFT | We transfer math skills to non-English languages simply by swapping in a few layers from a model fine-tuned on those languages into a model fine-tuned on math. | 11,849 | 2410.01335 | [
-0.0074081881903111935,
-0.07289569824934006,
0.07198745757341385,
0.014518056064844131,
-0.020392032340168953,
-0.03231361135840416,
-0.005319767631590366,
0.08664931356906891,
0.015207517892122269,
-0.060593072324991226,
-0.009835601784288883,
-0.03831799328327179,
0.06160057708621025,
0.05328541621565819,
0.01300698984414339,
0.02481149695813656,
0.00952006597071886,
0.10426167398691177,
-0.12471852451562881,
-0.11270342022180557,
-0.02188003994524479,
-0.04513302445411682,
-0.006629243027418852,
-0.021191824227571487,
0.09177088737487793,
0.0013533737510442734,
-0.03237439692020416,
-0.06333750486373901,
0.02262502908706665,
-0.02559422329068184,
-0.02057035081088543,
0.008822483010590076,
-0.09172677248716354,
0.06937939673662186,
-0.08936261385679245,
0.0905381366610527,
-0.052090730518102646,
0.016794392839074135,
0.02025637775659561,
-0.010923832654953003,
-0.04617183655500412,
-0.005808454472571611,
0.09643149375915527,
-0.021746402606368065,
0.046493083238601685,
0.05825158208608627,
-0.024765286594629288,
0.012662245891988277,
-0.01879592053592205,
0.03802547603845596,
-0.11820682883262634,
-0.04881174862384796,
-0.03872242197394371,
0.051948703825473785,
0.0058981808833777905,
0.010555417276918888,
0.11170002073049545,
0.023241061717271805,
-0.025059055536985397,
-0.03727263957262039,
-0.14884725213050842,
-0.04868350178003311,
-0.027848757803440094,
-0.010094931349158287,
0.014212731271982193,
-0.06762583553791046,
0.017557894811034203,
0.10272698104381561,
-0.03167323023080826,
0.06987526267766953,
-0.03674938529729843,
0.07074805349111557,
-0.04867085814476013,
0.049774426966905594,
0.04860767722129822,
0.004288964904844761,
0.040657490491867065,
0.010832958854734898,
0.02338484302163124,
-0.033124860376119614,
-0.0022861193865537643,
-0.013441287912428379,
-0.04057982563972473,
-0.027210548520088196,
0.15219369530677795,
-0.00843940582126379,
-0.0338931605219841,
0.020124033093452454,
0.056777458637952805,
-0.0457550473511219,
0.000528936565387994,
-0.01152896136045456,
0.04705698788166046,
0.06638309359550476,
0.07512685656547546,
0.08445491641759872,
0.01853506825864315,
0.022477397695183754,
-0.06195230036973953,
0.08003044873476028,
0.008156582713127136,
0.03269480913877487,
0.06781598925590515,
-0.04702401161193848,
-0.028022946789860725,
-0.022078758105635643,
0.04240474849939346,
0.04525471478700638,
-0.003221280872821808,
-0.16116955876350403,
0.056391358375549316,
0.04165007919073105,
0.007250072900205851,
0.018794743344187737,
0.0008273885468952358,
-0.024932388216257095,
0.019681181758642197,
-0.03362294286489487,
0.002155447145923972,
0.0480014868080616,
-0.052988577634096146,
0.10286504775285721,
0.04090188816189766,
0.009035283699631691,
0.06007300689816475,
0.06273134797811508,
-0.048969436436891556,
1.035553952408802e-33,
0.042124636471271515,
0.08988082408905029,
0.019651593640446663,
0.07599323987960815,
0.004801418632268906,
-0.04425619915127754,
0.046226393431425095,
0.0429544523358345,
0.02990836836397648,
0.008311888203024864,
-0.009864680469036102,
0.07533954083919525,
-0.0557357482612133,
0.015809183940291405,
0.04810864105820656,
0.017234958708286285,
0.050507497042417526,
-0.002472430467605591,
-0.008678333833813667,
-0.011852502822875977,
0.03727061673998833,
0.019488578662276268,
-0.015864603221416473,
-0.037681929767131805,
-0.03478206321597099,
0.09341011196374893,
0.07930565625429153,
-0.09927541017532349,
-0.05045026168227196,
-0.001126629300415516,
-0.10075941681861877,
-0.012026076205074787,
-0.03825024142861366,
0.023048782721161842,
0.04944508895277977,
-0.06889145821332932,
0.061739590018987656,
-0.08710939437150955,
0.04277907684445381,
0.014709322713315487,
-0.02569730579853058,
0.009433113038539886,
0.021587971597909927,
0.04303925484418869,
-0.03943905979394913,
-0.08981411904096603,
-0.021754615008831024,
0.014678079634904861,
0.01240403763949871,
-0.02957492135465145,
0.054275233298540115,
-0.018531110137701035,
-0.019080307334661484,
-0.0359070710837841,
0.01890125684440136,
0.039492517709732056,
0.0045707691460847855,
0.021641556173563004,
0.0349796786904335,
0.044464077800512314,
-0.0024407170712947845,
0.0686020478606224,
-0.0034932633861899376,
0.10160090029239655,
0.03179173171520233,
0.01548931747674942,
-0.013114732690155506,
0.06287258118391037,
0.07967785745859146,
-0.03228195011615753,
-0.12379072606563568,
-0.060155462473630905,
0.003787814872339368,
0.026826169341802597,
0.03219100832939148,
-0.05481391400098801,
-0.024475520476698875,
-0.11184491962194443,
0.07904281467199326,
0.01974192075431347,
0.10814468562602997,
0.03400779142975807,
0.010571069084107876,
-0.0847339853644371,
0.0025475535076111555,
-0.026407502591609955,
0.02239007130265236,
-0.07276543229818344,
0.02906917594373226,
0.011453467421233654,
-0.02435307949781418,
-0.05265381932258606,
-0.02830919809639454,
-0.052683036774396896,
0.012230103835463524,
-7.819050467043995e-34,
-0.04600505158305168,
-0.05714334547519684,
-0.06145218387246132,
0.09615568071603775,
-0.05372840538620949,
-0.08682388812303543,
0.08283928781747818,
0.007160789333283901,
-0.0063886940479278564,
-0.06922746449708939,
-0.04657071828842163,
-0.0467936247587204,
0.030739180743694305,
-0.0004068249254487455,
0.016912776976823807,
-0.02069138176739216,
-0.05384110286831856,
0.024008171632885933,
0.007478589657694101,
0.025111103430390358,
-0.03801218420267105,
0.07527892291545868,
-0.05586309731006622,
0.04547495022416115,
-0.062287699431180954,
0.021695533767342567,
-0.0480763204395771,
0.056365881115198135,
0.021517910063266754,
0.00765569694340229,
-0.011591587215662003,
-0.08703910559415817,
0.01424043532460928,
-0.06991114467382431,
-0.0605522021651268,
0.052063677459955215,
0.04327986761927605,
0.005334546789526939,
-0.03965224325656891,
0.07799817621707916,
0.03352810814976692,
-0.06678550690412521,
-0.011550000868737698,
-0.02343217097222805,
-0.007491310127079487,
0.015487369149923325,
-0.13310736417770386,
-0.030978789553046227,
-0.05163636431097984,
0.0010922813089564443,
-0.020898621529340744,
-0.014403467997908592,
-0.003621593350544572,
-0.0809176042675972,
-0.004939347971230745,
-0.09370161592960358,
0.016826307401061058,
-0.08582043647766113,
0.02717861346900463,
-0.009180829860270023,
0.024012025445699692,
-0.003267529886215925,
0.0363793708384037,
-0.08324363082647324,
0.01725509576499462,
0.00639510340988636,
-0.06429606676101685,
-0.03880227729678154,
-0.02679559774696827,
0.04054880142211914,
0.0067665548995137215,
-0.06438908725976944,
0.017568066716194153,
-0.043253425508737564,
-0.005981649272143841,
-0.007004549726843834,
0.02218341827392578,
0.024718113243579865,
0.026848696172237396,
-0.049358069896698,
-0.0012911239173263311,
-0.033895093947649,
0.018732495605945587,
-0.011338739655911922,
0.02422977052628994,
0.021205270662903786,
0.03195931017398834,
0.10573214292526245,
-0.011535506695508957,
-0.08226371556520462,
-0.061384011059999466,
0.021400609984993935,
0.027290362864732742,
0.030142148956656456,
-0.0759769156575203,
-4.2615880602170364e-8,
-0.003874334041029215,
0.09859925508499146,
0.03963581845164299,
0.0513850599527359,
0.0011359435738995671,
-0.06292879581451416,
-0.11238540709018707,
0.018368422985076904,
-0.019749928265810013,
0.022867951542139053,
0.03417592868208885,
0.005449663382023573,
-0.022553080692887306,
-0.05237651616334915,
0.0034525119699537754,
0.09851764142513275,
0.027370385825634003,
0.04674413055181503,
-0.016173165291547775,
-0.04240792617201805,
0.07044980674982071,
0.08077719807624817,
-0.010758435353636742,
0.005752161610871553,
0.049621500074863434,
-0.04512689262628555,
-0.02218012884259224,
0.13923825323581696,
0.12320576608181,
-0.038414351642131805,
-0.03769376128911972,
-0.05676552653312683,
0.008360396139323711,
0.01517996471375227,
0.05285019427537918,
0.04428265616297722,
0.036816589534282684,
0.04363059997558594,
0.010282379575073719,
-0.022010978311300278,
0.021141424775123596,
0.04345586895942688,
-0.05099586397409439,
0.05918950214982033,
0.09581143409013748,
-0.029325982555747032,
-0.04299568012356758,
-0.09032760560512543,
0.0357670933008194,
0.04129938408732414,
0.01330553088337183,
-0.0019440847681835294,
-0.0727062001824379,
0.05885719507932663,
0.03191180154681206,
0.04816179722547531,
0.0003866012557409704,
-0.03464126214385033,
-0.04301037639379501,
0.035302817821502686,
-0.010845466516911983,
-0.016248706728219986,
-0.06884291768074036,
0.02308998629450798
] | 0 | 0 | 0 | 0 |
|
Bayesian Optimization via Continual Variational Last Layer Training | https://openreview.net/forum?id=1jcnvghayD | [
"Paul Brunzema",
"Mikkel Jordahn",
"John Willes",
"Sebastian Trimpe",
"Jasper Snoek",
"James Harrison"
] | Spotlight | Gaussian Processes (GPs) are widely seen as the state-of-the-art surrogate models for Bayesian optimization (BO) due to their ability to model uncertainty and their performance on tasks where correlations are easily captured (such as those defined by Euclidean metrics) and their ability to be efficiently updated online. However, the performance of GPs depends on the choice of kernel, and kernel selection for complex correlation structures is often difficult or must be made bespoke. While Bayesian neural networks (BNNs) are a promising direction for higher capacity surrogate models, they have so far seen limited use due to poor performance on some problem types. In this paper, we propose an approach which shows competitive performance on many problem types, including some that BNNs typically struggle with. We build on variational Bayesian last layers (VBLLs), and connect training of these models to exact conditioning in GPs. We exploit this connection to develop an efficient online training algorithm that interleaves conditioning and optimization. Our findings suggest that VBLL networks significantly outperform GPs and other BNN architectures on tasks with complex input correlations, and match the performance of well-tuned GPs on established benchmark tasks. | Bayesian deep learning, bayesian optimization, uncertainty | We develop an efficient and expressive Bayesian neural network surrogate for Bayesian optimization | 11,786 | 2412.09477 | [
-0.08343546837568283,
-0.15534882247447968,
0.04414954036474228,
0.007594355382025242,
-0.012224316596984863,
-0.0027372415643185377,
0.003492174204438925,
-0.005369666498154402,
-0.005061769858002663,
-0.06002519652247429,
-0.06802228838205338,
0.020255325362086296,
0.056661155074834824,
0.04961729794740677,
-0.09054769575595856,
-0.025209292769432068,
0.143124520778656,
0.08802103996276855,
-0.05559874325990677,
-0.0032828242983669043,
-0.02784062922000885,
0.08201831579208374,
-0.02376823127269745,
0.032199662178754807,
-0.01225853618234396,
-0.0026913159526884556,
0.07654140144586563,
-0.020404715090990067,
0.005521758459508419,
0.010530304163694382,
0.07304776459932327,
-0.05503029376268387,
-0.059051841497421265,
-0.05570317804813385,
-0.03696460649371147,
0.04574926942586899,
-0.12082136422395706,
-0.029024934396147728,
0.023019418120384216,
-0.06406082957983017,
-0.00220604264177382,
-0.04438606649637222,
0.01379761565476656,
0.10420676320791245,
0.055581942200660706,
0.034621741622686386,
0.022359544411301613,
-0.06550757586956024,
0.08895087987184525,
-0.003347791265696287,
-0.06946252286434174,
-0.03276298940181732,
-0.02910522371530533,
0.0043816938996315,
0.03339795023202896,
0.02350851148366928,
-0.020785223692655563,
0.027841150760650635,
0.029524032026529312,
0.034783463925123215,
-0.0007954373140819371,
-0.11687316000461578,
-0.02088157646358013,
-0.06247597932815552,
0.001969980075955391,
0.042656295001506805,
0.03637566789984703,
0.06333363056182861,
0.04893500357866287,
0.05311916768550873,
-0.017469042912125587,
0.15751589834690094,
-0.07884124666452408,
0.022563589736819267,
-0.05963620916008949,
0.03735890984535217,
0.04648686200380325,
-0.007121497765183449,
0.041624441742897034,
-0.015744606032967567,
0.035134728997945786,
0.014676593244075775,
0.046546705067157745,
0.004661853425204754,
0.11612274497747421,
-0.05675327032804489,
-0.009820280596613884,
0.05690240114927292,
0.07487699389457703,
-0.027675412595272064,
0.04506249353289604,
0.04265166074037552,
0.01834152080118656,
-0.08712639659643173,
0.004750996362417936,
-0.012193284928798676,
-0.04546518996357918,
-0.02207479625940323,
0.05691317841410637,
0.06562164425849915,
0.0071669090539216995,
0.054251983761787415,
0.06517045199871063,
0.025875413790345192,
-0.016646800562739372,
-0.04703996330499649,
-0.005356519483029842,
0.11458125710487366,
0.030451184138655663,
-0.09731218963861465,
0.0019098111661151052,
-0.01972603052854538,
-0.010011616162955761,
-0.014676137827336788,
-0.028846533969044685,
0.07024233788251877,
0.007723737508058548,
-0.013272413983941078,
-0.01911870762705803,
0.14889103174209595,
-0.026566335931420326,
-0.0019909858237951994,
-0.02497524581849575,
-0.011304954066872597,
0.007690113037824631,
0.05201577767729759,
-0.02754533849656582,
3.329380200482812e-33,
0.018728498369455338,
0.02159930393099785,
0.07716520875692368,
0.0170427393168211,
-0.02305557020008564,
0.02732062339782715,
-0.023030560463666916,
0.021693605929613113,
-0.014065238647162914,
-0.00010831709369085729,
-0.043732285499572754,
0.02847379259765148,
-0.14487417042255402,
0.0070411404594779015,
0.12516295909881592,
0.019177064299583435,
-0.08268672972917557,
0.0006826932076364756,
-0.038140203803777695,
-0.05607043206691742,
0.07045865058898926,
-0.04173995554447174,
-0.026597853749990463,
0.007250379305332899,
0.0476539172232151,
-0.007982839830219746,
0.12007125467061996,
-0.03721783310174942,
-0.07430130243301392,
-0.014556379988789558,
-0.07471814006567001,
-0.0436212420463562,
-0.028657231479883194,
0.0749148353934288,
0.08069492876529694,
-0.025614051148295403,
0.00426082219928503,
-0.013314582407474518,
0.00007679512782488018,
-0.03606698662042618,
0.015391053631901741,
-0.0059750862419605255,
0.032237712293863297,
0.04097433388233185,
-0.06280303746461868,
-0.13171374797821045,
0.05552152916789055,
-0.028582792729139328,
0.003290296532213688,
-0.08969677239656448,
0.026448331773281097,
0.011275184340775013,
0.009725603275001049,
-0.057963307946920395,
-0.05280854552984238,
-0.017206300050020218,
0.0461878776550293,
0.036306656897068024,
0.008576609194278717,
0.020112572237849236,
0.035329535603523254,
0.015085109509527683,
-0.018412955105304718,
0.012444495223462582,
0.006742454599589109,
0.030055105686187744,
-0.007267516106367111,
0.036465588957071304,
0.033856626600027084,
0.007237880490720272,
0.01555904746055603,
0.014416114427149296,
0.06742394715547562,
-0.09305329620838165,
0.10689892619848251,
0.021357499063014984,
0.015753768384456635,
-0.022744813933968544,
-0.05307967960834503,
-0.0023427975829690695,
-0.01998062990605831,
0.0067636677995324135,
-0.04884373024106026,
0.0023404257372021675,
0.0007368593360297382,
-0.032750897109508514,
-0.0016565796686336398,
-0.03342435881495476,
-0.04703256115317345,
0.024172425270080566,
-0.06514925509691238,
-0.036928895860910416,
-0.010021451860666275,
-0.027338426560163498,
-0.06692106276750565,
-1.936686452381229e-33,
-0.035229384899139404,
-0.02530885674059391,
0.02640475519001484,
0.04705867916345596,
-0.03497718274593353,
-0.057206206023693085,
0.014672367833554745,
-0.03470434993505478,
-0.03738577291369438,
-0.06110241636633873,
-0.09458574652671814,
-0.049771614372730255,
0.09494731575250626,
0.04298924282193184,
-0.03480095416307449,
0.016382437199354172,
-0.09473102539777756,
-0.0015355735085904598,
0.02936924248933792,
-0.04738888889551163,
-0.02722517028450966,
0.04117560386657715,
0.006985615938901901,
-0.01707795076072216,
0.027258148416876793,
0.05515405908226967,
0.06335191428661346,
0.08865819126367569,
0.046088967472314835,
0.040176089853048325,
-0.0707959234714508,
-0.057287804782390594,
-0.05704580992460251,
-0.011213411577045918,
-0.0631784200668335,
0.1022927388548851,
0.06262680143117905,
0.045748986303806305,
-0.00306156137958169,
0.03887501731514931,
-0.0108147282153368,
-0.06495000422000885,
-0.008381870575249195,
-0.048145946115255356,
0.03675679489970207,
-0.03215932846069336,
0.03238360583782196,
0.007730698678642511,
-0.010279221460223198,
-0.009995880536735058,
0.020216621458530426,
0.09819389879703522,
-0.11025156825780869,
0.07236051559448242,
-0.07770678400993347,
-0.06338568031787872,
-0.09997467696666718,
-0.07924706488847733,
0.07985032349824905,
-0.008047910407185555,
0.01992645673453808,
-0.04058728367090225,
0.014973831363022327,
-0.03727605566382408,
-0.014779252000153065,
0.01817242242395878,
0.01963978260755539,
0.06435568630695343,
-0.014674320816993713,
0.04444955661892891,
-0.03186579421162605,
-0.030173875391483307,
0.0492522157728672,
0.016672741621732712,
-0.000998830539174378,
0.028155719861388206,
0.018778204917907715,
0.03346913307905197,
-0.024347834289073944,
-0.02443603053689003,
0.04072544351220131,
0.007706110831350088,
0.022393478080630302,
0.014069718308746815,
0.08225372433662415,
0.11364130675792694,
0.07369200885295868,
0.049500565975904465,
0.024016184732317924,
-0.037515077739953995,
-0.0658593699336052,
0.009980828501284122,
-0.023172641173005104,
0.04338116943836212,
-0.08125526458024979,
-4.358982508279041e-8,
-0.04830242320895195,
0.07422678172588348,
0.07101412862539291,
-0.05479208379983902,
0.09290105849504471,
-0.07998181879520416,
0.0588531494140625,
0.04164251312613487,
-0.03443233668804169,
0.0208978820592165,
0.03263597935438156,
-0.04693704470992088,
-0.007633867673575878,
-0.06483915448188782,
0.026510782539844513,
0.03444330394268036,
-0.006302340887486935,
-0.019741374999284744,
0.013195442035794258,
-0.05124429240822792,
0.05879967287182808,
0.03823235258460045,
0.00784284807741642,
0.01792348362505436,
0.03955848887562752,
-0.07272200286388397,
-0.043195690959692,
0.03676394745707512,
0.1129598543047905,
0.020493650808930397,
-0.01699574664235115,
-0.003276420058682561,
0.03925861045718193,
-0.006362887565046549,
-0.010881311260163784,
0.02529248595237732,
0.019013050943613052,
0.02466414123773575,
-0.05307750031352043,
-0.04659099504351616,
-0.03181128948926926,
0.01955939643085003,
-0.08334172517061234,
0.055104732513427734,
0.057348113507032394,
0.0004309756914153695,
0.04095114767551422,
-0.023084761574864388,
0.0491039864718914,
0.02163192629814148,
0.005345230456441641,
-0.03675076365470886,
-0.04679536446928978,
0.1091187596321106,
0.056683823466300964,
-0.011078815907239914,
0.0963178426027298,
-0.10025421530008316,
-0.02731207013130188,
0.0629182755947113,
-0.05821553245186806,
0.014602557756006718,
-0.11094105243682861,
0.01955810934305191
] | 0 | 0 | 0 | 0 |
|
Bayesian Optimization of Antibodies Informed by a Generative Model of Evolving Sequences | https://openreview.net/forum?id=E48QvQppIN | [
"Alan Nawzad Amin",
"Nate Gruver",
"Yilun Kuang",
"Yucen Lily Li",
"Hunter Elliott",
"Calvin McCarter",
"Aniruddh Raghu",
"Peyton Greenside",
"Andrew Gordon Wilson"
] | Spotlight | To build effective therapeutics, biologists iteratively mutate antibody sequences to improve binding and stability. Proposed mutations can be informed by previous measurements or by learning from large antibody databases to predict only typical antibodies. Unfortunately, the space of typical antibodies is enormous to search, and experiments often fail to find suitable antibodies on a budget. We introduce Clone-informed Bayesian Optimization (CloneBO), a Bayesian optimization procedure that efficiently optimizes antibodies in the lab by teaching a generative model how our immune system optimizes antibodies. Our immune system makes antibodies by iteratively evolving specific portions of their sequences to bind their target strongly and stably, resulting in a set of related, evolving sequences known as a *clonal family*. We train a large language model, CloneLM, on hundreds of thousands of clonal families and use it to design sequences with mutations that are most likely to optimize an antibody within the human immune system. We propose to guide our designs to fit previous measurements with a twisted sequential Monte Carlo procedure. We show that CloneBO optimizes antibodies substantially more efficiently than previous methods in realistic *in silico* experiments and designs stronger and more stable binders in *in vitro* wet lab experiments. | Bayesian optimization, generative model, antibody, biological sequence | null | 11,784 | 2412.07763 | [
-0.09708671271800995,
-0.10821616649627686,
0.030827803537249565,
-0.04576405882835388,
-0.021802248433232307,
-0.06340471655130386,
0.03235376998782158,
0.023624397814273834,
0.04325935244560242,
-0.03233359009027481,
-0.0407315231859684,
-0.014812059700489044,
0.04047445207834244,
0.0004722924204543233,
-0.06209998205304146,
0.07212209701538086,
0.028007306158542633,
0.07100862264633179,
0.019865291193127632,
-0.0612516850233078,
-0.03648589923977852,
0.025666719302535057,
0.014115114696323872,
-0.0029647015035152435,
-0.035339467227458954,
0.0007917072507552803,
0.02513355389237404,
-0.013641898520290852,
-0.02732006274163723,
-0.023670678958296776,
0.005432289559394121,
-0.013433772139251232,
-0.005907940678298473,
-0.06773467361927032,
-0.04664541408419609,
0.013878952711820602,
-0.10488881170749664,
-0.02494584210216999,
0.0012650948483496904,
-0.0076746949926018715,
-0.016893817111849785,
0.02025245502591133,
0.043374672532081604,
-0.0037146862596273422,
0.008460440672934055,
-0.02573530748486519,
-0.02620193362236023,
0.061442866921424866,
0.08901998400688171,
0.0008731044945307076,
-0.11660608649253845,
-0.06931452453136444,
-0.04021704941987991,
0.017511380836367607,
0.030717937275767326,
-0.05066562816500664,
0.014535456895828247,
-0.044971764087677,
0.038366395980119705,
-0.010237143374979496,
-0.03419814631342888,
-0.06749077886343002,
0.02914777584373951,
0.01953365094959736,
0.02258037030696869,
0.012894764542579651,
0.02014019340276718,
0.10251685976982117,
-0.012992560863494873,
0.06209307909011841,
-0.030929209664463997,
0.12288078665733337,
-0.059953510761260986,
0.07588966935873032,
0.04667908325791359,
0.07215713709592819,
0.03324636071920395,
0.05348685383796692,
0.04510900005698204,
-0.09005526453256607,
-0.06577103585004807,
0.006754790898412466,
0.04408454895019531,
0.02786019630730152,
0.058053117245435715,
-0.017321312800049782,
0.013418393209576607,
0.019899610430002213,
0.08360975235700607,
0.016330204904079437,
0.02817572094500065,
-0.003786211833357811,
0.02342512086033821,
-0.020093947649002075,
-0.014133667573332787,
0.045622169971466064,
0.03757930174469948,
-0.027202915400266647,
0.05579366907477379,
0.07302621752023697,
-0.02674012817442417,
0.057133883237838745,
0.11699070036411285,
-0.058067504316568375,
0.048895347863435745,
-0.109227754175663,
0.011607436463236809,
0.008516988717019558,
0.0010396537836641073,
-0.04207206889986992,
0.01986701413989067,
0.021634913980960846,
0.04875795170664787,
0.028437526896595955,
0.01078793779015541,
0.01663064770400524,
0.012844341807067394,
-0.01709560863673687,
-0.03174372762441635,
0.1242453083395958,
0.10380692780017853,
-0.00998864695429802,
-0.048364292830228806,
0.021623114123940468,
0.01170213520526886,
0.05864318460226059,
-0.08051750808954239,
-3.523442723572182e-35,
0.08842586725950241,
0.025831343606114388,
0.10514657944440842,
0.11105617880821228,
-0.029530106112360954,
-0.021162543445825577,
0.010004192590713501,
-0.03612443432211876,
-0.13570590317249298,
0.034419991075992584,
0.018118932843208313,
0.00672559579834342,
-0.09944537281990051,
0.09965585917234421,
0.07188506424427032,
-0.004600953310728073,
-0.03588259220123291,
0.06543358415365219,
-0.052824586629867554,
-0.000989670166745782,
0.033541008830070496,
-0.009055841714143753,
-0.040804747492074966,
-0.1028989627957344,
0.03340280055999756,
0.02880466729402542,
0.005709666758775711,
-0.009671985171735287,
-0.08514080941677094,
0.00496272835880518,
-0.11962977051734924,
0.02674688771367073,
-0.009662535041570663,
0.07358016818761826,
0.01159736979752779,
-0.08455853164196014,
0.039378535002470016,
-0.0771927759051323,
0.007884933613240719,
-0.014252298511564732,
0.07497037202119827,
-0.009295673109591007,
0.05068616941571236,
0.016506094485521317,
0.05709102749824524,
-0.055919159203767776,
-0.029866758733987808,
-0.07125874608755112,
0.020071592181921005,
-0.08207883685827255,
0.06277794390916824,
-0.023662202060222626,
0.03258330002427101,
-0.01206294447183609,
-0.028091702610254288,
0.019389312714338303,
-0.04507523402571678,
-0.014387072995305061,
0.005018563941121101,
0.05535485967993736,
0.009135737083852291,
0.006709841080009937,
-0.002270977245643735,
0.04044251888990402,
0.030630258843302727,
-0.007118510082364082,
-0.01233578473329544,
0.03307619318366051,
0.024696839973330498,
0.030813777819275856,
0.0395917147397995,
-0.09329753369092941,
-0.04058678820729256,
-0.06285671144723892,
0.05535212531685829,
-0.036258962005376816,
0.08051649481058121,
0.01091743167489767,
-0.048755574971437454,
-0.03171645104885101,
0.004510394763201475,
0.08231420814990997,
-0.10366537421941757,
0.018380846828222275,
0.016092471778392792,
-0.12241272628307343,
0.02833211049437523,
-0.04235824570059776,
-0.15986859798431396,
-0.03717130795121193,
-0.004856262821704149,
0.018606487661600113,
-0.03333150967955589,
-0.012813695706427097,
-0.03635844960808754,
-1.7813301802406785e-33,
-0.038990359753370285,
-0.09234119951725006,
0.08452680706977844,
0.035247694700956345,
-0.03660666197538376,
-0.10772077739238739,
0.012753976508975029,
-0.01563754677772522,
-0.01198731828480959,
-0.045527104288339615,
-0.02735566906630993,
-0.04837457463145256,
-0.009820365346968174,
-0.010934717953205109,
-0.02914958819746971,
0.025693226605653763,
0.010988864116370678,
0.05420287325978279,
-0.038014255464076996,
-0.00008467308362014592,
0.013673747889697552,
0.13428254425525665,
0.022803260013461113,
-0.04849570244550705,
-0.05053842440247536,
-0.009564418345689774,
-0.009332992136478424,
0.15042199194431305,
0.07080362737178802,
-0.020858848467469215,
-0.12658600509166718,
0.028709223493933678,
-0.03693024814128876,
0.04555550962686539,
-0.10554250329732895,
-0.00002220185888290871,
0.09789448231458664,
0.037599723786115646,
0.059795789420604706,
-0.020823772996664047,
-0.022413019090890884,
-0.025840017944574356,
-0.07628913968801498,
-0.03590567037463188,
0.1149059608578682,
-0.007340561598539352,
0.05393604189157486,
-0.020703841000795364,
0.05311109125614166,
-0.010481991805136204,
0.027013104408979416,
0.01429759245365858,
-0.0099748894572258,
0.011973457410931587,
-0.1132567822933197,
-0.07373253256082535,
-0.06978330761194229,
-0.042826637625694275,
0.08578801900148392,
-0.0021688230335712433,
-0.08104810863733292,
-0.048089802265167236,
0.0623798631131649,
0.01714012771844864,
-0.016671059653162956,
0.03230400010943413,
-0.009135991334915161,
0.014475969597697258,
0.01095900870859623,
-0.04476093500852585,
-0.03466803953051567,
0.0030030193738639355,
0.023248309269547462,
0.025692898780107498,
0.06366751343011856,
-0.004179341718554497,
-0.045991428196430206,
-0.03223054111003876,
0.0048693763092160225,
-0.011605601757764816,
-0.055058907717466354,
-0.015803856775164604,
-0.03562808036804199,
0.05477059260010719,
0.045602791011333466,
0.0065266783349215984,
0.05952620878815651,
0.036073461174964905,
0.08791064471006393,
-0.004316991660743952,
-0.09325525164604187,
-0.006135138217359781,
0.009032110683619976,
-0.020364869385957718,
-0.10340818017721176,
-3.872614584565781e-8,
0.03363002464175224,
0.025809455662965775,
0.07125459611415863,
0.0051052565686404705,
0.08018454164266586,
0.03955692797899246,
-0.10060831159353256,
0.0096665620803833,
-0.0054208035580813885,
-0.06517991423606873,
0.036518778651952744,
0.08512303978204727,
0.06718863546848297,
-0.008439403958618641,
0.021789714694023132,
0.028904026374220848,
-0.009593892842531204,
-0.0419604666531086,
-0.04434982314705849,
-0.019959399476647377,
0.007056737784296274,
0.09709220379590988,
0.023113463073968887,
0.0409870445728302,
0.03836161643266678,
-0.1127936840057373,
-0.013018584810197353,
-0.014684164896607399,
0.054799649864435196,
0.00434713764116168,
-0.04128620773553848,
0.0034968042746186256,
-0.07401611655950546,
0.04238960146903992,
-0.04067255184054375,
-0.010976854711771011,
-0.04214394837617874,
-0.05395490676164627,
0.04328044131398201,
-0.10714177787303925,
0.04369620606303215,
0.026793038472533226,
-0.06480032205581665,
-0.0446368008852005,
0.021471450105309486,
-0.01590718701481819,
-0.008683385327458382,
-0.04723317548632622,
0.055929627269506454,
-0.010785086080431938,
0.01095180120319128,
-0.019788755103945732,
-0.026700938120484352,
-0.03332741558551788,
0.034778062254190445,
0.02572173811495304,
0.022266538813710213,
-0.009425978176295757,
0.06172958388924599,
0.037123631685972214,
-0.00047241427819244564,
-0.0014436451019719243,
0.015231679193675518,
0.007613991387188435
] | https://github.com/alannawzadamin/clonebo | 8 | 0 | 0 | 0 |
Meta-Dynamical State Space Models for Integrative Neural Data Analysis | https://openreview.net/forum?id=SRpq5OBpED | [
"Ayesha Vermani",
"Josue Nassar",
"Hyungju Jeon",
"Matthew Dowling",
"Il Memming Park"
] | Spotlight | Learning shared structure across environments facilitates rapid learning and adaptive behavior in neural systems. This has been widely demonstrated and applied in machine learning to train models that are capable of generalizing to novel settings. However, there has been limited work exploiting the shared structure in neural activity during similar tasks for learning latent dynamics from neural recordings.
Existing approaches are designed to infer dynamics from a single dataset and cannot be readily adapted to account for statistical heterogeneities across recordings. In this work, we hypothesize that similar tasks admit a corresponding family of
related solutions and propose a novel approach for meta-learning this solution space from task-related neural activity of trained animals. Specifically, we capture the variabilities across recordings on a low-dimensional manifold which concisely parametrizes this family of dynamics, thereby facilitating rapid learning of latent dynamics given new recordings. We demonstrate the efficacy of our approach on
few-shot reconstruction and forecasting of synthetic dynamical systems, and neural recordings from the motor cortex during different arm reaching tasks. | neural dynamics, state-space model, meta learning | null | 11,740 | 2410.05454 | [
-0.06979741156101227,
-0.11878892034292221,
0.0975097119808197,
0.06351446360349655,
0.045377466827631,
0.02171015925705433,
0.011939232237637043,
-0.062514528632164,
0.058809809386730194,
-0.0489395447075367,
-0.0472712442278862,
-0.06763955950737,
-0.03934931382536888,
0.10450603812932968,
-0.0023461240343749523,
-0.062378913164138794,
-0.026929089799523354,
0.07759640365839005,
-0.05131195858120918,
0.03836710751056671,
-0.015248755924403667,
0.009817264042794704,
0.06183704733848572,
0.10890261083841324,
0.0007385583012364805,
0.05476528778672218,
0.02564140222966671,
0.030636653304100037,
-0.028398163616657257,
-0.03032951056957245,
0.10313772410154343,
-0.033251259475946426,
-0.07907968759536743,
-0.047057051211595535,
-0.05437926575541496,
0.04951727017760277,
-0.02269805036485195,
0.050164610147476196,
-0.0034361628349870443,
-0.00665541086345911,
0.05285561829805374,
0.07255726307630539,
0.03303755447268486,
0.029305750504136086,
0.011703193187713623,
0.05972186103463173,
0.01761225052177906,
-0.16418960690498352,
0.02015269733965397,
0.019021930173039436,
-0.12914538383483887,
0.023167777806520462,
0.025962578132748604,
0.09609156847000122,
0.0185999758541584,
0.05269286409020424,
0.022000014781951904,
-0.008417501114308834,
-0.012211596593260765,
-0.019282061606645584,
0.018519680947065353,
-0.06471829861402512,
0.021519549190998077,
-0.02013750933110714,
-0.0038048899732530117,
0.07238055020570755,
-0.037899166345596313,
0.11795797199010849,
-0.06791969388723373,
-0.055740732699632645,
-0.061192166060209274,
0.09888499975204468,
-0.047011081129312515,
-0.018570393323898315,
0.03605432063341141,
0.06839391589164734,
0.004348457790911198,
0.014274238608777523,
0.03488937020301819,
-0.04806356132030487,
0.03445272892713547,
0.045413509011268616,
-0.00028446235228329897,
0.03583183139562607,
0.05932591110467911,
0.032246485352516174,
-0.007646436803042889,
0.0360785536468029,
0.003581844735890627,
-0.041084401309490204,
-0.06345341354608536,
-0.04473400488495827,
-0.054980043321847916,
-0.015434861183166504,
0.041000545024871826,
0.03405657410621643,
0.01712171919643879,
-0.0792783722281456,
0.07528652250766754,
0.07760495692491531,
0.005600975826382637,
0.017280030995607376,
0.03829846531152725,
0.0827254131436348,
0.043725524097681046,
0.00010834540444193408,
0.000581576838158071,
0.0009624105296097696,
0.006553244777023792,
-0.04068460687994957,
0.030333707109093666,
0.00885780155658722,
0.012135830707848072,
0.0851229578256607,
0.06605632603168488,
-0.040462397038936615,
-0.037366971373558044,
-0.03472721204161644,
0.026090430095791817,
0.052476540207862854,
0.031619228422641754,
-0.018872419372200966,
-0.03478372469544411,
0.0020256140269339085,
0.08189170807600021,
-0.039308201521635056,
-0.08598025888204575,
2.129888867172457e-33,
0.022903019562363625,
-0.02859889157116413,
0.04962902143597603,
0.022577453404664993,
0.024692974984645844,
-0.06215030327439308,
-0.051823005080223083,
0.006685242988169193,
0.03488884121179581,
0.015002984553575516,
-0.03121846169233322,
0.07752616703510284,
0.007931321859359741,
0.054132841527462006,
0.016743868589401245,
0.03770848363637924,
-0.09888314455747604,
0.05939179286360741,
0.008442647755146027,
-0.01198151521384716,
0.03217793256044388,
0.004683326929807663,
-0.026985272765159607,
-0.02695399895310402,
-0.007972500286996365,
0.012434327974915504,
-0.022756556048989296,
0.023806599900126457,
-0.045352306216955185,
-0.004280465189367533,
-0.13819710910320282,
-0.0008941185660660267,
-0.002486284589394927,
0.024892162531614304,
0.02642321027815342,
0.026029450818896294,
-0.007798831444233656,
-0.042271923273801804,
0.0029522196855396032,
-0.01871630549430847,
0.08757971972227097,
0.03519239276647568,
0.0014634995022788644,
0.009447775781154633,
-0.07868172973394394,
-0.08826699107885361,
0.004903214052319527,
0.025757092982530594,
-0.02080691047012806,
-0.03309200331568718,
0.01265046838670969,
-0.012319521978497505,
0.0012253464665263891,
-0.13606096804141998,
-0.0872601717710495,
0.06723349541425705,
-0.021130576729774475,
-0.007990647107362747,
0.018321750685572624,
0.03255270794034004,
0.0025115215685218573,
-0.0623018778860569,
-0.013750403188169003,
-0.010694464668631554,
0.010995895601809025,
0.025825146585702896,
-0.029032710939645767,
0.0626245066523552,
0.08592493087053299,
0.005402435082942247,
-0.010766531340777874,
0.02749299444258213,
-0.052290625870227814,
-0.06673867255449295,
0.10769826173782349,
-0.004317670129239559,
0.0623803436756134,
-0.08089128136634827,
-0.10464102774858475,
0.01984589360654354,
0.007008960470557213,
-0.0362730547785759,
-0.0874449834227562,
0.0750517025589943,
-0.043614473193883896,
-0.06231294944882393,
0.098305843770504,
-0.07518832385540009,
-0.07280442863702774,
-0.0140110794454813,
-0.10127808898687363,
-0.07311233133077621,
-0.005069451406598091,
0.012381449341773987,
-0.03441758081316948,
-3.266653617871027e-33,
-0.030380601063370705,
-0.002587098628282547,
-0.006020217668265104,
0.02228757180273533,
-0.025305764749646187,
0.013582022860646248,
0.04236248880624771,
0.019957421347498894,
-0.032930318266153336,
-0.026332102715969086,
-0.050850775092840195,
-0.04041188582777977,
0.04402390867471695,
-0.014921752735972404,
-0.005296214949339628,
0.0007394223939627409,
-0.0030009520705789328,
0.015604843385517597,
0.06804922223091125,
-0.041629694402217865,
0.00911770761013031,
0.05828627943992615,
0.03263150900602341,
0.012423183768987656,
-0.025144707411527634,
0.06396449357271194,
0.004711073357611895,
0.07719571143388748,
-0.01640949584543705,
0.04090316966176033,
-0.0627029538154602,
-0.06844090670347214,
-0.015121178701519966,
-0.05464154854416847,
-0.0626600831747055,
0.10947851836681366,
0.018762798979878426,
-0.07137215882539749,
0.014802657067775726,
0.024631602689623833,
0.036454588174819946,
-0.06454558670520782,
-0.030784515663981438,
0.05437790974974632,
0.004679485224187374,
-0.03456124663352966,
-0.06653899699449539,
0.05986601859331131,
-0.0306679867208004,
0.041488129645586014,
-0.005671003833413124,
0.05986084043979645,
-0.11962908506393433,
-0.0649678036570549,
-0.04436064139008522,
-0.040198978036642075,
0.057093359529972076,
-0.04203623905777931,
0.05314478278160095,
0.012038202956318855,
-0.04069557785987854,
-0.06188196316361427,
-0.028648091480135918,
-0.020339470356702805,
-0.04450690373778343,
0.025970345363020897,
-0.049870241433382034,
-0.09382127970457077,
0.07155228406190872,
0.049365390092134476,
0.011502807959914207,
0.03396610915660858,
0.0009030292858369648,
-0.031247578561306,
0.08158264309167862,
-0.01717107743024826,
0.025243671610951424,
-0.052845340222120285,
0.0007544872350990772,
-0.024027487263083458,
0.027843700721859932,
-0.0867384597659111,
0.038068510591983795,
-0.005604498088359833,
0.009856448508799076,
0.11193718761205673,
-0.03264050930738449,
0.08151139318943024,
-0.028975144028663635,
-0.06013759970664978,
-0.0015870422357693315,
-0.007188953924924135,
-0.0009922365425154567,
0.020393889397382736,
-0.03125875070691109,
-4.331587177830443e-8,
-0.1268673688173294,
0.07919550687074661,
0.1269320696592331,
0.050430916249752045,
-0.015523131936788559,
0.00816095620393753,
0.0070140245370566845,
-0.005807993933558464,
0.0015793386846780777,
-0.038817327469587326,
0.15355972945690155,
-0.08620277047157288,
0.041210949420928955,
-0.021119598299264908,
-0.04564796760678291,
0.06447024643421173,
0.03458522632718086,
0.04921422526240349,
-0.05559907481074333,
0.05308954790234566,
0.028136275708675385,
0.05058765411376953,
0.0311095230281353,
-0.10177397727966309,
0.0668332502245903,
-0.07630670815706253,
-0.04970182850956917,
0.0765681117773056,
0.030181128531694412,
0.009446173906326294,
0.018270762637257576,
0.03484584391117096,
0.026725130155682564,
-0.052882224321365356,
-0.04861252009868622,
0.01311494130641222,
0.02582630328834057,
-0.02914152480661869,
-0.037832196801900864,
-0.03652626648545265,
-0.08888451009988785,
0.043733127415180206,
-0.050066184252500534,
-0.0026646279729902744,
-0.02840532548725605,
0.033321280032396317,
0.05267076939344406,
-0.033263105899095535,
0.06927237659692764,
-0.031538695096969604,
-0.016328835859894753,
-0.02147248201072216,
0.021526912227272987,
0.10268836468458176,
0.04253767058253288,
0.056959498673677444,
-0.029559796676039696,
-0.07355637848377228,
0.011795755475759506,
-0.01386327762156725,
0.006724923849105835,
0.0529790036380291,
-0.10136441141366959,
-0.009298731572926044
] | 0 | 0 | 0 | 0 |
|
AutoDAN-Turbo: A Lifelong Agent for Strategy Self-Exploration to Jailbreak LLMs | https://openreview.net/forum?id=bhK7U37VW8 | [
"Xiaogeng Liu",
"Peiran Li",
"G. Edward Suh",
"Yevgeniy Vorobeychik",
"Zhuoqing Mao",
"Somesh Jha",
"Patrick McDaniel",
"Huan Sun",
"Bo Li",
"Chaowei Xiao"
] | Spotlight | Jailbreak attacks serve as essential red-teaming tools, proactively assessing whether LLMs can behave responsibly and safely in adversarial environments. Despite diverse strategies (e.g., cipher, low-resource language, persuasions, and so on) that have been proposed and shown success, these strategies are still manually designed, limiting their scope and effectiveness as a red-teaming tool. In this paper, we propose AutoDAN-Turbo, a black-box jailbreak method that can automatically discover as many jailbreak strategies as possible from scratch, without any human intervention or predefined scopes (e.g., specified candidate strategies), and use them for red-teaming. As a result, AutoDAN-Turbo can significantly outperform baseline methods, achieving a 74.3% higher average attack success rate on public benchmarks. Notably, AutoDAN-Turbo achieves an 88.5 attack success rate on GPT-4-1106-turbo. In addition, AutoDAN-Turbo is a unified framework that can incorporate existing human-designed jailbreak strategies in a plug-and-play manner. By integrating human-designed strategies, AutoDAN-Turbo can even achieve a higher attack success rate of 93.4 on GPT-4-1106-turbo. | Large Language Model, Jailbreak Attack, LLM Agent | We propose a black-box jailbreak framework that can automatically discover as many jailbreak strategies as possible from scratch, without any human intervention or predefined scopes. | 11,606 | null | [
-0.05543596297502518,
-0.004315171390771866,
-0.04363085329532623,
-0.0006940593011677265,
-0.026179026812314987,
0.03729482367634773,
0.03763280063867569,
-0.05900857225060463,
-0.044390834867954254,
-0.031581200659275055,
-0.020937606692314148,
-0.06384500116109848,
0.015663882717490196,
-0.021506505087018013,
0.08642438054084778,
-0.0029526473954319954,
0.06616058945655823,
0.06904635578393936,
-0.058887649327516556,
-0.028878195211291313,
-0.020531021058559418,
-0.08135732263326645,
0.017653727903962135,
-0.0835992842912674,
0.023919710889458656,
-0.02442682534456253,
-0.07285049557685852,
-0.022827593609690666,
-0.024464471265673637,
-0.040931541472673416,
0.04465610161423683,
-0.004749069921672344,
0.005253348499536514,
0.03494524583220482,
-0.06377942115068436,
-0.01237938180565834,
-0.03545522689819336,
0.07546702772378922,
-0.014683048240840435,
-0.11693958193063736,
-0.06601791828870773,
0.03398914635181427,
0.002105849329382181,
0.04546554014086723,
0.01693958416581154,
-0.047781895846128464,
-0.04233948513865471,
0.04700840637087822,
0.015118680894374847,
-0.051062438637018204,
-0.03917643427848816,
0.001184915192425251,
0.05092494934797287,
0.0560448095202446,
-0.005092629697173834,
0.005437093321233988,
0.038062769919633865,
0.07076184451580048,
0.06753339618444443,
0.06409956514835358,
0.0013770307414233685,
0.07829057425260544,
-0.04922165721654892,
0.04106127843260765,
-0.03279852122068405,
0.017087822780013084,
0.09782605618238449,
0.05528421700000763,
0.018577327951788902,
0.00956939347088337,
-0.011084622703492641,
-0.04832446947693825,
-0.018507013097405434,
-0.02347288466989994,
-0.013000436127185822,
0.05538349226117134,
-0.06332232058048248,
-0.0376090444624424,
-0.013822540640830994,
-0.01290422584861517,
0.009007956832647324,
-0.03479984030127525,
0.029474547132849693,
0.0650569424033165,
0.06738656759262085,
0.07399225980043411,
-0.07011173665523529,
0.012759117409586906,
0.02558339759707451,
0.05823659524321556,
-0.011735656298696995,
0.08381911367177963,
0.03523522615432739,
-0.00555872218683362,
0.034619010984897614,
0.033889155834913254,
-0.031245386227965355,
-0.09474197775125504,
-0.06495184451341629,
0.071980781853199,
0.005385083612054586,
0.018950313329696655,
-0.001000847201794386,
-0.07298435270786285,
0.05018753930926323,
0.0014644380426034331,
0.050215594470500946,
-0.01583263836801052,
0.02914394624531269,
-0.007886146195232868,
-0.011376341804862022,
0.046360570937395096,
0.03469005972146988,
-0.021861132234334946,
0.012945227324962616,
0.0030301285441964865,
-0.08403411507606506,
0.09361513704061508,
-0.005346724297851324,
0.04959326237440109,
0.05205799639225006,
0.006457233335822821,
0.03555232658982277,
-0.02242688648402691,
0.09172706305980682,
0.03961057588458061,
-0.10815802961587906,
1.4430608341339374e-33,
0.01081005297601223,
0.016683295369148254,
-0.07377179712057114,
0.07388311624526978,
0.02015477977693081,
-0.07473483681678772,
0.033318277448415756,
0.07273902744054794,
-0.07013420760631561,
0.06051238253712654,
0.012474513612687588,
0.02776024304330349,
-0.028559837490320206,
0.035307589918375015,
0.10090538114309311,
0.0013130615698173642,
-0.045648831874132156,
-0.027307206764817238,
-0.022658655419945717,
-0.03722606226801872,
-0.0012980307219550014,
-0.019897278398275375,
0.04787486419081688,
-0.07804549485445023,
0.04792102426290512,
0.03880489617586136,
0.017380302771925926,
-0.01434777956455946,
0.048091188073158264,
0.0523945614695549,
-0.10943704843521118,
-0.047525908797979355,
-0.014187145978212357,
0.03738380968570709,
0.028464099392294884,
-0.030267102643847466,
-0.03957027941942215,
-0.06780242174863815,
-0.05078720673918724,
0.03932902589440346,
-0.04978862777352333,
0.0031865527853369713,
-0.05793089047074318,
-0.06240241229534149,
-0.020677443593740463,
-0.05513923615217209,
-0.055498912930488586,
0.005042674019932747,
-0.03132179006934166,
0.000488780380692333,
-0.0242621973156929,
0.02020931802690029,
-0.019206199795007706,
0.00031575700268149376,
-0.09893783181905746,
0.0018070625374093652,
-0.03510650619864464,
0.05751129239797592,
0.02193409390747547,
0.12923486530780792,
-0.014506109990179539,
-0.01814507134258747,
-0.11240135133266449,
0.02729174867272377,
0.018534133210778236,
-0.04398993030190468,
-0.0025816750712692738,
0.025522267445921898,
-0.022574959322810173,
0.03835337236523628,
-0.03599310666322708,
0.007259363308548927,
0.05473471060395241,
0.07229465246200562,
-0.027273647487163544,
-0.021923353895545006,
0.09420893341302872,
0.06113613024353981,
0.03037874586880207,
-0.1282908171415329,
0.05356321483850479,
0.019655952230095863,
0.017951473593711853,
0.04973889887332916,
-0.029697490856051445,
-0.010351445525884628,
-0.016942156478762627,
-0.1168699711561203,
-0.06746884435415268,
0.04746604710817337,
-0.05837073549628258,
-0.04100813716650009,
-0.01307515799999237,
0.005674047861248255,
-0.024964267387986183,
-2.5830107153457552e-33,
-0.03818297013640404,
-0.08123011142015457,
-0.05709478259086609,
0.02064369060099125,
0.04681863263249397,
-0.0608031190931797,
-0.008333477191627026,
0.012190002016723156,
-0.001358075882308185,
0.0008539519621990621,
-0.07494717836380005,
0.03779235854744911,
0.12463740259408951,
0.024555180221796036,
0.03827422484755516,
-0.0788111686706543,
-0.03253034129738808,
0.0030717283952981234,
-0.016281815245747566,
0.05577443540096283,
-0.028100373223423958,
0.07862462848424911,
-0.006315052974969149,
0.044509775936603546,
0.056555986404418945,
-0.022445885464549065,
-0.05004515498876572,
0.03458171337842941,
0.06362444907426834,
0.031004058197140694,
0.09673462808132172,
-0.03137071430683136,
-0.026505419984459877,
-0.044615183025598526,
-0.03922543302178383,
0.11504391580820084,
0.07161763310432434,
0.029728390276432037,
-0.05137915909290314,
0.07946392148733139,
0.11025048047304153,
-0.052301470190286636,
-0.09697160869836807,
-0.07911359518766403,
0.008727774024009705,
0.06513791531324387,
-0.08303115516901016,
0.07172854989767075,
-0.032556090503931046,
-0.07420577853918076,
0.04165327176451683,
0.024808155372738838,
-0.03729933500289917,
-0.013448390178382397,
-0.05473852902650833,
-0.08174873888492584,
0.08545532077550888,
-0.031288087368011475,
0.004684558138251305,
0.02597602643072605,
-0.02018504962325096,
-0.03435767441987991,
0.03493936359882355,
-0.007940180599689484,
0.029912704601883888,
-0.08873046934604645,
-0.007472075056284666,
0.06630205363035202,
-0.029646078124642372,
0.0062273722141981125,
0.009280131198465824,
-0.018996989354491234,
-0.049235306680202484,
0.05545984208583832,
-0.060962531715631485,
0.0008901008986867964,
-0.08855972439050674,
-0.1326129138469696,
0.012945897877216339,
-0.00871080718934536,
0.04504036530852318,
0.04619468003511429,
-0.022246740758419037,
0.022510871291160583,
0.03483721613883972,
0.03283151239156723,
0.06302618980407715,
0.11697594076395035,
0.01515248790383339,
-0.023209884762763977,
0.013032589107751846,
0.033482328057289124,
-0.03168405219912529,
0.13718760013580322,
-0.07638002932071686,
-4.4589882008949644e-8,
0.06098746135830879,
0.03662567958235741,
0.043437521904706955,
0.04813048616051674,
0.009134567342698574,
0.09071347862482071,
-0.05208699777722359,
-0.03962680697441101,
0.021737905219197273,
-0.03531929850578308,
0.09143176674842834,
-0.06124478951096535,
-0.023661652579903603,
-0.021536754444241524,
-0.05256929621100426,
0.06224675104022026,
-0.07239572703838348,
-0.005902606062591076,
-0.027097200974822044,
-0.027739686891436577,
-0.0278532225638628,
0.0330195389688015,
-0.031269095838069916,
-0.043997567147016525,
0.025393439456820488,
-0.01966932788491249,
-0.043230120092630386,
0.05298759415745735,
0.06456916779279709,
0.08269812911748886,
-0.04676825553178787,
-0.036972422152757645,
0.008407607674598694,
-0.025961514562368393,
-0.029972707852721214,
0.11433917284011841,
-0.03393822908401489,
-0.043242067098617554,
0.07825218886137009,
-0.01607508212327957,
-0.04250671714544296,
0.03022165037691593,
-0.037225283682346344,
0.007068314589560032,
-0.0598498173058033,
-0.010694019496440887,
-0.04978325217962265,
-0.12085968255996704,
0.07552968710660934,
-0.02998208999633789,
0.07587156444787979,
-0.08104898035526276,
-0.014848651364445686,
0.059237040579319,
0.09670984745025635,
0.07146181911230087,
-0.014686954207718372,
-0.041084446012973785,
0.10122932493686676,
0.02913816273212433,
0.021141668781638145,
-0.017192278057336807,
-0.029962003231048584,
0.0028548764530569315
] | 0 | 0 | 0 | 0 |
|
Towards Automated Knowledge Integration From Human-Interpretable Representations | https://openreview.net/forum?id=NTHMw8S1Ow | [
"Kasia Kobalczyk",
"Mihaela van der Schaar"
] | Spotlight | A significant challenge in machine learning, particularly in noisy and low-data environments, lies in effectively incorporating inductive biases to enhance data efficiency and robustness. Despite the success of informed machine learning methods, designing algorithms with explicit inductive biases remains largely a manual process. In this work, we explore how prior knowledge represented in its native formats, e.g. in natural language, can be integrated into machine learning models in an automated manner. Inspired by the learning to learn principles of meta-learning, we consider the approach of learning to integrate knowledge via conditional meta-learning, a paradigm we refer to as informed meta-learning. We introduce and motivate theoretically the principles of informed meta-learning enabling automated and controllable inductive bias selection. To illustrate our claims, we implement an instantiation of informed meta-learning--the Informed Neural Process, and empirically demonstrate the potential benefits and limitations of informed meta-learning in improving data efficiency and generalisation. | informed machine learning, knowledge integration, meta-learning, data efficiency, priors | We propose a new perspective on meta-learning as a paradigm enabling automated and controllable inductive bias specification, establishing a bridge between human-interpretable representations of knowledge and the hypothesis space of ML models. | 11,605 | 2402.16105 | [
-0.052816033363342285,
-0.02411128208041191,
-0.004670682363212109,
0.08730673789978027,
0.09088659286499023,
-0.01870087906718254,
0.0029884078539907932,
0.0011633102549239993,
-0.02044074609875679,
0.010548380203545094,
-0.11681447178125381,
-0.0034491457045078278,
-0.006749481428414583,
0.05606257542967796,
0.0028711718041449785,
0.03203597292304039,
0.0706268921494484,
0.05311998352408409,
-0.1018730029463768,
-0.0907844677567482,
0.013481046073138714,
0.038520943373441696,
0.018596667796373367,
0.06608660519123077,
0.03534308820962906,
-0.0403827503323555,
0.014482302591204643,
0.020426297560334206,
0.004784557037055492,
0.0045108371414244175,
0.05103324353694916,
-0.04594636335968971,
0.010146801359951496,
-0.040352366864681244,
-0.06591235846281052,
0.016792703419923782,
-0.026375388726592064,
0.10331235826015472,
0.002858469495549798,
-0.0331222303211689,
0.019372519105672836,
-0.05907406285405159,
-0.006310272496193647,
0.008970879949629307,
0.09233608096837997,
0.024077026173472404,
-0.0569177083671093,
-0.06253428012132645,
-0.018155768513679504,
0.0705723762512207,
-0.11948610842227936,
-0.016585486009716988,
-0.029603060334920883,
0.09830670058727264,
-0.013095316477119923,
0.05207258090376854,
-0.04541249945759773,
-0.04185467213392258,
-0.052294425666332245,
-0.02093854732811451,
0.00628445390611887,
-0.1418297439813614,
-0.04935584217309952,
0.04853055998682976,
-0.042578279972076416,
0.021984102204442024,
-0.013572655618190765,
0.05332032963633537,
-0.046763647347688675,
-0.03900102898478508,
-0.01397764217108488,
0.11799484491348267,
-0.05733230710029602,
0.07126527279615402,
0.026145396754145622,
0.08592944592237473,
0.03033958189189434,
-0.0425635427236557,
0.05985277146100998,
-0.027462951838970184,
-0.03388078883290291,
0.10965283960103989,
0.032608456909656525,
0.0070076435804367065,
0.10188528895378113,
-0.0383424274623394,
0.06861688941717148,
0.037416767328977585,
0.013177549466490746,
0.056538622826337814,
-0.0028207660652697086,
-0.04057535156607628,
0.01117087434977293,
0.03402536362409592,
0.01000079046934843,
0.0631629228591919,
-0.012228020466864109,
-0.12070859223604202,
-0.03863450512290001,
0.1052422821521759,
-0.026446491479873657,
0.061878953129053116,
-0.034786976873874664,
-0.06047854945063591,
-0.0521511547267437,
-0.0541522242128849,
0.04030179604887962,
0.041331660002470016,
0.01324223168194294,
-0.060122303664684296,
0.03401267156004906,
-0.01071664784103632,
-0.01461802888661623,
-0.05048806220293045,
0.0015185694210231304,
0.01711866445839405,
0.021523354575037956,
0.016481688246130943,
0.05874067172408104,
0.015563183464109898,
-0.0002765917161013931,
0.044019464403390884,
-0.007706940174102783,
0.08045616000890732,
0.057007260620594025,
-0.0928875282406807,
-0.06913265585899353,
-1.4359487259694779e-33,
0.031626053154468536,
0.006518266629427671,
0.10526078194379807,
0.022333519533276558,
0.03935462236404419,
-0.07861308008432388,
-0.08003649860620499,
0.00836936105042696,
-0.010962609201669693,
-0.01326385885477066,
0.06462652236223221,
0.03450610861182213,
-0.04288327321410179,
-0.0017292981501668692,
-0.027846336364746094,
0.04615635797381401,
-0.05601983889937401,
0.1013345941901207,
-0.034936100244522095,
-0.08472590148448944,
0.05330190435051918,
-0.03181377053260803,
0.0036645270884037018,
-0.007918265648186207,
-0.025565925985574722,
0.025552116334438324,
0.030215825885534286,
-0.01467506866902113,
-0.009000824764370918,
0.04242628812789917,
-0.08532244712114334,
-0.06705642491579056,
0.07394390553236008,
0.003065527183935046,
-0.04477451741695404,
-0.04333927854895592,
-0.05170832574367523,
0.0020806118845939636,
0.020898958668112755,
0.04921386390924454,
0.020890261977910995,
0.016558995470404625,
0.05553215742111206,
-0.01338393148034811,
-0.06965455412864685,
-0.0847259908914566,
-0.05515434965491295,
0.03402130678296089,
-0.0007593939080834389,
-0.05014890804886818,
0.07073833048343658,
-0.005827928427606821,
-0.052136246114969254,
-0.04813332483172417,
0.006526427809149027,
0.073637455701828,
-0.020464187487959862,
0.02629614993929863,
-0.013132582418620586,
-0.018419379368424416,
0.027356166392564774,
0.08152958750724792,
-0.004654388874769211,
0.08490633219480515,
-0.07226332277059555,
-0.011588526889681816,
0.014590144157409668,
0.04760123789310455,
0.0199127160012722,
-0.03517227619886398,
-0.06114007905125618,
0.06299614906311035,
-0.049041394144296646,
-0.02248798869550228,
0.0027615593280643225,
0.023578884080052376,
0.06061887741088867,
-0.10803469270467758,
0.007622049655765295,
0.03637662157416344,
-0.014509285800158978,
-0.026861965656280518,
-0.0575406514108181,
0.014161976054310799,
-0.048320140689611435,
-0.03214353322982788,
0.009290274232625961,
-0.022480271756649017,
-0.00607877504080534,
-0.03396335616707802,
-0.045427337288856506,
0.0031795972026884556,
0.004494645167142153,
0.010607628151774406,
-0.0008128051995299757,
-1.9778631470533203e-34,
-0.030878186225891113,
0.023411458358168602,
0.04658619686961174,
0.03190910443663597,
-0.034270793199539185,
0.05699953809380531,
-0.037689488381147385,
0.011466756463050842,
0.034478560090065,
-0.03611597791314125,
-0.06352853029966354,
-0.0578327514231205,
-0.0074112387374043465,
-0.015994912013411522,
-0.009230133146047592,
-0.0034861566964536905,
0.0087888203561306,
-0.006467063911259174,
-0.013357401825487614,
0.07420352101325989,
0.04404968395829201,
0.08481919020414352,
-0.11107246577739716,
0.045917682349681854,
0.013271521776914597,
0.009295480325818062,
-0.01987932249903679,
0.1050441637635231,
0.022882819175720215,
0.04724596440792084,
-0.08658511936664581,
0.008181361481547356,
-0.043901264667510986,
-0.13044992089271545,
-0.10686871409416199,
0.040165744721889496,
0.044975169003009796,
-0.08954499661922455,
-0.06601861119270325,
0.12547548115253448,
0.07281505316495895,
-0.004175981506705284,
-0.09556131809949875,
-0.05252304673194885,
-0.02702413685619831,
-0.04194325953722,
-0.013642054051160812,
0.006380417384207249,
0.04317929968237877,
0.026619987562298775,
0.022088132798671722,
0.02022610977292061,
-0.09911276400089264,
-0.00043695635395124555,
-0.0678810104727745,
-0.061447929590940475,
0.015509947203099728,
-0.0016388223739340901,
0.0933622345328331,
0.028986357152462006,
-0.022525127977132797,
-0.040442775934934616,
-0.002504906849935651,
-0.03300154209136963,
-0.027208203449845314,
-0.06454889476299286,
-0.025409623980522156,
0.016693739220499992,
0.009220619685947895,
0.009862200357019901,
0.08130483329296112,
0.015196426771581173,
-0.049449365586042404,
-0.07284954935312271,
0.03367089480161667,
0.010701196268200874,
0.041361793875694275,
-0.045482464134693146,
-0.05935591831803322,
0.0017526651499792933,
0.09752684086561203,
0.03106614761054516,
0.08451469987630844,
-0.055737536400556564,
0.06913378834724426,
0.07185398042201996,
0.012848100624978542,
0.08233805745840073,
0.020569005981087685,
0.013029470108449459,
-0.06925506889820099,
0.01566404290497303,
-0.04163798689842224,
0.06987566500902176,
-0.16531457006931305,
-3.538217896448259e-8,
-0.03941468521952629,
0.004830862395465374,
-0.0019253649516031146,
0.09929649531841278,
0.02740497887134552,
-0.001617496251128614,
-0.05182095989584923,
0.07878506928682327,
-0.008346443995833397,
-0.09603361040353775,
0.023934951052069664,
0.026660894975066185,
0.01875501684844494,
-0.03042539767920971,
-0.007627721410244703,
0.07107876986265182,
-0.02678769640624523,
0.009158045053482056,
-0.018139943480491638,
-0.02101726271212101,
0.05664270743727684,
0.04541397839784622,
-0.010125063359737396,
-0.067062608897686,
0.08312442898750305,
-0.07877758145332336,
-0.044655006378889084,
0.08098779618740082,
-0.004240795504301786,
0.05978946387767792,
-0.04041852802038193,
0.040517110377550125,
0.0032039200887084007,
-0.007971263490617275,
0.0753210112452507,
0.06286949664354324,
0.05799911916255951,
-0.030201271176338196,
-0.05614161863923073,
-0.09082615375518799,
0.015871888026595116,
0.09715458750724792,
-0.04481586813926697,
-0.0054581924341619015,
-0.03523414954543114,
-0.04225289076566696,
0.02786429226398468,
0.01904657483100891,
-0.07093881815671921,
-0.03781575709581375,
0.01663113757967949,
-0.061120111495256424,
0.06340797245502472,
0.040054816752672195,
0.0749758705496788,
0.016313834115862846,
0.020626898854970932,
-0.029762674123048782,
0.0009692033054307103,
0.014138402417302132,
0.049982257187366486,
0.07244884222745895,
0.009891797788441181,
0.01415463350713253
] | 0 | 0 | 0 | 0 |
|
BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval | https://openreview.net/forum?id=ykuc5q381b | [
"Hongjin SU",
"Howard Yen",
"Mengzhou Xia",
"Weijia Shi",
"Niklas Muennighoff",
"Han-yu Wang",
"Liu Haisu",
"Quan Shi",
"Zachary S Siegel",
"Michael Tang",
"Ruoxi Sun",
"Jinsung Yoon",
"Sercan O Arik",
"Danqi Chen",
"Tao Yu"
] | Spotlight | Existing retrieval benchmarks primarily consist of information-seeking queries (e.g., aggregated questions from search engines) where keyword or semantic-based retrieval is usually sufficient. However, many complex real-world queries require in-depth reasoning to identify relevant documents that go beyond surface form matching. For example, finding documentation for a coding question requires understanding the logic and syntax of the functions involved. To better benchmark retrieval on such challenging queries, we introduce BRIGHT, the first text retrieval benchmark that requires intensive reasoning to retrieve relevant documents. Our dataset consists of 1,398 real-world queries spanning diverse domains such as economics, psychology, mathematics, coding, and more. These queries are drawn from naturally occurring or carefully curated human data. Extensive evaluation reveals that even state-of-the-art retrieval models perform poorly on BRIGHT. The leading model on the MTEB leaderboard (Muennighoff et al., 2023), which achieves a score of 59.0 nDCG@10,1 produces a score of nDCG@10 of 18.0 on BRIGHT. We show that incorporating explicit reasoning about the query improves retrieval performance by up to 12.2 points. Moreover, incorporating retrieved documents from the top-performing retriever boosts question answering performance by over 6.6 points. We believe that BRIGHT paves the way for future research on retrieval systems in more realistic and challenging settings. | Retrieval benchmark, Reasoning | A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval | 11,583 | 2407.12883 | [
-0.06382066011428833,
-0.03852856531739235,
-0.0036612648982554674,
0.10545632243156433,
0.05599987134337425,
-0.022861311212182045,
0.021604059264063835,
0.08188391476869583,
0.010523500852286816,
-0.04197398200631142,
-0.057445354759693146,
-0.011757188476622105,
0.040008194744586945,
0.020262207835912704,
0.031199336051940918,
0.039993345737457275,
0.08494723588228226,
-0.016202982515096664,
-0.04657802730798721,
-0.03342819586396217,
0.05625540018081665,
0.04453302174806595,
0.04888864606618881,
-0.07753194123506546,
0.0064541045576334,
-0.010335121303796768,
0.018538961187005043,
-0.05013087019324303,
0.008841135539114475,
0.02134985849261284,
0.04788156598806381,
0.06228749081492424,
0.001863472512923181,
0.08408180624246597,
0.011351937428116798,
0.0017992558423429728,
-0.06634698808193207,
-0.0031957365572452545,
0.030194351449608803,
-0.016167834401130676,
-0.023504668846726418,
-0.04956824705004692,
0.024058876559138298,
0.05205417424440384,
0.05983788147568703,
0.028881484642624855,
-0.05567033588886261,
-0.018196742981672287,
0.02090022899210453,
-0.01094459556043148,
-0.13337352871894836,
0.010010768659412861,
0.004543858580291271,
0.0033146438654512167,
-0.011352348141372204,
0.06834262609481812,
-0.061990488320589066,
-0.03449781984090805,
-0.024133270606398582,
-0.12411480396986008,
-0.026267575100064278,
-0.11457566171884537,
0.03307527303695679,
-0.007401529233902693,
0.019191311672329903,
-0.04563959315419197,
0.016213547438383102,
0.014446808025240898,
0.020131302997469902,
-0.028721047565340996,
0.019134098663926125,
0.05529708042740822,
-0.05417850613594055,
0.04926192760467529,
0.031150758266448975,
0.06450244784355164,
0.023652654141187668,
-0.059606172144412994,
0.01645897515118122,
-0.035395074635744095,
-0.03447804972529411,
-0.01880207471549511,
0.014591099694371223,
0.06864152103662491,
0.12932467460632324,
-0.029048291966319084,
0.07775133103132248,
-0.014925277791917324,
0.018254274502396584,
0.009934872388839722,
0.00149889150634408,
-0.08525655418634415,
0.060664623975753784,
0.007367672864347696,
0.02033589966595173,
0.060827989131212234,
0.021618982776999474,
-0.04969421774148941,
-0.010072373785078526,
0.08308238536119461,
0.025418957695364952,
0.08470018208026886,
-0.02218650095164776,
-0.06357802450656891,
-0.03389866277575493,
0.037873998284339905,
0.05006290227174759,
0.11018926650285721,
0.00787199568003416,
-0.12384701520204544,
0.020458484068512917,
0.028634270653128624,
0.017572641372680664,
-0.00874353013932705,
-0.0017587305046617985,
-0.008556566201150417,
-0.012987480498850346,
-0.019895104691386223,
0.021121196448802948,
0.009232969023287296,
0.016607236117124557,
0.06750690937042236,
0.005937994923442602,
-0.036042992025613785,
-0.008841597475111485,
0.007126194424927235,
-0.018069511279463768,
2.213978222900514e-33,
0.055324602872133255,
0.029584283009171486,
0.0036872855853289366,
0.007296606432646513,
-0.0024507828056812286,
-0.01759672537446022,
0.032680902630090714,
0.09479863196611404,
-0.11066621541976929,
0.04225846007466316,
-0.09590831398963928,
0.09071438014507294,
-0.03546086698770523,
0.040946345776319504,
-0.032880328595638275,
-0.01392864529043436,
-0.07246812433004379,
0.022852541878819466,
-0.06393083930015564,
-0.008275829255580902,
0.06666246056556702,
-0.043650124222040176,
0.03562760353088379,
-0.012329944409430027,
-0.0017352112336084247,
0.02583560161292553,
-0.009632175788283348,
-0.03702147677540779,
0.011495147831737995,
-0.01592259109020233,
-0.0719185397028923,
-0.06987489014863968,
-0.03423381596803665,
-0.00037094863364472985,
0.0007816375000402331,
-0.027967052534222603,
-0.007115270476788282,
-0.06907187402248383,
0.0008781468495726585,
0.021863218396902084,
-0.10091809928417206,
0.062415167689323425,
0.04714947193861008,
-0.007700652349740267,
-0.047775719314813614,
0.0021668728440999985,
-0.08124740421772003,
0.031219959259033203,
0.017502518370747566,
-0.0022754163946956396,
0.015132205560803413,
0.03030119091272354,
0.019142521545290947,
-0.03909030184149742,
0.0179635901004076,
-0.06964119523763657,
-0.019949747249484062,
0.05865834653377533,
0.08635229617357254,
0.07800761610269547,
0.011428353376686573,
0.03999463841319084,
0.024957716464996338,
0.01900249719619751,
-0.061549920588731766,
0.06020190566778183,
0.01903342269361019,
0.06298670917749405,
0.06554435193538666,
0.017741592600941658,
0.029130473732948303,
0.026480551809072495,
0.08253634721040726,
-0.08647481352090836,
0.08713120222091675,
0.05753534287214279,
0.039875637739896774,
-0.09621540457010269,
-0.006109482608735561,
-0.04470934718847275,
0.01637606881558895,
-0.03657521679997444,
-0.012891178019344807,
-0.017919549718499184,
-0.054145392030477524,
0.0004756590642500669,
0.04448244720697403,
-0.13246439397335052,
0.0003117225132882595,
-0.026132119819521904,
-0.031869836151599884,
-0.000692421686835587,
-0.061533939093351364,
-0.11914554983377457,
-0.01010309997946024,
-8.68279636918852e-34,
0.011180901899933815,
-0.12924425303936005,
-0.016941623762249947,
0.0069632865488529205,
0.014166049659252167,
-0.04687204584479332,
-0.00913031492382288,
-0.05266641825437546,
-0.04549678415060043,
-0.038796503096818924,
-0.00975496880710125,
-0.046140387654304504,
0.04997457191348076,
-0.021310728043317795,
0.005438990890979767,
0.02322986349463463,
-0.04318820685148239,
-0.04246198385953903,
-0.10003439337015152,
0.06970041990280151,
0.0006030090735293925,
0.10882996022701263,
-0.10505139082670212,
0.023575931787490845,
-0.05104626342654228,
0.08310610800981522,
-0.03451135754585266,
-0.00459857564419508,
-0.056311529129743576,
0.019723063334822655,
-0.04092446714639664,
-0.045391932129859924,
-0.0568251758813858,
-0.04561055451631546,
-0.0909436047077179,
0.04184000566601753,
0.02023094706237316,
-0.04171059653162956,
-0.09869880974292755,
0.08409014344215393,
0.012903116643428802,
-0.031910039484500885,
0.004723102319985628,
-0.0023291679099202156,
-0.01603487879037857,
0.08328762650489807,
-0.1215931698679924,
-0.03692232444882393,
-0.021324357017874718,
0.025563424453139305,
0.05772276222705841,
-0.014710228890180588,
-0.026546206325292587,
0.02082023024559021,
0.01597204990684986,
-0.045092031359672546,
-0.12182185798883438,
-0.018301980569958687,
-0.03270523250102997,
0.04321752488613129,
-0.02837214060127735,
0.057486891746520996,
-0.013161584734916687,
0.07845360040664673,
-0.006056959740817547,
-0.07998422533273697,
-0.010792211629450321,
-0.03454136475920677,
-0.08111298829317093,
-0.015727272257208824,
0.02367374300956726,
-0.0898946225643158,
0.03671013563871384,
-0.014059450477361679,
0.05832146480679512,
0.12957483530044556,
0.05830882489681244,
0.054269399493932724,
-0.004241389222443104,
0.01356093306094408,
-0.010170198045670986,
-0.029720528051257133,
0.05543315038084984,
0.08185014873743057,
0.01408701203763485,
0.027301475405693054,
0.02493985928595066,
0.03282104432582855,
0.01993677020072937,
0.007611728738993406,
-0.05284159258008003,
-0.07212074846029282,
-0.023896703496575356,
0.019965047016739845,
-0.002135043265298009,
-4.287282351356225e-8,
-0.06983136385679245,
0.1163453757762909,
-0.03546120971441269,
0.054437052458524704,
0.030610162764787674,
-0.01936110109090805,
0.010227886028587818,
0.11576662957668304,
-0.07213205099105835,
0.02794097550213337,
0.0733565241098404,
-0.06658601760864258,
-0.03506355732679367,
0.022801844403147697,
0.040172845125198364,
0.11888275295495987,
0.01747618243098259,
-0.038706351071596146,
-0.031836699694395065,
-0.03815590962767601,
0.04703434929251671,
0.08200563490390778,
-0.004344909451901913,
0.03343972563743591,
0.06864021718502045,
0.02599271573126316,
-0.047329217195510864,
0.12604567408561707,
0.06430251896381378,
-0.012599486857652664,
0.0076695396564900875,
0.004983431659638882,
0.043118249624967575,
-0.051106926053762436,
0.11607471108436584,
0.015950756147503853,
-0.024562032893300056,
-0.011670056730508804,
-0.03792911395430565,
0.02344333566725254,
0.010287572629749775,
0.08584225177764893,
0.003769440110772848,
0.04946388676762581,
0.08094146847724915,
-0.025934649631381035,
-0.02845861203968525,
0.07565034925937653,
0.04306161403656006,
0.00821565743535757,
-0.0494045689702034,
-0.14346644282341003,
-0.021576223894953728,
0.03321540355682373,
0.0019198766676709056,
-0.00017146325262729079,
-0.06099619343876839,
0.04116387665271759,
0.006243343930691481,
-0.04751119390130043,
0.07448630034923553,
0.004878136329352856,
-0.0667278990149498,
0.04831152409315109
] | 0 | 0 | 1 | 0 |
|
Reinforcement Learning for Control of Non-Markovian Cellular Population Dynamics | https://openreview.net/forum?id=dsHpulHpOK | [
"Josiah C Kratz",
"Jacob Adamczyk"
] | Spotlight | Many organisms and cell types, from bacteria to cancer cells, exhibit a remarkable ability to adapt to fluctuating environments. Additionally, cells can leverage memory of past environments to better survive previously-encountered stressors. From a control perspective, this adaptability poses significant challenges in driving cell populations toward extinction, and is thus an open question with great clinical significance. In this work, we focus on drug dosing in cell populations exhibiting phenotypic plasticity. For specific dynamical models switching between resistant and susceptible states, exact solutions are known. However, when the underlying system parameters are unknown, and for complex memory-based systems, obtaining the optimal solution is currently intractable. To address this challenge, we apply reinforcement learning (RL) to identify informed dosing strategies to control cell populations evolving under novel non-Markovian dynamics. We find that model-free deep RL is able to recover exact solutions and control cell populations even in the presence of long-range temporal dynamics. To further test our approach in more realistic settings, we demonstrate performant RL-based control strategies in environments with dynamic memory strength. | optimal drug dosing, fractional differential equations, reinforcement learning, control theory | We train an RL agent to find performant drug dosing strategies in novel non-Markovian models relevant for cancer and bacterial systems. | 11,580 | 2410.08439 | [
-0.030192410573363304,
-0.023083191365003586,
-0.00539727509021759,
0.020620761439204216,
-0.03147488087415695,
0.056882940232753754,
0.0025278888642787933,
0.010050872340798378,
0.05350642278790474,
-0.007962903007864952,
-0.027476489543914795,
-0.04747757688164711,
0.016535453498363495,
0.062318507581949234,
-0.014696129597723484,
0.060396306216716766,
0.018172848969697952,
0.04455737769603729,
-0.04599912464618683,
0.01174191664904356,
0.03200610727071762,
0.07808030396699905,
0.06267648935317993,
0.0811711847782135,
-0.03408825024962425,
0.04038603603839874,
0.012772141955792904,
-0.006678776349872351,
-0.0300248172134161,
-0.04810026288032532,
0.09598816931247711,
-0.0584678091108799,
-0.03388874605298042,
-0.07401621341705322,
-0.027649395167827606,
0.035467226058244705,
-0.12762929499149323,
0.026365529745817184,
0.0016206419095396996,
-0.025593984872102737,
0.02235162816941738,
0.0302444938570261,
-0.008781344629824162,
0.043961044400930405,
0.040391914546489716,
-0.04075339064002037,
-0.03292357921600342,
-0.0225879717618227,
0.0770537331700325,
-0.032107047736644745,
-0.019573794677853584,
0.010134408250451088,
-0.0012233804445713758,
0.07453430444002151,
0.07521382719278336,
0.04911085590720177,
0.0012193139409646392,
0.03145793080329895,
-0.060886938124895096,
-0.0887303352355957,
-0.015102501958608627,
-0.09969717264175415,
0.052593495696783066,
0.0007343576871789992,
-0.0920519307255745,
0.031159264966845512,
-0.01831694133579731,
0.08958376199007034,
0.019199466332793236,
0.02042275480926037,
-0.04595566913485527,
0.06184712424874306,
0.00869553443044424,
0.01963757537305355,
0.05167385935783386,
-0.0351150780916214,
-0.0034722969867289066,
-0.0054115415550768375,
0.029529865831136703,
0.024444779381155968,
0.056453436613082886,
-0.05590223893523216,
0.03557353839278221,
0.023807944729924202,
0.02858639508485794,
-0.03277035802602768,
0.04664504528045654,
0.09255654364824295,
0.13450811803340912,
-0.03352828696370125,
0.04251615330576897,
0.09488160908222198,
0.03659205883741379,
-0.021883612498641014,
-0.05625615641474724,
-0.027899479493498802,
0.008762035518884659,
-0.10184421390295029,
-0.0064949579536914825,
0.04307262971997261,
-0.011847812682390213,
-0.008226314559578896,
0.001724403933621943,
0.06480240821838379,
0.04389984905719757,
-0.07932376861572266,
0.0016223269049078226,
0.01377071812748909,
0.0002674610586836934,
0.03503744304180145,
0.0995163694024086,
0.01783907599747181,
0.053026117384433746,
0.14034989476203918,
-0.0026716331485658884,
-0.013169263489544392,
0.025230402126908302,
-0.10791110247373581,
-0.0008615561528131366,
0.07807435840368271,
-0.029665183275938034,
-0.1357969343662262,
-0.04923761263489723,
-0.004966752137988806,
-0.036234255880117416,
0.034965142607688904,
-0.005381906870752573,
-9.745127348225084e-34,
0.031001029536128044,
-0.052345987409353256,
0.03698112443089485,
0.030987877398729324,
0.025341201573610306,
-0.0699077695608139,
0.01763751171529293,
-0.0393843874335289,
-0.038373515009880066,
-0.040082309395074844,
-0.056421149522066116,
-0.03443425893783569,
-0.10362263768911362,
0.05102560296654701,
0.04104040563106537,
-0.027370635420084,
-0.09018837660551071,
0.02867013029754162,
-0.0064483340829610825,
-0.11691044270992279,
-0.013802108354866505,
-0.025903647765517235,
-0.04759828746318817,
-0.07305362820625305,
-0.0024210112169384956,
0.007032441440969706,
0.021852189674973488,
0.052524495869874954,
0.031337324529886246,
-0.016957374289631844,
-0.0078045520931482315,
0.00295982020907104,
-0.01544140838086605,
0.04738359898328781,
0.05719459429383278,
0.09716425091028214,
-0.052041295915842056,
-0.009722613729536533,
0.04712335392832756,
-0.03501133620738983,
-0.00911899097263813,
0.0379117950797081,
0.061663489788770676,
0.031945858150720596,
-0.049607403576374054,
-0.09452883899211884,
0.0293085016310215,
-0.026158712804317474,
-0.14257784187793732,
-0.011203227564692497,
0.023976273834705353,
-0.014318025670945644,
-0.0056415158323943615,
-0.047054748982191086,
-0.04251081869006157,
-0.010252868756651878,
-0.05671155825257301,
0.0424942821264267,
-0.04981115460395813,
0.055308062583208084,
0.057537663727998734,
-0.015437998808920383,
-0.022795064374804497,
0.08223720639944077,
0.01909048669040203,
0.0940588191151619,
-0.07257223129272461,
0.020094525068998337,
0.04476698115468025,
0.012295862659811974,
0.037958916276693344,
-0.021550815552473068,
0.06238937750458717,
-0.0786295160651207,
0.022561345249414444,
0.0021099138539284468,
0.09190346300601959,
-0.026291362941265106,
-0.08744325488805771,
0.04526544734835625,
0.0607302188873291,
-0.10415513068437576,
-0.14610974490642548,
0.08600390702486038,
-0.019276553764939308,
-0.07993021607398987,
0.03923289105296135,
-0.07432937622070312,
-0.04349369555711746,
-0.02251330390572548,
0.03648143261671066,
-0.0936419814825058,
-0.011438606306910515,
0.03710196539759636,
-0.009781848639249802,
-4.920866849204241e-34,
0.062390927225351334,
-0.061422619968652725,
0.0352155901491642,
-0.04005352780222893,
0.03647557646036148,
0.03858191892504692,
-0.059975069016218185,
-0.027212241664528847,
0.010933959856629372,
-0.09496983140707016,
-0.030602259561419487,
0.09641068428754807,
-0.007151917088776827,
0.07240432500839233,
-0.062461357563734055,
-0.0943031832575798,
0.003136491635814309,
-0.002673933282494545,
-0.008195645175874233,
-0.02850322239100933,
-0.05525442212820053,
0.03023301623761654,
-0.05804731324315071,
0.03140099346637726,
-0.04506184160709381,
0.05439087003469467,
-0.038341350853443146,
0.0815286934375763,
-0.0007756304694339633,
-0.007604236714541912,
-0.1124931052327156,
-0.031937696039676666,
-0.05682060122489929,
0.038254816085100174,
-0.060530681163072586,
0.08600883930921555,
0.05021738260984421,
0.004510974977165461,
0.003981501795351505,
0.047857616096735,
-0.0016151522286236286,
-0.060205962508916855,
-0.09627805650234222,
0.05529449135065079,
0.06722729653120041,
0.020228924229741096,
0.03310742601752281,
0.01731390319764614,
0.032211340963840485,
0.025729160755872726,
-0.04731607064604759,
0.03339112922549248,
-0.05243632569909096,
-0.022155411541461945,
-0.035996951162815094,
-0.05321873724460602,
-0.031029507517814636,
-0.0294320210814476,
0.0370614230632782,
0.003955068998038769,
-0.11629214882850647,
-0.07542594522237778,
-0.04798461124300957,
-0.000863735331222415,
-0.10009772330522537,
0.048502109944820404,
0.019207458943128586,
0.017452960833907127,
0.06144392490386963,
0.0025108240079134703,
0.05839300528168678,
0.09430259466171265,
0.07927727699279785,
-0.06441677361726761,
0.05214259400963783,
-0.0649980753660202,
-0.10764090716838837,
-0.05310014262795448,
-0.00018479042046237737,
-0.07521568983793259,
0.011542471125721931,
0.000009108986887440551,
-0.012046407908201218,
0.016163352876901627,
-0.006789145991206169,
0.023844165727496147,
-0.013056217692792416,
-0.01097500417381525,
-0.015336072072386742,
-0.03453084081411362,
-0.013856252655386925,
-0.03943095728754997,
0.0014217038406059146,
0.0002185203047702089,
-0.031065909191966057,
-3.7163950139529334e-8,
0.0510961152613163,
0.06686847656965256,
0.08991905301809311,
0.016813119873404503,
0.09510903060436249,
0.04542388767004013,
-0.015456325374543667,
-0.06167103722691536,
0.07399947196245193,
0.0153106190264225,
0.07068109512329102,
0.059358373284339905,
0.041578080505132675,
-0.058570243418216705,
-0.08431178331375122,
0.12088266760110855,
0.00409220764413476,
-0.006499013863503933,
0.011235657148063183,
-0.0019296027021482587,
-0.05683257058262825,
-0.009915720671415329,
0.017678193747997284,
0.002842018147930503,
0.06901222467422485,
-0.005891203880310059,
0.00485588563606143,
-0.01345808245241642,
0.08994000405073166,
0.031900908797979355,
0.05266105756163597,
-0.023769807070493698,
0.043344900012016296,
0.0631154328584671,
-0.05257424712181091,
0.04467217996716499,
-0.0072752260603010654,
-0.04033108800649643,
-0.017240839079022408,
-0.019854243844747543,
-0.0167916901409626,
0.08234947174787521,
-0.08451154828071594,
0.007946583442389965,
-0.006572911515831947,
0.050266969949007034,
-0.0406012162566185,
-0.05203066021203995,
0.03222356364130974,
-0.06339848786592484,
0.02816280908882618,
0.001350719714537263,
-0.0025889999233186245,
0.06147409602999687,
0.0001232339855050668,
0.021571816876530647,
0.006702393759042025,
-0.00979697797447443,
0.06652922928333282,
0.05614910647273064,
0.005056519526988268,
0.00591965951025486,
-0.02529347501695156,
0.002157674403861165
] | https://github.com/JacobHA/RL4Dosing | 1 | 0 | 0 | 0 |
Online Reinforcement Learning in Non-Stationary Context-Driven Environments | https://openreview.net/forum?id=l6QnSQizmN | [
"Pouya Hamadanian",
"Arash Nasr-Esfahany",
"Malte Schwarzkopf",
"Siddhartha Sen",
"Mohammad Alizadeh"
] | Spotlight | We study online reinforcement learning (RL) in non-stationary environments, where a time-varying exogenous context process affects the environment dynamics. Online RL is challenging in such environments due to "catastrophic forgetting" (CF). The agent tends to forget prior knowledge as it trains on new experiences. Prior approaches to mitigate this issue assume task labels (which are often not available in practice), employ brittle regularization heuristics or use off-policy methods that suffer from instability and poor performance.
We present Locally Constrained Policy Optimization (LCPO), an online RL approach that combats CF by anchoring policy outputs on old experiences while optimizing the return on current experiences. To perform this anchoring, LCPO locally constrains policy optimization using samples from experiences that lie outside of the current context distribution. We evaluate LCPO in Mujoco, classic control and computer systems environments with a variety of synthetic and real context traces, and find that it outperforms a variety of baselines in the non-stationary setting, while achieving results on-par with a "prescient" agent trained offline across all context traces.
LCPO's source code is available at https://github.com/pouyahmdn/LCPO. | catastrophic forgetting, reinforcement learning, context-driven MDP, online learning, non-stationary | A reinforcement learning algorithm that solves catastrophic forgetting in non-stationary exogenous context-driven environments by constraining the policy optimization on out of distribution samples. | 11,502 | 2302.02182 | [
-0.02482660859823227,
-0.053388409316539764,
-0.004581865854561329,
0.07349028438329697,
0.05172885209321976,
0.012314239516854286,
0.011492365039885044,
0.006524972151964903,
-0.048423271626234055,
-0.015669330954551697,
-0.04717816039919853,
0.01652868092060089,
0.07756400853395462,
0.005073121748864651,
0.012195545248687267,
-0.004438130650669336,
0.09387978911399841,
0.05178084969520569,
-0.00550105283036828,
-0.03077871911227703,
-0.006930985022336245,
0.022008225321769714,
0.055178817361593246,
0.006232191808521748,
0.024756845086812973,
-0.014684001915156841,
-0.013941947370767593,
0.017968157306313515,
0.061095595359802246,
-0.01852811686694622,
0.03570037707686424,
0.03849057853221893,
-0.06055101379752159,
-0.0789199247956276,
-0.011302126571536064,
0.08089661598205566,
-0.12279161065816879,
0.037349458783864975,
-0.03028588928282261,
-0.03976759314537048,
-0.015134296379983425,
0.03229302167892456,
-0.012614230625331402,
0.014565136283636093,
0.002538982080295682,
-0.024029206484556198,
0.033201348036527634,
-0.03783337026834488,
-0.006893073208630085,
-0.009076462127268314,
-0.048719875514507294,
0.0019935027230530977,
-0.017061177641153336,
0.0033487624023109674,
0.007341650314629078,
0.08587684482336044,
0.0589623749256134,
0.044651828706264496,
-0.024291910231113434,
0.006162006873637438,
-0.03559476509690285,
-0.1286781132221222,
-0.032239992171525955,
-0.00919255893677473,
-0.1123935878276825,
0.038324058055877686,
0.024546006694436073,
0.06866582483053207,
0.07041485607624054,
0.014305472373962402,
-0.07318943738937378,
0.02609201893210411,
-0.01031911838799715,
-0.0018372440245002508,
0.03752024471759796,
0.009965545497834682,
-0.0407191663980484,
-0.07863714545965195,
-0.010032429359853268,
0.019319869577884674,
0.05659995228052139,
-0.002246713265776634,
-0.052849262952804565,
0.005971183534711599,
0.05569809302687645,
0.0025590804871171713,
0.02054937742650509,
-0.004721606615930796,
0.16559875011444092,
-0.01683603599667549,
0.08392554521560669,
-0.010595262981951237,
0.027834763750433922,
-0.011681610718369484,
0.06834359467029572,
-0.006602211855351925,
0.04834778234362602,
-0.04229162260890007,
-0.0014094518264755607,
0.09049469977617264,
0.021833857521414757,
0.045625168830156326,
-0.04261603206396103,
0.029558688402175903,
-0.025007594376802444,
-0.010975470766425133,
-0.01697246916592121,
0.09930674731731415,
0.05159624293446541,
-0.03689894825220108,
0.01742623932659626,
-0.014298017136752605,
0.011678216978907585,
0.016956523060798645,
0.004121918696910143,
-0.022332150489091873,
0.07869834452867508,
-0.07135741412639618,
-0.039596669375896454,
0.11713606864213943,
0.011611736379563808,
-0.06407366693019867,
0.05243956670165062,
-0.013819866813719273,
0.006598041392862797,
-0.04999183490872383,
-0.022362355142831802,
3.6509566543706634e-34,
0.05382384732365608,
-0.023642972111701965,
-0.002041079103946686,
-0.01708134077489376,
0.032075803726911545,
0.036445166915655136,
-0.0009434881503693759,
0.019324183464050293,
-0.008486553095281124,
-0.06474081426858902,
0.10975608974695206,
0.06720389425754547,
-0.0724530890583992,
0.04354763403534889,
0.0555991493165493,
-0.028885161504149437,
-0.06331028044223785,
0.06904400140047073,
0.043776072561740875,
-0.024553224444389343,
0.00852583535015583,
-0.0330849327147007,
-0.019505228847265244,
-0.10220857709646225,
0.03945530205965042,
0.04046699032187462,
0.035010918974876404,
0.025059370324015617,
0.07067826390266418,
-0.0006012067897245288,
-0.03942709416151047,
0.04011362046003342,
-0.0312710665166378,
0.12141615897417068,
0.02979329228401184,
0.013698756694793701,
-0.0904889851808548,
-0.04275700822472572,
-0.023666948080062866,
0.010929733514785767,
-0.03277100995182991,
0.05330411344766617,
0.0360918827354908,
0.04175515100359917,
-0.0737336128950119,
-0.20705708861351013,
0.07772393524646759,
0.003605129662901163,
-0.05000535398721695,
-0.04844691976904869,
-0.015619480982422829,
0.05720715969800949,
-0.004806524142622948,
-0.0665416270494461,
-0.10155681520700455,
-0.02834048867225647,
-0.019520971924066544,
0.008985253050923347,
-0.01875140890479088,
0.03270772472023964,
0.12543462216854095,
-0.004637491423636675,
0.009514355100691319,
0.08876580744981766,
0.05211205780506134,
0.04238002747297287,
-0.013683116994798183,
0.04916837811470032,
0.003146963892504573,
-0.015395306050777435,
0.028038527816534042,
-0.01930169388651848,
0.021504083648324013,
-0.01969662867486477,
0.12614567577838898,
-0.04181152209639549,
0.060668591409921646,
-0.058343157172203064,
-0.05671912804245949,
-0.04178736358880997,
0.0824926495552063,
-0.012247618287801743,
-0.05616821348667145,
0.0019107465632259846,
-0.016763310879468918,
-0.05039583146572113,
-0.0065346513874828815,
-0.0859733447432518,
-0.04379817470908165,
0.014445138163864613,
-0.047519251704216,
-0.09205437451601028,
-0.006718232296407223,
0.023832278326153755,
-0.01573227345943451,
-1.3819863720980346e-33,
0.023517796769738197,
-0.10662094503641129,
-0.0034216174390167,
0.01831076294183731,
0.0028419482987374067,
-0.02715965174138546,
-0.031110599637031555,
-0.0372474379837513,
-0.019300298765301704,
-0.09400168806314468,
-0.038297154009342194,
0.051105745136737823,
0.063876673579216,
0.09569177776575089,
-0.03618476539850235,
-0.05905603989958763,
-0.02106451988220215,
0.010952373035252094,
-0.04442864656448364,
0.015431328676640987,
0.061064038425683975,
-0.03219858556985855,
-0.056894782930612564,
0.0534895695745945,
0.018334457650780678,
0.021628733724355698,
-0.016599364578723907,
0.06005864217877388,
-0.028593948110938072,
0.03533126041293144,
-0.07476068288087845,
0.012901321053504944,
0.022247768938541412,
-0.016360541805624962,
-0.056764453649520874,
0.09911192953586578,
0.0054342057555913925,
0.05981489270925522,
-0.06543144583702087,
0.05182904377579689,
0.09646524488925934,
-0.030211828649044037,
-0.05729928985238075,
-0.05819582939147949,
0.030752209946513176,
0.007439875975251198,
-0.07680349051952362,
0.033114757388830185,
0.01972503401339054,
0.0234077125787735,
-0.030527541413903236,
0.07166430354118347,
-0.10619485378265381,
-0.06381874531507492,
-0.052461009472608566,
-0.06546127051115036,
-0.02556750923395157,
-0.04007727652788162,
-0.0108180595561862,
-0.025434346869587898,
-0.020421426743268967,
-0.027760306373238564,
-0.05116443336009979,
0.0022157449275255203,
0.029798198491334915,
-0.0015784366987645626,
0.02251112461090088,
0.03364867344498634,
-0.05697643384337425,
-0.0026261433959007263,
0.03349192813038826,
0.028060967102646828,
-0.11041779071092606,
-0.1031867265701294,
0.05344146117568016,
0.036793481558561325,
-0.028322387486696243,
-0.011798057705163956,
-0.04283016175031662,
-0.00873613078147173,
0.013841922394931316,
0.011885484680533409,
0.029615039005875587,
0.07424116134643555,
0.0077911680564284325,
0.07669540494680405,
-0.03658737242221832,
0.10336657613515854,
0.0353475958108902,
-0.005900945980101824,
-0.004725749604403973,
-0.03957877680659294,
-0.06461320072412491,
0.013409949839115143,
-0.05165775120258331,
-4.374787110350553e-8,
-0.07842524349689484,
0.03304646909236908,
0.06778980791568756,
0.004600792191922665,
0.08814957737922668,
0.0003290523018222302,
0.004415713716298342,
-0.01702306419610977,
-0.0006602483917959034,
0.007954669184982777,
0.020521067082881927,
-0.027400828897953033,
0.062374766916036606,
-0.11797751486301422,
0.014071333222091198,
0.08394818007946014,
0.01791924238204956,
-0.012800528667867184,
-0.03831391781568527,
0.007527302484959364,
0.06143435835838318,
-0.013563448563218117,
0.012880103662610054,
-0.003186360001564026,
-0.0009697931818664074,
-0.07165854424238205,
0.06617163121700287,
0.1251964569091797,
0.13047045469284058,
-0.006787530612200499,
0.00010922457295237109,
-0.037199169397354126,
0.044670794159173965,
-0.06349442899227142,
-0.08580243587493896,
0.0149040836840868,
0.08303496241569519,
-0.05554167553782463,
-0.0515216663479805,
-0.026785582304000854,
-0.03797265887260437,
-0.01262000110000372,
0.017198629677295685,
-0.029888341203331947,
0.017707498744130135,
0.05667141452431679,
-0.08997821807861328,
-0.054830729961395264,
0.06435639411211014,
-0.06970994919538498,
-0.02392277494072914,
0.03736748546361923,
-0.02434040978550911,
0.09383712708950043,
0.1292673796415329,
-0.03986595943570137,
0.02107113227248192,
-0.01964345946907997,
0.03833770379424095,
0.09860466420650482,
-0.033245302736759186,
0.016628876328468323,
-0.06921796500682831,
0.037793662399053574
] | https://github.com/lcpo-rl/lcpo | 2 | 0 | 0 | 0 |
CEB: Compositional Evaluation Benchmark for Fairness in Large Language Models | https://openreview.net/forum?id=IUmj2dw5se | [
"Song Wang",
"Peng Wang",
"Tong Zhou",
"Yushun Dong",
"Zhen Tan",
"Jundong Li"
] | Spotlight | As Large Language Models (LLMs) are increasingly deployed to handle various natural language processing (NLP) tasks, concerns regarding the potential negative societal impacts of LLM-generated content have also arisen. To evaluate the biases exhibited by LLMs, researchers have recently proposed a variety of datasets. However, existing bias evaluation efforts often focus on only a particular type of bias and employ inconsistent evaluation metrics, leading to difficulties in comparison across different datasets and LLMs. To address these limitations, we collect a variety of datasets designed for the bias evaluation of LLMs, and further propose CEB, a Compositional Evaluation Bechmark that covers different types of bias across different social groups and tasks. The curation of CEB is based on our newly proposed compositional taxonomy, which characterizes each dataset from three dimensions: bias types, social groups, and tasks. By combining the three dimensions, we develop a comprehensive evaluation strategy for the bias in LLMs. Our experiments demonstrate that the levels of bias vary across these dimensions, thereby providing guidance for the development of specific bias mitigation methods. | Fairness, Bias, Benchmark, Large Language Models | null | 11,458 | 2407.02408 | [
0.001167350448668003,
-0.0476197823882103,
0.006881395820528269,
0.003607192076742649,
0.06125320866703987,
0.04039697349071503,
-0.013949588872492313,
-0.04941674694418907,
0.05049670487642288,
-0.012888680212199688,
-0.055465180426836014,
-0.10997432470321655,
0.0562577024102211,
0.005043174605816603,
0.031120404601097107,
0.015527506358921528,
0.099579818546772,
0.014844603836536407,
-0.08044882118701935,
-0.0935673862695694,
0.016904693096876144,
0.08235161751508713,
0.004982592072337866,
0.007331510074436665,
-0.02831312268972397,
-0.09704926609992981,
-0.030964335426688194,
-0.04719981178641319,
0.004230649210512638,
0.029424555599689484,
0.03778763487935066,
-0.004609745927155018,
0.058996107429265976,
0.07166719436645508,
-0.07510291785001755,
0.014185808598995209,
-0.0361834354698658,
-0.022856634110212326,
-0.0025725453160703182,
-0.035460516810417175,
-0.03478096053004265,
0.0007619199459441006,
0.00615258514881134,
-0.013599030673503876,
0.07173196226358414,
0.0018632750725373626,
0.023231765255331993,
-0.008144767954945564,
-0.11378837376832962,
0.02543126605451107,
-0.02407808229327202,
-0.028145557269454002,
0.0426224023103714,
0.1195206493139267,
-0.04699503257870674,
0.009557980112731457,
0.006940394174307585,
0.0627717673778534,
-0.009714653715491295,
-0.037333521991968155,
-0.08749735355377197,
-0.17335467040538788,
-0.06433450430631638,
0.03227334842085838,
0.013477833941578865,
-0.038221463561058044,
-0.049068499356508255,
0.06076706200838089,
-0.09001324325799942,
0.024512525647878647,
0.01145104318857193,
0.021000726148486137,
0.03221145272254944,
0.06757529079914093,
0.04458627849817276,
0.05208877846598625,
0.01780429668724537,
-0.06463205069303513,
0.06327124685049057,
-0.10423019528388977,
0.031971707940101624,
0.020739661529660225,
0.03309476375579834,
0.037468165159225464,
0.10877913236618042,
0.004232633393257856,
-0.01188720017671585,
0.07805908471345901,
-0.06618261337280273,
0.08794397860765457,
-0.024439429864287376,
-0.010567580349743366,
0.117972731590271,
0.028188610449433327,
0.01512391772121191,
0.025428004562854767,
0.006597024854272604,
-0.06967918574810028,
0.04077623039484024,
0.08272381126880646,
0.006980757229030132,
0.06399289518594742,
-0.05417812615633011,
-0.05794411897659302,
-0.010652918368577957,
0.012785711325705051,
0.0021997930016368628,
0.015438801608979702,
-0.053343020379543304,
-0.02293977327644825,
-0.04512032866477966,
0.05278138071298599,
0.010065718553960323,
-0.005089002661406994,
0.013555433601140976,
-0.06344839185476303,
0.07649485021829605,
-0.025896240025758743,
0.025730552151799202,
0.02653416618704796,
-0.018206559121608734,
0.04395562410354614,
0.09397462755441666,
0.00018728303257375956,
0.02046557329595089,
0.029955770820379257,
-0.017991937696933746,
2.0785203150545e-33,
0.07135192304849625,
0.03810715675354004,
-0.024930188432335854,
-0.006517046596854925,
0.04644351452589035,
-0.032679010182619095,
-0.020892124623060226,
0.07510770112276077,
-0.06452690064907074,
-0.016435878351330757,
0.023111021146178246,
0.02556222677230835,
-0.0035308715887367725,
0.036211565136909485,
0.025038616731762886,
0.019667167216539383,
-0.03294698894023895,
0.09648428112268448,
-0.04923772066831589,
0.04217894747853279,
0.043925829231739044,
0.010715194046497345,
0.009221847169101238,
-0.023146571591496468,
-0.01684494875371456,
0.014807832427322865,
0.0201573446393013,
-0.03662801906466484,
-0.042929328978061676,
0.017758892849087715,
-0.0961904227733612,
-0.033412717282772064,
0.04363401234149933,
0.010906647890806198,
0.02797609008848667,
-0.004171501845121384,
-0.01519737672060728,
0.03260832652449608,
0.005784448701888323,
-0.029549406841397285,
-0.014613424427807331,
0.03861033543944359,
0.03539828211069107,
-0.004871249198913574,
-0.02028605155646801,
-0.07130301743745804,
0.004009652882814407,
-0.04953339323401451,
0.06839766353368759,
0.03152225911617279,
0.10847532749176025,
-0.012932123616337776,
0.09466491639614105,
0.013284167274832726,
0.016554567962884903,
0.025677770376205444,
-0.007922668941318989,
0.012698754668235779,
0.04250693693757057,
-0.022455425933003426,
-0.006647574715316296,
0.04732427000999451,
-0.012964475899934769,
0.0005913689383305609,
0.008280191570520401,
0.030628254637122154,
-0.024283207952976227,
-0.004627589136362076,
0.03868115320801735,
-0.04283370450139046,
0.05403098464012146,
0.045860838145017624,
-0.05899013206362724,
0.01934957318007946,
-0.01279730536043644,
0.0875278189778328,
-0.00820537842810154,
-0.011633561924099922,
0.01951419934630394,
-0.02290058322250843,
0.04236796125769615,
-0.05716206878423691,
0.004658116027712822,
-0.1478046029806137,
-0.06265932321548462,
-0.09696047753095627,
0.008135286159813404,
-0.035716667771339417,
0.06055392324924469,
0.03677015006542206,
-0.04401063174009323,
-0.025704924017190933,
-0.04762378707528114,
0.05448024719953537,
-0.01758120208978653,
-2.7227555041524242e-33,
-0.11330565810203552,
0.0016201324760913849,
-0.07209344208240509,
0.07166425883769989,
-0.03201134875416756,
-0.038424935191869736,
-0.0027265127282589674,
-0.02048577181994915,
0.017178550362586975,
0.013538680970668793,
-0.024575630202889442,
-0.06323045492172241,
0.0701439157128334,
-0.0039449892938137054,
0.06989695876836777,
-0.06924539804458618,
0.0043147774413228035,
0.07070640474557877,
0.02929491177201271,
0.03874754533171654,
0.0745207741856575,
0.10110115259885788,
-0.09426841139793396,
0.09822200238704681,
0.02332928031682968,
0.025397498160600662,
-0.047970499843358994,
-0.04408320412039757,
0.021311048418283463,
0.008076583035290241,
-0.0831340029835701,
-0.026962485164403915,
-0.0500643253326416,
-0.08018220961093903,
-0.01687994971871376,
-0.03617600351572037,
-0.0028015554416924715,
0.002087630331516266,
-0.04159114509820938,
0.04389873892068863,
0.017014646902680397,
0.016482267528772354,
-0.1441938579082489,
0.03988320007920265,
-0.04533480852842331,
0.011538127437233925,
-0.05555688962340355,
-0.0403873547911644,
0.06545116752386093,
-0.009164471179246902,
-0.0580865852534771,
-0.056354980915784836,
-0.006240117829293013,
0.087039515376091,
0.0041109900921583176,
-0.10893388837575912,
0.010497735813260078,
-0.11651317030191422,
-0.0038713980466127396,
0.052351102232933044,
-0.033987198024988174,
0.04074263572692871,
-0.0020728132221847773,
-0.06827723979949951,
0.0429609939455986,
-0.1058063879609108,
-0.062374915927648544,
-0.01517060399055481,
-0.015644580125808716,
0.02698023058474064,
0.07755530625581741,
-0.012386655434966087,
-0.04094149172306061,
-0.037655673921108246,
-0.05933826416730881,
0.00434950552880764,
0.05941222235560417,
-0.022352153435349464,
-0.0639757439494133,
0.025149591267108917,
0.0673341229557991,
0.011128036305308342,
0.08361995965242386,
-0.023499593138694763,
0.03683122247457504,
0.06977732479572296,
-0.01341298408806324,
0.06522028148174286,
-0.047939952462911606,
0.04082536697387695,
-0.04371140897274017,
-0.033945098519325256,
-0.027388283982872963,
0.06964650750160217,
-0.08494210243225098,
-4.200758851879982e-8,
-0.028897741809487343,
-0.013089246116578579,
-0.06514202058315277,
0.12701980769634247,
-0.10075568407773972,
-0.05402496084570885,
-0.07770554721355438,
0.0056589748710393906,
0.007378744427114725,
-0.027346381917595863,
0.03676970303058624,
0.012434748001396656,
-0.031884972006082535,
-0.052177075296640396,
0.05687599256634712,
0.051013898104429245,
-0.022219575941562653,
0.023878414183855057,
0.0002198176662204787,
-0.00013208235031925142,
0.08583616465330124,
0.04222923144698143,
-0.04695102944970131,
-0.09221950173377991,
0.013406249694526196,
-0.05442479997873306,
-0.04806222766637802,
0.04396491497755051,
0.022481190040707588,
-0.07288714498281479,
-0.02037753537297249,
0.02877114899456501,
-0.08976477384567261,
-0.03128306195139885,
0.08232854306697845,
0.06299573928117752,
-0.07814772427082062,
-0.009265457279980183,
0.08261660486459732,
0.10099263489246368,
0.07529018074274063,
0.09298902750015259,
-0.0035648434422910213,
0.020307548344135284,
0.019582128152251244,
0.0010187294101342559,
-0.07597469538450241,
0.02763274870812893,
-0.030526820570230484,
-0.09312321245670319,
-0.007827294990420341,
-0.07451876252889633,
-0.05860661342740059,
0.07133540511131287,
0.06729693710803986,
0.01103877928107977,
0.012804454192519188,
0.008691739290952682,
0.028109975159168243,
-0.0027591127436608076,
0.1003653034567833,
0.0057382467202842236,
-0.01682291552424431,
-0.03475189581513405
] | 0 | 0 | 1 | 0 |
|
TabWak: A Watermark for Tabular Diffusion Models | https://openreview.net/forum?id=71pur4y8gs | [
"Chaoyi Zhu",
"Jiayi Tang",
"Jeroen M. Galjaard",
"Pin-Yu Chen",
"Robert Birke",
"Cornelis Bos",
"Lydia Y. Chen"
] | Spotlight | Synthetic data offers alternatives for data augmentation and sharing. Till date, it remains unknown how to use watermarking techniques to trace and audit synthetic tables generated by tabular diffusion models to mitigate potential misuses. In this paper, we design TabWak, the first watermarking method to embed invisible signatures that control the sampling of Gaussian latent codes used to synthesize table rows via the diffusion backbone. TabWak has two key features. Different from existing image watermarking techniques, TabWak uses self-cloning and shuffling to embed the secret key in positional information of random seeds that control the Gaussian latents, allowing to use different seeds at each row for high inter-row diversity and enabling row-wise detectability. To further boost the robustness of watermark detection against post-editing attacks, TabWak uses a valid-bit mechanism that focuses on the tail of the latent code distribution for superior noise resilience. We provide theoretical guarantees on the row diversity and effectiveness of detectability. We evaluate TabWak on five datasets against baselines to show that the quality of watermarked tables remains nearly indistinguishable from non-watermarked tables while achieving high detectability in the presence of strong post-editing attacks, with a 100% true positive rate at a 0.1% false positive rate on synthetic tables with fewer than 300 rows. Our code is available at the following anonymized repository https://github.com/chaoyitud/TabWak. | Watermarking, Tabular data, Generative models, Tabular diffusion models | We propose Tabwak, the first watermarking method for tabular diffusion models that ensures high data quality, detectability, and robustness against post-editing attacks | 11,455 | null | [
-0.07113122940063477,
-0.05280769616365433,
-0.03185528144240379,
0.026068566367030144,
0.13385792076587677,
-0.039135899394750595,
-0.01134831365197897,
-0.1140279471874237,
-0.0015764526324346662,
0.019746942445635796,
0.028081635013222694,
0.006422596052289009,
0.09797109663486481,
-0.039650771766901016,
-0.09196119010448456,
-0.030660480260849,
-0.04834100976586342,
0.06960485875606537,
-0.10471262037754059,
-0.04029611498117447,
-0.04721260070800781,
-0.08071715384721756,
-0.007824291475117207,
-0.03334450349211693,
0.04305306077003479,
0.08574005216360092,
0.039615508168935776,
-0.014923302456736565,
-0.022049250081181526,
-0.07583450525999069,
0.03301498666405678,
-0.026462579146027565,
-0.002831041347235441,
0.05050176382064819,
-0.005985235795378685,
0.01236303336918354,
-0.014166567474603653,
-0.052482716739177704,
0.1017322912812233,
-0.009813832119107246,
-0.025536855682730675,
-0.07760979980230331,
-0.043504633009433746,
0.05406937748193741,
-0.013960420154035091,
-0.002736177761107683,
0.036777716130018234,
-0.008263763040304184,
-0.0700899213552475,
-0.0018175244331359863,
-0.007842072285711765,
0.04280564934015274,
-0.0936732217669487,
0.0061247460544109344,
-0.027708452194929123,
-0.11656162887811661,
0.04071914032101631,
-0.032011982053518295,
-0.010471698828041553,
0.07334335148334503,
-0.014681411907076836,
0.0005566747859120369,
0.048030585050582886,
-0.009796795435249805,
-0.018792375922203064,
0.025486599653959274,
-0.029940076172351837,
0.03860211744904518,
0.04042680189013481,
0.011599271558225155,
-0.01973150484263897,
0.08303428441286087,
-0.06567957252264023,
0.0734686553478241,
-0.10214745998382568,
0.04436220973730087,
-0.04703753814101219,
0.01740967482328415,
0.015460344962775707,
-0.101737841963768,
0.011413134634494781,
-0.059959735721349716,
0.03150924667716026,
0.07023743540048599,
0.09184971451759338,
0.013441145420074463,
0.04483211785554886,
0.058174699544906616,
0.025048911571502686,
-0.02190324291586876,
0.0008241822361014783,
0.05188475921750069,
0.06725148111581802,
0.03682773932814598,
0.012279163114726543,
0.04396446794271469,
0.03342777118086815,
-0.013564744964241982,
0.029385849833488464,
0.0747121050953865,
-0.07088421285152435,
-0.03058624640107155,
0.013861964456737041,
-0.010956203565001488,
0.11912526935338974,
-0.03267199173569679,
0.07946646958589554,
-0.02283540554344654,
0.061083681881427765,
-0.04480915516614914,
-0.005057969596236944,
0.04667195677757263,
-0.05707917362451553,
-0.04032554849982262,
0.0030773167964071035,
-0.01622343808412552,
-0.0620930939912796,
-0.023538850247859955,
0.011760449036955833,
-0.015247547067701817,
0.06306196004152298,
-0.02499311976134777,
-0.08186345547437668,
-0.02094367891550064,
-0.03198237717151642,
0.022524170577526093,
-0.04538705199956894,
3.424951931966495e-33,
0.021026799455285072,
0.012286178767681122,
0.00693130400031805,
-0.08650041371583939,
0.11095914989709854,
-0.005205559078603983,
-0.00400154571980238,
-0.043383967131376266,
-0.05346529930830002,
-0.06152515113353729,
-0.03836542367935181,
-0.006267039105296135,
-0.0864700898528099,
0.07115847617387772,
0.042222604155540466,
0.09137684851884842,
-0.07412975281476974,
0.0017273526173084974,
-0.038887206465005875,
-0.04160335287451744,
0.04792989417910576,
-0.09656991064548492,
-0.052183281630277634,
-0.013438660651445389,
0.0226079523563385,
0.022362930700182915,
0.02988237701356411,
-0.05331401526927948,
0.04381566867232323,
0.04519099369645119,
-0.07292456179857254,
-0.008459729142487049,
0.015350552275776863,
-0.014236257411539555,
0.03290712833404541,
-0.00025589001597836614,
0.0183738823980093,
-0.05791008099913597,
0.013367405161261559,
0.02538492903113365,
0.07955014705657959,
0.03553735464811325,
-0.0643719956278801,
-0.02445661649107933,
-0.0666300356388092,
0.031567491590976715,
0.025772305205464363,
0.0682036504149437,
0.0013872688869014382,
-0.06036050617694855,
0.05951112136244774,
0.008865504525601864,
-0.013187062926590443,
-0.025509705767035484,
-0.08285515010356903,
-0.05123551934957504,
0.02272411808371544,
-0.10629837960004807,
0.10997083783149719,
0.08554442971944809,
-0.024052271619439125,
0.012615370564162731,
0.03488540276885033,
0.031836237758398056,
-0.03366701304912567,
0.01661790907382965,
0.01945270411670208,
0.03058432787656784,
0.02518818899989128,
0.008266154676675797,
-0.004003985784947872,
-0.037494298070669174,
-0.03168819472193718,
-0.001968234311789274,
-0.05006071552634239,
-0.09788975119590759,
0.067593514919281,
-0.002038296079263091,
0.02385527454316616,
-0.056644801050424576,
0.010726927779614925,
-0.026272589340806007,
0.010631720535457134,
-0.08908364921808243,
-0.16855375468730927,
-0.04758087545633316,
-0.042641155421733856,
-0.05371418222784996,
-0.11771708726882935,
0.023359844461083412,
0.034001074731349945,
-0.03270149230957031,
0.029860971495509148,
-0.06019985303282738,
-0.035862404853105545,
-4.045057895461884e-33,
-0.03462160378694534,
0.045309558510780334,
-0.021176837384700775,
0.07074896991252899,
-0.019105631858110428,
0.0015634699957445264,
0.0535968653857708,
0.04959852620959282,
0.08356430381536484,
0.0031611009035259485,
-0.06922796368598938,
-0.0008437151554971933,
0.017635969445109367,
-0.01940244436264038,
-0.006770334206521511,
-0.003338760696351528,
-0.009420916438102722,
0.06637286394834518,
-0.029068687930703163,
-0.024889452382922173,
0.0006920702289789915,
0.005173492711037397,
-0.019067469984292984,
0.12692058086395264,
0.014749083667993546,
0.0632537379860878,
-0.005330522079020739,
-0.009353958070278168,
-0.008326218463480473,
-0.03336475417017937,
-0.03986496105790138,
0.05062846839427948,
-0.023182295262813568,
-0.04569045826792717,
-0.00027625096845440567,
-0.010913422331213951,
0.024322550743818283,
-0.012131698429584503,
-0.012688305228948593,
0.06816233694553375,
0.01702285371720791,
0.019690552726387978,
-0.011497650295495987,
0.051986295729875565,
-0.014715718105435371,
0.040431126952171326,
0.005638968665152788,
0.06790778040885925,
0.01618400774896145,
0.0015260432846844196,
0.03198971599340439,
-0.008249674923717976,
-0.04788001999258995,
0.031656987965106964,
-0.025531845167279243,
0.014558231458067894,
-0.07861008495092392,
0.052495721727609634,
-0.0030907802283763885,
0.01572967879474163,
-0.05789549648761749,
-0.012706342153251171,
0.0512084998190403,
-0.0480782613158226,
0.030205661430954933,
-0.05153683200478554,
0.017275869846343994,
-0.001416867016814649,
-0.03982839360833168,
0.024433758109807968,
0.025170648470520973,
-0.0705954059958458,
-0.0016932430444285274,
0.04565705358982086,
0.06755195558071136,
-0.04498446732759476,
-0.06481235474348068,
-0.007079600356519222,
-0.07878655940294266,
0.014658214524388313,
0.030013076961040497,
0.0007090013823471963,
0.05775246024131775,
-0.036889392882585526,
0.023894328624010086,
0.05700195953249931,
0.016635995358228683,
-0.04557199403643608,
-0.013349983841180801,
0.08223351836204529,
-0.04631263390183449,
-0.014612967148423195,
-0.0754709243774414,
0.12344755977392197,
0.010538211092352867,
-4.2060921856545974e-8,
-0.05314703658223152,
0.031579479575157166,
-0.004034434910863638,
-0.03487635403871536,
0.007718289736658335,
-0.025644317269325256,
0.040205445140600204,
0.13219618797302246,
0.06269572675228119,
-0.1392412781715393,
0.14642886817455292,
-0.025359485298395157,
-0.06592356413602829,
-0.06579021364450455,
-0.02933041751384735,
0.08875997364521027,
-0.053845375776290894,
-0.03819577023386955,
-0.03586919605731964,
0.01591556705534458,
-0.0484936386346817,
0.0755489245057106,
-0.0035842678043991327,
-0.011368484236299992,
0.020522460341453552,
0.05412571132183075,
-0.005306649953126907,
0.10776324570178986,
0.09977149963378906,
0.047592367976903915,
0.04897013679146767,
-0.02370082214474678,
0.08304299414157867,
0.010619252920150757,
-0.0058930665254592896,
0.0924411490559578,
0.030956817790865898,
-0.026334233582019806,
-0.02904689684510231,
0.025611143559217453,
0.00815503764897585,
-0.015845054760575294,
-0.05675715580582619,
-0.07117888331413269,
-0.035607531666755676,
0.021886873990297318,
0.02804543450474739,
-0.012592480517923832,
0.0075876908376812935,
0.01469749677926302,
0.06978129595518112,
-0.09468929469585419,
-0.02544364333152771,
0.09844677150249481,
0.0066700344905257225,
-0.0032404109369963408,
-0.028212307021021843,
0.01084627490490675,
0.1378849893808365,
-0.00263895932585001,
0.07091159373521805,
-0.0272331852465868,
0.06840036064386368,
0.10001805424690247
] | 0 | 0 | 0 | 0 |
|
Active Task Disambiguation with LLMs | https://openreview.net/forum?id=JAMxRSXLFz | [
"Kasia Kobalczyk",
"Nicolás Astorga",
"Tennison Liu",
"Mihaela van der Schaar"
] | Spotlight | Despite the impressive performance of large language models (LLMs) across various benchmarks, their ability to address ambiguously specified problems—frequent in real-world interactions—remains underexplored. To address this gap, we introduce a formal definition of task ambiguity and frame the problem of task disambiguation through the lens of Bayesian Experimental Design. By posing clarifying questions, LLM agents can acquire additional task specifications, progressively narrowing the space of viable solutions and reducing the risk of generating unsatisfactory outputs. Yet, generating effective clarifying questions requires LLM agents to engage in a form of meta-cognitive reasoning, an ability LLMs may presently lack. Our proposed approach of active task disambiguation enables LLM agents to generate targeted questions maximizing the information gain. Effectively, this approach shifts the load from implicit to explicit reasoning about the space of viable solutions. Empirical results demonstrate that this form of question selection leads to more effective task disambiguation in comparison to approaches relying on reasoning solely within the space of questions. | Task Ambiguity, Bayesian Experimental Design, Large Language Models, Active Learning | This paper formalizes task ambiguity in tasks specified in natural language and frames task disambiguation through Bayesian Experimental Design, leading to more effective strategies for LLMs to pose clarifying questions. | 11,335 | 2502.04485 | [
-0.002890742616727948,
-0.04293217882514,
0.015066975727677345,
0.024350106716156006,
0.10719875246286392,
-0.05083789676427841,
0.08204428106546402,
0.042960889637470245,
-0.0002874412457458675,
-0.01793285459280014,
-0.053206417709589005,
-0.10884184390306473,
0.015679843723773956,
-0.00045220355968922377,
0.055477213114500046,
0.04439227655529976,
0.1326570361852646,
0.02032671868801117,
-0.05728130042552948,
0.014437928795814514,
0.0927426889538765,
0.01784968003630638,
0.0630924254655838,
0.0020263884216547012,
-0.03783576563000679,
-0.043508823961019516,
0.024142244830727577,
0.0004134333285037428,
0.0035077338106930256,
0.04813763126730919,
0.07836694270372391,
0.08017618954181671,
-0.061193954199552536,
0.023527443408966064,
0.03503686934709549,
0.0010668260511010885,
-0.0919751301407814,
0.05979501083493233,
0.019347772002220154,
-0.035551827400922775,
-0.06055622920393944,
-0.04792878031730652,
0.01230448018759489,
0.03661935403943062,
0.005529964342713356,
-0.0590449683368206,
-0.05743763968348503,
0.03065398894250393,
-0.02755001373589039,
0.00689031183719635,
-0.15582332015037537,
-0.04273592308163643,
-0.031069491058588028,
-0.02457786537706852,
-0.005071623716503382,
0.01655137538909912,
-0.01640346087515354,
0.03366776555776596,
0.0035769406240433455,
-0.051610834896564484,
-0.07108470797538757,
-0.11979623138904572,
-0.003689801786094904,
-0.024663744494318962,
0.014717384241521358,
-0.0006062880856916308,
0.011292383074760437,
0.058214716613292694,
-0.03721442073583603,
0.02089371345937252,
-0.0185123011469841,
0.09404456615447998,
0.010941553860902786,
0.052639830857515335,
0.01249257754534483,
0.013676858507096767,
-0.0625455304980278,
-0.001317822840064764,
0.04944537952542305,
-0.027924316003918648,
0.005128834862262011,
-0.04065621644258499,
-0.02333001419901848,
0.060956958681344986,
0.11264916509389877,
0.05528672784566879,
-0.04026499390602112,
0.05366174131631851,
0.021075529977679253,
0.024226179346442223,
-0.05947878956794739,
-0.1285456120967865,
0.043777041137218475,
0.05442662537097931,
0.06622031331062317,
0.03284171223640442,
0.031943514943122864,
-0.05329303443431854,
-0.010043712332844734,
0.06825489550828934,
0.05346638336777687,
0.12553460896015167,
-0.0005419334047473967,
-0.03244658559560776,
-0.08100243657827377,
0.0002238098968518898,
-0.009374373592436314,
0.05193512141704559,
-0.0013832285767421126,
-0.11784514784812927,
-0.00987429078668356,
0.030671600252389908,
0.10470936447381973,
-0.028748586773872375,
-0.03401506319642067,
-0.04438779503107071,
0.09386145323514938,
0.024533694609999657,
0.05000614374876022,
-0.019909430295228958,
0.04978451505303383,
0.041569363325834274,
0.02178521826863289,
-0.027762778103351593,
0.07015929371118546,
-0.01268137525767088,
-0.07031701505184174,
1.99736583811266e-33,
0.10862936824560165,
0.005025871098041534,
0.05656838417053223,
0.12184154987335205,
0.025222372263669968,
-0.018533742055296898,
-0.01393868401646614,
0.05746667459607124,
-0.02448391541838646,
-0.009851330891251564,
0.018462950363755226,
-0.040938910096883774,
-0.016940345987677574,
0.0004059343773405999,
0.047447264194488525,
0.0376603826880455,
-0.05677243694663048,
0.10720741003751755,
-0.07778497785329819,
-0.026975296437740326,
0.00948549248278141,
0.014252548106014729,
0.05573606491088867,
-0.01126148272305727,
0.01238439604640007,
0.012715015560388565,
-0.015718910843133926,
-0.06092275679111481,
0.019104477018117905,
0.012759977020323277,
-0.12957637012004852,
-0.051192522048950195,
0.015873219817876816,
0.08425762504339218,
-0.05967951565980911,
0.01489970926195383,
-0.018088459968566895,
-0.09116029739379883,
0.018316064029932022,
-0.0346001498401165,
0.05038968846201897,
0.03467689827084541,
0.1381470412015915,
0.12586373090744019,
-0.05401548743247986,
-0.09149876981973648,
-0.007579305209219456,
0.019603971391916275,
-0.016394907608628273,
0.04101692885160446,
0.07622534781694412,
0.009068400599062443,
0.04115710034966469,
-0.06314418464899063,
0.027168262749910355,
-0.03973177820444107,
-0.046747006475925446,
-0.004305790178477764,
0.01913902349770069,
0.02910333313047886,
-0.040232084691524506,
0.011766517534852028,
-0.03255492076277733,
0.08690271526575089,
0.04814951494336128,
0.05275336280465126,
-0.008223330602049828,
0.0890001505613327,
0.08603806048631668,
-0.019459908828139305,
-0.03931953385472298,
-0.0436968170106411,
0.026367830112576485,
-0.008272085338830948,
0.01068124920129776,
0.040947094559669495,
0.013808834366500378,
-0.05043894425034523,
0.029582679271697998,
-0.04183749482035637,
0.07254504412412643,
-0.04684845730662346,
-0.040609605610370636,
0.025511259213089943,
-0.038271065801382065,
-0.048313576728105545,
0.019967572763562202,
-0.05817931890487671,
-0.014852532185614109,
-0.09130783379077911,
-0.0527777262032032,
-0.014812404289841652,
-0.10059709846973419,
-0.02450021728873253,
0.03026670590043068,
-1.186825932693222e-33,
-0.045567940920591354,
-0.0512923002243042,
-0.07381124794483185,
0.05127737671136856,
0.005835615564137697,
-0.03982003033161163,
0.030489865690469742,
-0.08198736608028412,
-0.03257749229669571,
-0.11042661964893341,
-0.1045374870300293,
-0.03374937176704407,
-0.05394838750362396,
-0.015580625273287296,
-0.05164254829287529,
-0.02348017878830433,
0.007033969275653362,
-0.011478260159492493,
-0.03402605280280113,
0.09040115773677826,
0.01908029615879059,
0.08267862349748611,
-0.016870254650712013,
-0.04323413968086243,
-0.01212090253829956,
0.04990468546748161,
0.036307577043771744,
0.003931615501642227,
-0.014382041059434414,
0.005656464956700802,
-0.013838354498147964,
-0.00373911508359015,
-0.0449538491666317,
-0.06677866727113724,
-0.03676905483007431,
0.1185736432671547,
0.043568339198827744,
-0.018571365624666214,
0.009025507606565952,
0.04980383813381195,
-0.0039059228729456663,
-0.07736353576183319,
-0.043408818542957306,
-0.015554164536297321,
-0.01888861134648323,
0.04522145539522171,
-0.05019262805581093,
-0.0031225834973156452,
0.01594644784927368,
0.022997163236141205,
0.019551603123545647,
-0.01134160440415144,
-0.008384544402360916,
-0.037516769021749496,
-0.07524498552083969,
-0.06020362302660942,
-0.032359108328819275,
-0.02560380846261978,
0.03300498053431511,
0.02784687466919422,
0.004804507829248905,
0.033543910831213,
0.04464523866772652,
-0.05324600264430046,
-0.005409355275332928,
-0.024100231006741524,
-0.046492431312799454,
0.043744370341300964,
0.005494334269315004,
-0.05583742633461952,
0.022124331444501877,
-0.07046683877706528,
0.017523543909192085,
-0.0348023883998394,
0.056345630437135696,
0.03774071857333183,
0.01812152937054634,
-0.04051888734102249,
-0.03269806131720543,
-0.031527865678071976,
0.019290367141366005,
-0.004751324187964201,
0.03347507491707802,
0.05551591515541077,
-0.043980155140161514,
-0.01840958371758461,
-0.026235604658722878,
0.12031444162130356,
0.030724167823791504,
-0.028579968959093094,
0.005932993721216917,
-0.01515384204685688,
0.009465414099395275,
0.09724077582359314,
-0.09888699650764465,
-3.9227458614732313e-8,
0.01545985322445631,
0.10556188225746155,
-0.0020856396295130253,
0.08131004124879837,
0.02807636931538582,
-0.019558856263756752,
-0.11591967940330505,
0.03650093451142311,
0.01867578737437725,
-0.052231889218091965,
0.0725809708237648,
-0.00809149257838726,
-0.03232015669345856,
0.0053600892424583435,
0.016699885949492455,
0.032601941376924515,
0.005251171067357063,
0.029949549585580826,
-0.051684267818927765,
-0.04187377542257309,
0.045322757214307785,
0.06711529940366745,
-0.048319049179553986,
0.029382100328803062,
0.0504329651594162,
-0.04750306159257889,
-0.03734886646270752,
0.0863378718495369,
0.07136905193328857,
0.03034401498734951,
0.00646850373595953,
0.004012304823845625,
-0.10820396989583969,
0.04801323264837265,
0.05281498655676842,
0.025966057553887367,
-0.06751757860183716,
-0.06571158766746521,
-0.00044781671022064984,
-0.05762028321623802,
-0.0008018147782422602,
0.04860646650195122,
-0.05401277542114258,
0.07785305380821228,
0.023913785815238953,
-0.0072638243436813354,
-0.08010877668857574,
-0.05461189150810242,
0.008879088796675205,
-0.04238741844892502,
-0.0057949465699493885,
-0.01952454075217247,
-0.04798685386776924,
0.03270690515637398,
0.025180786848068237,
0.07985755056142807,
0.017765911296010017,
0.021431585773825645,
-0.04115312919020653,
-0.02873590961098671,
0.0995488315820694,
0.08444231748580933,
-0.05076230689883232,
0.008535278029739857
] | https://github.com/kasia-kobalczyk/active-task-disambiguation | 5 | 0 | 0 | 0 |
Surprising Effectiveness of pretraining Ternary Language Model at Scale | https://openreview.net/forum?id=TJo6aQb7mK | [
"Ayush Kaushal",
"Tejas Vaidhya",
"Arnab Kumar Mondal",
"Tejas Pandey",
"Aaryan Bhagat",
"Irina Rish"
] | Spotlight | Rapid advancements in GPU computational power has outpaced memory capacity and bandwidth growth, creating bottlenecks in Large Language Model (LLM) inference. Post-training quantization is the leading method for addressing memory-related bottlenecks in LLM inference, but it suffers from significant performance degradation below 4-bit precision. This paper addresses these challenges by investigating the pretraining of low-bitwidth models specifically Ternary Language Models (TriLMs) as an alternative to traditional floating-point models (FloatLMs) and their post-training quantized versions (QuantLMs). We present Spectra LLM suite, the first open suite of LLMs spanning multiple bit-widths, including FloatLMs, QuantLMs, and TriLMs, ranging from 99M to 3.9B parameters trained on 300B tokens. Our comprehensive evaluation demonstrates that TriLMs offer superior scaling behavior in terms of model size (in bits). Surprisingly, at scales exceeding one billion parameters, TriLMs consistently outperform their QuantLM and FloatLM counterparts for a given bit size across various benchmarks. Notably, the 3.9B parameter TriLM matches the performance of the FloatLM 3.9B across all benchmarks, despite having fewer bits than FloatLM 830M. Overall, this research provides valuable insights into the feasibility and scalability of low-bitwidth language models, paving the way for the development of more efficient LLMs. | Large Language Models, low-bit language models, quantization-aware training, pretraining of large language models, and scaling laws | This paper introduces Spectra LLM suite, showcasing that Ternary Language Models (TriLMs) outperform traditional and quantized models in efficiency and scaling, especially beyond one billion parameters. | 11,310 | 2407.12327 | [
-0.008936787024140358,
-0.05028240755200386,
-0.06504230946302414,
0.03126521408557892,
-0.009827076457440853,
-0.02032100409269333,
-0.05335639417171478,
-0.04618977755308151,
0.025506803765892982,
-0.07572873681783676,
-0.02353597618639469,
-0.07090761512517929,
-0.03299865499138832,
0.03176356479525566,
0.049740977585315704,
-0.014738013967871666,
0.05726160109043121,
0.0572630874812603,
-0.08313778042793274,
-0.06016454100608826,
0.04963670298457146,
0.006283667869865894,
0.02081676758825779,
0.05051431432366371,
0.06765996664762497,
-0.039002642035484314,
0.001654391293413937,
-0.05883895233273506,
0.06638724356889725,
-0.05677996575832367,
0.02633478492498398,
0.07843032479286194,
-0.014595302753150463,
0.027112679556012154,
0.027955885976552963,
0.011317993514239788,
0.023239338770508766,
-0.054877541959285736,
0.011685486882925034,
-0.04421423748135567,
0.027846286073327065,
-0.03709346801042557,
0.049805354326963425,
0.040084533393383026,
0.07677629590034485,
0.0064105074852705,
0.023264164105057716,
0.01630321517586708,
-0.010922986082732677,
-0.047182466834783554,
0.04355209320783615,
0.011856691911816597,
-0.01959574967622757,
0.059702035039663315,
-0.13048946857452393,
-0.06036048382520676,
0.07303695380687714,
0.0436716228723526,
-0.04369935393333435,
0.0030760273803025484,
-0.0921797975897789,
-0.10055410861968994,
-0.018710840493440628,
0.0331449992954731,
-0.04214940592646599,
-0.01940419152379036,
0.032360225915908813,
0.010900854133069515,
-0.00974560622125864,
-0.011773507110774517,
-0.016846314072608948,
0.005827621556818485,
-0.057532329112291336,
0.07782605290412903,
-0.023355763405561447,
0.053162459284067154,
0.08483812212944031,
0.024889271706342697,
0.057294007390737534,
-0.028800563886761665,
0.013954809866845608,
-0.04092011973261833,
0.06673206388950348,
-0.008546165190637112,
0.08348618447780609,
-0.045786745846271515,
-0.033418379724025726,
0.01722470298409462,
-0.05904533341526985,
-0.025632211938500404,
-0.019204847514629364,
0.04074742645025253,
0.0016129532596096396,
0.03728137165307999,
0.03223923593759537,
0.04790058359503746,
0.018234338611364365,
-0.05127519741654396,
-0.03574030101299286,
0.0857023373246193,
0.06916771829128265,
0.07290516793727875,
0.02211289294064045,
0.03665412962436676,
-0.009867755696177483,
0.006200764328241348,
0.04050155356526375,
0.09866999834775925,
-0.0007704846211709082,
-0.03997170552611351,
0.10998106002807617,
0.020805958658456802,
-0.023959051817655563,
-0.002430015243589878,
0.0068419757299125195,
-0.002302940934896469,
-0.06542186439037323,
-0.041625138372182846,
0.019538529217243195,
0.01948736049234867,
-0.09420013427734375,
0.01786864921450615,
-0.022675275802612305,
0.041648950427770615,
-0.012111430987715721,
-0.05697441101074219,
-0.11430807411670685,
5.386841350571339e-33,
0.05138428136706352,
0.04629095643758774,
0.0010106678819283843,
-0.00731489947065711,
0.06646811962127686,
0.02399444207549095,
0.003974320832639933,
0.04253413528203964,
-0.02388014644384384,
0.030460670590400696,
-0.011875169351696968,
0.02029646560549736,
-0.01761881448328495,
0.038956716656684875,
0.07296694070100784,
-0.010353515855967999,
-0.06433483958244324,
0.023735199123620987,
-0.045176997780799866,
-0.055435653775930405,
0.0749787762761116,
0.0017559124389663339,
-0.003139826701954007,
-0.0016415503341704607,
0.025387054309248924,
0.047329433262348175,
0.02109532803297043,
-0.01105587650090456,
-0.06910181045532227,
0.04101526364684105,
-0.08321572095155716,
-0.0813940018415451,
0.015086679719388485,
-0.06212104111909866,
0.08831489831209183,
-0.02149989828467369,
-0.016439856961369514,
-0.03691745176911354,
0.016277160495519638,
-0.0151141919195652,
0.007155362982302904,
0.023188672959804535,
0.08218536525964737,
-0.10928444564342499,
-0.12560969591140747,
-0.05359181389212608,
0.07174156606197357,
0.0006864643655717373,
-0.09389255195856094,
-0.03826271370053291,
0.05923423916101456,
0.013177885673940182,
-0.13240136206150055,
0.031782858073711395,
0.026339884847402573,
-0.039621710777282715,
0.07246315479278564,
0.011044854298233986,
0.024256963282823563,
0.06101159751415253,
-0.01607627235352993,
0.01793837547302246,
-0.045404914766550064,
0.051661763340234756,
-0.01531936414539814,
0.015468496829271317,
-0.04857268184423447,
0.0676831305027008,
-0.02310628443956375,
0.047699060291051865,
-0.06847991049289703,
-0.010475791990756989,
0.03967071324586868,
-0.00005457336737890728,
0.09178730100393295,
-0.04330435395240784,
0.05215545743703842,
-0.11497735977172852,
0.016440683975815773,
0.03282634913921356,
0.03346990421414375,
0.004487166181206703,
0.0355304516851902,
-0.03126022592186928,
-0.0772925615310669,
-0.09904438257217407,
0.014300031587481499,
-0.04289045184850693,
-0.05078797787427902,
-0.05510146543383598,
-0.04447280988097191,
-0.04300545156002045,
0.07303594052791595,
-0.016650643199682236,
-0.03725850582122803,
-5.9340699703936075e-33,
-0.06766911596059799,
0.04932643473148346,
-0.025377485901117325,
0.12364855408668518,
-0.028985869139432907,
-0.09873820841312408,
0.06291081011295319,
0.059186942875385284,
0.03175323083996773,
-0.10314815491437912,
0.03482991084456444,
0.020631322637200356,
0.036706775426864624,
0.0407700315117836,
0.03904079273343086,
-0.03389006480574608,
-0.042844370007514954,
-0.010814289562404156,
-0.02044159360229969,
0.056739356368780136,
0.0022217596415430307,
0.019694769755005836,
-0.026713918894529343,
-0.04125896468758583,
-0.01525141205638647,
0.0285035353153944,
-0.03621739521622658,
0.013834034092724323,
-0.006885961163789034,
0.06927875429391861,
-0.030956029891967773,
-0.005949363578110933,
-0.0050456831231713295,
-0.053066279739141464,
-0.03182558715343475,
-0.006587565410882235,
0.10772848129272461,
0.06506281346082687,
-0.012993648648262024,
0.03970504179596901,
0.11496151983737946,
-0.07231290638446808,
-0.04492101073265076,
0.025027543306350708,
-0.0005871545872651041,
0.057524047791957855,
-0.06693647056818008,
-0.009307188913226128,
0.034805163741111755,
-0.005213776603341103,
0.023169973865151405,
0.06033395975828171,
0.05466031655669212,
0.05514101684093475,
0.0009284330299124122,
-0.1455799639225006,
-0.10781217366456985,
-0.02857482060790062,
0.00726421270519495,
-0.05381770059466362,
-0.056780554354190826,
-0.01885262131690979,
-0.00029725671629421413,
-0.08525210618972778,
0.007299109827727079,
-0.03539150208234787,
-0.06987089663743973,
0.002322995103895664,
-0.02149195782840252,
0.008349159732460976,
-0.0366373248398304,
-0.0319502167403698,
-0.05851207301020622,
-0.022972162812948227,
-0.1379275619983673,
-0.015402189455926418,
-0.02121448703110218,
-0.08003458380699158,
0.024226240813732147,
0.0036861489061266184,
0.03394592925906181,
0.08160199224948883,
-0.006379556376487017,
0.016249215230345726,
0.09766574949026108,
0.06346993893384933,
0.01786228083074093,
0.08162149041891098,
-0.028144221752882004,
0.02685203030705452,
-0.07765121012926102,
0.09833933413028717,
0.01368936151266098,
0.04773907735943794,
-0.058491259813308716,
-4.449001522743856e-8,
-0.040185753256082535,
0.01807956025004387,
-0.02009037509560585,
0.044080402702093124,
0.046083711087703705,
0.013859845697879791,
-0.0307371374219656,
0.07198537886142731,
0.020003441721200943,
-0.005649293307214975,
0.06412038207054138,
-0.013490553945302963,
-0.06974761933088303,
-0.0566607341170311,
0.051395807415246964,
0.07785031944513321,
0.018531857058405876,
-0.07228792458772659,
0.014542804099619389,
-0.09398024529218674,
-0.03600119426846504,
0.049160249531269073,
0.019587716087698936,
-0.041757531464099884,
0.02183273620903492,
-0.06521550565958023,
0.05206233635544777,
0.10774040222167969,
0.1319977343082428,
-0.0184804555028677,
-0.0006719503435306251,
0.00428828364238143,
-0.021822812035679817,
0.03979288414120674,
0.05532963573932648,
0.009654774330556393,
-0.017254596576094627,
0.08561644703149796,
0.059781529009342194,
-0.04391374811530113,
-0.060117267072200775,
0.05697789788246155,
-0.08617185056209564,
0.03980909287929535,
0.01581902988255024,
-0.049964919686317444,
-0.016678808256983757,
0.00615151459351182,
-0.024917304515838623,
-0.0022929206024855375,
-0.025298789143562317,
0.05498647689819336,
-0.06520124524831772,
0.08264099806547165,
0.04500683769583702,
0.05021039396524429,
-0.0632561519742012,
-0.042345814406871796,
-0.06694982200860977,
0.07637821137905121,
0.012672965414822102,
-0.09381622821092606,
-0.05853584036231041,
0.0008125849417410791
] | https://github.com/nolanoorg/spectrasuite | 46 | 0 | 0 | 0 |
Generating Freeform Endoskeletal Robots | https://openreview.net/forum?id=awvJBtB2op | [
"Muhan Li",
"Lingji Kong",
"Sam Kriegman"
] | Spotlight | The automatic design of embodied agents (e.g. robots) has existed for 31 years and is experiencing a renaissance of interest in the literature. To date however, the field has remained narrowly focused on two kinds of anatomically simple robots: (1) fully rigid, jointed bodies; and (2) fully soft, jointless bodies. Here we bridge these two extremes with the open ended creation of terrestrial endoskeletal robots: deformable soft bodies that leverage jointed internal skeletons to move efficiently across land. Simultaneous de novo generation of external and internal structures is achieved by (i) modeling 3D endoskeletal body plans as integrated collections of elastic and rigid cells that directly attach to form soft tissues anchored to compound rigid bodies; (ii) encoding these discrete mechanical subsystems into a continuous yet coherent latent embedding; (iii) optimizing the sensorimotor coordination of each decoded design using model-free reinforcement learning; and (iv) navigating this smooth yet highly non-convex latent manifold using evolutionary strategies. This yields an endless stream of novel species of ``higher robots'' that, like all higher animals, harness the mechanical advantages of both elastic tissues and skeletal levers for terrestrial travel. It also provides a plug-and-play experimental platform for benchmarking evolutionary design and representation learning algorithms in complex hierarchical embodied systems. | co-design, agent design, robots, morphology, evolution, locomotion | We introduce the multiphysics simulation, de novo design, and universal control of endoskeletal robots with minimal assumptions about the robots’ morphology and behavior. | 11,292 | 2412.01036 | [
0.025606391951441765,
-0.06961953639984131,
0.0359550304710865,
-0.030473437160253525,
0.0031680543906986713,
-0.08384852856397629,
-0.06907374411821365,
-0.02241802215576172,
0.026472575962543488,
0.13933339715003967,
0.025119172409176826,
-0.06977816671133041,
0.005528639070689678,
0.07792698591947556,
0.01883075013756752,
0.026354793459177017,
-0.007088173180818558,
0.07725319266319275,
-0.056182611733675,
0.13540199398994446,
0.05425272881984711,
0.07200737297534943,
0.04216788336634636,
0.0004348779038991779,
-0.0468989834189415,
-0.05298943817615509,
0.024056069552898407,
-0.021089836955070496,
0.0038092376198619604,
-0.09398391097784042,
-0.018454115837812424,
-0.06745400279760361,
-0.05528082698583603,
-0.10753513127565384,
0.007471016142517328,
0.07814471423625946,
-0.019872061908245087,
0.03439246863126755,
-0.06764961034059525,
-0.021641183644533157,
-0.010695414617657661,
0.02425672486424446,
0.03462638705968857,
0.042032357305288315,
0.05132411792874336,
0.05951379984617233,
-0.054117921739816666,
-0.056934937834739685,
0.0259525366127491,
-0.019489835947752,
-0.025854678824543953,
-0.03529557213187218,
0.050068240612745285,
0.031317226588726044,
-0.025416700169444084,
-0.022494526579976082,
-0.08475523442029953,
-0.02648078091442585,
-0.0036443453282117844,
-0.08933236449956894,
0.11234333366155624,
-0.06766151636838913,
0.08961424231529236,
-0.06160284951329231,
-0.02576887048780918,
-0.07156103849411011,
0.0001620898547116667,
0.01661473885178566,
-0.05829421430826187,
0.004849888384342194,
-0.02412484772503376,
-0.05435450002551079,
0.0011974910739809275,
-0.016784416511654854,
0.01961127109825611,
-0.024829279631376266,
0.01805282197892666,
0.08314590156078339,
0.03920911252498627,
0.02979026548564434,
0.0036162782926112413,
-0.023288805037736893,
0.014944999478757381,
0.002720962278544903,
-0.025943608954548836,
0.06131783500313759,
-0.03065062128007412,
0.08389440178871155,
-0.03924413025379181,
0.0006588543765246868,
0.023168962448835373,
-0.022973554208874702,
0.010322698391973972,
-0.06287072598934174,
0.010621550492942333,
-0.005435256287455559,
-0.024108709767460823,
0.02719859965145588,
0.03986538201570511,
0.006892720703035593,
0.010936466045677662,
-0.03767871484160423,
0.10030632466077805,
0.08173064887523651,
0.026547610759735107,
-0.10075509548187256,
0.011416565626859665,
-0.015346663072705269,
0.07761900126934052,
-0.0018915660912171006,
-0.02033752202987671,
-0.06037680804729462,
0.04211629554629326,
0.09764958173036575,
0.0015904100146144629,
0.005311976186931133,
-0.07041849941015244,
-0.002043954562395811,
0.09339005500078201,
0.047271449118852615,
-0.0031848822254687548,
-0.011308671906590462,
0.00530762504786253,
0.03329813480377197,
0.041203491389751434,
0.00032331087277270854,
-0.08380147814750671,
1.746415426627364e-33,
-0.0011945865117013454,
0.02790704555809498,
0.0052216569893062115,
-0.012314924038946629,
0.00810650922358036,
-0.07407350838184357,
-0.00011603914754232392,
0.033224448561668396,
-0.05048911273479462,
0.013603858649730682,
-0.11293400079011917,
0.039406586438417435,
0.011165685951709747,
0.1046016588807106,
0.04456419125199318,
-0.06271246075630188,
-0.04512843117117882,
0.03754272684454918,
0.014694618992507458,
0.004467251244932413,
0.055815842002630234,
0.025495879352092743,
0.03702158480882645,
-0.0492439866065979,
0.04334909841418266,
0.021046169102191925,
-0.009891286492347717,
-0.05689823627471924,
-0.0873783528804779,
-0.01667993888258934,
-0.09303423017263412,
0.004860964138060808,
-0.0045466432347893715,
0.09356267750263214,
0.03536858409643173,
0.012726628221571445,
0.004424940329045057,
-0.07556743174791336,
-0.031205488368868828,
-0.022684846073389053,
0.09204775094985962,
-0.002667040564119816,
0.00589066743850708,
0.0165589340031147,
-0.005769430659711361,
-0.03911140188574791,
0.022931070998311043,
0.06873738020658493,
-0.014990848489105701,
-0.05315601825714111,
0.019076667726039886,
0.10321852564811707,
0.00821950938552618,
-0.096064493060112,
-0.026595400646328926,
0.004815890919417143,
-0.06275918334722519,
-0.02718711644411087,
-0.03921795263886452,
0.046377018094062805,
-0.036519311368465424,
0.009820235893130302,
0.022616766393184662,
0.006309697404503822,
0.004605771973729134,
0.018060900270938873,
0.017642494291067123,
0.08980429172515869,
0.12219404429197311,
-0.03702772781252861,
0.01397786382585764,
-0.014340730383992195,
0.00429460359737277,
-0.03691268712282181,
0.0562373623251915,
0.03351346775889397,
0.06708463281393051,
-0.07214517146348953,
-0.053272366523742676,
0.004265515133738518,
0.008482152596116066,
0.1344718486070633,
-0.09081213921308517,
0.04332353547215462,
0.028123172000050545,
-0.01434152852743864,
0.02639089897274971,
-0.011193417012691498,
0.05090104043483734,
0.028392912819981575,
-0.11886591464281082,
-0.1565595120191574,
-0.08542898297309875,
0.03763116896152496,
-0.030360260978341103,
-2.768094504934632e-33,
0.040457092225551605,
-0.04166591539978981,
-0.008787917904555798,
0.011292438954114914,
0.030975619331002235,
-0.009127303957939148,
-0.04727097973227501,
-0.0714445561170578,
-0.007631349842995405,
-0.04230598732829094,
0.008821028284728527,
0.029141921550035477,
0.04155870899558067,
0.04880958050489426,
0.09564385563135147,
0.02488601952791214,
-0.04721922427415848,
-0.059873394668102264,
0.03170079365372658,
-0.08462903648614883,
0.054028503596782684,
0.05092041566967964,
-0.0016208235174417496,
0.03402089700102806,
0.060637205839157104,
0.04917670041322708,
-0.07749226689338684,
0.032585833221673965,
-0.03898387774825096,
0.05236563831567764,
-0.0715457871556282,
-0.03201091289520264,
0.010927866213023663,
0.05607270076870918,
0.006812988314777613,
0.03478063642978668,
-0.0699314996600151,
0.033962782472372055,
0.08457493036985397,
-0.059601351618766785,
0.00009885185863822699,
-0.053917303681373596,
-0.02971791848540306,
0.09608520567417145,
0.03152850270271301,
-0.08356018364429474,
-0.07258662581443787,
-0.04611510783433914,
-0.058608680963516235,
0.040142495185136795,
-0.04913121461868286,
0.05275886505842209,
0.006606139242649078,
-0.1272388994693756,
-0.011097072623670101,
-0.0446048304438591,
-0.013862389139831066,
0.019456228241324425,
0.07996036857366562,
-0.015457634814083576,
-0.06385918706655502,
0.006107795983552933,
-0.03470609337091446,
0.05909132584929466,
-0.023342173546552658,
0.055424414575099945,
-0.004828212782740593,
-0.07170401513576508,
-0.05742610618472099,
0.008609123528003693,
-0.01658707484602928,
-0.020079191774129868,
0.03549410402774811,
0.052539873868227005,
0.09173382818698883,
0.010980842635035515,
0.010011748410761356,
-0.07200417667627335,
0.02379576861858368,
-0.06565989553928375,
-0.035130083560943604,
-0.03293424844741821,
0.04464340955018997,
0.08675126731395721,
0.00692312978208065,
-0.007559539284557104,
-0.056137192994356155,
0.0606958232820034,
-0.045089688152074814,
0.04358764737844467,
0.06193721294403076,
-0.0008816228364594281,
0.04106587544083595,
0.06457668542861938,
-0.04255196452140808,
-4.692962818353408e-8,
0.004059251397848129,
0.019152807071805,
0.05507797375321388,
-0.058298371732234955,
-0.0007585565908811986,
0.06611599028110504,
-0.0085686556994915,
-0.014728636480867863,
-0.0538468062877655,
-0.06065740808844566,
0.03196486458182335,
-0.05181329697370529,
0.06075297296047211,
0.06559876352548599,
0.02309378795325756,
0.10265311598777771,
0.033440832048654556,
-0.027216969057917595,
-0.004123174585402012,
-0.0948793813586235,
-0.0434415377676487,
0.018327081575989723,
-0.0255447831004858,
-0.025219133123755455,
0.009485148824751377,
-0.05009151250123978,
-0.16810467839241028,
0.01625952310860157,
0.008646802976727486,
0.05302731320261955,
0.029870405793190002,
0.03170057386159897,
0.004270338453352451,
0.02414662577211857,
-0.05178319662809372,
0.032434623688459396,
0.011385536752641201,
-0.045217134058475494,
-0.08431199938058853,
0.028300443664193153,
0.057639360427856445,
0.12539933621883392,
-0.025750089436769485,
-0.04393090680241585,
0.020082298666238785,
-0.03421856835484505,
-0.027883002534508705,
-0.06600207835435867,
0.07245947420597076,
0.00405425950884819,
-0.02794169820845127,
-0.06334259361028671,
0.002431760774925351,
0.025120317935943604,
0.01157348696142435,
0.045841511338949203,
0.005136759020388126,
-0.06054859980940819,
0.03531979024410248,
0.11070241034030914,
-0.007841466926038265,
0.013910507783293724,
0.006188408005982637,
-0.033468831330537796
] | 0 | 0 | 0 | 0 |
|
Understanding Factual Recall in Transformers via Associative Memories | https://openreview.net/forum?id=hwSmPOAmhk | [
"Eshaan Nichani",
"Jason D. Lee",
"Alberto Bietti"
] | Spotlight | Large language models have demonstrated an impressive ability to perform factual recall. Prior work has found that transformers trained on factual recall tasks can store information at a rate proportional to their parameter count. In our work, we show that shallow transformers can use a combination of associative memories to obtain such near optimal storage capacity. We begin by proving that the storage capacities of both linear and MLP associative memories scale linearly with parameter count. We next introduce a synthetic factual recall task, and prove that a transformer with a single layer of self-attention followed by an MLP can obtain 100\% accuracy on the task whenever either the total number of self-attention parameters or MLP parameters scales (up to log factors) linearly with the number of facts. In particular, the transformer can trade off between using the value matrices or the MLP as an associative memory to store the dataset of facts. We complement these expressivity results with an analysis of the gradient flow trajectory of a simplified linear attention model trained on our factual recall task, where we show that the model exhibits sequential learning behavior. | transformers, associative memories, factual recall, storage capacity, training dynamics | null | 11,244 | 2412.06538 | [
-0.04826314002275467,
-0.07631666213274002,
-0.07164732366800308,
0.044341910630464554,
-0.009629787877202034,
0.049133580178022385,
0.007493080571293831,
0.03859908506274223,
0.06294477730989456,
-0.012467504478991032,
0.013371102511882782,
0.0075476448982954025,
0.03238700330257416,
0.07753031700849533,
0.0030913844238966703,
0.0795406773686409,
0.04486284777522087,
0.11430012434720993,
-0.11985655874013901,
-0.0486195906996727,
0.03686337172985077,
0.0008062227279879153,
0.07476777583360672,
0.0455758161842823,
0.03179469704627991,
0.015152734704315662,
-0.008262498304247856,
-0.08124291151762009,
-0.00018523199832998216,
-0.05426714941859245,
0.04402201250195503,
0.031831610947847366,
-0.04620364308357239,
0.07653658092021942,
-0.0057846722193062305,
0.06370948255062103,
-0.08126616477966309,
0.006713084410876036,
-0.007777708116918802,
-0.039426539093256,
0.06784410774707794,
-0.0069327326491475105,
-0.0009400439448654652,
0.038238827139139175,
0.06260303407907486,
0.05122032016515732,
0.038111619651317596,
-0.024335810914635658,
0.014440428465604782,
0.018786843866109848,
-0.02106194570660591,
0.03196640685200691,
-0.058514196425676346,
0.06790217012166977,
-0.06600049883127213,
0.043686509132385254,
0.05344175919890404,
0.012337743304669857,
-0.04438163340091705,
-0.014261976815760136,
-0.0681193470954895,
-0.07971787452697754,
-0.0056142620742321014,
-0.06083889678120613,
-0.1142454668879509,
-0.026523642241954803,
0.06179320067167282,
-0.0009158580214716494,
-0.05177759379148483,
0.0375308021903038,
-0.04529969394207001,
0.06697408854961395,
-0.035258110612630844,
-0.012345986440777779,
0.07496138662099838,
-0.019049350172281265,
0.0257912278175354,
-0.0921887680888176,
0.043404292315244675,
0.011125019751489162,
0.005025774240493774,
0.03466623276472092,
0.011350234039127827,
0.028839660808444023,
0.0841563493013382,
-0.05205581709742546,
0.04943329468369484,
0.024128539487719536,
-0.015570759773254395,
0.015412923879921436,
-0.050218429416418076,
-0.03457717224955559,
0.03722640126943588,
-0.04494817927479744,
0.0708928182721138,
0.03319317847490311,
-0.010061519220471382,
-0.032377999275922775,
-0.014451111666858196,
0.008449623361229897,
0.026721330359578133,
0.04938691854476929,
-0.009372975677251816,
0.020979082211852074,
-0.015523821115493774,
-0.06748990714550018,
0.006767027545720339,
0.06830503791570663,
0.00775877432897687,
-0.0782318264245987,
0.04865025356411934,
-0.004917727317661047,
0.04685274139046669,
-0.0036721420474350452,
0.04336588457226753,
-0.03948022425174713,
0.009403788484632969,
-0.03362806513905525,
-0.04979192093014717,
0.0204638559371233,
0.012626292183995247,
0.023872075602412224,
0.023075081408023834,
0.08448371291160583,
0.008744820952415466,
-0.06548010557889938,
0.003617919748649001,
1.2360238610239888e-33,
0.0009619694319553673,
0.06910311430692673,
0.01823064684867859,
0.035599734634160995,
-0.02535010688006878,
0.01895187795162201,
-0.00874151848256588,
0.05480034649372101,
-0.009003054350614548,
-0.017025211825966835,
-0.0510612390935421,
0.06112737953662872,
-0.08116992563009262,
0.09187432378530502,
-0.026969995349645615,
0.021183520555496216,
-0.07616861909627914,
0.04206673055887222,
-0.02327672392129898,
-0.06518847495317459,
0.03699919953942299,
0.04608495160937309,
0.03983785957098007,
-0.10966698825359344,
-0.029493942856788635,
-0.02751992829144001,
0.09560340642929077,
-0.00982714258134365,
-0.06919687986373901,
-0.02100358158349991,
-0.08118244260549545,
-0.0025333024095743895,
-0.017148077487945557,
0.006483588833361864,
0.045387256890535355,
-0.03260276839137077,
0.005268170032650232,
-0.05935262143611908,
0.06090563163161278,
-0.05095209181308746,
-0.01650223322212696,
0.08149440586566925,
-0.022294480353593826,
0.018081482499837875,
-0.11391941457986832,
-0.11674292385578156,
0.008737717755138874,
0.029749037697911263,
-0.028858065605163574,
-0.06626192480325699,
0.03414466232061386,
-0.009885518811643124,
-0.09429828077554703,
-0.021477624773979187,
0.04491876810789108,
-0.049018871039152145,
0.005416070111095905,
0.06934256106615067,
0.04680093005299568,
0.10542930662631989,
0.035598576068878174,
0.0737014189362526,
0.014403841458261013,
0.08823370933532715,
0.02271190658211708,
0.02148149162530899,
0.001890425686724484,
0.15714885294437408,
0.07582416385412216,
-0.01108493097126484,
0.007151376456022263,
-0.022212550044059753,
-0.028781944885849953,
-0.07345718145370483,
0.015547189861536026,
0.010532076470553875,
0.08394649624824524,
-0.08581702411174774,
-0.048017412424087524,
0.016383469104766846,
0.028255831450223923,
-0.06818857789039612,
0.01589912921190262,
0.008451313711702824,
-0.06208066642284393,
-0.021961571648716927,
0.06865008175373077,
-0.09320367872714996,
-0.047832638025283813,
-0.001422895467840135,
-0.026918714866042137,
-0.02948804572224617,
0.002424986334517598,
-0.013444170355796814,
-0.02856634184718132,
-1.7631171809716333e-33,
0.007559351623058319,
0.009812393225729465,
-0.07544251531362534,
0.093717060983181,
-0.03727519512176514,
-0.09452575445175171,
0.01680450327694416,
0.015515008941292763,
-0.05155494436621666,
-0.0018508529756218195,
-0.08217429369688034,
-0.05053943395614624,
0.028476718813180923,
-0.023472819477319717,
0.07635804265737534,
-0.06592906266450882,
-0.023892760276794434,
-0.010581408627331257,
-0.052008006721735,
0.008230064064264297,
-0.06590034067630768,
0.10038281232118607,
-0.07808288931846619,
-0.011753089725971222,
-0.11151520907878876,
0.012644680216908455,
-0.0906519889831543,
0.042795293033123016,
0.006517542060464621,
-0.011658570729196072,
-0.03403325751423836,
-0.011175266467034817,
-0.0031285507138818502,
-0.004607161041349173,
-0.04530831798911095,
0.03873120993375778,
0.07439640909433365,
-0.0302132535725832,
-0.04370527341961861,
0.069919653236866,
0.028269248083233833,
-0.06975159049034119,
-0.07180885225534439,
-0.013859120197594166,
-0.0468793623149395,
-0.0784103274345398,
-0.03908734768629074,
-0.025846127420663834,
0.051553916186094284,
0.048776768147945404,
-0.049106910824775696,
0.052443865686655045,
-0.062149520963430405,
0.04949641972780228,
-0.023128552362322807,
-0.03865550458431244,
-0.07449744641780853,
-0.04769032448530197,
0.08713260293006897,
-0.034874334931373596,
-0.05109519138932228,
0.017858043313026428,
0.008394814096391201,
-0.13384079933166504,
-0.004366485867649317,
-0.012463117018342018,
0.023380661383271217,
-0.06067425757646561,
0.0006671668379567564,
0.07624148577451706,
0.05154450982809067,
-0.0038777857553213835,
0.0653218999505043,
-0.04548925906419754,
0.017711985856294632,
-0.047905053943395615,
0.021480312570929527,
-0.008204933255910873,
-0.011751001700758934,
-0.013806325383484364,
-0.02419854886829853,
-0.0010236684465780854,
0.011579001322388649,
0.014376360923051834,
0.0024606327060610056,
0.017898300662636757,
0.07309926301240921,
0.05051824823021889,
-0.010816207155585289,
0.026233108714222908,
-0.010534719564020634,
0.010597268119454384,
-0.023930354043841362,
0.01277474220842123,
-0.03666563704609871,
-4.056786906403431e-8,
-0.12058241665363312,
0.03683944419026375,
0.048193298280239105,
0.04398678243160248,
0.04617363587021828,
-0.06260062754154205,
0.0044997925870120525,
0.13423402607440948,
-0.011091076768934727,
0.014041808433830738,
0.1004495918750763,
-0.02353941835463047,
-0.0757884606719017,
-0.0766594335436821,
0.024182550609111786,
0.06681471318006516,
0.06784648448228836,
-0.006427541375160217,
0.04139367491006851,
-0.023938674479722977,
0.013681275770068169,
0.0018946211785078049,
0.014367066323757172,
-0.02198307029902935,
0.0125164445489645,
-0.05592341721057892,
-0.0036606555804610252,
0.10598673671483994,
0.04995923861861229,
0.04778659716248512,
0.04514598101377487,
-0.0038936715573072433,
0.009709218516945839,
-0.07956653088331223,
0.012461110018193722,
0.09362095594406128,
0.10551398247480392,
0.01893034763634205,
0.050220467150211334,
-0.09671348333358765,
0.021780196577310562,
0.07091789692640305,
-0.05344649776816368,
0.11060964316129684,
0.03241109102964401,
0.01031262893229723,
-0.059040043503046036,
-0.11564525216817856,
-0.012662962079048157,
-0.0023664480540901423,
0.06016862019896507,
-0.010975155979394913,
-0.009657897986471653,
0.11949877440929413,
0.03570358455181122,
-0.026833966374397278,
0.0012129496317356825,
-0.074576735496521,
-0.031746815890073776,
0.05933060497045517,
0.011411486193537712,
0.08427882939577103,
-0.03878604620695114,
-0.01991291716694832
] | 0 | 0 | 0 | 0 |
|
Nesterov acceleration in benignly non-convex landscapes | https://openreview.net/forum?id=YwJkv2YqBq | [
"Kanan Gupta",
"Stephan Wojtowytsch"
] | Spotlight | While momentum-based optimization algorithms are commonly used in the notoriously non-convex optimization problems of deep learning, their analysis has historically been restricted to the convex and strongly convex setting. In this article, we partially close this gap between theory and practice and demonstrate that virtually identical guarantees can be obtained in optimization problems with a `benign' non-convexity. We show that these weaker geometric assumptions are well justified in overparametrized deep learning, at least locally. Variations of this result are obtained for a continuous time model of Nesterov's accelerated gradient descent algorithm (NAG), the classical discrete time version of NAG, and versions of NAG with stochastic gradient estimates with purely additive noise and with noise that exhibits both additive and multiplicative scaling. | Nonconvex optimization, stochastic optimization, stochastic acceleration, smooth convex optimization, deep learning, accelerated gradient descent | null | 11,191 | 2410.08395 | [
-0.056150127202272415,
-0.060000400990247726,
0.09365424513816833,
-0.023871010169386864,
0.032676439732313156,
-0.030925137922167778,
-0.02615220472216606,
-0.0626494437456131,
-0.003837844356894493,
-0.010015218518674374,
-0.034851737320423126,
0.031258612871170044,
0.013650022447109222,
0.004437597934156656,
-0.062179528176784515,
-0.02645142376422882,
0.03969082236289978,
0.09291761368513107,
-0.054574303328990936,
-0.015133596956729889,
-0.037433817982673645,
0.07651963084936142,
0.05818420276045799,
0.023076370358467102,
0.08437906950712204,
-0.0657878890633583,
0.002483794465661049,
0.02551215887069702,
0.02843167632818222,
-0.005358261521905661,
0.025846723467111588,
-0.055854085832834244,
0.011927392333745956,
-0.00510766077786684,
0.01867014169692993,
0.07167770713567734,
-0.07601572573184967,
0.027384592220187187,
-0.043447282165288925,
0.0065856617875397205,
-0.0041066743433475494,
0.014135251753032207,
0.039886798709630966,
-0.009114240296185017,
0.07745219767093658,
-0.015007716603577137,
0.03798312693834305,
-0.1043957993388176,
0.016478588804602623,
-0.030124150216579437,
-0.06545447558164597,
-0.03320982679724693,
-0.0337647944688797,
-0.011940599419176579,
-0.021492397412657738,
0.07079797983169556,
-0.019851965829730034,
-0.0009213322773575783,
0.017768923193216324,
0.006098985206335783,
0.04362858459353447,
-0.09635618329048157,
0.013080456294119358,
0.014173313975334167,
0.02117590606212616,
0.03401327133178711,
0.03928934782743454,
0.025207040831446648,
0.018501095473766327,
0.0795530155301094,
0.008394066244363785,
0.05828821659088135,
-0.03969307988882065,
0.01854921691119671,
-0.005555330775678158,
-0.04349888488650322,
0.07392749190330505,
0.000869731476996094,
-0.008568637073040009,
-0.06239936500787735,
0.048892248421907425,
0.016993500292301178,
0.03091597743332386,
-0.07068987935781479,
0.10297128558158875,
-0.03436571732163429,
-0.008755591697990894,
-0.006013131234794855,
-0.0018353962805122137,
-0.06917539983987808,
0.017189310863614082,
-0.023199651390314102,
-0.1127053052186966,
0.016196344047784805,
-0.008321656845510006,
-0.07190874218940735,
-0.01712329313158989,
-0.07635226845741272,
-0.05695166438817978,
0.11677236109972,
-0.006680505815893412,
0.018961813300848007,
0.013203354552388191,
0.031610265374183655,
0.05698094889521599,
-0.012015504762530327,
0.09615331888198853,
0.06320276111364365,
-0.0122310109436512,
0.04131525754928589,
0.09444161504507065,
0.030503399670124054,
0.0519602969288826,
0.03879973664879799,
-0.02053789049386978,
-0.04222414270043373,
0.02101256512105465,
-0.05062327906489372,
0.0359155610203743,
0.07100527733564377,
-0.12841933965682983,
-0.06526759266853333,
-0.0072967661544680595,
0.03266541287302971,
-0.01491797436028719,
0.02390875853598118,
-0.022974330931901932,
2.2331343727151017e-33,
-0.06995761394500732,
-0.02022452838718891,
0.03873362019658089,
-0.10876623541116714,
0.02383319102227688,
0.010508081875741482,
-0.03519916906952858,
-0.06697183847427368,
-0.0022350254002958536,
0.022522583603858948,
-0.03610030934214592,
-0.03798897564411163,
-0.09882090240716934,
0.0509590283036232,
0.08041319251060486,
-0.03425304591655731,
-0.08595889806747437,
0.008798087015748024,
0.0631486028432846,
-0.1147833839058876,
0.04657600820064545,
-0.06388811022043228,
-0.1222233772277832,
0.03135206922888756,
-0.005045819096267223,
-0.009727705270051956,
0.015945015475153923,
0.0333239883184433,
-0.021773317828774452,
0.035163819789886475,
-0.08309584110975266,
0.02845998853445053,
-0.03503132611513138,
0.06924527138471603,
0.019171809777617455,
0.09790987521409988,
0.003416175954043865,
-0.02973293513059616,
-0.0007298230193555355,
-0.00029334845021367073,
-0.03430163860321045,
0.03360728546977043,
0.05835464969277382,
-0.03816048800945282,
-0.12982463836669922,
-0.07881896197795868,
0.04984978958964348,
0.016972824931144714,
-0.049265798181295395,
-0.06403449922800064,
-0.01584651693701744,
0.011124697513878345,
-0.035319410264492035,
-0.05614422634243965,
0.007552322465926409,
0.008835555985569954,
0.02724209800362587,
0.06045389175415039,
0.004512832034379244,
0.07127819955348969,
0.03389817848801613,
-0.026093805208802223,
0.03829232230782509,
-0.04980434104800224,
-0.04422713816165924,
0.06603100895881653,
0.010144585743546486,
0.04972399026155472,
0.06622155755758286,
-0.084665946662426,
0.03834168240427971,
0.01666523702442646,
0.01553873810917139,
-0.0666792020201683,
0.05999431386590004,
-0.0021742014214396477,
0.026084428653120995,
-0.056947011500597,
0.022284094244241714,
0.05724041536450386,
-0.02364470064640045,
0.026366500183939934,
-0.033890530467033386,
-0.003732019569724798,
-0.001981589011847973,
-0.04464637488126755,
0.031211312860250473,
-0.014393804594874382,
0.018637174740433693,
0.03420380502939224,
-0.13357943296432495,
0.028469130396842957,
-0.012975084595382214,
0.016761912032961845,
-0.030668359249830246,
-3.0381815971093304e-33,
-0.026488520205020905,
0.009927728213369846,
-0.059729430824518204,
0.07698070257902145,
0.0013337214477360249,
-0.02775738760828972,
-0.04843852296471596,
-0.03700067102909088,
-0.10020633786916733,
-0.12447446584701538,
0.012126200832426548,
0.029328521341085434,
0.040547169744968414,
-0.06638412177562714,
0.010491396300494671,
-0.05915658921003342,
0.02231466956436634,
0.034541789442300797,
0.020668350160121918,
-0.015826091170310974,
0.05684872716665268,
0.034485917538404465,
-0.034724459052085876,
0.05903497710824013,
0.025388427078723907,
0.012522374279797077,
-0.09507805854082108,
0.13250133395195007,
-0.03180587664246559,
-0.008421296253800392,
-0.03383630886673927,
0.047791510820388794,
-0.06815893948078156,
-0.01750789023935795,
-0.047733720391988754,
0.08913706988096237,
-0.0053513916209340096,
0.03724167123436928,
-0.02791008912026882,
0.041148122400045395,
0.09855866432189941,
0.020667903125286102,
-0.009521705098450184,
0.004727077670395374,
0.07495887577533722,
0.017323646694421768,
0.08477772027254105,
-0.04458093270659447,
0.025874456390738487,
0.0019953909795731306,
0.017791491001844406,
0.011772405356168747,
-0.008187147788703442,
0.05404305085539818,
-0.05216227471828461,
0.007546965964138508,
-0.02637927047908306,
-0.08434366434812546,
0.1031845211982727,
-0.013201244175434113,
-0.07666247338056564,
0.0007259951089508832,
-0.08823754638433456,
0.03550302982330322,
-0.0166255421936512,
-0.01214217022061348,
-0.032329410314559937,
-0.009991761296987534,
0.026206862181425095,
0.031567253172397614,
-0.013391510583460331,
-0.01048223301768303,
0.003816169686615467,
0.011076977476477623,
-0.12644724547863007,
0.0272955521941185,
0.122165746986866,
0.028008216992020607,
0.014395486563444138,
-0.14764302968978882,
0.12390142679214478,
0.02835472673177719,
0.06681738048791885,
0.093141108751297,
0.01642272062599659,
0.008915916085243225,
0.013419101946055889,
0.004026469774544239,
0.02623303048312664,
-0.01018039882183075,
0.008460401557385921,
-0.013517448678612709,
-0.03215804696083069,
-0.059876374900341034,
-0.06937272846698761,
-4.645800899538699e-8,
-0.011167141608893871,
0.025294261053204536,
0.0019787831697613,
-0.07507757842540741,
0.10916679352521896,
0.01392588671296835,
0.012580743990838528,
-0.018828315660357475,
-0.0001624976866878569,
-0.025007043033838272,
0.08342432975769043,
-0.018594998866319656,
-0.047278545796871185,
-0.056500863283872604,
-0.07243967056274414,
0.10174638032913208,
-0.03772712126374245,
0.04293360561132431,
-0.00013812568795401603,
-0.010467877611517906,
-0.04019341990351677,
0.09862316399812698,
-0.034228261560201645,
0.0028013992123305798,
0.044162817299366,
-0.003173504024744034,
-0.030070584267377853,
0.0075221280567348,
0.07742100954055786,
-0.05150821805000305,
-0.074716717004776,
0.08773007988929749,
0.01644483208656311,
0.03245072811841965,
-0.03371702879667282,
0.07964841276407242,
0.025484168902039528,
0.05835963413119316,
-0.04417338967323303,
-0.09196532517671585,
-0.0642055869102478,
0.02192937396466732,
-0.004777631256729364,
-0.07128971070051193,
-0.06680797040462494,
0.07691934704780579,
-0.017227016389369965,
-0.013048077933490276,
-0.012989215552806854,
0.06828299909830093,
0.034221939742565155,
0.04627343267202377,
0.06624546647071838,
0.0859348475933075,
0.0745493471622467,
0.008505124598741531,
0.04905181750655174,
-0.09787646681070328,
-0.01177628617733717,
0.060088858008384705,
0.03760926052927971,
-0.006547350902110338,
-0.0646883174777031,
0.01271616667509079
] | 0 | 0 | 0 | 0 |
|
AnalogGenie: A Generative Engine for Automatic Discovery of Analog Circuit Topologies | https://openreview.net/forum?id=jCPak79Kev | [
"Jian Gao",
"Weidong Cao",
"Junyi Yang",
"Xuan Zhang"
] | Spotlight | The massive and large-scale design of foundational semiconductor integrated circuits (ICs) is crucial to sustaining the advancement of many emerging and future technologies, such as generative AI, 5G/6G, and quantum computing.
Excitingly, recent studies have shown the great capabilities of foundational models in expediting the design of digital ICs.
Yet, applying generative AI techniques to accelerate the design of analog ICs remains a significant challenge due to critical domain-specific issues, such as the lack of a comprehensive dataset and effective representation methods for analog circuits.
This paper proposes, $\textbf{AnalogGenie}$, a $\underline{\textbf{Gen}}$erat$\underline{\textbf{i}}$ve $\underline{\textbf{e}}$ngine for automatic design/discovery of $\underline{\textbf{Analog}}$ circuit topologies--the most challenging and creative task in the conventional manual design flow of analog ICs.
AnalogGenie addresses two key gaps in the field: building a foundational comprehensive dataset of analog circuit topology and developing a scalable sequence-based graph representation universal to analog circuits.
Experimental results show the remarkable generation performance of AnalogGenie in broadening the variety of analog ICs, increasing the number of devices within a single design, and discovering unseen circuit topologies far beyond any prior arts.
Our work paves the way to transform the longstanding time-consuming manual design flow of analog ICs to an automatic and massive manner powered by generative AI.
Our source code is available at https://github.com/xz-group/AnalogGenie. | Circuit Generation, Application of Generative Models, Electronic Design Automation | A versatile generative model capable of designing topologies for wide range of analog circuits. | 11,141 | 2503.00205 | [
-0.09900887310504913,
0.02159378118813038,
0.035588283091783524,
0.010709712281823158,
-0.037851735949516296,
-0.06975700706243515,
-0.02460283599793911,
0.0006256532506085932,
-0.08513911813497543,
0.0026573087088763714,
-0.08663449436426163,
-0.15984193980693817,
-0.0380697064101696,
-0.049040503799915314,
-0.00442005880177021,
0.02654244191944599,
0.006127134896814823,
0.04559529200196266,
0.06587975472211838,
-0.0777355208992958,
0.07044485211372375,
0.060632891952991486,
-0.015985149890184402,
-0.06619999557733536,
0.01936176046729088,
0.03954258933663368,
0.08989707380533218,
0.011461338959634304,
0.03595277667045593,
-0.06891951709985733,
0.04198964312672615,
0.06670979410409927,
0.003077693749219179,
-0.02443595789372921,
-0.04034587740898132,
0.05467865616083145,
0.03469695895910263,
0.025249473750591278,
0.08442112803459167,
-0.06635072827339172,
0.029097314924001694,
0.06505300849676132,
0.04490755870938301,
0.017338400706648827,
0.08700067549943924,
-0.022595025599002838,
-0.031834233552217484,
-0.05429703742265701,
-0.01019334141165018,
-0.1058054268360138,
0.045893605798482895,
0.01733044907450676,
0.004769803490489721,
0.05262109264731407,
0.03495572879910469,
0.0656312108039856,
0.07512621581554413,
-0.03988708183169365,
-0.017107263207435608,
0.03359982743859291,
-0.01114689465612173,
-0.07372037321329117,
-0.007653137668967247,
-0.04913773760199547,
-0.004445449449121952,
-0.011042174883186817,
0.05618634447455406,
0.03228895738720894,
0.06162690743803978,
-0.03196941688656807,
-0.03992296755313873,
0.04651951789855957,
-0.06290996074676514,
0.03409692645072937,
0.08629833161830902,
0.0897863507270813,
-0.0024272678419947624,
0.050529882311820984,
0.0832938477396965,
-0.08115798979997635,
-0.030938250944018364,
-0.01736554689705372,
-0.00045643976773135364,
0.03850438818335533,
0.08392813801765442,
0.04416138306260109,
-0.01087508536875248,
0.032126326113939285,
0.06131137162446976,
-0.05630422383546829,
-0.0874587669968605,
-0.04191676899790764,
0.028372179716825485,
0.0020208859350532293,
0.04242483153939247,
-0.012900627218186855,
0.06420265883207321,
-0.011919980868697166,
0.005405905656516552,
0.0707671046257019,
0.06606435030698776,
0.08394132554531097,
0.023520946502685547,
0.009919562377035618,
0.0042172446846961975,
-0.020136762410402298,
0.056518565863370895,
0.08249512314796448,
-0.03696051239967346,
-0.10972432792186737,
-0.018437502905726433,
0.018275734037160873,
0.06877264380455017,
-0.05938442423939705,
0.03182397410273552,
-0.023929089307785034,
-0.018754171207547188,
0.02563132345676422,
0.027908384799957275,
-0.028300780802965164,
-0.09157071262598038,
0.020669296383857727,
-0.07928960025310516,
0.1123027577996254,
0.0024993065744638443,
-0.021514486521482468,
-0.1055246964097023,
1.5658227601425976e-33,
-0.005801084451377392,
0.06647499650716782,
0.0416809543967247,
0.09678087383508682,
0.07237832248210907,
-0.020702047273516655,
0.014044320210814476,
0.014368497766554356,
0.01184967439621687,
0.04075414314866066,
0.006802443880587816,
-0.04113311320543289,
-0.08070199191570282,
0.12483759969472885,
0.053846728056669235,
-0.04016983509063721,
-0.10081017762422562,
-0.036122877150774,
-0.02365134097635746,
-0.08494680374860764,
0.05689283832907677,
0.004587425850331783,
-0.04007970169186592,
0.03722460940480232,
0.049909766763448715,
0.006957208272069693,
0.010845517739653587,
-0.033664070069789886,
-0.0894201397895813,
0.002776344073936343,
-0.066790871322155,
-0.028980979695916176,
0.01329964492470026,
-0.01457568071782589,
-0.04399729520082474,
0.017803743481636047,
0.04543953388929367,
-0.028255460783839226,
0.021837379783391953,
0.017416097223758698,
-0.008771728724241257,
-0.005115415435284376,
0.013859724625945091,
0.03674226254224777,
-0.04284245893359184,
-0.002156061353161931,
-0.0039092120714485645,
-0.008746794424951077,
0.07022909820079803,
-0.002399593126028776,
0.00839958991855383,
-0.05645820498466492,
0.022943969815969467,
0.009686334989964962,
0.021115360781550407,
-0.03695480898022652,
-0.033801984041929245,
0.010364735499024391,
0.09550206363201141,
0.11308984458446503,
-0.011188664473593235,
0.05456295609474182,
-0.09195270389318466,
0.10852792859077454,
-0.0995933935046196,
0.1274004727602005,
0.004088229965418577,
-0.050770606845617294,
0.0015589790418744087,
0.0626029297709465,
-0.005449662450700998,
0.03710450604557991,
-0.06511528789997101,
-0.08276788890361786,
0.08227314054965973,
-0.02379881776869297,
0.009307874366641045,
-0.05782831460237503,
0.006283617578446865,
0.00008577654079999775,
-0.04343666508793831,
0.03318735584616661,
-0.06232142448425293,
0.03474130481481552,
0.011459643952548504,
-0.06343299150466919,
-0.015994973480701447,
-0.005044050049036741,
-0.03529464825987816,
-0.08149302005767822,
-0.07603814452886581,
-0.057270657271146774,
0.03567911311984062,
0.02886503003537655,
-0.03801221400499344,
-3.427756955361145e-33,
-0.007673769723623991,
0.040921907871961594,
0.05919903516769409,
-0.010246915742754936,
-0.07623498141765594,
-0.042526647448539734,
-0.009698651731014252,
0.017236018553376198,
0.036960918456315994,
-0.07589820772409439,
0.032553501427173615,
0.06028981879353523,
0.02696942538022995,
-0.0023789959959685802,
-0.030299436300992966,
-0.0018788538873195648,
-0.07724789530038834,
-0.06006638705730438,
0.11857713758945465,
0.06186525523662567,
0.06819850206375122,
0.04006096348166466,
-0.11627232283353806,
-0.04023849219083786,
0.021933922544121742,
0.05484357476234436,
-0.04205016419291496,
0.07486535608768463,
0.01702435128390789,
-0.03705994412302971,
0.01983613707125187,
-0.015121272765100002,
0.028624221682548523,
0.0017668363871052861,
-0.07445204257965088,
0.0849609449505806,
0.1366516351699829,
0.03534311801195145,
-0.06261618435382843,
-0.004734031856060028,
-0.08320141583681107,
0.017071207985281944,
-0.08244495838880539,
0.002843308262526989,
0.0194461178034544,
-0.011999616399407387,
-0.025702401995658875,
0.08090271800756454,
0.05398133397102356,
0.04677863419055939,
0.007929385639727116,
-0.011473829858005047,
0.0018433618824929,
-0.06374587118625641,
-0.05227072536945343,
-0.05194729194045067,
0.0032456563785672188,
0.020172908902168274,
0.041438356041908264,
0.011691156774759293,
-0.028684906661510468,
-0.07166053354740143,
0.05263860523700714,
-0.06823546439409256,
-0.042745862156152725,
-0.02837706170976162,
-0.04517918452620506,
0.03564602509140968,
-0.004834271967411041,
-0.035091206431388855,
-0.08499085903167725,
0.00305622979067266,
0.028200024738907814,
-0.045320987701416016,
-0.0037617767229676247,
-0.0661250650882721,
0.00817090179771185,
0.009998350404202938,
0.06362547725439072,
-0.021844202652573586,
-0.03668050467967987,
0.0030818795785307884,
0.006578462664037943,
-0.056678593158721924,
-0.028782455250620842,
0.01721981354057789,
0.010392121970653534,
0.0263543538749218,
0.02226206846535206,
0.01743757165968418,
-0.09376414120197296,
0.07584647834300995,
-0.01322950515896082,
0.08071192353963852,
-0.06618554890155792,
-4.8859210011187315e-8,
0.10092423856258392,
0.017572440207004547,
0.03044038452208042,
-0.05829274281859398,
0.08161458373069763,
-0.03503108769655228,
-0.038350947201251984,
-0.030149484053254128,
0.0050342814065515995,
-0.11246328800916672,
0.046700552105903625,
0.0068816207349300385,
-0.025447247549891472,
0.019261430948972702,
0.027852756902575493,
0.004832690581679344,
-0.046731527894735336,
-0.01308752503246069,
-0.05297189950942993,
0.001771043404005468,
-0.00314150913618505,
0.05176975578069687,
-0.009603593498468399,
-0.0542648546397686,
0.09308101236820221,
-0.10201162099838257,
-0.012195965275168419,
-0.040832843631505966,
0.01817922107875347,
0.0758565217256546,
-0.05729638785123825,
0.05706336349248886,
0.03667042776942253,
0.048031870275735855,
0.0714823454618454,
0.03199559450149536,
0.008892593905329704,
-0.054309744387865067,
0.012631463818252087,
-0.1065370962023735,
-0.0340140201151371,
0.005730483680963516,
-0.028032245114445686,
-0.017675666138529778,
0.02540627121925354,
-0.0672513097524643,
0.028382916003465652,
0.0019734278321266174,
0.02378072403371334,
0.06611011177301407,
-0.06640229374170303,
0.02389892190694809,
-0.01380455307662487,
-0.026452386751770973,
0.02876608446240425,
0.004712392576038837,
-0.10060757398605347,
-0.031002074480056763,
0.04712308943271637,
-0.008247480727732182,
0.0016461104387417436,
-0.018000885844230652,
-0.009927024133503437,
0.0024629486724734306
] | https://github.com/xz-group/analoggenie | 24 | 1 | 1 | 0 |
Probabilistic Geometric Principal Component Analysis with application to neural data | https://openreview.net/forum?id=mkDam1xIzW | [
"Han-Lin Hsieh",
"Maryam Shanechi"
] | Spotlight | Dimensionality reduction is critical across various domains of science including neuroscience. Probabilistic Principal Component Analysis (PPCA) is a prominent dimensionality reduction method that provides a probabilistic approach unlike the deterministic approach of PCA and serves as a connection between PCA and Factor Analysis (FA). Despite their power, PPCA and its extensions are mainly based on linear models and can only describe the data in a Euclidean coordinate system around the mean of data. However, in many neuroscience applications, data may be distributed around a nonlinear geometry (i.e., manifold) rather than lying in the Euclidean space around the mean. We develop Probabilistic Geometric Principal Component Analysis (PGPCA) for such datasets as a new dimensionality reduction algorithm that can explicitly incorporate knowledge about a given nonlinear manifold that is first fitted from these data. Further, we show how in addition to the Euclidean coordinate system, a geometric coordinate system can be derived for the manifold to capture the deviations of data from the manifold and noise. We also derive a data-driven EM algorithm for learning the PGPCA model parameters. As such, PGPCA generalizes PPCA to better describe data distributions by incorporating a nonlinear manifold geometry. In simulations and brain data analyses, we show that PGPCA can effectively model the data distribution around various given manifolds and outperforms PPCA for such data. Moreover, PGPCA provides the capability to test whether the new geometric coordinate system better describes the data than the Euclidean one. Finally, PGPCA can perform dimensionality reduction and learn the data distribution both around and on the manifold. These capabilities make PGPCA valuable for enhancing the efficacy of dimensionality reduction for analysis of high-dimensional data that exhibit noise and are distributed around a nonlinear manifold, especially for neural data. | geometry, nonlinear manifold, factor analysis, dimensionality reduction, neural population activity | PGPCA generalizes PPCA by incorporating any nonlinear manifold with various distribution coordinates in its probabilistic model for high-dimensional brain data that exhibit noise. | 11,084 | null | [
-0.01473719347268343,
-0.08150844275951385,
0.046746790409088135,
-0.06501732021570206,
-0.016987137496471405,
0.016343489289283752,
0.029037144035100937,
-0.06033360958099365,
0.05324922129511833,
0.021709410473704338,
0.03892418369650841,
0.013180237263441086,
0.02659403719007969,
0.09216487407684326,
-0.07467314600944519,
0.017470218241214752,
0.0412694588303566,
0.05624490603804588,
-0.04982224106788635,
0.06591527163982391,
-0.018672075122594833,
0.04680470749735832,
-0.01865219697356224,
0.022952215746045113,
0.06620404869318008,
0.04810573533177376,
0.13135932385921478,
-0.02059715986251831,
-0.023290038108825684,
0.051268287003040314,
0.11643195897340775,
0.0008390018483623862,
0.01032499223947525,
-0.05578162893652916,
-0.05121493339538574,
0.029520651325583458,
-0.019317883998155594,
0.07048613578081131,
-0.043667279183864594,
0.018516933545470238,
0.027350306510925293,
0.09961213171482086,
0.0007898492040112615,
0.09992757439613342,
0.06625082343816757,
0.02714306302368641,
-0.029330726712942123,
-0.17821653187274933,
-0.004710622131824493,
-0.037912413477897644,
-0.061746951192617416,
-0.005771998781710863,
-0.07147657126188278,
0.02370530553162098,
0.005342165473848581,
-0.0020850019063800573,
0.04415947571396828,
-0.022612513974308968,
0.010804272256791592,
0.058408837765455246,
0.02394281141459942,
-0.010956479236483574,
-0.01626492477953434,
-0.01761838234961033,
0.0671158954501152,
0.015443717129528522,
0.01184175442904234,
0.012040313333272934,
-0.06529581546783447,
0.0125024588778615,
-0.09360821545124054,
0.030533935874700546,
-0.03207629173994064,
0.0025465867947787046,
-0.008300860412418842,
0.01683632656931877,
-0.06407929956912994,
-0.04134543612599373,
0.044234465807676315,
-0.03482942283153534,
0.02874608337879181,
0.1016363874077797,
0.03767476603388786,
-0.004089876543730497,
0.0999118909239769,
0.03567752242088318,
0.06031689792871475,
-0.03489364683628082,
-0.03779428079724312,
-0.030822936445474625,
0.03083289973437786,
0.014834755100309849,
0.017811322584748268,
-0.03490614891052246,
-0.03860677033662796,
-0.010619296692311764,
0.008327119052410126,
-0.061672426760196686,
0.08977022767066956,
0.047087620943784714,
0.010823573917150497,
-0.025872355327010155,
-0.0022246374282985926,
0.0616740845143795,
-0.026540551334619522,
-0.08774597197771072,
0.036771778017282486,
0.06815104931592941,
0.034536853432655334,
-0.06135791912674904,
0.012870672158896923,
-0.020610060542821884,
-0.08967165648937225,
0.07986222207546234,
0.0022772725205868483,
-0.07165881991386414,
0.02508487179875374,
0.01948712207376957,
0.10582975298166275,
0.009789022617042065,
-0.07066597044467926,
0.005430992692708969,
0.007616656832396984,
-0.015127323567867279,
0.024746425449848175,
0.0036561465822160244,
-0.06988000124692917,
4.902124050957861e-33,
0.03482011333107948,
-0.011969651095569134,
0.01211116649210453,
-0.07114595174789429,
0.02537061646580696,
0.006087723188102245,
-0.006025974173098803,
-0.04393788054585457,
0.12277405709028244,
0.1152891144156456,
-0.0979389101266861,
0.031191477552056313,
-0.004367456771433353,
0.029817283153533936,
0.014830873347818851,
0.03282264992594719,
0.00521168764680624,
0.043579526245594025,
-0.01619596593081951,
-0.05559321492910385,
0.04756588116288185,
-0.006647204048931599,
0.04249337315559387,
-0.04567936062812805,
-0.05307065322995186,
-0.027471337467432022,
-0.0760747641324997,
0.029520396143198013,
-0.05191178619861603,
-0.018265707418322563,
-0.08561942726373672,
0.031884342432022095,
-0.13017821311950684,
0.03833138570189476,
0.004716132301837206,
-0.08162865042686462,
0.007955028675496578,
-0.04370291903614998,
-0.007689361460506916,
-0.06405910849571228,
-0.014627590775489807,
0.034082379192113876,
0.02430807426571846,
-0.0628247931599617,
-0.060061015188694,
0.020660962909460068,
0.021432198584079742,
0.01731540821492672,
0.010369989089667797,
-0.07108237594366074,
0.0381026566028595,
-0.003430398413911462,
0.08758287876844406,
-0.030262373387813568,
-0.07071838527917862,
-0.013795615173876286,
-0.026515807956457138,
-0.008418023586273193,
0.0893971174955368,
0.00357144046574831,
-0.04392792284488678,
0.021604059264063835,
0.038038816303014755,
-0.10053489357233047,
-0.01034980546683073,
-0.032877787947654724,
-0.05004178732633591,
0.030693426728248596,
0.04803392291069031,
-0.003784353379160166,
0.004392268136143684,
0.044362716376781464,
-0.059418752789497375,
-0.030767466872930527,
0.027264520525932312,
-0.04229109361767769,
-0.056208621710538864,
-0.007091138511896133,
-0.014224853366613388,
0.060622602701187134,
-0.06271626800298691,
0.02240005135536194,
-0.03957512974739075,
-0.06196710839867592,
0.008797511458396912,
-0.04764846712350845,
0.04056572541594505,
-0.06882006675004959,
-0.05956585705280304,
-0.06752162426710129,
-0.08171401917934418,
-0.037251852452754974,
0.014672302640974522,
0.0012489660875871778,
0.030791301280260086,
-4.28585791322783e-33,
-0.08963173627853394,
0.029052734375,
0.04994402080774307,
0.012577771209180355,
0.043507736176252365,
-0.06283971667289734,
-0.020891565829515457,
0.05781739205121994,
0.01262263860553503,
-0.034458912909030914,
0.01692376285791397,
-0.06220600754022598,
0.05882345512509346,
-0.04060067981481552,
0.08276994526386261,
0.1077025830745697,
-0.11328785866498947,
0.04081632196903229,
-0.0007566705462522805,
-0.033445339649915695,
-0.030529893934726715,
0.009416289627552032,
-0.057429444044828415,
-0.02709314413368702,
0.033748745918273926,
-0.002869031159207225,
0.047268304973840714,
-0.029486093670129776,
-0.030741548165678978,
-0.00948538351804018,
-0.06490001082420349,
-0.006582389120012522,
-0.11391373723745346,
0.04146292060613632,
-0.13551300764083862,
0.008520994335412979,
-0.041243769228458405,
-0.03549885004758835,
-0.03129037097096443,
0.01742086187005043,
0.036782581359148026,
0.03213295340538025,
0.042205046862363815,
0.06275943666696548,
0.010465283878147602,
-0.03810986876487732,
0.03087903931736946,
0.06964610517024994,
0.005097654182463884,
0.058029647916555405,
-0.03622662276029587,
0.0740777850151062,
0.02604670263826847,
0.013767777010798454,
-0.008122135885059834,
-0.004162598866969347,
0.01761861890554428,
-0.0024695806205272675,
0.08852249383926392,
-0.06362228840589523,
0.016603948548436165,
-0.09010642766952515,
-0.015107658691704273,
0.07131720334291458,
-0.014700223691761494,
0.002090715803205967,
0.06667386740446091,
0.03716764599084854,
-0.008053579367697239,
0.04335898533463478,
0.020778415724635124,
0.041218481957912445,
-0.03162388876080513,
0.023467080667614937,
-0.013581438921391964,
-0.019939936697483063,
0.03628496453166008,
0.01482910756021738,
0.0261823832988739,
-0.011113442480564117,
0.16714511811733246,
-0.02728850021958351,
0.03621217608451843,
-0.03124772198498249,
0.06247240677475929,
0.0853821188211441,
0.05680006369948387,
-0.09564751386642456,
0.050310008227825165,
-0.020229563117027283,
0.014427417889237404,
0.018315138295292854,
-0.029482517391443253,
0.13009408116340637,
0.008555019274353981,
-4.0420808034014044e-8,
-0.04448474198579788,
-0.03325729817152023,
0.0023205913603305817,
-0.05685175955295563,
0.03849339857697487,
-0.0374557264149189,
-0.012255002744495869,
0.1068560853600502,
-0.021614009514451027,
-0.05222336947917938,
0.12299278378486633,
-0.06574708968400955,
-0.0738815888762474,
-0.03670534864068031,
0.017694752663373947,
0.05669418349862099,
0.030305728316307068,
0.08898286521434784,
0.017924999818205833,
0.01159440167248249,
-0.022408129647374153,
-0.02634206786751747,
0.01663394831120968,
0.00835627131164074,
0.10572908073663712,
-0.06413555145263672,
-0.015975547954440117,
0.11349622160196304,
0.030284041538834572,
-0.0716865062713623,
0.04967889189720154,
-0.027351941913366318,
0.032555412501096725,
-0.011921746656298637,
0.0035533206537365913,
0.033840395510196686,
0.06979452073574066,
-0.059117820113897324,
-0.04897322133183479,
-0.03437403216958046,
-0.07478737086057663,
0.030420193448662758,
-0.019681688398122787,
-0.0012464044848456979,
0.042876649647951126,
0.09728702157735825,
0.10790105909109116,
-0.03094668500125408,
0.03461020812392235,
0.09249856323003769,
-0.007887138985097408,
-0.055895011872053146,
0.025785688310861588,
0.0019767358899116516,
-0.014423225075006485,
0.04788767173886299,
0.015735140070319176,
0.01745377853512764,
0.0003811549686361104,
-0.02930361032485962,
-0.0586913600564003,
0.0351894311606884,
-0.02493138052523136,
-0.09816280752420425
] | 0 | 0 | 0 | 0 |
|
DataEnvGym: Data Generation Agents in Teacher Environments with Student Feedback | https://openreview.net/forum?id=00SnKBGTsz | [
"Zaid Khan",
"Elias Stengel-Eskin",
"Jaemin Cho",
"Mohit Bansal"
] | Spotlight | The process of creating training data to teach models is currently driven by humans, who manually analyze model weaknesses and plan how to create data that improves a student model. Recent approaches using large language models (LLMs) as annotators reduce human annotation effort, but still require humans to interpret feedback from evaluations and control the LLM to produce data the student needs. Automating this labor-intensive process by creating autonomous data generation agents – or teachers – is desirable, but requires environments that can simulate the feedback-driven, iterative, closed loop of data creation. To enable rapid and scalable testing for such agents and their modules, we introduce DataEnvGym, a testbed of teacher environments for data generation agents. DataEnvGym frames data generation as a sequential decision-making task, involving an agent consisting of a data generation policy (which generates a plan for creating training data) and a data generation engine (which transforms the plan into data), inside an environment that provides feedback from a student. The agent’s end goal is to improve student model performance. Students are iteratively trained and evaluated on generated data, with their feedback (in the form of errors or weak skills) being reported to the agent after each iteration. As a general-purpose testbed, DataEnvGym includes multiple instantiations of teacher environments across three levels of structure in the state representation and action space, with varying levels of scaffolding support. More structured environments are based on automatically-inferred skills and offer a higher degree of interpretability and control over the curriculum. We support developing and testing data generation agents in four diverse tasks covering text, images, and actions (mathematics, programming, visual question answering, and tool-use) and test multiple student and teacher models. We find that example agents in our teaching environments can iteratively improve students across diverse tasks and settings. Moreover, we show that environments can teach different skill levels and can be used to test variants of key modules, pointing to directions of future work in improving data generation agents, engines, and feedback mechanisms. Project page: https://DataEnvGym.github.io. | iterative data generation, llm agent, lifelong learning | null | 11,063 | 2410.06215 | [
-0.04942907765507698,
-0.041903626173734665,
-0.02018004097044468,
0.03674096614122391,
-0.003680122783407569,
-0.07864192128181458,
-0.019739774987101555,
-0.0428411103785038,
-0.03578775376081467,
-0.005750232841819525,
0.03118777647614479,
-0.05575105920433998,
0.08883450925350189,
0.0022587929852306843,
0.027280576527118683,
0.021715350449085236,
0.10382343083620071,
0.035128701478242874,
-0.05432802438735962,
-0.06349103897809982,
0.019237948581576347,
0.04023518040776253,
0.029135243967175484,
0.012077202089130878,
-0.008502522483468056,
0.03901543840765953,
0.02095429226756096,
-0.07506047189235687,
0.041277747601270676,
-0.0316784493625164,
0.06631603091955185,
-0.04814557731151581,
-0.01837088167667389,
0.09770617634057999,
0.03287949785590172,
-0.0042105079628527164,
-0.04168815165758133,
0.07462209463119507,
-0.05028292536735535,
-0.0061849383637309074,
-0.06506349891424179,
-0.04070063307881355,
0.00019210809841752052,
-0.020916588604450226,
0.016169626265764236,
-0.0510152243077755,
-0.046002715826034546,
-0.0809023380279541,
0.009265873581171036,
0.06411699950695038,
-0.10110525786876678,
-0.09202593564987183,
0.041117116808891296,
-0.044064268469810486,
-0.07703271508216858,
0.006246739998459816,
0.08994753658771515,
0.005169142037630081,
-0.027234919369220734,
-0.020349321886897087,
-0.054883550852537155,
-0.06038583442568779,
-0.04133395478129387,
0.006338066421449184,
-0.055560238659381866,
-0.057524386793375015,
-0.03364362567663193,
0.0733397975564003,
0.05635638162493706,
-0.03300165385007858,
-0.06977864354848862,
0.06450698524713516,
0.005511970724910498,
0.020225653424859047,
0.005306543782353401,
0.04226092994213104,
-0.040212515741586685,
-0.009450816549360752,
0.16845011711120605,
-0.06476301699876785,
-0.06615243852138519,
-0.03683178499341011,
0.03441895917057991,
0.021760590374469757,
0.053030066192150116,
0.03160801902413368,
0.12525981664657593,
0.07335470616817474,
0.015358936972916126,
0.000026433055609231815,
-0.000495200336445123,
0.022260313853621483,
0.03818582370877266,
0.07082617282867432,
0.023580312728881836,
0.09717985987663269,
0.04826708883047104,
-0.11560206115245819,
-0.00019743526354432106,
0.038282252848148346,
-0.031446490436792374,
0.1082623302936554,
0.02001180313527584,
-0.022634953260421753,
-0.04227237030863762,
-0.0646582767367363,
0.05816351994872093,
-0.032323289662599564,
0.009197323583066463,
-0.09296118468046188,
0.003963512368500233,
0.013790806755423546,
0.006032418925315142,
-0.0009992951527237892,
0.05009288713335991,
-0.05842255800962448,
-0.005461749620735645,
0.00991822499781847,
-0.041432611644268036,
0.050224460661411285,
0.04089004173874855,
0.02567555196583271,
-0.009419028647243977,
0.025499308481812477,
0.03698110207915306,
-0.023955393582582474,
-0.024681029841303825,
2.0060654146637067e-33,
0.09035007655620575,
0.04449162632226944,
0.004064869601279497,
0.09826832264661789,
0.05120636150240898,
-0.035005588084459305,
0.042238932102918625,
0.09088532626628876,
-0.08384748548269272,
-0.016185754910111427,
0.014238344505429268,
-0.024521877989172935,
-0.0416802316904068,
0.09089051187038422,
0.07199344784021378,
-0.02133825793862343,
-0.059064291417598724,
0.03183470293879509,
-0.054590582847595215,
0.05201949179172516,
0.07812321931123734,
-0.029651138931512833,
0.06922617554664612,
-0.052304957062006,
0.0567292645573616,
0.032949090003967285,
0.014998674392700195,
0.022005612030625343,
0.018699513748288155,
0.007859199307858944,
-0.0904352068901062,
-0.053336251527071,
0.01254728902131319,
0.07924873381853104,
-0.0439070500433445,
0.017360126599669456,
0.00012312093167565763,
-0.020082684233784676,
0.0012269895523786545,
0.04594961926341057,
0.021958598867058754,
-0.029759660363197327,
0.05903061851859093,
0.0213761106133461,
-0.0571359358727932,
-0.038922108709812164,
0.04002347216010094,
0.02481817454099655,
0.01551893912255764,
-0.054418452084064484,
0.03182855620980263,
0.03898637369275093,
-0.033225417137145996,
-0.04556655511260033,
-0.019406620413064957,
0.03755452111363411,
0.027650175616145134,
-0.0022502895444631577,
0.003858001669868827,
0.010332554578781128,
-0.0815339908003807,
0.0546458400785923,
0.009944412857294083,
0.10559108108282089,
-0.0026182655710726976,
-0.08089031279087067,
0.015373015776276588,
0.017368309199810028,
0.12716692686080933,
-0.057324234396219254,
-0.029066072776913643,
-0.01938115619122982,
-0.03409101441502571,
-0.011014308780431747,
0.05822193995118141,
-0.052886001765728,
0.012111607939004898,
-0.0966450572013855,
-0.01871906965970993,
-0.043565575033426285,
0.06520221382379532,
-0.07254558801651001,
-0.011390196159482002,
-0.043048713356256485,
-0.0031720767728984356,
-0.11745774745941162,
-0.04309128224849701,
-0.058809034526348114,
0.022020753473043442,
-0.05605513975024223,
-0.07066686451435089,
-0.005591368302702904,
-0.04384343698620796,
0.027015618979930878,
0.029397331178188324,
-1.7338670244323362e-33,
0.030532240867614746,
0.004044034983962774,
-0.10600315779447556,
0.11635944247245789,
0.07007075846195221,
-0.06953885406255722,
0.026899047195911407,
0.02509012445807457,
-0.029459495097398758,
-0.08641611784696579,
-0.07725243270397186,
-0.06245072931051254,
0.019719494506716728,
-0.007456415332853794,
-0.05559314042329788,
-0.05000627785921097,
0.002928233239799738,
-0.016992291435599327,
0.006949726492166519,
-0.02211739495396614,
-0.05840160325169563,
0.11563675105571747,
-0.06781337410211563,
0.0068830279633402824,
-0.036438003182411194,
0.0026467093266546726,
0.0122669767588377,
0.024927636608481407,
0.05532754957675934,
0.09350515156984329,
0.03853323683142662,
-0.01131434552371502,
0.019608009606599808,
-0.04154084622859955,
-0.034232985228300095,
0.05020364746451378,
0.083589106798172,
-0.04223810136318207,
-0.04670407995581627,
0.14671185612678528,
0.10023324191570282,
-0.05509630963206291,
-0.09367965906858444,
-0.012717719189822674,
0.015426388941705227,
0.036589521914720535,
-0.04468755051493645,
0.04228508099913597,
-0.018490342423319817,
-0.015442866832017899,
-0.03560449182987213,
-0.000029775090297334827,
-0.041983019560575485,
-0.031424231827259064,
-0.019638361409306526,
-0.06945867091417313,
0.04814564436674118,
-0.03710194304585457,
0.018722761422395706,
0.04289885610342026,
-0.09120187908411026,
-0.02842903509736061,
0.12092620134353638,
0.019760889932513237,
-0.03009585291147232,
-0.047832638025283813,
-0.04005326330661774,
0.029752377420663834,
0.00897249672561884,
0.03288191184401512,
0.05312132462859154,
0.005775259807705879,
-0.05539707466959953,
0.0484633669257164,
0.09132320433855057,
-0.026190483942627907,
-0.037029512226581573,
-0.035386040806770325,
0.06037935987114906,
-0.05610490217804909,
0.022451188415288925,
-0.004005401395261288,
0.01903044432401657,
0.04022768512368202,
0.02457638457417488,
0.03348063305020332,
0.014365771785378456,
0.0991799607872963,
-0.008456832729279995,
0.026500415056943893,
-0.022347189486026764,
0.020145075395703316,
-0.04533037915825844,
0.04558954015374184,
-0.02579200267791748,
-3.962030348247936e-8,
-0.01688849925994873,
-0.040495604276657104,
0.04283233359456062,
0.07512032985687256,
-0.008364755660295486,
-0.018002748489379883,
0.008909572847187519,
0.04780800640583038,
0.04882890731096268,
-0.006761820521205664,
0.04543272778391838,
0.025840817019343376,
-0.03761497512459755,
-0.04393377900123596,
0.0543731227517128,
0.10028044879436493,
0.02110227197408676,
0.016781257465481758,
0.008656715042889118,
-0.053731150925159454,
0.06388019025325775,
0.024409273639321327,
-0.019411128014326096,
-0.012425914406776428,
0.07066315412521362,
-0.03634818643331528,
0.0013448194367811084,
0.13498307764530182,
0.010708271525800228,
-0.004073635209351778,
0.05813384801149368,
-0.039173249155282974,
0.013363388366997242,
-0.05081566050648689,
0.05571066588163376,
0.07725384831428528,
0.00009348530147690326,
-0.020970819517970085,
0.06961533427238464,
-0.07492666691541672,
-0.029989132657647133,
0.06694075465202332,
-0.03926001116633415,
-0.05247076228260994,
0.026116661727428436,
-0.023232100531458855,
-0.07910317927598953,
-0.1264505684375763,
0.0465950109064579,
-0.09576454758644104,
-0.09499058872461319,
-0.05659724026918411,
-0.02905065193772316,
-0.0023677318822592497,
0.0426575243473053,
0.026957189664244652,
-0.038635894656181335,
-0.016764800995588303,
0.04175226017832756,
-0.003296507988125086,
-0.020774757489562035,
0.05095761641860008,
-0.008683201856911182,
-0.01725231669843197
] | https://github.com/codezakh/dataenvgym | 23 | 0 | 0 | 0 |
Correcting the Mythos of KL-Regularization: Direct Alignment without Overoptimization via Chi-Squared Preference Optimization | https://openreview.net/forum?id=hXm0Wu2U9K | [
"Audrey Huang",
"Wenhao Zhan",
"Tengyang Xie",
"Jason D. Lee",
"Wen Sun",
"Akshay Krishnamurthy",
"Dylan J Foster"
] | Spotlight | Language model alignment methods such as reinforcement learning from human feedback (RLHF) have led to impressive advances in language model capabilities, but are limited by a widely observed phenomenon known as *overoptimization*, where the quality of the language model degrades over the course of the alignment process. As the model optimizes performance on an offline reward model, it overfits to inaccuracies and drifts away from preferred responses covered by the data. To discourage such distribution shift, KL-regularization is widely employed in existing offline alignment methods, but overoptimization continues to harm performance. Lending theoretical insight into the source of these empirical observations, we first show that the KL-regularization is too weak to prevent overfitting, then ask: is it possible to design an efficient algorithm that is provably robust to overoptimization?
In this paper, we advance theoretical understanding of sample-efficient offline alignment and introduce a new algorithm called $\chi^2$-Preference Optimization ($\chi$PO). $\chi$PO is a one-line change to Direct Preference Optimization (DPO; Rafailov et al. 2023), that modifies only the logarithmic link function in the DPO objective. Despite this minimal change, $\chi$PO implicitly implements the principle of *pessimism in the face of uncertainty* via regularization with the $\chi^2$-divergence---which quantifies uncertainty more effectively than KL-regularization---and provably alleviates overoptimization, achieving sample-complexity guarantees based on *single-policy concentrability*, the gold standard in offline reinforcement learning. This guarantee makes $\chi$PO the first simple, yet general-purpose offline alignment algorithm that is provably robust to overoptimization. | Reinforcement Learning Theory, Offline Reinforcement Learning, single-policy concentrability, pessimism, RLHF | We propose a new theoretical algorithm for offline alignment/RLHF, Chi-Squared Preference Optimization, which is simple---a one-line change to DPO---yet enjoys the strongest known provable guarantees. | 11,056 | null | [
-0.10788565874099731,
-0.06648701429367065,
0.0290070753544569,
-0.03169881924986839,
-0.03917490690946579,
0.055068280547857285,
0.006745415739715099,
-0.031773678958415985,
0.05194443091750145,
0.05314473435282707,
-0.014536787755787373,
0.0155101353302598,
0.03299852833151817,
0.03383863717317581,
-0.006619039922952652,
-0.04317256435751915,
0.06369182467460632,
0.07898800820112228,
-0.14253608882427216,
-0.04338497668504715,
0.00019176139903720468,
0.005024915561079979,
0.03177649527788162,
0.004020133521407843,
0.03381863981485367,
0.022820480167865753,
-0.013764714822173119,
0.006973303388804197,
0.009708254598081112,
-0.05911035090684891,
-0.03361976146697998,
0.006354246288537979,
0.039179373532533646,
-0.0529056042432785,
-0.04071493074297905,
0.07393744587898254,
-0.12001211196184158,
0.002218235284090042,
-0.001207483932375908,
-0.006069808267056942,
-0.04231180250644684,
-0.008957130834460258,
-0.022298624739050865,
-0.05453462898731232,
0.05799289420247078,
0.010232399217784405,
-0.017238203436136246,
-0.017721161246299744,
0.014632767997682095,
0.029852956533432007,
-0.07370168715715408,
0.02890924923121929,
0.026132732629776,
0.04996882751584053,
0.023484952747821808,
0.029645739123225212,
0.02338007465004921,
0.04816465452313423,
0.04681975021958351,
-0.05139244720339775,
-0.05060424655675888,
-0.0991339311003685,
-0.061381127685308456,
0.041099913418293,
-0.07652243971824646,
-0.07216883450746536,
-0.015746397897601128,
-0.022502761334180832,
-0.021825674921274185,
0.08689006417989731,
-0.06878702342510223,
0.10360925644636154,
0.039524950087070465,
0.09988600760698318,
0.08412785828113556,
0.045120153576135635,
0.011401606723666191,
-0.04500100761651993,
0.03206590563058853,
-0.03955022245645523,
0.02570585161447525,
-0.0035782260820269585,
0.044759247452020645,
-0.014894668944180012,
0.06878587603569031,
-0.02339032106101513,
-0.012244919314980507,
-0.017329221591353416,
0.13100102543830872,
-0.04950597882270813,
-0.028316516429185867,
-0.04192502424120903,
0.06620480120182037,
-0.028937947005033493,
-0.005347562488168478,
0.09086974710226059,
0.0070634931325912476,
-0.047364067286252975,
-0.049947645515203476,
0.08383838087320328,
0.0606900155544281,
0.10806228965520859,
-0.023965800181031227,
-0.05713576078414917,
0.03494303673505783,
-0.07805517315864563,
-0.03593404218554497,
0.029571693390607834,
0.05199835076928139,
-0.11761363595724106,
0.02924801968038082,
0.014143242500722408,
-0.0015795682556927204,
0.0685667097568512,
-0.020282255485653877,
-0.03229933977127075,
0.06804437190294266,
-0.07384134829044342,
-0.0027103256434202194,
-0.015564439818263054,
-0.016259420663118362,
-0.03704902157187462,
-0.05455734208226204,
0.03204980120062828,
-0.019033990800380707,
0.017448533326387405,
0.01964888721704483,
3.112375861797356e-33,
-0.004855387378484011,
0.1091465950012207,
0.08380883187055588,
-0.04104795679450035,
-0.01926366239786148,
-0.08061186224222183,
-0.006943810731172562,
-0.022036880254745483,
-0.002740378025919199,
-0.01953142136335373,
-0.0035589775070548058,
-0.0276496522128582,
-0.031072180718183517,
0.059197936207056046,
-0.008046932518482208,
-0.015879902988672256,
-0.018450185656547546,
0.08780704438686371,
-0.03786401450634003,
0.011077242903411388,
0.1064649373292923,
0.0246344655752182,
0.021893655881285667,
-0.03607495129108429,
0.050695762038230896,
0.006292864214628935,
0.009459275752305984,
-0.0915018692612648,
-0.0471026636660099,
-0.001597278518602252,
0.022894851863384247,
-0.05703068897128105,
0.012738611549139023,
0.03334459662437439,
-0.03493034094572067,
-0.026221172884106636,
-0.038732562214136124,
-0.0898977741599083,
-0.01292150653898716,
0.03541117534041405,
-0.006275098770856857,
0.06621528416872025,
-0.03300419822335243,
0.03330225497484207,
-0.0026126070879399776,
-0.06739786267280579,
0.06143948435783386,
-0.03682804852724075,
-0.04405035078525543,
-0.09653739631175995,
0.003319547511637211,
0.05308143049478531,
-0.03191741928458214,
0.02864566072821617,
-0.07597848027944565,
0.04630580544471741,
0.07597877830266953,
-0.007922811433672905,
0.009966055862605572,
0.04902173951268196,
-0.041269220411777496,
-0.01506381668150425,
0.03137531131505966,
0.01950939930975437,
0.08431873470544815,
-0.04233585670590401,
-0.013779770582914352,
-0.03859927877783775,
-0.0360630601644516,
0.03087054006755352,
0.013273237273097038,
0.03307616338133812,
-0.03102779947221279,
-0.04365773871541023,
0.04136905074119568,
-0.010894554667174816,
0.04125863313674927,
-0.08201953768730164,
0.06146262213587761,
-0.02860344387590885,
-0.0076213316060602665,
0.015294047072529793,
-0.010309535078704357,
-0.06689538061618805,
-0.03538475185632706,
-0.12970730662345886,
0.06716600805521011,
-0.09280895441770554,
-0.007355367299169302,
0.0036099140997976065,
-0.046598274260759354,
0.03114907443523407,
-0.06691056489944458,
-0.004009356256574392,
0.009894145652651787,
-3.07546166576169e-33,
-0.025950884446501732,
-0.08400474488735199,
0.11476723104715347,
0.08796041458845139,
-0.0680372342467308,
-0.027938053011894226,
-0.0013333965325728059,
-0.025175483897328377,
-0.05288997292518616,
-0.021440569311380386,
-0.04018659144639969,
-0.009219623170793056,
0.08190661668777466,
-0.010807274840772152,
0.04884933680295944,
0.012127550318837166,
0.013029711320996284,
0.11779482662677765,
0.03690681234002113,
0.05274194851517677,
0.04533106088638306,
0.03606777265667915,
-0.09718722105026245,
0.08814266324043274,
0.057640980929136276,
0.009527376852929592,
-0.032856717705726624,
0.06490819901227951,
0.03630464896559715,
0.012630607932806015,
-0.07210110127925873,
0.011337225325405598,
-0.10676722228527069,
-0.0630091056227684,
-0.04233187437057495,
0.09043265134096146,
-0.07048985362052917,
0.06829409301280975,
-0.03719042241573334,
0.09450510144233704,
-0.006880260072648525,
0.010218370705842972,
-0.10763482004404068,
-0.05076958239078522,
0.023476311936974525,
-0.022588148713111877,
-0.0948030948638916,
-0.019689561799168587,
-0.03066975437104702,
-0.05565974861383438,
-0.01640956848859787,
0.07272449135780334,
-0.024679100140929222,
0.0839587077498436,
-0.029211226850748062,
0.003306369762867689,
-0.07562145590782166,
-0.0019731565844267607,
0.01830754056572914,
-0.020346997305750847,
-0.026447393000125885,
0.0040412223897874355,
0.057655442506074905,
0.008026988245546818,
0.018935931846499443,
0.03422058746218681,
0.08823417127132416,
0.05014108866453171,
0.0005747041432186961,
-0.047625139355659485,
0.02502492256462574,
-0.06019648164510727,
0.004260537680238485,
0.04117527976632118,
0.036996763199567795,
0.027093101292848587,
0.024561451748013496,
-0.016650453209877014,
-0.0073397341184318066,
-0.025625474750995636,
-0.005183142144232988,
0.007507661823183298,
-0.002361918333917856,
0.04693520441651344,
0.05744653567671776,
0.04969647899270058,
0.05829139053821564,
0.1549711674451828,
-0.015491453930735588,
-0.02301095984876156,
-0.002667189110070467,
0.008427930995821953,
0.039488937705755234,
0.03628701716661453,
-0.047881193459033966,
-4.6540929332650194e-8,
-0.11713798344135284,
-0.02448679506778717,
0.07593812048435211,
0.05817223712801933,
-0.009028904139995575,
-0.08368741720914841,
0.023255279287695885,
0.0289488323032856,
-0.023458193987607956,
-0.11563962697982788,
-0.01133188046514988,
0.05164489895105362,
0.01946779154241085,
-0.04125310480594635,
-0.02663525752723217,
0.127878800034523,
-0.03665453940629959,
0.05189390107989311,
0.021007174625992775,
0.06799538433551788,
0.002225374337285757,
0.06073477119207382,
-0.03605525195598602,
-0.03484882041811943,
0.02723289653658867,
-0.022258004173636436,
-0.01415647566318512,
0.11651641875505447,
0.04671734198927879,
-0.06006978079676628,
0.01193024031817913,
-0.012584876269102097,
0.03292134031653404,
-0.002691644476726651,
0.09402523189783096,
0.04282823204994202,
0.02444477006793022,
-0.0007260887650772929,
-0.08233250677585602,
-0.019557606428861618,
0.029801324009895325,
0.057690832763910294,
-0.05151138827204704,
-0.002188995713368058,
0.07148941606283188,
0.014650612138211727,
-0.06872306019067764,
-0.06816775351762772,
0.03461529687047005,
-0.1267363429069519,
0.03568509593605995,
-0.035210613161325455,
0.0019133572932332754,
-0.03658871725201607,
-0.0007226142915897071,
0.026046302169561386,
-0.027477549389004707,
-0.059234440326690674,
0.04747540131211281,
0.04631655663251877,
0.02211396023631096,
-0.029528941959142685,
-0.07067511975765228,
0.03957870975136757
] | 0 | 0 | 0 | 0 |
|
Exact Certification of (Graph) Neural Networks Against Label Poisoning | https://openreview.net/forum?id=d9aWa875kj | [
"Mahalakshmi Sabanayagam",
"Lukas Gosch",
"Stephan Günnemann",
"Debarghya Ghoshdastidar"
] | Spotlight | Machine learning models are highly vulnerable to label flipping, i.e., the adversarial modification (poisoning) of training labels to compromise performance. Thus, deriving robustness certificates is important to guarantee that test predictions remain unaffected and to understand worst-case robustness behavior. However, for Graph Neural Networks (GNNs), the problem of certifying label flipping has so far been unsolved. We change this by introducing an exact certification method, deriving both sample-wise and collective certificates. Our method leverages the Neural Tangent Kernel (NTK) to capture the training dynamics of wide networks enabling us to reformulate the bilevel optimization problem representing label flipping into a Mixed-Integer Linear Program (MILP). We apply our method to certify a broad range of GNN architectures in node classification tasks. Thereby, concerning the worst-case robustness to label flipping: $(i)$ we establish hierarchies of GNNs on different benchmark graphs; $(ii)$ quantify the effect of architectural choices such as activations, depth and skip-connections; and surprisingly, $(iii)$ uncover a novel phenomenon of the robustness plateauing for intermediate perturbation budgets across all investigated datasets and architectures. While we focus on GNNs, our certificates are applicable to sufficiently wide NNs in general through their NTK. Thus, our work presents the first exact certificate to a poisoning attack ever derived for neural networks, which could be of independent interest. The code is available at https://github.com/saper0/qpcert. | graph neural networks, robustness, certificates, provable robustness, neural networks, label poisoning, label flipping, poisoning, mixed-integer linear programming, neural tangent kernel, support vector machines | First exact sample-wise and collective certificates for (graph) neural networks against label poisoning based on the neural tangent kernel and reformulating the bilevel label poisoning problem into a mixed-integer linear program. | 11,038 | 2412.00537 | [
-0.0919480100274086,
-0.04085439071059227,
0.012838123366236687,
0.0610940158367157,
0.0765262097120285,
-0.0021930222865194082,
0.0026656542904675007,
-0.012576394714415073,
-0.04391897842288017,
-0.05042647942900658,
-0.050824154168367386,
-0.04685143753886223,
0.02288755215704441,
0.012067444622516632,
-0.10254078358411789,
-0.041809190064668655,
0.08703647553920746,
0.12879471480846405,
-0.03892290219664574,
-0.01690955087542534,
-0.05976599082350731,
0.052779678255319595,
0.006083939224481583,
0.006481915712356567,
0.039228405803442,
-0.03459719941020012,
-0.06171606108546257,
-0.012874267064034939,
0.015703966841101646,
-0.029765699058771133,
0.04873184859752655,
-0.013613458722829819,
-0.03950387239456177,
0.010378697887063026,
-0.0401153564453125,
0.06500127166509628,
-0.03740858659148216,
0.020999398082494736,
0.11762377619743347,
-0.0223331730812788,
0.04168214276432991,
-0.013592226430773735,
-0.008749837055802345,
0.01247235294431448,
0.050560541450977325,
0.025891127064824104,
-0.02177998051047325,
-0.07765504717826843,
-0.042206015437841415,
0.037336044013500214,
-0.021800199523568153,
-0.0426780991256237,
-0.03768773376941681,
0.016650285571813583,
-0.07716807723045349,
0.009798138402402401,
0.04068727046251297,
0.014354567974805832,
-0.012272139079868793,
0.12385335564613342,
-0.006953011266887188,
-0.09811583161354065,
-0.053421031683683395,
-0.02772381156682968,
0.06453453749418259,
0.011900033801794052,
0.012248249724507332,
0.009296724572777748,
-0.034653499722480774,
0.09188482910394669,
0.08004438877105713,
0.09298460185527802,
-0.07454860210418701,
0.03299617022275925,
0.015170838683843613,
0.09227776527404785,
0.04349258914589882,
0.038180772215127945,
0.03578007221221924,
-0.06049282103776932,
0.03894415125250816,
0.01811762899160385,
0.02718331664800644,
-0.03165087103843689,
0.10959019511938095,
-0.004306118935346603,
-0.049135975539684296,
0.0084261205047369,
0.033245690166950226,
0.011771743185818195,
-0.05758624151349068,
0.008412240073084831,
0.016656894236803055,
-0.0169406495988369,
0.06432302296161652,
-0.0009493949473835528,
0.031111886724829674,
-0.049297261983156204,
-0.009936711750924587,
0.1076623946428299,
-0.05557546392083168,
-0.016717230901122093,
0.00030912275542505085,
-0.08263882994651794,
0.05070345103740692,
-0.005482370965182781,
0.03309660777449608,
-0.011489005759358406,
0.08806347101926804,
-0.11351334303617477,
0.05929579213261604,
0.07457458227872849,
0.013028106652200222,
-0.020232439041137695,
0.003589936997741461,
0.03068867139518261,
-0.04129501059651375,
-0.022070538252592087,
0.02077707089483738,
0.0719979777932167,
-0.12129193544387817,
0.03490062057971954,
-0.01417077798396349,
-0.03187236562371254,
-0.03915843740105629,
0.036416929215192795,
-0.05089874193072319,
2.5859476145128877e-33,
0.022413909435272217,
0.049807388335466385,
0.03516034036874771,
-0.06362452358007431,
0.074860580265522,
-0.039818938821554184,
-0.02472873032093048,
0.008883962407708168,
0.023113839328289032,
0.02830439992249012,
-0.004494484513998032,
0.007957704365253448,
-0.06560259312391281,
0.03537047281861305,
0.011192962527275085,
0.05857190117239952,
0.005599867552518845,
0.008418330922722816,
-0.06701329350471497,
-0.11987841129302979,
0.06036438047885895,
-0.08155708760023117,
0.0018701081862673163,
-0.03848288953304291,
0.00454981392249465,
-0.018247613683342934,
0.003173146164044738,
0.02164030633866787,
0.002174186520278454,
0.023682070896029472,
-0.11156347393989563,
-0.04292329028248787,
0.09818216413259506,
0.1017058715224266,
0.025612058117985725,
0.040009818971157074,
-0.02314581908285618,
-0.03326904773712158,
0.012147151865065098,
-0.003537939628586173,
-0.07023739814758301,
0.015756065025925636,
0.02619747258722782,
-0.01549629494547844,
-0.009884882718324661,
0.01835845783352852,
-0.01568889617919922,
-0.03920504078269005,
-0.04019395262002945,
-0.08495637029409409,
-0.01582014188170433,
-0.004132692702114582,
-0.040437471121549606,
-0.026138346642255783,
-0.039665259420871735,
-0.03707759082317352,
0.05578092113137245,
0.07786549627780914,
-0.022276228293776512,
0.0978207141160965,
0.02822747640311718,
0.017710233107209206,
-0.054564159363508224,
0.033068954944610596,
-0.04948233440518379,
0.10331930220127106,
0.016529133543372154,
-0.01369400043040514,
0.05417531356215477,
-0.04025657847523689,
-0.07005175948143005,
0.04399879276752472,
-0.07789888978004456,
-0.06498738378286362,
0.04478796571493149,
-0.004464786034077406,
-0.00032886062399484217,
-0.010198895819485188,
-0.01981663517653942,
0.02293912135064602,
0.006038192193955183,
0.01721118949353695,
0.0034780907444655895,
-0.007526439148932695,
-0.058818455785512924,
-0.03807806223630905,
0.017767909914255142,
-0.04729108512401581,
-0.012728036381304264,
0.028842289000749588,
-0.04602028429508209,
0.005366895813494921,
0.02642522007226944,
0.03852356970310211,
-0.017273880541324615,
-1.3590277735650253e-33,
-0.054252054542303085,
0.08288558572530746,
-0.03982805088162422,
0.09166475385427475,
-0.05817351117730141,
-0.01095082238316536,
0.01835891604423523,
0.04034152999520302,
-0.041945621371269226,
-0.024085301905870438,
-0.0022205649875104427,
-0.03599632531404495,
0.05363500490784645,
-0.023974299430847168,
0.024318812415003777,
-0.06268206983804703,
-0.08397898077964783,
0.09672518074512482,
0.05860285088419914,
0.021156156435608864,
0.08657819032669067,
0.15350468456745148,
-0.07802402973175049,
0.06702892482280731,
0.0424908883869648,
0.01232246495783329,
-0.07830662280321121,
0.07015630602836609,
0.08490949124097824,
0.036777596920728683,
-0.013690284453332424,
-0.015674497932195663,
-0.04702409729361534,
-0.0038855180609971285,
0.0244953241199255,
0.06567300111055374,
0.06794144213199615,
-0.018601611256599426,
-0.018535463139414787,
0.10236455500125885,
0.03380320221185684,
-0.06712896376848221,
-0.1279076486825943,
0.01284701842814684,
0.01725456491112709,
-0.022513816133141518,
-0.023043157532811165,
-0.025036385282874107,
-0.025046629831194878,
0.0024372434709221125,
-0.004962472710758448,
0.016732262447476387,
-0.084041066467762,
0.045661561191082,
-0.046002097427845,
-0.06307369470596313,
0.001996680162847042,
0.05709592252969742,
0.10027004778385162,
0.033958520740270615,
-0.05353285372257233,
-0.0147604551166296,
-0.011587385088205338,
0.0024678728077560663,
-0.012370116077363491,
-0.0065361736342310905,
-0.014346538111567497,
0.034880056977272034,
0.03339824452996254,
0.060371775180101395,
0.038603998720645905,
0.017376460134983063,
0.00766974501311779,
-0.013789008371531963,
-0.08231721818447113,
0.05577453225851059,
0.009084437042474747,
0.010159008204936981,
-0.05277164652943611,
-0.031506218016147614,
0.05162177234888077,
-0.009891829453408718,
0.06563436985015869,
0.0943998470902443,
0.10902928560972214,
0.01995321922004223,
0.05863200128078461,
0.014192386530339718,
-0.025279896333813667,
0.02997468411922455,
-0.0026509009767323732,
-0.020926084369421005,
-0.02695615030825138,
0.051168836653232574,
-0.018864445388317108,
-4.504636663682504e-8,
-0.06606893986463547,
0.048112403601408005,
-0.03207014873623848,
0.024843139573931694,
0.09356709569692612,
0.009683923795819283,
0.04366837441921234,
0.070294588804245,
-0.05129186809062958,
0.047657862305641174,
0.045534782111644745,
-0.027212833985686302,
-0.055826470255851746,
-0.10009404271841049,
-0.09175089001655579,
0.01597943715751171,
-0.04261622577905655,
0.05436424911022186,
0.031139103695750237,
0.028542175889015198,
-0.004675041884183884,
-0.0071298000402748585,
0.08469127863645554,
-0.027979029342532158,
0.003225939115509391,
-0.14361871778964996,
0.0037069532554596663,
0.07249116152524948,
0.07947082072496414,
0.04192035645246506,
-0.08968047052621841,
-0.03714149817824364,
0.01412428542971611,
0.02587791346013546,
-0.017348188906908035,
0.1512221097946167,
-0.08393911272287369,
-0.020240401849150658,
0.02425367943942547,
0.07106148451566696,
-0.06785846501588821,
0.08862638473510742,
-0.05721118301153183,
-0.021177921444177628,
-0.017424870282411575,
0.0030952654778957367,
0.044650834053754807,
-0.04042777419090271,
0.05398350581526756,
-0.028823206201195717,
0.006041962653398514,
-0.05706791207194328,
-0.005394908599555492,
0.020834174007177353,
0.04715847596526146,
-0.013530312106013298,
-0.03177780658006668,
-0.06970981508493423,
-0.07542070001363754,
0.10403013229370117,
0.023052604869008064,
-0.006784639321267605,
-0.013176270760595798,
-0.06999284774065018
] | https://github.com/saper0/qpcert | 3 | 0 | 0 | 0 |
Better Instruction-Following Through Minimum Bayes Risk | https://openreview.net/forum?id=7xCSK9BLPy | [
"Ian Wu",
"Patrick Fernandes",
"Amanda Bertsch",
"Seungone Kim",
"Sina Khoshfetrat Pakazad",
"Graham Neubig"
] | Spotlight | General-purpose LLM judges capable of human-level evaluation provide not only a scalable and accurate way of evaluating instruction-following LLMs but also new avenues for supervising and improving their performance. One promising way of leveraging LLM judges for supervision is through Minimum Bayes Risk (MBR) decoding, which uses a reference-based evaluator to select a high-quality output from amongst a set of candidate outputs. In the first part of this work, we explore using MBR decoding as a method for improving the test-time performance of instruction-following LLMs. We find that MBR decoding with reference-based LLM judges substantially improves over greedy decoding, best-of-N decoding with reference-free judges and MBR decoding with lexical and embedding-based metrics on AlpacaEval and MT-Bench. These gains are consistent across LLMs with up to 70B parameters, demonstrating that smaller LLM judges can be used to supervise much larger LLMs. Then, seeking to retain the improvements from MBR decoding while mitigating additional test-time costs, we explore iterative self-training on MBR-decoded outputs. We find that self-training using Direct Preference Optimisation leads to significant performance gains, such that the self-trained models with greedy decoding generally match and sometimes exceed the performance of their base models with MBR decoding. | LLM, instruction-following, test time compute, decoding, MBR, minimal bayes risk, LLM judges, self-improvement | We investigate Minimum Bayes Risk decoding with LLM judges as utility metrics to improve instruction-following LLMs. | 10,961 | 2410.02902 | [
-0.011854260228574276,
-0.05990300700068474,
0.024223346263170242,
-0.010714806616306305,
0.02775891311466694,
0.011442981660366058,
0.01595117151737213,
-0.004133877344429493,
-0.0022646680008620024,
-0.06846460700035095,
-0.06117117404937744,
-0.04747825115919113,
0.044327761977910995,
-0.015905043110251427,
-0.05841732397675514,
-0.004025069996714592,
0.13008436560630798,
0.058343153446912766,
-0.03173483908176422,
-0.03245232254266739,
0.0701359435915947,
0.036179348826408386,
-0.012325059622526169,
0.015685338526964188,
0.009028051979839802,
-0.06326209008693695,
-0.018608925864100456,
0.008615497499704361,
0.07388297468423843,
-0.05330592766404152,
0.04457355663180351,
0.0414150170981884,
0.044032223522663116,
-0.007456694729626179,
-0.03397088497877121,
0.010441245511174202,
-0.005748044233769178,
-0.01778259687125683,
-0.06660972535610199,
-0.022289371117949486,
0.003141996683552861,
0.03662363067269325,
0.049445200711488724,
0.07101444154977798,
-0.02631569094955921,
-0.027629883959889412,
0.027253523468971252,
-0.1045844554901123,
-0.03552718088030815,
0.007569565903395414,
-0.02131044678390026,
0.03373149409890175,
0.012886805459856987,
-0.019620859995484352,
-0.057359807193279266,
0.005888356361538172,
0.09168772399425507,
-0.018890224397182465,
-0.03743845224380493,
-0.0006321545806713402,
-0.07620938867330551,
-0.09692556411027908,
0.020878249779343605,
0.006090039853006601,
-0.01163629349321127,
-0.01959901675581932,
0.05643944442272186,
0.10241152346134186,
-0.00010237968672299758,
0.01616845093667507,
-0.0707598328590393,
-0.029978573322296143,
-0.07139197736978531,
0.09447962790727615,
-0.02392355166375637,
0.03501806780695915,
0.0005578889395110309,
-0.04668578505516052,
0.09614630043506622,
-0.0319889634847641,
-0.042869389057159424,
-0.023236040025949478,
0.051595915108919144,
0.0022239473182708025,
0.02476103976368904,
-0.02251390367746353,
0.023861540481448174,
0.036256175488233566,
0.03751634433865547,
0.10955441743135452,
0.005528299603611231,
-0.08275878429412842,
-0.0052787247113883495,
0.03199620544910431,
-0.007544840686023235,
-0.020547671243548393,
-0.0045151179656386375,
-0.043782711029052734,
0.028555570170283318,
0.05908608436584473,
0.014024922624230385,
0.11002648621797562,
0.016399355605244637,
-0.05944947898387909,
-0.08226165920495987,
-0.005785144865512848,
0.03675168380141258,
0.036048922687768936,
0.05340401828289032,
-0.04001034051179886,
0.0400213748216629,
0.10049359500408173,
0.033543944358825684,
0.005093066953122616,
0.03715510293841362,
0.021608775481581688,
-0.012589927762746811,
-0.02484344132244587,
-0.021729618310928345,
0.04326638951897621,
-0.04628325626254082,
0.014190875925123692,
-0.02697732485830784,
0.04756814241409302,
0.025171402841806412,
-0.03253215178847313,
-0.03476947546005249,
2.571510707174867e-33,
0.10148778557777405,
0.03256741166114807,
-0.02893953211605549,
-0.02931697480380535,
-0.01849398761987686,
-0.013396136462688446,
0.051367878913879395,
0.12214304506778717,
-0.08390989154577255,
0.015780773013830185,
0.004979255609214306,
-0.037479106336832047,
-0.03269713371992111,
0.0018041129224002361,
0.0734328031539917,
0.052570417523384094,
-0.07581530511379242,
0.027027791365981102,
-0.11627034842967987,
-0.006316380109637976,
0.043144550174474716,
-0.010285403579473495,
-0.02193981222808361,
-0.1113516315817833,
0.03615324944257736,
-0.009119202382862568,
0.08458270132541656,
-0.05191485583782196,
-0.017734678462147713,
0.019647102802991867,
-0.13741233944892883,
-0.04997580870985985,
0.02977791801095009,
0.00975225493311882,
0.09622449427843094,
-0.07994528859853745,
-0.028886254876852036,
0.017813898622989655,
0.018598850816488266,
-0.03438357263803482,
-0.021235641092061996,
-0.014986120164394379,
0.020816069096326828,
-0.0316346175968647,
-0.04037823528051376,
-0.08704045414924622,
-0.01905367709696293,
-0.05915718525648117,
-0.07571104913949966,
-0.04707223176956177,
0.03236360847949982,
0.01010142918676138,
-0.01768549159169197,
0.0074699330143630505,
-0.015873843804001808,
0.026703650131821632,
0.029693126678466797,
0.07069280743598938,
0.01691252365708351,
0.12357809394598007,
0.029928365722298622,
0.11020362377166748,
0.018305141478776932,
0.1169511005282402,
0.03319213166832924,
-0.035302769392728806,
0.02221347950398922,
0.014091155491769314,
0.027252722531557083,
-0.07882218062877655,
-0.08919677138328552,
-0.04545975476503372,
0.055019013583660126,
-0.025353537872433662,
0.046945925801992416,
0.008006771095097065,
0.04404788091778755,
-0.06147342175245285,
0.00029532829648815095,
-0.023681029677391052,
0.08481579273939133,
0.041213516145944595,
-0.023745214566588402,
-0.04260512813925743,
-0.006023060530424118,
-0.09554147720336914,
0.037553299218416214,
-0.11683111637830734,
-0.012297889217734337,
0.044477880001068115,
-0.03319784626364708,
-0.05301615595817566,
-0.05415349826216698,
0.019333593547344208,
-0.03872138261795044,
-2.219042215829658e-33,
-0.0025627221912145615,
-0.001555493101477623,
0.012035705149173737,
0.08149780333042145,
-0.051310520619153976,
-0.10938730835914612,
0.051198482513427734,
0.015127230435609818,
-0.0006971507682465017,
-0.05131565406918526,
-0.12691433727741241,
-0.02222871221601963,
0.044671766459941864,
0.0019353167153894901,
0.03236320987343788,
-0.01109933853149414,
-0.05417940020561218,
0.03862864151597023,
-0.0006172864232212305,
-0.01886872388422489,
0.029741384088993073,
0.14590097963809967,
-0.0043982150964438915,
0.008544981479644775,
-0.0830431878566742,
0.03910284861922264,
0.044212810695171356,
0.09702395647764206,
0.05490689352154732,
0.04230155050754547,
-0.016219019889831543,
0.05515904724597931,
0.04194897040724754,
0.00505623547360301,
-0.06309971213340759,
-0.007964441552758217,
0.08860417455434799,
0.01370824221521616,
0.015782753005623817,
0.10917925834655762,
0.09949015080928802,
0.012039495632052422,
-0.05717545747756958,
-0.042186714708805084,
-0.009293454699218273,
-0.024845676496624947,
-0.10380593687295914,
-0.038776565343141556,
0.10474563390016556,
-0.030992088839411736,
-0.04906228557229042,
0.037046052515506744,
-0.06214604526758194,
0.07397940009832382,
-0.04130800440907478,
-0.16495585441589355,
-0.043704554438591,
-0.05673173442482948,
0.03916620463132858,
0.02866661176085472,
0.0008781976066529751,
0.023762589320540428,
0.0678064227104187,
-0.04551839828491211,
0.006848814897239208,
-0.04887179657816887,
-0.019818875938653946,
0.017053067684173584,
-0.05026776343584061,
0.04484076052904129,
0.03454066440463066,
-0.006038985680788755,
-0.0032349536195397377,
0.018587278202176094,
0.04592199996113777,
0.05186750367283821,
0.011602364480495453,
-0.02431032806634903,
-0.05992700532078743,
-0.02676902711391449,
-0.024743979796767235,
-0.015637988224625587,
-0.01682770624756813,
0.05391734093427658,
0.04090891405940056,
0.054573457688093185,
0.06305230408906937,
0.040854327380657196,
-0.03021516464650631,
-0.04017667472362518,
-0.04416107013821602,
0.012176637537777424,
0.019634457305073738,
-0.008162753656506538,
-0.07248925417661667,
-4.58720137430646e-8,
-0.030835315585136414,
-0.06819143891334534,
0.02580219693481922,
0.08567285537719727,
0.0179524514824152,
0.00013129149738233536,
-0.056211598217487335,
0.05599274858832359,
0.02450565993785858,
-0.0005811355076730251,
0.12088319659233093,
-0.07424388080835342,
-0.08415023237466812,
-0.08154281228780746,
0.08201485127210617,
0.0775403156876564,
-0.03594280779361725,
0.013644552789628506,
-0.03680724278092384,
0.011666739359498024,
0.03930850327014923,
0.04265594109892845,
0.040487632155418396,
-0.03342046961188316,
-0.029436003416776657,
-0.07292922586202621,
0.03108137473464012,
0.1092156395316124,
0.07195571810007095,
-0.005347213242202997,
0.03639162704348564,
0.03494926542043686,
-0.031659580767154694,
-0.05183141306042671,
0.015817562118172646,
0.00904405489563942,
-0.05489411950111389,
0.0016751594375818968,
0.05866925045847893,
0.05687320977449417,
-0.028644435107707977,
-0.01066615991294384,
-0.06104664131999016,
0.06198405101895332,
0.03037257306277752,
-0.05462753400206566,
-0.07430654019117355,
-0.03006771393120289,
-0.01762625388801098,
-0.08310004323720932,
0.04209250211715698,
-0.06446526199579239,
-0.019121598452329636,
0.04744714871048927,
0.0649617612361908,
0.07059217989444733,
0.008383299224078655,
-0.029160160571336746,
-0.03519554063677788,
0.08673957735300064,
-0.0011088392930105329,
0.0009406896424479783,
-0.011646068654954433,
-0.022187305614352226
] | 0 | 0 | 0 | 0 |
|
Union-over-Intersections: Object Detection beyond Winner-Takes-All | https://openreview.net/forum?id=HqLHY4TzGj | [
"Aritra Bhowmik",
"Pascal Mettes",
"Martin R. Oswald",
"Cees G. M. Snoek"
] | Spotlight | This paper revisits the problem of predicting box locations in object detection architectures. Typically, each box proposal or box query aims to directly maximize the intersection-over-union score with the ground truth, followed by a winner-takes-all non-maximum suppression where only the highest scoring box in each region is retained. We observe that both steps are sub-optimal: the first involves regressing proposals to the entire ground truth, which is a difficult task even with large receptive fields, and the second neglects valuable information from boxes other than the top candidate. Instead of regressing proposals to the whole ground truth, we propose a simpler approach—regress only to the area of intersection between the proposal and the ground truth. This avoids the need for proposals to extrapolate beyond their visual scope, improving localization accuracy. Rather than adopting a winner-takes-all strategy, we take the union over the regressed intersections of all boxes in a region to generate the final box outputs. Our plug-and-play method integrates seamlessly into proposal-based, grid-based, and query-based detection architectures with minimal modifications, consistently improving object localization and instance segmentation. We demonstrate its broad applicability and versatility across various detection and segmentation tasks. | localization based feature representation, intersection over union, object detection. | null | 10,907 | null | [
0.0656052976846695,
-0.028062911704182625,
0.005377473309636116,
0.024543115869164467,
0.1459340751171112,
0.0051252334378659725,
-0.036597881466150284,
-0.0670078918337822,
0.009767018258571625,
-0.05305933579802513,
-0.1310170292854309,
-0.08146966993808746,
0.03932567685842514,
0.06704772263765335,
-0.04372619092464447,
-0.03799569979310036,
0.15149888396263123,
0.06390713155269623,
-0.004748770501464605,
0.03418724983930588,
0.03336244076490402,
-0.008520822040736675,
0.06273168325424194,
-0.019832341000437737,
0.023352736607193947,
0.02152513898909092,
-0.016254397109150887,
-0.05213913321495056,
-0.005600788630545139,
-0.051928456872701645,
-0.0060363756492733955,
0.024425247684121132,
-0.011462483555078506,
0.09157589823007584,
-0.05027219280600548,
-0.014123671688139439,
-0.06771521270275116,
0.013232960365712643,
0.03303122520446777,
-0.021882405504584312,
-0.0324251651763916,
-0.003106171265244484,
0.06404848396778107,
0.006818288005888462,
0.0926244705915451,
0.099944107234478,
0.01750228740274906,
0.004502025432884693,
-0.007472250610589981,
-0.05049637705087662,
0.012481468729674816,
-0.007011177949607372,
0.003122207010164857,
0.022933589294552803,
-0.07023437321186066,
0.014826081693172455,
0.0403548926115036,
-0.02827521041035652,
0.047760043293237686,
0.06644076853990555,
-0.008095120079815388,
-0.028630264103412628,
0.01767205074429512,
0.011182435788214207,
-0.008893420919775963,
-0.053034957498311996,
0.015543222427368164,
0.02260730043053627,
-0.016022443771362305,
0.02455171011388302,
0.10179441422224045,
0.11120288074016571,
0.029006799682974815,
0.008630440570414066,
0.046558819711208344,
0.07108739018440247,
-0.045911990106105804,
-0.06265921890735626,
0.04281717538833618,
-0.024524517357349396,
-0.03385324776172638,
-0.05100904032588005,
0.03383464366197586,
0.0018753751646727324,
0.0703849047422409,
-0.009797061793506145,
-0.04729252681136131,
-0.006853294093161821,
0.0018570267129689455,
-0.01838570088148117,
-0.08545377105474472,
-0.08923637866973877,
0.020302066579461098,
0.036283377557992935,
0.04228251054883003,
-0.03619838133454323,
0.004983486142009497,
-0.08254942297935486,
-0.010309689678251743,
0.08321146667003632,
-0.08164472877979279,
0.029264265671372414,
0.03340977802872658,
-0.08391748368740082,
0.060952141880989075,
0.015007127076387405,
-0.027364078909158707,
0.012490415945649147,
0.04964666813611984,
-0.06083432212471962,
0.011270331218838692,
-0.016997380182147026,
-0.009068523533642292,
-0.03871176391839981,
-0.02066107839345932,
0.05752810463309288,
0.04100165516138077,
-0.020423930138349533,
-0.02573586255311966,
-0.057253606617450714,
0.003645949298515916,
-0.014962330460548401,
-0.06319352984428406,
-0.005542054306715727,
0.08378321677446365,
-0.027550145983695984,
-0.06641797721385956,
8.983571394430364e-34,
0.04249589517712593,
-0.03590330854058266,
0.04126721993088722,
0.012610492296516895,
0.07395537197589874,
0.06239558756351471,
-0.00026670045917853713,
0.07009847462177277,
-0.026334108784794807,
0.10074383020401001,
-0.05065819248557091,
-0.0012461479054763913,
-0.022988667711615562,
0.017368590459227562,
0.09499553591012955,
-0.041027966886758804,
-0.0014800115022808313,
-0.022280294448137283,
-0.14701735973358154,
0.0854518935084343,
-0.02673225849866867,
-0.02115735597908497,
-0.01623859815299511,
0.01796332187950611,
0.026503335684537888,
0.05894000828266144,
0.04649553447961807,
-0.019005754962563515,
-0.0173212718218565,
-0.00324119976721704,
-0.06454288214445114,
0.027772285044193268,
0.02512974478304386,
0.05881017819046974,
0.06276610493659973,
-0.05537226051092148,
-0.03066893108189106,
-0.029004685580730438,
0.0030919017735868692,
0.05639789626002312,
0.01039140485227108,
0.010119506157934666,
-0.046172622591257095,
-0.05596406012773514,
-0.0056770420633256435,
-0.0299770999699831,
-0.061264194548130035,
0.03272654861211777,
-0.049015212804079056,
0.06563425064086914,
0.08980931341648102,
0.05966300144791603,
-0.059636980295181274,
-0.018719475716352463,
-0.001691028126515448,
-0.03499036654829979,
-0.03524620458483696,
0.05905263125896454,
0.07134055346250534,
-0.03649314120411873,
0.018477216362953186,
-0.04077627509832382,
-0.050622232258319855,
0.05090372636914253,
-0.06723048537969589,
-0.002312090015038848,
0.036100294440984726,
-0.007795671466737986,
0.009395366534590721,
0.01849912479519844,
0.005249238573014736,
0.016849776729941368,
-0.0022621899843215942,
-0.004306904971599579,
0.029794150963425636,
-0.0033726308029145002,
-0.004923589061945677,
0.013388387858867645,
0.03401685506105423,
-0.03174133226275444,
0.002703557489439845,
-0.026872865855693817,
-0.017018631100654602,
-0.030043600127100945,
-0.014874636195600033,
0.036844536662101746,
0.0364927239716053,
-0.06672617048025131,
-0.015484856441617012,
-0.01668463461101055,
-0.05725947022438049,
-0.057462528347969055,
-0.027812203392386436,
-0.03380250558257103,
0.006354845128953457,
-1.5195624656629326e-34,
0.09157395362854004,
-0.027605390176177025,
-0.04862304404377937,
-0.06780005991458893,
-0.0393536202609539,
-0.06568250060081482,
0.029072649776935577,
-0.06661029160022736,
0.011410428211092949,
-0.024806473404169083,
-0.0868552178144455,
0.00683178286999464,
0.056726567447185516,
0.01667794957756996,
-0.01111385878175497,
0.009620476514101028,
-0.008188460022211075,
0.011113260872662067,
-0.08203878998756409,
-0.003773328149691224,
-0.0445716492831707,
0.11956420540809631,
-0.07976841181516647,
-0.06054789945483208,
-0.08093596249818802,
0.02310609631240368,
-0.01953287608921528,
0.0471162274479866,
-0.018754344433546066,
-0.04685099050402641,
-0.008898450993001461,
-0.03462740406394005,
-0.048840999603271484,
-0.016171298921108246,
-0.09051918983459473,
0.06981798261404037,
0.03829439729452133,
-0.051552411168813705,
-0.016759948804974556,
0.04849637672305107,
0.06282908469438553,
-0.05924154818058014,
-0.06918244808912277,
0.05890309065580368,
-0.02922223135828972,
-0.062171149998903275,
-0.08310229331254959,
0.10386430472135544,
-0.03757132589817047,
0.04189000651240349,
-0.1261041909456253,
0.06317777931690216,
-0.10230634361505508,
0.03613777086138725,
-0.01462195161730051,
0.02099856734275818,
-0.047142453491687775,
-0.0027338124345988035,
0.11085788160562515,
0.0025494955480098724,
-0.006983194034546614,
0.026096446439623833,
-0.051959194242954254,
0.0609203577041626,
0.03387324512004852,
-0.0007323576719500124,
-0.03873640298843384,
0.00744604179635644,
0.05154018476605415,
0.030952421948313713,
0.020131226629018784,
0.0022016833536326885,
-0.07526108622550964,
0.03566432744264603,
-0.03379308432340622,
0.08725271373987198,
0.025477658957242966,
0.031621091067790985,
0.10483110696077347,
0.016733430325984955,
-0.0189126655459404,
-0.04478895291686058,
0.02562279812991619,
0.0834815576672554,
0.15681017935276031,
0.02064838819205761,
-0.04043825715780258,
0.012719614431262016,
-0.009860990568995476,
0.021851249039173126,
-0.07663840055465698,
-0.018084200099110603,
0.019462984055280685,
0.007821658626198769,
-0.036908797919750214,
-4.3049432463249104e-8,
-0.08345741778612137,
0.06047435477375984,
-0.11188529431819916,
-0.0021956509444862604,
0.005013492424041033,
0.03152652457356453,
0.0015074523398652673,
0.13212622702121735,
-0.02392890676856041,
-0.013122729025781155,
0.07151743769645691,
-0.010646050795912743,
-0.043162837624549866,
-0.08592121303081512,
0.031892869621515274,
0.0643082931637764,
-0.060008611530065536,
0.04024556279182434,
-0.04060804843902588,
0.07795514911413193,
0.062020543962717056,
0.037061985582113266,
0.018297266215085983,
0.04999960958957672,
-0.02700481377542019,
-0.042389076203107834,
-0.028858274221420288,
0.11952279508113861,
0.04740617796778679,
0.006617770530283451,
-0.025588322430849075,
0.023420870304107666,
0.07581058144569397,
0.03708934411406517,
0.11108855903148651,
0.027181778103113174,
-0.06155100092291832,
0.06149198114871979,
0.011619294993579388,
-0.014247317798435688,
0.055745478719472885,
-0.05453033000230789,
-0.01750013791024685,
-0.008554147556424141,
-0.00811663456261158,
0.05268910154700279,
0.022348634898662567,
-0.12073621898889542,
0.06225144490599632,
-0.005661812610924244,
-0.026896784082055092,
-0.0005146560724824667,
-0.04955410584807396,
0.1177622377872467,
0.008460446260869503,
0.011620670557022095,
0.0017803962109610438,
-0.09198489785194397,
0.0049896398559212685,
0.12131525576114655,
-0.08239025622606277,
0.0625816211104393,
-0.05354498699307442,
0.016897855326533318
] | 0 | 0 | 0 | 0 |
|
MamKO: Mamba-based Koopman operator for modeling and predictive control | https://openreview.net/forum?id=hNjCVVm0EQ | [
"ZHAOYANG LI",
"Minghao Han",
"Xunyuan Yin"
] | Spotlight | The Koopman theory, which enables the transformation of nonlinear systems into linear representations, is a powerful and efficient tool to model and control nonlinear systems. However, the ability of the Koopman operator to model complex systems, particularly time-varying systems, is limited by the fixed linear state-space representation. To address the limitation, the large language model, Mamba, is considered a promising strategy for enhancing modeling capabilities while preserving the linear state-space structure.
In this paper, we propose a new framework, the Mamba-based Koopman operator (MamKO), which provides enhanced model prediction capability and adaptability, as compared to Koopman models with constant Koopman operators. Inspired by the Mamba structure, MamKO generates Koopman operators from online data; this enables the model to effectively capture the dynamic behaviors of the nonlinear system over time. A model predictive control system is then developed based on the proposed MamKO model. The modeling and control performance of the proposed method is evaluated through experiments on benchmark time-invariant and time-varying systems. The experimental results demonstrate the superiority of the proposed approach. Additionally, we perform ablation experiments to test the effectiveness of individual components of MamKO. This approach unlocks new possibilities for integrating large language models with control frameworks, and it achieves a good balance between advanced modeling capabilities and real-time control implementation efficiency. | Mamba; Koopman operator; model predictive control; nonlinear systems | null | 10,895 | null | [
-0.08070384711027145,
-0.03152446448802948,
-0.044891804456710815,
0.03522539511322975,
-0.02448577806353569,
0.032110050320625305,
-0.06493234634399414,
-0.03574403002858162,
0.006446666084229946,
0.046229295432567596,
0.01778753474354744,
-0.042819440364837646,
0.02710455097258091,
0.03283743932843208,
-0.011913028545677662,
0.04882748797535896,
0.03166695684194565,
0.010711944662034512,
0.012087516486644745,
-0.06877697259187698,
0.1120704784989357,
0.01744028925895691,
-0.02228551357984543,
-0.0012939729494974017,
-0.0008169860229827464,
0.08116057515144348,
-0.045006729662418365,
0.05416222661733627,
-0.06736894696950912,
-0.06933324784040451,
0.0263254102319479,
-0.001092904363758862,
0.055031631141901016,
-0.12777432799339294,
-0.11701760441064835,
-0.0072944010607898235,
-0.08269910514354706,
-0.007925202138721943,
-0.07808972895145416,
-0.0835970863699913,
0.04185808077454567,
-0.004168695770204067,
0.08480353653430939,
0.011086985468864441,
0.032350052148103714,
0.03020518645644188,
-0.014537974260747433,
-0.0770355835556984,
0.05127233266830444,
0.06477347761392593,
-0.14429156482219696,
-0.05199763551354408,
0.10894747823476791,
0.013186536729335785,
0.022867349907755852,
-0.011697794310748577,
0.0016477381577715278,
0.06572600454092026,
0.0730094388127327,
-0.08646179735660553,
-0.057738982141017914,
0.01467297412455082,
0.007300615310668945,
0.06868921220302582,
-0.05609778314828873,
-0.02889457903802395,
-0.027362674474716187,
0.030908988788723946,
-0.0588383823633194,
0.03100675344467163,
-0.07419463247060776,
-0.005775952711701393,
0.06786704063415527,
0.016193019226193428,
0.024636389687657356,
-0.051276955753564835,
0.08532159775495529,
0.01902235858142376,
0.015828801319003105,
0.06993219256401062,
0.04333958402276039,
0.06196510046720505,
-0.0748910903930664,
-0.08138033002614975,
0.027393538504838943,
0.06971770524978638,
0.02766190841794014,
0.026163257658481598,
0.13791728019714355,
0.017182573676109314,
-0.01627792976796627,
-0.07832708209753036,
0.060312964022159576,
-0.013266724534332752,
-0.013998775742948055,
0.009870942682027817,
0.0010891563724726439,
0.04459831491112709,
-0.04833061620593071,
0.04850360006093979,
0.0017505658324807882,
0.030908256769180298,
0.07149073481559753,
-0.027986668050289154,
0.05048401281237602,
-0.022215461358428,
0.040112946182489395,
-0.0052468678914010525,
0.003761876840144396,
0.004776408895850182,
0.010596376843750477,
-0.06407652795314789,
-0.016357503831386566,
0.019815718755126,
0.06690248101949692,
-0.019031519070267677,
0.0496361143887043,
-0.0912998765707016,
-0.000354577146936208,
0.02455950155854225,
0.04341550171375275,
-0.09102530032396317,
0.04169270023703575,
0.054544128477573395,
0.058974117040634155,
0.011298130266368389,
-0.04392025247216225,
-1.4342221268062115e-33,
-0.004037303850054741,
-0.02446579374372959,
0.017654983326792717,
0.0016514378366991878,
0.0544261671602726,
-0.05853705480694771,
-0.017629573121666908,
0.003703692462295294,
-0.0030691539868712425,
-0.015369030646979809,
-0.05018771067261696,
-0.026818472892045975,
-0.06062629073858261,
0.05363955348730087,
0.009710763581097126,
0.029505548998713493,
-0.02130645141005516,
0.00582666601985693,
0.0030053905211389065,
-0.06539912521839142,
0.036239925771951675,
0.051344312727451324,
-0.10107421875,
-0.039699736982584,
0.008396646939218044,
0.0618731863796711,
0.0032791621051728725,
-0.03519358113408089,
-0.03105529025197029,
-0.003084853757172823,
-0.0422980822622776,
0.0518355555832386,
-0.0620792917907238,
-0.01477036066353321,
-0.06687130033969879,
0.007527396082878113,
-0.07385178655385971,
-0.021944493055343628,
-0.028431136161088943,
-0.009216172620654106,
-0.014966419897973537,
-0.0818103700876236,
-0.0015259393258020282,
0.043229300528764725,
-0.09372762590646744,
-0.11861612647771835,
-0.008275275118649006,
-0.020025411620736122,
-0.02434476651251316,
0.01097843050956726,
-0.017688922584056854,
-0.011147962883114815,
-0.04756920039653778,
0.04664606973528862,
-0.049920618534088135,
0.000967648287769407,
-0.006337294355034828,
0.03694253787398338,
-0.007783170789480209,
0.08662162721157074,
-0.14564111828804016,
0.0010955907637253404,
0.16204629838466644,
0.038459256291389465,
0.023616177961230278,
-0.03703729063272476,
-0.007431323640048504,
0.03886452317237854,
0.0652526393532753,
0.0058060758747160435,
0.04696305841207504,
-0.058833248913288116,
0.02628311701118946,
-0.00046187781845219433,
0.058537617325782776,
-0.08302377909421921,
0.0292766522616148,
-0.0945580005645752,
-0.07060033082962036,
0.05268395319581032,
0.03651786223053932,
-0.02981170080602169,
0.016500918194651604,
0.05875622481107712,
0.00204224931076169,
-0.055134180933237076,
0.016237348318099976,
0.0363297276198864,
-0.0481310710310936,
-0.06747591495513916,
-0.12877823412418365,
-0.014213361777365208,
0.047669533640146255,
0.009126944467425346,
0.0038570936303585768,
-2.5234610898880597e-33,
0.05737824738025665,
0.06017355993390083,
0.05395352467894554,
0.05067581683397293,
-0.033177752047777176,
-0.05338317155838013,
-0.046536874026060104,
0.1051061674952507,
-0.09589917212724686,
-0.05634598061442375,
-0.04149655997753143,
-0.04935300350189209,
0.035240426659584045,
0.051760029047727585,
0.09511435031890869,
0.03102305345237255,
-0.041720934212207794,
-0.0678633451461792,
0.005143131595104933,
-0.015047991648316383,
-0.017084762454032898,
0.031108541414141655,
-0.024196960031986237,
0.0037929473910480738,
-0.03636658564209938,
0.0025170096196234226,
-0.03865065425634384,
0.05407387763261795,
0.029210545122623444,
0.04058403894305229,
-0.09711146354675293,
-0.019177328795194626,
-0.04726054519414902,
0.0339326485991478,
-0.043624017387628555,
-0.005631823558360338,
-0.030365927144885063,
0.014983883127570152,
-0.01783944107592106,
0.08149442821741104,
0.06888224929571152,
-0.009973004460334778,
-0.05176724120974541,
-0.03096221759915352,
-0.05039336904883385,
0.03359358012676239,
0.0027675735764205456,
0.017784060910344124,
0.057866912335157394,
-0.06035354360938072,
-0.07730306684970856,
0.06495113670825958,
-0.0739840641617775,
-0.0733141228556633,
-0.03640194237232208,
0.06171025335788727,
-0.005210330709815025,
-0.004975528456270695,
-0.0035716101992875338,
0.03112082928419113,
-0.06017830967903137,
-0.06409987807273865,
0.03782263398170471,
0.0397234670817852,
0.02195735089480877,
0.026117075234651566,
0.037690043449401855,
-0.005104124546051025,
-0.01399895828217268,
-0.0170053169131279,
-0.04884990677237511,
-0.0066391583532094955,
0.02213164232671261,
0.06751158833503723,
0.06219661235809326,
-0.1084255650639534,
-0.0370962917804718,
-0.0036195635329931974,
0.049048714339733124,
-0.00014327761891763657,
0.01284751109778881,
-0.025957178324460983,
0.06659536063671112,
0.010830337181687355,
-0.09028290957212448,
0.0957067683339119,
0.015600202605128288,
0.008516823872923851,
0.07419541478157043,
-0.03726546838879585,
0.0322408527135849,
0.06282130628824234,
-0.0021306045819073915,
0.06097491458058357,
-0.041303880512714386,
-4.68189682578668e-8,
-0.046056199818849564,
-0.024623455479741096,
0.020330607891082764,
0.0022785793989896774,
-0.021295597776770592,
-0.0012478376738727093,
0.01998949609696865,
-0.13978509604930878,
0.060471970587968826,
-0.04382944479584694,
0.0874766856431961,
-0.002107467269524932,
-0.001121348119340837,
-0.011397690512239933,
0.02903902530670166,
0.029469070956110954,
-0.002432019216939807,
0.00023217689886223525,
0.00837891548871994,
0.02084539458155632,
0.010646253824234009,
-0.04859514907002449,
-0.03220648318529129,
0.04657885804772377,
0.051161665469408035,
0.028326865285634995,
-0.11144746094942093,
0.1363510638475418,
0.05652308091521263,
0.050294794142246246,
-0.016410237178206444,
0.07819858938455582,
-0.028186261653900146,
0.09172823280096054,
-0.010316292755305767,
-0.04169376939535141,
0.032918814569711685,
-0.07451218366622925,
0.006206369958817959,
0.03727957606315613,
-0.021188970655202866,
0.02386409603059292,
-0.06551245599985123,
-0.0017742766067385674,
0.01636836677789688,
0.08340788632631302,
0.06985555589199066,
-0.1252318173646927,
0.016467861831188202,
-0.02103888988494873,
-0.007821785286068916,
0.06636317819356918,
-0.02446902170777321,
0.007039274089038372,
0.08036287128925323,
0.015462230890989304,
0.07041865587234497,
-0.008551386184990406,
0.05641898885369301,
-0.005975442007184029,
-0.06792771071195602,
0.022125033661723137,
-0.03923353552818298,
0.044500842690467834
] | 0 | 0 | 0 | 0 |
|
LoRA3D: Low-Rank Self-Calibration of 3D Geometric Foundation models | https://openreview.net/forum?id=LSp4KBhAom | [
"Ziqi Lu",
"Heng Yang",
"Danfei Xu",
"Boyi Li",
"Boris Ivanovic",
"Marco Pavone",
"Yue Wang"
] | Spotlight | Emerging 3D geometric foundation models, such as DUSt3R, offer a promising approach for in-the-wild 3D vision tasks.
However, due to the high-dimensional nature of the problem space and scarcity of high-quality 3D data,
these pre-trained models still struggle to generalize to many challenging circumstances,
such as limited view overlap or low lighting.
To address this, we propose LoRA3D, an efficient self-calibration pipeline to *specialize* the pre-trained models to target scenes using their own multi-view predictions.
Taking sparse RGB images as input, we leverage robust optimization techniques to refine multi-view predictions and align them into a global coordinate frame.
In particular, we incorporate prediction confidence into the geometric optimization process,
automatically re-weighting the confidence to better reflect point estimation accuracy.
We use the calibrated confidence to generate high-quality pseudo labels for the calibrating views and fine-tune the models using low-rank adaptation (LoRA) on the pseudo-labeled data.
Our method does not require any external priors or manual labels. It completes the self-calibration process on a **single standard GPU within just 5 minutes**.
Each low-rank adapter requires only **18MB** of storage.
We evaluated our method on **more than 160 scenes** from the Replica, TUM and Waymo Open datasets,
achieving up to **88\% performance improvement** on 3D reconstruction, multi-view pose estimation and novel-view rendering.
For more details, please visit our project page at https://520xyxyzq.github.io/lora3d/. | 3D foundation model, model specialization, robust optimization, low rank adaptation, self-supervised learning | An efficient test-time self-calibration pipeline to specialize a 3D foundation model to a target scene in 5 min on a single GPU. | 10,864 | 2412.07746 | [
-0.030458655208349228,
-0.08861272782087326,
-0.014678598381578922,
0.005360707174986601,
0.13546566665172577,
0.001936180517077446,
0.0018440092680975795,
-0.03455708175897598,
-0.016905661672353745,
-0.027710964903235435,
-0.08673989027738571,
-0.09149603545665741,
0.054922278970479965,
0.0823594480752945,
-0.013335475698113441,
-0.004896997008472681,
0.055375270545482635,
0.05723748356103897,
-0.07898976653814316,
-0.005585972685366869,
-0.01737000234425068,
-0.002783626550808549,
-0.0034911525435745716,
0.05085242539644241,
0.1073850691318512,
-0.019801121205091476,
0.04311519116163254,
-0.019501319155097008,
-0.06266927719116211,
-0.052110232412815094,
0.06528130918741226,
-0.006871180608868599,
0.0205313041806221,
-0.03331676870584488,
-0.00039468094473704696,
-0.011744127608835697,
-0.04326753318309784,
0.03551618754863739,
-0.05526934191584587,
0.0018884348683059216,
-0.010500953532755375,
0.03838573768734932,
-0.013595424592494965,
0.022775279358029366,
0.06095992028713226,
0.05243220925331116,
0.07490243017673492,
-0.06163615360856056,
0.058421455323696136,
-0.028351349756121635,
-0.016043970361351967,
-0.03993727266788483,
-0.09857578575611115,
-0.033222924917936325,
-0.033441122621297836,
0.05403716862201691,
0.060157597064971924,
-0.11518345028162003,
0.08316484838724136,
-0.015190456062555313,
0.038755763322114944,
-0.02787877805531025,
0.03211564943194389,
0.007617632858455181,
0.04578427970409393,
-0.010258269496262074,
0.013934380374848843,
-0.026782454922795296,
-0.024125168099999428,
0.018745141103863716,
-0.06302198022603989,
0.0032567281741648912,
0.04413196071982384,
-0.009775917045772076,
0.046512309461832047,
-0.0029289566446095705,
0.03275727108120918,
-0.008140710182487965,
-0.01359235867857933,
-0.047718845307826996,
0.01102309487760067,
0.05326464772224426,
0.006849945988506079,
0.054767049849033356,
-0.01786186173558235,
0.019058262929320335,
-0.01097609382122755,
0.09273312240839005,
0.03941073641180992,
-0.03467999026179314,
-0.0180243831127882,
-0.044910866767168045,
-0.11638510972261429,
0.03614518418908119,
-0.08187634497880936,
-0.026759477332234383,
0.028317853808403015,
-0.1262730062007904,
0.0065100714564323425,
0.06744320690631866,
0.00681687518954277,
-0.027103407308459282,
0.03915226086974144,
0.009512611664831638,
0.03183123096823692,
0.027627214789390564,
0.06570746004581451,
0.03771078586578369,
-0.006168081425130367,
-0.01735466532409191,
0.0825420543551445,
0.0063185677863657475,
-0.004170164465904236,
0.038947682827711105,
-0.01476704329252243,
0.0843762531876564,
0.0043528154492378235,
-0.0409148670732975,
0.025880571454763412,
0.018526723608374596,
-0.05889718979597092,
-0.01364897284656763,
0.08959675580263138,
0.03678065910935402,
0.011716501787304878,
-0.020673707127571106,
-0.07479596883058548,
7.5372325840965635e-34,
0.010600659064948559,
0.08094784617424011,
0.04286019131541252,
-0.022758403792977333,
0.04377470538020134,
-0.024130960926413536,
0.060625020414590836,
0.022012662142515182,
-0.0002353109302930534,
0.02979518473148346,
0.007357510272413492,
-0.06163426488637924,
0.01471781823784113,
0.06928904354572296,
-0.01625991426408291,
0.0438113808631897,
-0.026016922667622566,
-0.004655074793845415,
-0.07148636132478714,
0.13860252499580383,
0.015136297792196274,
-0.006688665132969618,
-0.029913615435361862,
-0.014365317299962044,
0.021387673914432526,
0.00874945055693388,
0.05965052545070648,
0.03991295024752617,
-0.04422502592206001,
-0.0033152312971651554,
0.01402695756405592,
-0.02189807966351509,
0.052162669599056244,
0.04278504475951195,
-0.007119506597518921,
-0.05081038549542427,
-0.04396529868245125,
0.036515217274427414,
-0.03348393738269806,
-0.021980322897434235,
-0.015413538552820683,
0.12392144650220871,
-0.022911356762051582,
-0.015774909406900406,
-0.06213023141026497,
0.03642062097787857,
-0.0033052656799554825,
0.11033112555742264,
-0.12127746641635895,
-0.021507123485207558,
0.07669883966445923,
-0.012641155160963535,
-0.0944751501083374,
-0.06045481935143471,
-0.14506596326828003,
-0.0701819434762001,
0.011828871443867683,
-0.035712018609046936,
0.07597871869802475,
-0.03534339368343353,
-0.007478226907551289,
-0.006367596797645092,
-0.06771719455718994,
0.06510679423809052,
-0.02341647818684578,
-0.034248437732458115,
0.11692408472299576,
0.06300488859415054,
0.020952133461833,
0.030382756143808365,
-0.08414621651172638,
-0.02883613109588623,
0.03524111211299896,
-0.05221591889858246,
0.0644373819231987,
-0.015033259987831116,
0.06047046184539795,
0.0031507902313023806,
0.08378240466117859,
0.08614244312047958,
-0.010702789761126041,
0.07360383868217468,
0.07591445744037628,
-0.06282340735197067,
-0.0700908824801445,
-0.05531886965036392,
0.04892844706773758,
-0.049343883991241455,
-0.0343807078897953,
-0.052340783178806305,
-0.04099804908037186,
-0.08372879773378372,
-0.01702241227030754,
-0.06645413488149643,
-0.0585390068590641,
1.7777373919599855e-33,
0.11004340648651123,
-0.008369688875973225,
-0.006641705054789782,
0.03806264325976372,
-0.017321009188890457,
-0.02770642191171646,
-0.02390025556087494,
0.040064163506031036,
0.001084702555090189,
-0.09984302520751953,
0.0016754961106926203,
0.0023176681715995073,
0.06678701192140579,
0.04543484374880791,
0.05996539816260338,
-0.006822260096669197,
-0.03644504025578499,
-0.02555660903453827,
-0.10464031249284744,
0.026994718238711357,
0.017106302082538605,
0.030448146164417267,
-0.07142075151205063,
-0.006949747446924448,
-0.0865786075592041,
0.08182697743177414,
-0.03496541827917099,
0.05913158506155014,
-0.028739867731928825,
0.012696584686636925,
-0.03167501091957092,
-0.013197579421103,
0.01629738137125969,
-0.06512364745140076,
-0.06205739080905914,
0.06601279228925705,
-0.0015648658154532313,
-0.022118395194411278,
-0.07721581310033798,
0.026870597153902054,
-0.027025146409869194,
0.009197758510708809,
-0.07128233462572098,
0.021047241985797882,
-0.037787891924381256,
0.003810109104961157,
0.020891597494482994,
-0.04820534586906433,
0.01869785413146019,
0.11021100729703903,
-0.09239793568849564,
0.04961808770895004,
-0.03407503664493561,
0.02736540138721466,
0.008130212314426899,
-0.08888700604438782,
-0.09184800088405609,
0.029682712629437447,
0.11630047857761383,
-0.012941508553922176,
-0.015508188866078854,
-0.04954936355352402,
-0.05996245890855789,
0.04362039268016815,
0.006871840450912714,
0.004580305889248848,
-0.06834965944290161,
-0.018163073807954788,
0.05655376613140106,
0.0843549594283104,
0.002655611839145422,
-0.048690397292375565,
0.040077950805425644,
0.0038765394128859043,
-0.05153973400592804,
-0.03420242667198181,
0.02721371501684189,
0.0036788033321499825,
0.03826203942298889,
-0.008835112676024437,
0.06154366210103035,
-0.03397763520479202,
0.04394056648015976,
0.14847338199615479,
0.11738255620002747,
0.048240553587675095,
0.05072683095932007,
0.039879944175481796,
0.08767681568861008,
0.024221738800406456,
0.008044266141951084,
-0.0045456900261342525,
0.039595577865839005,
0.021457331255078316,
-0.03677799180150032,
-3.7075949421705445e-8,
-0.0615077018737793,
0.05702202022075653,
0.03452741727232933,
0.04929746314883232,
-0.05082838609814644,
-0.0013754934770986438,
0.06289483606815338,
0.0712386965751648,
-0.021428577601909637,
-0.04211125522851944,
0.02624698169529438,
-0.03298282250761986,
-0.05419495701789856,
-0.025049053132534027,
0.03381317853927612,
0.07252886146306992,
0.03392190858721733,
0.06319382041692734,
-0.006198467221111059,
0.04233066737651825,
-0.04800202697515488,
0.048692815005779266,
0.05704500898718834,
-0.027985313907265663,
0.05628252774477005,
0.02258908376097679,
0.009656052105128765,
0.06278189271688461,
0.14704085886478424,
-0.02982369251549244,
0.0150448614731431,
-0.03975829482078552,
0.020252160727977753,
-0.04275282472372055,
-0.06434796005487442,
0.04204777255654335,
0.045970652252435684,
0.02212495729327202,
0.045304447412490845,
-0.08138300478458405,
-0.06937339156866074,
0.07429002970457077,
-0.04395768791437149,
0.003843507496640086,
0.04296054318547249,
0.042177703231573105,
-0.03010394237935543,
-0.06876672059297562,
-0.0006031806697137654,
-0.06088937819004059,
-0.0156762283295393,
-0.018850671127438545,
-0.017926566302776337,
0.07832243293523788,
0.058596864342689514,
0.031079277396202087,
0.00897165946662426,
-0.03589298576116562,
0.025013793259859085,
0.03225509822368622,
-0.045881759375333786,
-0.06988584995269775,
-0.04148203507065773,
-0.032470252364873886
] | 0 | 0 | 0 | 0 |
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.