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
|
|---|---|---|---|---|---|---|---|---|---|
Weighted Point Set Embedding for Multimodal Contrastive Learning Toward Optimal Similarity Metric | https://openreview.net/forum?id=uSz2K30RRd | [
"Toshimitsu Uesaka",
"Taiji Suzuki",
"Yuhta Takida",
"Chieh-Hsin Lai",
"Naoki Murata",
"Yuki Mitsufuji"
] | Spotlight | In typical multimodal contrastive learning, such as CLIP, encoders produce one
point in the latent representation space for each input. However, one-point representation
has difficulty in capturing the relationship and the similarity structure of a
huge amount of instances in the real world. For richer classes of the similarity, we
propose the use of weighted point sets, namely, sets of pairs of weight and vector,
as representations of instances. In this work, we theoretically show the benefit
of our proposed method through a new understanding of the contrastive loss of
CLIP, which we call symmetric InfoNCE. We clarify that the optimal similarity
that minimizes symmetric InfoNCE is the pointwise mutual information, and show
an upper bound of excess risk on downstream classification tasks of representations
that achieve the optimal similarity. In addition, we show that our proposed
similarity based on weighted point sets consistently achieves the optimal similarity.
To verify the effectiveness of our proposed method, we demonstrate pretraining of
text-image representation models and classification tasks on common benchmarks. | contrastive learning, representation learning, multimodal representation learning, theoretical analysis, InfoNCE, pointwise mutual information | We propose a new multimodal representation learning method and theoretically show benefits of our method. | 10,847 | 2404.19228 | [
-0.03860699385404587,
-0.19426579773426056,
-0.03070966713130474,
-0.032848894596099854,
0.09070336818695068,
0.02636544778943062,
-0.010997164994478226,
0.01382303237915039,
0.08303996175527573,
-0.07287126034498215,
-0.019487231969833374,
-0.03621785715222359,
0.09818987548351288,
0.11513516306877136,
-0.04471442103385925,
0.02505345270037651,
0.08811766654253006,
0.07497981190681458,
-0.08865487575531006,
0.011328119784593582,
0.02494080550968647,
-0.011008777655661106,
0.037708017975091934,
-0.012722247280180454,
0.08054196089506149,
0.04542069509625435,
0.039403729140758514,
0.021251244470477104,
-0.023768505081534386,
-0.013656732626259327,
0.03801921382546425,
-0.05616822466254234,
0.03369113802909851,
-0.012681511230766773,
-0.09992429614067078,
0.11779652535915375,
-0.04544123634696007,
0.046693768352270126,
-0.03933078423142433,
-0.0072898161597549915,
-0.022264821454882622,
0.00010365262278355658,
-0.02856864407658577,
0.08011649549007416,
0.0083395354449749,
0.04036495089530945,
-0.029631000012159348,
-0.007938198745250702,
0.034548692405223846,
-0.0037425365298986435,
-0.0016377803403884172,
0.057953864336013794,
-0.08503991365432739,
0.01992848329246044,
-0.06098635122179985,
-0.0022439267486333847,
0.025582771748304367,
-0.036402035504579544,
-0.00784212350845337,
-0.028672201558947563,
0.04964853450655937,
-0.10343889892101288,
0.04977409914135933,
-0.02500000037252903,
0.045563071966171265,
-0.04476131126284599,
0.06869146972894669,
-0.031092243269085884,
-0.019475551322102547,
0.00028693213243968785,
-0.05622772499918938,
0.034158993512392044,
-0.0801057294011116,
0.037898581475019455,
-0.033066526055336,
0.07876063883304596,
0.12292861938476562,
-0.06285843253135681,
-0.032835833728313446,
-0.06233574450016022,
0.019936030730605125,
-0.00154030486010015,
0.06563310325145721,
0.021988097578287125,
0.12829133868217468,
0.00959369819611311,
0.020106423646211624,
0.017729729413986206,
0.013708295300602913,
0.025833824649453163,
-0.043330542743206024,
-0.03293299674987793,
0.0044756303541362286,
0.000897283258382231,
0.04955434426665306,
-0.013119489885866642,
0.04219859093427658,
0.004059925209730864,
0.013678752817213535,
0.0786377489566803,
-0.023682935163378716,
0.019123824313282967,
0.014183063060045242,
0.02996990643441677,
0.043755609542131424,
-0.045058127492666245,
0.03306424617767334,
-0.08487308025360107,
0.06026913970708847,
-0.131223663687706,
0.03972099721431732,
-0.0062358081340789795,
-0.010134906508028507,
0.06857406347990036,
0.05003158748149872,
-0.03552867844700813,
0.05028502270579338,
-0.03433369845151901,
0.030640821903944016,
0.049262821674346924,
-0.06978510320186615,
0.0012139107566326857,
-0.02181629277765751,
0.0054012383334338665,
0.0133475661277771,
-0.03819519281387329,
-0.044776517897844315,
7.580429590807508e-35,
0.010330365039408207,
0.015057457610964775,
0.0553281269967556,
-0.015444680117070675,
0.021769391372799873,
-0.018074968829751015,
-0.10774979740381241,
0.044071074575185776,
-0.031030507758259773,
-0.034739114344120026,
0.002798675559461117,
0.022831745445728302,
-0.0013623257400467992,
0.14385828375816345,
-0.01989954523742199,
-0.012169672176241875,
-0.03895669803023338,
-0.03675594553351402,
0.01754041202366352,
-0.02296243980526924,
-0.03245958685874939,
-0.053413718938827515,
-0.0251461211591959,
-0.00900086760520935,
-0.05230587348341942,
-0.015286189503967762,
0.05027971789240837,
-0.11582453548908234,
0.0008963878499343991,
-0.023046541959047318,
-0.14972147345542908,
0.03944074362516403,
0.06720907241106033,
-0.04928187280893326,
0.03611643984913826,
-0.022971922531723976,
0.0022768760100007057,
0.016426270827651024,
0.0393022857606411,
-0.0687144547700882,
0.029780471697449684,
0.052372045814991,
-0.04823768511414528,
-0.03632987663149834,
-0.08814582973718643,
-0.057571891695261,
-0.010796603746712208,
0.021209249272942543,
-0.029230421409010887,
-0.08260445296764374,
0.06889370828866959,
-0.05732579529285431,
-0.030405644327402115,
-0.041221339255571365,
-0.009337100200355053,
-0.07583156228065491,
0.02468564175069332,
-0.01050320453941822,
0.03531044349074364,
0.01286696083843708,
-0.052049219608306885,
0.07567870616912842,
0.08134898543357849,
0.06387706100940704,
-0.039451658725738525,
0.0038564722053706646,
-0.01027332991361618,
0.027467813342809677,
0.07721717655658722,
0.05282106250524521,
-0.03988432511687279,
0.08257542550563812,
-0.059417784214019775,
-0.0879913717508316,
0.007643510587513447,
0.05440910905599594,
0.031309619545936584,
-0.08116868883371353,
-0.012854090891778469,
0.11079718917608261,
-0.05440331622958183,
-0.002122705802321434,
0.029643146321177483,
-0.11322241276502609,
-0.07423067092895508,
-0.02352779172360897,
0.08119715005159378,
-0.13519710302352905,
-0.030156463384628296,
-0.027907341718673706,
-0.04785677045583725,
-0.017139820381999016,
-0.06271553784608841,
0.0011230410309508443,
0.0439794659614563,
-8.827200340057268e-34,
-0.03291268274188042,
0.03542987257242203,
-0.03220486640930176,
0.047327492386102676,
0.044884130358695984,
-0.001489601214416325,
-0.022412072867155075,
0.015483061783015728,
-0.06896722316741943,
0.01203445065766573,
0.012095139361917973,
0.020003195852041245,
-0.02950264886021614,
-0.07755932956933975,
0.011700752191245556,
0.04885242506861687,
0.00218235794454813,
0.08475296944379807,
-0.05038703978061676,
-0.02461405098438263,
0.05971205607056618,
-0.0005386450211517513,
-0.03365124762058258,
0.08790179342031479,
-0.011468756943941116,
0.07279334962368011,
-0.0012504995102062821,
-0.03589139133691788,
-0.03228260949254036,
-0.0575815849006176,
-0.05179804936051369,
0.005365454126149416,
0.05026285722851753,
-0.022706707939505577,
-0.0593697763979435,
0.05187571421265602,
0.06020304188132286,
0.003564954735338688,
-0.045823223888874054,
0.09183985739946365,
0.07737951725721359,
0.0062627303414046764,
-0.04110213741660118,
-0.005992215592414141,
-0.017018115147948265,
-0.08800976723432541,
-0.023750469088554382,
-0.03650735318660736,
0.06245691329240799,
0.0621025450527668,
0.017998404800891876,
-0.0027388606686145067,
-0.08791675418615341,
0.04429824277758598,
0.003848337335512042,
-0.09685976058244705,
-0.03651414066553116,
0.01271025836467743,
0.038638681173324585,
0.05971301719546318,
-0.013669171370565891,
-0.007549648638814688,
-0.06908866763114929,
0.04480406641960144,
-0.013741845265030861,
-0.069731704890728,
0.006422432605177164,
0.012200521305203438,
0.04276224598288536,
0.03607025370001793,
-0.020287204533815384,
0.033111121505498886,
0.02742161974310875,
-0.05626485124230385,
0.058761950582265854,
0.015536992810666561,
0.055164482444524765,
0.024708185344934464,
-0.029649851843714714,
-0.03164353966712952,
0.03178888559341431,
-0.03376384079456329,
0.015206549316644669,
0.0547211691737175,
0.05884420871734619,
0.13104960322380066,
0.059231702238321304,
0.03376691788434982,
0.07103502005338669,
-0.00020709109958261251,
0.008425694890320301,
0.001528513035736978,
-0.009333034977316856,
0.00819855835288763,
-0.06739978492259979,
-4.395664632284024e-8,
-0.08561206609010696,
-0.040624234825372696,
0.04698559269309044,
-0.03601371496915817,
-0.08527473360300064,
-0.06846711039543152,
-0.005383996292948723,
0.0512072816491127,
-0.03851639851927757,
-0.013703147880733013,
0.08961985260248184,
-0.01875300146639347,
-0.010620790533721447,
-0.014377378858625889,
-0.00007787509821355343,
0.01448488887399435,
0.03347468376159668,
0.001952681108377874,
0.04009236767888069,
0.07692623138427734,
-0.005873818416148424,
-0.0032005759421736,
0.0032519912347197533,
-0.015857063233852386,
0.0022787037305533886,
-0.04629719257354736,
-0.0042540705762803555,
0.026660656556487083,
0.13717839121818542,
-0.057841312140226364,
-0.007430723402649164,
0.03528422862291336,
0.015135454945266247,
-0.10123647004365921,
0.01388286892324686,
0.02072412148118019,
0.00029986834852024913,
-0.044456928968429565,
-0.08829820156097412,
0.06127862632274628,
0.02722691372036934,
-0.005859653931111097,
0.030810393393039703,
-0.00020247903012204915,
0.10926225781440735,
0.09379370510578156,
-0.015881413593888283,
-0.0012230221182107925,
0.001257154275663197,
-0.0633428618311882,
-0.025512874126434326,
0.010466818697750568,
0.020161746069788933,
0.06340917199850082,
0.05696461722254753,
-0.03874994069337845,
-0.0030666848178952932,
0.036614954471588135,
0.09256720542907715,
-0.026800764724612236,
-0.04554860293865204,
0.02417466975748539,
-0.06283076107501984,
0.01649363711476326
] |
Can LLMs Really Learn to Translate a Low-Resource Language from One Grammar Book? | https://openreview.net/forum?id=aMBSY2ebPw | [
"Seth Aycock",
"David Stap",
"Di Wu",
"Christof Monz",
"Khalil Sima'an"
] | Spotlight | Extremely low-resource (XLR) languages lack substantial corpora for training NLP models, motivating the use of all available resources such as dictionaries and grammar books. Machine Translation from One Book (Tanzer et al., 2024) suggests that prompting long-context LLMs with one grammar book enables English–Kalamang translation, an XLR language unseen by LLMs—a noteworthy case of linguistics helping an NLP task. We investigate the source of this translation ability, finding almost all improvements stem from the book’s parallel examples rather than its grammatical explanations. We find similar results for Nepali and Guarani, seen low-resource languages, and we achieve performance comparable to an LLM with a grammar book by simply fine-tuning an encoder-decoder translation model. We then investigate where grammar books help by testing two linguistic tasks, grammaticality judgment and gloss prediction, and we explore what kind of grammatical knowledge helps by introducing a typological feature prompt that achieves leading results on these more relevant tasks. We thus emphasise the importance of task-appropriate data for XLR languages: parallel examples for translation, and grammatical data for linguistic tasks. As we find no evidence that long-context LLMs can make effective use of grammatical explanations for XLR translation, we conclude data collection for multilingual XLR tasks such as translation is best focused on parallel data over linguistic description. | llms, translation, low-resource, grammar, long-context, linguistics | We show LLMs fail to exploit grammatical explanations for translation, but benefit from typological knowledge for linguistic tasks. | 10,831 | 2409.19151 | [
-0.04531392082571983,
0.001891577267087996,
0.0309612900018692,
-0.014733274467289448,
0.03737197443842888,
-0.06339138001203537,
-0.0040644025430083275,
0.01578792929649353,
0.05401873588562012,
0.014766072854399681,
-0.012323632836341858,
-0.08508756011724472,
0.031282875686883926,
-0.028490616008639336,
0.054751381278038025,
-0.012886888347566128,
0.06313962489366531,
0.04679269343614578,
-0.07421647757291794,
-0.10040566325187683,
0.07211871445178986,
0.05101373791694641,
0.03132365271449089,
0.005471763666719198,
0.058893345296382904,
-0.022980282083153725,
-0.043857309967279434,
-0.04508248716592789,
0.03448301553726196,
0.009633840061724186,
-0.005139870569109917,
0.01141215581446886,
-0.05710960924625397,
0.09117002785205841,
-0.041784580796957016,
0.0994691401720047,
-0.03005646914243698,
0.013637593947350979,
0.01847781427204609,
-0.012394435703754425,
-0.03463614732027054,
0.012736150063574314,
-0.004139523487538099,
-0.030773211270570755,
0.043471455574035645,
0.045798324048519135,
0.01059806440025568,
-0.022578343749046326,
-0.06129230186343193,
-0.005119387526065111,
-0.08800825476646423,
-0.017512978985905647,
0.0245620496571064,
-0.008817775174975395,
0.028844086453318596,
-0.027635548263788223,
-0.03073083609342575,
0.02048763819038868,
-0.009750774130225182,
-0.07287489622831345,
-0.08086787909269333,
-0.07815361768007278,
-0.05329334735870361,
0.033110830932855606,
0.021176708862185478,
-0.07898223400115967,
0.039397917687892914,
0.07564526051282883,
-0.09095734357833862,
0.01724998466670513,
-0.06356392800807953,
0.029246000573039055,
-0.004483710043132305,
0.11670850217342377,
-0.06940647214651108,
-0.015950152650475502,
-0.0032542708795517683,
0.022547557950019836,
0.01513354480266571,
-0.06978286802768707,
0.026953479275107384,
0.05879514291882515,
0.019515885040163994,
0.05043772608041763,
0.001267836894840002,
-0.016662705689668655,
0.03203698247671127,
0.04788036271929741,
0.03760500252246857,
-0.006141304038465023,
0.03137865662574768,
-0.054625142365694046,
0.016805246472358704,
0.05610791966319084,
-0.01424234639853239,
0.057577986270189285,
0.00519300764426589,
-0.029789384454488754,
-0.00966996792703867,
0.07410448044538498,
0.05692388862371445,
0.09583496302366257,
0.06875074654817581,
-0.10044929385185242,
-0.12007218599319458,
-0.06854205578565598,
-0.008960659615695477,
0.023189548403024673,
0.03905976563692093,
-0.08727819472551346,
0.013148230500519276,
0.08723378926515579,
0.024904889985919,
0.020536823198199272,
0.03594397008419037,
-0.04506077617406845,
0.037228360772132874,
-0.047466251999139786,
0.0568450503051281,
0.01978495344519615,
-0.07844512909650803,
0.01483192853629589,
0.08772267401218414,
0.010183671489357948,
0.020200151950120926,
-0.09073703736066818,
-0.009804295375943184,
6.7943426540734365e-34,
0.10275860130786896,
0.08203031867742538,
-0.023636169731616974,
-0.001792746945284307,
0.048178210854530334,
-0.05662195757031441,
0.02618742361664772,
0.058477453887462616,
-0.060859210789203644,
0.011991997249424458,
-0.0012852043146267533,
0.034253012388944626,
-0.04657438024878502,
0.042803775519132614,
-0.02167171984910965,
0.06770538538694382,
0.021065888926386833,
0.009369511157274246,
-0.03337240219116211,
0.04012781009078026,
0.03949671611189842,
0.0074121830984950066,
0.031118255108594894,
-0.07322085648775101,
0.026365531608462334,
-0.005198672879487276,
0.06957294046878815,
-0.12655988335609436,
0.022279856726527214,
0.01003906037658453,
-0.09072448313236237,
-0.06008805334568024,
0.02824525348842144,
0.007476421073079109,
-0.00007429988181684166,
-0.09621215611696243,
-0.05947836861014366,
-0.03839414566755295,
-0.0014001845847815275,
-0.0492573007941246,
-0.00044333323603495955,
-0.006696460768580437,
-0.0015306472778320312,
0.03957473486661911,
-0.016524817794561386,
0.008715575560927391,
-0.005725028924643993,
-0.0024702083319425583,
0.004785379394888878,
-0.04530400410294533,
0.075218066573143,
0.010053936392068863,
-0.08879218995571136,
-0.03402455151081085,
0.0817321389913559,
0.07297901809215546,
-0.018044155091047287,
0.04857909306883812,
-0.031059980392456055,
0.040067873895168304,
0.004909672308713198,
0.03123519942164421,
0.025202970951795578,
0.06304703652858734,
0.042819418013095856,
-0.018723921850323677,
-0.03909244015812874,
0.05387412756681442,
0.025142693892121315,
-0.03560452163219452,
-0.06700117886066437,
-0.055591803044080734,
0.011517766863107681,
0.009833980351686478,
0.13373985886573792,
0.014955762773752213,
0.011782984249293804,
-0.11693673580884933,
0.021444136276841164,
0.008201148360967636,
0.01837611384689808,
-0.062248531728982925,
0.04856562241911888,
-0.031210826709866524,
-0.053559765219688416,
-0.08443396538496017,
0.05487737059593201,
-0.035935766994953156,
0.026410292834043503,
-0.00415307842195034,
-0.025761082768440247,
-0.05143483728170395,
-0.10479302704334259,
0.005416348110884428,
0.01979653909802437,
-1.6188182699104895e-33,
-0.009139993228018284,
-0.06312046200037003,
-0.0388474278151989,
0.09680385887622833,
-0.048656873404979706,
-0.0385514572262764,
0.013198665343225002,
0.0277890432626009,
0.012245445512235165,
-0.044560570269823074,
-0.027403641492128372,
-0.0430891178548336,
0.06052679941058159,
0.10141163319349289,
0.030422378331422806,
-0.05190291255712509,
0.050792574882507324,
0.056131064891815186,
0.005152364727109671,
0.09234829246997833,
-0.014920218847692013,
0.15721175074577332,
-0.08714758604764938,
0.028449054807424545,
-0.0209251269698143,
-0.007091077044606209,
-0.053169772028923035,
0.06043930724263191,
-0.08825252950191498,
0.02244626358151436,
-0.006186298094689846,
-0.020128589123487473,
-0.02154594473540783,
-0.02664806880056858,
-0.11030586808919907,
0.01964191161096096,
0.04097217321395874,
-0.0315251387655735,
-0.06065855920314789,
0.07686920464038849,
0.06748486310243607,
0.03644084557890892,
0.03426836431026459,
-0.0761115625500679,
-0.06696749478578568,
-0.02020028606057167,
-0.06174904853105545,
-0.09461012482643127,
0.04938749596476555,
0.004750568885356188,
0.020800471305847168,
-0.014843599870800972,
-0.033579736948013306,
-0.02555026486515999,
0.05713233724236488,
-0.10583465546369553,
-0.012407161295413971,
-0.1562778204679489,
-0.010478230193257332,
-0.0880943089723587,
-0.07783085107803345,
-0.012545223347842693,
0.0542229525744915,
-0.010542068630456924,
0.045262448489665985,
-0.0276498906314373,
0.008594742976129055,
-0.004722172394394875,
0.08109211921691895,
-0.05173086002469063,
0.05458882451057434,
-0.07162001729011536,
-0.009248541668057442,
0.03407609462738037,
0.011319588869810104,
0.0477638766169548,
0.02220248617231846,
-0.0774575024843216,
-0.02029980719089508,
-0.12196377664804459,
-0.007156782783567905,
-0.014725982211530209,
-0.0022655839566141367,
-0.009781216271221638,
0.034644003957509995,
0.06079140678048134,
0.03868691250681877,
0.07265456765890121,
0.023371713235974312,
0.021924439817667007,
-0.0024658571928739548,
0.06095017120242119,
0.0008177096606232226,
0.11328556388616562,
-0.03067619912326336,
-4.171344514247721e-8,
-0.08879624307155609,
0.025437787175178528,
-0.04343170300126076,
0.01771220751106739,
0.022190948948264122,
-0.0433703176677227,
-0.033311646431684494,
0.06277907639741898,
-0.025014089420437813,
-0.043858565390110016,
0.046583883464336395,
0.029977360740303993,
-0.06492684036493301,
-0.013031805865466595,
0.012626351788640022,
0.07519519329071045,
0.06222953647375107,
-0.006680521182715893,
-0.009749931283295155,
-0.013532964512705803,
0.0671052485704422,
0.03533884882926941,
-0.0025815817061811686,
0.009838123805820942,
0.04083384573459625,
-0.014053563587367535,
-0.04326393082737923,
0.0652119442820549,
0.08095906674861908,
-0.0717262551188469,
-0.025604451075196266,
0.010261956602334976,
-0.02047962322831154,
-0.007343166042119265,
0.04406783729791641,
0.061937060207128525,
-0.039764732122421265,
-0.011758891865611076,
0.08937028050422668,
0.030933251604437828,
0.09369377046823502,
0.049029767513275146,
-0.05497816577553749,
0.0735917016863823,
0.05397308990359306,
-0.03885721042752266,
-0.05403458699584007,
-0.10132644325494766,
-0.01880502700805664,
-0.08354680985212326,
0.07816759496927261,
-0.08909814804792404,
0.03596825152635574,
0.028314361348748207,
0.0356130413711071,
0.043121322989463806,
-0.07559502869844437,
0.00701082032173872,
0.025802286341786385,
0.011362585239112377,
0.028133025392889977,
0.021237626671791077,
-0.04653957486152649,
0.01098068617284298
] |
Retri3D: 3D Neural Graphics Representation Retrieval | https://openreview.net/forum?id=q3EbOXb4y1 | [
"Yushi Guan",
"Daniel Kwan",
"Jean Sebastien Dandurand",
"Xi Yan",
"Ruofan Liang",
"Yuxuan Zhang",
"Nilesh Jain",
"Nilesh Ahuja",
"Selvakumar Panneer",
"Nandita Vijaykumar"
] | Spotlight | Learnable 3D Neural Graphics Representations (3DNGR) have emerged as promising 3D representations for reconstructing 3D scenes from 2D images. Numerous works, including Neural Radiance Fields (NeRF), 3D Gaussian Splatting (3DGS), and their variants, have significantly enhanced the quality of these representations. The ease of construction from 2D images, suitability for online viewing/sharing, and applications in game/art design downstream tasks make it a vital 3D representation, with potential creation of large numbers of such 3D models. This necessitates large data stores, local or online, to save 3D visual data in these formats. However, no existing framework enables accurate retrieval of stored 3DNGRs. In this work, we propose, Retri3D, a framework that enables accurate and efficient retrieval of 3D scenes represented as NGRs from large data stores using text queries. We introduce a novel Neural Field Artifact Analysis technique, combined with a Smart Camera Movement Module, to select clean views and navigate pre-trained 3DNGRs. These techniques enable accurate retrieval by selecting the best viewing directions in the 3D scene for high-quality visual feature embeddings. We demonstrate that Retri3D is compatible with any NGR representation. On the LERF and ScanNet++ datasets, we show significant improvement in retrieval accuracy compared to existing techniques, while being orders of magnitude faster and storage efficient. | Neural Graphics Representation; 3D Retrieval; Database | A framework for the retrieval of neural graphics representation | 10,814 | null | [
-0.06494531035423279,
-0.07782367616891861,
0.02856508456170559,
-0.004332990385591984,
0.05098776891827583,
-0.0041488404385745525,
-0.04046414792537689,
-0.0767269879579544,
0.001445619622245431,
-0.053937047719955444,
-0.025846611708402634,
-0.042130276560783386,
0.002338543999940157,
0.08453138172626495,
-0.012288954108953476,
-0.002631677780300379,
0.08238635212182999,
0.09289396554231644,
-0.047972794622182846,
0.014453449286520481,
-0.008302085101604462,
-0.04192608967423439,
0.0027935148682445288,
-0.030306290835142136,
0.07281922549009323,
0.0287527646869421,
0.04477639123797417,
-0.002277644816786051,
0.014858639799058437,
0.00780258234590292,
0.01750969886779785,
0.030415745452046394,
-0.11453139781951904,
0.04506968706846237,
-0.0358818881213665,
0.06041579693555832,
-0.05953656882047653,
0.01118991244584322,
-0.06178350746631622,
-0.05205955728888512,
-0.0053032333962619305,
0.018119163811206818,
-0.026294102892279625,
0.03600401058793068,
0.1082155779004097,
0.06405100971460342,
-0.01077542919665575,
-0.09092805534601212,
0.05238117277622223,
0.012444760650396347,
-0.09177745878696442,
-0.06480501592159271,
-0.02578587457537651,
0.03184910863637924,
0.07670795172452927,
0.10540152341127396,
0.02466944232583046,
-0.046979378908872604,
-0.0030877559911459684,
0.06355688720941544,
0.01918288879096508,
-0.004723642952740192,
0.046411510556936264,
-0.037197310477495193,
-0.01832890696823597,
-0.00260655814781785,
0.02539585717022419,
-0.051423702389001846,
-0.022875018417835236,
-0.07529119402170181,
-0.03976103290915489,
0.06094072759151459,
-0.05327894166111946,
-0.052621953189373016,
-0.02863362617790699,
0.06678096204996109,
-0.05146497115492821,
-0.0659460797905922,
0.003577020950615406,
-0.07546009868383408,
0.03428904339671135,
0.009759471751749516,
0.0382327176630497,
0.06585810333490372,
0.07875456660985947,
-0.030791111290454865,
0.033882077783346176,
0.07284996658563614,
0.004576798994094133,
-0.014896643348038197,
-0.037395134568214417,
0.0010354631813243032,
0.011686384677886963,
-0.04737585037946701,
0.013521580956876278,
-0.02120114304125309,
0.01895006000995636,
-0.0709645077586174,
0.029986431822180748,
-0.006641545798629522,
0.000802316761109978,
-0.019625242799520493,
0.0951409637928009,
0.01869424805045128,
0.03947103023529053,
0.029517540708184242,
0.10066942125558853,
0.06210193783044815,
0.02835032157599926,
0.0561247281730175,
0.03209758922457695,
0.06581589579582214,
-0.04138325899839401,
-0.026297366246581078,
0.01859709620475769,
0.054477863013744354,
-0.08427154272794724,
0.0074149793945252895,
0.037983689457178116,
0.04383393004536629,
0.04758913442492485,
0.0034124029334634542,
-0.041715897619724274,
-0.06393883377313614,
0.06933308392763138,
-0.05245884880423546,
-0.028871867805719376,
2.680481421218993e-33,
0.014378322288393974,
0.06911049038171768,
0.013618256896734238,
0.01948280818760395,
0.02671617642045021,
-0.02482510171830654,
-0.03091658465564251,
0.09900176525115967,
0.008832457475364208,
-0.013497698120772839,
-0.06853598356246948,
0.0361458845436573,
-0.019974684342741966,
0.12276346236467361,
0.030251767486333847,
-0.042287811636924744,
-0.07708793133497238,
-0.0532943531870842,
-0.007150063291192055,
0.08203361183404922,
0.01164379995316267,
-0.07595406472682953,
0.026937929913401604,
-0.008084872737526894,
-0.05919639393687248,
0.0008296354208141565,
-0.04274207726120949,
-0.008316516876220703,
-0.052421752363443375,
-0.027900107204914093,
0.0012259094510227442,
-0.01852651871740818,
0.012953629717230797,
-0.03645414113998413,
0.011883017607033253,
-0.04501333087682724,
-0.0029300751630216837,
-0.05205761268734932,
-0.027989711612462997,
0.025707319378852844,
0.03194381296634674,
0.050463560968637466,
-0.03479618951678276,
-0.055414337664842606,
-0.06628408282995224,
0.03093712031841278,
0.0757349506020546,
-0.014442580752074718,
-0.04468139633536339,
0.00020250312809366733,
0.09717322885990143,
0.03397795930504799,
-0.03811998292803764,
-0.0689624771475792,
0.008310215547680855,
-0.044203873723745346,
0.0026538230013102293,
-0.007931644096970558,
0.06408687680959702,
-0.0031200735829770565,
-0.023295123130083084,
0.013552883639931679,
0.018341148272156715,
0.044378094375133514,
0.0012098539154976606,
0.005821208469569683,
0.07346702367067337,
0.07374592125415802,
0.032692886888980865,
0.012307769618928432,
-0.006842194590717554,
0.004726835992187262,
0.03810394927859306,
-0.04197298362851143,
-0.013367471285164356,
-0.027418458834290504,
0.022718094289302826,
-0.00985187478363514,
0.030948596075177193,
-0.045773543417453766,
-0.06433860212564468,
-0.04643687605857849,
-0.0018588119419291615,
-0.014375765807926655,
-0.1042604073882103,
-0.06984398514032364,
0.06006055325269699,
-0.11647168546915054,
-0.0039649526588618755,
-0.07559134066104889,
-0.0450221486389637,
-0.07845283299684525,
-0.0665321797132492,
-0.09851743280887604,
-0.03202839940786362,
-1.778031816560668e-33,
0.03686882182955742,
-0.024833299219608307,
-0.07828544825315475,
0.006169429048895836,
-0.030581006780266762,
-0.025864191353321075,
-0.008380189538002014,
0.07462208718061447,
0.00795500073581934,
-0.05348622053861618,
-0.004981929436326027,
0.07683947682380676,
0.027353746816515923,
-0.03679496794939041,
0.050711773335933685,
-0.012161225080490112,
-0.0352250337600708,
-0.06637164950370789,
-0.061862245202064514,
0.0005611606175079942,
0.04417310282588005,
0.040783416479825974,
-0.06421227008104324,
-0.02055497281253338,
-0.06217392906546593,
0.13553893566131592,
-0.011053759604692459,
-0.03945015370845795,
-0.023153888061642647,
0.01311072614043951,
-0.009492659941315651,
-0.020921172574162483,
0.03454916924238205,
-0.06648366898298264,
-0.015273175202310085,
-0.0055290344171226025,
0.10511264204978943,
-0.03721965104341507,
-0.05989508330821991,
0.0011276027653366327,
0.07874056696891785,
0.0130428122356534,
-0.056070927530527115,
0.0808076411485672,
-0.05867531895637512,
0.038230881094932556,
-0.019194893538951874,
0.029743783175945282,
0.028938518837094307,
0.07299089431762695,
-0.041036009788513184,
0.03441794216632843,
-0.08792699128389359,
-0.0937577411532402,
0.032741472125053406,
-0.08543043583631516,
-0.08748547732830048,
-0.04061511904001236,
0.07539169490337372,
0.002918959828093648,
0.0255128163844347,
-0.04480789601802826,
-0.08625342696905136,
-0.0026500504463911057,
-0.006410268601030111,
0.040119022130966187,
-0.03272656351327896,
-0.0029654288664460182,
-0.00865212082862854,
0.0510864220559597,
-0.0499364510178566,
0.034475769847631454,
0.04075421765446663,
-0.016946710646152496,
-0.03118530847132206,
0.0039163315668702126,
0.02264656499028206,
0.05341734364628792,
0.07319853454828262,
0.010835894383490086,
0.09573160111904144,
-0.0643724650144577,
0.027490539476275444,
0.15768155455589294,
0.10344573855400085,
0.04970357567071915,
-0.059220295399427414,
0.017702611163258553,
0.04007713496685028,
-0.07461325824260712,
-0.012670081108808517,
0.04162529483437538,
-0.004687326028943062,
0.09908056259155273,
0.05650263652205467,
-4.0433395298578034e-8,
-0.0525226816534996,
0.09029284119606018,
0.044537805020809174,
-0.04380122572183609,
-0.013347804546356201,
-0.018826283514499664,
0.010737291537225246,
0.10325848311185837,
-0.022423021495342255,
-0.01579405926167965,
0.03543764352798462,
-0.121865414083004,
-0.03019857034087181,
-0.06116705387830734,
0.11242716014385223,
0.13434135913848877,
0.023067796602845192,
-0.03687397390604019,
0.0009789452888071537,
0.04011259600520134,
-0.0011549657210707664,
0.06609567999839783,
0.037961445748806,
-0.030091824010014534,
-0.026657309383153915,
0.04194953292608261,
-0.0637013167142868,
0.05167977139353752,
0.07959649711847305,
-0.027840115129947662,
0.059883106499910355,
0.009827587753534317,
0.1639007031917572,
-0.007187677547335625,
-0.008436528034508228,
-0.009075512178242207,
0.021973976865410805,
-0.0286119282245636,
0.011713298037648201,
-0.09185361862182617,
-0.018700843676924706,
0.05429144203662872,
-0.0014411727897822857,
-0.0032036886550486088,
-0.007288321387022734,
-0.02065851166844368,
0.07999330759048462,
-0.08889639377593994,
0.048772286623716354,
0.015331896021962166,
-0.12208007276058197,
-0.047766607254743576,
-0.014419091865420341,
0.08831886947154999,
0.06215779855847359,
-0.014950989745557308,
-0.02764781564474106,
0.00853500235825777,
0.050933267921209335,
0.018549537286162376,
0.012051265686750412,
0.008242662064731121,
-0.0666976273059845,
0.021176794543862343
] |
Test-time Adaptation for Cross-modal Retrieval with Query Shift | https://openreview.net/forum?id=BmG88rONaU | [
"Haobin Li",
"Peng Hu",
"Qianjun Zhang",
"Xi Peng",
"XitingLiu",
"Mouxing Yang"
] | Spotlight | The success of most existing cross-modal retrieval methods heavily relies on the assumption that the given queries follow the same distribution of the source domain.
However, such an assumption is easily violated in real-world scenarios due to the complexity and diversity of queries, thus leading to the query shift problem.
Specifically, query shift refers to the online query stream originating from the domain that follows a different distribution with the source one.
In this paper, we observe that query shift would not only diminish the uniformity (namely, within-modality scatter) of the query modality but also amplify the gap between query and gallery modalities.
Based on the observations, we propose a novel method dubbed Test-time adaptation for Cross-modal Retrieval (TCR).
In brief, TCR employs a novel module to refine the query predictions (namely, retrieval results of the query) and a joint objective to prevent query shift from disturbing the common space, thus achieving online adaptation for the cross-modal retrieval models with query shift.
Expensive experiments demonstrate the effectiveness of the proposed TCR against query shift.
Code is available at https://github.com/XLearning-SCU/2025-ICLR-TCR. | Test-time adaptation, Cross-modal retrieval, Query shift | We propose a novel test-time adaptation method for achieving robust cross-modal retrieval against query shift. | 10,708 | 2410.15624 | [
-0.0855100154876709,
-0.1013781949877739,
0.019116373732686043,
0.08312695473432541,
0.12119214236736298,
0.0010098087368533015,
0.0027453401125967503,
-0.04789607599377632,
-0.01649695262312889,
-0.11289260536432266,
0.00035556056536734104,
-0.03695829212665558,
0.07755459100008011,
0.0132235586643219,
0.0429476834833622,
-0.06368858367204666,
0.09823133796453476,
0.06586674600839615,
-0.09198599308729172,
-0.023333553224802017,
0.034194216132164,
0.030495833605527878,
0.027843786403536797,
-0.029004927724599838,
0.031510040163993835,
-0.016499508172273636,
-0.05423765257000923,
-0.013183820061385632,
0.04655614122748375,
-0.014344691298902035,
0.003946701996028423,
0.09003830701112747,
-0.0965999960899353,
0.09377558529376984,
-0.06396348774433136,
-0.05669966712594032,
-0.10241837054491043,
0.03577510640025139,
0.030022742226719856,
-0.059026386588811874,
-0.0295814611017704,
-0.014052974060177803,
-0.010692107491195202,
0.030107788741588593,
-0.04987896978855133,
-0.0129321264103055,
-0.04182971641421318,
-0.02210810035467148,
0.03153901919722557,
0.06017019972205162,
-0.13028565049171448,
-0.011111038736999035,
-0.03357662633061409,
0.039916910231113434,
0.004362060688436031,
0.011223619803786278,
0.009544309228658676,
-0.004084853921085596,
0.007121735252439976,
0.029203051701188087,
-0.01377594843506813,
-0.10157427191734314,
0.01460175309330225,
0.04354044795036316,
0.026382973417639732,
-0.0486360527575016,
0.03602159023284912,
-0.008982202038168907,
0.06432108581066132,
-0.08936364948749542,
-0.0856122374534607,
0.04260321334004402,
-0.09668592363595963,
0.0845055803656578,
-0.003725469345226884,
0.05895473062992096,
-0.03513726592063904,
-0.030100800096988678,
-0.060258813202381134,
-0.03645556792616844,
0.024241313338279724,
-0.0019319964339956641,
0.03570806235074997,
0.039700947701931,
0.1496419608592987,
-0.055919088423252106,
0.030732890591025352,
-0.03773648291826248,
-0.04754010587930679,
-0.03297320753335953,
-0.039615046232938766,
0.01334447506815195,
0.00379541446454823,
0.013350091874599457,
-0.01007323432713747,
0.0010769834043458104,
0.03329901769757271,
-0.04941270127892494,
0.08146682381629944,
0.09638126194477081,
0.048349861055612564,
0.08182088285684586,
-0.0048279124312102795,
-0.05415484309196472,
-0.043167680501937866,
-0.02005285583436489,
0.05707710608839989,
0.043291784822940826,
-0.0017134620575234294,
-0.03146984800696373,
0.010954669676721096,
0.11171350628137589,
0.05251595005393028,
-0.10613271594047546,
-0.045354366302490234,
-0.015458272770047188,
-0.002177290851250291,
-0.02057700604200363,
0.03465498611330986,
-0.02408500947058201,
0.08376851677894592,
0.06397160142660141,
0.001795756397768855,
-0.08060476928949356,
0.047731611877679825,
0.035558391362428665,
-0.04428315907716751,
1.5309534650998814e-33,
0.08874455839395523,
-0.010294563136994839,
0.05106011778116226,
0.03357332944869995,
0.005955425091087818,
0.005217705387622118,
0.010865144431591034,
0.08150159567594528,
-0.020702801644802094,
-0.044566165655851364,
-0.02263447269797325,
0.04312436655163765,
-0.007498032879084349,
-0.015916947275400162,
0.04466976225376129,
0.018360702320933342,
-0.0322553850710392,
0.005138622131198645,
0.011770859360694885,
0.029629383236169815,
-0.018790652975440025,
-0.033711474388837814,
0.04477274417877197,
0.046208251267671585,
-0.003951475024223328,
0.02270139753818512,
0.034658920019865036,
-0.01182530727237463,
-0.031861741095781326,
0.014865580014884472,
-0.09401474893093109,
-0.022091126069426537,
-0.04106242582201958,
0.007632039487361908,
0.0703621432185173,
-0.04548356682062149,
-0.020557167008519173,
-0.044397834688425064,
-0.048333462327718735,
0.023950938135385513,
-0.07167405635118484,
-0.04526522383093834,
-0.01815595105290413,
-0.022703275084495544,
-0.08271993696689606,
-0.03772328048944473,
-0.08964093774557114,
-0.00012080613669240847,
-0.01948416978120804,
-0.01623387075960636,
0.08585700392723083,
0.021201765164732933,
-0.03355608135461807,
-0.06526012718677521,
0.04818020015954971,
-0.026108816266059875,
-0.0049564721994102,
0.03815654665231705,
0.0655067041516304,
0.07700013369321823,
0.023269779980182648,
-0.02095174230635166,
0.022706326097249985,
0.034623704850673676,
0.01864907518029213,
0.03450663387775421,
-0.0182934682816267,
0.025684135034680367,
0.059510938823223114,
0.046271294355392456,
0.025555351749062538,
0.02457139641046524,
-0.028165310621261597,
0.009481928311288357,
0.011376336216926575,
-0.001418215804733336,
-0.019241200760006905,
-0.038615331053733826,
0.08766157925128937,
-0.038750745356082916,
0.044805873185396194,
-0.018948113545775414,
0.03286353871226311,
-0.03995179384946823,
-0.054653432220220566,
-0.07813195884227753,
0.030140025541186333,
-0.10813818126916885,
-0.046106159687042236,
-0.08961698412895203,
0.042359333485364914,
0.08243189007043839,
-0.015994049608707428,
-0.019038992002606392,
0.07588028907775879,
-1.369909545175778e-33,
0.013372663408517838,
-0.05083790794014931,
-0.0007743376190774143,
0.04983832687139511,
0.02882254496216774,
-0.08641006052494049,
0.03212537616491318,
0.004551301710307598,
-0.01042596623301506,
-0.012128759175539017,
-0.02069641649723053,
-0.07944531738758087,
-0.01491378340870142,
-0.0602618046104908,
-0.026858801022171974,
0.1174197643995285,
0.04452468827366829,
-0.059585802257061005,
-0.1264234185218811,
0.11032062768936157,
0.034576065838336945,
0.08682772517204285,
-0.09959029406309128,
-0.013274746015667915,
-0.07176241278648376,
0.021958965808153152,
-0.02160840481519699,
0.030515730381011963,
-0.062002167105674744,
-0.04716480150818825,
-0.003850372740998864,
-0.055644597858190536,
-0.055236805230379105,
-0.019774852320551872,
-0.006113475654274225,
0.07989625632762909,
0.0071349553763866425,
-0.048364557325839996,
-0.04513317719101906,
0.1501050889492035,
-0.001639818656258285,
0.02425757423043251,
0.010401884093880653,
-0.005939775612205267,
0.012203431688249111,
0.0019144391408190131,
-0.10885272920131683,
0.000788543839007616,
0.029866622760891914,
0.03519245982170105,
0.0336468406021595,
0.0032302222680300474,
-0.006772768683731556,
-0.03505218029022217,
0.048349618911743164,
-0.05909198895096779,
-0.09379994869232178,
0.020658204331994057,
0.016586273908615112,
0.057514309883117676,
-0.023678038269281387,
-0.005982891656458378,
-0.01666041649878025,
-0.028693007305264473,
0.05037596821784973,
-0.06472805142402649,
0.015801846981048584,
-0.06105085834860802,
-0.042473405599594116,
0.011801482178270817,
0.0412290096282959,
-0.08510410040616989,
0.04191352427005768,
-0.04733400419354439,
0.07481028139591217,
0.02744997851550579,
0.11645754426717758,
0.04104607552289963,
-0.002317796926945448,
-0.010498310439288616,
-0.03328942507505417,
0.03751276060938835,
0.03636612370610237,
0.030873939394950867,
-0.03784814476966858,
0.14192283153533936,
-0.023875931277871132,
0.061809588223695755,
0.0815793052315712,
-0.09607644379138947,
0.00791123602539301,
-0.04696238785982132,
-0.010778931900858879,
0.05422934889793396,
0.08963438868522644,
-3.924686353684592e-8,
-0.0606493204832077,
0.0799560621380806,
-0.1035175696015358,
-0.013700613752007484,
0.0352354571223259,
0.01454152725636959,
0.006406006868928671,
0.07430215179920197,
-0.03293708339333534,
-0.041645485907793045,
0.06797268986701965,
-0.016527744010090828,
0.061589568853378296,
0.0261674877256155,
0.0173663217574358,
-0.03873367980122566,
-0.03168872371315956,
-0.011586135253310204,
-0.023099638521671295,
-0.07253286242485046,
-0.024259166792035103,
0.05119328200817108,
0.08243589848279953,
-0.01634237729012966,
-0.019480209797620773,
-0.014601771719753742,
-0.051801227033138275,
0.0581912100315094,
0.05988536402583122,
-0.04741918295621872,
-0.02447894774377346,
0.0027817613445222378,
-0.01586875692009926,
-0.03885680437088013,
0.020469602197408676,
0.038898032158613205,
-0.011140350252389908,
-0.021162087097764015,
-0.02443975768983364,
0.014464637264609337,
0.07295281440019608,
-0.03118082322180271,
-0.06896796077489853,
0.06976325064897537,
0.04181651398539543,
-0.004319443833082914,
0.08773826062679291,
-0.019533369690179825,
0.03456803038716316,
-0.07028442621231079,
-0.020004360005259514,
-0.09924515336751938,
0.006069559138268232,
0.031734175980091095,
0.07375869154930115,
-0.017169862985610962,
-0.012579831294715405,
0.045827716588974,
0.08873900771141052,
0.020475920289754868,
0.0015955273993313313,
-0.0006126367370598018,
-0.04957832396030426,
0.028559135273098946
] |
Universal generalization guarantees for Wasserstein distributionally robust models | https://openreview.net/forum?id=0h6v4SpLCY | [
"Tam Le",
"Jerome Malick"
] | Spotlight | Distributionally robust optimization has emerged as an attractive way to train robust machine learning models, capturing data uncertainty and distribution shifts. Recent statistical analyses have proved that generalization guarantees of robust models based on the Wasserstein distance have generalization guarantees that do not suffer from the curse of dimensionality. However, these results are either approximate, obtained in specific cases, or based on assumptions difficult to verify in practice. In contrast, we establish exact generalization guarantees that cover a wide range of cases, with arbitrary transport costs and parametric loss functions, including deep learning objectives with nonsmooth activations. We complete our analysis with an excess bound on the robust objective and an extension to Wasserstein robust models with entropic regularizations. | generalization guarantees, optimal transport, distributionally robust optimization, nonsmooth analysis | Exact generalization guarantees for Wasserstein distributionally robust models with dimension-free sample rates. | 10,606 | 2402.11981 | [
-0.05110188573598862,
-0.09663243591785431,
0.03965427353978157,
0.04352704435586929,
0.03989091515541077,
0.07718656212091446,
-0.002825560513883829,
-0.004957592114806175,
-0.03826536983251572,
-0.01081027276813984,
0.016526373103260994,
-0.02700929343700409,
0.033984772861003876,
0.06107595935463905,
-0.10939183086156845,
-0.06192133575677872,
0.12026342004537582,
0.03925560042262077,
-0.17778326570987701,
0.03916283696889877,
-0.031139757484197617,
0.07198850065469742,
-0.002277475083246827,
0.02345641516149044,
0.05483792722225189,
-0.05290104076266289,
0.04028138890862465,
-0.021233564242720604,
-0.01525465864688158,
-0.015357245691120625,
0.005487538408488035,
-0.10961561650037766,
-0.024372266605496407,
-0.014305810444056988,
0.02126385271549225,
0.10361890494823456,
-0.057379912585020065,
0.06540283560752869,
0.04915874823927879,
0.0019363807514309883,
-0.05360855907201767,
0.018913742154836655,
0.02875843271613121,
-0.014327801764011383,
-0.03531121090054512,
-0.012661240994930267,
0.025125423446297646,
-0.04796254634857178,
0.025011632591485977,
0.045663874596357346,
0.04521968215703964,
0.0069664656184613705,
-0.022596288472414017,
0.057369280606508255,
0.006790890824049711,
-0.036453697830438614,
0.01863596774637699,
-0.005957876332104206,
-0.032278504222631454,
0.05611832067370415,
0.019430510699748993,
-0.11897804588079453,
-0.0007164367125369608,
-0.004602879751473665,
0.03062315285205841,
0.001518342411145568,
0.011957318522036076,
0.00996587984263897,
-0.01724398136138916,
0.04720507189631462,
-0.04190841317176819,
0.11160908639431,
-0.10699259489774704,
0.03461321443319321,
0.06126193329691887,
0.026278609409928322,
0.06286705285310745,
0.02431556023657322,
0.05768970027565956,
-0.031527984887361526,
0.07466856390237808,
0.035031504929065704,
-0.005157557316124439,
0.01828010194003582,
0.10099393129348755,
0.04269980266690254,
-0.052978336811065674,
0.09368188679218292,
0.022400610148906708,
-0.047055136412382126,
-0.0033608735539019108,
-0.05051799863576889,
-0.04879390448331833,
-0.024044154211878777,
-0.07366432249546051,
-0.031743038445711136,
0.03311162441968918,
-0.07286371290683746,
0.004394694231450558,
0.07682682573795319,
-0.0005444855778478086,
0.011923586018383503,
0.0312747098505497,
-0.008465440943837166,
0.02055666409432888,
-0.09515359252691269,
0.07954899966716766,
0.015522054396569729,
-0.02054193615913391,
-0.034080248326063156,
0.11066000163555145,
-0.002898614155128598,
0.06545734405517578,
0.0050884694792330265,
-0.05931788682937622,
-0.03693916276097298,
0.022136393934488297,
0.004191880580037832,
-0.030440310016274452,
-0.00745039526373148,
-0.07043289393186569,
0.013134189881384373,
0.05848001688718796,
0.05052647367119789,
0.009335069917142391,
0.039995189756155014,
0.034755170345306396,
1.4900192620553953e-33,
-0.0014828346902504563,
0.011640738695859909,
0.021104777231812477,
0.025225616991519928,
-0.012975259684026241,
-0.015338573604822159,
-0.04788701981306076,
-0.04588508978486061,
0.013073808513581753,
-0.0072058080695569515,
-0.02964291349053383,
-0.0027861683629453182,
-0.11129369586706161,
0.03650616854429245,
0.024949047714471817,
0.04605744034051895,
-0.02896476909518242,
0.030184004455804825,
0.04330787435173988,
-0.03395824879407883,
0.057216159999370575,
-0.09993921965360641,
-0.03357403725385666,
-0.005055123474448919,
-0.008672504685819149,
-0.04726910963654518,
-0.016223790124058723,
-0.00999419204890728,
0.022993164137005806,
-0.002975889015942812,
-0.11714746803045273,
0.020366325974464417,
0.004994635470211506,
-0.005012004170566797,
0.03326437622308731,
-0.010076156817376614,
-0.022745054215192795,
-0.0224677212536335,
0.009845804423093796,
-0.032548289746046066,
-0.026831062510609627,
0.07067668437957764,
0.0006320097018033266,
-0.027471905574202538,
-0.056978534907102585,
-0.0180814266204834,
0.05047265812754631,
-0.07227720320224762,
-0.04430532455444336,
-0.18858250975608826,
0.011825734749436378,
0.013950521126389503,
0.02500104531645775,
-0.07351013273000717,
-0.0143533730879426,
-0.03188430517911911,
0.040207285434007645,
0.019503749907016754,
0.021437834948301315,
0.05519804731011391,
0.004480639472603798,
0.04967699944972992,
0.03133562207221985,
0.02315942570567131,
0.0568305104970932,
0.05138995498418808,
0.03186923637986183,
0.00507727125659585,
-0.02878192625939846,
0.024441514164209366,
0.014648191630840302,
-0.045609913766384125,
-0.000014491411093331408,
-0.08857595920562744,
0.0005100301350466907,
0.05602482706308365,
0.007487264461815357,
0.0030423488933593035,
-0.005552647635340691,
-0.023177165538072586,
-0.06940753012895584,
0.02960534580051899,
-0.04107135161757469,
0.02543489821255207,
-0.054668694734573364,
-0.02392200566828251,
0.04645693674683571,
-0.06604170054197311,
-0.08648090809583664,
-0.02302587404847145,
0.008294006809592247,
0.06646142154932022,
-0.04131123051047325,
0.01107876468449831,
0.011527187190949917,
-2.5371956393514642e-33,
-0.07931850105524063,
0.05859936401247978,
-0.027972888201475143,
-0.0076616969890892506,
-0.021170830354094505,
0.04865102469921112,
-0.03824472054839134,
0.09570538997650146,
-0.01451919600367546,
-0.0710483267903328,
-0.042442481964826584,
-0.01094631664454937,
0.019889798015356064,
-0.022313790395855904,
0.012153385207057,
0.010093658231198788,
0.023844948038458824,
0.06250260770320892,
-0.006321549415588379,
-0.0556202158331871,
0.08908606320619583,
0.028891494497656822,
-0.08255566656589508,
0.060207173228263855,
0.01094140112400055,
0.010462329722940922,
-0.1246640533208847,
0.1467842161655426,
-0.010556896217167377,
0.05583271384239197,
-0.026833126321434975,
0.010898495092988014,
-0.07116030156612396,
0.005427285097539425,
-0.0418173223733902,
0.11644800007343292,
-0.04152350872755051,
0.08411631733179092,
-0.05299433320760727,
0.12307492643594742,
-0.021251285448670387,
0.008792711421847343,
-0.09715206921100616,
0.04952104017138481,
0.021787377074360847,
0.007808544673025608,
0.03264183923602104,
-0.07997702807188034,
0.06368719041347504,
-0.06263051927089691,
-0.02679300494492054,
0.02564220502972603,
-0.0492139607667923,
0.060624346137046814,
-0.05770425498485565,
0.03669110685586929,
0.023135956376791,
-0.11791621148586273,
0.0472259484231472,
0.028302600607275963,
-0.10629156976938248,
-0.058313459157943726,
-0.03932385891675949,
0.027459047734737396,
-0.0808790847659111,
0.008045732975006104,
-0.049998991191387177,
-0.05757616087794304,
0.03195364028215408,
0.02743341773748398,
0.0028440780006349087,
-0.12904493510723114,
0.05632881820201874,
0.07949741929769516,
-0.03206532076001167,
0.009337256662547588,
0.08319581300020218,
0.032969214022159576,
0.015762947499752045,
0.08869447559118271,
0.07155698537826538,
0.04186086729168892,
0.05421895906329155,
0.09621075540781021,
0.10639549791812897,
-0.002382498001679778,
0.06513257324695587,
-0.06356753408908844,
0.0162588469684124,
-0.03412826731801033,
-0.030909862369298935,
0.04537096619606018,
-0.013091139495372772,
-0.0034881578758358955,
-0.02865486964583397,
-3.519938474028095e-8,
-0.030680520460009575,
-0.00022768678900320083,
0.04851224645972252,
-0.08900271356105804,
0.0012231322471052408,
-0.046907875686883926,
0.01627066731452942,
0.050372641533613205,
-0.039016153663396835,
0.015786726027727127,
0.026503143832087517,
-0.015411704778671265,
0.007441617548465729,
-0.013829436153173447,
-0.0671999379992485,
0.043358106166124344,
-0.00923407543450594,
0.10704512149095535,
0.00488610053434968,
-0.025075146928429604,
-0.015318545512855053,
-0.001370919169858098,
0.027619879692792892,
-0.10142280161380768,
0.021031709387898445,
-0.05747310817241669,
0.032223109155893326,
-0.049184400588274,
-0.010873733088374138,
0.05189155787229538,
-0.03150612860918045,
-0.017109094187617302,
0.04001269489526749,
0.02638881839811802,
-0.1057809367775917,
0.1249847263097763,
-0.02690926380455494,
0.010325766168534756,
-0.04042152315378189,
-0.02693631686270237,
-0.0340261310338974,
0.05807379260659218,
-0.10718126595020294,
0.004134536255151033,
-0.03222079202532768,
0.06002277508378029,
0.031475406140089035,
0.005253473296761513,
0.03976842015981674,
0.05354657024145126,
0.050635308027267456,
0.001474583288654685,
0.10596796870231628,
0.01801048032939434,
0.0638241395354271,
-0.03943176940083504,
-0.004452281631529331,
-0.05971536040306091,
-0.0225693229585886,
0.006282818038016558,
0.015757471323013306,
0.03399360552430153,
-0.028592035174369812,
-0.015811949968338013
] |
Conformal Prediction Sets Can Cause Disparate Impact | https://openreview.net/forum?id=fZK6AQXlUU | [
"Jesse C. Cresswell",
"Bhargava Kumar",
"Yi Sui",
"Mouloud Belbahri"
] | Spotlight | Conformal prediction is a statistically rigorous method for quantifying uncertainty in models by having them output sets of predictions, with larger sets indicating more uncertainty. However, prediction sets are not inherently actionable; many applications require a single output to act on, not several. To overcome this limitation, prediction sets can be provided to a human who then makes an informed decision. In any such system it is crucial to ensure the fairness of outcomes across protected groups, and researchers have proposed that Equalized Coverage be used as the standard for fairness. By conducting experiments with human participants, we demonstrate that providing prediction sets can lead to disparate impact in decisions. Disquietingly, we find that providing sets that satisfy Equalized Coverage actually increases disparate impact compared to marginal coverage. Instead of equalizing coverage, we propose to equalize set sizes across groups which empirically leads to lower disparate impact. | Conformal Prediction, Fairness, Uncertainty Quantification, Trustworthy ML, Human Subject Experiments | We demonstrate that providing conformal prediction sets to human decision makers can increase the unfairness of outcomes, and that applying Equalized Coverage increases unfairness more than marginal coverage. | 10,559 | 2410.01888 | [
-0.03747725114226341,
0.003425384871661663,
0.05574510619044304,
0.06223631650209427,
0.07150620967149734,
-0.02384028211236,
0.012603476643562317,
0.03542761504650116,
0.0006729742744937539,
0.04846298322081566,
-0.08233343809843063,
-0.06426392495632172,
0.08004002273082733,
0.0175656508654356,
0.002765666926279664,
-0.010792604647576809,
0.09673219174146652,
0.060468848794698715,
-0.1211896762251854,
-0.028846660628914833,
-0.017941351979970932,
-0.01748371496796608,
-0.016773996874690056,
0.04795423895120621,
-0.025159908458590508,
-0.11398402601480484,
0.00913976039737463,
0.014145788736641407,
-0.011959915049374104,
-0.02959977090358734,
-0.003419154556468129,
-0.060434918850660324,
0.05468727648258209,
-0.03050190955400467,
-0.11173752695322037,
-0.07663081586360931,
-0.023070545867085457,
-0.040314339101314545,
0.004073033574968576,
-0.010545792989432812,
-0.026232410222291946,
0.03460860252380371,
0.032987795770168304,
0.06842216104269028,
-0.0458560548722744,
0.05590994656085968,
0.08532635867595673,
-0.008008548058569431,
-0.02355932630598545,
-0.05597926303744316,
0.0026908153668045998,
-0.02623952552676201,
0.03696252405643463,
-0.023413022980093956,
-0.06889880448579788,
-0.024045677855610847,
0.005068439058959484,
0.026031866669654846,
-0.05760010704398155,
-0.03558678179979324,
-0.10762382298707962,
-0.17139236629009247,
-0.07643455266952515,
0.04741741716861725,
0.000557433464564383,
0.031642623245716095,
-0.013006583787500858,
0.0811275839805603,
-0.06949884444475174,
-0.015338592231273651,
0.03110354207456112,
0.09366718679666519,
-0.026349354535341263,
0.02935148775577545,
0.1442621350288391,
0.10466416925191879,
0.024074159562587738,
0.022441808134317398,
0.03237751126289368,
-0.007021049968898296,
-0.02420607954263687,
-0.058812130242586136,
0.02822466753423214,
0.059613507241010666,
0.08013367652893066,
-0.02957535721361637,
0.021702494472265244,
0.04627826064825058,
0.039085663855075836,
0.0003834104572888464,
-0.03485696390271187,
0.0019511538557708263,
0.06776202470064163,
0.02664330042898655,
-0.0728536918759346,
0.050061099231243134,
0.023811768740415573,
-0.07301589846611023,
0.04854503273963928,
0.09094763547182083,
-0.02618308737874031,
0.04059936851263046,
0.024899862706661224,
-0.022956643253564835,
0.05205542966723442,
-0.058948107063770294,
0.04539500176906586,
-0.05344407632946968,
0.012367164716124535,
-0.02264690212905407,
0.05466993525624275,
-0.0074433982372283936,
0.035975053906440735,
0.008431934751570225,
0.0031162414234131575,
-0.016498878598213196,
0.02819686196744442,
0.018884213641285896,
-0.03660999611020088,
0.02776792272925377,
-0.025619762018322945,
-0.0013861008919775486,
0.09659675508737564,
-0.021037554368376732,
0.06732513010501862,
0.0003029094368685037,
-0.082449771463871,
3.524646342543703e-33,
-0.026623226702213287,
-0.07307779788970947,
0.043991569429636,
0.00010121295053977519,
0.07891777157783508,
-0.041053470224142075,
-0.061753347516059875,
0.008479590527713299,
-0.0016917017055675387,
0.026802700012922287,
-0.05155817046761513,
0.0023298163432627916,
0.034514881670475006,
0.05424143001437187,
0.034604497253894806,
0.04371451959013939,
-0.03490285575389862,
0.09408903867006302,
-0.12404309958219528,
0.016740357503294945,
0.057337529957294464,
-0.047797463834285736,
-0.03706036135554314,
-0.016137147322297096,
-0.019920451566576958,
-0.04009679704904556,
-0.05536497384309769,
0.031244395300745964,
0.029745783656835556,
-0.0016263191355392337,
-0.09174143522977829,
0.02004418522119522,
0.09228614717721939,
0.043333690613508224,
0.012133392505347729,
0.042120084166526794,
-0.06664396822452545,
0.018572701141238213,
0.005214664153754711,
0.008246351964771748,
-0.039639249444007874,
-0.04179687052965164,
0.039953529834747314,
-0.032071035355329514,
0.04875456914305687,
-0.07206323742866516,
-0.055367838591337204,
-0.03241651505231857,
-0.042738284915685654,
-0.003170148003846407,
0.035486284643411636,
0.07000216841697693,
-0.005321267060935497,
-0.07284026592969894,
-0.07709991186857224,
-0.03162446618080139,
0.005560065619647503,
-0.025581132620573044,
0.006658888887614012,
-0.03364821523427963,
-0.0682498961687088,
-0.011760622262954712,
-0.041797541081905365,
-0.016997741535305977,
-0.03398922458291054,
0.03435805067420006,
-0.0012213867157697678,
-0.08013825863599777,
0.03615693375468254,
0.017981162294745445,
0.03780786320567131,
0.03236977383494377,
-0.07711885124444962,
0.044847000390291214,
-0.06282906234264374,
0.04559740051627159,
0.0514092892408371,
0.12249137461185455,
0.09986679255962372,
-0.02598046511411667,
0.03270404040813446,
-0.04808376729488373,
-0.06931661814451218,
-0.08015429228544235,
-0.016634758561849594,
-0.03603039309382439,
0.045463841408491135,
-0.00763764139264822,
-0.06408759951591492,
-0.026771487668156624,
-0.060958076268434525,
-0.010024964809417725,
-0.01972312480211258,
0.013555978424847126,
-0.009389203041791916,
-5.768266492210124e-33,
-0.11790241301059723,
0.007034613750874996,
0.007073141634464264,
0.02626010775566101,
0.022311897948384285,
-0.01924671232700348,
-0.000358121789759025,
-0.05948026850819588,
0.05992436408996582,
-0.0027279271744191647,
0.01894524320960045,
0.015045478008687496,
0.09869126975536346,
-0.06082213297486305,
-0.035296760499477386,
-0.06764291226863861,
0.04928727075457573,
0.004089368041604757,
-0.00108357984572649,
-0.017862973734736443,
0.07359674572944641,
0.0013402299955487251,
-0.06208224222064018,
0.09626731276512146,
-0.0017468581208959222,
-0.017735689878463745,
-0.05897808074951172,
0.019154466688632965,
0.008011567406356335,
-0.011818417347967625,
-0.1353205442428589,
-0.10189484059810638,
-0.030170075595378876,
-0.05708007887005806,
0.02322492189705372,
-0.026441514492034912,
0.03176575154066086,
0.06921083480119705,
-0.07389294356107712,
0.036238569766283035,
-0.01100074965506792,
-0.025099560618400574,
-0.13093321025371552,
0.02238447032868862,
0.005168448202311993,
-0.013669895008206367,
-0.001600863179191947,
-0.08654241263866425,
0.04682027921080589,
-0.07769764959812164,
-0.05004410445690155,
0.04717375710606575,
-0.02121947892010212,
0.06973030418157578,
-0.02138935960829258,
-0.07356876134872437,
0.03517024964094162,
-0.045310888439416885,
0.036274462938308716,
0.122581347823143,
-0.0669846162199974,
0.0100889652967453,
-0.033489976078271866,
0.08397470414638519,
-0.03961899131536484,
-0.016044670715928078,
-0.09268317371606827,
-0.018997227773070335,
0.02356199361383915,
0.02563299611210823,
-0.020762594416737556,
-0.03687988221645355,
0.0022308018524199724,
-0.048530712723731995,
-0.01190339494496584,
0.012227274477481842,
0.06480555981397629,
0.02174261212348938,
-0.039361484348773956,
0.11954610794782639,
0.03182116523385048,
-0.008695333264768124,
0.06851222366094589,
-0.027345508337020874,
0.04777069762349129,
0.020042292773723602,
0.04609682783484459,
0.01132054440677166,
0.055805984884500504,
0.012090275064110756,
-0.08669125288724899,
0.0389067567884922,
-0.021583352237939835,
0.04391494393348694,
-0.10833720862865448,
-4.586243917970023e-8,
0.006571460980921984,
0.08191706240177155,
0.027633102610707283,
0.108064204454422,
-0.03924473002552986,
-0.034596607089042664,
-0.08064428716897964,
-0.00958925113081932,
-0.09650634974241257,
0.028948035091161728,
0.04420170187950134,
0.010875657200813293,
0.006319803651422262,
0.030114104971289635,
0.07486777007579803,
0.07608701288700104,
-0.03065299615263939,
0.04987325891852379,
-0.03403136506676674,
0.02543003298342228,
0.04503488913178444,
-0.002828113501891494,
0.018123449757695198,
-0.048087142407894135,
0.11067546904087067,
-0.017720045521855354,
-0.10986804962158203,
0.03776846453547478,
-0.012665815651416779,
0.033780667930841446,
-0.014810558408498764,
0.0007526854169555008,
-0.01664913445711136,
-0.016700319945812225,
0.018923256546258926,
0.028129931539297104,
-0.04597063362598419,
-0.024802785366773605,
0.031145401298999786,
0.029017364606261253,
0.021798232570290565,
0.01758570969104767,
0.0067628007382154465,
0.05001169443130493,
0.08277389407157898,
-0.006667337846010923,
0.0053332652896642685,
0.03828660026192665,
0.013080465607345104,
-0.053037501871585846,
0.03205064311623573,
0.0031713054049760103,
0.03913683816790581,
0.030994068831205368,
0.12099628150463104,
-0.01585913822054863,
0.022775471210479736,
-0.05282389372587204,
-0.01585202105343342,
0.005533372983336449,
0.08560647070407867,
-0.09228429943323135,
-0.046825312077999115,
-0.00007552326133009046
] |
In Search of Forgotten Domain Generalization | https://openreview.net/forum?id=Fk3eod9aaD | [
"Prasanna Mayilvahanan",
"Roland S. Zimmermann",
"Thaddäus Wiedemer",
"Evgenia Rusak",
"Attila Juhos",
"Matthias Bethge",
"Wieland Brendel"
] | Spotlight | Out-of-Domain (OOD) generalization is the ability of a model trained on one or more domains to generalize to unseen domains. In the ImageNet era of computer vision, evaluation sets for measuring a model's OOD performance were designed to be strictly OOD with respect to style. However, the emergence of foundation models and expansive web-scale datasets has obfuscated this evaluation process, as datasets cover a broad range of domains and risk test domain contamination. In search of the forgotten domain generalization, we create large-scale datasets subsampled from LAION---LAION-Natural and LAION-Rendition---that are strictly OOD to corresponding ImageNet and DomainNet test sets in terms of style. Training CLIP models on these datasets reveals that a significant portion of their performance is explained by in-domain examples. This indicates that the OOD generalization challenges from the ImageNet era still prevail and that training on web-scale data merely creates the illusion of OOD generalization. Furthermore, through a systematic exploration of combining natural and rendition datasets in varying proportions, we identify optimal mixing ratios for model generalization across these domains. Our datasets and results re-enable meaningful assessment of OOD robustness at scale---a crucial prerequisite for improving model robustness. | Out-of-Distribution Robustness, OOD generalization, Out-of-Domain Robustness, Evaluation | CLIP's high performance on style-centric domain shifts is significantly influenced by the presence of such images in its training set. | 10,412 | 2410.08258 | [
-0.00536266528069973,
-0.1004866436123848,
0.03010193444788456,
0.03508147969841957,
0.1901978701353073,
-0.06183547526597977,
-0.02276744693517685,
-0.02326446771621704,
0.03964986279606819,
-0.08215878903865814,
0.0018301453674212098,
-0.05560087412595749,
0.05249118432402611,
0.021952971816062927,
-0.008304792456328869,
-0.07132086157798767,
0.07625631242990494,
0.07262858003377914,
-0.14655807614326477,
0.0023773489519953728,
0.034114137291908264,
0.02551448717713356,
0.06660734117031097,
-0.03226831555366516,
0.009341352619230747,
-0.037053246051073074,
0.035171009600162506,
0.010038979351520538,
0.03249654546380043,
-0.05880570039153099,
0.04602903872728348,
0.04159059748053551,
0.010668443515896797,
0.04550662264227867,
0.0148337222635746,
0.04598380625247955,
-0.015294276177883148,
0.0558607317507267,
-0.012879609130322933,
-0.02148117870092392,
0.011751776561141014,
-0.031766146421432495,
0.0713111013174057,
0.017922818660736084,
0.027985893189907074,
0.03040970116853714,
-0.035091232508420944,
-0.07873167097568512,
0.04330385476350784,
-0.03368954360485077,
-0.03203512728214264,
-0.03107631579041481,
-0.0823284313082695,
0.06505007296800613,
-0.03177298605442047,
-0.0415206216275692,
0.044030141085386276,
-0.040225621312856674,
-0.022368216887116432,
0.03361278772354126,
-0.07867322862148285,
-0.06600971519947052,
-0.03777991980314255,
0.03869713097810745,
0.0014987471513450146,
0.02480991929769516,
0.0375450998544693,
-0.015503116883337498,
-0.02992645464837551,
-0.03073567897081375,
-0.03628284111618996,
0.04575568810105324,
-0.029289107769727707,
0.10475806891918182,
0.017508821561932564,
0.1221139058470726,
0.03924737870693207,
0.017376059666275978,
0.010070648044347763,
-0.11199446767568588,
0.03868847340345383,
0.01944263093173504,
0.10024514049291611,
0.04791596904397011,
0.048051465302705765,
0.07179126143455505,
0.0003290208405815065,
0.0029846651013940573,
-0.002963276579976082,
0.018325842916965485,
-0.10027145594358444,
-0.06135372444987297,
0.06349699199199677,
0.0324023962020874,
0.06726361066102982,
-0.01824229583144188,
0.021901460364460945,
-0.12050508707761765,
-0.023102762177586555,
0.09415426850318909,
-0.03526435047388077,
0.029984185472130775,
0.0561833381652832,
-0.05595272779464722,
0.0213452260941267,
0.0013947529951110482,
0.06815943866968155,
0.06336043030023575,
0.11009781062602997,
-0.0712708830833435,
0.037703629583120346,
0.06775321066379547,
0.013713905587792397,
-0.08997125178575516,
0.029189161956310272,
-0.07553984224796295,
-0.058245476335287094,
-0.012591495178639889,
-0.01732633449137211,
0.044739946722984314,
-0.02910279855132103,
0.021048082038760185,
-0.004761309362947941,
-0.0006272061727941036,
0.003626269521191716,
-0.019785454496741295,
-0.050109997391700745,
2.038719545703596e-33,
0.04988608881831169,
0.0014300309121608734,
0.05193595960736275,
-0.011646476574242115,
0.0677349716424942,
-0.04106758162379265,
0.0011684257769957185,
-0.004017242230474949,
-0.023016463965177536,
0.03128804638981819,
-0.021486695855855942,
0.02979537658393383,
-0.07881636917591095,
0.11696744710206985,
0.0563884973526001,
0.0448596328496933,
-0.024014320224523544,
0.0802987590432167,
0.01827152632176876,
0.048476506024599075,
0.018149979412555695,
-0.007438612636178732,
0.020876338705420494,
-0.02705278806388378,
0.02418205514550209,
-0.07705032080411911,
0.0018241129582747817,
0.03566644713282585,
-0.05001934617757797,
0.02854355424642563,
-0.10298366844654083,
0.027067290619015694,
0.0436948798596859,
0.09838757663965225,
0.04676077887415886,
-0.026416517794132233,
-0.008487986400723457,
-0.02761232666671276,
0.026567397639155388,
-0.008725518360733986,
-0.030967099592089653,
0.009110964834690094,
-0.0500846765935421,
0.042689137160778046,
-0.06239784508943558,
-0.04519437253475189,
-0.019502870738506317,
0.024860283359885216,
-0.0315147303044796,
-0.008473067544400692,
0.05453278869390488,
0.00772925466299057,
-0.035433463752269745,
-0.0394832119345665,
-0.09054222702980042,
0.003762147156521678,
0.04013033211231232,
-0.0382797010242939,
0.02774050459265709,
-0.004898590501397848,
0.0036025261506438255,
0.04235716536641121,
-0.04636864364147186,
0.0589069202542305,
-0.05666584149003029,
-0.01550372876226902,
0.020826106891036034,
-0.03128727898001671,
-0.048860304057598114,
-0.0008833120227791369,
-0.020398909226059914,
0.07482839375734329,
-0.10865151137113571,
-0.007294377777725458,
0.037745315581560135,
-0.02456355094909668,
0.06616050004959106,
-0.019520800560712814,
0.011661609634757042,
0.06365373730659485,
-0.014495912939310074,
0.029903419315814972,
0.021658027544617653,
-0.0784570723772049,
-0.09802264720201492,
-0.017835592851042747,
0.05884690582752228,
-0.05910082906484604,
-0.03758940100669861,
-0.050849027931690216,
-0.07146104425191879,
-0.035699088126420975,
-0.05684126168489456,
-0.05579923465847969,
0.01588054932653904,
-2.3764322768681234e-33,
-0.07517993450164795,
0.06184748187661171,
-0.061567727476358414,
0.07006923854351044,
-0.03618694096803665,
-0.068463034927845,
0.011050032451748848,
0.04206382483243942,
-0.03344659507274628,
-0.07350112497806549,
0.02938711829483509,
-0.006729189306497574,
-0.062261682003736496,
-0.04620775580406189,
-0.038477614521980286,
-0.021793382242321968,
-0.017822880297899246,
-0.007714300416409969,
-0.012594400905072689,
0.05037430673837662,
0.05231801047921181,
0.05738355219364166,
-0.06952593475580215,
0.026648476719856262,
-0.031110771000385284,
0.03830351680517197,
-0.08861075341701508,
0.07552310079336166,
-0.011278596706688404,
-0.021419620141386986,
-0.022359615191817284,
0.010953934863209724,
-0.024836579337716103,
0.011600186116993427,
-0.05732393264770508,
0.04742655158042908,
0.0748746246099472,
-0.057587187737226486,
-0.027381157502532005,
0.09625158458948135,
-0.007391584571450949,
-0.017461055889725685,
-0.11041869223117828,
0.03217101842164993,
-0.055587511509656906,
-0.08263567090034485,
-0.042743414640426636,
0.026008129119873047,
0.06221894919872284,
-0.011448112316429615,
-0.07802213728427887,
0.10273057967424393,
-0.04674810171127319,
0.011048351414501667,
-0.031386107206344604,
-0.08710319548845291,
-0.004073264542967081,
-0.009107563644647598,
0.014406095258891582,
0.015238835476338863,
-0.019201703369617462,
-0.01519894227385521,
-0.1160610243678093,
0.03435881435871124,
-0.042229413986206055,
-0.019094625487923622,
-0.06190112233161926,
0.02875596098601818,
-0.05307261645793915,
0.030103111639618874,
-0.043930474668741226,
-0.07489106804132462,
-0.0034390625078231096,
0.01301902811974287,
-0.06357500702142715,
-0.0688151866197586,
0.08912425488233566,
0.027712911367416382,
0.06591708213090897,
-0.015481598675251007,
0.056529056280851364,
-0.0034328708425164223,
0.03404998779296875,
0.029475757852196693,
0.0792921930551529,
0.16834649443626404,
0.041417691856622696,
0.03426460549235344,
0.013775404542684555,
0.006164369639009237,
-0.0966244712471962,
-0.03954695910215378,
-0.04677768051624298,
0.09575475752353668,
-0.02785407193005085,
-4.2650000864341564e-8,
-0.047685761004686356,
0.0919964611530304,
0.04103158041834831,
0.02561323717236519,
-0.017786609008908272,
-0.030118973925709724,
-0.01949474774301052,
0.0816747322678566,
-0.043565038591623306,
-0.039017993956804276,
0.017861660569906235,
0.06288272887468338,
-0.07476753741502762,
-0.014717653393745422,
-0.02025545760989189,
0.016046615317463875,
0.002900240011513233,
0.06808333843946457,
-0.00004751464439323172,
0.04529029130935669,
-0.0018184728687629104,
0.002452405169606209,
0.03403976187109947,
-0.094034843146801,
0.08435340970754623,
-0.052382223308086395,
-0.01628412865102291,
0.07530885934829712,
0.03499313443899155,
0.04038340225815773,
0.013254077173769474,
0.011825557798147202,
-0.056876979768276215,
-0.0779242143034935,
0.02670631930232048,
0.09520839154720306,
-0.03401415795087814,
0.013925337232649326,
0.0029990391340106726,
-0.0177536029368639,
-0.004713553003966808,
0.06683874130249023,
0.010885988362133503,
0.014088797383010387,
0.040664248168468475,
-0.013244525529444218,
0.02905461937189102,
-0.008817930705845356,
0.05291362479329109,
-0.0058869351632893085,
-0.002467996906489134,
-0.016249116510152817,
-0.034228745847940445,
0.12189044803380966,
0.025882884860038757,
-0.07686350494623184,
0.030162254348397255,
-0.013821238651871681,
0.03344665467739105,
0.11555396765470505,
0.005279727280139923,
-0.06052050739526749,
-0.02832789346575737,
0.007687751203775406
] |
TOP-ERL: Transformer-based Off-Policy Episodic Reinforcement Learning | https://openreview.net/forum?id=N4NhVN30ph | [
"Ge Li",
"Dong Tian",
"Hongyi Zhou",
"Xinkai Jiang",
"Rudolf Lioutikov",
"Gerhard Neumann"
] | Spotlight | This work introduces Transformer-based Off-Policy Episodic Reinforcement Learning (TOP-ERL), a novel algorithm that enables off-policy updates in the ERL framework. In ERL, policies predict entire action trajectories over multiple time steps instead of single actions at every time step. These trajectories are typically parameterized by trajectory generators such as Movement Primitives (MP), allowing for smooth and efficient exploration over long horizons while capturing high-level temporal correlations. However, ERL methods are often constrained to on-policy frameworks due to the difficulty of evaluating state-action values for entire action sequences, limiting their sample efficiency and preventing the use of more efficient off-policy architectures. TOP-ERL addresses this shortcoming by segmenting long action sequences and estimating the state-action values for each segment using a transformer-based critic architecture alongside an n-step return estimation. These contributions result in efficient and stable training that is reflected in the empirical results conducted on sophisticated robot learning environments. TOP-ERL significantly outperforms state-of-the-art RL methods. Thorough ablation studies additionally show the impact of key design choices on the model performance. | Value of sequences of actions, Reinforcement Learning, Transformer, Robot Manipulation, Movement Primitives. | We propose a novel transformer-based RL method that learns values of consecutive actions. | 10,360 | null | [
-0.06321164220571518,
-0.08145280182361603,
0.05505343899130821,
0.011195753701031208,
0.04564186558127403,
0.045290764421224594,
-0.04302468150854111,
-0.0075330534018576145,
0.010734409093856812,
-0.003652711631730199,
-0.04433969035744667,
0.010091557167470455,
-0.049768511205911636,
0.04167387634515762,
0.027221819385886192,
0.05297841504216194,
0.024586550891399384,
0.047160834074020386,
-0.05141759663820267,
-0.07375891506671906,
0.11017382144927979,
0.024929950013756752,
0.09619023650884628,
0.056410640478134155,
0.029860613867640495,
0.057938288897275925,
0.03646596521139145,
-0.05699852108955383,
-0.019134972244501114,
-0.03688419610261917,
0.06245604529976845,
-0.060900114476680756,
0.03131067752838135,
-0.09825439006090164,
-0.07370176166296005,
0.08149436861276627,
-0.11215688288211823,
-0.02528027817606926,
-0.03897087648510933,
-0.036097049713134766,
-0.009037642739713192,
0.0010845998767763376,
0.05062197148799896,
-0.0005079366965219378,
0.05186624825000763,
0.046777382493019104,
0.02918712981045246,
-0.1179126501083374,
0.03445381298661232,
0.04055895283818245,
0.044210419058799744,
0.01776256039738655,
0.02147676795721054,
0.01061355508863926,
-0.004283974412828684,
0.09529253095388412,
0.0576537661254406,
-0.001709260162897408,
0.05956517904996872,
-0.05259209871292114,
-0.05570729821920395,
-0.08648697286844254,
-0.025674667209386826,
-0.08111501485109329,
-0.19172920286655426,
-0.010811789892613888,
0.012417868711054325,
-0.020928766578435898,
-0.020402247086167336,
0.017244553193449974,
-0.0800805613398552,
-0.0006323304260149598,
0.02719179354608059,
0.05358199402689934,
-0.009145796298980713,
-0.031139453873038292,
0.06109968200325966,
-0.026188602671027184,
0.006676404271274805,
-0.01595020852982998,
0.016627194359898567,
-0.019114311784505844,
0.00699234614148736,
0.029819050803780556,
0.11204990744590759,
-0.007893688976764679,
0.005173930432647467,
0.03213949501514435,
0.05045585706830025,
0.029677119106054306,
0.07481194287538528,
-0.04069630056619644,
0.039777275174856186,
-0.06949075311422348,
0.012770413421094418,
0.014941924251616001,
-0.0814003050327301,
-0.09130195528268814,
-0.04383796080946922,
0.05872320383787155,
0.006873230449855328,
0.008733980357646942,
-0.0412176214158535,
0.0501849502325058,
-0.03247284144163132,
-0.08743805438280106,
0.05515449866652489,
0.027514604851603508,
0.027934236451983452,
-0.010881626978516579,
0.03436194360256195,
-0.06551270186901093,
0.12644775211811066,
0.021238679066300392,
0.009552176110446453,
0.03151264414191246,
-0.02973218820989132,
-0.022654855623841286,
-0.02508176490664482,
0.07443774491548538,
0.017863303422927856,
-0.0872475728392601,
0.02783450484275818,
0.08657802641391754,
0.026855461299419403,
-0.050477880984544754,
-0.00674099987372756,
3.6438804079674536e-33,
0.027322184294462204,
-0.008368443697690964,
0.03193846344947815,
0.009665765799582005,
-0.01117937732487917,
0.021393634378910065,
0.08727961033582687,
-0.016030391678214073,
-0.013650177977979183,
-0.0025883938651531935,
0.028451822698116302,
0.06549732387065887,
-0.05149037763476372,
0.09176428616046906,
0.008812637999653816,
0.01459325198084116,
-0.01279369741678238,
0.06217004731297493,
-0.009014111943542957,
-0.0569336824119091,
0.06236722692847252,
-0.031386032700538635,
-0.005238605663180351,
-0.06777460128068924,
0.020834891125559807,
-0.031107140704989433,
0.009136845357716084,
-0.010534626431763172,
-0.02033950574696064,
-0.0006970837130211294,
-0.026053206995129585,
0.03225649148225784,
-0.07305149734020233,
0.09486404061317444,
0.048060547560453415,
-0.007259149570018053,
-0.03239350765943527,
0.025830809026956558,
-0.0008551464998163283,
0.0005715725128538907,
0.0034327206667512655,
0.024477407336235046,
0.012712177820503712,
0.0019879192113876343,
-0.07527778297662735,
-0.12210997194051743,
0.020993957296013832,
0.009648574516177177,
-0.044100433588027954,
-0.00435689277946949,
0.028779421001672745,
0.01255199033766985,
-0.01744200848042965,
-0.05327485874295235,
-0.033848393708467484,
0.013422223739326,
0.016950054094195366,
0.11309825628995895,
-0.03805805742740631,
0.029465919360518456,
0.05248128995299339,
0.04948803782463074,
0.016757436096668243,
0.04492739588022232,
-0.013820812106132507,
0.08749470114707947,
-0.01379936933517456,
0.09266427904367447,
0.03684481978416443,
0.0015555290738120675,
0.044503167271614075,
0.05848587676882744,
0.06510980427265167,
-0.05217321962118149,
0.0554809644818306,
-0.04626841098070145,
0.024623777717351913,
-0.03859449550509453,
-0.029447270557284355,
0.04003920033574104,
-0.06620467454195023,
0.02210244908928871,
-0.034876804798841476,
0.014912312850356102,
0.04895440489053726,
-0.040074776858091354,
-0.0026168478652834892,
-0.030911345034837723,
-0.0511457622051239,
-0.002907818416133523,
-0.08566182851791382,
-0.07149189710617065,
-0.037183307111263275,
0.08259204775094986,
0.009212035685777664,
-3.197623079460973e-33,
0.0656108632683754,
-0.021782590076327324,
0.04236392676830292,
0.0328843891620636,
-0.025407882407307625,
-0.005794439930468798,
-0.02056378498673439,
0.02878321334719658,
0.03474820405244827,
-0.02742627263069153,
-0.04362240806221962,
-0.023769840598106384,
-0.0007803700282238424,
0.009897002950310707,
0.10719078034162521,
-0.0663510337471962,
-0.034936223179101944,
-0.007248444017022848,
-0.0007049559499137104,
0.01863686554133892,
0.003795847063884139,
0.03530221804976463,
-0.08455899357795715,
0.05244757980108261,
-0.035410720854997635,
0.009768891148269176,
-0.026152141392230988,
0.027691271156072617,
0.022603895515203476,
0.014980530366301537,
-0.06814082711935043,
-0.029548820108175278,
-0.0024449373595416546,
0.013046720065176487,
-0.12675411999225616,
0.07271776348352432,
0.016398580744862556,
0.0005591866793110967,
-0.05104197561740875,
0.029465923085808754,
0.02700551226735115,
-0.04969141259789467,
-0.009298364631831646,
0.0003698021173477173,
0.012226306833326817,
-0.0712781772017479,
-0.03941021487116814,
0.052342139184474945,
-0.05066107586026192,
-0.009660978801548481,
-0.08103794604539871,
0.06723468005657196,
-0.13013163208961487,
-0.06643584370613098,
-0.04560451954603195,
-0.06728902459144592,
-0.047719452530145645,
0.017818985506892204,
0.03213074430823326,
-0.0352436862885952,
-0.04155721142888069,
0.08436354994773865,
0.015242389403283596,
0.021419070661067963,
0.023764964193105698,
0.07184793800115585,
-0.024430157616734505,
0.010746375657618046,
0.04419960826635361,
0.007783539593219757,
0.021500209346413612,
-0.0006395124946720898,
0.027157917618751526,
-0.027978982776403427,
0.04996400326490402,
-0.04397297650575638,
0.010732214897871017,
-0.050485506653785706,
-0.0032500550150871277,
-0.04718302935361862,
-0.02493988163769245,
0.009206701070070267,
-0.017624137923121452,
0.09441302716732025,
-0.029859626665711403,
0.0825219601392746,
-0.03841255232691765,
0.09027586132287979,
0.028936775401234627,
0.03388548269867897,
0.014468052424490452,
-0.01523405872285366,
0.041638556867837906,
0.08128734678030014,
-0.04976556822657585,
-4.573201195512411e-8,
-0.06329483538866043,
0.053122229874134064,
0.06985674798488617,
0.0011616625124588609,
0.06714395433664322,
0.0034448120277374983,
0.04783966764807701,
0.06660526990890503,
-0.026646655052900314,
-0.025113128125667572,
0.08655212819576263,
0.028779905289411545,
0.022964555770158768,
-0.08189722150564194,
-0.03402213752269745,
0.06311874091625214,
0.08761272579431534,
-0.0071705919690430164,
0.019149240106344223,
-0.0015305301640182734,
0.0257361251860857,
-0.033000439405441284,
0.030092831701040268,
-0.08884500712156296,
0.008157392032444477,
-0.09359090775251389,
-0.07779431343078613,
0.033661145716905594,
0.03974456340074539,
-0.022633297368884087,
0.018341049551963806,
-0.03954142704606056,
0.04185768589377403,
0.060449760407209396,
0.04815766215324402,
0.06897158175706863,
0.05593531206250191,
-0.04899488389492035,
0.009905570186674595,
0.0032148396130651236,
-0.01698426902294159,
0.0825515016913414,
0.005312346387654543,
-0.0023746357765048742,
-0.08136570453643799,
0.02391178347170353,
-0.08484842628240585,
-0.1510278433561325,
0.021354345604777336,
-0.1031084805727005,
0.05089786648750305,
-0.03375284746289253,
-0.025941845029592514,
0.08480504900217056,
0.07863516360521317,
0.03909265249967575,
0.05514943599700928,
-0.14861100912094116,
0.06070741266012192,
0.1286722868680954,
-0.005197500344365835,
-0.04915094003081322,
-0.1190500259399414,
0.005738452076911926
] |
On the Expressiveness of Rational ReLU Neural Networks With Bounded Depth | https://openreview.net/forum?id=uREg3OHjLL | [
"Gennadiy Averkov",
"Christopher Hojny",
"Maximilian Merkert"
] | Spotlight | To confirm that the expressive power of ReLU neural networks grows with their depth, the function $F_n = \max (0,x_1,\ldots,x_n )$ has been considered in the literature.
A conjecture by Hertrich, Basu, Di Summa, and Skutella [NeurIPS 2021] states that any ReLU network that exactly represents $F_n$ has at least $\lceil \log_2 (n+1) \rceil$ hidden layers.
The conjecture has recently been confirmed for networks with integer weights by Haase, Hertrich, and Loho [ICLR 2023].
We follow up on this line of research and show that, within ReLU networks whose weights are decimal fractions, $F_n$ can only be represented by networks with at least $\lceil \log_3 (n+1) \rceil$ hidden layers.
Moreover, if all weights are $N$-ary fractions, then $F_n$ can only be represented by networks with at least $\Omega( \frac{\ln n}{\ln \ln N})$ layers.
These results are a partial confirmation of the above conjecture for rational ReLU networks, and provide the first non-constant lower bound on the depth of practically relevant ReLU networks. | expressive power, depth, exact representations, ReLU networks, mixed volumes, lattice polytopes, number theory | $\max \{0,x_1,\ldots,x_n\}$ requires depth $\Omega(\log n)$ to be exactly described with ReLUs and weights being decimal fractions. | 10,347 | 2502.06283 | [
-0.12013143301010132,
-0.0839027687907219,
0.011677623726427555,
-0.002228289609774947,
0.046260129660367966,
0.02636757865548134,
-0.002794896950945258,
-0.02562783472239971,
0.08283761888742447,
-0.07152434438467026,
-0.1291407346725464,
0.003360767150297761,
0.030336132273077965,
0.021126655861735344,
-0.03679005429148674,
0.018003787845373154,
0.0225372277200222,
0.08683644235134125,
-0.09563034027814865,
0.022977296262979507,
0.057842474430799484,
0.009605607949197292,
-0.02521595172584057,
0.016674378886818886,
0.07372067868709564,
-0.14216962456703186,
-0.040884848684072495,
-0.021648580208420753,
0.018182314932346344,
-0.009948045015335083,
0.02416572906076908,
-0.04146507754921913,
-0.0056424676440656185,
0.028319282457232475,
0.004856855142861605,
0.05398645997047424,
0.015879346057772636,
-0.030671564862132072,
-0.0026967742014676332,
0.016426900401711464,
-0.0053386748768389225,
0.06460726261138916,
0.015069014392793179,
0.02593420259654522,
0.05561654269695282,
-0.03359108045697212,
0.02672496810555458,
-0.0008149943314492702,
-0.11280977725982666,
-0.00704929931089282,
-0.017047766596078873,
0.025165854021906853,
-0.017793383449316025,
0.07040371745824814,
-0.00008158705168170854,
-0.036764733493328094,
-0.04630037397146225,
-0.0026416887994855642,
-0.056547388434410095,
0.0909838154911995,
0.0189572311937809,
-0.05045708641409874,
0.02114759013056755,
-0.056248679757118225,
0.007481906563043594,
0.014488117769360542,
-0.018535062670707703,
-0.01179108303040266,
0.040795180946588516,
0.08369331806898117,
0.09804316610097885,
-0.016435351222753525,
-0.11443767696619034,
0.005363098811358213,
0.11492042243480682,
0.00505228852853179,
0.08647212386131287,
0.012158136814832687,
-0.006810077466070652,
-0.03915998712182045,
0.012738345190882683,
-0.029988225549459457,
-0.027407722547650337,
-0.021237412467598915,
0.05288229137659073,
-0.08697552233934402,
-0.10378200560808182,
0.020830335095524788,
0.02697877027094364,
-0.0011447336291894317,
-0.07517383247613907,
0.004768881481140852,
-0.013391276821494102,
0.01725270226597786,
0.0808042585849762,
-0.06259127706289291,
-0.051501695066690445,
-0.037611592561006546,
-0.022769153118133545,
0.06536045670509338,
-0.05723986029624939,
-0.01239419262856245,
0.06296861916780472,
-0.06352784484624863,
0.030494341626763344,
0.011319701559841633,
0.06627316772937775,
0.09026394784450531,
0.004204912111163139,
-0.0543978177011013,
0.018377000465989113,
-0.0663348138332367,
0.05451635643839836,
0.030304258689284325,
0.09046904742717743,
-0.05885753408074379,
0.028353029862046242,
-0.023139728233218193,
-0.047332026064395905,
0.013947076164186,
-0.015697848051786423,
0.030647743493318558,
-0.030584678053855896,
-0.013566456735134125,
0.06252636015415192,
0.013948743231594563,
-0.0680008977651596,
5.396487016403805e-33,
0.0571417510509491,
0.0841326043009758,
0.04431699588894844,
-0.022226529195904732,
0.0894591361284256,
-0.006538699846714735,
0.07993338257074356,
-0.03633812442421913,
0.06885819137096405,
0.058603230863809586,
-0.08699194341897964,
0.10922139883041382,
-0.00014273705892264843,
0.001573452027514577,
0.05337778106331825,
-0.027471421286463737,
0.056126974523067474,
-0.018154090270400047,
-0.03809158131480217,
-0.1194562166929245,
0.02179192565381527,
-0.03661215305328369,
0.008037243038415909,
-0.06012702360749245,
0.01966935023665428,
-0.037090789526700974,
-0.006584289483726025,
-0.029357995837926865,
-0.023525262251496315,
0.022640570998191833,
-0.08003216981887817,
0.007165179587900639,
0.04284461960196495,
0.02540515922009945,
-0.004243754781782627,
0.019713953137397766,
0.008666508831083775,
-0.004361138679087162,
0.03847944363951683,
-0.008228638209402561,
0.018123770132660866,
0.04250583425164223,
-0.05282578244805336,
-0.025294369086623192,
0.02422088384628296,
-0.07598188519477844,
0.005469390191137791,
0.09473466873168945,
-0.09485488384962082,
-0.05341402068734169,
0.012930861674249172,
0.054511070251464844,
-0.060510050505399704,
-0.08382032066583633,
0.04829571023583412,
-0.07100451737642288,
0.019092928618192673,
0.12476196885108948,
0.049401767551898956,
0.09740367531776428,
-0.018146542832255363,
-0.0001401046320097521,
-0.02204711362719536,
0.04351193830370903,
-0.018781015649437904,
0.01930386945605278,
0.017315438017249107,
-0.00032435631146654487,
0.07464884221553802,
0.007050539366900921,
0.011956504546105862,
0.06551692634820938,
-0.059164345264434814,
-0.0769219845533371,
0.038861483335494995,
-0.013992095366120338,
0.003231251612305641,
-0.13324593007564545,
0.008409795351326466,
0.09443489462137222,
-0.034810442477464676,
0.017112988978624344,
0.0190846286714077,
0.04362946003675461,
-0.08218744397163391,
-0.018047545105218887,
0.059043776243925095,
0.013847075402736664,
-0.019692616537213326,
-0.019344698637723923,
-0.010228353552520275,
-0.07086274772882462,
0.047571104019880295,
-0.08245780318975449,
-0.03255690634250641,
-4.274396475965328e-33,
-0.06316116452217102,
0.03038584068417549,
-0.060215938836336136,
0.03969857096672058,
-0.059671152383089066,
-0.03371081128716469,
-0.03892991319298744,
0.06461191177368164,
-0.007707381155341864,
-0.031991757452487946,
0.054996345192193985,
-0.04573513939976692,
0.03399840369820595,
-0.021613188087940216,
0.03300819918513298,
-0.05349378660321236,
-0.07271920889616013,
0.024571068584918976,
0.05662650614976883,
-0.06135374680161476,
0.07646381109952927,
0.04309388995170593,
-0.049241531640291214,
0.10199093818664551,
-0.018799971789121628,
0.031007902696728706,
-0.10587155818939209,
0.03225041925907135,
0.030646629631519318,
0.0346541702747345,
-0.04109395295381546,
-0.13396893441677094,
-0.05590037256479263,
-0.052295077592134476,
0.020194612443447113,
-0.03341594338417053,
0.06781189888715744,
0.07523780316114426,
-0.05247380584478378,
0.009411518462002277,
0.06120244786143303,
-0.06741245836019516,
-0.00011086103040724993,
0.013801416382193565,
0.010031264275312424,
-0.05482535436749458,
-0.011549538001418114,
-0.029173189774155617,
-0.013105279766023159,
-0.011075744405388832,
0.03262333944439888,
-0.023201853036880493,
-0.013027437962591648,
0.03280124440789223,
-0.03823722153902054,
0.032224029302597046,
-0.024418838322162628,
0.05100135877728462,
0.08870217949151993,
-0.0038051465526223183,
-0.04958405718207359,
-0.05072220042347908,
-0.05450922250747681,
-0.003828550223261118,
-0.004279812797904015,
-0.04688779264688492,
-0.008946237154304981,
0.049379728734493256,
-0.0032739266753196716,
0.030286885797977448,
0.03625630959868431,
-0.007232744712382555,
0.03921766206622124,
0.08524266630411148,
-0.12730012834072113,
0.02276323363184929,
0.02076243981719017,
0.014847395941615105,
-0.022169332951307297,
0.02206135354936123,
0.007076501380652189,
0.011973458342254162,
-0.0037749430630356073,
0.006972246337682009,
0.16140906512737274,
-0.024953151121735573,
0.055437251925468445,
-0.010396899655461311,
0.019905149936676025,
-0.03198220580816269,
0.02674306184053421,
0.014578902162611485,
-0.0051462166011333466,
0.023464437574148178,
0.006803054362535477,
-4.547224463635757e-8,
0.011166826821863651,
0.08680011332035065,
-0.003436262020841241,
-0.011908196844160557,
0.12604571878910065,
-0.035688724368810654,
0.11286837607622147,
0.07146520167589188,
-0.037473298609256744,
0.07662054896354675,
0.05072921887040138,
0.0498071014881134,
-0.002187351230531931,
-0.016294369474053383,
-0.00039429168100468814,
0.014350315555930138,
-0.000429368345066905,
-0.022288747131824493,
0.03607596457004547,
-0.08731341361999512,
-0.05082497373223305,
0.0410337969660759,
0.010089735500514507,
-0.023784112185239792,
-0.042299047112464905,
-0.07885509729385376,
-0.07145961374044418,
-0.007475895807147026,
0.002651410875841975,
0.0523921363055706,
-0.02391764521598816,
0.022884102538228035,
0.08731552958488464,
0.0012109536910429597,
0.0018003735458478332,
0.08630598336458206,
-0.039484478533267975,
0.06065714731812477,
-0.11955985426902771,
-0.01226489245891571,
-0.06347964704036713,
-0.00960766151547432,
0.0505911149084568,
-0.053240057080984116,
0.04614663124084473,
-0.12354516983032227,
-0.02996862307190895,
-0.0452931709587574,
0.0040033431723713875,
-0.023238690569996834,
-0.023399773985147476,
0.027785426005721092,
0.013677818700671196,
-0.013978367671370506,
0.02029690146446228,
0.038781411945819855,
-0.019019590690732002,
-0.08978410810232162,
-0.06464763730764389,
0.07412931323051453,
0.041464291512966156,
0.04106835648417473,
-0.013949951156973839,
-0.00024264348030555993
] |
On the Optimization and Generalization of Two-layer Transformers with Sign Gradient Descent | https://openreview.net/forum?id=97rOQDPmk2 | [
"Bingrui Li",
"Wei Huang",
"Andi Han",
"Zhanpeng Zhou",
"Taiji Suzuki",
"Jun Zhu",
"Jianfei Chen"
] | Spotlight | The Adam optimizer is widely used for transformer optimization in practice, which makes understanding the underlying optimization mechanisms an important problem.
However, due to the Adam's complexity, theoretical analysis of how it optimizes transformers remains a challenging task.
Fortunately, Sign Gradient Descent (SignGD) serves as an effective surrogate for Adam.
Despite its simplicity, theoretical understanding of how SignGD optimizes transformers still lags behind.
In this work, we study how SignGD optimizes a two-layer transformer -- consisting of a softmax attention layer with trainable query-key parameterization followed by a linear layer -- on
a linearly separable noisy dataset.
We identify four stages in the training dynamics, each exhibiting intriguing behaviors.
Based on the training dynamics, we prove the fast convergence but poor generalization of the learned transformer on the noisy dataset.
We also show that Adam behaves similarly to SignGD in terms of both optimization and generalization in this setting.
Additionally, we find that the poor generalization of SignGD is not solely due to data noise,
suggesting that both SignGD and Adam requires high-quality data for real-world tasks.
Finally, experiments on synthetic and real-world datasets empirically support our theoretical results. | Sign Gradient Descent; Transformer; Training Dynamics; Theory | We study the optimization dynamics and generalization properties of a two-layer transformer trained on a signal-noise dataset. | 10,289 | 2410.04870 | [
-0.08291740715503693,
-0.08158556371927261,
0.06278339773416519,
0.020271439105272293,
0.029962273314595222,
-0.005131040699779987,
0.034650009125471115,
-0.029657328501343727,
-0.019273146986961365,
-0.07478844374418259,
-0.030342107638716698,
0.006026478949934244,
0.01157713495194912,
0.027753222733736038,
-0.03221726790070534,
0.005947068799287081,
0.0598757341504097,
0.14419229328632355,
-0.10895588248968124,
-0.02727893367409706,
0.026804834604263306,
-0.0002574262907728553,
0.04180852696299553,
-0.03730284050107002,
0.0546347051858902,
-0.027549488469958305,
-0.03676622360944748,
-0.070229172706604,
0.02169858291745186,
0.004380352329462767,
0.059993769973516464,
-0.062079910188913345,
-0.0038547555450350046,
0.05521778017282486,
-0.10531525313854218,
0.0386430062353611,
-0.026303401216864586,
0.022998619824647903,
-0.02278231829404831,
-0.03138703852891922,
0.01939072459936142,
-0.012673020362854004,
-0.010820555500686169,
0.011106264777481556,
0.06530730426311493,
0.03143841400742531,
0.06695248186588287,
-0.04955923557281494,
0.034707289189100266,
-0.033950526267290115,
-0.0019482641946524382,
-0.08612511307001114,
-0.026366280391812325,
0.06435097754001617,
0.011497984640300274,
0.0487515963613987,
-0.002468290738761425,
0.045849211513996124,
0.025832125917077065,
0.03431079909205437,
-0.018209246918559074,
-0.07683621346950531,
0.07422322034835815,
-0.05056476593017578,
-0.03938519209623337,
-0.043842609971761703,
0.07955974340438843,
-0.004513793624937534,
-0.051171861588954926,
-0.006852359976619482,
-0.014302884228527546,
0.06962575018405914,
-0.08554436266422272,
-0.010567705146968365,
0.1098586693406105,
0.0572616346180439,
0.06170089542865753,
-0.03516869619488716,
-0.001254257862456143,
-0.03620792180299759,
0.014306467026472092,
-0.035830043256282806,
-0.0017953144852072,
0.02042086236178875,
0.12453419715166092,
-0.04957083985209465,
-0.01227748766541481,
-0.02539040520787239,
-0.050804708153009415,
0.03537149727344513,
-0.12304391711950302,
-0.03449739143252373,
-0.025968972593545914,
-0.02401532046496868,
0.06003011390566826,
0.05096219480037689,
-0.026564573869109154,
-0.05538417398929596,
-0.05879809334874153,
0.05887959152460098,
-0.05516335740685463,
0.005513711366802454,
0.06493331491947174,
-0.021480169147253036,
0.03667869418859482,
-0.08819426596164703,
0.0368424728512764,
-0.0003381648566573858,
0.015738701447844505,
-0.09171595424413681,
0.04412198066711426,
0.06642085313796997,
0.02206183783710003,
-0.04856839403510094,
0.10265883058309555,
-0.003481562016531825,
-0.01718195341527462,
0.017478978261351585,
-0.047957733273506165,
0.042909443378448486,
-0.02187531627714634,
0.04645480588078499,
-0.07218310981988907,
0.056766021996736526,
0.043570298701524734,
-0.0023603965528309345,
-0.03202488273382187,
2.809960287252145e-33,
0.008193783462047577,
0.07137154787778854,
0.019251082092523575,
-0.006665111519396305,
0.009350692853331566,
-0.0037954533472657204,
0.023621074855327606,
0.041191890835762024,
-0.005945863667875528,
0.05846136808395386,
-0.051106639206409454,
-0.003432673402130604,
-0.06053294241428375,
0.0931953713297844,
0.052170075476169586,
0.005786716938018799,
-0.024089979007840157,
-0.007142374757677317,
-0.00952843576669693,
-0.037083111703395844,
0.030110184103250504,
-0.008427901193499565,
0.021553440019488335,
-0.037959448993206024,
-0.0359957292675972,
-0.029644601047039032,
0.052674345672130585,
-0.010747061111032963,
-0.029896806925535202,
-0.04617859423160553,
-0.12495775520801544,
0.00044789386447519064,
-0.003967856522649527,
0.00408327579498291,
0.058903299272060394,
-0.06408672034740448,
0.033443428575992584,
0.009391912259161472,
-0.022418759763240814,
-0.0055657317861914635,
-0.008158632554113865,
0.05730176344513893,
-0.029292453080415726,
0.010866199620068073,
-0.048468444496393204,
-0.024771256372332573,
0.04322321340441704,
0.007141421549022198,
0.05547749996185303,
-0.06596175581216812,
-0.02029765397310257,
0.018650420010089874,
-0.05263052135705948,
-0.0392485186457634,
0.0312926322221756,
0.02209634706377983,
0.09732882678508759,
0.06660617142915726,
-0.02005624584853649,
0.031952641904354095,
-0.012973463162779808,
0.00637860270217061,
0.007773094344884157,
0.045966897159814835,
-0.02900165505707264,
-0.012662937864661217,
0.03776878863573074,
0.04786597564816475,
0.03893714398145676,
-0.03598489239811897,
0.042940981686115265,
-0.008982501924037933,
-0.0068527949042618275,
-0.08577418327331543,
0.09413483738899231,
-0.012253228574991226,
0.06216731667518616,
-0.024843081831932068,
-0.0022697679232805967,
0.010784202255308628,
-0.06076338514685631,
0.048095423728227615,
-0.042794447392225266,
-0.007115571293979883,
-0.07426949590444565,
-0.0155332675203681,
0.001205095206387341,
-0.05661327391862869,
-0.03154245764017105,
0.053660690784454346,
-0.10678274929523468,
0.014160792343318462,
-0.07352706789970398,
-0.012030575424432755,
-0.026297956705093384,
-2.2544841051909192e-33,
-0.03482207655906677,
0.019547374919056892,
-0.08040712028741837,
0.049367114901542664,
0.011148206889629364,
-0.007994516752660275,
-0.037878137081861496,
0.0342898853123188,
-0.054901354014873505,
0.01623084768652916,
0.016456712037324905,
-0.07706829160451889,
0.03492589294910431,
-0.07344114035367966,
0.057007819414138794,
-0.016168907284736633,
-0.027195485308766365,
0.04970283806324005,
0.01427436713129282,
0.012289926409721375,
0.01761230081319809,
0.17021149396896362,
-0.05027728155255318,
0.03345326706767082,
-0.06008034199476242,
0.011064276099205017,
-0.09220141917467117,
0.14092929661273956,
0.018700391054153442,
-0.018853845074772835,
-0.05960353463888168,
-0.021479889750480652,
-0.0242159366607666,
0.005497426725924015,
-0.049249317497015,
0.06173477694392204,
-0.0183817520737648,
0.010750208981335163,
-0.021178647875785828,
0.10548423230648041,
0.0384046696126461,
-0.03935744985938072,
-0.039927784353494644,
0.04712453484535217,
-0.0010598128428682685,
-0.02810019813477993,
-0.0023250775411725044,
-0.0802205353975296,
-0.05060506612062454,
0.10523410141468048,
0.047856613993644714,
0.043591007590293884,
-0.08216552436351776,
-0.019274307414889336,
-0.05850076302886009,
-0.04419504478573799,
-0.06981583684682846,
0.007323107682168484,
0.1623941957950592,
0.006974277552217245,
-0.01990128681063652,
-0.024381641298532486,
0.011472665704786777,
-0.062465980648994446,
-0.06704386323690414,
0.01068932469934225,
0.02936135046184063,
-0.03063069097697735,
0.04590053856372833,
0.0816102921962738,
0.03551635146141052,
-0.007546315435320139,
0.05292334407567978,
0.0341116338968277,
-0.04464416205883026,
-0.003928883001208305,
-0.03442991152405739,
-0.05181264504790306,
-0.038062095642089844,
-0.10162640362977982,
0.0008985089953057468,
0.015719667077064514,
-0.012158650904893875,
0.040967535227537155,
0.05631623417139053,
0.08292816579341888,
0.050485387444496155,
0.06889425218105316,
0.05977039784193039,
-0.012982858344912529,
0.007778610568493605,
0.02163386344909668,
-0.05705804005265236,
-0.005253979936242104,
-0.0784630998969078,
-4.3058104637339056e-8,
-0.1192401722073555,
0.0885409563779831,
0.00011246588110225275,
-0.027185188606381416,
0.07580931484699249,
-0.037675194442272186,
0.059808917343616486,
0.1817326843738556,
-0.03276507556438446,
-0.002004998503252864,
0.08930416405200958,
0.00941696297377348,
-0.020154086872935295,
-0.053243789821863174,
0.016327448189258575,
0.05091759189963341,
-0.05254945531487465,
-0.0007251253700815141,
0.000943745719268918,
-0.046183835715055466,
-0.007919734343886375,
0.08218643814325333,
-0.07376568764448166,
-0.027618927881121635,
0.06418617814779282,
-0.0688798800110817,
0.015604577958583832,
0.07272080332040787,
0.043304137885570526,
0.014780602417886257,
0.014376627281308174,
0.016482926905155182,
0.027498625218868256,
-0.09149061143398285,
-0.019071731716394424,
0.10441799461841583,
0.03961312770843506,
-0.010151825845241547,
-0.03285403922200203,
0.010532691143453121,
-0.011672984808683395,
0.10304746776819229,
-0.024821331724524498,
0.020077310502529144,
-0.06423646211624146,
0.02072281949222088,
0.007113451138138771,
-0.08047156780958176,
0.05025064945220947,
-0.023428093641996384,
0.13344375789165497,
-0.05993950739502907,
0.012607820332050323,
0.026368726044893265,
0.026764025911688805,
-0.03399144858121872,
0.033124253153800964,
-0.08907017856836319,
0.0014993296936154366,
0.09291020035743713,
-0.011507773771882057,
0.015822190791368484,
-0.10289857536554337,
0.02817358635365963
] |
Emergent Orientation Maps —— Mechanisms, Coding Efficiency and Robustness | https://openreview.net/forum?id=rySLejeB1k | [
"Haixin Zhong",
"Haoyu Wang",
"Wei P Dai",
"Yuchao Huang",
"Mingyi Huang",
"Rubin Wang",
"Anna Wang Roe",
"yuguo yu"
] | Spotlight | Extensive experimental studies have shown that in lower mammals, neuronal orientation preference in the primary visual cortex is organized in disordered "salt-and-pepper" organizations. In contrast, higher-order mammals display a continuous variation in orientation preference, forming pinwheel-like structures. Despite these observations, the spiking mechanisms underlying the emergence of these distinct topological structures and their functional roles in visual processing remain poorly understood. To address this, we developed a self-evolving spiking neural network model with Hebbian plasticity, trained using physiological parameters characteristic of rodents, cats, and primates, including retinotopy, neuronal morphology, and connectivity patterns. Our results identify critical factors, such as the degree of input visual field overlap, neuronal connection range, and the balance between localized connectivity and long-range competition, that determine the emergence of either salt-and-pepper or pinwheel-like topologies. Furthermore, we demonstrate that pinwheel structures exhibit lower wiring costs and enhanced sparse coding capabilities compared to salt-and-pepper organizations. They also maintain greater coding robustness against noise in naturalistic visual stimuli. These findings suggest that such topological structures confer significant computational advantages in visual processing and highlight their potential application in the design of brain-inspired deep learning networks and algorithms. | Vision, Energy Efficient Coding, Neural Network, Sensory Coding, Spiking Mechanisms | null | 10,274 | null | [
0.0013471691636368632,
-0.18476173281669617,
0.052665214985609055,
0.040279705077409744,
-0.03780590742826462,
0.030178051441907883,
-0.018431512638926506,
-0.016962630674242973,
0.09252002835273743,
-0.01936183124780655,
-0.009078579023480415,
-0.03802119567990303,
0.007702099625021219,
0.03070753440260887,
0.037078458815813065,
-0.04848557710647583,
-0.011597727425396442,
0.018168553709983826,
0.0006646759575232863,
0.02292298525571823,
0.05184316262602806,
-0.07025289535522461,
0.0021183404605835676,
0.010536422953009605,
-0.05443919822573662,
0.08893304318189621,
0.025473393499851227,
-0.010671432130038738,
-0.011080536060035229,
-0.06325061619281769,
0.06101304292678833,
0.018710948526859283,
-0.018539991229772568,
-0.023321403190493584,
-0.033632420003414154,
0.014477320946753025,
-0.01531994715332985,
-0.027428096160292625,
-0.01157830748707056,
-0.018228014931082726,
0.01777559146285057,
0.03120638243854046,
0.08324023336172104,
0.027644651010632515,
0.03650294244289398,
0.04435563087463379,
0.027228394523262978,
-0.045695479959249496,
-0.02296307496726513,
-0.05394004285335541,
-0.03839018568396568,
0.013684292323887348,
0.06743496656417847,
-0.021253375336527824,
0.0492384135723114,
0.1761060208082199,
0.02394212782382965,
0.01821964420378208,
0.02041735127568245,
-0.0021663836669176817,
0.058095164597034454,
-0.01139827910810709,
0.04907660558819771,
-0.008668957278132439,
0.03560008481144905,
-0.006408669054508209,
0.020780695602297783,
0.02962305024266243,
0.043532922863960266,
-0.030752336606383324,
0.049762215465307236,
0.04911923035979271,
0.001256785006262362,
0.006036645267158747,
0.11702582985162735,
0.03781942278146744,
-0.027160227298736572,
-0.005789485294371843,
0.025910641998052597,
-0.035898931324481964,
-0.016679640859365463,
0.09808625280857086,
0.015165728516876698,
0.06769785284996033,
0.06730714440345764,
0.010068324394524097,
-0.03854189068078995,
0.07069768756628036,
-0.11152181029319763,
-0.012843271717429161,
-0.06368687003850937,
-0.01906426064670086,
-0.09079383313655853,
-0.10150860995054245,
-0.003872794099152088,
0.026570726186037064,
0.020507635548710823,
-0.050339117646217346,
-0.04276283085346222,
0.0007699548732489347,
0.022131972014904022,
-0.08598259836435318,
0.004825284704566002,
0.032458823174238205,
0.1369284838438034,
-0.04173363000154495,
0.06864573806524277,
0.036394450813531876,
-0.028316354379057884,
0.02027677185833454,
0.0009493360412307084,
0.004737398587167263,
-0.06300856918096542,
0.07208885997533798,
-0.02039075829088688,
-0.015740739181637764,
-0.0731249675154686,
0.000697220501024276,
0.10668709874153137,
0.06834684312343597,
0.07576190680265427,
-0.02125754952430725,
-0.07478359341621399,
-0.07358720898628235,
0.009421863593161106,
-0.025132298469543457,
-0.13706642389297485,
3.1055959226225e-34,
-0.03349177539348602,
-0.0019433435518294573,
0.051066044718027115,
-0.029012491926550865,
0.054037872701883316,
-0.01916065439581871,
-0.014181115664541721,
-0.03078806959092617,
0.022720275446772575,
0.053434815257787704,
-0.10718261450529099,
0.09240244328975677,
-0.04968647658824921,
0.09834471344947815,
0.05216654762625694,
-0.09243438392877579,
-0.02735724113881588,
-0.02558334544301033,
-0.018310589715838432,
-0.1359715759754181,
-0.031583938747644424,
0.03963472694158554,
-0.005159808322787285,
-0.014871370978653431,
-0.0376204177737236,
-0.009626675397157669,
-0.01398018654435873,
-0.0840720683336258,
-0.08662855625152588,
0.02319418452680111,
-0.03220465034246445,
0.0766291692852974,
-0.04914269223809242,
-0.005386537406593561,
0.04736560583114624,
-0.02137271873652935,
0.06338853389024734,
-0.03961503133177757,
0.005763662047684193,
0.004093253519386053,
0.02540128491818905,
0.009967483580112457,
-0.030051743611693382,
0.024911407381296158,
-0.03473011404275894,
0.07888609170913696,
0.06707072257995605,
0.030124487355351448,
-0.07135606557130814,
-0.05349106714129448,
0.03495423495769501,
0.005002699792385101,
0.07136073708534241,
-0.02801172249019146,
0.013408546335995197,
-0.006025776267051697,
-0.004341211635619402,
-0.016387125477194786,
-0.027023009955883026,
0.0344298854470253,
-0.04151390865445137,
0.050135284662246704,
0.008785003796219826,
0.02701818197965622,
0.07551105320453644,
-0.006266767159104347,
0.020624447613954544,
-0.015200739726424217,
0.048863790929317474,
-0.023957915604114532,
0.04593292623758316,
0.059909358620643616,
-0.09792668372392654,
-0.08746146410703659,
-0.044791556894779205,
0.014895935542881489,
0.029589075595140457,
0.04456137493252754,
-0.06245512515306473,
-0.03582083433866501,
-0.0032981140539050102,
-0.037322353571653366,
-0.03081537038087845,
0.01398359052836895,
-0.03557625785470009,
0.035474322736263275,
0.13095448911190033,
-0.028950128704309464,
-0.05120205879211426,
0.031367797404527664,
0.020917147397994995,
0.011232654564082623,
0.12325721234083176,
-0.06749173998832703,
-0.03112713433802128,
-7.508302106919392e-34,
-0.07589074224233627,
-0.051475007086992264,
0.017652742564678192,
0.000632435898296535,
-0.016283776611089706,
0.0032530815806239843,
0.05567809194326401,
-0.06066332757472992,
-0.05315915122628212,
-0.06888565421104431,
-0.024637343361973763,
0.0377572663128376,
0.015538210049271584,
-0.0076217674650251865,
0.07479152083396912,
-0.005488290451467037,
-0.04145383834838867,
0.08046268671751022,
0.0759257823228836,
-0.08703602850437164,
-0.003383458824828267,
0.04001143202185631,
-0.07882753014564514,
-0.028665633872151375,
-0.022848933935165405,
0.10417662560939789,
-0.05621343106031418,
0.002347359200939536,
0.025251947343349457,
0.0285146776586771,
-0.03589744493365288,
0.05844986438751221,
-0.04111102223396301,
-0.03012857399880886,
0.09032954275608063,
0.0638309195637703,
-0.08760202676057816,
-0.05044545233249664,
-0.07194812595844269,
-0.01980111002922058,
-0.07117372006177902,
0.0011070560431107879,
0.02319912426173687,
0.05189330875873566,
0.06315845996141434,
-0.04875223711133003,
-0.04881562292575836,
-0.012294420972466469,
-0.04085203632712364,
0.10011165589094162,
-0.0406169667840004,
0.0354112945497036,
-0.08081936836242676,
-0.14856548607349396,
-0.08989846706390381,
0.01317664422094822,
-0.011247480288147926,
-0.015389623120427132,
0.11335824429988861,
0.004527868237346411,
-0.024266544729471207,
-0.04437338188290596,
-0.04454727843403816,
0.01862117275595665,
-0.04987237602472305,
-0.05155099555850029,
0.04383176565170288,
-0.0550743043422699,
0.06297183781862259,
-0.012353137135505676,
0.0165314469486475,
0.01101198885589838,
0.059039317071437836,
-0.05642641335725784,
0.012106441892683506,
0.023549145087599754,
0.013736921362578869,
0.031908709555864334,
-0.0015038998099043965,
-0.027809126302599907,
-0.03736389800906181,
0.025255171582102776,
-0.05809332802891731,
0.0750628262758255,
0.014279531314969063,
0.05614204704761505,
-0.027020802721381187,
0.05136809125542641,
0.05160360783338547,
-0.015015191398561,
0.061410315334796906,
-0.08434531092643738,
-0.05985630676150322,
0.026015030220150948,
0.015084648504853249,
-4.389714547414769e-8,
-0.06691551953554153,
0.02432427369058132,
0.02569563128054142,
0.04546544328331947,
0.0631515234708786,
0.03616289049386978,
0.03805408254265785,
0.025994839146733284,
-0.01870088465511799,
-0.029965665191411972,
0.05468056723475456,
0.015721384435892105,
-0.04569563269615173,
0.022040490061044693,
0.03213019669055939,
0.10547146201133728,
0.004155831877142191,
0.0033954887185245752,
-0.012583701871335506,
0.06819965690374374,
-0.033675000071525574,
0.015186765231192112,
-0.0476306676864624,
-0.028934337198734283,
-0.03643856942653656,
-0.03772168606519699,
-0.03257919102907181,
0.07655363529920578,
0.04543003439903259,
-0.026891853660345078,
0.08737382292747498,
-0.0015942861791700125,
0.04601859673857689,
0.025386817753314972,
-0.01552411075681448,
0.014216292649507523,
-0.047690946608781815,
0.04479816555976868,
-0.028473855927586555,
-0.07607705891132355,
-0.05931038782000542,
0.029523443430662155,
0.025668660178780556,
0.0141607029363513,
-0.10568815469741821,
-0.052953366190195084,
0.08445774763822556,
0.025082677602767944,
0.011453935876488686,
-0.005883756093680859,
-0.1325351595878601,
0.08471370488405228,
-0.027058390900492668,
0.0438757985830307,
0.00985793024301529,
0.016629545018076897,
0.027354585006833076,
-0.04784988611936569,
-0.022374801337718964,
0.028398841619491577,
0.032853323966264725,
0.04125294089317322,
-0.061918530613183975,
0.01998859830200672
] |
How Much is Unseen Depends Chiefly on Information About the Seen | https://openreview.net/forum?id=uqWM9hBDAE | [
"Seongmin Lee",
"Marcel Boehme"
] | Spotlight | The *missing mass* refers to the proportion of data points in an *unknown* population of classifier inputs that belong to classes *not* present in the classifier's training data, which is assumed to be a random sample from that unknown population.
We find that *in expectation* the missing mass is entirely determined by the number $f_k$ of classes that *do* appear in the training data the same number of times *and an exponentially decaying error*.
While this is the first precise characterization of the expected missing mass in terms of the sample, the induced estimator suffers from an impractically high variance. However, our theory suggests a large search space of nearly unbiased estimators that can be searched effectively and efficiently. Hence, we cast distribution-free estimation as an optimization problem to find a distribution-specific estimator with a minimized mean-squared error (MSE), given only the sample.
In our experiments, our search algorithm discovers estimators that have a substantially smaller MSE than the state-of-the-art Good-Turing estimator. This holds for over 93\% of runs when there are at least as many samples as classes. Our estimators' MSE is roughly 80\% of the Good-Turing estimator's. | Good-Turing frequency estimation, Multinomial probability estimation, Unseen events, Missing mass, Probability mass | Given a random sample from an unknown multinomial distribution, we develop a distribution-free methodology to generate a distribution-specific estimator for the proportion of data that belong to classes not observed in the training sample. | 10,263 | 2402.05835 | [
-0.10311497747898102,
-0.06051197275519371,
0.008419708348810673,
0.10303782671689987,
0.10700494050979614,
0.005116429179906845,
-0.006661063525825739,
-0.014002195559442043,
0.05551464855670929,
0.028525374829769135,
-0.008750404231250286,
-0.031223496422171593,
0.07618547976016998,
-0.09334205836057663,
-0.02129819057881832,
-0.06709983944892883,
0.05271431803703308,
0.05494474992156029,
-0.04499809816479683,
0.005897852126508951,
-0.02831023931503296,
0.049131494015455246,
0.04822925105690956,
0.03392165154218674,
0.011908818036317825,
-0.03756876662373543,
-0.032635029405355453,
-0.09525873512029648,
-0.06075694039463997,
-0.01114194467663765,
0.017030814662575722,
-0.007342093158513308,
0.04307160899043083,
-0.0008944877190515399,
0.00912071205675602,
-0.011512357741594315,
-0.032999347895383835,
0.07093744724988937,
0.03606056421995163,
0.02765621617436409,
0.00012403467553667724,
-0.00025381866726092994,
0.00517222099006176,
0.0556907095015049,
0.023753024637699127,
-0.0066432044841349125,
0.019473113119602203,
-0.08977170288562775,
-0.014231679029762745,
0.010079498402774334,
-0.034307174384593964,
-0.01019578892737627,
-0.08142133057117462,
0.050486162304878235,
-0.03329053893685341,
-0.09465108066797256,
0.07435041666030884,
-0.018284417688846588,
-0.02513454481959343,
0.08197276294231415,
-0.01044261921197176,
-0.09678491950035095,
0.05261073261499405,
0.0009410748607479036,
-0.004909081384539604,
-0.022061629220843315,
0.06542449444532394,
0.013621262274682522,
0.07572496682405472,
0.02522587776184082,
0.016480427235364914,
0.10309741646051407,
-0.11761419475078583,
0.13830862939357758,
0.05057847872376442,
-0.02423943392932415,
0.052348021417856216,
0.028140300884842873,
0.04338798299431801,
0.013482814654707909,
-0.01816701330244541,
-0.06075221672654152,
-0.03348454833030701,
0.03607986494898796,
0.10923531651496887,
-0.0057234326377511024,
-0.029671072959899902,
0.046861715614795685,
0.05657880753278732,
-0.05975370481610298,
-0.0769241601228714,
0.0022752212826162577,
-0.08123672753572464,
0.02664899080991745,
0.048349518328905106,
0.042900051921606064,
0.032858725637197495,
0.0021434251684695482,
0.024952242150902748,
0.098390594124794,
-0.07350006699562073,
0.07744429260492325,
0.0447239987552166,
-0.036120202392339706,
-0.029172077775001526,
-0.011844180524349213,
0.02250099740922451,
-0.0007960210205055773,
0.010431844741106033,
-0.052135784178972244,
0.011044737882912159,
-0.04230355843901634,
-0.05587846785783768,
0.04849914833903313,
0.005402819253504276,
-0.021706728264689445,
0.0633992999792099,
-0.03070066124200821,
-0.08028800785541534,
0.07306356728076935,
-0.0744754895567894,
0.00047809482202865183,
0.007837343961000443,
0.02166251838207245,
0.033229172229766846,
0.044536393135786057,
-0.1308298110961914,
2.0691674207837856e-33,
-0.02274804748594761,
0.006680978462100029,
0.0641334131360054,
-0.052095212042331696,
0.04799789935350418,
-0.0060113659128546715,
0.07980231195688248,
0.03861610218882561,
0.07558591663837433,
0.05140308290719986,
-0.02114911749958992,
-0.02388998493552208,
0.005248085595667362,
0.039098985493183136,
0.05935376137495041,
-0.00673326663672924,
0.0021722756791859865,
0.0995478555560112,
-0.06650446355342865,
-0.0796351209282875,
0.07396215200424194,
-0.007530421018600464,
0.03958149999380112,
-0.09541735798120499,
-0.01778355985879898,
0.027216123417019844,
0.012796225026249886,
-0.05579159036278725,
-0.004895955789834261,
0.015129905194044113,
-0.06294960528612137,
0.024786461144685745,
0.06565625965595245,
-0.023362278938293457,
0.002285099821165204,
-0.0773918479681015,
0.011012214235961437,
0.03136110678315163,
-0.02900615893304348,
0.013950545340776443,
-0.04792763665318489,
0.022607827559113503,
0.07807911187410355,
0.001415778766386211,
-0.09839683026075363,
-0.10317188501358032,
0.04693315923213959,
-0.037327662110328674,
-0.000545020739082247,
-0.025037486106157303,
0.002579799387603998,
-0.004352982621639967,
-0.04085046425461769,
-0.033852413296699524,
0.026735588908195496,
0.055331721901893616,
0.01412468496710062,
0.04811040312051773,
-0.018561627715826035,
0.004796163644641638,
-0.01840745285153389,
0.00831722654402256,
0.01410466805100441,
0.09434759616851807,
0.03180284425616264,
0.015636485069990158,
-0.013127109967172146,
-0.004596642684191465,
0.04467727616429329,
-0.01126656774431467,
0.0937318354845047,
-0.05486995726823807,
-0.07060660421848297,
-0.010311664082109928,
0.07425638288259506,
0.02725512720644474,
0.08442986756563187,
-0.003996776416897774,
0.0018038582056760788,
-0.041675351560115814,
0.08783721178770065,
-0.0990559533238411,
-0.040399447083473206,
-0.09093981236219406,
-0.009021284058690071,
0.017860891297459602,
0.027406778186559677,
-0.08721106499433517,
0.028236741200089455,
0.02228124998509884,
-0.049455784261226654,
-0.03737032413482666,
-0.14358945190906525,
-0.022850599139928818,
-0.05697886645793915,
-2.3406737387161094e-33,
-0.07936916500329971,
0.047537025064229965,
-0.0029675248079001904,
0.0689447894692421,
-0.030484840273857117,
0.015259144827723503,
-0.0272484440356493,
0.017603090032935143,
-0.031693752855062485,
-0.05113659054040909,
-0.04558113217353821,
0.03724610432982445,
0.019223550334572792,
0.03560776263475418,
-0.07611400634050369,
0.06978318095207214,
0.03241664543747902,
0.05562252551317215,
0.011055096052587032,
-0.022473888471722603,
0.03366930037736893,
0.0028129995334893465,
-0.036236632615327835,
0.00229728315025568,
-0.016270775347948074,
0.05697355419397354,
-0.12270414084196091,
0.10550844669342041,
-0.04966115206480026,
-0.027295205742120743,
-0.05328134447336197,
-0.06762886792421341,
-0.044236306101083755,
-0.039264168590307236,
-0.04689626023173332,
0.028542904183268547,
0.07505696266889572,
0.10358063131570816,
-0.06877239793539047,
0.10199114680290222,
0.0018049972131848335,
0.016596296802163124,
-0.07476203888654709,
-0.010578223504126072,
-0.007037476170808077,
-0.03162245452404022,
0.0007267994806170464,
0.04150130972266197,
0.09483101218938828,
0.030855979770421982,
-0.017905695363879204,
-0.005336212459951639,
-0.11779610067605972,
0.029290974140167236,
0.001702166860923171,
-0.009056352078914642,
-0.05039579048752785,
0.0017117186216637492,
-0.03674345463514328,
0.024618225172162056,
-0.05163422226905823,
0.02393324114382267,
-0.09473855048418045,
0.0327519029378891,
-0.050864703953266144,
-0.04403708875179291,
0.021498097106814384,
0.05899664759635925,
0.023335183039307594,
0.05848529934883118,
0.07276632636785507,
-0.01405727956444025,
-0.05108420178294182,
-0.01105019636452198,
-0.03459974005818367,
0.13213121891021729,
0.03744984790682793,
-0.006682177539914846,
-0.013029753230512142,
-0.04064618796110153,
0.03864618018269539,
-0.03475905954837799,
0.02159501612186432,
-0.05286155268549919,
0.07498118281364441,
0.031352151185274124,
0.08659245818853378,
-0.04753180593252182,
0.042508091777563095,
-0.03001135215163231,
-0.07339303940534592,
0.05005325376987457,
-0.0173831544816494,
0.024975987151265144,
-0.003860528813675046,
-5.2379171222582954e-8,
0.010969962924718857,
0.09383174031972885,
0.03999488055706024,
0.007511053700000048,
0.1610480397939682,
-0.020348716527223587,
-0.07359537482261658,
0.05211496353149414,
-0.08201751857995987,
-0.07655259966850281,
0.06326095014810562,
-0.03666713088750839,
-0.0006352558848448098,
0.022952357307076454,
0.029327256605029106,
0.027731185778975487,
-0.05596308782696724,
0.02376898005604744,
-0.03130344673991203,
0.06395640224218369,
0.006118114106357098,
0.0052819461561739445,
0.04467770457267761,
-0.06084781512618065,
0.044592272490262985,
-0.07949107140302658,
0.005444114096462727,
0.016972625628113747,
0.007510301191359758,
0.010610074736177921,
-0.060936082154512405,
0.010387316346168518,
-0.013259400613605976,
-0.023065868765115738,
0.08047355711460114,
0.06105615943670273,
-0.059205979108810425,
-0.04817740246653557,
-0.09650679677724838,
-0.056820765137672424,
-0.019406113773584366,
-0.00662239408120513,
-0.0069639962166547775,
0.0032289389055222273,
0.0806548073887825,
0.009315161034464836,
-0.014996650628745556,
-0.06361521780490875,
-0.006476661656051874,
-0.04348674416542053,
0.05533406510949135,
-0.03237982094287872,
-0.02728426270186901,
0.02651924267411232,
0.01237521879374981,
0.05803912878036499,
-0.005641148891299963,
-0.03149249404668808,
-0.04493602737784386,
-0.02832372672855854,
0.027110662311315536,
-0.06176301836967468,
-0.051787640899419785,
-0.017813868820667267
] |
Harnessing Diversity for Important Data Selection in Pretraining Large Language Models | https://openreview.net/forum?id=bMC1t7eLRc | [
"Chi Zhang",
"Huaping Zhong",
"Kuan Zhang",
"Chengliang Chai",
"Rui Wang",
"Xinlin Zhuang",
"Tianyi Bai",
"Qiu Jiantao",
"Lei Cao",
"Ju Fan",
"Ye Yuan",
"Guoren Wang",
"Conghui He"
] | Spotlight | Data selection is of great significance in pretraining large language models, given the variation in quality within the large-scale available training corpora.
To achieve this, researchers are currently investigating the use of data influence to measure the importance of data instances, $i.e.,$ a high influence score indicates that incorporating this instance to the training set is likely to enhance the model performance. Consequently, they select the top-$k$ instances with the highest scores. However, this approach has several limitations.
(1) Calculating the accurate influence of all available data is time-consuming.
(2) The selected data instances are not diverse enough, which may hinder the pretrained model's ability to generalize effectively to various downstream tasks.
In this paper, we introduce $\texttt{Quad}$, a data selection approach that considers both quality and diversity by using data influence to achieve state-of-the-art pretraining results.
To compute the influence ($i.e.,$ the quality) more accurately and efficiently, we incorporate the attention layers to capture more semantic details, which can be accelerated through the Kronecker product.
For the diversity, $\texttt{Quad}$ clusters the dataset into similar data instances within each cluster and diverse instances across different clusters. For each cluster, if we opt to select data from it, we take some samples to evaluate the influence to prevent processing all instances. Overall, we favor clusters with highly influential instances (ensuring high quality) or clusters that have been selected less frequently (ensuring diversity), thereby well balancing between quality and diversity. Experiments on Slimpajama and FineWeb over 7B large language models demonstrate that $\texttt{Quad}$ significantly outperforms other data selection methods with a low FLOPs consumption. Further analysis also validates the effectiveness of our influence calculation. | LLMs, data selection, influence function, diversity | null | 10,241 | 2409.16986 | [
0.0249224454164505,
-0.0672997459769249,
0.0049482001923024654,
0.0549488291144371,
0.06566168367862701,
0.08204388618469238,
-0.004373564384877682,
0.009945773519575596,
0.036773208528757095,
-0.017571331933140755,
-0.02653254196047783,
-0.03776135295629501,
0.055138442665338516,
-0.00029101839754730463,
-0.0021316891070455313,
0.012136470526456833,
0.16636963188648224,
0.01446378044784069,
-0.12736737728118896,
-0.11114350706338882,
0.03658584877848625,
0.0013224828289821744,
0.03340527415275574,
-0.009329408407211304,
0.043443888425827026,
-0.049786277115345,
0.03406604006886482,
0.03218759968876839,
0.06456972658634186,
-0.03675394877791405,
-0.061490755528211594,
0.062129467725753784,
-0.020424554124474525,
0.019737688824534416,
-0.1085205227136612,
0.07023274898529053,
-0.036188505589962006,
-0.004171897657215595,
0.05160066485404968,
0.017169475555419922,
0.024803156033158302,
0.022409845143556595,
0.038153477013111115,
0.025148805230855942,
0.045840661972761154,
-0.02951030805706978,
-0.003809701418504119,
-0.03598502278327942,
0.03859681263566017,
0.02239449881017208,
-0.06201712414622307,
-0.044664714485406876,
0.024721460416913033,
0.037504054605960846,
-0.02563672512769699,
-0.06001100689172745,
0.026851212605834007,
0.0050252871587872505,
-0.05252908542752266,
-0.03423469886183739,
-0.083356112241745,
-0.10211549699306488,
-0.012208948843181133,
-0.05295490846037865,
-0.009695573709905148,
-0.06682375073432922,
0.012987124733626842,
0.0308667104691267,
-0.047193508595228195,
0.044534578919410706,
0.047323089092969894,
0.09679258614778519,
0.047924481332302094,
0.06315094232559204,
0.02402196079492569,
0.059424519538879395,
0.01922883279621601,
0.020001458004117012,
0.07700495421886444,
-0.03460681438446045,
0.001461769687011838,
0.022191664204001427,
0.0030968787614256144,
0.003241892671212554,
0.13803592324256897,
0.038991089910268784,
0.10463804006576538,
-0.024895191192626953,
0.04844300448894501,
0.06586695462465286,
-0.08928839862346649,
0.012639274820685387,
0.09583543241024017,
-0.03321830928325653,
-0.04779569059610367,
0.0610106997191906,
-0.0393267497420311,
-0.08982862532138824,
0.04533753544092178,
0.06913502514362335,
0.06830350309610367,
0.07164786010980606,
0.006970163434743881,
-0.020227273926138878,
-0.015128896571695805,
-0.05609847977757454,
0.03491966426372528,
0.03513671085238457,
0.0050179180689156055,
-0.04797697812318802,
0.009664302691817284,
0.07443966716527939,
-0.028355393558740616,
-0.01093963347375393,
0.060867179185152054,
0.03443169221282005,
0.005494601558893919,
-0.005501552484929562,
-0.01550673320889473,
0.09920971840620041,
-0.03424687311053276,
0.027243316173553467,
0.058423031121492386,
-0.03454054892063141,
-0.01666819304227829,
-0.023150822147727013,
-0.05250421166419983,
-5.010493242502152e-34,
0.09913536161184311,
0.022540809586644173,
0.05780262500047684,
0.04149307683110237,
0.030030477792024612,
-0.0651894137263298,
-0.03608287125825882,
0.04829396679997444,
-0.08285931497812271,
-0.001498284749686718,
-0.05289746820926666,
0.030747702345252037,
-0.04386115446686745,
0.03768444433808327,
0.029526524245738983,
0.029839269816875458,
-0.09640190750360489,
0.07464373856782913,
-0.0567299947142601,
0.023264212533831596,
0.07269911468029022,
-0.031198225915431976,
0.023993805050849915,
-0.027777377516031265,
0.02905743382871151,
-0.03614001348614693,
0.03304917737841606,
-0.042223863303661346,
-0.06741886585950851,
0.010325413197278976,
-0.08758777379989624,
-0.07568217068910599,
0.02477811835706234,
-0.00446078646928072,
-0.012325490824878216,
-0.04135344922542572,
-0.061985164880752563,
-0.04169236496090889,
0.061027802526950836,
0.033904217183589935,
0.01082237996160984,
-0.0019791191443800926,
0.05377766489982605,
0.015635279938578606,
-0.036265358328819275,
-0.0002479669637978077,
-0.012571570463478565,
0.01473049633204937,
-0.00802451279014349,
-0.04210064187645912,
0.07255323231220245,
-0.021719466894865036,
0.019452014937996864,
0.02499973401427269,
-0.01906461827456951,
0.018799373880028725,
-0.004698307253420353,
0.04249824956059456,
0.0746784582734108,
-0.0003888327337335795,
-0.02602413482964039,
-0.0008791760774329305,
-0.0339927114546299,
0.13881273567676544,
0.08834344148635864,
0.029319413006305695,
0.02426258660852909,
0.049905937165021896,
0.039757706224918365,
-0.021697157993912697,
-0.05635789781808853,
-0.036306481808423996,
0.0026191878132522106,
-0.08126332610845566,
0.05242941156029701,
0.0031277521047741175,
0.045827534049749374,
-0.11249154061079025,
0.06874267011880875,
-0.01940540410578251,
-0.05209711566567421,
-0.04068134352564812,
-0.005927794612944126,
-0.09633436799049377,
-0.06492366641759872,
0.023140117526054382,
0.0551716573536396,
-0.0883365198969841,
0.03040713630616665,
-0.02298591285943985,
-0.022743135690689087,
0.04726821929216385,
-0.05532939359545708,
-0.0768568143248558,
-0.024183399975299835,
8.433154429852482e-34,
-0.05778549984097481,
0.024138012900948524,
0.039339564740657806,
0.045106980949640274,
-0.00607158662751317,
-0.019619593396782875,
0.03881269320845604,
0.0559297613799572,
-0.07792698591947556,
-0.03734564036130905,
-0.01322103850543499,
-0.0856228619813919,
0.08720318973064423,
-0.035630397498607635,
0.04229620471596718,
0.04517447575926781,
-0.050589632242918015,
0.07165078073740005,
-0.015006539411842823,
0.08456380665302277,
0.0036318639758974314,
0.03573611006140709,
-0.007634609472006559,
0.005080288276076317,
-0.009085391648113728,
-0.023407522588968277,
-0.0768548846244812,
0.06839625537395477,
0.013874953612685204,
-0.05848006531596184,
-0.04127822443842888,
-0.025789622217416763,
-0.08164379000663757,
-0.04956991225481033,
-0.09622012823820114,
0.006632047705352306,
0.05772741511464119,
-0.04141788184642792,
-0.016558123752474785,
0.16179686784744263,
-0.010974211618304253,
-0.020174508914351463,
-0.10498254001140594,
-0.0019531776197254658,
-0.07897455990314484,
0.002695508301258087,
-0.11894765496253967,
-0.01879478618502617,
0.04036322608590126,
-0.02697872184216976,
-0.021963652223348618,
0.026682548224925995,
-0.05943181738257408,
0.09056347608566284,
-0.030248474329710007,
-0.0627136304974556,
0.03355587273836136,
-0.053630780428647995,
0.024256883189082146,
0.007165768649429083,
-0.11865414679050446,
-0.003152357181534171,
0.04645206406712532,
-0.05803883820772171,
-0.011964726261794567,
-0.06939776986837387,
-0.03233177959918976,
-0.0008344380185008049,
0.008984560146927834,
0.019950488582253456,
-0.001655225409194827,
-0.01840977370738983,
0.057576362043619156,
-0.003002721816301346,
-0.03149045258760452,
-0.003491555340588093,
0.049365658313035965,
-0.035201989114284515,
0.008210555650293827,
-0.0437663272023201,
0.030025431886315346,
0.031939852982759476,
0.019388921558856964,
-0.032697588205337524,
0.07835742831230164,
0.10260659456253052,
0.09493476897478104,
0.042219128459692,
0.023628007620573044,
0.0027644382789731026,
-0.04686456546187401,
0.035262659192085266,
-0.017347827553749084,
0.02728264220058918,
-0.038875143975019455,
-4.7385494639229364e-8,
-0.07923604547977448,
0.0499388612806797,
0.002142887329682708,
0.0946473628282547,
-0.014875167049467564,
-0.09987907856702805,
-0.10026893019676208,
0.13296765089035034,
-0.03742431849241257,
0.010490966029465199,
-0.03414167836308479,
0.04194367304444313,
-0.03817170113325119,
-0.04119586572051048,
0.026614222675561905,
0.0838538259267807,
-0.0017652901588007808,
0.021690772846341133,
0.014696498401463032,
-0.02756938338279724,
0.024900441989302635,
0.02536560408771038,
0.0915074422955513,
0.01000631507486105,
0.046492353081703186,
-0.06886720657348633,
-0.025213170796632767,
0.06545483320951462,
0.04615414887666702,
-0.0037705248687416315,
-0.014917794615030289,
-0.021499160677194595,
-0.03823977708816528,
-0.04782712459564209,
0.02483842521905899,
0.097142793238163,
-0.05461294203996658,
-0.035403311252593994,
-0.006937295664101839,
0.017213329672813416,
0.053199004381895065,
0.0013008200330659747,
-0.040880363434553146,
0.030460607260465622,
0.0236800666898489,
-0.04123656079173088,
-0.09110566973686218,
-0.07843147218227386,
0.0399802140891552,
-0.06359940767288208,
0.021458519622683525,
-0.06906863301992416,
-0.02878992073237896,
0.053272731602191925,
0.0647464171051979,
0.021163860335946083,
-0.06971697509288788,
-0.0362212248146534,
0.03106147050857544,
-0.0034385088365525007,
0.04704560711979866,
0.01639518328011036,
0.0013818343868479133,
0.009105674922466278
] |
Adaptive Gradient Clipping for Robust Federated Learning | https://openreview.net/forum?id=03OkC0LKDD | [
"Youssef Allouah",
"Rachid Guerraoui",
"Nirupam Gupta",
"Ahmed Jellouli",
"Geovani Rizk",
"John Stephan"
] | Spotlight | Robust federated learning aims to maintain reliable performance despite the presence of adversarial or misbehaving workers. While state-of-the-art (SOTA) robust distributed gradient descent (Robust-DGD) methods were proven theoretically optimal, their empirical success has often relied on pre-aggregation gradient clipping.
However, existing static clipping strategies yield inconsistent results: enhancing robustness against some attacks while being ineffective or even detrimental against others.
To address this limitation, we propose a principled adaptive clipping strategy, Adaptive Robust Clipping (ARC), which dynamically adjusts clipping thresholds based on the input gradients. We prove that ARC not only preserves the theoretical robustness guarantees of SOTA Robust-DGD methods but also provably improves asymptotic convergence when the model is well-initialized. Extensive experiments on benchmark image classification tasks confirm these theoretical insights, demonstrating that ARC significantly enhances robustness, particularly in highly heterogeneous and adversarial settings. | Federated learning, robustness, Byzantine resilience | null | 10,208 | null | [
-0.1024138554930687,
-0.06669361889362335,
0.005282706581056118,
0.09826532006263733,
0.07080008834600449,
-0.005504261236637831,
-0.057915717363357544,
-0.10011976957321167,
-0.0005779513157904148,
-0.08338283747434616,
-0.02788463793694973,
0.021759245544672012,
0.011399677954614162,
0.013656986877322197,
-0.03610219806432724,
-0.02716882899403572,
0.098417267203331,
0.12276355177164078,
-0.0954722911119461,
-0.06469884514808655,
-0.009060434997081757,
-0.03255070000886917,
0.013275818899273872,
0.04045310243964195,
0.09008340537548065,
-0.05610745772719383,
-0.030640283599495888,
-0.09022126346826553,
-0.005377121269702911,
-0.03426413610577583,
0.03924684599041939,
-0.07176743447780609,
-0.014947985298931599,
-0.009601390920579433,
-0.10337422788143158,
0.05083293467760086,
-0.015850719064474106,
0.0703170970082283,
0.07347075641155243,
-0.018628843128681183,
-0.05035572499036789,
-0.008809290826320648,
-0.003206933382898569,
-0.11777695268392563,
-0.02189267985522747,
-0.01743471622467041,
0.03250720724463463,
-0.06789013743400574,
-0.013025647960603237,
0.025203831493854523,
0.00801362656056881,
-0.04783633351325989,
-0.055743370205163956,
0.047954630106687546,
-0.021453257650136948,
0.012147781439125538,
0.045823510736227036,
0.0076589761301875114,
0.011957811191678047,
0.0329381488263607,
-0.005475626327097416,
-0.13067930936813354,
0.034211572259664536,
0.07371657341718674,
-0.010202515870332718,
-0.0350935235619545,
0.047772813588380814,
0.024332933127880096,
-0.0023819683119654655,
0.013827311806380749,
0.013629605062305927,
0.08895505964756012,
-0.10281313955783844,
0.03402436897158623,
0.05940704420208931,
0.11241784691810608,
0.049991246312856674,
0.0052736420184373856,
0.06662965565919876,
-0.08233042061328888,
0.06621582061052322,
0.06474903970956802,
0.05098176375031471,
0.022058317437767982,
0.0957578495144844,
0.0042055449448525906,
-0.06759967654943466,
0.021210655570030212,
0.041454557329416275,
-0.026716260239481926,
0.008679871447384357,
0.01969338394701481,
-0.0018397466046735644,
0.04478611797094345,
0.019627558067440987,
-0.05463410168886185,
0.0133437504991889,
-0.06729105859994888,
-0.06227400526404381,
0.10557607561349869,
-0.03856730833649635,
0.022861473262310028,
0.000547492818441242,
0.026165558025240898,
0.08175665140151978,
0.01271540392190218,
0.07793071120977402,
-0.04390458017587662,
0.025112047791481018,
-0.04007269814610481,
0.10900670289993286,
0.06546216458082199,
0.0394040048122406,
0.044370636343955994,
0.00481773167848587,
-0.026723382994532585,
-0.03765622153878212,
-0.0008549962076358497,
-0.0034810490906238556,
0.04099233075976372,
-0.023345138877630234,
-0.006894765887409449,
0.0004417991149239242,
0.04105227813124657,
0.07094287127256393,
0.015858924016356468,
-0.05258192867040634,
4.017006926831418e-33,
0.036993540823459625,
0.031079966574907303,
0.02949039824306965,
-0.09644772112369537,
0.04969684034585953,
-0.10434751957654953,
-0.005852927453815937,
0.031117932870984077,
-0.1082843542098999,
0.013232032768428326,
0.008226304315030575,
0.006119262892752886,
-0.034839969128370285,
0.10682736337184906,
0.0319431871175766,
-0.007363580632954836,
-0.0022868183441460133,
0.05445616692304611,
0.005008029751479626,
-0.022013003006577492,
0.02962232381105423,
-0.08079977333545685,
-0.052326153963804245,
-0.019642706960439682,
0.006208821199834347,
0.007423532661050558,
0.05019741505384445,
0.02070043236017227,
0.01785821095108986,
0.08699709922075272,
-0.07701827585697174,
-0.01660953275859356,
0.04808034747838974,
0.07129466533660889,
0.020312102511525154,
0.015383790247142315,
-0.01945018582046032,
-0.02673952840268612,
-0.02140401117503643,
0.053857412189245224,
0.024511190131306648,
0.048536885529756546,
0.007692434825003147,
-0.005062875337898731,
-0.11125030368566513,
-0.05680019408464432,
-0.034767694771289825,
0.0682072713971138,
-0.12512414157390594,
-0.010106280446052551,
-0.01887311227619648,
0.015958189964294434,
-0.0548405721783638,
-0.06326670199632645,
-0.0708317682147026,
-0.04560535028576851,
-0.04218549653887749,
-0.026376647874712944,
-0.01319724041968584,
0.056626543402671814,
-0.03595266491174698,
0.024058235809206963,
-0.0696311667561531,
0.09596744179725647,
-0.04999261349439621,
0.005375191103667021,
0.02652980014681816,
0.04465833678841591,
-0.008797339163720608,
-0.016017522662878036,
-0.02191699668765068,
0.0736980065703392,
-0.0640067383646965,
-0.056709252297878265,
0.055407147854566574,
-0.055328670889139175,
0.11178018152713776,
-0.04134027659893036,
0.009227930568158627,
-0.005475177429616451,
-0.07077042013406754,
0.014403742738068104,
-0.006728812586516142,
-0.05178651586174965,
-0.12185941636562347,
-0.0007898072362877429,
0.06676231324672699,
-0.01198135782033205,
-0.07811956852674484,
0.06116233766078949,
-0.07098355889320374,
-0.029571395367383957,
0.006094386335462332,
-0.004680117592215538,
0.040450241416692734,
-3.568125307186727e-33,
-0.03448593616485596,
0.06194949522614479,
-0.08615996688604355,
0.03088824264705181,
0.00667652627453208,
0.053396075963974,
-0.011997673660516739,
0.03971836715936661,
-0.08089154213666916,
-0.07788950949907303,
-0.0685315951704979,
0.04271722957491875,
-0.0053464993834495544,
-0.05053430050611496,
-0.015811849385499954,
-0.07423426955938339,
-0.041615962982177734,
0.04581010714173317,
-0.03937048465013504,
0.020834844559431076,
0.06822158396244049,
0.006596807390451431,
-0.007891514338552952,
0.10324007272720337,
0.001253790338523686,
0.06553931534290314,
-0.11795809864997864,
0.0628923773765564,
0.01751410961151123,
0.005257034674286842,
-0.0011887241853401065,
0.017051417380571365,
-0.021214880049228668,
-0.05749546363949776,
-0.06265532225370407,
0.07403464615345001,
-0.014848579652607441,
-0.0007842718041501939,
-0.035910677164793015,
0.07383976876735687,
0.0020038362126797438,
-0.03076234646141529,
-0.08351098746061325,
0.03863918408751488,
0.0245998315513134,
-0.03444305807352066,
0.019775617867708206,
0.022170500829815865,
0.007324843667447567,
0.012978374026715755,
-0.05918054282665253,
-0.0038874472957104445,
-0.06081179901957512,
0.11188225448131561,
-0.05927233397960663,
-0.034146495163440704,
0.04182842746376991,
-0.0001532814058009535,
-0.01684868335723877,
0.03823975473642349,
-0.03244216740131378,
0.004782660864293575,
-0.059618886560201645,
0.039075400680303574,
0.039970800280570984,
-0.023727551102638245,
0.002255467465147376,
0.0610089972615242,
0.011749956756830215,
0.058633845299482346,
0.03927897661924362,
0.023033112287521362,
0.05679897964000702,
0.009900784119963646,
0.0012036800617352128,
-0.022682353854179382,
-0.027688413858413696,
-0.07125474512577057,
-0.03573039174079895,
0.042061276733875275,
0.031235868111252785,
-0.022724656388163567,
0.06731311976909637,
0.041417837142944336,
0.1026337593793869,
0.060598816722631454,
0.048262666910886765,
0.0851195678114891,
0.010029839351773262,
0.04623274505138397,
0.002684238599613309,
-0.014794986695051193,
-0.03556645289063454,
-0.01596066914498806,
0.027341490611433983,
-4.128661501567876e-8,
-0.054820477962493896,
0.07701665163040161,
-0.02799222245812416,
-0.007828256115317345,
0.01613471657037735,
-0.023083698004484177,
-0.008853956125676632,
0.03503090888261795,
-0.02883388102054596,
-0.06434615701436996,
0.028920471668243408,
-0.020322559401392937,
-0.0031448518857359886,
-0.1040230542421341,
-0.012606589123606682,
0.06579098850488663,
-0.013022436760365963,
0.07242988049983978,
-0.03331325575709343,
0.007760987617075443,
-0.035370536148548126,
-0.04835357144474983,
-0.003475002944469452,
-0.10300028324127197,
0.0006884449394419789,
-0.05248359590768814,
-0.008290212601423264,
0.03869613632559776,
0.10318262130022049,
0.03866429626941681,
-0.10394081473350525,
-0.004758465103805065,
0.0062424177303910255,
0.02429516799747944,
0.024271968752145767,
0.10945919901132584,
0.023835564032197,
-0.04455888271331787,
0.018207145854830742,
-0.05432022362947464,
-0.016300974413752556,
0.030339177697896957,
-0.0030686950776726007,
-0.009890011511743069,
0.005222553852945566,
0.02316748909652233,
0.027034660801291466,
-0.035153672099113464,
0.030241012573242188,
-0.01772230491042137,
0.007016644347459078,
-0.0207852553576231,
-0.004139792174100876,
0.13712574541568756,
0.07424531877040863,
-0.017675714567303658,
0.0032389925327152014,
-0.08693499863147736,
0.004076685756444931,
0.11493225395679474,
0.04787652939558029,
-0.03534463793039322,
-0.04797486215829849,
-0.02789921686053276
] |
Imputation for prediction: beware of diminishing returns. | https://openreview.net/forum?id=D1Y2XFgsPI | [
"Marine Le Morvan",
"Gael Varoquaux"
] | Spotlight | Missing values are prevalent across various fields, posing challenges for training and deploying predictive models. In this context, imputation is a common practice, driven by the hope that accurate imputations will enhance predictions. However, recent theoretical and empirical studies indicate that simple constant imputation can be consistent and competitive. This empirical study aims at clarifying
*if* and *when* investing in advanced imputation methods yields significantly better predictions. Relating imputation and predictive accuracies across combinations of imputation and predictive models on 19 datasets, we show that imputation accuracy matters less i) when using expressive models, ii) when incorporating missingness indicators as complementary inputs, iii) matters much more for generated linear outcomes than for real-data outcomes. Interestingly, we also show that the use of the missingness indicator is beneficial to the prediction performance, even in MCAR scenarios. Overall, on real-data with powerful models, imputation quality has only a minor effect on prediction performance. Thus, investing in better imputations for improved predictions often offers limited benefits. | imputation, missing | We show that improving imputation often offers limited benefits for predictive performances with missing values: its beneficial impact is reduced with expressive models, the use of missingness indicators, and real-world (non-linear) data. | 10,201 | 2407.19804 | [
-0.08379006385803223,
-0.007605255581438541,
0.03534706309437752,
0.12116111069917679,
0.08822330832481384,
-0.000980482785962522,
-0.034815218299627304,
0.03457148000597954,
0.030001159757375717,
-0.06487555056810379,
-0.03100859373807907,
-0.059090979397296906,
0.07522738724946976,
-0.02084636315703392,
-0.002688890090212226,
-0.017379483208060265,
-0.012347538955509663,
0.007308420725166798,
-0.06794542074203491,
-0.06997993588447571,
-0.02790852077305317,
0.05966082587838173,
-0.03768622875213623,
0.0742340013384819,
0.032365892082452774,
-0.06719328463077545,
-0.006441075820475817,
0.02099277265369892,
-0.025593817234039307,
0.06479769945144653,
-0.044830676168203354,
-0.03026324138045311,
0.0032146507874131203,
-0.007278931327164173,
-0.07413897663354874,
-0.008369198068976402,
-0.04206113889813423,
0.017221027985215187,
0.03817464038729668,
-0.035870350897312164,
-0.03352682664990425,
-0.03668389096856117,
0.01754417084157467,
0.08844012022018433,
-0.03337312117218971,
-0.016080515459179878,
0.037393879145383835,
-0.1306941956281662,
-0.055980730801820755,
0.01781224086880684,
-0.1225438192486763,
-0.044502533972263336,
0.02173248864710331,
0.04253232106566429,
-0.021421557292342186,
-0.0024838498793542385,
-0.09275100380182266,
-0.00926227681338787,
-0.055754952132701874,
-0.017400743439793587,
-0.08044588565826416,
-0.05496679246425629,
-0.02054021507501602,
-0.003220944432541728,
-0.015465947799384594,
-0.04021753743290901,
0.027878202497959137,
0.033680565655231476,
0.001851047039963305,
0.06194515526294708,
0.0034586479887366295,
0.0381513312458992,
-0.0888688862323761,
0.07480507344007492,
0.034320492297410965,
0.05453536659479141,
0.0841091200709343,
0.07700278609991074,
0.0068183657713234425,
0.024650515988469124,
-0.10349629074335098,
-0.0058799791149795055,
-0.053422749042510986,
0.01437608152627945,
0.008178924210369587,
-0.03508434444665909,
0.03710496798157692,
-0.004902660381048918,
-0.04692152887582779,
-0.022981183603405952,
-0.03105001524090767,
-0.013682977296411991,
0.0273766927421093,
-0.01736925169825554,
0.03247066214680672,
0.1275217980146408,
0.03787989541888237,
-0.05723080784082413,
-0.012920846231281757,
0.03462689369916916,
-0.033396270126104355,
0.0946044847369194,
-0.025450410321354866,
-0.03244908154010773,
0.0016641541151329875,
-0.0828016921877861,
0.06318120658397675,
0.002684934064745903,
-0.005508598405867815,
0.003491802141070366,
0.01414124108850956,
0.011061673052608967,
0.03865586593747139,
-0.021868234500288963,
0.04452569782733917,
0.015468629077076912,
-0.038720179349184036,
-0.03229907900094986,
-0.00605902960523963,
0.08213290572166443,
-0.016520168632268906,
0.004283488728106022,
0.13041375577449799,
0.025875946506857872,
0.008964410051703453,
-0.002453635213896632,
-0.08416040986776352,
5.1761721917528454e-33,
-0.01998433656990528,
-0.021672623232007027,
0.058897122740745544,
0.002813402796164155,
-0.009933356195688248,
-0.07843046635389328,
-0.0343145951628685,
-0.028496744111180305,
0.05110473185777664,
-0.028270026668906212,
-0.060430560261011124,
0.02717166766524315,
-0.014463959261775017,
0.05102169141173363,
-0.02631411887705326,
0.06192466616630554,
0.08804035186767578,
0.03028368204832077,
-0.06603066623210907,
0.06040123477578163,
0.10357904434204102,
-0.014395970851182938,
-0.042062561959028244,
-0.07805410027503967,
0.04062006250023842,
0.09378952533006668,
0.0036023191642016172,
0.020175466313958168,
-0.025182604789733887,
0.010374307632446289,
-0.05652407184243202,
-0.003287513507530093,
0.08787117153406143,
-0.003529641078785062,
-0.035027094185352325,
-0.031969573348760605,
-0.019295530393719673,
-0.04025649651885033,
0.022276852279901505,
0.03855080157518387,
-0.046994857490062714,
0.00010705347085604444,
-0.03516902029514313,
-0.05058950185775757,
0.013073512353003025,
-0.09206153452396393,
0.06555033475160599,
-0.018484540283679962,
-0.04581853747367859,
0.051761727780103683,
0.08003636449575424,
0.01523149199783802,
-0.018074432387948036,
-0.01476376224309206,
-0.11795692890882492,
0.03773786872625351,
-0.0026545198634266853,
-0.07708790898323059,
-0.015468277037143707,
-0.058427855372428894,
-0.05161353200674057,
0.022494986653327942,
-0.013384492136538029,
-0.03439231961965561,
-0.05904724821448326,
0.03221318498253822,
0.012930935248732567,
0.0068229916505515575,
0.010054661892354488,
-0.03531394526362419,
0.019214536994695663,
-0.061631232500076294,
-0.0786842331290245,
0.004644566681236029,
0.08283385634422302,
-0.03300639986991882,
0.03552481159567833,
-0.07620294392108917,
0.07344581931829453,
0.026528140529990196,
0.023797156289219856,
0.017232777550816536,
0.012707263231277466,
-0.046030040830373764,
0.019550789147615433,
-0.011578088626265526,
0.0842885822057724,
-0.07153865694999695,
-0.02198997139930725,
-0.05333545431494713,
-0.06230325251817703,
-0.011621976271271706,
-0.08579153567552567,
0.012456201016902924,
-0.027272837236523628,
-6.207512364255934e-33,
-0.03686774894595146,
0.008530148304998875,
0.021426748484373093,
0.06365243345499039,
-0.05122782289981842,
-0.042416054755449295,
0.032730817794799805,
-0.02417891100049019,
0.06955327838659286,
-0.10819076746702194,
0.04204835370182991,
-0.04774584248661995,
0.063522569835186,
-0.0017537715611979365,
-0.08011627942323685,
0.05305761471390724,
-0.048854947090148926,
-0.08622007817029953,
-0.017472457140684128,
-0.004684160929173231,
0.08377708494663239,
0.054056260734796524,
-0.06580308824777603,
0.03646277263760567,
-0.06821098178625107,
0.034657180309295654,
-0.034178510308265686,
0.0389503575861454,
-0.08647046983242035,
-0.11200614273548126,
-0.0038900284562259912,
-0.06431242823600769,
-0.002321387641131878,
-0.1095336303114891,
-0.023802705109119415,
-0.004526430740952492,
0.04706890508532524,
0.024387963116168976,
-0.10926643013954163,
0.08304772526025772,
0.08991295844316483,
0.0005166569026187062,
-0.09242335706949234,
0.057629555463790894,
-0.0441240556538105,
0.01350515428930521,
0.0064989603124558926,
0.011297018267214298,
0.14398719370365143,
0.023626545444130898,
0.010932167991995811,
0.011905664578080177,
-0.06806128472089767,
0.0313548743724823,
-0.003844604827463627,
-0.05689027160406113,
0.010783853009343147,
0.002110919216647744,
0.014404426328837872,
0.04091034084558487,
-0.10198963433504105,
0.015097036026418209,
0.018225587904453278,
0.007713413797318935,
-0.014475997537374496,
-0.05061773955821991,
0.008539503440260887,
0.043512724339962006,
0.05374623462557793,
-0.013661839067935944,
-0.01065690815448761,
0.01993621699512005,
-0.039466552436351776,
-0.042352061718702316,
0.03252853825688362,
0.10256308317184448,
0.02456802688539028,
-0.022390326485037804,
-0.0024671887513250113,
0.04662838205695152,
0.040932584553956985,
0.0202786922454834,
-0.01735720969736576,
-0.05279146879911423,
0.03331824019551277,
0.08931432664394379,
0.03821173682808876,
-0.01574103534221649,
-0.0059448955580592155,
-0.021131565794348717,
-0.12058647722005844,
0.0544908381998539,
0.0020289032254368067,
0.06528092175722122,
-0.06260669976472855,
-4.47716352880434e-8,
0.04582061618566513,
0.14237508177757263,
0.032097529619932175,
0.07178066670894623,
-0.0014892506878823042,
-0.06003858894109726,
-0.03903917223215103,
0.06215585768222809,
-0.05162734538316727,
0.029093340039253235,
0.06071086972951889,
-0.06993789970874786,
-0.07720723003149033,
0.01982438564300537,
0.08024793118238449,
0.03496839478611946,
0.07241357862949371,
0.026531362906098366,
-0.0358346551656723,
0.018441572785377502,
0.02805347368121147,
0.05220860615372658,
-0.027567069977521896,
-0.05309367552399635,
0.11735055595636368,
-0.037848446518182755,
-0.007662071380764246,
0.049310699105262756,
-0.0402936227619648,
0.0524274967610836,
-0.03860874101519585,
0.006283230148255825,
0.046665698289871216,
-0.02813463658094406,
-0.007561775855720043,
0.08721425384283066,
0.10076262056827545,
-0.028310758993029594,
-0.09330200403928757,
-0.032662130892276764,
0.003277750685811043,
0.033334244042634964,
-0.014844421297311783,
-0.002630973234772682,
0.04591644927859306,
-0.018653174862265587,
0.0013734729727730155,
0.0032489080913364887,
0.03759472444653511,
-0.03387856110930443,
0.02812006324529648,
-0.026303090155124664,
0.03686881437897682,
0.05372697860002518,
0.12239199876785278,
0.004301173612475395,
0.03024463541805744,
0.05259045585989952,
0.0011545753804966807,
0.019561907276511192,
0.09409040957689285,
-0.08800555020570755,
-0.016963208094239235,
0.08369795233011246
] |
Joint Gradient Balancing for Data Ordering in Finite-Sum Multi-Objective Optimization | https://openreview.net/forum?id=rdAbEn5DZt | [
"Hansi Yang",
"James Kwok"
] | Spotlight | In finite-sum optimization problems, the sample orders for parameter updates can significantly influence the convergence rate of optimization algorithms. While numerous sample ordering techniques have been proposed in the context of single-objective optimization, the problem of sample ordering in finite-sum multi-objective optimization has not been thoroughly explored. To address this gap, we propose a sample ordering method called JoGBa, which finds the sample orders for multiple objectives by jointly performing online vector balancing on the gradients of all objectives. Our theoretical analysis demonstrates that this approach outperforms the standard baseline of random ordering and accelerates the convergence rate for the MGDA algorithm. Empirical evaluation across various datasets with different multi-objective optimization algorithms further demonstrates that JoGBa can achieve faster convergence and superior final performance than other data ordering strategies. | multi-objective optimization | null | 10,113 | null | [
-0.057454194873571396,
-0.04591664299368858,
0.002194654894992709,
-0.008429549634456635,
0.006936517544090748,
-0.039600495249032974,
-0.04080570489168167,
-0.06919126957654953,
-0.01724306307733059,
0.011224731802940369,
-0.10623138397932053,
0.03130223602056503,
0.044334299862384796,
-0.045095086097717285,
0.0001559925585752353,
0.047276392579078674,
0.04731505736708641,
0.08832445740699768,
-0.026227669790387154,
-0.04269825294613838,
0.024674896150827408,
0.06839575618505478,
0.023648198693990707,
-0.05394301190972328,
0.035363439470529556,
-0.021944507956504822,
-0.004953868221491575,
-0.04274126514792442,
0.030727215111255646,
0.007443426642566919,
0.08110471069812775,
0.03240577504038811,
0.028685515746474266,
-0.026136448606848717,
-0.03430289030075073,
-0.0116270137950778,
-0.1087782233953476,
-0.02106904238462448,
-0.001172548159956932,
0.05236559733748436,
-0.013947851955890656,
0.07252486795186996,
0.001460658386349678,
-0.04997663572430611,
-0.007051417138427496,
0.030775563791394234,
-0.015479274094104767,
-0.035195522010326385,
0.05921376124024391,
0.04555170610547066,
-0.045739270746707916,
-0.03526626527309418,
-0.15272894501686096,
0.0934109091758728,
-0.04724447429180145,
-0.02814587950706482,
0.0388052798807621,
-0.01666182465851307,
0.043689124286174774,
0.04438094422221184,
-0.03377785533666611,
-0.12059188634157181,
0.017589544877409935,
0.02600828744471073,
0.08484107255935669,
-0.11264324933290482,
0.03875347226858139,
0.026130463927984238,
-0.04073194041848183,
0.03502941504120827,
0.002652344759553671,
0.0307527557015419,
-0.048103250563144684,
0.0466962493956089,
-0.05415835604071617,
0.12124636769294739,
0.0638219490647316,
-0.053472768515348434,
-0.014738308265805244,
-0.009417758323252201,
-0.040784887969493866,
0.05153113231062889,
0.039176471531391144,
-0.0013210881734266877,
0.11369394510984421,
-0.05570349469780922,
-0.02860778383910656,
-0.05268441513180733,
0.0681159719824791,
0.03706429526209831,
0.01800568588078022,
0.06351225078105927,
0.006756452843546867,
0.09207732230424881,
-0.02710520662367344,
-0.05649682879447937,
0.021885229274630547,
-0.006586740724742413,
-0.0074719772674143314,
0.05639071390032768,
-0.011724387295544147,
0.05122770741581917,
0.029453014954924583,
-0.004932926036417484,
0.005995890591293573,
-0.10386922955513,
0.05312000587582588,
0.0075922152027487755,
0.01639648899435997,
-0.09572113305330276,
0.07815845310688019,
0.04616633430123329,
0.041815146803855896,
-0.005256309639662504,
-0.09885482490062714,
0.0040192874148488045,
-0.03628106042742729,
0.010110796429216862,
-0.02218012884259224,
0.03774021565914154,
-0.0687313973903656,
-0.052123721688985825,
0.04802830517292023,
-0.00006726846913807094,
-0.020024379715323448,
0.09880610555410385,
-0.018457507714629173,
3.606129406890796e-33,
-0.05879887193441391,
-0.08029492944478989,
0.09600210189819336,
-0.03213030472397804,
-0.019822433590888977,
-0.0006578377797268331,
0.0680733397603035,
0.008834836073219776,
-0.043128713965415955,
0.032721616327762604,
0.01494535617530346,
0.03618382662534714,
-0.03339515998959541,
0.014404571615159512,
0.015260075218975544,
-0.057464100420475006,
-0.015368747524917126,
0.09455439448356628,
0.018227193504571915,
0.014329700730741024,
0.0876191109418869,
-0.054281506687402725,
0.019598595798015594,
-0.02600904181599617,
0.00462872302159667,
0.007079910486936569,
0.05249616876244545,
0.008745159953832626,
-0.024080010131001472,
0.03880220279097557,
-0.058382656425237656,
-0.03591448813676834,
-0.044459018856287,
0.04282348230481148,
0.010017769411206245,
0.02146792970597744,
0.03698241338133812,
-0.03288448974490166,
-0.06508012115955353,
0.04868146404623985,
-0.05835120752453804,
0.0236787348985672,
0.036479443311691284,
-0.07541467994451523,
-0.07879097014665604,
-0.0475265197455883,
-0.05824185162782669,
0.04318751394748688,
0.09826137870550156,
-0.041177645325660706,
-0.05824420601129532,
-0.006401787977665663,
-0.03326931223273277,
0.028087394312024117,
-0.0036812706384807825,
0.04352631792426109,
-0.057740334421396255,
-0.021092766895890236,
0.03743811696767807,
0.042389754205942154,
-0.09913653135299683,
0.003279821015894413,
-0.030593810603022575,
0.0322345532476902,
0.06994469463825226,
0.05724300816655159,
0.028664255514740944,
-0.05772322416305542,
0.06753131747245789,
0.0024081014562398195,
0.03245376795530319,
0.03375575691461563,
-0.04119211807847023,
-0.03745587170124054,
0.0027845907025039196,
0.007730573881417513,
0.07541435956954956,
0.025916406884789467,
0.0009810240007936954,
0.006535337772220373,
0.04816843196749687,
0.14338712394237518,
0.04067092016339302,
-0.05651307478547096,
0.02474900148808956,
-0.022012606263160706,
-0.0151445297524333,
-0.019006187096238136,
-0.08381606638431549,
-0.00438314862549305,
-0.058716416358947754,
0.05059344321489334,
-0.031074246391654015,
0.019735202193260193,
-0.009750349447131157,
-2.8177771410141208e-33,
-0.0390981063246727,
-0.0459350124001503,
0.03867271542549133,
0.051841359585523605,
0.12569569051265717,
-0.07475647330284119,
-0.0580962598323822,
-0.10200812667608261,
-0.01284835021942854,
0.02479635365307331,
0.010731110349297523,
-0.08029939979314804,
-0.003477766877040267,
-0.03362763300538063,
0.04600546881556511,
0.024001242592930794,
-0.05075419321656227,
0.025243988260626793,
-0.027002092450857162,
0.006310522090643644,
-0.002616018755361438,
0.06417207419872284,
0.02756825089454651,
0.0050056809559464455,
-0.009952492080628872,
-0.030977535992860794,
0.057354606688022614,
0.02904498763382435,
0.011593319475650787,
-0.016715476289391518,
-0.02357693947851658,
-0.04448806494474411,
-0.09824498742818832,
0.061809100210666656,
-0.02674686349928379,
0.04837963730096817,
0.012868775986135006,
0.020159726962447166,
0.015723809599876404,
0.15501724183559418,
0.046702779829502106,
0.05436103790998459,
0.020025718957185745,
-0.0569012425839901,
-0.00001976824933080934,
0.01929084211587906,
-0.07102350145578384,
0.002228293800726533,
0.02000010572373867,
0.02098538540303707,
-0.032198529690504074,
0.05565633624792099,
-0.10814479738473892,
0.01699039526283741,
0.008423619903624058,
-0.08252188563346863,
0.021703798323869705,
0.03367073833942413,
0.054798733443021774,
-0.029422469437122345,
0.054330013692379,
0.03249535337090492,
-0.006149658467620611,
-0.04579685255885124,
0.0072249737568199635,
0.07242132723331451,
0.08219237625598907,
-0.07021084427833557,
-0.021885765716433525,
0.0877898633480072,
-0.0445445291697979,
-0.012452046386897564,
0.03626590967178345,
0.03091980703175068,
-0.06047448888421059,
-0.008107026107609272,
0.05428481847047806,
-0.018741469830274582,
0.0026402613148093224,
-0.11394360661506653,
0.06198786199092865,
0.052773695439100266,
0.04746774956583977,
0.08927836269140244,
0.013028922490775585,
0.03766236454248428,
0.055713143199682236,
0.07516362518072128,
0.08817390352487564,
0.026810701936483383,
-0.009548427537083626,
-0.05603085085749626,
0.07759980112314224,
-0.05082682892680168,
-0.08094779402017593,
-3.537543946663391e-8,
-0.03866958990693092,
-0.06236657872796059,
0.03273024782538414,
0.026484748348593712,
0.05869579687714577,
-0.0002624386106617749,
-0.02974829077720642,
0.02364961802959442,
0.01745021902024746,
-0.09786610305309296,
0.07122205942869186,
0.03646761178970337,
-0.037450727075338364,
-0.02012757398188114,
-0.03924844041466713,
-0.024371597915887833,
-0.08988881856203079,
-0.06073928251862526,
0.012122836895287037,
-0.05190683528780937,
-0.07412688434123993,
0.05625837668776512,
0.06560125201940536,
-0.017813395708799362,
0.07804867625236511,
-0.05381627753376961,
0.03791069984436035,
0.04084233194589615,
0.14143334329128265,
-0.05141940340399742,
-0.04944942146539688,
0.005445721093565226,
-0.034680724143981934,
-0.002006549621000886,
0.0072978767566382885,
-0.018908128142356873,
0.0018353027990087867,
0.04788556322455406,
-0.08116523176431656,
-0.08764021098613739,
-0.05083028972148895,
0.05105949193239212,
0.04111374914646149,
-0.024934353306889534,
0.06702212244272232,
0.010406430810689926,
-0.04138479381799698,
-0.002524588955566287,
0.017172085121273994,
0.009966546669602394,
-0.00881972536444664,
-0.062160126864910126,
0.03649048134684563,
0.06984605640172958,
0.05868000537157059,
-0.0004710517532657832,
-0.04191107675433159,
-0.09209661185741425,
0.03326989710330963,
-0.01950015313923359,
0.025362608954310417,
-0.08657578378915787,
-0.1329389363527298,
-0.03662532567977905
] |
Learning from negative feedback, or positive feedback or both | https://openreview.net/forum?id=4FVGowGzQb | [
"Abbas Abdolmaleki",
"Bilal Piot",
"Bobak Shahriari",
"Jost Tobias Springenberg",
"Tim Hertweck",
"Michael Bloesch",
"Rishabh Joshi",
"Thomas Lampe",
"Junhyuk Oh",
"Nicolas Heess",
"Jonas Buchli",
"Martin Riedmiller"
] | Spotlight | Existing preference optimization methods often assume scenarios where paired preference feedback (preferred/positive vs. dis-preferred/negative examples) is available. This requirement limits their applicability in scenarios where only unpaired feedback—for example, either positive or negative— is available. To address this, we introduce a novel approach that decouples learning from positive and negative feedback. This decoupling enables control over the influence of each feedback type and, importantly, allows learning even when only one feedback type is present. A key contribution is demonstrating stable learning from negative feedback alone, a capability not well-addressed by current methods. Our approach builds upon the probabilistic framework introduced in (Dayan and Hinton, 1997), which uses expectation-maximization (EM) to directly optimize the probability of positive outcomes (as opposed to classic expected reward maximization). We address a key limitation in current EM-based methods: they solely maximize the likelihood of positive examples, while neglecting negative ones. We show how to extend EM algorithms to explicitly incorporate negative examples, leading to a theoretically grounded algorithm that offers an intuitive and versatile way to learn from both positive and negative feedback. We evaluate our approach for training language models based on human feedback as well as training policies for sequential decision-making problems, where learned value functions are available. | Preference Optimization, Policy Optimization, Negative Feedback, Positive feedback, Reinforcement Learning, Probabilistic Inference | A new policy optimization algorithm that learns from different type and number of feedback (positive, negative, or both) to optimize policies. | 10,103 | 2410.04166 | [
-0.042857032269239426,
-0.0509546622633934,
0.03067089430987835,
-0.018114427104592323,
0.00505793234333396,
0.006343581713736057,
0.03971165791153908,
0.00476253405213356,
0.03314674645662308,
-0.01762000471353531,
-0.05275052785873413,
-0.002801733324304223,
0.09728001803159714,
0.06509478390216827,
-0.02004697360098362,
0.06284534186124802,
0.1926858276128769,
0.08582029491662979,
-0.06187868118286133,
0.01078562717884779,
-0.02400987595319748,
-0.05200459808111191,
0.01997135579586029,
-0.11410639435052872,
-0.002158534014597535,
-0.08786062896251678,
-0.008574460633099079,
0.01580711267888546,
0.02308131940662861,
-0.04079681262373924,
0.021081577986478806,
-0.09759018570184708,
0.05743918567895889,
-0.08126404136419296,
-0.07647207379341125,
-0.006841643247753382,
-0.10862214118242264,
-0.009432892315089703,
-0.06322728097438812,
-0.0047194818034768105,
-0.08127951622009277,
0.038689129054546356,
-0.04024577513337135,
-0.027324046939611435,
-0.018270552158355713,
-0.007352412212640047,
0.026843395084142685,
-0.025246860459446907,
0.013071153312921524,
0.002016910118982196,
0.043244753032922745,
0.027923962101340294,
-0.0669146478176117,
0.0037668184377253056,
0.029369188472628593,
0.034298814833164215,
0.02169734053313732,
0.02779928594827652,
-0.07210566103458405,
-0.04774821549654007,
-0.014878368005156517,
-0.11429207772016525,
-0.02867891639471054,
-0.07688084244728088,
-0.014954617246985435,
-0.04637802392244339,
0.015349098481237888,
0.0709315687417984,
-0.07335719466209412,
0.087955541908741,
0.012464027851819992,
0.01974955014884472,
0.01072663627564907,
0.010516311042010784,
0.04671500250697136,
0.12269003689289093,
-0.007842015475034714,
-0.024153156206011772,
-0.023451782763004303,
0.004831531550735235,
-0.003241824684664607,
0.012115970253944397,
-0.012107244692742825,
0.02089986577630043,
0.1205591931939125,
-0.03476912900805473,
-0.017647502943873405,
0.03477651998400688,
0.10681183636188507,
0.059999022632837296,
-0.10456614941358566,
0.02811501733958721,
-0.04862771928310394,
-0.03652362897992134,
-0.0037855810951441526,
0.029710274189710617,
-0.024865861982107162,
-0.0484224371612072,
-0.04829276725649834,
0.1517048329114914,
-0.020350858569145203,
0.07961522042751312,
-0.015751134604215622,
-0.07764414697885513,
0.024145949631929398,
-0.11561254411935806,
-0.05752645432949066,
0.0671854317188263,
0.06217195466160774,
-0.14761485159397125,
0.0005413609324023128,
0.005277018062770367,
-0.032484833151102066,
0.00032669611391611397,
-0.07097291946411133,
-0.010663234628736973,
0.07076580822467804,
0.07647047936916351,
-0.01687566563487053,
0.022857550531625748,
-0.03175390511751175,
-0.03964950889348984,
0.06669719517230988,
-0.015182734467089176,
0.013944650068879128,
0.03419005498290062,
0.00726752495393157,
2.6146024933660888e-33,
-0.015475182794034481,
-0.03823184221982956,
0.08615140616893768,
-0.028599703684449196,
0.07964333146810532,
0.01287011057138443,
-0.040223442018032074,
0.016402510926127434,
-0.040138911455869675,
0.006050044670701027,
0.10457585752010345,
0.010349983349442482,
-0.0027289865538477898,
0.06658114492893219,
0.05396636947989464,
-0.03133894503116608,
-0.03958417847752571,
-0.0024973819963634014,
-0.014473914168775082,
-0.08355524390935898,
0.06568853557109833,
-0.005794898606836796,
0.010851585306227207,
-0.0853789895772934,
-0.004779803100973368,
-0.06166897714138031,
0.0305061973631382,
-0.0037658680230379105,
-0.015395799651741982,
-0.02580021694302559,
-0.09135953336954117,
0.022790689021348953,
0.03687005862593651,
-0.03802228346467018,
-0.005474086385220289,
0.043696850538253784,
-0.03723111003637314,
0.0009993432322517037,
-0.03894520550966263,
-0.008220459334552288,
0.01704803667962551,
0.025917647406458855,
0.009639094583690166,
0.0033580809831619263,
-0.045912567526102066,
-0.04249675199389458,
0.03611929342150688,
-0.05345969647169113,
-0.07689926028251648,
-0.06563716381788254,
0.08565741032361984,
-0.016836214810609818,
0.009302244521677494,
0.010808792896568775,
-0.026811916381120682,
0.0018025505123659968,
-0.00857551395893097,
0.11160841584205627,
-0.00027316060732118785,
-0.05105163902044296,
-0.002502432093024254,
0.025024104863405228,
-0.00464037386700511,
-0.07118341326713562,
-0.05617973953485489,
0.10398019850254059,
-0.03175105154514313,
0.006296223495155573,
0.026285264641046524,
-0.013169055804610252,
0.009498141705989838,
0.086924247443676,
-0.03319592773914337,
-0.12078636884689331,
0.03779691830277443,
0.06154431030154228,
-0.019956905394792557,
-0.017603399232029915,
0.12128914147615433,
-0.049247417598962784,
0.057586561888456345,
-0.02056548371911049,
0.03196072578430176,
-0.043427255004644394,
0.009954866953194141,
-0.023187849670648575,
-0.002176102716475725,
-0.0821358785033226,
0.0007164868293330073,
0.0018391445046290755,
-0.09150756150484085,
0.039559949189424515,
-0.043261248618364334,
0.05325843393802643,
0.07097896933555603,
-3.270637441694361e-33,
-0.08056692034006119,
-0.003547657746821642,
-0.0077076866291463375,
-0.0076919300481677055,
0.030719244852662086,
0.028484882786870003,
-0.04784741252660751,
-0.08374785631895065,
-0.023185009136795998,
-0.04311990737915039,
-0.04095026105642319,
0.013992386870086193,
0.044670604169368744,
-0.013289771042764187,
-0.04246964305639267,
-0.022068552672863007,
-0.05129353329539299,
0.06415645778179169,
0.0384720116853714,
-0.023756304755806923,
-0.00704602338373661,
0.14055989682674408,
-0.05003321170806885,
0.08406257629394531,
0.021344901993870735,
0.02085033245384693,
0.013621923513710499,
0.06521470099687576,
-0.003802438033744693,
-0.03896908089518547,
-0.09582246840000153,
0.031009815633296967,
-0.015198977664113045,
0.05706356465816498,
-0.05159304663538933,
0.04130002111196518,
-0.02026645839214325,
0.021613508462905884,
-0.08556728065013885,
0.1093883365392685,
0.017982246354222298,
-0.022703034803271294,
0.0011809084098786116,
0.01960168033838272,
-0.002011962002143264,
-0.058143869042396545,
-0.06295713782310486,
0.03913484886288643,
0.08779557794332504,
0.00397964334115386,
-0.03610549494624138,
0.0418781042098999,
-0.09175439178943634,
0.045282743871212006,
-0.03629285842180252,
-0.0875798687338829,
0.014819946140050888,
-0.01352748367935419,
0.14242039620876312,
0.09789352118968964,
-0.06775914877653122,
0.01640230417251587,
-0.0228122565895319,
0.020474962890148163,
-0.004718329291790724,
-0.01128747034817934,
0.04994774982333183,
-0.0076464503072202206,
0.0190545953810215,
0.032209306955337524,
0.03677982836961746,
0.07669631391763687,
-0.02555057220160961,
-0.07905944436788559,
0.03825557231903076,
0.05297590419650078,
0.012208127416670322,
0.017229778692126274,
-0.0423368401825428,
-0.07178255915641785,
0.018633944913744926,
0.033598802983760834,
0.042272794991731644,
0.038320574909448624,
0.036230240017175674,
0.08067996799945831,
-0.013799814507365227,
0.0617511123418808,
0.03391506150364876,
0.005964798852801323,
-0.021530862897634506,
0.03207719326019287,
0.04725128784775734,
0.06939244270324707,
-0.007472094614058733,
-4.768449812786457e-8,
-0.024688268080353737,
-0.0019303561421111226,
0.043979667127132416,
0.05487948656082153,
0.04953595623373985,
-0.008740333840250969,
-0.01252765767276287,
0.05237358435988426,
-0.024525005370378494,
-0.07988519966602325,
0.05372185260057449,
0.07784867286682129,
-0.009895427152514458,
-0.07556948810815811,
-0.05026095733046532,
0.061493441462516785,
0.04137701541185379,
-0.009328031912446022,
0.008992563933134079,
0.007875471375882626,
-0.0009496839484199882,
-0.03600518777966499,
0.05604512244462967,
-0.015262891538441181,
0.024112243205308914,
-0.07620811462402344,
0.045576322823762894,
0.09427392482757568,
0.021609121933579445,
0.0031715775839984417,
-0.004878409672528505,
-0.01794511079788208,
-0.011235935613512993,
-0.015071163885295391,
0.028166944161057472,
0.06615156680345535,
-0.03571171313524246,
-0.04004306346178055,
0.02259032428264618,
-0.10587511956691742,
0.015769682824611664,
0.054612897336483,
-0.03184565529227257,
-0.033519446849823,
-0.01641528680920601,
0.07866168022155762,
0.003863542340695858,
-0.0739649161696434,
-0.026131125167012215,
0.01998390443623066,
0.06068453565239906,
-0.05677270516753197,
0.007781718857586384,
0.017354769632220268,
0.017506221309304237,
0.03979487344622612,
-0.024261895567178726,
-0.03693489357829094,
-0.0016521482029929757,
0.0987916961312294,
0.05564792454242706,
0.0015366277657449245,
-0.05316612869501114,
0.00701895821839571
] |
CausalRivers - Scaling up benchmarking of causal discovery for real-world time-series | https://openreview.net/forum?id=wmV4cIbgl6 | [
"Gideon Stein",
"Maha Shadaydeh",
"Jan Blunk",
"Niklas Penzel",
"Joachim Denzler"
] | Spotlight | Causal discovery, or identifying causal relationships from observational data, is a notoriously challenging task, with numerous methods proposed to tackle it.
Despite this, in-the-wild evaluation of these methods is still lacking, as works frequently rely on synthetic data evaluation and sparse real-world examples under critical theoretical assumptions.
Real-world causal structures, however, are often complex, evolving over time, non-linear, and influenced by unobserved factors, making
it hard to decide on a proper causal discovery strategy.
To bridge this gap, we introduce CausalRivers, the largest in-the-wild causal discovery benchmarking kit for time-series data to date.
CausalRivers features an extensive dataset on river discharge that covers the eastern German territory (666 measurement stations) and the state of Bavaria (494 measurement stations).
It spans the years 2019 to 2023 with a 15-minute temporal resolution.
Further, we provide additional data from a flood around the Elbe River, as an event with a pronounced distributional shift.
Leveraging multiple sources of information and time-series meta-data, we constructed two distinct causal ground truth graphs (Bavaria and eastern Germany).
These graphs can be sampled to generate thousands of subgraphs to benchmark causal discovery across diverse and challenging settings.
To demonstrate the utility of CausalRivers, we evaluate several causal discovery approaches through a set of experiments to identify areas for improvement.
CausalRivers has the potential to facilitate robust evaluations and comparisons of causal discovery methods.
Besides this primary purpose, we also expect that this dataset will be relevant for connected areas of research, such as time-series forecasting and anomaly detection.
Based on this, we hope to push benchmark-driven method development that fosters advanced techniques for causal discovery, as is the case for many other areas of machine learning. | Causal Discovery, Benchmarking, Time-series | The largest in-the-wild causal discovery benchmark kit to this date, including high-resolution ts and ground truth causal graphs with over 1000 nodes. | 10,051 | 2503.17452 | [
-0.06597191095352173,
-0.025462118908762932,
0.06843986362218857,
0.051985785365104675,
0.0903526172041893,
0.0010996406199410558,
-0.09908487647771835,
-0.029142023995518684,
-0.0055955396965146065,
-0.021446581929922104,
-0.06160404160618782,
-0.08194243907928467,
-0.06961369514465332,
0.07928889244794846,
-0.0681103840470314,
-0.04440727084875107,
0.025794286280870438,
-0.04144342988729477,
-0.012191396206617355,
-0.04072564095258713,
-0.013670886866748333,
0.019657740369439125,
-0.06901505589485168,
0.001524367369711399,
0.04373052716255188,
0.008119537495076656,
-0.014695823192596436,
-0.02995355799794197,
-0.021539172157645226,
-0.017359955236315727,
-0.018670663237571716,
0.058196261525154114,
-0.09365276992321014,
-0.10246855020523071,
0.010773156769573689,
0.009538179263472557,
0.10719918459653854,
0.014277786016464233,
-0.034187451004981995,
-0.004034645855426788,
0.07628865540027618,
-0.03368682041764259,
0.027745375409722328,
0.059772931039333344,
0.017782967537641525,
0.08262618631124496,
0.06084175780415535,
0.0449746772646904,
-0.015426555648446083,
0.025689374655485153,
-0.06423080712556839,
0.02402864769101143,
0.017124846577644348,
0.05489007756114006,
0.03865252062678337,
-0.010266687721014023,
0.026964096352458,
-0.023707443848252296,
-0.00008133579103741795,
0.006853038910776377,
-0.03478747978806496,
-0.005640092305839062,
-0.12931175529956818,
-0.013590937480330467,
0.0022730540949851274,
0.049887094646692276,
-0.045005667954683304,
0.11111921072006226,
-0.005061452277004719,
-0.011567060835659504,
0.014172401279211044,
0.05548274517059326,
-0.03967529162764549,
0.021961113438010216,
0.060873664915561676,
-0.03005218133330345,
0.022697720676660538,
0.04853542894124985,
0.05466843023896217,
-0.09181742370128632,
0.042468298226594925,
0.02510445937514305,
0.02270437590777874,
-0.00618519214913249,
0.04373698681592941,
-0.0059988293796777725,
0.08648978173732758,
0.05745420977473259,
0.0049397884868085384,
0.062446992844343185,
0.07102307677268982,
-0.032105665653944016,
0.09570788592100143,
-0.0036700251512229443,
0.02074212022125721,
0.083511121571064,
0.008723217993974686,
0.0012291088933125138,
0.09955316036939621,
0.08215329796075821,
0.04866311699151993,
0.007241554092615843,
0.01220662146806717,
0.03649909794330597,
0.062435269355773926,
-0.029721342027187347,
-0.04659370332956314,
0.05402180925011635,
-0.005320310592651367,
-0.07666532695293427,
0.004952230956405401,
-0.019842270761728287,
0.08265307545661926,
-0.05027712509036064,
0.026255367323756218,
-0.04981682449579239,
-0.004825545009225607,
-0.038798362016677856,
-0.013527757488191128,
-0.06185347959399223,
0.006242512259632349,
0.008089396171271801,
0.002307380549609661,
0.0769592672586441,
0.09543411433696747,
0.09918078035116196,
-0.021573375910520554,
-1.6079206107589015e-34,
0.058762598782777786,
-0.040216416120529175,
0.0382966585457325,
-0.02983364835381508,
-0.030628962442278862,
-0.007268137764185667,
-0.08731242269277573,
-0.0012008275371044874,
0.010832750238478184,
0.033215753734111786,
-0.04729534685611725,
-0.018803240731358528,
-0.0873718336224556,
-0.01591985672712326,
0.031750090420246124,
0.027203764766454697,
-0.06888968497514725,
0.033274710178375244,
-0.019707122817635536,
-0.006461052689701319,
0.037777576595544815,
-0.10400696098804474,
-0.01948932744562626,
-0.007555557880550623,
0.025955675169825554,
-0.006042760331183672,
0.08519352972507477,
-0.0080433189868927,
-0.051172591745853424,
0.01767241768538952,
-0.029275966808199883,
-0.014743288047611713,
0.047047264873981476,
0.03488181158900261,
0.0030003595165908337,
-0.002041531028226018,
-0.047575294971466064,
-0.07170578092336655,
-0.04040931910276413,
0.008694547228515148,
0.011626103892922401,
0.009692487306892872,
-0.09178449958562851,
0.019515642896294594,
-0.02045884169638157,
-0.10295840352773666,
0.04211656376719475,
-0.01297038234770298,
0.038013871759176254,
-0.03184555470943451,
0.019188465550541878,
-0.004659506492316723,
0.005350525490939617,
-0.08872389793395996,
-0.05742823705077171,
0.1471000760793686,
0.020055469125509262,
0.049798235297203064,
0.009139888919889927,
0.01794399879872799,
-0.018877742812037468,
0.006165222264826298,
-0.0068395971320569515,
0.007506111636757851,
0.011450078338384628,
0.09187816828489304,
0.012413701973855495,
0.08811727166175842,
0.009318209253251553,
0.010793943889439106,
-0.035725727677345276,
0.028781481087207794,
0.022763093933463097,
-0.03767836466431618,
0.08756060898303986,
-0.02612486109137535,
0.03538619726896286,
-0.043789930641651154,
-0.023505156859755516,
0.06170324608683586,
-0.08063457906246185,
-0.057515766471624374,
0.024530721828341484,
0.02840351313352585,
0.0034097705502063036,
-0.07463289797306061,
-0.022308066487312317,
-0.03602597862482071,
-0.06578775495290756,
-0.09042557328939438,
-0.04139385744929314,
0.06936682015657425,
-0.010492718778550625,
0.028132054954767227,
0.01089069340378046,
-3.6948436900408654e-34,
0.005671454593539238,
-0.022755783051252365,
0.05570739135146141,
-0.039784301072359085,
-0.003734259633347392,
-0.046540435403585434,
-0.0388970784842968,
0.031058404594659805,
0.06667265295982361,
-0.0014770899433642626,
-0.022844485938549042,
-0.03514837846159935,
0.007188454270362854,
0.004001728259027004,
0.005403562448918819,
0.03313811123371124,
0.03988596051931381,
-0.02359670400619507,
-0.06749829649925232,
-0.029409296810626984,
0.005808271002024412,
-0.017457103356719017,
-0.04592384770512581,
-0.004134196788072586,
-0.01257164217531681,
-0.006504579912871122,
0.027946162968873978,
0.04311139136552811,
-0.032840318977832794,
-0.01472298614680767,
-0.04992377758026123,
-0.023203538730740547,
0.0022758073173463345,
-0.09295769035816193,
-0.09330447763204575,
0.13150370121002197,
0.13506996631622314,
-0.003920077811926603,
-0.04428111016750336,
-0.007868780754506588,
0.05991640314459801,
0.08066929876804352,
0.010911798104643822,
-0.06764189898967743,
0.004925963934510946,
0.05719167739152908,
-0.08705273270606995,
0.13275350630283356,
0.001060187118127942,
0.045906443148851395,
-0.00912304874509573,
0.043252237141132355,
-0.0277289766818285,
0.08299732208251953,
-0.019703317433595657,
-0.07862762361764908,
0.07119029015302658,
-0.06011519581079483,
-0.04087904840707779,
0.017515992745757103,
-0.08741076290607452,
-0.018983319401741028,
-0.032890789210796356,
-0.00065011897822842,
0.009633801877498627,
-0.029549656435847282,
0.02087494358420372,
-0.07897115498781204,
-0.007101882714778185,
0.02567448653280735,
0.04099126160144806,
0.058009374886751175,
-0.09701138734817505,
-0.08180945366621017,
-0.029109574854373932,
0.0177396759390831,
-0.01491458062082529,
0.04709886014461517,
-0.06579470634460449,
0.004065851680934429,
0.06885719299316406,
0.026182299479842186,
0.08020467311143875,
0.013386738486588001,
-0.0201112013310194,
0.0925615206360817,
0.07564220577478409,
-0.010300051420927048,
0.052051398903131485,
0.06050662696361542,
-0.10837391018867493,
-0.03826318308711052,
-0.1221223920583725,
0.05627517029643059,
-0.0022248965688049793,
-4.172449408201828e-8,
-0.005355729255825281,
0.05276067554950714,
-0.04373737797141075,
0.004890243988484144,
0.01455775834619999,
-0.054220035672187805,
-0.049562569707632065,
0.10026174783706665,
-0.05151762068271637,
-0.0818188488483429,
0.06352505832910538,
-0.009837402030825615,
-0.03599667176604271,
0.006418139673769474,
0.039549291133880615,
-0.051612358540296555,
-0.02803725190460682,
-0.09763515740633011,
-0.09068070352077484,
-0.004218986723572016,
0.09453929215669632,
0.04728815704584122,
-0.02458878420293331,
0.05146816745400429,
0.09299972653388977,
-0.07434964925050735,
-0.024123620241880417,
0.04842499643564224,
0.10067295283079147,
-0.05410551652312279,
0.02030405029654503,
0.051768604665994644,
-0.02707342617213726,
-0.030867187306284904,
0.008115756325423717,
0.016570059582591057,
-0.05217505991458893,
0.02847505547106266,
-0.07347327470779419,
-0.016020039096474648,
0.016444524750113487,
0.07034460455179214,
-0.06621004641056061,
0.01582668349146843,
0.05667509883642197,
-0.034844495356082916,
-0.05547715350985527,
-0.027119267731904984,
0.05803002789616585,
-0.02719937637448311,
0.021575910970568657,
-0.0408773347735405,
0.022697046399116516,
0.07565756887197495,
0.11789218336343765,
0.07682932913303375,
0.005757580511271954,
-0.056668393313884735,
-0.0047783562913537025,
-0.025937173515558243,
0.1175118163228035,
-0.014663362875580788,
-0.06260645389556885,
0.029265932738780975
] |
Signature Kernel Conditional Independence Tests in Causal Discovery for Stochastic Processes | https://openreview.net/forum?id=Nx4PMtJ1ER | [
"Georg Manten",
"Cecilia Casolo",
"Emilio Ferrucci",
"Søren Wengel Mogensen",
"Cristopher Salvi",
"Niki Kilbertus"
] | Spotlight | Inferring the causal structure underlying stochastic dynamical systems from observational data holds great promise in domains ranging from science and health to finance. Such processes can often be accurately modeled via stochastic differential equations (SDEs), which naturally imply causal relationships via `which variables enter the differential of which other variables'. In this paper, we develop conditional independence (CI) constraints on coordinate processes over selected intervals that are Markov with respect to the acyclic dependence graph (allowing self-loops) induced by a general SDE model. We then provide a sound and complete causal discovery algorithm, capable of handling both fully and partially observed data, and uniquely recovering the underlying or induced ancestral graph by exploiting time directionality assuming a CI oracle. Finally, to make our algorithm practically usable, we also propose a flexible, consistent signature kernel-based CI test to infer these constraints from data. We extensively benchmark the CI test in isolation and as part of our causal discovery algorithms, outperforming existing approaches in SDE models and beyond. | causality, dynamical systems, stochastic processes, causal discovery, signature kernel | We develop a kernel-based conditional independence test on ‘path-space’ and constraint-based causal discovery algorithms for SDE solutions that make use of the test for robust causal discovery. | 9,971 | 2402.18477 | [
-0.11360736936330795,
-0.07307422161102295,
-0.013388300314545631,
0.034096430987119675,
0.12233765423297882,
-0.04115471988916397,
-0.00007076085603330284,
-0.029599720612168312,
0.03502514585852623,
-0.044111501425504684,
-0.0006384160951711237,
-0.05505328252911568,
0.01913462020456791,
0.011750591918826103,
0.007209331262856722,
-0.06229542940855026,
-0.023523002862930298,
-0.02019052393734455,
-0.0004752483800984919,
-0.01598724164068699,
-0.06016603112220764,
0.060366157442331314,
-0.07762215286493301,
0.01787421479821205,
0.03826471418142319,
-0.005973129533231258,
0.1096632331609726,
-0.009591869078576565,
0.007551397196948528,
0.01644936203956604,
0.041393011808395386,
0.006028828211128712,
-0.08240644633769989,
-0.044945523142814636,
0.06361610442399979,
0.016110699623823166,
0.07029037922620773,
-0.015899401158094406,
-0.032855939120054245,
-0.004441300872713327,
0.045627254992723465,
-0.015839004889130592,
0.04442606866359711,
0.048481520265340805,
0.01686498522758484,
0.03706420958042145,
0.017670052126049995,
0.028739532455801964,
-0.02350894920527935,
-0.003545773681253195,
-0.04114534333348274,
0.0011504284339025617,
-0.0011683146003633738,
-0.012296339496970177,
0.08805912733078003,
-0.0028769432101398706,
0.01715523563325405,
-0.04182149097323418,
-0.027556898072361946,
0.02061857469379902,
0.0036494566593319178,
-0.08254552632570267,
-0.03925100713968277,
-0.009318668395280838,
-0.007117031142115593,
0.10345906764268875,
-0.002508606528863311,
0.10442283749580383,
0.018636807799339294,
0.04260309413075447,
-0.04444420337677002,
0.08639039844274521,
-0.11763271689414978,
0.030143331736326218,
0.03451382368803024,
0.10708731412887573,
0.05551261082291603,
0.000025792069209273905,
0.007504709996283054,
-0.14661304652690887,
-0.04226576164364815,
0.07193980365991592,
0.03308505937457085,
-0.03321891650557518,
0.07744240015745163,
-0.03471537306904793,
0.07372686266899109,
0.03356693685054779,
0.05825851857662201,
0.0780009925365448,
0.0511472187936306,
-0.0131918303668499,
0.026288628578186035,
-0.048281121999025345,
0.0555105023086071,
0.04797283932566643,
0.03788356855511665,
-0.05727126821875572,
0.08194754272699356,
0.07501255720853806,
-0.014627687633037567,
0.019494561478495598,
0.02381632663309574,
0.07905523478984833,
0.01042409148067236,
-0.057601142674684525,
-0.023693140596151352,
-0.02733396552503109,
0.02957165241241455,
-0.08432284742593765,
-0.004680468235164881,
0.0134262191131711,
0.058100562542676926,
0.020349275320768356,
0.008533891290426254,
-0.047736991196870804,
-0.03631667047739029,
-0.036126960068941116,
0.05123298615217209,
-0.029512016102671623,
-0.0163840614259243,
-0.008879205211997032,
-0.028954196721315384,
0.012691629119217396,
0.004483855329453945,
-0.023956112563610077,
-0.03830257058143616,
1.2084314270258395e-33,
0.04372977837920189,
-0.00673358840867877,
0.11433916538953781,
-0.03720654919743538,
-0.010919311083853245,
-0.008570433594286442,
-0.0941726416349411,
-0.04454614967107773,
-0.007455287966877222,
0.025562448427081108,
-0.023635951802134514,
0.0027229811530560255,
-0.08830275386571884,
-0.02681782841682434,
0.008759352378547192,
0.08304566144943237,
-0.03224663436412811,
0.0634130910038948,
0.011660103686153889,
-0.045776043087244034,
0.07038619369268417,
-0.10525412857532501,
-0.06167175620794296,
-0.03655845299363136,
0.03818528354167938,
-0.02266848459839821,
-0.0033943569287657738,
-0.04741137847304344,
0.011196172796189785,
0.012151443399488926,
-0.07632903009653091,
-0.005334933754056692,
0.038247495889663696,
0.07682598382234573,
0.0003212339070159942,
0.005852102767676115,
-0.11761268973350525,
-0.02811230719089508,
-0.07092858850955963,
-0.010347609408199787,
-0.005657981149852276,
-0.06812289357185364,
-0.0370129719376564,
-0.02940835803747177,
-0.057059165090322495,
-0.05781637132167816,
-0.026694297790527344,
-0.04906860366463661,
0.011526591144502163,
-0.0877033993601799,
0.019205743446946144,
0.036974236369132996,
0.053153879940509796,
-0.14071780443191528,
-0.0810164287686348,
-0.001299448311328888,
-0.04550408944487572,
-0.019395019859075546,
-0.04122724384069443,
0.039600975811481476,
0.014077815227210522,
0.007513629738241434,
-0.02324562333524227,
0.016656868159770966,
-0.014601758681237698,
0.1539813131093979,
-0.060332734137773514,
-0.014550600200891495,
0.07534508407115936,
0.04424745589494705,
-0.005629385821521282,
0.01707625947892666,
-0.0272910688072443,
-0.10449473559856415,
0.04301341250538826,
-0.00845000334084034,
-0.010927120223641396,
-0.04095814377069473,
0.031450770795345306,
0.03243838623166084,
-0.07346048951148987,
-0.09003446251153946,
-0.014197171665728092,
0.05232377350330353,
0.004220934119075537,
-0.06513908505439758,
-0.056541066616773605,
-0.04605698958039284,
-0.08042965829372406,
-0.08025050163269043,
0.009477164596319199,
0.01899227499961853,
0.039140570908784866,
0.03695632889866829,
0.043312303721904755,
-4.097345353554398e-33,
0.009607486426830292,
0.04304369166493416,
0.13822782039642334,
-0.07037591189146042,
0.00967783760279417,
-0.017386509105563164,
-0.034498997032642365,
-0.007115001790225506,
0.04105279594659805,
-0.06659913808107376,
-0.040155939757823944,
-0.07933147996664047,
0.011406783014535904,
0.01566079445183277,
0.0014608220662921667,
0.04135835915803909,
-0.0371827594935894,
-0.022647954523563385,
-0.030067821964621544,
0.021469594910740852,
-0.018326183781027794,
0.08059156686067581,
0.016795311123132706,
0.006759710144251585,
-0.002267469884827733,
0.004894909914582968,
-0.015788530930876732,
0.09051783382892609,
-0.03366491198539734,
0.009241261519491673,
-0.09877154976129532,
0.033415473997592926,
0.009394383057951927,
-0.04385329782962799,
-0.028398945927619934,
0.07415807247161865,
0.019327225163578987,
-0.028300898149609566,
-0.04263715445995331,
0.035785846412181854,
0.025222856551408768,
0.06256270408630371,
-0.05538453161716461,
0.0009878199780359864,
0.04645054042339325,
0.02409268543124199,
0.03311845287680626,
0.1269778311252594,
0.05324449762701988,
-0.027656393125653267,
0.015460222028195858,
0.11108729988336563,
-0.0010823167394846678,
0.06414025276899338,
-0.08885643631219864,
-0.0012467056512832642,
0.015297629870474339,
-0.06675002723932266,
0.02202633209526539,
-0.02575059048831463,
-0.11722619831562042,
-0.04195690155029297,
-0.004567243624478579,
0.07442423701286316,
-0.00043808601913042367,
-0.09276720881462097,
0.02923249639570713,
-0.097433902323246,
0.05571094527840614,
-0.005274870432913303,
0.0838065817952156,
0.030365947633981705,
-0.10080993920564651,
-0.05132226273417473,
0.06352437287569046,
-0.007971299812197685,
0.04412279278039932,
-0.06012163311243057,
-0.049279697239398956,
-0.032895177602767944,
-0.01906425878405571,
0.03012474998831749,
0.030173737555742264,
0.016894454136490822,
-0.07007250934839249,
0.09611833840608597,
0.04597434401512146,
-0.06495200842618942,
0.07832818478345871,
0.02746284194290638,
-0.08242571353912354,
0.0025037953164428473,
-0.07686968147754669,
0.10302763432264328,
-0.08407343924045563,
-4.236926898215643e-8,
-0.012885180301964283,
0.01805252581834793,
0.00443065632134676,
0.03759000450372696,
0.09043074399232864,
-0.03806383162736893,
0.04448463395237923,
0.07455478608608246,
-0.00013515098544303328,
0.008300146088004112,
0.004809089004993439,
0.014411895535886288,
-0.0029495980124920607,
-0.05623544752597809,
-0.0604405514895916,
-0.026852650567889214,
-0.03520051762461662,
0.006375056691467762,
-0.0517268069088459,
-0.02573634311556816,
0.054593294858932495,
-0.03714926913380623,
0.007891387678682804,
0.09411071240901947,
0.06016412004828453,
-0.040166791528463364,
0.030114151537418365,
0.056329477578401566,
0.1115930899977684,
0.07355564832687378,
0.008253678679466248,
-0.011530943214893341,
-0.042725708335638046,
-0.01535793300718069,
0.0003681971866171807,
0.05418072268366814,
0.054244447499513626,
0.014373289421200752,
-0.061165034770965576,
-0.032171040773391724,
-0.018235431984066963,
0.12043250352144241,
-0.04001520574092865,
0.055847037583589554,
0.009138686582446098,
0.015036044642329216,
-0.0022395330015569925,
-0.0325658917427063,
0.02588571421802044,
-0.051334816962480545,
-0.01612110249698162,
-0.030085718259215355,
0.04920480027794838,
-0.003865449922159314,
0.052171677350997925,
0.06080888584256172,
0.019838830456137657,
0.04074818640947342,
-0.06487873196601868,
0.023008307442069054,
0.07321023941040039,
-0.0299236923456192,
0.022829554975032806,
-0.01266954094171524
] |
Provably Reliable Conformal Prediction Sets in the Presence of Data Poisoning | https://openreview.net/forum?id=ofuLWn8DFZ | [
"Yan Scholten",
"Stephan Günnemann"
] | Spotlight | Conformal prediction provides model-agnostic and distribution-free uncertainty quantification through prediction sets that are guaranteed to include the ground truth with any user-specified probability. Yet, conformal prediction is not reliable under poisoning attacks where adversaries manipulate both training and calibration data, which can significantly alter prediction sets in practice. As a solution, we propose reliable prediction sets (RPS): the first efficient method for constructing conformal prediction sets with provable reliability guarantees under poisoning. To ensure reliability under training poisoning, we introduce smoothed score functions that reliably aggregate predictions of classifiers trained on distinct partitions of the training data. To ensure reliability under calibration poisoning, we construct multiple prediction sets, each calibrated on distinct subsets of the calibration data. We then aggregate them into a majority prediction set, which includes a class only if it appears in a majority of the individual sets. Both proposed aggregations mitigate the influence of datapoints in the training and calibration data on the final prediction set. We experimentally validate our approach on image classification tasks, achieving strong reliability while maintaining utility and preserving coverage on clean data. Overall, our approach represents an important step towards more trustworthy uncertainty quantification in the presence of data poisoning. | Conformal prediction, Certifiable robustness, Adversarial robustness | We propose the first efficient method for making conformal prediction sets more reliable against data poisoning and label flipping attacks. | 9,924 | 2410.09878 | [
-0.10999629646539688,
-0.08558919280767441,
0.022117359563708305,
0.04868372529745102,
0.12557242810726166,
-0.05388963222503662,
0.02986006997525692,
-0.05306556820869446,
-0.02912190929055214,
-0.020593924447894096,
-0.06834053248167038,
-0.0784182921051979,
0.15091538429260254,
0.052212905138731,
-0.04759378731250763,
-0.03513224795460701,
0.012895451858639717,
0.08759181201457977,
-0.08956187218427658,
0.010103492997586727,
-0.059137750416994095,
0.009088296443223953,
-0.010225002653896809,
0.04445665702223778,
0.03402459993958473,
-0.01853729598224163,
0.019167035818099976,
-0.023049145936965942,
-0.04495128244161606,
-0.06469281017780304,
-0.0054887160658836365,
-0.05901508778333664,
-0.028370056301355362,
-0.04099414497613907,
-0.03782699629664421,
0.025442374870181084,
-0.02519591897726059,
0.006006005220115185,
0.061952631920576096,
0.0009318498778156936,
-0.02763275057077408,
0.046170447021722794,
0.028201080858707428,
0.07914429157972336,
-0.0032664122991263866,
0.09420496225357056,
0.03713923320174217,
-0.03449965640902519,
-0.009074680507183075,
-0.005286261439323425,
-0.010471496731042862,
-0.001823417260311544,
-0.005449265241622925,
-0.0002633139956742525,
-0.022215358912944794,
-0.03568548336625099,
0.03972027078270912,
-0.011690986342728138,
-0.03289518877863884,
-0.006830201484262943,
-0.014104321599006653,
-0.0844593495130539,
-0.0015226638643071055,
-0.002719580428674817,
0.014090965501964092,
-0.008608677424490452,
0.06375371664762497,
0.03436654433608055,
0.016801204532384872,
0.07375983148813248,
-0.047408413141965866,
0.07821347564458847,
-0.019381247460842133,
0.027293305844068527,
0.021555934101343155,
0.11899887770414352,
0.03910453990101814,
-0.0363963320851326,
0.007946527563035488,
-0.027062080800533295,
-0.04119375720620155,
0.026269035413861275,
0.06666414439678192,
-0.0014710346003994346,
0.13612964749336243,
-0.016156384721398354,
0.01560045313090086,
0.017337452620267868,
0.068084716796875,
0.03804151341319084,
-0.0356251560151577,
0.05455339699983597,
0.08484004437923431,
-0.002863074652850628,
-0.0019780003931373358,
0.050173643976449966,
-0.020536387339234352,
-0.11530065536499023,
0.01553107425570488,
0.08112926036119461,
-0.12339477986097336,
0.03361911326646805,
0.03224940598011017,
-0.07384815067052841,
0.12964653968811035,
0.0036441616248339415,
0.008535013534128666,
0.0253310464322567,
0.03670915961265564,
-0.05843653157353401,
0.05775130167603493,
-0.01889352686703205,
-0.016873246058821678,
-0.024267209693789482,
0.05734698474407196,
0.04672747850418091,
-0.030154859647154808,
0.0049858554266393185,
-0.06526069343090057,
0.04930407181382179,
-0.029899397864937782,
0.0012590473052114248,
0.05323360860347748,
-0.021976357325911522,
0.07760942727327347,
-0.03888530284166336,
-0.02278844639658928,
4.557119765069381e-33,
-0.012757871299982071,
-0.007895192131400108,
0.04483187198638916,
-0.015244774520397186,
0.04427707567811012,
-0.04435194283723831,
-0.04228246957063675,
-0.00034729167236946523,
-0.010719860903918743,
0.04109090566635132,
-0.020840561017394066,
-0.015332669951021671,
-0.04074179008603096,
0.06419732421636581,
0.10242735594511032,
0.07467738538980484,
-0.007160309236496687,
-0.0011742310598492622,
-0.12122349441051483,
-0.028475210070610046,
0.056251801550388336,
-0.0959479883313179,
0.009120271541178226,
-0.07824965566396713,
-0.005537890829145908,
-0.00014359113993123174,
0.07535606622695923,
0.08755510300397873,
-0.01694723591208458,
0.00296796765178442,
-0.05877600610256195,
0.009585482999682426,
0.060127243399620056,
0.04825975000858307,
-0.013630506582558155,
-0.053333282470703125,
-0.053467970341444016,
0.020276550203561783,
-0.02692502737045288,
0.03711741045117378,
0.0061858901754021645,
0.019327158108353615,
-0.030997375026345253,
0.026568206027150154,
0.025217728689312935,
-0.03233103081583977,
-0.06963079422712326,
0.06554168462753296,
-0.03464667499065399,
-0.07354892790317535,
0.041850656270980835,
-0.013669093139469624,
-0.03677013888955116,
-0.06156110391020775,
-0.13145560026168823,
0.02806716412305832,
0.03613731265068054,
-0.009043881669640541,
-0.006964020896703005,
-0.07452432066202164,
-0.0005430808523669839,
-0.0014492347836494446,
-0.05104626342654228,
-0.007866277359426022,
-0.04256045073270798,
0.04301220551133156,
0.007986783050000668,
-0.041056446731090546,
0.007413364481180906,
-0.010337482206523418,
0.025974547490477562,
0.04639352858066559,
-0.057307519018650055,
-0.03215545043349266,
0.024189135059714317,
0.018697094172239304,
0.08810214698314667,
0.04323885217308998,
0.03636232763528824,
0.04849390685558319,
0.02123355306684971,
-0.024901054799556732,
0.0011590125504881144,
-0.03744683787226677,
-0.03746224567294121,
-0.03189312294125557,
0.04420137032866478,
-0.023925943300127983,
-0.13250765204429626,
0.04846172407269478,
-0.12303520739078522,
-0.046319667249917984,
-0.011479043401777744,
-0.032860178500413895,
-0.0722656175494194,
-3.4147119068028946e-33,
-0.03720895200967789,
0.05440045893192291,
0.03050968050956726,
0.07456806302070618,
-0.05218205600976944,
-0.049935564398765564,
-0.05016838386654854,
0.0704021006822586,
0.003503379411995411,
-0.08121809363365173,
-0.03592468798160553,
0.013058540411293507,
0.03988289460539818,
-0.043706368654966354,
-0.036562494933605194,
-0.011652599088847637,
-0.04024706035852432,
0.03159157559275627,
-0.0927523523569107,
0.01725609041750431,
0.05920393764972687,
0.05322274565696716,
-0.0623319074511528,
0.08501636981964111,
-0.027886006981134415,
0.05983379855751991,
-0.04941905662417412,
0.047573547810316086,
-0.0022192082833498716,
0.04682030901312828,
-0.03540148586034775,
-0.03434460610151291,
0.01740443892776966,
-0.0935588926076889,
-0.012019927613437176,
-0.03519289568066597,
0.038395289331674576,
-0.0006794860819354653,
-0.06573735177516937,
0.09657732397317886,
0.05419180542230606,
0.019461005926132202,
-0.13129852712154388,
0.007389959879219532,
0.004882609471678734,
-0.09441491216421127,
0.0645095705986023,
-0.018087996169924736,
-0.005439666099846363,
0.004363537300378084,
-0.040163446217775345,
0.0185922309756279,
-0.09821469336748123,
0.08944353461265564,
0.012889329344034195,
-0.005027990322560072,
-0.0499514639377594,
-0.01676683872938156,
0.0796150490641594,
0.06382403522729874,
-0.07066984474658966,
0.017346756532788277,
-0.017223458737134933,
0.09968642890453339,
0.015699602663517,
-0.022242190316319466,
-0.01995212212204933,
0.04856061190366745,
0.06666156649589539,
0.09248635917901993,
0.01874714531004429,
0.030334264039993286,
0.012192235328257084,
-0.022570600733160973,
-0.026023030281066895,
-0.010968067683279514,
0.03543725609779358,
-0.029859526082873344,
0.03514457494020462,
0.04172574728727341,
0.05284556373953819,
-0.052226483821868896,
0.09624183923006058,
0.08088278770446777,
0.10399246215820312,
-0.014818419702351093,
0.07875905930995941,
-0.035674404352903366,
0.1210750862956047,
0.0035672199446707964,
-0.04063960537314415,
-0.015871120616793633,
-0.0300194900482893,
0.05371912941336632,
-0.021848594769835472,
-4.346081539097213e-8,
-0.03986385092139244,
0.08777856081724167,
-0.000865055073518306,
0.03036164492368698,
0.01874616928398609,
-0.0044570001773536205,
-0.03409791737794876,
0.03143206611275673,
-0.010931435972452164,
-0.026730697602033615,
0.085807666182518,
-0.016199076548218727,
-0.06450680643320084,
-0.07563844323158264,
0.018628111109137535,
0.043390676379203796,
0.01703852042555809,
0.12154754996299744,
-0.07696139067411423,
0.010400183498859406,
-0.053318530321121216,
-0.05012577772140503,
0.10240574181079865,
-0.042737435549497604,
0.03332850709557533,
-0.05018318444490433,
-0.06334827095270157,
-0.0071707009337842464,
0.01239387784153223,
0.07343680411577225,
-0.024055400863289833,
-0.050673894584178925,
0.07156261801719666,
0.013612685725092888,
-0.010846937075257301,
0.10831671953201294,
-0.05784188583493233,
-0.03277132287621498,
-0.0351400226354599,
0.04622620344161987,
0.00924011878669262,
0.040555499494075775,
-0.028267232701182365,
0.02318323776125908,
0.02651577815413475,
-0.01592828333377838,
0.03979751467704773,
-0.052314527332782745,
0.0295961145311594,
0.0035156551748514175,
0.0030612621922045946,
0.026126863434910774,
-0.055126603692770004,
0.10270160436630249,
0.023083744570612907,
-0.02169017307460308,
-0.016920190304517746,
-0.078338123857975,
-0.0057152085937559605,
0.06785093992948532,
0.014079033397138119,
-0.054356012493371964,
0.02438395284116268,
-0.036038417369127274
] |
Weak-to-Strong Preference Optimization: Stealing Reward from Weak Aligned Model | https://openreview.net/forum?id=f7KxfUrRSb | [
"Wenhong Zhu",
"Zhiwei He",
"Xiaofeng Wang",
"Pengfei Liu",
"Rui Wang"
] | Spotlight | Aligning language models (LMs) with human preferences has become a key area of research, enabling these models to meet diverse user needs better. Inspired by weak-to-strong generalization, where a strong LM fine-tuned on labels generated by a weaker model can consistently outperform its weak supervisor, we extend this idea to model alignment. In this work, we observe that the alignment behavior in weaker models can be effectively transferred to stronger models and even exhibit an amplification effect. Based on this insight, we propose a method called Weak-to-Strong Preference Optimization (WSPO), which achieves strong model alignment by learning the distribution differences before and after the alignment of the weak model. Experiments demonstrate that WSPO delivers outstanding performance, improving the win rate of Qwen2-7B-Instruct on Arena-Hard from 39.70 to 49.60, achieving a remarkable 47.04 length-controlled win rate on AlpacaEval 2, and scoring 7.33 on MT-bench. Our results suggest that using the weak model to elicit a strong model with a high alignment ability is feasible. The code is available at https://github.com/zwhong714/weak-to-strong-preference-optimization. | weak-to-strong, model alignment | We propose a method called WSPO, which effectively transfers the alignment ability from a weak model to a strong model, resulting in an amplified alignment phenomenon. | 9,884 | 2410.18640 | [
-0.0627899318933487,
-0.09099368751049042,
-0.00783071480691433,
-0.019373225048184395,
-0.022481046617031097,
0.03977295383810997,
0.00664013484492898,
0.010300994850695133,
0.01409847941249609,
0.022688547149300575,
-0.019654667004942894,
-0.08241415023803711,
0.05756068602204323,
0.06834634393453598,
0.03215869143605232,
0.00781029649078846,
0.13319315016269684,
0.04672519490122795,
-0.12949210405349731,
0.0015060455771163106,
-0.04420941695570946,
-0.009747173637151718,
-0.032403189688920975,
0.008623123168945312,
0.007915303111076355,
0.046231575310230255,
-0.04385652765631676,
0.04671228677034378,
0.041961513459682465,
-0.07278626412153244,
0.017430396750569344,
-0.05499839037656784,
0.029509272426366806,
-0.01293821819126606,
-0.07379386574029922,
0.06128969043493271,
-0.17784132063388824,
-0.07355847209692001,
0.010353957302868366,
-0.06603977084159851,
-0.05233866348862648,
0.038475580513477325,
0.025900796055793762,
-0.05362346023321152,
0.018691308796405792,
0.025609716773033142,
0.022363079711794853,
0.016749637201428413,
0.01864146627485752,
0.019569681957364082,
-0.09135986119508743,
-0.044713493436574936,
0.03783572092652321,
0.03350925073027611,
-0.02599925920367241,
0.061759255826473236,
0.011630957014858723,
0.03473621979355812,
0.035228531807661057,
-0.06635948270559311,
-0.05466204881668091,
-0.01933114416897297,
-0.01257159374654293,
0.047392021864652634,
-0.000037461286410689354,
-0.07941936701536179,
0.014795136637985706,
0.05660519376397133,
-0.04791568964719772,
0.06984582543373108,
-0.00148291967343539,
0.08417725563049316,
0.0030913804657757282,
0.0037520250771194696,
0.07664033025503159,
0.09269250929355621,
0.02329423278570175,
-0.024414118379354477,
0.024256713688373566,
-0.030773887410759926,
-0.013420464470982552,
0.012475584633648396,
0.03227923810482025,
-0.01962742954492569,
0.09004581719636917,
-0.018046416342258453,
-0.010578435845673084,
0.03046330064535141,
0.06425237655639648,
-0.06432686001062393,
-0.03731124475598335,
-0.005579110234975815,
0.04753325507044792,
0.018345924094319344,
-0.0038136704824864864,
0.11993536353111267,
0.03856854513287544,
-0.03648672625422478,
-0.07416412234306335,
0.12120233476161957,
-0.006545953918248415,
0.11763471364974976,
0.000436133035691455,
-0.07287629693746567,
0.05736706033349037,
-0.10179455578327179,
-0.010442991741001606,
0.03436775133013725,
0.007255530916154385,
-0.08405321091413498,
0.03692041337490082,
0.05563419312238693,
-0.0005393428727984428,
0.08382951468229294,
-0.03995772823691368,
0.02021906152367592,
0.020506104454398155,
-0.026605036109685898,
-0.049208492040634155,
0.01897505111992359,
-0.03502122312784195,
-0.049485109746456146,
-0.02493557520210743,
0.014317845925688744,
0.007124696858227253,
0.06742303818464279,
-0.011217717081308365,
2.238764072300579e-33,
0.04272812232375145,
0.05358530208468437,
0.06761015951633453,
0.00842010322958231,
0.09795830398797989,
-0.05068841576576233,
0.019582973793148994,
0.03365188091993332,
-0.0755205750465393,
-0.0006314084166660905,
-0.04227154701948166,
0.04666578397154808,
-0.01811041310429573,
0.04928618669509888,
0.05549570545554161,
-0.0397273413836956,
-0.043223556131124496,
0.03227188438177109,
-0.06416334211826324,
-0.017377182841300964,
0.07325673848390579,
0.024070780724287033,
0.014143451116979122,
-0.05599338933825493,
0.025357011705636978,
-0.0003851211804430932,
0.0432434044778347,
-0.015183103270828724,
-0.09389614313840866,
0.007002213504165411,
-0.11139648407697678,
-0.09827416390180588,
0.019600186496973038,
0.0024206151720136404,
0.023129671812057495,
-0.056712787598371506,
-0.0717528685927391,
-0.06699883192777634,
-0.0042993417009711266,
-0.02820017747581005,
-0.010783117264509201,
0.043671123683452606,
0.01917150802910328,
-0.025509189814329147,
-0.03790457174181938,
0.0018028505146503448,
0.03521019220352173,
-0.01437671110033989,
0.042726531624794006,
-0.07547299563884735,
0.06717486679553986,
-0.0016897792229428887,
-0.028128471225500107,
0.008586287498474121,
-0.10281237214803696,
0.0022854404523968697,
0.04082195833325386,
0.07913179695606232,
0.019175831228494644,
0.06686628609895706,
-0.05042669549584389,
0.03948827460408211,
0.013821976259350777,
0.05193847045302391,
0.02852236106991768,
-0.030897190794348717,
0.023886676877737045,
0.0027622119523584843,
0.044520776718854904,
0.0011381844524294138,
0.027003269642591476,
-0.05441537871956825,
-0.06365539878606796,
-0.031717170029878616,
0.06331559270620346,
0.029388774186372757,
0.09490401297807693,
-0.10055185109376907,
0.07719969749450684,
-0.05689144507050514,
-0.00414822855964303,
0.036184392869472504,
0.010907390154898167,
-0.03788250312209129,
-0.07382281124591827,
-0.08917313069105148,
0.05240719020366669,
-0.09861861914396286,
0.010110127739608288,
0.013263575732707977,
-0.03709493204951286,
0.041187550872564316,
-0.01556312758475542,
-0.00892258994281292,
0.037093792110681534,
-1.6433238373762802e-33,
-0.0200969148427248,
-0.03304252028465271,
0.08223595470190048,
0.08019260317087173,
-0.03111911378800869,
-0.06339128315448761,
-0.001037965645082295,
0.015784749761223793,
-0.07924475520849228,
-0.0047684842720627785,
0.007437379565089941,
-0.005075662396848202,
0.042325444519519806,
-0.015516531653702259,
0.018174946308135986,
-0.02177431620657444,
0.03603716194629669,
0.04586423933506012,
0.03215834125876427,
0.03161783888936043,
0.05018513649702072,
0.1181095615029335,
-0.12247884273529053,
0.06515012681484222,
-0.008208458311855793,
0.04231984168291092,
-0.036007534712553024,
0.015561907552182674,
0.03368444740772247,
0.007900849916040897,
-0.0967186689376831,
-0.0013861358165740967,
-0.0931253656744957,
-0.05200733616948128,
-0.06430953741073608,
0.06329424679279327,
-0.05775517225265503,
0.028960904106497765,
-0.000795857748016715,
0.13943558931350708,
-0.02669983170926571,
-0.016814397647976875,
-0.11045008152723312,
-0.010048866271972656,
0.013020104728639126,
0.019401442259550095,
-0.07714802771806717,
-0.06479494273662567,
-0.01864432357251644,
-0.03636278957128525,
-0.019648782908916473,
0.05404861271381378,
-0.05105576291680336,
0.047218967229127884,
-0.06257312744855881,
-0.07861017435789108,
-0.04758458212018013,
-0.013232814148068428,
0.015136933885514736,
0.026264162734150887,
-0.02336510829627514,
-0.007011956535279751,
0.01257559098303318,
0.022752946242690086,
0.026282884180545807,
0.007864398881793022,
0.038111601024866104,
-0.013033087365329266,
-0.028806492686271667,
-0.04042520746588707,
0.051130522042512894,
-0.06314553320407867,
0.03833502158522606,
0.05156315118074417,
-0.025782644748687744,
-0.010644912719726562,
-0.014008441008627415,
-0.009499817155301571,
0.03342042863368988,
-0.05156278237700462,
0.06174534931778908,
0.04254339262843132,
0.02618326060473919,
0.0021748635917901993,
0.0030562556348741055,
0.04704704135656357,
0.054737817496061325,
0.10227212309837341,
-0.0045921532437205315,
-0.01785867102444172,
-0.016866382211446762,
0.004578304942697287,
0.038366351276636124,
0.07437308132648468,
-0.045382060110569,
-4.99791319441556e-8,
-0.11693193763494492,
0.02652686834335327,
0.034588705748319626,
0.1079348549246788,
-0.07144543528556824,
-0.07128188759088516,
-0.017490608617663383,
0.028997423127293587,
0.02506427839398384,
0.018141141161322594,
0.024197064340114594,
0.002349143149331212,
-0.06722234189510345,
-0.018796579912304878,
-0.03883153572678566,
0.11359765380620956,
-0.030566496774554253,
0.06318175792694092,
-0.015529428608715534,
0.005907075945287943,
-0.016758929938077927,
0.03997031971812248,
0.058431971818208694,
-0.018891770392656326,
0.047514427453279495,
-0.07844572514295578,
-0.036317378282547,
0.08370896428823471,
0.06914819031953812,
0.03403095528483391,
-0.029420888051390648,
-0.04339532554149628,
0.0035401564091444016,
-0.008751175366342068,
0.025067592039704323,
0.10198230296373367,
-0.09669242799282074,
-0.041471026837825775,
-0.011233197525143623,
0.015932941809296608,
0.020239675417542458,
0.11640316247940063,
-0.07141372561454773,
0.00907287560403347,
-0.0097195478156209,
0.026566598564386368,
0.01100254151970148,
-0.07961036264896393,
0.05659855157136917,
-0.11707790195941925,
0.04312334209680557,
-0.06135270372033119,
0.01587345078587532,
-0.02334732562303543,
-0.005246602930128574,
0.012471825815737247,
-0.008940797299146652,
-0.01530472468584776,
0.07153641432523727,
0.0033111730590462685,
0.07902181893587112,
-0.0434899777173996,
-0.04482042416930199,
0.05884947255253792
] |
Benchmarking Predictive Coding Networks -- Made Simple | https://openreview.net/forum?id=sahQq2sH5x | [
"Luca Pinchetti",
"Chang Qi",
"Oleh Lokshyn",
"Cornelius Emde",
"Amine M'Charrak",
"Mufeng Tang",
"Simon Frieder",
"Bayar Menzat",
"Gaspard Oliviers",
"Rafal Bogacz",
"Thomas Lukasiewicz",
"Tommaso Salvatori"
] | Spotlight | In this work, we tackle the problems of efficiency and scalability for predictive coding networks (PCNs) in machine learning. To do so, we propose a library that focuses on performance and simplicity, and use it to implement a large set of standard benchmarks for the community to use for their experiments. As most works in the field propose their own tasks and architectures, do not compare one against each other, and focus on small-scale tasks, a simple and fast open-source library, and a comprehensive set of benchmarks, would address all of these concerns. Then, we perform extensive tests on such benchmarks using both existing algorithms for PCNs, as well as adaptations of other methods popular in the bio-plausible deep learning community. All of this has allowed us to (i) test architectures much larger than commonly used in the literature, on more complex datasets; (ii) reach new state-of-the-art results in all of the tasks and dataset provided; (iii) clearly highlight what the current limitations of PCNs are, allowing us to state important future research directions. With the hope of galvanizing community efforts towards one of the main open problems in the field, scalability, we will release the code, tests, and benchmarks. | cognitive science, predictive coding, computational neuroscience | null | 9,795 | 2407.01163 | [
-0.09407634288072586,
-0.12921583652496338,
0.03014698252081871,
0.06324220448732376,
0.06654072552919388,
-0.006676611490547657,
-0.09069152921438217,
-0.08267122507095337,
-0.06873905658721924,
-0.06443627178668976,
-0.0778588131070137,
-0.07650887966156006,
-0.09378449618816376,
0.027074314653873444,
-0.061613138765096664,
0.014326250180602074,
0.056392889469861984,
0.05019709840416908,
-0.08060923963785172,
-0.006504962686449289,
-0.009178781881928444,
0.0011312972055748105,
0.018468685448169708,
0.03265148773789406,
0.08926248550415039,
0.016382798552513123,
-0.008244005031883717,
-0.00530474865809083,
-0.0298957247287035,
0.020264919847249985,
0.11160590499639511,
-0.03939590975642204,
-0.005346132908016443,
0.051233552396297455,
0.011777113191783428,
0.022427136078476906,
-0.07782207429409027,
-0.0299625713378191,
-0.0018620436312630773,
-0.02944597415626049,
0.019477542489767075,
-0.013302916660904884,
0.03638201951980591,
0.07417289912700653,
0.050579749047756195,
0.06830088049173355,
0.028936732560396194,
-0.027563178911805153,
-0.022140581160783768,
-0.08965357393026352,
-0.04719411954283714,
-0.03550240397453308,
0.013093373738229275,
0.023498089984059334,
-0.050782330334186554,
-0.005926485639065504,
0.004236176609992981,
-0.013211758807301521,
0.0011832022573798895,
0.0024563597980886698,
0.02549525536596775,
-0.08934513479471207,
0.011970029212534428,
0.020861929282546043,
-0.04387901350855827,
0.0066790711134672165,
0.053178660571575165,
0.04770496115088463,
-0.0007265578606165946,
-0.04064677655696869,
0.010089767165482044,
0.10717351734638214,
-0.07163795828819275,
0.07172244042158127,
0.07247662544250488,
0.07287654280662537,
0.07806530594825745,
-0.003814862109720707,
0.05993543565273285,
-0.08072672039270401,
0.022532742470502853,
-0.0022507545072585344,
0.015141269192099571,
0.02652958407998085,
0.0996548980474472,
-0.01876571588218212,
0.014114823192358017,
0.06981842964887619,
-0.0753428190946579,
0.007742166984826326,
-0.04271258786320686,
-0.03928690031170845,
0.07274920493364334,
-0.06767700612545013,
0.024851469323039055,
0.01358549389988184,
0.034254834055900574,
-0.0757698193192482,
-0.1008412092924118,
0.09202536940574646,
-0.09624908864498138,
-0.006030538119375706,
0.04971451312303543,
-0.03944750502705574,
0.04452010616660118,
0.033508144319057465,
0.08461052179336548,
0.010237161070108414,
0.0679924264550209,
-0.0566217415034771,
0.011101333424448967,
0.06374672055244446,
0.015242156572639942,
-0.030977431684732437,
0.06373210996389389,
-0.029782798141241074,
-0.04242882877588272,
-0.00613160477951169,
0.04923087731003761,
0.05962187424302101,
-0.07696240395307541,
-0.03244272619485855,
-0.06430268287658691,
-0.0013911558780819178,
0.028486305847764015,
0.04646032303571701,
-0.07602252811193466,
4.338737893941555e-33,
0.03768452629446983,
0.02263611927628517,
0.004434932488948107,
-0.016461148858070374,
0.05859414488077164,
-0.04231619834899902,
-0.003964390140026808,
-0.013948054984211922,
-0.04662209004163742,
0.0028086805250495672,
-0.06497496366500854,
0.004829287528991699,
-0.037404485046863556,
0.11643191426992416,
0.0027710343711078167,
0.018951652571558952,
-0.03713035583496094,
-0.013263655826449394,
-0.019163817167282104,
-0.003022652119398117,
0.018499506637454033,
-0.038011547178030014,
-0.022211706265807152,
-0.0016206859145313501,
-0.03178670257329941,
-0.052076492458581924,
0.03137116879224777,
-0.08299744129180908,
0.01677642948925495,
-0.010147721506655216,
-0.11461251974105835,
-0.017532626166939735,
0.028759952634572983,
0.07362004369497299,
0.04659834876656532,
-0.040298085659742355,
0.020660657435655594,
-0.02990070730447769,
-0.018192032352089882,
0.06192668899893761,
-0.013207578100264072,
0.02664707787334919,
-0.017927763983607292,
-0.059410035610198975,
0.017255324870347977,
-0.07619807124137878,
-0.05174814909696579,
0.005028482526540756,
-0.07426479458808899,
-0.02443540282547474,
0.1359899789094925,
-0.013099170289933681,
-0.04197775572538376,
-0.009618591517210007,
0.026421532034873962,
-0.0295486468821764,
0.04396452382206917,
0.00561353238299489,
0.09542851895093918,
0.12738899886608124,
0.01954810880124569,
0.0111438212916255,
-0.03607146441936493,
0.001158153754658997,
-0.00018391066987533122,
0.029665684327483177,
0.005812039133161306,
0.0652288943529129,
0.03933950141072273,
-0.04227787256240845,
0.019123980775475502,
0.017242442816495895,
0.015936722978949547,
-0.06747552007436752,
0.06836243718862534,
0.08652034401893616,
0.025728972628712654,
-0.1004885584115982,
-0.06281712651252747,
0.08011551201343536,
-0.01777939684689045,
0.0559266172349453,
-0.022494902834296227,
0.04096120968461037,
-0.039963558316230774,
-0.03759274631738663,
0.03299762308597565,
-0.06705227494239807,
-0.0590677410364151,
0.05941411480307579,
-0.07578510791063309,
-0.03678862005472183,
0.025039561092853546,
-0.051817670464515686,
0.02315347269177437,
-3.358699968328197e-33,
-0.0180895384401083,
0.04072662815451622,
-0.04054943844676018,
0.07662077993154526,
-0.03816952928900719,
-0.07045876979827881,
-0.03406283259391785,
-0.01398697029799223,
-0.020718097686767578,
-0.03503841534256935,
0.01795562356710434,
-0.1059001013636589,
0.07945125550031662,
0.004029892850667238,
0.07734867185354233,
-0.01054317131638527,
-0.09171973168849945,
0.009125600568950176,
-0.021059691905975342,
0.031146839261054993,
0.012888428755104542,
0.13503555953502655,
-0.046345796436071396,
0.024212684482336044,
-0.0490557923913002,
-0.0054962667636573315,
-0.034322962164878845,
0.05329829826951027,
-0.014490519650280476,
-0.12903061509132385,
-0.06940032541751862,
0.004183912184089422,
-0.07585250586271286,
-0.038842763751745224,
0.04756709188222885,
0.1114678606390953,
0.04715472087264061,
-0.030481746420264244,
-0.02546754665672779,
0.04844895005226135,
0.09663599729537964,
0.0026462979149073362,
-0.07010836899280548,
-0.0015545974019914865,
0.0006990924593992531,
0.003029139479622245,
-0.07344919443130493,
0.011194807477295399,
0.050719574093818665,
0.06785927712917328,
-0.013021113350987434,
0.042537279427051544,
-0.07680563628673553,
0.03991629555821419,
0.017300263047218323,
-0.043943628668785095,
-0.00423757079988718,
0.0021670209243893623,
0.07395622879266739,
0.04630104824900627,
-0.07804988324642181,
-0.09481724351644516,
0.0025455025024712086,
0.04056017845869064,
0.014891035854816437,
-0.05768318474292755,
0.0005315374583005905,
0.04806804656982422,
0.04701923951506615,
0.016949035227298737,
0.022898010909557343,
-0.023336604237556458,
0.026535188779234886,
0.03811074048280716,
-0.04736081883311272,
-0.00803182739764452,
0.03991347923874855,
-0.026641778647899628,
0.006811006926000118,
0.007997267879545689,
0.03646580129861832,
-0.014865908771753311,
-0.01359559502452612,
-0.034827977418899536,
0.13172197341918945,
0.08978468924760818,
0.07188867032527924,
0.03021152876317501,
0.03596467524766922,
-0.003448738483712077,
-0.021001365035772324,
0.05781625583767891,
0.0014669089578092098,
0.017269793897867203,
-0.006671291310340166,
-4.669090714060076e-8,
-0.05265875905752182,
0.027637936174869537,
-0.034655265510082245,
-0.03351516276597977,
0.03719056770205498,
-0.00018468299822416157,
0.024855291470885277,
0.09270640462636948,
-0.007442818488925695,
0.08543837815523148,
0.1607625037431717,
-0.006527288351207972,
-0.05227706953883171,
-0.027713803574442863,
0.01843114197254181,
0.0778280571103096,
-0.007849783636629581,
-0.0039077382534742355,
0.017642993479967117,
-0.0215066559612751,
0.0020018701907247305,
0.07641233503818512,
-0.021534159779548645,
-0.019129153341054916,
-0.03739359229803085,
-0.13598649203777313,
0.005570299457758665,
0.03728863596916199,
-0.009137862361967564,
0.016730602830648422,
-0.025168832391500473,
0.027780916541814804,
0.04947582259774208,
-0.09681633859872818,
0.025273501873016357,
0.05533083528280258,
-0.02391265518963337,
0.055599141865968704,
0.044963039457798004,
-0.013252018950879574,
-0.0067103104665875435,
0.015022209845483303,
-0.06183243170380592,
-0.04594966396689415,
-0.012560959905385971,
-0.06373023986816406,
0.02200673148036003,
0.050788793712854385,
-0.02296627126634121,
-0.07413426786661148,
0.01403034571558237,
-0.01444817241281271,
-0.04143166169524193,
0.06859951466321945,
0.07215195894241333,
0.014567169360816479,
-0.03274955600500107,
-0.038662780076265335,
-0.0006312891491688788,
0.09646101295948029,
0.0011864708503708243,
0.04755135625600815,
-0.07372966408729553,
0.0006622338551096618
] |
LiFT: Learning to Fine-Tune via Bayesian Parameter Efficient Meta Fine-Tuning | https://openreview.net/forum?id=7nyJBVCTGQ | [
"Minyoung Kim",
"Timothy Hospedales"
] | Spotlight | We tackle the problem of parameter-efficient fine-tuning (PEFT) of a pre-trained large deep model on many different but related tasks. Instead of the simple but strong baseline strategy of task-wise independent fine-tuning, we aim to meta-learn the core shared information that can be used for unseen test tasks to improve the prediction performance further. That is, we propose a method for {\em learning-to-fine-tune} (LiFT). LiFT introduces a novel hierarchical Bayesian model that can be superior to both existing general meta learning algorithms like MAML and recent LoRA zoo mixing approaches such as LoRA-Retriever and model-based clustering. In our Bayesian model, the parameters of the task-specific LoRA modules are regarded as random variables where these task-wise LoRA modules are governed/regularized by higher-level latent random variables, which represents the prior of the LoRA modules that capture the shared information across all training tasks. To make the posterior inference feasible, we propose a novel SGLD-Gibbs sampling algorithm that is computationally efficient. To represent the posterior samples from the SGLD-Gibbs, we propose an online EM algorithm that maintains a Gaussian mixture representation for the posterior in an online manner in the course of iterative posterior sampling. We demonstrate the effectiveness of LiFT on NLP and vision multi-task meta learning benchmarks. | Bayesian methods, Parameter efficient fine-tuning, meta learning | We propose a novel hierarchical Bayesian model for meta PEFT that can be superior to both existing general meta learning algorithms like MAML and recent LoRA zoo mixing approaches such as Retrievers and model-based clustering. | 9,775 | null | [
-0.06888409703969955,
-0.12950707972049713,
0.07445162534713745,
0.0378006249666214,
0.0888778418302536,
-0.02312607504427433,
0.03275201469659805,
-0.05430670455098152,
-0.03806984797120094,
-0.044388365000486374,
-0.05048283562064171,
-0.0097722876816988,
-0.01241142489016056,
-0.0006125963991507888,
0.015798835083842278,
-0.0002636100398376584,
0.10122881084680557,
0.09844719618558884,
-0.06360600143671036,
-0.03213569521903992,
0.03592660650610924,
0.048115380108356476,
0.09861492365598679,
-0.011958828195929527,
-0.032745543867349625,
-0.00042865704745054245,
0.03632790967822075,
-0.008513694629073143,
-0.010483266785740852,
-0.038660064339637756,
0.005458309780806303,
0.030819371342658997,
-0.015172597020864487,
-0.04926864802837372,
-0.019174350425601006,
0.06930077075958252,
-0.05378945916891098,
0.052667614072561264,
-0.04298021271824837,
0.02203996852040291,
0.060972802340984344,
0.01711304299533367,
0.051906779408454895,
0.01829199679195881,
0.011935526505112648,
-0.010472203604876995,
-0.002761252922937274,
-0.0430106520652771,
0.042895060032606125,
0.031052706763148308,
-0.06454944610595703,
-0.019877340644598007,
-0.07942358404397964,
0.05071956291794777,
-0.0039036779198795557,
0.05554456263780594,
0.015684805810451508,
-0.14832574129104614,
0.02463568188250065,
-0.02041430026292801,
-0.04188085347414017,
-0.061134230345487595,
0.008258065208792686,
-0.0686996802687645,
-0.0047117373906075954,
-0.006277757231146097,
0.026833074167370796,
0.052470650523900986,
0.0017894472694024444,
-0.05001448094844818,
-0.012280001305043697,
0.08565173298120499,
-0.042397014796733856,
0.0658174604177475,
-0.008563529700040817,
0.05617998167872429,
0.011485532857477665,
-0.00035760452738031745,
0.05945767089724541,
-0.017252521589398384,
-0.026750992983579636,
0.0165947824716568,
0.02238897606730461,
-0.016662927344441414,
0.11531751602888107,
0.00902929250150919,
0.014045570977032185,
0.03468053787946701,
0.02886134199798107,
-0.014322924427688122,
0.010218765586614609,
-0.028110817074775696,
-0.04763801768422127,
-0.01905975677073002,
-0.028638271614909172,
0.02922804094851017,
0.027238313108682632,
-0.12250165641307831,
0.015648866072297096,
0.14426735043525696,
0.0038885909598320723,
0.0682419091463089,
0.0720512792468071,
0.010744821280241013,
-0.06487182527780533,
0.001311784377321601,
0.012231552973389626,
0.10810624063014984,
0.008702199906110764,
-0.0525747649371624,
-0.0024133925326168537,
0.05010366812348366,
0.06903083622455597,
-0.007833238691091537,
-0.020815076306462288,
0.026156364008784294,
0.0802888497710228,
-0.037964291870594025,
-0.055918291211128235,
0.0571046844124794,
-0.032548896968364716,
-0.03264915570616722,
-0.0025854548439383507,
0.0070220367051661015,
0.027056433260440826,
-0.010981443338096142,
-0.1258251816034317,
1.0055252145330774e-33,
0.12231898307800293,
-0.04183333367109299,
0.09309735894203186,
0.02436627633869648,
-0.011074203997850418,
0.04083138331770897,
-0.013559237122535706,
0.033547427505254745,
-0.06708665937185287,
0.03859667107462883,
-0.048574890941381454,
0.022936031222343445,
-0.05693991482257843,
0.017017263919115067,
0.021333426237106323,
0.035142313688993454,
-0.04359684884548187,
0.04957317188382149,
-0.006552246864885092,
-0.09127157926559448,
0.06525643914937973,
0.011875930242240429,
0.0003379334229975939,
-0.036840829998254776,
-0.0004214290529489517,
0.034031108021736145,
0.09423070400953293,
-0.07337334007024765,
-0.025905419141054153,
0.031814999878406525,
-0.09230876713991165,
-0.05436324700713158,
0.016616875305771828,
0.06031723693013191,
-0.01972055248916149,
-0.00898747332394123,
-0.01770108938217163,
-0.05405917391180992,
0.017082197591662407,
-0.0438329316675663,
0.011946247890591621,
-0.011055747978389263,
0.07852814346551895,
-0.07612467557191849,
-0.09592487663030624,
-0.09770093113183975,
-0.01690962351858616,
-0.026807721704244614,
-0.0033842376433312893,
-0.015934903174638748,
0.061832208186388016,
-0.09802845865488052,
0.004556270781904459,
-0.045527927577495575,
-0.10664183646440506,
0.08004961162805557,
0.04237215593457222,
0.03629246726632118,
0.03287086263298988,
0.03673641383647919,
0.01714853197336197,
-0.002120569348335266,
-0.05688684061169624,
0.0871635302901268,
0.018981538712978363,
0.03939424827694893,
0.0039864396676421165,
0.1051255539059639,
0.06962800770998001,
-0.022272201254963875,
-0.00408948166295886,
0.020977670326828957,
-0.013565890491008759,
-0.03364958241581917,
0.044079504907131195,
0.017301958054304123,
0.02706422284245491,
-0.0817825049161911,
-0.003727139439433813,
0.06197197362780571,
0.05994509533047676,
0.048557013273239136,
-0.06513023376464844,
0.019792577251791954,
0.005326033569872379,
0.01672464609146118,
-0.019373685121536255,
-0.06275264918804169,
-0.022853264585137367,
0.004078398458659649,
-0.05175163224339485,
-0.017894446849822998,
-0.011951159685850143,
-0.05519777163863182,
-0.05566312000155449,
1.0364021450513172e-34,
-0.021937834098935127,
-0.022334350273013115,
-0.015196457505226135,
0.08182331174612045,
0.04895855113863945,
-0.09744922071695328,
0.00798312108963728,
0.007519757375121117,
-0.06412366032600403,
-0.026149261742830276,
-0.05983920395374298,
-0.0687282606959343,
0.0008179456926882267,
-0.06515221297740936,
0.04749155417084694,
0.007135160733014345,
-0.03348103538155556,
0.006224943324923515,
-0.043846357613801956,
0.01465549785643816,
0.0010239052353426814,
0.01995064876973629,
0.011737232096493244,
0.07438699901103973,
-0.04298175498843193,
-0.015772292390465736,
-0.005659193731844425,
0.11077956855297089,
-0.009722255170345306,
0.04198930785059929,
-0.06050726771354675,
0.0035444111563265324,
-0.014847943559288979,
-0.008540291339159012,
-0.09895620495080948,
0.03431513160467148,
0.05699719488620758,
0.0013493119040504098,
-0.025608820840716362,
0.009160678833723068,
0.04647595062851906,
-0.0048212930560112,
-0.093804731965065,
0.00385722192004323,
-0.0036132498644292355,
-0.019742054864764214,
-0.013805089518427849,
0.002664535539224744,
0.030536454170942307,
-0.009534702636301517,
-0.08048567920923233,
-0.022520804777741432,
-0.07607981562614441,
-0.003171877469867468,
-0.05162736028432846,
-0.05313332751393318,
-0.027891041710972786,
-0.04059788957238197,
0.04822424054145813,
0.034366536885499954,
-0.007700849324464798,
-0.03899924084544182,
-0.02728269062936306,
-0.044541023671627045,
0.010982396081089973,
-0.045820388942956924,
0.01096396241337061,
-0.03205849975347519,
0.014297500252723694,
0.029835492372512817,
-0.05755121260881424,
-0.012337079271674156,
0.05433163046836853,
-0.02128736488521099,
0.04190405085682869,
0.019022813066840172,
0.01570955291390419,
-0.008938406594097614,
0.005108347162604332,
-0.05312212184071541,
0.04284607991576195,
-0.07755279541015625,
-0.02223017066717148,
0.03183012083172798,
0.05557439848780632,
0.0640421062707901,
0.07389256358146667,
0.11688864231109619,
0.049848418682813644,
-0.031196333467960358,
-0.07522384077310562,
0.041738398373126984,
0.010278205387294292,
0.015275816433131695,
-0.12292758375406265,
-3.841427798079167e-8,
-0.06625251471996307,
0.07958543300628662,
0.0075080497190356255,
0.05767263099551201,
0.016041748225688934,
0.020980816334486008,
0.01807090826332569,
0.15198104083538055,
0.007075953297317028,
-0.03568157181143761,
0.13473092019557953,
-0.035939816385507584,
-0.021926339715719223,
-0.03074786253273487,
0.06717468798160553,
0.10437977313995361,
0.0482955239713192,
0.10172483325004578,
-0.03780844807624817,
-0.010953651741147041,
0.034144386649131775,
0.06669145822525024,
0.11494585871696472,
0.06421004980802536,
0.07234261184930801,
-0.1116529181599617,
-0.07746832817792892,
0.07258876413106918,
0.10071104764938354,
-0.002024676650762558,
-0.052357882261276245,
-0.027848443016409874,
-0.015631243586540222,
0.014158564619719982,
0.0023033570032566786,
0.09744618833065033,
-0.005434948485344648,
-0.011691689491271973,
0.010542426258325577,
-0.03993234783411026,
0.02914590761065483,
0.11232659220695496,
-0.07327180355787277,
0.013133223168551922,
0.009155599400401115,
0.03463977575302124,
-0.025575267150998116,
-0.015561213716864586,
0.05209822207689285,
-0.023982873186469078,
0.016350004822015762,
-0.04564337804913521,
-0.014736803248524666,
0.05538352206349373,
-0.01850602962076664,
0.0865980014204979,
0.011478973552584648,
-0.07108474522829056,
0.06213167682290077,
0.012386164627969265,
0.05843454971909523,
-0.020141402259469032,
-0.1014157086610794,
0.027142561972141266
] |
Lightweight Neural App Control | https://openreview.net/forum?id=BL4WBIfyrz | [
"Filippos Christianos",
"Georgios Papoudakis",
"Thomas Coste",
"Jianye HAO",
"Jun Wang",
"Kun Shao"
] | Spotlight | This paper introduces a novel mobile phone control architecture, Lightweight Multi-modal App Control (LiMAC), for efficient interactions and control across various Android apps. LiMAC takes as input a textual goal and a sequence of past mobile observations, such as screenshots and corresponding UI trees, to generate precise actions. To address the computational constraints inherent to smartphones, we introduce a small Action Transformer (AcT) integrated with a fine-tuned vision-language model (VLM) for real-time decision-making and task execution. We evaluate LiMAC on two open-source mobile control datasets, demonstrating the superior performance of our small-form-factor approach against fine-tuned versions of open-source VLMs, such as Florence2 and Qwen2-VL. It also significantly outperforms prompt engineering baselines utilising closed-source foundation models like GPT-4o. More specifically, LiMAC increases the overall action accuracy by up to 19% compared to fine-tuned VLMs, and up to 42% compared to prompt-engineering baselines. | vision-language model, multi-modal, android control, app agent | null | 9,741 | 2410.17883 | [
-0.04336881637573242,
-0.04904515668749809,
0.050089459866285324,
-0.005890659987926483,
0.061467185616493225,
0.008348085917532444,
-0.011485543102025986,
0.05014146864414215,
0.017007289454340935,
0.005293459631502628,
0.009967987425625324,
-0.08815906196832657,
0.03262815624475479,
-0.014493472874164581,
0.004305725917220116,
-0.016810964792966843,
0.08043954521417618,
-0.0029488971922546625,
-0.11323060095310211,
0.014372793026268482,
-0.002395526273176074,
0.070729561150074,
0.06401875615119934,
-0.008376965299248695,
0.07420951873064041,
0.024335134774446487,
-0.07665277272462845,
-0.08486253768205643,
0.04715276136994362,
-0.04875291883945465,
0.020037267357110977,
0.013655743561685085,
0.006695664022117853,
0.03376294672489166,
-0.11311902850866318,
0.004949558060616255,
-0.013609671033918858,
-0.07928721606731415,
-0.09015599638223648,
-0.11637132614850998,
-0.07054707407951355,
-0.006616941187530756,
0.04750656336545944,
0.07739870995283127,
0.008633210323750973,
0.025878669694066048,
0.054916705936193466,
-0.006467370316386223,
-0.007469537202268839,
-0.017684821039438248,
0.012182212434709072,
-0.03290749713778496,
0.0734536200761795,
0.012745467945933342,
-0.12855784595012665,
0.04087246581912041,
-0.016281399875879288,
0.04271603003144264,
0.10886809974908829,
0.05685633048415184,
0.024743076413869858,
0.026800505816936493,
-0.07391457259654999,
0.018923675641417503,
-0.08494745194911957,
0.03091408684849739,
0.07055091857910156,
-0.08291570097208023,
0.03311147168278694,
-0.12196841835975647,
-0.04265356808900833,
0.04700406268239021,
0.04768078774213791,
-0.02402096427977085,
-0.06368008255958557,
-0.04117169976234436,
-0.028365381062030792,
-0.06201522797346115,
-0.021503178402781487,
-0.046790625900030136,
0.06295931339263916,
-0.06589274853467941,
-0.07727587968111038,
0.054279621690511703,
0.10317375510931015,
-0.042607229202985764,
0.016594072803854942,
0.08120298385620117,
0.03479434922337532,
0.044668179005384445,
-0.0337907075881958,
-0.0022974403109401464,
0.06696093827486038,
-0.07162287086248398,
-0.035748157650232315,
0.0064285509288311005,
-0.06553936749696732,
-0.08361069113016129,
-0.10787694901227951,
0.08588336408138275,
-0.026456549763679504,
0.008999516256153584,
-0.020112276077270508,
0.005414038430899382,
0.044268686324357986,
0.012475910596549511,
0.041827257722616196,
-0.010759130120277405,
0.00970078632235527,
0.025771940127015114,
0.02987261861562729,
-0.041123632341623306,
0.026746151968836784,
-0.02758985385298729,
0.0819961279630661,
0.030150584876537323,
0.01828174851834774,
-0.01922190748155117,
-0.005908660590648651,
0.0066752079874277115,
-0.029838353395462036,
-0.00300515815615654,
-0.08739205449819565,
-0.10329204052686691,
0.06478144973516464,
0.004115740768611431,
-0.09534624963998795,
1.3763219586950097e-33,
0.0586635060608387,
0.06581506878137589,
0.05282090976834297,
0.03450348600745201,
0.03687307983636856,
-0.04600411653518677,
0.0707017183303833,
0.0322238951921463,
0.012308908626437187,
-0.023707687854766846,
0.019494544714689255,
-0.017586713656783104,
-0.05273999646306038,
0.08438336849212646,
0.10837806016206741,
-0.024588782340288162,
-0.036395829170942307,
-0.010944993235170841,
0.0025412077084183693,
-0.008232778869569302,
-0.022496193647384644,
-0.06261543184518814,
0.012866511940956116,
-0.018651291728019714,
0.016113972291350365,
0.01840486377477646,
0.0859205424785614,
0.0004997673677280545,
-0.022981956601142883,
-0.009999455884099007,
-0.10467708110809326,
-0.003908700775355101,
0.0041694906540215015,
0.0051922849379479885,
0.04769698157906532,
-0.019862418994307518,
-0.033177971839904785,
-0.06077862158417702,
0.09329384565353394,
0.010343953035771847,
-0.10263488441705704,
0.030471650883555412,
-0.028959065675735474,
-0.0035894461907446384,
-0.07023872435092926,
-0.024078581482172012,
-0.01307010743767023,
0.07156623899936676,
-0.08747170865535736,
0.002510951366275549,
0.04269871860742569,
0.03394303098320961,
0.011807008646428585,
-0.03925193101167679,
-0.010613000020384789,
-0.01592758297920227,
0.029549237340688705,
0.053003452718257904,
-0.027008023113012314,
0.06122677028179169,
0.003305017249658704,
0.03749047592282295,
-0.05951118841767311,
0.13291791081428528,
-0.03512611985206604,
0.014028577134013176,
0.01793782413005829,
0.03028620220720768,
0.0032165185548365116,
0.008432039991021156,
-0.04802309349179268,
0.028240913525223732,
0.04642104730010033,
-0.013685062527656555,
0.06837679445743561,
0.025548135861754417,
0.10943436622619629,
-0.11590351909399033,
-0.009506615810096264,
0.027667464688420296,
-0.02071853168308735,
0.02689758874475956,
0.03502495959401131,
0.007197223603725433,
0.02176877111196518,
-0.09743696451187134,
0.039573006331920624,
-0.05700821056962013,
-0.06355465203523636,
0.029734276235103607,
-0.11891327053308487,
-0.02076926827430725,
0.009697641246020794,
-0.004876800812780857,
-0.004505347926169634,
-9.765678295555319e-34,
-0.017336560413241386,
-0.04039836674928665,
-0.013883007690310478,
0.03773018717765808,
0.022642146795988083,
-0.08832093328237534,
0.025946613401174545,
0.010898089967668056,
0.008370507508516312,
-0.015751099213957787,
-0.029035985469818115,
0.0270280372351408,
0.043855518102645874,
-0.02553221955895424,
0.08117728680372238,
-0.023790590465068817,
-0.001573092071339488,
-0.05036858841776848,
0.0038263031747192144,
-0.014292282052338123,
-0.03543736785650253,
0.09639573842287064,
-0.04454950988292694,
0.010738004930317402,
-0.09899454563856125,
-0.04016319289803505,
-0.048121608793735504,
0.03360873833298683,
0.005703939590603113,
-0.03190777450799942,
-0.036765776574611664,
-0.015817470848560333,
0.036264196038246155,
-0.029080042615532875,
-0.014741732738912106,
0.02051180601119995,
0.027848821133375168,
-0.029794778674840927,
0.0194734837859869,
0.012810233980417252,
0.07653047144412994,
-0.05863210931420326,
-0.012815303169190884,
0.011553737334907055,
0.01978760026395321,
0.011699804104864597,
-0.07369603961706161,
0.009920487180352211,
-0.05110270529985428,
-0.017619648948311806,
0.03324189782142639,
0.06415878236293793,
-0.09653466939926147,
0.009895172901451588,
-0.004047572612762451,
-0.06672217696905136,
0.02141069807112217,
-0.01871906965970993,
0.03731287270784378,
-0.04254237934947014,
0.015446686185896397,
0.016260569915175438,
-0.020075619220733643,
-0.05795387551188469,
0.028703881427645683,
0.013190190307796001,
0.037484265863895416,
-0.03374907374382019,
0.053845200687646866,
0.03521678224205971,
0.04646025970578194,
-0.07363653928041458,
0.04293625429272652,
0.05177801102399826,
0.02428881824016571,
-0.0024029212072491646,
-0.009645655751228333,
-0.019976342096924782,
-0.026935040950775146,
-0.04731707647442818,
0.08749058842658997,
0.04817308112978935,
0.10100962221622467,
0.02542094513773918,
0.08324487507343292,
0.11451785266399384,
-0.005969516932964325,
0.01720448024570942,
0.04013904184103012,
0.09716621786355972,
-0.07675835490226746,
0.11664139479398727,
0.059585437178611755,
0.0950334444642067,
-0.07083616405725479,
-4.426441790883473e-8,
0.032347407191991806,
0.07750046253204346,
0.04178141430020332,
-0.0670306384563446,
-0.03304513171315193,
0.0019925672095268965,
0.0019139749929308891,
0.0034511529374867678,
0.09684182703495026,
0.017868604511022568,
0.030343923717737198,
-0.04458733648061752,
-0.03519798815250397,
-0.024462444707751274,
0.012720844708383083,
0.031151998788118362,
0.03952425718307495,
0.0590101033449173,
-0.03037961944937706,
0.021936997771263123,
0.05570231005549431,
0.06858307868242264,
-0.017950095236301422,
-0.03308269754052162,
0.059608206152915955,
-0.07776587456464767,
-0.023658426478505135,
0.0348404124379158,
0.03707567974925041,
0.002914775861427188,
0.06745094060897827,
0.0395030714571476,
0.014287404716014862,
-0.06504498422145844,
-0.012993473559617996,
0.05357193946838379,
0.04628065228462219,
-0.07240185886621475,
0.07558750361204147,
-0.0051711248233914375,
-0.03450022265315056,
0.050558775663375854,
-0.03255019709467888,
-0.0025832983665168285,
-0.08558367937803268,
-0.012012885883450508,
-0.0008624070324003696,
-0.21721680462360382,
0.05638095736503601,
-0.0021725650876760483,
-0.011038222350180149,
-0.006190885324031115,
-0.017718486487865448,
0.043252069503068924,
0.043882936239242554,
0.04876711964607239,
0.05227933079004288,
-0.11198422312736511,
0.0005813016905449331,
0.03423471748828888,
0.008650810457766056,
0.027033351361751556,
-0.0487675778567791,
0.008097777143120766
] |
Progressive Compositionality in Text-to-Image Generative Models | https://openreview.net/forum?id=S85PP4xjFD | [
"Xu Han",
"Linghao Jin",
"Xiaofeng Liu",
"Paul Pu Liang"
] | Spotlight | Despite the impressive text-to-image (T2I) synthesis capabilities of diffusion models, they often struggle to understand compositional relationships between objects and attributes, especially in complex settings. Existing approaches through building compositional architectures or generating difficult negative captions often assume a fixed prespecified compositional structure, which limits generalization to new distributions. In this paper, we argue that curriculum training is crucial to equipping generative models with a fundamental understanding of compositionality. To achieve this, we leverage large-language models (LLMs) to automatically compose complex scenarios and harness Visual-Question Answering (VQA) checkers to automatically curate a contrastive dataset, ConPair, consisting of 15k pairs of high-quality contrastive images. These pairs feature minimal visual discrepancies and cover a wide range of attribute categories, especially complex and natural scenarios. To learn effectively from these error cases (i.e., hard negative images), we propose EvoGen, a new multi-stage curriculum for contrastive learning of diffusion models. Through extensive experiments across a wide range of compositional scenarios, we showcase the effectiveness of our proposed framework on compositional T2I benchmarks. | compositional text-to-image generation, contrastive learning, compositional understanding, T2I generation | Contrastively improving compositional text-to-image generation in diffusion models via generating fine-grained hard negative images. | 9,669 | 2410.16719 | [
0.03170916810631752,
-0.08364788442850113,
0.04858384281396866,
0.06982060521841049,
0.049466148018836975,
-0.013401915319263935,
-0.05914755538105965,
-0.011098549701273441,
0.03282348811626434,
-0.05928638577461243,
0.002072314964607358,
-0.14502009749412537,
0.0573158785700798,
0.09316190332174301,
-0.048253219574689865,
0.008429770357906818,
0.1373860388994217,
0.06326895207166672,
-0.08752229064702988,
-0.013191456906497478,
-0.054050739854574203,
0.03123370185494423,
0.11017455160617828,
-0.056075017899274826,
0.016857385635375977,
0.01361177396029234,
-0.00015839803381823003,
-0.05604488030076027,
0.03152933344244957,
-0.05553916469216347,
0.012075153179466724,
0.010832761414349079,
-0.05378689616918564,
0.0970873162150383,
-0.01661297306418419,
0.10418187826871872,
-0.01388738863170147,
0.05697060003876686,
-0.0002789823047351092,
-0.05089082196354866,
-0.023398302495479584,
-0.007875320501625538,
0.0043535116128623486,
-0.02064739353954792,
0.08976767957210541,
-0.08755824714899063,
-0.019130190834403038,
-0.03376094251871109,
0.03650284186005592,
0.015219987370073795,
-0.05382528901100159,
-0.01569388248026371,
-0.05532805994153023,
0.0513114333152771,
-0.025862962007522583,
0.0875508189201355,
0.07105641812086105,
0.006246235221624374,
0.01083921268582344,
-0.024094047024846077,
-0.03812723606824875,
-0.04550868272781372,
-0.032594457268714905,
-0.0071021756157279015,
0.01171313039958477,
-0.014085983857512474,
0.009736550971865654,
0.07421324402093887,
-0.08206077665090561,
0.07845759391784668,
-0.04482879862189293,
0.04513953626155853,
-0.01710853911936283,
0.01793856918811798,
0.013671395368874073,
0.10357964783906937,
0.0028884036000818014,
-0.10368724167346954,
0.032369185239076614,
-0.09375058114528656,
0.044899363070726395,
0.022150278091430664,
0.09198720008134842,
0.011460857465863228,
0.12571030855178833,
0.03030153177678585,
0.014937530271708965,
0.0699305608868599,
-0.05459870025515556,
-0.00802221242338419,
-0.08782617747783661,
-0.02347428910434246,
0.08540979027748108,
-0.009785539470613003,
0.08304793387651443,
0.05772820860147476,
-0.007696034852415323,
-0.06431088596582413,
0.023158559575676918,
0.06338581442832947,
0.021676944568753242,
0.03620095178484917,
-0.016027165576815605,
-0.0796298235654831,
0.010567517951130867,
-0.033421728760004044,
0.0376552939414978,
-0.001126885530538857,
0.009340304881334305,
-0.10222237557172775,
0.042490165680646896,
0.03341687470674515,
-0.004152378067374229,
-0.03645988553762436,
0.009030361659824848,
-0.03714781999588013,
0.058901939541101456,
-0.03175028786063194,
0.02554589882493019,
-0.0002125406317645684,
0.03009820729494095,
0.06235847622156143,
-0.03789482265710831,
0.0022382643073797226,
0.052415523678064346,
-0.07757977396249771,
0.016044972464442253,
-1.2793718682494915e-33,
0.08685602992773056,
-0.03235606849193573,
0.0658339336514473,
0.10013942420482635,
0.10089244693517685,
0.014209193177521229,
-0.05120884254574776,
0.07276076078414917,
-0.05507887527346611,
0.011289776302874088,
0.010143570601940155,
0.0026514616329222918,
-0.10601552575826645,
0.1010354533791542,
0.06930271536111832,
-0.036155108362436295,
-0.05118121951818466,
0.04155103862285614,
-0.0316944494843483,
0.003602968994528055,
-0.03995155170559883,
-0.009683361276984215,
0.03201479837298393,
-0.04954379051923752,
0.008974052965641022,
-0.024428660050034523,
0.10803074389696121,
-0.05933608487248421,
-0.05826294422149658,
-0.01918906159698963,
-0.17691931128501892,
-0.013073666952550411,
-0.020746348425745964,
0.03064902313053608,
0.02387818694114685,
-0.050766270607709885,
-0.0275166854262352,
-0.021779567003250122,
0.03417966887354851,
-0.03624941036105156,
-0.022617967799305916,
0.026176204904913902,
-0.07225014269351959,
0.022151418030261993,
-0.07443379610776901,
0.0018677747575566173,
0.05872219055891037,
0.03913164883852005,
-0.01440418977290392,
-0.055404603481292725,
0.04746529087424278,
-0.004794062580913305,
-0.02396894432604313,
-0.0682411715388298,
-0.0003819615230895579,
-0.016355836763978004,
0.005201053339987993,
-0.0039608594961464405,
0.03270585834980011,
-0.03526541590690613,
0.04109177738428116,
0.08146961033344269,
-0.05159473419189453,
0.10084180533885956,
0.012281030416488647,
0.04971175640821457,
-0.05767181143164635,
0.05261319875717163,
0.02525394968688488,
-0.000464519951492548,
-0.007566861342638731,
0.002352259587496519,
-0.027798667550086975,
-0.07182382792234421,
0.08859482407569885,
0.05426372215151787,
0.010075144469738007,
-0.06897182017564774,
0.02903466299176216,
0.03794721141457558,
-0.020977426320314407,
-0.04377356544137001,
-0.02688199281692505,
-0.11434865742921829,
-0.07564615458250046,
-0.04672184959053993,
0.04093177989125252,
-0.07691072672605515,
0.015160912647843361,
-0.028082380071282387,
-0.04793381690979004,
-0.05682847648859024,
-0.08014144003391266,
-0.00609229551628232,
0.06425417214632034,
7.466939399449833e-34,
-0.03324630856513977,
0.0482456348836422,
-0.09359272569417953,
0.026171326637268066,
-0.016573842614889145,
-0.0564783550798893,
-0.0004685770836658776,
0.07164779305458069,
0.015883224084973335,
-0.058948930352926254,
0.01704726554453373,
-0.013214613310992718,
-0.033905234187841415,
-0.06489711999893188,
0.0008398090722039342,
-0.030889468267560005,
0.013928430154919624,
0.011770403943955898,
-0.03681052848696709,
0.03254232555627823,
-0.0015105303609743714,
0.10633936524391174,
-0.05066550523042679,
0.022323161363601685,
-0.06252647191286087,
0.08915883302688599,
-0.04502568393945694,
0.054431501775979996,
-0.04935312643647194,
0.011352024041116238,
-0.01852879859507084,
0.02699601650238037,
0.01795235089957714,
-0.014821325428783894,
-0.12758657336235046,
0.08468902111053467,
0.04404179006814957,
-0.06268702447414398,
-0.026786912232637405,
0.07982717454433441,
0.0696081891655922,
-0.07220640033483505,
-0.03146446496248245,
0.0430510975420475,
-0.04263606667518616,
-0.04391781985759735,
-0.044639844447374344,
-0.0513806976377964,
0.05904721841216087,
0.05765017122030258,
-0.041608333587646484,
-0.03146986663341522,
-0.0728127583861351,
0.016904596239328384,
0.00668691610917449,
-0.1049938052892685,
-0.05276286229491234,
-0.0597098208963871,
0.06444931775331497,
0.00504679000005126,
-0.053886231034994125,
-0.006050430238246918,
0.02268979698419571,
-0.07551059126853943,
-0.03201786428689957,
-0.07088878750801086,
-0.06554558128118515,
-0.011576693505048752,
0.025211259722709656,
0.003569861873984337,
0.03176122531294823,
-0.0012587960809469223,
0.007292147260159254,
0.008929853327572346,
0.035236749798059464,
-0.04893796145915985,
0.060356806963682175,
0.01720983162522316,
0.0070816148072481155,
-0.0958041250705719,
0.04274910315871239,
-0.055295296013355255,
-0.0014041735557839274,
0.0695645809173584,
0.09189443290233612,
0.06833163648843765,
-0.009204715490341187,
0.045064810663461685,
0.07435190677642822,
0.0675797238945961,
-0.006865471601486206,
-0.029965439811348915,
-0.009670947678387165,
0.06338359415531158,
-0.03857364505529404,
-3.503659229409095e-8,
-0.06403780728578568,
0.043974000960588455,
-0.020673613995313644,
0.02249244973063469,
-0.012722166255116463,
-0.06351137161254883,
-0.03429252654314041,
0.06679698824882507,
-0.0026524493005126715,
0.0017627261113375425,
0.029826555401086807,
0.04967448115348816,
0.019071923568844795,
-0.08282297104597092,
-0.02529112622141838,
0.09748077392578125,
0.01381854247301817,
0.0006411096546798944,
0.02010800875723362,
-0.0005094162188470364,
0.055348481982946396,
-0.00022035853180568665,
-0.006949336268007755,
0.017223356291651726,
0.0061621349304914474,
-0.016296906396746635,
-0.05831129103899002,
0.016824809834361076,
0.08966401219367981,
-0.009878702461719513,
0.07956762611865997,
0.029310325160622597,
-0.0275065116584301,
-0.04606848955154419,
-0.0010501547949388623,
0.06902822852134705,
0.039722975343465805,
-0.02857682667672634,
0.007843030616641045,
-0.040281981229782104,
-0.0197153240442276,
0.02057402767241001,
-0.018769122660160065,
0.02596616931259632,
-0.00037027528742328286,
0.018111534416675568,
-0.01320337038487196,
-0.02377503179013729,
0.05644732341170311,
0.029078280553221703,
-0.050163350999355316,
-0.07236351072788239,
-0.00013334120740182698,
0.07303246110677719,
0.03587620332837105,
-0.04838589206337929,
0.03737983852624893,
0.02271895669400692,
0.057477422058582306,
0.04922694340348244,
0.04174695909023285,
0.060572754591703415,
0.0005763969966210425,
-0.02825980633497238
] |
On Quantizing Neural Representation for Variable-Rate Video Coding | https://openreview.net/forum?id=44cMlQSreK | [
"Junqi Shi",
"Zhujia Chen",
"Hanfei Li",
"Qi Zhao",
"Ming Lu",
"Tong Chen",
"Zhan Ma"
] | Spotlight | This work introduces NeuroQuant, a novel post-training quantization (PTQ) approach tailored to non-generalized Implicit Neural Representations for variable-rate Video Coding (INR-VC). Unlike existing methods that require extensive weight retraining for each target bitrate, we hypothesize that variable-rate coding can be achieved by adjusting quantization parameters (QPs) of pre-trained weights. Our study reveals that traditional quantization methods, which assume inter-layer independence, are ineffective for non-generalized INR-VC models due to significant dependencies across layers. To address this, we redefine variable-rate INR-VC as a mixed-precision quantization problem and establish a theoretical framework for sensitivity criteria aimed at simplified, fine-grained rate control. Additionally, we propose network-wise calibration and channel-wise quantization strategies to minimize quantization-induced errors, arriving at a unified formula for representation-oriented PTQ calibration. Our experimental evaluations demonstrate that NeuroQuant significantly outperforms existing techniques in varying bitwidth quantization and compression efficiency, accelerating encoding by up to eight times and enabling quantization down to INT2 with minimal reconstruction loss. This work introduces variable-rate INR-VC for the first time and lays a theoretical foundation for future research in rate-distortion optimization, advancing the field of video coding technology. The materials
will be available at https://github.com/Eric-qi/NeuroQuant. | Variable Rate, Video Coding, Quantization, Neural Representation | null | 9,659 | 2502.11729 | [
-0.017059609293937683,
-0.10827846825122833,
0.04269442334771156,
0.02721005491912365,
0.0128486854955554,
0.08518745750188828,
-0.01806897670030594,
-0.09185829013586044,
0.07566457986831665,
-0.07164350897073746,
0.0011800980428233743,
-0.07902569323778152,
-0.07130307704210281,
0.05457471311092377,
-0.07071392983198166,
0.007884083315730095,
0.07928142696619034,
0.11383802443742752,
-0.08806264400482178,
-0.05319537967443466,
0.06520571559667587,
-0.07824119180440903,
-0.02378539741039276,
-0.01619189977645874,
0.10639632493257523,
-0.034617263823747635,
-0.013118835166096687,
0.024133507162332535,
0.0555395781993866,
0.0014989098999649286,
0.025270283222198486,
-0.036630693823099136,
0.015438362956047058,
-0.017202164977788925,
-0.07125589996576309,
0.01576955057680607,
-0.019684499129652977,
-0.05156109109520912,
-0.017655717208981514,
-0.03816260024905205,
0.03341855853796005,
0.03726715222001076,
-0.010127437300980091,
0.011223995126783848,
0.01942373253405094,
0.053532253950834274,
0.057316236197948456,
-0.0027902426663786173,
-0.0750732570886612,
-0.029866669327020645,
0.09217240661382675,
0.07322338968515396,
-0.008901267312467098,
0.00433692941442132,
0.0015719999792054296,
0.03926835209131241,
0.01363296527415514,
0.04369312524795532,
-0.06543684005737305,
0.014852036722004414,
-0.06782693415880203,
-0.0019400367746129632,
0.07099876552820206,
0.02060634456574917,
-0.009841613471508026,
0.010324732400476933,
0.07773842662572861,
0.04644741117954254,
-0.007468604017049074,
-0.004226307850331068,
-0.06868422776460648,
-0.014518745243549347,
-0.04543616250157356,
0.0728553980588913,
-0.004040091298520565,
0.011249998584389687,
0.05664655566215515,
0.02951839193701744,
0.054785486310720444,
-0.07094771414995193,
0.030456338077783585,
0.011855337768793106,
0.030652418732643127,
-0.001576100243255496,
0.10786992311477661,
-0.02019372768700123,
0.01164937112480402,
0.08601278066635132,
0.013760510832071304,
0.020671671256422997,
-0.14341402053833008,
0.017109166830778122,
0.012048784643411636,
-0.03722313418984413,
0.06629538536071777,
-0.013191662728786469,
-0.03602946549654007,
-0.02339860238134861,
-0.012483043596148491,
-0.010279307141900063,
0.04738558083772659,
-0.029172129929065704,
-0.016444126144051552,
0.057663995772600174,
0.056759752333164215,
-0.018763164058327675,
0.10913941264152527,
0.11433201283216476,
0.018746014684438705,
0.01869964227080345,
0.014027277007699013,
0.02133386768400669,
0.006696135737001896,
-0.09072349965572357,
0.07356181740760803,
0.0625334084033966,
-0.04418124631047249,
0.0007539117359556258,
0.040025219321250916,
-0.014523505233228207,
-0.0896553099155426,
-0.10721895843744278,
-0.14076727628707886,
0.018117215484380722,
0.051148053258657455,
-0.054881639778614044,
-0.06064188852906227,
3.5466681271804044e-33,
-0.013746929354965687,
0.055005088448524475,
0.0061127603985369205,
-0.035584621131420135,
0.04091383516788483,
-0.004572407808154821,
0.061091627925634384,
0.035267412662506104,
-0.00674153259024024,
0.04159555584192276,
0.016515005379915237,
-0.10893480479717255,
-0.10466951131820679,
0.0793711245059967,
0.032897740602493286,
0.023304911330342293,
-0.12639042735099792,
-0.009574844501912594,
0.017348533496260643,
-0.0346565768122673,
0.02417122945189476,
-0.04088234156370163,
-0.060955457389354706,
0.04497378692030907,
-0.012005051597952843,
-0.014082256704568863,
0.049459125846624374,
-0.0017677668947726488,
-0.05643410608172417,
-0.030132023617625237,
-0.03258553892374039,
-0.056640855967998505,
0.050942059606313705,
-0.07663960009813309,
0.007915813475847244,
-0.06190866231918335,
0.0302291102707386,
0.031253982335329056,
-0.014126021414995193,
0.034479252994060516,
0.04332463815808296,
0.08542614430189133,
0.010541750118136406,
-0.04052755981683731,
-0.0980479046702385,
-0.04814997687935829,
0.03274970501661301,
-0.00803618784993887,
-0.13454939424991608,
-0.059515099972486496,
0.05612197518348694,
0.02074376493692398,
0.0031277742236852646,
-0.03390033543109894,
0.03141489252448082,
-0.04078613966703415,
0.056102581322193146,
0.01446375623345375,
-0.022923164069652557,
0.08394519984722137,
0.007480634842067957,
-0.027272284030914307,
-0.007414359133690596,
0.031366147100925446,
0.002355178352445364,
-0.038749031722545624,
0.06751357018947601,
0.052512601017951965,
-0.06419394165277481,
0.015438633970916271,
-0.08453201502561569,
0.07282620668411255,
0.017715178430080414,
-0.09513931721448898,
0.05521020293235779,
0.020872583612799644,
0.019063830375671387,
-0.048840172588825226,
-0.03840135037899017,
0.020560679957270622,
-0.001932719605974853,
0.05255010724067688,
0.008646583184599876,
0.029552657157182693,
-0.033443938940763474,
0.0038673090748488903,
0.05090971663594246,
-0.0007807934307493269,
-0.03241780772805214,
0.03530791029334068,
-0.0065593961626291275,
-0.02995252050459385,
0.009225825779139996,
-0.06069313734769821,
-0.05986820533871651,
-2.1481487025445428e-33,
-0.04206375032663345,
0.14510318636894226,
-0.02143169566988945,
0.06428150087594986,
-0.05157559737563133,
-0.02335086278617382,
0.037855930626392365,
-0.00023820882779546082,
0.11554106324911118,
-0.10122570395469666,
-0.0269989762455225,
-0.09296032041311264,
-0.04355042055249214,
0.022458059713244438,
-0.018683915957808495,
0.02294760011136532,
-0.1155545637011528,
0.05241764709353447,
-0.02319871447980404,
-0.005765875335782766,
0.02223099395632744,
0.022119862958788872,
0.00829085148870945,
-0.017430556938052177,
0.022606385871767998,
0.02768828719854355,
-0.004588944837450981,
0.09638535231351852,
0.058991286903619766,
-0.03125054016709328,
-0.007648021448403597,
-0.03298741206526756,
-0.002473706379532814,
-0.09400970488786697,
-0.04182753339409828,
0.023733465000987053,
0.09644908457994461,
-0.01890774443745613,
-0.03567848727107048,
0.07863704860210419,
0.07953793555498123,
-0.026022620499134064,
-0.12224480509757996,
0.0690690129995346,
0.06480798125267029,
-0.015389029867947102,
-0.029742825776338577,
-0.04426811635494232,
-0.018312811851501465,
0.047179292887449265,
-0.01886744238436222,
0.033569417893886566,
-0.03154635429382324,
0.08683738857507706,
-0.017491964623332024,
-0.04137471318244934,
-0.09364112466573715,
0.0509757436811924,
0.11992022395133972,
0.03623897209763527,
0.0023529608733952045,
-0.009411264210939407,
-0.02004534937441349,
-0.10833847522735596,
-0.04055356606841087,
-0.03135732188820839,
0.018696114420890808,
0.03749026358127594,
0.032307013869285583,
0.023275623098015785,
-0.009605013765394688,
0.0013972260057926178,
0.05479499697685242,
0.014942752197384834,
0.003855957416817546,
-0.05330035835504532,
0.008358635008335114,
0.008628667332231998,
0.006594196427613497,
0.02574753388762474,
-0.03289945796132088,
0.029434848576784134,
0.019089259207248688,
-0.005380749236792326,
0.1019582748413086,
0.09749128669500351,
0.04215935617685318,
0.08095309138298035,
0.01817895844578743,
-0.018731828778982162,
-0.046233341097831726,
0.05521376058459282,
0.03873709216713905,
0.009038444608449936,
-0.10287534445524216,
-4.258461316908324e-8,
-0.03878891468048096,
0.038946498185396194,
-0.04431286081671715,
-0.018358703702688217,
-0.011830322444438934,
-0.003554900409653783,
0.043047066777944565,
0.0062300898134708405,
0.026275603100657463,
-0.03567860275506973,
0.1012222021818161,
0.007217312231659889,
-0.05439671874046326,
-0.03552720323204994,
-0.04213419184088707,
0.060661233961582184,
-0.004152646288275719,
-0.007418324239552021,
0.0339038148522377,
0.023543687537312508,
-0.0069679757580161095,
0.10392553359270096,
-0.0385311134159565,
-0.06827468425035477,
0.008731858804821968,
-0.03599897772073746,
0.017187166959047318,
0.08536360412836075,
0.08556773513555527,
0.0043672313913702965,
-0.012038023211061954,
0.058597397059202194,
0.059517376124858856,
-0.05711079016327858,
0.030433403328061104,
-0.0016654691426083446,
-0.008158627897500992,
0.022299444302916527,
0.025916285812854767,
-0.0330016165971756,
-0.06676145642995834,
-0.036087166517972946,
-0.011112750507891178,
-0.003404174465686083,
0.031979307532310486,
0.013705567456781864,
-0.008788570761680603,
-0.013544557616114616,
-0.01340656727552414,
-0.03620579093694687,
0.02576139196753502,
0.1062113419175148,
-0.04229065775871277,
0.046028729528188705,
0.055911749601364136,
-0.05335424095392227,
0.06807844340801239,
-0.06046834588050842,
-0.049663152545690536,
0.032844871282577515,
-0.04154394567012787,
0.002283194800838828,
-0.045453350991010666,
-0.04281413555145264
] |
Can Watermarked LLMs be Identified by Users via Crafted Prompts? | https://openreview.net/forum?id=ujpAYpFDEA | [
"Aiwei Liu",
"Sheng Guan",
"Yiming Liu",
"Leyi Pan",
"Yifei Zhang",
"Liancheng Fang",
"Lijie Wen",
"Philip S. Yu",
"Xuming Hu"
] | Spotlight | Text watermarking for Large Language Models (LLMs) has made significant progress in detecting LLM outputs and preventing misuse. Current watermarking techniques offer high detectability, minimal impact on text quality, and robustness to text editing.
However, current researches lack investigation into the imperceptibility of watermarking techniques in LLM services.
This is crucial as LLM providers may not want to disclose the presence of watermarks in real-world scenarios, as it could reduce user willingness to use the service and make watermarks more vulnerable to attacks. This work is the first to investigate the imperceptibility of watermarked LLMs. We design an identification algorithm called Water-Probe that detects watermarks through well-designed prompts to the LLM. Our key motivation is that current watermarked LLMs expose consistent biases under the same watermark key, resulting in similar differences across prompts under different watermark keys. Experiments show that almost all mainstream watermarking algorithms are easily identified with our well-designed prompts,
while Water-Probe demonstrates a minimal false positive rate for non-watermarked LLMs.
Finally, we propose that the key to enhancing the imperceptibility of watermarked LLMs is to increase the randomness of watermark key selection. Based on this, we introduce the Water-Bag strategy, which significantly improves watermark imperceptibility by merging multiple watermark keys. | Large Language Models, Watermark, Identification | We present Water-Probe, a method to identify watermarked LLMs using crafted prompts, and propose the Water-Bag strategy to improve watermark imperceptibility. | 9,641 | 2410.03168 | [
-0.0620604082942009,
-0.011408415623009205,
0.049994539469480515,
0.0175493024289608,
0.11136402934789658,
-0.030908968299627304,
0.03416651859879494,
-0.06619108468294144,
0.028320571407675743,
-0.03112180531024933,
-0.08261597901582718,
0.002490786137059331,
0.15488845109939575,
0.005010500084608793,
-0.0162058025598526,
-0.018295206129550934,
0.02771228551864624,
0.03732511028647423,
-0.014066494069993496,
-0.06629439443349838,
-0.005142584443092346,
-0.038791049271821976,
0.040731970220804214,
0.021210238337516785,
-0.012625942006707191,
0.016212692484259605,
-0.002480852883309126,
0.012125326320528984,
-0.034686461091041565,
-0.0011943428544327617,
0.03764977306127548,
-0.0284236092120409,
0.02243248000741005,
0.022099284455180168,
-0.02240174263715744,
-0.014934377744793892,
0.04813918098807335,
0.008827474899590015,
0.015954993665218353,
-0.0860162302851677,
-0.055357180535793304,
-0.07183296233415604,
-0.0439632385969162,
0.09518632292747498,
0.03510087728500366,
-0.00356239452958107,
0.03896196931600571,
0.06473424285650253,
-0.1516944169998169,
0.07901522517204285,
0.005018405616283417,
-0.09495791047811508,
-0.03941335156559944,
-0.007461098022758961,
-0.10371945798397064,
-0.06055038049817085,
-0.01650969497859478,
-0.0012925878399983048,
-0.04707882180809975,
0.03694790229201317,
-0.06374497711658478,
-0.03139227628707886,
-0.02536291442811489,
0.05159372463822365,
0.05230709910392761,
0.029572201892733574,
0.004281109198927879,
0.05052046477794647,
-0.0005819812067784369,
0.005932642146945,
0.010260806418955326,
0.023646587505936623,
0.006151533219963312,
0.06752480566501617,
-0.06050460785627365,
0.05333857983350754,
-0.05540319159626961,
0.01732967048883438,
0.007825170643627644,
-0.050331003963947296,
0.022382168099284172,
-0.07056772708892822,
0.03541048616170883,
0.06698131561279297,
0.05650120973587036,
0.008194692432880402,
0.01851261779665947,
-0.021331196650862694,
0.016413608565926552,
0.023904765024781227,
-0.009617790579795837,
-0.08748218417167664,
0.11832212656736374,
0.001623388146981597,
0.003416813910007477,
0.04736177250742912,
-0.030008956789970398,
-0.02666603960096836,
0.009088436141610146,
0.03707834333181381,
-0.07556357234716415,
-0.0222379919141531,
0.007073529995977879,
-0.07196277379989624,
0.12220262736082077,
0.08203189820051193,
0.024762745946645737,
0.0015072198584675789,
0.0766012892127037,
-0.050944071263074875,
-0.013046294450759888,
0.014966639690101147,
-0.006383771076798439,
-0.02116595022380352,
0.031156010925769806,
-0.07462296634912491,
0.021907709538936615,
-0.05999738350510597,
0.017464149743318558,
0.01105739176273346,
0.045001886785030365,
-0.00031084733200259507,
-0.0033247678074985743,
-0.007993094623088837,
-0.0014888112200424075,
0.031098684296011925,
-0.021495388820767403,
3.589222859390094e-33,
0.07026412338018417,
0.07330618798732758,
-0.0769665539264679,
-0.021545473486185074,
0.04546426609158516,
-0.004671524744480848,
-0.013226820155978203,
0.02345111221075058,
-0.06567434221506119,
-0.08622013032436371,
0.059277359396219254,
-0.010331826284527779,
0.014976675622165203,
0.08404146134853363,
0.05728442966938019,
0.0407090038061142,
0.04262261465191841,
0.0051750182174146175,
-0.024269601330161095,
-0.057213302701711655,
-0.0072766696102917194,
-0.05422430858016014,
0.0340634286403656,
-0.06962570548057556,
0.044145215302705765,
0.060299552977085114,
0.06334599107503891,
-0.01211595255881548,
0.05360458046197891,
0.013192176818847656,
-0.030503392219543457,
-0.08027496933937073,
0.09931313246488571,
-0.004483046475797892,
0.021738344803452492,
0.049622200429439545,
0.0014940723776817322,
-0.052549175918102264,
0.05434561148285866,
0.0031171576119959354,
-0.0017589300405234098,
0.007409309968352318,
-0.027904801070690155,
-0.02005348727107048,
-0.11299850791692734,
0.027169538661837578,
-0.07648281008005142,
-0.0013874153373762965,
-0.0037089744582772255,
-0.0353664793074131,
0.06475690752267838,
0.04368308186531067,
-0.055724501609802246,
0.01727769337594509,
-0.08729825913906097,
-0.049947962164878845,
-0.010235349647700787,
-0.06635434925556183,
0.010082836262881756,
0.022454051300883293,
-0.02762368693947792,
0.042611103504896164,
0.011167816817760468,
0.03135716915130615,
-0.0022301606368273497,
-0.06678611040115356,
0.0480688251554966,
0.02905227616429329,
-0.007474605925381184,
-0.013164825737476349,
-0.02112969011068344,
-0.058611322194337845,
-0.02244589850306511,
0.06778497248888016,
-0.11608333140611649,
-0.08939363807439804,
0.06379063427448273,
0.018115108832716942,
0.03365587443113327,
-0.024623693898320198,
0.03905317559838295,
-0.07333028316497803,
0.09675978124141693,
-0.027588464319705963,
-0.0638517513871193,
-0.09037446230649948,
-0.0026303583290427923,
-0.09438253939151764,
-0.07161920517683029,
0.016367781907320023,
0.014862819574773312,
-0.03400694578886032,
0.02489508129656315,
0.01589326560497284,
-0.0352812334895134,
-4.494755748996428e-33,
-0.07288873195648193,
-0.04899583011865616,
-0.07208544760942459,
0.13304774463176727,
-0.0435883067548275,
-0.028616828843951225,
0.008388107642531395,
0.06879554688930511,
0.020124228671193123,
-0.03599162399768829,
-0.07161006331443787,
0.03990296274423599,
-0.008943520486354828,
-0.020023934543132782,
-0.047264568507671356,
-0.025503063574433327,
0.012133277021348476,
0.08065102994441986,
-0.0185538437217474,
-0.03150765225291252,
0.025899818167090416,
0.011089125648140907,
-0.038279060274362564,
0.07296893000602722,
-0.0017468570731580257,
0.022928984835743904,
0.02719016931951046,
-0.050793495029211044,
-0.023515766486525536,
-0.01602398417890072,
-0.039465997368097305,
0.09938277304172516,
0.0013806758215650916,
-0.061542268842458725,
-0.0320904366672039,
0.0022470036055892706,
0.035457175225019455,
0.04741332679986954,
-0.02071979083120823,
0.08587510883808136,
0.1004457101225853,
-0.009545301087200642,
-0.007589929271489382,
-0.014445425942540169,
-0.03042738139629364,
-0.014011172577738762,
-0.0576791949570179,
0.03751181811094284,
0.07908757776021957,
-0.006564215291291475,
-0.011141300201416016,
-0.09050726890563965,
0.010206947103142738,
0.004745265934616327,
-0.04769709333777428,
-0.04235769808292389,
-0.04543749988079071,
-0.0033031897619366646,
0.008909856900572777,
0.0007074230234138668,
-0.009086353704333305,
0.010952973738312721,
0.010094950906932354,
-0.028319712728261948,
0.08860807120800018,
-0.052278973162174225,
-0.00036159451701678336,
0.06676550954580307,
-0.055177200585603714,
-0.03816645219922066,
0.1490333378314972,
-0.09888302534818649,
-0.07320866733789444,
0.02955758199095726,
0.06751705706119537,
-0.013174186460673809,
-0.02823442779481411,
-0.02074330300092697,
-0.0995946153998375,
0.013569027185440063,
0.08387303352355957,
0.0527733638882637,
-0.03769472986459732,
0.07053827494382858,
0.07439184188842773,
-0.0012601643102243543,
-0.001759681268595159,
0.04356609284877777,
-0.06729385256767273,
0.07020137459039688,
0.0202848669141531,
0.06860379874706268,
-0.05542810633778572,
0.07014548778533936,
0.013768981210887432,
-4.056237301597321e-8,
-0.03451595827937126,
-0.05106740817427635,
-0.06498666852712631,
-0.008849730715155602,
-0.034775007516145706,
0.04138676077127457,
-0.05927965044975281,
0.074987031519413,
0.05993509292602539,
-0.15603090822696686,
0.052828989923000336,
-0.06061478704214096,
-0.13208229839801788,
-0.0867181345820427,
-0.032280489802360535,
0.035290926694869995,
-0.081594318151474,
0.014034165069460869,
-0.08534806966781616,
-0.03813672065734863,
-0.018340708687901497,
0.0810268372297287,
0.0028004853520542383,
-0.053425904363393784,
0.00966724194586277,
0.02914479561150074,
0.028300950303673744,
0.09695027768611908,
0.0064847711473703384,
0.062388937920331955,
-0.0015311937313526869,
-0.02307097055017948,
0.008414323441684246,
-0.0004797453584615141,
0.05494983121752739,
0.07656276226043701,
-0.022534120827913284,
-0.0006693910690955818,
-0.00919873546808958,
0.07060223817825317,
0.02424389310181141,
-0.0035831143613904715,
0.0125805102288723,
0.012210597284138203,
0.0012180386111140251,
-0.008442831225693226,
0.029252104461193085,
-0.09928063303232193,
0.025412028655409813,
-0.07325220853090286,
0.028429720550775528,
-0.08776628971099854,
0.04338722303509712,
0.055369604378938675,
-0.026096027344465256,
0.013239894062280655,
0.02601487934589386,
0.0499175600707531,
0.07958260178565979,
0.01015628594905138,
0.11952821910381317,
-0.01817389763891697,
0.05733273923397064,
0.0441184900701046
] |
Temporal Heterogeneous Graph Generation with Privacy, Utility, and Efficiency | https://openreview.net/forum?id=tj5xJInWty | [
"Xinyu He",
"Dongqi Fu",
"Hanghang Tong",
"Ross Maciejewski",
"Jingrui He"
] | Spotlight | Nowadays, temporal heterogeneous graphs attract much research and industrial attention for building the next-generation Relational Deep Learning models and applications, due to their informative structures and features. While providing timely and precise services like personalized recommendations and question answering, this rich information also introduces extra exposure risk for each node in the graph. The distinctive local topology, the abundant heterogeneous features, and the time dimension of the graph data are more prone to expose sensitive information and narrow down the scope of victim candidates, which calls for well-defined protection techniques on graphs. To this end, we propose a Temporal Heterogeneous Graph Generator balancing Privacy, Utility, and Efficiency, named THePUff. More specifically, we first propose a differential privacy algorithm to perturb the input temporal heterogeneous graph for protecting privacy, and then utilize both the perturbed graph and the original one in a generative adversarial setting for THePUff to learn and generate privacy-guaranteed and utility-preserved graph data in an efficient manner. We further propose 6 new metrics in the temporal setting to measure heterogeneous graph utility and privacy. Finally, based on temporal heterogeneous graph datasets with up to 1 million nodes and 20 million edges, the experiments show that THePUff generates utilizable temporal heterogeneous graphs with privacy protected, compared with state-of-the-art baselines. | Temporal Graph, Heterogeneous Graph, Graph Generation | null | 9,626 | null | [
-0.12120416760444641,
-0.0013633393682539463,
-0.00988810695707798,
0.05375369265675545,
0.06531398743391037,
0.021477140486240387,
-0.07461380213499069,
-0.05198022350668907,
0.0116756372153759,
-0.0241833608597517,
0.005442121997475624,
-0.04837944731116295,
0.026086345314979553,
0.005125661380589008,
-0.005404421128332615,
0.056653399020433426,
0.081334188580513,
0.059755124151706696,
-0.0671970471739769,
-0.055660463869571686,
-0.004400711972266436,
0.0008952628122642636,
0.0016595333581790328,
0.0014872063184157014,
0.04428410902619362,
-0.006554813124239445,
0.05103512108325958,
-0.03589503467082977,
0.05185195431113243,
0.014631121419370174,
0.1050306037068367,
-0.03632591292262077,
-0.03046148084104061,
-0.008989371359348297,
-0.043806467205286026,
0.024191731587052345,
-0.048415351659059525,
-0.0031850074883550406,
0.030714018270373344,
-0.05559701845049858,
-0.005140431225299835,
-0.0582890547811985,
-0.021211525425314903,
0.0637345239520073,
-0.03626565262675285,
0.03506110981106758,
0.010067558847367764,
0.013296096585690975,
-0.03545520454645157,
0.0009622303186915815,
0.010874910280108452,
-0.03190186992287636,
-0.007456241641193628,
0.020561642944812775,
0.012020186521112919,
-0.07309482246637344,
-0.02392655983567238,
-0.03841995820403099,
-0.03740130737423897,
0.04866429790854454,
-0.04364612698554993,
-0.10371969640254974,
-0.007734273094683886,
-0.05722489953041077,
-0.06401389092206955,
0.02284344844520092,
-0.007199151907116175,
0.0248564463108778,
-0.026598632335662842,
0.0750860944390297,
0.015671469271183014,
0.09307931363582611,
-0.06066211313009262,
-0.007468897849321365,
0.05833631008863449,
0.08894209563732147,
-0.024913450703024864,
-0.07227443158626556,
-0.01382515486329794,
-0.06491290032863617,
0.010321084409952164,
0.0117538096383214,
0.025452544912695885,
-0.014749618247151375,
0.02722388692200184,
0.02092043310403824,
0.0036056702956557274,
0.05332312732934952,
0.018238812685012817,
0.023281298577785492,
-0.06929737329483032,
0.09243510663509369,
-0.014482080936431885,
-0.02569723315536976,
0.033835507929325104,
-0.017347579821944237,
-0.03699730709195137,
-0.008132407441735268,
-0.00017214110994245857,
0.07867299020290375,
-0.0037751442287117243,
0.0446607731282711,
0.022955510765314102,
0.03065556101500988,
0.07520650327205658,
0.11598118394613266,
-0.04371853917837143,
0.03579021617770195,
-0.019882017746567726,
0.02720373496413231,
0.021067896857857704,
0.0552949421107769,
0.020142771303653717,
-0.010133018717169762,
0.016757819801568985,
-0.016949506476521492,
-0.06748849898576736,
-0.01692459173500538,
0.04557386785745621,
0.0945446565747261,
-0.013044411316514015,
0.031703826040029526,
-0.005805314518511295,
0.042714256793260574,
-0.06466303020715714,
-0.0025410810485482216,
-0.003957793116569519,
9.626418949181341e-34,
0.09688735753297806,
0.023899691179394722,
-0.02171938866376877,
-0.05812634527683258,
0.003664975054562092,
0.0688803568482399,
-0.08411721140146255,
0.05244065821170807,
-0.017441757023334503,
0.023822667077183723,
-0.014390750788152218,
0.018666811287403107,
0.002187456004321575,
0.07399647682905197,
0.05670576170086861,
0.04559582471847534,
0.017749212682247162,
0.07056161016225815,
0.11256439238786697,
-0.0984833613038063,
0.1261693239212036,
-0.03546983748674393,
0.0010008711833506823,
0.08839447051286697,
0.07137839496135712,
-0.01579669490456581,
0.05878124013543129,
0.01319859828799963,
0.08745444566011429,
0.0035209436900913715,
-0.10523334890604019,
0.03597346320748329,
0.015273530036211014,
0.052191443741321564,
0.007297285832464695,
0.004078451544046402,
-0.055650029331445694,
-0.03398313745856285,
0.029467744752764702,
-0.017569858580827713,
0.019098907709121704,
0.022630956023931503,
0.01269183773547411,
-0.044276028871536255,
-0.06981804966926575,
0.011362116783857346,
-0.04684232920408249,
0.0017989423358812928,
-0.018295902758836746,
-0.04749736189842224,
-0.0022746145259588957,
-0.026415156200528145,
-0.0369446724653244,
0.00033978073042817414,
-0.06710293889045715,
-0.029320016503334045,
0.05920780450105667,
0.05094769224524498,
0.009982731193304062,
0.02847048081457615,
0.018823182210326195,
0.0731540322303772,
0.04140384867787361,
-0.02305964007973671,
-0.010210194624960423,
-0.011648481711745262,
0.004891043994575739,
0.018030494451522827,
0.08038678020238876,
-0.055730823427438736,
0.016729485243558884,
0.0582650788128376,
-0.02030939795076847,
-0.043142493814229965,
-0.015059908851981163,
-0.008632324635982513,
0.00848039798438549,
-0.01701010949909687,
-0.024944987148046494,
0.015601078048348427,
-0.013609707355499268,
-0.01420378778129816,
-0.015053709968924522,
0.011741414666175842,
-0.06262723356485367,
-0.07364823669195175,
0.009415028616786003,
0.049944400787353516,
0.00864444486796856,
0.031420521438121796,
0.007032924797385931,
-0.052493639290332794,
-0.02270434983074665,
0.01580556482076645,
-0.04134434461593628,
-1.2135876227931261e-33,
-0.03716485574841499,
0.052132152020931244,
-0.04167677089571953,
0.04091800004243851,
0.07677780091762543,
-0.05652887374162674,
-0.0723625048995018,
0.021578848361968994,
-0.03148436173796654,
0.04704860970377922,
-0.08813463151454926,
-0.04620736837387085,
0.08899220824241638,
-0.0878729522228241,
0.06458429992198944,
0.04392705485224724,
-0.023868300020694733,
0.0005858539952896535,
-0.1658753603696823,
0.03723152354359627,
-0.028989264741539955,
0.08714063465595245,
-0.1136481985449791,
0.05481665953993797,
0.09901619702577591,
-0.020048584789037704,
-0.015574986115098,
0.010730471462011337,
0.046571481972932816,
0.03970697149634361,
-0.10310913622379303,
-0.032144103199243546,
-0.0591621994972229,
-0.15463940799236298,
-0.012369243428111076,
-0.001163211651146412,
0.04805789515376091,
0.017887013033032417,
0.02191789261996746,
0.0242609903216362,
0.006944486405700445,
-0.0002967044129036367,
-0.12276318669319153,
0.018074041232466698,
-0.04355859011411667,
0.09464410692453384,
-0.09515240788459778,
-0.009621317498385906,
-0.01563280262053013,
0.014314522966742516,
0.0011790848802775145,
0.09357427060604095,
0.039901524782180786,
-0.013176162727177143,
-0.05668368190526962,
-0.08546264469623566,
0.0795765295624733,
0.022913634777069092,
0.04748653247952461,
0.05837269127368927,
-0.015164833515882492,
-0.06238510459661484,
-0.042983341962099075,
-0.004395037889480591,
-0.021900564432144165,
-0.01812773011624813,
-0.012325813993811607,
-0.0420522578060627,
-0.012935863807797432,
0.046302054077386856,
0.0631854310631752,
-0.003607411403208971,
-0.03748365491628647,
-0.02279430814087391,
0.04335023835301399,
0.008023766800761223,
0.07953820377588272,
-0.00385935977101326,
-0.032001953572034836,
0.03059571050107479,
0.04762056842446327,
0.02302873693406582,
0.051741451025009155,
-0.009697864763438702,
0.05028350278735161,
-0.06203026697039604,
0.0640459880232811,
0.005769672803580761,
0.028863605111837387,
0.029957415536046028,
0.008106103166937828,
-0.018521778285503387,
-0.2167540192604065,
0.04369153454899788,
0.018960215151309967,
-4.127106478790665e-8,
-0.0808391124010086,
0.04405400529503822,
-0.011929133906960487,
0.047158438712358475,
0.044270213693380356,
-0.0007102791569195688,
0.055212587118148804,
0.07993236929178238,
-0.013349317014217377,
0.0798492580652237,
0.1393493413925171,
-0.022750305011868477,
0.0075656031258404255,
-0.11601080745458603,
-0.02887631207704544,
0.014318599365651608,
-0.015204629860818386,
-0.08910726755857468,
0.033646322786808014,
-0.011707938276231289,
-0.05368107184767723,
-0.007156179286539555,
-0.08682727068662643,
-0.014208784326910973,
0.006705609615892172,
-0.04565314203500748,
-0.037769440561532974,
-0.03981080651283264,
0.07045619934797287,
0.1287880390882492,
-0.04932870715856552,
-0.03656855225563049,
0.01398006733506918,
-0.049594245851039886,
-0.043196842074394226,
0.16092804074287415,
-0.027935640886425972,
-0.02805328741669655,
-0.014136198908090591,
-0.01680191233754158,
0.08419772982597351,
0.05204230919480324,
-0.034652020782232285,
-0.006718406453728676,
-0.06474719941616058,
0.0006857038242742419,
-0.0060350773856043816,
-0.07963863760232925,
0.06135864555835724,
-0.0011345811653882265,
-0.039168406277894974,
-0.05676009878516197,
0.04051235318183899,
0.05036207288503647,
0.03615174815058708,
0.0020346851088106632,
0.000298769649816677,
-0.09762359410524368,
0.07904062420129776,
0.06711827963590622,
0.06952536851167679,
-0.01005373802036047,
-0.012964682653546333,
-0.02258998714387417
] |
Attention with Markov: A Curious Case of Single-layer Transformers | https://openreview.net/forum?id=SqZ0KY4qBD | [
"Ashok Vardhan Makkuva",
"Marco Bondaschi",
"Adway Girish",
"Alliot Nagle",
"Martin Jaggi",
"Hyeji Kim",
"Michael Gastpar"
] | Spotlight | Attention-based transformers have achieved tremendous success across a variety of disciplines including natural languages. To deepen our understanding of their sequential modeling capabilities, there is a growing interest in using Markov input processes to study them. A key finding is that when trained on first-order Markov chains, transformers with two or more layers consistently develop an induction head mechanism to estimate the in-context bigram conditional distribution. In contrast, single-layer transformers, unable to form an induction head, directly learn the Markov kernel but often face a surprising challenge: they become trapped in local minima representing the unigram distribution, whereas deeper models reliably converge to the ground-truth bigram. While single-layer transformers can theoretically model first-order Markov chains, their empirical failure to learn this simple kernel in practice remains a curious phenomenon. To explain this contrasting behavior of single-layer models, in this paper we introduce a new framework for a principled analysis of transformers via Markov chains. Leveraging our framework, we theoretically characterize the loss landscape of single-layer transformers and show the existence of global minima (bigram) and bad local minima (unigram) contingent on data properties and model architecture. We precisely delineate the regimes under which these local optima occur. Backed by experiments, we demonstrate that our theoretical findings are in congruence with the empirical results. Finally, we outline several open problems in this arena. | Markov chains, Transformers, Optimization, Landscape | We theoretically characterize the loss landscape of single-layer transformers and show the existence of global minima (bigram) and bad local minima (unigram) contingent upon the specific data characteristics and the transformer architecture. | 9,590 | null | [
-0.0763641968369484,
-0.012081832624971867,
0.06201443448662758,
-0.012520947493612766,
0.08317795395851135,
0.04939410462975502,
-0.04794789478182793,
0.00007604635902680457,
0.10124155133962631,
-0.05142589658498764,
-0.020328596234321594,
-0.029005063697695732,
0.043038491159677505,
0.0036056770477443933,
0.01897544413805008,
-0.028066007420420647,
0.047223880887031555,
0.06395961344242096,
-0.03549940511584282,
-0.10816260427236557,
0.043076977133750916,
0.08376374840736389,
-0.0027594948187470436,
-0.016589893028140068,
0.015688739717006683,
0.03980623558163643,
-0.014754300005733967,
-0.0528787225484848,
0.0059033045545220375,
-0.03603431582450867,
0.0034206255804747343,
-0.004358978010714054,
-0.01614448055624962,
0.003582105040550232,
0.01742134615778923,
0.117337167263031,
-0.06431631743907928,
-0.008512280881404877,
-0.0520651675760746,
-0.006377991754561663,
0.03925922140479088,
0.03379647806286812,
-0.007110322825610638,
0.043868325650691986,
0.027349000796675682,
-0.03953160345554352,
0.06024741008877754,
-0.07344415038824081,
-0.0361534021794796,
-0.0632995069026947,
-0.00032887584529817104,
0.007372405380010605,
-0.01727958954870701,
0.1338878571987152,
-0.025042425841093063,
0.019090237095952034,
0.01492305938154459,
-0.03818926587700844,
-0.04042280092835426,
-0.00040881475433707237,
-0.12456995993852615,
-0.07982662320137024,
-0.006443632300943136,
-0.07364091277122498,
0.012922278605401516,
0.010500130243599415,
0.0451766699552536,
0.037897612899541855,
-0.05279496684670448,
0.04953473433852196,
-0.08523229509592056,
0.05479865148663521,
0.005023986101150513,
0.03572116419672966,
0.04898981750011444,
-0.023215537890791893,
0.04190628230571747,
-0.08814507722854614,
0.02822200581431389,
-0.0927625522017479,
0.019711775705218315,
-0.00965930987149477,
0.051637303084135056,
-0.042830757796764374,
0.09154535830020905,
0.014155188575387001,
-0.027546098455786705,
0.006856762804090977,
0.04643857106566429,
-0.015440079383552074,
-0.08402224630117416,
-0.07334066182374954,
0.02998109720647335,
0.06270522624254227,
0.07677273452281952,
0.0505257211625576,
-0.05224548280239105,
-0.05127691850066185,
-0.001481605926528573,
0.08848148584365845,
0.07038957625627518,
-0.02582198940217495,
0.06991071999073029,
-0.04140601307153702,
0.001704343012534082,
-0.03001646138727665,
-0.0612536258995533,
0.02640089951455593,
-0.0017214508261531591,
-0.09270228445529938,
0.04842488840222359,
0.012365546077489853,
0.06440072506666183,
-0.03172138333320618,
-0.009631737135350704,
-0.0560329332947731,
0.04044828191399574,
-0.03315969929099083,
-0.12935443222522736,
0.002689867978915572,
-0.05562638118863106,
0.02018607221543789,
-0.02372223138809204,
0.09113682061433792,
0.004632268566638231,
-0.0063345362432301044,
0.015751104801893234,
2.0727201687881536e-33,
0.018257496878504753,
0.010825511068105698,
0.059895627200603485,
0.03608948737382889,
-0.018152745440602303,
0.02626206912100315,
0.06622256338596344,
-0.005660475231707096,
0.00015792487829457968,
0.04734610766172409,
-0.016461657360196114,
0.06920197606086731,
-0.07006726413965225,
0.031120665371418,
-0.0025095539167523384,
-0.009448516182601452,
-0.010723743587732315,
0.040219444781541824,
0.021454399451613426,
-0.02280435711145401,
0.01243828609585762,
0.04639164358377457,
-0.001529143424704671,
-0.053219024091959,
-0.005326686892658472,
0.0017330521950498223,
0.07744649052619934,
-0.08272942155599594,
-0.0302236657589674,
0.0016388125950470567,
-0.11744320392608643,
0.004112219903618097,
-0.030262678861618042,
0.0304634477943182,
0.02511708252131939,
-0.061156779527664185,
-0.006806582678109407,
-0.029944129288196564,
-0.025858908891677856,
-0.13546359539031982,
-0.06646029651165009,
0.04144139587879181,
-0.0585772842168808,
-0.006764153018593788,
-0.08431081473827362,
-0.08107543736696243,
0.0386313870549202,
-0.027639638632535934,
0.028910821303725243,
-0.08718667179346085,
0.07851427048444748,
-0.058044444769620895,
-0.038343481719493866,
-0.03458778187632561,
0.03970802202820778,
0.03996323421597481,
0.04607608541846275,
0.08133730292320251,
0.05075087770819664,
0.07462145388126373,
0.016739919781684875,
0.035947490483522415,
-0.051725659519433975,
0.05902257561683655,
0.014459710568189621,
0.014589045196771622,
0.028141258284449577,
0.07286045700311661,
0.0538744181394577,
-0.05634110048413277,
-0.003351029707118869,
0.023428160697221756,
0.03401926904916763,
-0.068033367395401,
0.08143550157546997,
0.00669701537117362,
-0.009188773110508919,
-0.052676282823085785,
0.00883167888969183,
0.0695764496922493,
0.001904444769024849,
-0.024176623672246933,
-0.03192705661058426,
0.011919345706701279,
-0.004397393204271793,
0.02841532602906227,
0.035363927483558655,
-0.04404700919985771,
-0.07311709225177765,
0.025281604379415512,
-0.0025049522519111633,
-0.017420632764697075,
-0.00637771375477314,
0.016644932329654694,
0.0093186991289258,
-2.0672882828614008e-33,
-0.052650898694992065,
0.044772204011678696,
-0.040474195033311844,
0.05808283016085625,
-0.05757753551006317,
-0.09384710341691971,
-0.009637740440666676,
0.0744919553399086,
-0.016215423122048378,
-0.004481576848775148,
-0.037118807435035706,
-0.03147753328084946,
0.04118616133928299,
-0.03246123343706131,
0.07252439111471176,
-0.03581821173429489,
0.008523966185748577,
-0.006127084605395794,
0.005955812055617571,
0.0109287453815341,
0.010030295699834824,
0.06738229095935822,
-0.14941661059856415,
0.01836596615612507,
-0.061713527888059616,
0.06064615398645401,
-0.1136707216501236,
0.11897432059049606,
0.03459479287266731,
-0.036000169813632965,
-0.06790249049663544,
-0.014765692874789238,
-0.024178188294172287,
0.08644811809062958,
-0.06929744780063629,
0.025588152930140495,
-0.0026470699813216925,
0.006921145599335432,
-0.046177204698324203,
0.0335422046482563,
0.0029200403951108456,
-0.055612802505493164,
-0.03890001028776169,
-0.00957549549639225,
-0.062256962060928345,
-0.07749570906162262,
-0.04531319439411163,
0.0029267235659062862,
-0.011900928802788258,
-0.029019102454185486,
-0.041842058300971985,
0.08511912077665329,
-0.11281707882881165,
-0.01001915242522955,
-0.06855326890945435,
-0.015902714803814888,
-0.06011319160461426,
-0.027414066717028618,
0.08378671854734421,
-0.022516343742609024,
-0.04990072548389435,
-0.05696592107415199,
-0.010951431468129158,
-0.0477069690823555,
0.012965315021574497,
0.02290167845785618,
-0.011883484199643135,
-0.07252142578363419,
0.09068364650011063,
0.0717175155878067,
0.10388407856225967,
0.07408328354358673,
0.02341950684785843,
-0.02397076040506363,
0.044754695147275925,
0.0359107181429863,
-0.003372360020875931,
-0.07547148317098618,
-0.027583450078964233,
-0.05050761625170708,
-0.03399308770895004,
0.011247116141021252,
-0.019105324521660805,
0.04413990676403046,
0.05586903914809227,
0.06788593530654907,
0.10453363507986069,
-0.014074933715164661,
0.05470173433423042,
-0.019212298095226288,
-0.03705127164721489,
0.07575912028551102,
-0.015725353732705116,
0.08658663928508759,
-0.07754150032997131,
-4.321971402987401e-8,
-0.07929615676403046,
0.04570871591567993,
0.03833117336034775,
-0.02520640566945076,
0.0965113416314125,
-0.035386115312576294,
0.06259133666753769,
0.1081596314907074,
-0.019848987460136414,
0.03875450789928436,
0.04449533671140671,
-0.011412892490625381,
0.025390129536390305,
-0.06461641192436218,
-0.021913111209869385,
0.11428277939558029,
0.034625571221113205,
-0.0026756038423627615,
-0.005092603620141745,
-0.05123632773756981,
0.003627792466431856,
0.10579065978527069,
0.009780416265130043,
0.057966385036706924,
-0.012014003470540047,
-0.03909243270754814,
0.015593011863529682,
0.04631897807121277,
0.053924158215522766,
-0.007911520078778267,
-0.008160147815942764,
0.001684150891378522,
0.0038630045019090176,
-0.03423628583550453,
-0.01836473122239113,
0.12171932309865952,
0.08650782704353333,
0.004509868565946817,
0.021619319915771484,
-0.03747592866420746,
0.00494163203984499,
0.04740159958600998,
-0.09585174173116684,
0.04101872444152832,
0.000757084577344358,
0.032332900911569595,
-0.018899375572800636,
-0.12189631909132004,
0.018860016018152237,
-0.02054271101951599,
0.03323427960276604,
-0.029191000387072563,
-0.055647190660238266,
0.05191513150930405,
0.012039409019052982,
0.010853998363018036,
0.050406377762556076,
-0.07837279886007309,
-0.02832418866455555,
0.10235960781574249,
-0.02622242644429207,
0.05883866921067238,
-0.048853799700737,
0.006488356273621321
] |
The Computational Complexity of Circuit Discovery for Inner Interpretability | https://openreview.net/forum?id=QogcGNXJVw | [
"Federico Adolfi",
"Martina G. Vilas",
"Todd Wareham"
] | Spotlight | Many proposed applications of neural networks in machine learning, cognitive/brain science, and society hinge on the feasibility of inner interpretability via circuit discovery. This calls for empirical and theoretical explorations of viable algorithmic options. Despite advances in the design and testing of heuristics, there are concerns about their scalability and faithfulness at a time when we lack understanding of the complexity properties of the problems they are deployed to solve. To address this, we study circuit discovery with classical and parameterized computational complexity theory: (1) we describe a conceptual scaffolding to reason about circuit finding queries in terms of affordances for description, explanation, prediction and control; (2) we formalize a comprehensive set of queries for mechanistic explanation, and propose a formal framework for their analysis; (3) we use it to settle the complexity of many query variants and relaxations of practical interest on multi-layer perceptrons. Our findings reveal a challenging complexity landscape. Many queries are intractable, remain fixed-parameter intractable relative to model/circuit features, and inapproximable under additive, multiplicative, and probabilistic approximation schemes. To navigate this landscape, we prove there exist transformations to tackle some of these hard problems with better-understood heuristics, and prove the tractability or fixed-parameter tractability of more modest queries which retain useful affordances. This framework allows us to understand the scope and limits of interpretability queries, explore viable options, and compare their resource demands on existing and future architectures. | inner interpretability, mechanistic interpretability, circuit discovery, computational complexity, parameterized complexity | Parameterized complexity of circuit queries explains interpretability outcomes and viable algorithmic options for circuit discovery. | 9,565 | 2410.08025 | [
-0.04574814438819885,
-0.014070602133870125,
0.027562595903873444,
0.07075739651918411,
0.027870438992977142,
-0.04476654529571533,
-0.015175757929682732,
-0.008066060952842236,
-0.005361916963011026,
0.004637585021555424,
-0.08986379206180573,
-0.12274960428476334,
-0.012906487099826336,
0.016424087807536125,
-0.06941448897123337,
0.054599739611148834,
0.007399349007755518,
0.033057086169719696,
-0.0336553230881691,
-0.014830955304205418,
0.048998598009347916,
0.005740732420235872,
-0.011864439584314823,
-0.02507004141807556,
-0.012866285629570484,
0.06445334106683731,
-0.03934437036514282,
-0.04278931766748428,
-0.03523621708154678,
0.006720404606312513,
0.00011396603076718748,
-0.011496301740407944,
0.0024617696180939674,
0.038322947919368744,
0.004154932219535112,
0.08097822964191437,
-0.005582829937338829,
0.03156735375523567,
0.02984161488711834,
-0.030395394191145897,
-0.019527001306414604,
-0.0011487002484500408,
0.037216924130916595,
0.023227781057357788,
0.11210568249225616,
0.009172099642455578,
0.025874817743897438,
-0.07257023453712463,
-0.031219780445098877,
-0.09512799233198166,
-0.07704415172338486,
-0.02699034847319126,
0.01545755285769701,
0.027867788448929787,
0.006123775150626898,
0.020904213190078735,
-0.01301281712949276,
0.06710951775312424,
-0.04401348903775215,
0.0220674816519022,
-0.04360569640994072,
-0.1177864745259285,
0.04982221499085426,
-0.008110436610877514,
0.02134134992957115,
0.04489533230662346,
0.049744293093681335,
0.005431370344012976,
0.047339312732219696,
-0.02168213576078415,
0.011000742204487324,
0.06384249031543732,
-0.07234601676464081,
-0.026523245498538017,
0.0994693711400032,
0.047965433448553085,
0.05439883470535278,
0.023124543949961662,
-0.010293793864548206,
-0.04132414981722832,
-0.04358500987291336,
-0.02186097763478756,
-0.018151624128222466,
0.043895311653614044,
0.12915970385074615,
-0.013886181637644768,
0.008979774080216885,
-0.024409376084804535,
0.026424970477819443,
0.043686069548130035,
-0.11210398375988007,
-0.11755786836147308,
-0.09091338515281677,
-0.08156579732894897,
0.027263188734650612,
0.017886314541101456,
0.0037186595145612955,
-0.057857267558574677,
-0.051456306129693985,
0.04753059148788452,
0.023950278759002686,
0.09602472186088562,
0.06676482409238815,
-0.07927509397268295,
0.0070465258322656155,
0.023730482906103134,
0.04526389762759209,
0.04060489684343338,
0.029044371098279953,
-0.12989723682403564,
-0.0047402167692780495,
0.0613696426153183,
0.023149380460381508,
0.02133157104253769,
0.04988378658890724,
-0.016402559354901314,
0.001150449737906456,
-0.004125294741243124,
0.10280748456716537,
0.0493495874106884,
-0.042385391891002655,
0.025101758539676666,
-0.01719294674694538,
0.019426096230745316,
0.04199912026524544,
-0.03537258133292198,
-0.08969217538833618,
1.2428294223022729e-33,
-0.03844795003533363,
0.0406220406293869,
0.028996875509619713,
0.04357512667775154,
0.0734994187951088,
0.007111560553312302,
0.0005417725187726319,
0.049810584634542465,
0.06608466058969498,
0.04504915699362755,
-0.023554544895887375,
0.04073144868016243,
-0.03627486154437065,
0.012703567743301392,
0.10414256155490875,
-0.005000434350222349,
-0.09294328838586807,
-0.037854183465242386,
-0.02333993650972843,
-0.12986257672309875,
0.0852288231253624,
-0.09355374425649643,
-0.029588332399725914,
-0.03558256849646568,
0.018354464322328568,
-0.021442487835884094,
0.03368837758898735,
0.010852574370801449,
-0.05840509757399559,
-0.023015625774860382,
-0.09306961297988892,
0.05347147956490517,
0.023220865055918694,
0.011269869282841682,
0.017417754977941513,
-0.015172725543379784,
0.05532894656062126,
-0.05943342298269272,
0.08437664806842804,
-0.007081281393766403,
-0.04281502962112427,
-0.01240332517772913,
0.027634140104055405,
-0.009139001369476318,
-0.07442871481180191,
-0.022801300510764122,
-0.00522928312420845,
0.011306524276733398,
-0.023092124611139297,
-0.05133181810379028,
-0.04095753654837608,
-0.04827320948243141,
0.08277271687984467,
-0.08760888874530792,
0.04548807069659233,
0.015963280573487282,
-0.0008378289057873189,
0.045436758548021317,
0.0786081999540329,
0.15823133289813995,
-0.028060758486390114,
0.06161298230290413,
-0.02994469366967678,
0.057125311344861984,
-0.04988465458154678,
0.1118653416633606,
-0.08085906505584717,
0.0639362707734108,
0.07811503857374191,
-0.04193985089659691,
-0.018545610830187798,
0.030093224719166756,
-0.01869446225464344,
-0.10063431411981583,
0.11842893064022064,
-0.01709800958633423,
0.008910110220313072,
-0.05107228830456734,
0.04140634462237358,
0.0010837074369192123,
-0.03570389375090599,
-0.04204889014363289,
-0.026701118797063828,
0.08530818670988083,
-0.033851832151412964,
-0.04432111978530884,
0.041688915342092514,
-0.06542442739009857,
-0.07735038548707962,
-0.018339361995458603,
-0.0862591341137886,
0.03742985054850578,
0.06382542848587036,
-0.03415730223059654,
0.006451520137488842,
-2.629983653275606e-33,
-0.08349782973527908,
-0.057900622487068176,
0.014420191757380962,
0.027755415067076683,
0.02435978688299656,
-0.05930605158209801,
-0.10492338985204697,
-0.03842107579112053,
0.02799052931368351,
-0.052761197090148926,
-0.012293740175664425,
0.0020406392868608236,
0.028964774683117867,
-0.046607501804828644,
-0.007506397552788258,
0.05264182388782501,
-0.09507498890161514,
0.004700645804405212,
0.1060156524181366,
0.07245463132858276,
0.012429486028850079,
0.06503759324550629,
-0.1247783824801445,
-0.008633042685687542,
-0.0535905659198761,
0.08033092319965363,
-0.04559505358338356,
0.0011903116246685386,
-0.0012699693907052279,
0.03940983861684799,
-0.10154977440834045,
-0.024313585832715034,
-0.038335807621479034,
0.006262746173888445,
-0.010484177619218826,
0.06202000379562378,
0.012114237062633038,
-0.011774834245443344,
-0.024253299459815025,
0.026166027411818504,
0.0321517214179039,
-0.0012065119808539748,
-0.0672312006354332,
-0.01573188602924347,
0.031163068488240242,
-0.010484673082828522,
0.009754524566233158,
0.03479459881782532,
0.04956808313727379,
0.03560073673725128,
0.019216496497392654,
0.018379803746938705,
-0.011530349962413311,
-0.056913793087005615,
-0.044495198875665665,
-0.0009948021033778787,
-0.04367559775710106,
0.044519126415252686,
0.15620046854019165,
-0.08551695942878723,
-0.09293406456708908,
-0.04023006930947304,
0.013873089104890823,
0.039175767451524734,
0.010338830761611462,
-0.015343147329986095,
0.002362574217841029,
0.025693152099847794,
0.0643748864531517,
0.021558063104748726,
-0.02733173407614231,
-0.014421016909182072,
0.021920926868915558,
-0.03666143864393234,
-0.01902540773153305,
-0.008313155733048916,
-0.027865225449204445,
-0.0868491679430008,
-0.024196060374379158,
-0.02413034625351429,
0.07223434746265411,
0.0061096316203475,
0.011929038912057877,
-0.04585565999150276,
0.006463861092925072,
0.03555970638990402,
0.041132885962724686,
0.08506190776824951,
0.017911236733198166,
-0.028769569471478462,
-0.027774328365921974,
0.0887223407626152,
-0.0537588894367218,
0.031281664967536926,
-0.08304289728403091,
-4.371660722313209e-8,
0.006113563664257526,
0.005649714730679989,
0.010914056561887264,
-0.03333856910467148,
0.17127597332000732,
-0.06641929596662521,
0.021985720843076706,
0.045103006064891815,
-0.08135267347097397,
0.029036179184913635,
0.1011514663696289,
0.06829141825437546,
0.031547557562589645,
-0.03330644220113754,
-0.016006987541913986,
0.06909333169460297,
-0.017453519627451897,
0.015940073877573013,
-0.02701047994196415,
0.008726331405341625,
-0.029415156692266464,
-0.005638115108013153,
-0.0020268247462809086,
0.03284657374024391,
0.11063080281019211,
-0.10049279779195786,
-0.0529031977057457,
0.029816411435604095,
0.013655691407620907,
0.045574285089969635,
0.024818075820803642,
0.02901238575577736,
0.040914103388786316,
0.02170732244849205,
0.01428063865751028,
0.030409950762987137,
-0.02797248773276806,
-0.05676109716296196,
-0.02595236897468567,
-0.08769289404153824,
-0.07121354341506958,
0.028690729290246964,
-0.032701656222343445,
0.053831104189157486,
0.031341955065727234,
-0.06302862614393234,
0.02285543642938137,
-0.04258595407009125,
-0.04712974280118942,
0.013842524029314518,
0.021906711161136627,
0.0009566738153807819,
-0.029297852888703346,
0.02319631166756153,
0.01169249415397644,
0.05113469064235687,
-0.011930372565984726,
-0.001812477596104145,
-0.03223132714629173,
0.03888693451881409,
0.02238444611430168,
0.08049910515546799,
-0.01493990421295166,
-0.026624545454978943
] |
Topological Schrödinger Bridge Matching | https://openreview.net/forum?id=WzCEiBILHu | [
"Maosheng Yang"
] | Spotlight | Given two boundary distributions, the \emph{Schrödinger Bridge} (SB) problem seeks the “most likely” random evolution between them with respect to a reference process. It has revealed rich connections to recent machine learning methods for generative modeling and distribution matching. While these methods perform well in Euclidean domains, they are not directly applicable to topological domains such as graphs and simplicial complexes, which are crucial for data defined over network entities, such as node signals and edge flows. In this work, we propose the \emph{Topological Schrödinger Bridge problem} ($\mathcal{T}$SBP) for matching signal distributions on a topological domain. We set the reference process to follow some linear tractable \emph{topology-aware} stochastic dynamics such as topological heat diffusion. For the case of Gaussian boundary distributions, we derive a \emph{closed-form} topological SB ($\mathcal{T}$SB) in terms of its time-marginal and stochastic differential. In the general case, leveraging the well-known result, we show that the optimal process follows the forward-backward topological dynamics governed by some unknowns. Building on these results, we develop $\mathcal{T}$SB-based models for matching topological signals by parameterizing the unknowns in the optimal process as \emph{(topological) neural networks} and learning them through \emph{likelihood training}. We validate the theoretical results and demonstrate the practical applications of $\mathcal{T}$SB-based models on both synthetic and real-world networks, emphasizing the role of topology. Additionally, we discuss the connections of $\mathcal{T}$SB-based models to other emerging models, and outline future directions for topological signal matching. | Schrödinger Bridge, Topological Signal Distribution Matching, Topological Stochastic Dynamics, Topological Generative Models | We investigate Schrödinger bridge based models for topological signal matching. | 9,513 | null | [
-0.05957631394267082,
-0.09449761360883713,
0.03655562549829483,
0.01455705612897873,
-0.0074217235669493675,
0.006044524721801281,
0.0010962230153381824,
-0.07267061620950699,
0.03427103906869888,
-0.028278948739171028,
-0.008249755017459393,
-0.0005628680228255689,
0.020509840920567513,
0.0032307810615748167,
0.06553235650062561,
0.03287450224161148,
0.026778340339660645,
-0.039866551756858826,
0.006868938449770212,
-0.009187091141939163,
0.030904412269592285,
0.07424640655517578,
-0.04342461749911308,
-0.04715893045067787,
0.024967383593320847,
0.003787641879171133,
0.08525083214044571,
0.036450937390327454,
0.007612003479152918,
-0.023200716823339462,
0.10642971843481064,
0.01766391098499298,
-0.11203261464834213,
-0.09754481911659241,
0.01521897129714489,
0.03976183757185936,
-0.09041246771812439,
0.08125531673431396,
0.053364627063274384,
-0.05430849641561508,
0.06611611694097519,
0.03140846639871597,
0.03371609002351761,
0.07035595178604126,
0.019157379865646362,
0.05156463384628296,
-0.04630677402019501,
-0.08702352643013,
-0.017143964767456055,
-0.0822799801826477,
0.03329693153500557,
0.008060432970523834,
-0.014745960012078285,
-0.010004411451518536,
0.05183012783527374,
0.07823125272989273,
0.05402076616883278,
-0.02085014246404171,
0.0016251378692686558,
-0.02090613730251789,
0.036316048353910446,
-0.08221517503261566,
-0.039525777101516724,
-0.00569114601239562,
-0.006667096167802811,
-0.027830764651298523,
0.026578964665532112,
0.11730559915304184,
0.036507707089185715,
0.013905465602874756,
-0.013216829858720303,
0.0648251622915268,
-0.03410852700471878,
0.0486147440969944,
0.08481280505657196,
0.005196497309952974,
-0.03671984747052193,
0.023073093965649605,
-0.04794229194521904,
-0.022862091660499573,
-0.036676254123449326,
-0.042015839368104935,
0.0331878624856472,
0.033891115337610245,
0.005660654976963997,
0.04388595372438431,
-0.11651773005723953,
0.016041357070207596,
0.0012049759970977902,
-0.010678417049348354,
-0.06829938292503357,
0.03610076382756233,
-0.03450581058859825,
-0.062455274164676666,
0.01626528985798359,
0.027567485347390175,
-0.00590894790366292,
-0.042304787784814835,
0.0695139616727829,
0.061951227486133575,
-0.04149208590388298,
-0.03393935784697533,
0.06871970742940903,
0.031232459470629692,
0.059097934514284134,
-0.03826725110411644,
-0.005744109395891428,
0.0407218262553215,
0.04586106166243553,
-0.0916001945734024,
0.04104294255375862,
-0.03947284445166588,
0.04591154307126999,
0.03603459522128105,
-0.047874342650175095,
-0.06352182477712631,
0.07250497490167618,
-0.008241066709160805,
0.014108029194176197,
0.07185159623622894,
-0.0837327390909195,
0.04547721892595291,
-0.09282689541578293,
-0.0035505201667547226,
-0.0335199199616909,
0.0006070590461604297,
0.0483439676463604,
7.766833263106042e-35,
0.0900910347700119,
-0.020238425582647324,
-0.0333402194082737,
-0.05053987354040146,
0.16364699602127075,
0.012828672304749489,
-0.000275243102805689,
-0.020375484600663185,
0.05021696910262108,
0.04610961303114891,
-0.07138781249523163,
0.00450462382286787,
-0.06916828453540802,
-0.03669878467917442,
-0.0009509812807664275,
0.018855730071663857,
-0.07971958816051483,
0.023442214354872704,
0.003418870037421584,
-0.043271973729133606,
0.08882920444011688,
-0.004179174546152353,
-0.029380979016423225,
0.05012085661292076,
-0.07937926054000854,
-0.01687910407781601,
0.003476677695289254,
-0.07640472054481506,
0.03407086059451103,
0.016195695847272873,
-0.0751376822590828,
0.08975744992494583,
-0.02420371025800705,
0.03694973886013031,
0.0708993598818779,
0.024538211524486542,
0.015857180580496788,
0.015713050961494446,
-0.11310823261737823,
-0.08800730854272842,
-0.030452104285359383,
-0.006820760667324066,
-0.010182074271142483,
0.05976102128624916,
-0.017554158344864845,
-0.09990300238132477,
0.04749634489417076,
-0.06059839203953743,
0.03224760666489601,
-0.07083351910114288,
0.04315304011106491,
-0.07269306480884552,
-0.04894205927848816,
0.018419522792100906,
-0.03393194079399109,
0.010251577012240887,
0.02337559126317501,
0.002905540633946657,
0.03039056621491909,
0.1150176078081131,
0.03055165708065033,
0.021540949121117592,
0.0015443883603438735,
0.014164690859615803,
0.039354387670755386,
0.05371784418821335,
0.0051767234690487385,
-0.06938287615776062,
0.07788794487714767,
-0.04881685599684715,
0.04541812092065811,
0.042032431811094284,
-0.0732307955622673,
-0.05863334983587265,
0.06001969426870346,
-0.02838665060698986,
-0.02404269017279148,
0.005772316362708807,
-0.019137687981128693,
0.07666095346212387,
-0.0244740042835474,
-0.062221258878707886,
-0.05670492723584175,
0.05457574874162674,
0.06675383448600769,
-0.11761593073606491,
-0.0034846134949475527,
-0.11899417638778687,
-0.07391158491373062,
-0.03702659532427788,
-0.08259817212820053,
-0.022188615053892136,
0.03689708560705185,
0.09914664924144745,
0.0012640261556953192,
-1.871613840184592e-33,
-0.025809863582253456,
0.033834341913461685,
0.07384857535362244,
0.03505270928144455,
-0.04587814211845398,
-0.007101775612682104,
-0.004686400294303894,
-0.028083879500627518,
0.02801041305065155,
-0.005779293831437826,
-0.04708315059542656,
0.019581973552703857,
0.06261245161294937,
-0.0758703425526619,
-0.04367905855178833,
-0.06902243942022324,
0.08885028213262558,
-0.021656610071659088,
-0.010996957309544086,
0.0004398493329063058,
-0.032463908195495605,
0.05097785219550133,
-0.08251398801803589,
-0.016583357006311417,
0.01100992038846016,
0.08071912080049515,
0.025195494294166565,
0.038048382848501205,
-0.01633562706410885,
0.031969498842954636,
-0.04491109028458595,
-0.052573636174201965,
-0.07154190540313721,
0.0025142119266092777,
-0.05561172962188721,
0.13832125067710876,
0.037643782794475555,
0.07506909966468811,
0.00811254046857357,
0.017998892813920975,
0.05572075396776199,
-0.02952353097498417,
-0.06683468818664551,
0.03906233236193657,
0.035616252571344376,
-0.022298092022538185,
-0.0781887099146843,
-0.04624880105257034,
-0.008307326585054398,
-0.025476055219769478,
0.007577130571007729,
0.08751959353685379,
-0.05905855819582939,
-0.03943539038300514,
-0.08137772232294083,
-0.022291412577033043,
-0.02175336889922619,
0.06498578935861588,
-0.002681960351765156,
0.07061193138360977,
-0.01861434057354927,
-0.043863359838724136,
0.02323482744395733,
0.11811886727809906,
-0.02061392553150654,
-0.0319204218685627,
-0.038161750882864,
-0.04073761776089668,
0.015461189672350883,
0.039416491985321045,
-0.07330186665058136,
-0.00033185366191901267,
-0.04465121775865555,
-0.01640230603516102,
0.10607016086578369,
-0.06766355782747269,
0.0663580596446991,
0.010194968432188034,
0.00817836169153452,
-0.02316652052104473,
-0.032367922365665436,
0.03874046728014946,
0.0017935362411662936,
-0.009905020706355572,
0.0995156392455101,
-0.013325958512723446,
0.014240479096770287,
-0.030702998861670494,
0.05037551000714302,
-0.04130126163363457,
0.016009576618671417,
0.0021571717225015163,
-0.055366549640893936,
0.018487878143787384,
-0.06295912712812424,
-4.316321522424005e-8,
-0.0657259151339531,
-0.015721255913376808,
0.04769520089030266,
-0.06348281353712082,
0.0714353546500206,
0.025857770815491676,
0.060315921902656555,
0.021576836705207825,
-0.010553242638707161,
-0.030647654086351395,
0.0008836049819365144,
-0.010631442070007324,
-0.08796817809343338,
-0.029269373044371605,
-0.057934895157814026,
0.0538894385099411,
-0.02265818417072296,
-0.027113694697618484,
0.018827013671398163,
-0.060288917273283005,
-0.053246524184942245,
0.0018568746745586395,
0.031081119552254677,
-0.006775822024792433,
-0.000711820728611201,
-0.049859266728162766,
-0.060963716357946396,
-0.08783324807882309,
0.0764295756816864,
0.014380398206412792,
-0.0684937909245491,
-0.006857970263808966,
0.036754216998815536,
0.06288827955722809,
0.031395453959703445,
0.01835598237812519,
0.01956038549542427,
-0.03903849795460701,
-0.04615350812673569,
0.03171736001968384,
0.02499217353761196,
0.09236196428537369,
-0.008639151230454445,
-0.005079069174826145,
0.03140329569578171,
0.08913110196590424,
0.11878938972949982,
0.1278526484966278,
0.09159450978040695,
0.056511394679546356,
0.028803369030356407,
0.02534867450594902,
0.013881567865610123,
-0.024330172687768936,
0.003561069257557392,
0.01545676589012146,
0.007920053787529469,
-0.10996414721012115,
0.036930423229932785,
-0.005457709077745676,
-0.01322928350418806,
0.00417093513533473,
-0.06916846334934235,
-0.004735860973596573
] |
ThunderKittens: Simple, Fast, and Adorable Kernels | https://openreview.net/forum?id=0fJfVOSUra | [
"Benjamin Frederick Spector",
"Simran Arora",
"Aaryan Singhal",
"Arjun Parthasarathy",
"Daniel Y Fu",
"Christopher Re"
] | Spotlight | The challenge of mapping AI architectures to GPU hardware is creating a critical bottleneck in AI progress. Despite substantial efforts, hand-written custom kernels fail to meet their theoretical performance thresholds, even on well-established operations like linear attention. The diverse capabilities of GPUs suggests we might we need a wide variety of techniques to achieve high performance. However, our work explores if a small number of key abstractions can drastically simplify the process. We present ThunderKittens (TK), a framework for writing performant AI kernels while remaining easy to use. Our abstractions map to the three levels of the GPU hierarchy: (1) at the warp-level, we provide 16x16 matrix tiles as basic data structures and PyTorch-like operations, (2) at the thread-block level, we provide templates for asynchronously overlapping operations, and (3) at the grid-level, TK helps hide block launch, tear-down, and memory costs. We show the value of TK by providing simple & diverse kernels that match or outperform prior art. We match CuBLAS and FlashAttention-3 on GEMM and attention inference performance and outperform the strongest baselines by $10-40$\% on attention backwards, $8\times$ on state space models, and $14\times$ on linear attention. | Systems, Kernels, Efficiency, Efficient Models, IO Awareness, GPUs | ThunderKittens (TK) is a framework that simplifies the creation of high-performance AI kernels through key abstractions, enabling efficient implementation of ML architectures on GPU hardware and surpassing previous approaches in hardware utilization. | 9,493 | 2410.20399 | [
-0.05681906268000603,
-0.007247030269354582,
-0.008467723615467548,
0.03016752004623413,
0.03839721903204918,
-0.10951179265975952,
-0.05464699864387512,
-0.007211415562778711,
-0.029725315049290657,
-0.0568610243499279,
-0.0890832245349884,
-0.041845958679914474,
-0.07550720125436783,
0.013120387680828571,
-0.01659892126917839,
0.01794888637959957,
0.07787720859050751,
0.05882914364337921,
-0.0341087281703949,
0.00466347299516201,
-0.02010328508913517,
-0.09392363578081131,
-0.029660409316420555,
-0.002955265808850527,
-0.000674879178404808,
0.08245599269866943,
-0.0005403012037277222,
-0.10699595510959625,
0.008671673946082592,
0.012657465413212776,
0.03525911271572113,
-0.0014736874727532268,
0.00398428738117218,
0.0810861587524414,
-0.002292904770001769,
0.056675296276807785,
-0.047675102949142456,
-0.04416719079017639,
-0.03506552428007126,
-0.1105988621711731,
0.02554328739643097,
-0.017669739201664925,
-0.036943089216947556,
0.06811677664518356,
0.07300154119729996,
-0.003583786776289344,
0.05640563741326332,
-0.05177885293960571,
0.017207998782396317,
-0.024298997595906258,
-0.058255597949028015,
-0.03759181126952171,
-0.044286079704761505,
-0.00958456564694643,
-0.020441750064492226,
0.010586938820779324,
0.024497099220752716,
-0.075140081346035,
0.06199904531240463,
0.023060807958245277,
0.026921803131699562,
-0.08915995061397552,
0.035703614354133606,
-0.0037648139987140894,
-0.034746136516332626,
-0.02524261362850666,
0.08549445122480392,
0.004079488571733236,
0.06780246645212173,
0.013402269221842289,
0.039426080882549286,
0.0813964307308197,
-0.019559741020202637,
0.05141628533601761,
-0.03862221911549568,
0.036578383296728134,
0.014085627160966396,
-0.04610154405236244,
0.004654086660593748,
-0.05090845376253128,
0.020379340276122093,
-0.026224078610539436,
0.017485329881310463,
0.04103084281086922,
0.05206221714615822,
-0.03435849770903587,
0.018558865413069725,
0.04720417037606239,
0.03286557272076607,
0.008496171794831753,
0.06539113074541092,
-0.030181968584656715,
0.053610909730196,
-0.03460194170475006,
0.08419342339038849,
-0.015566875226795673,
0.0328495055437088,
-0.1507190465927124,
-0.08054773509502411,
0.08312182128429413,
-0.05970831215381622,
0.005435583181679249,
0.040683336555957794,
0.03304840624332428,
0.0015686404658481479,
0.040723372250795364,
-0.024876657873392105,
0.016350455582141876,
-0.10014384239912033,
-0.02256309799849987,
0.04042065143585205,
0.051547277718782425,
-0.005423109978437424,
0.0028832817915827036,
0.04595187306404114,
-0.039362404495477676,
-0.059404406696558,
0.029652269557118416,
0.11595257371664047,
0.1254374384880066,
-0.0075652445666491985,
-0.007761848624795675,
-0.03426136076450348,
0.05175749585032463,
0.037355031818151474,
-0.0434073805809021,
-0.10322404652833939,
3.4449728048129066e-33,
0.03508575260639191,
0.03893042728304863,
-0.03311869129538536,
0.025506874546408653,
0.10757079720497131,
-0.046293336898088455,
0.05546307936310768,
-0.032677143812179565,
-0.022442763671278954,
0.04019849747419357,
-0.041826941072940826,
0.026641681790351868,
-0.0432267002761364,
0.08644986152648926,
0.09970904141664505,
-0.022714227437973022,
-0.03785499185323715,
0.050137948244810104,
-0.03417353704571724,
-0.0038598477840423584,
0.006371085997670889,
0.030108138918876648,
-0.013788980431854725,
-0.030790964141488075,
0.012232107110321522,
0.033740583807229996,
-0.013602049089968204,
-0.11117717623710632,
0.019056282937526703,
0.033372823148965836,
-0.09583429247140884,
-0.05992107093334198,
-0.033089641481637955,
0.06970168650150299,
0.010298920795321465,
0.002588629489764571,
-0.019673101603984833,
-0.1290631890296936,
0.07242748886346817,
0.004600759595632553,
-0.08019714057445526,
0.012626944109797478,
-0.047928232699632645,
-0.12651391327381134,
-0.09470081329345703,
0.061300430446863174,
0.0317816436290741,
0.02659740298986435,
-0.010050088167190552,
-0.040385156869888306,
0.05225715786218643,
0.036580778658390045,
-0.039366479963064194,
-0.06256595253944397,
0.0008932471973821521,
-0.07423825562000275,
0.02358156442642212,
0.03743381053209305,
0.1575283259153366,
0.16471083462238312,
-0.06080042943358421,
0.01957256905734539,
-0.046229369938373566,
0.07906641066074371,
-0.08051867038011551,
-0.01730213686823845,
-0.004965095315128565,
0.08642660826444626,
-0.015110242180526257,
0.0411636158823967,
-0.082290880382061,
-0.01764966920018196,
-0.007207654882222414,
-0.031043177470564842,
0.05492951720952988,
0.023004913702607155,
0.04846091940999031,
-0.10056465864181519,
-0.036181628704071045,
-0.04086427763104439,
-0.030192503705620766,
0.050052303820848465,
-0.007734790910035372,
-0.018048103898763657,
-0.05296050384640694,
-0.05729075148701668,
-0.051018647849559784,
0.007307101972401142,
-0.0736110657453537,
0.01750328578054905,
-0.11748044192790985,
-0.0823754072189331,
0.09733391553163528,
-0.011024409905076027,
-0.05493093654513359,
-3.016574908915263e-33,
0.03761802241206169,
-0.076055146753788,
-0.05448246747255325,
0.09261593967676163,
-0.009226901456713676,
0.007842215709388256,
0.01847986876964569,
-0.06398384273052216,
-0.010005063377320766,
0.018292879685759544,
-0.08569401502609253,
0.055419210344552994,
0.0069322725757956505,
0.044498782604932785,
0.04476747661828995,
-0.027627797797322273,
-0.003579811193048954,
0.024597367271780968,
-0.06964684277772903,
-0.026301659643650055,
-0.02672019973397255,
0.10909032076597214,
-0.005084410309791565,
-0.011865288950502872,
-0.05020638182759285,
0.06420601159334183,
-0.03738084062933922,
0.019517505541443825,
0.02377396821975708,
0.003183522494509816,
-0.00704484386369586,
0.0021571123506873846,
0.037496183067560196,
0.03492967039346695,
0.055553726851940155,
0.054956261068582535,
0.05361820384860039,
0.001438357518054545,
-0.0023389800917357206,
0.03397154435515404,
0.129114031791687,
-0.02589922398328781,
-0.0001541103993076831,
-0.001689313561655581,
-0.03270978853106499,
0.08546262979507446,
-0.07935818284749985,
0.025784973055124283,
-0.050210993736982346,
-0.042997006326913834,
-0.035255059599876404,
0.06093919649720192,
-0.038413289934396744,
-0.00535999471321702,
-0.015531927347183228,
-0.09713795781135559,
-0.014867620542645454,
0.00531901977956295,
0.024149294942617416,
0.007362294010818005,
-0.02559308148920536,
-0.08146104961633682,
0.034077540040016174,
-0.061054445803165436,
0.03936169296503067,
-0.03181534633040428,
0.0077955941669642925,
-0.059662025421857834,
-0.053921982645988464,
0.051055703312158585,
0.00363510032184422,
-0.03891434147953987,
-0.01313333585858345,
0.0007041512872092426,
-0.02275272086262703,
0.08535227924585342,
0.041918959468603134,
-0.056706465780735016,
0.054416317492723465,
-0.014748684130609035,
-0.02416123077273369,
0.03320088982582092,
0.03981954976916313,
0.040723130106925964,
0.01914927549660206,
0.07096151262521744,
0.022132670506834984,
0.048548582941293716,
0.06397069245576859,
0.03592723608016968,
-0.027610640972852707,
0.055974170565605164,
0.021845534443855286,
0.06874044239521027,
-0.001150096533820033,
-4.696048705454814e-8,
0.008825760334730148,
0.04996151477098465,
0.03148505464196205,
0.04559408873319626,
-0.0015058295102789998,
-0.07368957251310349,
0.08337756991386414,
0.07494984567165375,
-0.01429994497448206,
-0.015152888372540474,
0.11528245359659195,
-0.12331817299127579,
-0.00288413860835135,
-0.038216471672058105,
0.06661141663789749,
0.035575494170188904,
-0.0037830842193216085,
-0.05124478042125702,
-0.01403819303959608,
-0.04738778620958328,
-0.008485049940645695,
0.055160991847515106,
0.02909630723297596,
0.004340897314250469,
-0.051319755613803864,
-0.03377451002597809,
0.052042149007320404,
0.011614891700446606,
0.05521508678793907,
0.03896879777312279,
-0.014066344127058983,
-0.011207296513020992,
0.017313536256551743,
0.04276922345161438,
0.033492740243673325,
-0.013405028730630875,
0.028038958087563515,
0.04720795154571533,
0.049540624022483826,
0.019666941836476326,
-0.002575257560238242,
0.0020034192129969597,
-0.0037572437431663275,
-0.003760391613468528,
-0.0067809163592755795,
0.03374907746911049,
-0.09595184773206711,
-0.05657698214054108,
-0.031626731157302856,
0.03647137060761452,
-0.006464627105742693,
0.01785106584429741,
-0.06148514524102211,
0.1085745096206665,
0.021015480160713196,
0.012003462761640549,
-0.05267225578427315,
-0.05366840958595276,
0.05254263058304787,
0.034293580800294876,
0.03583883121609688,
-0.04685834422707558,
-0.04065471887588501,
-0.009749635122716427
] |
gRNAde: Geometric Deep Learning for 3D RNA inverse design | https://openreview.net/forum?id=lvw3UgeVxS | [
"Chaitanya K. Joshi",
"Arian Rokkum Jamasb",
"Ramon Viñas Torné",
"Charles Harris",
"Simon V Mathis",
"Alex Morehead",
"Rishabh Anand",
"Pietro Lio"
] | Spotlight | Computational RNA design tasks are often posed as inverse problems, where sequences are designed based on adopting a single desired secondary structure without considering 3D conformational diversity. We introduce gRNAde, a geometric RNA design pipeline operating on 3D RNA backbones to design sequences that explicitly account for structure and dynamics. gRNAde uses a multi-state Graph Neural Network and autoregressive decoding to generates candidate RNA sequences conditioned on one or more 3D backbone structures where the identities of the bases are unknown. On a single-state fixed backbone re-design benchmark of 14 RNA structures from the PDB identified by Das et al. (2010), gRNAde obtains higher native sequence recovery rates (56% on average) compared to Rosetta (45% on average), taking under a second to produce designs compared to the reported hours for Rosetta. We further demonstrate the utility of gRNAde on a new benchmark of multi-state design for structurally flexible RNAs, as well as zero-shot ranking of mutational fitness landscapes in a retrospective analysis of a recent ribozyme. Experimental wet lab validation on 10 different structured RNA backbones finds that gRNAde has a success rate of 50% at designing pseudoknotted RNA structures, a significant advance over 35% for Rosetta. Open source code and tutorials are available at: github.com/chaitjo/geometric-rna-design | RNA Structure, RNA Design, Geometric Deep Learning, Graph Neural Networks | GNN-based, wet-lab validated 3D RNA design pipeline; obtains SOTA performance for single-state, multi-state design, mutant fitness ranking. | 9,487 | 2305.14749 | [
-0.11082303524017334,
-0.08978942781686783,
-0.00751268258318305,
-0.057957980781793594,
-0.05639392510056496,
-0.022141795605421066,
-0.14653265476226807,
-0.017270293086767197,
-0.03079194203019142,
-0.03918030858039856,
-0.10808862745761871,
-0.017997419461607933,
-0.06415078043937683,
0.07833513617515564,
-0.06500158458948135,
0.026372354477643967,
-0.03593705594539642,
-0.0027535154949873686,
0.003149600001052022,
0.010217597708106041,
0.00312167895026505,
0.05426839366555214,
0.011054468341171741,
-0.005824207328259945,
0.05118165537714958,
0.00995084922760725,
0.047612205147743225,
-0.0036266986280679703,
-0.03387463837862015,
-0.06387051194906235,
0.10695765167474747,
-0.038999032229185104,
0.00953956600278616,
-0.03732350841164589,
-0.004338556434959173,
0.07616789638996124,
-0.0549442283809185,
-0.04306197538971901,
0.04665611684322357,
0.04151871055364609,
0.008859763853251934,
0.07544448226690292,
0.051177021116018295,
0.03486870601773262,
0.08009538799524307,
-0.010656075552105904,
-0.029623419046401978,
-0.007176781073212624,
0.017551811411976814,
0.012996436096727848,
0.011362073943018913,
-0.027912506833672523,
-0.06860578805208206,
-0.007686718832701445,
0.09245559573173523,
0.09264066815376282,
0.00037108123069629073,
-0.01567268744111061,
-0.014425198547542095,
-0.05419083312153816,
0.08244676142930984,
-0.08569687604904175,
-0.017030086368322372,
-0.06027740612626076,
0.032593075186014175,
0.026295870542526245,
0.029024848714470863,
0.036477331072092056,
0.009682189673185349,
0.043952565640211105,
-0.007705893833190203,
0.03283943235874176,
-0.08265566825866699,
0.07050245255231857,
0.04083872586488724,
0.03865104168653488,
0.058905135840177536,
0.011067124083638191,
0.04952450469136238,
-0.021924301981925964,
-0.04568321630358696,
0.030411649495363235,
0.05808586627244949,
0.015619519166648388,
0.030129844322800636,
0.03573838621377945,
-0.06075551360845566,
0.10780992358922958,
0.023834751918911934,
0.013569105416536331,
0.03598981723189354,
-0.025315746665000916,
0.028648262843489647,
-0.059675879776477814,
-0.03939273953437805,
0.0590636320412159,
0.03380506485700607,
-0.06258788704872131,
0.011632693000137806,
0.005960690323263407,
0.010829368606209755,
0.05331768840551376,
0.07067809998989105,
0.011134324595332146,
-0.006480901036411524,
0.0017214439576491714,
0.09892125427722931,
0.0379219613969326,
-0.017574679106473923,
-0.04671814292669296,
0.018260419368743896,
0.06443920731544495,
0.06181173026561737,
0.04988874867558479,
0.008777681738138199,
-0.016677234321832657,
-0.05841248854994774,
-0.07005822658538818,
0.03462260216474533,
0.09420840442180634,
-0.07843083143234253,
-0.010113188996911049,
-0.03141014277935028,
0.01879633404314518,
-0.05869468301534653,
-0.02748161554336548,
-0.0733867660164833,
2.510788342431234e-33,
0.004155092407017946,
0.06614813208580017,
0.04032604768872261,
0.0009566976805217564,
0.07781991362571716,
-0.043735213577747345,
-0.018075179308652878,
-0.0029148103203624487,
-0.09178190678358078,
0.031285855919122696,
-0.11537313461303711,
-0.03242550045251846,
-0.031059298664331436,
0.06676401197910309,
-0.02826574072241783,
0.003209744580090046,
0.0009975371649488807,
0.027900515124201775,
-0.1410759538412094,
-0.03624381124973297,
0.011548864655196667,
0.010132042691111565,
-0.013384543359279633,
-0.029061784967780113,
0.007941857911646366,
-0.01074195746332407,
0.0384005643427372,
0.005157453939318657,
-0.06342414766550064,
-0.012895994819700718,
-0.08410295099020004,
0.014541125856339931,
-0.04107361659407616,
0.11425164341926575,
-0.0038007183466106653,
-0.03925099968910217,
0.02711373008787632,
-0.07398540526628494,
-0.005427919328212738,
0.023446258157491684,
0.0246691033244133,
0.059514425694942474,
0.019700516015291214,
0.0295969620347023,
-0.04139789566397667,
0.015765126794576645,
0.06822267174720764,
-0.02070706896483898,
-0.0764315128326416,
-0.0712398961186409,
0.07640102505683899,
0.041257344186306,
-0.024260805919766426,
-0.024914629757404327,
0.01339595764875412,
-0.016027817502617836,
-0.04452717304229736,
0.013507803902029991,
0.10809338092803955,
0.07697267085313797,
0.037859875708818436,
0.01601642370223999,
0.004166750703006983,
0.022561347112059593,
-0.009001584723591805,
0.05923931300640106,
-0.037719130516052246,
-0.012622129172086716,
0.052653636783361435,
0.08722462505102158,
0.011912224814295769,
-0.0011057130759581923,
0.016356220468878746,
-0.03992215171456337,
0.050555676221847534,
-0.009989042766392231,
0.07362571358680725,
-0.10497567057609558,
-0.04063042998313904,
0.07593558728694916,
-0.06163511052727699,
0.03500712662935257,
-0.020624946802854538,
-0.023902561515569687,
-0.030702054500579834,
-0.11663097888231277,
0.1167428269982338,
-0.07638750970363617,
-0.00894643273204565,
0.02481764368712902,
-0.03962549939751625,
-0.09838611632585526,
-0.014596854336559772,
-0.06874047219753265,
0.0791880339384079,
-2.6341504134072865e-33,
0.04230966791510582,
-0.007232590112835169,
-0.03351166471838951,
-0.047085657715797424,
-0.042013559490442276,
-0.0055449362844228745,
-0.05551431328058243,
-0.05272766575217247,
-0.046709638088941574,
-0.07061786949634552,
0.01396912895143032,
0.0086735300719738,
0.07442103326320648,
-0.023589059710502625,
0.030962172895669937,
0.023322908207774162,
-0.08314245939254761,
-0.06542310863733292,
0.010055594146251678,
-0.019179295748472214,
0.07246048748493195,
0.05267777666449547,
-0.0974239856004715,
0.02007693238556385,
-0.020599903538823128,
0.01464171800762415,
-0.0533938854932785,
0.02324298955500126,
0.013402123004198074,
0.017919305711984634,
-0.04388956353068352,
-0.046479251235723495,
-0.05660765618085861,
0.07119584083557129,
-0.04274671897292137,
0.09169047325849533,
-0.0003524508501868695,
0.004540545400232077,
0.06375479698181152,
-0.05657396838068962,
-0.03057684190571308,
0.030036527663469315,
-0.07397909462451935,
0.12195771932601929,
0.03434230014681816,
0.05984405800700188,
-0.01586860604584217,
-0.017449142411351204,
0.036481697112321854,
0.03224455565214157,
0.008421815000474453,
0.040240224450826645,
0.04402990639209747,
0.010261351242661476,
-0.021472521126270294,
-0.11224734038114548,
-0.03160041198134422,
-0.012183409184217453,
0.05273193493485451,
0.02043810486793518,
-0.08667110651731491,
-0.05839455500245094,
-0.023785782977938652,
-0.031625133007764816,
-0.025928637012839317,
0.05154531076550484,
0.011317530646920204,
-0.0008564916206523776,
0.018919022753834724,
-0.006162455305457115,
-0.002345288172364235,
-0.05190486088395119,
0.06440926343202591,
0.06792855262756348,
0.06718085706233978,
-0.05850852653384209,
-0.06099560484290123,
0.001210982445627451,
-0.0933275818824768,
-0.016464123502373695,
-0.023461291566491127,
-0.006843490060418844,
-0.036831531673669815,
0.1120748296380043,
0.08342938870191574,
-0.015472926199436188,
0.04870058596134186,
0.08209013938903809,
0.004544347990304232,
0.05160512775182724,
0.006692415568977594,
0.050945404917001724,
-0.06621410697698593,
0.0587984062731266,
0.09072521328926086,
-3.966519912523836e-8,
-0.06208391860127449,
0.02349299006164074,
0.09004230052232742,
-0.034987833350896835,
0.03609120100736618,
0.0732126533985138,
0.010900174267590046,
0.07202252000570297,
0.015636015683412552,
-0.04217403382062912,
0.06562186032533646,
-0.04743479937314987,
-0.03608822077512741,
-0.013779123313724995,
0.034889888018369675,
0.0891231968998909,
0.010788777843117714,
-0.034308385103940964,
-0.0019859387539327145,
0.033626262098550797,
-0.07303013652563095,
0.12060906738042831,
-0.02435629442334175,
-0.01995817944407463,
-0.06698974967002869,
-0.029038138687610626,
-0.06388776749372482,
-0.015443435870110989,
0.05051908269524574,
-0.011083918623626232,
0.021857015788555145,
0.03881686553359032,
0.10172832757234573,
-0.00815918855369091,
-0.09683028608560562,
0.04997866228222847,
0.08070650696754456,
0.008071578107774258,
0.007662680000066757,
-0.01356717012822628,
0.0349123515188694,
0.05395831912755966,
-0.054811786860227585,
-0.03174176812171936,
-0.10374275594949722,
-0.012883899733424187,
-0.014306134544312954,
0.004659403581172228,
0.022342819720506668,
-0.004101394210010767,
0.03171100094914436,
-0.0019218111410737038,
0.005902916193008423,
0.0064981491304934025,
-0.014886570163071156,
0.062006108462810516,
-0.062283288687467575,
-0.09198690950870514,
0.030319439247250557,
0.05745093524456024,
-0.0047068665735423565,
-0.09513070434331894,
0.024964788928627968,
-0.06989525258541107
] |
A Second-Order Perspective on Model Compositionality and Incremental Learning | https://openreview.net/forum?id=OZVTqoli2N | [
"Angelo Porrello",
"Lorenzo Bonicelli",
"Pietro Buzzega",
"Monica Millunzi",
"Simone Calderara",
"Rita Cucchiara"
] | Spotlight | The fine-tuning of deep pre-trained models has revealed compositional properties, with multiple specialized modules that can be arbitrarily composed into a single, multi-task model. However, identifying the conditions that promote compositionality remains an open issue, with recent efforts concentrating mainly on linearized networks. We conduct a theoretical study that attempts to demystify compositionality in standard non-linear networks through the second-order Taylor approximation of the loss function. The proposed formulation highlights the importance of staying within the pre-training basin to achieve composable modules. Moreover, it provides the basis for two dual incremental training algorithms: the one from the perspective of multiple models trained individually, while the other aims to optimize the composed model as a whole. We probe their application in incremental classification tasks and highlight some valuable skills. In fact, the pool of incrementally learned modules not only supports the creation of an effective multi-task model but also enables unlearning and specialization in certain tasks. Code available at <https://github.com/aimagelab/mammoth> | Continual Learning, Model Compositionality, Ensemble Learning, Task Arithmetic | We explore compositionality in fine-tuning non-linear deep models, revealing that staying within the pre-training basin is key to creating effective incremental learners. | 9,444 | 2405.16350 | [
-0.03380879387259483,
-0.10742539167404175,
0.03812717646360397,
0.045444510877132416,
0.04715918377041817,
0.014220619574189186,
-0.056267306208610535,
-0.05762515589594841,
-0.0054822457022964954,
-0.08757966011762619,
-0.03780876100063324,
-0.011893809773027897,
0.01747460849583149,
0.07927890121936798,
-0.025749634951353073,
-0.008805690333247185,
0.10729281604290009,
0.12884128093719482,
-0.09187006950378418,
-0.024663448333740234,
-0.03320704773068428,
0.052586380392313004,
0.010667262598872185,
0.05138702318072319,
0.035704076290130615,
-0.04336177557706833,
-0.06313806772232056,
-0.07070757448673248,
0.03311651200056076,
-0.029571998864412308,
0.048347264528274536,
0.00905552413314581,
-0.00720617501065135,
0.06887371838092804,
-0.09483037889003754,
0.034317366778850555,
-0.05122803524136543,
0.03987809270620346,
0.0032733732368797064,
-0.04845988005399704,
0.0340677909553051,
0.009573151357471943,
-0.0033138457220047712,
-0.011564811691641808,
0.024814320728182793,
-0.07016053795814514,
-0.011920392513275146,
-0.12238547950983047,
0.05031628906726837,
0.022446615621447563,
-0.012690127827227116,
-0.01691175065934658,
-0.02566220425069332,
0.07828481495380402,
0.07974477857351303,
-0.024923883378505707,
-0.011782859452068806,
0.022114915773272514,
-0.03645094484090805,
-0.01693568006157875,
0.012949454598128796,
-0.0931679755449295,
-0.0007614773930981755,
-0.050614628940820694,
-0.0066179996356368065,
-0.013751990161836147,
0.006376223638653755,
0.02457631751894951,
-0.053781576454639435,
0.060054875910282135,
0.019260652363300323,
0.060608699917793274,
-0.026622142642736435,
0.00309232366271317,
0.0734890028834343,
0.015642313286662102,
0.113706573843956,
0.015211572870612144,
0.008599254302680492,
-0.03630587086081505,
0.009597464464604855,
0.07278375327587128,
0.0020474838092923164,
-0.07555652409791946,
0.09299169480800629,
-0.009911852888762951,
-0.042387738823890686,
0.057469069957733154,
0.035357024520635605,
-0.04900354892015457,
-0.09327542781829834,
0.0284431129693985,
0.012599888257682323,
-0.019977284595370293,
-0.0051123471930623055,
0.005283365957438946,
0.029443521052598953,
-0.05455772951245308,
-0.007665565237402916,
0.1017293781042099,
-0.02028377167880535,
0.010983009822666645,
0.03462473675608635,
0.0029402170330286026,
0.08051702380180359,
-0.033858560025691986,
0.02108277752995491,
0.047641877084970474,
0.051969025284051895,
-0.08224046230316162,
0.023996569216251373,
0.026896286755800247,
0.025441275909543037,
0.026353780180215836,
-0.011716709472239017,
0.03838668018579483,
0.036272287368774414,
-0.03218547999858856,
0.0319519080221653,
0.10984017699956894,
-0.025422487407922745,
0.019762348383665085,
0.04798747971653938,
0.009135819971561432,
0.024435287341475487,
-0.03679904341697693,
-0.04082491993904114,
2.1024509923143417e-33,
0.03767918050289154,
0.012443678453564644,
0.07908700406551361,
0.013760273344814777,
0.07636646926403046,
-0.04635468125343323,
0.046970974653959274,
0.011489694938063622,
0.0056235152296721935,
0.001915701199322939,
-0.034285325556993484,
0.032639313489198685,
-0.04866021126508713,
0.1506170779466629,
0.02124182879924774,
-0.03873239830136299,
0.034812018275260925,
0.006682312581688166,
0.08753005415201187,
-0.015200266614556313,
0.016253238543868065,
0.010876446962356567,
0.0016238342504948378,
-0.0025520033668726683,
0.029428165405988693,
0.01905764453113079,
0.07442205399274826,
-0.047648344188928604,
-0.10671057552099228,
0.03204219788312912,
-0.0956975594162941,
0.0004454385780263692,
-0.023738212883472443,
0.04100920259952545,
0.04950555041432381,
-0.0017733689164742827,
-0.020428309217095375,
0.04875459149479866,
0.0560360886156559,
-0.024101711809635162,
-0.01598924584686756,
0.008158785291016102,
-0.021672818809747696,
-0.03583415597677231,
-0.09661048650741577,
-0.11487207561731339,
-0.010411561466753483,
0.08375568687915802,
-0.023078441619873047,
-0.050263773649930954,
0.005021801684051752,
0.03250672668218613,
-0.020682426169514656,
-0.039810311049222946,
-0.04306064173579216,
0.03297513350844383,
0.002285180613398552,
0.044108711183071136,
0.07396989315748215,
-0.005315412301570177,
-0.03201068192720413,
0.05166182294487953,
-0.09241686761379242,
0.10522117465734482,
0.0008165929466485977,
0.019916491582989693,
-0.055396921932697296,
0.049906715750694275,
0.09321298450231552,
-0.017667924985289574,
-0.026232222095131874,
0.021035494282841682,
-0.06360447406768799,
-0.07616277039051056,
0.10930237174034119,
-0.008762638084590435,
0.01891764998435974,
-0.22546082735061646,
0.0024899279233068228,
0.061267733573913574,
-0.05274980142712593,
0.1270834356546402,
0.005467984359711409,
-0.05047145113348961,
-0.05531461909413338,
-0.05739252269268036,
0.00798160582780838,
0.0023876691702753305,
0.020782284438610077,
0.0016921738861128688,
-0.10637441277503967,
0.010409004054963589,
-0.016996411606669426,
-0.03370746225118637,
0.04738753288984299,
-1.0594809562487973e-33,
0.005365962628275156,
0.03667445480823517,
-0.07256095111370087,
0.020728135481476784,
0.012072712182998657,
-0.012749910354614258,
-0.04528779536485672,
-0.0052046640776097775,
-0.12921008467674255,
-0.0090979328379035,
0.03007606603205204,
-0.06759602576494217,
-0.03699949011206627,
-0.03998137637972832,
0.023159489035606384,
0.04671035334467888,
-0.10475385934114456,
0.03452371805906296,
0.0462217777967453,
-0.01716374233365059,
0.03753211349248886,
0.08473923802375793,
-0.09252376854419708,
0.028249463066458702,
-0.02027026377618313,
0.0516386441886425,
-0.04585734009742737,
0.10272262245416641,
-0.0075385659001767635,
-0.020595207810401917,
-0.07522514462471008,
-0.02324269525706768,
0.0008832264575175941,
-0.05015643313527107,
-0.019101154059171677,
0.0786004513502121,
-0.011854014359414577,
-0.007919756695628166,
-0.01995004527270794,
0.03485892713069916,
0.0352599173784256,
-0.03646743670105934,
-0.045703988522291183,
0.0553208589553833,
0.012675787322223186,
-0.03704383596777916,
-0.012088809162378311,
-0.009743399918079376,
0.02513730525970459,
0.03730429336428642,
-0.0051758503541350365,
-0.00500181969255209,
-0.08891888707876205,
0.02415095828473568,
-0.041714537888765335,
-0.06274020671844482,
0.04328193515539169,
-0.033112991601228714,
0.09588874876499176,
-0.03313642367720604,
-0.02317081019282341,
-0.02628326788544655,
0.014551730826497078,
0.01920735463500023,
0.04776272922754288,
-0.051852207630872726,
-0.004768914543092251,
0.02860902063548565,
0.012704330496490002,
0.042715515941381454,
-0.004849707242101431,
0.05369862914085388,
0.0027436460368335247,
-0.021012423560023308,
-0.021264294162392616,
-0.06919354200363159,
-0.006939906161278486,
-0.007674780674278736,
-0.00769696244969964,
-0.03358626365661621,
0.013302328065037727,
-0.015222262591123581,
-0.012876643799245358,
0.05665458366274834,
0.09511210769414902,
0.05563293397426605,
0.03650057315826416,
0.08657269924879074,
0.026509540155529976,
-0.039636656641960144,
-0.008746886625885963,
0.03378676995635033,
-0.04491700604557991,
0.02878761850297451,
-0.03148739039897919,
-4.77838213441828e-8,
-0.09560611099004745,
0.04318036884069443,
-0.08494000881910324,
0.025576291605830193,
0.012307649478316307,
-0.051669903099536896,
0.007194075733423233,
0.11360350251197815,
0.0006500128074549139,
0.009245072491466999,
0.06330222636461258,
0.01617213524878025,
0.0666157677769661,
-0.10526742786169052,
-0.017269110307097435,
0.03309791907668114,
0.058578964322805405,
-0.020197857171297073,
0.0008479218813590705,
-0.03132551908493042,
0.07514537870883942,
0.0305495448410511,
-0.025502556934952736,
-0.056036703288555145,
0.038719091564416885,
-0.10946948081254959,
-0.03553921729326248,
0.11951805651187897,
0.05881061777472496,
0.05122660472989082,
-0.051437996327877045,
0.038644611835479736,
-0.06513698399066925,
-0.010112266056239605,
0.027137551456689835,
0.07936176657676697,
0.08205003291368484,
-0.047381822019815445,
-0.04756513610482216,
-0.04836023226380348,
-0.03276108577847481,
0.074736088514328,
-0.06332641839981079,
0.00811469554901123,
0.019237689673900604,
-0.06178253889083862,
-0.043937310576438904,
-0.06916031241416931,
0.019815942272543907,
-0.0509122796356678,
0.040227364748716354,
0.008388455025851727,
0.012540954165160656,
0.09896764159202576,
0.020383751019835472,
-0.024932796135544777,
0.02910691127181053,
-0.06398335099220276,
-0.016148418188095093,
0.07734272629022598,
0.03374622017145157,
-0.020333264023065567,
-0.05258064344525337,
-0.03197718784213066
] |
Lean-STaR: Learning to Interleave Thinking and Proving | https://openreview.net/forum?id=SOWZ59UyNc | [
"Haohan Lin",
"Zhiqing Sun",
"Sean Welleck",
"Yiming Yang"
] | Spotlight | Traditional language model-based theorem proving assumes that by training on a sufficient amount of formal proof data, a model will learn to prove theorems. Our key observation is that a wealth of informal information that is not present in formal proofs can be useful for learning to prove theorems. For instance, humans think through steps of a proof, but this thought process is not visible in the resulting code. We present Lean-STaR, a framework for training language models to produce informal thoughts prior to each step of a proof, thereby boosting the model's theorem-proving capabilities. Lean-STaR uses retrospective ground-truth tactics to generate synthetic thoughts for training the language model. At inference time, the trained model directly generates the thoughts prior to the prediction of the tactics in each proof step. Building on the self-taught reasoner framework, we then apply expert iteration to further fine-tune the model on the correct proofs it samples and verifies using the Lean solver. Lean-STaR significantly outperform base models (43.4% → 46.3%, Pass@64). We also analyze the impact of the augmented thoughts on various aspects of the theorem proving process, providing insights into their effectiveness. | Automated Theorem Proving, Chain-of-Thought, Reinforcement Learning, Reasoning | null | 9,410 | null | [
-0.0721626728773117,
-0.013771619647741318,
0.010121272876858711,
0.004595175385475159,
0.04197559133172035,
-0.024653447791934013,
0.008299726992845535,
0.10975032299757004,
0.050484031438827515,
0.020687228068709373,
-0.10134318470954895,
-0.036411721259355545,
0.06011482700705528,
0.014834202826023102,
0.031829096376895905,
-0.0004148852895013988,
0.0036292928270995617,
0.02249438874423504,
-0.005372172687202692,
-0.0959562212228775,
-0.015950001776218414,
-0.05706197768449783,
0.037696849554777145,
0.005656723864376545,
-0.008812549524009228,
-0.022636910900473595,
-0.015572450123727322,
-0.029140479862689972,
0.016511421650648117,
-0.05037685111165047,
0.02539162151515484,
0.08736510574817657,
0.026251323521137238,
-0.004036231432110071,
-0.0510537289083004,
0.07756577432155609,
-0.014308050274848938,
0.03397645428776741,
-0.01696660742163658,
-0.027594558894634247,
-0.04831095412373543,
-0.05341232940554619,
-0.0016279573319479823,
-0.027799829840660095,
0.049623265862464905,
0.019476791843771935,
0.015868879854679108,
0.03744416683912277,
-0.03487415984272957,
0.04906522110104561,
-0.14799782633781433,
-0.07367265224456787,
-0.035991325974464417,
-0.056261129677295685,
0.027412639930844307,
0.0048430548049509525,
0.15406344830989838,
0.07011539489030838,
-0.021339058876037598,
-0.035828039050102234,
-0.014174765907227993,
-0.08613911271095276,
-0.11607342213392258,
-0.022915083914995193,
-0.00481440918520093,
-0.0000793272047303617,
0.022348089143633842,
0.08931191265583038,
-0.04423636198043823,
0.032955776900053024,
-0.005555023904889822,
0.07931078970432281,
-0.0838642492890358,
-0.029987024143338203,
-0.0146294254809618,
0.043302807956933975,
0.06062788516283035,
-0.035625725984573364,
0.06700965762138367,
-0.05084583908319473,
-0.0756850391626358,
0.07190702110528946,
-0.04875895753502846,
0.013033402152359486,
0.06330706924200058,
-0.016686873510479927,
0.04809863120317459,
0.06251031905412674,
0.005730571690946817,
0.046548791229724884,
-0.010245461948215961,
-0.02538202330470085,
-0.043149206787347794,
-0.02774345688521862,
0.10641062259674072,
0.10538865625858307,
-0.03201264142990112,
-0.10762467980384827,
0.006730704568326473,
0.11027892678976059,
-0.022310083732008934,
0.11127810180187225,
0.06694475561380386,
-0.13140231370925903,
0.0221796166151762,
-0.010196276940405369,
0.04379802569746971,
0.030128583312034607,
0.06148695573210716,
-0.0598335899412632,
0.024448849260807037,
0.0321677140891552,
0.0027618552558124065,
0.02035924419760704,
0.00861793477088213,
0.05570462718605995,
0.03375350683927536,
-0.01756509765982628,
-0.0268105398863554,
-0.03244202211499214,
-0.006207781378179789,
0.11684821546077728,
0.053368549793958664,
0.02629423700273037,
0.04339289665222168,
-0.059143342077732086,
-0.06898770481348038,
7.291163045561072e-34,
0.09922614693641663,
0.07147745043039322,
0.07688052952289581,
0.036695439368486404,
0.049286823719739914,
-0.028906289488077164,
0.0025541658978909254,
0.019263284280896187,
-0.02809559926390648,
0.009938064962625504,
0.050054848194122314,
-0.020575053989887238,
-0.01180118229240179,
0.015854619443416595,
0.02475030906498432,
0.008480483666062355,
-0.07816935330629349,
-0.006101724226027727,
-0.07496467232704163,
-0.006452282425016165,
0.09963243454694748,
-0.07192093878984451,
0.0013117213966324925,
-0.05484721064567566,
0.014892544597387314,
0.013613766059279442,
0.07936923950910568,
0.002592572709545493,
-0.04660988226532936,
-0.02062351629137993,
-0.056467343121767044,
-0.06667883694171906,
-0.027058195322752,
0.06346192210912704,
0.03562726825475693,
-0.06195635348558426,
0.0008748963009566069,
-0.06676974147558212,
0.047970615327358246,
0.010560394264757633,
-0.024696439504623413,
-0.0046036336570978165,
0.02722446247935295,
0.04416251182556152,
-0.04202704876661301,
-0.02400311455130577,
-0.032696086913347244,
-0.04073921591043472,
-0.004056322388350964,
-0.0348920002579689,
0.020798902958631516,
0.029613031074404716,
-0.014168879017233849,
-0.09020375460386276,
-0.019755573943257332,
0.020575987175107002,
-0.014793704263865948,
0.03685122728347778,
0.044848982244729996,
0.03676283732056618,
0.05147137865424156,
0.057582151144742966,
-0.015117032453417778,
0.06924420595169067,
-0.050520408898591995,
0.08329609036445618,
-0.08395993709564209,
0.07483391463756561,
0.12688568234443665,
-0.03998060151934624,
-0.04730069637298584,
0.0002918571117334068,
-0.04254576936364174,
-0.039691586047410965,
0.07666727155447006,
-0.03319014981389046,
0.0055541121400892735,
-0.11543568968772888,
0.03252936527132988,
-0.015733765438199043,
0.06755609065294266,
-0.0006255103508010507,
-0.018212640658020973,
-0.04257616400718689,
0.000976958661340177,
-0.054126981645822525,
-0.019526271149516106,
-0.07145129889249802,
0.0045683616772294044,
-0.0325225293636322,
-0.08723904937505722,
-0.10272905975580215,
-0.006494987290352583,
-0.0251537524163723,
-0.05009179562330246,
-1.1510153238148865e-33,
-0.010698159225285053,
-0.04966292530298233,
-0.032159581780433655,
0.07606003433465958,
-0.015833932906389236,
-0.019315620884299278,
0.004418864846229553,
0.02072761580348015,
-0.013687988743185997,
-0.05561934411525726,
-0.0609176866710186,
-0.04449627920985222,
0.028712334111332893,
0.019309714436531067,
0.05323687940835953,
-0.07085561752319336,
-0.0549238920211792,
0.015932047739624977,
0.006270051468163729,
0.054764144122600555,
0.012231919914484024,
0.037562061101198196,
-0.06679342687129974,
0.009469271637499332,
-0.0008065203437581658,
-0.008603440597653389,
-0.07661961764097214,
0.11445987224578857,
0.020389743149280548,
0.03334657847881317,
-0.005673597101122141,
-0.06697680056095123,
-0.05746231973171234,
-0.05520033836364746,
-0.05398320034146309,
0.0022213547490537167,
0.06562493741512299,
0.009196280501782894,
-0.057866111397743225,
0.11031875014305115,
0.07245249301195145,
-0.09078115969896317,
-0.02645070105791092,
-0.033759258687496185,
-0.08867304772138596,
-0.022691264748573303,
-0.05891527608036995,
-0.01412216667085886,
-0.03383677452802658,
0.024758322164416313,
0.017260584980249405,
-0.0006566679221577942,
-0.024616844952106476,
0.02196267433464527,
-0.05164448544383049,
-0.05422086641192436,
-0.008121568709611893,
-0.012733579613268375,
0.02690861001610756,
0.01614646427333355,
-0.029626883566379547,
-0.02267158031463623,
0.038124945014715195,
-0.027263185009360313,
0.05366357043385506,
-0.06931235641241074,
-0.023775476962327957,
0.050060369074344635,
-0.001339130220003426,
0.03114939294755459,
0.0563596673309803,
0.0008369560819119215,
-0.04785946011543274,
0.04066232964396477,
0.10473067313432693,
0.05602350831031799,
0.0024027302861213684,
-0.04302305728197098,
-0.032580774277448654,
-0.0838315337896347,
0.01635461114346981,
-0.04188784584403038,
0.05327853932976723,
0.010400073602795601,
0.07711443305015564,
0.029386790469288826,
0.03921571373939514,
0.0602654404938221,
-0.02282029576599598,
0.08190786838531494,
-0.06714583188295364,
0.03582518547773361,
-0.04874632507562637,
0.11467866599559784,
-0.07472582906484604,
-4.339050718726867e-8,
-0.06088460609316826,
0.05928085371851921,
0.06592191755771637,
0.13296204805374146,
0.08295878022909164,
-0.025860225781798363,
0.0023526595905423164,
0.004557624459266663,
-0.02641320787370205,
-0.04487147927284241,
0.04082155600190163,
0.02091381326317787,
-0.047185592353343964,
-0.07698790729045868,
-0.006514451932162046,
0.0514473095536232,
0.004722035024315119,
-0.01484211627393961,
-0.034552160650491714,
-0.03635489195585251,
0.07291945815086365,
-0.022868994623422623,
-0.029151007533073425,
0.00657978281378746,
0.020501326769590378,
-0.04092580825090408,
0.005026205908507109,
0.14139039814472198,
0.034120578318834305,
-0.00500988494604826,
-0.00903678871691227,
0.0357712022960186,
0.015335926786065102,
-0.08163845539093018,
0.07406456768512726,
0.07409236580133438,
0.06930530816316605,
0.017742682248353958,
0.0429033525288105,
-0.016500618308782578,
-0.007143172901123762,
0.06564567238092422,
-0.02517867647111416,
0.0015170160913839936,
0.026868991553783417,
-0.037393201142549515,
-0.06045966222882271,
-0.10571497678756714,
-0.020685574039816856,
-0.020294347777962685,
0.017092028632760048,
-0.0035868340637534857,
-0.019871503114700317,
0.061953458935022354,
0.02024654671549797,
0.06887175142765045,
-0.024049879983067513,
-0.0012575850123539567,
-0.04309042915701866,
0.07942665368318558,
0.043642524629831314,
0.03062995709478855,
0.0522945299744606,
0.03328985348343849
] |
Spectral Compressive Imaging via Unmixing-driven Subspace Diffusion Refinement | https://openreview.net/forum?id=Q150eWkQ4I | [
"Haijin Zeng",
"Benteng Sun",
"Yongyong Chen",
"Jingyong Su",
"Yong Xu"
] | Spotlight | Spectral Compressive Imaging (SCI) reconstruction is inherently ill-posed, offering multiple plausible solutions from a single observation. Traditional deterministic methods typically struggle to effectively recover high-frequency details. Although diffusion models offer promising solutions to this challenge, their application is constrained by the limited training data and high computational demands associated with multispectral images (MSIs), complicating direct training. To address these issues, we propose a novel Predict-and-unmixing-driven-Subspace-Refine framework (PSR-SCI). This framework begins with a cost-effective predictor that produces an initial, rough estimate of the MSI. Subsequently, we introduce a unmixing-driven reversible spectral embedding module that decomposes the MSI into subspace images and spectral coefficients. This decomposition facilitates the adaptation of pre-trained RGB diffusion models and focuses refinement processes on high-frequency details, thereby enabling efficient diffusion generation with minimal MSI data. Additionally, we design a high-dimensional guidance mechanism with imaging consistency to enhance the model's efficacy. The refined subspace image is then reconstructed back into an MSI using the reversible embedding, yielding the final MSI with full spectral resolution. Experimental results on the standard KAIST and zero-shot datasets NTIRE, ICVL, and Harvard show that PSR-SCI enhances visual quality and delivers PSNR and SSIM metrics comparable to existing diffusion, transformer, and deep unfolding techniques. This framework provides a robust alternative to traditional deterministic SCI reconstruction methods. Code and models are available at [https://github.com/SMARK2022/PSR-SCI](https://github.com/SMARK2022/PSR-SCI). | Spectral compressive imaging, subspace, diffusion, fine-tune | null | 9,358 | null | [
-0.019648687914013863,
-0.0773438960313797,
0.03491885960102081,
0.09678402543067932,
0.07325856387615204,
0.022957710549235344,
-0.03629283979535103,
-0.06846439093351364,
-0.016871057450771332,
-0.022032029926776886,
-0.04476030915975571,
-0.02283683605492115,
0.004375944845378399,
0.04217056185007095,
-0.02335367724299431,
-0.048591624945402145,
0.0055306656286120415,
0.03213367983698845,
-0.03314884379506111,
-0.020456472411751747,
0.014793419279158115,
-0.06420131772756577,
-0.031168198212981224,
-0.02325231395661831,
0.05880292132496834,
0.04050420969724655,
0.05791209638118744,
-0.018358251079916954,
-0.004701781552284956,
0.02381771244108677,
0.10370390862226486,
0.05479181557893753,
-0.03534527122974396,
0.01002851314842701,
-0.05481543764472008,
0.02075245790183544,
0.027816157788038254,
0.05225711315870285,
0.019685084000229836,
0.030398834496736526,
-0.01386900432407856,
0.08640418946743011,
-0.04259946197271347,
0.028717953711748123,
-0.008851001039147377,
-0.06158753111958504,
0.01264569628983736,
-0.08969134092330933,
0.054250869899988174,
-0.03811889886856079,
0.0323500819504261,
0.02716927044093609,
-0.05232551693916321,
0.07146964967250824,
-0.009067944251000881,
0.01963936723768711,
0.008292961865663528,
-0.07174862176179886,
0.07457032054662704,
-0.021574337035417557,
-0.04751652479171753,
-0.07834047824144363,
0.018358852714300156,
-0.006799166090786457,
0.01928102783858776,
0.031047038733959198,
-0.04337090998888016,
0.0017207636265084147,
0.01513840164989233,
0.030240017920732498,
-0.0893736407160759,
0.0038731589447706938,
-0.04230537638068199,
-0.01575743965804577,
-0.01176544837653637,
-0.00032159616239368916,
0.06292828917503357,
0.06314114481210709,
0.04948849231004715,
0.06258242577314377,
0.08258232474327087,
0.0621488019824028,
0.0394979864358902,
-0.016227178275585175,
0.08276671916246414,
0.001740695908665657,
0.04423468932509422,
0.05435776337981224,
0.040047407150268555,
-0.02134002186357975,
-0.06756255775690079,
-0.032097507268190384,
-0.08052504062652588,
0.03782457858324051,
-0.022930648177862167,
-0.04295438155531883,
0.033884305506944656,
-0.06665346026420593,
0.023156743496656418,
0.05199238657951355,
0.025708293542265892,
-0.021646279841661453,
-0.027242545038461685,
0.03129100054502487,
-0.04926057159900665,
-0.10869969427585602,
0.09643811732530594,
0.03965824097394943,
0.01315297745168209,
0.011534477584064007,
0.08005741238594055,
0.047100577503442764,
0.009058602154254913,
0.03961893543601036,
0.04789464920759201,
-0.05872901901602745,
0.001663591479882598,
0.01912546157836914,
0.05618692934513092,
0.016912780702114105,
-0.0387398824095726,
-0.03921593353152275,
-0.06449597328901291,
0.0051663657650351524,
0.09534669667482376,
-0.0614146813750267,
-0.09496048837900162,
1.1376741976170228e-33,
-0.0594043955206871,
0.04744840785861015,
0.031505435705184937,
0.057500194758176804,
0.04684166982769966,
-0.014547090046107769,
-0.03866148740053177,
0.02007484622299671,
0.04463590681552887,
0.009278640151023865,
0.019586799666285515,
0.01520482636988163,
-0.05473417416214943,
0.13886143267154694,
-0.03784741833806038,
-0.0030262309592217207,
-0.013858783058822155,
0.08329089730978012,
-0.02889014407992363,
-0.10660284757614136,
-0.016527604311704636,
-0.020313343033194542,
0.0259335208684206,
-0.07253649830818176,
-0.014848473481833935,
-0.04553196579217911,
0.05532224848866463,
-0.0874335840344429,
0.07022560387849808,
-0.0012577543966472149,
-0.09982650727033615,
-0.05938202515244484,
0.06818477809429169,
0.041408177465200424,
0.02147625759243965,
0.012343795038759708,
0.010978727601468563,
0.01608487218618393,
0.00038239845889620483,
0.007475679274648428,
0.004829243756830692,
0.08639673888683319,
-0.019266216084361076,
-0.02908831089735031,
-0.05263142287731171,
-0.08754780888557434,
-0.014780696481466293,
0.02275744453072548,
-0.09191299229860306,
-0.0035526545252650976,
-0.04228856414556503,
-0.024360423907637596,
-0.0363142192363739,
-0.01534451637417078,
-0.10531787574291229,
-0.04913652688264847,
0.08877785503864288,
-0.043770186603069305,
0.010403339751064777,
-0.02002287097275257,
0.006273685954511166,
-0.001039649243466556,
-0.026202119886875153,
0.0221435334533453,
0.016763344407081604,
0.013891863636672497,
-0.012790261767804623,
0.04029709845781326,
0.005605918355286121,
0.1013830155134201,
-0.08890695869922638,
0.09560783952474594,
0.006682374514639378,
-0.0613093376159668,
0.07732386887073517,
0.005143206100910902,
0.03614995256066322,
0.05391940474510193,
-0.026478148996829987,
0.07964041829109192,
-0.03278210014104843,
0.03907551243901253,
-0.06456096470355988,
-0.02821221947669983,
-0.06405611336231232,
0.003627448109909892,
0.013996544294059277,
-0.06871312111616135,
-0.028004618361592293,
-0.16210679709911346,
-0.020711693912744522,
-0.030067503452301025,
-0.06321322917938232,
0.02858021855354309,
0.03442364186048508,
3.5554401619379195e-34,
0.08547239005565643,
0.09316074103116989,
0.047937359660863876,
-0.07721386104822159,
-0.01667763851583004,
-0.04617968201637268,
0.03573689982295036,
0.07588161528110504,
0.09350727498531342,
-0.09961781650781631,
0.015225694514811039,
-0.0694006159901619,
0.0028198857326060534,
0.038999784737825394,
0.015499923378229141,
0.0034252083860337734,
-0.057090576738119125,
0.03289170563220978,
-0.03523165360093117,
-0.019952740520238876,
-0.058732617646455765,
0.04063435271382332,
0.05743813142180443,
0.05746595934033394,
-0.062325675040483475,
0.11035420000553131,
0.006209093611687422,
0.0525604784488678,
0.017757486552000046,
-0.05481145530939102,
-0.012753698974847794,
-0.04893624037504196,
-0.015098432078957558,
-0.06854556500911713,
-0.05769290030002594,
0.11727303266525269,
-0.016786916181445122,
0.00690943980589509,
-0.024552177637815475,
0.09042758494615555,
0.05855991691350937,
0.015628954395651817,
-0.03912564367055893,
0.05521917715668678,
0.01608506590127945,
-0.05795256420969963,
0.02495548687875271,
-0.01666904054582119,
0.1298598200082779,
0.019285624846816063,
-0.002790004713460803,
0.01304722111672163,
-0.11233111470937729,
0.03247857466340065,
0.04084460809826851,
-0.07163048535585403,
-0.08147228509187698,
-0.018089480698108673,
0.10288240015506744,
-0.034376636147499084,
-0.010359533131122589,
0.011784622445702553,
-0.05246163159608841,
-0.025933586061000824,
-0.041788239032030106,
-0.001154167577624321,
-0.005192313343286514,
-0.026506273075938225,
0.021601328626275063,
0.10733800381422043,
-0.06782929599285126,
0.04990229383111,
0.0011484979186207056,
0.013860687613487244,
0.11163614690303802,
-0.03290478140115738,
-0.026034049689769745,
-0.017834177240729332,
-0.018671328201889992,
0.056976474821567535,
0.06657055765390396,
-0.04819733649492264,
-0.01624724268913269,
-0.0125584090128541,
0.07805837690830231,
0.06264560669660568,
-0.024092454463243484,
0.003316600574180484,
0.06340447068214417,
-0.059769514948129654,
0.0030696499161422253,
-0.02488703653216362,
0.05935588479042053,
0.03652545437216759,
0.006277742795646191,
-3.894568223472561e-8,
-0.09616435319185257,
0.01144698727875948,
0.002107167849317193,
-0.030202776193618774,
-0.02976406179368496,
-0.09132178127765656,
0.09350618720054626,
-0.014149690046906471,
0.012046964839100838,
-0.06530787795782089,
0.056251410394907,
-0.12805905938148499,
-0.06304030865430832,
-0.024761859327554703,
-0.050735194236040115,
0.0777142122387886,
-0.0011057502124458551,
0.010897135362029076,
0.033723413944244385,
-0.006226003635674715,
-0.03398815169930458,
-0.01909075491130352,
0.08329323679208755,
-0.006710509769618511,
0.10602796822786331,
0.005146773066371679,
-0.007966207340359688,
0.017883718013763428,
0.04167016223073006,
0.014376587234437466,
0.011286694556474686,
-0.029671110212802887,
0.046003393828868866,
-0.018659595400094986,
-0.08776736259460449,
0.0862903892993927,
0.0361359529197216,
0.004100656136870384,
-0.0630171000957489,
-0.07717191427946091,
-0.08879125863313675,
0.10342028737068176,
-0.05179041624069214,
0.0026917406357824802,
0.022755062207579613,
0.02148810587823391,
0.07081156969070435,
0.05946710333228111,
0.01800817809998989,
-0.03967348113656044,
0.04377425089478493,
-0.0404612235724926,
0.03804989904165268,
0.04086996987462044,
0.035154566168785095,
-0.021921277046203613,
-0.022427784278988838,
0.0002919829566963017,
0.04063420742750168,
0.016570813953876495,
-0.03846019133925438,
-0.006920217536389828,
-0.028734080493450165,
0.04556579515337944
] |
CBQ: Cross-Block Quantization for Large Language Models | https://openreview.net/forum?id=eW4yh6HKz4 | [
"Xin Ding",
"Xiaoyu Liu",
"Zhijun Tu",
"Yun Zhang",
"Wei Li",
"Jie Hu",
"Hanting Chen",
"Yehui Tang",
"Zhiwei Xiong",
"Baoqun Yin",
"Yunhe Wang"
] | Spotlight | Post-training quantization (PTQ) has played a pivotal role in compressing large language models (LLMs) at ultra-low costs. Although current PTQ methods have achieved promising results by addressing outliers and employing layer- or block-wise loss optimization techniques, they still suffer from significant performance degradation at ultra-low bits precision. To dissect this issue, we conducted an in-depth analysis of quantization errors specific to LLMs and surprisingly discovered that, unlike traditional sources of quantization errors, the growing number of model parameters, combined with the reduction in quantization bits, intensifies inter-layer and intra-layer dependencies, which severely impact quantization accuracy. This finding highlights a critical challenge in quantizing LLMs. To address this, we propose CBQ, a cross-block reconstruction-based PTQ method for LLMs. CBQ leverages a cross-block dependency to establish long-range dependencies across multiple blocks and integrates an adaptive LoRA-Rounding technique to manage intra-layer dependencies. To further enhance performance, CBQ incorporates a coarse-to-fine pre-processing mechanism for processing weights and activations. Extensive experiments show that CBQ achieves superior low-bit quantization (W4A4, W4A8, W2A16) and outperforms existing state-of-the-art methods across various LLMs and datasets. Notably, CBQ only takes 4.3 hours to quantize a weight-only quantization of a 4-bit LLAMA1-65B model, achieving a commendable trade off between performance and efficiency. | Large Language Model Compression, ultra-low bits precision | null | 9,338 | 2312.07950 | [
-0.04847251623868942,
-0.008909446187317371,
0.05152526870369911,
0.02444203570485115,
0.04363204166293144,
0.05266420170664787,
-0.055420804768800735,
-0.031061630696058273,
0.03545272350311279,
-0.08619146794080734,
-0.022697312757372856,
-0.048577047884464264,
-0.02560388669371605,
0.010642972774803638,
0.02066851034760475,
0.007688022218644619,
0.0690353736281395,
0.12421681731939316,
-0.10222209244966507,
-0.07850321382284164,
0.027896158397197723,
0.0447559654712677,
-0.04409576207399368,
0.029890382662415504,
0.1011694073677063,
-0.01220400258898735,
-0.0329451821744442,
0.007916655391454697,
0.07010047882795334,
-0.009197525680065155,
0.02434702217578888,
0.03655914217233658,
0.02790953777730465,
0.012268034741282463,
-0.024278270080685616,
0.06370119750499725,
-0.04088564217090607,
-0.028482412919402122,
0.027443693950772285,
-0.06083802506327629,
0.002186019904911518,
0.024083901196718216,
0.04957342892885208,
0.012205375358462334,
0.0514044314622879,
-0.017698019742965698,
0.011473620310425758,
0.03277059644460678,
-0.012023679912090302,
-0.0005314338486641645,
0.022776976227760315,
0.03751078620553017,
-0.03817327693104744,
0.09466834366321564,
-0.030562208965420723,
-0.056087594479322433,
0.022763272747397423,
0.057564716786146164,
-0.06885629892349243,
0.04222476854920387,
-0.0970422551035881,
-0.04406594857573509,
0.024414388462901115,
0.0004969428409822285,
0.023109572008252144,
-0.028018325567245483,
0.009840159676969051,
0.05586618557572365,
-0.023118985816836357,
0.05378970503807068,
-0.047003813087940216,
-0.010189793072640896,
-0.07051322609186172,
0.12700335681438446,
0.0012290949234738946,
0.09055410325527191,
0.04374537989497185,
0.020879779011011124,
0.02832351252436638,
-0.049201928079128265,
-0.019040238112211227,
0.031575024127960205,
0.04736606776714325,
-0.02713867835700512,
0.1143583282828331,
-0.06749813258647919,
-0.031358856707811356,
0.006339394953101873,
-0.004369833506643772,
0.012347782962024212,
-0.011823716573417187,
-0.023626767098903656,
0.02329099178314209,
0.04724014177918434,
0.08274528384208679,
-0.0029133802745491266,
0.023809028789401054,
-0.00173101550899446,
0.019435815513134003,
0.08535882085561752,
0.07764904946088791,
0.06267450004816055,
-0.04406673088669777,
-0.02097867615520954,
-0.035035666078329086,
-0.046891190111637115,
0.05599924549460411,
0.056986819952726364,
-0.015332251787185669,
-0.061507418751716614,
0.0439533106982708,
0.04789314791560173,
0.01850838214159012,
-0.050806332379579544,
0.0628761425614357,
-0.04368313029408455,
0.02013777196407318,
-0.067730613052845,
0.0268834438174963,
-0.010414361953735352,
-0.0992618054151535,
-0.015007955953478813,
-0.005883645266294479,
0.03738444298505783,
0.0029458568897098303,
-0.043449290096759796,
-0.07139457762241364,
1.3281310984151115e-33,
0.07987198233604431,
0.08696840703487396,
0.01977473683655262,
0.01585526019334793,
0.0074955858290195465,
0.009754225611686707,
0.10728361457586288,
0.009494929574429989,
-0.043605986982584,
0.0672903060913086,
-0.024293741211295128,
-0.037236012518405914,
-0.05232910066843033,
-0.005790525581687689,
0.04398706927895546,
0.004866565577685833,
-0.05258623883128166,
0.02237613871693611,
-0.048440344631671906,
-0.04011989012360573,
0.04251592606306076,
-0.07658998668193817,
0.024562058970332146,
0.020818792283535004,
0.02340092696249485,
0.057526927441358566,
0.1071595847606659,
-0.0798802599310875,
-0.0881541520357132,
0.03199975937604904,
-0.09428243339061737,
-0.07290017604827881,
0.06925756484270096,
-0.036682989448308945,
0.05163224786520004,
-0.03170548751950264,
0.02119273506104946,
0.005841445177793503,
0.00012991092808078974,
-0.021464301273226738,
-0.02707308903336525,
0.01769847422838211,
0.027247559279203415,
-0.06340807676315308,
-0.09308302402496338,
-0.05657846853137016,
-0.012771709822118282,
-0.03365882858633995,
-0.06003595143556595,
-0.09231038391590118,
0.08172738552093506,
-0.0260462686419487,
-0.056436650454998016,
0.05949735268950462,
0.040413204580545425,
-0.032076504081487656,
0.0954681783914566,
-0.0016137477941811085,
0.039558086544275284,
0.09400024265050888,
0.003453186247497797,
0.03184564784169197,
-0.048488058149814606,
0.11552875488996506,
0.04174988344311714,
0.012493159621953964,
-0.021239152178168297,
0.039284598082304,
-0.05129723623394966,
0.03251860663294792,
-0.06308925896883011,
-0.027260107919573784,
0.02181428112089634,
0.016620783135294914,
0.0776214450597763,
-0.05343342944979668,
0.013667034916579723,
-0.08312191814184189,
0.02132744900882244,
-0.02114994265139103,
0.028282664716243744,
0.04636377841234207,
0.028687234967947006,
-0.056839536875486374,
-0.07903733104467392,
-0.08799499273300171,
0.03231554850935936,
-0.046792689710855484,
-0.029386458918452263,
-0.0028338695410639048,
-0.03461458534002304,
-0.07591530680656433,
0.02888464741408825,
-0.06219948083162308,
-0.049517922103405,
-9.13207214678272e-34,
-0.055930908769369125,
0.06450161337852478,
0.0005769138806499541,
0.043186794966459274,
-0.07733834534883499,
-0.16952767968177795,
-0.0009591482230462134,
0.0376766175031662,
0.06764808297157288,
-0.021227873861789703,
-0.055884331464767456,
-0.08470558375120163,
0.04763401672244072,
0.011880713514983654,
0.04621077701449394,
0.06039481610059738,
-0.040526289492845535,
0.02438592165708542,
-0.01813367009162903,
0.08783182501792908,
-0.023757651448249817,
0.00016448037058580667,
-0.0309344120323658,
-0.0046479045413434505,
-0.02395741455256939,
0.05227842926979065,
-0.029283443465828896,
0.056275624781847,
0.0688190683722496,
-0.021054964512586594,
-0.07038277387619019,
-0.06805309653282166,
-0.06299291551113129,
-0.07791884988546371,
-0.07332593947649002,
0.008566580712795258,
0.04070815443992615,
0.045569535344839096,
-0.002681420184671879,
-0.010722195729613304,
0.08429133147001266,
-0.052891481667757034,
-0.12911000847816467,
0.03853258118033409,
0.013946679420769215,
0.02051352523267269,
-0.11408258229494095,
0.01771484687924385,
-0.006685564294457436,
0.01940026320517063,
-0.008129922673106194,
0.04128800332546234,
0.058402951806783676,
0.10265012085437775,
-0.021347718313336372,
-0.0813673660159111,
-0.09549304842948914,
0.007606348022818565,
0.02385850064456463,
-0.009795858524739742,
-0.06307455152273178,
-0.03335052728652954,
0.042800359427928925,
-0.13932090997695923,
0.04959486797451973,
-0.01992204785346985,
-0.008155643939971924,
0.030242176726460457,
-0.026114447042346,
0.0024719643406569958,
0.01883738301694393,
0.008027277886867523,
-0.01739581860601902,
0.0037044882774353027,
-0.023960860446095467,
-0.026672644540667534,
-0.026360085234045982,
-0.00462415162473917,
0.04742675647139549,
0.06531865894794464,
0.016562609001994133,
0.06224801018834114,
0.023467008024454117,
0.001871247310191393,
0.09702234715223312,
0.10974732041358948,
0.01143902912735939,
0.08785893023014069,
-0.007992863655090332,
-0.014111434109508991,
-0.06680642068386078,
0.04742078855633736,
0.05325637757778168,
0.012443392537534237,
-0.07070594280958176,
-4.050591684290339e-8,
-0.0906592607498169,
0.03672647476196289,
-0.07344318181276321,
-0.026348207145929337,
0.054956283420324326,
-0.05262196436524391,
-0.029668904840946198,
0.08842901885509491,
0.02223721332848072,
-0.0424359031021595,
0.04332716390490532,
-0.03113289549946785,
-0.07299017161130905,
-0.04858112335205078,
-0.057336121797561646,
0.06553247570991516,
-0.004568160977214575,
0.012550952844321728,
0.00968136452138424,
-0.04683874174952507,
-0.02248987928032875,
0.10323502868413925,
0.017494628205895424,
0.009760879911482334,
0.029326293617486954,
-0.056047774851322174,
0.010352877900004387,
0.10286726802587509,
0.15269844233989716,
0.002818317385390401,
-0.02290240488946438,
-0.0005680916947312653,
0.023887570947408676,
0.02567780204117298,
0.03292142599821091,
0.06993306428194046,
0.03227139636874199,
-0.023892749100923538,
0.014577028341591358,
-0.04229984059929848,
0.007951282896101475,
-0.023999448865652084,
-0.07822015881538391,
0.035703711211681366,
0.05221393331885338,
-0.04857170209288597,
-0.06273459643125534,
0.007840260863304138,
0.00282100448384881,
-0.08975055813789368,
0.0006798270042054355,
0.01896553859114647,
-0.04052553325891495,
0.03827033191919327,
0.0267080906778574,
-0.0034982997458428144,
0.022855276241898537,
-0.05484975501894951,
0.03455917164683342,
-0.019428610801696777,
-0.02521887607872486,
-0.05143677443265915,
-0.016630742698907852,
0.02107580378651619
] |
Iterative Label Refinement Matters More than Preference Optimization under Weak Supervision | https://openreview.net/forum?id=q5EZ7gKcnW | [
"Yaowen Ye",
"Cassidy Laidlaw",
"Jacob Steinhardt"
] | Spotlight | Language model (LM) post-training relies on two stages of human supervision: task demonstrations for supervised finetuning (SFT), followed by preference comparisons for reinforcement learning from human feedback (RLHF). As LMs become more capable, the tasks they are given become harder to supervise. Will post-training remain effective under unreliable supervision? To test this, we simulate unreliable demonstrations and comparison feedback using small LMs and time-constrained humans. We find that in the presence of unreliable supervision, SFT still retains some effectiveness, but DPO (a common RLHF algorithm) fails to improve the model beyond SFT. To address this, we propose *iterative label refinement* (ILR) as an alternative to RLHF. ILR improves the SFT data by using comparison feedback to decide whether human demonstrations should be replaced by model-generated alternatives, then retrains the model via SFT on the updated data. SFT+ILR outperforms SFT+DPO on several tasks with unreliable supervision (math, coding, and safe instruction-following). Our findings suggest that as LMs are used for complex tasks where human supervision is unreliable, RLHF may no longer be the best use of human comparison feedback; instead, it is better to direct feedback towards improving the training *data* rather than continually training the *model*. Our code and data are available at https://github.com/helloelwin/iterative-label-refinement. | unreliable human supervision, language model post-training, scalable oversight | We find SFT+DPO breaks down under unreliable supervision and show that it is better to direct unreliable feedback towards improving the training *data* rather than continually training the *model*. | 9,316 | 2501.07886 | [
-0.040012799203395844,
-0.08652250468730927,
0.06978694349527359,
0.018627364188432693,
0.03751658275723457,
0.03181608021259308,
-0.010623774491250515,
0.07766333222389221,
-0.005793588701635599,
-0.047811586409807205,
-0.04810335114598274,
0.0002778152993414551,
0.007451324723660946,
0.0458931103348732,
0.011718552559614182,
-0.034451063722372055,
0.16608604788780212,
0.03395582735538483,
-0.04912789538502693,
-0.07296224683523178,
-0.022815877571702003,
0.0274319127202034,
0.058673515915870667,
-0.017820969223976135,
-0.06470981985330582,
0.005473265424370766,
-0.08201004564762115,
-0.04750852659344673,
0.06373488903045654,
0.014522121287882328,
0.00965936854481697,
0.002716990653425455,
-0.029421692714095116,
-0.031141242012381554,
-0.10985933244228363,
0.07509474456310272,
-0.07133170962333679,
-0.03610098361968994,
0.035684410482645035,
-0.02927507273852825,
-0.1001310870051384,
0.027300339192152023,
0.0155396219342947,
0.00424755085259676,
0.023405110463500023,
0.009770656935870647,
0.007911738008260727,
-0.07205496728420258,
-0.08216282725334167,
0.06742382794618607,
-0.026789722964167595,
-0.04681601747870445,
0.08750404417514801,
-0.016586050391197205,
0.018795911222696304,
0.046724747866392136,
0.08869250863790512,
0.030476467683911324,
0.046557631343603134,
-0.013019821606576443,
-0.05157794803380966,
-0.0905645564198494,
-0.08451730012893677,
-0.03824026137590408,
-0.019906427711248398,
-0.049268659204244614,
-0.011047896929085255,
-0.0033639271277934313,
0.03762377053499222,
0.05195719748735428,
0.004518386907875538,
0.06845533847808838,
0.016067907214164734,
0.03491288423538208,
0.037670206278562546,
0.05518732964992523,
-0.010924646630883217,
-0.05451207607984543,
0.04959423840045929,
-0.04648914933204651,
0.018870333209633827,
-0.009834113530814648,
-0.006008253898471594,
-0.0248686783015728,
0.08045197278261185,
-0.014424417167901993,
0.008959689177572727,
0.08359769731760025,
0.054327595978975296,
0.02637816220521927,
0.01602795720100403,
-0.0048347474075853825,
-0.007275567855685949,
-0.021840278059244156,
-0.0802563726902008,
0.015926862135529518,
-0.036567989736795425,
-0.026315337046980858,
-0.08043485879898071,
0.05026562884449959,
0.01600763015449047,
0.1225997656583786,
-0.013020590879023075,
-0.03592658415436745,
-0.046182382851839066,
-0.04692057520151138,
0.041503556072711945,
0.06326618045568466,
0.031297434121370316,
-0.0780632495880127,
0.0060577294789254665,
0.015007499605417252,
-0.011007649824023247,
0.04637986049056053,
-0.0016647716984152794,
-0.04417363926768303,
-0.003266597166657448,
-0.10346473008394241,
-0.029881808906793594,
0.02300753816962242,
-0.02121168188750744,
-0.017915181815624237,
0.054846540093421936,
0.02466323971748352,
0.04587214067578316,
-0.04913860559463501,
-0.041661784052848816,
3.905737937660442e-33,
0.05816204845905304,
0.06320323795080185,
0.04222675785422325,
-0.03681844100356102,
0.03783246874809265,
-0.03654668852686882,
0.04087786749005318,
0.047971297055482864,
-0.03444080799818039,
-0.05366094410419464,
0.08040865510702133,
-0.08675505965948105,
-0.08255331963300705,
0.04460262134671211,
0.017959019169211388,
-0.03715784102678299,
0.019245712086558342,
0.047935716807842255,
-0.06935123354196548,
-0.03060678206384182,
0.055579934269189835,
0.0007356366841122508,
0.015790661796927452,
-0.0899321585893631,
0.08579006791114807,
-0.01562926173210144,
0.044896237552165985,
0.05841624364256859,
0.009623926132917404,
0.005478746257722378,
-0.09538984298706055,
-0.015453796833753586,
0.030757097527384758,
0.0949445366859436,
-0.02110612764954567,
-0.015889590606093407,
-0.0036995538976043463,
-0.08859807252883911,
0.03415963798761368,
0.010681974701583385,
0.04447420313954353,
0.0034860537853091955,
0.06241742894053459,
-0.013224795460700989,
-0.030223075300455093,
-0.036323584616184235,
0.05342327058315277,
-0.032651957124471664,
-0.06632232666015625,
-0.00031673695775680244,
0.006463055033236742,
0.0858617052435875,
-0.0374789796769619,
-0.06112989783287048,
-0.022806966677308083,
0.03668845072388649,
-0.014191470108926296,
0.06905853003263474,
-0.02632453478872776,
0.05660322681069374,
0.024931099265813828,
0.0646706372499466,
-0.03667360916733742,
0.12545107305049896,
0.05299443006515503,
0.06787854433059692,
0.002478070091456175,
0.0038727938663214445,
0.09613153338432312,
-0.023905381560325623,
-0.05199999362230301,
-0.034758299589157104,
0.008423099294304848,
-0.045309532433748245,
0.1291453093290329,
-0.05920923501253128,
0.021783336997032166,
-0.07974233478307724,
0.03987311199307442,
-0.041425611823797226,
0.0014187254710122943,
-0.031785428524017334,
-0.049158647656440735,
0.0028002311009913683,
0.004587567411363125,
-0.1140376627445221,
0.055491115897893906,
-0.0467301681637764,
0.0232772845774889,
-0.051366664469242096,
-0.04905719682574272,
-0.06689705699682236,
-0.002411803463473916,
0.07670528441667557,
0.029291795566678047,
-4.010346649308056e-33,
-0.015519731678068638,
-0.03736025467514992,
0.0024613665882498026,
0.0659172534942627,
-0.07030443847179413,
-0.03911815956234932,
0.018932273611426353,
-0.03312300145626068,
-0.002203269861638546,
-0.08976773917675018,
-0.033119309693574905,
-0.07548964023590088,
-0.016629643738269806,
0.00969546940177679,
0.00196545640937984,
-0.03433793783187866,
-0.03064749389886856,
-0.03899947926402092,
0.05365200713276863,
0.017773738130927086,
0.05293272063136101,
0.08782089501619339,
-0.014166691340506077,
0.09623744338750839,
-0.07593114674091339,
0.029391437768936157,
0.017870334908366203,
0.07663097232580185,
0.057709984481334686,
-0.003422088921070099,
0.003236837452277541,
-0.05684412270784378,
0.06370419263839722,
-0.047897834330797195,
-0.02317066304385662,
0.061631783843040466,
0.05235927551984787,
0.04245645925402641,
-0.0742245614528656,
0.1348555088043213,
0.014834064058959484,
-0.011875665746629238,
-0.03973542898893356,
-0.00630365451797843,
-0.0268411822617054,
-0.0060959444381296635,
-0.11351313441991806,
-0.037276796996593475,
0.008063663728535175,
-0.026486128568649292,
-0.04299177601933479,
0.010720310732722282,
-0.07407095283269882,
-0.03434264659881592,
-0.06563738733530045,
-0.11358579248189926,
0.007547176908701658,
-0.039377421140670776,
0.04715657979249954,
-0.012652809731662273,
-0.08818306028842926,
0.010272370651364326,
0.03587953373789787,
0.00946483202278614,
0.040471095591783524,
-0.06512674689292908,
-0.023486129939556122,
0.0074107227846980095,
0.026053769513964653,
0.014018326997756958,
0.07560953497886658,
0.033622607588768005,
0.004071422386914492,
-0.003944672644138336,
0.03582566976547241,
-0.04850774258375168,
-0.010323475115001202,
-0.040530975908041,
-0.013789179734885693,
-0.052717287093400955,
0.04154585301876068,
0.007046112325042486,
0.007350321859121323,
0.0494198203086853,
0.05445927381515503,
0.08888166397809982,
-0.012820389121770859,
0.11344091594219208,
-0.04613623395562172,
-0.0320916473865509,
-0.05550382286310196,
0.0013169439043849707,
0.0810927301645279,
0.0579923652112484,
-0.053760867565870285,
-4.896608984950035e-8,
-0.08141610771417618,
0.029165184125304222,
0.09083238244056702,
0.05634738877415657,
0.018284639343619347,
-0.07359201461076736,
0.01594134233891964,
0.009555472992360592,
-0.038609929382801056,
-0.012040385976433754,
0.0021668116096407175,
0.025151824578642845,
0.03666520118713379,
-0.050211355090141296,
-0.013549664989113808,
0.03388699144124985,
0.027038663625717163,
-0.0016644889255985618,
0.008001775480806828,
0.024448782205581665,
0.009966370649635792,
-0.002105128951370716,
0.06203626096248627,
-0.06439689546823502,
0.0716097429394722,
-0.01645701937377453,
0.02444075606763363,
0.06844539940357208,
0.05042967572808266,
0.03506030514836311,
0.02575303055346012,
-0.04027175158262253,
0.02301889657974243,
0.023495664820075035,
0.027755463495850563,
0.012180399149656296,
0.07577171176671982,
-0.09899573028087616,
-0.014069225639104843,
-0.0012358642416074872,
0.016456447541713715,
0.10103056579828262,
-0.047893863171339035,
0.02905663289129734,
0.04256248101592064,
0.0023275308776646852,
-0.022619768977165222,
-0.15871480107307434,
0.008866426534950733,
-0.14002755284309387,
-0.020551588386297226,
-0.045903611928224564,
-0.0780162364244461,
0.04416865110397339,
0.12060652673244476,
0.0490175224840641,
-0.03372650593519211,
-0.010135123506188393,
-0.013467641547322273,
0.1185036152601242,
0.04111633449792862,
0.003605055855587125,
-0.029899336397647858,
-0.01770290918648243
] |
uniINF: Best-of-Both-Worlds Algorithm for Parameter-Free Heavy-Tailed MABs | https://openreview.net/forum?id=2pNLknCTvG | [
"Yu Chen",
"Jiatai Huang",
"Yan Dai",
"Longbo Huang"
] | Spotlight | In this paper, we present a novel algorithm, `uniINF`, for the Heavy-Tailed Multi-Armed Bandits (HTMAB) problem, demonstrating robustness and adaptability in both stochastic and adversarial environments. Unlike the stochastic MAB setting where loss distributions are stationary with time, our study extends to the adversarial setup, where losses are generated from heavy-tailed distributions that depend on both arms and time. Our novel algorithm `uniINF` enjoys the so-called Best-of-Both-Worlds (BoBW) property, performing optimally in both stochastic and adversarial environments *without* knowing the exact environment type. Moreover, our algorithm also possesses a Parameter-Free feature, *i.e.*, it operates *without* the need of knowing the heavy-tail parameters $(\sigma, \alpha)$ a-priori.
To be precise, `uniINF` ensures nearly-optimal regret in both stochastic and adversarial environments, matching the corresponding lower bounds when $(\sigma, \alpha)$ is known (up to logarithmic factors). To our knowledge, `uniINF` is the first parameter-free algorithm to achieve the BoBW property for the heavy-tailed MAB problem. Technically, we develop innovative techniques to achieve BoBW guarantees for Parameter-Free HTMABs, including a refined analysis for the dynamics of log-barrier, an auto-balancing learning rate scheduling scheme, an adaptive skipping-clipping loss tuning technique, and a stopping-time analysis for logarithmic regret. | Heavy Tailed, Multi-Armed Bandits, Parameter-Free, Best-of-Both-Worlds | null | 9,301 | 2410.03284 | [
-0.09279125183820724,
-0.03262387588620186,
-0.018094999715685844,
0.0024323677644133568,
0.07991654425859451,
-0.019005876034498215,
0.03138066828250885,
-0.015339489094913006,
-0.035702917724847794,
0.041025158017873764,
-0.05008146911859512,
-0.03914657235145569,
0.06520093232393265,
-0.04729501157999039,
-0.008596229366958141,
0.019439127296209335,
0.05790197104215622,
0.019031230360269547,
-0.04755297303199768,
-0.001570283668115735,
-0.035977091640233994,
0.041548456996679306,
0.04855391010642052,
-0.003439731663092971,
-0.008419484831392765,
-0.10457494109869003,
-0.014084967784583569,
-0.01074080541729927,
-0.03539900481700897,
-0.03800123184919357,
0.03535657003521919,
-0.0027719559147953987,
-0.0202894676476717,
-0.05126401036977768,
-0.045695431530475616,
0.06707783788442612,
-0.12281369417905807,
-0.014698712155222893,
0.014214199967682362,
-0.04352400824427605,
0.022965412586927414,
0.01895877532660961,
-0.03952881321310997,
-0.014875064603984356,
-0.03454352542757988,
-0.03994095325469971,
0.008292126469314098,
0.018158694729208946,
0.06423501670360565,
0.02077753283083439,
-0.009149529039859772,
-0.0010852148989215493,
-0.050169117748737335,
0.03045646846294403,
-0.007020356599241495,
0.0497150681912899,
0.032214391976594925,
0.02968640811741352,
-0.005775519646704197,
0.048218052834272385,
-0.0005883529665879905,
-0.12350073456764221,
-0.0035019507631659508,
-0.035769909620285034,
0.07246291637420654,
-0.04805818572640419,
0.002886832458898425,
0.06134359538555145,
-0.044588830322027206,
0.10391059517860413,
0.046900779008865356,
0.0014701802283525467,
-0.11221005022525787,
0.0369117371737957,
0.031846605241298676,
0.07440169155597687,
0.04476521536707878,
-0.05177817866206169,
-0.01968281716108322,
0.04228377714753151,
0.013947157189249992,
0.00740027753636241,
0.029395287856459618,
-0.012764732353389263,
0.08857196569442749,
-0.016075506806373596,
-0.0535370334982872,
-0.05243530124425888,
0.22140412032604218,
-0.05552568659186363,
-0.044327329844236374,
-0.03385268524289131,
-0.031920772045850754,
0.05994946137070656,
0.06381835788488388,
-0.032061200588941574,
0.0611637681722641,
0.01063258945941925,
-0.06681699305772781,
0.14961226284503937,
0.0021277759224176407,
0.004435480106621981,
0.04737566038966179,
-0.038275301456451416,
0.08762209862470627,
-0.03912772983312607,
0.016746101900935173,
-0.008567475713789463,
0.04500187188386917,
-0.06995871663093567,
0.020184479653835297,
0.04262761026620865,
0.06901999562978745,
-0.007170440629124641,
-0.04455427825450897,
0.006502991542220116,
-0.02615855634212494,
0.03040909767150879,
-0.11220989376306534,
0.0643366202712059,
-0.03481750190258026,
0.016517141833901405,
0.06889626383781433,
-0.00784782599657774,
0.030817991122603416,
0.04074579477310181,
-0.006674408912658691,
-2.158920271559906e-34,
0.02465333789587021,
-0.02472233586013317,
0.04846147820353508,
-0.10270939767360687,
0.031880829483270645,
0.02089325152337551,
0.03159964829683304,
0.006945781875401735,
-0.04723449796438217,
0.003106927964836359,
0.017408955842256546,
0.030197588726878166,
-0.11595295369625092,
0.06656599044799805,
0.10283760726451874,
-0.03122534602880478,
0.03634849190711975,
0.008988110348582268,
0.07600095123052597,
-0.008403090760111809,
0.10527939349412918,
-0.03323381021618843,
-0.06915438920259476,
-0.09259157627820969,
-0.012368196621537209,
-0.05766376107931137,
0.006126469932496548,
-0.02092936635017395,
0.016604963690042496,
0.06722865998744965,
-0.10844040662050247,
0.008441193960607052,
-0.07042387127876282,
-0.008929726667702198,
0.04269944503903389,
-0.0023059931118041277,
-0.15050937235355377,
-0.02118939347565174,
-0.04218561202287674,
-0.025630323216319084,
-0.010962622240185738,
0.031801506876945496,
0.040922366082668304,
0.007300228346139193,
-0.045378319919109344,
-0.12050522118806839,
0.008105414919555187,
0.024400874972343445,
-0.06443094462156296,
-0.08182673901319504,
-0.03660358116030693,
0.00601992616429925,
0.022911034524440765,
-0.02560976892709732,
-0.008540929295122623,
0.013445823453366756,
0.0823085606098175,
0.05641188472509384,
-0.02544569782912731,
0.097272127866745,
0.012863341718912125,
-0.0041344743221998215,
-0.0034170160070061684,
0.025878261774778366,
-0.022713977843523026,
0.032222699373960495,
-0.00886258203536272,
0.033164262771606445,
0.05800464749336243,
-0.04944545403122902,
0.027573812752962112,
0.012753648683428764,
0.06804902851581573,
-0.04210694134235382,
0.05163809657096863,
-0.023397553712129593,
0.11606808751821518,
0.013547898270189762,
0.06276607513427734,
-0.08306922763586044,
-0.014838453382253647,
0.026948213577270508,
-0.03024757094681263,
0.02370275743305683,
-0.045845478773117065,
-0.02003760263323784,
0.00836834218353033,
-0.11074601113796234,
-0.05780647322535515,
-0.0168023519217968,
-0.07887884229421616,
0.02459804341197014,
-0.08548843860626221,
-0.01709110662341118,
0.015641797333955765,
-1.024903330741879e-33,
-0.02159915678203106,
-0.059750884771347046,
0.032238882035017014,
0.02043403685092926,
0.024486176669597626,
-0.03242552652955055,
-0.05793197453022003,
0.00038294127443805337,
-0.05057645961642265,
-0.005123637616634369,
-0.0854819193482399,
0.05496140196919441,
-0.010485773906111717,
-0.04100285470485687,
0.0644669309258461,
-0.0405595600605011,
0.06723800301551819,
0.08293350040912628,
-0.005338180810213089,
-0.026371072977781296,
0.07285013049840927,
0.06184179708361626,
-0.03462396562099457,
0.0338297039270401,
0.021160373464226723,
-0.04872926324605942,
-0.018162600696086884,
0.09663105756044388,
-0.039159372448921204,
0.022688208147883415,
-0.08835878223180771,
0.028326036408543587,
-0.07597444206476212,
-0.01286135334521532,
-0.046748772263526917,
0.003998982720077038,
0.07925371080636978,
0.0690954402089119,
-0.06674382835626602,
0.1185683161020279,
0.039262738078832626,
-0.007746266666799784,
-0.008369151502847672,
-0.050023097544908524,
0.001000169781036675,
0.043678417801856995,
-0.08733141422271729,
0.000723924720659852,
0.09888894855976105,
-0.010029343888163567,
0.0443742461502552,
0.02553982473909855,
-0.0285060852766037,
0.10940411686897278,
-0.027985811233520508,
-0.10532023757696152,
-0.005716520361602306,
-0.03699757158756256,
0.09220332652330399,
0.05906442552804947,
-0.018728550523519516,
0.0028628830332309008,
-0.1112181693315506,
0.059297990053892136,
-0.02766365371644497,
-0.03573458641767502,
-0.035423628985881805,
-0.0014526288723573089,
0.009562397375702858,
-0.01727544143795967,
-0.02666565030813217,
0.038348257541656494,
-0.024560140445828438,
0.01482678484171629,
-0.030750634148716927,
0.0453968308866024,
0.008482338860630989,
-0.017486875876784325,
0.034886401146650314,
-0.06025942414999008,
0.005512353498488665,
-0.006466179620474577,
0.05674304440617561,
0.02558102458715439,
0.05847791209816933,
0.022481635212898254,
0.048997342586517334,
-0.029355766251683235,
-0.004640981089323759,
-0.03802594542503357,
0.02672160044312477,
0.005483478307723999,
-0.013019583187997341,
0.01742134802043438,
-0.06231173127889633,
-4.719491286664379e-8,
-0.025445975363254547,
0.02723698504269123,
0.07017683982849121,
-0.015272886492311954,
0.0927559956908226,
0.009592002257704735,
-0.04914254695177078,
0.0161170344799757,
-0.0055356272496283054,
0.0036966949701309204,
0.07388303428888321,
0.03380695730447769,
0.01961120031774044,
-0.0009767961455509067,
-0.08904477208852768,
0.025259584188461304,
-0.046616241335868835,
-0.00699171656742692,
-0.028713520616292953,
0.0063034771010279655,
0.021583396941423416,
0.017895318567752838,
0.05192868039011955,
-0.06470995396375656,
0.0305795781314373,
-0.06408458203077316,
0.0095303850248456,
0.005858023650944233,
0.056926432996988297,
0.12174627184867859,
-0.09276483207941055,
-0.0410640686750412,
-0.006038804072886705,
-0.002083010273054242,
-0.009741353802382946,
0.12609098851680756,
-0.06161161884665489,
-0.04306803271174431,
-0.05288507044315338,
-0.007636735215783119,
0.008598030544817448,
0.06670929491519928,
0.0036373273469507694,
-0.02705611288547516,
0.04900825768709183,
0.011627556756138802,
0.03742789477109909,
-0.027752630412578583,
0.07739909738302231,
-0.0749683827161789,
0.03389019891619682,
-0.029133329167962074,
0.044449396431446075,
0.048052046447992325,
0.06356510519981384,
0.033963389694690704,
-0.005585516337305307,
-0.08785208314657211,
0.018935365602374077,
0.05441264435648918,
0.06240980327129364,
-0.036414846777915955,
-0.08055514842271805,
0.019607558846473694
] |
Exploring Local Memorization in Diffusion Models via Bright Ending Attention | https://openreview.net/forum?id=p4cLtzk4oe | [
"Chen Chen",
"Daochang Liu",
"Mubarak Shah",
"Chang Xu"
] | Spotlight | Text-to-image diffusion models have achieved unprecedented proficiency in generating realistic images. However, their inherent tendency to memorize and replicate training data during inference raises significant concerns, including potential copyright infringement. In response, various methods have been proposed to evaluate, detect, and mitigate memorization. Our analysis reveals that existing approaches significantly underperform in handling local memorization, where only specific image regions are memorized, compared to global memorization, where the entire image is replicated. Also, they cannot locate the local memorization regions, making it hard to investigate locally. To address these, we identify a novel "bright ending" (BE) anomaly in diffusion models prone to memorizing training images. BE refers to a distinct cross-attention pattern observed in text-to-image diffusion models, where memorized image patches exhibit significantly greater attention to the final text token during the last inference step than non-memorized patches. This pattern highlights regions where the generated image replicates training data and enables efficient localization of memorized regions. Equipped with this, we propose a simple yet effective method to integrate BE into existing frameworks, significantly improving their performance by narrowing the performance gap caused by local memorization. Our results not only validate the successful execution of the new localization task but also establish new state-of-the-art performance across all existing tasks, underscoring the significance of the BE phenomenon. | Diffusion Models, Local Memorization, Bright Ending Attention | null | 9,283 | 2410.21665 | [
0.03355773165822029,
-0.07962819933891296,
0.017934439703822136,
0.06312672793865204,
0.12954330444335938,
-0.01727382466197014,
-0.02524862252175808,
-0.03436613827943802,
0.10424255579710007,
-0.05225076898932457,
0.04619009047746658,
0.057687677443027496,
0.04097902774810791,
0.02718646638095379,
-0.024731097742915154,
-0.022866450250148773,
0.03644716367125511,
0.054871536791324615,
-0.08708560466766357,
-0.10928982496261597,
0.007757925894111395,
-0.034889210015535355,
0.13887232542037964,
0.010147967375814915,
0.026875440031290054,
0.04877447336912155,
-0.01793949492275715,
-0.03153763711452484,
0.033433299511671066,
-0.03516722470521927,
0.05867784097790718,
0.029719870537519455,
-0.0015607320237904787,
0.05826224014163017,
-0.028863366693258286,
0.04614182934165001,
-0.11385364830493927,
0.004436261486262083,
0.016811270266771317,
-0.059331607073545456,
0.015245859511196613,
-0.015740327537059784,
-0.0009664399549365044,
0.05231720209121704,
0.05262739211320877,
-0.0032507269643247128,
0.01855073682963848,
0.0221632719039917,
-0.013070035725831985,
-0.0022417926229536533,
-0.038748711347579956,
-0.053933221846818924,
-0.11675828695297241,
0.0323931984603405,
-0.04146880656480789,
0.04459817335009575,
0.09549533575773239,
0.044827837496995926,
0.02345959283411503,
0.011189858429133892,
-0.07828232645988464,
-0.045156486332416534,
-0.039825230836868286,
-0.032808635383844376,
-0.03458678722381592,
0.021430950611829758,
0.04200716316699982,
0.06649883836507797,
-0.09478753805160522,
-0.005094524938613176,
0.05162271857261658,
0.07843860238790512,
-0.06717867404222488,
0.0259786918759346,
0.040820661932229996,
0.027043107897043228,
-0.009515268728137016,
-0.043717287480831146,
0.01726689375936985,
-0.09620082378387451,
0.02624870091676712,
-0.04687308147549629,
0.1127668097615242,
-0.038839515298604965,
0.1266944855451584,
0.005083967931568623,
0.05868842080235481,
0.030528586357831955,
0.028968123719096184,
-0.04365377128124237,
-0.06690867990255356,
-0.022679375484585762,
-0.04137425869703293,
-0.03635884076356888,
0.003161580301821232,
0.04052235931158066,
-0.011774713173508644,
-0.020889639854431152,
-0.00031301291892305017,
0.05357151851058006,
0.00921434536576271,
-0.021511336788535118,
-0.007264475338160992,
-0.020183639600872993,
0.08691257238388062,
-0.020978519693017006,
0.09604109823703766,
0.04582681879401207,
0.06386101990938187,
-0.014835086651146412,
0.08958763629198074,
0.03306133300065994,
-0.028318146243691444,
0.010746710002422333,
0.06495271623134613,
-0.09230422973632812,
0.05761340633034706,
-0.011983831413090229,
-0.0061582052148878574,
0.07292807102203369,
0.05984613671898842,
0.02317766100168228,
-0.06912950426340103,
0.0009139123139902949,
-0.01225539855659008,
-0.08691069483757019,
-0.02270757406949997,
1.633118893372212e-33,
0.04295440763235092,
-0.011845572851598263,
-0.011934531852602959,
0.06588957458734512,
0.07117772847414017,
-0.017995798960328102,
-0.004542461130768061,
-0.013205072842538357,
-0.08868463337421417,
-0.09075701236724854,
-0.01896817423403263,
0.004646974615752697,
-0.03682524338364601,
0.1206812933087349,
0.020456552505493164,
-0.04476155340671539,
-0.04879319667816162,
0.04473360255360603,
-0.03033517114818096,
0.03767988458275795,
-0.0007397672743536532,
-0.019644640386104584,
0.051513779908418655,
-0.07952248305082321,
-0.04896370321512222,
-0.023758191615343094,
0.01788952760398388,
-0.0669093206524849,
0.03143783286213875,
0.01851491816341877,
-0.11516408622264862,
-0.008581078611314297,
0.00779388565570116,
0.07003982365131378,
0.004006451461464167,
-0.011952878907322884,
0.022312669083476067,
-0.03964726999402046,
0.024335235357284546,
-0.01960141398012638,
-0.044591907411813736,
0.06949672847986221,
-0.02906409651041031,
-0.02636796049773693,
-0.126749187707901,
0.0003995326696895063,
0.04222318157553673,
0.022488411515951157,
0.025738369673490524,
-0.017842521890997887,
0.011159971356391907,
-0.02893141657114029,
-0.015473824925720692,
-0.0757284015417099,
-0.05589059740304947,
-0.007671455852687359,
0.06705489754676819,
-0.052809830754995346,
0.027648670598864555,
-0.007305949926376343,
0.059996407479047775,
0.08811862021684647,
-0.02260868065059185,
0.10549203306436539,
-0.007494048215448856,
0.0013629820896312594,
-0.008585828356444836,
0.07839410752058029,
0.03279302641749382,
0.025769876316189766,
-0.026284344494342804,
-0.003084559692069888,
-0.03609592467546463,
-0.0435284823179245,
0.01853122189640999,
-0.046165548264980316,
0.05283829942345619,
-0.0014797091716900468,
0.037190817296504974,
-0.024990027770400047,
0.00835131760686636,
-0.08921041339635849,
-0.01472196914255619,
-0.06084202229976654,
-0.08570974320173264,
0.017146429046988487,
0.010775119997560978,
-0.13425768911838531,
-0.0833270400762558,
-0.008610920049250126,
0.009247714653611183,
-0.00825012382119894,
-0.012792566791176796,
-0.02273630164563656,
-0.02304035797715187,
-1.853450432083472e-33,
-0.012525362893939018,
0.0765899270772934,
0.025274639949202538,
0.0602116696536541,
-0.04867253452539444,
-0.008375964127480984,
0.029423316940665245,
0.0622791163623333,
-0.013158675283193588,
-0.09397301822900772,
-0.01843395084142685,
-0.010252149775624275,
-0.06033727154135704,
-0.02127736248075962,
-0.03690160810947418,
-0.11084397882223129,
0.08132985234260559,
0.06084217503666878,
-0.06257051974534988,
0.0113872941583395,
-0.01921248435974121,
0.011000851169228554,
-0.024486001580953598,
0.028723785653710365,
-0.052359215915203094,
0.06445696204900742,
-0.07346281409263611,
0.022054003551602364,
0.011718322522938251,
-0.07508920878171921,
-0.07196009159088135,
0.0415036715567112,
-0.02192927710711956,
-0.03325928747653961,
-0.08462976664304733,
0.07843559980392456,
0.029490428045392036,
0.010288071818649769,
-0.027060983702540398,
0.0809076800942421,
0.08317963778972626,
-0.03925004601478577,
-0.0664534717798233,
0.05639395862817764,
-0.04553276300430298,
-0.10250964760780334,
-0.08572057634592056,
0.06397776305675507,
-0.0033508494962006807,
0.03867171332240105,
-0.01859639212489128,
-0.002538662636652589,
-0.08887343853712082,
0.010095706209540367,
-0.1234118714928627,
-0.020528173074126244,
-0.07637472450733185,
-0.019270481541752815,
0.06066490709781647,
-0.04217041656374931,
-0.027951238676905632,
-0.004956965334713459,
-0.03298381716012955,
-0.10157612711191177,
-0.05043556168675423,
-0.07455487549304962,
-0.016038790345191956,
0.006994547322392464,
0.007405182812362909,
0.08129442483186722,
0.020503606647253036,
-0.024248600006103516,
-0.04466189816594124,
-0.011493765749037266,
0.00728018069639802,
-0.012915367260575294,
-0.003773801028728485,
-0.00564579525962472,
-0.03795960918068886,
-0.05845910683274269,
0.04990303888916969,
0.035003770142793655,
-0.03720977529883385,
0.05218455567955971,
0.06880960613489151,
0.052037905901670456,
0.017808059230446815,
-0.00332471402361989,
0.023677997291088104,
-0.004151500761508942,
-0.024951176717877388,
0.014493634924292564,
-0.009542456828057766,
0.06887742877006531,
-0.053520895540714264,
-3.8518688683097935e-8,
-0.08961563557386398,
-0.005970061291009188,
-0.0020632643718272448,
0.03685443475842476,
0.02908124029636383,
-0.04058040305972099,
0.02628256380558014,
0.1377975344657898,
0.006231233011931181,
-0.05931846797466278,
0.06343497335910797,
-0.03401980921626091,
-0.05868229269981384,
-0.06701166182756424,
-0.05054837465286255,
0.10315153747797012,
0.05875155329704285,
0.04107069596648216,
-0.020270593464374542,
-0.01641290821135044,
-0.07112560421228409,
0.02803436852991581,
0.08808164298534393,
-0.06883248686790466,
-0.002141565317288041,
-0.026164544746279716,
-0.03296451270580292,
0.0691833347082138,
0.06458008289337158,
-0.03990495204925537,
0.0401175394654274,
-0.008596111088991165,
-0.011990584433078766,
0.016969040036201477,
0.02986813150346279,
0.0654013529419899,
0.05780943110585213,
0.032041847705841064,
-0.009874800220131874,
-0.0095760403200984,
0.035518161952495575,
0.018060855567455292,
0.04758178070187569,
0.07797211408615112,
0.0014812161680310965,
0.004476851783692837,
0.0692124292254448,
-0.05748958885669708,
0.047358233481645584,
-0.017286740243434906,
0.05245395004749298,
-0.03529173135757446,
-0.04736756905913353,
0.06689387559890747,
0.008373401127755642,
-0.06798635423183441,
0.07066100090742111,
0.042206596583127975,
0.027446212247014046,
0.09020372480154037,
0.07765233516693115,
0.02097412198781967,
-0.024437086656689644,
0.023665832355618477
] |
EmbedLLM: Learning Compact Representations of Large Language Models | https://openreview.net/forum?id=Fs9EabmQrJ | [
"Richard Zhuang",
"Tianhao Wu",
"Zhaojin Wen",
"Andrew Li",
"Jiantao Jiao",
"Kannan Ramchandran"
] | Spotlight | With hundreds of thousands of language models available on Huggingface today, efficiently evaluating and utilizing these models across various downstream tasks has become increasingly critical. Many existing methods repeatedly learn task-specific representations of Large Language Models (LLMs), which leads to inefficiencies in both time and computational resources. To address this, we propose EmbedLLM, a framework designed to learn compact vector representations of LLMs that facilitate downstream applications involving many models, such as model routing. We introduce an encoder-decoder approach for learning such embedding, along with a systematic framework to evaluate their effectiveness. Empirical results show that EmbedLLM outperforms prior methods in model routing. Additionally, we demonstrate that our method can forecast a model's performance on multiple benchmarks, without incurring additional inference cost. Extensive probing experiments validate that the learned embeddings capture key model characteristics, e.g. whether the model is specialized for coding tasks, even without being explicitly trained on them. We open source our dataset, code and embedder to facilitate further research and application. | Large Language Models, Representation Learning, Model Routing | null | 9,255 | 2410.02223 | [
-0.02798040211200714,
-0.18391132354736328,
0.037606336176395416,
0.06512275338172913,
0.06251940131187439,
0.028927618637681007,
-0.07273084670305252,
0.003024798585101962,
0.0036037154495716095,
-0.04280128702521324,
-0.08339272439479828,
0.0038744104094803333,
0.03213394805788994,
0.025422925129532814,
-0.0437704436480999,
0.07489240914583206,
0.12010583281517029,
0.09053730964660645,
-0.020737137645483017,
-0.06783168017864227,
0.003642158815637231,
-0.008391049690544605,
0.04716802015900612,
-0.03132729232311249,
0.059925954788923264,
-0.015457346104085445,
-0.004017180297523737,
0.007656927220523357,
0.026298532262444496,
0.008880392648279667,
0.0879548117518425,
-0.019138650968670845,
-0.08457093685865402,
0.06292792409658432,
-0.028723454102873802,
0.06177784502506256,
-0.07484131306409836,
-0.0224023275077343,
-0.07040469348430634,
-0.040469955652952194,
-0.006807140540331602,
-0.006752423942089081,
-0.006762369070202112,
0.029342366382479668,
0.027341414242982864,
-0.03186878189444542,
0.016015999019145966,
-0.0824456438422203,
-0.04746711999177933,
0.05710763856768608,
-0.04070965573191643,
-0.02904306724667549,
-0.0190565325319767,
0.03773822635412216,
-0.012033752165734768,
-0.025504963472485542,
0.020058264955878258,
-0.022999094799160957,
-0.06605610996484756,
-0.013032844290137291,
-0.05388246104121208,
-0.16323629021644592,
0.021739494055509567,
-0.02744273841381073,
-0.0900721549987793,
-0.03083731420338154,
0.023486191406846046,
0.10705140978097916,
-0.057048700749874115,
0.059211164712905884,
-0.03598553687334061,
0.033960144966840744,
-0.034144796431064606,
0.06024450808763504,
0.0038875832688063383,
0.00025757838739082217,
0.05512640252709389,
-0.04960145056247711,
0.0900174155831337,
-0.0685078352689743,
-0.017680391669273376,
0.020685799419879913,
0.0001343615003861487,
-0.0087239108979702,
0.18065212666988373,
0.014423437416553497,
0.08703657239675522,
0.02724299393594265,
-0.003540357109159231,
0.026069987565279007,
-0.025200674310326576,
-0.06237059459090233,
0.0788283497095108,
0.0029151923954486847,
0.0033872900530695915,
0.044407352805137634,
0.009884699247777462,
-0.043382707983255386,
-0.029342886060476303,
0.04748135432600975,
-0.028325287625193596,
0.11771414428949356,
0.028296807780861855,
-0.07411482185125351,
-0.005070139653980732,
0.0020029000006616116,
0.006096815224736929,
0.10937865823507309,
-0.025173768401145935,
-0.0722164586186409,
-0.03245119750499725,
0.041056156158447266,
0.030004505068063736,
0.06617985665798187,
0.040287718176841736,
-0.03381543979048729,
0.018654096871614456,
-0.06796509772539139,
0.007391099818050861,
0.09837016463279724,
-0.06749434024095535,
0.07727959752082825,
0.016270119696855545,
-0.004857472609728575,
-0.07308577001094818,
-0.056433673948049545,
-0.0012090792879462242,
9.456365324235666e-34,
0.06536176800727844,
0.030901912599802017,
0.022351570427417755,
0.05474579334259033,
0.07211361080408096,
0.006953657604753971,
0.025893522426486015,
0.0924709364771843,
-0.06174330785870552,
0.008501843549311161,
-0.06765207648277283,
-0.0020366956014186144,
0.014617274515330791,
0.07014159113168716,
-0.041995614767074585,
-0.027119001373648643,
-0.06977637857198715,
0.0217710230499506,
-0.03336743637919426,
-0.050230711698532104,
0.038765862584114075,
-0.04828429967164993,
-0.0286580640822649,
-0.014995737932622433,
0.029192626476287842,
0.022367868572473526,
0.09822561591863632,
-0.03255915641784668,
-0.042795579880476,
-0.004746588878333569,
-0.10436698794364929,
-0.05200915411114693,
0.026816433295607567,
0.04193712770938873,
0.04290397837758064,
-0.02108786441385746,
0.030307475477457047,
-0.00593042466789484,
-0.0044400691986083984,
-0.016121800988912582,
-0.002254484687000513,
-0.0004129412118345499,
-0.023312505334615707,
-0.033454280346632004,
-0.10792222619056702,
-0.10095828026533127,
-0.02569306641817093,
-0.02259029820561409,
-0.017179105430841446,
-0.02640078030526638,
0.10300561785697937,
-0.008668052032589912,
0.000640909478534013,
-0.0231794323772192,
0.024035155773162842,
-0.07455981522798538,
0.0019870970863848925,
0.050870612263679504,
0.02895190194249153,
0.001803408726118505,
-0.004894391633570194,
-0.0007683951989747584,
0.025424540042877197,
0.04034636914730072,
0.056104790419340134,
0.00817304477095604,
-0.024131659418344498,
0.12637606263160706,
0.008561265654861927,
-0.03463584557175636,
-0.01364624872803688,
0.04137350246310234,
0.009944810532033443,
-0.037155911326408386,
0.057895343750715256,
0.02267725020647049,
0.0001900232891784981,
-0.17996391654014587,
-0.005989554338157177,
0.04259949550032616,
0.012132187373936176,
-0.019562218338251114,
0.03272373601794243,
-0.0477605015039444,
-0.023040669038891792,
-0.07918187975883484,
0.015921201556921005,
-0.13216596841812134,
0.03237733617424965,
0.049197956919670105,
-0.08197769522666931,
-0.08373410254716873,
-0.04523187875747681,
-0.016782747581601143,
0.0391862578690052,
-1.1417572956117032e-33,
0.016563046723604202,
0.041190680116415024,
-0.016461947932839394,
0.03618764132261276,
-0.046571120619773865,
-0.08020444959402084,
0.0010963284876197577,
0.06449281424283981,
-0.00892944447696209,
-0.05571582168340683,
-0.022391675040125847,
-0.020247094333171844,
0.11316855251789093,
-0.023504221811890602,
0.10495609045028687,
-0.02554074488580227,
-0.02562974952161312,
-0.018499765545129776,
0.005658050067722797,
0.01985214091837406,
-0.09098631143569946,
0.047994405031204224,
-0.07007215917110443,
-0.006692822091281414,
0.024307768791913986,
0.08346467465162277,
0.030671775341033936,
0.04072939231991768,
-0.05892021581530571,
-0.018844911828637123,
-0.031971465796232224,
-0.045406293123960495,
0.008897144347429276,
-0.05405430868268013,
-0.07532518357038498,
0.09673219174146652,
0.048307742923498154,
-0.010344011709094048,
-0.023908017203211784,
0.03636804595589638,
0.07530897110700607,
-0.05242065712809563,
-0.036719467490911484,
0.0181911438703537,
-0.03844179958105087,
-0.00007767000352032483,
-0.07482404261827469,
0.020456483587622643,
0.03669607266783714,
-0.02266789972782135,
-0.005885849706828594,
0.04904197156429291,
-0.0727754756808281,
0.012786252424120903,
0.02790156751871109,
-0.06643863022327423,
-0.05010558292269707,
-0.05012330040335655,
0.03938012570142746,
-0.006481601856648922,
0.0186463575810194,
0.009571747854351997,
0.04560006782412529,
-0.04447462782263756,
0.004467782098799944,
-0.08231592178344727,
-0.02180398814380169,
0.0008349122363142669,
-0.025722425431013107,
-0.05482574179768562,
0.05154997110366821,
-0.030912011861801147,
-0.02770974114537239,
0.08284419775009155,
0.027393635362386703,
-0.041639018803834915,
0.027417819947004318,
-0.011512578465044498,
-0.056507088243961334,
-0.03312836214900017,
0.04983256012201309,
-0.048518892377614975,
0.02474323846399784,
-0.030164752155542374,
0.08543661236763,
0.07840174436569214,
0.01439757738262415,
0.0637981966137886,
-0.008125431835651398,
0.023709366098046303,
-0.0032008588314056396,
0.03090711124241352,
-0.010934784077107906,
0.04552345722913742,
-0.007214543875306845,
-3.3966887968972514e-8,
-0.08790148049592972,
0.07440870255231857,
0.027292124927043915,
-0.00967575516551733,
0.011678321287035942,
0.024751150980591774,
0.012167025357484818,
0.10494011640548706,
-0.022060593590140343,
-0.007616053801029921,
0.06780634075403214,
0.0328550711274147,
-0.04476103559136391,
-0.051602303981781006,
0.01584123820066452,
0.06313549727201462,
0.018397338688373566,
-0.02888602577149868,
0.01333878468722105,
0.010318640619516373,
-0.010375918820500374,
0.06709519028663635,
0.03373432159423828,
0.0013339644065126777,
0.019276531413197517,
-0.032797809690237045,
0.003575626527890563,
0.11979607492685318,
0.09704697877168655,
-0.012725794687867165,
-0.06902088969945908,
0.036810439079999924,
0.0151798902079463,
-0.08550024032592773,
-0.018149107694625854,
0.12401612848043442,
0.03889081999659538,
-0.06476818770170212,
0.009725198149681091,
-0.061508603394031525,
0.07029745727777481,
0.039224348962306976,
-0.01701427809894085,
0.01899024285376072,
0.03263351693749428,
0.061225224286317825,
-0.02819790318608284,
-0.033694006502628326,
0.009414265863597393,
-0.023224877193570137,
-0.03471309319138527,
-0.015576575882732868,
-0.03477946296334267,
0.08680922538042068,
0.019608350470662117,
0.040365856140851974,
-0.06979411095380783,
-0.06363088637590408,
0.033603306859731674,
0.04534877836704254,
-0.03348521143198013,
0.014879154972732067,
-0.021605877205729485,
0.0363958440721035
] |
InverseBench: Benchmarking Plug-and-Play Diffusion Priors for Inverse Problems in Physical Sciences | https://openreview.net/forum?id=U3PBITXNG6 | [
"Hongkai Zheng",
"Wenda Chu",
"Bingliang Zhang",
"Zihui Wu",
"Austin Wang",
"Berthy Feng",
"Caifeng Zou",
"Yu Sun",
"Nikola Borislavov Kovachki",
"Zachary E Ross",
"Katherine Bouman",
"Yisong Yue"
] | Spotlight | Plug-and-play diffusion priors (PnPDP) have emerged as a promising research direction for solving inverse problems.
However, current studies primarily focus on natural image restoration, leaving the performance of these algorithms in scientific inverse problems largely unexplored. To address this gap, we introduce \textsc{InverseBench}, a framework that evaluates diffusion models across five distinct scientific inverse problems. These problems present unique structural challenges that differ from existing benchmarks, arising from critical scientific applications such as optical tomography, medical imaging, black hole imaging, seismology, and fluid dynamics. With \textsc{InverseBench}, we benchmark 14 inverse problem algorithms that use plug-and-play diffusion priors against strong, domain-specific baselines, offering valuable new insights into the strengths and weaknesses of existing algorithms. To facilitate further research and development, we open-source the codebase, along with datasets and pre-trained models, at [https://devzhk.github.io/InverseBench/](https://devzhk.github.io/InverseBench/). | inverse problem, benchmark, diffusion model | null | 9,220 | 2503.11043 | [
-0.10902376472949982,
-0.07827092707157135,
0.1033761203289032,
0.024081457406282425,
0.018715158104896545,
-0.10599584877490997,
-0.024360790848731995,
0.003927310463041067,
0.007319146301597357,
0.04391612485051155,
-0.02934001386165619,
-0.0027584738563746214,
-0.05131726711988449,
0.04138774052262306,
-0.09049925953149796,
0.004530619829893112,
0.0401039682328701,
0.027384400367736816,
-0.04044942557811737,
-0.04405942186713219,
-0.15127380192279816,
0.0080144377425313,
-0.05605531111359596,
-0.03522622957825661,
0.12418165802955627,
-0.026032736524939537,
0.037811171263456345,
-0.03769892454147339,
-0.025250276550650597,
-0.011378483846783638,
0.0960180014371872,
0.003157888539135456,
0.05885060504078865,
0.000016508784028701484,
-0.08817661553621292,
0.09724684804677963,
-0.010327769443392754,
-0.056898415088653564,
-0.03987186774611473,
0.013036356307566166,
-0.0007025997620075941,
0.07683568447828293,
0.06114736199378967,
0.0019016334554180503,
0.12431913614273071,
-0.02152780070900917,
-0.049352340400218964,
-0.05572133883833885,
0.056784991174936295,
0.010059294290840626,
0.020838230848312378,
0.0009575722506269813,
-0.11592192947864532,
0.049951981753110886,
0.026042398065328598,
0.07139580696821213,
0.10044188052415848,
-0.03383396938443184,
-0.003949081990867853,
0.013991067185997963,
0.030060216784477234,
-0.05161167308688164,
-0.11515115201473236,
-0.012749141082167625,
0.03216499090194702,
0.03134364262223244,
0.02443525940179825,
-0.014471413567662239,
-0.03583420068025589,
0.0976160317659378,
-0.025895895436406136,
0.0920671597123146,
-0.048747651278972626,
-0.027583735063672066,
0.030170554295182228,
-0.03088989295065403,
0.01794174686074257,
0.038451146334409714,
0.013450132682919502,
-0.0597415566444397,
0.10668619722127914,
-0.02911384217441082,
0.025217782706022263,
-0.021031463518738747,
0.057562075555324554,
0.01485398318618536,
-0.04192797839641571,
0.057328153401613235,
0.025831298902630806,
0.050287049263715744,
-0.025485264137387276,
-0.04701385274529457,
-0.027372771874070168,
-0.04192855954170227,
0.020206332206726074,
-0.012794550508260727,
0.07618378847837448,
-0.026394251734018326,
0.07346877455711365,
0.08466550707817078,
-0.03634197637438774,
-0.0037065136712044477,
-0.041604574769735336,
0.004512493498623371,
-0.009501651860773563,
-0.02771313115954399,
0.06560663133859634,
0.10582967847585678,
-0.009740312583744526,
0.00256111822091043,
0.0643208771944046,
-0.037568215280771255,
0.06026589870452881,
-0.0012734804768115282,
0.035374075174331665,
-0.08275692164897919,
0.08333522826433182,
0.024265261366963387,
-0.020084690302610397,
0.03548318147659302,
-0.045250315219163895,
-0.0165425892919302,
-0.006163740996271372,
-0.01520142238587141,
0.03662976995110512,
0.006856732536107302,
-0.0033900649286806583,
2.1850300214725844e-33,
0.009738803841173649,
-0.010815291665494442,
0.07384635508060455,
0.040255989879369736,
0.0073735942132771015,
0.020278535783290863,
0.005171129945665598,
-0.07201042771339417,
-0.07503584772348404,
-0.03745191544294357,
-0.05399109423160553,
0.050647567957639694,
-0.09917476773262024,
0.02374451979994774,
-0.04408159479498863,
-0.004071470815688372,
0.02503647468984127,
0.07356152683496475,
-0.059934332966804504,
-0.05412665009498596,
0.059594713151454926,
-0.041962794959545135,
-0.07832428812980652,
-0.08734069019556046,
-0.0025832823012024164,
-0.013751188293099403,
0.022252678871154785,
-0.0799255445599556,
0.004327361471951008,
0.0016211319016292691,
-0.13381560146808624,
0.04391814395785332,
-0.02010248601436615,
0.10603543370962143,
-0.0066728065721690655,
0.08039731532335281,
0.0518263503909111,
-0.0462239570915699,
-0.02095220796763897,
-0.011055520735681057,
0.00013990150182507932,
0.10504152625799179,
-0.041186243295669556,
-0.059548765420913696,
-0.041622355580329895,
-0.07912866771221161,
-0.00046823080629110336,
0.021849514916539192,
0.022391511127352715,
-0.011424126103520393,
0.011391796171665192,
-0.03396402299404144,
-0.0396067313849926,
-0.05310184881091118,
-0.0251664686948061,
-0.037783991545438766,
0.008775865659117699,
-0.02441505528986454,
0.03547423332929611,
0.059181924909353256,
0.1510361284017563,
0.037154145538806915,
-0.03813345357775688,
0.02694624848663807,
-0.00021913880482316017,
0.06566020846366882,
-0.00030339721706695855,
-0.018302926793694496,
-0.015464678406715393,
0.05187755078077316,
-0.053550172597169876,
0.012872648425400257,
0.008338378742337227,
-0.06685607135295868,
0.12365716695785522,
0.014967885799705982,
0.05905159190297127,
-0.046332359313964844,
-0.013392171822488308,
-0.05036771297454834,
-0.11201160401105881,
-0.015802718698978424,
-0.013421540148556232,
-0.05786672979593277,
-0.07973798364400864,
-0.0020265039056539536,
0.04057677462697029,
-0.03670116141438484,
-0.1074170470237732,
-0.09959173202514648,
-0.004236428998410702,
-0.037738483399152756,
-0.023726753890514374,
0.00991109199821949,
0.03917447850108147,
-1.0506987277512086e-33,
-0.02438378520309925,
0.018290042877197266,
-0.037357743829488754,
0.046018149703741074,
-0.030715642496943474,
-0.025218944996595383,
0.024200567975640297,
0.05980636179447174,
0.05338708311319351,
-0.08757099509239197,
0.007734953425824642,
0.021648423746228218,
0.016414429992437363,
-0.013990778475999832,
0.02316778153181076,
-0.05093483626842499,
-0.026522109284996986,
-0.04176124185323715,
-0.009028129279613495,
0.016465501859784126,
-0.042812786996364594,
0.09051623940467834,
0.009696400724351406,
0.00453941710293293,
-0.09385140985250473,
0.05958849564194679,
-0.02942366898059845,
0.05126652121543884,
-0.001931779203005135,
-0.04265187308192253,
-0.08939900249242783,
0.01981489174067974,
-0.03787253051996231,
-0.020679021254181862,
-0.10823799669742584,
0.18756310641765594,
-0.00026884800172410905,
0.03942624852061272,
-0.08983208984136581,
0.01611965522170067,
0.010544100776314735,
0.0218778345733881,
-0.07813312113285065,
0.007451927289366722,
0.031628284603357315,
-0.06351339817047119,
-0.017112575471401215,
0.07082699239253998,
0.058545030653476715,
-0.0005002287216484547,
-0.03741568699479103,
-0.01191561110317707,
-0.014290612190961838,
0.060336481779813766,
-0.06364212185144424,
0.005034672562032938,
-0.014983597211539745,
-0.019269632175564766,
0.06363733112812042,
0.03234436735510826,
-0.02778119221329689,
0.021748561412096024,
0.04063789173960686,
0.017603978514671326,
-0.053745757788419724,
0.06983267515897751,
-0.03709960728883743,
0.060372285544872284,
-0.023326225578784943,
0.009221535176038742,
-0.04152951389551163,
0.04189518094062805,
0.062277860939502716,
0.01316927932202816,
0.04719449207186699,
-0.06361322849988937,
0.013834592886269093,
0.06330888718366623,
0.001481364481151104,
0.03362780809402466,
0.05016779154539108,
0.03701598197221756,
0.032578110694885254,
0.05226533114910126,
0.012210795655846596,
-0.009155589155852795,
-0.0064285267144441605,
-0.047923460602760315,
-0.02227371744811535,
0.023260485380887985,
-0.022248461842536926,
0.014738576486706734,
0.03300100564956665,
-0.017418161034584045,
0.0309612937271595,
-3.675605597663889e-8,
-0.07259498536586761,
0.041830360889434814,
0.10274729877710342,
-0.0034587348345667124,
0.016705429181456566,
-0.004251748789101839,
0.06188882514834404,
0.08713005483150482,
-0.030664725229144096,
-0.06896211206912994,
0.020548781380057335,
0.004183453973382711,
0.015865866094827652,
-0.04547629877924919,
-0.048781901597976685,
0.14919304847717285,
0.015529154799878597,
-0.008338704705238342,
-0.0680587887763977,
0.005655401851981878,
-0.015460778027772903,
0.0639973133802414,
0.058763328939676285,
-0.06483964622020721,
0.045256201177835464,
-0.016565682366490364,
-0.07135531306266785,
0.012825251556932926,
0.030876420438289642,
-0.0008660270832479,
-0.03807196393609047,
0.006411450915038586,
0.06229523569345474,
0.04973233491182327,
0.007542699575424194,
0.07453218102455139,
0.03153552487492561,
-0.005278551485389471,
-0.1201825737953186,
-0.030807554721832275,
-0.07199589163064957,
0.04663537070155144,
-0.025578880682587624,
-0.014076168648898602,
0.023547695949673653,
0.07191427052021027,
0.03789586201310158,
0.00047546293353661895,
0.014336472377181053,
0.03564581274986267,
0.049140218645334244,
-0.03549487516283989,
-0.026521677151322365,
0.020807970315217972,
0.04348478838801384,
0.05525616556406021,
-0.004575065337121487,
0.0038726795464754105,
0.01947425864636898,
0.0647493451833725,
0.006885826122015715,
-0.032447248697280884,
0.026323895901441574,
0.05443654954433441
] |
What Does It Mean to Be a Transformer? Insights from a Theoretical Hessian Analysis | https://openreview.net/forum?id=3ddi7Uss2A | [
"Weronika Ormaniec",
"Felix Dangel",
"Sidak Pal Singh"
] | Spotlight | The Transformer architecture has inarguably revolutionized deep learning, overtaking classical architectures like multi-layer perceptions (MLPs) and convolutional neural networks (CNNs). At its core, the attention block differs in form and functionality from most other architectural components in deep learning—to the extent that, in comparison to MLPs/CNNs, Transformers are more often accompanied by adaptive optimizers, layer normalization, learning rate warmup, etc. The root causes behind these outward manifestations and the precise mechanisms that govern them remain poorly understood. In this work, we bridge this gap by providing a fundamental understanding of what distinguishes the Transformer from the other architectures—grounded in a theoretical comparison of the (loss) Hessian. Concretely, for a single self-attention layer, (a) we first entirely derive the Transformer’s Hessian and express it in matrix derivatives; (b) we then characterize it in terms of data, weight, and attention moment dependencies; and (c) while doing so further highlight the important structural differences to the Hessian of classical networks. Our results suggest that various common architectural and optimization choices in Transformers can be traced back to their highly non-linear dependencies on the data and weight matrices, which vary heterogeneously across parameters. Ultimately, our findings provide a deeper understanding of the Transformer’s unique optimization landscape and the challenges it poses. | Hessian, Transformers | null | 9,197 | 2410.10986 | [
-0.11823861300945282,
-0.05357462540268898,
0.014165354892611504,
-0.028153659775853157,
0.04823067784309387,
-0.006832333747297525,
-0.021566856652498245,
0.03462738171219826,
0.07671293616294861,
-0.03143417090177536,
-0.003947980236262083,
0.04768439754843712,
-0.003843792248517275,
0.025226419791579247,
-0.04519205912947655,
0.00205300678499043,
0.06655753403902054,
0.08722398430109024,
-0.12848664820194244,
-0.029640546068549156,
0.029192540794610977,
-0.011519373394548893,
-0.023217877373099327,
-0.0023053206969052553,
0.04193059355020523,
0.0046606967225670815,
-0.016971420496702194,
-0.031815674155950546,
-0.04296112433075905,
0.010148654691874981,
0.02968299575150013,
-0.03528989851474762,
-0.03672296553850174,
0.07093886286020279,
-0.11097550392150879,
0.09997571259737015,
0.0021976358257234097,
-0.0040560937486588955,
-0.010057060979306698,
-0.015506611205637455,
0.012368499301373959,
-0.020367510616779327,
-0.0027676827739924192,
0.007523429114371538,
0.08542092144489288,
0.03820083662867546,
0.0720839574933052,
-0.06646910309791565,
-0.004786312580108643,
-0.06392859667539597,
-0.0010225072037428617,
-0.05432285740971565,
-0.062006767839193344,
0.09117960184812546,
-0.015167642384767532,
0.07965623587369919,
0.03839452192187309,
0.015559695661067963,
-0.06389961391687393,
0.008330405689775944,
-0.06530918180942535,
-0.05792537331581116,
0.02223496325314045,
-0.008634808473289013,
-0.00961912889033556,
-0.007753708399832249,
0.04890688881278038,
-0.04690414294600487,
-0.030443526804447174,
-0.02602248080074787,
-0.042519450187683105,
0.0730462595820427,
0.015820616856217384,
-0.041719499975442886,
0.12249397486448288,
-0.06559224426746368,
0.09541284292936325,
0.04802750423550606,
0.08899654448032379,
-0.07841888815164566,
0.08708846569061279,
0.05025709047913551,
0.015310882590711117,
0.018232841044664383,
0.08676354587078094,
-0.044134579598903656,
-0.040109749883413315,
0.04327882081270218,
-0.02175496332347393,
-0.023682458326220512,
-0.07245407998561859,
-0.07626424729824066,
-0.06897687166929245,
-0.0462443046271801,
0.046507298946380615,
-0.004071182571351528,
-0.043823596090078354,
-0.03953104466199875,
-0.07647553831338882,
0.051483046263456345,
0.019461138173937798,
-0.08676540106534958,
0.06683757156133652,
0.03267238661646843,
0.0007856659940443933,
-0.035022713243961334,
0.021650444716215134,
0.1158592477440834,
-0.010479125194251537,
-0.08966297656297684,
-0.005867095664143562,
0.01213058177381754,
0.017434606328606606,
-0.06835289299488068,
0.05148885399103165,
-0.0357353575527668,
0.016702421009540558,
0.027585774660110474,
0.024074174463748932,
0.04199765622615814,
-0.05245905742049217,
0.015429788269102573,
-0.00610977690666914,
0.041692353785037994,
0.03946990892291069,
0.0190634373575449,
-0.09134173393249512,
3.106551654631016e-33,
-0.035649098455905914,
0.028233468532562256,
0.02906143292784691,
0.02686740830540657,
-0.005547563079744577,
-0.03620566427707672,
0.05818832665681839,
-0.008416779339313507,
0.028890427201986313,
-0.00912413652986288,
-0.02421451173722744,
0.0801156535744667,
-0.037067677825689316,
0.08620698750019073,
0.010910707525908947,
0.011097278445959091,
-0.06812962144613266,
0.006971036549657583,
0.06260436028242111,
-0.09967806190252304,
0.031907301396131516,
0.02554052509367466,
0.03333635628223419,
-0.019482918083667755,
-0.07654687017202377,
-0.06903867423534393,
-0.004163268022239208,
-0.022344913333654404,
-0.08570744842290878,
-0.02158316597342491,
-0.09572984278202057,
0.05113495513796806,
-0.05774891749024391,
-0.04569308087229729,
0.01504574716091156,
-0.03241400048136711,
0.00742877135053277,
-0.024201905354857445,
0.031119097024202347,
-0.08126483857631683,
-0.039246927946805954,
0.08538908511400223,
-0.025903334841132164,
0.005058954004198313,
-0.03889866918325424,
-0.006748702377080917,
0.049804337322711945,
-0.004774394910782576,
-0.0029341033659875393,
-0.08970486372709274,
0.028954265639185905,
-0.02476619742810726,
-0.028854038566350937,
-0.015391766093671322,
0.05319659411907196,
0.0810665562748909,
0.08097973465919495,
0.10881009697914124,
0.030134866014122963,
0.00414247764274478,
0.0016434465069323778,
0.04743043705821037,
-0.018692951649427414,
0.04107220098376274,
0.05192749574780464,
0.011987674050033092,
0.046548910439014435,
0.09468957781791687,
0.015611555427312851,
0.0001948426361195743,
-0.009103396907448769,
0.035457395017147064,
-0.02949153073132038,
-0.038419097661972046,
0.055697113275527954,
0.0017077206866815686,
0.017595170065760612,
-0.015002045780420303,
-0.014708828181028366,
0.028424713760614395,
-0.04019472748041153,
0.01952660083770752,
-0.00531381880864501,
-0.002831295831128955,
-0.08116910606622696,
0.06058517470955849,
0.04109439626336098,
-0.08592413365840912,
0.009038261137902737,
0.01701998896896839,
-0.047272708266973495,
0.017576981335878372,
-0.020277969539165497,
-0.007279920391738415,
-0.039352938532829285,
-3.1055583619045715e-33,
-0.05024782568216324,
0.010795706883072853,
-0.09003684669733047,
0.06072124093770981,
-0.022107571363449097,
-0.024755673483014107,
-0.05639560893177986,
0.041258785873651505,
-0.04876890406012535,
-0.03133818507194519,
0.035912737250328064,
-0.04202844947576523,
-0.011581279337406158,
-0.038293831050395966,
0.03458273783326149,
-0.02226690575480461,
-0.0648588165640831,
-0.022218963131308556,
-0.06602281332015991,
-0.056662898510694504,
0.035230234265327454,
0.13499628007411957,
-0.08960918337106705,
-0.0007017332827672362,
-0.1099649965763092,
0.004508424550294876,
-0.1493908315896988,
0.12496881186962128,
0.043529778718948364,
-0.015992343425750732,
-0.03837409242987633,
0.016148341819643974,
-0.010081551037728786,
0.02742079831659794,
0.011086313053965569,
0.09363989531993866,
-0.04771332070231438,
-0.03495779260993004,
-0.004822582472115755,
0.05343448743224144,
0.02443414367735386,
-0.021837063133716583,
0.006817552261054516,
0.01784401759505272,
-0.03829614445567131,
-0.03544046729803085,
-0.010262243449687958,
-0.044214919209480286,
-0.06006241589784622,
0.0494534894824028,
-0.02710115909576416,
0.030145863071084023,
-0.09699353575706482,
-0.011463222093880177,
-0.08746707439422607,
0.017440877854824066,
0.02005857042968273,
-0.0340648889541626,
0.14517904818058014,
0.007202332839369774,
-0.03587985783815384,
-0.06010214611887932,
-0.02145540341734886,
-0.06700436770915985,
-0.04480419307947159,
0.013542236760258675,
-0.013697371818125248,
-0.03369446098804474,
0.13212871551513672,
0.08106779307126999,
0.03670180216431618,
-0.03308669105172157,
0.020955704152584076,
-0.007147317286580801,
-0.0032160114496946335,
-0.05645805224776268,
0.0795043483376503,
-0.017615100368857384,
-0.06566417217254639,
-0.030401775613427162,
0.005640987306833267,
-0.020600028336048126,
0.02915254235267639,
0.016829609870910645,
0.027676204219460487,
0.07298903167247772,
0.06744051724672318,
-0.024851640686392784,
0.04358505457639694,
-0.0710175484418869,
0.03297867625951767,
0.05549270287156105,
-0.0180745180696249,
0.043542373925447464,
-0.02569485828280449,
-4.939030162631752e-8,
-0.08580508083105087,
0.0337817519903183,
-0.01628274843096733,
-0.03491954877972603,
0.0319303497672081,
-0.08994430303573608,
0.06386880576610565,
0.11365024000406265,
-0.07450820505619049,
0.07089574635028839,
0.06842547655105591,
-0.03130412846803665,
-0.03993832319974899,
-0.04558058828115463,
-0.020375419408082962,
0.07930516451597214,
0.02741658315062523,
0.01695055514574051,
0.011391418054699898,
-0.036324046552181244,
-0.01729857362806797,
0.06957186758518219,
0.008242138661444187,
-0.06932570040225983,
0.055970437824726105,
-0.056276511400938034,
-0.033052247017621994,
-0.019075579941272736,
0.002724109450355172,
0.06380148231983185,
-0.02191231958568096,
0.05967704579234123,
-0.015270935371518135,
-0.04168232902884483,
0.020068543031811714,
0.12119394540786743,
0.10698714107275009,
-0.04346797242760658,
0.0045152646489441395,
0.014909405261278152,
0.041339050978422165,
0.054033201187849045,
-0.0552675798535347,
0.0673912763595581,
0.018810361623764038,
0.03466826677322388,
-0.0038554577622562647,
-0.03562898188829422,
-0.011257747188210487,
0.01266764011234045,
0.12030451744794846,
0.020513882860541344,
-0.0029840944334864616,
0.05788823217153549,
0.03820659592747688,
-0.034559816122055054,
0.04505069926381111,
-0.058516014367341995,
0.007396025583148003,
0.10342297703027725,
0.03508424013853073,
0.07523185014724731,
-0.05276595428586006,
-0.004701272584497929
] |
Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks | https://openreview.net/forum?id=4NTrco82W0 | [
"Rui Hu",
"Yifan Zhang",
"Zhuoran Li",
"Longbo Huang"
] | Spotlight | Generative Flow Networks (GFlowNets) are a novel class of generative models designed to sample from unnormalized distributions and have found applications in various important tasks, attracting great research interest in their training algorithms. In general, GFlowNets are trained by fitting the forward flow to the backward flow on sampled training objects. Prior work focused on the choice of training objects, parameterizations, sampling and resampling strategies, and backward policies, aiming to enhance credit assignment, exploration, or exploitation of the training process. However, the choice of regression loss, which can highly influence the exploration and exploitation behavior of the under-training policy, has been overlooked. Due to the lack of theoretical understanding for choosing an appropriate regression loss, most existing algorithms train the flow network by minimizing the squared error of the forward and backward flows in log-space, i.e., using the quadratic regression loss. In this work, we rigorously prove that distinct regression losses correspond to specific divergence measures, enabling us to design and analyze regression losses according to the desired properties of the corresponding divergence measures. Specifically, we examine two key properties: zero-forcing and zero-avoiding, where the former promotes exploitation and higher rewards, and the latter encourages exploration and enhances diversity. Based on our theoretical framework, we propose three novel regression losses, namely, Shifted-Cosh, Linex(1/2), and Linex(1). We evaluate them across three benchmarks: hyper-grid, bit-sequence generation, and molecule generation. Our proposed losses are compatible with most existing training algorithms, and significantly improve the performances of the algorithms concerning convergence speed, sample diversity, and robustness. | GFlowNet, Generative Models, f-Divergence, Loss Function | null | 9,174 | 2410.02596 | [
-0.10196391493082047,
-0.04557789862155914,
0.0999336987733841,
0.0575561448931694,
0.07848659157752991,
0.025934018194675446,
-0.05548078566789627,
-0.030232617631554604,
0.06593557447195053,
-0.0305438544601202,
-0.07734059542417526,
-0.014829769730567932,
0.03831600770354271,
-0.008546072989702225,
-0.05267893895506859,
-0.02738800272345543,
0.038124337792396545,
0.10064371675252914,
-0.04146653413772583,
-0.020365271717309952,
0.058378446847200394,
0.05938978120684624,
0.017960086464881897,
0.017260270193219185,
0.06293325126171112,
-0.013822514563798904,
-0.0061165206134319305,
-0.04779494181275368,
0.013029715046286583,
-0.03412974625825882,
0.05933130159974098,
0.03668174892663956,
-0.07389768958091736,
0.008511025458574295,
-0.04470917582511902,
0.038170866668224335,
-0.07006105780601501,
0.042533375322818756,
0.02969944104552269,
0.07140745967626572,
0.040318042039871216,
0.030688557773828506,
0.026413287967443466,
0.019521957263350487,
0.0519724115729332,
-0.01613081432878971,
-0.016411779448390007,
-0.08862181752920151,
0.0012429662747308612,
0.02991788275539875,
0.014752884395420551,
-0.04571432247757912,
-0.013334259390830994,
0.03743661195039749,
0.04330003634095192,
-0.0008487949380651116,
0.010535887442529202,
0.010037745349109173,
-0.014904730953276157,
0.07233574986457825,
0.009113792330026627,
-0.1854802370071411,
-0.060868702828884125,
-0.05285418778657913,
-0.07960408180952072,
-0.03440890833735466,
-0.029135407879948616,
0.113620325922966,
-0.0068972473964095116,
0.05970010906457901,
-0.006048216484487057,
0.06930478662252426,
-0.14411567151546478,
-0.01511644758284092,
0.0463147796690464,
0.05241692066192627,
0.062212806195020676,
0.034359160810709,
-0.031607359647750854,
-0.03948937729001045,
0.044907551258802414,
0.05720606818795204,
0.013139352202415466,
-0.0459367036819458,
0.017206113785505295,
-0.0067273033782839775,
-0.02600366808474064,
0.056221943348646164,
0.11830783635377884,
0.017549963667988777,
-0.09745340049266815,
0.07837114483118057,
-0.02071402035653591,
0.004829940851777792,
0.0715814009308815,
0.0090958746150136,
0.003455215832218528,
-0.05911411717534065,
-0.004405271261930466,
0.07890108972787857,
0.034423649311065674,
0.08970405906438828,
0.027639631181955338,
-0.032059989869594574,
0.09977923333644867,
-0.05521440505981445,
0.05803072080016136,
-0.010705902241170406,
0.07227169722318649,
-0.021038439124822617,
0.035570185631513596,
0.020192448049783707,
0.055569518357515335,
-0.048752281814813614,
-0.056585993617773056,
-0.019745511934161186,
-0.06592285633087158,
-0.03246726840734482,
-0.07515303045511246,
0.12165624648332596,
-0.11458050459623337,
0.024239230901002884,
-0.005299507640302181,
-0.008503182791173458,
0.02542884647846222,
-0.03350439667701721,
-0.06953923404216766,
9.117410609821097e-34,
0.027545390650629997,
-0.0024037903640419245,
0.03120323270559311,
-0.00412330124527216,
0.06907209008932114,
0.04621447995305061,
0.04071657732129097,
-0.008663399145007133,
0.044030968099832535,
-0.012078284285962582,
-0.025413289666175842,
-0.01588878221809864,
-0.13805416226387024,
0.08112915605306625,
0.040182702243328094,
-0.042024627327919006,
-0.04266451671719551,
0.06116604432463646,
0.05005158111453056,
-0.005980398505926132,
0.012819409370422363,
-0.06613413244485855,
-0.05082366615533829,
-0.036597322672605515,
0.025370674207806587,
-0.02354443445801735,
0.023269401863217354,
-0.010567784309387207,
-0.0035070169251412153,
0.008237923495471478,
-0.06109597533941269,
-0.03184102475643158,
0.01116859633475542,
-0.05817674100399017,
0.01592434197664261,
0.05355479568243027,
-0.0257974062114954,
0.004143646452575922,
-0.008001578040421009,
-0.0042588678188622,
-0.06323179602622986,
0.03549111261963844,
-0.009797578677535057,
-0.03169558197259903,
-0.11547232419252396,
-0.1592399775981903,
0.07492022216320038,
-0.05777473747730255,
0.020657673478126526,
-0.015600734390318394,
0.01558708306401968,
-0.0004977067583240569,
-0.009654724970459938,
-0.0413159541785717,
-0.05233247950673103,
0.006151488516479731,
0.0018518794095143676,
0.06276319175958633,
-0.01821885071694851,
0.013942351564764977,
-0.029911868274211884,
0.028223995119333267,
-0.04721101000905037,
0.01254195161163807,
-0.004325141664594412,
0.053770650178194046,
-0.008687874302268028,
0.019980870187282562,
0.03193975239992142,
-0.007839052937924862,
-0.024131979793310165,
-0.05440666526556015,
-0.06960566341876984,
-0.05222441256046295,
0.10071784257888794,
0.016617583110928535,
0.032113343477249146,
-0.031108533963561058,
0.03620314970612526,
-0.020363032817840576,
-0.07349344342947006,
0.06602626293897629,
-0.049633149057626724,
-0.011181030422449112,
-0.012063302099704742,
0.02308572456240654,
0.08517194539308548,
-0.10177414119243622,
-0.011423927731812,
0.002577083185315132,
-0.11725723743438721,
-0.050145380198955536,
-0.06360764056444168,
0.028364164754748344,
0.06421530246734619,
-1.6998151571409223e-33,
-0.06878428906202316,
0.09166895598173141,
-0.04925767332315445,
0.023028520867228508,
0.03007160685956478,
0.006112917326390743,
0.02641179785132408,
0.013237935490906239,
0.023315785452723503,
-0.09679310023784637,
-0.028285730630159378,
0.004877264611423016,
-0.00489064259454608,
0.016810495406389236,
0.036928847432136536,
-0.011044090613722801,
-0.046424299478530884,
0.012394298799335957,
-0.025269854813814163,
-0.04300333559513092,
0.04010305926203728,
0.13952802121639252,
-0.06960836052894592,
-0.004360274411737919,
-0.013496972620487213,
0.009802966378629208,
-0.060152798891067505,
0.14063167572021484,
0.005043894983828068,
-0.03640580549836159,
-0.02617635950446129,
-0.02817385457456112,
-0.06977864354848862,
-0.007887592539191246,
-0.029853783547878265,
0.0837351456284523,
0.057821933180093765,
0.0640515685081482,
-0.04669691249728203,
0.11012782156467438,
0.07531554996967316,
-0.030707361176609993,
-0.032158076763153076,
0.06868157535791397,
-0.018071385100483894,
-0.0251096710562706,
0.01143088098615408,
-0.06984929740428925,
0.07186223566532135,
0.05147641524672508,
-0.08366277068853378,
0.036427080631256104,
-0.06605074554681778,
0.09884408861398697,
-0.04567122086882591,
-0.09448777139186859,
0.07112026959657669,
-0.008820664137601852,
0.04878564924001694,
0.02729216404259205,
-0.047993700951337814,
-0.011159082874655724,
-0.06336454302072525,
-0.0032905321568250656,
-0.06296383589506149,
-0.010077748447656631,
-0.020357178524136543,
0.028486842289566994,
0.04962383955717087,
0.0356077179312706,
-0.0036920104175806046,
0.060574062168598175,
0.023906271904706955,
0.04602314159274101,
0.02812471240758896,
0.0041347392834723,
-0.005260726902633905,
0.01305494736880064,
-0.03909918665885925,
-0.006941895000636578,
0.03535546362400055,
0.09523028135299683,
0.004840290639549494,
0.0382666140794754,
0.13297702372074127,
0.004109079018235207,
0.03955290466547012,
0.00486131152138114,
0.0017927115550264716,
-0.019496778026223183,
-0.06666959077119827,
-0.06583822518587112,
0.027494776993989944,
0.016420168802142143,
-0.06390678882598877,
-4.3744513789079065e-8,
-0.10226057469844818,
0.07211216539144516,
0.06047713756561279,
0.039238959550857544,
0.06406677514314651,
-0.015954017639160156,
-0.07542279362678528,
0.09227876365184784,
-0.032481081783771515,
-0.029784008860588074,
0.07490324229001999,
-0.004976851399987936,
-0.04554443061351776,
-0.030418381094932556,
0.025222763419151306,
0.02369449846446514,
0.011508075520396233,
0.01723623462021351,
0.01210994552820921,
0.004542804788798094,
0.004693901631981134,
0.024318771436810493,
-0.04858881235122681,
-0.014889110811054707,
0.0374603234231472,
-0.07454735785722733,
0.05799901857972145,
0.035431165248155594,
0.04498022794723511,
-0.001494247349910438,
0.010437865741550922,
0.05117564648389816,
0.023045821115374565,
-0.015472352504730225,
-0.04981078952550888,
0.059423670172691345,
0.01102319173514843,
-0.04591351002454758,
-0.0035157364327460527,
-0.01254526898264885,
-0.0741163045167923,
0.03700559586286545,
-0.013407111167907715,
-0.045608606189489365,
-0.03371967375278473,
0.035950519144535065,
0.024701641872525215,
-0.0270838662981987,
0.026692206040024757,
-0.060742296278476715,
0.03228587284684181,
-0.007246828638017178,
0.011118707247078419,
0.06806004047393799,
0.06454341858625412,
-0.0750722587108612,
-0.05076492205262184,
-0.038523461669683456,
-0.0241185761988163,
0.02888312004506588,
-0.04478541761636734,
0.04247210547327995,
-0.054136015474796295,
-0.015580059960484505
] |
Adam Exploits ℓ∞-geometry of Loss Landscape via Coordinate-wise Adaptivity | https://openreview.net/forum?id=PUnD86UEK5 | [
"Shuo Xie",
"Mohamad Amin Mohamadi",
"Zhiyuan Li"
] | Spotlight | Adam outperforms SGD when training language models. Yet this advantage is not well-understood theoretically -- previous convergence analysis for Adam and SGD mainly focuses on the number of steps $T$ and is already minimax-optimal in non-convex cases, which are both $\widetilde{O}(T^{-1/4})$. In this work, we argue that the exploitation of nice $\ell_\infty$-geometry is the key advantage of Adam over SGD. More specifically, we give a new convergence analysis for Adam under novel assumptions that loss is smooth under $\ell_\infty$-geometry rather than the more common $\ell_2$-geometry, which yields a much better empirical smoothness constant for GPT-2 and ResNet models. Our experiments confirm that Adam performs much worse when the favorable $\ell_\infty$-geometry is changed while SGD provably remains unaffected. We also extend the convergence analysis to blockwise Adam under novel blockwise smoothness assumptions. | Adam, coordinate-wise adaptivity, adaptive algorithms, infinity norm | null | 9,170 | null | [
-0.1125955879688263,
-0.09720256179571152,
0.07878101617097855,
0.006114942487329245,
0.0944899395108223,
0.020582672208547592,
0.037264399230480194,
-0.006370672956109047,
0.0366101935505867,
-0.02222343347966671,
-0.021550564095377922,
0.046811580657958984,
-0.005518107675015926,
0.013283329084515572,
-0.0329899862408638,
0.004558903630822897,
0.07565829157829285,
0.12930604815483093,
-0.06295990198850632,
-0.02433023415505886,
0.029855290427803993,
0.05446984991431236,
0.0340438149869442,
-0.016609933227300644,
0.043403517454862595,
-0.048247553408145905,
-0.042039208114147186,
-0.046520158648490906,
0.05429413169622421,
-0.022182432934641838,
0.059469159692525864,
0.023866241797804832,
-0.05688664689660072,
0.002718366915360093,
-0.05099567770957947,
0.04041048884391785,
0.016491778194904327,
0.03612197935581207,
-0.01623515971004963,
-0.005077023059129715,
-0.003140784101560712,
0.04176485165953636,
0.044395919889211655,
-0.003663019509986043,
0.0881965234875679,
-0.051123376935720444,
-0.009214833378791809,
-0.08052251487970352,
0.016289914026856422,
-0.017343709245324135,
-0.0357629768550396,
-0.0307918768376112,
-0.02250029146671295,
0.007474665995687246,
-0.017054276540875435,
-0.009607979096472263,
-0.049630019813776016,
0.04744905233383179,
0.020864218473434448,
0.008922304026782513,
0.0988486260175705,
-0.14726190268993378,
-0.029513906687498093,
0.022954726591706276,
-0.030344752594828606,
-0.028672421351075172,
0.03887172043323517,
0.011210192926228046,
-0.09574878960847855,
0.11132939904928207,
-0.0063509223982691765,
0.08298119902610779,
-0.089106485247612,
0.06624877452850342,
0.0687088668346405,
0.050356425344944,
0.08963363617658615,
-0.0006018094136379659,
-0.029910463839769363,
-0.06067394092679024,
0.041319310665130615,
0.02990580163896084,
0.047085314989089966,
-0.057422567158937454,
0.02023150585591793,
0.04146120324730873,
-0.0018255185568705201,
0.0023933385964483023,
0.03581647574901581,
-0.01518007181584835,
-0.059608541429042816,
0.03294645622372627,
-0.008574982173740864,
0.007587949279695749,
0.08522593975067139,
-0.004123631399124861,
-0.02997957542538643,
-0.06705931574106216,
-0.03732922673225403,
0.15153124928474426,
-0.0030589643865823746,
0.01156886201351881,
-0.0013699762057512999,
0.007663675118237734,
0.05425245314836502,
-0.02864629030227661,
0.0731423869729042,
-0.027347685769200325,
0.036284394562244415,
-0.033787619322538376,
0.0883110985159874,
0.0025277272798120975,
0.036504246294498444,
0.05912739038467407,
0.011617369949817657,
-0.07621313631534576,
-0.02839205227792263,
0.013094558380544186,
0.0341239757835865,
0.08402492851018906,
-0.0734063908457756,
0.09293512254953384,
0.007962745614349842,
0.06843326985836029,
0.0033421977423131466,
0.048637546598911285,
0.010766551829874516,
3.028846702595863e-33,
0.02920367382466793,
0.05804521217942238,
0.03129507601261139,
-0.06526895612478256,
0.08089958131313324,
0.02096952311694622,
0.01999346911907196,
0.0038491638842970133,
0.0294988751411438,
0.0025860550813376904,
-0.08239221572875977,
-0.05514996498823166,
-0.05406605452299118,
0.06227409094572067,
0.0651322454214096,
0.00893835723400116,
0.015552142634987831,
0.03300584480166435,
0.018818264827132225,
-0.009712792932987213,
0.035428211092948914,
-0.04218558967113495,
-0.012505877763032913,
-0.021018214523792267,
-0.05812864005565643,
-0.016963133588433266,
-0.0023383914958685637,
0.003991516772657633,
-0.044134605675935745,
0.013171958737075329,
-0.11291711032390594,
-0.01648826152086258,
-0.010164251551032066,
0.009057087823748589,
0.026431677863001823,
0.0006468196515925229,
0.012630851939320564,
-0.009389667771756649,
-0.014610715210437775,
-0.009192020632326603,
-0.05659978836774826,
0.06801312416791916,
0.006218282505869865,
-0.044374026358127594,
-0.1306845098733902,
-0.10739748924970627,
0.08122923225164413,
0.044317252933979034,
-0.03149692714214325,
-0.09761971235275269,
0.020384330302476883,
0.04950552061200142,
-0.09296300262212753,
-0.09365264326334,
0.0004514473839662969,
0.0006870812503620982,
0.02899646759033203,
0.08070667833089828,
0.049903448671102524,
0.06608156859874725,
0.024419767782092094,
-0.004548666533082724,
-0.012893975712358952,
0.001478478079661727,
-0.08620850741863251,
-0.01645134761929512,
-0.038137078285217285,
0.007567353080958128,
0.041744157671928406,
-0.04612691327929497,
-0.025029823184013367,
-0.02484614960849285,
-0.02756371907889843,
-0.014636593870818615,
0.07701445370912552,
-0.037151485681533813,
0.019235052168369293,
-0.1432815045118332,
0.03754038363695145,
0.05182661861181259,
-0.06876353174448013,
0.026800334453582764,
-0.022890368476510048,
-0.11592457443475723,
-0.14103107154369354,
-0.09132596850395203,
0.09223804622888565,
-0.07034928351640701,
0.03934146836400032,
0.018637917935848236,
-0.05354444310069084,
-0.06273853033781052,
-0.018397172912955284,
-0.015152382664382458,
0.027258137241005898,
-3.157113707421714e-33,
-0.03635869175195694,
-0.021560557186603546,
-0.05377264693379402,
0.0628521665930748,
-0.039687592536211014,
-0.011773837730288506,
0.0012515386333689094,
0.0530083030462265,
-0.062226828187704086,
-0.11524992436170578,
0.018874265253543854,
-0.0014490046305581927,
0.03354499489068985,
-0.03615782409906387,
0.07434675842523575,
-0.08061318844556808,
-0.010145035572350025,
0.0024102558381855488,
0.028650114312767982,
0.01681794971227646,
0.11367614567279816,
0.09660914540290833,
-0.057168371975421906,
0.019164886325597763,
0.01792537048459053,
-0.003630128689110279,
-0.08982352167367935,
0.13313829898834229,
-0.07655375450849533,
-0.019837601110339165,
-0.0017525848234072328,
-0.050704069435596466,
-0.09163381159305573,
0.04009656608104706,
-0.05291356146335602,
0.07303678244352341,
-0.006287176162004471,
0.04696938395500183,
-0.048666272312402725,
0.10677871108055115,
0.0844951868057251,
-0.010085156187415123,
0.03011641651391983,
0.08757438510656357,
-0.0036844864953309298,
0.0013630009489133954,
-0.04513087123632431,
-0.04819544777274132,
-0.04394657909870148,
0.0336139090359211,
-0.0190642811357975,
0.03007194772362709,
-0.027143364772200584,
0.048514943569898605,
-0.03394177928566933,
-0.01442609541118145,
-0.023753633722662926,
-0.013990212231874466,
0.04680727422237396,
0.008340083993971348,
-0.06700372695922852,
-0.08145999908447266,
-0.06804230809211731,
0.06380871683359146,
0.03206496313214302,
0.010053029283881187,
-0.04990784823894501,
-0.009945094585418701,
0.026379317045211792,
0.04263131320476532,
-0.0059064351953566074,
-0.03916962072253227,
-0.061409369111061096,
0.0469907745718956,
-0.043430525809526443,
0.00303675071336329,
0.04734424129128456,
-0.03676781803369522,
-0.05187151953577995,
-0.1128343865275383,
-0.0029568762984126806,
0.03302594646811485,
0.009623900055885315,
0.01451587863266468,
0.06992288678884506,
0.06485266238451004,
0.01576891541481018,
0.0005844967672601342,
0.06803072243928909,
0.023789016529917717,
0.009405102580785751,
-0.020207474008202553,
-0.034261059015989304,
-0.02049836330115795,
-0.003549163928255439,
-4.376926199256559e-8,
-0.0852295383810997,
0.015642842277884483,
0.016640089452266693,
0.03479212522506714,
0.04608466848731041,
-0.08151023834943771,
-0.0019499288173392415,
0.06277865171432495,
0.002361861988902092,
-0.008688779547810555,
0.008998163044452667,
0.026658393442630768,
-0.044166453182697296,
-0.04870355501770973,
-0.012495405040681362,
0.07745568454265594,
-0.049715422093868256,
0.03488682582974434,
0.02304685302078724,
0.012791418470442295,
-0.019391203299164772,
0.0789017453789711,
-0.010020419955253601,
-0.024839282035827637,
-0.0021617889869958162,
-0.06511963903903961,
0.03476389870047569,
0.05797204747796059,
0.10450565069913864,
-0.002314658835530281,
-0.005068759433925152,
0.019671183079481125,
-0.021824561059474945,
-0.01787162572145462,
-0.02277201972901821,
0.06721968203783035,
-0.04390886798501015,
0.059338413178920746,
-0.01820843480527401,
0.04629300534725189,
-0.012375416234135628,
0.08843116462230682,
-0.013973981142044067,
-0.08088681101799011,
-0.03549158573150635,
-0.012239988893270493,
-0.011415024287998676,
-0.035997211933135986,
0.006350807845592499,
-0.02262614667415619,
0.01607482321560383,
0.004411764442920685,
0.0071685947477817535,
0.09008228778839111,
0.07754281908273697,
-0.04749380797147751,
-0.013775741681456566,
-0.07318393886089325,
-0.006502173840999603,
0.10003586113452911,
-0.013653392903506756,
0.015105708502233028,
-0.05977320671081543,
-0.03381473571062088
] |
CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph | https://openreview.net/forum?id=mOpNrrV2zH | [
"Haitao Lin",
"Guojiang Zhao",
"Odin Zhang",
"Yufei Huang",
"Lirong Wu",
"Cheng Tan",
"Zicheng Liu",
"Zhifeng Gao",
"Stan Z. Li"
] | Spotlight | Structure-based drug design (SBDD) aims to generate potential drugs that can bind to a target protein and is greatly expedited by the aid of AI techniques in generative models. However, a lack of systematic understanding persists due to the diverse settings, complex implementation, difficult reproducibility, and task singularity. Firstly, the absence of standardization can lead to unfair comparisons and inconclusive insights. To address this dilemma, we propose CBGBench, a comprehensive benchmark for SBDD, that unifies the task as a generative graph completion, analogous to fill-in-the-blank of the 3D complex binding graph. By categorizing existing methods based on their attributes, CBGBench facilitates a modular and extensible framework that implements cutting-edge methods. Secondly, a single de novo molecule generation task can hardly reflect their capabilities. To broaden the scope, we adapt these models to a range of tasks essential in drug design, considered sub-tasks within the graph fill-in-the-blank tasks. These tasks include the generative designation of de novo molecules, linkers, fragments, scaffolds, and sidechains, all conditioned on the structures of protein pockets. Our evaluations are conducted with fairness, encompassing comprehensive perspectives on interaction, chemical properties, geometry authenticity, and substructure validity. We further provide insights with analysis from empirical studies. Our results indicate that there is potential for further improvements on many tasks, with optimization in network architectures, and effective incorporation of chemical prior knowledge. Finally, to lower the barrier to entry and facilitate further developments in the field, we also provide a single [codebase](https://github.com/EDAPINENUT/CBGBench) that unifies the discussed models, data pre-processing, training, sampling, and evaluation. | Molecule Generation Benchmark, Target-Aware Drug Design, Generative Model | null | 9,115 | 2406.10840 | [
-0.055153653025627136,
-0.08855557441711426,
-0.0005950747872702777,
-0.012937819585204124,
-0.002237633801996708,
-0.033563051372766495,
-0.07831308990716934,
0.03815457224845886,
0.019298365339636803,
-0.01573282852768898,
-0.06266960501670837,
-0.07759435474872589,
0.03673301264643669,
0.05230323225259781,
0.020341696217656136,
0.09969949722290039,
0.014077099971473217,
0.0038045316468924284,
-0.02264227718114853,
-0.03246862441301346,
0.043271906673908234,
0.06811545044183731,
0.013719293288886547,
0.004003620706498623,
0.04473991319537163,
0.0558604896068573,
0.006196147762238979,
-0.03886378929018974,
0.03498440235853195,
-0.023332303389906883,
0.09071612358093262,
0.07238246500492096,
-0.008984137326478958,
-0.009610245004296303,
0.0024703205563127995,
0.051892586052417755,
-0.06636348366737366,
0.053140025585889816,
0.05452508106827736,
-0.025165170431137085,
0.03223301097750664,
0.0387100987136364,
0.024303849786520004,
0.07856523990631104,
0.06571654230356216,
-0.04454057663679123,
-0.09679840505123138,
0.0058113629929721355,
0.06124395877122879,
0.038049668073654175,
-0.038661278784275055,
-0.03924975171685219,
-0.07507327944040298,
0.060396358370780945,
0.06718502193689346,
0.045760855078697205,
-0.025140421465039253,
-0.03776341304183006,
-0.016206534579396248,
-0.007162267807871103,
0.015407538041472435,
-0.07439084351062775,
0.009978112764656544,
-0.0597182959318161,
0.05207521468400955,
-0.008499004878103733,
0.04294397309422493,
0.004225366748869419,
0.03184541314840317,
0.06857042014598846,
0.0018897693371400237,
0.04941391572356224,
-0.07586370408535004,
0.10682021081447601,
0.04201532527804375,
0.08412431180477142,
0.042918018996715546,
0.04152233898639679,
-0.03482086956501007,
-0.10451921820640564,
0.010427531786262989,
0.021990535780787468,
0.06497745960950851,
0.07728613913059235,
0.06717236340045929,
0.03015006147325039,
-0.007991372607648373,
0.03543539717793465,
0.00046452388050965965,
0.05139422416687012,
-0.03045661933720112,
-0.0066134389489889145,
-0.011996660381555557,
-0.01818910241127014,
-0.04858521744608879,
0.04056580364704132,
0.0939924567937851,
0.0012427764013409615,
-0.008343875408172607,
0.03341929242014885,
-0.028057212010025978,
0.035385411232709885,
0.022247638553380966,
-0.0863453820347786,
-0.007231364492326975,
-0.006397458724677563,
0.04708891361951828,
0.02025476284325123,
-0.011498603969812393,
-0.04740623012185097,
0.009582556784152985,
0.032100312411785126,
0.07109708338975906,
0.00030764189432375133,
0.04810429364442825,
0.011485408991575241,
0.01530130673199892,
-0.0412137508392334,
0.06345940381288528,
0.025302881374955177,
-0.04416024312376976,
0.043980762362480164,
-0.011897710151970387,
-0.003548435168340802,
-0.12739500403404236,
0.029388435184955597,
-0.10482310503721237,
-2.610421452814756e-34,
0.06482569873332977,
0.028809817507863045,
0.06569259613752365,
0.13655495643615723,
0.06686331331729889,
-0.0390782356262207,
-0.06924093514680862,
0.023222321644425392,
-0.07892906665802002,
0.05389677733182907,
-0.08222758769989014,
-0.02742830477654934,
-0.040786582976579666,
0.09527937322854996,
0.02262352779507637,
-0.02266119047999382,
-0.015001548454165459,
0.04690232872962952,
-0.04961864650249481,
-0.06102456897497177,
0.0007579405792057514,
-0.008952881209552288,
-0.027456235140562057,
0.016184920445084572,
0.025983396917581558,
0.07646286487579346,
0.01753988489508629,
0.015139768831431866,
0.0043786317110061646,
-0.021784210577607155,
-0.1059187799692154,
0.016814960166811943,
0.017798570916056633,
0.020569007843732834,
-0.06320124119520187,
-0.04690198972821236,
0.03198655694723129,
-0.07379969954490662,
-0.028074560686945915,
-0.037022218108177185,
-0.0316523052752018,
0.02268129214644432,
-0.06877215206623077,
-0.014658733271062374,
-0.06359735131263733,
0.008699205704033375,
0.024607516825199127,
-0.03553055226802826,
-0.01850053481757641,
-0.009621115401387215,
0.007605841848999262,
-0.0015427599428221583,
0.09033854305744171,
0.041948795318603516,
-0.07638373970985413,
-0.012519254349172115,
-0.0413123220205307,
-0.009007547982037067,
0.0769166648387909,
0.03853324428200722,
0.016880886629223824,
0.06448418647050858,
-0.05104311555624008,
0.08433470875024796,
0.017245890572667122,
0.08861469477415085,
-0.05055281147360802,
-0.021803105250000954,
0.023996692150831223,
0.03031870536506176,
-0.023656705394387245,
0.056920647621154785,
-0.0046314578503370285,
-0.0814611092209816,
0.1046367809176445,
-0.007176024839282036,
-0.015145423822104931,
-0.10382930934429169,
0.002095063915476203,
0.04498136788606644,
-0.06139026954770088,
-0.017679104581475258,
-0.05409252271056175,
0.028602182865142822,
-0.04606414958834648,
-0.15769629180431366,
0.050539854913949966,
-0.03801216930150986,
-0.06528168171644211,
-0.04834559187293053,
-0.06785868853330612,
-0.07866396009922028,
-0.015274682082235813,
-0.00838716235011816,
0.02034502848982811,
6.9974841491073554e-34,
0.02290298603475094,
0.009285032749176025,
0.05835850164294243,
0.004718821030110121,
0.0005961307906545699,
-0.016882775351405144,
-0.03183350712060928,
-0.11197911947965622,
0.04674045369029045,
-0.012497134506702423,
0.0172700397670269,
-0.04957353696227074,
-0.004170096945017576,
-0.046207234263420105,
-0.0190652497112751,
-0.005301419645547867,
-0.07808773964643478,
-0.03832787275314331,
-0.06138356029987335,
0.05660219117999077,
0.04419584199786186,
0.06292729079723358,
-0.05958309397101402,
0.02203034609556198,
0.0035834666341543198,
0.07730377465486526,
-0.006656070705503225,
-0.004708553198724985,
-0.0016594991320744157,
-0.004667055793106556,
-0.03291661664843559,
0.01521019171923399,
-0.05802913010120392,
-0.04185948148369789,
-0.062172871083021164,
0.1115545928478241,
0.014023059979081154,
-0.016053175553679466,
0.034143175929784775,
-0.03555925190448761,
-0.04650562256574631,
0.011595278978347778,
-0.18569980561733246,
0.0517130121588707,
0.06320706009864807,
0.05773179233074188,
-0.03505796194076538,
0.092007577419281,
0.04721886292099953,
0.01907040923833847,
-0.01774907298386097,
0.03758299723267555,
-0.007371980231255293,
-0.05030746012926102,
-0.02780798077583313,
-0.11671054363250732,
0.0016775986878201365,
-0.05485840141773224,
0.0654580295085907,
-0.03822831064462662,
-0.10056447982788086,
-0.043678656220436096,
0.03375588729977608,
-0.03320150077342987,
-0.060235969722270966,
0.00012943112233188003,
0.04579974338412285,
0.03497190400958061,
0.05255228653550148,
-0.01408869493752718,
0.005636121146380901,
0.020563969388604164,
0.08683649450540543,
0.0588524155318737,
0.10974910855293274,
-0.023606643080711365,
0.013676619157195091,
-0.024426674470305443,
-0.03760882094502449,
-0.04052306339144707,
0.012465320527553558,
-0.03835132345557213,
-0.009503710083663464,
0.04580390080809593,
0.06324465572834015,
0.08945311605930328,
-0.0045262835919857025,
0.07773557305335999,
0.0334550216794014,
0.05004655569791794,
-0.048897042870521545,
0.060926247388124466,
-0.0737682431936264,
0.1561001092195511,
-0.014230486936867237,
-3.891728184157728e-8,
0.0000765910153859295,
0.017032504081726074,
0.01783357746899128,
-0.03625677525997162,
0.01857793517410755,
0.047426264733076096,
-0.0288371779024601,
0.11533336341381073,
-0.004641545470803976,
-0.0038972515612840652,
0.054454926401376724,
0.008294605650007725,
-0.10104196518659592,
-0.009580730460584164,
-0.009424787014722824,
-0.015278520993888378,
0.01823822595179081,
0.02883720211684704,
-0.0014739714097231627,
-0.0029464734252542257,
-0.02036004513502121,
0.021514683961868286,
-0.02942632883787155,
-0.01574835367500782,
0.12524208426475525,
-0.15687233209609985,
-0.06434903293848038,
-0.01858273521065712,
0.007433222606778145,
0.03577934205532074,
0.033906299620866776,
0.03478643670678139,
0.020956188440322876,
0.07054228335618973,
0.0013514148304238915,
0.05576259642839432,
0.03808958828449249,
0.0071329534985125065,
-0.08312411606311798,
0.0066911932080984116,
0.012831243686378002,
0.0698811411857605,
-0.056635450571775436,
-0.02380409464240074,
-0.028891829773783684,
-0.018313728272914886,
-0.04697166010737419,
0.05391920730471611,
0.035525497049093246,
-0.048167143017053604,
-0.0276348777115345,
-0.03495672345161438,
-0.005686348769813776,
-0.03286449611186981,
-0.009101522155106068,
0.08447028696537018,
-0.05066492781043053,
-0.008135279640555382,
0.06272570043802261,
-0.06994056701660156,
-0.008830025792121887,
-0.05996442213654518,
-0.03150888904929161,
-0.03413568064570427
] |
Fair Clustering in the Sliding Window Model | https://openreview.net/forum?id=VGQugiuCQs | [
"Vincent Cohen-Addad",
"Shaofeng H.-C. Jiang",
"Qiaoyuan Yang",
"Yubo Zhang",
"Samson Zhou"
] | Spotlight | We study streaming algorithms for proportionally fair clustering, a notion originally suggested by Chierichetti et al. (2017), in the sliding window model. We show that although there exist efficient streaming algorithms in the insertion-only model, surprisingly no algorithm can achieve finite ratio without violating the fairness constraint in sliding window. Hence, the problem of fair clustering is a rare separation between the insertion-only streaming model and the sliding window model. On the other hand, we show that if the fairness constraint is relaxed by a multiplicative $(1+\varepsilon)$ factor, there exists a $(1 + \varepsilon)$-approximate sliding window algorithm that uses $\text{poly}(k\varepsilon^{-1}\log n)$ space. This achieves essentially the best parameters (up to degree in the polynomial) provided the aforementioned lower bound. We also implement a number of empirical evaluations on real datasets to complement our theoretical results. | fair clustering, sliding window model | null | 9,108 | 2503.05173 | [
-0.05888081341981888,
-0.07455436885356903,
-0.002691390924155712,
0.014644802547991276,
0.03923855721950531,
0.027015404775738716,
0.01204886008054018,
-0.05813213437795639,
0.0682329460978508,
0.007257696706801653,
-0.03719653561711311,
0.007544196676462889,
0.07800345867872238,
0.06662809103727341,
-0.062167152762413025,
-0.07107014954090118,
0.0693061575293541,
0.019842714071273804,
-0.07058315724134445,
-0.04071958363056183,
-0.024162890389561653,
-0.07598905265331268,
-0.04131454974412918,
0.02923295833170414,
0.01821303553879261,
-0.05995464324951172,
0.010379177518188953,
-0.0030027260072529316,
0.020596489310264587,
0.019451308995485306,
-0.007951656356453896,
-0.016490617766976357,
0.038226332515478134,
0.0030338705983012915,
0.010082609951496124,
-0.03892478719353676,
-0.03168410062789917,
-0.011616898700594902,
-0.06066030263900757,
0.03192168101668358,
0.07624656707048416,
0.004215073771774769,
-0.028218431398272514,
-0.01811268739402294,
-0.01358081679791212,
0.009678572416305542,
-0.02890978939831257,
0.02523597702383995,
-0.09696008265018463,
-0.02944289520382881,
-0.06423363089561462,
0.0792379155755043,
0.004914012271910906,
0.1328226625919342,
-0.08185972273349762,
-0.018583858385682106,
-0.008019459433853626,
-0.00022306309256237,
-0.000633643358014524,
0.008240308612585068,
-0.018872642889618874,
-0.08573734760284424,
-0.03503834456205368,
0.00644852826371789,
-0.012071650475263596,
0.028741957619786263,
-0.05251292139291763,
0.09363626688718796,
0.022649267688393593,
0.06772644072771072,
0.006922988221049309,
0.08297103643417358,
-0.011565100401639938,
0.02881101332604885,
0.003844590624794364,
-0.0480855293571949,
0.03192133456468582,
-0.013600597158074379,
-0.003618039656430483,
0.04651038348674774,
0.019417334347963333,
-0.019290348514914513,
-0.039537496864795685,
-0.036334674805402756,
0.038858816027641296,
-0.020677514374256134,
0.051229529082775116,
-0.02075236663222313,
0.1065668836236,
-0.011368351988494396,
-0.08193746954202652,
0.08865788578987122,
0.007850192487239838,
-0.029421770945191383,
0.03604020178318024,
0.013553979806602001,
0.029396360740065575,
0.010558239184319973,
0.09674252569675446,
0.09429928660392761,
-0.010662098415195942,
0.031033474951982498,
0.059710219502449036,
-0.06136520579457283,
0.056987833231687546,
-0.03850612789392471,
-0.04064030945301056,
0.02739678882062435,
-0.00567347276955843,
-0.06522392481565475,
0.03427385166287422,
0.025123355910182,
0.053335610777139664,
0.08331213891506195,
0.0019081364152953029,
-0.02878914214670658,
0.06516891717910767,
-0.04165993630886078,
-0.03700094297528267,
0.02179688774049282,
0.026339605450630188,
-0.00156608154065907,
-0.02299129217863083,
0.04420223832130432,
0.04809712991118431,
-0.05065644159913063,
-0.0660536140203476,
1.2756123067077756e-33,
0.02742447704076767,
-0.042367104440927505,
0.03854374587535858,
-0.09556080400943756,
0.08068756014108658,
0.008806624449789524,
0.0061413878574967384,
0.04900427162647247,
-0.050519172102212906,
0.01607353240251541,
-0.051941726356744766,
0.051768966019153595,
0.006033807527273893,
0.0043357317335903645,
0.04470274597406387,
-0.06929336488246918,
0.018256254494190216,
0.08319486677646637,
0.019354676827788353,
-0.12137691676616669,
0.09739284217357635,
-0.08441811800003052,
-0.036073047667741776,
-0.03936146944761276,
-0.0908857211470604,
0.02257799729704857,
0.0024963899049907923,
-0.07446331530809402,
0.09122493118047714,
0.04182620719075203,
-0.13078399002552032,
0.03358367085456848,
0.03068152256309986,
0.011776176281273365,
0.030194956809282303,
0.03856465592980385,
-0.03301636129617691,
-0.0864543691277504,
0.028456509113311768,
-0.05128754675388336,
-0.027715926989912987,
0.025336071848869324,
0.051888227462768555,
-0.0865866169333458,
-0.06734558939933777,
-0.03986884281039238,
0.06341586261987686,
0.013804084621369839,
-0.03721153363585472,
0.009223553352057934,
0.060273703187704086,
-0.039465442299842834,
-0.03241093084216118,
0.005776898469775915,
-0.02641366980969906,
0.01580948941409588,
0.006006212439388037,
0.07741402089595795,
0.03602193295955658,
0.1576891988515854,
-0.08175701647996902,
-0.024342557415366173,
-0.06510888040065765,
0.01710893213748932,
-0.016688209027051926,
-0.03064679354429245,
0.02232474274933338,
0.015852347016334534,
0.03836086764931679,
-0.028166575357317924,
0.01793534681200981,
0.03221452981233597,
-0.07309863716363907,
0.04121709614992142,
-0.027559451758861542,
0.05326884239912033,
0.06917249411344528,
0.03655461221933365,
-0.03578254207968712,
-0.027864059433341026,
-0.04549689218401909,
-0.047247905284166336,
-0.0076744393445551395,
-0.05348018556833267,
-0.06763166189193726,
-0.047385312616825104,
0.029897557571530342,
0.018324697390198708,
-0.07091470062732697,
-0.09231120347976685,
0.002233098493888974,
-0.0003403561713639647,
0.052078649401664734,
0.07723601907491684,
0.021805813536047935,
-1.2346477979495574e-33,
-0.007393374107778072,
-0.05973716452717781,
0.020190849900245667,
0.09636896103620529,
0.05910889431834221,
-0.026638520881533623,
-0.010970010422170162,
0.01739618182182312,
0.08551742881536484,
-0.014037602581083775,
-0.03735462203621864,
-0.003103111404925585,
0.11006493121385574,
0.05360724404454231,
0.000410116947023198,
-0.019262809306383133,
0.09368531405925751,
0.04837590083479881,
0.047212932258844376,
0.015981456264853477,
0.09013266116380692,
-0.015819456428289413,
-0.0601046159863472,
0.039074745029211044,
0.05991218611598015,
0.03268738090991974,
-0.02371135726571083,
0.04988129809498787,
-0.0934576764702797,
0.09354601800441742,
-0.09062834829092026,
-0.05334958806633949,
-0.14558112621307373,
-0.07525043934583664,
0.030397232621908188,
-0.06568322330713272,
0.027839498594403267,
0.01706002466380596,
-0.016183119267225266,
0.02497909590601921,
0.0010360615560784936,
-0.0049325255677104,
-0.008750650100409985,
0.02429068461060524,
0.08838173002004623,
0.031231330707669258,
-0.036280445754528046,
-0.037953317165374756,
0.023166652768850327,
-0.08789928257465363,
-0.06099402531981468,
-0.05241154134273529,
0.04380267485976219,
0.10955796390771866,
-0.0018425433663651347,
0.00940834078937769,
0.016007643193006516,
0.05456637963652611,
-0.0400165319442749,
-0.0316847488284111,
0.03310314193367958,
-0.053081829100847244,
-0.07882320135831833,
0.022693568840622902,
-0.022887540981173515,
-0.07539071887731552,
0.003532544942572713,
-0.10460864752531052,
-0.06971140205860138,
0.0020059230737388134,
0.011961165815591812,
-0.027671458199620247,
0.017208106815814972,
0.06375353038311005,
-0.10613393038511276,
0.032101068645715714,
0.09355683624744415,
0.00036734691821038723,
-0.027988029643893242,
0.07706847786903381,
-0.04618801176548004,
0.06881135702133179,
0.03446635976433754,
-0.060811758041381836,
0.03266166150569916,
0.021019188687205315,
0.09809532016515732,
-0.02646910771727562,
-0.04532454535365105,
-0.010015363804996014,
-0.05005102977156639,
0.020941954106092453,
0.0351409837603569,
0.02583266980946064,
-0.07867497950792313,
-4.396160591113585e-8,
-0.07506758719682693,
-0.00581355532631278,
-0.04237338528037071,
0.010597585700452328,
0.048261258751153946,
-0.028127597644925117,
0.027829186990857124,
0.03464452549815178,
-0.03553570806980133,
-0.09750176966190338,
0.09024210274219513,
-0.051479700952768326,
0.021721476688981056,
0.07366588711738586,
-0.01541242003440857,
0.055580660700798035,
-0.014553995802998543,
-0.03665410354733467,
0.003988454584032297,
0.03916020318865776,
0.01531713455915451,
0.05161397159099579,
0.04524863511323929,
0.001583441044203937,
-0.03783321753144264,
-0.01685451716184616,
0.05022156611084938,
0.009045103564858437,
0.026973508298397064,
-0.05469402298331261,
-0.05220770835876465,
-0.005055947694927454,
0.05847335234284401,
0.02198229543864727,
-0.010310796089470387,
0.12450692057609558,
-0.09381140023469925,
0.05267731472849846,
-0.08183745294809341,
-0.02206575497984886,
0.08061937242746353,
0.022001424804329872,
0.011948028579354286,
-0.0554935559630394,
0.07758520543575287,
0.04743028059601784,
0.013256180100142956,
0.048419367522001266,
-0.01691792532801628,
-0.010425901040434837,
0.05023280531167984,
-0.0022644181735813618,
-0.027665281668305397,
-0.05199659243226051,
0.10848294943571091,
-0.03327338397502899,
-0.02117338962852955,
-0.11801613867282867,
0.027151132002472878,
-0.01505085825920105,
0.016806164756417274,
-0.008707240223884583,
-0.03806447982788086,
0.013459987007081509
] |
NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models | https://openreview.net/forum?id=lgsyLSsDRe | [
"Chankyu Lee",
"Rajarshi Roy",
"Mengyao Xu",
"Jonathan Raiman",
"Mohammad Shoeybi",
"Bryan Catanzaro",
"Wei Ping"
] | Spotlight | Decoder-only large language model (LLM)-based embedding models are beginning to outperform BERT or T5-based embedding models in general-purpose text embedding tasks, including dense vector-based retrieval. In this work, we introduce the NV-Embed model, incorporating architectural designs, training procedures, and curated datasets to significantly enhance the performance of LLM as a versatile embedding model, while maintaining its simplicity and reproducibility.For model architecture, we propose a latent attention layer to obtain pooled embeddings, which consistently improves retrieval and downstream task accuracy compared to mean pooling or using the last <EOS> token embedding from LLMs. To enhance representation learning, we remove the causal attention mask of LLMs during contrastive training. For training algorithm, we introduce a two-stage contrastive instruction-tuning method. It first applies contrastive training with instructions on retrieval datasets, utilizing in-batch negatives and curated hard negative examples. At stage-2, it blends various non-retrieval into instruction tuning, which not only enhances non-retrieval task accuracy but also improves retrieval performance. For training data, we utilize the hard-negative mining, synthetic data generation and existing public available datasets to boost the performance of embedding model. By combining these techniques, our NV-Embed- v1 model secured the No.1 position on the Massive Text Embedding Benchmark (MTEB) (as of May 24, 2024), across 56 embedding tasks. NV-Embed-v2 has reclaimed and maintained the top spot on MTEB since August 30, 2024, demonstrating the sustained effectiveness of the proposed methods over time. Also, it achieved the highest scores in the Long Doc section and the second-highest scores in the QA section of the AIR Benchmark, which covers a range of out-of-domain information retrieval topics beyond those in MTEB. We further provide the analysis of model compression techniques for generalist embedding models. We open-source the model at: https://huggingface.co/nvidia/NV-Embed-v2 . | LLM, embedding model, retriever | null | 9,106 | null | [
0.007477000821381807,
-0.09742795675992966,
0.04992618411779404,
0.04080026224255562,
0.06483044475317001,
0.040854405611753464,
-0.042727310210466385,
0.03067142330110073,
0.04478118196129799,
-0.0863129198551178,
-0.07456149160861969,
-0.03346043825149536,
0.020314252004027367,
0.03960631787776947,
-0.06677769124507904,
0.004896540194749832,
0.18675348162651062,
0.05947771295905113,
-0.03954756632447243,
-0.0414721854031086,
0.045859478414058685,
0.05886061489582062,
0.11172432452440262,
-0.05049477145075798,
0.0051639145240187645,
-0.033234477043151855,
-0.04632555693387985,
-0.049032386392354965,
0.054982636123895645,
-0.02107296884059906,
0.08143621683120728,
0.0067632002755999565,
-0.037959616631269455,
0.11836481094360352,
-0.032344259321689606,
0.055289722979068756,
-0.10152773559093475,
0.02839585207402706,
-0.05019744485616684,
-0.03447974845767021,
-0.006893258076161146,
0.010587853379547596,
-0.01904970407485962,
0.08606525510549545,
0.09179361164569855,
-0.0492677241563797,
-0.022465554997324944,
-0.07025658339262009,
0.014654356054961681,
0.01378380786627531,
-0.01722670905292034,
0.004011041484773159,
-0.024409113451838493,
0.0019396031275391579,
-0.0731610357761383,
-0.005698195658624172,
0.030338259413838387,
0.013907120563089848,
-0.048273682594299316,
-0.08631982654333115,
-0.03981677070260048,
-0.09344818443059921,
0.06883794814348221,
-0.058541204780340195,
0.01733042299747467,
-0.026767801493406296,
0.008420279249548912,
0.09037397801876068,
-0.06724581867456436,
0.014904279261827469,
-0.024867812171578407,
0.047813888639211655,
-0.036493733525276184,
0.01889084093272686,
-0.016046954318881035,
0.0844334214925766,
0.04090425744652748,
-0.011048507876694202,
0.11297349631786346,
-0.06472822278738022,
0.010974589735269547,
0.024551324546337128,
0.019539451226592064,
0.02633056417107582,
0.13533984124660492,
0.03180808946490288,
0.08404116332530975,
0.0211280919611454,
-0.002466856501996517,
0.015322642400860786,
-0.014511514455080032,
-0.06260856986045837,
0.06624557822942734,
-0.009953765198588371,
0.07615994662046432,
0.06409827619791031,
0.030913377180695534,
-0.013965080492198467,
-0.02384772337973118,
0.04303934425115585,
0.009677594527602196,
0.13313087821006775,
0.029572248458862305,
-0.03242810443043709,
-0.030175894498825073,
-0.004165797960013151,
0.04189341515302658,
0.07784382253885269,
-0.0062579927034676075,
-0.15905427932739258,
0.012820531614124775,
0.1004343256354332,
0.047552283853292465,
-0.008103436790406704,
0.06704190373420715,
-0.07315456122159958,
0.015557818114757538,
-0.050641726702451706,
0.05049896240234375,
0.0603087842464447,
-0.03738441318273544,
0.06408732384443283,
-0.041247665882110596,
0.009515687823295593,
0.014064611867070198,
-0.02674286812543869,
-0.013543600216507912,
1.0707881098777538e-33,
-0.0002428272709948942,
0.006507315207272768,
0.0018224178347736597,
0.03984716907143593,
0.0324060395359993,
-0.0003423159068915993,
0.04465468227863312,
0.09323662519454956,
-0.09556460380554199,
0.03519091382622719,
-0.013856681063771248,
-0.009370827116072178,
-0.0303586907684803,
0.06808958947658539,
-0.007334460038691759,
-0.0001113702091970481,
-0.07869914174079895,
-0.03469650074839592,
-0.008224195800721645,
-0.017462873831391335,
-0.014438391663134098,
0.024040576070547104,
0.019804829731583595,
-0.052746061235666275,
0.01957196183502674,
-0.00868284236639738,
0.09946435689926147,
-0.09906051307916641,
-0.07795726507902145,
-0.009610245004296303,
-0.1739509105682373,
0.019406462088227272,
0.020507624372839928,
0.015121580101549625,
0.03632509708404541,
-0.04829007387161255,
0.01368656288832426,
0.020273538306355476,
0.040432158857584,
-0.021933622658252716,
0.018746059387922287,
0.04871046915650368,
0.008921269327402115,
-0.004769604653120041,
-0.08792832493782043,
-0.09547331929206848,
0.02050476334989071,
0.014546791091561317,
-0.013813887722790241,
-0.021032588556408882,
0.06663256883621216,
0.03134067729115486,
-0.05088387802243233,
-0.054948821663856506,
0.08670971542596817,
-0.04911351576447487,
0.0532376803457737,
0.06956460326910019,
-0.0026374084409326315,
0.004793578293174505,
-0.017116565257310867,
0.035969000309705734,
0.024122929200530052,
0.09934572130441666,
-0.0015937575371935964,
0.016858907416462898,
-0.035500675439834595,
0.08335208892822266,
0.022081097587943077,
-0.02722763828933239,
-0.015014142729341984,
0.04170263558626175,
0.04143573343753815,
-0.015936611220240593,
0.00603917520493269,
-0.01974930800497532,
-0.0039780824445188046,
-0.11886035650968552,
0.019411159679293633,
0.02505366876721382,
0.06052665784955025,
-0.03570432960987091,
-0.05395008251070976,
-0.04044262692332268,
-0.07523954659700394,
-0.06502795964479446,
0.07649815082550049,
-0.10594478994607925,
0.06726151704788208,
-0.017711590975522995,
-0.03363405540585518,
-0.09264848381280899,
-0.035625945776700974,
-0.047008804976940155,
0.009531408548355103,
7.285799852585446e-34,
0.0017201323062181473,
-0.009397324174642563,
-0.05286957696080208,
-0.017371661961078644,
-0.017646903172135353,
-0.03547709807753563,
0.03898516669869423,
0.05160557106137276,
-0.03985214978456497,
-0.008124768733978271,
-0.02985941804945469,
-0.03689970448613167,
0.03264515846967697,
-0.018443936482071877,
0.03253898024559021,
0.058314282447099686,
-0.032711297273635864,
0.08953738212585449,
-0.011720510199666023,
0.001405189512297511,
-0.041079480201005936,
0.056518252938985825,
0.004109673202037811,
0.05806368216872215,
0.031816091388463974,
0.08536762744188309,
-0.025241004303097725,
0.043280813843011856,
-0.03736007958650589,
-0.10643038898706436,
-0.028921231627464294,
0.013541508466005325,
0.026903586462140083,
-0.005603113677352667,
-0.07410632073879242,
0.06988346576690674,
0.07028958201408386,
-0.014109251089394093,
-0.05450950190424919,
0.011789204552769661,
0.08155253529548645,
-0.018381094560027122,
-0.0022403893526643515,
0.06237446144223213,
-0.021672537550330162,
0.02127186767756939,
-0.1407143771648407,
-0.021368615329265594,
0.07243703305721283,
0.016801968216896057,
-0.050332989543676376,
-0.02667420171201229,
-0.07947936654090881,
0.010504835285246372,
-0.022414524108171463,
-0.08968029916286469,
-0.008820255286991596,
-0.05231628194451332,
0.059308260679244995,
-0.008177407085895538,
0.02924881875514984,
-0.006367134395986795,
0.024311842396855354,
-0.11866167187690735,
-0.0006556552834808826,
-0.07937591522932053,
-0.017105890437960625,
0.02729848027229309,
-0.007864361628890038,
-0.002876277081668377,
-0.003892680862918496,
-0.06619377434253693,
0.011613738723099232,
0.0016153823817148805,
0.04588335007429123,
0.017962362617254257,
0.03627672791481018,
-0.04260764643549919,
-0.03368310630321503,
-0.08373305946588516,
-0.0052997516468167305,
-0.02625303715467453,
-0.028775224462151527,
0.0268833227455616,
0.05700080096721649,
0.04469776898622513,
0.07977298647165298,
0.019907470792531967,
0.008581213653087616,
0.005726370494812727,
-0.0575004443526268,
-0.008995281532406807,
-0.014224864542484283,
0.05871067941188812,
0.02750253677368164,
-3.679802773604024e-8,
-0.03070368990302086,
-0.005803680047392845,
-0.013943424448370934,
0.028822224587202072,
-0.004899968858808279,
-0.08066217601299286,
-0.06338971853256226,
0.08423059433698654,
-0.05410701781511307,
-0.005034302361309528,
0.06272099167108536,
-0.023982107639312744,
-0.05301200598478317,
-0.06386561691761017,
0.03133606165647507,
0.06050702556967735,
0.038542356342077255,
-0.004993352573364973,
0.03256621211767197,
-0.06061847135424614,
-0.006426186300814152,
0.07641863077878952,
0.05232749134302139,
-0.0009409960475750268,
0.005791614297777414,
-0.028504420071840286,
-0.014809377491474152,
0.06500396877527237,
0.11751151829957962,
-0.061489537358284,
0.04780345410108566,
0.04759952053427696,
-0.025468191131949425,
-0.08456401526927948,
-0.035587772727012634,
0.08301464468240738,
0.033374372869729996,
-0.025332029908895493,
0.015851551666855812,
-0.009072261862456799,
0.03735659271478653,
0.05276180058717728,
-0.009498963132500648,
0.03678246960043907,
0.021367540583014488,
-0.02143145352602005,
-0.03754621744155884,
-0.07826515287160873,
0.036017730832099915,
-0.017336413264274597,
0.030893532559275627,
-0.06334209442138672,
-0.0534374862909317,
0.0854586586356163,
0.004446194041520357,
0.018979165703058243,
0.018913837149739265,
-0.010522402822971344,
0.021116720512509346,
-0.007383540738373995,
-0.010900340043008327,
0.06576584279537201,
-0.020196421071887016,
0.019784091040492058
] |
Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient | https://openreview.net/forum?id=SUc1UOWndp | [
"George Wang",
"Jesse Hoogland",
"Stan van Wingerden",
"Zach Furman",
"Daniel Murfet"
] | Spotlight | We introduce refined variants of the Local Learning Coefficient (LLC), a measure of model complexity grounded in singular learning theory, to study the development of internal structure in transformer language models during training. By applying these refined LLCs (rLLCs) to individual components of a two-layer attention-only transformer, we gain novel insights into the progressive differentiation and specialization of attention heads. Our methodology reveals how attention heads differentiate into distinct functional roles over the course of training, analyzes the types of data these heads specialize to process, and discovers a previously unidentified multigram circuit. These findings demonstrate that rLLCs provide a principled, quantitative toolkit for developmental interpretability, which aims to understand models through their evolution across the learning process. This work advances the field of developmental interpretability by providing a mathematically rigorous approach to understanding neural networks through the lens of their learning process. More broadly, this work takes a step towards establishing the correspondence between data distributional structure, geometric properties of the loss landscape, learning dynamics, and emergent computational structures in neural networks. | Developmental Interpretability, Mechanistic Interpretability, Singular Learning Theory, Learning Dynamics, Stagewise development, Model complexity | The paper studies how language model attention heads develop during training using new, refined variants of the Local Learning Coefficient (LLC) | 9,016 | 2410.02984 | [
0.018519647419452667,
-0.07459819316864014,
0.02019980177283287,
0.035900406539440155,
0.07315979897975922,
0.04693511128425598,
-0.058210112154483795,
-0.013629808090627193,
0.09155675023794174,
-0.005319081246852875,
0.027354493737220764,
-0.04242487996816635,
-0.011410482227802277,
0.05257200449705124,
-0.07367130368947983,
-0.0014259542804211378,
-0.10884818434715271,
0.05929289385676384,
-0.075608029961586,
-0.04471985623240471,
0.055381204932928085,
0.042441222816705704,
0.04149225354194641,
0.00536651536822319,
0.04264353588223457,
0.08320541679859161,
-0.05918349698185921,
-0.030456896871328354,
0.01737731136381626,
-0.04059566929936409,
0.06651220470666885,
0.009541460312902927,
0.036547742784023285,
0.08331166952848434,
-0.08691363781690598,
0.09426439553499222,
0.008531327359378338,
0.058053795248270035,
0.010329528711736202,
-0.00004006822928204201,
0.034261275082826614,
0.0036295794416218996,
-0.024851594120264053,
0.00445582065731287,
0.09294391423463821,
-0.027005326002836227,
0.03309592604637146,
-0.07053286582231522,
-0.08783221989870071,
-0.06833033263683319,
-0.016735529527068138,
-0.07648365199565887,
-0.03450204059481621,
0.08280634880065918,
0.028542086482048035,
0.0782499685883522,
-0.006840876303613186,
0.042109597474336624,
-0.007169356569647789,
0.012448972091078758,
-0.10852932929992676,
-0.11142748594284058,
-0.0006932255928404629,
-0.029572103172540665,
-0.003920810297131538,
0.013358036056160927,
0.008025528863072395,
-0.006833258550614119,
-0.06227825582027435,
-0.03221901133656502,
-0.06267187744379044,
0.09035828709602356,
-0.003550655907019973,
-0.01744461990892887,
0.09924455732107162,
-0.00042391105671413243,
0.09841273725032806,
0.023287653923034668,
0.05315001681447029,
-0.03857864439487457,
-0.034857336431741714,
0.03379281237721443,
0.046404361724853516,
-0.067467100918293,
0.1018410325050354,
0.05500923469662666,
0.008229468017816544,
-0.028460439294576645,
-0.034701768308877945,
0.04340483248233795,
-0.101103775203228,
-0.07066018879413605,
-0.06333664804697037,
-0.052413925528526306,
0.03451480343937874,
0.03061152808368206,
-0.05753167346119881,
-0.0512191504240036,
0.005698542110621929,
-0.0032466589473187923,
0.023738332092761993,
-0.015028810128569603,
0.06893037259578705,
0.005572188179939985,
0.021584149450063705,
-0.05931101366877556,
0.016464604064822197,
0.010562875308096409,
-0.001075398176908493,
-0.06618651747703552,
-0.003718725172802806,
-0.0025605475530028343,
-0.020438531413674355,
0.027625776827335358,
0.0085211843252182,
-0.005691403057426214,
-0.054765891283750534,
-0.033415116369724274,
0.02958209626376629,
0.0036393331829458475,
0.04100894182920456,
0.06178444251418114,
-0.03475608304142952,
0.005314686335623264,
0.011925308033823967,
-0.05749450623989105,
-0.06867082417011261,
5.205428317599393e-34,
-0.007371915969997644,
0.03918606415390968,
-0.046753913164138794,
0.08450648188591003,
0.03105890564620495,
-0.06068047136068344,
0.025527210906147957,
-0.011691078543663025,
0.020880838856101036,
0.028369098901748657,
-0.048445772379636765,
0.07632596790790558,
-0.027758251875638962,
0.03338209539651871,
0.025660820305347443,
0.05553457513451576,
-0.0405549518764019,
0.02392515167593956,
-0.007210630923509598,
-0.040346600115299225,
0.016891876235604286,
0.045776527374982834,
0.0401228591799736,
-0.08007004857063293,
-0.01179596222937107,
-0.05583781749010086,
0.0024841066915541887,
-0.03738394007086754,
-0.0934344083070755,
0.003179829567670822,
-0.10767889767885208,
0.04553871229290962,
-0.05806169658899307,
-0.00951696652919054,
0.04986802861094475,
-0.05332519859075546,
0.06124274432659149,
-0.04918578267097473,
0.06998303532600403,
-0.042027927935123444,
0.010290347971022129,
0.038103558123111725,
-0.04111936688423157,
-0.0043394071981310844,
-0.06351491063833237,
0.03145389258861542,
0.03796781972050667,
-0.012292376719415188,
-0.032740190625190735,
-0.04263279587030411,
-0.040683358907699585,
-0.015842439606785774,
0.02051585353910923,
-0.04518668353557587,
0.014855386689305305,
0.1379108726978302,
-0.01510576717555523,
0.05479058250784874,
-0.01939467526972294,
0.033533014357089996,
0.024147288873791695,
0.08719625324010849,
-0.01327342540025711,
0.031945161521434784,
0.032146625220775604,
0.014067892916500568,
-0.009348451159894466,
0.07076792418956757,
0.13541646301746368,
-0.04468904063105583,
0.03985103592276573,
-0.0006937467260286212,
0.005035029258579016,
-0.04160572215914726,
0.09203533828258514,
-0.05473187193274498,
0.05861184000968933,
-0.05378174036741257,
0.0733519047498703,
0.027010679244995117,
-0.02206510305404663,
-0.00968805979937315,
-0.052771225571632385,
-0.024235453456640244,
-0.06564320623874664,
-0.05814896523952484,
0.07556843757629395,
-0.07727537304162979,
-0.010990848764777184,
-0.008720975369215012,
-0.04069431126117706,
-0.0060157785192132,
-0.01184688601642847,
-0.04980282485485077,
-0.04760726913809776,
-2.7443458456281756e-33,
-0.09142345190048218,
0.005298189353197813,
-0.06049753725528717,
0.0011402698000892997,
-0.06562887132167816,
-0.0006833153311163187,
-0.008259386755526066,
0.06115995720028877,
-0.01045492198318243,
-0.05666692554950714,
0.017430519685149193,
-0.05350806191563606,
-0.018358957022428513,
0.022829290479421616,
0.05036267265677452,
-0.01814888045191765,
-0.03821710869669914,
0.07088597863912582,
0.031834330409765244,
0.00014851291780360043,
0.04490064084529877,
0.10682081431150436,
-0.10325197875499725,
0.02638564445078373,
-0.0631457269191742,
0.027995485812425613,
-0.08116665482521057,
0.03999420255422592,
0.012111296877264977,
-0.0022717146202921867,
-0.09881123900413513,
0.002389429835602641,
-0.03426918014883995,
0.0257418155670166,
-0.06464124470949173,
-0.0034824879840016365,
-0.05544229596853256,
-0.045185454189777374,
-0.028510285541415215,
0.024606093764305115,
0.02586280182003975,
-0.03836821764707565,
-0.034565266221761703,
-0.001543840509839356,
0.015872890129685402,
-0.06820426881313324,
0.05891070514917374,
-0.0008278014720417559,
-0.039319269359111786,
0.02309253253042698,
0.0005899906391277909,
0.0342581644654274,
-0.04686937481164932,
-0.0323159322142601,
-0.1358371376991272,
0.00598116172477603,
0.08420831710100174,
-0.08900783956050873,
0.1419987976551056,
-0.05054910108447075,
0.0035097100771963596,
-0.01683308556675911,
0.013396053574979305,
0.04974224418401718,
-0.02927997335791588,
0.026694122701883316,
0.009915398433804512,
-0.04453319311141968,
0.06626471132040024,
-0.004572877660393715,
0.04408039525151253,
0.0004587464209180325,
0.010857848450541496,
-0.06281167268753052,
-0.01273089274764061,
-0.010411584749817848,
-0.07467921078205109,
-0.06480904668569565,
-0.08222854137420654,
-0.02926451340317726,
0.032431721687316895,
0.018320171162486076,
-0.0010122881503775716,
0.010068007744848728,
0.0836818516254425,
0.02814042940735817,
0.07416991144418716,
0.046638209372758865,
0.059783004224300385,
-0.052197348326444626,
0.027318714186549187,
0.06939224898815155,
-0.1070166602730751,
0.060021042823791504,
-0.06840646266937256,
-4.694203781241413e-8,
-0.085660919547081,
0.002249300479888916,
-0.048962999135255814,
0.002783541101962328,
0.05188065022230148,
-0.04608369618654251,
-0.020866742357611656,
0.08114846050739288,
-0.0811978280544281,
0.08740920573472977,
0.05095868930220604,
0.04484450817108154,
0.008970733731985092,
-0.07326485961675644,
0.03961048275232315,
0.11165950447320938,
0.054245930165052414,
0.06952723860740662,
-0.0005369037389755249,
-0.016288170590996742,
0.03627430647611618,
0.07744497060775757,
0.040597621351480484,
-0.015626391395926476,
0.0038546426221728325,
-0.08952640742063522,
-0.07135612517595291,
0.059537824243307114,
-0.011461152695119381,
-0.030136888846755028,
0.10105274617671967,
0.06561724096536636,
0.0065857768058776855,
-0.003791846102103591,
0.016620079055428505,
0.07642631232738495,
0.0652487501502037,
-0.031603775918483734,
-0.030624285340309143,
0.0657651275396347,
-0.0223231241106987,
0.03597962111234665,
0.0018735502380877733,
0.06411012262105942,
-0.01947917230427265,
0.009668407030403614,
-0.02229461632668972,
-0.04469597712159157,
-0.026599770411849022,
-0.05428365617990494,
0.10999545454978943,
0.0466005876660347,
-0.07844378799200058,
0.019842827692627907,
0.04796343296766281,
-0.0028394043911248446,
0.008077924139797688,
-0.05972539260983467,
-0.03710338473320007,
0.0743914544582367,
0.015540572814643383,
0.11054325103759766,
0.023562490940093994,
-0.05018257349729538
] |
VLMaterial: Procedural Material Generation with Large Vision-Language Models | https://openreview.net/forum?id=wHebuIb6IH | [
"Beichen Li",
"Rundi Wu",
"Armando Solar-Lezama",
"Changxi Zheng",
"Liang Shi",
"Bernd Bickel",
"Wojciech Matusik"
] | Spotlight | Procedural materials, represented as functional node graphs, are ubiquitous in computer graphics for photorealistic material appearance design. They allow users to perform intuitive and precise editing to achieve desired visual appearances. However, creating a procedural material given an input image requires professional knowledge and significant effort. In this work, we leverage the ability to convert procedural materials into standard Python programs and fine-tune a large pre-trained vision-language model (VLM) to generate such programs from input images. To enable effective fine-tuning, we also contribute an open-source procedural material dataset and propose to perform program-level augmentation by prompting another pre-trained large language model (LLM). Through extensive evaluation, we show that our method outperforms previous methods on both synthetic and real-world examples. | generative model, procedural material, appearance modeling | null | 8,994 | 2501.18623 | [
0.00990014523267746,
-0.032267630100250244,
0.04888958856463432,
0.004005033057183027,
0.026452189311385155,
-0.03219535946846008,
-0.06855522096157074,
-0.02773907780647278,
-0.06699608266353607,
-0.031655553728342056,
-0.07404350489377975,
-0.09776382148265839,
0.03466538339853287,
-0.004104468040168285,
0.06003925949335098,
0.008911887183785439,
0.04712056741118431,
0.04996572434902191,
0.00482546491548419,
-0.029966946691274643,
0.010498364455997944,
-0.004376987926661968,
0.030081668868660927,
-0.05692777782678604,
0.11957964301109314,
-0.037710294127464294,
0.015243140049278736,
-0.01575411856174469,
0.08263658732175827,
-0.059907909482717514,
0.017329517751932144,
0.046509988605976105,
0.031772173941135406,
0.06753473728895187,
0.00795094482600689,
0.11160019040107727,
-0.07285401970148087,
-0.007308694068342447,
-0.07804429531097412,
-0.01845693588256836,
-0.07423121482133865,
0.023077325895428658,
-0.016251638531684875,
-0.018058335408568382,
0.07227745652198792,
-0.02565503492951393,
-0.011412912048399448,
-0.018584053963422775,
0.01827111467719078,
0.027188576757907867,
-0.065155528485775,
-0.13214923441410065,
-0.06856213510036469,
0.005067013669759035,
-0.004617783240973949,
0.043558478355407715,
0.12621669471263885,
-0.04712069779634476,
0.04911487549543381,
-0.04826344549655914,
-0.05660118907690048,
0.0218505822122097,
-0.015109747648239136,
-0.026410451158881187,
-0.04236583784222603,
0.01944214105606079,
0.05508323386311531,
-0.023297589272260666,
0.024587424471974373,
-0.06357142329216003,
-0.006287523545324802,
0.029064597561955452,
0.01572297140955925,
0.022159449756145477,
0.009936421178281307,
0.041290149092674255,
-0.007078663911670446,
-0.056202881038188934,
-0.0019376251148059964,
-0.05712958797812462,
0.02082592435181141,
0.02423543483018875,
0.014123925007879734,
0.0013411765685305,
0.005707289557904005,
0.05034313350915909,
-0.048594389110803604,
0.0293212179094553,
0.04823765903711319,
0.04735184460878372,
-0.08014849573373795,
-0.03789440914988518,
-0.007623536046594381,
-0.03845078870654106,
-0.044762417674064636,
0.051438335329294205,
0.06103173643350601,
-0.15545019507408142,
-0.004282546695321798,
0.05905129760503769,
0.03168129548430443,
-0.05790823698043823,
0.028360828757286072,
-0.053121380507946014,
0.047288428992033005,
-0.004632794763892889,
0.09992112219333649,
0.016172854229807854,
-0.017376188188791275,
-0.023711657151579857,
0.08378683775663376,
0.01225512195378542,
-0.01538217905908823,
-0.012469803914427757,
-0.005554236471652985,
0.028953714296221733,
0.039787985384464264,
-0.04060642793774605,
-0.024590281769633293,
0.011921369470655918,
0.010119653306901455,
0.061126142740249634,
-0.004300747532397509,
0.0022443742491304874,
0.016935288906097412,
-0.09219034761190414,
-0.06689295917749405,
-1.0151791455601977e-33,
0.07271738350391388,
0.04398393630981445,
0.0868452712893486,
0.018376419320702553,
0.11106166988611221,
-0.034234460443258286,
0.035646528005599976,
-0.010818302631378174,
-0.07515516877174377,
-0.044554125517606735,
-0.05527427792549133,
-0.03034682758152485,
-0.09705974906682968,
0.19869086146354675,
0.037965092808008194,
-0.029427185654640198,
0.006164546590298414,
-0.010183255188167095,
-0.0029282609466463327,
0.04571232572197914,
-0.0427890308201313,
-0.03794829919934273,
0.03935757279396057,
-0.0029399110935628414,
-0.024843212217092514,
0.0926436260342598,
0.026868609711527824,
-0.023424580693244934,
-0.017931722104549408,
-0.01978158950805664,
-0.09427782893180847,
0.013941459357738495,
0.06092935800552368,
0.10084860771894455,
-0.009745362214744091,
-0.0241644736379385,
0.003863940015435219,
-0.0571674183011055,
0.06058381497859955,
0.014303788542747498,
0.019940191879868507,
0.09764526784420013,
0.043063800781965256,
0.004858505446463823,
-0.0907638669013977,
0.03361610695719719,
0.026684816926717758,
0.07346659153699875,
-0.03639448434114456,
-0.010760724544525146,
0.0585574135184288,
0.05502425134181976,
-0.019101211801171303,
-0.0194325502961874,
-0.07424116879701614,
0.020575063303112984,
0.09338510036468506,
-0.007925360463559628,
0.009461384266614914,
-0.03522783890366554,
0.050375912338495255,
0.05471579357981682,
0.07482864707708359,
0.07750873267650604,
-0.011726880446076393,
0.0005501523846760392,
0.02184281125664711,
0.012123228050768375,
0.03771046921610832,
-0.015007589012384415,
-0.0971650555729866,
0.035779643803834915,
-0.034180935472249985,
-0.030301449820399284,
0.001491817645728588,
-0.02519896812736988,
0.010634728707373142,
-0.13506631553173065,
0.04140208661556244,
0.0691310465335846,
-0.0462610088288784,
-0.009289577603340149,
-0.03537331521511078,
-0.0875554084777832,
0.008170578628778458,
-0.07042296975851059,
0.07090187072753906,
-0.011891254223883152,
0.016399722546339035,
-0.09696182608604431,
-0.03914019092917442,
-0.058774225413799286,
0.02493738755583763,
-0.03862196207046509,
0.019483376294374466,
1.2049862872640946e-34,
-0.01998433843255043,
0.05178595334291458,
-0.01596469059586525,
0.08335746079683304,
-0.015252645127475262,
-0.08657263964414597,
-0.027990611270070076,
0.020979726687073708,
0.011923118494451046,
-0.0038570102769881487,
0.017421597614884377,
-0.000574922189116478,
0.02148800902068615,
-0.0694291815161705,
-0.01903846673667431,
-0.10725118964910507,
-0.010049342177808285,
0.03368750214576721,
-0.009282425977289677,
-0.0019523208029568195,
-0.030368857085704803,
0.04516369104385376,
-0.03256206214427948,
-0.03837503120303154,
-0.08969327062368393,
0.027805767953395844,
-0.0506652407348156,
-0.012942853383719921,
-0.0700615867972374,
0.07070077955722809,
-0.027152474969625473,
-0.035662297159433365,
0.037559833377599716,
0.007061625365167856,
0.02101844735443592,
0.02126886509358883,
0.11076805740594864,
-0.05969783663749695,
-0.017788277938961983,
0.029138293117284775,
0.06666470319032669,
-0.051795534789562225,
-0.0809452012181282,
0.05093671381473541,
-0.14129026234149933,
0.014703287743031979,
-0.0421566516160965,
0.014720316976308823,
-0.0036436179652810097,
0.014081262052059174,
-0.06191692501306534,
0.010809218510985374,
0.0036264348309487104,
-0.025650136172771454,
-0.013327721506357193,
-0.11817772686481476,
0.0006384879816323519,
0.012019582092761993,
0.09087184816598892,
-0.059998489916324615,
-0.09674723446369171,
-0.04653056338429451,
0.0013918783515691757,
-0.0984172597527504,
0.008462296798825264,
-0.0033809335436671972,
-0.10147889703512192,
-0.05723851919174194,
0.005856735166162252,
-0.014145136810839176,
0.11080741882324219,
-0.00719674164429307,
-0.016857938840985298,
0.09260130673646927,
-0.002412687987089157,
-0.0366450697183609,
0.027015622705221176,
0.03483305871486664,
0.0025647254660725594,
-0.032554931938648224,
0.047156743705272675,
-0.04330291599035263,
-0.0008551637292839587,
0.1285627782344818,
0.053317952901124954,
0.07971420139074326,
-0.00458815973252058,
0.05295221135020256,
0.01835743524134159,
0.07462514191865921,
-0.054849881678819656,
0.11531901359558105,
0.04704577475786209,
0.06776237487792969,
-0.03736255317926407,
-3.57936649209023e-8,
-0.06851471960544586,
0.05081858113408089,
0.015306629240512848,
-0.000986084807664156,
0.03245314583182335,
-0.0380793921649456,
0.03987041115760803,
0.09470395743846893,
-0.05050545930862427,
-0.01141203660517931,
0.03661597892642021,
0.009924481622874737,
0.05096439644694328,
-0.09968885034322739,
0.022682160139083862,
0.09417273849248886,
0.035896673798561096,
0.1060650572180748,
-0.022014224901795387,
0.00922250747680664,
-0.008362755179405212,
0.030594129115343094,
-0.016607243567705154,
-0.0002991194778587669,
-0.03416941687464714,
-0.036211222410202026,
-0.05304064229130745,
0.08200469613075256,
-0.019925152882933617,
0.03110710345208645,
0.034848980605602264,
-0.007244682405143976,
-0.0012787735322490335,
0.06897813081741333,
-0.03688054159283638,
0.008874589577317238,
-0.01682317815721035,
-0.021920980885624886,
-0.02719063311815262,
-0.02260783314704895,
0.021849796175956726,
0.0567750446498394,
0.01390832383185625,
-0.01733304373919964,
-0.016765320673584938,
0.007056190632283688,
0.025922438129782677,
-0.08925753086805344,
0.04414025694131851,
-0.007680854294449091,
-0.03080916590988636,
-0.029784707352519035,
-0.02337263524532318,
-0.009008260443806648,
0.03334394097328186,
0.046237703412771225,
-0.021315019577741623,
0.04667271301150322,
0.054475534707307816,
0.09120545536279678,
-0.05635371804237366,
0.01778538152575493,
0.03759270906448364,
-0.058678049594163895
] |
To Trust or Not to Trust? Enhancing Large Language Models' Situated Faithfulness to External Contexts | https://openreview.net/forum?id=K2jOacHUlO | [
"Yukun Huang",
"Sanxing Chen",
"Hongyi Cai",
"Bhuwan Dhingra"
] | Spotlight | Large Language Models (LLMs) are often augmented with external contexts, such as those used in retrieval-augmented generation (RAG). However, these contexts can be inaccurate or intentionally misleading, leading to conflicts with the model’s internal knowledge. We argue that robust LLMs should demonstrate situated faithfulness, dynamically calibrating their trust in external information based on their confidence in the internal knowledge and the external context to resolve knowledge conflicts. To benchmark this capability, we evaluate LLMs across several QA datasets, including a newly created dataset featuring in-the-wild incorrect contexts sourced from Reddit posts. We show that when provided with both correct and incorrect contexts, both open-source and proprietary models tend to overly rely on external information, regardless of its factual accuracy. To enhance situated faithfulness, we propose two approaches: Self-Guided Confidence Reasoning (SCR) and Rule-Based Confidence Reasoning (RCR). SCR enables models to self-access the confidence of external information relative to their own internal knowledge to produce the most accurate answer. RCR, in contrast, extracts explicit confidence signals from the LLM and determines the final answer using predefined rules.
Our results show that for LLMs with strong reasoning capabilities, such as GPT-4o and GPT-4o mini, SCR outperforms RCR, achieving improvements of up to 24.2\% over a direct input augmentation baseline. Conversely, for a smaller model like Llama-3-8B, RCR outperforms SCR. Fine-tuning SCR with our proposed Confidence Reasoning Direct Preference Optimization (CR-DPO) method improves performance on both seen and unseen datasets, yielding an average improvement of 8.9\% on Llama-3-8B. In addition to quantitative results, we offer insights into the relative strengths of SCR and RCR. Our findings highlight promising avenues for improving situated faithfulness in LLMs. | Large Language Model, Knowledge Conflict, Retrieval Augmented Generation, Confidence Estimation, Reasoning | We benchmark the challenge of ensuring large language models remain situationally faithful to potentially incorrect external information and propose Self-Guided Confidence Reasoning to enhance LLM's reliability. | 8,980 | 2410.14675 | [
-0.06650435924530029,
-0.07361345738172531,
-0.01899372786283493,
0.07512860745191574,
0.0366087444126606,
-0.01892862468957901,
0.019283829256892204,
0.0526120588183403,
0.05612339824438095,
-0.07589081674814224,
-0.03646181523799896,
-0.09012613445520401,
0.13447821140289307,
0.0017473321640864015,
0.055878497660160065,
0.011553443036973476,
0.11105971038341522,
0.03198619559407234,
-0.09872007369995117,
-0.11209650337696075,
-0.027722639963030815,
0.060188744217157364,
0.019379829987883568,
-0.018886743113398552,
-0.004097894765436649,
-0.0782826840877533,
-0.009758435189723969,
0.03881927207112312,
0.0096258744597435,
0.000504770316183567,
0.06017277389764786,
0.04758908599615097,
-0.03872327134013176,
-0.031140616163611412,
0.004985594656318426,
-0.0005918406532146037,
-0.061662714928388596,
0.05755635350942612,
-0.008094347082078457,
-0.08780572563409805,
-0.04810472950339317,
-0.005492653232067823,
0.06280824542045593,
0.04991421848535538,
0.07275921106338501,
0.049461428076028824,
-0.005091124214231968,
0.06936357915401459,
-0.04723713919520378,
0.005359651520848274,
-0.1407945603132248,
-0.034464459866285324,
-0.005718090105801821,
0.05326635763049126,
-0.04770283401012421,
0.013050363399088383,
0.010185236111283302,
0.06469212472438812,
-0.027555769309401512,
-0.04083450883626938,
0.03860553354024887,
-0.11794453859329224,
-0.08978094160556793,
0.04146776720881462,
-0.03107333555817604,
0.036714356392621994,
0.006019714288413525,
0.07918377220630646,
-0.05398297682404518,
-0.004325920715928078,
-0.0261219572275877,
0.00921968836337328,
-0.012912364676594734,
0.09003492444753647,
-0.023387953639030457,
0.025695305317640305,
0.015547475777566433,
-0.047143854200839996,
0.008368026465177536,
-0.029213905334472656,
0.0210108682513237,
0.07260022312402725,
0.03524796664714813,
0.06174541637301445,
0.07161314785480499,
0.02949092537164688,
0.07609035074710846,
-0.0066071623004972935,
0.01850534975528717,
0.02566993236541748,
0.009471761994063854,
-0.09888669103384018,
0.02926599606871605,
-0.007547020446509123,
0.06351665407419205,
0.03589523211121559,
0.03396041691303253,
-0.06975813210010529,
-0.023280199617147446,
0.06849300861358643,
0.041264522820711136,
0.05718814581632614,
-0.0540449284017086,
-0.07115573436021805,
-0.015838181599974632,
-0.0010043338406831026,
-0.017774110659956932,
0.09784986078739166,
0.04401849955320358,
-0.03939574956893921,
-0.04628875479102135,
0.06569134443998337,
0.005797233898192644,
-0.020910026505589485,
0.09219229966402054,
0.017984669655561447,
0.01856212690472603,
0.04178315028548241,
0.0536254346370697,
-0.03713470697402954,
-0.030374323949217796,
0.06460395455360413,
0.1410244256258011,
0.019716257229447365,
0.008588332682847977,
-0.06604357063770294,
-0.02914619632065296,
1.92527442261566e-33,
0.051006950438022614,
0.0376146100461483,
0.06630844622850418,
0.04301019757986069,
0.008721033111214638,
0.011940432712435722,
-0.051453329622745514,
0.05346405878663063,
-0.00883133988827467,
-0.010234640911221504,
0.04675321280956268,
0.05041151121258736,
-0.012043568305671215,
-0.01623985357582569,
0.06484410911798477,
0.15663892030715942,
-0.12309956550598145,
0.012696368619799614,
-0.014379248954355717,
0.025067461654543877,
0.10837280750274658,
-0.053332313895225525,
0.018388886004686356,
-0.043996334075927734,
-0.0017549732001498342,
0.03232415392994881,
0.11837717890739441,
0.021232211962342262,
-0.026268012821674347,
0.012706248089671135,
-0.07719338685274124,
-0.052279990166425705,
-0.00003626824400271289,
0.009426011703908443,
0.058729033917188644,
-0.05051556974649429,
-0.02783806063234806,
-0.07145241647958755,
0.029704337939620018,
0.014571531675755978,
0.0347178615629673,
0.0025142200756818056,
0.041637737303972244,
0.11558803170919418,
-0.04400237277150154,
-0.05772934854030609,
-0.09184586256742477,
-0.0361490473151207,
-0.04290406033396721,
-0.036887045949697495,
0.06345570832490921,
0.08585639297962189,
-0.007449569646269083,
-0.030820783227682114,
0.003215357195585966,
0.002177610294893384,
-0.01787104457616806,
0.07634071260690689,
-0.0030307190027087927,
0.010310162790119648,
0.07538359612226486,
0.05469924211502075,
-0.013610662892460823,
0.08258510380983353,
-0.004794790409505367,
0.023130929097533226,
-0.05003204569220543,
0.007119768764823675,
0.018843621015548706,
-0.04523870348930359,
-0.022456279024481773,
-0.03584464266896248,
-0.03073236346244812,
-0.030855096876621246,
-0.030146438628435135,
-0.04963982477784157,
-0.012401514686644077,
-0.006026938557624817,
0.05523096024990082,
-0.043606799095869064,
0.07977073639631271,
-0.0211362075060606,
-0.04589414224028587,
0.005333366338163614,
-0.08866465091705322,
-0.08111731708049774,
0.0292965117841959,
-0.061141688376665115,
-0.015804214403033257,
-0.0018069262150675058,
0.04901173338294029,
-0.050143782049417496,
-0.019358130171895027,
0.029208092018961906,
-0.060739316046237946,
-2.0530760051469824e-33,
0.00778689282014966,
-0.06453682482242584,
0.050614386796951294,
0.03136463835835457,
-0.04268411174416542,
-0.04786580055952072,
0.009500388987362385,
0.02449614368379116,
0.04914436489343643,
-0.03288016840815544,
-0.08795356005430222,
-0.013233856298029423,
0.060656722635030746,
0.045013677328825,
0.00523755419999361,
-0.06321042031049728,
0.025041578337550163,
-0.0026531892362982035,
-0.07946190983057022,
0.03786800056695938,
0.022087564691901207,
-0.02029292657971382,
-0.11142032593488693,
0.07131484895944595,
-0.006262862589210272,
0.014143237844109535,
-0.044153131544589996,
-0.046826623380184174,
-0.07142095267772675,
0.014285244978964329,
-0.024531634524464607,
-0.019078586250543594,
-0.049505263566970825,
-0.06352882087230682,
-0.09726958721876144,
0.04322464019060135,
-0.0033207733649760485,
-0.06726834177970886,
-0.02310548909008503,
0.021846024319529533,
0.01184907741844654,
0.0038411756977438927,
-0.11439947038888931,
-0.08071460574865341,
-0.09508729726076126,
-0.048230092972517014,
-0.10620778053998947,
-0.02145562693476677,
0.01234122458845377,
0.015362067148089409,
0.04364713653922081,
-0.02872614748775959,
-0.030868805944919586,
0.0011311951093375683,
-0.00014117915998212993,
-0.046273574233055115,
-0.028701454401016235,
-0.006023111287504435,
0.06902013719081879,
0.032927628606557846,
-0.06952761858701706,
0.010365497320890427,
-0.02616833709180355,
-0.026537416502833366,
0.05031062662601471,
-0.05019976198673248,
-0.03813062608242035,
0.04608375206589699,
0.022185122594237328,
-0.03612067177891731,
0.07096260786056519,
-0.027339326217770576,
-0.037456147372722626,
-0.043003879487514496,
0.05201920121908188,
0.0010952355805784464,
0.028624992817640305,
-0.06821714341640472,
-0.03499689698219299,
-0.06815489381551743,
0.03565223515033722,
-0.04054585471749306,
0.09196531772613525,
0.06889477372169495,
-0.00038901178049854934,
0.017980234697461128,
0.06012770161032677,
0.04872853308916092,
0.011605527251958847,
0.01899033598601818,
0.003697033040225506,
-0.0026769039686769247,
-0.06270794570446014,
0.06254741549491882,
-0.052451156079769135,
-4.5608182119849516e-8,
-0.030484182760119438,
0.014132230542600155,
-0.025051148608326912,
0.14073801040649414,
0.006544703152030706,
-0.03726247698068619,
-0.03227374702692032,
0.04581217095255852,
-0.047994595021009445,
-0.014635330066084862,
-0.03813992813229561,
-0.008433089591562748,
-0.031717631965875626,
-0.04891527071595192,
-0.012880130670964718,
0.08566858619451523,
0.02586967684328556,
0.07322193682193756,
-0.025250406935811043,
-0.02821507304906845,
0.10793714970350266,
0.03720153123140335,
-0.04658022150397301,
0.010000064969062805,
0.06674040108919144,
-0.008042741566896439,
-0.054256558418273926,
0.10136701166629791,
0.0705975815653801,
0.04625943675637245,
0.015305387787520885,
-0.09103454649448395,
-0.020837049931287766,
-0.06821978092193604,
-0.03506118059158325,
0.09556199610233307,
0.036685049533843994,
-0.05212421342730522,
0.053520672023296356,
-0.07941769063472748,
0.015034650452435017,
0.07484626024961472,
-0.03577610105276108,
0.055436234921216965,
0.021549509838223457,
-0.003990907687693834,
-0.02547488361597061,
-0.04900640249252319,
-0.003389235120266676,
-0.003437975887209177,
-0.009758306667208672,
-0.09467676281929016,
0.005392789375036955,
0.054371047765016556,
0.048579417169094086,
0.002658478682860732,
-0.027851523831486702,
0.015340261161327362,
0.010090578347444534,
0.051453597843647,
0.010654586367309093,
0.02608276903629303,
0.0061694905161857605,
0.046831775456666946
] |
Representative Guidance: Diffusion Model Sampling with Coherence | https://openreview.net/forum?id=gWgaypDBs8 | [
"Anh-Dung Dinh",
"Daochang Liu",
"Chang Xu"
] | Spotlight | The diffusion sampling process faces a persistent challenge stemming from its incoherence, attributable to varying noise directions across different timesteps.
Our Representative Guidance (RepG) offers a new perspective to handle this issue by reformulating the sampling process with a coherent direction towards a representative target.
In this formulation, while the classic classifier guidance improves feature discernment by steering the model away from ambiguous features, it fails to provide a favourable representative target since the class label is overly compact and leads to sacrificed diversity and the adversarial generation problem.
In contrast, we leverage self-supervised representations as the coherent target and treat sampling as a downstream task, which refines image details and corrects errors rather than settling for simpler samples.
Our representative guidance achieves superior performance and illustrates the potential of pre-trained self-supervised models in image sampling. Our findings demonstrate that RepG not only substantially enhances vanilla diffusion sampling but also surpasses state-of-the-art benchmarks when combined with classifier-free guidance. | generative models, diffusion model | The paper investigates the inconsistency property of the diffusion model and propose a guidance method based on representative information to fix the inconsistency problem | 8,945 | null | [
-0.012910928577184677,
-0.06872513890266418,
0.04882867634296417,
0.07023827731609344,
0.14053411781787872,
-0.04159514605998993,
0.003256979864090681,
-0.09887918084859848,
-0.006879385095089674,
-0.1338595449924469,
-0.06262367218732834,
-0.09559662640094757,
0.01363753154873848,
0.0014149986673146486,
-0.03142867982387543,
-0.029119716957211494,
0.08682799339294434,
0.03017040528357029,
-0.060921043157577515,
-0.022095026448369026,
0.013285939581692219,
-0.0016413343837484717,
0.0025794629473239183,
0.04638798162341118,
0.0160868838429451,
-0.031172089278697968,
0.06443909555673599,
-0.06588345766067505,
0.05582624301314354,
-0.07886113226413727,
0.06360120326280594,
-0.006778033450245857,
-0.1293778121471405,
-0.0057745627127587795,
-0.018579423427581787,
0.04431072622537613,
-0.007745308335870504,
0.06284796446561813,
0.021489841863512993,
-0.01621384732425213,
0.07807937264442444,
0.0010768177453428507,
-0.034355804324150085,
0.05651888996362686,
0.05339512228965759,
0.011792267672717571,
0.04889489710330963,
-0.0364784337580204,
0.019320201128721237,
0.03674130514264107,
0.004328847862780094,
-0.0010590238962322474,
-0.08875194936990738,
0.03249237313866615,
-0.034385427832603455,
0.03742389380931854,
0.0991835743188858,
0.09113676100969315,
0.047810789197683334,
0.0044351741671562195,
0.01942916214466095,
-0.10124176740646362,
-0.003789197653532028,
0.009495428763329983,
-0.016080481931567192,
-0.047757357358932495,
0.010278175584971905,
0.08604854345321655,
0.0023312370758503675,
0.0032769760582596064,
0.05621590092778206,
0.08645831048488617,
-0.03625671938061714,
0.06503589451313019,
0.06420908123254776,
0.06416362524032593,
-0.039689548313617706,
-0.004516678862273693,
0.10102635622024536,
-0.06802169978618622,
0.007388170342892408,
0.018809054046869278,
0.06965307891368866,
0.004477499518543482,
0.08875301480293274,
0.017699452117085457,
-0.019611069932579994,
0.04561452195048332,
0.02700732834637165,
-0.03811468929052353,
-0.051199160516262054,
-0.01624142937362194,
-0.05790301412343979,
-0.022428257390856743,
-0.017066815868020058,
0.011661574244499207,
0.03886253759264946,
-0.07077447324991226,
-0.01954628713428974,
0.10307564586400986,
-0.024379167705774307,
0.03350911661982536,
0.0005839768564328551,
0.01991875469684601,
0.030206894502043724,
-0.08297760784626007,
0.07117895781993866,
0.014368923380970955,
0.03401875123381615,
-0.004488827660679817,
0.023305822163820267,
-0.0043625254184007645,
-0.010049216449260712,
-0.01934608444571495,
0.12614348530769348,
-0.023203985765576363,
-0.05013458803296089,
0.04449240863323212,
0.04547682777047157,
0.0021749064326286316,
-0.05803878605365753,
-0.012300215661525726,
-0.058457233011722565,
-0.003411582438275218,
0.04149240255355835,
0.005640930030494928,
-0.0533231645822525,
2.1175301971124916e-33,
-0.00985682662576437,
-0.0033965676557272673,
0.07940295338630676,
0.011431338265538216,
0.03610403463244438,
-0.036082107573747635,
-0.02565581537783146,
0.017680475488305092,
0.008853207342326641,
0.05016738176345825,
0.05044781416654587,
-0.07124000042676926,
-0.06268493831157684,
0.11734671145677567,
0.007188572082668543,
0.012626647017896175,
-0.11131787300109863,
0.07261992990970612,
-0.03482013940811157,
-0.013901133090257645,
-0.0015213064616546035,
0.06433698534965515,
-0.07721611857414246,
-0.07132823765277863,
0.0008504783036187291,
0.0004666756431106478,
0.01044868677854538,
-0.018392402678728104,
0.02092246524989605,
0.03067261166870594,
-0.10597234219312668,
0.005336828529834747,
-0.029320785775780678,
0.004776148125529289,
0.041687414050102234,
0.01718662679195404,
0.006100554019212723,
-0.029320107772946358,
-0.021443724632263184,
-0.004452073480933905,
0.07649362832307816,
0.03485623747110367,
0.03453461453318596,
0.004937875550240278,
-0.09439994394779205,
0.0055876243859529495,
-0.022854773327708244,
-0.02644541673362255,
-0.10512463748455048,
-0.008949022740125656,
-0.03792967274785042,
-0.0028537658508867025,
-0.014352889731526375,
-0.05072108656167984,
-0.06467457115650177,
-0.05297606810927391,
0.007146845106035471,
-0.008751902729272842,
0.020596586167812347,
0.01055060513317585,
0.06565168499946594,
0.03846117481589317,
-0.04710673168301582,
0.09210994094610214,
0.05406646803021431,
0.04324731603264809,
-0.03195064887404442,
0.07622397691011429,
0.03449420630931854,
0.01746957190334797,
-0.0704331174492836,
0.033939532935619354,
-0.0831262618303299,
-0.07487299293279648,
0.05596331134438515,
-0.05145587399601936,
0.11301051825284958,
0.02995259128510952,
0.041657812893390656,
-0.00018738338258117437,
-0.016339534893631935,
-0.07068733125925064,
-0.027832724153995514,
-0.046661376953125,
-0.0863100066781044,
0.011157351545989513,
0.034465398639440536,
-0.05077086761593819,
-0.014928172342479229,
0.02251676842570305,
-0.020030487328767776,
0.03161856159567833,
-0.007460960652679205,
0.02874000556766987,
-0.05781189724802971,
-1.5405746108548733e-33,
0.060633584856987,
0.07292719930410385,
0.04443385824561119,
0.004490203224122524,
0.011000044643878937,
0.01804780773818493,
0.03791392594575882,
0.025757301598787308,
0.07957128435373306,
-0.07780761271715164,
-0.08215245604515076,
-0.018402397632598877,
0.0015238336054608226,
-0.02640499174594879,
-0.044840507209300995,
-0.0410371869802475,
0.0019442606717348099,
0.025542793795466423,
-0.01655951514840126,
0.05281321331858635,
-0.030151691287755966,
0.07771016657352448,
0.012110242620110512,
0.025352103635668755,
-0.05197128653526306,
0.058397188782691956,
0.0036170491948723793,
0.02229538932442665,
0.04545018821954727,
-0.07362474501132965,
-0.031311240047216415,
-0.015370593406260014,
0.045325931161642075,
-0.09177756309509277,
-0.09686768800020218,
0.06131916120648384,
0.0167950801551342,
0.03420470654964447,
-0.0005289461114443839,
0.1289188116788864,
0.0038123009726405144,
-0.04141835868358612,
-0.08980043977499008,
0.04428429529070854,
-0.06808701157569885,
-0.059463657438755035,
-0.04817958548665047,
0.03652277588844299,
0.007791413925588131,
0.07073862850666046,
-0.06044788286089897,
0.044222112745046616,
-0.1430845856666565,
0.05956157669425011,
-0.054666925221681595,
-0.07665862888097763,
-0.009333089925348759,
-0.01637857034802437,
0.07208039611577988,
0.042679574340581894,
-0.0955614522099495,
0.004511634353548288,
-0.015593077056109905,
-0.07901854813098907,
-0.0014777048490941525,
-0.0766678974032402,
-0.06371020525693893,
-0.03156602010130882,
0.04275857284665108,
0.03700777515769005,
-0.004399649333208799,
-0.041673507541418076,
-0.0013245211448520422,
-0.005891417618840933,
-0.052037857472896576,
-0.11435306072235107,
0.03856511414051056,
-0.02497762069106102,
-0.00044161625555716455,
-0.016641661524772644,
0.05338646471500397,
-0.010337690822780132,
0.014017869718372822,
0.020840415731072426,
0.04182594642043114,
0.04259895160794258,
0.01841021701693535,
0.0035714968107640743,
0.015769867226481438,
0.03438267111778259,
0.021849937736988068,
-0.017352120950818062,
0.004193814471364021,
0.098710797727108,
-0.005155711434781551,
-4.537761455480904e-8,
-0.11064179986715317,
0.09078483283519745,
0.08217133581638336,
0.0639113336801529,
-0.002136131515726447,
-0.05643561854958534,
0.00977195892482996,
0.062184564769268036,
-0.013644049875438213,
-0.014565967954695225,
0.08270946890115738,
-0.01637909933924675,
-0.04308958724141121,
-0.030309027060866356,
0.00410763593390584,
0.010063293389976025,
0.031810011714696884,
0.05311714485287666,
-0.02384716458618641,
-0.006538721267133951,
-0.04531721770763397,
-0.04650167375802994,
0.024176202714443207,
-0.10362841188907623,
0.08588431775569916,
0.0015690414002165198,
0.03692477196455002,
0.07181766629219055,
0.06775601953268051,
0.00871074479073286,
-0.06030791252851486,
0.05542891100049019,
-0.040594395250082016,
0.004074887838214636,
-0.007481682114303112,
0.1099359318614006,
-0.032103367149829865,
0.04727955907583237,
-0.03768115118145943,
-0.04671551287174225,
-0.07792001217603683,
0.07659673690795898,
-0.04136093705892563,
-0.0266665481030941,
0.00005255557698546909,
0.014600863680243492,
0.06430062651634216,
-0.06546549499034882,
0.026671763509511948,
0.0038335295394062996,
0.0054581561125814915,
-0.09108792990446091,
0.06032487377524376,
0.06991350650787354,
0.07812274247407913,
0.013322563841938972,
-0.039188697934150696,
-0.027209734544157982,
0.012564229778945446,
0.09121085703372955,
0.0640253946185112,
0.034005019813776016,
-0.014038045890629292,
-0.005763544701039791
] |
AIR-BENCH 2024: A Safety Benchmark based on Regulation and Policies Specified Risk Categories | https://openreview.net/forum?id=UVnD9Ze6mF | [
"Yi Zeng",
"Yu Yang",
"Andy Zhou",
"Jeffrey Ziwei Tan",
"Yuheng Tu",
"Yifan Mai",
"Kevin Klyman",
"Minzhou Pan",
"Ruoxi Jia",
"Dawn Song",
"Percy Liang",
"Bo Li"
] | Spotlight | Foundation models (FMs) provide societal benefits but also amplify risks. Governments, companies, and researchers have proposed regulatory frameworks, acceptable use policies, and safety benchmarks in response. However, existing public benchmarks often define safety categories based on previous literature, intuitions, or common sense, leading to disjointed sets of categories for risks specified in recent regulations and policies, which makes it challenging to evaluate and compare FMs across these benchmarks. To bridge this gap, we introduce AIR-BENCH 2024, the first AI safety benchmark aligned with emerging government regulations and company policies, following the regulation-based safety categories grounded in the AI Risks taxonomy, AIR 2024. AIR 2024 decomposes 8 government regulations and 16 company policies into a four-tiered safety taxonomy with 314 granular risk categories in the lowest tier. AIR-BENCH 2024 contains 5,694 diverse prompts spanning these categories, with manual curation and human auditing to ensure quality. We evaluate leading language models on AIR-BENCH 2024 uncovering insights into their alignment with specified safety concerns. By bridging the gap between public benchmarks and practical AI risks, AIR-BENCH 2024 provides a foundation for assessing model safety across jurisdictions, fostering the development of safer and more responsible AI systems. | AI Safety, Regulation, Policy, Safety Alignment, Foundation Models | A LLM Safety evaluation benchmark focus on 314 Risk categories specified in 8 regulation and 16 AI companies' policies | 8,850 | null | [
-0.029796283692121506,
-0.04356447979807854,
-0.03157838433980942,
0.003145224181935191,
0.12407326698303223,
0.04690844565629959,
0.043618012219667435,
-0.004652957431972027,
-0.03782682865858078,
0.026023566722869873,
-0.04816026985645294,
-0.04224282503128052,
0.0005615391419269145,
0.021158535033464432,
-0.0122675197198987,
-0.0114324651658535,
0.04392741248011589,
-0.025160711258649826,
-0.13822244107723236,
-0.014024576172232628,
0.057999998331069946,
0.06461981683969498,
-0.00031883068731985986,
-0.009723411872982979,
-0.06541844457387924,
-0.05896740034222603,
-0.07541970908641815,
-0.0017736367881298065,
-0.04793902859091759,
0.002263049827888608,
0.01954519748687744,
-0.041336704045534134,
0.021573666483163834,
0.02352234162390232,
0.0394224189221859,
-0.031096914783120155,
0.0028049834072589874,
-0.04090626910328865,
0.03198075294494629,
-0.018872613087296486,
-0.09150435775518417,
-0.08550303429365158,
-0.046406395733356476,
0.02569757215678692,
0.02046084962785244,
-0.013505390845239162,
-0.0060129715129733086,
-0.06688355654478073,
-0.015409499406814575,
0.02141408436000347,
-0.08368594199419022,
-0.0017285217763856053,
0.07104882597923279,
-0.002468348480761051,
0.01660294085741043,
-0.05405619740486145,
-0.0073820073157548904,
-0.03628050908446312,
-0.054101623594760895,
-0.06095486879348755,
-0.006241138558834791,
-0.10034841299057007,
-0.026829862967133522,
0.023045755922794342,
0.014602473005652428,
-0.0033373828046023846,
-0.09152426570653915,
-0.00882735662162304,
0.008791927248239517,
0.04489006847143173,
-0.05108907073736191,
0.002822274575009942,
0.010791742242872715,
0.02972337044775486,
0.009276819415390491,
0.07177481800317764,
0.04084078222513199,
0.03925801441073418,
0.07503090798854828,
-0.12376409024000168,
0.032404717057943344,
-0.007373135536909103,
0.034026820212602615,
0.06490162760019302,
0.018914125859737396,
0.03686326742172241,
0.014169931411743164,
-0.020338525995612144,
0.031176531687378883,
0.07973304390907288,
-0.05606875196099281,
-0.051599081605672836,
0.1568506509065628,
0.03484164923429489,
0.05717373266816139,
0.03634723275899887,
-0.013494228944182396,
-0.05855388939380646,
-0.04759405180811882,
0.028549887239933014,
0.012073197402060032,
0.024136818945407867,
-0.030612139031291008,
-0.04821028932929039,
-0.02832183800637722,
-0.06429903209209442,
0.02236533910036087,
-0.04910288751125336,
0.0027596259023994207,
0.026551077142357826,
0.004985084757208824,
0.0925237163901329,
0.0503903403878212,
-0.08480534702539444,
0.005184215493500233,
0.02052777260541916,
-0.08414427936077118,
-0.01771017350256443,
0.05465954169631004,
0.017210152000188828,
-0.0005833369214087725,
0.05065348744392395,
0.06583130359649658,
0.0187999214977026,
0.057494696229696274,
0.04059356078505516,
-0.052537914365530014,
1.8878727637663873e-33,
-0.0022431586403399706,
0.04887224733829498,
-0.015089583583176136,
-0.005000736098736525,
-0.008573794737458229,
-0.04744052141904831,
-0.05435195565223694,
0.05010642111301422,
-0.028110837563872337,
0.09932669997215271,
-0.07372483611106873,
-0.0025810792576521635,
-0.041066721081733704,
0.018360400572419167,
0.06192147359251976,
-0.008173882961273193,
-0.03731900081038475,
0.0633997917175293,
-0.05308426171541214,
-0.016567867249250412,
0.04275107383728027,
-0.09074153006076813,
0.011130842380225658,
-0.0024591530673205853,
0.07949717342853546,
0.027997344732284546,
0.05299382284283638,
-0.02597225271165371,
0.040221743285655975,
0.06923794746398926,
-0.06628016382455826,
0.0651806890964508,
0.0576104037463665,
-0.026495767757296562,
0.013663957826793194,
-0.026893531903624535,
-0.11542024463415146,
-0.019705625250935555,
-0.013022752478718758,
0.0039374749176204205,
-0.09016227722167969,
0.019377747550606728,
0.008487873710691929,
0.024124247953295708,
0.045573409646749496,
-0.03902639076113701,
-0.03120705671608448,
-0.045442089438438416,
-0.04120022431015968,
-0.034791000187397,
-0.041523054242134094,
0.04463580623269081,
0.07038632035255432,
-0.026007914915680885,
-0.04908955469727516,
-0.06408432871103287,
0.012329264543950558,
-0.01858040876686573,
-0.06107622757554054,
0.021386241540312767,
-0.021003523841500282,
0.09755514562129974,
-0.012169500812888145,
0.05888846889138222,
-0.035215917974710464,
0.025792814791202545,
0.01746194250881672,
0.06593887507915497,
0.04621046036481857,
0.001663132687099278,
-0.03171756863594055,
0.02378135174512863,
-0.02176859602332115,
-0.029134342446923256,
-0.030102018266916275,
-0.018811358138918877,
0.040476251393556595,
-0.05217413231730461,
-0.012685908004641533,
-0.07241101562976837,
-0.0011845347471535206,
0.031782280653715134,
0.07135780155658722,
0.0031702069099992514,
-0.012945848517119884,
-0.07667003571987152,
0.022143486887216568,
0.0650271400809288,
-0.012589091435074806,
0.011240244843065739,
-0.0900692567229271,
-0.05438869073987007,
0.03251216560602188,
0.07640565186738968,
0.016944393515586853,
-3.950588557274082e-33,
-0.029375966638326645,
-0.05063817277550697,
-0.06844400614500046,
-0.030130717903375626,
-0.10118670016527176,
-0.03818269446492195,
-0.05594228580594063,
-0.06481345742940903,
0.06885036826133728,
0.02453755773603916,
-0.044485870748758316,
-0.026901571080088615,
0.048660214990377426,
0.017101628705859184,
0.05912432819604874,
-0.07995327562093735,
-0.08357863873243332,
-0.054019998759031296,
0.016465768218040466,
0.05130688473582268,
0.10653813183307648,
0.03357294574379921,
-0.12770143151283264,
0.08418294787406921,
-0.09180683642625809,
0.046771254390478134,
-0.07358308881521225,
-0.028736848384141922,
0.06229376047849655,
-0.005049864761531353,
-0.07226522266864777,
0.02485165186226368,
-0.013630851171910763,
0.12275328487157822,
-0.04636499658226967,
-0.0702456384897232,
0.05159426108002663,
-0.12884537875652313,
-0.04556804522871971,
0.025872843340039253,
0.04994704946875572,
0.03856915235519409,
-0.021913984790444374,
-0.036340318620204926,
-0.022496260702610016,
0.007682148367166519,
-0.001009295112453401,
-0.05148778483271599,
-0.007404828444123268,
-0.06275710463523865,
0.010097960010170937,
-0.06526952981948853,
-0.0985100269317627,
0.05651981383562088,
-0.013200919143855572,
-0.003284233622252941,
0.010607164353132248,
-0.08095623552799225,
-0.012752121314406395,
0.08099937438964844,
-0.02687646448612213,
0.07656127959489822,
0.0011613511014729738,
0.019986771047115326,
-0.03579412400722504,
-0.023771274834871292,
-0.0021655161399394274,
-0.021371297538280487,
0.01944141648709774,
-0.057353269308805466,
0.003308519022539258,
-0.05680495873093605,
-0.10066355019807816,
0.05622049421072006,
-0.027544235810637474,
0.010194046422839165,
0.013810734264552593,
-0.03929128870368004,
-0.06916084885597229,
0.053363170474767685,
0.051317326724529266,
-0.043942052870988846,
0.02047676406800747,
0.09995557367801666,
0.10373461991548538,
0.07314631342887878,
0.004282524809241295,
0.012438146397471428,
-0.010808447375893593,
0.08731825649738312,
-0.13300955295562744,
-0.008825341239571571,
-0.0783083513379097,
0.0720222070813179,
-0.07385788857936859,
-4.4552258771091147e-8,
0.01899304986000061,
0.06885648518800735,
0.029720712453126907,
0.04174954444169998,
-0.02178414911031723,
-0.03578248620033264,
-0.04933999106287956,
-0.00047799633466638625,
-0.034743908792734146,
-0.034016501158475876,
0.06321373581886292,
0.04743547365069389,
-0.01814728043973446,
-0.0026459370274096727,
-0.017580846324563026,
0.0494391955435276,
-0.0860222727060318,
0.095621757209301,
-0.011943851597607136,
0.05520137771964073,
0.05899091809988022,
0.05214636027812958,
-0.024595320224761963,
-0.028139442205429077,
0.0361035093665123,
-0.0444486103951931,
0.004170027561485767,
0.09800808876752853,
0.09262372553348541,
0.10543908923864365,
-0.021529987454414368,
0.04389317333698273,
-0.007388306315988302,
-0.011312228627502918,
-0.031284261494874954,
0.03204523026943207,
0.014139545150101185,
-0.022033998742699623,
0.056753404438495636,
0.039399515837430954,
0.06513893604278564,
0.05463474616408348,
-0.026784390211105347,
0.05269383266568184,
0.04964643344283104,
-0.05120503902435303,
-0.07773460447788239,
-0.0469466894865036,
0.08117539435625076,
-0.05460673198103905,
-0.009898698888719082,
-0.04179784655570984,
0.04242274537682533,
0.1343424916267395,
0.06311380863189697,
0.015208500437438488,
-0.050070833414793015,
-0.06586754322052002,
-0.005086986813694239,
0.028225412592291832,
0.06203456223011017,
0.0029462911188602448,
0.008508272469043732,
0.023107530549168587
] |
DARE the Extreme: Revisiting Delta-Parameter Pruning For Fine-Tuned Models | https://openreview.net/forum?id=avSocG0oFA | [
"Wenlong Deng",
"Yize Zhao",
"Vala Vakilian",
"Minghui Chen",
"Xiaoxiao Li",
"Christos Thrampoulidis"
] | Spotlight | Storing open-source fine-tuned models separately introduces redundancy and increases response times in applications utilizing multiple models. Delta-parameter pruning (DPP), particularly the random drop and rescale (DARE) method proposed by Yu et al., addresses this by pruning the majority of delta parameters—the differences between fine-tuned and pre-trained model weights—while typically maintaining minimal performance loss. However, DARE fails when either the pruning rate or the magnitude of the delta parameters is large. We highlight two key reasons for this failure: (1) an excessively large rescaling factor as pruning rates increase, and (2) high mean and variance in the delta parameters. To push DARE’s limits, we introduce DAREx (DARE the eXtreme), which features two algorithmic improvements: (1) DAREx-q, a rescaling factor modification that significantly boosts performance at high pruning rates (e.g., > 30% on COLA and SST2 for encoder models, with even greater gains in decoder models), and (2) DAREx-L2, which combines DARE with AdamR, an in-training method that applies appropriate delta regularization before DPP. We also demonstrate that DAREx-q can be seamlessly combined with vanilla parameter-efficient fine-tuning techniques like LoRA and can facilitate structural DPP. Additionally, we revisit the application of importance-based pruning techniques within DPP, demonstrating that they outperform random-based methods when delta parameters are large. Through this comprehensive study, we develop a pipeline for selecting the most appropriate DPP method under various practical scenarios. | Delta parameter pruning, Efficiency, Large Language Models | null | 8,801 | 2410.09344 | [
-0.05856107920408249,
-0.05438357591629028,
-0.013450867496430874,
0.051694199442863464,
0.0020100960973650217,
-0.02442682720720768,
-0.06513035297393799,
0.028757460415363312,
0.027170058339834213,
-0.09447586536407471,
-0.005126306787133217,
-0.0026516274083405733,
0.009205739013850689,
-0.024498043581843376,
0.03291485831141472,
-0.08473693579435349,
0.02906227298080921,
0.11712267249822617,
-0.15250708162784576,
-0.10145794600248337,
0.007532806601375341,
-0.04592444375157356,
-0.012059520930051804,
-0.01738477684557438,
0.06095841899514198,
-0.07902857661247253,
-0.021428080275654793,
-0.0010375281563028693,
0.05080884322524071,
-0.09269732981920242,
0.04804093390703201,
0.04651862755417824,
-0.006456695031374693,
-0.0159324798732996,
-0.0915919691324234,
-0.00029594654915854335,
-0.03603527694940567,
-0.061006251722574234,
-0.03727084398269653,
-0.05728885903954506,
-0.013647475279867649,
0.0353848822414875,
0.029564060270786285,
0.05990948900580406,
0.02144077979028225,
-0.022217610850930214,
0.03787631914019585,
-0.020497698336839676,
0.009358496405184269,
-0.05895712971687317,
-0.0023972929921001196,
0.014737019315361977,
-0.03559555858373642,
0.07205744087696075,
-0.017400959506630898,
-0.0128807807341218,
0.004397011827677488,
-0.012926413677632809,
-0.04297700896859169,
-0.030075255781412125,
-0.052132878452539444,
-0.017685189843177795,
-0.07768142223358154,
-0.0015060020377859473,
-0.006067777052521706,
-0.10883678495883942,
0.07830211520195007,
-0.013754870742559433,
0.021014008671045303,
0.04342016577720642,
-0.015250486321747303,
0.05740663781762123,
-0.038339994847774506,
0.08996295183897018,
0.09397957473993301,
0.020442793145775795,
0.058280471712350845,
-0.0010816419962793589,
0.03403768315911293,
0.034682497382164,
-0.030188890174031258,
-0.04760099574923515,
0.01840001717209816,
-0.026479732245206833,
0.05142755061388016,
-0.05681774765253067,
0.03019557148218155,
0.02927261032164097,
0.12133264541625977,
-0.051898062229156494,
0.0032224534079432487,
0.07873493432998657,
-0.04839422181248665,
0.0012356588849797845,
-0.05427606403827667,
0.06010466441512108,
-0.03588198125362396,
-0.03876091167330742,
-0.08382592350244522,
0.10012457519769669,
0.030601300299167633,
0.054472118616104126,
0.048987727612257004,
-0.03694133087992668,
0.036504197865724564,
-0.07025499641895294,
0.047643378376960754,
0.046883054077625275,
0.008885991759598255,
-0.04647054895758629,
0.025392835959792137,
0.0654720813035965,
0.013706653378903866,
0.02859587036073208,
0.033741507679224014,
-0.008345642127096653,
-0.06037001684308052,
-0.07067406177520752,
-0.010739341378211975,
0.027136176824569702,
-0.046254415065050125,
0.05602870509028435,
-0.04541277140378952,
0.005810687784105539,
0.03929723799228668,
0.04158909618854523,
-0.08497107028961182,
4.984455675289454e-33,
0.006376841105520725,
0.06274732202291489,
0.032003920525312424,
-0.07426726818084717,
0.050658829510211945,
-0.0008904281421564519,
0.05300436541438103,
0.04548722133040428,
-0.05501219630241394,
0.025347445160150528,
-0.060756441205739975,
0.007303144317120314,
-0.05205973982810974,
0.05172937735915184,
0.044962406158447266,
-0.008153540082275867,
-0.0033024107106029987,
0.05391635373234749,
-0.042623311281204224,
-0.027000483125448227,
0.10350643843412399,
0.0048465607687830925,
0.03098255954682827,
-0.037999559193849564,
0.02654302306473255,
0.02888590656220913,
0.05130968242883682,
-0.0005106360767967999,
-0.12880994379520416,
0.0310188177973032,
-0.0596257708966732,
0.003359174821525812,
-0.02916456200182438,
0.013940428383648396,
0.014895250089466572,
-0.07333901524543762,
0.03159422054886818,
0.0024792689364403486,
-0.05956830084323883,
-0.06105643883347511,
-0.02098989300429821,
0.07342872023582458,
-0.013257825747132301,
-0.03902469202876091,
-0.05652395635843277,
-0.06821730732917786,
0.04299140349030495,
-0.01234609354287386,
-0.0075879329815506935,
0.0038596701342612505,
0.02087237872183323,
0.018326535820961,
0.01921261101961136,
0.004162367898970842,
-0.02204694040119648,
0.009733493439853191,
0.09714695811271667,
0.013683007098734379,
-0.007345674093812704,
0.05741414427757263,
-0.07245303690433502,
0.018157431855797768,
-0.04936306178569794,
0.12006320059299469,
0.0903632789850235,
0.023809485137462616,
0.029457978904247284,
0.0635908767580986,
-0.017675157636404037,
0.03891436383128166,
-0.028636164963245392,
-0.01196814700961113,
-0.004361671861261129,
-0.08031723648309708,
0.05893727391958237,
-0.07348692417144775,
0.16566938161849976,
-0.052118703722953796,
0.042753398418426514,
-0.05419144034385681,
-0.05193064361810684,
0.08780037611722946,
-0.034678880125284195,
-0.01994485966861248,
-0.05166281759738922,
-0.08183801919221878,
0.02095673978328705,
-0.022787081077694893,
-0.023282617330551147,
0.002320921281352639,
-0.0493604876101017,
0.0056833806447684765,
0.02980678528547287,
-0.0764511451125145,
-0.016454437747597694,
-4.4860214586278335e-33,
-0.04311723634600639,
0.0028348788619041443,
-0.01686820760369301,
0.1233505979180336,
-0.0029944225680083036,
-0.08360644429922104,
0.03060331754386425,
0.07349462062120438,
0.0043241544626653194,
-0.115058034658432,
-0.02846292220056057,
-0.036481842398643494,
0.04467395693063736,
-0.005583249498158693,
0.05713726207613945,
0.000265038397628814,
-0.026254335418343544,
-0.03328128904104233,
0.02206786349415779,
-0.02759028784930706,
-0.0006036755512468517,
0.10813824832439423,
-0.06499665975570679,
0.00012444723688531667,
-0.03313279524445534,
-0.01850356161594391,
-0.04342015087604523,
0.053847454488277435,
-0.019966911524534225,
-0.10175929218530655,
-0.06652342528104782,
0.010269521735608578,
-0.007767448201775551,
-0.001029760343953967,
-0.03516566753387451,
0.018673386424779892,
0.025548135861754417,
0.027256794273853302,
0.02802199311554432,
0.11621148139238358,
0.06005985289812088,
-0.0014068343443796039,
-0.05866358429193497,
0.02011244185268879,
0.02969072200357914,
-0.030611183494329453,
-0.0718470886349678,
-0.07354652136564255,
-0.003729891264811158,
0.01145988516509533,
-0.02259480208158493,
-0.010798992589116096,
-0.03300799801945686,
0.01196789089590311,
0.007496203761547804,
-0.06541220098733902,
-0.02502020262181759,
-0.023776322603225708,
0.03097951039671898,
0.039036162197589874,
-0.010244251228868961,
-0.07930585741996765,
-0.00253697601146996,
-0.020381653681397438,
-0.027023443952202797,
-0.05633292347192764,
0.07527239620685577,
0.02777421846985817,
0.0472664050757885,
0.05008498951792717,
-0.022900495678186417,
0.04641122743487358,
0.007021599914878607,
0.06993556022644043,
-0.10836669057607651,
-0.029202217236161232,
-0.040881071239709854,
-0.0027874114457517862,
0.010195761919021606,
0.007155098486691713,
-0.002157441806048155,
0.02801649644970894,
-0.036074474453926086,
-0.02951633930206299,
0.09081350266933441,
0.07032253593206406,
0.05821764096617699,
0.06496578454971313,
0.06654448807239532,
0.008799358271062374,
-0.07008996605873108,
0.08433100581169128,
0.0736258253455162,
0.1281483918428421,
-0.023560045287013054,
-5.0024279829585794e-8,
-0.050325486809015274,
0.07858346402645111,
-0.05668387562036514,
0.11517643183469772,
0.02919081412255764,
-0.008557697758078575,
-0.007288947235792875,
0.094273142516613,
-0.010682892985641956,
0.0355338454246521,
0.11755898594856262,
-0.022732490673661232,
-0.03411611542105675,
0.04015139117836952,
0.0647541731595993,
0.05877048894762993,
0.04631510749459267,
-0.009871314279735088,
-0.057025790214538574,
-0.04197699949145317,
0.007708120159804821,
0.09552621096372604,
-0.007921985350549221,
-0.04532580077648163,
0.0502643920481205,
-0.0309308972209692,
-0.011278935708105564,
0.14715255796909332,
0.047105513513088226,
0.015331524424254894,
-0.030087046325206757,
0.005671946797519922,
0.046775199472904205,
-0.011471430771052837,
-0.04948049411177635,
0.04705503582954407,
0.045744821429252625,
0.0510203093290329,
0.015125779435038567,
0.0027932412922382355,
0.06729026138782501,
0.020603884011507034,
-0.09484583139419556,
0.021851148456335068,
-0.024184105917811394,
-0.027792658656835556,
0.0006570507539436221,
-0.03032698854804039,
0.06070166453719139,
-0.024009479209780693,
0.007956924848258495,
-0.012800277210772038,
-0.09490511566400528,
-0.01393952127546072,
-0.00452926242724061,
0.042457371950149536,
-0.041434671729803085,
0.012148589827120304,
-0.02123892679810524,
-0.012409144081175327,
0.04986465349793434,
-0.08312033116817474,
-0.12745578587055206,
0.012862282805144787
] |
CREIMBO: Cross-Regional Ensemble Interactions in Multi-view Brain Observations | https://openreview.net/forum?id=28abpUEICJ | [
"Noga Mudrik",
"Ryan Ly",
"Oliver Ruebel",
"Adam Shabti Charles"
] | Spotlight | Modern recordings of neural activity provide diverse observations of neurons across brain areas, behavioral conditions, and subjects; presenting an exciting opportunity to reveal the fundamentals of brain-wide dynamics. Current analysis methods, however, often fail to fully harness the richness of such data, as they provide either uninterpretable representations (e.g., via deep networks) or oversimplify models (e.g., by assuming stationary dynamics or analyzing each session independently). Here, instead of regarding asynchronous neural recordings that lack alignment in neural identity or brain areas as a limitation, we leverage these diverse views into the brain to learn a unified model of neural dynamics. Specifically, we assume that brain activity is driven by multiple hidden global sub-circuits. These sub-circuits represent global basis interactions between neural ensembles—functional groups of neurons—such that the time-varying decomposition of these sub-circuits defines how the ensembles' interactions evolve over time non-stationarily and non-linearly.
We discover the neural ensembles underlying non-simultaneous observations, along with their non-stationary evolving interactions, with our new model, **CREIMBO** (**C**ross-**R**egional **E**nsemble **I**nteractions in **M**ulti-view **B**rain **O**bservations). CREIMBO identifies the hidden composition of per-session neural ensembles through novel graph-driven dictionary learning and models the ensemble dynamics on a low-dimensional manifold spanned by a sparse time-varying composition of the global sub-circuits. Thus, CREIMBO disentangles overlapping temporal neural processes while preserving interpretability due to the use of a shared underlying sub-circuit basis. Moreover, CREIMBO distinguishes session-specific computations from global (session-invariant) ones by identifying session covariates and variations in sub-circuit activations. We demonstrate CREIMBO's ability to recover true components in synthetic data, and uncover meaningful brain dynamics in human high-density electrode recordings, including cross-subject neural mechanisms as well as inter- vs. intra-region dynamical motifs. Furthermore, using mouse whole-brain recordings, we show CREIMBO's ability to discover dynamical interactions that capture task and behavioral variables and meaningfully align with the biological importance of the brain areas they represent. | computational neuroscience, multi-regional brain interactions, sparsity, cross-session variability, dynamical systems modeling, neural dynamics, non-simultaneous neural recordings | We propose a framework for uncovering latent multi-regional neural sub-circuits by leveraging the richness and variability of multi-session neural data. | 8,800 | 2405.17395 | [
-0.05545981600880623,
-0.14178475737571716,
0.0530068576335907,
0.02495177835226059,
0.05759715288877487,
0.0663352683186531,
-0.03164523467421532,
-0.081930972635746,
0.08265503495931625,
-0.062005747109651566,
-0.09451047331094742,
-0.11851657181978226,
-0.0923442393541336,
0.0010384592460468411,
-0.07030287384986877,
-0.0369175523519516,
-0.01923254318535328,
0.01747906394302845,
-0.05222688987851143,
0.011450003832578659,
-0.01893453672528267,
-0.03222587704658508,
0.00042436030344106257,
0.09029444307088852,
0.0034383570309728384,
0.05232664942741394,
0.017367219552397728,
-0.0018541094614192843,
-0.0017092585330829024,
-0.04584112390875816,
0.08105886727571487,
0.0033945818431675434,
-0.02851758524775505,
0.018771113827824593,
-0.053372468799352646,
0.06381642818450928,
-0.06729377806186676,
0.014257102273404598,
0.04560008645057678,
-0.011219378560781479,
0.02368660271167755,
0.06502825021743774,
-0.0014615897089242935,
0.029978737235069275,
0.010189194232225418,
-0.004827951081097126,
0.029983656480908394,
-0.08722051978111267,
-0.04414106905460358,
-0.02801292948424816,
-0.06569655239582062,
0.019905313849449158,
-0.0042227185331285,
0.12077024579048157,
0.0200080294162035,
0.11029305309057236,
-0.032629698514938354,
0.0572931207716465,
0.011036144569516182,
0.015018030069768429,
-0.03424764797091484,
0.012804976664483547,
0.018262064084410667,
0.01773255504667759,
0.03569042310118675,
0.09234903752803802,
-0.023552969098091125,
0.09817127138376236,
-0.05379388481378555,
0.04900893196463585,
0.00009358100942336023,
0.05723356828093529,
-0.07039424031972885,
-0.09786406904459,
0.05449248105287552,
-0.011810936033725739,
0.013951160944998264,
-0.015769656747579575,
0.05988630652427673,
-0.06941308081150055,
0.007366053760051727,
0.049010664224624634,
0.03439531847834587,
-0.07613787055015564,
0.12728096544742584,
-0.025430534034967422,
-0.027894172817468643,
0.01975560374557972,
0.011432485654950142,
0.005038904491811991,
-0.0752803236246109,
-0.02096468210220337,
-0.0046731261536479,
-0.025080375373363495,
0.0545012429356575,
0.03364565223455429,
0.013510661199688911,
0.049804698675870895,
0.11320702731609344,
0.07151837646961212,
-0.0014732818817719817,
0.057725876569747925,
0.050242580473423004,
0.0213769543915987,
0.03186418116092682,
-0.01735658571124077,
0.02818925864994526,
0.014994350261986256,
0.020815342664718628,
-0.0017097904346883297,
-0.021860256791114807,
0.023738114163279533,
0.012692841701209545,
0.0950143113732338,
0.049970462918281555,
-0.002269660821184516,
0.03141256049275398,
0.01937580481171608,
0.05847404524683952,
0.05180109664797783,
0.001012675347737968,
-0.0413273423910141,
-0.07549770176410675,
0.004874584265053272,
0.04141944274306297,
0.00611883495002985,
-0.1197420284152031,
9.342955833320239e-34,
0.020639142021536827,
-0.11148642748594284,
0.015013504773378372,
0.03920818492770195,
0.024506473913788795,
-0.02740810066461563,
-0.03308775648474693,
-0.07783311605453491,
0.022238466888666153,
0.08680025488138199,
-0.04233880341053009,
0.1277524083852768,
0.036721959710121155,
0.08025551587343216,
-0.05167045444250107,
0.036597371101379395,
-0.09266805648803711,
0.018033191561698914,
0.031623393297195435,
-0.06606123596429825,
0.011318349279463291,
0.024017639458179474,
-0.03821999579668045,
-0.04430342838168144,
-0.04038463905453682,
-0.022894421592354774,
0.0010502614313736558,
-0.0033694785088300705,
-0.055346470326185226,
0.02166803926229477,
-0.06947626173496246,
0.01612365059554577,
-0.076996810734272,
0.005638542119413614,
0.044768642634153366,
-0.00011227236973354593,
0.052019428461790085,
-0.03448471054434776,
-0.010543429292738438,
0.0030279685743153095,
0.033174045383930206,
-0.005598206538707018,
-0.07254038006067276,
-0.04082448035478592,
-0.021719934418797493,
-0.03724776208400726,
0.05432809516787529,
0.03780833259224892,
-0.006334912497550249,
-0.0641382560133934,
-0.07563337683677673,
-0.02487700805068016,
0.030109550803899765,
-0.07055430114269257,
-0.04167773574590683,
0.04943600296974182,
-0.021834082901477814,
0.013207817450165749,
0.061892662197351456,
0.08831518888473511,
-0.04081015661358833,
0.07127774506807327,
-0.023867426440119743,
-0.0045165098272264,
0.04548059403896332,
0.04947691038250923,
-0.08691989630460739,
0.05759672075510025,
0.09573137760162354,
-0.04601849615573883,
0.0289379321038723,
0.054454922676086426,
-0.01402481272816658,
-0.08007186651229858,
0.08637583255767822,
-0.04243616759777069,
0.05631242319941521,
-0.004795153625309467,
-0.07981384545564651,
0.03251161798834801,
-0.0072908843867480755,
-0.10651811212301254,
-0.04119575396180153,
0.04185275733470917,
-0.06056070700287819,
-0.025429774075746536,
0.06519628316164017,
-0.05952391028404236,
-0.07594569027423859,
-0.025907589122653008,
-0.052558571100234985,
-0.061644237488508224,
0.09090041369199753,
-0.05798742175102234,
-0.03178872913122177,
-2.6721078606793074e-33,
-0.028367651626467705,
-0.08829449862241745,
0.0028008439112454653,
-0.03536978363990784,
-0.043141014873981476,
0.009240502491593361,
0.01011863723397255,
-0.007257688790559769,
-0.06182195991277695,
-0.03753046318888664,
-0.0011658257571980357,
-0.01914868876338005,
0.033280789852142334,
-0.001295767375268042,
0.06487246602773666,
-0.0912017971277237,
0.004319013096392155,
0.05106391757726669,
0.08058056235313416,
-0.019799979403614998,
0.03997870534658432,
0.05928322300314903,
-0.027660991996526718,
-0.007310478948056698,
-0.020623445510864258,
0.07514326274394989,
-0.003160106483846903,
0.14102360606193542,
-0.000590363226365298,
0.01173989288508892,
-0.08790156245231628,
-0.023517807945609093,
-0.05737825110554695,
-0.03164137154817581,
-0.08577064424753189,
0.10742390155792236,
-0.03716573119163513,
-0.03865000233054161,
-0.03533657267689705,
-0.04193342477083206,
0.02448474057018757,
-0.03482432663440704,
-0.02188582345843315,
0.02499903365969658,
0.005532245151698589,
-0.019036948680877686,
-0.09130861610174179,
0.09250105917453766,
-0.06796664744615555,
0.020663311704993248,
-0.025939641520380974,
0.04281891509890556,
-0.08898615092039108,
-0.043304089456796646,
-0.06780240684747696,
0.03751260042190552,
0.011640146374702454,
-0.04486836865544319,
0.05975155159831047,
-0.00919851940125227,
-0.026548128575086594,
-0.0018307287245988846,
-0.03263142332434654,
-0.015403647907078266,
0.027233466506004333,
0.0076209004037082195,
0.019111888483166695,
-0.03536175936460495,
0.10213345289230347,
0.046716928482055664,
0.009625067003071308,
-0.003446388989686966,
-0.030221683904528618,
-0.014899689704179764,
0.01648203656077385,
0.016291556879878044,
-0.02419143170118332,
-0.05576423928141594,
0.00598507234826684,
-0.03433048725128174,
0.005628540646284819,
-0.02411852590739727,
0.023027436807751656,
0.02543838880956173,
0.0026100073009729385,
0.027286671102046967,
-0.027514144778251648,
0.05899237468838692,
0.014034006744623184,
-0.051906030625104904,
-0.0003261203528381884,
0.004576563835144043,
-0.04993179067969322,
0.021760448813438416,
-0.03650690242648125,
-4.310286882969194e-8,
0.0035155299119651318,
0.07380632311105728,
0.06430201232433319,
0.05833732709288597,
0.04908619076013565,
-0.040219713002443314,
0.04581887274980545,
-0.04213670641183853,
-0.04860864579677582,
0.0164140947163105,
0.13923144340515137,
-0.05567431077361107,
0.07231592386960983,
-0.028883835300803185,
0.030029255896806717,
0.06486736238002777,
0.0026821710634976625,
0.025063367560505867,
-0.04249420762062073,
-0.006436185445636511,
0.05157026648521423,
0.020407812669873238,
0.05520752817392349,
-0.016502564772963524,
0.06622876226902008,
-0.020849568769335747,
-0.07509701699018478,
0.06276257336139679,
-0.05632355436682701,
-0.02072155475616455,
0.0006877679843455553,
0.10514489561319351,
0.020774899050593376,
-0.0016093782614916563,
0.0005417460924945772,
0.04139217361807823,
0.020525222644209862,
0.009929107502102852,
-0.0758490264415741,
-0.05362730845808983,
-0.10130544751882553,
-0.023016951978206635,
-0.01860802061855793,
0.010787969455122948,
0.016694827005267143,
-0.08140154182910919,
-0.0012864203890785575,
-0.04910778999328613,
0.10793186724185944,
-0.004784492775797844,
-0.018210148438811302,
0.025034388527274132,
0.027962757274508476,
0.0834958404302597,
0.04277581721544266,
0.045166030526161194,
-0.046941570937633514,
-0.07297752052545547,
0.025971544906497,
0.042108889669179916,
0.028576159849762917,
0.039203330874443054,
-0.13150890171527863,
-0.0525212287902832
] |
Learning vector fields of differential equations on manifolds with geometrically constrained operator-valued kernels | https://openreview.net/forum?id=OwpLQrpdwE | [
"Daning Huang",
"Hanyang He",
"John Harlim",
"Yan Li"
] | Spotlight | We address the problem of learning ordinary differential equations (ODEs) on manifolds. Existing machine learning methods, particularly those using neural networks, often struggle with high computational demands. To overcome this issue, we introduce a geometrically constrained operator-valued kernel that allows us to represent vector fields on tangent bundles of smooth manifolds. The construction of the kernel imposes the geometric constraints that are estimated from the data and ensures the computational feasibility for learning high dimensional systems of ODEs. Once the vector fields are estimated, e.g., by the kernel ridge regression, we need an ODE solver that guarantees the solution to stay on (or close to) the manifold. To overcome this issue, we propose a geometry-preserving ODE solver that approximates the exponential maps corresponding to the ODE solutions. We deduce a theoretical error bound for the proposed solver that guarantees the approximate solutions to lie on the manifold in the limit of large data. We verify the effectiveness of the proposed approach on high-dimensional dynamical systems, including the cavity flow problem, the beating and travelling waves in Kuramoto-Sivashinsky equations, and the reaction-diffusion dynamics. | Dynamics on manifolds, Operator-valued kernel, Geometry-preserving time integration, Ordinary differential equations | We present a geometrically constrained kernel and a geometry-preserving ODE solver for learning ODEs on manifolds, improving computational efficiency and long-term prediction accuracy in high-dimensional systems. | 8,739 | null | [
-0.10028740018606186,
-0.04242604225873947,
0.02174762636423111,
0.007670873310416937,
-0.007096387445926666,
0.030360858887434006,
-0.010058668442070484,
-0.010631927289068699,
0.016228845342993736,
-0.004776916932314634,
0.004912505391985178,
0.018241191282868385,
-0.04186568036675453,
0.014843554235994816,
-0.04336823150515556,
0.0357346385717392,
-0.007417685352265835,
0.09182801842689514,
-0.0416305772960186,
-0.029384052380919456,
-0.049418121576309204,
0.07047707587480545,
-0.023527074605226517,
0.07828377932310104,
-0.005731179378926754,
0.04472889378666878,
0.09554049372673035,
0.04063442721962929,
-0.061116721481084824,
0.0012458593118935823,
0.06471393257379532,
-0.04128079116344452,
-0.1088453009724617,
-0.01783786527812481,
-0.10195503383874893,
0.12348012626171112,
-0.030844971537590027,
0.07091978192329407,
-0.07526765763759613,
-0.07133124768733978,
-0.017370328307151794,
0.03215359151363373,
0.017176633700728416,
0.06336791068315506,
0.12456954270601273,
0.014554239809513092,
0.002245982876047492,
-0.06938473880290985,
0.016419701278209686,
-0.026683159172534943,
-0.0975198820233345,
0.01541014015674591,
-0.016706140711903572,
-0.015085428953170776,
0.03301543369889259,
-0.0935075506567955,
-0.005462494678795338,
-0.004076172132045031,
0.07176082581281662,
0.07672248035669327,
0.014818446710705757,
-0.09506238251924515,
-0.0014584299642592669,
0.029386596754193306,
-0.05054379254579544,
0.020028185099363327,
0.07156874239444733,
0.07139502465724945,
-0.03173414617776871,
0.09030256420373917,
-0.09808357805013657,
0.04034605622291565,
-0.05683078244328499,
-0.015133294276893139,
0.004115027841180563,
-0.0038664438761770725,
0.02257191389799118,
0.03813963383436203,
0.006902918219566345,
0.033164866268634796,
0.05307658389210701,
0.021540697664022446,
0.005662036128342152,
-0.030164245516061783,
0.0022984277456998825,
0.02505975216627121,
-0.050969112664461136,
0.017621401697397232,
0.03784489259123802,
-0.04411112889647484,
0.06644734740257263,
-0.05440716817975044,
0.007891431450843811,
-0.04649082198739052,
0.0656743049621582,
-0.04833601787686348,
-0.013248339295387268,
0.01862310618162155,
0.050766315311193466,
0.04867592826485634,
-0.023571861907839775,
-0.060227036476135254,
-0.027076637372374535,
0.12084121257066727,
0.0849945917725563,
-0.04010709375143051,
0.03422271087765694,
0.049278128892183304,
0.016944650560617447,
-0.08806237578392029,
0.041237011551856995,
-0.001167968730442226,
-0.05662892013788223,
0.034302886575460434,
0.0039243982173502445,
-0.07282303273677826,
0.042570315301418304,
-0.06655588001012802,
0.04005223512649536,
0.05198417976498604,
-0.05522984638810158,
-0.035689592361450195,
-0.043356217443943024,
-0.033469073474407196,
0.027463190257549286,
-0.02418384701013565,
-0.060406479984521866,
-2.180227943378483e-33,
0.017618240788578987,
0.039579764008522034,
0.03228222206234932,
0.011810041032731533,
0.012544694356620312,
-0.006396034732460976,
-0.01675618439912796,
0.03252112865447998,
0.044694166630506516,
0.058629799634218216,
-0.1053759828209877,
-0.029221337288618088,
-0.043210554867982864,
0.04269058629870415,
-0.004664091859012842,
-0.04089074209332466,
0.007950140163302422,
0.01450413279235363,
-0.07473152875900269,
-0.05687742680311203,
0.05146166309714317,
-0.13596901297569275,
-0.04428219050168991,
-0.0023513431660830975,
-0.04118119552731514,
0.05156698822975159,
-0.011643794365227222,
0.028037479147315025,
0.01790432259440422,
-0.022936109453439713,
-0.01816495507955551,
-0.06591196358203888,
-0.006966329179704189,
0.09167122840881348,
0.0617007240653038,
0.04238929972052574,
-0.010365337133407593,
0.03759012743830681,
0.008789150044322014,
-0.022435497492551804,
-0.08575444668531418,
0.00344051793217659,
0.001681761583313346,
0.05643901228904724,
-0.052592579275369644,
0.014427175745368004,
0.04322504997253418,
0.06653765588998795,
0.005979763809591532,
-0.08006784319877625,
-0.04049399495124817,
-0.02268531732261181,
-0.03673987090587616,
-0.07723180949687958,
0.004025401081889868,
0.050529222935438156,
0.002024667337536812,
-0.052557725459337234,
0.029242809861898422,
-0.007237102370709181,
0.013062568381428719,
-0.024733323603868484,
0.021701090037822723,
-0.09593412280082703,
-0.01980748400092125,
0.04647265374660492,
0.026543764397501945,
0.025272876024246216,
0.009764714166522026,
-0.012778012081980705,
0.004338813479989767,
0.060097746551036835,
-0.003961550537496805,
-0.05188649147748947,
0.07120446860790253,
0.07498359680175781,
0.05116741731762886,
-0.018024994060397148,
-0.006307900883257389,
0.0920964926481247,
-0.025917373597621918,
-0.07866255193948746,
-0.02140248753130436,
0.020962731912732124,
0.052016694098711014,
-0.04841812700033188,
-0.0006138470489531755,
-0.07987260073423386,
-0.005457266699522734,
-0.06979179382324219,
-0.09984331578016281,
-0.05370790883898735,
0.0018799081444740295,
-0.009496532380580902,
0.021280018612742424,
2.4973776967627785e-34,
0.0774250254034996,
0.0445110909640789,
0.03384149819612503,
0.009015661664307117,
-0.013168711215257645,
-0.0018865579040721059,
-0.025271087884902954,
0.07968677580356598,
-0.004506566561758518,
-0.09763091802597046,
-0.027870090678334236,
0.023583419620990753,
0.056070342659950256,
-0.0022639702074229717,
-0.008788015693426132,
0.07373359799385071,
-0.07128582149744034,
-0.05081920325756073,
-0.0015287345740944147,
0.011354539543390274,
-0.09881514310836792,
0.07959087193012238,
-0.002735244110226631,
0.014626587741076946,
-0.021534612402319908,
0.042933542281389236,
-0.02769266627728939,
0.04056344926357269,
-0.060206763446331024,
0.031725168228149414,
0.0021646388340741396,
-0.13550499081611633,
-0.06612233072519302,
0.01409449614584446,
-0.05427103117108345,
0.12316801398992538,
0.0007005722145549953,
-0.00009591125854058191,
0.000015192863429547288,
0.08089956641197205,
-0.0019236428197473288,
-0.03432321548461914,
-0.062402427196502686,
0.026110222563147545,
-0.019181374460458755,
0.0383804515004158,
0.03415543958544731,
-0.03449280560016632,
0.01435606088489294,
0.08275087922811508,
0.020999830216169357,
0.11870401352643967,
-0.05121797323226929,
-0.012901835143566132,
-0.07347966730594635,
0.01049428153783083,
-0.04133981093764305,
-0.02687664143741131,
0.06388683617115021,
-0.05044016242027283,
-0.03921310976147652,
-0.09396211802959442,
0.010274630971252918,
0.07139254361391068,
-0.032784588634967804,
0.023588888347148895,
-0.08484417200088501,
0.028728995472192764,
-0.002956175711005926,
0.0839712843298912,
-0.005170674063265324,
-0.06642487645149231,
-0.0029419369529932737,
0.02187010459601879,
-0.006618529558181763,
-0.09529463946819305,
-0.009113945066928864,
-0.036009859293699265,
0.009697159752249718,
-0.028023961931467056,
0.10526961088180542,
0.016767388209700584,
0.11814694106578827,
0.04456141218543053,
0.02873173914849758,
0.015788989141583443,
0.0593867190182209,
-0.07165440171957016,
0.035375189036130905,
-0.019413284957408905,
-0.01909944787621498,
-0.017441509291529655,
-0.06488236784934998,
0.04425676912069321,
-0.02979641780257225,
-3.963022621178425e-8,
-0.02488606609404087,
0.020964249968528748,
0.022986100986599922,
-0.07224267721176147,
0.08403626084327698,
-0.01251782476902008,
0.01018347218632698,
0.02987927384674549,
-0.010989702306687832,
0.0013531938893720508,
0.03125681355595589,
-0.01305767334997654,
-0.09813210368156433,
-0.09643511474132538,
-0.023876285180449486,
0.03973010182380676,
-0.026604389771819115,
0.02316470630466938,
0.01337161660194397,
-0.0004618943785317242,
0.014981584623456001,
0.031068937852978706,
0.08126609772443771,
-0.0009665900724940002,
0.12227786332368851,
-0.08669371902942657,
-0.012755213305354118,
-0.013554122298955917,
0.13006733357906342,
0.024490216746926308,
-0.07454343140125275,
0.023930078372359276,
-0.007545034401118755,
0.021782726049423218,
0.033527523279190063,
0.02649378776550293,
-0.00037533053546212614,
-0.07260626554489136,
-0.052953120321035385,
0.02449491061270237,
-0.030048690736293793,
0.0729527622461319,
-0.03179364278912544,
0.020065724849700928,
0.026276947930455208,
0.14816522598266602,
0.039689209312200546,
-0.03289046138525009,
0.0531945563852787,
0.071861132979393,
0.05615771561861038,
0.03360966965556145,
0.03302418440580368,
0.08013589680194855,
0.06322908401489258,
-0.03843166306614876,
-0.01390758529305458,
-0.019674597308039665,
0.0006531922263093293,
-0.027134478092193604,
-0.1119500920176506,
0.06783746927976608,
0.08489640057086945,
-0.06933268159627914
] |
Budgeted Online Continual Learning by Adaptive Layer Freezing and Frequency-based Sampling | https://openreview.net/forum?id=dOAkHmsjRX | [
"Minhyuk Seo",
"Hyunseo Koh",
"Jonghyun Choi"
] | Spotlight | The majority of online continual learning (CL) advocates single-epoch training and imposes restrictions on the size of replay memory. However, single-epoch training would incur a different amount of computations per CL algorithm, and the additional storage cost to store logit or model in addition to replay memory is largely ignored in calculating the storage budget. Arguing different computational and storage budgets hinder fair comparison among CL algorithms in practice, we propose to use floating point operations (FLOPs) and total memory size in Byte as a metric for computational and memory budgets, respectively, to compare and develop CL algorithms in the same ‘total resource budget.’ To improve a CL method in a limited total budget, we propose adaptive layer freezing that does not update the layers for less informative batches to reduce computational costs with a negligible loss of accuracy. In addition, we propose a memory retrieval method that allows the model to learn the same amount of knowledge as using random retrieval in fewer iterations. Empirical validations on the CIFAR-10/100, CLEAR-10/100, and ImageNet-1K datasets demonstrate that the proposed approach outperforms the state-of-the-art methods within the same total budget. Furthermore, we validate its effectiveness in the Multi-modal Concept incremental Learning setup with multimodal large language models, such as LLaVA-1.5-7B. Code is available at https://github.com/snumprlab/budgeted-cl. | Continual Learning, Lifelong Learning, Efficient Training, Layer Freezing | null | 8,656 | 2410.15143 | [
-0.03263207897543907,
-0.02140667289495468,
-0.028281327337026596,
0.05617091804742813,
0.01953323930501938,
-0.011068711057305336,
-0.028941484168171883,
-0.006736107170581818,
0.026775497943162918,
0.00954030454158783,
-0.019517099484801292,
0.013513141311705112,
0.007669774349778891,
-0.04917867109179497,
-0.07152426242828369,
-0.06807450205087662,
0.02047399803996086,
0.05093150958418846,
-0.023727847263216972,
-0.049739986658096313,
0.02418765053153038,
-0.023608515039086342,
0.10564007610082626,
0.01575762778520584,
-0.003378856461495161,
-0.014480219222605228,
-0.053633593022823334,
-0.06237851083278656,
0.06491853296756744,
-0.007325871381908655,
0.01660218834877014,
0.07161341607570648,
-0.021587561815977097,
0.004224319942295551,
-0.004906350281089544,
0.018551882356405258,
-0.0472244918346405,
0.01335691288113594,
0.008468825370073318,
0.0007798151345923543,
-0.01821226440370083,
0.01997113600373268,
0.016440505161881447,
0.04149039089679718,
0.06520392000675201,
0.007819769904017448,
0.013373584486544132,
-0.010678226128220558,
0.005684204399585724,
0.08426104485988617,
0.029116185382008553,
0.011335873045027256,
-0.0444832481443882,
0.0753883495926857,
0.0015190131962299347,
0.017422474920749664,
0.040124792605638504,
0.08449157327413559,
-0.08953104168176651,
0.0631253719329834,
-0.07534252852201462,
-0.09450045228004456,
0.012488915584981441,
-0.008769143372774124,
-0.03563680127263069,
-0.019643692299723625,
0.06250002980232239,
0.053879573941230774,
0.05646033212542534,
-0.012040325440466404,
-0.04804530367255211,
0.09889519959688187,
-0.07573842257261276,
0.018054455518722534,
0.00463955570012331,
0.05272125452756882,
0.06596887856721878,
-0.008645706810057163,
0.03916531428694725,
-0.015376895666122437,
-0.009209667332470417,
-0.03675892949104309,
0.032993167638778687,
-0.08216375857591629,
0.12369029968976974,
-0.07702434062957764,
0.0810617208480835,
0.010399491526186466,
0.07582614570856094,
-0.05983489006757736,
0.008793148212134838,
0.055990226566791534,
0.039328351616859436,
-0.04744701087474823,
0.0001412850251654163,
0.06867587566375732,
0.003751839743927121,
0.0007212507189251482,
-0.062117718160152435,
0.04725927487015724,
0.022730836644768715,
0.0770874097943306,
0.012020392343401909,
0.01202320959419012,
0.012078317813575268,
-0.061461955308914185,
0.05715392902493477,
0.10640111565589905,
0.02360476925969124,
-0.08002247661352158,
0.06468220055103302,
0.03609636425971985,
-0.003730209544301033,
-0.008212369866669178,
0.09075509011745453,
-0.021786006167531013,
-0.07122734934091568,
-0.04997065290808678,
-0.025643331930041313,
0.119630828499794,
0.00010787116480059922,
0.045115407556295395,
-0.013414517976343632,
-0.008107248693704605,
-0.026011379435658455,
-0.05785970762372017,
-0.02563486434519291,
2.647625803100533e-33,
0.009087363258004189,
0.036467596888542175,
0.0103200888261199,
-0.07058791816234589,
0.012902677059173584,
-0.016087137162685394,
0.09574824571609497,
0.11319337040185928,
-0.038855571299791336,
-0.06252589821815491,
-0.017735736444592476,
-0.0007380909519270062,
-0.032194286584854126,
0.06714284420013428,
0.1181027963757515,
-0.07766799628734589,
-0.046763841062784195,
0.05868461728096008,
0.0442526750266552,
-0.06975021958351135,
0.06915678083896637,
-0.04979651793837547,
0.017615122720599174,
-0.10367666184902191,
0.02367481216788292,
-0.00862538255751133,
0.005083517637103796,
0.017035579308867455,
0.028445672243833542,
0.0129308570176363,
-0.042878951877355576,
-0.009970179758965969,
-0.008054141886532307,
-0.038934629410505295,
0.06034713611006737,
-0.07143037766218185,
-0.008715583011507988,
0.005292767658829689,
-0.015584311448037624,
-0.054786261171102524,
-0.03841105103492737,
0.008186312392354012,
-0.00968296080827713,
-0.04121346399188042,
-0.10815964639186859,
-0.18350063264369965,
0.03156661614775658,
0.001457418780773878,
-0.05606807395815849,
-0.0347452275454998,
0.041505858302116394,
0.009920691139996052,
-0.1297432780265808,
-0.01474528294056654,
-0.010166994296014309,
-0.022940024733543396,
0.04919932037591934,
0.015080864541232586,
0.01416807435452938,
0.07264944911003113,
0.05546637997031212,
0.004263094160705805,
-0.028343413025140762,
0.09566053003072739,
-0.018677635118365288,
0.08583210408687592,
0.07130695134401321,
0.040656205266714096,
-0.0077499765902757645,
-0.025684304535388947,
0.0353909432888031,
-0.047719720751047134,
0.05108025670051575,
-0.06985721737146378,
0.06199509650468826,
-0.03150596097111702,
0.03615385666489601,
-0.098940409719944,
-0.06637590378522873,
0.041106779128313065,
0.07807954400777817,
-0.028281543403863907,
-0.03281952813267708,
-0.049655068665742874,
-0.05500001087784767,
-0.031035901978611946,
0.062298282980918884,
0.01367910299450159,
-0.02076597698032856,
0.017514366656541824,
-0.060733236372470856,
-0.05296837165951729,
0.026383522897958755,
-0.09741360694169998,
-0.03919181227684021,
-2.37683727140618e-33,
-0.001579386182129383,
0.0030757682397961617,
-0.0006474286783486605,
0.07887956500053406,
-0.0016273664077743888,
-0.028455259278416634,
-0.030616924166679382,
-0.0055721597746014595,
-0.05280771106481552,
-0.13257646560668945,
-0.04884974658489227,
-0.017313459888100624,
-0.030572760850191116,
0.03392370790243149,
-0.028226662427186966,
0.014726320281624794,
-0.10738731175661087,
0.017274457961320877,
-0.015535660088062286,
0.048061221837997437,
-0.021008921787142754,
-0.004468114580959082,
-0.029061105102300644,
0.022518623620271683,
-0.013816801831126213,
0.04456597939133644,
-0.053752828389406204,
0.06293010711669922,
0.051947444677352905,
-0.013616010546684265,
-0.04981701448559761,
-0.10194388031959534,
-0.03269834443926811,
-0.07480432093143463,
-0.000879646569956094,
0.04483741521835327,
0.06908214837312698,
0.01986723393201828,
-0.042848389595746994,
0.109151192009449,
0.13139060139656067,
-0.03400859236717224,
0.0029628919437527657,
-0.010293879546225071,
0.04493529722094536,
-0.05380169302225113,
-0.14568176865577698,
0.030116695910692215,
0.033771611750125885,
0.02634843811392784,
0.027340449392795563,
0.038062505424022675,
-0.11721126735210419,
-0.02770650014281273,
0.0036154051776975393,
-0.049300745129585266,
-0.10697924345731735,
0.039562009274959564,
0.021265782415866852,
-0.05117947980761528,
-0.025389576330780983,
-0.09054894000291824,
-0.05431732162833214,
-0.02872915007174015,
0.034397754818201065,
-0.008861119858920574,
-0.004835224244743586,
0.03074401617050171,
-0.05918777361512184,
0.04092409089207649,
0.025084367021918297,
0.03449350595474243,
-0.009884499944746494,
-0.0361967459321022,
-0.019937042146921158,
0.04138614609837532,
0.0019724308513104916,
-0.028021913021802902,
-0.003410130040720105,
0.007073456421494484,
-0.07166381180286407,
0.06655482947826385,
-0.050675466656684875,
0.035725802183151245,
0.029803656041622162,
0.05606182664632797,
0.0008487231680192053,
0.042675767093896866,
-0.005118229892104864,
-0.006500654388219118,
-0.029571443796157837,
-0.002932671457529068,
0.0034972275607287884,
0.03306973725557327,
0.003077390603721142,
-4.6528576547189004e-8,
-0.027479110285639763,
0.02985406294465065,
0.09487223625183105,
0.053694333881139755,
0.12696608901023865,
-0.024096034467220306,
0.017132094129920006,
0.06365831196308136,
-0.006366307847201824,
-0.026325957849621773,
0.10001977533102036,
0.013496382161974907,
-0.021658357232809067,
-0.04634162038564682,
-0.018768122419714928,
0.08962816745042801,
0.025575656443834305,
-0.08247984200716019,
0.006074665579944849,
-0.06185627356171608,
-0.004872624296694994,
-0.014498216100037098,
0.037194207310676575,
-0.010972254909574986,
-0.024767916649580002,
-0.03805636242032051,
0.08256445825099945,
0.13346560299396515,
0.05569130554795265,
-0.031845416873693466,
0.000010900319466600195,
-0.0430934764444828,
0.08484343439340591,
-0.07494228333234787,
0.04298265278339386,
-0.0022328041959553957,
0.03250744566321373,
0.0006486145430244505,
0.019526662304997444,
-0.003576100105419755,
-0.02304995059967041,
0.026033833622932434,
0.04323570430278778,
0.008491299115121365,
0.07044429332017899,
-0.027611490339040756,
-0.05753450468182564,
-0.01783943735063076,
-0.02319006435573101,
0.011235007084906101,
-0.012187029235064983,
0.015121299773454666,
-0.024928070604801178,
0.12635895609855652,
0.08493363857269287,
-0.026692410930991173,
-0.007059665396809578,
-0.060304880142211914,
0.0056832339614629745,
0.10065185278654099,
0.016549671068787575,
-0.033021338284015656,
-0.08485405147075653,
0.04505576193332672
] |
Control-oriented Clustering of Visual Latent Representation | https://openreview.net/forum?id=pPQPQ7Yd58 | [
"Han Qi",
"Haocheng Yin",
"Heng Yang"
] | Spotlight | We initiate a study of the geometry of the visual representation space ---the information channel from the vision encoder to the action decoder--- in an image-based control pipeline learned from behavior cloning. Inspired by the phenomenon of *neural collapse* (NC) in image classification, we empirically demonstrate the prevalent emergence of a similar *law of clustering* in the visual representation space. Specifically,
- In discrete image-based control (e.g., Lunar Lander), the visual representations cluster according to the natural discrete action labels;
- In continuous image-based control (e.g., Planar Pushing and Block Stacking), the clustering emerges according to ``control-oriented'' classes that are based on (a) the relative pose between the object and the target in the input or (b) the relative pose of the object induced by expert actions in the output. Each of the classes corresponds to one relative pose orthant (REPO).
Beyond empirical observation, we show such a law of clustering can be leveraged as an algorithmic tool to improve test-time performance when training a policy with limited expert demonstrations. Particularly, we pretrain the vision encoder using NC as a regularization to encourage control-oriented clustering of the visual features. Surprisingly, such an NC-pretrained vision encoder, when finetuned end-to-end with the action decoder, boosts the test-time performance by 10% to 35%. Real-world vision-based planar pushing experiments confirmed the surprising advantage of control-oriented visual representation pretraining. | neural collapse, learning from demonstration, vision-based learning control | We discover a control-oriented law of clustering (similar to neural collapse) in the visual representation space of an end-to-end image-based control pipeline trained from behavior cloning. | 8,620 | 2410.05063 | [
-0.008008933626115322,
-0.10526681691408157,
0.022141244262456894,
-0.004181853495538235,
0.06358713656663895,
0.05499027669429779,
-0.014815196394920349,
-0.04176894575357437,
0.07523829489946365,
-0.0031728886533528566,
0.02664836496114731,
-0.024022841826081276,
0.03810107707977295,
-0.0010913218138739467,
-0.06569085270166397,
-0.018571248278021812,
0.071625255048275,
0.07933031767606735,
-0.06990525871515274,
-0.006456311792135239,
0.0250567477196455,
-0.040681805461645126,
0.09747378528118134,
0.024967115372419357,
-0.045045409351587296,
0.12199150025844574,
0.0031175492331385612,
-0.0791858583688736,
0.02082129940390587,
-0.07644342631101608,
0.009823795408010483,
0.020071621984243393,
-0.032213956117630005,
0.0057599302381277084,
-0.055402807891368866,
0.1789517104625702,
-0.011689108796417713,
0.01215267926454544,
-0.09250559657812119,
0.0025462512858211994,
-0.003530640620738268,
0.020902235060930252,
0.011149306781589985,
-0.015331897884607315,
0.03328356146812439,
0.1260867863893509,
0.027143804356455803,
-0.026479583233594894,
-0.0685063824057579,
-0.053262412548065186,
-0.02722068503499031,
-0.016322415322065353,
0.012756045907735825,
-0.059285204857587814,
0.0019422521581873298,
-0.025286510586738586,
0.08934386819601059,
-0.028029024600982666,
0.046373866498470306,
0.02020684815943241,
0.0692247748374939,
0.01828155294060707,
0.000004352204541646643,
-0.02392585575580597,
-0.012867392972111702,
0.06416168063879013,
0.04098783805966377,
-0.008499816060066223,
0.029698122292757034,
-0.0794408768415451,
0.0024074495304375887,
0.030809424817562103,
-0.039385851472616196,
0.0008949454058893025,
0.010867152363061905,
-0.006408965215086937,
0.009985722601413727,
-0.03198124095797539,
0.007009370252490044,
-0.08805444836616516,
0.05227388069033623,
-0.006120732054114342,
-0.06423336267471313,
0.025880763307213783,
0.07058104872703552,
0.004954177886247635,
-0.03340047225356102,
0.055478353053331375,
0.0660354420542717,
-0.010548876598477364,
-0.12175988405942917,
0.005809476599097252,
-0.0006621280335821211,
-0.012715798802673817,
0.05414516106247902,
-0.06097725033760071,
0.011884801089763641,
-0.08062659949064255,
0.047775376588106155,
0.04424053803086281,
-0.04239986836910248,
0.012004447169601917,
0.030560409650206566,
-0.01462655607610941,
0.03896438702940941,
-0.02905188500881195,
0.057526398450136185,
-0.03899421542882919,
0.07857498526573181,
-0.04919308423995972,
-0.03187451884150505,
-0.05740662291646004,
-0.0483422577381134,
0.07375941425561905,
-0.018270473927259445,
-0.015149575658142567,
0.010194920003414154,
-0.039876144379377365,
-0.03499668091535568,
-0.026604996994137764,
0.08926109969615936,
-0.07738077640533447,
0.03649058938026428,
-0.020473292097449303,
0.11426615715026855,
-0.031401801854372025,
-0.09372306615114212,
-3.276499485085118e-33,
0.04460901767015457,
-0.029718104749917984,
0.027389131486415863,
-0.020469417795538902,
0.09272380918264389,
-0.0141633665189147,
-0.00032521632965654135,
0.026743875816464424,
0.027350008487701416,
0.0035486838314682245,
0.013390012085437775,
-0.06622856855392456,
0.014055276289582253,
0.13457952439785004,
0.05922161042690277,
-0.032574839890003204,
-0.0010833732085302472,
0.03911992162466049,
-0.0984848365187645,
-0.054231226444244385,
-0.03885851055383682,
0.06322711706161499,
-0.037235669791698456,
-0.0787251889705658,
-0.06542760133743286,
-0.011483526788651943,
0.012423591688275337,
-0.050196461379528046,
-0.0025770931970328093,
-0.04043986275792122,
-0.042726919054985046,
0.011127573437988758,
-0.022970374673604965,
0.017051255330443382,
0.04951529577374458,
0.07167862355709076,
0.023286590352654457,
-0.04901089146733284,
0.010218647308647633,
-0.0030658442992717028,
0.003925445955246687,
0.027485646307468414,
-0.04921616241335869,
-0.08274949342012405,
0.0026948871091008186,
-0.0042969039641320705,
0.07477253675460815,
-0.0014892873587086797,
-0.09466801583766937,
-0.07682307809591293,
0.03672531619668007,
-0.003490690840408206,
0.04081474244594574,
-0.09231407195329666,
0.01969967409968376,
-0.040999241173267365,
0.008494938723742962,
0.03180961683392525,
0.038801148533821106,
0.04499608278274536,
-0.059068918228149414,
0.025799604132771492,
0.04403136670589447,
0.06048361957073212,
-0.044905275106430054,
0.013899717479944229,
0.009483047761023045,
0.05090421810746193,
0.08285117149353027,
0.028887087479233742,
-0.04404187574982643,
0.11600678414106369,
-0.1258867383003235,
-0.05202397331595421,
0.08188754320144653,
0.00699248630553484,
0.018358496949076653,
-0.03358620032668114,
-0.09060843288898468,
0.09232260286808014,
-0.04911436140537262,
-0.0803539976477623,
-0.005987728014588356,
-0.024643849581480026,
-0.024242857471108437,
0.012607707642018795,
0.06804296374320984,
-0.0366414450109005,
-0.0517762191593647,
-0.010227321647107601,
-0.06189529225230217,
-0.09007589519023895,
-0.002265959745272994,
0.0062661077827215195,
0.04983764514327049,
1.5325634331830271e-33,
-0.006751112174242735,
0.027220064774155617,
-0.018984539434313774,
0.02176726795732975,
-0.024454260244965553,
-0.0724170058965683,
-0.022684553638100624,
-0.004666759632527828,
-0.07487610727548599,
-0.00718417763710022,
-0.04436261206865311,
-0.019225087016820908,
-0.002408645348623395,
-0.019242143258452415,
-0.03554757684469223,
-0.027305899187922478,
0.0006449504871852696,
0.03476409241557121,
0.06474003195762634,
0.001502383267506957,
0.08355849236249924,
0.05452265590429306,
0.009350325912237167,
0.02906903438270092,
-0.05111837759613991,
0.09404291957616806,
0.015603035688400269,
0.09911619871854782,
0.06043107435107231,
0.03838290274143219,
-0.002947262255474925,
-0.09194143861532211,
-0.038597285747528076,
0.027744721621274948,
-0.06477510184049606,
0.06964117288589478,
0.015253391116857529,
0.012588872574269772,
-0.03821275383234024,
0.010713220573961735,
0.0023045982234179974,
-0.020979231223464012,
-0.036271195858716965,
0.023315947502851486,
0.014746755361557007,
-0.09902556985616684,
0.01633460260927677,
0.027611171826720238,
-0.08347993344068527,
0.05033426731824875,
-0.09222031384706497,
0.03624071925878525,
-0.020556490868330002,
-0.010427472181618214,
-0.020780282095074654,
0.056406259536743164,
0.01258818432688713,
0.03785603120923042,
0.053748104721307755,
-0.03128238022327423,
-0.00891013815999031,
-0.04837178438901901,
0.0174320749938488,
0.017905786633491516,
-0.03723590821027756,
-0.04565253481268883,
0.031034380197525024,
0.01987628825008869,
0.056507740169763565,
0.0429733544588089,
0.03363141417503357,
0.024722054600715637,
0.016766130924224854,
0.05670585483312607,
0.06698770821094513,
-0.021484043449163437,
0.04527828469872475,
0.007756291888654232,
0.04334711283445358,
-0.011123140342533588,
-0.025664368644356728,
-0.04054884985089302,
0.03287696838378906,
0.04090841859579086,
0.0354158841073513,
0.16684900224208832,
-0.06378963589668274,
0.07572077214717865,
0.05083524435758591,
-0.012959924526512623,
-0.016597861424088478,
-0.019019171595573425,
0.007661313284188509,
0.06730090081691742,
-0.08085464686155319,
-4.5454498831531964e-8,
-0.12211397290229797,
0.005012664943933487,
0.0925440639257431,
-0.04700950160622597,
-0.0099485469982028,
-0.021105101332068443,
0.027353163808584213,
0.02298910729587078,
-0.07095510512590408,
-0.061998359858989716,
0.11211149394512177,
0.026612112298607826,
-0.06191858649253845,
-0.047096069902181625,
0.01786063238978386,
0.030201327055692673,
0.010862966068089008,
0.00998767465353012,
-0.01802917569875717,
0.08574700355529785,
-0.012357505969703197,
-0.0841279998421669,
0.02932402864098549,
0.03343391790986061,
-0.0068001640029251575,
-0.07928864657878876,
-0.028013646602630615,
0.11502217501401901,
0.029549986124038696,
-0.01928066276013851,
0.03699351102113724,
0.05064411088824272,
0.06389953196048737,
0.015634609386324883,
0.04245070368051529,
0.025143222883343697,
-0.04968167468905449,
-0.0031035798601806164,
-0.012100176885724068,
0.010689924471080303,
-0.021907314658164978,
0.02576526068150997,
0.04195542633533478,
0.021543722599744797,
0.05722317472100258,
0.11681538820266724,
-0.005734073929488659,
-0.09304902702569962,
0.03699866309762001,
-0.07712214440107346,
-0.030966630205512047,
-0.02043001912534237,
-0.0453215129673481,
0.15224875509738922,
0.027130989357829094,
0.03886930271983147,
-0.02034679800271988,
-0.07564724236726761,
0.061127688735723495,
0.03454176336526871,
-0.002197669353336096,
0.0162999015301466,
-0.0965256616473198,
0.022171996533870697
] |
When do GFlowNets learn the right distribution? | https://openreview.net/forum?id=9GsgCUJtic | [
"Tiago Silva",
"Rodrigo Barreto Alves",
"Eliezer de Souza da Silva",
"Amauri H Souza",
"Vikas Garg",
"Samuel Kaski",
"Diego Mesquita"
] | Spotlight | Generative Flow Networks (GFlowNets) are an emerging class of sampling methods for distributions over discrete and compositional objects, e.g., graphs. In spite of their remarkable success in problems such as drug discovery and phylogenetic inference, the question of when and whether GFlowNets learn to sample from the target distribution remains underexplored. To tackle this issue, we first assess the extent to which a violation of the detailed balance of the underlying flow network might hamper the correctness of GFlowNet's sampling distribution. In particular, we demonstrate that the impact of an imbalanced edge on the model's accuracy is influenced by the total amount of flow passing through it and, as a consequence, is unevenly distributed across the network. We also argue that, depending on the parameterization, imbalance may be inevitable. In this regard, we consider the problem of sampling from distributions over graphs with GFlowNets parameterized by graph neural networks (GNNs) and show that the representation limits of GNNs delineate which distributions these GFlowNets can approximate. Lastly, we address these limitations by proposing a theoretically sound and computationally tractable metric for assessing GFlowNets, experimentally showing it is a better proxy for correctness than popular evaluation protocols. | GFlowNets | We examine the limitations and the stability of GFlowNets in face of balance violations, introducing a novel metric for assessing their accuracy. | 8,538 | null | [
-0.10838951170444489,
-0.05514270067214966,
0.055573102086782455,
-0.039989713579416275,
0.04498707130551338,
-0.04551217705011368,
-0.02841508574783802,
0.00002266372393933125,
0.06984935700893402,
-0.02140628732740879,
-0.03025316447019577,
-0.07818499952554703,
0.006996217183768749,
0.06460516154766083,
-0.05584566667675972,
-0.036144305020570755,
0.07735055685043335,
0.05302910879254341,
-0.06876324117183685,
0.018272677436470985,
0.030864745378494263,
0.09480420500040054,
0.006506793666630983,
-0.025156579911708832,
0.0019812092650681734,
-0.026558885350823402,
-0.03172933682799339,
-0.07282126694917679,
0.03574267402291298,
-0.029618846252560616,
0.09594739228487015,
0.1123272180557251,
-0.0667283833026886,
0.02975834719836712,
-0.04081875830888748,
-0.030077168717980385,
-0.06839334219694138,
0.04780077934265137,
0.0832362025976181,
0.0506419911980629,
0.06008047237992287,
0.05823579803109169,
0.023778339847922325,
0.03546113520860672,
0.05535045638680458,
-0.00154192466288805,
-0.02423006109893322,
-0.004843219183385372,
0.0030468094628304243,
0.07281085103750229,
-0.0057509467005729675,
-0.05124456807971001,
-0.05085230618715286,
0.10494203865528107,
0.0218161940574646,
-0.055623527616262436,
-0.009685981087386608,
-0.0532880537211895,
-0.012905657291412354,
0.038807518780231476,
0.008230920881032944,
-0.10171728581190109,
-0.05684330686926842,
0.00337077584117651,
-0.03758925944566727,
-0.020612841472029686,
0.030633606016635895,
0.07923396676778793,
0.04015818610787392,
0.008283191360533237,
0.07925520092248917,
0.10597150027751923,
-0.16382455825805664,
0.05885152891278267,
0.06955374032258987,
0.07750197499990463,
0.04568999633193016,
-0.0015104797203093767,
-0.01714026927947998,
-0.05856707692146301,
0.0608355775475502,
0.046886980533599854,
0.04728187620639801,
-0.024695124477148056,
0.018543999642133713,
0.005297315772622824,
-0.061235785484313965,
0.09710221737623215,
-0.0018181701889261603,
0.009725349955260754,
-0.07397226244211197,
0.06068750470876694,
-0.020249925553798676,
-0.057827696204185486,
0.03881518915295601,
0.06475631147623062,
0.014791200868785381,
-0.08291806280612946,
0.003710223361849785,
0.09614595770835876,
-0.048039160668849945,
-0.0028370830696076155,
0.0889083668589592,
-0.03888026624917984,
0.03442481905221939,
-0.006571998354047537,
0.004058889113366604,
0.03073335438966751,
0.07618843764066696,
-0.0608493946492672,
-0.011578314937651157,
0.05347331613302231,
0.03853503242135048,
0.04662042111158371,
-0.0005159168504178524,
-0.044040486216545105,
0.02827112190425396,
-0.03149620443582535,
-0.07756925374269485,
0.10028091818094254,
-0.12714903056621552,
0.04066052287817001,
-0.07346965372562408,
-0.04415770247578621,
0.021644605323672295,
0.022741511464118958,
-0.057929638773202896,
1.3273736392428004e-33,
0.009877569042146206,
-0.055180035531520844,
0.0765785202383995,
0.011776255443692207,
0.10749893635511398,
0.03773961961269379,
-0.017355531454086304,
-0.019904322922229767,
0.024056987836956978,
0.025948232039809227,
-0.11047349125146866,
-0.01121614035218954,
-0.09374402463436127,
0.07086838781833649,
-0.00041019803029485047,
0.03583713248372078,
-0.049167271703481674,
0.0572209432721138,
-0.007120735477656126,
-0.02452462911605835,
0.010223855264484882,
-0.1295143961906433,
0.0008290132973343134,
0.0005089680198580027,
-0.03668705374002457,
0.005883955396711826,
-0.04893232509493828,
0.05592482537031174,
0.017437821254134178,
-0.009950638748705387,
-0.09472822397947311,
-0.004620762541890144,
0.031986016780138016,
0.00890715979039669,
-0.012568637728691101,
-0.014068596996366978,
0.029738176614046097,
-0.041026629507541656,
0.011432358995079994,
-0.05133886635303497,
-0.03672976419329643,
0.04726849123835564,
0.051669031381607056,
-0.002639253856614232,
-0.03361433371901512,
-0.032752588391304016,
0.0025242269039154053,
-0.03690868616104126,
0.016252348199486732,
0.013813936151564121,
0.0441800132393837,
-0.024849044159054756,
0.04937585070729256,
0.009830614551901817,
-0.019888389855623245,
0.03838934749364853,
-0.01183366496115923,
0.051247961819171906,
0.04626314342021942,
0.09670499712228775,
0.01798298768699169,
0.037054166197776794,
-0.0877605676651001,
0.050726715475320816,
0.11560375988483429,
0.07969648391008377,
-0.08531727641820908,
0.020957008004188538,
0.03832219913601875,
0.028813490644097328,
-0.01868467964231968,
-0.0030003751162439585,
-0.12853971123695374,
-0.017989883199334145,
0.005540526006370783,
0.02123803086578846,
-0.006453316193073988,
0.014023257419466972,
-0.011881568469107151,
0.10000848770141602,
-0.14367389678955078,
0.020005585625767708,
-0.03624371439218521,
0.03708063066005707,
-0.07551145553588867,
-0.006488345097750425,
0.07209713011980057,
-0.045319005846977234,
0.031725361943244934,
0.022019896656274796,
-0.06687252968549728,
-0.05766898766160011,
-0.03757588937878609,
-0.00004150170207140036,
0.03562569618225098,
-1.783854921701054e-33,
-0.10325542092323303,
0.08941187709569931,
0.023119069635868073,
0.07784494012594223,
0.003134396392852068,
-0.03123614564538002,
0.004171699285507202,
0.013387266546487808,
0.0029115607030689716,
-0.026384560391306877,
-0.0234605111181736,
0.06985796242952347,
-0.026904035359621048,
-0.06628323346376419,
0.05404603108763695,
-0.05891941860318184,
-0.012900296598672867,
0.011142715811729431,
0.003473399206995964,
-0.019759533926844597,
-0.01921868324279785,
0.03807733207941055,
-0.12330812960863113,
-0.034943096339702606,
-0.006618305575102568,
0.04434613883495331,
-0.037716977298259735,
0.031227443367242813,
0.046336181461811066,
0.006161577999591827,
-0.05330921709537506,
-0.01848256215453148,
-0.09906381368637085,
-0.036725036799907684,
0.015556314028799534,
0.05287352576851845,
0.05661037191748619,
0.058191731572151184,
0.01685837283730507,
0.03168494254350662,
-0.0007057985640130937,
0.009380972944200039,
-0.14084060490131378,
0.053911298513412476,
-0.00603119982406497,
0.012382281012833118,
0.004597839433699846,
-0.0018678881460800767,
0.01548693049699068,
-0.005830747075378895,
-0.04500742256641388,
0.10961362719535828,
-0.028462011367082596,
0.04383845627307892,
0.016257859766483307,
-0.05606423690915108,
0.04956560581922531,
0.013941390439867973,
0.0772199034690857,
-0.04678349196910858,
-0.13644321262836456,
-0.056754328310489655,
-0.0881832018494606,
-0.07089423388242722,
-0.04472990706562996,
-0.020157989114522934,
0.026390517130494118,
-0.0023677616845816374,
0.039877261966466904,
0.0730551928281784,
0.026275957003235817,
0.03202735632658005,
0.06582463532686234,
-0.01348194945603609,
0.03951476514339447,
-0.041807908564805984,
-0.0010592446196824312,
-0.02379307895898819,
0.01121467724442482,
-0.07759416103363037,
0.011306495405733585,
-0.004344453569501638,
0.003110229503363371,
0.020170656964182854,
0.09647233039140701,
-0.07108725607395172,
0.012361085042357445,
-0.03991169109940529,
0.07936375588178635,
0.053312260657548904,
-0.017260422930121422,
-0.028351204469799995,
-0.05249303579330444,
0.0033338575158268213,
-0.06328652054071426,
-4.387882412970612e-8,
-0.07271620631217957,
0.06787419319152832,
0.018940817564725876,
-0.0009430648060515523,
0.052964817732572556,
0.08540108054876328,
0.0001667275937506929,
0.11108699440956116,
0.003970228601247072,
0.02495328150689602,
0.09378286451101303,
0.011961973272264004,
-0.015390822663903236,
-0.05531520023941994,
-0.0011086675804108381,
0.013234218582510948,
-0.022042345255613327,
0.037814393639564514,
-0.010968003422021866,
-0.00526110315695405,
-0.04780995473265648,
-0.01444877777248621,
-0.021374322474002838,
0.010275653563439846,
0.033376194536685944,
-0.12484122067689896,
0.009350976906716824,
-0.012354138307273388,
0.021664373576641083,
0.028146425262093544,
-0.01705760322511196,
0.01629210263490677,
-0.011739399284124374,
0.042822930961847305,
-0.00827003177255392,
0.025654882192611694,
-0.061636198312044144,
-0.02653731033205986,
-0.0225428007543087,
-0.024678699672222137,
-0.06144584342837334,
0.07173272222280502,
-0.045512303709983826,
-0.038917019963264465,
-0.005772753618657589,
-0.012000979855656624,
0.031180864199995995,
0.047442272305488586,
0.019398821517825127,
-0.00816824659705162,
-0.00008262661867775023,
-0.021073851734399796,
-0.006289063021540642,
0.021944932639598846,
0.0019271274795755744,
-0.02356966771185398,
-0.037461988627910614,
-0.10172491520643234,
0.005267367232590914,
-0.02344534546136856,
-0.0016509088454768062,
0.029184341430664062,
-0.03814122453331947,
-0.06980448961257935
] |
R2-Guard: Robust Reasoning Enabled LLM Guardrail via Knowledge-Enhanced Logical Reasoning | https://openreview.net/forum?id=CkgKSqZbuC | [
"Mintong Kang",
"Bo Li"
] | Spotlight | As large language models (LLMs) become increasingly prevalent across various applications, it is critical to establish safety guardrails to moderate input/output content of LLMs and ensure compliance with safety policies. Existing guardrail models, such as OpenAI Mod and LlamaGuard, treat various safety categories (e.g., self-harm, self-harm/instructions) independently and fail to explicitly capture the intercorrelations among them. This has led to limitations such as ineffectiveness due to inadequate training on long-tail data from correlated safety categories, susceptibility to jailbreaking attacks, and inflexibility regarding new safety categories.
To address these limitations, we propose $R^2$-Guard, a robust reasoning enabled LLM guardrail via knowledge-enhanced logical reasoning. Specifically, $R^2$-Guard comprises two parts: data-driven guardrail models and reasoning components. The data-driven guardrail models provide unsafety probabilities of moderated content on different safety categories.
We then encode safety knowledge among different categories as first-order logical rules and embed them into a probabilistic graphic model (PGM) based reasoning component. The unsafety probabilities of different categories from data-driven guardrail models are sent to the reasoning component for final inference. We employ two types of PGMs: Markov logic networks (MLNs) and probabilistic circuits (PCs), and optimize PCs to achieve precision-efficiency balance via improved graph structure. We also propose different methods to optimize the weights of knowledge. To further perform stress tests for guardrail models, we employ a pairwise construction method to construct a new safety benchmark TwinSafety, which features principled categories and presents new challenges for moderation. We show that $R^2$-Guard is effective even given unrepresentative categories or challenging jailbreaking prompts. We demonstrate the effectiveness of $R^2$-Guard by comparisons with eight strong guardrail models on six standard moderation datasets, and demonstrate the robustness of $R^2$-Guard against four SOTA jailbreaking attacks. $R^2$-Guard significantly surpasses SOTA method LlamaGuard by 12.6% on standard moderation datasets and by 59.9% against jailbreaking attacks.
We further reveal that $R^2$-Guard can effectively adapt to safety category updates by simply editing the PGM reasoning graph. | LLM guardrail model, content moderation | We propose an effective, robust, and flexible LLM guardrail model via knowledge-enhanced logical reasoning. | 8,497 | null | [
0.03611089289188385,
-0.03823345899581909,
-0.03987131267786026,
0.022367477416992188,
0.12535996735095978,
0.030362876132130623,
0.06797149032354355,
0.017549889162182808,
-0.03615593537688255,
0.03087685815989971,
0.004626305308192968,
-0.028071805834770203,
0.10238524526357651,
0.08651142567396164,
0.09308561682701111,
0.07146663963794708,
0.1301543116569519,
0.07174668461084366,
-0.12089379131793976,
-0.02411317080259323,
0.038676097989082336,
0.01004913542419672,
0.011434678919613361,
0.00037305665318854153,
-0.09588886797428131,
-0.008085227571427822,
0.01912662573158741,
0.017779259011149406,
-0.005207723472267389,
-0.04366111382842064,
0.02095668390393257,
-0.08375663310289383,
-0.007861092686653137,
0.0697220116853714,
-0.025522973388433456,
-0.009322970174252987,
-0.008635981939733028,
0.018809054046869278,
-0.021928343921899796,
-0.048233821988105774,
-0.1159978061914444,
-0.00003166485839756206,
-0.00603181216865778,
0.01780950091779232,
0.02977382019162178,
-0.006950877141207457,
-0.03440959379076958,
0.003655781736597419,
-0.009917243383824825,
-0.016746118664741516,
-0.05756375938653946,
-0.0006196741596795619,
0.010000615380704403,
-0.023563219234347343,
0.07655877619981766,
-0.04355516657233238,
-0.050774671137332916,
-0.012758417055010796,
-0.04745526611804962,
-0.0007839825702831149,
0.0006759728421457112,
-0.06102560833096504,
-0.004401565995067358,
0.004255134146660566,
-0.06991138309240341,
-0.03332589194178581,
-0.04962579905986786,
0.0786716490983963,
0.07805239409208298,
0.05373223125934601,
-0.04442516714334488,
0.03886136785149574,
-0.03123490698635578,
-0.02493133768439293,
-0.0749371200799942,
0.02282494306564331,
0.0027098385617136955,
-0.021792128682136536,
0.004065078217536211,
-0.10172412544488907,
-0.0085060503333807,
0.003505377098917961,
0.03975188732147217,
0.07821553945541382,
0.0716576799750328,
0.052190493792295456,
-0.010475106537342072,
0.07993005216121674,
0.05018165335059166,
0.04630260542035103,
0.0215306393802166,
0.04074283689260483,
0.1351933479309082,
0.06643593311309814,
0.09912405908107758,
0.013137005269527435,
-0.015964888036251068,
-0.1255517601966858,
-0.07396630197763443,
0.08744718879461288,
-0.006823397241532803,
0.07009750604629517,
-0.028535369783639908,
0.004670200869441032,
0.06929880380630493,
-0.05771893262863159,
-0.012549424543976784,
0.01794649288058281,
-0.02932480163872242,
-0.033330902457237244,
0.06938762217760086,
0.10247141122817993,
0.04730315878987312,
-0.03331267833709717,
-0.0423862561583519,
-0.009998041205108166,
-0.10308472812175751,
-0.019239237532019615,
0.06731683760881424,
0.020614314824342728,
0.021644074469804764,
0.047574058175086975,
0.06797687709331512,
0.0468173548579216,
0.08137304335832596,
-0.06618089973926544,
-0.04755672067403793,
3.015189294949231e-33,
0.06741165369749069,
0.038175079971551895,
-0.02359599806368351,
-0.03825409337878227,
0.03400896489620209,
-0.016286617144942284,
-0.05410135164856911,
0.01552263367921114,
-0.03472556918859482,
0.06523178517818451,
-0.04578845575451851,
-0.02034507505595684,
-0.055299062281847,
0.027851887047290802,
0.11350800096988678,
0.06801693886518478,
-0.07217494398355484,
-0.01883961446583271,
-0.046381495893001556,
-0.04698869213461876,
0.02034248784184456,
-0.050624627619981766,
-0.0008197049610316753,
0.054927580058574677,
-0.012228433042764664,
0.06124771386384964,
0.0868450328707695,
-0.010268778540194035,
0.054009296000003815,
0.0446467250585556,
-0.07388170063495636,
-0.030137652531266212,
0.006151469890028238,
0.028406765311956406,
0.07048968225717545,
-0.0582597553730011,
-0.07315441966056824,
-0.07202430814504623,
0.009666206315159798,
-0.011197184212505817,
-0.029915405437350273,
-0.003144432557746768,
0.01724405772984028,
0.01617266982793808,
-0.033365391194820404,
-0.037074413150548935,
-0.020642174407839775,
-0.003673933446407318,
-0.0570998415350914,
-0.04797011613845825,
0.04530173912644386,
0.04378509148955345,
0.030805552378296852,
-0.04952476918697357,
-0.11846234649419785,
-0.07595592737197876,
-0.036914438009262085,
0.059669990092515945,
0.010047058574855328,
0.09066072851419449,
-0.09443448483943939,
0.059226009994745255,
0.033173609524965286,
0.032942648977041245,
0.007886354811489582,
0.022599423304200172,
-0.07762938737869263,
0.028258798643946648,
0.0724741742014885,
-0.015557964332401752,
-0.06813574582338333,
-0.007710774429142475,
0.0742821991443634,
-0.02534753456711769,
-0.06270138174295425,
-0.01424577459692955,
-0.012280822731554508,
-0.05587800219655037,
0.05158313736319542,
-0.03404808044433594,
0.029862206429243088,
-0.008309331722557545,
-0.007868096232414246,
0.0364573709666729,
0.005671185906976461,
-0.05511496588587761,
0.021200116723775864,
-0.06933899223804474,
-0.06659527868032455,
0.020233333110809326,
-0.030049193650484085,
-0.05432488024234772,
0.04161342978477478,
0.014612635597586632,
0.0048424797132611275,
-3.354598227727797e-33,
-0.02443568781018257,
-0.02606569044291973,
-0.03180236369371414,
0.05893772467970848,
-0.04494111239910126,
-0.09549679607152939,
-0.02835177257657051,
0.006310933269560337,
0.01773194968700409,
-0.035034678876399994,
-0.08729773759841919,
-0.03390313684940338,
0.07282420247793198,
0.0016131416196003556,
0.027010319754481316,
-0.09164564311504364,
-0.058246299624443054,
0.03633707016706467,
-0.09573289006948471,
0.06928442418575287,
-0.015141395851969719,
0.049167782068252563,
-0.13579989969730377,
0.05309907719492912,
-0.0016077874461188912,
0.07140090316534042,
-0.013533170334994793,
0.001723546301946044,
0.016810109838843346,
0.06207473576068878,
0.0031207443680614233,
-0.02118041180074215,
0.01888030581176281,
-0.01755928434431553,
-0.0822351798415184,
-0.051867611706256866,
0.03794166073203087,
-0.0502757802605629,
-0.05728476122021675,
0.0011771282879635692,
0.04060475155711174,
0.0016297108959406614,
-0.0678156241774559,
-0.06043252721428871,
-0.029382431879639626,
0.005523968487977982,
-0.006606339011341333,
0.01153956726193428,
0.012620871886610985,
-0.04123115539550781,
-0.031846072524785995,
-0.05218726024031639,
0.012070772238075733,
-0.004961416590958834,
0.04893277958035469,
-0.08930443972349167,
-0.0015670418506488204,
-0.026402639225125313,
0.013523286208510399,
0.028849419206380844,
-0.041913773864507675,
-0.029533226042985916,
0.03397655114531517,
0.012642443180084229,
0.031439054757356644,
-0.03283146023750305,
-0.07941663265228271,
0.0004027254180982709,
0.011733702383935452,
-0.042893532663583755,
0.033144522458314896,
0.03550253063440323,
-0.024707099422812462,
-0.02687428519129753,
0.02667270042002201,
-0.0191253200173378,
-0.013851463794708252,
-0.04618648439645767,
-0.038287825882434845,
0.023228701204061508,
0.05218232423067093,
-0.09887497872114182,
0.09057027846574783,
0.08286163210868835,
0.05065607279539108,
-0.0022293729707598686,
0.04259498044848442,
0.04798278957605362,
0.05654031038284302,
0.05753454938530922,
-0.11311186850070953,
0.041902415454387665,
-0.05203884467482567,
0.0921022966504097,
-0.08342079818248749,
-4.251166885183011e-8,
-0.05924602597951889,
-0.012964518740773201,
-0.03213949501514435,
0.0349680632352829,
0.07273232936859131,
0.03872998058795929,
-0.05896599218249321,
-0.02909703552722931,
-0.048627596348524094,
-0.04414677247405052,
0.09946371614933014,
0.022742191329598427,
-0.06702589243650436,
-0.06909889727830887,
-0.050096262246370316,
0.0803508311510086,
0.04563051462173462,
0.00803366955369711,
-0.023263605311512947,
0.02279089391231537,
-0.003492280375212431,
-0.021925613284111023,
-0.05232309177517891,
0.04001801460981369,
0.02571619115769863,
-0.037745386362075806,
-0.07289179414510727,
0.06823841482400894,
0.07097667455673218,
0.08609452098608017,
-0.01470174454152584,
-0.08514154702425003,
0.08511956036090851,
0.050096187740564346,
0.03122691437602043,
0.058258481323719025,
-0.008760126307606697,
-0.00843549333512783,
0.062298309057950974,
-0.04792029410600662,
0.059930186718702316,
0.04407929256558418,
-0.019663607701659203,
0.05763977766036987,
0.016925102099776268,
0.006259530317038298,
-0.04931515082716942,
-0.12131165713071823,
-0.014249877072870731,
0.0006441540317609906,
-0.011602361686527729,
-0.018965985625982285,
-0.009522552601993084,
0.09808646887540817,
0.04492872580885887,
0.04718076437711716,
-0.014219173230230808,
-0.02539512887597084,
-0.0016688217874616385,
0.03619766980409622,
0.050421927124261856,
-0.007413583807647228,
0.06381905823945999,
-0.04431500658392906
] |
NetFormer: An interpretable model for recovering dynamical connectivity in neuronal population dynamics | https://openreview.net/forum?id=bcTjW5kS4W | [
"Ziyu Lu",
"Wuwei Zhang",
"Trung Le",
"Hao Wang",
"Uygar Sümbül",
"Eric Todd SheaBrown",
"Lu Mi"
] | Spotlight | Neuronal dynamics are highly nonlinear and nonstationary. Traditional methods for extracting the underlying network structure from neuronal activity recordings mainly concentrate on modeling static connectivity, without accounting for key nonstationary aspects of biological neural systems, such as ongoing synaptic plasticity and neuronal modulation. To bridge this gap, we introduce the NetFormer model, an interpretable approach applicable to such systems. In NetFormer, the activity of each neuron across a series of historical time steps is defined as a token. These tokens are then linearly mapped through a query and key mechanism to generate a state- (and hence time-) dependent attention matrix that directly encodes nonstationary connectivity structures. We analyze our formulation from the perspective of nonstationary and nonlinear networked dynamical systems, and show both via an analytical expansion and targeted simulations how it can approximate the underlying ground truth. Next, we demonstrate NetFormer's ability to model a key feature of biological networks, spike-timing-dependent plasticity, whereby connection strengths continually change in response to local activity patterns. We further demonstrate that NetFormer can capture task-induced connectivity patterns on activity generated by task-trained recurrent neural networks. Thus informed, we apply NetFormer to a multi-modal dataset of real neural recordings, which contains neural activity, cell type, and behavioral state information. We show that the NetFormer effectively predicts neural dynamics and identifies cell-type specific, state-dependent dynamic connectivity that matches patterns measured in separate ground-truth physiology experiments, demonstrating its ability to help decode complex neural interactions based on population activity observations alone. | neuronal dynamics, dynamical connectivity, interpretability, attention mechanism, transformer | We proposed NetFormer, an interpretable transformer-inspired model to capture neuronal population dynamics and uncover dynamical inter-neuron connectivity. | 8,471 | null | [
-0.0517653152346611,
-0.15404534339904785,
0.022746769711375237,
0.0680929571390152,
-0.00033458825782872736,
0.06632154434919357,
-0.018731193616986275,
-0.04412568360567093,
0.09401916712522507,
-0.05438585206866264,
-0.022306639701128006,
-0.02388979308307171,
-0.010878737084567547,
0.0749322846531868,
-0.039708077907562256,
0.011881587095558643,
0.008623301982879639,
0.01939137652516365,
-0.032480183988809586,
0.0066428580321371555,
0.06691308319568634,
-0.002417883835732937,
-0.03657645359635353,
0.033668793737888336,
-0.003758571809157729,
0.0318216010928154,
-0.0019470845581963658,
-0.011791585944592953,
-0.00897891167551279,
0.0010189516469836235,
0.026318904012441635,
-0.030049825087189674,
-0.015819067135453224,
0.03273569792509079,
-0.05571143701672554,
0.023373695090413094,
-0.06384944915771484,
0.024137502536177635,
-0.013704778626561165,
-0.026704492047429085,
0.10785483568906784,
0.018382200971245766,
-0.025480179116129875,
0.03418973088264465,
0.04308885708451271,
-0.0016629173187538981,
0.05350600928068161,
-0.13484185934066772,
-0.05979098752140999,
-0.009097449481487274,
-0.04623734951019287,
0.03199945017695427,
0.019212618470191956,
-0.007330751977860928,
0.09485551714897156,
0.0606706403195858,
0.0005268378299660981,
-0.004224379081279039,
-0.052273910492658615,
-0.012185816653072834,
-0.027341220527887344,
0.012502286583185196,
0.02508372999727726,
-0.07681749016046524,
0.035575058311223984,
0.08980070054531097,
0.004315538797527552,
0.08226452022790909,
0.012928600423038006,
0.003512639319524169,
-0.06467590481042862,
0.078359454870224,
-0.0698745846748352,
-0.07683774828910828,
0.011511254124343395,
0.018446987494826317,
0.05789719521999359,
0.058466967195272446,
0.07179859280586243,
-0.07918424904346466,
0.04627781733870506,
0.047156624495983124,
0.04171045497059822,
-0.004767189733684063,
0.09747279435396194,
0.04578705504536629,
0.06735112518072128,
0.05930086970329285,
0.02438104897737503,
0.010157782584428787,
-0.11975692957639694,
-0.02928566187620163,
-0.019277522340416908,
-0.02351551130414009,
0.058876004070043564,
0.0107260188087821,
0.013373687863349915,
-0.05088049918413162,
0.05518455430865288,
0.05137959122657776,
-0.016089066863059998,
0.02148696407675743,
0.018404116854071617,
0.018341096118092537,
0.03667181357741356,
0.011001169681549072,
0.048113495111465454,
0.0572400763630867,
0.03276393562555313,
0.031970612704753876,
0.009597301483154297,
-0.013336473144590855,
-0.007210466545075178,
0.08754708617925644,
0.04330457001924515,
-0.018294574692845345,
0.03399749845266342,
-0.02100612036883831,
0.07743078470230103,
0.04591279476881027,
0.039634596556425095,
-0.04135320335626602,
-0.05801231786608696,
-0.04666367173194885,
0.010953010059893131,
-0.02896958775818348,
-0.02998924069106579,
1.9872356482024645e-33,
-0.03083709254860878,
-0.01316136121749878,
0.002137698233127594,
0.01678643375635147,
0.0571712888777256,
-0.08789131045341492,
-0.015699680894613266,
-0.0157320573925972,
0.046927183866500854,
0.08581644296646118,
-0.08732486516237259,
0.03893535956740379,
-0.007426883559674025,
0.09067223221063614,
0.0021288974676281214,
0.019956419244408607,
-0.032706569880247116,
-0.013162762857973576,
0.028419682756066322,
-0.12516281008720398,
-0.02245352230966091,
-0.019491005688905716,
-0.019585901871323586,
-0.04273959621787071,
0.03237901255488396,
-0.10395052284002304,
-0.003528678324073553,
0.004709272179752588,
-0.013089420273900032,
-0.033245399594306946,
-0.07276836037635803,
0.02797401323914528,
-0.02968156524002552,
-0.04529036581516266,
0.06277664750814438,
-0.03510156273841858,
0.058123551309108734,
-0.06427096575498581,
0.03097517043352127,
-0.015598112717270851,
0.04252389073371887,
0.04041200503706932,
-0.025394875556230545,
-0.01736069656908512,
-0.09017887711524963,
-0.07314807921648026,
0.016309505328536034,
0.015331929549574852,
-0.03946525231003761,
-0.06474002450704575,
-0.025977542623877525,
0.00039958825800567865,
0.015708887949585915,
-0.10617952048778534,
-0.04789760708808899,
-0.0055174813605844975,
-0.03049369901418686,
0.058517396450042725,
0.03336649760603905,
0.11952389776706696,
0.04175347089767456,
-0.02479594200849533,
0.04417850077152252,
0.006065238267183304,
0.07693041115999222,
0.08779477328062057,
-0.0944083034992218,
0.05815558880567551,
0.09882240742444992,
-0.0001931442238856107,
0.035733167082071304,
0.04115460440516472,
-0.004544200841337442,
-0.10433376580476761,
0.023650866001844406,
-0.004152405075728893,
0.008776852861046791,
-0.03735891357064247,
-0.09282676130533218,
0.01649267040193081,
0.03809529170393944,
-0.09626275300979614,
-0.09240449219942093,
0.08684957772493362,
-0.042617317289114,
-0.01720479689538479,
0.06780409067869186,
-0.09799936413764954,
-0.10420612245798111,
-0.005928108002990484,
0.015577870421111584,
-0.029509644955396652,
0.043287042528390884,
-0.019838903099298477,
-0.030586885288357735,
-2.2579567726426375e-33,
-0.04165711626410484,
0.06325531005859375,
-0.025727469474077225,
-0.05341561511158943,
-0.004344469401985407,
0.00885468628257513,
-0.013517586514353752,
0.05271168425679207,
-0.05699454993009567,
-0.01828673668205738,
-0.02070779539644718,
-0.06386420875787735,
0.08097688108682632,
0.02059350349009037,
0.024994613602757454,
-0.07039197534322739,
-0.046095941215753555,
0.019244592636823654,
0.03855743259191513,
0.01734781824052334,
-0.025662047788500786,
0.04867451265454292,
-0.1128063052892685,
-0.049786556512117386,
0.017079252749681473,
0.08484286814928055,
-0.09755361080169678,
0.12448360025882721,
-0.028973177075386047,
-0.027532055974006653,
-0.1086319088935852,
0.019697802141308784,
-0.014441282488405704,
-0.062455710023641586,
-0.048224762082099915,
0.12920568883419037,
0.013326871208846569,
-0.05172222852706909,
-0.025097768753767014,
-0.0660548210144043,
0.05595488101243973,
-0.0533776693046093,
0.014115624129772186,
0.06628209352493286,
0.052591145038604736,
-0.007008291315287352,
-0.04800795763731003,
0.04203109070658684,
-0.08169257640838623,
0.06866846978664398,
0.0627359002828598,
0.06220580264925957,
-0.08184610307216644,
-0.022699950262904167,
-0.09998626261949539,
0.06651389598846436,
0.005333548411726952,
-0.09234334528446198,
0.038318827748298645,
-0.04056689515709877,
-0.07176568359136581,
-0.06212776526808739,
0.020525893196463585,
-0.03540770336985588,
0.020228488370776176,
-0.029636381193995476,
0.041800156235694885,
-0.05380002781748772,
0.06334032118320465,
0.012313815765082836,
0.07609226554632187,
0.04557327553629875,
0.02596225030720234,
0.02238616533577442,
0.059809643775224686,
-0.057398341596126556,
-0.05555885657668114,
-0.009123032912611961,
-0.004109934438019991,
-0.035711389034986496,
-0.0034907718654721975,
0.036506619304418564,
-0.022727442905306816,
-0.021415166556835175,
-0.013703154399991035,
0.03563809022307396,
0.07146011292934418,
0.010803443379700184,
0.019986826926469803,
-0.04436495155096054,
0.030448472127318382,
-0.0028896364383399487,
-0.0746677815914154,
0.09299162030220032,
-0.0313592366874218,
-4.5506030943442966e-8,
-0.08324937522411346,
0.09877727925777435,
-0.041407398879528046,
0.024591991677880287,
0.04428740218281746,
-0.05458776280283928,
0.11234766989946365,
-0.02250073477625847,
-0.015955669805407524,
0.031538818031549454,
0.1106458380818367,
-0.041107166558504105,
0.02506706863641739,
-0.06104990839958191,
-0.0009749457822181284,
0.0701759085059166,
0.05418761074542999,
-0.00190664769615978,
-0.015299067832529545,
-0.058827631175518036,
-0.03260692581534386,
0.041871704161167145,
-0.017430009320378304,
0.055355530232191086,
0.03259503096342087,
-0.033469367772340775,
-0.029210209846496582,
0.09720246493816376,
0.020013583824038506,
-0.038287509232759476,
0.05524125695228577,
0.0640982910990715,
0.032910726964473724,
0.02906777337193489,
-0.10001889616250992,
0.03301582485437393,
-0.027541672810912132,
-0.0207197405397892,
-0.01760600320994854,
0.0022234043572098017,
-0.023430241271853447,
0.05151347815990448,
-0.00042259469046257436,
0.034680213779211044,
0.0015539205633103848,
-0.03479129448533058,
-0.002473231637850404,
-0.031789086759090424,
0.03289686515927315,
-0.08686068654060364,
0.03491298109292984,
0.012687277048826218,
-0.03795485571026802,
0.0258976761251688,
-0.009174005128443241,
-0.019526565447449684,
0.02116565965116024,
0.008951567113399506,
0.0015096230199560523,
0.04915722832083702,
-0.00250772456638515,
0.05881049856543541,
-0.10733305662870407,
-0.05591408908367157
] |
Revisiting Random Walks for Learning on Graphs | https://openreview.net/forum?id=SG1R2H3fa1 | [
"Jinwoo Kim",
"Olga Zaghen",
"Ayhan Suleymanzade",
"Youngmin Ryou",
"Seunghoon Hong"
] | Spotlight | We revisit a simple model class for machine learning on graphs, where a random walk on a graph produces a machine-readable record, and this record is processed by a deep neural network to directly make vertex-level or graph-level predictions. We call these stochastic machines random walk neural networks (RWNNs), and through principled analysis, show that we can design them to be isomorphism invariant while capable of universal approximation of graph functions in probability. A useful finding is that almost any kind of record of random walks guarantees probabilistic invariance as long as the vertices are anonymized. This enables us, for example, to record random walks in plain text and adopt a language model to read these text records to solve graph tasks. We further establish a parallelism to message passing neural networks using tools from Markov chain theory, and show that over-smoothing in message passing is alleviated by construction in RWNNs, while over-squashing manifests as probabilistic under-reaching. We empirically demonstrate RWNNs on a range of problems, verifying our theoretical analysis and demonstrating the use of language models for separating strongly regular graphs where 3-WL test fails, and transductive classification on arXiv citation network. Code is available at https://github.com/jw9730/random-walk. | Graph machine learning, random walk, invariance, universal approximation, markov chain | From the perspectives of invariance and universality, we revisit a simple idea where a random walk on a graph produces a machine-readable record which is processed by a deep neural network to directly make vertex-level or graph-level predictions. | 8,462 | 2407.01214 | [
-0.15121878683567047,
-0.0861879512667656,
0.048640962690114975,
0.07257618010044098,
0.06578496098518372,
-0.029270658269524574,
0.014474469237029552,
-0.036612849682569504,
-0.0061267465353012085,
-0.005434409715235233,
-0.009624979458749294,
0.001249883440323174,
0.025125766173005104,
0.06702523678541183,
-0.018209857866168022,
0.02270318754017353,
0.056618932634592056,
0.10299491137266159,
-0.06365242600440979,
-0.01215224526822567,
0.019327979534864426,
0.059192460030317307,
0.06573451310396194,
0.016551870852708817,
0.062150705605745316,
-0.04531634598970413,
0.013441148214042187,
-0.005631406791508198,
0.01114027202129364,
0.01922387257218361,
0.09798403084278107,
-0.04693719744682312,
-0.03397928178310394,
-0.00008231905667344108,
0.006320149637758732,
0.09175070375204086,
-0.05153285712003708,
-0.005459161940962076,
0.06359612196683884,
0.016785435378551483,
0.02982572838664055,
0.01685592718422413,
0.03256085515022278,
0.06872343271970749,
0.05613624304533005,
-0.014650220051407814,
0.014823483303189278,
-0.035408660769462585,
-0.049345675855875015,
0.05377636477351189,
-0.05827941372990608,
-0.06290000677108765,
-0.03987915813922882,
0.031524207442998886,
-0.002151233609765768,
-0.034243129193782806,
-0.0036793197505176067,
0.01950879581272602,
-0.05852600932121277,
-0.03746652975678444,
0.01932489313185215,
-0.10058949887752533,
-0.04873717948794365,
-0.026264721527695656,
-0.0512252114713192,
-0.007103827781975269,
-0.008828011341392994,
0.10394049435853958,
0.00955111812800169,
0.08416314423084259,
0.04953935369849205,
0.07581251114606857,
-0.08058109879493713,
0.09183690696954727,
0.01000227965414524,
0.020066218450665474,
-0.0012882562587037683,
0.02802308276295662,
-0.05990643426775932,
-0.05608614161610603,
-0.054205723106861115,
-0.02243775501847267,
0.13168251514434814,
-0.00581998098641634,
0.06823980808258057,
-0.0017437073402106762,
-0.02880006842315197,
0.01792764663696289,
0.0001704516471363604,
-0.03612356260418892,
-0.09651404619216919,
-0.014789752662181854,
0.019631637260317802,
-0.04576593264937401,
-0.015005499124526978,
0.0709705725312233,
0.021377235651016235,
-0.06745731830596924,
0.005559407640248537,
0.11019141972064972,
-0.04011159762740135,
0.06862622499465942,
0.062654510140419,
-0.06572569906711578,
0.031107638031244278,
0.059505391865968704,
-0.06221386790275574,
0.017467088997364044,
0.0042357551865279675,
-0.0803871899843216,
0.041288577020168304,
0.0938870757818222,
0.021130597218871117,
0.06833354383707047,
0.036031026393175125,
-0.05983295291662216,
-0.0653342679142952,
-0.06198390573263168,
0.012955866754055023,
0.09312949329614639,
-0.10220745205879211,
0.044441647827625275,
-0.02235260419547558,
-0.03221989423036575,
-0.03446952626109123,
0.00015795623767189682,
0.024833664298057556,
1.759702553553487e-33,
0.09977734088897705,
0.0586455799639225,
0.04633861035108566,
-0.008446799591183662,
0.11370561271905899,
0.04211484268307686,
-0.007005088031291962,
-0.01599397137761116,
-0.006713624577969313,
-0.012930541299283504,
-0.05982962250709534,
-0.023648742586374283,
0.017866216599941254,
0.019097918644547462,
-0.02026524394750595,
-0.0093399528414011,
-0.008927814662456512,
-0.02021826058626175,
0.04310276731848717,
-0.08669392019510269,
0.09563162177801132,
-0.047653112560510635,
0.03634796664118767,
-0.023528216406702995,
0.0319974385201931,
0.03831956535577774,
0.028886578977108,
-0.02714354731142521,
0.04216936603188515,
0.00396695826202631,
-0.03728964552283287,
0.033306851983070374,
0.029092973098158836,
0.07035556435585022,
-0.001297623966820538,
-0.033236611634492874,
-0.014115732163190842,
-0.05851815640926361,
-0.004535522311925888,
-0.09053311496973038,
0.01935707964003086,
-0.006020307540893555,
0.030290020629763603,
-0.01504385657608509,
-0.07111308723688126,
-0.021314464509487152,
-0.01605970412492752,
-0.04180377349257469,
-0.008694094605743885,
-0.05794139951467514,
0.05211203545331955,
-0.020491963252425194,
-0.014140569604933262,
0.005459675565361977,
0.015626899898052216,
-0.05664852634072304,
0.0897059366106987,
0.06376196444034576,
-0.005094190128147602,
0.06393659859895706,
0.05828985944390297,
0.06483933329582214,
0.04424084722995758,
0.03402646258473396,
0.020489517599344254,
0.04882083460688591,
-0.07272405922412872,
0.03818454593420029,
0.10777921974658966,
-0.11125419288873672,
0.032735370099544525,
0.022220253944396973,
-0.016989294439554214,
-0.025326114147901535,
0.06739319115877151,
0.04160096123814583,
-0.03911969065666199,
-0.17997892200946808,
-0.013058523647487164,
0.05895271897315979,
-0.0990639254450798,
-0.0652664378285408,
-0.03245053440332413,
-0.03872285038232803,
-0.04278945177793503,
-0.07807964086532593,
0.03350216522812843,
-0.06781890243291855,
-0.0046989829279482365,
-0.0008502285345457494,
-0.04238274320960045,
-0.05747818946838379,
0.005484460387378931,
0.007708458695560694,
0.02178516797721386,
-2.0805203084898288e-33,
-0.06112891435623169,
0.03930508345365524,
0.01240635383874178,
0.057973697781562805,
-0.027040354907512665,
-0.09309487044811249,
-0.03557467460632324,
-0.0353798046708107,
0.011262299492955208,
-0.004968484863638878,
-0.07285302877426147,
-0.027354935184121132,
0.020766284316778183,
0.014015610329806805,
0.0691717118024826,
-0.05487070605158806,
0.04693901166319847,
-0.0025242965202778578,
-0.022902837023139,
0.01962949149310589,
-0.011089558713138103,
0.08076930046081543,
-0.14535808563232422,
0.0548069104552269,
0.07712849974632263,
0.01934046298265457,
0.025895601138472557,
0.05648372322320938,
0.023537227883934975,
0.016464143991470337,
-0.06166188046336174,
-0.04420223832130432,
-0.05796435475349426,
-0.04739752784371376,
-0.03693825751543045,
0.022630395367741585,
0.012619021348655224,
0.02516450732946396,
-0.04217039793729782,
0.02439689263701439,
0.07834959775209427,
-0.047589607536792755,
-0.14246723055839539,
-0.0065679410472512245,
0.008536032401025295,
0.03814024478197098,
-0.11771703511476517,
0.012601018883287907,
-0.08586350083351135,
0.000019410725144552998,
-0.021054977551102638,
0.06743394583463669,
-0.00329209235496819,
0.019217628985643387,
-0.030207164585590363,
-0.06460083276033401,
-0.0684373527765274,
0.02971990779042244,
-0.005578986369073391,
-0.008670846931636333,
-0.056241218000650406,
-0.014628811739385128,
-0.021453149616718292,
0.01839226484298706,
0.003988959826529026,
-0.050783101469278336,
-0.0376460999250412,
-0.03923457860946655,
0.05607733130455017,
0.028556041419506073,
0.030619965866208076,
-0.03457021340727806,
-0.04659920930862427,
0.03558832034468651,
0.060867950320243835,
-0.00535796070471406,
0.052458327263593674,
-0.002508096396923065,
-0.027988210320472717,
-0.03731747344136238,
0.02946172095835209,
-0.0376906581223011,
0.06170077249407768,
0.01565304957330227,
0.08187311142683029,
-0.009612481109797955,
-0.0004092361486982554,
-0.011329717002809048,
0.06205694004893303,
0.021797116845846176,
0.012136249803006649,
0.029743826016783714,
-0.0624454990029335,
0.043181754648685455,
-0.03981968015432358,
-3.926828640032909e-8,
-0.13672949373722076,
0.04856725037097931,
0.06293930858373642,
0.040181271731853485,
0.1585184931755066,
0.006990169640630484,
0.08625580370426178,
0.08840350061655045,
-0.03534402325749397,
0.06384199112653732,
0.0752069279551506,
0.003275595372542739,
-0.007664406672120094,
-0.07324746251106262,
-0.046336203813552856,
0.02729552611708641,
0.024463459849357605,
0.010005536489188671,
0.015983425080776215,
0.015551942400634289,
0.012072996236383915,
-0.017072554677724838,
-0.03652067855000496,
0.04796018451452255,
0.00843868125230074,
-0.08409550040960312,
-0.011880917474627495,
0.02304229699075222,
0.03207562863826752,
0.030059602111577988,
-0.037090692669153214,
0.012438485398888588,
0.04939112067222595,
0.03449145331978798,
-0.009593944065272808,
0.12800545990467072,
0.10542497783899307,
-0.007218748331069946,
0.02812410145998001,
-0.035966906696558,
0.03870178014039993,
0.04159952327609062,
-0.050330743193626404,
0.00398213928565383,
-0.04247007146477699,
-0.039620500057935715,
0.005198867991566658,
-0.07324817031621933,
0.046833865344524384,
-0.04496289789676666,
-0.038953304290771484,
-0.02947055548429489,
0.0007842476479709148,
-0.0298773180693388,
-0.0036706896498799324,
0.02578582800924778,
-0.03498992696404457,
-0.027156949043273926,
-0.04261452704668045,
0.09961209446191788,
0.007821543142199516,
-0.02150077186524868,
-0.029544338583946228,
-0.049782559275627136
] |
DeFT: Decoding with Flash Tree-attention for Efficient Tree-structured LLM Inference | https://openreview.net/forum?id=2c7pfOqu9k | [
"Jinwei Yao",
"Kaiqi Chen",
"Kexun Zhang",
"Jiaxuan You",
"Binhang Yuan",
"Zeke Wang",
"Tao Lin"
] | Spotlight | Large language models (LLMs) are increasingly employed for complex tasks that process multiple generation calls in a tree structure with shared prefixes of tokens, including few-shot prompting, multi-step reasoning, speculative decoding, etc. However, existing inference systems for tree-based applications are inefficient due to improper partitioning of queries and KV cache during attention calculation.This leads to two main issues: (1) a lack of memory access (IO) reuse for KV cache of shared prefixes, and (2) poor load balancing.As a result, there is redundant KV cache IO between GPU global memory and shared memory, along with low GPU utilization. To address these challenges, we propose DeFT(Decoding with Flash Tree-Attention), a hardware-efficient attention algorithm with prefix-aware and load-balanced KV cache partitions. DeFT reduces the number of read/write operations of KV cache during attention calculation through **KV-Guided Grouping**, a method that avoids repeatedly loading KV cache of shared prefixes in attention computation. Additionally, we propose **Flattened Tree KV Splitting**, a mechanism that ensures even distribution of the KV cache across partitions with little computation redundancy, enhancing GPU utilization during attention computations. By reducing 73-99% KV cache IO and nearly 100% IO for partial results during attention calculation, DeFT achieves up to 2.23/3.59$\times$ speedup in the end-to-end/attention latency across three practical tree-based workloads compared to state-of-the-art attention algorithms. Our code is available at https://github.com/LINs-lab/DeFT. | LLM inference, attention, memory-efficiency, tree-based decoding | We propose DeFT, a hardware-efficient tree attention algorithm to improve the tree-based decoding (e.g. multi-step reasoning, speculative decoding, etc) efficiency with IO-awareness for shared prefixes and load-balancing.. | 8,454 | 2404.00242 | [
-0.00010926272079814225,
-0.03407592698931694,
0.02021639049053192,
0.034246642142534256,
0.044330570846796036,
-0.08808555454015732,
-0.025120988488197327,
0.0488552488386631,
0.031700585037469864,
-0.03225325793027878,
-0.050779227167367935,
-0.03024829551577568,
-0.03622524440288544,
0.01068823877722025,
0.015346303582191467,
-0.014822694472968578,
0.022188754752278328,
0.07184261083602905,
-0.05307944491505623,
-0.0832071378827095,
0.06326735019683838,
-0.054078880697488785,
0.007739201188087463,
0.022142203524708748,
0.031730927526950836,
0.020404323935508728,
-0.02752847597002983,
-0.04703407362103462,
0.04141994193196297,
0.005593128502368927,
0.05605578050017357,
0.027069902047514915,
-0.006886765360832214,
0.07665248960256577,
-0.048214301466941833,
0.039222393184900284,
-0.09768041223287582,
-0.1304933875799179,
-0.034893669188022614,
-0.050886619836091995,
-0.023532915860414505,
0.04504666477441788,
-0.017145726829767227,
0.04616055265069008,
0.1125643327832222,
0.03251843899488449,
0.04353872314095497,
0.011902821250259876,
-0.059963028877973557,
-0.05575619637966156,
0.002068949630483985,
-0.05257561430335045,
-0.05641775578260422,
0.0504445806145668,
-0.03298920765519142,
0.04790335148572922,
0.04534552991390228,
-0.057364873588085175,
-0.02253081649541855,
0.026309754699468613,
-0.026354705914855003,
-0.09601273387670517,
0.014075817540287971,
0.0024366509169340134,
-0.04739799350500107,
0.0064482917077839375,
0.1466883271932602,
-0.0098528191447258,
0.032493315637111664,
-0.010090026073157787,
-0.016221309080719948,
0.09352569282054901,
-0.043037187308073044,
0.0005855967756360769,
-0.06620445847511292,
0.07024651765823364,
0.014254115521907806,
-0.04983251914381981,
0.06488654762506485,
-0.0704769417643547,
-0.0033215947914868593,
-0.01618696004152298,
0.03205201402306557,
-0.054901111871004105,
0.11678434163331985,
-0.03392969071865082,
0.03519497066736221,
0.034776076674461365,
0.012982827611267567,
0.02844994328916073,
-0.03432857617735863,
-0.08565570414066315,
0.08877813071012497,
0.026982735842466354,
-0.024219460785388947,
0.03718908876180649,
0.008324853144586086,
-0.04548367112874985,
-0.08841698616743088,
0.04131950065493584,
-0.017872311174869537,
0.11199932545423508,
0.06400961428880692,
-0.03388335555791855,
-0.005812176037579775,
0.019775552675127983,
-0.030979402363300323,
0.059823177754879,
-0.055656906217336655,
-0.047016046941280365,
0.03540425002574921,
0.06965301185846329,
0.031263720244169235,
-0.0029730526730418205,
0.005736825987696648,
-0.010605079121887684,
-0.015264258719980717,
-0.048361506313085556,
0.05424699932336807,
0.06968683004379272,
-0.012789195403456688,
-0.0005382314557209611,
-0.07056327909231186,
0.028711916878819466,
0.04863801971077919,
-0.09779582172632217,
-0.05791965499520302,
1.922887067057537e-33,
0.06298740208148956,
-0.010638990439474583,
0.0038205510936677456,
0.03379196301102638,
0.03356704115867615,
0.016581827774643898,
0.05993040278553963,
0.042260680347681046,
-0.06698321551084518,
-0.017836157232522964,
0.028685392811894417,
-0.023601001128554344,
0.008251107297837734,
0.022614814341068268,
0.05116843432188034,
0.0385022833943367,
-0.012303940951824188,
0.010394922457635403,
-0.046699944883584976,
-0.0007167596486397088,
0.031187167391180992,
0.0504017099738121,
-0.009047586470842361,
-0.06709648668766022,
0.06444565206766129,
0.036286622285842896,
0.007795909885317087,
-0.07670154422521591,
0.024561144411563873,
0.024308040738105774,
-0.06255318969488144,
-0.08308970183134079,
-0.0574052669107914,
0.06297800689935684,
0.03278226777911186,
-0.057440876960754395,
0.019048165529966354,
-0.09468691051006317,
0.025002356618642807,
-0.0219923947006464,
0.021858589723706245,
0.04527459293603897,
0.020740892738103867,
-0.09208700805902481,
-0.12082290649414062,
-0.02160867117345333,
-0.055263180285692215,
0.020730512216687202,
-0.018093815073370934,
0.033696357160806656,
0.05680251494050026,
0.02806110866367817,
-0.044264014810323715,
-0.010619805194437504,
-0.017222443595528603,
-0.04435233399271965,
0.0061778416857123375,
0.03800320625305176,
0.09331460297107697,
0.13913950324058533,
-0.012151870876550674,
-0.009357416070997715,
-0.014550416730344296,
0.10087317228317261,
-0.008157334290444851,
0.05471421405673027,
-0.039356522262096405,
0.0896245613694191,
-0.016332820057868958,
0.05137048661708832,
-0.0672827884554863,
-0.02996126003563404,
0.024947406724095345,
-0.07046358287334442,
0.015467062592506409,
0.0460670068860054,
0.0375419445335865,
-0.09141837060451508,
-0.04342937469482422,
-0.006913982331752777,
-0.006243555806577206,
0.004681424703449011,
0.01312270574271679,
0.053468093276023865,
-0.02442314103245735,
-0.08358490467071533,
0.03810514882206917,
-0.024651367217302322,
-0.04669049382209778,
-0.006728185806423426,
-0.051351554691791534,
-0.05870872363448143,
0.08709906786680222,
-0.041615501046180725,
-0.06184221804141998,
-1.6835345603820336e-33,
0.0037534653674811125,
-0.03316954895853996,
-0.025168849155306816,
0.09969160705804825,
-0.019668186083436012,
-0.042742155492305756,
-0.02364349737763405,
-0.07064400613307953,
-0.06515046954154968,
-0.055235620588064194,
-0.10603777319192886,
0.008237389847636223,
0.02396504208445549,
0.03137001395225525,
0.048380959779024124,
-0.034132953733205795,
0.021665634587407112,
-0.01478861179202795,
-0.07867518067359924,
0.08855096250772476,
-0.023039057850837708,
0.07622844725847244,
-0.008035033009946346,
-0.02647187002003193,
-0.024671019986271858,
0.0361454002559185,
-0.026756830513477325,
-0.01584235392510891,
0.015578821301460266,
-0.02677851915359497,
-0.03841957449913025,
0.004429731052368879,
0.02230275608599186,
-0.042889658361673355,
0.0037650507874786854,
0.040659960359334946,
0.0720941498875618,
0.004677110817283392,
-0.02003215067088604,
0.08346385508775711,
0.12923675775527954,
-0.045277614146471024,
-0.029999002814292908,
0.028030389919877052,
-0.031235359609127045,
0.06151435151696205,
-0.10448044538497925,
-0.0018355014035478234,
-0.03454280272126198,
0.014976405538618565,
0.00751907704398036,
0.019481763243675232,
-0.013097664341330528,
0.04253397136926651,
-0.022265689447522163,
-0.07924216985702515,
-0.06901510059833527,
0.04324343428015709,
0.026217060163617134,
-0.06614566594362259,
0.05013492330908775,
-0.05646770820021629,
-0.022856691852211952,
-0.11926931887865067,
0.02324654534459114,
-0.06457637250423431,
-0.01680108532309532,
-0.07820003479719162,
-0.012314585037529469,
-0.034332890063524246,
0.0461905263364315,
-0.008348173461854458,
-0.016079632565379143,
-0.0006688556168228388,
-0.01790553890168667,
0.07701512426137924,
0.03151366859674454,
-0.11355295777320862,
0.03253955394029617,
0.011325965635478497,
-0.037218958139419556,
0.01997196301817894,
0.020099960267543793,
0.01169020589441061,
0.04081741347908974,
-0.02112770639359951,
0.03621198236942291,
0.09766370058059692,
0.012026781216263771,
-0.05870097130537033,
-0.07390034198760986,
0.08929172903299332,
0.038426872342824936,
0.05050184577703476,
-0.03540534898638725,
-3.745934407106688e-8,
0.0036580930463969707,
0.0006874979590065777,
-0.006985617335885763,
0.13404177129268646,
0.053047168999910355,
-0.09500502049922943,
-0.004467502236366272,
0.10261145979166031,
-0.0077958120964467525,
0.01748867891728878,
0.1162385493516922,
-0.07398553937673569,
-0.03190821409225464,
-0.05237885192036629,
0.06647820770740509,
0.078731007874012,
0.005851174704730511,
-0.015741240233182907,
-0.04732079431414604,
-0.03476683050394058,
-0.04761383682489395,
0.0476287342607975,
0.007421367801725864,
0.06256614625453949,
0.006582857109606266,
-0.02959037385880947,
0.11591649055480957,
0.07063448429107666,
0.08398088067770004,
-0.032158926129341125,
0.03706970438361168,
0.013205112889409065,
-0.003353867679834366,
-0.014338544569909573,
-0.0030246470123529434,
0.0174830611795187,
0.04362054541707039,
0.09298712015151978,
0.10190244019031525,
0.05173497647047043,
0.0029231065418571234,
-0.07407260686159134,
0.005020450334995985,
0.08729302138090134,
-0.027385041117668152,
-0.015632377937436104,
-0.08769777417182922,
-0.029167387634515762,
0.01065713819116354,
0.013539724983274937,
-0.01808714121580124,
0.0022046093363314867,
-0.06129071116447449,
0.06781332939863205,
-0.05922294408082962,
0.023309748619794846,
0.013531598262488842,
-0.028090890496969223,
0.022918544709682465,
0.017435580492019653,
0.10142716020345688,
-0.06621374189853668,
-0.07336563616991043,
0.007431795354932547
] |
How to Find the Exact Pareto Front for Multi-Objective MDPs? | https://openreview.net/forum?id=S4dItvpvAv | [
"Yining Li",
"Peizhong Ju",
"Ness Shroff"
] | Spotlight | Multi-Objective Markov Decision Processes (MO-MDPs) are receiving increasing attention, as real-world decision-making problems often involve conflicting objectives that cannot be addressed by a single-objective MDP.
The Pareto front identifies the set of policies that cannot be dominated, providing a foundation for finding Pareto optimal solutions that can efficiently adapt to various preferences.
However, finding the Pareto front is a highly challenging problem. Most existing methods either (i) rely on traversing the *continuous preference space*, which is impractical and results in approximations that are difficult to evaluate against the true Pareto front, or (ii) focus solely on deterministic Pareto optimal policies, from which there are no known techniques to characterize the full Pareto front. Moreover, finding the structure of the Pareto front itself remains unclear even in the context of dynamic programming, where the MDP is fully known in advance.
In this work, we address the challenge of efficiently discovering the Pareto front, involving both deterministic and stochastic Pareto optimal policies.
By investigating the geometric structure of the Pareto front in MO-MDPs, we uncover a key property: the Pareto front is on the boundary of a convex polytope whose vertices all correspond to deterministic policies, and neighboring vertices of the Pareto front differ by only one state-action pair of the deterministic policy, almost surely.
This insight transforms the global comparison across all policies into a localized search among deterministic policies that differ by only one state-action pair, drastically reducing the complexity of searching for the exact Pareto front.
We develop an efficient algorithm that identifies the vertices of the Pareto front by solving a single-objective MDP only once and then traversing the edges of the Pareto front, making it more efficient than existing methods. Furthermore, the entire Pareto front can be found in $V$ iterations, where $V$ represents the number of vertices on the Pareto front.
Our empirical studies demonstrate the effectiveness of our theoretical strategy in discovering the Pareto front efficiently. | Multi-objective optimization, Markov decision Process | null | 8,449 | 2410.15557 | [
-0.026949599385261536,
-0.05789583921432495,
-0.0392720066010952,
-0.09303233027458191,
0.0032938611693680286,
-0.04826068505644798,
0.03714974969625473,
-0.04428140074014664,
-0.06913915276527405,
0.04340796917676926,
-0.06770237535238266,
-0.03451799228787422,
0.021764682605862617,
0.012054272927343845,
-0.0036671492271125317,
0.0976552739739418,
0.07352413982152939,
0.0002026717847911641,
0.017506709322333336,
0.07639702409505844,
-0.011943337507545948,
0.031248394399881363,
-0.01675776205956936,
0.01677561178803444,
-0.016960082575678825,
0.051823846995830536,
0.03543432429432869,
-0.03296029567718506,
-0.08396649360656738,
0.009531446732580662,
0.07329855114221573,
-0.02265835739672184,
0.06635463982820511,
-0.07291163504123688,
0.008474132046103477,
0.13142433762550354,
-0.09970743954181671,
-0.005122100003063679,
0.021084247156977654,
-0.017129860818386078,
0.004548672121018171,
0.046960629522800446,
0.003842407139018178,
0.005866555497050285,
-0.04668637737631798,
-0.020739197731018066,
0.03064652532339096,
0.05006539821624756,
-0.015944287180900574,
-0.039074625819921494,
0.038194816559553146,
-0.02399456314742565,
-0.10716182738542557,
0.050385523587465286,
0.03482265770435333,
0.032939497381448746,
-0.03495492413640022,
-0.02148565836250782,
-0.046079035848379135,
-0.017482396215200424,
-0.048064205795526505,
-0.05543159693479538,
-0.04958811029791832,
-0.027415381744503975,
0.05029664188623428,
-0.05105966702103615,
0.004561114124953747,
-0.013934968039393425,
-0.021127251908183098,
0.03933184593915939,
-0.015189725905656815,
-0.0547986775636673,
-0.028972748667001724,
-0.027608735486865044,
-0.00043064175406470895,
0.0056437039747834206,
0.11332288384437561,
-0.000764639931730926,
0.02023153007030487,
-0.05162782967090607,
-0.03960032761096954,
0.031222248449921608,
0.02022503688931465,
0.057872943580150604,
0.10454822331666946,
-0.03299954533576965,
-0.05430902540683746,
0.01613769121468067,
0.12474863976240158,
-0.031686946749687195,
-0.012401656247675419,
0.02971087582409382,
-0.02548699639737606,
-0.015764130279421806,
0.028626736253499985,
0.010301928967237473,
-0.02587970718741417,
-0.07417513430118561,
0.00941417831927538,
0.07981400936841965,
-0.03635556995868683,
-0.03419290482997894,
0.03977116197347641,
-0.031158674508333206,
0.025892820209264755,
-0.07274690270423889,
-0.011719555594027042,
0.013256336562335491,
-0.03709844872355461,
-0.027446644380688667,
0.030850138515233994,
0.07499495893716812,
0.06333979964256287,
0.0634656623005867,
-0.032523103058338165,
-0.01589364930987358,
0.012079511769115925,
-0.04355522245168686,
0.041032057255506516,
0.03809057176113129,
-0.130776047706604,
-0.0036620162427425385,
0.0760033130645752,
0.0008056702208705246,
0.0353267677128315,
0.049924369901418686,
-0.0038256559055298567,
1.9004321862209542e-33,
-0.07766374945640564,
-0.018664525821805,
0.06955995410680771,
-0.09542509913444519,
-0.003298300551250577,
-0.03591320663690567,
0.05724042281508446,
-0.007403511554002762,
0.007273510564118624,
0.037994805723428726,
0.021968724206089973,
-0.006794403772801161,
-0.049418333917856216,
-0.021953528746962547,
0.06723625957965851,
-0.044350024312734604,
-0.05940038710832596,
0.11476849019527435,
-0.02778058685362339,
-0.1446392685174942,
-0.0007697201217524707,
0.08364463597536087,
-0.045421477407217026,
-0.030101720243692398,
0.09487094730138779,
0.001784760388545692,
0.019585425034165382,
-0.07395698130130768,
-0.00611660024151206,
0.07036884874105453,
-0.0037231610622256994,
0.03506120294332504,
-0.013244850561022758,
0.0566953606903553,
-0.04072033241391182,
0.04488999769091606,
-0.12084219604730606,
-0.031191498041152954,
-0.06138468533754349,
-0.054872024804353714,
-0.0763988047838211,
-0.04186144471168518,
0.04412783309817314,
0.015124350786209106,
-0.1046837568283081,
-0.04371337592601776,
0.011126792058348656,
0.03702965006232262,
-0.006540229078382254,
-0.01875220611691475,
0.025715138763189316,
-0.046865493059158325,
0.052047278732061386,
-0.05658750236034393,
-0.11496274173259735,
-0.0599980391561985,
-0.027602234855294228,
0.07557011395692825,
0.06741927564144135,
0.035332899540662766,
-0.030470166355371475,
-0.017722629010677338,
-0.018577532842755318,
-0.015115291811525822,
0.009673702530562878,
0.04359905794262886,
-0.01827014796435833,
-0.05376609042286873,
0.18255221843719482,
-0.006839606445282698,
0.006390358787029982,
-0.012929016724228859,
0.119951993227005,
-0.09246451407670975,
0.0030412268824875355,
-0.018049711361527443,
0.06809411942958832,
0.015742335468530655,
0.020374847576022148,
-0.00004393098424770869,
-0.028415001928806305,
0.05574949458241463,
-0.007714119739830494,
-0.037296049296855927,
0.10381872951984406,
-0.003485938301309943,
0.0421781986951828,
-0.06067177653312683,
-0.053066495805978775,
0.05978884920477867,
-0.04731316491961479,
0.006926227826625109,
-0.04647267982363701,
-0.01610638201236725,
0.02838953398168087,
-3.14383282530933e-33,
-0.03054802119731903,
-0.098556749522686,
-0.006684652529656887,
0.04819468408823013,
0.028674904257059097,
-0.03696298971772194,
-0.08347748219966888,
-0.09313192218542099,
0.012082586996257305,
-0.02141593210399151,
-0.020788736641407013,
-0.0027952799573540688,
0.014064885675907135,
-0.003675124142318964,
-0.06197718530893326,
0.021604079753160477,
0.00910482369363308,
-0.03642118722200394,
-0.0024190321564674377,
0.02925919182598591,
0.09495797008275986,
0.05575175955891609,
-0.05657605826854706,
0.09815414249897003,
-0.004511911887675524,
0.00990895926952362,
-0.03686625137925148,
0.059163857251405716,
0.02563055045902729,
0.03957284614443779,
-0.07361112534999847,
-0.04174625128507614,
-0.11057537794113159,
-0.0076476153917610645,
-0.008560562506318092,
0.005971810780465603,
-0.0983075499534607,
0.09492409229278564,
-0.03751339763402939,
0.13511310517787933,
0.009066927246749401,
-0.10217257589101791,
0.010539003647863865,
0.061274610459804535,
0.02042977698147297,
-0.009030637331306934,
0.025205833837389946,
0.035071518272161484,
0.020996369421482086,
-0.020585069432854652,
0.006295616738498211,
0.047460608184337616,
0.0025494524743407965,
0.09994177520275116,
-0.06450928747653961,
-0.0379461869597435,
0.021093185991048813,
-0.038703981786966324,
0.014744551852345467,
-0.006081299390643835,
0.0008579747518524528,
0.058576036244630814,
-0.005541651509702206,
0.05967968702316284,
0.03208160400390625,
0.014658967033028603,
-0.028670625761151314,
-0.06153791770339012,
-0.009892495349049568,
0.015335050411522388,
-0.05850932002067566,
-0.001891721272841096,
0.048640619963407516,
-0.016693854704499245,
0.04609346389770508,
0.08636193722486496,
0.014388613402843475,
-0.00434436509385705,
0.026506511494517326,
0.02385552041232586,
0.019614646211266518,
0.09356383979320526,
0.052666012197732925,
0.02612921968102455,
-0.019687971100211143,
-0.042407743632793427,
0.027012832462787628,
0.030772607773542404,
0.012739954516291618,
0.019618630409240723,
-0.029708759859204292,
-0.015640845522284508,
0.06489714235067368,
-0.024950485676527023,
-0.0611744150519371,
-4.35988098956841e-8,
-0.026711903512477875,
-0.03015844151377678,
0.08166194707155228,
-0.00992827583104372,
0.07448872178792953,
0.08049163222312927,
-0.06718116998672485,
0.06035463884472847,
0.017120718955993652,
-0.05942380428314209,
0.06420999020338058,
0.09019295871257782,
0.042305562645196915,
-0.04885927215218544,
-0.06704734265804291,
0.0020958285313099623,
-0.03423447906970978,
-0.04179764539003372,
-0.04967259243130684,
-0.05052777752280235,
-0.02007850632071495,
0.025332028046250343,
0.034818410873413086,
-0.020444288849830627,
0.051611993461847305,
0.11989187449216843,
-0.10307923704385757,
-0.05020595341920853,
0.06852996349334717,
0.11810494214296341,
0.05449730157852173,
0.001596106681972742,
0.053287722170352936,
0.022165076807141304,
-0.0791596993803978,
0.09274809062480927,
0.00694787735119462,
-0.026614747941493988,
-0.015411625616252422,
-0.030286207795143127,
0.0046923537738621235,
-0.016824785619974136,
-0.04248816892504692,
-0.010821499861776829,
-0.03218389302492142,
0.04688360169529915,
-0.03932129591703415,
-0.021653849631547928,
0.013430802151560783,
-0.04329344630241394,
-0.002321036299690604,
-0.0682796910405159,
-0.0013498220359906554,
0.04211253672838211,
0.0315597839653492,
-0.06978143751621246,
-0.008042792789638042,
-0.11583027243614197,
0.06350237131118774,
0.03106735087931156,
0.023198142647743225,
0.01151799876242876,
-0.0027626107912510633,
0.057423125952482224
] |
Enhancing Learning with Label Differential Privacy by Vector Approximation | https://openreview.net/forum?id=IwPXYk6BV9 | [
"Puning Zhao",
"Jiafei Wu",
"Zhe Liu",
"Li Shen",
"Zhikun Zhang",
"Rongfei Fan",
"Le Sun",
"Qingming Li"
] | Spotlight | Label differential privacy (DP) is a framework that protects the privacy of labels in training datasets, while the feature vectors are public. Existing approaches protect the privacy of labels by flipping them randomly, and then train a model to make the output approximate the privatized label. However, as the number of classes K increases, stronger randomization is needed, thus the performances of these methods become significantly worse. In this paper, we propose a vector approximation approach for learning with label local differential privacy, which is easy to implement and introduces little additional computational overhead. Instead of flipping each label into a single scalar, our method converts each label into a random vector with K components, whose expectations reflect class conditional probabilities. Intuitively, vector approximation retains more information than scalar labels. A brief theoretical analysis shows that the performance of our method only decays slightly with K. Finally, we conduct experiments on both synthesized and real datasets, which validate our theoretical analysis as well as the practical performance of our method. | label differential privacy | null | 8,446 | 2405.15150 | [
-0.11617784202098846,
-0.025174980983138084,
-0.03364846110343933,
0.013080085627734661,
0.04036165401339531,
0.012449718080461025,
0.07490014284849167,
-0.04043693467974663,
0.03860132768750191,
-0.02843848243355751,
0.038077425211668015,
0.007706360425800085,
0.021240253001451492,
-0.05440380424261093,
-0.015136541798710823,
-0.007464060094207525,
0.08398561179637909,
0.09650071710348129,
-0.06909706443548203,
-0.02608819492161274,
-0.09207823872566223,
0.045781537890434265,
0.056597255170345306,
0.045507002621889114,
0.044629357755184174,
0.002495085820555687,
0.12201355397701263,
0.01353814173489809,
-0.0003339949471410364,
-0.016937896609306335,
0.08095423132181168,
-0.004612256772816181,
-0.027966247871518135,
-0.05088476464152336,
-0.046231452375650406,
-0.031236669048666954,
-0.10805904865264893,
0.04388546198606491,
-0.007143182214349508,
0.005669578909873962,
0.00824626162648201,
0.012235449627041817,
-0.023365557193756104,
0.04317368566989899,
-0.027748441323637962,
0.0495733879506588,
0.04301225766539574,
-0.05197035148739815,
-0.0235389806330204,
-0.006148212123662233,
-0.017284808680415154,
-0.047251224517822266,
-0.05928029865026474,
-0.013619658537209034,
-0.0628223866224289,
-0.06990059465169907,
-0.01878691092133522,
-0.037446457892656326,
-0.030531199648976326,
0.10104630142450333,
-0.07631441205739975,
-0.08549581468105316,
0.025490786880254745,
-0.004205468110740185,
-0.03195463865995407,
0.013938075862824917,
0.045547232031822205,
0.053801022469997406,
-0.03993212431669235,
-0.005914910696446896,
-0.017493905499577522,
0.15585437417030334,
-0.024947820231318474,
0.023264069110155106,
0.01894841156899929,
0.051302410662174225,
-0.013631707988679409,
0.008402140811085701,
0.02805398218333721,
-0.05256691202521324,
-0.03652973845601082,
-0.011142413131892681,
0.05238895118236542,
-0.0018999067833647132,
0.07150329649448395,
0.007530474103987217,
0.01848508231341839,
-0.0066606164909899235,
0.043569814413785934,
0.031122423708438873,
-0.014991868287324905,
0.00562457088381052,
0.0028558524791151285,
-0.05981035530567169,
-0.028606995940208435,
-0.014964785426855087,
-0.006232380401343107,
-0.02358626388013363,
0.053558267652988434,
0.04165044054389,
-0.03718617558479309,
0.05785071849822998,
0.005082917399704456,
-0.0019701847340911627,
0.07564795762300491,
0.039283785969018936,
0.024201037362217903,
-0.029034020379185677,
0.09518029540777206,
-0.02184293419122696,
0.00697964895516634,
0.0021705711260437965,
-0.11449131369590759,
0.014547009021043777,
-0.005409166682511568,
-0.04584687575697899,
0.0182894766330719,
-0.03132084012031555,
0.049639519304037094,
0.02363463304936886,
-0.06063675880432129,
0.005798605736345053,
0.022046493366360664,
-0.023968802765011787,
-0.09042181074619293,
0.025868110358715057,
-0.08368054032325745,
2.819254957818295e-33,
0.029085783287882805,
-0.020450009033083916,
-0.04455293342471123,
-0.03484932333230972,
0.004781061317771673,
0.014772599563002586,
-0.04096328094601631,
0.03907344117760658,
0.0008109089103527367,
0.07428542524576187,
0.0330398790538311,
0.028965378180146217,
0.03935081139206886,
0.0003395219682715833,
0.01772378757596016,
0.04860278591513634,
0.024126533418893814,
0.06785323470830917,
0.0009818404214456677,
-0.04179590195417404,
0.1592777967453003,
-0.012029498815536499,
0.07704134285449982,
-0.038300130516290665,
0.00791880488395691,
-0.04720157012343407,
0.03526626527309418,
-0.0062508960254490376,
0.07048925012350082,
0.018743878230452538,
-0.060324836522340775,
0.016247591003775597,
0.10445916652679443,
0.050393037497997284,
0.08976899832487106,
0.015513327904045582,
-0.011256465688347816,
-0.019986143335700035,
0.056746236979961395,
0.015046913176774979,
0.009426194243133068,
-0.042527683079242706,
0.046505194157361984,
-0.05857396125793457,
-0.08182816207408905,
0.08746939152479172,
-0.024194233119487762,
0.03264620900154114,
0.00189525680616498,
-0.048073362559080124,
-0.029238617047667503,
0.0035253739915788174,
-0.09557577222585678,
0.017405351623892784,
-0.054335445165634155,
0.011687645688652992,
0.01919056475162506,
0.035133183002471924,
-0.04769867658615112,
0.0013641490368172526,
-0.084369957447052,
0.12206698954105377,
0.04820755869150162,
-0.02759777009487152,
-0.07161165028810501,
0.027985425665974617,
0.044675104320049286,
-0.024393340572714806,
0.10053756088018417,
-0.07296498864889145,
0.05409403517842293,
0.07837030291557312,
-0.08906012773513794,
-0.04545791447162628,
0.018344206735491753,
-0.010866842232644558,
0.016072222962975502,
-0.021821733564138412,
-0.018789920955896378,
0.001775323529727757,
0.014354707673192024,
0.01104788389056921,
-0.007515535224229097,
0.051907625049352646,
-0.04637745022773743,
-0.06282185763120651,
-0.010268290527164936,
0.007857338525354862,
-0.06206680089235306,
0.05986674129962921,
-0.06492234021425247,
-0.06422127038240433,
-0.018178323283791542,
-0.02741992101073265,
-0.0877307802438736,
-3.1067371623332555e-33,
-0.12887538969516754,
0.03858843073248863,
-0.016918394714593887,
0.0002660895697772503,
-0.0004221973067615181,
-0.0555216446518898,
-0.04295668005943298,
0.05733473598957062,
0.02631303481757641,
-0.041153017431497574,
-0.05636655539274216,
0.018877824768424034,
0.03902480751276016,
-0.04203534871339798,
0.02685563825070858,
0.08095420897006989,
-0.027017593383789062,
0.07069674134254456,
-0.03932332992553711,
0.038609277456998825,
-0.05039146542549133,
0.07412992417812347,
-0.07014888525009155,
0.11870120465755463,
0.05345749109983444,
-0.019157832488417625,
-0.030171984806656837,
0.08690357953310013,
0.06039279326796532,
-0.05753680318593979,
-0.11628402024507523,
-0.0403405986726284,
-0.04435208812355995,
-0.11790309101343155,
-0.07416248321533203,
-0.028713781386613846,
0.007884956896305084,
0.027358515188097954,
-0.0056667025201022625,
0.036648351699113846,
-0.03807250037789345,
-0.014351376332342625,
-0.12597103416919708,
-0.006959145423024893,
-0.024659452959895134,
-0.048008859157562256,
-0.06218142434954643,
0.0005808363202959299,
0.036765746772289276,
0.031042668968439102,
-0.018460284918546677,
0.03019649349153042,
-0.0299392007291317,
0.07604921609163284,
-0.042205046862363815,
-0.043205518275499344,
0.004900829866528511,
0.009949419647455215,
0.055229540914297104,
0.08658112585544586,
0.0031400825828313828,
0.0004977902281098068,
0.016806676983833313,
0.014069933444261551,
-0.000680100463796407,
-0.052798546850681305,
-0.05583049729466438,
-0.0041910940781235695,
-0.0012240444775670767,
0.028326457366347313,
0.08202555030584335,
-0.026196347549557686,
-0.046083319932222366,
0.03267715126276016,
-0.04914179444313049,
-0.009902308695018291,
0.042482271790504456,
-0.06755104660987854,
-0.011143586598336697,
0.09070548415184021,
0.12936371564865112,
-0.04265361279249191,
0.05941664054989815,
-0.025745853781700134,
0.07339548319578171,
-0.04273037612438202,
0.09603147953748703,
-0.07583829760551453,
-0.004124025348573923,
0.0020424064714461565,
0.01378393080085516,
0.03309934586286545,
-0.0755670890212059,
0.08183775097131729,
-0.03631244972348213,
-3.721063279726877e-8,
0.012774049304425716,
-0.040224384516477585,
-0.026736414059996605,
0.07201043516397476,
0.057285234332084656,
0.05286955088376999,
0.012871953658759594,
0.05465664714574814,
-0.14882981777191162,
0.0036331205628812313,
0.053086210042238235,
-0.021950503811240196,
-0.038843389600515366,
-0.046494804322719574,
-0.00775222247466445,
0.0673079788684845,
-0.009795677848160267,
-0.005148063413798809,
0.0318625271320343,
0.038234543055295944,
-0.08187295496463776,
0.00782740768045187,
0.014846346341073513,
-0.038115959614515305,
0.03221794590353966,
-0.03400644287467003,
0.05485011264681816,
-0.010118507780134678,
0.15031029284000397,
0.09020917862653732,
-0.03390544280409813,
-0.030639011412858963,
-0.019933361560106277,
0.03096822276711464,
-0.006322851404547691,
0.13689914345741272,
-0.046448688954114914,
-0.043066300451755524,
0.008002720773220062,
0.08048903197050095,
0.021663310006260872,
0.003116343868896365,
-0.02650669775903225,
0.027701597660779953,
-0.004846920724958181,
0.027797825634479523,
0.04434137046337128,
-0.12141022831201553,
0.01763695292174816,
-0.045193735510110855,
0.02778026834130287,
-0.01041124016046524,
-0.037859007716178894,
0.05748765170574188,
0.06740737706422806,
-0.01816120184957981,
-0.04708311706781387,
-0.028179675340652466,
-0.023657716810703278,
0.11090634018182755,
0.007441060151904821,
0.01957274228334427,
-0.04672190174460411,
0.010022743605077267
] |
VisualPredicator: Learning Abstract World Models with Neuro-Symbolic Predicates for Robot Planning | https://openreview.net/forum?id=QOfswj7hij | [
"Yichao Liang",
"Nishanth Kumar",
"Hao Tang",
"Adrian Weller",
"Joshua B. Tenenbaum",
"Tom Silver",
"Joao F. Henriques",
"Kevin Ellis"
] | Spotlight | Broadly intelligent agents should form task-specific abstractions that selectively expose the essential elements of a task, while abstracting away the complexity of the raw sensorimotor space. In this work, we present Neuro-Symbolic Predicates, a first-order abstraction language that combines the strengths of symbolic and neural knowledge representations. We outline an online algorithm for inventing such predicates and learning abstract world models. We compare our approach to hierarchical reinforcement learning, vision-language model planning, and symbolic predicate invention approaches, on both in- and out-of-distribution tasks across five simulated robotic domains. Results show that our approach offers better sample complexity, stronger out-of-distribution generalization, and improved interpretability. | learning abstractions for planning, neuro-symbolic ai, concept learning | learning neuro-symbolic predicates from interaction allows from improve sample efficiency, generalization and interpretability | 8,442 | 2410.23156 | [
0.06730424612760544,
-0.08478541672229767,
0.10042896121740341,
-0.023569360375404358,
0.06132135167717934,
0.013737105764448643,
-0.009105856530368328,
-0.008628403767943382,
-0.015567204914987087,
0.020023703575134277,
0.0017929015448316932,
-0.06908553838729858,
0.04024989902973175,
0.11581673473119736,
0.07921189069747925,
0.08108445256948471,
0.07133118063211441,
0.01670253276824951,
-0.07474607229232788,
-0.07384005934000015,
0.08465207368135452,
0.01790301315486431,
0.027477756142616272,
0.031077604740858078,
-0.022279290482401848,
0.07706611603498459,
0.07602830231189728,
-0.08666422963142395,
0.030117791146039963,
-0.013480160385370255,
0.016154397279024124,
-0.0581350140273571,
-0.003112001111730933,
-0.04060325399041176,
-0.0018962667090818286,
0.1581210196018219,
-0.08182080835103989,
-0.00888106506317854,
-0.020566795021295547,
-0.05699111148715019,
-0.00693001551553607,
-0.006905324291437864,
0.07345885783433914,
0.015527448616921902,
0.041931141167879105,
0.07956217974424362,
0.019218314439058304,
-0.058674123138189316,
0.03840700536966324,
-0.04236414283514023,
-0.0928020104765892,
-0.04761890321969986,
0.027921976521611214,
-0.05894012749195099,
0.026536833494901657,
0.0793372094631195,
0.016155248507857323,
-0.0037962652277201414,
0.011878373101353645,
-0.0904856026172638,
-0.00797655526548624,
-0.06708910316228867,
0.04397188499569893,
-0.02518903836607933,
-0.004689646884799004,
0.027860792353749275,
0.0017216247506439686,
-0.012225130572915077,
0.012360172346234322,
0.016666950657963753,
-0.0020702502224594355,
0.06590716540813446,
0.007849078625440598,
0.03610514849424362,
-0.016817988827824593,
0.019545244053006172,
0.03787841647863388,
-0.04575139656662941,
0.05991742014884949,
-0.038195572793483734,
-0.04010185971856117,
0.03312712535262108,
-0.046714458614587784,
0.08078110218048096,
0.1255144625902176,
0.03288303315639496,
-0.007703087758272886,
0.08824145793914795,
0.0553990863263607,
0.0209254901856184,
-0.037782300263643265,
-0.09211274236440659,
0.016102692112326622,
-0.025936776772141457,
-0.013636012561619282,
0.011396752670407295,
-0.05746813118457794,
-0.10622482746839523,
-0.04627414047718048,
0.024044912308454514,
0.030214564874768257,
0.0372859425842762,
0.03090105950832367,
-0.032501280307769775,
-0.04402632638812065,
-0.031084779649972916,
0.03920101746916771,
0.010956713929772377,
-0.0201103612780571,
-0.045284759253263474,
0.03790849819779396,
-0.020000088959932327,
0.0913795456290245,
0.020960381254553795,
-0.018730470910668373,
0.055095162242650986,
0.023281464353203773,
0.02986510470509529,
0.008874005638062954,
0.037214018404483795,
0.04939629137516022,
-0.016830191016197205,
-0.00679797725751996,
0.04931136965751648,
0.06547070294618607,
-0.05838128179311752,
0.004646791145205498,
-2.6756020176371267e-33,
0.050452250987291336,
-0.004990614019334316,
0.1051139310002327,
0.06526967883110046,
0.07035153359174728,
0.01676291972398758,
0.0456734336912632,
0.040802907198667526,
0.002650298411026597,
0.05977647006511688,
-0.032941896468400955,
0.023709019646048546,
-0.09655741602182388,
0.14065741002559662,
0.0013200133107602596,
-0.033938754349946976,
0.01612078957259655,
0.015655111521482468,
-0.06341666728258133,
0.012173588387668133,
0.03849693015217781,
0.017169279977679253,
-0.07095009833574295,
-0.02888810634613037,
-0.025007564574480057,
0.03764718398451805,
-0.006576863117516041,
-0.03560608625411987,
-0.003967614844441414,
-0.021380729973316193,
-0.09207358956336975,
0.018194900825619698,
-0.09253241866827011,
0.10829703509807587,
0.004435580689460039,
-0.0017240148736163974,
-0.055263448506593704,
-0.06856685131788254,
0.021761292591691017,
-0.05411496013402939,
0.02405550517141819,
0.031440261751413345,
-0.016832441091537476,
0.011270452290773392,
0.0026886097621172667,
-0.04027526080608368,
-0.03272383660078049,
0.011829103343188763,
0.007186441216617823,
-0.03270130977034569,
0.05988123640418053,
0.019828014075756073,
0.0036033098585903645,
-0.08180948346853256,
0.040937770158052444,
-0.015760116279125214,
-0.013676578179001808,
0.07062916457653046,
0.01064650435000658,
0.035286616533994675,
0.018372371792793274,
0.002560078864917159,
-0.009675974026322365,
0.07890018820762634,
0.008456110022962093,
0.047189921140670776,
0.01088409312069416,
0.000699471915140748,
0.14771297574043274,
-0.009537585079669952,
-0.007912001572549343,
0.07145208865404129,
-0.005971635226160288,
-0.024371806532144547,
0.02198190800845623,
0.009553500451147556,
0.04212165251374245,
-0.10715558379888535,
-0.003219564678147435,
0.06461483985185623,
-0.06730357557535172,
-0.024080853909254074,
-0.016268229112029076,
0.025514885783195496,
0.063979871571064,
-0.01707809418439865,
0.03539016470313072,
-0.02974664978682995,
-0.029211590066552162,
-0.06869077682495117,
-0.09323783963918686,
-0.035791944712400436,
-0.049388110637664795,
-0.00036666871164925396,
-0.05893167480826378,
4.383620013130377e-34,
-0.06449640542268753,
-0.003253607777878642,
0.10823012888431549,
-0.028517920523881912,
-0.026327090337872505,
-0.11114007979631424,
-0.009937622584402561,
-0.03193940222263336,
0.012213515117764473,
-0.04798907786607742,
-0.12056435644626617,
-0.04710480943322182,
0.00465090898796916,
-0.0008890120661817491,
0.02527018077671528,
0.04135102033615112,
-0.011704280972480774,
-0.017344994470477104,
0.023652076721191406,
-0.022772090509533882,
-0.02937460131943226,
0.12844732403755188,
-0.10774458199739456,
0.0026981751434504986,
-0.04357421398162842,
0.06855206936597824,
0.026211045682430267,
0.04434097930788994,
-0.007593480404466391,
0.09689196199178696,
-0.0470794253051281,
-0.10722658038139343,
-0.03184540942311287,
-0.03115849941968918,
-0.02423716150224209,
0.01962445303797722,
-0.011983674019575119,
0.011300439946353436,
-0.09219931811094284,
0.06929054856300354,
-0.012451707385480404,
-0.11316695064306259,
-0.023839030414819717,
-0.037982698529958725,
-0.04075876250863075,
-0.044356029480695724,
-0.037852201610803604,
0.008204909041523933,
-0.007896113209426403,
0.05800565332174301,
-0.04314980283379555,
0.06720194965600967,
-0.08684802055358887,
-0.07146578282117844,
-0.06408340483903885,
-0.062199633568525314,
-0.03932788968086243,
0.001997716026380658,
0.09409523755311966,
-0.03150573745369911,
-0.06678645312786102,
-0.0037044251803308725,
0.0022755898535251617,
0.08169679343700409,
-0.07434727996587753,
0.011818697676062584,
-0.050681617110967636,
0.03962940722703934,
0.012110135518014431,
-0.026628823950886726,
-0.008698268793523312,
-0.00781421922147274,
0.006397842895239592,
-0.003706590970978141,
0.07917128503322601,
0.0064649381674826145,
0.042273830622434616,
0.01830487884581089,
0.052237894386053085,
-0.01712222583591938,
0.024585336446762085,
-0.005467492621392012,
0.0541779100894928,
0.08389592170715332,
0.05545235797762871,
0.07261954993009567,
-0.000012129566130170133,
0.015399015508592129,
0.0379917249083519,
0.02190030738711357,
-0.02586054615676403,
0.006755673326551914,
-0.0236637182533741,
0.06348355114459991,
-0.11779062449932098,
-3.7029003863153775e-8,
-0.07134582847356796,
0.04652857035398483,
0.09953875839710236,
-0.05482669174671173,
0.07765068113803864,
0.012231707572937012,
-0.01643390767276287,
-0.015455285087227821,
-0.09239555895328522,
0.004456559661775827,
0.0383131168782711,
0.03268589451909065,
-0.01312584150582552,
-0.05168202146887779,
-0.05082615464925766,
0.015207752585411072,
0.0005467284354381263,
-0.02018595114350319,
0.02207891270518303,
0.008590002544224262,
0.06630745530128479,
-0.03050195798277855,
-0.038143958896398544,
-0.07126658409833908,
0.04497639462351799,
-0.07032269984483719,
-0.10247842222452164,
0.05598670244216919,
0.03461715206503868,
0.0557711198925972,
0.06863700598478317,
0.025927739217877388,
0.09903183579444885,
0.00913211889564991,
0.04760699346661568,
-0.03710373863577843,
-0.01883438229560852,
-0.04494462534785271,
-0.043877556920051575,
-0.08020445704460144,
-0.017117131501436234,
0.06865907460451126,
0.06540663540363312,
-0.005995531100779772,
0.01607748121023178,
0.01892216131091118,
-0.024318022653460503,
-0.12340612709522247,
-0.03331043943762779,
-0.010110510513186455,
-0.06420698016881943,
0.032984111458063126,
-0.03132152557373047,
0.07198228687047958,
0.06958060711622238,
-0.028847314417362213,
0.05211332067847252,
-0.08987240493297577,
-0.007981792092323303,
0.032153498381376266,
-0.010149733163416386,
0.03164418786764145,
0.01824968121945858,
-0.012034807354211807
] |
Multi-Draft Speculative Sampling: Canonical Decomposition and Theoretical Limits | https://openreview.net/forum?id=N1L5TgtkAw | [
"Ashish J Khisti",
"MohammadReza Ebrahimi",
"Hassan Dbouk",
"Arash Behboodi",
"Roland Memisevic",
"Christos Louizos"
] | Spotlight | We consider multi-draft speculative sampling, where the proposal sequences are sampled independently from different draft models. At each step, a token-level draft selection scheme takes a list of valid tokens as input and produces an output token whose distribution matches that of the target model. Previous works have demonstrated that the optimal scheme (which maximizes the probability of accepting one of the input tokens) can be cast as a solution to a linear program. In this work we show that the optimal scheme can be decomposed into a two-step solution: in the first step an importance sampling (IS) type scheme is used to select one intermediate token; in the second step (single-draft) speculative sampling is applied to generate the output token. For the case of two identical draft models we further 1) establish a necessary and sufficient condition on the distributions of the target and draft models for the acceptance probability to equal one and 2) provide an explicit expression for the optimal acceptance probability. Our theoretical analysis also motives a new class of token-level selection schemes based on weighted importance sampling. Our experimental results demonstrate consistent improvements in the achievable block efficiency and token rates over baseline schemes in a number of scenarios. | speculative decoding, multi draft speculative sampling, large language models, weighted importance sampling, optimal transport | We show the optimal multi-draft speculative sampling can be decomposed into a two-step solution: importance sampling followed by (single-draft) speculative sampling, and provide an explicit expression for the optimal acceptance probability. | 8,423 | null | [
-0.053820013999938965,
-0.011402127332985401,
0.0012418716214597225,
-0.00994180329144001,
0.1163771003484726,
-0.04158688709139824,
0.016950951889157295,
0.05089946091175079,
0.0619254931807518,
0.019139500334858894,
-0.07442156225442886,
0.0017135378438979387,
0.05028195679187775,
-0.033517591655254364,
0.06783464550971985,
0.000014579300113837235,
0.03764902427792549,
0.042626168578863144,
-0.062349408864974976,
0.025612832978367805,
-0.009625124745070934,
-0.02425403892993927,
0.006189034786075354,
0.0003371683997102082,
0.039009664207696915,
-0.07629320025444031,
0.0290422011166811,
0.003109806915745139,
0.010922065936028957,
0.01297063659876585,
0.02335648611187935,
0.10370157659053802,
-0.00982043705880642,
-0.08780283480882645,
0.015691213309764862,
-0.026586312800645828,
-0.03866736218333244,
-0.042245250195264816,
0.060095421969890594,
-0.015447473153471947,
-0.007108474615961313,
0.013517280109226704,
-0.03868848830461502,
0.05028849467635155,
0.03871067240834236,
-0.022723045200109482,
0.007442883215844631,
-0.016520800068974495,
-0.027214981615543365,
0.06591064482927322,
-0.025884829461574554,
-0.001341317780315876,
-0.031266722828149796,
-0.02118690311908722,
-0.02936229296028614,
-0.06828548014163971,
0.004298079758882523,
-0.05637195706367493,
0.03515229746699333,
-0.037613097578287125,
-0.11635121703147888,
-0.07995954155921936,
-0.05138082057237625,
-0.07197984308004379,
0.04923940449953079,
-0.03776423633098602,
-0.01660175621509552,
0.035912055522203445,
0.048600077629089355,
0.050283387303352356,
-0.07889950275421143,
0.06163373962044716,
-0.06231599301099777,
-0.003203887026757002,
0.0009282657410949469,
0.050082821398973465,
0.020281292498111725,
-0.0077974204905331135,
-0.03270166739821434,
-0.08997723460197449,
-0.0014282921329140663,
0.01652936264872551,
0.00431794673204422,
-0.07742921262979507,
0.0097179701551795,
0.015940995886921883,
0.056547630578279495,
0.04969397559762001,
0.1590486764907837,
0.025767937302589417,
-0.05942796915769577,
0.012940854765474796,
0.025009796023368835,
0.026828454807400703,
0.0151597885414958,
0.022833626717329025,
0.01501201931387186,
-0.050631843507289886,
-0.0030061646830290556,
0.11365970224142075,
-0.01152726262807846,
0.039260249584913254,
0.07907267659902573,
-0.05615067854523659,
0.04991864040493965,
-0.11707841604948044,
0.008347230032086372,
0.02109246701002121,
0.027944808825850487,
-0.0860808715224266,
-0.030074240639805794,
0.0020975794177502394,
0.04194754734635353,
0.05320912227034569,
-0.017718665301799774,
0.009303178638219833,
0.05376967042684555,
0.04013166204094887,
0.017925895750522614,
0.08086813986301422,
0.06312664598226547,
0.01574036478996277,
0.04837711900472641,
0.01961694471538067,
-0.0441543310880661,
-0.014843097887933254,
0.013515991158783436,
2.2224875163035212e-33,
-0.002144685946404934,
0.03125551715493202,
0.01882491074502468,
0.02855323813855648,
0.005387295037508011,
-0.05028591677546501,
0.05428995192050934,
-0.03719896078109741,
-0.008651713840663433,
-0.028332754969596863,
0.016050266101956367,
-0.014557212591171265,
0.02136031910777092,
0.13837358355522156,
-0.007586429826915264,
-0.02711898274719715,
-0.02341429889202118,
0.11783944070339203,
-0.00391634926199913,
0.0004167184524703771,
0.023851072415709496,
0.026743963360786438,
-0.03847964107990265,
-0.04365020990371704,
0.026352526620030403,
0.026299338787794113,
-0.0003366899327374995,
-0.0002399469813099131,
-0.05650267004966736,
0.033494800329208374,
-0.055781081318855286,
-0.018021848052740097,
-0.027428658679127693,
-0.02706814743578434,
0.03500139340758324,
0.05126672983169556,
0.03925666958093643,
-0.08105826377868652,
0.027532726526260376,
-0.026910575106739998,
-0.03817487135529518,
0.02370905503630638,
0.04725728556513786,
0.06242392957210541,
-0.09363871067762375,
-0.08096400648355484,
-0.03670322895050049,
-0.03031074069440365,
-0.022369520738720894,
-0.09639183431863785,
0.03170798718929291,
0.027249624952673912,
-0.06633991748094559,
-0.05280892923474312,
-0.08375068753957748,
-0.06895293295383453,
-0.01809779182076454,
-0.011750136502087116,
0.03983624652028084,
0.05567661672830582,
-0.06858734041452408,
0.09696827083826065,
-0.056031741201877594,
0.031429920345544815,
-0.014949191361665726,
0.0698801800608635,
-0.0356760174036026,
-0.008689096197485924,
0.10792666673660278,
0.0057837688364088535,
-0.014280172064900398,
-0.0077900062315166,
-0.08798317611217499,
-0.05447057634592056,
-0.012914466671645641,
-0.016607841476798058,
0.1469831019639969,
0.016947753727436066,
0.0059959981590509415,
-0.03083466924726963,
0.008851838298141956,
-0.03181608021259308,
-0.061591390520334244,
-0.045147232711315155,
0.002961829537525773,
-0.009766780771315098,
0.006929733790457249,
-0.0011512511409819126,
-0.07192183285951614,
-0.08776263892650604,
-0.09746062010526657,
0.02487284131348133,
-0.046883586794137955,
-0.004540157038718462,
0.04575253650546074,
-2.8932817145613287e-33,
0.001019311253912747,
0.02928525023162365,
0.04532903805375099,
0.06669073551893234,
-0.03462726250290871,
-0.06905437260866165,
-0.01088915765285492,
-0.13801737129688263,
0.014130299910902977,
-0.04970379173755646,
-0.12319476902484894,
-0.022411661222577095,
0.03105759248137474,
-0.03288876637816429,
0.025877216830849648,
-0.016249699518084526,
-0.07585973292589188,
0.031155046075582504,
0.025103870779275894,
0.059667378664016724,
0.04124772548675537,
-0.020437344908714294,
-0.04750268906354904,
0.07847041636705399,
0.02671886421740055,
0.05135708302259445,
-0.06810787320137024,
0.08015935868024826,
0.041895218193531036,
-0.13899794220924377,
-0.0755186378955841,
-0.05743055045604706,
0.013682377524673939,
-0.003939592279493809,
0.0047081876546144485,
-0.006531342398375273,
0.03144235536456108,
0.09819870442152023,
-0.0046757725067436695,
0.1430305391550064,
0.03986561298370361,
-0.04208812490105629,
0.0024421652778983116,
0.04505423083901405,
-0.0306667760014534,
0.018583811819553375,
0.04215969517827034,
0.029609667137265205,
0.023067068308591843,
-0.031033959239721298,
0.010018683969974518,
0.08460500836372375,
-0.030051864683628082,
0.11001119017601013,
-0.01606155000627041,
-0.08788085728883743,
-0.02073945291340351,
-0.02658383548259735,
0.04084527865052223,
0.044106192886829376,
0.024787548929452896,
0.031081492081284523,
0.06230831891298294,
-0.08428098261356354,
0.06683482974767685,
-0.0791238397359848,
0.023953914642333984,
-0.040385689586400986,
-0.05780555680394173,
0.02166120707988739,
0.040150247514247894,
-0.05234699323773384,
0.06061793491244316,
0.02670709230005741,
0.005998542532324791,
-0.02726881392300129,
0.0664425790309906,
-0.04353739321231842,
0.0071222856640815735,
0.011338641867041588,
-0.05302164703607559,
0.07841572165489197,
-0.001905205543152988,
0.014316797256469727,
0.06066134572029114,
0.07055048644542694,
0.050452977418899536,
0.019665183499455452,
0.007179091684520245,
-0.058098405599594116,
0.004828011617064476,
0.0059442645870149136,
0.10687204450368881,
0.05344022065401077,
-0.08023006469011307,
-4.9158707327023876e-8,
0.0018232184229418635,
-0.017691241577267647,
-0.006992234382778406,
0.04060491546988487,
0.07652093470096588,
0.028579166159033775,
-0.04338500648736954,
-0.07330410927534103,
-0.0062693459913134575,
-0.08859783411026001,
0.09025976061820984,
-0.006209830287843943,
-0.06880434602499008,
-0.05000457167625427,
-0.026893744245171547,
-0.0038063013926148415,
-0.013747166842222214,
-0.020731767639517784,
-0.04786871373653412,
-0.012740065343677998,
-0.061687078326940536,
0.040135644376277924,
0.009211823344230652,
-0.09164212644100189,
0.029389595612883568,
0.035644832998514175,
0.061647653579711914,
0.0809018686413765,
0.10114014893770218,
0.028615809977054596,
-0.022124573588371277,
0.015088397078216076,
0.010903619229793549,
-0.0173204243183136,
0.04813402518630028,
0.003195498138666153,
0.009370414540171623,
-0.022480618208646774,
0.020670445635914803,
-0.05278514698147774,
0.08033469319343567,
0.029590563848614693,
-0.10460910946130753,
-0.01631757989525795,
0.02473331429064274,
-0.06254743039608002,
-0.0281322430819273,
-0.082399882376194,
-0.01701975055038929,
-0.03496720641851425,
0.008275824598968029,
-0.08556055277585983,
0.046643443405628204,
0.11020798981189728,
-0.009617535397410393,
0.043745558708906174,
-0.09174881130456924,
-0.025780612602829933,
0.026516323909163475,
-0.016242649406194687,
0.12834793329238892,
-0.035362374037504196,
-0.07836829870939255,
0.0417308583855629
] |
First-Person Fairness in Chatbots | https://openreview.net/forum?id=TlAdgeoDTo | [
"Tyna Eloundou",
"Alex Beutel",
"David G. Robinson",
"Keren Gu",
"Anna-Luisa Brakman",
"Pamela Mishkin",
"Meghan Shah",
"Johannes Heidecke",
"Lilian Weng",
"Adam Tauman Kalai"
] | Spotlight | Evaluating chatbot fairness is crucial given their rapid proliferation, yet typical chatbot tasks (e.g., resume writing, entertainment) diverge from the institutional decision-making tasks (e.g., resume screening) which have traditionally been central to discussion of algorithmic fairness. The open-ended nature and diverse use-cases of chatbots necessitate novel methods for bias assessment. This paper addresses these challenges by introducing a scalable counterfactual approach to evaluate "first-person fairness," meaning fairness toward chatbot users based on demographic characteristics. Our method employs a Language Model as a Research Assistant (LMRA) to yield quantitative measures of harmful stereotypes and qualitative analyses of demographic differences in chatbot responses. We apply this approach to assess biases in six of our language models across millions of interactions, covering sixty-six tasks in nine domains and spanning two genders and four races. Independent human annotations corroborate the LMRA-generated bias evaluations. This study represents the first large-scale fairness evaluation based on real-world chat data. We highlight that post-training reinforcement learning techniques significantly mitigate these biases. This evaluation provides a practical methodology for ongoing bias monitoring and mitigation. | fairness, large language models, chatbots | A methodology for evaluating bias in open-ended chat | 8,413 | 2410.19803 | [
-0.056226711720228195,
-0.060385603457689285,
0.025934576988220215,
0.034743569791316986,
0.0364236906170845,
0.009372963570058346,
0.009712114930152893,
-0.013857119716703892,
-0.001941031776368618,
-0.0199867095798254,
-0.0736301839351654,
-0.08077674359083176,
0.016557393595576286,
0.020182164385914803,
0.022824447602033615,
-0.006176518276333809,
0.10303476452827454,
-0.030685897916555405,
-0.05073441192507744,
-0.02499370463192463,
0.04764637351036072,
0.030734682455658913,
0.10491552203893661,
-0.009295538067817688,
-0.0345739983022213,
-0.1131056398153305,
-0.009457895532250404,
-0.042825307697057724,
-0.03654683753848076,
0.030297866091132164,
0.03847835585474968,
0.009713718667626381,
0.051733437925577164,
0.057121727615594864,
-0.0340677835047245,
0.009866921231150627,
-0.06316649913787842,
-0.019112229347229004,
0.013673992827534676,
-0.010437713004648685,
-0.07837732136249542,
-0.07649169862270355,
-0.0025196068454533815,
-0.02327682077884674,
0.06884414702653885,
-0.03891099989414215,
0.0011110069463029504,
-0.002321552485227585,
-0.038312602788209915,
0.0032299573067575693,
-0.0012708886060863733,
0.029260648414492607,
0.09766767919063568,
0.09958646446466446,
-0.06203199550509453,
0.02526976354420185,
0.007590608671307564,
0.07754681259393692,
0.06636634469032288,
0.009513754397630692,
-0.09432104229927063,
-0.09985334426164627,
-0.04403600096702576,
0.020250340923666954,
-0.03765811398625374,
-0.007804994937032461,
-0.047589246183633804,
0.0179989505559206,
-0.051196616142988205,
0.018712513148784637,
0.018256323412060738,
0.06263566762208939,
-0.01874106377363205,
0.0947975367307663,
0.08393578231334686,
0.01624554954469204,
0.0366484709084034,
-0.05607954040169716,
0.07191048562526703,
-0.026119226589798927,
-0.024669701233506203,
-0.04304629564285278,
0.06602600961923599,
0.07010596990585327,
0.0430469885468483,
0.01963498629629612,
0.006101893726736307,
0.04295658320188522,
-0.048267293721437454,
0.055812228471040726,
-0.03326176851987839,
-0.00026740197790786624,
0.11222577840089798,
0.051436085253953934,
0.021452635526657104,
0.01855417899787426,
-0.0159629937261343,
0.0025459600146859884,
-0.009986800141632557,
0.06704559922218323,
0.01150030829012394,
-0.016463100910186768,
-0.04537304490804672,
0.01467756275087595,
0.028048120439052582,
-0.009125694632530212,
0.06708044558763504,
-0.011723463423550129,
-0.047470808029174805,
-0.01706871949136257,
-0.024001983925700188,
-0.04149765893816948,
0.03250293806195259,
0.0668531209230423,
0.07289271056652069,
-0.005007695406675339,
0.08361542969942093,
0.009223890490829945,
-0.002077842829748988,
0.05687665939331055,
0.08977506309747696,
-0.021574968472123146,
0.09122022986412048,
0.001919388654641807,
0.025843236595392227,
0.008051808923482895,
-0.033715635538101196,
-2.8323024400171937e-34,
0.07617545872926712,
0.013665412552654743,
-0.00776613550260663,
0.030991462990641594,
0.02597917802631855,
-0.01805194281041622,
-0.030764495953917503,
0.08795472234487534,
-0.03242829442024231,
-0.012335320934653282,
-0.030343851074576378,
0.05836118012666702,
-0.08062570542097092,
0.029437104240059853,
0.0236430075019598,
0.03615043684840202,
-0.05787355825304985,
0.10943634063005447,
-0.018830209970474243,
0.011649826541543007,
0.06767532974481583,
0.00014950688637327403,
0.03105962462723255,
0.06800912320613861,
-0.011700429953634739,
0.030933573842048645,
0.04465433210134506,
-0.04751516506075859,
0.018204323947429657,
0.00752616161480546,
-0.11787179112434387,
-0.03551172837615013,
0.020516633987426758,
-0.006079787854105234,
0.04135088250041008,
-0.018938373774290085,
-0.050743937492370605,
0.033136606216430664,
-0.035281725227832794,
0.04642581567168236,
-0.06271152198314667,
0.0017683515325188637,
0.04700149968266487,
-0.06330753117799759,
-0.02564002387225628,
-0.1062195673584938,
-0.007821540348231792,
-0.06733749061822891,
0.005539958365261555,
0.05239510163664818,
0.019333133473992348,
0.07859288901090622,
0.06037166714668274,
0.03529296815395355,
-0.056527234613895416,
-0.06224307045340538,
0.02144256979227066,
0.021785419434309006,
0.011694339103996754,
-0.06490238755941391,
0.021155277267098427,
-0.0010026072850450873,
-0.023664629086852074,
0.01575431600213051,
0.005483607761561871,
-0.053077273070812225,
0.005196726880967617,
0.02322872169315815,
0.010297444649040699,
0.0034484807401895523,
0.05819728597998619,
0.054686639457941055,
-0.029697975143790245,
0.026043903082609177,
-0.05949850007891655,
0.08387748897075653,
0.036506764590740204,
0.015853669494390488,
0.04019523784518242,
-0.04542740434408188,
0.08659825474023819,
-0.028644073754549026,
-0.08539048582315445,
-0.1349196434020996,
-0.0683530941605568,
-0.12132945656776428,
-0.0005388605641201138,
-0.07503210753202438,
0.0573708601295948,
0.04885292053222656,
-0.061616528779268265,
-0.02807774394750595,
-0.04293390363454819,
0.026206061244010925,
-0.019083494320511818,
-2.448791897931653e-34,
-0.06180549040436745,
-0.019195949658751488,
-0.039750512689352036,
0.04161681979894638,
-0.0030420292168855667,
-0.06227504462003708,
0.0737411230802536,
0.006380914710462093,
0.07252322882413864,
0.011431201361119747,
-0.05353002995252609,
-0.0438118502497673,
0.09764238446950912,
0.01867031864821911,
0.09612946957349777,
-0.055941324681043625,
-0.011190474033355713,
0.0245262011885643,
-0.016469676047563553,
-0.0075093647465109825,
0.04407314956188202,
0.08487961441278458,
-0.10655892640352249,
0.012443293817341328,
0.003641568822786212,
0.002649158239364624,
-0.014346480369567871,
-0.05011637136340141,
0.022069497033953667,
0.017139440402388573,
-0.0636632889509201,
0.01814230903983116,
-0.07030981034040451,
0.003555148374289274,
0.018571501597762108,
-0.03334743157029152,
0.03854703903198242,
-0.02518540993332863,
-0.013152428902685642,
0.01846816949546337,
0.08068454265594482,
-0.02880892902612686,
-0.175988107919693,
-0.04133449122309685,
0.000341794133419171,
-0.015448409132659435,
-0.0650680884718895,
-0.06638769805431366,
0.007881414145231247,
-0.0655675083398819,
-0.033262211829423904,
-0.017512068152427673,
0.032220788300037384,
0.004427051171660423,
-0.055228106677532196,
-0.16919249296188354,
0.04851324483752251,
-0.01486394926905632,
0.05504880100488663,
0.036764055490493774,
-0.05017542466521263,
-0.023984460160136223,
-0.005266543012112379,
0.0007073397864587605,
-0.036391984671354294,
-0.11385337263345718,
-0.009911905974149704,
0.014813440851867199,
-0.029639557003974915,
-0.0314280167222023,
0.12681414186954498,
-0.05149317905306816,
-0.020024042576551437,
0.005347045138478279,
-0.045447517186403275,
0.011956539005041122,
0.03288573771715164,
-0.0031715307850390673,
-0.0029373790603131056,
0.0013295997632667422,
0.025701018050312996,
-0.019663922488689423,
0.08937739580869675,
0.012450640089809895,
0.01677515357732773,
0.11400272697210312,
0.004278797190636396,
0.08119619637727737,
-0.036360111087560654,
0.022544587031006813,
-0.0254236850887537,
-0.025141814723610878,
-0.037892576307058334,
0.02343131974339485,
-0.1274012327194214,
-3.545012461358965e-8,
-0.01656310074031353,
-0.030507076531648636,
0.021133773028850555,
0.13646963238716125,
-0.04282969981431961,
-0.01625479944050312,
-0.07215546071529388,
-0.0017069129971787333,
0.03306138515472412,
-0.012719418853521347,
0.05197152495384216,
0.027902478352189064,
0.04502509906888008,
-0.07900289446115494,
0.08905436098575592,
0.07588858157396317,
-0.016377920284867287,
0.03203912451863289,
0.02085406333208084,
0.02913457341492176,
0.09946291148662567,
0.037080343812704086,
-0.029654288664460182,
-0.04680747911334038,
-0.055136509239673615,
-0.0747920349240303,
-0.05242026224732399,
0.08481132984161377,
-0.024827811866998672,
-0.021317342296242714,
-0.04247364029288292,
0.026516469195485115,
-0.02254517935216427,
-0.10095985233783722,
0.08825128525495529,
0.032701198011636734,
-0.04887651279568672,
-0.06210353225469589,
0.008735686540603638,
0.08263532817363739,
0.03687179088592529,
0.08332245796918869,
-0.009181245230138302,
-0.04275944083929062,
0.029855048283934593,
0.006995729636400938,
-0.07156165689229965,
-0.024645399302244186,
-0.035147301852703094,
-0.06116439774632454,
-0.043199557811021805,
-0.04600571095943451,
0.014220994897186756,
0.05156237632036209,
0.05228688567876816,
-0.03938300535082817,
0.019857948645949364,
0.0032067103311419487,
0.020002225413918495,
0.05622154101729393,
0.12521113455295563,
0.07130202651023865,
-0.04658099263906479,
-0.03880317509174347
] |
Can Large Language Models Understand Symbolic Graphics Programs? | https://openreview.net/forum?id=Yk87CwhBDx | [
"Zeju Qiu",
"Weiyang Liu",
"Haiwen Feng",
"Zhen Liu",
"Tim Z. Xiao",
"Katherine M. Collins",
"Joshua B. Tenenbaum",
"Adrian Weller",
"Michael J. Black",
"Bernhard Schölkopf"
] | Spotlight | Against the backdrop of enthusiasm for large language models (LLMs), there is a growing need to scientifically assess their capabilities and shortcomings. This is nontrivial in part because it is difficult to find tasks which the models have not encountered during training. Utilizing symbolic graphics programs, we propose a domain well-suited to test multiple spatial-semantic reasoning skills of LLMs. Popular in computer graphics, these programs procedurally generate visual data. While LLMs exhibit impressive skills in general program synthesis and analysis, symbolic graphics programs offer a new layer of evaluation: they allow us to test an LLM's ability to answer semantic questions about the images or 3D geometries without a vision encoder. To semantically understand the symbolic programs, LLMs would need to possess the ability to "imagine" and reason how the corresponding graphics content would look with only the symbolic description of the local curvatures and strokes. We use this task to evaluate LLMs by creating a large benchmark for the semantic visual understanding of symbolic graphics programs, built procedurally with minimal human effort. Particular emphasis is placed on transformations of images that leave the image level semantics invariant while introducing significant changes to the underlying program. We evaluate commercial and open-source LLMs on our benchmark to assess their ability to reason about visual output of programs, finding that LLMs considered stronger at reasoning generally perform better. Lastly, we introduce a novel method to improve this ability -- Symbolic Instruction Tuning (SIT), in which the LLM is finetuned with pre-collected instruction data on symbolic graphics programs. Interestingly, we find that SIT not only improves LLM's understanding on symbolic programs, but it also improves general reasoning ability on various other benchmarks. | Large Language Models, Symbolic Graphics Programs | Assessing Large Language Model Reasoning over Symbolic Graphics Programs | 8,410 | 2408.08313 | [
0.03035205602645874,
-0.05933142080903053,
0.006488444283604622,
-0.04769665747880936,
0.04832131415605545,
-0.09393693506717682,
-0.04279973730444908,
0.013708560727536678,
-0.009211640805006027,
-0.015934759750962257,
-0.06728764623403549,
-0.1070595309138298,
0.0467475950717926,
0.09021919965744019,
0.04043571650981903,
0.03140627220273018,
0.029907606542110443,
0.04777565971016884,
-0.04229508340358734,
-0.021692492067813873,
0.03322939947247505,
-0.07047691941261292,
-0.009375996887683868,
-0.012870714999735355,
0.06511243432760239,
0.05151620879769325,
0.03816470503807068,
-0.019986309111118317,
0.00699175801128149,
-0.017048699781298637,
0.013937151059508324,
0.051958464086055756,
-0.02192840911448002,
0.05979296565055847,
0.006674450356513262,
0.09617144614458084,
-0.007801742758601904,
0.002512914827093482,
0.007083551026880741,
-0.07935933023691177,
-0.10735826939344406,
0.019067781046032906,
0.04723460227251053,
-0.004905507434159517,
0.09048411250114441,
-0.005197024438530207,
-0.017978208139538765,
-0.03366900607943535,
-0.04083156958222389,
0.020872412249445915,
-0.11488763242959976,
-0.11793283373117447,
-0.092770054936409,
-0.06855932623147964,
-0.016765180975198746,
0.05047565698623657,
0.12054960429668427,
-0.03930983692407608,
-0.01709870435297489,
-0.04118267819285393,
-0.04573492333292961,
-0.03407410904765129,
-0.01157973799854517,
0.05349072441458702,
-0.03316863253712654,
-0.03152059018611908,
0.0519103966653347,
-0.03275679051876068,
0.06857077032327652,
-0.04812900349497795,
-0.09194764494895935,
0.021668342873454094,
-0.03732134401798248,
-0.0006431436049751937,
-0.027866553515195847,
0.03519928827881813,
-0.0029087055008858442,
-0.045992299914360046,
0.010702081955969334,
-0.1610974222421646,
0.035126470029354095,
0.0363861583173275,
0.008011015132069588,
0.0767071321606636,
0.05974499508738518,
0.06369222700595856,
0.0047726682387292385,
0.06786029785871506,
0.0379025861620903,
0.06416913121938705,
0.01786857284605503,
-0.03508499264717102,
0.0034644952975213528,
0.03170636296272278,
0.09378144890069962,
0.04046407714486122,
0.04832234978675842,
-0.1200009360909462,
-0.025977469980716705,
0.0793830081820488,
0.024383116513490677,
-0.000446868158178404,
0.07335548847913742,
-0.05481232330203056,
0.02360251359641552,
0.01227273978292942,
0.06108958274126053,
-0.07015427201986313,
0.0059809451922774315,
-0.07550808787345886,
0.055307019501924515,
0.021996259689331055,
-0.01173537876456976,
0.02502022683620453,
-0.05110699683427811,
-0.0013115081237629056,
0.0004663975560106337,
-0.03636742755770683,
0.025460632517933846,
-0.005367584992200136,
-0.008301363326609135,
0.03662602975964546,
-0.005855323746800423,
0.04949609935283661,
0.0778997465968132,
-0.06864307820796967,
-0.07844220101833344,
-2.1376764254111586e-34,
0.04182396084070206,
0.03843710571527481,
0.06669731438159943,
0.05425151437520981,
0.07663662731647491,
-0.008872762322425842,
0.025355743244290352,
0.03116041049361229,
-0.01347309909760952,
0.04016231745481491,
-0.05994749814271927,
0.033410776406526566,
-0.02446659840643406,
0.13386473059654236,
0.0617198720574379,
0.010906863026320934,
-0.004477550275623798,
0.02564958482980728,
-0.1140737310051918,
0.00543496198952198,
-0.016007238999009132,
0.018756648525595665,
-0.01330520585179329,
-0.04156753793358803,
-0.06093670427799225,
0.06461203098297119,
-0.0077971494756639,
-0.022434549406170845,
-0.006562912371009588,
-0.018773797899484634,
-0.0631551444530487,
-0.04418254643678665,
-0.03309479355812073,
0.0888592079281807,
0.05657445266842842,
-0.06177125126123428,
-0.0006526302313432097,
-0.04859720543026924,
0.03307848423719406,
0.0034485089126974344,
0.01585289090871811,
0.039985865354537964,
-0.027642812579870224,
0.03331233933568001,
-0.05261564254760742,
0.019526507705450058,
-0.01315310038626194,
0.030877763405442238,
-0.09779722988605499,
-0.01800461858510971,
0.06248851493000984,
0.059272777289152145,
-0.0256644394248724,
-0.017908165231347084,
-0.03770992159843445,
-0.026777692139148712,
0.02321966178715229,
-0.043646398931741714,
-0.015809128060936928,
0.029809553176164627,
0.00030882793362252414,
0.052584074437618256,
0.02247164212167263,
0.07904165983200073,
-0.051921527832746506,
0.0017056064680218697,
-0.003785234410315752,
0.047859299927949905,
0.07585808634757996,
0.054150279611349106,
-0.065212681889534,
0.019714368507266045,
0.0062465057708323,
-0.05406852811574936,
0.005291431676596403,
-0.026086878031492233,
0.032737847417593,
-0.0683065801858902,
-0.0010196455987170339,
0.033358894288539886,
-0.007229190319776535,
-0.03827016055583954,
-0.050300054252147675,
-0.1202123835682869,
-0.0036760764196515083,
-0.06011404097080231,
0.04902741685509682,
-0.1414998471736908,
0.0033851477783173323,
-0.09236054867506027,
-0.09947162121534348,
-0.05649052932858467,
-0.0267577413469553,
-0.0362333245575428,
0.012104387395083904,
-1.805453161029475e-33,
-0.06110267713665962,
-0.014934775419533253,
0.0032952993642538786,
0.07380547374486923,
-0.07794741541147232,
-0.07481853663921356,
0.04187933728098869,
0.02545689046382904,
-0.054281704127788544,
-0.05834117531776428,
-0.029539229348301888,
0.02440398372709751,
-0.01277540810406208,
-0.0102576008066535,
-0.009453886188566685,
-0.056457217782735825,
-0.022306371480226517,
-0.02818943001329899,
-0.061714574694633484,
0.07263704389333725,
0.0001922555675264448,
0.1652071624994278,
-0.10793814063072205,
-0.016730917617678642,
-0.0984584391117096,
0.09277114272117615,
0.008578951470553875,
-0.05623539909720421,
-0.0361558198928833,
0.08519455790519714,
0.03380695357918739,
-0.03005099855363369,
-0.029840314760804176,
-0.008355697616934776,
0.006830292288213968,
-0.023032784461975098,
0.09144573658704758,
-0.055767372250556946,
-0.065559983253479,
0.07324652373790741,
0.045372698456048965,
-0.023630734533071518,
0.017034608870744705,
0.034112658351659775,
-0.045329753309488297,
0.042947299778461456,
-0.049635306000709534,
0.018454659730196,
-0.018923325464129448,
0.03731515258550644,
-0.022567501291632652,
-0.01137443445622921,
0.022171733900904655,
-0.04263168200850487,
0.027494730427861214,
-0.10814138501882553,
-0.061550039798021317,
-0.00862678699195385,
0.018149109557271004,
-0.0055892751552164555,
-0.07628442347049713,
-0.06005405634641647,
0.013880654238164425,
-0.048595476895570755,
-0.012050092220306396,
0.013980266638100147,
-0.08846606314182281,
-0.055721357464790344,
0.008171550929546356,
0.006499890703707933,
0.06885138154029846,
-0.008006349205970764,
-0.07147207856178284,
0.04901103675365448,
0.04803721606731415,
0.015579712577164173,
0.04267844557762146,
0.001053023966960609,
0.006912153214216232,
-0.02763836830854416,
0.060883354395627975,
-0.06584500521421432,
0.05656973272562027,
0.09700850397348404,
0.023688020184636116,
0.06568247824907303,
-0.04131513833999634,
0.02734897844493389,
0.06053551286458969,
0.03591904416680336,
-0.009154273197054863,
0.10727015882730484,
0.030699264258146286,
0.08236624300479889,
-0.007388690020889044,
-4.0784819077543943e-8,
-0.06484942883253098,
0.014306594617664814,
0.024707550182938576,
-0.012969094328582287,
-0.025175567716360092,
0.021834900602698326,
-0.013451323844492435,
0.04976038634777069,
-0.0565304309129715,
-0.036452073603868484,
0.08635637909173965,
-0.011853629723191261,
-0.06615912914276123,
0.004436148796230555,
0.04520060494542122,
0.07634640485048294,
0.04207776114344597,
0.01492720004171133,
-0.004173098132014275,
0.03544751554727554,
0.004774824716150761,
0.04390089213848114,
-0.026691526174545288,
0.021954327821731567,
-0.04179872199892998,
-0.05728556588292122,
-0.05558081343770027,
0.10360094904899597,
0.05349278450012207,
-0.017859110608696938,
0.08876661956310272,
-0.01429822389036417,
0.07012820988893509,
0.025846226140856743,
0.04777436703443527,
0.0036642574705183506,
-0.055520545691251755,
-0.008639859966933727,
0.01405508816242218,
-0.03772900626063347,
0.010433352552354336,
0.05458815023303032,
0.01926514133810997,
0.011231479234993458,
0.04886287450790405,
-0.01934860274195671,
0.0016278540715575218,
-0.05939323827624321,
-0.033309753984212875,
0.03154437616467476,
-0.10683030635118484,
0.022861650213599205,
-0.08224665373563766,
0.06924714893102646,
0.00675880117341876,
0.005097911227494478,
-0.06328555941581726,
0.053959839046001434,
-0.009564782492816448,
0.10523365437984467,
-0.024166453629732132,
0.027409816160798073,
-0.023019736632704735,
-0.06751056015491486
] |
Token Statistics Transformer: Linear-Time Attention via Variational Rate Reduction | https://openreview.net/forum?id=lXRDQsiP2v | [
"Ziyang Wu",
"Tianjiao Ding",
"Yifu Lu",
"Druv Pai",
"Jingyuan Zhang",
"Weida Wang",
"Yaodong Yu",
"Yi Ma",
"Benjamin David Haeffele"
] | Spotlight | The attention operator is arguably the key distinguishing factor of transformer architectures, which have demonstrated state-of-the-art performance on a variety of tasks. However, transformer attention operators often impose a significant computational burden, with the computational complexity scaling quadratically with the number of tokens. In this work, we propose a novel transformer attention operator whose computational complexity scales linearly with the number of tokens. We derive our network architecture by extending prior work which has shown that a transformer style architecture naturally arises by "white-box" architecture design, where each layer of the network is designed to implement an incremental optimization step of a maximal coding rate reduction objective (MCR$^2$). Specifically, we derive a novel variational form of the MCR$^2$ objective and show that the architecture that results from unrolled gradient descent of this variational objective leads to a new attention module called Token Statistics Self-Attention ($\texttt{TSSA}$). $\texttt{TSSA}$ has $\textit{linear computational and memory complexity}$ and radically departs from the typical attention architecture that computes pairwise similarities between tokens. Experiments on vision, language, and long sequence tasks show that simply swapping $\texttt{TSSA}$ for standard self-attention, which we refer to as the Token Statistics Transformer ($\texttt{ToST}$), achieves competitive performance with conventional transformers while being significantly more computationally efficient and interpretable. Our results also somewhat call into question the conventional wisdom that pairwise similarity style attention mechanisms are critical to the success of transformer architectures. | white-box deep neural networks, representation learning, transformer | We build a novel transformer-like architecture with linear-complexity attention by algorithmic unrolling of a compression-based learning objective. | 8,362 | 2412.17810 | [
-0.007524509914219379,
-0.019911162555217743,
0.030390918254852295,
-0.0038300452288240194,
-0.012592889368534088,
-0.014641251415014267,
-0.047568414360284805,
-0.0019687169697135687,
0.07021129131317139,
-0.018089599907398224,
-0.040203120559453964,
0.02358938194811344,
0.00991315022110939,
0.022794457152485847,
-0.0007690974744036794,
0.045878615230321884,
0.06707587093114853,
0.11137843877077103,
-0.1059308722615242,
-0.07207399606704712,
0.03411662206053734,
-0.010563206858932972,
0.008339319378137589,
-0.044882163405418396,
0.060859717428684235,
-0.00023840999347157776,
-0.0169341042637825,
-0.10855512320995331,
0.044389110058546066,
0.02078787237405777,
0.05536462739109993,
-0.02361600659787655,
0.03965023532509804,
0.046590425074100494,
-0.11064831167459488,
0.013503636233508587,
-0.035807084292173386,
-0.033771026879549026,
0.02124636434018612,
-0.03202001377940178,
0.027774019166827202,
0.04019808769226074,
-0.007718426641076803,
0.02397695556282997,
0.03973871469497681,
0.032521896064281464,
0.09493698924779892,
-0.07876298576593399,
-0.00611510407179594,
-0.05970311164855957,
0.015423036180436611,
-0.03361747786402702,
-0.07696530967950821,
0.02215537056326866,
-0.004451252985745668,
0.01994829624891281,
-0.022125791758298874,
0.04124218225479126,
0.0054242429323494434,
0.011324035003781319,
-0.07126054167747498,
-0.07135175168514252,
0.03223466873168945,
-0.044842034578323364,
-0.01635082997381687,
-0.029216405004262924,
0.09499740600585938,
-0.044581763446331024,
0.011646676808595657,
-0.013043810613453388,
-0.039030008018016815,
0.07129465043544769,
-0.046847153455019,
0.07660632580518723,
0.07654941827058792,
0.0029802711214870214,
0.11773351579904556,
-0.021050486713647842,
0.07412950694561005,
-0.04673754423856735,
0.046616118401288986,
-0.02964012511074543,
0.07433062046766281,
0.037513263523578644,
0.1319519430398941,
-0.058090612292289734,
-0.017936086282134056,
0.022058583796024323,
-0.011250617913901806,
0.0107492133975029,
-0.09111694991588593,
-0.0733565092086792,
-0.003681164002045989,
-0.07102076709270477,
0.009430821985006332,
-0.05684806779026985,
-0.06993760913610458,
-0.005695330444723368,
-0.029289091005921364,
0.08189775049686432,
-0.017396997660398483,
-0.0464947335422039,
0.05727847293019295,
0.004346013069152832,
0.04527904838323593,
-0.04719916731119156,
0.09391827881336212,
0.0988842025399208,
0.0030263527296483517,
0.007407786324620247,
0.036263011395931244,
0.04899992048740387,
-0.022566840052604675,
-0.04995293915271759,
0.04325895011425018,
-0.01706756465137005,
-0.009502808563411236,
0.014253389090299606,
0.041477739810943604,
0.07357103377580643,
0.0519186407327652,
-0.022400476038455963,
-0.07636694610118866,
0.02447887323796749,
0.003979455213993788,
0.00685418164357543,
-0.022241372615098953,
2.4324691945977758e-33,
-0.0648529902100563,
0.07108791917562485,
0.02294817939400673,
-0.03551873937249184,
0.0889093354344368,
-0.003035066882148385,
0.03613870590925217,
-0.01095642987638712,
-0.013580928556621075,
0.004011048935353756,
0.009311988949775696,
0.02707749232649803,
-0.05671599507331848,
0.09139729291200638,
0.039824143052101135,
-0.07654855400323868,
-0.0184091217815876,
0.01722242683172226,
0.03581508621573448,
-0.12981925904750824,
0.0069648795761168,
-0.0030443519353866577,
0.01457796897739172,
0.014999082311987877,
-0.03152989223599434,
-0.06474930793046951,
0.08280845731496811,
-0.13804084062576294,
-0.013151573948562145,
-0.023830007761716843,
-0.11466862261295319,
0.02340761199593544,
-0.06024909391999245,
-0.0011138340923935175,
0.07124920934438705,
-0.08393973857164383,
0.037974439561367035,
-0.05907741189002991,
-0.004205025732517242,
0.018882494419813156,
-0.03273198381066322,
0.07633309066295624,
-0.04124797508120537,
-0.015405996702611446,
-0.09207969903945923,
-0.1124889925122261,
0.01361661683768034,
0.03149080276489258,
0.00817872490733862,
-0.05953178554773331,
0.05913732573390007,
-0.02873973920941353,
-0.031317416578531265,
0.023859474807977676,
0.09662631154060364,
-0.018441639840602875,
0.050715312361717224,
0.04910294711589813,
0.062279049307107925,
0.12973061203956604,
0.0027004883158951998,
0.013100082986056805,
0.013417846523225307,
0.0785791203379631,
0.09329220652580261,
0.013432026840746403,
-0.019816994667053223,
0.044706813991069794,
0.0029932865872979164,
-0.015264857560396194,
-0.0031095207668840885,
0.07998430728912354,
-0.001655528205446899,
-0.07148823887109756,
0.020731383934617043,
0.049016084522008896,
0.0979292094707489,
-0.06334681808948517,
-0.05828534811735153,
-0.013798014260828495,
-0.03321067616343498,
0.04196736961603165,
-0.03454718366265297,
-0.001457864884287119,
-0.03475644439458847,
0.03390365466475487,
0.03236348181962967,
-0.05119607225060463,
-0.025395894423127174,
-0.0019026785157620907,
-0.02035614848136902,
-0.027646398171782494,
-0.028550460934638977,
-0.0007115051266737282,
-0.07032544165849686,
-1.3117798801596773e-33,
-0.061138443648815155,
0.04297741502523422,
-0.03794095665216446,
0.04496043547987938,
-0.028151964768767357,
-0.06184271350502968,
0.005333427805453539,
-0.01970251277089119,
0.01308108028024435,
0.022053010761737823,
-0.039650335907936096,
-0.08084487169981003,
0.033890996128320694,
-0.03470749035477638,
0.06238320842385292,
-0.02079908363521099,
-0.053634434938430786,
0.07880516350269318,
-0.06671059131622314,
-0.00431212130934,
-0.031241804361343384,
0.1423352062702179,
-0.041368268430233,
-0.03829570114612579,
-0.07473240047693253,
0.033342886716127396,
-0.11774609982967377,
0.10698123276233673,
0.073697030544281,
-0.07411709427833557,
-0.08826766908168793,
0.011987640522420406,
0.009610900655388832,
-0.06003126874566078,
-0.04703497141599655,
0.06497995555400848,
0.01110861450433731,
0.059262026101350784,
-0.029805146157741547,
0.07658807933330536,
0.04367448017001152,
-0.06720039993524551,
-0.02764812484383583,
0.030951282009482384,
-0.0014923468697816133,
-0.004038254264742136,
-0.021420111879706383,
-0.019043361768126488,
-0.03136714547872543,
0.04142875596880913,
-0.002583131194114685,
0.015702832490205765,
-0.05877937749028206,
0.03748229518532753,
-0.05625038221478462,
-0.05177517235279083,
0.009147311560809612,
-0.020595375448465347,
0.1392839252948761,
-0.022739438340067863,
0.007578482385724783,
-0.012174747884273529,
0.038489390164613724,
-0.07954727113246918,
0.02514730393886566,
-0.052091456949710846,
0.04648036137223244,
-0.02634691447019577,
0.019077572971582413,
0.05078485980629921,
0.052939143031835556,
0.01715978980064392,
0.018626747652888298,
0.004892374854534864,
0.029597226530313492,
0.006321343593299389,
0.07355796545743942,
-0.03210749104619026,
-0.042652130126953125,
-0.028273912146687508,
-0.03020142763853073,
0.02874099463224411,
-0.0112025560811162,
-0.06564271450042725,
0.035111621022224426,
0.052428096532821655,
0.09521259367465973,
0.013641982339322567,
0.07274572551250458,
0.0010739516001194715,
-0.023051658645272255,
0.042801860719919205,
0.020924517884850502,
0.049408577382564545,
-0.05401260405778885,
-4.712068246703893e-8,
-0.08851484954357147,
-0.022786688059568405,
-0.08973027020692825,
-0.02468065358698368,
0.06627412140369415,
-0.0699579194188118,
0.021508418023586273,
0.12245742976665497,
-0.03181014209985733,
-0.005237501114606857,
0.10008314996957779,
-0.03169215843081474,
-0.045185234397649765,
-0.07138300687074661,
0.006073981989175081,
0.016263920813798904,
0.0327720083296299,
-0.031653910875320435,
0.037347063422203064,
-0.03944579139351845,
-0.055617719888687134,
0.041380688548088074,
0.013109512627124786,
-0.056488897651433945,
0.010469546541571617,
-0.06626903265714645,
0.01557762548327446,
0.0577489472925663,
-0.0016169995069503784,
-0.012211139313876629,
-0.029607122763991356,
0.09551621973514557,
0.032705068588256836,
-0.042878493666648865,
-0.042003292590379715,
0.04195728152990341,
0.09181419759988785,
0.030681485310196877,
-0.03704184666275978,
-0.013943290337920189,
0.013463301584124565,
0.028598101809620857,
-0.04509511590003967,
0.09279219806194305,
-0.0016469553811475635,
-0.02024722471833229,
0.01364525780081749,
-0.04534938931465149,
0.0015066653722897172,
-0.08074235171079636,
0.09928102046251297,
0.0006902319146320224,
-0.01900605671107769,
0.033398229628801346,
-0.007105947006493807,
-0.014364433474838734,
0.05355185270309448,
-0.05629527568817139,
-0.02695746161043644,
0.08175591379404068,
0.02793230675160885,
0.03680330887436867,
-0.08709385991096497,
-0.043547775596380234
] |
Nonlinear Sequence Embedding by Monotone Variational Inequality | https://openreview.net/forum?id=U834XHJuqk | [
"Jonathan Yuyang Zhou",
"Yao Xie"
] | Spotlight | In the wild, we often encounter collections of sequential data such as electrocardiograms, motion capture, genomes, and natural language, and sequences may be multichannel or symbolic with nonlinear dynamics. We introduce a method to learn low-dimensional representations of nonlinear sequence and time-series data without supervision which has provable recovery guarantees. The learned representation can be used for downstream machine-learning tasks such as clustering and classification. The method assumes that the observed sequences arise from a common domain, with each sequence following its own autoregressive model, and these models are related through low-rank regularization. We cast the problem as a convex matrix parameter recovery problem using monotone variational inequalities (VIs) and encode the common domain assumption via low-rank constraint across the learned representations, which can learn a subspace approximately spanning the entire domain as well as faithful representations for the dynamics of each individual sequence incorporating the domain information in totality. We show the competitive performance of our method on real-world time-series data with baselines and demonstrate its effectiveness for symbolic text modeling and RNA sequence clustering. | Monotone Variational Inequality, Convex Optimization, Sequence Data, Time Series, Representation Learning | We introduce a convex method to learn low dimensional representations of nonlinear time-series and sequences with recovery guarantees. | 8,337 | null | [
-0.06517373770475388,
-0.11693039536476135,
0.02275627665221691,
0.042754270136356354,
0.018652131780982018,
0.0877254530787468,
-0.05561217665672302,
-0.03725790977478027,
0.03138462081551552,
-0.06010860949754715,
-0.013922516256570816,
0.0014159068232402205,
-0.008475780487060547,
0.08949538320302963,
-0.07377266883850098,
0.01238697487860918,
0.022477947175502777,
0.016453715041279793,
-0.026158686727285385,
-0.013054415583610535,
0.02423509582877159,
0.07936321943998337,
0.051642272621393204,
0.07102607935667038,
0.0049286456778645515,
-0.011164293624460697,
0.08935266733169556,
0.0914604514837265,
-0.009315784089267254,
0.029608426615595818,
0.04847836121916771,
0.02395995520055294,
0.0256502628326416,
-0.02697562985122204,
-0.0395718477666378,
0.14303842186927795,
-0.06646142154932022,
0.033696502447128296,
0.006489569786936045,
0.01724669709801674,
0.07122547924518585,
0.0652787834405899,
0.0029967273585498333,
0.055941496044397354,
0.07233072817325592,
-0.03314458206295967,
0.005170270800590515,
-0.11659438163042068,
-0.025832677260041237,
0.0068453773856163025,
-0.018912428990006447,
0.043896038085222244,
-0.034645915031433105,
0.08861076086759567,
-0.01163135003298521,
-0.01886075921356678,
-0.014769772067666054,
-0.03264741227030754,
0.02321617677807808,
-0.06671144813299179,
0.04689443111419678,
-0.07894743978977203,
0.013274863362312317,
-0.08349159359931946,
-0.021913306787610054,
-0.03407485410571098,
0.01970931701362133,
0.12569954991340637,
-0.023477794602513313,
0.06043049693107605,
-0.06779094785451889,
0.0369202196598053,
-0.05289245769381523,
0.06362032890319824,
-0.0048336852341890335,
0.040487486869096756,
0.03619728237390518,
-0.02631276473402977,
0.05345002934336662,
-0.012671304866671562,
0.029425354674458504,
-0.017597367987036705,
-0.006185884587466717,
0.035697560757398605,
0.04432714730501175,
0.058575015515089035,
0.05672312155365944,
0.037708912044763565,
0.027038004249334335,
-0.01212447788566351,
-0.03847457841038704,
-0.040621548891067505,
0.09788155555725098,
0.020537620410323143,
0.033892739564180374,
0.03137553483247757,
0.04055405780673027,
-0.005392951425164938,
0.08163964748382568,
0.09253210574388504,
0.040006380528211594,
0.06696000695228577,
0.046716704964637756,
0.09237273037433624,
0.037920672446489334,
-0.05594755336642265,
0.05355514585971832,
0.023279987275600433,
0.03357941284775734,
-0.07304967194795609,
0.05256039649248123,
0.027852781116962433,
0.0039044988807290792,
0.1105925664305687,
-0.007829640060663223,
-0.014113753102719784,
0.05829976499080658,
-0.05399106442928314,
0.038163747638463974,
0.08281919360160828,
-0.010509602725505829,
-0.05866111069917679,
0.00812135636806488,
0.06226440146565437,
0.035100195556879044,
-0.016365712508559227,
-0.005121775437146425,
-1.1012699181737739e-34,
0.010702311061322689,
-0.005906430538743734,
0.08567411452531815,
-0.038480378687381744,
-0.013923613354563713,
-0.01322948932647705,
-0.04717526212334633,
-0.004356780555099249,
-0.021361371502280235,
-0.008544239215552807,
-0.044369809329509735,
0.030453985556960106,
-0.06262681633234024,
0.09446701407432556,
-0.0842360183596611,
-0.029716407880187035,
-0.05701165646314621,
0.03577835485339165,
-0.0139883728697896,
-0.08315668255090714,
-0.04397638514637947,
-0.0901869535446167,
-0.07200080901384354,
-0.007138976361602545,
-0.04206639900803566,
-0.010881229303777218,
0.06141028553247452,
-0.054480139166116714,
0.00696814339607954,
-0.018557868897914886,
-0.11513345688581467,
0.024942616000771523,
0.040642861276865005,
0.04739377275109291,
0.08720381557941437,
0.04392005130648613,
0.0869712382555008,
-0.009330620989203453,
0.037743888795375824,
0.006430118344724178,
0.012076988816261292,
0.021529855206608772,
-0.007735279388725758,
-0.07708199322223663,
-0.056383490562438965,
-0.13590918481349945,
0.009609637781977654,
0.013787812553346157,
-0.07443682104349136,
-0.05434658005833626,
0.002242506016045809,
-0.04254285618662834,
0.021506857126951218,
-0.07427895814180374,
-0.03455556556582451,
-0.04255494102835655,
-0.0049205198884010315,
0.01085570827126503,
0.03766835108399391,
0.07463311403989792,
-0.059264589101076126,
0.013382473960518837,
0.07696670293807983,
0.008523060940206051,
-0.0037329536862671375,
-0.014487489126622677,
-0.03746376186609268,
0.04285912215709686,
0.04352395609021187,
0.017641384154558182,
0.03441832587122917,
0.02252236008644104,
-0.06686951965093613,
-0.07469770312309265,
0.05905481427907944,
0.05127706006169319,
0.056466419249773026,
-0.028631065040826797,
-0.1318165808916092,
0.10221977531909943,
-0.03461041301488876,
-0.06508056819438934,
0.012197941541671753,
-0.04234687238931656,
-0.08550436049699783,
-0.06724059581756592,
0.07916762679815292,
-0.0888344869017601,
-0.037353627383708954,
-0.10388306528329849,
-0.08593692630529404,
0.014704582281410694,
-0.057117849588394165,
0.04865162447094917,
0.051149290055036545,
-8.403882324629603e-35,
0.013644088990986347,
0.0069879493676126,
-0.04282284155488014,
-0.05187581479549408,
0.01164101343601942,
-0.06627606600522995,
-0.0017413091845810413,
0.042969733476638794,
-0.037074629217386246,
-0.043912217020988464,
-0.007768846116960049,
-0.08436913043260574,
0.03460252657532692,
-0.019081884995102882,
0.05596667155623436,
0.014284353703260422,
-0.05196055769920349,
0.05984126031398773,
0.03679317981004715,
0.052379023283720016,
-0.0023195301182568073,
0.03287671133875847,
-0.04959184303879738,
0.05993904173374176,
0.038036201149225235,
0.06412868201732635,
0.029341639950871468,
0.10451459884643555,
-0.07089674472808838,
-0.041579701006412506,
-0.04512808099389076,
-0.07111566513776779,
-0.007433517836034298,
0.029194612056016922,
-0.11720706522464752,
0.03321191668510437,
0.022534899413585663,
-0.0038630464114248753,
-0.013787300325930119,
0.04327363520860672,
0.0033901117276400328,
0.03155796229839325,
-0.0762489065527916,
0.007978894747793674,
0.008328300900757313,
-0.04123881086707115,
-0.013268235139548779,
-0.0071898289024829865,
0.08559296280145645,
0.03619829937815666,
-0.037374433130025864,
0.05563044920563698,
-0.010662923566997051,
0.04170405492186546,
-0.08196815103292465,
-0.053130850195884705,
0.02488737925887108,
0.008039151318371296,
0.03463549166917801,
-0.0731825903058052,
-0.03535142540931702,
-0.02599056251347065,
0.019371338188648224,
-0.053450632840394974,
-0.0107920216396451,
-0.06726089119911194,
0.03794846311211586,
-0.052359625697135925,
-0.014823677949607372,
0.016544818878173828,
-0.026250578463077545,
0.038040805608034134,
0.0035644101444631815,
0.00728306919336319,
0.04091325402259827,
-0.057863976806402206,
-0.022372299805283546,
-0.02381708286702633,
-0.07847892493009567,
-0.05688179284334183,
0.07170490175485611,
-0.011351601220667362,
0.04664352908730507,
0.04347776621580124,
-0.033104993402957916,
0.05903032422065735,
0.040500346571207047,
0.04571234807372093,
-0.022150469943881035,
-0.026011357083916664,
-0.031770333647727966,
-0.040918875485658646,
-0.05120908468961716,
0.00568965170532465,
-0.036703310906887054,
-3.9365772863675375e-8,
-0.09353259950876236,
0.018793869763612747,
0.02369610033929348,
0.00030407123267650604,
0.016049953177571297,
-0.04751654341816902,
-0.02886190079152584,
0.08257822692394257,
0.03057657741010189,
-0.0299663208425045,
0.062215764075517654,
-0.04229367896914482,
-0.04211108759045601,
-0.045579854398965836,
-0.06839603930711746,
0.05276183411478996,
-0.003976480104029179,
0.04386075213551521,
0.007331809028983116,
0.020915091037750244,
-0.018263425678014755,
0.06446143239736557,
0.03420712426304817,
0.005478819366544485,
0.06252534687519073,
-0.05125460773706436,
-0.0030283010564744473,
0.04186127334833145,
0.0782388299703598,
-0.07184766232967377,
-0.01760922744870186,
0.039733998477458954,
0.08442838490009308,
-0.060263775289058685,
-0.11583960801362991,
0.06273774802684784,
0.043778195977211,
-0.03003348596394062,
-0.14525900781154633,
-0.07852693647146225,
0.006853956263512373,
0.09241801500320435,
-0.06164773181080818,
-0.06358422338962555,
0.02317996136844158,
-0.029332110658288002,
0.021098226308822632,
0.0015594852156937122,
0.10416999459266663,
0.005194611847400665,
0.07630092650651932,
-0.006630684249103069,
0.05770500749349594,
0.03987499326467514,
0.06803092360496521,
0.04104970395565033,
-0.0314701683819294,
-0.01868349500000477,
0.019608203321695328,
-0.0004609022580552846,
-0.009142492897808552,
0.03858010098338127,
-0.022138185799121857,
0.002178531140089035
] |
X-ALMA: Plug & Play Modules and Adaptive Rejection for Quality Translation at Scale | https://openreview.net/forum?id=csbf1p8xUq | [
"Haoran Xu",
"Kenton Murray",
"Philipp Koehn",
"Hieu Hoang",
"Akiko Eriguchi",
"Huda Khayrallah"
] | Spotlight | Large language models (LLMs) have achieved remarkable success across various NLP tasks with a focus on English due to English-centric pre-training and limited multilingual data. In this work, we focus on the problem of translation, and
while some multilingual LLMs claim to support for hundreds of languages, models often fail to provide high-quality responses for mid- and low-resource languages, leading to imbalanced performance heavily skewed in favor of high-resource languages. We introduce **X-ALMA**, a model designed to ensure top-tier performance across 50 diverse languages, regardless of their resource levels. X-ALMA surpasses state-of-the-art open-source multilingual LLMs, such as Aya-101 and Aya-23, in every single translation direction on the FLORES-200 and WMT'23 test datasets according to COMET-22. This is achieved by plug-and-play language-specific module architecture to prevent language conflicts during training and a carefully designed training regimen with novel optimization methods to maximize the translation performance. After the final stage of training regimen, our proposed **A**daptive **R**ejection **P**reference **O**ptimization (**ARPO**) surpasses existing preference optimization methods in translation tasks. | Large Language Model, Machine Translation, Multilingual | We present X-ALMA, a multilingual machine translation model that prioritizes quality over quantity by delivering top-tier performance across 50 diverse languages, regardless of their resource levels | 8,256 | null | [
-0.024033961817622185,
-0.055758994072675705,
0.04577721655368805,
0.007426721043884754,
0.04484286159276962,
-0.004633079282939434,
-0.01045497041195631,
-0.017210058867931366,
0.040484338998794556,
-0.06437871605157852,
-0.028830651193857193,
-0.04431844502687454,
0.03310634195804596,
-0.02206527628004551,
0.03892321512103081,
-0.032537009567022324,
0.062307387590408325,
0.07817260921001434,
-0.05902446061372757,
-0.06748750805854797,
0.027875453233718872,
0.061219215393066406,
0.07820330560207367,
0.05187603831291199,
0.08701265603303909,
-0.025950951501727104,
-0.07896683365106583,
-0.028165876865386963,
0.03975529968738556,
-0.08202693611383438,
-0.00848577730357647,
0.05846043676137924,
-0.02149188704788685,
0.03548220545053482,
-0.04191554710268974,
0.06520766019821167,
-0.0822543278336525,
-0.07736342400312424,
-0.021558044478297234,
-0.025201132521033287,
0.014211979694664478,
0.018859293311834335,
0.04871901124715805,
-0.03098156489431858,
0.05131321772933006,
-0.008010735735297203,
-0.006830551661550999,
-0.005224323831498623,
-0.010362147353589535,
0.06721186637878418,
-0.09724859148263931,
-0.06386609375476837,
-0.011242717504501343,
0.043654900044202805,
-0.026973212137818336,
0.031374283134937286,
0.06325831264257431,
0.03511229157447815,
0.0009055478149093688,
-0.010392357595264912,
-0.08585266023874283,
-0.05310971289873123,
-0.01667366921901703,
0.009555285796523094,
-0.028050711378455162,
-0.053442128002643585,
0.052006304264068604,
0.09277131408452988,
-0.043108582496643066,
0.03290553763508797,
-0.008989306166768074,
0.04922236129641533,
-0.028776362538337708,
0.09385576099157333,
0.0027298363856971264,
0.0443517230451107,
0.01827208697795868,
-0.010097258724272251,
0.05588405206799507,
0.009985287673771381,
0.0028292678762227297,
0.0020565318409353495,
0.07701604813337326,
-0.036742113530635834,
0.07861393690109253,
-0.050191719084978104,
0.033138077706098557,
0.009712806902825832,
0.011965512298047543,
0.028439372777938843,
-0.051373668015003204,
-0.006408728193491697,
0.06010537967085838,
0.08128420263528824,
-0.009175372309982777,
0.07643865048885345,
0.040805768221616745,
-0.0537889301776886,
0.002986279781907797,
0.08214295655488968,
0.07041892409324646,
0.10526539385318756,
0.045042283833026886,
-0.0056648836471140385,
-0.07924461364746094,
-0.05020024999976158,
0.04128965362906456,
0.011994708329439163,
-0.03602069988846779,
-0.06108986586332321,
0.048938412219285965,
0.008897390216588974,
0.030950048938393593,
-0.041440047323703766,
0.00546377943828702,
-0.0018364343559369445,
0.025634488090872765,
-0.0388789176940918,
-0.04876377806067467,
0.07531187683343887,
-0.07823711633682251,
0.011474202387034893,
0.05469631403684616,
0.019754881039261818,
0.006477647460997105,
-0.000436400092439726,
-0.056083742529153824,
-6.114739089928926e-35,
0.07097691297531128,
0.06569091975688934,
-0.00935228168964386,
-0.011606506071984768,
-0.009729905985295773,
-0.061868347227573395,
0.03128557652235031,
0.07481403648853302,
-0.13846781849861145,
-0.04520824924111366,
-0.017282700166106224,
0.023442909121513367,
-0.04009123891592026,
0.022249219939112663,
0.03626272454857826,
-0.0006104417843744159,
0.02464216947555542,
0.05092201381921768,
-0.013257026672363281,
0.08017086237668991,
0.09884249418973923,
-0.04156158119440079,
0.010557210072875023,
-0.06988877803087234,
0.025622591376304626,
0.058603957295417786,
0.04386056587100029,
-0.05520366504788399,
-0.06248660385608673,
0.007693492341786623,
-0.09201564639806747,
-0.07501424849033356,
0.03246643766760826,
0.08357401192188263,
0.0402168333530426,
-0.09566407650709152,
-0.08298727124929428,
-0.013736075721681118,
-0.0665382370352745,
-0.03858920559287071,
-0.050499171018600464,
0.06195445731282234,
-0.012733659707009792,
-0.02514549158513546,
0.024337438866496086,
-0.05696835368871689,
0.029042700305581093,
-0.038892850279808044,
-0.026234691962599754,
-0.04910838603973389,
0.0034980312921106815,
0.007596980314701796,
-0.0751919150352478,
0.056300025433301926,
0.061590246856212616,
0.029423581436276436,
0.06736018508672714,
0.051955465227365494,
-0.026089871302247047,
0.048069026321172714,
0.011822870932519436,
-0.03563247248530388,
0.006844582036137581,
0.12486504018306732,
0.06764193624258041,
-0.00005966243406874128,
-0.0048409621231257915,
0.009986571967601776,
0.03745615854859352,
-0.04262707754969597,
-0.035422783344984055,
-0.069246307015419,
0.03164997324347496,
0.04522261023521423,
0.027656231075525284,
-0.035079263150691986,
0.042636092752218246,
-0.0897664949297905,
0.05153666064143181,
0.010019907727837563,
0.01931403949856758,
-0.03329245746135712,
0.026490001007914543,
-0.06521428376436234,
-0.06595262885093689,
-0.06399250775575638,
0.028752831742167473,
-0.0347190722823143,
0.013610899448394775,
0.03390553221106529,
-0.0009945641504600644,
-0.0007747791823931038,
-0.030547551810741425,
-0.024148249998688698,
0.02445085346698761,
1.0050315987412282e-33,
-0.022357093170285225,
-0.004405404906719923,
-0.047321878373622894,
0.09372633695602417,
-0.041108448058366776,
-0.0933523029088974,
0.06012372300028801,
0.07883811742067337,
0.025069834664463997,
-0.019397038966417313,
-0.03410547599196434,
-0.08367551118135452,
0.051533691585063934,
-0.019081344828009605,
0.021591348573565483,
-0.059404365718364716,
0.052660081535577774,
0.05617290362715721,
0.02732967957854271,
0.09890668839216232,
-0.05132334306836128,
0.1350851207971573,
-0.04655976966023445,
0.021087639033794403,
0.009572443552315235,
0.009794041514396667,
-0.0790078267455101,
0.025907296687364578,
-0.023580776527523994,
0.002099857432767749,
-0.03370103985071182,
-0.019236203283071518,
-0.06435500085353851,
-0.024821685627102852,
-0.021847503259778023,
0.01920255646109581,
0.028490934520959854,
0.05559466779232025,
-0.03025643341243267,
0.108645960688591,
0.10444910824298859,
-0.013910479843616486,
-0.05267050117254257,
-0.017933722585439682,
-0.00403948500752449,
0.026649171486496925,
-0.14118601381778717,
-0.025721000507473946,
0.006707826163619757,
-0.021267486736178398,
-0.018771084025502205,
-0.04658748209476471,
-0.004307722207158804,
0.05074005573987961,
0.036753226071596146,
-0.172962948679924,
-0.0370384119451046,
-0.054958928376436234,
-0.02519400231540203,
-0.033307041972875595,
-0.008472676388919353,
-0.0437612421810627,
0.018230048939585686,
-0.02183881402015686,
0.012548179365694523,
-0.04660246521234512,
0.024729767814278603,
0.035411376506090164,
0.026233715936541557,
0.019759178161621094,
0.021003004163503647,
-0.04042309522628784,
-0.013038493692874908,
0.05652804300189018,
-0.02917839214205742,
0.015478025190532207,
-0.04555952176451683,
-0.0546584390103817,
0.04948559030890465,
-0.06641155481338501,
-0.0452992245554924,
0.06306880712509155,
0.001330144819803536,
0.051236871629953384,
0.07737112790346146,
0.0951799601316452,
0.013739177957177162,
0.10481253266334534,
0.012751222588121891,
0.03134465962648392,
0.0032430707942694426,
0.06129683926701546,
-0.03389246389269829,
0.008583799004554749,
-0.04902072623372078,
-4.0908766152369935e-8,
-0.0639333724975586,
0.0061143613420426846,
-0.06575752049684525,
0.07447443157434464,
-0.011069468222558498,
-0.07814302295446396,
-0.05014205724000931,
0.03937619924545288,
0.058276575058698654,
0.0467500165104866,
0.021319229155778885,
0.01509001012891531,
-0.03787129744887352,
-0.05064988508820534,
0.006136017385870218,
0.06917507201433182,
0.008822750300168991,
0.09123998135328293,
-0.02290232852101326,
-0.022039243951439857,
0.024093909189105034,
0.10082148015499115,
0.060141731053590775,
-0.04813537746667862,
0.011044681072235107,
-0.0637553334236145,
-0.009674843400716782,
0.08468884229660034,
0.11773893982172012,
-0.11476612836122513,
-0.052287615835666656,
-0.07283985614776611,
-0.03451760858297348,
-0.0454152375459671,
0.00882396288216114,
0.10533643513917923,
-0.008969170041382313,
0.018238471820950508,
0.022495297715067863,
0.04662148654460907,
0.06469589471817017,
0.0013867516536265612,
-0.06680071353912354,
0.024294663220643997,
0.016724852845072746,
-0.021646080538630486,
-0.08166973292827606,
-0.1111917644739151,
0.014429038390517235,
-0.11519013345241547,
0.03258476033806801,
-0.04187178239226341,
-0.02715284749865532,
0.015370950102806091,
0.05688326805830002,
0.08406563848257065,
-0.010152342729270458,
-0.040742937475442886,
0.063704714179039,
0.03971259295940399,
0.026707302778959274,
-0.06983929872512817,
-0.035948462784290314,
-0.01728503592312336
] |
Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification | https://openreview.net/forum?id=oI5tZaWkF9 | [
"Hsun-Yu Kuo",
"Yin-Hsiang Liao",
"Yu-Chieh Chao",
"Wei-Yun Ma",
"Pu-Jen Cheng"
] | Spotlight | Synthetic data augmentation via Large Language Models (LLMs) allows researchers to leverage additional training data, thus enhancing the performance of downstream tasks, especially when real-world data is scarce. However, the generated data can deviate from the real-world data, and this misalignment can bring about deficient results while applying the trained model to applications. Therefore, we proposed efficient weighted-loss approaches to align synthetic data with real-world distribution by emphasizing high-quality and diversified data generated by LLMs using merely a tiny amount of real-world data. We empirically assessed the effectiveness of our methods on multiple text classification tasks, and the results showed that leveraging our approaches on a BERT-level model robustly outperformed standard cross-entropy and other data weighting approaches, providing potential solutions to effectively leveraging synthetic data from any suitable data generator. | data weighing, data augmentation, distillation, data-efficient training, NLP in resource-constrained settings, fine-tuning, weighted loss | Weighted loss for training on data generated by LLM | 8,195 | 2410.21526 | [
-0.1288820505142212,
-0.10733351111412048,
0.014849225990474224,
0.09431128948926926,
0.08482339233160019,
0.03646634891629219,
-0.052175287157297134,
0.01757248304784298,
0.0045468490570783615,
-0.028081893920898438,
-0.049879614263772964,
-0.06559651345014572,
0.049765631556510925,
0.011714939028024673,
0.00832614116370678,
0.01441874634474516,
0.10061430186033249,
0.08440287411212921,
-0.07629603892564774,
-0.02041877992451191,
0.05597576126456261,
0.05162401497364044,
0.10282355546951294,
0.05078069865703583,
0.11936147511005402,
-0.037498556077480316,
-0.01539518404752016,
-0.009735269472002983,
-0.0017072164919227362,
-0.008136174641549587,
0.03608144819736481,
0.040670305490493774,
-0.08989309519529343,
-0.016442572697997093,
-0.02980247139930725,
0.06859071552753448,
-0.07830583304166794,
0.045318469405174255,
0.008599851280450821,
0.0064475517719984055,
0.011456728912889957,
-0.04647461697459221,
-0.02096869796514511,
0.03380507975816727,
0.04407323896884918,
-0.004085985012352467,
-0.022639093920588493,
-0.0315420888364315,
0.026796547695994377,
0.05421515181660652,
-0.04316234961152077,
0.009442231617867947,
-0.02981233038008213,
0.09956233948469162,
-0.07175356894731522,
0.030201174318790436,
0.08544066548347473,
0.018913108855485916,
-0.0328601635992527,
-0.03481248766183853,
0.0014748312532901764,
-0.09203313291072845,
0.028026334941387177,
0.004189095925539732,
-0.04880375415086746,
-0.06210342049598694,
-0.02772017940878868,
0.11788439750671387,
-0.012811651453375816,
0.028709009289741516,
0.015918919816613197,
0.1024923324584961,
-0.06331069022417068,
0.08700002729892731,
0.019686589017510414,
-0.01001391839236021,
0.03209901973605156,
-0.019910374656319618,
0.050041601061820984,
0.009172883816063404,
-0.005777567159384489,
0.0234131570905447,
0.036082617938518524,
0.008955704979598522,
0.13821524381637573,
0.024570249021053314,
0.04279793053865433,
0.02252238057553768,
-0.04376673698425293,
0.06394657492637634,
-0.04974280297756195,
-0.04478588327765465,
0.10443489253520966,
0.010546383447945118,
-0.03460187837481499,
0.0216913390904665,
-0.007629521656781435,
-0.07048521935939789,
0.024532433599233627,
0.02351429872214794,
-0.00459663663059473,
0.08526305109262466,
0.016098253428936005,
-0.03701944276690483,
-0.0023682909086346626,
-0.03480588272213936,
0.028303520753979683,
-0.00578049011528492,
0.011314239352941513,
-0.07694383710622787,
0.08012694120407104,
0.039750780910253525,
0.024935398250818253,
-0.06710273772478104,
0.010462802834808826,
0.026984816417098045,
0.04081549495458603,
-0.06025043874979019,
-0.06802430003881454,
0.06799693405628204,
-0.12460410594940186,
0.044673752039670944,
-0.02994018793106079,
-0.013258138671517372,
-0.011508855037391186,
0.02028561197221279,
-0.005317832343280315,
3.7660859027048925e-33,
0.019238894805312157,
0.040816035121679306,
0.042730748653411865,
0.01767919957637787,
0.018518958240747452,
-0.0578157939016819,
-0.025546744465827942,
0.04705880209803581,
-0.06618630141019821,
0.0011146427132189274,
0.00280574569478631,
0.04562857002019882,
-0.022794771939516068,
0.10381662100553513,
0.055927034467458725,
-0.008069958537817001,
-0.06888892501592636,
0.04356931895017624,
0.03212462738156319,
-0.026211077347397804,
0.0010421880288049579,
-0.0302052553743124,
0.027413057163357735,
-0.05527316778898239,
0.005349323153495789,
0.04481067508459091,
0.062066856771707535,
0.0013622920960187912,
-0.06217212229967117,
0.004821183159947395,
-0.10290442407131195,
0.004710093606263399,
0.06924515962600708,
-0.035316016525030136,
0.03559023514389992,
-0.01868254877626896,
-0.04080517590045929,
-0.032384954392910004,
0.01603727415204048,
0.05751003697514534,
-0.03960482403635979,
0.07767313718795776,
0.02394440770149231,
0.006911180913448334,
-0.03892653435468674,
-0.007682538125663996,
0.010160606354475021,
0.006395936477929354,
0.02388884499669075,
-0.04660353437066078,
0.0647280141711235,
-0.02219559997320175,
-0.033845141530036926,
0.03913348913192749,
-0.007741569075733423,
0.019531186670064926,
-0.02192547731101513,
-0.0016654465580359101,
0.014815702103078365,
0.00841255858540535,
-0.004536748863756657,
0.025823606178164482,
0.0065209874883294106,
0.08999087661504745,
0.004441902507096529,
-0.031534384936094284,
0.04767642170190811,
0.012029319070279598,
-0.019936444237828255,
0.033089570701122284,
-0.003766808658838272,
0.012539748102426529,
-0.03720470517873764,
-0.005569246597588062,
0.0538470484316349,
-0.01712210476398468,
0.05115622282028198,
-0.08241304010152817,
0.046318646520376205,
0.054851558059453964,
0.05541990324854851,
0.0033732156734913588,
0.05038093030452728,
-0.09834165126085281,
-0.04981224238872528,
-0.05033062770962715,
0.03157228231430054,
-0.08671125769615173,
-0.04471265897154808,
0.09300544112920761,
-0.08162476122379303,
-0.05041548237204552,
-0.04697982966899872,
-0.019872387871146202,
0.005676952190697193,
-2.2950590477744352e-33,
-0.09828048944473267,
-0.00767316110432148,
-0.04731752723455429,
0.06211775913834572,
-0.07407312095165253,
-0.022110693156719208,
-0.07084187120199203,
0.06857574731111526,
-0.01244301162660122,
-0.011606490239501,
-0.03477625176310539,
-0.033573925495147705,
0.015431798063218594,
-0.01983216218650341,
0.017803175374865532,
-0.01162867620587349,
-0.0055508678779006,
0.0761289894580841,
-0.08109057694673538,
0.0027877953834831715,
0.04956815019249916,
0.09115374833345413,
-0.09685637056827545,
0.05335560068488121,
0.029094291850924492,
0.021396230906248093,
-0.03825034573674202,
0.10808490961790085,
-0.030323760583996773,
-0.037548523396253586,
-0.02606317028403282,
-0.06518051028251648,
-0.017887337133288383,
-0.020598120987415314,
-0.09370604902505875,
0.06905671954154968,
0.09119497239589691,
-0.01862114481627941,
0.009693690575659275,
0.04167748615145683,
-0.004197598900645971,
0.009758668951690197,
-0.11264848709106445,
-0.014229284599423409,
-0.06147270277142525,
-0.06281041353940964,
-0.1352357268333435,
-0.03284542262554169,
0.0689876601099968,
0.012344938702881336,
-0.009499164298176765,
-0.017656393349170685,
-0.07484062761068344,
0.0657319501042366,
-0.014439010061323643,
-0.144622802734375,
-0.055059049278497696,
-0.03515543416142464,
-0.06441425532102585,
-0.02458728663623333,
-0.06898382306098938,
0.031675342470407486,
0.11452198773622513,
-0.07552159577608109,
-0.01161582488566637,
-0.08608751744031906,
0.023014258593320847,
-0.00607337849214673,
-0.02546049654483795,
0.01264030858874321,
0.0391555055975914,
-0.01927497796714306,
0.028583602979779243,
0.037975993007421494,
-0.011824521236121655,
-0.026000455021858215,
0.041926685720682144,
-0.01470115128904581,
-0.07054486125707626,
-0.013726498931646347,
0.0663973018527031,
-0.018314246088266373,
0.0840856209397316,
0.07621113955974579,
0.05351300910115242,
0.09971414506435394,
0.03256634250283241,
0.041343122720718384,
-0.0012098705628886819,
0.017173757776618004,
-0.022674378007650375,
0.027567783370614052,
-0.0365372858941555,
0.07996000349521637,
-0.036015018820762634,
-4.101920580978913e-8,
-0.01567208580672741,
0.01066551636904478,
-0.010165159590542316,
0.054427530616521835,
-0.07756882905960083,
-0.024250194430351257,
-0.07412006705999374,
0.06772511452436447,
-0.057031456381082535,
-0.0308960173279047,
0.05066351965069771,
0.009397616609930992,
0.002446188125759363,
0.0017948931781575084,
0.02410840429365635,
0.01887260563671589,
-0.008317462168633938,
0.02438119798898697,
-0.00512858247384429,
-0.024956833571195602,
-0.001262968173250556,
0.10218048095703125,
0.02549438737332821,
-0.018903451040387154,
0.08853578567504883,
-0.03942735865712166,
0.008170662447810173,
0.029556522145867348,
0.05291574448347092,
-0.027043677866458893,
0.00365168577991426,
-0.055838245898485184,
-0.14783059060573578,
0.006688100751489401,
0.054482508450746536,
0.04136352613568306,
0.06206002086400986,
-0.01926816813647747,
-0.08160895109176636,
-0.007600752171128988,
0.049157168716192245,
0.075544074177742,
-0.05797991901636124,
0.07953985780477524,
0.0531974658370018,
-0.07022087275981903,
-0.05253785848617554,
-0.03677942976355553,
0.08387316018342972,
-0.07111523300409317,
0.05840315669775009,
-0.0654618889093399,
0.014246348291635513,
0.03487018868327141,
0.023908425122499466,
0.02880893275141716,
-0.06633447855710983,
0.012595140375196934,
0.011711862869560719,
0.011422851122915745,
0.05428427830338478,
-0.07699131220579147,
-0.03485219553112984,
0.04482357203960419
] |
Systems with Switching Causal Relations: A Meta-Causal Perspective | https://openreview.net/forum?id=J9VogDTa1W | [
"Moritz Willig",
"Tim Tobiasch",
"Florian Peter Busch",
"Jonas Seng",
"Devendra Singh Dhami",
"Kristian Kersting"
] | Spotlight | Most work on causality in machine learning assumes that causal relationships are driven by a constant underlying process. However, the flexibility of agents' actions or tipping points in the environmental process can change the qualitative dynamics of the system. As a result, new causal relationships may emerge, while existing ones change or disappear, resulting in an altered causal graph. To analyze these qualitative changes on the causal graph, we propose the concept of meta-causal states, which groups classical causal models into clusters based on equivalent qualitative behavior and consolidates specific mechanism parameterizations. We demonstrate how meta-causal states can be inferred from observed agent behavior, and discuss potential methods for disentangling these states from unlabeled data. Finally, we direct our analysis towards the application of a dynamical system, showing that meta-causal states can also emerge from inherent system dynamics, and thus constitute more than a context-dependent framework in which mechanisms emerge only as a result of external factors. | Meta-Causality, Meta-Causal Reasoning, Agent Behavior, System Dynamics | This paper introduces meta-causal states to capture changing causal dynamics emerging from agents and environmental behaviors in dynamical systems. | 8,182 | 2410.13054 | [
-0.013055377639830112,
-0.06041238084435463,
0.01874598301947117,
0.036083851009607315,
0.11893100291490555,
-0.009448936209082603,
-0.00675604585558176,
-0.039645303040742874,
0.0034420883748680353,
-0.011894498951733112,
-0.05039585009217262,
-0.03884962946176529,
0.03322341665625572,
0.05972592532634735,
-0.018099399283528328,
0.03166431933641434,
-0.002129604108631611,
0.008939643390476704,
-0.07840291410684586,
-0.039671193808317184,
0.01578904129564762,
-0.021620430052280426,
-0.03688947483897209,
0.04938848689198494,
-0.049251507967710495,
0.04334123805165291,
0.0032200459390878677,
0.06661359965801239,
0.006338866893202066,
-0.021591216325759888,
0.03453359007835388,
0.021226122975349426,
-0.16746440529823303,
-0.07858902961015701,
0.015105746686458588,
0.08680473268032074,
0.028810348361730576,
0.07479619234800339,
-0.006337945815175772,
0.018529172986745834,
0.056287214159965515,
-0.020043589174747467,
-0.01983315870165825,
-0.002099944045767188,
0.029132157564163208,
0.053627632558345795,
-0.03586534038186073,
0.033147312700748444,
-0.06870583444833755,
-0.01508402545005083,
-0.1541600078344345,
0.008278295397758484,
-0.06176371872425079,
0.04593224078416824,
0.0704096257686615,
0.026743920519948006,
0.12181327491998672,
0.02740597352385521,
-0.0377085916697979,
-0.0541280098259449,
-0.0083972392603755,
-0.037692852318286896,
-0.07377497851848602,
0.014661205932497978,
-0.011875434778630733,
0.12063088268041611,
-0.10465100407600403,
0.08564992249011993,
-0.00797747727483511,
0.024494528770446777,
-0.023842468857765198,
0.09301906824111938,
-0.07763984799385071,
-0.023619238287210464,
0.05883917212486267,
0.04371064528822899,
-0.004460600204765797,
0.05383786931633949,
0.034273143857717514,
-0.08513163030147552,
0.030337810516357422,
0.055644407868385315,
-0.023718634620308876,
0.0040636551566421986,
0.07254786789417267,
-0.051416996866464615,
-0.014999989420175552,
-0.05706241354346275,
0.07318060845136642,
0.05001690983772278,
-0.027275903150439262,
0.00482436316087842,
0.07140582799911499,
-0.017995234578847885,
0.0516907274723053,
0.052104055881500244,
0.0568348690867424,
-0.08266382664442062,
0.07843540608882904,
0.11991322040557861,
-0.017418179661035538,
0.023233536630868912,
0.007178914733231068,
0.054967984557151794,
0.03985043615102768,
-0.05639601871371269,
0.004585325717926025,
0.004509438760578632,
0.017047610133886337,
-0.02984561212360859,
0.0008571788785047829,
0.009438588283956051,
0.05698681250214577,
-0.009162361733615398,
-0.04364893212914467,
-0.05094057321548462,
0.06995107978582382,
-0.0430331826210022,
-0.014196500182151794,
-0.026498377323150635,
-0.003956828266382217,
-0.03356972709298134,
-0.01840181089937687,
0.0456705316901207,
0.018064014613628387,
-0.029284533113241196,
-0.08874811977148056,
1.1344816284285864e-33,
0.0520615316927433,
-0.04334297031164169,
0.05141303315758705,
-0.004859402310103178,
0.021093666553497314,
-0.053032055497169495,
-0.0946550965309143,
-0.05844626575708389,
0.06600643694400787,
-0.011750332079827785,
0.009858197532594204,
0.052983563393354416,
-0.048937998712062836,
0.012112254276871681,
0.07011118531227112,
0.008880829438567162,
0.010140222497284412,
0.05476687476038933,
0.027915140613913536,
-0.13533112406730652,
0.01921524666249752,
-0.012270807288587093,
-0.051045265048742294,
0.03470524773001671,
-0.04267054796218872,
0.042576707899570465,
-0.003462412627413869,
0.030631467700004578,
-0.03687627986073494,
0.008288816548883915,
0.025505149737000465,
0.04644189774990082,
0.013030266389250755,
0.06748867779970169,
0.003599993884563446,
-0.001278176554478705,
-0.05933021008968353,
-0.06250680983066559,
-0.0043904464691877365,
-0.007904153317213058,
-0.000213700914173387,
-0.020069319754838943,
0.03715858235955238,
-0.0285397507250309,
-0.036878179758787155,
-0.11465270072221756,
0.008762918412685394,
-0.07312826067209244,
-0.07844938337802887,
-0.07456851750612259,
0.013773110695183277,
0.023016294464468956,
0.07369887828826904,
-0.10516589134931564,
-0.0203729085624218,
0.0587671734392643,
0.01954723708331585,
0.015672093257308006,
-0.050175752490758896,
-0.017610009759664536,
0.014885744079947472,
0.011773774400353432,
-0.02190478891134262,
0.08037359267473221,
0.013861477375030518,
0.07663560658693314,
0.004844521638005972,
0.0044010053388774395,
0.020106561481952667,
0.04345157369971275,
-0.042821530252695084,
0.09314454346895218,
-0.04518453776836395,
-0.010720225982367992,
0.0286755058914423,
-0.01219622790813446,
-0.047210369259119034,
-0.10548976063728333,
-0.05111272633075714,
0.019862480461597443,
-0.057268425822257996,
-0.080593541264534,
-0.09308882802724838,
0.05002501234412193,
-0.03898414969444275,
-0.04716112092137337,
-0.020318204537034035,
-0.06786170601844788,
-0.06152592599391937,
-0.05508872866630554,
-0.05981471389532089,
-0.023319130763411522,
0.03350624069571495,
0.050867047160863876,
0.02438206598162651,
-4.3021293264371026e-33,
-0.08763658255338669,
0.008244440890848637,
0.0689709261059761,
-0.005406188778579235,
-0.00725911557674408,
0.025458894670009613,
-0.04401421174407005,
-0.03309285640716553,
0.01484384760260582,
-0.011703481897711754,
-0.05753905698657036,
-0.03899478539824486,
0.058706242591142654,
0.021284515038132668,
-0.0558750219643116,
0.030122367665171623,
0.03476352244615555,
0.022107195109128952,
0.05172028765082359,
-0.045722514390945435,
0.05436576530337334,
0.05115339532494545,
-0.026663783937692642,
0.03679531812667847,
0.006554527208209038,
-0.008534382097423077,
0.010583349503576756,
0.05205953121185303,
0.02065633237361908,
0.04153551161289215,
-0.06541477143764496,
-0.004636378958821297,
-0.04353522136807442,
-0.027592897415161133,
-0.05102907866239548,
0.12202847003936768,
0.0003631610597949475,
-0.0711340457201004,
-0.047113168984651566,
0.024215195327997208,
0.0319363959133625,
-0.06634610891342163,
-0.037548135966062546,
-0.007011828012764454,
0.027739984914660454,
-0.03973347693681717,
-0.05802915245294571,
0.09232991188764572,
0.024368997663259506,
-0.09873379021883011,
-0.0002647441579028964,
0.0626024529337883,
0.009286080487072468,
-0.022639289498329163,
-0.08847491443157196,
0.004781763069331646,
0.11634344607591629,
0.00011451409227447584,
0.041068412363529205,
0.03301914036273956,
-0.03258717060089111,
-0.07086450606584549,
-0.027375569567084312,
0.017688235267996788,
-0.05104426294565201,
-0.04093256965279579,
0.029443705454468727,
-0.11102854460477829,
0.02646583318710327,
0.003802649909630418,
0.05594189837574959,
0.042795054614543915,
-0.0814402848482132,
-0.0699663758277893,
0.03573233261704445,
-0.024639487266540527,
-0.057890884578228,
-0.03360108286142349,
-0.09820431470870972,
-0.04564359411597252,
0.05337996035814285,
0.007040258031338453,
0.058485135436058044,
-0.003992670681327581,
-0.04699770733714104,
0.028596578165888786,
-0.01846982352435589,
0.03917958214879036,
0.0446757972240448,
0.05466635897755623,
-0.06941786408424377,
-0.004805319011211395,
-0.05560852587223053,
0.04655463621020317,
-0.1281294971704483,
-4.429912436876293e-8,
-0.02819390594959259,
0.08212579041719437,
0.06093902140855789,
0.07795485854148865,
0.0687817707657814,
0.02565973810851574,
0.02579420618712902,
0.07492393255233765,
-0.005339747294783592,
-0.001403009402565658,
0.06624831259250641,
0.03472520783543587,
0.03532528877258301,
0.016435161232948303,
-0.039065614342689514,
0.06183909997344017,
0.06775243580341339,
-0.014985398389399052,
-0.011132195591926575,
-0.00991127546876669,
0.046705424785614014,
-0.0018371318001300097,
-0.02161029353737831,
0.012444837018847466,
0.06941846758127213,
-0.16859771311283112,
-0.037144917994737625,
0.06331035494804382,
0.056651510298252106,
0.05697719752788544,
0.015454739332199097,
0.06020830199122429,
0.010451316833496094,
0.006906589493155479,
-0.07038751989603043,
0.0348983034491539,
-0.014094945043325424,
-0.05726678669452667,
-0.07200106978416443,
-0.0669025182723999,
0.021834762766957283,
0.11932229995727539,
-0.11096645891666412,
0.05908485874533653,
0.002803621580824256,
0.027633903548121452,
0.02186570130288601,
-0.01947164162993431,
0.02188907563686371,
-0.020986124873161316,
0.008744697086513042,
-0.01613512448966503,
-0.0224752314388752,
0.025221746414899826,
0.08400510251522064,
-0.0009703016257844865,
0.030535627156496048,
-0.014383369125425816,
0.05321316421031952,
0.023823026567697525,
0.05814775452017784,
0.02281779982149601,
0.0062476941384375095,
-0.047399748116731644
] |
Multi-Robot Motion Planning with Diffusion Models | https://openreview.net/forum?id=AUCYptvAf3 | [
"Yorai Shaoul",
"Itamar Mishani",
"Shivam Vats",
"Jiaoyang Li",
"Maxim Likhachev"
] | Spotlight | Diffusion models have recently been successfully applied to a wide range of robotics applications for learning complex multi-modal behaviors from data. However, prior works have mostly been confined to single-robot and small-scale environments due to the high sample complexity of learning multi-robot diffusion models. In this paper, we propose a method for generating collision-free multi-robot trajectories that conform to underlying data distributions while using only single-robot data. Our algorithm, Multi-robot Multi-model planning Diffusion (MMD), does so by combining learned diffusion models with classical search-based techniques---generating data-driven motions under collision constraints. Scaling further, we show how to compose multiple diffusion models to plan in large environments where a single diffusion model fails to generalize well. We demonstrate the effectiveness of our approach in planning for dozens of robots in a variety of simulated scenarios motivated by logistics environments. | Multi-Agent Planning, Robotics, Generative Models | This paper presents a method combining diffusion models with search-based techniques to generate scalable multi-robot trajectories using only single-robot data. | 8,083 | 2410.03072 | [
0.06870345026254654,
-0.10564040392637253,
0.01575051061809063,
-0.015253587625920773,
0.04674779251217842,
-0.08128706365823746,
-0.08889500796794891,
-0.03531533479690552,
0.005384978372603655,
0.032148923724889755,
0.041535619646310806,
-0.03081071749329567,
0.013244186528027058,
0.06903162598609924,
0.03811755031347275,
0.02278592623770237,
0.006885239854454994,
0.03947916999459267,
-0.0921071246266365,
-0.014521488919854164,
-0.0295850969851017,
0.024923313409090042,
0.05713378265500069,
-0.007408725563436747,
-0.01343932468444109,
0.040794000029563904,
0.0736946165561676,
-0.09291195869445801,
-0.045172762125730515,
-0.04741716384887695,
0.08514678478240967,
-0.03849632292985916,
-0.04594019055366516,
0.004806420765817165,
0.02886515110731125,
0.05647902190685272,
-0.0639130026102066,
0.004028122406452894,
-0.006102232728153467,
-0.009272970259189606,
0.00012692446762230247,
0.0663478672504425,
0.07699687033891678,
-0.030896836891770363,
0.04224935919046402,
0.03155174478888512,
-0.027997175231575966,
-0.07632670551538467,
0.1062692329287529,
0.014898892492055893,
-0.047530800104141235,
-0.0668552815914154,
0.04464002698659897,
0.057328324764966965,
-0.02876557968556881,
-0.056645698845386505,
-0.0013116293121129274,
-0.02137955091893673,
0.07188273966312408,
-0.07750087976455688,
-0.0425596758723259,
-0.06321214884519577,
0.043817054480314255,
-0.02477421425282955,
-0.0660409927368164,
-0.02619986981153488,
-0.0008113018120639026,
0.0438775010406971,
-0.038499049842357635,
0.06569370627403259,
0.02833661250770092,
-0.0010434818686917424,
0.011920277960598469,
0.024585429579019547,
0.0455801859498024,
-0.0006369632319547236,
-0.01840241067111492,
-0.03891498222947121,
-0.04975350573658943,
0.012987828813493252,
0.004366777371615171,
-0.02941754460334778,
-0.006127083674073219,
0.04859653860330582,
0.04331377148628235,
-0.017749233171343803,
0.0007705694297328591,
0.07913843542337418,
0.0823175460100174,
-0.034341391175985336,
-0.022990619763731956,
0.04334482550621033,
0.008226302452385426,
-0.041430339217185974,
-0.07228823006153107,
-0.005904701538383961,
-0.053739845752716064,
-0.06783735007047653,
0.05547739192843437,
-0.006099603604525328,
-0.043568532913923264,
0.012247899547219276,
0.014793855138123035,
0.05200663208961487,
0.009020213037729263,
-0.0346994511783123,
0.006176639813929796,
-0.021499471738934517,
0.048433177173137665,
0.006609327159821987,
0.04321861267089844,
-0.04200560972094536,
0.09560997784137726,
0.05989323928952217,
-0.03216739743947983,
-0.007967053912580013,
0.010241416282951832,
0.03556203469634056,
-0.04748305678367615,
0.09325975924730301,
-0.00953346211463213,
-0.11534490436315536,
-0.004213524982333183,
0.04087240248918533,
0.01534639485180378,
-0.04184608906507492,
-0.013373990543186665,
-2.246670557506805e-33,
0.030995702371001244,
-0.04535974934697151,
0.034758228808641434,
0.054045237600803375,
0.10199027508497238,
-0.01541394367814064,
-0.011601210571825504,
-0.041459303349256516,
0.004212940111756325,
-0.007365694269537926,
-0.023678481578826904,
0.015559262596070766,
-0.05721813440322876,
0.02422192692756653,
0.004245759919285774,
-0.029611853882670403,
0.0102295046672225,
0.02023465745151043,
0.0038376408629119396,
0.07795562595129013,
0.05888286978006363,
-0.00030157031142152846,
-0.0627279207110405,
-0.025512872263789177,
0.018872777000069618,
0.03897363319993019,
0.012227385304868221,
-0.008857307024300098,
0.08269060403108597,
-0.003836633637547493,
-0.11834194511175156,
0.025254635140299797,
-0.08688351511955261,
0.15849649906158447,
0.03156537562608719,
0.11522796005010605,
-0.15451698005199432,
-0.009743141010403633,
-0.02663514018058777,
0.03718343749642372,
0.011147424578666687,
-0.0630016177892685,
-0.018002083525061607,
-0.032584596425294876,
-0.09781695157289505,
-0.025194654241204262,
-0.002776609966531396,
-0.008497570641338825,
-0.06539124250411987,
-0.07713644206523895,
0.004362284205853939,
0.02696037106215954,
-0.011173196136951447,
-0.09865177422761917,
-0.020105894654989243,
-0.06208496913313866,
0.03889257460832596,
-0.06049139425158501,
0.002090058522298932,
0.02190946415066719,
0.0288997869938612,
-0.016980813816189766,
-0.024941178038716316,
0.04533851519227028,
0.06467290967702866,
-0.023004380986094475,
-0.009725959040224552,
0.10476723313331604,
0.16757982969284058,
-0.018298475071787834,
0.021463431417942047,
0.020322630181908607,
0.005842692218720913,
-0.08636064827442169,
0.05205945670604706,
-0.03589543327689171,
0.053134188055992126,
-0.0451279915869236,
0.0282968208193779,
0.01678316481411457,
-0.05431528016924858,
-0.05668342486023903,
-0.05131968483328819,
-0.020415399223566055,
0.06260058283805847,
0.013505429029464722,
-0.06027562543749809,
-0.027424281463027,
-0.09438005834817886,
-0.07602612674236298,
-0.08678003400564194,
-0.046027541160583496,
-0.1102260947227478,
0.023076770827174187,
-0.0328185111284256,
-1.245724260812165e-33,
0.003499317681416869,
0.008082134649157524,
0.07686395943164825,
0.030569028109312057,
-0.01136811077594757,
-0.01016371138393879,
-0.03928164765238762,
-0.059992466121912,
0.07597735524177551,
-0.035787373781204224,
-0.11037179082632065,
-0.020377246662974358,
0.03338974714279175,
-0.05864952504634857,
0.05081925541162491,
0.008216236717998981,
0.09236078709363937,
-0.05552444979548454,
0.005409735254943371,
-0.038796309381723404,
-0.039646606892347336,
0.03339851275086403,
-0.05406134948134422,
0.010193340480327606,
-0.0017688700463622808,
0.01054114755243063,
-0.048973068594932556,
-0.027018168941140175,
0.006347415503114462,
0.07501336932182312,
-0.05689083784818649,
-0.06664687395095825,
0.069083072245121,
-0.016239991411566734,
-0.027280403301119804,
0.06331997364759445,
0.001774124102666974,
0.05900822952389717,
0.03724585101008415,
0.014259111136198044,
0.00239061308093369,
-0.10490158945322037,
-0.058926187455654144,
0.05572536587715149,
0.03212830424308777,
-0.07468771934509277,
-0.0215875543653965,
0.061558011919260025,
0.00816577859222889,
0.039362095296382904,
-0.039692770689725876,
0.10799282044172287,
-0.1114264726638794,
-0.10705430805683136,
-0.019587475806474686,
0.002008049516007304,
-0.026940012350678444,
0.008144179359078407,
0.05576305091381073,
-0.03441295027732849,
-0.06252407282590866,
-0.030724506825208664,
0.03443899750709534,
0.1193811371922493,
-0.08031239360570908,
0.014784686267375946,
-0.02503899298608303,
-0.0260208398103714,
-0.031452931463718414,
-0.008968259207904339,
0.00410205265507102,
0.051094118505716324,
0.05082577466964722,
0.009290066547691822,
0.050130072981119156,
-0.07464194297790527,
-0.001330797327682376,
0.0038814737927168608,
0.0438481867313385,
0.02203313075006008,
0.021142277866601944,
0.03416074067354202,
0.03253103047609329,
0.10428323596715927,
-0.007367908488959074,
0.04886886104941368,
0.0006418694974854589,
0.04062870517373085,
-0.009418375790119171,
0.024694401770830154,
0.03869817405939102,
0.021522745490074158,
0.06328456848859787,
0.033625997602939606,
-0.13561315834522247,
-3.450439223229296e-8,
-0.04728863015770912,
0.03403091803193092,
0.03503427654504776,
-0.003933166153728962,
0.060903094708919525,
0.04682503640651703,
0.036148253828287125,
0.05091725289821625,
-0.023235447704792023,
0.030849991366267204,
0.03560120612382889,
0.05929357931017876,
-0.03961886838078499,
0.03760876879096031,
-0.03849516808986664,
0.051642999053001404,
0.035133637487888336,
-0.07864397019147873,
0.02145679108798504,
0.061274170875549316,
-0.02501257136464119,
-0.06082697585225105,
0.017650360241532326,
-0.000927418761420995,
0.013428322039544582,
-0.005743908230215311,
-0.14570803940296173,
-0.04699171707034111,
0.0725177600979805,
-0.00014665606431663036,
-0.043379079550504684,
-0.07621448487043381,
-0.00225256965495646,
-0.03956891596317291,
0.040836505591869354,
0.0068519036285579205,
0.007604673970490694,
-0.04237747937440872,
-0.0127668846398592,
-0.06319163739681244,
0.00366410193964839,
0.10187550634145737,
0.044512227177619934,
0.019943326711654663,
0.024745794013142586,
0.07810422033071518,
-0.044648002833127975,
-0.06903621554374695,
0.010434133000671864,
-0.0021554341074079275,
-0.05531731992959976,
0.014516047202050686,
0.014577249996364117,
0.07076892256736755,
0.13237501680850983,
0.021739322692155838,
0.02979762852191925,
-0.06847988069057465,
0.07870769500732422,
0.017577137798070908,
-0.009982055053114891,
0.0004535022017080337,
-0.039556339383125305,
0.052430588752031326
] |
Graph Neural Networks Can (Often) Count Substructures | https://openreview.net/forum?id=sZQRUrvLn4 | [
"Paolo Pellizzoni",
"Till Hendrik Schulz",
"Karsten Borgwardt"
] | Spotlight | Message passing graph neural networks (GNNs) are known to have limited expressive power in their ability to distinguish some non-isomorphic graphs.
Because of this, it is well known that they are unable to detect or count arbitrary graph substructures (i.e., solving the subgraph isomorphism problem), a task that is of great importance for several types of graph-structured data.
However, we observe that GNNs are in fact able to count graph patterns quite accurately across several real-world graph datasets.
Motivated by this observation, we provide an analysis of the subgraph-counting capabilities of GNNs beyond the worst case, deriving several sufficient conditions for GNNs to be able to count subgraphs and, more importantly, to be able to sample-efficiently learn to count subgraphs.
Moreover, we develop novel dynamic programming algorithms for solving the subgraph isomorphism problem on restricted classes of pattern and target graphs, and show that message-passing GNNs can efficiently simulate these dynamic programs.
Finally, we empirically validate that our sufficient conditions for GNNs to count subgraphs hold on many real-world datasets, providing a theoretically-grounded explanation to our motivating observations. | graph neural networks, subgraphs, expressivity | We provide a theoretical analysis of the subgraph-counting capabilities of graph neural networks beyond the worst case. | 8,022 | null | [
-0.08163964003324509,
-0.06816643476486206,
0.06606186181306839,
0.061533499509096146,
0.025178633630275726,
-0.05575285479426384,
-0.0006366338348016143,
-0.0455106645822525,
-0.02278679423034191,
-0.07638131827116013,
-0.03502000495791435,
-0.06933605670928955,
0.032585207372903824,
0.11292222887277603,
-0.01084598246961832,
0.04536943510174751,
0.09184426814317703,
0.059199705719947815,
0.0019537117332220078,
-0.015736885368824005,
0.023933270946145058,
0.028103092685341835,
0.01794024184346199,
0.007812271825969219,
0.06443751603364944,
-0.02859070524573326,
-0.05628371238708496,
-0.04561256244778633,
0.09137263894081116,
0.04004474729299545,
0.060248784720897675,
-0.03995522856712341,
-0.04684975743293762,
0.05173250287771225,
0.0029075308702886105,
0.01995522901415825,
-0.03209804370999336,
-0.02394143119454384,
0.08061689883470535,
0.0479726679623127,
-0.0011428501456975937,
0.023096831515431404,
0.02775464579463005,
0.06360314786434174,
0.08324284106492996,
0.06545045971870422,
-0.01701555773615837,
0.029768716543912888,
-0.06536593288183212,
0.04654368385672569,
-0.036538418382406235,
-0.022824512794613838,
-0.09415718168020248,
0.04400770366191864,
0.07698510587215424,
-0.016648976132273674,
-0.010091079398989677,
-0.005640394054353237,
0.014353561215102673,
0.07595790177583694,
0.023873789235949516,
-0.09304259717464447,
0.02405269630253315,
-0.08419710397720337,
-0.09738878160715103,
0.07453804463148117,
0.027476457878947258,
0.10111881047487259,
0.07393968105316162,
0.09697818756103516,
0.06700564175844193,
0.08263137936592102,
-0.09771429747343063,
0.06348512321710587,
-0.0074924323707818985,
0.05356649309396744,
0.010868411511182785,
0.006938552018254995,
-0.032839346677064896,
-0.1092953085899353,
0.0022745979949831963,
-0.009965001605451107,
0.01978400908410549,
0.016838017851114273,
0.09132330864667892,
-0.032028473913669586,
-0.01608734019100666,
-0.011086485348641872,
-0.014036137610673904,
0.016068696975708008,
-0.09577742218971252,
-0.0007281387806870043,
0.004879043437540531,
-0.022816771641373634,
0.016201069578528404,
0.011466028168797493,
0.05162842944264412,
0.012380915693938732,
0.007063091732561588,
0.08697136491537094,
-0.0872778668999672,
0.0579158291220665,
0.059308506548404694,
-0.045993879437446594,
-0.04004037752747536,
0.07380066066980362,
-0.025448719039559364,
0.01068155374377966,
0.015098903328180313,
-0.07851383835077286,
-0.01863866299390793,
0.11017310619354248,
0.027853738516569138,
0.025093749165534973,
0.05276069790124893,
-0.07403270155191422,
0.005899390671402216,
-0.04053640738129616,
0.01391083188354969,
0.08246543258428574,
-0.02820710651576519,
0.05485767871141434,
0.038488566875457764,
-0.009483429603278637,
-0.014837340451776981,
-0.008967483416199684,
-0.06696394085884094,
1.8722726680340294e-33,
0.06568583101034164,
0.02585460990667343,
0.052836909890174866,
0.03141496703028679,
0.09917853772640228,
-0.03221338242292404,
0.03133107349276543,
0.04941403493285179,
-0.009448368102312088,
0.058189842849969864,
-0.07370785623788834,
-0.0077104102820158005,
0.029585976153612137,
0.059176355600357056,
0.07560959458351135,
0.015489092096686363,
0.03415483236312866,
-0.04638608917593956,
-0.04841892421245575,
-0.09153233468532562,
0.06183348223567009,
-0.053165290504693985,
0.035742003470659256,
0.03503040596842766,
0.0018906618934124708,
0.031240301206707954,
-0.030714411288499832,
0.01605515368282795,
0.019367244094610214,
-0.0273267962038517,
-0.06206020340323448,
0.02309517189860344,
0.045692551881074905,
0.07996731251478195,
0.044772855937480927,
-0.07184723019599915,
-0.01225681696087122,
-0.03008136712014675,
0.008302654139697552,
-0.060396600514650345,
0.023439601063728333,
-0.002030798699706793,
-0.01247161440551281,
0.00983919482678175,
-0.04459376260638237,
-0.014481943100690842,
-0.03464866429567337,
0.0019083624938502908,
0.001911337487399578,
-0.06453481316566467,
0.03425820544362068,
-0.02069729007780552,
-0.04013249650597572,
0.0386049747467041,
0.03647608309984207,
-0.07334655523300171,
0.03808317705988884,
0.04302399978041649,
0.041154567152261734,
0.07101903110742569,
-0.008826169185340405,
0.0034990727435797453,
0.03806062042713165,
0.060330260545015335,
-0.02512511797249317,
0.07522539049386978,
-0.1140962764620781,
0.0023714075796306133,
0.07097001373767853,
-0.06462836265563965,
0.039289698004722595,
0.07484190911054611,
-0.11656387895345688,
-0.05254356190562248,
-0.00737570459023118,
0.003456676146015525,
-0.019289502874016762,
-0.12277980148792267,
-0.1027575209736824,
0.0684691071510315,
-0.05834992602467537,
-0.028121814131736755,
0.05191385746002197,
-0.01197243481874466,
-0.007731000892817974,
0.003566654399037361,
0.06990256905555725,
-0.0449109748005867,
0.026678474619984627,
-0.04672892391681671,
-0.04163096845149994,
-0.09383316338062286,
-0.036017753183841705,
-0.09926265478134155,
0.03959745541214943,
-1.5042817737862122e-33,
-0.004725954495370388,
0.11163868755102158,
-0.02070505917072296,
0.035035792738199234,
-0.005384540650993586,
-0.09569543600082397,
-0.03860504925251007,
-0.04158218204975128,
-0.03856373578310013,
0.0756903812289238,
-0.08302033692598343,
-0.022026916965842247,
0.022221015766263008,
-0.028308283537626266,
-0.010269283317029476,
-0.011945119127631187,
-0.05436282977461815,
-0.02830353006720543,
-0.01351437158882618,
0.012207810766994953,
0.0032189826015383005,
0.11402540653944016,
-0.13936969637870789,
0.010967066511511803,
0.02976849116384983,
0.05144777148962021,
-0.054961878806352615,
0.033262934535741806,
0.047693151980638504,
0.01515358779579401,
-0.0392986461520195,
-0.030477579683065414,
-0.04005284979939461,
-0.0365668423473835,
0.09059704095125198,
0.029241492971777916,
0.08209531009197235,
0.02171417884528637,
-0.04085536301136017,
-0.041159737855196,
-0.0027577427681535482,
-0.020283140242099762,
-0.04292255640029907,
0.03805207833647728,
0.054602187126874924,
0.040209706872701645,
-0.010890762321650982,
0.03904764726758003,
-0.09946482628583908,
0.0008396424236707389,
-0.02529878169298172,
0.00004350191738922149,
-0.0023354515433311462,
0.0014231043169274926,
-0.01662014052271843,
-0.0756370946764946,
-0.03830039128661156,
0.09562204033136368,
0.013836062513291836,
-0.01301182247698307,
-0.05506131052970886,
-0.06574665755033493,
0.002454062458127737,
-0.03227711468935013,
-0.013523445464670658,
-0.03740018978714943,
-0.03088824264705181,
-0.03723972290754318,
-0.007010419853031635,
0.017654098570346832,
0.026261216029524803,
-0.06354985386133194,
-0.07352498173713684,
0.023811709135770798,
-0.017450399696826935,
0.03685001656413078,
-0.012457984499633312,
0.032995257526636124,
-0.0855325385928154,
-0.011779924854636192,
-0.019235726445913315,
-0.08089260756969452,
0.024965401738882065,
0.0430319681763649,
0.0449112169444561,
0.018976708874106407,
-0.0035903325770050287,
0.026264367625117302,
0.03240499645471573,
0.02628117986023426,
-0.005153287202119827,
0.009712633676826954,
-0.029325541108846664,
0.05022358149290085,
0.036925237625837326,
-3.909199008944597e-8,
-0.1567375659942627,
0.0010409518145024776,
0.031474098563194275,
-0.02491898089647293,
0.09905333817005157,
0.03818909451365471,
0.05465547740459442,
0.10990191996097565,
-0.03839388117194176,
0.02723449096083641,
0.05450347810983658,
-0.032763078808784485,
-0.013394242152571678,
-0.01690412499010563,
-0.054607540369033813,
-0.035605691373348236,
0.05009472370147705,
-0.022946635261178017,
0.03340116888284683,
0.02113136649131775,
-0.02559833787381649,
-0.00834933202713728,
0.0021614779252558947,
0.03846339136362076,
-0.030612407252192497,
-0.1535479575395584,
-0.015112478286027908,
0.023426584899425507,
-0.04342885687947273,
0.03241022303700447,
0.03014932945370674,
0.0105037996545434,
0.10391150414943695,
-0.0021514086984097958,
0.03422589227557182,
0.1022905632853508,
0.010981258936226368,
0.00399926770478487,
-0.06793563067913055,
-0.07121707499027252,
-0.03174343332648277,
0.035520099103450775,
-0.06344565749168396,
0.009300293400883675,
-0.008224493823945522,
-0.06255701184272766,
-0.010933241806924343,
-0.05859364569187164,
0.04496989771723747,
-0.01708403043448925,
-0.006622915156185627,
-0.05560007691383362,
-0.06459881365299225,
-0.040150124579668045,
-0.023096075281500816,
0.057288385927677155,
-0.05201905593276024,
-0.06078838184475899,
0.05782945454120636,
0.053592901676893234,
-0.040208712220191956,
0.023272614926099777,
-0.019842112436890602,
-0.08761437982320786
] |
Towards hyperparameter-free optimization with differential privacy | https://openreview.net/forum?id=2kGKsyhtvh | [
"Ruixuan Liu",
"Zhiqi Bu"
] | Spotlight | Differential privacy (DP) is a privacy-preserving paradigm that protects the training data when training deep learning models. Critically, the performance of models is determined by the training hyperparameters, especially those of the learning rate schedule, thus requiring fine-grained hyperparameter tuning on the data. In practice, it is common to tune the learning rate hyperparameters through the grid search that (1) is computationally expensive as multiple runs are needed, and (2) increases the risk of data leakage as the selection of hyperparameters is data-dependent. In this work, we adapt the automatic learning rate schedule to DP optimization for any models and optimizers, so as to significantly mitigate or even eliminate the cost of hyperparameter tuning when applied together with automatic per-sample gradient clipping. Our hyperparameter-free DP optimization is almost as computationally efficient as the standard non-DP optimization, and achieves state-of-the-art DP performance on various language and vision tasks. | Differential privacy, optimization, hyper-parameter tuning | We introduce a hyperparameter-free differential privacy training method that automatically adjusts the learning rate, reducing the need for extra tuning efforts in a privatized, efficient, and scalable manner for language and vision tasks. | 8,016 | 2503.00703 | [
-0.15727250277996063,
-0.044225871562957764,
0.019490566104650497,
0.05713558569550514,
0.00842832401394844,
-0.0006639519706368446,
-0.03601863980293274,
-0.06376765668392181,
0.032536253333091736,
-0.021682679653167725,
0.020558388903737068,
-0.02978181652724743,
-0.02797924354672432,
-0.02098691090941429,
-0.025340743362903595,
0.0189240500330925,
0.06329015642404556,
0.12050949037075043,
-0.10966968536376953,
-0.03472248464822769,
0.0047172741033136845,
-0.02487415075302124,
0.037797968834638596,
0.0012238665949553251,
0.0545118972659111,
-0.02152084931731224,
0.05923242121934891,
-0.03831013664603233,
0.010541113093495369,
-0.0003879486757796258,
0.0977512001991272,
-0.09015432745218277,
-0.03487599268555641,
0.03196446970105171,
0.021537085995078087,
-0.015112618915736675,
-0.06598223745822906,
0.027012400329113007,
0.012595843523740768,
-0.030349405482411385,
-0.015588926151394844,
-0.056472014635801315,
-0.06790713965892792,
-0.0026774993166327477,
0.003943468909710646,
0.022098859772086143,
0.05329618230462074,
0.004995489493012428,
-0.007694761268794537,
-0.00553507637232542,
-0.02703496813774109,
-0.051481183618307114,
-0.06583093851804733,
0.014799021184444427,
-0.03244125470519066,
-0.022418811917304993,
-0.014342918992042542,
-0.032665204256772995,
-0.0431949682533741,
0.04835628345608711,
-0.08781731128692627,
-0.11386092007160187,
-0.03548712655901909,
0.016142966225743294,
-0.04100700467824936,
-0.0006265976116992533,
0.04780365526676178,
0.012584480457007885,
0.0323714017868042,
-0.00113636942114681,
-0.019101792946457863,
0.17435060441493988,
-0.018211303278803825,
-0.011924338527023792,
0.0053078592754900455,
0.06889183819293976,
0.0014436041237786412,
-0.0362478643655777,
0.04325736686587334,
-0.08542805910110474,
0.021659931167960167,
0.0012538466835394502,
0.03011818788945675,
0.0340832956135273,
0.0812729224562645,
-0.05005892738699913,
0.007793329190462828,
0.031998589634895325,
0.051088329404592514,
-0.0013615259667858481,
-0.013452079147100449,
0.00232185167260468,
-0.053418196737766266,
-0.019058501347899437,
-0.015440373681485653,
-0.01559437531977892,
-0.03206072002649307,
-0.06887161731719971,
-0.020568018779158592,
0.04205955192446709,
-0.052359145134687424,
0.06135289743542671,
0.04956226423382759,
0.04115549847483635,
0.030776970088481903,
0.05778877064585686,
0.023399528115987778,
0.038668375462293625,
0.0446845106780529,
0.017366627231240273,
0.08791155368089676,
0.049323778599500656,
-0.00812064204365015,
-0.0018068463541567326,
0.01435244083404541,
-0.019400501623749733,
-0.015502563677728176,
-0.040723446756601334,
0.07904966175556183,
0.07371702045202255,
-0.06696462631225586,
-0.007357615977525711,
-0.00915317703038454,
0.032506268471479416,
-0.013675541616976261,
-0.00027651211712509394,
-0.10533930361270905,
3.2601318282758713e-33,
0.004785774741321802,
0.033428631722927094,
-0.031135929748415947,
-0.11370082199573517,
-0.010313654318451881,
0.0372471958398819,
0.03859155997633934,
0.08647982776165009,
-0.0800657719373703,
0.04052085429430008,
-0.048633504658937454,
-0.05898777022957802,
-0.014357922598719597,
0.0418158695101738,
0.016499172896146774,
0.02071608044207096,
-0.0023049325682222843,
0.08747798204421997,
0.04672853276133537,
-0.01962977834045887,
0.12010674923658371,
-0.007779921870678663,
-0.010580195114016533,
0.0010648880852386355,
0.010009673424065113,
-0.02210436947643757,
0.08763574808835983,
-0.0162101611495018,
0.04455764964222908,
0.04973317310214043,
-0.05742994695901871,
0.04448249936103821,
0.011411993764340878,
0.04763513430953026,
0.12513037025928497,
0.050541047006845474,
-0.03553168475627899,
0.030970754101872444,
0.09376578778028488,
0.027162974700331688,
0.03854375705122948,
-0.014432541094720364,
0.05523024871945381,
-0.05367456004023552,
-0.11582669615745544,
-0.03961880877614021,
-0.02939499542117119,
0.05767581984400749,
-0.01864808052778244,
-0.051582563668489456,
-0.00991595908999443,
0.028956778347492218,
-0.09482342004776001,
-0.0426076203584671,
-0.04623163864016533,
-0.0295247845351696,
0.04612495005130768,
0.028319116681814194,
0.0007742690504528582,
0.06759565323591232,
-0.032700590789318085,
0.06328420341014862,
0.008774391375482082,
0.00038795938598923385,
-0.021738486364483833,
-0.03388337790966034,
0.018948392942547798,
0.009122896008193493,
0.03491373732686043,
-0.04707546532154083,
-0.008388328365981579,
0.030700912699103355,
-0.032953035086393356,
-0.04251715913414955,
0.040653131902217865,
-0.03341267257928848,
0.08722613751888275,
0.01974647305905819,
-0.002847393974661827,
-0.04307352378964424,
0.039556264877319336,
0.05893454700708389,
-0.014446943067014217,
0.030862463638186455,
-0.025491269305348396,
-0.06852005422115326,
-0.01871076598763466,
0.024704914540052414,
-0.03634944558143616,
0.01367761567234993,
-0.03753171116113663,
-0.06853488087654114,
-0.05031388998031616,
-0.02355370856821537,
-0.06310241669416428,
-2.7905624264116693e-33,
-0.07806398719549179,
0.03212393447756767,
-0.08458635956048965,
0.10135340690612793,
0.032804615795612335,
-0.019590236246585846,
-0.034717023372650146,
0.019279148429632187,
-0.012668130919337273,
-0.03427503630518913,
-0.04638497903943062,
0.03140491247177124,
0.025710007175803185,
-0.040744103491306305,
0.00783002283424139,
0.07872834801673889,
-0.03997472673654556,
0.044309113174676895,
-0.09289205819368362,
0.025889622047543526,
-0.08307570964097977,
0.086830273270607,
-0.08468049764633179,
0.10878872871398926,
-0.013410093262791634,
-0.03446230664849281,
-0.07037311047315598,
0.08128543943166733,
0.061626043170690536,
0.023744814097881317,
-0.06805811077356339,
0.01390924584120512,
-0.1058640331029892,
-0.1509707123041153,
-0.04990926384925842,
-0.006736801005899906,
0.009793740697205067,
0.039422210305929184,
0.0018227107357233763,
0.0936659574508667,
0.04918332025408745,
-0.0022219489328563213,
-0.08758840709924698,
0.011391356587409973,
0.0019543073140084743,
0.0155755914747715,
-0.03826337680220604,
-0.03976352885365486,
-0.024634897708892822,
0.05012087896466255,
-0.06945991516113281,
0.034269675612449646,
-0.003896346315741539,
0.059502068907022476,
-0.07467296719551086,
-0.03042205050587654,
0.019900400191545486,
0.03094252198934555,
0.03596305847167969,
0.08438363671302795,
-0.002880495274439454,
0.006833303719758987,
-0.04426560178399086,
0.0030424129217863083,
-0.026376133784651756,
-0.06407878547906876,
-0.029438626021146774,
0.02195451408624649,
-0.014021321199834347,
0.015896771103143692,
0.0224846750497818,
-0.06947176903486252,
0.009284453466534615,
0.04310676455497742,
-0.024930525571107864,
-0.012997802346944809,
0.0558139942586422,
-0.005905788857489824,
-0.000127572740893811,
0.08969210833311081,
0.1385139673948288,
-0.00362524576485157,
0.02734312228858471,
-0.03484988585114479,
0.04478653892874718,
0.005852625705301762,
0.12082487344741821,
0.0039125909097492695,
-0.04192475974559784,
0.00523274065926671,
-0.006896874867379665,
-0.012591331265866756,
-0.06678435206413269,
0.04296356067061424,
-0.026946336030960083,
-3.766255929349427e-8,
0.022844597697257996,
0.03749896585941315,
0.02595169097185135,
0.07334193587303162,
0.049320630729198456,
-0.010277721099555492,
0.027638845145702362,
0.08211012929677963,
-0.08577726781368256,
-0.028015512973070145,
0.1411319077014923,
-0.019142525270581245,
-0.04427454620599747,
-0.04776385799050331,
-0.0312710665166378,
0.08349990844726562,
0.0070000821724534035,
-0.03978128731250763,
-0.0004114357871003449,
0.0013618131633847952,
-0.08149122446775436,
0.0263910461217165,
-0.02550339139997959,
-0.01341711264103651,
0.0674336776137352,
0.008915678597986698,
0.02960910275578499,
-0.021002862602472305,
0.1408471167087555,
0.09213397651910782,
-0.02815314568579197,
-0.008467000909149647,
0.008663513697683811,
-0.021158546209335327,
-0.05242732912302017,
0.14714039862155914,
-0.034123267978429794,
-0.04064609482884407,
0.03769151121377945,
0.02286931872367859,
0.07025925815105438,
0.05842147767543793,
0.010276712477207184,
0.009633256122469902,
-0.01287938468158245,
-0.01774878241121769,
0.04304833337664604,
-0.12685996294021606,
0.01527565997093916,
-0.0011805958347395062,
0.03225918114185333,
-0.022941045463085175,
-0.014549088664352894,
0.07430889457464218,
0.02682315558195114,
-0.04456544294953346,
-0.01690078154206276,
-0.07543475925922394,
0.033489737659692764,
0.11554298549890518,
0.026963310316205025,
0.003553708316758275,
-0.11417510360479355,
-0.0012304135598242283
] |
AgentTrek: Agent Trajectory Synthesis via Guiding Replay with Web Tutorials | https://openreview.net/forum?id=EEgYUccwsV | [
"Yiheng Xu",
"Dunjie Lu",
"Zhennan Shen",
"Junli Wang",
"Zekun Wang",
"Yuchen Mao",
"Caiming Xiong",
"Tao Yu"
] | Spotlight | Graphical User Interface (GUI) agents hold great potential for automating complex tasks across diverse digital environments, from web applications to desktop software. However, the development of such agents is hindered by the lack of high-quality, multi-step trajectory data required for effective training. Existing approaches rely on expensive and labor-intensive human annotation, making them unsustainable at scale. To address this challenge, we propose AgentTrek, a scalable data synthesis pipeline that generates high-quality web agent trajectories by leveraging web tutorials. Our method automatically gathers tutorial-like texts from the internet, transforms them into task goals with step-by-step instructions, and employs a visual-language model (VLM) agent to simulate their execution in a real digital environment. A VLM-based evaluator ensures the correctness of the generated trajectories. We demonstrate that training GUI agents with these synthesized trajectories significantly improves their grounding and planning performance over the current models. Moreover, our approach is more cost-efficient compared to traditional human annotation methods. This work underscores the potential of guided replay with web tutorials as a viable strategy for large-scale GUI agent training, paving the way for more capable and autonomous digital agents. | Data Synthesis, GUI Agent, Large Language Model | null | 8,010 | 2412.09605 | [
-0.06242460384964943,
-0.050716765224933624,
-0.0316799022257328,
0.0037319448310881853,
0.05382338911294937,
-0.059223275631666183,
-0.025438541546463966,
-0.02552979625761509,
-0.011093630455434322,
-0.052645400166511536,
-0.05100589245557785,
-0.09663610905408859,
0.05163697525858879,
0.022902850061655045,
0.044460490345954895,
0.003777370322495699,
0.06146200746297836,
0.0680917352437973,
-0.0445052832365036,
-0.019046515226364136,
0.02948928251862526,
0.01706377975642681,
0.051294684410095215,
-0.01603701151907444,
0.016291821375489235,
0.028897181153297424,
0.0011392241576686502,
-0.056773215532302856,
0.042683087289333344,
0.0008062467677518725,
0.02371460758149624,
-0.02153215929865837,
-0.07463255524635315,
0.02748480625450611,
0.015104771591722965,
0.049285825341939926,
-0.05592527985572815,
0.03659830987453461,
-0.03520852327346802,
0.000866310962010175,
-0.04909702390432358,
-0.022401148453354836,
0.012714561074972153,
0.014710522256791592,
0.04116467759013176,
-0.059752244502305984,
-0.06420302391052246,
-0.08483926206827164,
0.032983094453811646,
0.09460705518722534,
-0.09989012032747269,
-0.10008151084184647,
0.04557698220014572,
0.007952816784381866,
-0.023895524442195892,
0.05653272941708565,
0.026137856766581535,
0.0005499780527316034,
0.046900175511837006,
-0.01662084460258484,
-0.0006053152028471231,
-0.07277259975671768,
0.01322077214717865,
-0.01996438391506672,
-0.07153958827257156,
-0.0735124722123146,
0.03626938536763191,
-0.0025255861692130566,
0.053723808377981186,
-0.0009011341026052833,
-0.11832810193300247,
0.008026937022805214,
0.019129473716020584,
0.03820435702800751,
-0.044632624834775925,
0.02608019858598709,
-0.05536976456642151,
-0.010729181580245495,
0.021009044721722603,
-0.0668933093547821,
-0.027460236102342606,
-0.04710621014237404,
-0.01716834120452404,
0.10595633834600449,
0.1007724478840828,
0.023005781695246696,
0.0778113305568695,
0.0684107169508934,
0.04834461212158203,
0.03753741830587387,
-0.04638572782278061,
-0.01784701645374298,
0.10977841168642044,
-0.0014404773246496916,
-0.025488976389169693,
0.0821889117360115,
-0.025606174021959305,
-0.10201027989387512,
-0.028072133660316467,
0.08141421526670456,
-0.026414044201374054,
0.055108338594436646,
0.013573195785284042,
-0.010126453824341297,
-0.017919447273015976,
0.012778002768754959,
0.06653574109077454,
0.07768013328313828,
0.0061398702673614025,
-0.05563540756702423,
-0.011699321679770947,
0.013241383247077465,
0.03916071727871895,
-0.008977152407169342,
0.10660987347364426,
0.010267527773976326,
-0.008876455947756767,
0.009966172277927399,
-0.012547435238957405,
0.04182076454162598,
0.07636673748493195,
0.024359505623579025,
-0.055457618087530136,
0.024155769497156143,
0.10954928398132324,
-0.03700058534741402,
0.026902135461568832,
1.6378408909136635e-33,
0.16229301691055298,
0.07935381680727005,
0.08398100733757019,
0.07709246873855591,
0.06448827683925629,
-0.012718874961137772,
0.015525361523032188,
0.008380137383937836,
-0.030188361182808876,
-0.02865721844136715,
0.05122948810458183,
-0.02810772694647312,
-0.048421382904052734,
0.10876603424549103,
0.03376176580786705,
-0.059717658907175064,
-0.046728674322366714,
0.06628388911485672,
-0.031242044642567635,
-0.028929097577929497,
0.013030333444476128,
-0.002815697807818651,
0.028480179607868195,
-0.054513029754161835,
0.027413345873355865,
0.06360834091901779,
0.012355160899460316,
0.001167573849670589,
0.01368041429668665,
-0.01640254259109497,
-0.08768653869628906,
-0.037082239985466,
0.017076067626476288,
0.07684065401554108,
-0.038405582308769226,
-0.03411654755473137,
-0.02887372486293316,
-0.07111525535583496,
0.0024122849572449923,
0.07218226045370102,
-0.030261913314461708,
-0.003480302868410945,
0.0031050522811710835,
-0.030266283079981804,
-0.05355903133749962,
-0.09515430778265,
-0.01984846033155918,
0.0636519342660904,
0.00943477638065815,
-0.033853378146886826,
-0.015034008771181107,
0.058030787855386734,
0.03057493455708027,
-0.01424476969987154,
-0.050576046109199524,
0.0013275196542963386,
0.046355411410331726,
0.0019164129626005888,
0.006758525036275387,
0.013648373074829578,
-0.004618460778146982,
0.017530398443341255,
-0.05425005406141281,
0.10161512345075607,
-0.0024092751555144787,
-0.011587081477046013,
0.06051066517829895,
0.09562907367944717,
0.03773584961891174,
0.005897377151995897,
-0.04066683351993561,
0.03428647294640541,
0.03500808775424957,
-0.0038879106286913157,
0.05647408962249756,
-0.03570474684238434,
0.021047066897153854,
-0.13225173950195312,
0.024933135136961937,
-0.043935827910900116,
-0.007306229323148727,
-0.042496275156736374,
-0.0007914809975773096,
-0.05758250132203102,
0.026560483500361443,
-0.06462190300226212,
-0.022767743095755577,
-0.08060263842344284,
-0.014736433513462543,
0.0024915675166994333,
-0.09893503040075302,
-0.03323681652545929,
-0.10265541821718216,
0.017110787332057953,
0.03704318404197693,
-1.5733973518657087e-33,
0.017652662470936775,
-0.013594919815659523,
-0.08296985179185867,
0.07778172194957733,
0.002592642791569233,
-0.04535006359219551,
-0.015049390494823456,
0.008206496946513653,
0.017873326316475868,
-0.07965889573097229,
-0.07724444568157196,
-0.004452064633369446,
-0.005177509505301714,
0.010797222144901752,
0.0023725845385342836,
0.0023074522614479065,
0.0014692040858790278,
-0.017789149656891823,
-0.02784464880824089,
-0.06765797734260559,
-0.008028763346374035,
0.0663365051150322,
-0.08196666091680527,
-0.0439295619726181,
0.012418895959854126,
0.006382999010384083,
0.06523530930280685,
0.043919630348682404,
0.015985028818249702,
0.020776530727744102,
0.05379243940114975,
-0.01650814339518547,
0.017122235149145126,
-0.06655960530042648,
0.008542404510080814,
0.08287330716848373,
0.0530649833381176,
-0.01389270182698965,
-0.028398700058460236,
0.09502733498811722,
0.08451139181852341,
-0.09733444452285767,
-0.02956392802298069,
-0.03411204367876053,
-0.014833134599030018,
0.030347801744937897,
-0.09849094599485397,
0.05166436731815338,
0.002660268684849143,
0.014257330447435379,
-0.015250968746840954,
0.03554375097155571,
-0.07350349426269531,
-0.0746704712510109,
-0.04333661124110222,
-0.10547806322574615,
0.010979793965816498,
-0.0522698350250721,
0.04661739617586136,
-0.08186152577400208,
-0.09466618299484253,
0.010806172154843807,
0.05432409048080444,
0.01581519842147827,
-0.024169841781258583,
-0.042111095041036606,
-0.086554154753685,
0.03036058321595192,
-0.025327278301119804,
0.000037655099731637165,
0.04812092334032059,
-0.030817575752735138,
-0.025700485333800316,
0.011551999486982822,
0.06611035764217377,
-0.06274331361055374,
0.04000314325094223,
-0.005533092655241489,
0.0228588804602623,
-0.06744714826345444,
0.044818535447120667,
0.05022699385881424,
0.056651897728443146,
-0.006461028475314379,
0.01065127458423376,
0.13178564608097076,
-0.06107853725552559,
0.13012783229351044,
0.0004487427941057831,
0.0552593395113945,
-0.02589118294417858,
0.052770599722862244,
-0.0336870402097702,
0.07173887640237808,
-0.08127540349960327,
-3.777016743811146e-8,
-0.09899971634149551,
0.0921633318066597,
0.003072418039664626,
0.05885923653841019,
-0.024129537865519524,
0.01899472437798977,
0.01476480346173048,
0.053523797541856766,
0.0583849661052227,
-0.021106742322444916,
0.056251779198646545,
0.0034981670323759317,
0.023360474035143852,
-0.03104528971016407,
-0.017248287796974182,
0.061309508979320526,
-0.015981292352080345,
-0.004912871401757002,
-0.003820278448984027,
-0.025330204516649246,
0.052237074822187424,
0.03807472810149193,
-0.05311393737792969,
-0.04561077430844307,
0.038691598922014236,
-0.015389704145491123,
-0.017133554443717003,
0.09293655306100845,
0.025309478864073753,
-0.0096813403069973,
0.03893999382853508,
-0.010958222672343254,
0.024182427674531937,
-0.016565904021263123,
0.010609067045152187,
0.05968847498297691,
0.0013205637224018574,
-0.0635053962469101,
0.01636023260653019,
-0.08660236746072769,
0.013687679544091225,
0.06915178149938583,
-0.0001376158179482445,
-0.037607815116643906,
0.04188689589500427,
-0.028741132467985153,
-0.09422047436237335,
-0.10439186543226242,
0.05640597641468048,
-0.11530420184135437,
-0.10903108865022659,
-0.07784207165241241,
0.016575060784816742,
0.033852849155664444,
0.09095127880573273,
0.03486940637230873,
-0.024510672315955162,
-0.019006015732884407,
0.12368255853652954,
0.07866054773330688,
-0.03436660394072533,
0.033935073763132095,
0.004398834891617298,
0.004778270609676838
] |
Bilinear MLPs enable weight-based mechanistic interpretability | https://openreview.net/forum?id=gI0kPklUKS | [
"Michael T Pearce",
"Thomas Dooms",
"Alice Rigg",
"Jose Oramas",
"Lee Sharkey"
] | Spotlight | A mechanistic understanding of how MLPs do computation in deep neural net-
works remains elusive. Current interpretability work can extract features from
hidden activations over an input dataset but generally cannot explain how MLP
weights construct features. One challenge is that element-wise nonlinearities
introduce higher-order interactions and make it difficult to trace computations
through the MLP layer. In this paper, we analyze bilinear MLPs, a type of
Gated Linear Unit (GLU) without any element-wise nonlinearity that neverthe-
less achieves competitive performance. Bilinear MLPs can be fully expressed in
terms of linear operations using a third-order tensor, allowing flexible analysis of
the weights. Analyzing the spectra of bilinear MLP weights using eigendecom-
position reveals interpretable low-rank structure across toy tasks, image classifi-
cation, and language modeling. We use this understanding to craft adversarial
examples, uncover overfitting, and identify small language model circuits directly
from the weights alone. Our results demonstrate that bilinear layers serve as an
interpretable drop-in replacement for current activation functions and that weight-
based interpretability is viable for understanding deep-learning models. | interpretability, mechanistic interpretability, bilinear, feature extraction, weight-based, eigenvector, eigendecomposition, tensor network | The close-to-linear structure of bilinear MLPs enables weight-based analysis that reveals interpretable low rank structure across multiple modalities. | 7,965 | 2410.08417 | [
-0.10143394768238068,
-0.10467594116926193,
0.07625649869441986,
0.018157942220568657,
0.02985827811062336,
-0.01558675430715084,
-0.013196511194109917,
-0.08638869971036911,
0.004271361511200666,
-0.06065789982676506,
-0.05564652383327484,
-0.04275380074977875,
-0.02996835857629776,
0.03661532327532768,
-0.07589979469776154,
0.06514838337898254,
0.07224947214126587,
0.10428524017333984,
-0.17392466962337494,
-0.010057278908789158,
0.10673344135284424,
-0.011140258051455021,
-0.06168074533343315,
0.028603417798876762,
0.01879844442009926,
0.0377790629863739,
-0.022261684760451317,
-0.05366884917020798,
0.004442580509930849,
-0.05230426415801048,
0.03550753742456436,
-0.02322549559175968,
0.01216243114322424,
0.02793073281645775,
-0.09037800133228302,
0.06658317148685455,
-0.031350284814834595,
-0.0010088360868394375,
0.020209401845932007,
-0.07143144309520721,
0.004464340396225452,
-0.007963499054312706,
-0.005006634164601564,
0.003504946595057845,
0.09814314544200897,
-0.03424416854977608,
0.02858692966401577,
-0.12033987045288086,
0.0012890915386378765,
0.031663477420806885,
-0.049441397190093994,
0.000022466256268671714,
0.026360759511590004,
0.0858040601015091,
-0.03711472451686859,
-0.003196877893060446,
0.1146429181098938,
0.06404921412467957,
0.013348238542675972,
-0.021798254922032356,
-0.01544254645705223,
-0.023831743746995926,
0.07402276247739792,
-0.016753608360886574,
-0.01198500394821167,
-0.026613395661115646,
0.02839229628443718,
0.0163424015045166,
-0.01741570420563221,
0.020095692947506905,
0.054245367646217346,
0.021847177296876907,
-0.06243119016289711,
-0.0013560348888859153,
0.0758209154009819,
0.054094348102808,
0.09935709089040756,
-0.02045191079378128,
0.06918051093816757,
-0.03653797507286072,
-0.027312131598591805,
-0.0016671463381499052,
-0.001481659710407257,
0.01931472308933735,
0.13614530861377716,
-0.01864822953939438,
-0.03807235136628151,
0.019760776311159134,
0.0494985431432724,
0.011741682887077332,
-0.05455106496810913,
-0.05464279651641846,
-0.03604862466454506,
0.014117141254246235,
0.09840475022792816,
0.05406821519136429,
-0.02074495702981949,
-0.08929876983165741,
-0.05693027377128601,
0.06720275431871414,
-0.007857775315642357,
0.07533378154039383,
0.038060300052165985,
0.00873484369367361,
-0.03572801128029823,
0.013662032783031464,
0.059975605458021164,
0.01346422266215086,
-0.00038932680035941303,
-0.06657861918210983,
0.021160483360290527,
0.02770047076046467,
0.0432007759809494,
0.047905437648296356,
0.057761602103710175,
-0.008047505281865597,
0.028727110475301743,
-0.041420020163059235,
0.04141988977789879,
0.03870990127325058,
-0.049655135720968246,
0.06370648741722107,
-0.013427347876131535,
0.018827786669135094,
-0.03449735417962074,
0.0688948705792427,
-0.04556430131196976,
3.7796242915485036e-33,
-0.011694170534610748,
0.01901400089263916,
0.06203349307179451,
-0.0015250525902956724,
0.029027072712779045,
-0.10545909404754639,
0.0196884386241436,
0.013856346718966961,
-0.03338353708386421,
-0.006389075890183449,
-0.036662016063928604,
0.05755173787474632,
-0.04487663879990578,
0.07619614899158478,
0.030039042234420776,
0.019039016216993332,
-0.032922785729169846,
-0.03511372581124306,
0.027173174545168877,
-0.03293255716562271,
-0.008171533234417439,
-0.020414777100086212,
-0.024900473654270172,
-0.038093797862529755,
0.05648689344525337,
-0.02170448750257492,
0.050103556364774704,
-0.03856462985277176,
-0.10214491188526154,
0.00804335530847311,
-0.06677265465259552,
-0.011599291115999222,
-0.0038789717946201563,
-0.0210019052028656,
-0.02095658890902996,
-0.06014111265540123,
0.036815620958805084,
-0.10185922682285309,
0.02571677602827549,
-0.02538417838513851,
0.014461472630500793,
-0.002434981521219015,
-0.043298911303281784,
-0.00622608931735158,
-0.05472814664244652,
-0.03218945488333702,
-0.011188874952495098,
0.02749246172606945,
-0.09348217397928238,
-0.059712447226047516,
-0.007324590813368559,
0.06919469684362411,
0.012481463141739368,
-0.08227898180484772,
0.03214681148529053,
-0.018522795289754868,
0.00511977169662714,
0.04217683523893356,
0.08219727873802185,
0.0908103957772255,
-0.017040062695741653,
0.10429858416318893,
0.04328630119562149,
0.10526511818170547,
-0.017899414524435997,
-0.028333673253655434,
-0.038943324238061905,
0.0803849920630455,
0.04754331335425377,
-0.02737439237535,
-0.06353054940700531,
0.03146138787269592,
-0.009762275964021683,
-0.0612344816327095,
0.08331278711557388,
-0.036563027650117874,
0.030805502086877823,
-0.130464568734169,
0.04583887383341789,
0.060427118092775345,
0.012869080528616905,
0.012929975986480713,
0.07874980568885803,
0.042617812752723694,
-0.0999138355255127,
-0.05351990461349487,
0.035843949764966965,
-0.09951965510845184,
-0.001192379044368863,
0.0442974753677845,
-0.08457989245653152,
-0.01955842413008213,
-0.020341262221336365,
0.03072335012257099,
-0.03431512415409088,
-3.098896247671286e-33,
-0.05160873010754585,
0.02583528496325016,
-0.02186865545809269,
-0.004249724093824625,
-0.04478590562939644,
-0.0318271666765213,
0.011168009601533413,
-0.0337311327457428,
0.028496913611888885,
0.010989086702466011,
0.03686899319291115,
-0.02603180892765522,
-0.040407706052064896,
-0.037069544196128845,
0.1351676732301712,
-0.028354382142424583,
-0.06446798890829086,
0.02381374128162861,
0.05084564536809921,
0.017711088061332703,
0.05312807857990265,
0.10672780126333237,
-0.06198885291814804,
0.08934035152196884,
-0.04790080338716507,
0.023563439026474953,
-0.01781349629163742,
-0.0010868124663829803,
0.029037896543741226,
-0.04904048889875412,
-0.0012328906450420618,
0.0490804985165596,
-0.0021803872659802437,
0.023258568719029427,
-0.041227132081985474,
0.08621279150247574,
-0.0004886846872977912,
-0.04487408697605133,
0.019848952069878578,
-0.0028622527606785297,
0.0827292874455452,
0.0037361448630690575,
-0.07361346483230591,
-0.0329292006790638,
-0.05046284198760986,
-0.00826883502304554,
-0.014898656867444515,
0.017101701349020004,
-0.0007481384091079235,
0.010555490851402283,
-0.02755623310804367,
0.02015080489218235,
-0.09891673922538757,
0.05470583960413933,
-0.03615758568048477,
-0.01509089209139347,
-0.030077645555138588,
0.019418245181441307,
0.10118654370307922,
0.004404078703373671,
-0.0335400253534317,
0.008276580832898617,
0.009275262244045734,
-0.05242474377155304,
-0.0029626935720443726,
-0.027480650693178177,
0.016866488382220268,
-0.0011175520485267043,
0.02478337101638317,
-0.005335323978215456,
0.0525798425078392,
-0.02364317886531353,
0.04619242250919342,
0.04673917219042778,
-0.03543971851468086,
-0.029085172340273857,
0.0016500659985467792,
-0.036281973123550415,
0.004935136530548334,
-0.07123570889234543,
0.13239476084709167,
-0.02856745570898056,
0.059440650045871735,
0.013680495321750641,
0.06670455634593964,
0.029884228482842445,
-0.05803397297859192,
0.07006865739822388,
0.051972389221191406,
-0.003980912268161774,
0.0061359223909676075,
-0.024666093289852142,
-0.13063165545463562,
0.035933930426836014,
0.022742627188563347,
-4.438562584141437e-8,
-0.058460693806409836,
0.03799781948328018,
0.04715451970696449,
-0.01432148925960064,
-0.0004070280701853335,
-0.04369122534990311,
0.01023311261087656,
0.05691302940249443,
-0.013292538933455944,
0.03870934247970581,
0.09435895830392838,
-0.03784157708287239,
-0.05230649933218956,
-0.12696240842342377,
-0.037534672766923904,
0.1332535743713379,
-0.01979001797735691,
-0.012855676002800465,
0.046787139028310776,
-0.06501878798007965,
-0.008500183932483196,
0.028035946190357208,
-0.013896855525672436,
-0.004018029198050499,
0.01476428285241127,
-0.1062735766172409,
-0.11304952204227448,
0.10266277939081192,
0.08804035931825638,
-0.0008026352734304965,
0.028678640723228455,
0.012210502289235592,
0.007166903465986252,
0.004538169130682945,
0.0426601842045784,
0.08466134965419769,
-0.00206917617470026,
-0.05571434274315834,
-0.04265128821134567,
0.016024120151996613,
0.0004943170351907611,
0.09926099330186844,
0.004366498440504074,
-0.03275752812623978,
0.04444693773984909,
-0.012753073126077652,
-0.07062122225761414,
-0.08605276793241501,
-0.003690047888085246,
-0.01714758388698101,
0.05266324058175087,
-0.03721855580806732,
0.03261639550328255,
0.11532275378704071,
-0.022777505218982697,
0.03637305274605751,
-0.009083968587219715,
-0.057752229273319244,
0.005591421853750944,
0.07132621854543686,
0.0143976379185915,
-0.012027367949485779,
0.00660064909607172,
-0.023665547370910645
] |
Towards a Unified and Verified Understanding of Group-Operation Networks | https://openreview.net/forum?id=8xxEBAtD7y | [
"Wilson Wu",
"Louis Jaburi",
"jacob drori",
"Jason Gross"
] | Spotlight | A recent line of work in mechanistic interpretability has focused on reverse-engineering the computation performed by neural networks trained on the binary operation of finite groups. We investigate the internals of one-hidden-layer neural networks trained on this task, revealing previously unidentified structure and producing a more complete description of such models in a step towards unifying the explanations of previous works (Chughtai et al., 2023; Stander et al., 2024). Notably, these models approximate equivariance in each input argument. We verify that our explanation applies to a large fraction of networks trained on this task by translating it into a compact proof of model performance, a quantitative evaluation of the extent to which we faithfully and concisely explain model internals. In the main text, we focus on the symmetric group S5. For models trained on this group, our explanation yields a guarantee of model accuracy that runs 3x faster than brute force and gives a >=95% accuracy bound for 45% of the models we trained. We were unable to obtain nontrivial non-vacuous accuracy bounds using only explanations from previous works. | mechanistic interpretability, verification, proof, guarantees, interpretability, equivariance, group theory, representation theory | We investigate how neural networks compute group operations, finding an explanation that unifies those of previous works, then verify this explanation by translating it into a compact proof of model performance. | 7,930 | 2410.07476 | [
-0.14825721085071564,
-0.03587949648499489,
0.07277905941009521,
0.10058163851499557,
0.07953499257564545,
-0.08859151601791382,
-0.018478229641914368,
-0.0526268295943737,
0.034877557307481766,
-0.023444002494215965,
-0.07133423537015915,
-0.04204024001955986,
0.00626546423882246,
-0.0025263363495469093,
-0.15532037615776062,
-0.058074284344911575,
0.053514327853918076,
0.0572948083281517,
-0.13181769847869873,
-0.04941904917359352,
0.01905304566025734,
-0.011564964428544044,
0.020416846498847008,
0.04502245783805847,
0.06175268068909645,
-0.060975853353738785,
-0.09758200496435165,
0.002474590204656124,
-0.018769143149256706,
-0.04843509942293167,
0.012156721204519272,
0.00549622206017375,
0.023753732442855835,
0.02687721699476242,
-0.050059761852025986,
0.12374792248010635,
0.02486264519393444,
-0.018432356417179108,
0.0616568848490715,
0.0013368501095101237,
0.014786867424845695,
0.0063188583590090275,
0.03585078567266464,
0.040738411247730255,
0.10455354303121567,
0.05093858018517494,
0.04026677459478378,
-0.029958343133330345,
-0.02153804525732994,
-0.01723380573093891,
-0.021668877452611923,
0.0110645592212677,
-0.024918382987380028,
0.045474905520677567,
-0.028987254947423935,
-0.04361093416810036,
0.03498583659529686,
0.002118731616064906,
-0.09398087114095688,
-0.06043392792344093,
0.037930455058813095,
-0.08328965306282043,
-0.05216455087065697,
0.0038861942011862993,
-0.015631478279829025,
0.058510228991508484,
0.03818737715482712,
0.0163804292678833,
-0.013839609920978546,
0.019618738442659378,
0.08001603186130524,
0.014960129745304585,
-0.11597013473510742,
0.029674148187041283,
0.047684088349342346,
0.0603659525513649,
0.075531966984272,
0.02582559734582901,
-0.020164797082543373,
-0.09928982704877853,
0.0023450495209544897,
-0.012251104228198528,
0.019024938344955444,
0.019193708896636963,
0.06265733391046524,
-0.023270081728696823,
-0.023642972111701965,
0.04769820719957352,
0.09639210999011993,
0.05235344544053078,
-0.10143551975488663,
-0.019746266305446625,
0.006889495998620987,
-0.027307283133268356,
0.11677370220422745,
0.05288862809538841,
0.031020257622003555,
-0.0403120294213295,
-0.04229258373379707,
0.11467454582452774,
-0.024207929149270058,
0.07897506654262543,
0.05880124121904373,
-0.06939805299043655,
0.04484494775533676,
0.022998610511422157,
0.03229862451553345,
0.01626753993332386,
0.04399581626057625,
-0.08866456896066666,
0.026958923786878586,
0.07322061061859131,
-0.02879451960325241,
0.014058616012334824,
0.035352837294340134,
-0.04830490052700043,
-0.00504844356328249,
0.031953904777765274,
0.0013518619816750288,
0.04218454658985138,
-0.05235612019896507,
0.034309402108192444,
0.033352721482515335,
0.03889012709259987,
0.06557535380125046,
0.0053583597764372826,
-0.1114557683467865,
1.7622498865459173e-33,
-0.006884426344186068,
0.04145508259534836,
0.08049312233924866,
0.06281571835279465,
0.042390525341033936,
-0.010516532696783543,
-0.011125740595161915,
0.0503196120262146,
0.02830517292022705,
0.06651420891284943,
-0.12579438090324402,
0.052733998745679855,
0.006874747574329376,
0.013262997381389141,
0.07089660316705704,
0.05697810649871826,
-0.04416610673069954,
-0.028280168771743774,
-0.033695828169584274,
-0.07935772836208344,
0.04474034532904625,
-0.027997491881251335,
-0.08598577976226807,
-0.06907784193754196,
-0.017271818593144417,
-0.00755044212564826,
0.010709737427532673,
-0.024089409038424492,
-0.03489784896373749,
0.009125697426497936,
-0.07311083376407623,
0.03599148988723755,
0.038092829287052155,
0.025565236806869507,
0.03960874304175377,
-0.007907840423285961,
0.0781799927353859,
-0.04842820391058922,
0.0807952880859375,
-0.014189849607646465,
0.036390867084264755,
-0.010148173198103905,
-0.030748270452022552,
-0.04839799180626869,
-0.05454478785395622,
-0.033997923135757446,
-0.02992580644786358,
0.014044484123587608,
-0.041659753769636154,
-0.053510673344135284,
0.016939256340265274,
0.09008219838142395,
0.022366760298609734,
-0.0950072780251503,
0.006206033751368523,
-0.017529623582959175,
0.06789616495370865,
0.0820787101984024,
0.02016768604516983,
0.10795528441667557,
0.025262385606765747,
0.07707905024290085,
-0.01799655333161354,
0.08852044492959976,
-0.0645526871085167,
0.05186278000473976,
-0.031150447204709053,
0.056466564536094666,
0.09846700727939606,
-0.028768008574843407,
-0.10701126605272293,
0.017705952748656273,
-0.045331571251153946,
-0.028279712423682213,
0.0762707069516182,
0.0065566012635827065,
0.02601606585085392,
-0.09910047054290771,
0.023674895986914635,
0.018695484846830368,
0.010559937916696072,
0.03051496483385563,
-0.02192637510597706,
0.014639968052506447,
-0.09930478036403656,
-0.012635484337806702,
0.04424642026424408,
-0.08339504897594452,
-0.01984269730746746,
-0.01679004542529583,
-0.04464290291070938,
-0.08313780277967453,
0.01940692774951458,
0.011953037232160568,
-0.031083572655916214,
-2.381113499449281e-33,
-0.1034746989607811,
-0.012048792093992233,
-0.031387150287628174,
0.03371765837073326,
-0.03879433870315552,
-0.05683759227395058,
-0.04661955684423447,
-0.0076547530479729176,
-0.023630036041140556,
0.03249257057905197,
0.021158138290047646,
-0.01148701086640358,
0.021694831550121307,
0.0016530074644833803,
0.04047281667590141,
-0.055417995899915695,
-0.03925969824194908,
-0.006355184596031904,
0.08627088367938995,
0.026210317388176918,
0.06657042354345322,
0.08367336541414261,
-0.06768798828125,
0.047870684415102005,
0.03571971133351326,
0.05865243822336197,
-0.02365121617913246,
0.044807810336351395,
0.06464076787233353,
0.0009659621864557266,
-0.04734668880701065,
-0.045081738382577896,
-0.03221074119210243,
-0.02480144612491131,
-0.02692381851375103,
0.03455154970288277,
-0.0006550262914970517,
0.0584726519882679,
-0.005686140153557062,
-0.006271978374570608,
0.08327007293701172,
-0.030962785705924034,
-0.10332081466913223,
0.004697368945926428,
-0.021253548562526703,
-0.015813978388905525,
0.023250380530953407,
0.01674473099410534,
-0.04291438311338425,
0.01758429780602455,
0.06783420592546463,
0.01264636218547821,
-0.016030311584472656,
0.0029541486874222755,
-0.016124527901411057,
-0.010066544637084007,
0.005452388897538185,
-0.05721128359436989,
0.09910131990909576,
0.001727348892018199,
-0.07009711861610413,
-0.008894295431673527,
-0.016068603843450546,
0.04512730613350868,
-0.05434580147266388,
-0.004520368296653032,
-0.033433228731155396,
0.014216196723282337,
0.060715727508068085,
0.0043514445424079895,
0.03678246960043907,
-0.052381813526153564,
-0.016602594405412674,
0.0032499325461685658,
0.020304640755057335,
0.0218228567391634,
-0.006892469711601734,
-0.03189261257648468,
-0.03524767979979515,
-0.04225662350654602,
0.052566464990377426,
-0.02660963498055935,
0.08459341526031494,
0.04773188382387161,
0.04474177584052086,
-0.030389396473765373,
0.041028719395399094,
0.06846556812524796,
0.01895853504538536,
0.03059328906238079,
-0.027317874133586884,
0.008597792126238346,
-0.04646879434585571,
0.06957747787237167,
-0.0516967736184597,
-4.483192483917264e-8,
-0.02824496477842331,
0.08832421153783798,
0.07144090533256531,
0.005565768573433161,
0.060255713760852814,
0.0012807021848857403,
-0.023378320038318634,
0.020130548626184464,
-0.026812637224793434,
-0.008208186365664005,
0.09400603920221329,
-0.039601780474185944,
-0.040469273924827576,
-0.04859507828950882,
-0.04906105622649193,
0.1040278822183609,
-0.01592913456261158,
-0.006754729896783829,
-0.009488534182310104,
0.005104388576000929,
0.026975778862833977,
-0.0008091193158179522,
0.005876959301531315,
-0.056725699454545975,
0.06798382103443146,
-0.08570625633001328,
-0.16798381507396698,
0.04991818964481354,
0.0020545474253594875,
0.0272966418415308,
-0.02663085237145424,
-0.010961653664708138,
0.0063409055583179,
0.033799827098846436,
-0.009264505468308926,
0.10133159905672073,
-0.014278650283813477,
-0.013647119514644146,
0.02250487171113491,
-0.05958592891693115,
-0.061610445380210876,
0.03357895463705063,
-0.00978002231568098,
0.044486574828624725,
0.00313937827013433,
-0.05740341916680336,
0.003035709960386157,
-0.0686168447136879,
-0.04385749250650406,
-0.0316547155380249,
0.01399894431233406,
-0.03617514669895172,
-0.06585758924484253,
0.06704576313495636,
-0.06512586772441864,
0.02257217839360237,
-0.023300839588046074,
-0.09825915843248367,
-0.049733102321624756,
0.06104195490479469,
-0.031399019062519073,
0.05984930321574211,
-0.0271394532173872,
-0.0411125048995018
] |
Approximation algorithms for combinatorial optimization with predictions | https://openreview.net/forum?id=AEFVa6VMu1 | [
"Antonios Antoniadis",
"Marek Elias",
"Adam Polak",
"Moritz Venzin"
] | Spotlight | We initiate a systematic study of utilizing predictions to improve over approximation guarantees of classic algorithms, without increasing the running time. We propose a generic method for a wide class of optimization problems that ask to select a feasible subset of input items of minimal (or maximal) total weight. This gives simple (near-)linear-time algorithms for, e.g., Vertex Cover, Steiner Tree, Minimum Weight Perfect Matching, Knapsack, and Maximum Clique. Our algorithms produce an optimal solution when provided with perfect predictions and their approximation ratio smoothly degrades with increasing prediction error. With small enough prediction error we achieve approximation guarantees that are beyond the reach without predictions in given time bounds, as exemplified by the NP-hardness and APX-hardness of many of the above problems. Although we show our approach to be optimal for this class of problems as a whole, there is a potential for exploiting specific structural properties of individual problems to obtain improved bounds; we demonstrate this on the Steiner Tree problem. We conclude with an empirical evaluation of our approach. | Approximation Algorithm, Predictions, ML-augmented, Combinatorial Optimization | We give near-linear time learning-augmented (1+eta/OPT)-approximation algorithms for a number of problems, including Vertex Cover, Steiner Tree, Minimum Weight Perfect Matching, Knapsack, and Maximum Clique. | 7,901 | 2411.16600 | [
-0.06701923161745071,
0.06428487598896027,
0.04707321897149086,
0.0310877226293087,
0.0660788044333458,
-0.0730946809053421,
0.015600834041833878,
0.014484972693026066,
-0.07162884622812271,
0.05354335159063339,
-0.11429807543754578,
-0.03858262673020363,
0.04170646890997887,
0.08823707699775696,
-0.012130229733884335,
0.07018791139125824,
0.1183343231678009,
0.03714781999588013,
0.019642654806375504,
-0.006507076323032379,
-0.027139008045196533,
-0.0355476438999176,
-0.02834262140095234,
-0.029239578172564507,
0.012681132182478905,
-0.038977935910224915,
-0.007629276253283024,
-0.028211280703544617,
-0.018845222890377045,
0.03572119027376175,
-0.04861290380358696,
-0.07218294590711594,
0.04326283931732178,
-0.03076806291937828,
-0.0853753387928009,
0.03525015339255333,
-0.09884537011384964,
-0.01414048857986927,
0.009147577919065952,
0.013240664266049862,
0.030627990141510963,
0.081154003739357,
-0.009009137749671936,
0.06536027789115906,
-0.060995109379291534,
0.035340338945388794,
0.029867904260754585,
0.03565359115600586,
-0.03985459730029106,
-0.04052899032831192,
-0.006952264811843634,
-0.06930163502693176,
-0.04175907000899315,
-0.005395463202148676,
-0.0651547983288765,
-0.06279762834310532,
0.0035296690184623003,
-0.05494856461882591,
0.013820651918649673,
-0.006028896663337946,
0.018844230100512505,
-0.10855655372142792,
-0.018648913130164146,
-0.03681424260139465,
0.022373972460627556,
-0.02036622166633606,
0.09135496616363525,
-0.0032482510432600975,
-0.059722594916820526,
0.08160774409770966,
0.07511813193559647,
0.024554632604122162,
-0.10208939760923386,
0.06106145679950714,
0.1150360032916069,
0.05893765762448311,
0.1210055723786354,
0.001977961976081133,
-0.05236401408910751,
0.014919995330274105,
-0.07555446028709412,
0.01698880083858967,
-0.004049329552799463,
-0.02508324384689331,
0.05186512693762779,
0.009080621413886547,
-0.013672132976353168,
0.07568799704313278,
0.08355206251144409,
-0.01997583545744419,
-0.08786854147911072,
-0.06593078374862671,
-0.05362141877412796,
0.03234456107020378,
-0.04692573845386505,
0.005976109765470028,
-0.009044775739312172,
-0.002758672460913658,
0.011932811699807644,
0.057390596717596054,
-0.06550758332014084,
0.05316556245088577,
0.0828060582280159,
-0.07109417021274567,
0.0335056409239769,
0.003763222834095359,
-0.01106006558984518,
0.03940776735544205,
-0.019658556208014488,
-0.07972162216901779,
0.0350593663752079,
0.056382402777671814,
0.059633150696754456,
0.03391340747475624,
-0.004026395734399557,
0.014569842256605625,
0.022099705412983894,
-0.007464529015123844,
-0.013643462210893631,
0.1196351945400238,
-0.009889868088066578,
0.04178101196885109,
0.04543437063694,
-0.02391190081834793,
0.025604454800486565,
0.028831033036112785,
-0.025084111839532852,
1.3165286018305109e-33,
-0.020065799355506897,
0.007691825740039349,
0.08041292428970337,
-0.008516106754541397,
0.06465485692024231,
-0.040449339896440506,
0.05271640792489052,
-0.04680762067437172,
-0.004585954360663891,
-0.01743357814848423,
-0.0307108536362648,
-0.06446333974599838,
0.04306729510426521,
0.016615726053714752,
0.09753619134426117,
-0.009401867166161537,
0.02906970866024494,
0.016460003331303596,
-0.012970357201993465,
-0.05149035528302193,
0.060327641665935516,
-0.09265938401222229,
0.02205345220863819,
-0.0026048573199659586,
0.03479744866490364,
-0.02856532856822014,
0.03860706463456154,
-0.018783118575811386,
0.06786473840475082,
-0.03882100433111191,
-0.08656401932239532,
0.03427724912762642,
0.06598221510648727,
0.07353989034891129,
-0.0007250620401464403,
-0.005476955324411392,
0.05859578400850296,
-0.025479352101683617,
-0.011319915764033794,
-0.008971857838332653,
0.001467732130549848,
0.021805429831147194,
0.03602733835577965,
0.0065714833326637745,
0.009860293939709663,
-0.047217171639204025,
-0.030530592426657677,
0.06729553639888763,
-0.07620366662740707,
0.030471928417682648,
-0.032008714973926544,
0.03563130646944046,
-0.02743932232260704,
0.018871070817112923,
-0.061306703835725784,
-0.006908876821398735,
0.034233927726745605,
-0.0006457757554017007,
0.02416113018989563,
0.0409579761326313,
0.07572482526302338,
-0.009531617164611816,
0.02719561569392681,
-0.033480122685432434,
-0.04947034642100334,
0.0705735981464386,
0.0013529635034501553,
-0.07784789800643921,
0.09692185372114182,
-0.036174431443214417,
-0.00045844478881917894,
-0.036433521658182144,
0.021536337211728096,
-0.04164106026291847,
0.060150034725666046,
0.016775889322161674,
0.09574683010578156,
-0.03728390857577324,
-0.015721920877695084,
-0.020706111565232277,
-0.027364833280444145,
0.03982497751712799,
0.07546431571245193,
-0.047036148607730865,
-0.045127928256988525,
-0.12198781967163086,
0.0471319705247879,
0.03966706246137619,
-0.03312041610479355,
-0.05364908650517464,
-0.1278337985277176,
0.07009102404117584,
-0.04474849998950958,
-0.034326743334531784,
0.012754257768392563,
-4.184947233339506e-34,
0.016928846016526222,
-0.03546939045190811,
0.08751882612705231,
0.022061051800847054,
0.08077168464660645,
-0.09539658576250076,
-0.06854863464832306,
-0.10389139503240585,
0.014217409305274487,
-0.01301299873739481,
-0.047768495976924896,
-0.0470828153192997,
0.09379653632640839,
-0.021718159317970276,
-0.058577485382556915,
0.03944678232073784,
-0.049946531653404236,
0.006612905766814947,
-0.02522009238600731,
0.008224588818848133,
0.06417565047740936,
0.08038900792598724,
-0.052817560732364655,
0.09106643497943878,
0.030749782919883728,
-0.005976235494017601,
-0.010980703867971897,
0.060852739959955215,
-0.05996523052453995,
0.004518331028521061,
-0.05421726033091545,
-0.07181942462921143,
-0.07502420991659164,
-0.03374185040593147,
0.015166874043643475,
0.040224745869636536,
0.040967825800180435,
0.13354694843292236,
-0.013078187592327595,
0.05743942782282829,
0.07092645764350891,
-0.02958737313747406,
-0.06837649643421173,
-0.007504109293222427,
-0.018485574051737785,
0.01220038440078497,
-0.037647370249032974,
-0.06885985285043716,
-0.013504082337021828,
-0.026075981557369232,
-0.021450944244861603,
0.015849357470870018,
0.011784199625253677,
-0.0009817847749218345,
0.015828855335712433,
-0.03146785870194435,
-0.09120714664459229,
0.06584887206554413,
0.12637074291706085,
-0.0385165736079216,
-0.04499930888414383,
0.038726773113012314,
-0.016763603314757347,
0.10046707093715668,
0.0074516902677714825,
-0.02913830801844597,
-0.012203466147184372,
-0.059963785111904144,
-0.013203471899032593,
0.05802437290549278,
-0.08868416398763657,
0.043269701302051544,
0.003025611862540245,
0.01937292329967022,
-0.020763428881764412,
0.07035133987665176,
0.044022951275110245,
0.07928311824798584,
0.00006678055069642141,
0.016289513558149338,
-0.021880902349948883,
0.01572645641863346,
0.06288478523492813,
-0.0052878460846841335,
0.06774549186229706,
0.02608306333422661,
0.08042583614587784,
0.04527073726058006,
0.04119814932346344,
-0.009531065821647644,
-0.06301276385784149,
-0.043959811329841614,
0.05779438838362694,
0.016824979335069656,
-0.03316768258810043,
-4.3736164911933884e-8,
-0.04132688418030739,
0.0024873274378478527,
-0.06840813905000687,
0.023512259125709534,
0.08583290874958038,
-0.02439124882221222,
0.07379931956529617,
0.06900454312562943,
-0.02653348445892334,
0.0460515022277832,
0.0355084054172039,
0.027881689369678497,
0.04960070550441742,
0.08078023046255112,
-0.047457847744226456,
0.025417545810341835,
-0.048016466200351715,
-0.04437613487243652,
-0.0021980097517371178,
-0.06410960108041763,
-0.056907545775175095,
0.012822577729821205,
0.037598978728055954,
0.06322220712900162,
0.020281296223402023,
-0.03239525109529495,
-0.03367675840854645,
-0.03737030178308487,
0.04041251912713051,
0.110287144780159,
-0.09402308613061905,
-0.007596749346703291,
0.049467891454696655,
-0.04464277625083923,
0.04915964975953102,
0.058913424611091614,
-0.09283911436796188,
0.04929586127400398,
-0.07929044961929321,
-0.005777823273092508,
0.041465289890766144,
-0.0032644872553646564,
-0.03623298928141594,
-0.037956610321998596,
0.10226073116064072,
-0.05334283784031868,
-0.04514344036579132,
0.057381097227334976,
0.006047502160072327,
-0.07380253076553345,
0.0205567367374897,
0.004853852093219757,
-0.040366604924201965,
-0.062155093997716904,
0.03786635771393776,
0.011654563248157501,
-0.002256094478070736,
-0.0923316478729248,
0.08410719782114029,
-0.06936241686344147,
0.0004714336828328669,
-0.061525970697402954,
-0.014680256135761738,
-0.006673465482890606
] |
Bayesian Experimental Design Via Contrastive Diffusions | https://openreview.net/forum?id=h8yg0hT96f | [
"Jacopo Iollo",
"Christophe Heinkelé",
"Pierre Alliez",
"Florence Forbes"
] | Spotlight | Bayesian Optimal Experimental Design (BOED) is a powerful tool to reduce the cost of running a sequence of experiments.
When based on the Expected Information Gain (EIG), design optimization corresponds to the maximization of some intractable expected *contrast* between prior and posterior distributions.
Scaling this maximization to high dimensional and complex settings has been an issue due to BOED inherent computational complexity.
In this work, we introduce an *pooled posterior* distribution with cost-effective sampling properties and provide a tractable access to the EIG contrast maximization via a new EIG gradient expression. Diffusion-based samplers are used to compute the dynamics of the pooled posterior and ideas from bi-level optimization are leveraged to derive an efficient joint sampling-optimization loop, without resorting to lower bound approximations of the EIG. The resulting efficiency gain allows to extend BOED to the well-tested generative capabilities of diffusion models.
By incorporating generative models into the BOED framework, we expand its scope and its use in scenarios that were previously impractical. Numerical experiments and comparison with state-of-the-art methods show the potential of the approach. | Bayesian Optimal Experimental Design, Conditional Diffusion Models, score based sampling, Bayesian Inverse Problems, Experimental Design, Sampling as Optimization | The paper introduces an efficient BOED method leveraging diffusion-based samplers and bi-level optimization ideas to jointly sample the introduced pooled posterior and maximize Expected Information Gain, enabling larger-scale applications. | 7,853 | 2410.11826 | [
-0.025442341342568398,
-0.05645754560828209,
0.0844120979309082,
0.0017532756319269538,
0.031655147671699524,
-0.06619438529014587,
-0.03287527710199356,
0.030177194625139236,
0.05673372372984886,
-0.0048325443640351295,
-0.08766636997461319,
-0.007797685917466879,
0.020964032039046288,
-0.008252131752669811,
-0.06751313805580139,
0.04511558264493942,
0.08504389226436615,
0.1275750994682312,
-0.05374155193567276,
-0.016598481684923172,
-0.03203829005360603,
0.04650289565324783,
0.023896193131804466,
-0.09566152095794678,
0.011036546900868416,
0.0299864299595356,
0.06801677495241165,
-0.04263347387313843,
0.04419314116239548,
-0.007212736178189516,
0.07830996811389923,
-0.012209254316985607,
0.03412245959043503,
-0.07622231543064117,
0.011292563751339912,
-0.006681602913886309,
-0.05010242015123367,
-0.005599756725132465,
0.007637247908860445,
0.049651194363832474,
-0.022741667926311493,
0.050763342529535294,
0.06471401453018188,
0.025522656738758087,
-0.00024831280461512506,
-0.07781918346881866,
0.007881076075136662,
-0.001272033085115254,
0.03455706685781479,
0.042362045496702194,
0.020806318148970604,
-0.056819137185811996,
-0.07747302204370499,
-0.019507817924022675,
0.020509235560894012,
-0.001868128776550293,
0.06515491753816605,
-0.030656732618808746,
0.06409315019845963,
-0.01343680452555418,
-0.10015622526407242,
-0.06970583647489548,
0.030460136011242867,
-0.0234474316239357,
0.05337054654955864,
-0.0018968060612678528,
0.0651804730296135,
0.0630221888422966,
-0.0014292034320533276,
0.06802709400653839,
-0.01176601741462946,
0.09627760946750641,
-0.09778353571891785,
-0.00664040632545948,
0.012310031801462173,
0.05507062375545502,
0.02143671177327633,
-0.013830325566232204,
-0.013765241019427776,
-0.08285795152187347,
0.012293574400246143,
-0.013975313864648342,
0.02989291399717331,
-0.023513831198215485,
0.06536688655614853,
0.025901736691594124,
-0.06868711858987808,
0.04432745277881622,
0.0729685053229332,
0.03170246630907059,
-0.037495847791433334,
0.015291223302483559,
-0.188145250082016,
-0.042253248393535614,
-0.006774309556931257,
0.0021375061478465796,
0.004936771467328072,
-0.01801333762705326,
0.049132347106933594,
0.07271841168403625,
0.051224470138549805,
0.0669863224029541,
0.09719885140657425,
-0.006876594386994839,
-0.007884115912020206,
-0.14295579493045807,
0.09286657720804214,
0.05696937069296837,
0.03226088359951973,
-0.02800953947007656,
0.009798417799174786,
-0.007341224700212479,
0.08242800086736679,
-0.0034038731828331947,
-0.018743842840194702,
-0.055977191776037216,
0.030464963987469673,
-0.008913930505514145,
0.008232642896473408,
0.026447748765349388,
0.05299577862024307,
-0.004917622543871403,
-0.04053206741809845,
0.00528170820325613,
0.022844167426228523,
0.023158393800258636,
-0.042888201773166656,
3.521298938708744e-34,
0.0007674107328057289,
-0.04204782843589783,
0.04811907187104225,
0.10835950076580048,
0.02727884240448475,
0.033739302307367325,
0.02091948874294758,
-0.06386459618806839,
-0.05260176211595535,
-0.03262653201818466,
-0.02107403613626957,
0.031475044786930084,
-0.07981374114751816,
0.0518379770219326,
0.09101852774620056,
-0.016186397522687912,
-0.019311681389808655,
0.10191350430250168,
-0.0062802741304039955,
-0.01331693772226572,
0.01133342832326889,
-0.01976349763572216,
-0.04505413398146629,
-0.039408888667821884,
-0.061931073665618896,
-0.0011206027120351791,
0.03326217830181122,
-0.007413812447339296,
-0.05791030824184418,
0.00659539271146059,
-0.12586311995983124,
0.02040311135351658,
-0.06082746013998985,
0.01976710557937622,
0.014506792649626732,
-0.004616913385689259,
0.05305356904864311,
-0.014756271615624428,
-0.05459263175725937,
0.022878214716911316,
0.023104563355445862,
0.019563568755984306,
0.07064507901668549,
0.014192381873726845,
-0.12776440382003784,
-0.10358269512653351,
-0.005293929949402809,
-0.05622636899352074,
0.06974218785762787,
-0.08155427873134613,
-0.01416055765002966,
-0.027518002316355705,
0.045137837529182434,
-0.05601789429783821,
0.01795044355094433,
0.040766943246126175,
0.014565961435437202,
0.009413031861186028,
0.018128421157598495,
0.12015315890312195,
-0.033185601234436035,
0.13597485423088074,
0.00048174819676205516,
0.05687893554568291,
0.028576822951436043,
0.08034753799438477,
-0.05333438515663147,
-0.0006988316308706999,
0.04876156151294708,
-0.004175735171884298,
0.00938214547932148,
0.019377566874027252,
0.012653890065848827,
-0.08718662708997726,
0.05290994793176651,
-0.011880750767886639,
0.021554483100771904,
0.06192852556705475,
0.03797407075762749,
-0.03612954169511795,
-0.000779106980189681,
-0.013117978349328041,
-0.06310795247554779,
0.0230802521109581,
-0.0012362285051494837,
0.0359242744743824,
-0.01651536487042904,
-0.06339989602565765,
-0.11478588730096817,
-0.05471205338835716,
-0.04665949568152428,
-0.0394572950899601,
-0.027285724878311157,
-0.0523189939558506,
0.028565697371959686,
-1.58612207821336e-33,
-0.05928588658571243,
0.007382125593721867,
0.0823979452252388,
-0.004131062887609005,
0.004250098951160908,
-0.042741987854242325,
-0.002741081640124321,
-0.02243710868060589,
0.028225205838680267,
-0.08222042769193649,
-0.04941016808152199,
-0.020369205623865128,
-0.03700738027691841,
-0.02162354439496994,
0.03600022569298744,
0.027055539190769196,
0.0028588720597326756,
0.01357265654951334,
0.011676719412207603,
-0.024734828621149063,
-0.01572251319885254,
0.05579708144068718,
0.016513768583536148,
0.03312385454773903,
-0.05347195640206337,
0.028179358690977097,
-0.03330414369702339,
0.003338714363053441,
0.010789786465466022,
-0.07205107808113098,
-0.11441101133823395,
0.02510177157819271,
-0.057143326848745346,
0.04330163449048996,
-0.1293371468782425,
0.12631751596927643,
0.04166645184159279,
0.04358600080013275,
-0.02946314588189125,
0.020396845415234566,
0.024547751992940903,
-0.015092522837221622,
-0.07262054830789566,
-0.0366881862282753,
0.07102902233600616,
-0.03390861302614212,
-0.016950247809290886,
-0.03485271334648132,
0.003002579091116786,
0.005372317042201757,
-0.03594985976815224,
0.10169483721256256,
-0.04789334535598755,
0.03078385628759861,
-0.10728978365659714,
-0.12270782887935638,
0.005361043848097324,
-0.046311743557453156,
0.1309848129749298,
0.03967372328042984,
-0.0034642009995877743,
0.005548230838030577,
-0.05074788257479668,
-0.040968019515275955,
-0.06315430253744125,
0.0060445950366556644,
0.018284566700458527,
0.023001980036497116,
0.055756550282239914,
-0.030086973682045937,
-0.029989922419190407,
-0.06071888655424118,
0.05666838213801384,
0.03666846081614494,
0.09986671060323715,
0.0026511410251259804,
0.037894561886787415,
-0.012948759831488132,
-0.01414549257606268,
0.01850328966975212,
0.010262107476592064,
0.04154997318983078,
-0.04222983121871948,
0.01424048375338316,
0.047749314457178116,
-0.01684648171067238,
-0.023949170485138893,
-0.017621414735913277,
0.05506948009133339,
-0.09972912818193436,
-0.02247820794582367,
-0.05638529732823372,
0.12081032246351242,
0.05690041184425354,
-0.11277370154857635,
-4.857501423316535e-8,
-0.04468792676925659,
0.01975603774189949,
0.07115723192691803,
0.05819302052259445,
0.05521421507000923,
-0.03569186478853226,
-0.06792459636926651,
0.02923433668911457,
-0.01667984388768673,
-0.07905356585979462,
0.06393061578273773,
-0.006254991982132196,
-0.00689032394438982,
-0.038364212960004807,
-0.0043150694109499454,
0.05662108212709427,
-0.004497914109379053,
-0.027693774551153183,
-0.007950156927108765,
-0.01994611695408821,
-0.06046043708920479,
0.0823286846280098,
0.011889766901731491,
-0.016898253932595253,
0.06444685906171799,
-0.038955338299274445,
-0.0003929173981305212,
-0.03834368661046028,
0.11409011483192444,
0.004274971317499876,
-0.025783833116292953,
0.004087424837052822,
-0.08654238283634186,
0.0670834332704544,
0.02240990847349167,
0.00301336869597435,
-0.053583670407533646,
-0.0076340301893651485,
-0.05939187481999397,
-0.009974167682230473,
-0.060899022966623306,
0.024486038833856583,
-0.034978773444890976,
0.019737400114536285,
0.048417240381240845,
0.07069702446460724,
-0.03098091296851635,
0.03078378736972809,
0.04867112264037132,
0.015911675989627838,
0.05458460748195648,
-0.05319987237453461,
-0.002631777198985219,
-0.007086475379765034,
0.020192859694361687,
0.00009468100324738771,
0.07845360785722733,
0.04771467298269272,
-0.0012351038167253137,
-0.00964801013469696,
0.04698741436004639,
0.09219583868980408,
-0.08610375970602036,
0.02735931985080242
] |
MorphoDiff: Cellular Morphology Painting with Diffusion Models | https://openreview.net/forum?id=PstM8YfhvI | [
"Zeinab Navidi",
"Jun Ma",
"Esteban Miglietta",
"Le Liu",
"Anne E Carpenter",
"Beth A Cimini",
"Benjamin Haibe-Kains",
"BO WANG"
] | Spotlight | Understanding cellular responses to external stimuli is critical for parsing biological mechanisms and advancing therapeutic development. High-content image-based assays provide a cost-effective approach to examine cellular phenotypes induced by diverse interventions, which offers valuable insights into biological processes and cellular states. We introduce MorphoDiff, a generative pipeline to predict high-resolution cell morphological responses under different conditions based on perturbation encoding. To the best of our knowledge, MorphoDiff is the first framework capable of producing guided, high-resolution predictions of cell morphology that generalize across both chemical and genetic interventions. The model integrates perturbation embeddings as guiding signals within a 2D latent diffusion model. The comprehensive computational, biological, and visual validations across three open-source Cell Painting datasets show that MorphoDiff can generate high-fidelity images and produce meaningful biology signals under various interventions. We envision the model will facilitate efficient in silico exploration of perturbational landscapes towards more effective drug discovery studies. | Generative Modelling, Latent Diffusion Model, Cell Painting, Morphology, Drug Response Prediction, Cellular Phenotype, Machine Learning | We present MorphoDiff as the first generative framework capable of producing guided, high-resolution predictions of cellular morphology that generalizes across both chemical and genetic interventions. | 7,849 | null | [
0.014268849976360798,
-0.10929560661315918,
0.05253593996167183,
-0.021010611206293106,
0.02232612669467926,
-0.0697835311293602,
-0.08145154267549515,
0.005198410712182522,
0.09248370677232742,
-0.053156714886426926,
0.008237449452280998,
-0.04859858378767967,
0.015533571131527424,
0.08667107671499252,
-0.04536726325750351,
0.041137926280498505,
-0.06437370926141739,
0.05589495599269867,
-0.08324291557073593,
-0.059599630534648895,
0.01694888062775135,
-0.005505749955773354,
-0.012417545542120934,
-0.017696890980005264,
0.06860655546188354,
0.024034183472394943,
0.02130352146923542,
0.01113372016698122,
-0.01574047841131687,
-0.052802544087171555,
0.04723142459988594,
0.00037623386015184224,
-0.0620795339345932,
0.028429148718714714,
-0.013279912993311882,
0.03494793921709061,
-0.049924999475479126,
0.06187295913696289,
-0.005401303060352802,
-0.06873833388090134,
0.06483296304941177,
0.03361129015684128,
-0.02120349556207657,
0.04353950917720795,
0.05013737827539444,
-0.07928908616304398,
-0.03793973848223686,
-0.08258142322301865,
0.0031581150833517313,
0.022051919251680374,
-0.03885084390640259,
-0.08123376965522766,
-0.06178678944706917,
0.05705837532877922,
0.00752191012725234,
0.010589788667857647,
0.00233199424110353,
-0.003965876530855894,
0.05330532044172287,
-0.015230182558298111,
-0.03304673731327057,
-0.0681665688753128,
0.005588687025010586,
-0.01480830181390047,
0.01655907928943634,
0.0036219139583408833,
0.029958762228488922,
-0.02124149538576603,
-0.05807201936841011,
0.009818452410399914,
-0.008175390772521496,
0.08134101331233978,
-0.01614864356815815,
0.008078995160758495,
0.02620295062661171,
0.049370501190423965,
0.005326454062014818,
0.07170120626688004,
0.047640111297369,
-0.07092319428920746,
0.06773923337459564,
0.08494412899017334,
0.06260847300291061,
0.048430003225803375,
0.03834649547934532,
0.04490665718913078,
0.02237360179424286,
0.05881531164050102,
-0.03772846609354019,
-0.005241067614406347,
0.016751987859606743,
0.026697194203734398,
-0.0973680391907692,
0.03574381768703461,
-0.056663379073143005,
0.0003073124389629811,
0.03200923651456833,
-0.10036911070346832,
0.025296201929450035,
0.06219514459371567,
0.01579010859131813,
-0.04310044273734093,
0.02474488876760006,
0.005526433233171701,
0.10252877324819565,
-0.08602092415094376,
0.08188140392303467,
0.03569020330905914,
0.007370509207248688,
0.01820746809244156,
0.05186880752444267,
-0.0361754409968853,
0.07289966940879822,
0.11525966227054596,
0.058751244097948074,
-0.06600827723741531,
0.0288645401597023,
0.01533389650285244,
0.03494469076395035,
0.0027317535132169724,
-0.037836071103811264,
-0.009485363028943539,
-0.07858782261610031,
-0.031469907611608505,
0.014900227077305317,
0.04437382146716118,
-0.10536588728427887,
1.4072378274636205e-33,
0.05596395954489708,
0.026984116062521935,
0.07138406485319138,
0.13563239574432373,
0.04315318539738655,
-0.01411004364490509,
-0.02759176678955555,
-0.10454434156417847,
-0.05271388590335846,
-0.07251857221126556,
-0.07200823724269867,
0.005478139501065016,
-0.10536953806877136,
0.16251641511917114,
-0.01165519654750824,
-0.03912005573511124,
-0.025131570175290108,
0.06398620456457138,
-0.0018452228978276253,
-0.03294498845934868,
-0.013085062615573406,
-0.06163133308291435,
-0.08124218136072159,
-0.1007653996348381,
-0.032263725996017456,
-0.03674681857228279,
-0.06538403034210205,
-0.06169894337654114,
0.002234339714050293,
0.03481440246105194,
-0.13004036247730255,
0.010875103995203972,
0.020985081791877747,
-0.021399086341261864,
0.0033980407752096653,
-0.002420290606096387,
0.0676925852894783,
-0.10411632061004639,
-0.01591574214398861,
-0.023857783526182175,
0.020005712285637856,
0.024857353419065475,
-0.0019133850000798702,
-0.020183611661195755,
-0.023814788088202477,
0.03616718202829361,
-0.04198087379336357,
0.044690124690532684,
-0.05085613578557968,
0.004779651761054993,
0.013470209203660488,
-0.038948290050029755,
0.06912964582443237,
-0.07876477390527725,
-0.0980469137430191,
-0.0014227539068087935,
-0.03960300236940384,
-0.08887236565351486,
0.012031210586428642,
-0.03419100493192673,
0.0728551372885704,
0.03173094615340233,
-0.04093163087964058,
0.048004068434238434,
0.021749284118413925,
0.012287321500480175,
-0.09618467092514038,
0.019008774310350418,
-0.06435944885015488,
0.10985108464956284,
-0.03267604857683182,
0.037534378468990326,
-0.02027582749724388,
-0.07509511709213257,
0.04966053366661072,
0.038905493915081024,
0.059661153703927994,
0.04011611267924309,
0.012679778039455414,
0.017655013129115105,
-0.04318076744675636,
-0.04476739838719368,
-0.10685081034898758,
-0.05537496507167816,
-0.046900760382413864,
-0.03720363974571228,
0.05790453031659126,
-0.020403742790222168,
-0.13729095458984375,
-0.03766293078660965,
-0.0017086889129132032,
-0.02881520241498947,
-0.006185983773320913,
0.046268247067928314,
-0.04891806095838547,
-2.3311568262492568e-33,
0.006544850766658783,
0.04072686284780502,
0.049544867128133774,
-0.025113172829151154,
0.03260882943868637,
0.042520955204963684,
0.04539288580417633,
0.04983685165643692,
0.11233857274055481,
-0.05683935061097145,
0.0304556991904974,
0.027086252346634865,
-0.01972513645887375,
-0.05290110781788826,
0.005372367333620787,
-0.0700325146317482,
0.036374226212501526,
-0.019097160547971725,
0.002909140894189477,
0.006793109234422445,
-0.07925708591938019,
0.0409306176006794,
-0.019147362560033798,
0.08874011784791946,
-0.05423068627715111,
0.07867994159460068,
-0.022644637152552605,
0.01415021438151598,
-0.06205551326274872,
0.03069271333515644,
-0.06974007934331894,
0.0013323314487934113,
-0.021334366872906685,
-0.036673299968242645,
-0.044258225709199905,
0.07079263031482697,
0.06446437537670135,
-0.054053086787462234,
0.02639596350491047,
-0.015428462997078896,
-0.026913436129689217,
-0.06685240566730499,
-0.11990365386009216,
0.08097370713949203,
-0.01115407794713974,
-0.015176602639257908,
0.0502174012362957,
0.05222225934267044,
0.06184442341327667,
0.05967725068330765,
-0.024394549429416656,
0.07160606235265732,
-0.028824584558606148,
0.04990046098828316,
-0.02939690090715885,
-0.012604247778654099,
-0.03329407796263695,
-0.07154738903045654,
0.012762751430273056,
-0.009009197354316711,
-0.14528602361679077,
-0.042470939457416534,
-0.03293445333838463,
-0.019126955419778824,
-0.07045633345842361,
0.01898208074271679,
-0.027500730007886887,
-0.01852554827928543,
0.03644011542201042,
0.00457370700314641,
0.002350633032619953,
0.06821206212043762,
0.028383763507008553,
0.02769523486495018,
0.04344520717859268,
-0.12849178910255432,
-0.008620680309832096,
-0.045774299651384354,
-0.01641446352005005,
0.01602882519364357,
0.039501119405031204,
0.025888923555612564,
0.010955126024782658,
-0.02927299775183201,
0.0652872696518898,
0.019067876040935516,
0.028219183906912804,
0.04445836320519447,
0.016790097579360008,
0.012256380170583725,
0.07311020791530609,
-0.035568926483392715,
-0.01969054713845253,
0.0646270364522934,
-0.019612587988376617,
-3.878722054651007e-8,
0.02009664475917816,
0.031165385618805885,
0.07781561464071274,
-0.06926557421684265,
-0.034737128764390945,
-0.03432910889387131,
0.017479440197348595,
0.05832558870315552,
0.017202025279402733,
-0.016507882624864578,
0.06470217555761337,
0.004668138921260834,
-0.047297146171331406,
-0.03548351302742958,
0.04181394726037979,
0.11320653557777405,
0.03744630888104439,
-0.0005445390124805272,
-0.02914634346961975,
-0.021294550970196724,
-0.10970088839530945,
-0.018653629347682,
-0.0050913686864078045,
-0.0012141632614657283,
0.09730721265077591,
-0.03374886512756348,
-0.05242355540394783,
-0.019992077723145485,
0.026566797867417336,
0.018481209874153137,
0.09461966156959534,
0.020120378583669662,
0.03292073681950569,
0.034548915922641754,
-0.028363151475787163,
0.08169858157634735,
-0.005276826675981283,
-0.03279092162847519,
-0.04320716857910156,
0.020689085125923157,
-0.02242874912917614,
0.06050818786025047,
0.010445425286889076,
-0.01848052628338337,
0.031628672033548355,
0.038033708930015564,
0.09701306372880936,
-0.00889193918555975,
0.02648504078388214,
-0.006732223089784384,
0.020450977608561516,
-0.04383934661746025,
-0.016251463443040848,
0.054525818675756454,
0.007377798203378916,
-0.02019202522933483,
-0.03508384898304939,
-0.005546751897782087,
0.10538842529058456,
-0.01653340272605419,
0.020845595747232437,
0.014499965123832226,
0.08192924410104752,
-0.014667683281004429
] |
Uncertainty Modeling in Graph Neural Networks via Stochastic Differential Equations | https://openreview.net/forum?id=TYSQYx9vwd | [
"Richard Bergna",
"Sergio Calvo Ordoñez",
"Felix Opolka",
"Pietro Lio",
"José Miguel Hernández-Lobato"
] | Spotlight | We propose a novel Stochastic Differential Equation (SDE) framework to address the problem of learning uncertainty-aware representations for graph-structured data. While Graph Neural Ordinary Differential Equations (GNODEs) have shown promise in learning node representations, they lack the ability to quantify uncertainty. To address this, we introduce Latent Graph Neural Stochastic Differential Equations (LGNSDE), which enhance GNODE by embedding randomness through a Bayesian prior-posterior mechanism for epistemic uncertainty and Brownian motion for aleatoric uncertainty. By leveraging the existence and uniqueness of solutions to graph-based SDEs, we prove that the variance of the latent space bounds the variance of model outputs, thereby providing theoretically sensible guarantees for the uncertainty estimates. Furthermore, we show mathematically that LGNSDEs are robust to small perturbations in the input, maintaining stability over time. Empirical results across several benchmarks demonstrate that our framework is competitive in out-of-distribution detection, robustness to noise perturbations, and active learning, underscoring the ability of LGNSDEs to quantify uncertainty reliably. | Graph Neural Networks, Stochastic Differential Equations, Uncertainty Quantification, Bayesian Machine Learning | We introduce Latent Graph Neural SDEs (LGNSDEs), a model that quantifies uncertainty in graph-structured data while maintaining robustness to input perturbations, supported by theoretical guarantees and empirical evaluation. | 7,838 | 2408.16115 | [
-0.07437864691019058,
-0.09903212636709213,
0.052172545343637466,
0.05685301497578621,
-0.0022966263350099325,
-0.009784543886780739,
-0.04842337220907211,
-0.00947472732514143,
0.014821382239460945,
0.003630717284977436,
-0.024751219898462296,
-0.006414578761905432,
0.060229748487472534,
-0.001830396126024425,
-0.0689370259642601,
0.0543486513197422,
0.0451519712805748,
0.07784031331539154,
-0.03951847180724144,
-0.00640777125954628,
-0.007188132498413324,
0.051440585404634476,
-0.023276006802916527,
-0.012669187039136887,
0.07384011149406433,
0.019830161705613136,
0.0460219569504261,
0.0192248672246933,
-0.0026763041969388723,
-0.022534195333719254,
0.09918076545000076,
-0.08834517002105713,
-0.07278915494680405,
-0.019289525225758553,
0.03288410231471062,
0.017947997897863388,
-0.07004042714834213,
-0.036427684128284454,
0.04928625375032425,
0.020291650667786598,
0.051575951278209686,
0.04239244759082794,
0.02632715553045273,
0.05858771130442619,
0.07438431680202484,
-0.05307963117957115,
-0.006936711259186268,
-0.08386684954166412,
0.030333952978253365,
0.022079486399888992,
0.024125469848513603,
-0.08257374912500381,
-0.04868023097515106,
-0.08775017410516739,
0.016413100063800812,
0.021307403221726418,
0.02265797182917595,
-0.00448157312348485,
-0.00420139916241169,
0.04328211024403572,
0.08607631176710129,
-0.05889054387807846,
0.028667541220784187,
-0.013580954633653164,
-0.03471202030777931,
0.08963994681835175,
0.02329789474606514,
0.03640348091721535,
0.02272855117917061,
0.06103774532675743,
0.012027008458971977,
0.09880449622869492,
-0.07954254746437073,
0.049682509154081345,
0.00022899718896951526,
-0.038795046508312225,
0.09721498191356659,
-0.006627344060689211,
0.016376439481973648,
-0.014084463939070702,
0.02142282761633396,
0.06434198468923569,
0.028042657300829887,
-0.002005737042054534,
0.04300408810377121,
0.05554885417222977,
0.00027463032165542245,
0.05858562886714935,
0.0067997765727341175,
-0.020399469882249832,
-0.059841983020305634,
-0.04089104384183884,
-0.07093553990125656,
-0.020445220172405243,
0.0025509679690003395,
0.011052546091377735,
0.04325026646256447,
-0.03667271137237549,
0.007949593476951122,
0.0534663051366806,
-0.014808518812060356,
0.036105040460824966,
0.021063707768917084,
0.010821754112839699,
-0.011931142769753933,
-0.019399959594011307,
0.005750042386353016,
0.018234798684716225,
0.0020491252653300762,
-0.07346995174884796,
-0.021327005699276924,
0.04168080911040306,
-0.0033426557201892138,
0.08244261890649796,
-0.0144581887871027,
-0.05204329639673233,
0.03964102268218994,
-0.10335295647382736,
0.010486088693141937,
0.02204390987753868,
-0.08346851170063019,
0.07166555523872375,
-0.027707701548933983,
-0.024950724095106125,
0.00397471571341157,
-0.03661437705159187,
0.04102925583720207,
1.1747245856934912e-33,
0.04832594469189644,
-0.04092766344547272,
0.05705215781927109,
-0.007479381747543812,
0.0803016647696495,
0.05564851686358452,
-0.0651869997382164,
-0.014587489888072014,
-0.023391850292682648,
0.039049938321113586,
-0.09033267199993134,
-0.03327050805091858,
-0.08593408763408661,
-0.027390452101826668,
0.05336836352944374,
0.008533361367881298,
-0.034075796604156494,
0.005207787733525038,
0.05557028576731682,
-0.08461400866508484,
0.0482703372836113,
-0.12154728174209595,
-0.05205506086349487,
0.031160417944192886,
-0.005449522752314806,
0.0740894228219986,
0.06575503945350647,
0.04222436994314194,
0.06670483201742172,
-0.01828545145690441,
-0.1398012340068817,
0.010712665505707264,
0.018747646361589432,
0.033725347369909286,
0.002221018308773637,
-0.023712877184152603,
-0.03231462463736534,
-0.015469597652554512,
-0.02879207208752632,
-0.12544415891170502,
0.03455309942364693,
0.04208613559603691,
0.014884723350405693,
0.01717790588736534,
-0.08006685227155685,
-0.04007972404360771,
0.030187735334038734,
-0.16654221713542938,
0.019864674657583237,
-0.08705146610736847,
-0.05122189596295357,
0.025830376893281937,
0.0330844521522522,
0.026073390617966652,
0.009250136092305183,
-0.04138832166790962,
0.09250520914793015,
0.01058986596763134,
-0.046259548515081406,
-0.005297858268022537,
-0.0359027199447155,
0.014781709760427475,
0.05420181527733803,
-0.024294309318065643,
0.07966703176498413,
0.14323852956295013,
-0.044040583074092865,
0.05261094123125076,
0.005578374955803156,
-0.04219266399741173,
-0.01937154494225979,
-0.010801964439451694,
-0.022171970456838608,
-0.10250912606716156,
0.00014113495126366615,
-0.002592402044683695,
-0.020319201052188873,
-0.04168001934885979,
-0.03769593685865402,
0.004148898646235466,
-0.06787628680467606,
-0.13376553356647491,
-0.039847809821367264,
0.10917703062295914,
-0.02879825234413147,
-0.025483692064881325,
0.039794258773326874,
-0.03173490986227989,
0.01550516951829195,
-0.012619253247976303,
-0.022355202585458755,
-0.06973335891962051,
0.03938299044966698,
0.00665199663490057,
-0.018930749967694283,
-2.225835471151457e-33,
-0.031119249761104584,
0.10983507335186005,
0.047354117035865784,
0.028063323348760605,
0.024518216028809547,
-0.0346502922475338,
-0.03836366906762123,
0.03417773172259331,
-0.0032160549890249968,
-0.04152551293373108,
-0.10386361926794052,
-0.05610385909676552,
-0.03654683008790016,
-0.05944123491644859,
0.03817221149802208,
-0.014760762453079224,
-0.02216249704360962,
0.020304609090089798,
-0.0025475542061030865,
-0.06395427882671356,
-0.0027321865782141685,
0.027252530679106712,
-0.02002882771193981,
0.025525381788611412,
0.10984185338020325,
0.0796424075961113,
-0.02841917797923088,
0.055113594979047775,
-0.034900471568107605,
0.0016131688607856631,
-0.026852915063500404,
-0.03949201479554176,
-0.009512328542768955,
-0.04781471565365791,
-0.014155189506709576,
0.06138165295124054,
0.039980851113796234,
0.02677682228386402,
0.011215408332645893,
0.030626334249973297,
-0.00841427966952324,
-0.060906268656253815,
-0.08006533980369568,
-0.01936407759785652,
0.052908822894096375,
0.06668946892023087,
-0.004496690351516008,
-0.013129552826285362,
-0.04387173429131508,
0.012296135537326336,
-0.021359242498874664,
0.09479329735040665,
-0.026432424783706665,
-0.0009261497179977596,
-0.07403674721717834,
-0.027156593278050423,
-0.020236926153302193,
0.005129636265337467,
0.0921202227473259,
0.01358078420162201,
-0.05124691128730774,
-0.06539370119571686,
-0.018351323902606964,
0.014547516591846943,
-0.09560198336839676,
-0.004903534427285194,
-0.09202603250741959,
0.007879109121859074,
0.08073549717664719,
0.05067825689911842,
0.034623902291059494,
-0.026658689603209496,
-0.03654590994119644,
0.04709101840853691,
0.07484033703804016,
-0.07624227553606033,
0.029850032180547714,
0.013364934362471104,
0.06051279231905937,
0.003694183425977826,
0.044401369988918304,
0.0504331961274147,
0.09430193901062012,
0.04549770429730415,
0.02873973734676838,
0.014945607632398605,
0.020774882286787033,
-0.04074569419026375,
0.06523286551237106,
0.032839544117450714,
-0.003951093647629023,
0.06284819543361664,
-0.09385406225919724,
0.030554551631212234,
-0.07804006338119507,
-3.638892920321268e-8,
-0.14347213506698608,
0.0769871398806572,
0.06583597511053085,
0.0006913240649737418,
0.09025116264820099,
-0.01689746230840683,
0.07613500207662582,
0.04067395254969597,
0.02858714386820793,
0.0005808990099467337,
0.07199912518262863,
-0.04548094421625137,
-0.066188283264637,
-0.13187472522258759,
-0.04545878246426582,
0.06553321331739426,
-0.023489192128181458,
-0.008694297634065151,
0.035933367908000946,
-0.004188486374914646,
-0.0032214121893048286,
-0.0014048266457393765,
0.003496347228065133,
0.0016804502811282873,
0.03465375304222107,
-0.06848723441362381,
-0.005412339232861996,
0.08569113165140152,
0.015284723602235317,
0.08745177090167999,
-0.002272080397233367,
-0.002767283236607909,
0.09319643676280975,
0.06609570235013962,
0.01850363425910473,
0.0905589684844017,
0.05397571250796318,
-0.05216449499130249,
-0.025466304272413254,
-0.013616299256682396,
-0.06692946702241898,
0.07176187634468079,
-0.03688247129321098,
0.017861705273389816,
-0.08868245035409927,
0.08175136893987656,
0.04447184130549431,
0.022777298465371132,
0.08709927648305893,
-0.033285532146692276,
-0.010754641145467758,
-0.0233649630099535,
-0.08382224291563034,
0.013455120846629143,
-0.011980550363659859,
-0.010240903124213219,
0.011948416009545326,
-0.10077707469463348,
-0.05582553148269653,
-0.003111241152510047,
-0.0426400825381279,
0.013470230624079704,
0.0002320190833415836,
-0.0035710951779037714
] |
Simplifying Deep Temporal Difference Learning | https://openreview.net/forum?id=7IzeL0kflu | [
"Matteo Gallici",
"Mattie Fellows",
"Benjamin Ellis",
"Bartomeu Pou",
"Ivan Masmitja",
"Jakob Nicolaus Foerster",
"Mario Martin"
] | Spotlight | $Q$-learning played a foundational role in the field reinforcement learning (RL).
However, TD algorithms with off-policy data, such as $Q$-learning, or nonlinear function approximation like deep neural networks require several additional tricks to stabilise training, primarily a large replay buffer and target networks. Unfortunately, the delayed updating of frozen network parameters in the target network harms the sample efficiency and, similarly, the large replay buffer introduces memory and implementation overheads. In this paper, we investigate whether it is possible to accelerate and simplify off-policy TD training while maintaining its stability. Our key theoretical result demonstrates for the first time that regularisation techniques such as LayerNorm can yield provably convergent TD algorithms without the need for a target network or replay buffer, even with off-policy data. Empirically, we find that online, parallelised sampling enabled by vectorised environments stabilises training without the need for a large replay buffer. Motivated by these findings, we propose PQN, our simplified deep online $Q$-Learning algorithm. Surprisingly, this simple algorithm is competitive with more complex methods like: Rainbow in Atari, PPO-RNN in Craftax, QMix in Smax, and can be up to 50x faster than traditional DQN without sacrificing sample efficiency. In an era where PPO has become the go-to RL algorithm, PQN reestablishes off-policy $Q$-learning as a viable alternative. | Reinforcement Learning, TD, Theory, Q-learning, Parallelisation, Network Normalisation | null | 7,833 | 2407.04811 | [
-0.06948775798082352,
-0.03988448902964592,
0.03479202091693878,
0.032928746193647385,
0.019182434305548668,
0.07339256256818771,
-0.03599783405661583,
-0.0793192908167839,
0.01414407603442669,
-0.019799228757619858,
-0.08573073893785477,
0.04112553969025612,
-0.04069188982248306,
-0.0028832475654780865,
-0.03638472408056259,
-0.004781379830092192,
0.05328138545155525,
0.1290876567363739,
-0.07762206345796585,
-0.02711150050163269,
-0.027818385511636734,
0.02865426056087017,
0.08900929987430573,
-0.0183677077293396,
0.05370771139860153,
0.023265328258275986,
-0.027539918199181557,
-0.028745600953698158,
0.053017232567071915,
-0.02496468834578991,
0.03728056699037552,
-0.06466094404459,
-0.04850330203771591,
-0.07247479259967804,
-0.10625452548265457,
0.06016520410776138,
-0.12336675077676773,
0.009236054494976997,
-0.0231896061450243,
0.020690424367785454,
0.013036102056503296,
0.00641470355913043,
-0.02205764874815941,
0.019268041476607323,
0.10129380226135254,
-0.012485003098845482,
0.014656784944236279,
-0.007905847392976284,
0.029489634558558464,
0.029127778485417366,
0.03041895292699337,
-0.01875041238963604,
-0.015975940972566605,
0.06324034929275513,
0.07664141058921814,
0.03757274150848389,
0.06278436630964279,
0.10144484788179398,
-0.03484763577580452,
-0.004838734399527311,
-0.01476483978331089,
-0.12874490022659302,
0.04839393123984337,
-0.026313431560993195,
-0.08283469825983047,
-0.02354460582137108,
-0.005433349870145321,
0.10243014246225357,
0.06543479859828949,
0.018116997554898262,
-0.02802143804728985,
0.09448607265949249,
-0.029302673414349556,
0.02375704050064087,
0.029295790940523148,
0.05367102846503258,
-0.010024162009358406,
-0.023951081559062004,
-0.010041198693215847,
-0.060130126774311066,
0.09109693765640259,
-0.06356319040060043,
-0.0536038801074028,
-0.03608263283967972,
0.09286771714687347,
-0.0026523119304329157,
-0.018735472112894058,
0.040203772485256195,
0.03879506140947342,
-0.04542539641261101,
0.0268937386572361,
0.0656195878982544,
0.03983193635940552,
-0.04985533654689789,
-0.018633823841810226,
-0.03373018279671669,
0.030314810574054718,
-0.0548899807035923,
-0.01786068081855774,
0.1143457219004631,
0.03702981397509575,
0.043817635625600815,
-0.015505950897932053,
0.04377451911568642,
0.02786543034017086,
-0.03782561421394348,
0.05518404021859169,
0.08093685656785965,
0.020608970895409584,
-0.0651240274310112,
0.0504457950592041,
0.02994605526328087,
0.03750627860426903,
0.017262866720557213,
-0.026326892897486687,
0.006059793755412102,
-0.016380295157432556,
-0.06816910952329636,
0.010195615701377392,
0.08866823464632034,
-0.009354163892567158,
-0.02492000162601471,
-0.057405985891819,
0.0423198863863945,
0.03369463235139847,
0.0028055799193680286,
-0.010194691829383373,
3.7003360955595864e-33,
0.04890786111354828,
0.052165936678647995,
0.03148520365357399,
-0.06168149784207344,
0.0289760772138834,
0.008020931854844093,
0.11208188533782959,
-0.033954914659261703,
-0.011725760996341705,
-0.005226471461355686,
0.005393342114984989,
-0.012197847478091717,
-0.07555428147315979,
0.09334857761859894,
0.04422149062156677,
-0.03592175990343094,
-0.08381395041942596,
0.006255524232983589,
0.0365305170416832,
-0.040694259107112885,
0.0446992926299572,
-0.054935626685619354,
-0.047401174902915955,
-0.02696591056883335,
0.0003738933301065117,
-0.01682938262820244,
0.01816442608833313,
0.055184975266456604,
0.06262939423322678,
0.007960856892168522,
-0.014974272809922695,
0.0022350347135215998,
-0.04501233994960785,
0.025357237085700035,
0.054778072983026505,
0.0056232549250125885,
0.013165839947760105,
-0.03157685324549675,
-0.019271627068519592,
-0.010861647315323353,
-0.032660793513059616,
0.04252925142645836,
0.020209038630127907,
0.0025314234662801027,
-0.058572959154844284,
-0.19759544730186462,
0.021266989409923553,
-0.0023144769947975874,
-0.05863947048783302,
-0.07860224694013596,
0.04114789515733719,
0.0292572770267725,
-0.06560403853654861,
-0.10578130185604095,
0.01973610743880272,
-0.03204481303691864,
0.030793530866503716,
0.0716991126537323,
-0.05247533693909645,
0.10464581102132797,
0.09359226375818253,
-0.027593493461608887,
-0.06018740311264992,
0.08518920838832855,
-0.04131331294775009,
0.11575924605131149,
-0.020296761766076088,
0.03305850923061371,
0.016812417656183243,
-0.045168910175561905,
0.02314949408173561,
0.03820662200450897,
0.01565004140138626,
-0.07259432226419449,
0.1046842634677887,
0.027743307873606682,
0.044138044118881226,
-0.09424823522567749,
-0.03483251482248306,
0.00029809336410835385,
0.028042588382959366,
0.05064086616039276,
-0.05693927779793739,
-0.01248963549733162,
-0.041793495416641235,
-0.0872468501329422,
0.011341963894665241,
-0.033567123115062714,
0.010471913032233715,
0.0036774042528122663,
-0.11131912469863892,
-0.01619739457964897,
0.00745015311986208,
-0.025391899049282074,
-0.0008001040550880134,
-2.017639258573506e-33,
0.013650617562234402,
-0.008444794453680515,
-0.034771230071783066,
0.07871649414300919,
-0.022435341030359268,
-0.03039274737238884,
-0.06051982566714287,
-0.03289211168885231,
-0.03912391886115074,
-0.07899245619773865,
-0.05701077729463577,
-0.019281776621937752,
0.0375993549823761,
0.0016124845715239644,
0.027109291404485703,
-0.0449136383831501,
-0.021104758605360985,
0.02438095211982727,
-0.01572805643081665,
0.03366106376051903,
0.03970237076282501,
-0.006622445769608021,
-0.01910698227584362,
0.07848469913005829,
-0.025657780468463898,
-0.00974542647600174,
-0.020069755613803864,
0.08245717734098434,
0.056592293083667755,
-0.006522499956190586,
-0.033888258039951324,
-0.13898171484470367,
-0.057515159249305725,
-0.07003190368413925,
-0.033946942538022995,
0.10444668680429459,
0.10798707604408264,
0.05162821710109711,
-0.06901320070028305,
0.05730057507753372,
0.07035612314939499,
-0.09208739548921585,
-0.05879365652799606,
0.01945991814136505,
0.00414283899590373,
0.04442837834358215,
-0.07637334614992142,
-0.01000794768333435,
-0.03826134651899338,
0.035373639315366745,
0.020671488717198372,
0.03795694559812546,
-0.07610893994569778,
0.01656307838857174,
-0.0017274885904043913,
-0.0832267552614212,
-0.015797557309269905,
0.06774333119392395,
0.041193000972270966,
0.019041825085878372,
-0.0034372012596577406,
-0.0046544624492526054,
-0.06647606194019318,
-0.0034138273913413286,
0.04598715156316757,
0.04875343292951584,
0.0010337206767871976,
0.07040158659219742,
0.02843148447573185,
0.008839980699121952,
-0.010773733258247375,
0.007016919553279877,
0.008455484174191952,
0.041847094893455505,
-0.012973643839359283,
-0.028785506263375282,
-0.030333183705806732,
-0.047133758664131165,
-0.0315604992210865,
0.03162024915218353,
0.03546811640262604,
0.01300953608006239,
0.021377120167016983,
0.0023390138521790504,
0.10123275220394135,
0.11093506217002869,
-0.01123654656112194,
0.10520344972610474,
0.008806261233985424,
-0.011453828774392605,
-0.0011408443097025156,
-0.016789285466074944,
0.015485196374356747,
-0.022179508581757545,
-0.023279402405023575,
-4.533021069619281e-8,
-0.07662070542573929,
0.0458250530064106,
0.05531121417880058,
0.03143922612071037,
0.04660746455192566,
-0.07157675921916962,
0.04191717505455017,
-0.018165891990065575,
-0.0010096841724589467,
-0.04177350923418999,
0.05090194568037987,
0.060997337102890015,
0.01843974180519581,
-0.09588905423879623,
-0.03504927456378937,
0.09679052233695984,
0.020981786772608757,
-0.0650104358792305,
0.03138016164302826,
0.010093123652040958,
0.008269254118204117,
0.05954073369503021,
-0.04222191870212555,
-0.09036076068878174,
0.021960826590657234,
-0.024240709841251373,
0.03133753314614296,
0.06853935122489929,
0.11429005861282349,
-0.029520567506551743,
-0.005252692382782698,
-0.042271845042705536,
0.030334582552313805,
-0.05249542370438576,
-0.00460350327193737,
0.0651257336139679,
0.04201144352555275,
-0.004956772550940514,
0.05053392052650452,
-0.05132177844643593,
-0.03185499832034111,
-0.005028437823057175,
-0.0025571377482265234,
-0.0334567129611969,
0.005384320858865976,
-0.02390652522444725,
-0.08313745260238647,
-0.05583544448018074,
0.03484293445944786,
-0.09221423417329788,
0.05504719167947769,
-0.05053955316543579,
0.04089827090501785,
0.028269806876778603,
0.0854680985212326,
-0.039172299206256866,
0.030918274074792862,
-0.09593675285577774,
0.017398029565811157,
0.09129153192043304,
-0.004974670242518187,
-0.023762647062540054,
-0.055113088339567184,
0.043296873569488525
] |
The Superposition of Diffusion Models Using the Itô Density Estimator | https://openreview.net/forum?id=2o58Mbqkd2 | [
"Marta Skreta",
"Lazar Atanackovic",
"Joey Bose",
"Alexander Tong",
"Kirill Neklyudov"
] | Spotlight | The Cambrian explosion of easily accessible pre-trained diffusion models suggests a demand for methods that combine multiple different pre-trained diffusion models without incurring the significant computational burden of re-training a larger combined model. In this paper, we cast the problem of combining multiple pre-trained diffusion models at the generation stage under a novel proposed framework termed superposition. Theoretically, we derive superposition from rigorous first principles stemming from the celebrated continuity equation and design two novel algorithms tailor-made for combining diffusion models in SuperDiff. SuperDiff leverages a new scalable Itô density estimator for the log likelihood of the diffusion SDE which incurs *no additional overhead* compared to the well-known Hutchinson's estimator needed for divergence calculations. We demonstrate that SuperDiff is scalable to large pre-trained diffusion models as superposition is performed *solely through composition during inference*, and also enjoys painless implementation as it combines different pre-trained vector fields through an automated re-weighting scheme. Notably, we show that SuperDiff is efficient during inference time, and mimics traditional composition operators such as the logical OR and the logical AND. We empirically demonstrate the utility of using SuperDiff for generating more diverse images on CIFAR-10, more faithful prompt conditioned image editing using Stable Diffusion, as well as improved conditional molecule generation and unconditional *de novo* structure design of proteins. https://github.com/necludov/super-diffusion | generative modelling, protein generation, image generation, diffusion models | The principled way to combine the outputs of several diffusion models | 7,783 | null | [
-0.05272304266691208,
-0.10374680161476135,
0.06887903809547424,
0.06350535899400711,
0.03328101709485054,
-0.019828543066978455,
-0.08718899637460709,
-0.0375029481947422,
0.0058935112319886684,
0.06355246156454086,
0.023550402373075485,
-0.020785942673683167,
-0.0435144305229187,
0.07905807346105576,
-0.0484275221824646,
-0.04074952006340027,
0.002031467854976654,
0.03640105575323105,
-0.0970626249909401,
-0.012857011519372463,
0.03098013438284397,
0.04194684326648712,
-0.07030387967824936,
-0.029497651383280754,
0.14362841844558716,
0.05718369781970978,
-0.006913016550242901,
-0.08271624892950058,
0.034862443804740906,
0.03192220628261566,
0.10345686227083206,
0.03863440826535225,
-0.023197930306196213,
0.024557197466492653,
-0.018823325634002686,
0.03678501024842262,
-0.05114798620343208,
0.014797691255807877,
0.04692644625902176,
-0.00010993358591804281,
-0.015498160384595394,
-0.0031114905141294003,
-0.03149044141173363,
-0.01051128190010786,
0.03666571155190468,
0.002245801268145442,
0.006229423452168703,
-0.0732882171869278,
0.0132363336160779,
0.016705654561519623,
0.020757384598255157,
-0.06081506982445717,
-0.03330090269446373,
0.006688424851745367,
0.07703869044780731,
0.035290323197841644,
0.03677262365818024,
0.014974578283727169,
-0.012664040550589561,
-0.008489856496453285,
-0.0943489521741867,
-0.09582030028104782,
-0.0374465249478817,
-0.040777429938316345,
0.02788073942065239,
0.03242183104157448,
-0.0326659195125103,
0.09688127785921097,
-0.012535534799098969,
0.13731348514556885,
0.0255066379904747,
0.08852025866508484,
-0.12619683146476746,
-0.00009714112093206495,
0.0281965434551239,
-0.0405602790415287,
0.0322331078350544,
0.015168419107794762,
0.02789016254246235,
-0.12916548550128937,
0.025411952286958694,
0.03775881975889206,
0.0901917815208435,
-0.08134680241346359,
0.07266481965780258,
0.033968664705753326,
0.019858215004205704,
0.07339151203632355,
0.05888080969452858,
-0.054789911955595016,
0.0011811468284577131,
0.062274109572172165,
-0.055365972220897675,
-0.028237130492925644,
0.020264780148863792,
0.08058805763721466,
0.007028681226074696,
-0.0060046096332371235,
0.06927438825368881,
0.05803954601287842,
0.061191998422145844,
0.004963825456798077,
-0.04969216510653496,
-0.06498591601848602,
0.07942880690097809,
-0.11912710964679718,
0.06864339858293533,
0.03815113380551338,
0.03761909157037735,
0.02927693910896778,
0.03920675814151764,
0.01846080645918846,
0.013401571661233902,
0.016288531944155693,
-0.05361386388540268,
-0.01027242373675108,
0.013486797921359539,
-0.005477176513522863,
0.036789972335100174,
0.03031453862786293,
-0.028779378160834312,
0.03314061090350151,
-0.02645442821085453,
0.058026719838380814,
0.041286733001470566,
-0.035571128129959106,
-0.057664256542921066,
3.420900884559974e-33,
-0.010044422000646591,
-0.02055952697992325,
0.0250637698918581,
0.0370182991027832,
0.039728399366140366,
-0.023076321929693222,
-0.004929135553538799,
-0.02603728696703911,
0.013209720142185688,
-0.024218736216425896,
-0.08499366044998169,
0.041552525013685226,
-0.13687726855278015,
-0.024477265775203705,
0.02015460841357708,
-0.03511371091008186,
-0.12921307981014252,
0.05695352330803871,
0.0250008013099432,
-0.04329834505915642,
0.06401089578866959,
-0.06961613893508911,
-0.013902487233281136,
-0.0447150357067585,
-0.037986546754837036,
-0.05027681589126587,
0.09548158943653107,
-0.08932870626449585,
-0.0005492122145369649,
0.04592427983880043,
-0.11094661056995392,
-0.057473648339509964,
-0.09185168892145157,
-0.007372546941041946,
0.002019248204305768,
-0.02021046169102192,
0.0009128176025114954,
-0.027720555663108826,
-0.0872408077120781,
-0.02112361416220665,
0.0012682833475992084,
0.033876337110996246,
-0.009106369689106941,
-0.03575078397989273,
-0.13967111706733704,
-0.06968148052692413,
0.010668948292732239,
0.04376360774040222,
-0.013814970850944519,
-0.014967065304517746,
0.035161152482032776,
-0.036769792437553406,
0.0007457547471858561,
-0.04288680478930473,
-0.06208506599068642,
0.06549570709466934,
0.03882899507880211,
-0.03624894097447395,
0.0073043592274188995,
0.01093879621475935,
0.02782154269516468,
0.08195197582244873,
-0.010858113877475262,
0.09370551258325577,
0.04867533966898918,
0.008786463178694248,
-0.03853750601410866,
0.0957675576210022,
0.044065624475479126,
0.011414024978876114,
-0.019992368295788765,
-0.04107030853629112,
0.02079644240438938,
-0.1447654515504837,
0.034325409680604935,
-0.016499793156981468,
0.022305509075522423,
-0.02638903632760048,
-0.01745612733066082,
-0.003510681912302971,
-0.026077108457684517,
0.01798195205628872,
-0.014326103031635284,
-0.0871208980679512,
-0.00028975296299904585,
-0.010091138072311878,
-0.02830551378428936,
-0.11218012869358063,
-0.04106241837143898,
-0.0286472849547863,
-0.04806819558143616,
-0.08169325441122055,
0.02358928881585598,
-0.06461077183485031,
-0.01748483255505562,
-3.667973634055495e-33,
-0.02589794062077999,
0.07109079509973526,
0.10197506844997406,
0.01976512372493744,
-0.041283801198005676,
0.01687430590391159,
0.030040288344025612,
-0.011267081834375858,
0.011506570503115654,
-0.0812535360455513,
-0.037726785987615585,
0.051336999982595444,
0.0681171864271164,
-0.01368170976638794,
0.014512242749333382,
-0.048526421189308167,
-0.001323668984696269,
-0.017658084630966187,
0.0741044133901596,
-0.005739414598792791,
-0.057671722024679184,
0.04316335171461105,
-0.028877737000584602,
0.057450953871011734,
-0.03571915253996849,
0.024365317076444626,
-0.03607663884758949,
0.15238143503665924,
0.01479316782206297,
-0.006143098697066307,
-0.028736477717757225,
-0.03040037304162979,
0.006811012513935566,
-0.02783329039812088,
-0.08013074845075607,
0.11296125501394272,
0.011459292843937874,
0.07355248183012009,
-0.06002981588244438,
-0.00037744425935670733,
-0.015795119106769562,
-0.049463771283626556,
-0.04637007415294647,
0.06673124432563782,
0.008311618119478226,
-0.0210159569978714,
-0.02265305072069168,
0.025235751643776894,
0.023982830345630646,
-0.029844287782907486,
-0.00080100167542696,
0.050537772476673126,
-0.07292778789997101,
0.012348583899438381,
-0.02135384827852249,
-0.03747077286243439,
-0.022404493764042854,
-0.03813600167632103,
0.02248520404100418,
-0.08132188767194748,
0.023161092773079872,
-0.03246990591287613,
0.04024558886885643,
-0.0170894768089056,
0.00949733518064022,
-0.05734080448746681,
-0.11068344861268997,
-0.013140277937054634,
-0.023456694558262825,
0.026289695873856544,
-0.004921755753457546,
0.028408024460077286,
0.008732747286558151,
0.024473046883940697,
0.029656236991286278,
-0.01803601160645485,
0.04525778442621231,
-0.08649418503046036,
-0.017164254561066628,
0.01277500856667757,
-0.021396664902567863,
0.0688415914773941,
0.016567878425121307,
-0.02243279665708542,
0.0357247032225132,
0.041893381625413895,
0.0327676497399807,
0.005422673653811216,
0.01765262335538864,
-0.025184590369462967,
-0.08891739696264267,
0.026244334876537323,
0.03340911492705345,
0.06700929999351501,
-0.10142029821872711,
-4.786838303516561e-8,
-0.060847606509923935,
0.06740220636129379,
0.06445977836847305,
0.009933732450008392,
0.1296364665031433,
-0.03049241192638874,
0.00585132883861661,
0.11460955440998077,
0.012449202127754688,
-0.07553880661725998,
0.010262374766170979,
-0.023149216547608376,
-0.029480496421456337,
-0.061404719948768616,
0.04610034078359604,
0.054181504994630814,
-0.02346865087747574,
-0.04084114357829094,
-0.05111485347151756,
-0.05164644494652748,
0.03496670350432396,
0.04708337038755417,
0.03323134034872055,
-0.010621431283652782,
0.07172583043575287,
-0.01565827615559101,
-0.01816984824836254,
0.032404787838459015,
0.08230175077915192,
-0.020445173606276512,
-0.033727895468473434,
0.03643449395895004,
-0.02854880504310131,
0.09716346114873886,
0.0015506688505411148,
0.1024409681558609,
0.04003845900297165,
0.05916789174079895,
0.01602991670370102,
-0.055525459349155426,
-0.026699291542172432,
0.01039921585470438,
-0.055829741060733795,
0.05729994550347328,
0.01282009482383728,
0.05387958884239197,
0.03937509283423424,
-0.020802248269319534,
0.07966353744268417,
0.05770881846547127,
-0.0010529012652114034,
-0.014305119402706623,
-0.02135556749999523,
0.05158381536602974,
0.038516733795404434,
-0.001575642148964107,
-0.016892053186893463,
-0.022344930097460747,
0.006522893439978361,
0.08573400974273682,
0.0009750662720762193,
-0.0009934043046087027,
0.0075113167986273766,
0.0887017697095871
] |
Differentiable Integer Linear Programming | https://openreview.net/forum?id=FPfCUJTsCn | [
"Zijie Geng",
"Jie Wang",
"Xijun Li",
"Fangzhou Zhu",
"Jianye HAO",
"Bin Li",
"Feng Wu"
] | Spotlight | Machine learning (ML) techniques have shown great potential in generating high-quality solutions for integer linear programs (ILPs).
However, existing methods typically rely on a *supervised learning* paradigm, leading to (1) *expensive training cost* due to repeated invocations of traditional solvers to generate training labels, and (2) *plausible yet infeasible solutions* due to the misalignment between the training objective (minimizing prediction loss) and the inference objective (generating high-quality solutions).
To tackle this challenge, we propose **DiffILO** (**Diff**erentiable **I**nteger **L**inear Programming **O**ptimization), an *unsupervised learning paradigm for learning to solve ILPs*.
Specifically, through a novel probabilistic modeling, DiffILO reformulates ILPs---discrete and constrained optimization problems---into continuous, differentiable (almost everywhere), and unconstrained optimization problems.
This reformulation enables DiffILO to simultaneously solve ILPs and train the model via straightforward gradient descent, providing two major advantages.
First, it significantly reduces the training cost, as the training process does not need the aid of traditional solvers at all.
Second, it facilitates the generation of feasible and high-quality solutions, as the model *learns to solve ILPs* in an end-to-end manner, thus aligning the training and inference objectives.
Experiments on commonly used ILP datasets demonstrate that DiffILO not only achieves an average training speedup of $13.2$ times compared to supervised methods, but also outperforms them by generating heuristic solutions with significantly higher feasibility ratios and much better solution qualities. | Integer Linear Programming, Learning to Optimize | We propose a differentiable approach for learning to solve integer linear programming (ILP) problems with unsupervised learning. | 7,722 | 2109.15122 | [
-0.08136141300201416,
0.0066976784728467464,
-0.010941173881292343,
0.009903800673782825,
0.06509669870138168,
-0.018734823912382126,
0.04528556019067764,
-0.025061484426259995,
-0.08953847736120224,
-0.008384335786104202,
-0.07444364577531815,
-0.049754854291677475,
0.036048486828804016,
0.03509433940052986,
-0.043947335332632065,
0.02828127145767212,
0.12740106880664825,
0.11343151330947876,
-0.06897778809070587,
0.01794348657131195,
-0.00812016986310482,
0.03094697743654251,
-0.024409402161836624,
0.026768863201141357,
0.04390951991081238,
-0.03180238604545593,
0.08232638984918594,
-0.04676533117890358,
0.033337969332933426,
-0.03849133849143982,
0.022590676322579384,
-0.04050327464938164,
0.035410407930612564,
-0.0606093555688858,
0.011702439747750759,
0.061724815517663956,
-0.06044818460941315,
0.03555529937148094,
0.009510994888842106,
0.011372453533113003,
0.004623929969966412,
0.07226210087537766,
-0.020520560443401337,
-0.028314445167779922,
0.08836958557367325,
-0.0021741760428994894,
0.005482850130647421,
-0.07549382001161575,
-0.029104173183441162,
0.01845845952630043,
-0.022537516430020332,
-0.011783260852098465,
-0.05141453817486763,
-0.04296833649277687,
-0.11488541215658188,
-0.054921094328165054,
-0.010385561734437943,
0.018030598759651184,
0.026457462459802628,
-0.0060242596082389355,
-0.08383484184741974,
-0.08596181124448776,
-0.025955792516469955,
-0.07381139695644379,
-0.08181164413690567,
-0.09330672770738602,
0.03311280161142349,
0.016209403052926064,
-0.017886029556393623,
0.022519506514072418,
-0.060970183461904526,
0.11996795982122421,
-0.07204750180244446,
0.06545691192150116,
0.05476308614015579,
0.021380502730607986,
0.1272214651107788,
-0.011602126061916351,
0.03446107730269432,
-0.004334293305873871,
-0.01401596050709486,
0.036634210497140884,
0.011011717841029167,
-0.03857613727450371,
0.12586453557014465,
-0.024561064317822456,
-0.02991640567779541,
0.0839286595582962,
0.05592575669288635,
-0.00912352092564106,
-0.005376600194722414,
0.05761297792196274,
-0.029650837182998657,
-0.0010853003477677703,
0.02786092646420002,
-0.0026487866416573524,
0.0034413770772516727,
-0.04357904568314552,
0.007703696843236685,
0.15320058166980743,
-0.056557547301054,
0.09164571017026901,
0.01282606553286314,
-0.03791699931025505,
-0.013505727984011173,
-0.05745510384440422,
0.08060072362422943,
0.009652267210185528,
0.04917430505156517,
-0.13516603410243988,
0.058988869190216064,
-0.022780176252126694,
0.038655806332826614,
-0.0034198679495602846,
0.002377910539507866,
-0.051877424120903015,
-0.0020958289969712496,
-0.03969994932413101,
0.024503273889422417,
0.0032905193511396646,
-0.0550740584731102,
-0.024118369445204735,
0.014018852263689041,
0.11102857440710068,
-0.033514805138111115,
-0.011633042246103287,
-0.05208701640367508,
2.907726473834271e-34,
-0.03435779735445976,
-0.018346738070249557,
0.05532316863536835,
0.0016720622079446912,
0.05467872694134712,
-0.011425424367189407,
0.041898325085639954,
0.0495288148522377,
0.02306080237030983,
0.027713920921087265,
-0.01639716885983944,
-0.046603891998529434,
-0.04068652540445328,
0.06338504701852798,
0.06760091334581375,
-0.025536246597766876,
-0.02180616371333599,
0.05949113517999649,
-0.02150033973157406,
-0.06759720295667648,
0.03937356546521187,
-0.05971599742770195,
-0.011530730873346329,
-0.05327033996582031,
0.06659580767154694,
0.013320613652467728,
0.07654870301485062,
-0.012167779728770256,
0.00968533381819725,
-0.005286162719130516,
-0.061941634863615036,
-0.012298386543989182,
0.013456638902425766,
0.07715819031000137,
0.029800662770867348,
0.009125121869146824,
0.0035460470244288445,
0.005384562071412802,
0.01965896412730217,
-0.010469749569892883,
-0.07646191120147705,
0.008101032115519047,
0.0926475077867508,
0.004329845774918795,
-0.04756003990769386,
0.009087498299777508,
0.03680703416466713,
0.01766487956047058,
-0.0664098709821701,
-0.09013129025697708,
-0.026614904403686523,
0.025668902322649956,
-0.04379493370652199,
-0.06364196538925171,
-0.06763937324285507,
0.03250904008746147,
0.0061609577387571335,
0.09377645701169968,
0.03333039954304695,
0.052645739167928696,
0.02679472602903843,
0.025333654135465622,
0.035577476024627686,
0.006077371537685394,
-0.03397275507450104,
0.08172620832920074,
-0.03207041695713997,
0.05094333738088608,
0.08772952109575272,
0.012642362155020237,
-0.0335066057741642,
-0.020911650732159615,
0.04035302251577377,
-0.031005820259451866,
0.061866577714681625,
-0.006013322155922651,
0.05471888184547424,
0.03152776509523392,
0.010608140379190445,
-0.0383843258023262,
0.005433916114270687,
0.06703851372003555,
-0.02522212266921997,
-0.0536523163318634,
0.021856920793652534,
-0.08786862343549728,
-0.0027877939864993095,
-0.04499083384871483,
-0.03149934113025665,
-0.04027660936117172,
-0.07447897642850876,
0.01274239644408226,
-0.04504064843058586,
0.03403821960091591,
0.0552302710711956,
-1.1805040689732021e-33,
-0.055718064308166504,
0.0775650292634964,
-0.04149425029754639,
-0.012727337889373302,
-0.049970969557762146,
-0.016958575695753098,
-0.01575341634452343,
-0.02775673381984234,
0.05411184951663017,
-0.01263831090182066,
-0.046319253742694855,
-0.028427165001630783,
0.044080495834350586,
0.02775559201836586,
-0.04659964144229889,
-0.0021880122367292643,
-0.08717374503612518,
0.11115920543670654,
-0.0038188761100172997,
0.057815879583358765,
-0.026498807594180107,
0.12082628160715103,
-0.050073977559804916,
-0.008212253451347351,
0.02315600775182247,
0.022511256858706474,
-0.06082845479249954,
0.07838033139705658,
-0.001406898139975965,
0.035809945315122604,
-0.047826070338487625,
-0.009776492603123188,
-0.11763305217027664,
-0.07370312511920929,
-0.06463932245969772,
0.013231324031949043,
-0.0009913509711623192,
-0.03503299131989479,
-0.07419987767934799,
0.1246771365404129,
0.05162997171282768,
-0.1052572950720787,
-0.022789668291807175,
0.005593820009380579,
0.039777033030986786,
0.05184726044535637,
-0.021026618778705597,
-0.0693918913602829,
0.06375721096992493,
0.016731495037674904,
-0.03679826855659485,
0.058153241872787476,
-0.07339370995759964,
0.07951075583696365,
-0.06748267263174057,
-0.027255522087216377,
-0.0356721468269825,
0.0068505872040987015,
0.07163166254758835,
0.015426380559802055,
-0.06440914422273636,
0.0047650644555687904,
0.02366812340915203,
0.0458688847720623,
-0.015264341607689857,
-0.017648132517933846,
0.021598471328616142,
-0.016108883544802666,
-0.03408505395054817,
0.04594654217362404,
-0.015440219081938267,
0.007671938743442297,
0.03058924525976181,
0.012305069714784622,
-0.09507958590984344,
-0.021872935816645622,
0.015758905559778214,
-0.009133923798799515,
-0.050187528133392334,
-0.00953705795109272,
0.09201221168041229,
-0.03145666420459747,
0.03812330588698387,
0.06483583152294159,
-0.005501695908606052,
-0.012708091177046299,
0.10076963901519775,
0.044588908553123474,
0.03741561993956566,
0.016989773139357567,
-0.006799394264817238,
0.0006375576485879719,
0.02585943415760994,
0.0629490315914154,
-0.04988647624850273,
-4.7308422068681466e-8,
-0.027071500197052956,
0.02823157235980034,
-0.019608404487371445,
-0.03976426646113396,
0.03722768649458885,
0.022798262536525726,
-0.13145563006401062,
0.013898838311433792,
-0.0028311258647590876,
-0.005681950598955154,
0.059098560363054276,
-0.017031094059348106,
0.011997071094810963,
-0.07881078869104385,
-0.03592156246304512,
0.08000010997056961,
-0.010256144218146801,
-0.035735588520765305,
0.04632864519953728,
0.026009133085608482,
-0.017531419172883034,
0.02383841574192047,
0.02685646340250969,
-0.006123451981693506,
0.07871901243925095,
-0.1541104018688202,
-0.04112223535776138,
0.047007378190755844,
0.11331375688314438,
0.03871512785553932,
-0.042224787175655365,
0.03703751042485237,
0.07665158808231354,
0.01685469225049019,
-0.01936095766723156,
0.09250685572624207,
-0.008304259739816189,
-0.04725728929042816,
-0.05796092748641968,
-0.1098715141415596,
-0.017228443175554276,
0.053835898637771606,
-0.07363929599523544,
-0.02850406989455223,
-0.0016625436255708337,
0.009080203250050545,
0.037759292870759964,
-0.10490768402814865,
0.03280165418982506,
-0.03746124729514122,
0.012252467684447765,
0.020425481721758842,
0.07225808501243591,
-0.010729842819273472,
0.10875746607780457,
-0.0254799984395504,
-0.06870873272418976,
-0.08440527319908142,
0.022083140909671783,
0.0843983069062233,
-0.0031858226284384727,
-0.026911653578281403,
0.02159229665994644,
-0.018228719010949135
] |
Streaming Algorithms For ℓp Flows and ℓp Regression | https://openreview.net/forum?id=Kpjvm2mB0K | [
"Amit Chakrabarti",
"Jeffrey Jiang",
"David Woodruff",
"Taisuke Yasuda"
] | Spotlight | We initiate the study of one-pass streaming algorithms for underdetermined $\ell_p$ linear regression problems of the form
$$
\min_{\mathbf A\mathbf x = \mathbf b} \lVert\mathbf x\rVert_p \,, \qquad
\text{where } \mathbf A \in \mathbb R^{n \times d} \text{ with } n \ll d \,,
$$
which generalizes basis pursuit ($p = 1$) and least squares solutions to
underdetermined linear systems ($p = 2$). We study the column-arrival
streaming model, in which the columns of $\mathbf A$ are presented one by one in a
stream. When $\mathbf A$ is the incidence matrix of a graph, this corresponds to an
edge insertion graph stream, and the regression problem captures $\ell_p$
flows which includes transshipment ($p = 1$), electrical flows ($p = 2$), and
max flow ($p = \infty$) on undirected graphs as special cases. Our goal is to
design algorithms which use space much less than the entire stream, which has
a length of $d$.
For the task of estimating the cost of the $\ell_p$ regression problem for
$p\in[2,\infty]$, we show a streaming algorithm which constructs a sparse
instance supported on $\tilde O(\varepsilon^{-2}n)$ columns of $\mathbf A$
which approximates the cost up to a $(1\pm\varepsilon)$ factor, which
corresponds to $\tilde O(\varepsilon^{-2}n^2)$ bits of space in general and
an $\tilde O(\varepsilon^{-2}n)$ space semi-streaming algorithm for
constructing $\ell_p$ flow sparsifiers on graphs. This extends to $p\in(1,
2)$ with $\tilde O(\varepsilon^{2}n^{q/2})$ columns, where $q$ is the H\"older
conjugate exponent of $p$. For $p = 2$, we show that $\Omega(n^2)$ bits of
space are required in general even for outputting a constant factor
solution. For $p = 1$, we show that the cost cannot be estimated even to an
$o(\sqrt n)$ factor in $\mathrm{poly}(n)$ space.
On the other hand, if we are interested in outputting a solution $\mathbf
x$, then we show that $(1+\varepsilon)$-approximations require $\Omega(d)$
space for $p > 1$, and in general, $\kappa$-approximations require
$\tilde\Omega(d/\kappa^{2q})$ space for $p > 1$. We complement these lower
bounds with the first sublinear space upper bounds for this problem, showing
that we can output a $\kappa$-approximation using space only
$\mathrm{poly}(n) \cdot \tilde O(d/\kappa^q)$ for $p > 1$, as well as a
$\sqrt n$-approximation using $\mathrm{poly}(n, \log d)$ space for $p = 1$. | Regression, Streaming, Online algorithms, Flows | We give new streaming algorithms for underconstrained regression when we see the columns one at a time, and obtain related results for flows. | 7,681 | null | [
-0.06843648850917816,
-0.08579462766647339,
0.0017889923183247447,
-0.009272018447518349,
0.1416814774274826,
0.01780005916953087,
0.006851287093013525,
-0.04213326796889305,
0.038476668298244476,
0.05303467810153961,
-0.019375029951334,
0.06728747487068176,
0.049206361174583435,
0.006497025024145842,
-0.027438396587967873,
0.03858254477381706,
0.04374048113822937,
-0.03122912161052227,
-0.007045330014079809,
0.02860955521464348,
0.04666212573647499,
-0.06565514951944351,
-0.11251748353242874,
-0.019399479031562805,
0.05821816623210907,
0.034155115485191345,
0.04884826019406319,
-0.007876010611653328,
0.021791892126202583,
-0.04449867457151413,
-0.007105616852641106,
0.03280654549598694,
-0.0411822646856308,
-0.028160512447357178,
-0.023566285148262978,
0.03980090841650963,
0.014599239453673363,
0.05088373273611069,
-0.05654372274875641,
0.07919295877218246,
0.09046020358800888,
-0.0014915401116013527,
0.02765342965722084,
0.04193174093961716,
0.050020817667245865,
-0.051352761685848236,
-0.05712852254509926,
-0.04150412231683731,
-0.03582993522286415,
-0.04758478328585625,
-0.036427147686481476,
0.016319623216986656,
-0.03164078667759895,
0.009210978634655476,
-0.08158549666404724,
-0.1335294097661972,
-0.012408867478370667,
-0.02697622962296009,
0.04321204125881195,
0.06621038168668747,
0.02215810865163803,
-0.05017377436161041,
0.01686847396194935,
0.02124951221048832,
-0.045522138476371765,
0.05204378440976143,
-0.06084031984210014,
0.09215423464775085,
-0.016107819974422455,
0.11156576126813889,
-0.08738571405410767,
0.030394509434700012,
-0.0655718743801117,
-0.020226873457431793,
0.061128657311201096,
-0.09642371535301208,
0.12260987609624863,
0.03282216563820839,
-0.04308664798736572,
0.01951940916478634,
0.006830978207290173,
0.0010843994095921516,
-0.049825116991996765,
-0.1077447384595871,
0.007306947838515043,
-0.024616427719593048,
-0.04920608922839165,
-0.01863025687634945,
0.04695750027894974,
-0.025202227756381035,
-0.04675211384892464,
0.012196316383779049,
-0.033391982316970825,
0.11931531131267548,
0.008753438480198383,
-0.044242583215236664,
0.05646216496825218,
0.042122066020965576,
0.05673559382557869,
0.04463128000497818,
-0.02220803312957287,
0.0002295192243764177,
0.006009261589497328,
0.0647633969783783,
-0.01197020709514618,
-0.07054104655981064,
-0.029653940349817276,
0.053100090473890305,
0.035883828997612,
-0.1261938363313675,
0.014961373060941696,
-0.03617130219936371,
0.021085793152451515,
0.03221018984913826,
0.0216086246073246,
-0.03680836409330368,
0.03034708835184574,
-0.01064636092633009,
0.04792165011167526,
0.00479224743321538,
-0.0010120752267539501,
-0.016618501394987106,
0.06590130925178528,
-0.00488821929320693,
0.04599911719560623,
0.09897612780332565,
-0.02734091505408287,
1.3007004785611926e-33,
-0.03062676079571247,
0.07192579656839371,
0.0017523751594126225,
-0.04352686554193497,
0.013603526167571545,
0.039217665791511536,
0.000041658197005745023,
0.038218043744564056,
0.034399084746837616,
0.03390602767467499,
-0.09412147104740143,
-0.0333249568939209,
0.01733427494764328,
0.08660820871591568,
0.01894993521273136,
-0.030763279646635056,
0.04891914129257202,
0.026231972500681877,
0.0035545185673981905,
-0.07104342430830002,
0.08493649959564209,
-0.075619176030159,
-0.02417881228029728,
-0.06468633562326431,
-0.02942151390016079,
-0.09468478709459305,
-0.004084882792085409,
-0.08898081630468369,
-0.011276057921350002,
0.027269791811704636,
-0.04227230325341225,
0.04722891375422478,
0.010499640367925167,
0.04980493709445,
-0.04549993947148323,
-0.0185768511146307,
0.04377111792564392,
-0.015719495713710785,
0.0206456296145916,
-0.056929465383291245,
-0.060891956090927124,
0.046144548803567886,
0.01570187509059906,
-0.07505796104669571,
-0.04400201141834259,
-0.07844587415456772,
-0.00024281252990476787,
0.013756283558905125,
0.0024243835359811783,
-0.09167671948671341,
0.027495678514242172,
-0.003402487840503454,
-0.018766483291983604,
0.0533689446747303,
0.011882682330906391,
0.013553849421441555,
-0.043946459889411926,
-0.012503417208790779,
0.053942564874887466,
0.06094241514801979,
0.019521331414580345,
0.0016391637036576867,
0.021024174988269806,
-0.05013395845890045,
0.03509100154042244,
0.027234142646193504,
0.014896360225975513,
-0.061738550662994385,
0.10936858505010605,
-0.0401027575135231,
-0.03497776761651039,
-0.005591155495494604,
0.054526202380657196,
-0.03699003532528877,
0.09507113695144653,
-0.015027348883450031,
0.08877504616975784,
-0.01417575217783451,
0.03725360706448555,
0.010385851375758648,
-0.11109393835067749,
-0.009374091401696205,
0.05476291850209236,
-0.020063715055584908,
-0.014067924581468105,
-0.0000938569355639629,
0.026448266580700874,
-0.09342752397060394,
-0.06947969645261765,
-0.07051049172878265,
-0.012240679934620857,
0.03499831631779671,
-0.03657924756407738,
-0.005659590009599924,
0.07133830338716507,
-7.0616528245153094e-34,
0.0378030464053154,
0.04540647566318512,
0.015084979124367237,
-0.03038199618458748,
0.04343068599700928,
-0.08204174041748047,
0.09749254584312439,
-0.027146367356181145,
0.06899139285087585,
-0.04605884104967117,
-0.026207145303487778,
-0.030155887827277184,
-0.0829954743385315,
0.04193751513957977,
0.08932968974113464,
-0.0425175204873085,
0.040896907448768616,
-0.04651441425085068,
0.020803699269890785,
-0.06695672869682312,
-0.026844630017876625,
0.04019707441329956,
-0.011175859719514847,
-0.03707680478692055,
0.021779309958219528,
0.03069242462515831,
0.02006777562201023,
0.05373048409819603,
-0.06517532467842102,
-0.042654797434806824,
-0.0093048932030797,
0.02029993198812008,
-0.08071357756853104,
0.026554273441433907,
-0.042517829686403275,
0.12879063189029694,
-0.05371871218085289,
0.034217577427625656,
-0.03188982605934143,
0.09137854725122452,
-0.03760683909058571,
0.021253198385238647,
-0.019239218905568123,
0.03387549892067909,
0.03351214528083801,
0.01427058968693018,
-0.0028667128644883633,
-0.012021386064589024,
0.024890664964914322,
-0.006101679988205433,
-0.0544380322098732,
0.06087090075016022,
-0.005444245412945747,
0.08497568964958191,
-0.003185911802574992,
-0.034716639667749405,
0.008444303646683693,
0.04124118760228157,
0.0002365016407566145,
-0.06922522932291031,
-0.025148948654532433,
0.014325224794447422,
-0.09098881483078003,
0.044076286256313324,
0.06940267980098724,
-0.009966513141989708,
0.012250901199877262,
-0.058497317135334015,
-0.00448597501963377,
0.01976730115711689,
0.00426475889980793,
0.010546959936618805,
-0.03561538830399513,
0.017081856727600098,
0.027142778038978577,
0.02734113484621048,
-0.021455172449350357,
0.007722151931375265,
-0.05814667418599129,
-0.010040856897830963,
-0.10431762039661407,
0.02872033417224884,
0.024229977279901505,
0.018558645620942116,
-0.03460916131734848,
-0.04666946455836296,
0.08359355479478836,
-0.06952310353517532,
0.006986213382333517,
-0.06271433085203171,
0.0018978131702169776,
-0.030343415215611458,
-0.03510960564017296,
-0.0483858548104763,
0.015357183292508125,
-4.6527066643875514e-8,
-0.12843184173107147,
0.02430589497089386,
-0.04031450301408768,
-0.09231820702552795,
0.09593244642019272,
-0.027033310383558273,
0.038458749651908875,
0.041650593280792236,
-0.00719463499262929,
-0.03474046289920807,
0.014892934821546078,
0.0027897916734218597,
0.008934048004448414,
-0.00845919456332922,
-0.039577651768922806,
0.029808584600687027,
0.01498633436858654,
-0.007812114432454109,
0.012192066758871078,
-0.012923511676490307,
-0.009281952865421772,
0.05850560590624809,
-0.032944608479738235,
0.0016636059153825045,
0.033695515245199203,
-0.089103102684021,
0.0063904146663844585,
-0.03043697215616703,
0.17571865022182465,
0.04867821931838989,
-0.08249031752347946,
0.03066563978791237,
0.15554434061050415,
0.057274121791124344,
0.09819089621305466,
0.07065550982952118,
0.030685219913721085,
-0.028465643525123596,
-0.10304933041334152,
0.022375140339136124,
-0.0006576866726391017,
0.026612913236021996,
-0.008717685006558895,
-0.045542843639850616,
0.10547823458909988,
0.08153466135263443,
0.013804294168949127,
-0.019949031993746758,
0.021278085187077522,
-0.044006120413541794,
0.02757854387164116,
-0.02440560981631279,
0.023073095828294754,
0.07314174622297287,
0.0956575945019722,
-0.03126852214336395,
-0.07793402671813965,
-0.08533930033445358,
0.11010105162858963,
0.0038568235468119383,
-0.05025612190365791,
0.06293876469135284,
-0.012988485395908356,
0.029385672882199287
] |
Generalized Principal-Agent Problem with a Learning Agent | https://openreview.net/forum?id=LqTz13JS2P | [
"Tao Lin",
"Yiling Chen"
] | Spotlight | Generalized principal-agent problems, including Stackelberg games, contract design, and Bayesian persuasion, are a class of economic problems where an agent best responds to a principal's committed strategy.
We study repeated generalized principal-agent problems under the assumption that the principal does not have commitment power and the agent uses algorithms to learn to respond to the principal. We reduce this problem to a one-shot generalized principal-agent problem where the agent approximately best responds. Using this reduction, we show that: (1) if the agent uses contextual no-regret learning algorithms with regret $\mathrm{Reg}(T)$, then the principal can guarantee utility at least $U^* - \Theta\big(\sqrt{\tfrac{\mathrm{Reg}(T)}{T}}\big)$, where $U^*$ is the principal's optimal utility in the classic model with a best-responding agent.
(2) If the agent uses contextual no-swap-regret learning algorithms with swap-regret $\mathrm{SReg}(T)$, then the principal cannot obtain utility more than $U^* + O(\frac{\mathrm{SReg(T)}}{T})$.
But (3) if the agent uses mean-based learning algorithms (which can be no-regret but not no-swap-regret), then the principal can sometimes do significantly better than $U^*$.
These results not only refine previous results in Stackelberg games and contract design, but also lead to new results for Bayesian persuasion with a learning agent and all generalized principal-agent problems where the agent does not have private information. | principal-agent problems, Bayesian persuasion, no-regret learning, no-swap-regret | We study generalized principal-agent problem with a learning agent, which not only sharpens previous works on contract design and Stackelberg games but also leads to new results for Bayesian persuasion. | 7,667 | 2402.09721 | [
-0.06555911898612976,
0.017003880813717842,
-0.006198566872626543,
-0.04662807658314705,
-0.04415598139166832,
0.011004950851202011,
0.04103222116827965,
0.025953883305191994,
0.1033487543463707,
0.06221111863851547,
-0.0019851757679134607,
0.07239298522472382,
0.038552671670913696,
0.05111462250351906,
-0.0015968165826052427,
0.035541947931051254,
0.09240683913230896,
0.07883965224027634,
-0.0517321415245533,
0.047916024923324585,
0.038849275559186935,
-0.01651037484407425,
0.023845475167036057,
-0.02009519934654236,
0.0012519621523097157,
-0.060740187764167786,
0.03865218162536621,
0.03509093448519707,
-0.008797703310847282,
-0.003273169044405222,
0.004294798243790865,
0.03054572269320488,
-0.027432654052972794,
-0.07951684296131134,
0.026357376947999,
0.06790705025196075,
-0.10511410981416702,
0.05116089805960655,
-0.02525203861296177,
0.06437349319458008,
0.028550369665026665,
0.05129067599773407,
-0.024884194135665894,
0.04764816164970398,
0.023301802575588226,
-0.05894548445940018,
0.0354599729180336,
-0.04446106404066086,
0.009434187784790993,
-0.07392139732837677,
0.0003571647685021162,
0.010129773989319801,
-0.02353730797767639,
-0.0023985065054148436,
0.022106720134615898,
-0.0057822419330477715,
-0.04997755587100983,
0.03575022518634796,
-0.00918424129486084,
-0.013352627865970135,
-0.029016194865107536,
-0.12933200597763062,
-0.056294165551662445,
-0.06222090497612953,
0.006149123422801495,
0.03231487423181534,
-0.022050822153687477,
0.011383521370589733,
-0.059974100440740585,
0.07685975730419159,
0.002233203500509262,
0.02358711138367653,
-0.04678067937493324,
0.011285177432000637,
0.024047061800956726,
-0.02278781123459339,
0.006820064503699541,
-0.009410358965396881,
0.06900917738676071,
0.02925456315279007,
-0.007995205000042915,
-0.019332280382514,
-0.03485822677612305,
0.01319805160164833,
0.07972127199172974,
-0.05088980868458748,
0.04887448251247406,
0.030486777424812317,
0.10862310230731964,
-0.06080044433474541,
0.0070764594711363316,
-0.0006232367595657706,
0.014287689700722694,
-0.04678720980882645,
0.0030083099845796824,
0.018786679953336716,
-0.02618429623544216,
-0.10183215141296387,
-0.10283597558736801,
0.08141757547855377,
-0.00372533337213099,
0.006323748733848333,
-0.030442185699939728,
-0.014049911871552467,
0.01049412228167057,
-0.023849597200751305,
-0.06481792777776718,
0.07124001532793045,
0.09314657002687454,
-0.034964412450790405,
-0.0046163625083863735,
0.014157897792756557,
0.031805120408535004,
0.05581722781062126,
-0.0524667352437973,
-0.015655191615223885,
-0.016215724870562553,
-0.01916818507015705,
-0.018983518704771996,
0.05230621248483658,
-0.012501542456448078,
-0.038358062505722046,
0.03268182650208473,
0.04633375257253647,
-0.006198101211339235,
-0.009443341754376888,
-0.034110404551029205,
-8.563819810495995e-34,
0.04926793649792671,
0.03300749137997627,
0.042525313794612885,
-0.13883252441883087,
-0.021555962041020393,
0.03649836406111717,
0.02859184332191944,
-0.003939878195524216,
-0.028095118701457977,
-0.015553043223917484,
0.02241249941289425,
0.04342130571603775,
-0.058477845042943954,
0.0857066810131073,
0.07778780162334442,
0.016628487035632133,
-0.016222890466451645,
-0.018001340329647064,
0.1147654727101326,
-0.07782400399446487,
0.033520765602588654,
0.11024246364831924,
0.05125609040260315,
-0.1215541809797287,
-0.025313125923275948,
-0.03626752272248268,
-0.021199988201260567,
-0.026943642646074295,
0.04246358200907707,
-0.01688011921942234,
-0.061770666390657425,
0.03654151409864426,
-0.05606679618358612,
0.08197560161352158,
-0.010899152606725693,
0.002567962044849992,
-0.05538225546479225,
-0.03141678124666214,
-0.007313760928809643,
0.0033407604787498713,
-0.062220729887485504,
0.038640379905700684,
0.06518004089593887,
0.08204558491706848,
-0.0019252592464908957,
-0.07010723650455475,
0.04724753275513649,
-0.02424025721848011,
-0.0013565249973908067,
-0.014425572007894516,
0.014019478112459183,
0.011070458218455315,
-0.06126304715871811,
-0.06454990059137344,
-0.060639411211013794,
-0.01612834632396698,
0.009955327026546001,
0.09862779825925827,
0.011660137213766575,
0.007867148146033287,
0.04381929710507393,
-0.0983714684844017,
0.037694595754146576,
0.0023267127107828856,
-0.036870986223220825,
0.07038211077451706,
-0.06274408847093582,
0.015784045681357384,
0.020951541140675545,
-0.10041742026805878,
0.01037982665002346,
-0.012560351751744747,
0.009061767719686031,
-0.08412516862154007,
0.06959090381860733,
-0.0021864466834813356,
0.12085866183042526,
0.06663861125707626,
0.031313274055719376,
-0.03481954708695412,
-0.059313785284757614,
-0.07243598997592926,
-0.059251002967357635,
-0.006909704767167568,
-0.023218519985675812,
-0.01552656851708889,
0.037300847470760345,
-0.0877608209848404,
0.0710553228855133,
0.045704394578933716,
-0.10188920050859451,
0.007748747244477272,
-0.04353875666856766,
0.09270601719617844,
0.055413804948329926,
-3.1121534362256615e-34,
-0.0020712735131382942,
-0.08405311405658722,
0.012971128337085247,
0.07178438454866409,
0.04426885396242142,
-0.038219861686229706,
-0.06762044131755829,
-0.10589506477117538,
-0.07479386776685715,
-0.06586777418851852,
-0.06968708336353302,
0.039469845592975616,
0.07547405362129211,
0.041372861713171005,
0.009822357445955276,
-0.04835304990410805,
0.0371757373213768,
-0.010286067612469196,
-0.02571517415344715,
-0.029172765091061592,
0.0684945285320282,
0.09728681296110153,
-0.02980862185359001,
0.06700121611356735,
0.05615527927875519,
0.018331274390220642,
-0.0824233666062355,
0.10193368792533875,
-0.0040800124406814575,
0.01890541985630989,
-0.011055038310587406,
-0.012300757691264153,
-0.07231134921312332,
0.010165542364120483,
0.0019399701850488782,
0.09142421931028366,
-0.008905271999537945,
0.0928167849779129,
-0.10888965427875519,
0.17467083036899567,
0.03484440967440605,
-0.12437963485717773,
0.013547942042350769,
0.010574779473245144,
0.055432289838790894,
-0.040030959993600845,
0.036023061722517014,
-0.08362928032875061,
0.05579637363553047,
0.007820096798241138,
0.028121760115027428,
0.046485692262649536,
-0.014536436647176743,
0.01977480575442314,
-0.02675509639084339,
-0.03869078308343887,
0.08395807445049286,
-0.029011320322752,
0.08690089732408524,
0.016372667625546455,
0.038119666278362274,
-0.014977051876485348,
-0.07676825672388077,
0.05144844576716423,
-0.027504917234182358,
0.05283266305923462,
0.014098300598561764,
-0.05976955592632294,
0.003013158915564418,
0.013077279552817345,
-0.00800546444952488,
-0.005972180515527725,
-0.02056996524333954,
-0.044344350695610046,
0.04828127846121788,
0.10772635042667389,
-0.014138700440526009,
0.03355948254466057,
0.008413049392402172,
-0.034866925328969955,
-0.04127711430191994,
0.021804504096508026,
0.03730041906237602,
-0.052707742899656296,
0.01244774553924799,
0.013200080953538418,
0.047102317214012146,
0.000067079410655424,
0.03821069747209549,
-0.0589228980243206,
-0.000285771326161921,
0.0024698395282030106,
0.044493936002254486,
-0.011582068167626858,
-0.057225439697504044,
-4.794810948283157e-8,
-0.033697210252285004,
-0.02786976471543312,
0.05912015587091446,
0.033478230237960815,
0.10933749377727509,
-0.07890686392784119,
-0.04757285863161087,
-0.03930709511041641,
-0.06010042130947113,
0.024611223489046097,
0.06339506804943085,
0.06938706338405609,
-0.029740622267127037,
-0.08369874209165573,
-0.015820566564798355,
0.06955090165138245,
0.03621584177017212,
-0.0754554495215416,
0.0099201500415802,
-0.03220319002866745,
0.008408263325691223,
-0.05144037306308746,
0.017651023343205452,
-0.07037397474050522,
-0.046288516372442245,
-0.042358942329883575,
0.02457299642264843,
0.07996606826782227,
0.08400211483240128,
0.11580058187246323,
0.006591880228370428,
-0.04975118860602379,
0.042724717408418655,
-0.015137029811739922,
-0.040895551443099976,
-0.016849515959620476,
0.039427898824214935,
0.012359793297946453,
0.0277559757232666,
-0.04484891891479492,
0.025584116578102112,
-0.038692865520715714,
-0.05134466290473938,
-0.03717456012964249,
-0.003981363493949175,
0.06553162634372711,
-0.027005059644579887,
-0.06526973098516464,
0.08379482477903366,
0.05524113401770592,
0.07051670551300049,
0.0008100985432974994,
-0.016168493777513504,
0.014970354735851288,
0.06817632168531418,
-0.08826956897974014,
0.03494137525558472,
-0.05241083726286888,
-0.0328894704580307,
-0.0007637267699465156,
0.019442671909928322,
-0.013840037398040295,
-0.06743352860212326,
-0.016719240695238113
] |
Targeted Attack Improves Protection against Unauthorized Diffusion Customization | https://openreview.net/forum?id=agHddsQhsL | [
"Boyang Zheng",
"Chumeng Liang",
"Xiaoyu Wu"
] | Spotlight | Diffusion models build a new milestone for image generation yet raising public concerns, for they can be fine-tuned on unauthorized images for customization. Protection based on adversarial attacks rises to encounter this unauthorized diffusion customization, by adding protective watermarks to images and poisoning diffusion models. However, current protection, leveraging untargeted attacks, does not appear to be effective enough. In this paper, we propose a simple yet effective improvement for the protection against unauthorized diffusion customization by introducing targeted attacks. We show that by carefully selecting the target, targeted attacks significantly outperform untargeted attacks in poisoning diffusion models and degrading the customization image quality. Extensive experiments validate the superiority of our method on two mainstream customization methods of diffusion models, compared to existing protections. To explain the surprising success of targeted attacks, we delve into the mechanism of attack-based protections and propose a hypothesis based on our observation, which enhances the comprehension of attack-based protections. To the best of our knowledge, we are the first to both reveal the vulnerability of diffusion models to targeted attacks and leverage targeted attacks to enhance protection against unauthorized diffusion customization. | Protection, Unauthorized Diffusion Customization, Adversarial Attack, Diffusion Model, Privacy | We propose a stronger image protection method against unauthorized diffusion customization based on targeted attack. | 7,651 | 2310.04687 | [
-0.03840409219264984,
0.039575692266225815,
0.0335966981947422,
0.024131258949637413,
0.1531752347946167,
-0.047396108508110046,
-0.011297558434307575,
-0.07271571457386017,
0.02942117303609848,
-0.017959298565983772,
0.00506976991891861,
0.012812426313757896,
0.09284289926290512,
-0.007940428331494331,
-0.009550042450428009,
-0.05390916392207146,
-0.002607859205454588,
0.0742807388305664,
-0.11124927550554276,
-0.029417963698506355,
-0.07994875311851501,
-0.07818028330802917,
-0.012748586013913155,
-0.03226246312260628,
-0.02543099783360958,
-0.00881278794258833,
0.013136890716850758,
-0.03370688855648041,
-0.02062126249074936,
-0.07062861323356628,
0.06842957437038422,
-0.04714718088507652,
0.005494030192494392,
0.04229233041405678,
-0.03848293051123619,
-0.02707919292151928,
-0.05151214078068733,
-0.03592315688729286,
-0.005730750970542431,
-0.02894221991300583,
-0.00964680127799511,
-0.03930921107530594,
-0.06654299050569534,
0.0506930947303772,
-0.02560301683843136,
-0.024792848154902458,
0.0795925110578537,
0.03646690025925636,
-0.04387551546096802,
-0.008080950006842613,
0.011000007390975952,
-0.009678011760115623,
-0.09787945449352264,
-0.010601556859910488,
-0.03982562944293022,
-0.11733873188495636,
0.033366136252880096,
0.046466950327157974,
0.03387201204895973,
0.02340664714574814,
-0.05791018530726433,
-0.04952691122889519,
0.0006712435861118138,
0.044895075261592865,
-0.028988288715481758,
0.047475412487983704,
0.043417878448963165,
0.030119504779577255,
-0.012160000391304493,
0.04906619340181351,
0.06808897107839584,
0.03235149756073952,
-0.03387746587395668,
0.02361147478222847,
0.02186880260705948,
0.05288831144571304,
-0.028742866590619087,
0.0013898087199777365,
-0.03234623372554779,
-0.1265159398317337,
0.09779035300016403,
-0.03222741559147835,
0.057963985949754715,
0.00011366197577444836,
0.0735265240073204,
0.022212205454707146,
-0.07727699726819992,
-0.02886684238910675,
0.13105198740959167,
0.016441866755485535,
-0.07061313837766647,
0.04630054533481598,
0.016338326036930084,
0.027119725942611694,
0.06842335313558578,
-0.002517815213650465,
-0.0293202456086874,
-0.04977855458855629,
-0.02526513673365116,
0.07607052475214005,
-0.08363142609596252,
-0.07236450165510178,
-0.023530196398496628,
-0.0479394793510437,
0.14282892644405365,
-0.0015738018555566669,
0.06959905475378036,
0.0004848046228289604,
0.0035862650256603956,
0.04815889522433281,
0.02717139944434166,
0.030536577105522156,
-0.0005208813236095011,
-0.0694466307759285,
0.07764876633882523,
-0.07152491062879562,
0.01741188019514084,
0.004848484415560961,
0.015610984526574612,
-0.008993471972644329,
0.08671551197767258,
0.022863587364554405,
-0.017493680119514465,
-0.03387423977255821,
-0.06199083477258682,
-0.01244949921965599,
-0.06596944481134415,
1.841161740671579e-33,
-0.04511089250445366,
0.00026685104239732027,
-0.021915866062045097,
0.014096536673605442,
0.04809361696243286,
-0.05300730839371681,
-0.0003962223418056965,
-0.025038477033376694,
-0.047287922352552414,
-0.07116815447807312,
0.015323523432016373,
0.009670604020357132,
-0.029934825375676155,
0.15810757875442505,
0.13624973595142365,
0.04736756905913353,
0.012808924540877342,
0.0491093248128891,
0.030104074627161026,
0.015923144295811653,
0.006737197749316692,
-0.044471375644207,
-0.053265783935785294,
-0.03640452027320862,
-0.05108623951673508,
-0.03369755297899246,
0.04731564596295357,
-0.03683421015739441,
0.02745923027396202,
0.02835046872496605,
-0.10735507309436798,
-0.014187497086822987,
0.07699906826019287,
0.06576453894376755,
0.015932956710457802,
0.007372725289314985,
0.0076087419874966145,
-0.026131322607398033,
-0.00047419837210327387,
0.056060437113046646,
0.018288953229784966,
-0.004284263588488102,
-0.05403074994683266,
0.0336977019906044,
-0.0733371302485466,
0.05988721549510956,
-0.048183467239141464,
0.018328549340367317,
-0.0269213505089283,
-0.07772369682788849,
0.017488138750195503,
0.024143097922205925,
0.01657627336680889,
-0.05899649113416672,
-0.08152400702238083,
-0.051524434238672256,
-0.029264086857438087,
-0.09718546271324158,
0.03954373300075531,
-0.026479488238692284,
0.04028618335723877,
0.07329587638378143,
0.006485606078058481,
0.056474871933460236,
-0.04728396236896515,
-0.07364611327648163,
0.048326946794986725,
-0.015605243854224682,
-0.0442352220416069,
-0.0019742646254599094,
-0.07861088216304779,
0.01625477336347103,
0.0018290102016180754,
-0.027305440977215767,
-0.049110978841781616,
-0.08282556384801865,
0.0683523640036583,
0.10491631925106049,
0.057073239237070084,
-0.060575176030397415,
-0.00815382320433855,
-0.06343559175729752,
0.06615807116031647,
-0.03503637760877609,
-0.07671485096216202,
-0.016049601137638092,
-0.021070541813969612,
-0.06345357745885849,
-0.1062939241528511,
0.014238333329558372,
0.03235047683119774,
-0.0018570987740531564,
-0.03948221355676651,
0.03172694519162178,
-0.032410845160484314,
-3.498636325979852e-33,
-0.08362415432929993,
0.058359213173389435,
-0.004662369843572378,
0.05467645451426506,
-0.06072143092751503,
0.005403106100857258,
-0.04401203617453575,
0.07492082566022873,
0.05102305859327316,
-0.06810026615858078,
-0.016741327941417694,
0.04405238479375839,
-0.043158721178770065,
-0.05546065419912338,
-0.02799873985350132,
-0.08143418282270432,
0.02830871008336544,
0.05263489857316017,
-0.06315977871417999,
-0.028454117476940155,
0.013102009892463684,
0.07545138150453568,
-0.0704386755824089,
0.045346807688474655,
-0.009707370772957802,
0.01963784731924534,
-0.0699782744050026,
0.05037017911672592,
0.08263880014419556,
-0.024385491386055946,
-0.05000462383031845,
0.035832878202199936,
-0.0059714061208069324,
-0.03664764389395714,
-0.007747578900307417,
0.0782342180609703,
0.0024972609244287014,
0.04368320479989052,
-0.01116859819740057,
0.035061463713645935,
0.013829140923917294,
-0.024988345801830292,
-0.05654526129364967,
0.04052278771996498,
-0.021548939868807793,
-0.07125668227672577,
0.021007176488637924,
-0.010752897709608078,
0.03505174070596695,
-0.001043527270667255,
0.06000075116753578,
-0.01340689416974783,
0.025578293949365616,
0.04163646697998047,
-0.06884138286113739,
-0.018745161592960358,
0.002821490168571472,
0.008505568839609623,
0.11976543068885803,
0.05755553022027016,
0.07685818523168564,
-0.041012365370988846,
-0.017613260075449944,
-0.0886104479432106,
-0.06867726892232895,
-0.023797892034053802,
-0.03658026456832886,
0.054404016584157944,
-0.010369952768087387,
0.050738632678985596,
-0.0071749198250472546,
-0.028143396601080894,
-0.010916858911514282,
0.03113432042300701,
-0.015573185868561268,
-0.13067787885665894,
0.016345219686627388,
0.0025665259454399347,
-0.03311220556497574,
0.05773365497589111,
0.04289022460579872,
0.0044030845165252686,
0.001978489337489009,
-0.016423793509602547,
0.09378883242607117,
-0.016264868900179863,
-0.008209173567593098,
-0.008396543562412262,
-0.014349459670484066,
0.03678205981850624,
-0.019801197573542595,
-0.011845977045595646,
-0.017843952402472496,
0.09284305572509766,
0.010987241752445698,
-3.944664328514591e-8,
-0.016880594193935394,
0.02500550076365471,
0.06833523511886597,
0.04339289665222168,
0.03542196378111839,
0.057672079652547836,
-0.018060384318232536,
0.03260425105690956,
0.02224496193230152,
-0.10777167230844498,
0.029026158154010773,
0.0026194220408797264,
-0.035502370446920395,
-0.08170846849679947,
-0.08060923963785172,
0.09715916216373444,
-0.0613100491464138,
0.020856861025094986,
-0.019073987379670143,
-0.01013994961977005,
-0.10867892950773239,
0.03495539724826813,
0.011797022074460983,
-0.07334698736667633,
0.05002674087882042,
0.019583621993660927,
-0.019841501489281654,
0.00016620181850157678,
0.043467484414577484,
0.10599418729543686,
-0.04779321327805519,
-0.024869874119758606,
0.01063737366348505,
0.054389312863349915,
-0.060977622866630554,
0.10882050544023514,
-0.0527205653488636,
-0.024586204439401627,
0.010905222967267036,
-0.06007075309753418,
-0.010992384515702724,
-0.02326713688671589,
0.05268071964383125,
0.003176393685862422,
-0.02067570760846138,
0.004895837511867285,
0.03164536878466606,
-0.04164499044418335,
0.02869446575641632,
0.007673373445868492,
0.059209492057561874,
-0.03325992077589035,
0.012136769481003284,
0.1231372132897377,
-0.00956482719630003,
-0.06104028597474098,
0.04370470345020294,
0.04311254248023033,
0.07992953807115555,
0.11649280041456223,
0.07920751720666885,
0.0011914960341528058,
0.06408216059207916,
0.04706940799951553
] |
High-dimensional Analysis of Knowledge Distillation: Weak-to-Strong Generalization and Scaling Laws | https://openreview.net/forum?id=1xzqz73hvL | [
"Muhammed Emrullah Ildiz",
"Halil Alperen Gozeten",
"Ege Onur Taga",
"Marco Mondelli",
"Samet Oymak"
] | Spotlight | A growing number of machine learning scenarios rely on knowledge distillation where one uses the output of a surrogate model as labels to supervise the training of a target model. In this work, we provide a sharp characterization of this process for ridgeless, high-dimensional regression, under two settings: *(i)* model shift, where the surrogate model is arbitrary, and *(ii)* distribution shift, where the surrogate model is the solution of empirical risk minimization with out-of-distribution data. In both cases, we characterize the precise risk of the target model through non-asymptotic bounds in terms of sample size and data distribution under mild conditions. As a consequence, we identify the form of the optimal surrogate model, which reveals the benefits and limitations of discarding weak features in a data-dependent fashion. In the context of weak-to-strong (W2S) generalization, this has the interpretation that *(i)* W2S training, with the surrogate as the weak model, can provably outperform training with strong labels under the same data budget, but *(ii)* it is unable to improve the data scaling law. We validate our results on numerical experiments both on ridgeless regression and on neural network architectures. | empirical risk minimization, high-dimensional statistics, scaling laws, weak to strong generalization, knowledge distillation | This paper provides a sharp characterization of a two-stage learning process, where the first-stage (surrogate) model's output supervises the second stage, thus revealing the form of optimal surrogates and when it is beneficial to discard features. | 7,640 | 2410.18837 | [
-0.026790183037519455,
-0.09694401919841766,
0.02685234136879444,
0.05593647435307503,
0.11153438687324524,
-0.0037189165595918894,
-0.014287471771240234,
0.01717202365398407,
0.0233218502253294,
-0.013313276693224907,
0.01637374423444271,
0.05831398442387581,
0.045417819172143936,
-0.002277552615851164,
-0.03777332976460457,
0.009106766432523727,
0.08189410716295242,
0.03338146582245827,
-0.1291782110929489,
0.04697011038661003,
0.049584608525037766,
0.030624711886048317,
-0.0250899288803339,
0.03175355866551399,
0.026427747681736946,
-0.03263593465089798,
-0.01978698931634426,
-0.021005434915423393,
0.009527713060379028,
-0.02929377369582653,
0.04114285483956337,
-0.01401753444224596,
-0.025455480441451073,
-0.009241236373782158,
0.010912880301475525,
0.04111583158373833,
-0.019544342532753944,
0.06351131200790405,
0.06614217162132263,
0.02809767611324787,
0.06365113705396652,
0.03897523880004883,
-0.015698006376624107,
0.10991814732551575,
0.061526522040367126,
-0.003438318148255348,
0.059938423335552216,
-0.03641168773174286,
-0.03247359022498131,
0.026816166937351227,
-0.059100404381752014,
-0.02627035230398178,
-0.05777842551469803,
0.06958857923746109,
-0.018380768597126007,
-0.08396546542644501,
0.014419196173548698,
-0.0004231555503793061,
0.01812523789703846,
0.057781584560871124,
-0.0015707623679190874,
-0.09248717874288559,
-0.02819589152932167,
-0.020567787811160088,
0.003970209509134293,
-0.06482139974832535,
0.00007795899728080258,
0.06358344852924347,
-0.05496467649936676,
0.06720156222581863,
-0.031052231788635254,
0.05881750211119652,
-0.0849452018737793,
0.017045849934220314,
0.05966043472290039,
0.003587070619687438,
0.03549503907561302,
0.04031995311379433,
0.016369841992855072,
-0.0854136049747467,
0.03475483879446983,
0.03586924448609352,
0.06129094958305359,
-0.09733211249113083,
0.10483745485544205,
0.017096897587180138,
-0.023211516439914703,
-0.000699617899954319,
0.06333859264850616,
-0.02310303784906864,
-0.03755509480834007,
-0.04003680869936943,
-0.01124544721096754,
0.006852895487099886,
0.03353572636842728,
-0.018824579194188118,
0.019586676731705666,
-0.09249890595674515,
-0.03628372773528099,
0.13189861178398132,
-0.055418118834495544,
0.09135696291923523,
0.03420630842447281,
-0.00042087314068339765,
0.054498765617609024,
-0.09571206569671631,
0.02777674049139023,
0.0067542982287704945,
0.031511131674051285,
-0.052881769835948944,
0.04339584708213806,
0.10828115046024323,
0.01734502799808979,
0.026215169578790665,
-0.008988830260932446,
-0.04362146928906441,
0.01282990537583828,
-0.023149743676185608,
-0.08315328508615494,
0.016631072387099266,
-0.04839412122964859,
-0.01349964551627636,
0.02006710134446621,
0.025355054065585136,
-0.09318933635950089,
0.045102883130311966,
-0.038248974829912186,
2.8972401917877228e-33,
0.055040668696165085,
0.05574168637394905,
-0.014892259612679482,
0.01814248040318489,
0.009841907769441605,
-0.06281275302171707,
0.021987924352288246,
-0.01518482156097889,
0.013117312453687191,
0.07775913923978806,
-0.015701899304986,
0.0025214648339897394,
-0.0465550534427166,
0.033448539674282074,
0.061271097511053085,
0.04556433483958244,
0.017337772995233536,
0.10389593243598938,
0.004512756131589413,
-0.06358311325311661,
0.025869149714708328,
-0.05093451961874962,
-0.05060556158423424,
-0.06710901111364365,
-0.04313100874423981,
-0.05714680254459381,
-0.023423073813319206,
0.012811372987926006,
-0.049709904938936234,
0.013075991533696651,
-0.04336579889059067,
0.007671902887523174,
0.09277576208114624,
0.0019847583025693893,
0.01331100519746542,
-0.04247189685702324,
-0.07442663609981537,
-0.0690210685133934,
0.023755745962262154,
-0.01786622777581215,
-0.04499662667512894,
0.038128290325403214,
0.06168149411678314,
0.021323498338460922,
-0.010556088760495186,
-0.03963689133524895,
0.067585788667202,
-0.05470386520028114,
-0.08368488401174545,
-0.1334940791130066,
-0.010450301691889763,
0.027961116284132004,
-0.022064460441470146,
-0.07422688603401184,
-0.06589394807815552,
0.06620412319898605,
0.044374529272317886,
-0.018894163891673088,
0.02421390824019909,
0.01876104064285755,
-0.09529516845941544,
-0.0007173270569182932,
-0.042249102145433426,
0.08100904524326324,
0.005039793439209461,
0.0018676817417144775,
-0.031206203624606133,
-0.017164725810289383,
0.0598529614508152,
0.01143008191138506,
0.0029672652017325163,
-0.02365763671696186,
-0.03991193324327469,
-0.056322045624256134,
0.08651436120271683,
-0.0350290983915329,
0.06556341797113419,
-0.0754055455327034,
0.04138076677918434,
0.019630052149295807,
0.03039279393851757,
-0.029313262552022934,
-0.029651369899511337,
-0.06362147629261017,
0.0022097667679190636,
0.03518672659993172,
-0.02013751119375229,
-0.10725025087594986,
-0.04476282745599747,
-0.029287267476320267,
-0.09133348613977432,
0.017066849395632744,
-0.09074075520038605,
-0.07996270060539246,
0.06429779529571533,
-2.4306131992204228e-33,
-0.0989677906036377,
0.05231815204024315,
0.0013735416578128934,
0.04037627950310707,
-0.020259959623217583,
-0.02391035109758377,
-0.04119659215211868,
0.07016979157924652,
-0.07166008651256561,
-0.05877004191279411,
0.012775424867868423,
-0.0663859099149704,
0.019305048510432243,
-0.008931715041399002,
-0.03783693164587021,
0.02851164899766445,
-0.014142889529466629,
0.051852595061063766,
0.016915863379836082,
0.01140167098492384,
0.021746806800365448,
0.04927482828497887,
-0.03243125230073929,
0.10179037600755692,
-0.04599678888916969,
-0.0051368288695812225,
-0.0616484060883522,
0.1479988545179367,
0.03251149132847786,
-0.006434507668018341,
-0.06194351613521576,
0.002170227002352476,
0.024023378267884254,
-0.0047461893409490585,
-0.10132399946451187,
0.05091745778918266,
-0.06063074246048927,
-0.03477758914232254,
-0.09417922794818878,
0.11040612310171127,
0.015062426216900349,
0.008670247159898281,
-0.0958268791437149,
-0.04951389878988266,
-0.0010286032920703292,
-0.039792947471141815,
0.02721087634563446,
-0.03148797154426575,
0.06391748040914536,
0.03127795830368996,
0.03491051495075226,
0.06613888591527939,
-0.05085873231291771,
0.13039788603782654,
-0.018729958683252335,
-0.044816236943006516,
-0.013495579361915588,
-0.03912947699427605,
-0.016223382204771042,
0.03048356994986534,
-0.026931408792734146,
0.012926600873470306,
-0.034834589809179306,
0.02614927478134632,
0.04165484756231308,
-0.046946048736572266,
-0.03006299026310444,
0.13160790503025055,
0.015351396054029465,
0.025060050189495087,
-0.028888247907161713,
-0.041892148554325104,
0.06272095441818237,
-0.049994856119155884,
-0.06093365326523781,
-0.03409069404006004,
-0.036892376840114594,
-0.07709113508462906,
-0.03474495932459831,
0.01920882984995842,
0.017494341358542442,
0.01920507103204727,
0.052880898118019104,
0.0493270568549633,
0.11093676090240479,
0.04860125854611397,
0.07487012445926666,
0.01585681363940239,
-0.013461335562169552,
-0.012498383410274982,
-0.04158034175634384,
-0.050294194370508194,
-0.006646649446338415,
0.042814645916223526,
-0.013193854130804539,
-4.847761303494735e-8,
-0.059531860053539276,
0.10164281725883484,
0.07773225009441376,
-0.03212679177522659,
0.08463642746210098,
0.006073689088225365,
-0.052663080394268036,
0.06475451588630676,
-0.07507745921611786,
0.057489920407533646,
0.013624838553369045,
0.006327373906970024,
-0.04399023577570915,
-0.0528532974421978,
-0.006701948121190071,
0.11493650823831558,
-0.015284904278814793,
0.059855762869119644,
-0.005090402439236641,
-0.005740555003285408,
0.021725168451666832,
0.023030303418636322,
0.07991696894168854,
-0.04377586022019386,
0.09565867483615875,
-0.10498369485139847,
-0.041160948574543,
0.08229825645685196,
0.06786636263132095,
0.013806043192744255,
-0.019309740513563156,
-0.019418371841311455,
0.020989414304494858,
0.0173147264868021,
-0.01825256273150444,
0.0787571519613266,
-0.011503606103360653,
-0.04953743517398834,
-0.0016621145186945796,
0.058094847947359085,
0.008660104125738144,
0.06679514050483704,
-0.0420365110039711,
0.06218951195478439,
-0.030607974156737328,
0.06692172586917877,
-0.01714547537267208,
-0.023408085107803345,
0.02575814351439476,
-0.02982129342854023,
0.08509152382612228,
-0.034416746348142624,
0.08437307178974152,
0.06388339400291443,
0.05164938420057297,
0.03274916112422943,
-0.06083175167441368,
0.0485088974237442,
-0.015867268666625023,
0.006009009201079607,
0.038466159254312515,
-0.034550271928310394,
-0.03191554546356201,
0.01009285356849432
] |
BlendRL: A Framework for Merging Symbolic and Neural Policy Learning | https://openreview.net/forum?id=60i0ksMAhd | [
"Hikaru Shindo",
"Quentin Delfosse",
"Devendra Singh Dhami",
"Kristian Kersting"
] | Spotlight | Humans can leverage both symbolic reasoning and intuitive responses. In contrast, reinforcement learning policies are typically encoded in either opaque systems like neural networks or symbolic systems that rely on predefined symbols and rules. This disjointed approach severely limits the agents’ capabilities, as they often lack either the flexible low-level reaction characteristic of neural agents or the interpretable reasoning of symbolic agents.
To overcome this challenge, we introduce *BlendRL*, a neuro-symbolic RL framework that harmoniously integrates both paradigms.
We empirically demonstrate that BlendRL agents outperform both neural and symbolic baselines in standard Atari environments, and showcase their robustness to environmental changes. Additionally, we analyze the interaction between neural and symbolic policies, illustrating how their hybrid use helps agents overcome each other's limitations. | Neuro-Symbolic AI, Differentiable Reasoning, Reinforcement Learning, Interpretable AI, First-order logic | We propose a framework that jointly learns symbolic and neural policies for reinforcement learning. | 7,559 | 2410.11689 | [
0.009082437492907047,
-0.08881255984306335,
0.09726477414369583,
-0.039337314665317535,
-0.0009050287189893425,
0.04300875589251518,
0.0027244454249739647,
-0.06195869669318199,
-0.01521108578890562,
-0.0025274890940636396,
-0.020573653280735016,
-0.04664139449596405,
0.02076675184071064,
0.008923550136387348,
0.05845833569765091,
0.09867927432060242,
-0.002794817788526416,
0.039585962891578674,
-0.09498104453086853,
-0.007916971109807491,
0.0025442459154874086,
-0.05255252867937088,
0.037916820496320724,
-0.010358835570514202,
-0.018312647938728333,
0.09248713403940201,
0.0019473519641906023,
-0.008361224085092545,
0.05502878874540329,
-0.007039031479507685,
0.10840184986591339,
-0.037323959171772,
-0.04598456993699074,
-0.05869627371430397,
-0.04932033270597458,
0.09928935766220093,
-0.06359019875526428,
0.00004270498175173998,
-0.0008844753610901535,
-0.026077529415488243,
-0.06769290566444397,
0.017494216561317444,
0.005276655778288841,
0.019549164921045303,
0.00949707068502903,
0.001612829859368503,
0.0422402061522007,
-0.0841040089726448,
0.05055606737732887,
-0.059656932950019836,
-0.05513857305049896,
-0.040500275790691376,
0.02766530215740204,
-0.02913353033363819,
0.05238977447152138,
0.11136502772569656,
0.019602196291089058,
0.08114489167928696,
-0.02094864286482334,
-0.05063517019152641,
-0.045342862606048584,
-0.06357492506504059,
0.0004154592170380056,
0.0002960990241263062,
-0.05045989528298378,
0.06278447806835175,
-0.02141304686665535,
0.06956727802753448,
0.04935140907764435,
-0.0023797841276973486,
-0.07414032518863678,
0.02765970304608345,
0.014970819465816021,
-0.05234858766198158,
-0.015090558677911758,
0.024805264547467232,
-0.0018821401754394174,
-0.03710576891899109,
0.05997881665825844,
-0.06795533746480942,
-0.043118853121995926,
-0.07311836630105972,
-0.07708919793367386,
0.029078997671604156,
0.07227805256843567,
0.051834478974342346,
-0.012035488151013851,
0.046602070331573486,
0.06276657432317734,
0.036917973309755325,
-0.011575430631637573,
-0.018200254067778587,
0.10693274438381195,
-0.012857652269303799,
0.06626871973276138,
0.0517059862613678,
-0.06321366131305695,
-0.07219532877206802,
-0.06359170377254486,
0.09798791259527206,
0.012299258261919022,
0.08417874574661255,
-0.02102452516555786,
0.016979962587356567,
-0.006610094103962183,
-0.07857572287321091,
0.07319363951683044,
-0.01621408388018608,
0.03669094294309616,
-0.03219984471797943,
0.02600773610174656,
0.015360853634774685,
0.08113239705562592,
0.08026469498872757,
-0.022389616817235947,
0.007541677914559841,
0.02984127588570118,
-0.023610860109329224,
0.03130245581269264,
0.02588569186627865,
0.006788704544305801,
0.00034903461346402764,
-0.001622445648536086,
0.07079437375068665,
0.06740119308233261,
-0.011579896323382854,
-0.03340720757842064,
-4.262419198720906e-34,
0.07024943083524704,
0.04424579441547394,
0.04954304173588753,
0.021795039996504784,
0.07909427583217621,
-0.022279895842075348,
-0.015224965289235115,
-0.034346628934144974,
0.025830792263150215,
0.011054771952331066,
-0.03345436975359917,
0.021855616942048073,
-0.10563775151968002,
0.11378323286771774,
0.010235775262117386,
-0.017807764932513237,
-0.06798054277896881,
0.006798866204917431,
-0.0704505667090416,
-0.08101774752140045,
-0.0422123558819294,
-0.008722580038011074,
-0.03596401587128639,
0.0006523434421978891,
-0.032034751027822495,
0.042150624096393585,
-0.015226814895868301,
0.07080595195293427,
0.06720054894685745,
-0.01533435843884945,
0.012449467554688454,
0.0409020371735096,
-0.043166592717170715,
0.12513835728168488,
0.032507434487342834,
-0.004887010902166367,
0.005372676067054272,
-0.033830154687166214,
0.03363804519176483,
0.06292956322431564,
0.011945105157792568,
-0.005605795886367559,
-0.009562675841152668,
0.05266163498163223,
0.00475698197260499,
-0.05727668106555939,
-0.009299580939114094,
0.027228765189647675,
-0.03509596735239029,
-0.05145459249615669,
-0.013299216516315937,
0.10186123847961426,
0.02633984573185444,
-0.09719128161668777,
-0.014864517375826836,
-0.038143325597047806,
-0.0017666603671386838,
0.028875023126602173,
-0.06173417344689369,
0.045078061521053314,
-0.012686402536928654,
0.02110462263226509,
0.036756813526153564,
0.04188963770866394,
0.0011742089409381151,
0.11952969431877136,
-0.04099380224943161,
0.06131431832909584,
0.09328316897153854,
0.03734075650572777,
-0.03926926106214523,
0.07023799419403076,
-0.046197324991226196,
-0.00843552965670824,
0.03363490849733353,
-0.085549496114254,
0.02008603885769844,
-0.11436881870031357,
-0.02867264673113823,
-0.009903950616717339,
-0.016962511464953423,
-0.017819102853536606,
-0.10160595178604126,
0.030171172693371773,
0.011250603012740612,
-0.06733158975839615,
0.03288072347640991,
-0.05020836368203163,
-0.027647290378808975,
-0.04149718955159187,
-0.12338416278362274,
-0.11567160487174988,
-0.0018027964979410172,
-0.0069213807582855225,
-0.015755660831928253,
-1.9792525721924426e-33,
-0.030660532414913177,
-0.05741070955991745,
-0.028420226648449898,
0.04898332804441452,
-0.07023666799068451,
-0.04832950979471207,
-0.022141387686133385,
-0.03896501660346985,
0.013565709814429283,
-0.02312135323882103,
-0.03145017847418785,
0.06815381348133087,
0.1090187281370163,
-0.0487540140748024,
-0.05345402657985687,
-0.07082702964544296,
-0.018566377460956573,
-0.027585575357079506,
0.04924102872610092,
-0.04923168569803238,
0.0030581604223698378,
0.04940473288297653,
-0.057488974183797836,
0.031906452029943466,
-0.01161279622465372,
0.03309348598122597,
-0.000402019388275221,
0.07453688979148865,
0.018635837361216545,
0.04202452674508095,
0.0012765794526785612,
-0.04785671457648277,
0.013625653460621834,
-0.0645294114947319,
-0.010608960874378681,
0.06689885258674622,
0.01574464701116085,
-0.030256181955337524,
-0.1354227364063263,
0.003642147174105048,
-0.0027224174700677395,
-0.09473846107721329,
-0.025228915736079216,
0.031209813430905342,
-0.0005758197512477636,
0.01609737239778042,
-0.1104699894785881,
0.0026652237866073847,
-0.0476655475795269,
-0.005764198489487171,
0.03632141649723053,
-0.028948059305548668,
-0.024119576439261436,
-0.09029852598905563,
-0.08335434645414352,
-0.004908264614641666,
0.01253510732203722,
0.03118593990802765,
0.06887780874967575,
0.036904092878103256,
-0.07825477421283722,
-0.03204577416181564,
-0.061700381338596344,
0.0035140637774020433,
-0.07255160808563232,
0.0027739887591451406,
-0.006690031383186579,
0.03522922843694687,
0.014692396856844425,
-0.04825486242771149,
0.08015823364257812,
0.010183756239712238,
0.0031838531140238047,
0.05337543785572052,
0.08241640776395798,
-0.020177101716399193,
-0.06340906023979187,
0.007221716921776533,
-0.0010117435595020652,
-0.012761499732732773,
-0.0203048437833786,
0.012462923303246498,
0.05917489901185036,
0.03744189068675041,
0.00590835977345705,
0.05996726080775261,
-0.0799354612827301,
0.11724280565977097,
0.021361883729696274,
-0.012506579048931599,
0.01669337972998619,
0.02358047105371952,
-0.047661274671554565,
0.0518047958612442,
-0.13043901324272156,
-3.85192642227139e-8,
0.0008780976058915257,
0.09692738205194473,
0.1145569458603859,
0.04933018609881401,
0.06412970274686813,
0.01924450322985649,
-0.001265486585907638,
-0.03319329768419266,
-0.06410697102546692,
0.011831958778202534,
0.03357039764523506,
0.04913764446973801,
-0.02653871662914753,
-0.03316376358270645,
-0.009526641108095646,
0.09264957159757614,
-0.02303045615553856,
0.006019902415573597,
-0.015407254919409752,
-0.028521297499537468,
0.05039277300238609,
0.0053211092017591,
-0.018517376855015755,
-0.03702876716852188,
0.04969986900687218,
-0.08095195889472961,
-0.0711965337395668,
0.13930638134479523,
0.04537266120314598,
0.01627505198121071,
0.061464764177799225,
0.015016961842775345,
0.08720988035202026,
-0.05965064465999603,
-0.05815448239445686,
0.05956941843032837,
0.0684218779206276,
-0.03358167037367821,
0.004887166898697615,
0.015311217866837978,
-0.029979096725583076,
-0.002672227332368493,
0.02974400855600834,
-0.04591362550854683,
-0.07252994924783707,
-0.0231328047811985,
-0.009587354958057404,
-0.0944296270608902,
0.016218487173318863,
-0.0236226674169302,
0.05052642524242401,
0.013589839451014996,
-0.023147059604525566,
0.0981714054942131,
0.044395945966243744,
-0.10205545276403427,
-0.07267587631940842,
-0.020322676748037338,
-0.042603522539138794,
0.03477104380726814,
-0.0052337646484375,
0.060354165732860565,
-0.045245423913002014,
-0.04714236408472061
] |
Uncovering Overfitting in Large Language Model Editing | https://openreview.net/forum?id=t8qcGXaepr | [
"Mengqi Zhang",
"Xiaotian Ye",
"Qiang Liu",
"Shu Wu",
"Pengjie Ren",
"Zhumin Chen"
] | Spotlight | Knowledge editing has been proposed as an effective method for updating and correcting the internal knowledge of Large Language Models (LLMs). However, existing editing methods often struggle with complex tasks, such as multi-hop reasoning. In this paper, we identify and investigate the phenomenon of Editing Overfit, where edited models assign disproportionately high probabilities to the edit target, hindering the generalization of new knowledge in complex scenarios. We attribute this issue to the current editing paradigm, which places excessive emphasis on the direct correspondence between the input prompt and the edit target for each edit sample. To further explore this issue, we introduce a new benchmark, EVOKE (EValuation of Editing Overfit in Knowledge Editing), along with fine-grained evaluation metrics. Through comprehensive experiments and analysis, we demonstrate that Editing Overfit is prevalent in current editing methods and that common overfitting mitigation strategies are ineffective in knowledge editing. To overcome this, inspired by LLMs’ knowledge recall mechanisms, we propose a new plug-and-play strategy called Learn the Inference (LTI), which introduce a Multi-stage Inference Constraint module to guide the edited models in recalling new knowledge similarly to how unedited LLMs leverage knowledge through in-context learning. Extensive experimental results across a wide range of tasks validate the effectiveness of LTI in mitigating Editing Overfit. | Large language models, Knowledge editing, Editing overfit | We explore the overfitting problem in knowledge editing and develop a new benchmark along with a novel overfitting mitigation strategy. | 7,510 | 2410.07819 | [
-0.02935866080224514,
-0.04122529178857803,
0.02865581400692463,
0.08265192806720734,
0.07035013288259506,
0.023238258436322212,
-0.04955536872148514,
0.010543598793447018,
0.0269918292760849,
0.07209303975105286,
-0.05147084221243858,
-0.01525893434882164,
0.03692341595888138,
-0.02310798689723015,
0.005715053062886,
0.019986771047115326,
0.02179260365664959,
0.030579818412661552,
-0.11798881739377975,
-0.07429137825965881,
0.033886414021253586,
0.021501250565052032,
0.09131664782762527,
0.02514435164630413,
-0.015414305962622166,
-0.06415528059005737,
-0.018588067963719368,
-0.006170975510030985,
-0.01779664307832718,
-0.024377109482884407,
-0.030426619574427605,
0.0858871191740036,
-0.00043663126416504383,
0.015093047171831131,
-0.04788976535201073,
0.02426491305232048,
-0.03987731412053108,
0.06628099828958511,
0.017334235832095146,
-0.07394523918628693,
-0.042227208614349365,
-0.012958652339875698,
-0.03330779820680618,
0.00685069989413023,
0.04336928203701973,
-0.002130800159648061,
-0.03623334318399429,
-0.031344782561063766,
-0.017804451286792755,
0.1229838877916336,
-0.07250117510557175,
-0.10278398543596268,
-0.04115419462323189,
0.03663177043199539,
-0.01796692982316017,
0.0001324759650742635,
0.00911514088511467,
0.025305213406682014,
0.01631844975054264,
-0.12025002390146255,
-0.04727371782064438,
-0.09402147680521011,
-0.04874774441123009,
0.012030964717268944,
-0.06331571936607361,
-0.06623878329992294,
-0.012935641221702099,
0.05597592890262604,
-0.07540112733840942,
-0.0031131382565945387,
0.023097727447748184,
0.03591885045170784,
-0.05226825550198555,
0.028869329020380974,
-0.05556655675172806,
0.0498557947576046,
-0.04552029073238373,
-0.01819206029176712,
-0.04184657335281372,
-0.03579368442296982,
0.03999296948313713,
-0.003308251267299056,
-0.0156637541949749,
-0.018876658752560616,
0.11279767006635666,
0.00435435539111495,
0.07287213951349258,
0.005685917101800442,
0.08336558938026428,
0.03993630036711693,
-0.03545103222131729,
-0.04582985118031502,
0.06280065327882767,
-0.00007212642958620563,
0.045011021196842194,
0.06592927128076553,
-0.009439337067306042,
-0.0382465198636055,
-0.03770696371793747,
0.06996720284223557,
0.011686905287206173,
0.11000187695026398,
-0.007131129037588835,
-0.0769149586558342,
0.010034476406872272,
-0.025835691019892693,
0.05541730299592018,
0.04301767796278,
0.030118945986032486,
-0.11509104818105698,
0.021014323458075523,
0.031101571395993233,
0.06763552129268646,
-0.02911473996937275,
0.013297876343131065,
0.010522770695388317,
0.044030964374542236,
0.03771718963980675,
-0.051655787974596024,
0.08081544935703278,
-0.011427313089370728,
0.029041415080428123,
0.04381011798977852,
0.02902643196284771,
-0.0560792051255703,
-0.028893381357192993,
-0.0516589879989624,
7.713390744060601e-35,
0.13390971720218658,
0.08202727884054184,
0.005264332052320242,
0.04515773430466652,
0.07432576268911362,
0.03910506144165993,
-0.0482795387506485,
0.03984237462282181,
-0.047714900225400925,
-0.0594254806637764,
0.05482788383960724,
-0.04833870008587837,
-0.045752398669719696,
0.02764890342950821,
0.043127384036779404,
0.03654267266392708,
-0.060744430869817734,
0.06861846894025803,
-0.049895983189344406,
0.03627583011984825,
0.04821772873401642,
-0.07980265468358994,
-0.047439444810152054,
-0.06725041568279266,
-0.02539266273379326,
0.10508706420660019,
0.06393378227949142,
0.004975176881998777,
-0.019882801920175552,
-0.00853321235626936,
-0.02262197621166706,
-0.06335565447807312,
0.05601084977388382,
0.05669163540005684,
-0.007298447657376528,
-0.06327248364686966,
0.001467387774027884,
-0.10614427179098129,
0.05946894362568855,
-0.05278884992003441,
-0.038788583129644394,
-0.005425712093710899,
0.0013314688112586737,
0.0110443951562047,
-0.0866430327296257,
-0.04845225438475609,
-0.022589121013879776,
-0.019292527809739113,
-0.055846478790044785,
-0.023525632917881012,
0.12318496406078339,
0.07552313804626465,
0.005504939705133438,
0.022577397525310516,
-0.04872734099626541,
0.04722500592470169,
0.022758029401302338,
-0.03968026861548424,
0.024548638612031937,
0.07937002927064896,
0.045316342264413834,
0.07127703726291656,
-0.033256109803915024,
0.10392294079065323,
0.028417153283953667,
0.021955246105790138,
-0.03074703738093376,
0.09592827409505844,
0.09790636599063873,
-0.037171147763729095,
-0.08920559287071228,
-0.04223671555519104,
-0.04272785782814026,
-0.007507590111345053,
-0.029525095596909523,
-0.045048683881759644,
-0.017693428322672844,
-0.048431314527988434,
0.07749998569488525,
-0.033080559223890305,
0.004469188861548901,
-0.052613548934459686,
0.002842737827450037,
-0.031835898756980896,
-0.012038675136864185,
-0.061557043343782425,
-0.010279345326125622,
-0.015031494200229645,
-0.0034058678429573774,
-0.02269754745066166,
0.02537696063518524,
-0.02047213912010193,
-0.060090627521276474,
-0.017947040498256683,
0.060051072388887405,
-7.737993473481327e-35,
-0.004211538005620241,
-0.054410770535469055,
-0.023749902844429016,
0.06750969588756561,
-0.0109872380271554,
-0.08667518198490143,
-0.0276638250797987,
0.03728513419628143,
0.0778827890753746,
-0.12608657777309418,
-0.021282242611050606,
-0.03675961494445801,
0.027852337807416916,
-0.03175477311015129,
0.015408450737595558,
-0.013913819566369057,
-0.0016317201079800725,
0.010707572102546692,
0.003786776913329959,
0.05073367431759834,
0.010669298470020294,
-0.041214603930711746,
-0.012383369728922844,
0.07405457645654678,
-0.027003809809684753,
0.04364843666553497,
-0.03883541002869606,
0.06775039434432983,
0.007428511045873165,
0.0039559355936944485,
-0.05228911340236664,
-0.021687446162104607,
-0.0343492291867733,
-0.010436434298753738,
-0.04377603903412819,
-0.022666465491056442,
0.08080586045980453,
-0.04929220676422119,
-0.03719639778137207,
0.08314384520053864,
0.09038680791854858,
-0.02016354724764824,
-0.0549619235098362,
-0.0052450536750257015,
-0.01319667138159275,
0.008348728530108929,
-0.0935477539896965,
0.01836865395307541,
0.039026305079460144,
-0.027142537757754326,
-0.06700186431407928,
-0.06573771685361862,
-0.07200179249048233,
0.03449337184429169,
0.02315916121006012,
-0.07019715011119843,
-0.019236885011196136,
-0.0521845817565918,
0.04584797844290733,
-0.032659970223903656,
-0.0658896341919899,
0.06856491416692734,
0.05109510198235512,
-0.007018487434834242,
0.05414598807692528,
-0.004371738992631435,
-0.04542801156640053,
0.010035553947091103,
0.0034803934395313263,
0.02740877866744995,
0.020858565345406532,
0.0015234372112900019,
-0.054192010313272476,
-0.04817637428641319,
0.08296012878417969,
-0.026976723223924637,
0.042366873472929,
-0.0830199345946312,
0.008831221610307693,
-0.13209743797779083,
-0.036561135202646255,
-0.04475705698132515,
0.023160947486758232,
0.028292115777730942,
0.04203900322318077,
0.08814306557178497,
0.004869275726377964,
0.1576385498046875,
-0.007919983938336372,
0.031854405999183655,
-0.08743319660425186,
0.020564202219247818,
-0.003044914221391082,
0.053489334881305695,
-0.14118683338165283,
-3.5849836876877816e-8,
-0.030569642782211304,
0.06800693273544312,
-0.012883051298558712,
0.07748804986476898,
0.0771106705069542,
-0.04285644739866257,
-0.04221363738179207,
0.11220566183328629,
0.048258401453495026,
-0.05517954006791115,
0.05349675565958023,
0.03285753354430199,
-0.014226117171347141,
-0.036299608647823334,
-0.058823950588703156,
0.08222238719463348,
-0.049104321748018265,
0.09982211887836456,
-0.01112484373152256,
0.023635296151041985,
0.05671117827296257,
0.049668800085783005,
0.016560953110456467,
0.0073902723379433155,
0.04140615463256836,
-0.07091468572616577,
-0.046651411801576614,
0.09088042378425598,
0.034141313284635544,
0.006788867525756359,
-0.02462293952703476,
-0.04422944411635399,
0.06781090795993805,
0.054810866713523865,
0.03523701801896095,
0.043098535388708115,
0.08165013045072556,
0.0003165343077853322,
0.009731147438287735,
-0.010745118372142315,
0.062433261424303055,
0.026345625519752502,
-0.022067608311772346,
0.014910445548593998,
-0.03336396813392639,
-0.06563390046358109,
0.01336620282381773,
-0.06432168185710907,
0.025920378044247627,
-0.07496824860572815,
-0.008048446848988533,
0.022381840273737907,
0.052764810621738434,
-0.0058947657234966755,
0.01752522401511669,
0.09531527757644653,
-0.02523968555033207,
0.022552410140633583,
0.025129370391368866,
-0.031240439042448997,
0.04367624223232269,
-0.03747060149908066,
0.002301180502399802,
0.06105956807732582
] |
Advantage-Guided Distillation for Preference Alignment in Small Language Models | https://openreview.net/forum?id=xsx3Fpo3UD | [
"Shiping Gao",
"Fanqi Wan",
"Jiajian Guo",
"Xiaojun Quan",
"Qifan Wang"
] | Spotlight | Alignment techniques enable Large Language Models (LLMs) to generate outputs that align with human preferences and play a crucial role in their effectiveness. However, their impact often diminishes when applied to Small Language Models (SLMs), likely due to the limited capacity of these models. Instead of directly applying existing alignment techniques to SLMs, we propose to utilize a well-aligned teacher LLM to guide the alignment process for these models, thereby facilitating the transfer of the teacher's knowledge of human preferences to the student model. To achieve this, we first explore a straightforward approach, Dual-Constrained Knowledge Distillation (DCKD), that employs knowledge distillation with two KL-divergence constraints from the aligned teacher to the unaligned student. To further enhance the student's ability to distinguish between preferred and dispreferred responses, we then propose Advantage-Guided Distillation for Preference Alignment (ADPA), which leverages an advantage function from the aligned teacher to deliver more nuanced, distribution-level reward signals for the student's alignment. Our experimental results show that these two approaches appreciably improve the alignment of SLMs and narrow the performance gap with larger counterparts. Among them, ADPA demonstrates superior performance and achieves even greater effectiveness when integrated with DCKD. Our code is available at https://github.com/SLIT-AI/ADPA . | Preference Alignment; Large language model; Knowledge Distillation; Advantage Function | null | 7,508 | 2502.17927 | [
-0.0381377749145031,
-0.05777842923998833,
0.09492708742618561,
-0.0029664274770766497,
-0.033355630934238434,
0.0038807920645922422,
-0.012415890581905842,
0.0025693592615425587,
0.04917315021157265,
0.020936427637934685,
0.010585968382656574,
-0.005226854234933853,
0.04361825808882713,
0.026312608271837234,
0.05116528645157814,
0.01441369391977787,
0.03947092965245247,
0.1211477667093277,
-0.12557640671730042,
-0.03362727537751198,
-0.0036544655449688435,
-0.024451322853565216,
-0.051401298493146896,
-0.008903122507035732,
-0.004551185294985771,
0.05071425810456276,
0.042009759694337845,
0.0194874070584774,
0.03567112982273102,
-0.05472666397690773,
0.011299561709165573,
-0.019674239680171013,
-0.029620610177516937,
0.04281600937247276,
-0.05437299609184265,
0.05909162014722824,
-0.1097937673330307,
0.03559447079896927,
-0.014016788452863693,
-0.021530991420149803,
-0.06393613666296005,
0.038104619830846786,
-0.006734798662364483,
0.02737751044332981,
0.033609505742788315,
0.015625158324837685,
-0.05192674323916435,
0.0022987755946815014,
0.027610789984464645,
0.022166257724165916,
-0.047050561755895615,
-0.008069257251918316,
-0.0020707601215690374,
0.049358759075403214,
0.013585672713816166,
0.05603170022368431,
-0.024671077728271484,
0.031221479177474976,
0.019579105079174042,
-0.06060236692428589,
-0.06658878922462463,
-0.1038561686873436,
-0.05841445550322533,
0.03509495034813881,
-0.026588913053274155,
-0.06902537494897842,
-0.019656546413898468,
0.07457529008388519,
-0.018896454945206642,
0.05445423349738121,
-0.06873225420713425,
0.09611909836530685,
0.04547261819243431,
0.05456560105085373,
0.06421884894371033,
0.05575011298060417,
0.04066081345081329,
-0.0028747429605573416,
0.03242041543126106,
-0.022001653909683228,
0.023318147286772728,
0.0740489587187767,
0.062206584960222244,
-0.025842079892754555,
0.13140642642974854,
-0.010707239620387554,
0.005396406631916761,
-0.012016106396913528,
0.09115434437990189,
-0.04340536892414093,
-0.0705932155251503,
-0.03874966502189636,
-0.006208496168255806,
0.006397297140210867,
0.030146688222885132,
0.06729467958211899,
0.008230534382164478,
0.03134802356362343,
-0.08534019440412521,
0.06696899980306625,
0.05102642998099327,
0.12588953971862793,
-0.024796009063720703,
-0.09641432762145996,
-0.016875851899385452,
-0.12287186086177826,
0.012011180631816387,
0.03840534761548042,
0.03704440966248512,
-0.1308104246854782,
0.0321197584271431,
0.05443784222006798,
0.03517576679587364,
0.06540501117706299,
-0.04271627590060234,
-0.048087649047374725,
0.0453454852104187,
-0.0480634979903698,
0.018332980573177338,
-0.04318050295114517,
-0.019536443054676056,
-0.06174524873495102,
-0.014903431758284569,
-0.0003629218554124236,
-0.03149779513478279,
-0.01376763079315424,
-0.031496014446020126,
1.1258147453141721e-33,
0.019083935767412186,
0.1299402415752411,
0.020334208384156227,
0.051037199795246124,
0.013513615354895592,
-0.01913304068148136,
-0.0195261649787426,
0.01046321727335453,
-0.12232806533575058,
0.04310174286365509,
-0.003217455232515931,
0.04731475189328194,
-0.0826411098241806,
0.0378296934068203,
0.016187796369194984,
-0.00069995759986341,
-0.0455874539911747,
0.05443543568253517,
-0.07786126434803009,
0.003617767710238695,
0.03462507575750351,
0.027305882424116135,
0.017351990565657616,
-0.04347650334239006,
-0.04062962532043457,
-0.035775937139987946,
0.04462006315588951,
-0.014254110865294933,
-0.05348971113562584,
-0.015108379535377026,
-0.03688341751694679,
-0.07000769674777985,
0.004291227553039789,
-0.022227922454476357,
-0.005681355018168688,
-0.08768542110919952,
-0.024496017023921013,
-0.06720296293497086,
0.011427171528339386,
-0.04828092083334923,
-0.01663350686430931,
0.054372187703847885,
0.032833222299814224,
0.016536224633455276,
-0.05807988718152046,
0.03092360682785511,
0.062369074672460556,
0.008233349770307541,
0.020888613536953926,
-0.10720475018024445,
-0.023976344615221024,
0.031175268813967705,
-0.05296362563967705,
-0.050778768956661224,
-0.03964759409427643,
0.018796849995851517,
0.03061656281352043,
0.03270592913031578,
0.005980352871119976,
-0.01799684204161167,
-0.05817270278930664,
0.04698975011706352,
-0.0022784676402807236,
0.02641657367348671,
0.06934625655412674,
-0.037578895688056946,
-0.0439029335975647,
0.03299666941165924,
0.1063586100935936,
0.012405535206198692,
-0.06349441409111023,
-0.04658591374754906,
-0.07487040013074875,
-0.08084968477487564,
-0.009666071273386478,
0.002999313874170184,
0.053985126316547394,
-0.09393744170665741,
0.1133197769522667,
-0.031975146383047104,
0.0434240959584713,
-0.07695727795362473,
-0.05111437290906906,
-0.08606090396642685,
-0.0919952243566513,
-0.055370792746543884,
0.041839323937892914,
-0.08544007688760757,
0.021083956584334373,
-0.04125123843550682,
0.012911365367472172,
0.04588524252176285,
-0.02953481301665306,
0.031479090452194214,
0.05950581282377243,
-1.712931286373199e-33,
-0.011424725875258446,
-0.007623626384884119,
0.027988005429506302,
0.08171699941158295,
0.04656536877155304,
0.05183681473135948,
0.036864567548036575,
0.01454127486795187,
-0.03393268957734108,
-0.05028163269162178,
0.02423076517879963,
-0.05444476753473282,
-0.04050596430897713,
0.012934918515384197,
-0.011121603660285473,
0.0813932865858078,
-0.006309432443231344,
0.08027244359254837,
0.041749339550733566,
0.022660478949546814,
0.010562688112258911,
0.10835682600736618,
-0.10429921746253967,
0.0846184492111206,
0.012392979115247726,
0.014742882922291756,
0.020990217104554176,
0.056472048163414,
-0.051221128553152084,
0.03391091898083687,
-0.013288886286318302,
0.017823711037635803,
-0.021112674847245216,
-0.03460165858268738,
-0.11772742867469788,
0.012908406555652618,
-0.06164269521832466,
-0.07565322518348694,
-0.013992822729051113,
0.13978634774684906,
-0.035367708653211594,
-0.03194732591509819,
-0.01703089103102684,
-0.05753285437822342,
-0.016261929646134377,
0.04612129554152489,
-0.033355459570884705,
-0.05549285560846329,
0.029586657881736755,
-0.0069068861193954945,
0.040983591228723526,
0.03899388387799263,
-0.07225947827100754,
0.030904507264494896,
-0.021842211484909058,
-0.027232730761170387,
-0.005985385738313198,
-0.0493016391992569,
0.06506115943193436,
0.052490122616291046,
-0.02491481974720955,
-0.03144500032067299,
0.09862799197435379,
-0.08874624222517014,
-0.022474505007267,
-0.03555009141564369,
0.03381508216261864,
0.09530377388000488,
-0.0267995186150074,
-0.04795265942811966,
-0.004362705163657665,
-0.00795813463628292,
0.02338276244699955,
-0.04746460169553757,
0.008368209935724735,
-0.028718819841742516,
0.012396732345223427,
-0.02701904997229576,
-0.0388069711625576,
-0.0433821938931942,
-0.011636268347501755,
0.013943579979240894,
0.00031328314798884094,
0.020471680909395218,
0.04381963610649109,
0.05126942694187164,
0.08354389667510986,
0.10672637075185776,
0.01541863288730383,
0.043130457401275635,
-0.019279437139630318,
0.011982541531324387,
0.06682558357715607,
0.054148003458976746,
-0.04094703868031502,
-4.3097127644387e-8,
-0.12328599393367767,
-0.03886152431368828,
0.006951283663511276,
0.1059514582157135,
-0.07267498970031738,
-0.0583176389336586,
-0.027122696861624718,
0.05543585866689682,
-0.03810090571641922,
-0.016951708123087883,
0.047016315162181854,
0.06137534976005554,
-0.03514600545167923,
-0.01564241573214531,
0.03415350243449211,
0.13456645607948303,
-0.02793755941092968,
0.05991816148161888,
-0.002798809204250574,
0.024471687152981758,
0.0998305231332779,
0.013246366754174232,
0.0038694872055202723,
0.028890322893857956,
0.03906784951686859,
-0.07226914167404175,
-0.02703971602022648,
0.13783659040927887,
0.038861315697431564,
-0.019783608615398407,
-0.023383699357509613,
-0.046545833349227905,
-0.0023803943768143654,
-0.06292320042848587,
0.00938345119357109,
0.11473648995161057,
-0.05993284657597542,
-0.04012395814061165,
0.02060549333691597,
0.02194424346089363,
0.015153354033827782,
0.07070273905992508,
-0.06552108377218246,
0.06973910331726074,
-0.04945025220513344,
0.010040433146059513,
-0.017134467139840126,
-0.04086149111390114,
0.015446273609995842,
0.017640767619013786,
0.0460156574845314,
-0.042186882346868515,
0.02647736296057701,
-0.04236500710248947,
0.03714067488908768,
0.01163728442043066,
-0.07408878207206726,
0.0024760295636951923,
0.012703720480203629,
-0.03762713819742203,
0.06955759972333908,
0.03925796225667,
-0.03875114396214485,
0.0070721665397286415
] |
SVBench: A Benchmark with Temporal Multi-Turn Dialogues for Streaming Video Understanding | https://openreview.net/forum?id=Hz4BYVY8YM | [
"Zhenyu Yang",
"Yuhang Hu",
"Zemin Du",
"Dizhan Xue",
"Shengsheng Qian",
"Jiahong Wu",
"Fan Yang",
"Weiming Dong",
"Changsheng Xu"
] | Spotlight | Despite the significant advancements of Large Vision-Language Models (LVLMs) on established benchmarks, there remains a notable gap in suitable evaluation regarding their applicability in the emerging domain of long-context streaming video understanding. Current benchmarks for video understanding typically emphasize isolated single-instance text inputs and fail to evaluate the capacity to sustain temporal reasoning throughout the entire duration of video streams. To address these limitations, we introduce SVBench, a pioneering benchmark with temporal multi-turn question-answering chains specifically designed to thoroughly assess the capabilities of streaming video understanding of current LVLMs. We design a semi-automated annotation pipeline to obtain 49,979 Question-Answer (QA) pairs of 1,353 streaming videos, which includes generating QA chains that represent a series of consecutive multi-turn dialogues over video segments and constructing temporal linkages between successive QA chains. Our experimental results, obtained from 14 models in dialogue and streaming evaluations, reveal that while the closed-source GPT-4o outperforms others, most open-source LVLMs struggle with long-context streaming video understanding. We also construct a StreamingChat model, which significantly outperforms open-source LVLMs on our SVBench and achieves comparable performance on diverse vision-language benchmarks. We expect SVBench to advance the research of streaming video understanding by providing a comprehensive and in-depth analysis of current LVLMs. Our benchmark and model can be accessed at https://yzy-bupt.github.io/SVBench. | Multimodal large language model, Streaming video analysis, Video understanding | A benchmark with temporal multi-turn dialogues specifically designed to thoroughly assess the capabilities of long-context streaming video understanding of current LVLMs. | 7,487 | 2502.10810 | [
-0.030707379803061485,
-0.06721515953540802,
0.00505982106551528,
-0.05669769272208214,
0.15251773595809937,
0.020906519144773483,
-0.030429352074861526,
0.013316205702722073,
0.08676061779260635,
-0.0805647000670433,
-0.12016259878873825,
-0.08108998090028763,
-0.09143990278244019,
0.11716286092996597,
-0.004043920431286097,
-0.01605108007788658,
0.027341343462467194,
-0.010575037449598312,
-0.10947999358177185,
-0.02152066119015217,
0.020649567246437073,
-0.035063426941633224,
0.0804566964507103,
0.009147881530225277,
0.03556228429079056,
-0.006925147492438555,
-0.04278728738427162,
-0.07897546142339706,
0.05206475034356117,
0.013277757912874222,
0.05414123088121414,
0.034533023834228516,
-0.011704432778060436,
0.08031152933835983,
-0.028965914621949196,
0.016441645100712776,
-0.033481664955616,
-0.009587298147380352,
-0.0777633860707283,
-0.007040127646178007,
0.019154228270053864,
-0.00753792142495513,
0.07996299862861633,
0.002199288457632065,
0.0589122511446476,
-0.015484742820262909,
-0.0436655730009079,
-0.08065397292375565,
0.03259861096739769,
0.02490118332207203,
-0.10223612934350967,
-0.008432808332145214,
0.022160224616527557,
0.04100821912288666,
-0.07391050457954407,
0.06141950935125351,
0.0044881426729261875,
0.05315465107560158,
0.03492708504199982,
0.0035901509691029787,
0.00834078062325716,
-0.13213397562503815,
-0.019777284935116768,
0.04187257960438728,
-0.06856770068407059,
-0.034865450114011765,
0.033657919615507126,
0.0026643015444278717,
-0.02967766299843788,
0.043024737387895584,
-0.0696612074971199,
0.06832386553287506,
-0.02435210905969143,
0.07473263889551163,
-0.016493912786245346,
0.009317835792899132,
0.02882993593811989,
-0.03264646232128143,
0.007316218223422766,
-0.10216069966554642,
0.07493133842945099,
-0.018585503101348877,
0.06988997012376785,
0.03217393532395363,
0.1297583132982254,
-0.002514281775802374,
0.020114336162805557,
0.06012215465307236,
-0.059582874178886414,
0.05412214249372482,
-0.1268714815378189,
-0.039955031126737595,
0.09592334181070328,
-0.03167116269469261,
0.0805990919470787,
0.05682436376810074,
0.004475978668779135,
-0.09586931765079498,
-0.002781978342682123,
0.0637204647064209,
0.042931362986564636,
0.06198696047067642,
-0.0011345709208399057,
-0.06861312687397003,
0.004233236890286207,
0.02169337496161461,
0.03245296701788902,
0.08660256862640381,
-0.010604079812765121,
-0.05741399899125099,
0.0572444386780262,
0.040802597999572754,
0.09346619248390198,
-0.03891872242093086,
0.02981300838291645,
0.010811518877744675,
0.0015515269478783011,
0.0017981941346079111,
0.024227967485785484,
0.041573211550712585,
0.0231850054115057,
-0.027230912819504738,
-0.014492795802652836,
-0.05136736109852791,
0.030612749978899956,
-0.00841441284865141,
0.05375954136252403,
5.713323151568721e-34,
0.05786997452378273,
-0.013897623866796494,
0.0470719039440155,
0.04280310124158859,
0.007293131668120623,
0.0612063892185688,
-0.042977020144462585,
0.08496899157762527,
-0.07596195489168167,
-0.04997841641306877,
0.036169975996017456,
-0.0608539842069149,
-0.08255228400230408,
-0.000375175237422809,
-0.011386490426957607,
0.009129473939538002,
-0.07649475336074829,
0.006362245883792639,
-0.07260610908269882,
-0.02475198544561863,
0.03819165751338005,
-0.020252447575330734,
-0.000851410673931241,
0.036037079989910126,
0.06635000556707382,
-0.05004403740167618,
0.0827236920595169,
-0.11021438241004944,
-0.018591737374663353,
-0.012924513779580593,
-0.10623553395271301,
-0.0531870499253273,
-0.005182371009141207,
0.026712022721767426,
0.09319402277469635,
-0.04525237902998924,
-0.02013910748064518,
-0.04781894385814667,
-0.002595324767753482,
-0.047940488904714584,
-0.062483880668878555,
-0.010913774371147156,
-0.011351135559380054,
-0.023870883509516716,
-0.0696534812450409,
-0.03608681634068489,
-0.010444064624607563,
0.08601444959640503,
-0.07373703271150589,
0.0013070356799289584,
0.028565842658281326,
0.07495167851448059,
-0.026560481637716293,
-0.052459705621004105,
-0.005258290097117424,
0.0029590479098260403,
0.029249973595142365,
0.022398214787244797,
-0.014752554707229137,
0.0219960305839777,
0.04216285049915314,
0.0812351182103157,
-0.020686816424131393,
0.05759429559111595,
-0.00838504359126091,
0.03978446498513222,
0.02051195129752159,
0.084637850522995,
0.026094039902091026,
0.04263472557067871,
-0.01984941028058529,
-0.002081414684653282,
0.04911251366138458,
-0.05014386773109436,
0.07019298523664474,
0.04970216006040573,
-0.017976466566324234,
-0.06605654954910278,
0.06060617044568062,
0.04214467108249664,
0.017846006900072098,
-0.06700531393289566,
0.04405484348535538,
-0.03573925048112869,
-0.0018697618506848812,
-0.06481847167015076,
0.05783158540725708,
-0.11295672506093979,
-0.018747003749012947,
-0.01583445630967617,
-0.04533662274479866,
-0.026501189917325974,
0.017410852015018463,
-0.022106874734163284,
0.0393298976123333,
5.396717982676642e-34,
0.030736248940229416,
0.017555667087435722,
-0.05626034736633301,
0.054210033267736435,
0.045030876994132996,
-0.047205280512571335,
-0.03824520483613014,
0.036785226315259933,
0.042258888483047485,
-0.04622096195816994,
0.008788076229393482,
-0.06088564917445183,
-0.06964366883039474,
-0.0053177191875875,
0.025629710406064987,
-0.052411410957574844,
0.006439859513193369,
-0.07193119078874588,
-0.04608292505145073,
0.062304265797138214,
0.009357625618577003,
0.03167813643813133,
-0.0616178922355175,
0.005820019636303186,
-0.018191466107964516,
0.04481435939669609,
0.037254393100738525,
0.07057791203260422,
-0.02948615700006485,
-0.033344022929668427,
-0.0018161620246246457,
-0.003757427679374814,
-0.03786301612854004,
-0.053036585450172424,
-0.08807792514562607,
0.04451613873243332,
0.07980242371559143,
-0.08183769881725311,
-0.04535365849733353,
0.06539520621299744,
0.11830391734838486,
-0.0825023353099823,
-0.04786543548107147,
-0.029638132080435753,
-0.03594299033284187,
-0.022138863801956177,
-0.10826889425516129,
0.011245977133512497,
-0.07382992655038834,
0.026488017290830612,
-0.04871679097414017,
0.021901195868849754,
-0.06085909157991409,
0.04143747314810753,
-0.006125498563051224,
-0.06995005160570145,
-0.0983586236834526,
0.03718991577625275,
0.023244909942150116,
-0.029811909422278404,
-0.024796634912490845,
0.02774050459265709,
0.007867677137255669,
0.0025647461879998446,
0.07877462357282639,
-0.04918413609266281,
0.002048179740086198,
-0.019224610179662704,
-0.01883271522819996,
-0.024043496698141098,
0.0388663150370121,
-0.041634272783994675,
-0.005401998292654753,
-0.0016294674715027213,
0.02926149219274521,
0.014481820166110992,
-0.0019684238359332085,
-0.043878473341464996,
0.03295939788222313,
-0.059898074716329575,
-0.011160651221871376,
-0.037378761917352676,
0.08525528758764267,
0.050330888479948044,
0.07857265323400497,
0.03449055925011635,
0.06194812431931496,
0.08580197393894196,
0.03526993840932846,
0.10787422209978104,
0.00302445818670094,
0.0201075728982687,
-0.03819819167256355,
0.08371328562498093,
-0.006785918492823839,
-3.37534373784365e-8,
-0.08759620040655136,
0.055058594793081284,
-0.08710221201181412,
0.030826134607195854,
-0.004415886476635933,
-0.03239170461893082,
-0.013828352093696594,
0.09204568713903427,
0.015606033615767956,
-0.008791545405983925,
0.049969520419836044,
0.020661761984229088,
-0.01650596410036087,
-0.035046566277742386,
-0.03498751297593117,
0.09162262827157974,
-0.011794454418122768,
-0.02729135937988758,
-0.006943338550627232,
0.017461517825722694,
0.0951700210571289,
0.021130016073584557,
0.0023828567937016487,
0.024590732529759407,
-0.10980527848005295,
0.01266559213399887,
-0.0003984069044236094,
0.10823488235473633,
0.11499156802892685,
-0.05812807008624077,
-0.010313494130969048,
0.036131806671619415,
-0.0243354644626379,
-0.0573880672454834,
-0.0023917299695312977,
0.01786193624138832,
-0.02999097853899002,
-0.020922575145959854,
0.07033152133226395,
0.008886740542948246,
0.026594167575240135,
0.0431932657957077,
-0.02099599316716194,
0.03493274375796318,
-0.023687710985541344,
0.031241774559020996,
-0.06146093085408211,
-0.04808276519179344,
0.007756710052490234,
-0.016493510454893112,
-0.04910683631896973,
-0.0716274157166481,
-0.004740322940051556,
0.09743896126747131,
0.05766189843416214,
-0.0063081178814172745,
0.024295508861541748,
-0.005440305452793837,
0.03212980180978775,
0.0576062835752964,
-0.02105529233813286,
0.042640794068574905,
0.024523891508579254,
0.0329231321811676
] |
Knowledge Distillation with Multi-granularity Mixture of Priors for Image Super-Resolution | https://openreview.net/forum?id=cWHonXThtM | [
"Simiao Li",
"Yun Zhang",
"Wei Li",
"Hanting Chen",
"Wenjia Wang",
"Bingyi Jing",
"Shaohui Lin",
"Jie Hu"
] | Spotlight | Knowledge distillation (KD) is a promising yet challenging model compression approach that transmits rich learning representations from robust but resource-demanding teacher models to efficient student models. Previous methods for image super-resolution (SR) are often tailored to specific teacher-student architectures, limiting their potential for improvement and hindering broader applications. This work presents a novel KD framework for SR models, the multi-granularity Mixture of Priors Knowledge Distillation (MiPKD), which can be universally applied to a wide range of architectures at both feature and block levels. The teacher’s knowledge is effectively integrated with the student's feature via the Feature Prior Mixer, and the reconstructed feature propagates dynamically in the training phase with the Block Prior Mixer. Extensive experiments illustrate the significance of the proposed MiPKD technique. | Image Super-Resolution, Knowledge Distillation, Model Compression | null | 7,476 | 2404.02573 | [
0.0009318246738985181,
-0.03429847210645676,
0.11792535334825516,
-0.03733557462692261,
0.001997187966480851,
-0.030372969806194305,
-0.043976228684186935,
-0.036824654787778854,
0.019613424316048622,
-0.021311188116669655,
-0.021855542436242104,
0.0915280133485794,
0.05150753632187843,
0.05006491020321846,
-0.06797581911087036,
0.03164856880903244,
0.05714616924524307,
0.10527341812849045,
-0.11469939351081848,
-0.07996100932359695,
0.058209631592035294,
-0.07614153623580933,
-0.029079603031277657,
-0.029424138367176056,
0.10386396944522858,
0.06538847088813782,
0.0401843786239624,
-0.04766590893268585,
0.1050071194767952,
-0.0010452336864545941,
-0.0004154347989242524,
0.015641601756215096,
0.0067301723174750805,
0.05253620073199272,
-0.07196558266878128,
0.03941895440220833,
-0.0005862324615009129,
0.11948782950639725,
-0.03400907292962074,
-0.01691213622689247,
0.016571534797549248,
0.0732426717877388,
0.02202431485056877,
-0.00027105279150418937,
0.04461148753762245,
0.04087822511792183,
0.014113524928689003,
-0.041655611246824265,
0.025678545236587524,
-0.04046984016895294,
-0.03339892625808716,
-0.06114601716399193,
-0.06556349992752075,
0.04846804961562157,
0.02114175446331501,
0.0672837421298027,
-0.011770512908697128,
-0.06793250143527985,
-0.024640901014208794,
0.07503253221511841,
-0.10923831909894943,
-0.031139343976974487,
-0.03713733330368996,
0.017516473308205605,
0.09696361422538757,
0.030177392065525055,
-0.015900220721960068,
0.06286939978599548,
0.05557745322585106,
-0.009635946713387966,
-0.03393873944878578,
0.023433910682797432,
0.055765628814697266,
-0.00986203271895647,
0.049429699778556824,
-0.002273996127769351,
0.009573249146342278,
0.02344590239226818,
0.05731853470206261,
-0.07044447213411331,
0.027883317321538925,
-0.03323593735694885,
0.03679535537958145,
-0.10870286822319031,
0.09713512659072876,
-0.002569829346612096,
0.052890241146087646,
-0.0029213798698037863,
0.0037774110678583384,
-0.03768086060881615,
-0.07834669202566147,
-0.05187922716140747,
-0.03560592606663704,
-0.01428279373794794,
0.01625388115644455,
0.005940273869782686,
0.09868881106376648,
-0.08227070420980453,
-0.002198378322646022,
0.05824730917811394,
-0.022987741976976395,
0.007841677404940128,
0.05831488221883774,
-0.07167764008045197,
0.04572880640625954,
-0.09645070135593414,
0.049168769270181656,
0.08366642147302628,
-0.011914427392184734,
-0.041061241179704666,
0.055166300386190414,
0.01521229837089777,
0.043240971863269806,
0.004547515418380499,
0.016324881464242935,
-0.021752241998910904,
0.05429467558860779,
-0.008474919945001602,
-0.05939735099673271,
-0.06661001592874527,
0.02516307681798935,
-0.022248905152082443,
-0.04937049001455307,
0.015669679269194603,
-0.006435910705476999,
-0.03309336304664612,
-0.08431562781333923,
2.7415125369007093e-33,
0.06600873917341232,
0.058618444949388504,
0.000793912447988987,
0.04794599488377571,
0.014584317803382874,
0.027651524171233177,
0.032520320266485214,
0.01942325569689274,
0.026412589475512505,
0.07780771702528,
0.0003562447091098875,
-0.054350145161151886,
-0.08612047135829926,
0.08886288106441498,
-0.008622286841273308,
0.01598481833934784,
-0.08900589495897293,
0.04780895635485649,
0.0066989390179514885,
0.048442017287015915,
0.060540035367012024,
-0.05683470517396927,
0.00044817107846029103,
-0.033956289291381836,
-0.08089033514261246,
0.008015411905944347,
0.028906548395752907,
-0.0015579957980662584,
0.009588426910340786,
-0.028225092217326164,
-0.04610838741064072,
-0.03394986689090729,
-0.034710004925727844,
0.05162275210022926,
-0.046017780900001526,
-0.06865154206752777,
0.03878161311149597,
-0.04737861081957817,
0.014480427838861942,
0.0017316850135102868,
0.009315802715718746,
0.01504606194794178,
-0.016840707510709763,
-0.05054846405982971,
-0.06897394359111786,
-0.006954558659344912,
0.0669712945818901,
0.026994679123163223,
-0.04704277962446213,
0.0024783993139863014,
0.010550593957304955,
-0.032918598502874374,
-0.08419431746006012,
-0.09478902071714401,
-0.03900318592786789,
0.026206202805042267,
0.05842740833759308,
0.015035996213555336,
0.04720240458846092,
0.016842445358633995,
-0.022963061928749084,
0.00012221939687151462,
-0.03192504122853279,
0.10338234156370163,
0.06087411940097809,
-0.020658688619732857,
-0.03347841277718544,
0.06493564695119858,
0.02975248172879219,
-0.031005417928099632,
-0.058017950505018234,
-0.01528282929211855,
0.016203081235289574,
-0.08190809935331345,
0.030245276167988777,
-0.0633435919880867,
0.002913377247750759,
-0.02713618054986,
-0.006038045510649681,
0.037962909787893295,
-0.03358529508113861,
-0.03305116295814514,
-0.06623220443725586,
-0.0664977952837944,
-0.051518071442842484,
0.1257280856370926,
-0.019530508667230606,
-0.07622258365154266,
0.002951911184936762,
-0.10452930629253387,
-0.04069467633962631,
0.03319847211241722,
0.0024068234488368034,
0.017150836065411568,
0.010294212028384209,
-1.7324628596960805e-33,
0.06979816406965256,
0.061261195689439774,
-0.006847769021987915,
0.04696594178676605,
0.07884098589420319,
-0.03506898880004883,
0.006634208839386702,
0.06438300758600235,
-0.0392594039440155,
-0.09844426065683365,
0.04723906144499779,
-0.021976498886942863,
-0.084810771048069,
-0.07809329032897949,
-0.03476303815841675,
0.044973064213991165,
-0.01575472205877304,
-0.0423472598195076,
-0.021455613896250725,
-0.03143216297030449,
-0.04480745270848274,
0.0008282852359116077,
-0.06967826932668686,
0.06269856542348862,
-0.00478197680786252,
-0.004391833208501339,
-0.04976207762956619,
0.09813180565834045,
0.03718987852334976,
0.08106333017349243,
0.03581409901380539,
-0.05141805112361908,
0.03922196105122566,
-0.06276851892471313,
-0.09784980118274689,
0.051086101680994034,
0.03538991138339043,
-0.0058847451582551,
-0.05976215749979019,
0.051449719816446304,
-0.010768369771540165,
-0.060257818549871445,
-0.044097330421209335,
0.020309660583734512,
0.0236574187874794,
-0.015289265662431717,
0.011948592960834503,
-0.025338763371109962,
0.08527339994907379,
0.02463713474571705,
-0.0008734416915103793,
-0.03635525330901146,
-0.055068712681531906,
-0.027204226702451706,
-0.006776595488190651,
0.02026531472802162,
-0.010632036253809929,
-0.01213084440678358,
0.08151280879974365,
0.017482737079262733,
0.03268926590681076,
-0.14212776720523834,
-0.010074645280838013,
-0.07637608051300049,
-0.005271630827337503,
0.011884568259119987,
-0.041613563895225525,
0.07843334227800369,
-0.07688857614994049,
0.09427623450756073,
-0.07888839393854141,
0.0011956884991377592,
0.035652972757816315,
-0.016852308064699173,
0.04111098498106003,
-0.03367423266172409,
0.027691084891557693,
0.024043647572398186,
0.019892005249857903,
0.015274356119334698,
-0.031835440546274185,
-0.016834499314427376,
-0.015245910733938217,
0.07819639891386032,
0.11817264556884766,
0.08675341308116913,
0.003968002274632454,
0.028946638107299805,
0.06178140640258789,
-0.022137295454740524,
-0.04403636232018471,
0.09519680589437485,
0.037371374666690826,
-0.040335867553949356,
0.039787475019693375,
-3.9688643482804764e-8,
-0.10358256846666336,
0.019102448597550392,
-0.056744884699583054,
-0.018141526728868484,
0.01873372681438923,
-0.06810946762561798,
-0.04339952394366264,
0.07603073120117188,
0.03748203441500664,
-0.06540471315383911,
0.05601558834314346,
-0.0429287925362587,
-0.09049848467111588,
-0.02095184288918972,
0.06660828739404678,
0.032592836767435074,
0.024387378245592117,
0.09860417991876602,
-0.018276622518897057,
-0.035112302750349045,
0.04359595477581024,
-0.07695832848548889,
0.016452789306640625,
0.0029980777762830257,
0.03241177275776863,
-0.07846469432115555,
0.04942384362220764,
0.10426449775695801,
0.04841584339737892,
0.0420759953558445,
-0.0554855614900589,
-0.011252118274569511,
0.05048682913184166,
-0.04677734896540642,
0.05620691925287247,
0.04860404133796692,
0.03822316601872444,
-0.005477114114910364,
-0.009169328026473522,
-0.06818050891160965,
-0.01930055394768715,
-0.009743166156113148,
-0.04214756190776825,
0.02004864625632763,
0.05688714608550072,
0.0640641450881958,
-0.004946444649249315,
-0.04521343484520912,
0.021954631432890892,
0.14885616302490234,
0.03569244593381882,
0.042751505970954895,
0.03252685070037842,
0.005909333471208811,
0.09565147757530212,
-0.006623567081987858,
0.03704822063446045,
-0.006382733583450317,
0.045906394720077515,
-0.016922567039728165,
0.029056699946522713,
0.03178858011960983,
-0.045475710183382034,
0.046879418194293976
] |
Controlling Language and Diffusion Models by Transporting Activations | https://openreview.net/forum?id=l2zFn6TIQi | [
"Pau Rodriguez",
"Arno Blaas",
"Michal Klein",
"Luca Zappella",
"Nicholas Apostoloff",
"marco cuturi",
"Xavier Suau"
] | Spotlight | The increasing capabilities of large generative models and their ever more widespread deployment have raised concerns about their reliability, safety, and potential misuse. To address these issues, recent works have proposed to control model generation by steering model activations in order to effectively induce or prevent the emergence of concepts or behaviors in the generated output.
In this paper we introduce Activation Transport (AcT), a general framework to steer activations guided by optimal transport theory that generalizes many previous activation-steering works. AcT is modality-agnostic and provides fine-grained control over the model behavior with negligible computational overhead, while minimally impacting model abilities. We experimentally show the effectiveness and versatility of our approach by addressing key challenges in large language models (LLMs) and text-to-image diffusion models (T2Is). For LLMs, we show that AcT can effectively mitigate toxicity, induce arbitrary concepts, and increase their truthfulness. In T2Is, we show how AcT enables fine-grained style control and concept negation. | controllability, generative models, toxicity, diffusion | We propose an inference-time intervention framework based on Optimal Transport that generalizes previous methods and allows interpretable control of both LLMs and Diffusion models. | 7,460 | 2410.23054 | [
0.03436309099197388,
-0.15389959514141083,
0.05381038039922714,
0.09230973571538925,
0.06525862962007523,
-0.01839521899819374,
0.0061002797447144985,
0.009718538261950016,
0.09281934797763824,
-0.03425135090947151,
0.0021620120387524366,
-0.0311098862439394,
0.06711667776107788,
0.050001416355371475,
-0.01808607205748558,
0.00001835354123613797,
0.10144516825675964,
0.07419592142105103,
-0.16415110230445862,
-0.04081225395202637,
0.06218694895505905,
0.020990675315260887,
0.0011424677213653922,
-0.04319165647029877,
-0.019504418596625328,
0.02680780366063118,
-0.058594994246959686,
-0.07270701974630356,
0.04322995990514755,
-0.08710861951112747,
0.007123442832380533,
0.029560452327132225,
-0.07057058066129684,
-0.00034135906025767326,
-0.06993121653795242,
0.05206747353076935,
-0.06318497657775879,
0.009244389832019806,
0.014905628748238087,
-0.07879524677991867,
0.018491441383957863,
-0.04949076846241951,
0.02276603691279888,
0.043565329164266586,
0.01581466756761074,
-0.08776994794607162,
-0.002275970298796892,
-0.025776775553822517,
-0.013820585794746876,
0.010686484165489674,
-0.049474939703941345,
-0.04090534895658493,
-0.01185176707804203,
0.04813725873827934,
-0.004286524374037981,
-0.007735266815871,
0.0989885926246643,
0.036180365830659866,
-0.07980557531118393,
-0.005892076529562473,
-0.0956251248717308,
-0.046045415103435516,
-0.0425722599029541,
-0.03358481824398041,
-0.02975793182849884,
0.015160249546170235,
0.01283649168908596,
0.05408395081758499,
-0.04478073492646217,
0.09632764756679535,
0.011187124066054821,
0.03476094454526901,
-0.042066507041454315,
-0.012641708366572857,
0.058510921895504,
0.019983584061264992,
0.010131222195923328,
-0.008067825809121132,
0.05062377452850342,
-0.08774364739656448,
0.04885212704539299,
0.0023453342728316784,
0.07218499481678009,
0.010400429368019104,
0.08284568786621094,
0.057903487235307693,
-0.06353073567152023,
0.057168249040842056,
0.07006526738405228,
0.018539587035775185,
-0.08194421231746674,
-0.03639630228281021,
0.08383382856845856,
0.014002522453665733,
0.027293583378195763,
0.003924640826880932,
-0.04486797749996185,
-0.07105591893196106,
-0.02020525000989437,
0.07446020096540451,
0.001272131223231554,
0.030561130493879318,
-0.022093677893280983,
-0.04510672762989998,
0.05723310634493828,
-0.09160307794809341,
0.06974229216575623,
-0.003211595816537738,
-0.015013676136732101,
-0.005188782699406147,
0.059827711433172226,
0.035079874098300934,
-0.016562219709157944,
-0.01114574819803238,
0.01670653745532036,
-0.021780099719762802,
0.011011372320353985,
-0.04560868814587593,
0.03630172461271286,
0.007623653393238783,
-0.019433869048953056,
0.01644221507012844,
-0.041958827525377274,
-0.0016512904549017549,
0.001799732563085854,
-0.039084259420633316,
-0.03882605582475662,
-4.292323591334833e-34,
-0.04901964217424393,
0.002621015999466181,
-0.011666577309370041,
0.11712370067834854,
0.06297481060028076,
-0.020092716440558434,
-0.034591615200042725,
-0.06853697448968887,
-0.03254120051860809,
-0.07311560958623886,
-0.06807775795459747,
-0.013451135717332363,
-0.08992757648229599,
0.10808228701353073,
0.10636662691831589,
-0.018670545890927315,
-0.01801491342484951,
0.05358535796403885,
-0.010686610825359821,
-0.040101516991853714,
0.038766324520111084,
-0.04046553000807762,
-0.031772468239068985,
-0.07308446615934372,
-0.036636173725128174,
0.027697447687387466,
0.04585794359445572,
-0.05645746365189552,
-0.005425138399004936,
0.020818842574954033,
-0.14908654987812042,
0.001921704038977623,
0.031915854662656784,
0.036604706197977066,
0.018704604357481003,
-0.0011866374406963587,
-0.05251478776335716,
-0.010322380810976028,
0.06063982844352722,
0.0000297690221486846,
-0.02007812075316906,
0.03016877919435501,
-0.03134217858314514,
0.01224710326641798,
-0.08822917193174362,
0.03852643072605133,
0.015496988780796528,
0.017404630780220032,
-0.027445482090115547,
-0.08864905685186386,
0.05130580812692642,
0.009348487481474876,
0.06687583774328232,
-0.032369617372751236,
0.0175892636179924,
-0.028943169862031937,
-0.012977978214621544,
-0.007886504754424095,
0.0372091643512249,
0.006537369452416897,
0.04566456377506256,
0.069149911403656,
0.012145867571234703,
0.11102919280529022,
0.08890119940042496,
-0.010435046628117561,
-0.0537785105407238,
0.022567540407180786,
-0.031536251306533813,
-0.018068617209792137,
-0.040979400277137756,
0.036895330995321274,
-0.07241570204496384,
-0.05852840468287468,
0.02146783098578453,
-0.04649178311228752,
0.034313514828681946,
-0.09828081727027893,
0.00836502481251955,
0.04185573384165764,
0.00020527656306512654,
-0.0815201923251152,
-0.0060371458530426025,
-0.02760695107281208,
-0.013779226690530777,
-0.08318247646093369,
-0.03093741461634636,
-0.044035207480192184,
-0.09056871384382248,
-0.04587024077773094,
0.010018797591328621,
-0.029294395819306374,
-0.038148168474435806,
0.011240885592997074,
-0.00046878340071998537,
-1.3702446529012585e-33,
-0.041187141090631485,
0.07060103118419647,
-0.017036909237504005,
0.06947923451662064,
-0.0787283256649971,
-0.02866446040570736,
0.04491105675697327,
0.01955096796154976,
0.0532853864133358,
-0.06836908310651779,
-0.04818403348326683,
0.03515377268195152,
-0.011062403209507465,
-0.061002496629953384,
0.053233832120895386,
-0.11454694718122482,
0.005626032594591379,
0.007937324233353138,
0.0161359254270792,
-0.013015302829444408,
0.01800517365336418,
0.07075121998786926,
-0.1445537954568863,
0.05368774011731148,
-0.09022925049066544,
0.04410368949174881,
-0.1016896516084671,
0.08212647587060928,
-0.015492926351726055,
0.019936857745051384,
-0.05438624694943428,
0.08858247846364975,
0.027839845046401024,
-0.04521890729665756,
-0.05833710357546806,
0.10738833248615265,
0.0418851338326931,
0.05320725962519646,
-0.020519336685538292,
0.04873619228601456,
0.05153469368815422,
-0.05199737474322319,
-0.05856511741876602,
0.049083881080150604,
-0.003587825922295451,
-0.03790638595819473,
-0.06949156522750854,
-0.04086422547698021,
0.041330303996801376,
0.013138722628355026,
0.055697545409202576,
0.0028175462502986193,
-0.049964215606451035,
-0.019954759627580643,
-0.010441996157169342,
-0.029841264709830284,
-0.011821197345852852,
-0.0714070200920105,
0.053972650319337845,
-0.003518813755363226,
-0.010406360030174255,
-0.043526627123355865,
-0.026067500934004784,
-0.14211183786392212,
-0.08768943697214127,
-0.026869652792811394,
-0.029742609709501266,
-0.01046166755259037,
0.04440005123615265,
0.028364555910229683,
0.035042837262153625,
-0.0009707281133159995,
-0.026155704632401466,
0.014055456034839153,
-0.0021009130869060755,
-0.09960107505321503,
0.020443277433514595,
-0.004845683928579092,
-0.020805178210139275,
-0.04682552441954613,
0.07535714656114578,
0.04918806999921799,
0.0326380655169487,
0.025003617629408836,
0.0751185268163681,
0.021338921040296555,
-0.05060246214270592,
0.023073915392160416,
0.02696472406387329,
0.0369839146733284,
0.002309520496055484,
0.008714454248547554,
0.016904115676879883,
0.06508258730173111,
-0.0832885131239891,
-4.6376250395496754e-8,
-0.0327545590698719,
0.02289155125617981,
0.06252136826515198,
0.007404729258269072,
0.0023903625551611185,
-0.0010516764596104622,
-0.0006538835004903376,
0.06393734365701675,
0.0516180582344532,
-0.01356799528002739,
-0.008328505791723728,
0.00690695084631443,
-0.027786115184426308,
-0.056600622832775116,
0.008716732263565063,
0.14488981664180756,
0.001749245566315949,
0.03709017485380173,
-0.023858172819018364,
-0.05796387419104576,
-0.03214431554079056,
0.017737675458192825,
-0.014242418110370636,
0.038526155054569244,
0.046490542590618134,
-0.046201594173908234,
-0.030319053679704666,
0.04212821274995804,
0.1135893389582634,
0.013403129763901234,
0.0602928102016449,
-0.013263421133160591,
-0.07636184245347977,
0.06674893945455551,
-0.05007563531398773,
0.1028708666563034,
-0.010129370726644993,
-0.02911805734038353,
-0.020349469035863876,
-0.001297749113291502,
0.02057824656367302,
0.08467318117618561,
-0.01710871048271656,
0.022267397493124008,
-0.003981930669397116,
0.05098192021250725,
-0.038780320435762405,
-0.08097358793020248,
0.04144488275051117,
0.040944892913103104,
0.02229660376906395,
0.025279957801103592,
-0.014462956227362156,
0.12697800993919373,
-0.0013705578166991472,
0.0012230821885168552,
0.057880569249391556,
0.02067488245666027,
0.05931173637509346,
0.08646716922521591,
0.07024633884429932,
0.05423743277788162,
0.017835943028330803,
-0.023791642859578133
] |
Efficient and Accurate Explanation Estimation with Distribution Compression | https://openreview.net/forum?id=LiUfN9h0Lx | [
"Hubert Baniecki",
"Giuseppe Casalicchio",
"Bernd Bischl",
"Przemyslaw Biecek"
] | Spotlight | We discover a theoretical connection between explanation estimation and distribution compression that significantly improves the approximation of feature attributions, importance, and effects. While the exact computation of various machine learning explanations requires numerous model inferences and becomes impractical, the computational cost of approximation increases with an ever-increasing size of data and model parameters. We show that the standard i.i.d. sampling used in a broad spectrum of algorithms for post-hoc explanation leads to an approximation error worthy of improvement. To this end, we introduce Compress Then Explain (CTE), a new paradigm of sample-efficient explainability. It relies on distribution compression through kernel thinning to obtain a data sample that best approximates its marginal distribution. CTE significantly improves the accuracy and stability of explanation estimation with negligible computational overhead. It often achieves an on-par explanation approximation error 2-3x faster by using fewer samples, i.e. requiring 2-3x fewer model evaluations. CTE is a simple, yet powerful, plug-in for any explanation method that now relies on i.i.d. sampling. | explainable ai, feature attributions, feature importance, sampling, kernel thinning | null | 7,423 | 2406.18334 | [
-0.09441870450973511,
0.0035202328581362963,
0.06560882925987244,
0.06051993370056152,
0.15807923674583435,
-0.04989880695939064,
-0.010963440872728825,
0.044725410640239716,
0.06855776160955429,
0.001577621209435165,
-0.06504732370376587,
-0.06522979587316513,
-0.012424186803400517,
-0.015962738543748856,
-0.02571273408830166,
-0.03151727467775345,
0.06064986065030098,
0.10669741034507751,
-0.0981975719332695,
-0.05204159393906593,
0.07241389155387878,
-0.008069823496043682,
0.03581871837377548,
0.062255822122097015,
0.14522381126880646,
-0.000739209121093154,
0.002969387685880065,
-0.022927511483430862,
0.03813296556472778,
0.009440058842301369,
0.061222512274980545,
0.002827236894518137,
0.011633926071226597,
0.023339450359344482,
-0.07426434010267258,
0.04668654873967171,
-0.052429571747779846,
0.09836167842149734,
0.02026125229895115,
0.006022683810442686,
-0.0055668954737484455,
0.06512726843357086,
-0.0023015381302684546,
0.05257657170295715,
0.03315522521734238,
0.005079774186015129,
0.04235389083623886,
-0.052394382655620575,
0.020922034978866577,
0.05266113579273224,
-0.053572315722703934,
0.0677468478679657,
-0.03947412967681885,
-0.03262648731470108,
0.01740092784166336,
-0.08326640725135803,
0.0007127253920771182,
0.012300088070333004,
-0.012313900515437126,
0.04000946879386902,
-0.034966472536325455,
-0.11110444366931915,
-0.016515174880623817,
0.07994894683361053,
0.005640701856464148,
-0.03469332680106163,
0.07104334980249405,
0.06586681306362152,
-0.044288188219070435,
0.012174803763628006,
-0.031436383724212646,
0.09567799419164658,
-0.10845036804676056,
0.04983523115515709,
-0.0315617211163044,
0.12679938971996307,
0.034096527844667435,
0.10634686052799225,
-0.013095749542117119,
-0.0027734912000596523,
-0.010888333432376385,
0.0059287576004862785,
0.03453211486339569,
-0.046866681426763535,
0.02880556508898735,
-0.05564141646027565,
0.015925614163279533,
0.01315914187580347,
-0.023018844425678253,
0.054850075393915176,
-0.00784294307231903,
-0.014774816110730171,
-0.044517748057842255,
-0.01874207891523838,
0.007638254202902317,
0.06836956739425659,
-0.028634898364543915,
-0.10051634907722473,
-0.07465030997991562,
0.11950161308050156,
-0.05693139508366585,
0.10488902777433395,
0.03294594958424568,
-0.08771119266748428,
-0.02078128233551979,
-0.1332378089427948,
0.04608508199453354,
-0.020851854234933853,
0.03051770105957985,
-0.019345639273524284,
0.015185120515525341,
0.01903611049056053,
-0.025354454293847084,
-0.030469605699181557,
0.048475973308086395,
0.003662391100078821,
0.025298191234469414,
-0.01614725962281227,
0.04128464683890343,
0.0007970027509145439,
-0.02028997242450714,
-0.02689843811094761,
-0.08636816591024399,
0.02963569201529026,
0.01425429992377758,
0.0045853727497160435,
-0.0754789412021637,
1.3325222126639058e-33,
0.045235734432935715,
0.06378795951604843,
0.04483494535088539,
0.043662432581186295,
0.07400023192167282,
-0.0030606871005147696,
-0.02400747686624527,
0.015539305284619331,
0.03422774374485016,
0.1016388013958931,
-0.03032354637980461,
-0.004569568671286106,
-0.01548114325851202,
0.034279756247997284,
-0.03242389112710953,
0.035894423723220825,
-0.09492101520299911,
0.11924710124731064,
-0.0754668340086937,
-0.03874405473470688,
-0.022752031683921814,
-0.09270661324262619,
-0.021671516820788383,
-0.044692013412714005,
-0.007753161247819662,
-0.055310316383838654,
0.005073636770248413,
-0.04371711611747742,
-0.027507757768034935,
-0.019740339368581772,
-0.0993679091334343,
0.005050409585237503,
0.033742379397153854,
-0.07300888001918793,
0.03299036994576454,
-0.09770108014345169,
0.021958542987704277,
-0.05768115818500519,
-0.0041994149796664715,
-0.009576212614774704,
0.006202719174325466,
-0.015035254880785942,
0.004506805911660194,
-0.016528591513633728,
-0.05908626690506935,
-0.037212274968624115,
-0.025833433493971825,
-0.06704524159431458,
-0.04397623986005783,
0.0049186223186552525,
0.08387931436300278,
0.029762454330921173,
-0.004650598857551813,
-0.010270061902701855,
0.009349900297820568,
0.012373468838632107,
0.021523993462324142,
-0.010897442698478699,
0.032564371824264526,
0.08825214207172394,
-0.03505831956863403,
0.04533755034208298,
-0.03848186507821083,
0.023407701402902603,
0.046449214220047,
0.04055910184979439,
-0.018430138006806374,
0.08502689749002457,
0.06819111108779907,
0.035574693232774734,
-0.016969377174973488,
0.029993006959557533,
-0.03116948902606964,
-0.0819917842745781,
0.04851556941866875,
-0.006291755009442568,
0.006145638879388571,
-0.053851932287216187,
0.03249715268611908,
0.002331347204744816,
-0.03138125315308571,
-0.044826049357652664,
0.02130098268389702,
-0.006083495914936066,
-0.06723278015851974,
-0.04086102172732353,
0.04423029348254204,
-0.006658574100583792,
-0.03577878326177597,
-0.09020125865936279,
-0.03846289590001106,
-0.0316779725253582,
0.01256878674030304,
-0.06868597865104675,
-0.01683611422777176,
-1.1920155565811101e-33,
-0.08881080150604248,
0.03774172067642212,
-0.007226631045341492,
-0.02144552767276764,
-0.026836056262254715,
0.0023345141671597958,
-0.06635349988937378,
0.012805270031094551,
0.002419372322037816,
-0.061226144433021545,
-0.07445845752954483,
-0.026607463136315346,
0.046655405312776566,
0.020155426114797592,
-0.01574934646487236,
0.04220551624894142,
-0.00872083567082882,
0.00958938430994749,
0.011135033331811428,
0.01816943660378456,
-0.013917479664087296,
0.09819147735834122,
-0.054323840886354446,
0.007712846621870995,
0.022238366305828094,
0.02117942087352276,
-0.026118047535419464,
0.12262085825204849,
0.035258613526821136,
-0.03553111106157303,
-0.11014421284198761,
-0.023634828627109528,
-0.05309561267495155,
-0.05048808082938194,
-0.0803181529045105,
0.02428109012544155,
0.011106123216450214,
0.09342212229967117,
-0.025392696261405945,
0.06643302738666534,
0.08588199317455292,
-0.0281971525400877,
-0.07084040343761444,
-0.054460469633340836,
-0.032619379460811615,
-0.06513309478759766,
0.006300248671323061,
-0.02396918088197708,
0.054115697741508484,
0.0632496029138565,
-0.010358314029872417,
0.08469908684492111,
-0.08034208416938782,
0.10045202076435089,
-0.029059724882245064,
-0.07827742397785187,
-0.10483638942241669,
-0.020333051681518555,
0.04884635657072067,
-0.04362647235393524,
-0.10549232363700867,
0.05826282873749733,
0.0004820544272661209,
0.03083905763924122,
0.0004764367185998708,
-0.02794530615210533,
-0.0158242154866457,
0.018014857545495033,
0.03317795693874359,
0.039060354232788086,
0.025056099519133568,
-0.051038749516010284,
-0.010046721436083317,
0.03585078939795494,
0.03930458799004555,
0.0136066609993577,
0.048649758100509644,
-0.03334800526499748,
-0.03953446447849274,
-0.02309725061058998,
0.09617706388235092,
0.011244677007198334,
0.06589232385158539,
-0.00934856291860342,
0.029633695259690285,
0.016959646716713905,
0.014889216050505638,
0.03138532117009163,
0.02031274139881134,
0.06165789067745209,
-0.061304494738578796,
0.04295867308974266,
-0.004516877233982086,
0.15803499519824982,
-0.018297763541340828,
-4.2392944266111954e-8,
-0.00015403433644678444,
0.06969007104635239,
0.0016060678754001856,
-0.021994860842823982,
0.09586463123559952,
-0.018714461475610733,
-0.009540731087327003,
0.06769437342882156,
-0.08993225544691086,
-0.08398380130529404,
0.06300874799489975,
-0.026551969349384308,
-0.0642358735203743,
0.010705355554819107,
0.0109989820048213,
0.09528777748346329,
-0.06211501732468605,
0.026029672473669052,
-0.018323462456464767,
-0.0017589681083336473,
-0.025537624955177307,
0.008340551517903805,
0.01957423985004425,
-0.012949799187481403,
0.07769627124071121,
-0.05740990862250328,
0.029792454093694687,
0.09656742215156555,
0.04755691811442375,
0.029242662712931633,
-0.06011440232396126,
0.015820305794477463,
-0.04749911651015282,
-0.02556917630136013,
0.03099021501839161,
0.0686669796705246,
0.011616017669439316,
-0.05344843491911888,
0.024872805923223495,
-0.011427556164562702,
-0.043343301862478256,
0.041157957166433334,
-0.016299499198794365,
-0.009668015874922276,
0.041242294013500214,
0.014593432657420635,
0.025959936901926994,
-0.0906912088394165,
-0.019989468157291412,
-0.0418158695101738,
-0.006113697774708271,
0.038769446313381195,
-0.039782993495464325,
0.06346886605024338,
-0.028509994968771935,
0.00475160451605916,
-0.07225842773914337,
0.060314495116472244,
-0.026935238391160965,
0.04004298150539398,
0.044447772204875946,
0.09127171337604523,
0.020731741562485695,
0.009413345716893673
] |
Interleaved Scene Graphs for Interleaved Text-and-Image Generation Assessment | https://openreview.net/forum?id=rDLgnYLM5b | [
"Dongping Chen",
"Ruoxi Chen",
"Shu Pu",
"Zhaoyi Liu",
"Yanru Wu",
"Caixi Chen",
"Benlin Liu",
"Yue Huang",
"Yao Wan",
"Pan Zhou",
"Ranjay Krishna"
] | Spotlight | Many real-world user queries (e.g. *"How do to make egg fried rice?"*) could benefit from systems capable of generating responses with both textual steps with accompanying images, similar to a cookbook.
Models designed to generate interleaved text and images face challenges in ensuring consistency within and across these modalities.
To address these challenges, we present ISG, a comprehensive evaluation framework for interleaved text-and-image generation. ISG leverages a scene graph structure to capture relationships between text and image blocks, evaluating responses on four levels of granularity: holistic, structural, block-level, and image-specific. This multi-tiered evaluation allows for a nuanced assessment of consistency, coherence, and accuracy, and provides interpretable question-answer feedback.
In conjunction with ISG, we introduce a benchmark, ISG-Bench, encompassing 1,150 samples across 8 categories and 21 subcategories. This benchmark dataset includes complex language-vision dependencies and golden answers to evaluate models effectively on vision-centric tasks such as style transfer, a challenging area for current models.
Using ISG-Bench, we demonstrate that recent unified vision-language models perform poorly on generating interleaved content. While compositional approaches that combine separate language and image models show a 111% improvement over unified models at the holistic level, their performance remains suboptimal at both block and image levels.
To facilitate future work, we develop ISG-Agent, a baseline agent employing a *"plan-execute-refine"* pipeline to invoke tools, achieving a 122% performance improvement. | Interleaved Text-and-Image Generation, Generative Models, Multimodal Large Language Model, Scene Graphs, Automatic Evaluation, Benchmark | This paper introduces a fine-grained automatic evaluation framework and a new benchmark for interleaved text-and-image generation, offering valuable insights for future research in accurate interleaved generation. | 7,414 | 2411.17188 | [
-0.013791707344353199,
-0.03286363556981087,
0.05012102425098419,
0.053443510085344315,
0.11825156211853027,
-0.004916995298117399,
-0.05494173243641853,
-0.025759097188711166,
0.026150738820433617,
-0.06172255426645279,
-0.042815688997507095,
-0.12683194875717163,
0.05731774866580963,
0.0644402727484703,
0.00643492816016078,
0.013313602656126022,
0.09266915917396545,
-0.00290189147926867,
-0.09981553256511688,
-0.047108542174100876,
0.0400240384042263,
0.03759082034230232,
0.09770581126213074,
-0.07118992507457733,
0.06483311951160431,
0.020479796454310417,
-0.01849011890590191,
-0.011861876584589481,
0.03433234617114067,
-0.036175355315208435,
0.01420578919351101,
0.022257333621382713,
-0.033307287842035294,
0.12405871599912643,
-0.035774726420640945,
0.11831043660640717,
0.028169669210910797,
0.0059749106876552105,
0.04061766713857651,
-0.02414911985397339,
-0.03051532246172428,
0.033758699893951416,
0.03545398637652397,
0.015327262692153454,
0.12038764357566833,
0.004714669659733772,
-0.10991424322128296,
-0.04579445347189903,
-0.04400760680437088,
-0.0010790596716105938,
-0.10808113217353821,
0.005550865549594164,
-0.05507936328649521,
0.00026705616619437933,
0.008107371628284454,
0.08861415833234787,
0.07127908617258072,
-0.026966264471411705,
0.04798683151602745,
0.017661312595009804,
-0.05277632176876068,
-0.10039582848548889,
0.00708391796797514,
0.020843498408794403,
0.012396848760545254,
-0.018391866236925125,
-0.006908447947353125,
-0.01314657274633646,
-0.04557710886001587,
-0.047352924942970276,
-0.06594271957874298,
0.017687072977423668,
0.02694871835410595,
0.0763249322772026,
-0.0020411289297044277,
0.08769267052412033,
0.019229434430599213,
-0.04461728036403656,
0.028386814519762993,
-0.09003031253814697,
0.02255251817405224,
-0.012897858396172523,
0.08293377608060837,
-0.008129755966365337,
0.06083446741104126,
-0.013479669578373432,
0.011363587342202663,
0.06993880122900009,
-0.08346018195152283,
0.010439167730510235,
-0.07465267926454544,
-0.04934448003768921,
0.04834454134106636,
-0.014789457432925701,
0.04141820967197418,
0.03769974410533905,
0.021003244444727898,
-0.15326757729053497,
-0.017353959381580353,
0.04094111546874046,
0.03583940491080284,
0.03506748005747795,
0.01248208899050951,
-0.09973607957363129,
0.02197054959833622,
0.006449840497225523,
0.08418800681829453,
0.029858658090233803,
0.00042952809599228203,
-0.07030313462018967,
0.020230291411280632,
0.005984912626445293,
0.005676307249814272,
-0.0061773802153766155,
0.02076938934624195,
-0.06747931241989136,
0.06255290657281876,
-0.0364854671061039,
0.036439232528209686,
-0.0005808601854369044,
0.028355805203318596,
0.018036874011158943,
-0.0016696156235411763,
0.00898545328527689,
-0.0473080649971962,
-0.0543212890625,
0.0230864267796278,
1.5802927282592353e-33,
0.03239237144589424,
0.03613701090216637,
0.09225115925073624,
0.07461170107126236,
0.0770527794957161,
-0.011052962392568588,
-0.035852886736392975,
0.05025387927889824,
-0.0829470083117485,
0.01088075339794159,
-0.027286184951663017,
-0.039113130420446396,
-0.06910941749811172,
0.12967343628406525,
0.07505404949188232,
0.002439766190946102,
-0.06135578081011772,
0.010564463213086128,
-0.029092513024806976,
0.05234543979167938,
-0.05851106345653534,
-0.006775970105081797,
0.039789143949747086,
-0.00031919433968141675,
0.05371164530515671,
0.001339394017122686,
0.07323697954416275,
-0.057479050010442734,
-0.04354773089289665,
-0.03909461572766304,
-0.10570823401212692,
-0.03134802728891373,
0.05345199629664421,
0.014693225733935833,
-0.040303442627191544,
-0.043664898723363876,
-0.011375409550964832,
-0.042851414531469345,
0.008521951735019684,
-0.006717517040669918,
-0.04221523180603981,
0.10369323939085007,
0.022582339122891426,
0.003665467258542776,
-0.10091184079647064,
0.05433550104498863,
0.006133798975497484,
0.03760196268558502,
-0.0410354845225811,
0.0339546762406826,
0.028639158234000206,
-0.005804418586194515,
0.028793824836611748,
-0.019450802356004715,
-0.010736797004938126,
0.008704716339707375,
0.06363106518983841,
-0.011132337152957916,
0.06627310067415237,
0.015373778529465199,
0.10230328887701035,
0.06967782974243164,
-0.052765797823667526,
0.06902586668729782,
0.035508010536432266,
0.05247223377227783,
-0.012280669063329697,
0.00594696868211031,
-0.010454975068569183,
0.017186759039759636,
-0.03609636425971985,
-0.009391950443387032,
-0.05185813829302788,
-0.05308246612548828,
0.009092454798519611,
0.0015738965012133121,
-0.048520058393478394,
-0.04570459574460983,
-0.01680738478899002,
0.036870718002319336,
-0.013844122178852558,
-0.0010138023644685745,
0.018959635868668556,
-0.08000145852565765,
-0.05740657448768616,
-0.031241655349731445,
0.05283732712268829,
-0.09918511658906937,
0.018835093826055527,
0.004550587385892868,
-0.025086333975195885,
-0.02360270731151104,
-0.014967030845582485,
-0.03255463391542435,
0.05928054824471474,
1.1140530748558557e-34,
0.014456956647336483,
0.04102357476949692,
-0.08612654358148575,
0.07684123516082764,
0.014246070757508278,
-0.04873370751738548,
0.028295420110225677,
0.028566233813762665,
0.010868902318179607,
-0.07289665937423706,
-0.0002338908234378323,
0.02488795667886734,
-0.015408097766339779,
-0.05291426181793213,
-0.01501538697630167,
-0.0223495252430439,
0.017111940309405327,
-0.024303006008267403,
-0.05376790836453438,
0.05871046707034111,
0.03560702130198479,
0.09512752294540405,
-0.057418785989284515,
0.01845455914735794,
-0.0557188019156456,
0.12323960661888123,
-0.04648154228925705,
-0.03611327335238457,
-0.027890773490071297,
-0.026595154777169228,
0.007443401962518692,
-0.07246246933937073,
0.05312183126807213,
-0.04618292674422264,
-0.04585985094308853,
0.022130198776721954,
0.05816986411809921,
-0.07877875119447708,
-0.07121112942695618,
0.06937309354543686,
0.05151977762579918,
-0.0005685337237082422,
-0.07523489743471146,
0.049162134528160095,
-0.08341538906097412,
-0.01086061354726553,
-0.05506684631109238,
0.008031758479773998,
-0.0261958297342062,
0.10439574718475342,
-0.09682754427194595,
-0.0008718863246031106,
-0.07745141535997391,
0.01930222101509571,
-0.018628019839525223,
-0.10083353519439697,
-0.06784463673830032,
-0.00761999748647213,
0.022497640922665596,
-0.029338456690311432,
-0.05614139512181282,
-0.016729846596717834,
-0.013566548004746437,
-0.09529396891593933,
0.06021970137953758,
-0.06183571368455887,
-0.06729935109615326,
-0.07913952320814133,
-0.019503727555274963,
0.0468599759042263,
-0.0273763258010149,
-0.04941317066550255,
0.013292763382196426,
0.05840864032506943,
0.05929955095052719,
0.027972429990768433,
0.019172903150320053,
0.0007594992639496922,
0.017557285726070404,
-0.09787734597921371,
0.023168055340647697,
-0.030942542478442192,
-0.02050098404288292,
0.16233910620212555,
0.08437491208314896,
0.07932901382446289,
-0.004515680950134993,
0.037064775824546814,
0.05343659967184067,
0.016705142334103584,
-0.01959266886115074,
0.020889461040496826,
-0.011171285063028336,
0.06958232820034027,
-0.006617917213588953,
-4.0491123343144864e-8,
-0.07365617156028748,
0.004817802459001541,
-0.026246534660458565,
0.004557427018880844,
-0.06156462803483009,
-0.04927948862314224,
0.03816233202815056,
0.03357831388711929,
0.012544653378427029,
-0.000012805903679691255,
0.03159193694591522,
0.011850615032017231,
-0.015456128865480423,
-0.043802931904792786,
-0.05365389958024025,
0.02523738704621792,
0.0377701111137867,
0.08214525878429413,
0.023116664960980415,
-0.018915947526693344,
0.0330599769949913,
0.012475922703742981,
-0.0215463824570179,
0.016242623329162598,
0.029575377702713013,
-0.01812925934791565,
-0.03577382117509842,
0.04937487468123436,
0.08076421916484833,
-0.04948596656322479,
0.04773279279470444,
-0.01310790702700615,
-0.006124658044427633,
0.02803526446223259,
0.027018366381525993,
0.055001795291900635,
-0.034783735871315,
0.0012131318217143416,
0.017200222238898277,
-0.04063628241419792,
0.0381721630692482,
0.08458243310451508,
-0.046064771711826324,
0.033906884491443634,
0.03105887398123741,
0.011027560569345951,
0.027838123962283134,
-0.05072132870554924,
0.0188596211373806,
-0.027471201494336128,
-0.10167978703975677,
-0.09265851974487305,
-0.00527846859768033,
0.04016382247209549,
0.03961930051445961,
-0.05390956997871399,
0.03766225650906563,
0.03689192607998848,
0.13014574348926544,
0.12459080666303635,
0.06356444954872131,
0.004103117156773806,
-0.02462884970009327,
-0.03803025186061859
] |
ODE-based Smoothing Neural Network for Reinforcement Learning Tasks | https://openreview.net/forum?id=S5Yo6w3n3f | [
"Yinuo Wang",
"Wenxuan Wang",
"Xujie Song",
"Tong Liu",
"Yuming Yin",
"Liangfa Chen",
"Likun Wang",
"Jingliang Duan",
"Shengbo Eben Li"
] | Spotlight | The smoothness of control actions is a significant challenge faced by deep reinforcement learning (RL) techniques in solving optimal control problems. Existing RL-trained policies tend to produce non-smooth actions due to high-frequency input noise and unconstrained Lipschitz constants in neural networks. This article presents a Smooth ODE (SmODE) network capable of simultaneously addressing both causes of unsmooth control actions, thereby enhancing policy performance and robustness under noise condition. We first design a smooth ODE neuron with first-order low-pass filtering expression, which can dynamically filter out high frequency noises of hidden state by a learnable state-based system time constant. Additionally, we construct a state-based mapping function, $g$, and theoretically demonstrate its capacity to control the ODE neuron's Lipschitz constant. Then, based on the above neuronal structure design, we further advanced the SmODE network serving as RL policy approximators. This network is compatible with most existing RL algorithms, offering improved adaptability compared to prior approaches. Various experiments show that our SmODE network demonstrates superior anti-interference capabilities and smoother action outputs than the multi-layer perception and smooth network architectures like LipsNet. | Reinforcement Learning, Smooth Control, Low-pass Filter, Neural ODE | This paper proposes a neural unit structure with smooth properties, and based on it, proposes a smoothing policy neural network. This work is the first time that Neural ODE method is used for action smoothing in deep reinforcement learning. | 7,391 | null | [
-0.15257707238197327,
-0.07234440743923187,
0.07876811921596527,
0.03655253350734711,
-0.05866679176688194,
0.014015884138643742,
-0.01873665489256382,
-0.10336101055145264,
0.004831337369978428,
0.0215029027312994,
-0.0024567199870944023,
-0.027885988354682922,
-0.0040042526088654995,
0.020228073000907898,
-0.1289975941181183,
0.0672503113746643,
0.055978115648031235,
0.09426611661911011,
-0.05755285546183586,
-0.04043935239315033,
0.05229872837662697,
0.019433993846178055,
0.07503767311573029,
0.023092061281204224,
-0.01822390966117382,
0.07159176468849182,
-0.046664200723171234,
-0.024111032485961914,
-0.009432431310415268,
-0.017609449103474617,
0.07341772317886353,
-0.148096963763237,
0.02769039198756218,
-0.04940343648195267,
-0.10790471732616425,
0.04149096459150314,
-0.09225719422101974,
0.013356387615203857,
-0.03618985414505005,
-0.08109807223081589,
0.00042614853009581566,
0.06731776893138885,
0.02271660976111889,
0.015145001001656055,
0.07247114181518555,
0.01195304375141859,
-0.003943488467484713,
-0.07401265949010849,
0.009232415817677975,
0.01399500947445631,
0.00011142191215185449,
-0.014467068947851658,
0.06753687560558319,
0.0248857531696558,
0.09028312563896179,
0.02885335311293602,
-0.05765010789036751,
0.013093673624098301,
-0.0692281499505043,
-0.01943930611014366,
-0.0916968360543251,
-0.09689651429653168,
0.03544896841049194,
-0.03083427995443344,
-0.10830690711736679,
0.019328417256474495,
-0.013661512173712254,
0.06697611510753632,
0.0253073051571846,
0.04204358905553818,
-0.012794439680874348,
0.06280844658613205,
0.014195065945386887,
-0.030132409185171127,
0.014353379607200623,
0.005664176307618618,
0.0281927902251482,
0.00050888437544927,
0.040902040898799896,
0.05824140086770058,
0.055506326258182526,
-0.09613372385501862,
-0.08297152072191238,
0.038726773113012314,
0.06642866134643555,
-0.020933061838150024,
-0.05993462726473808,
0.0434347540140152,
0.04895101860165596,
0.03178698569536209,
0.006857897154986858,
0.048767104744911194,
0.02263757213950157,
-0.06835731863975525,
0.010154664516448975,
0.0009549710084684193,
0.024146825075149536,
-0.01974823698401451,
-0.08630706369876862,
0.059116072952747345,
0.01379476860165596,
0.05682564154267311,
-0.03168017417192459,
0.01900401897728443,
0.01642397604882717,
-0.027324460446834564,
0.054359856992959976,
0.026013946160674095,
0.020504189655184746,
-0.05025511607527733,
-0.017517466098070145,
-0.007737179286777973,
-0.009291604161262512,
0.06375071406364441,
0.012879564426839352,
0.015303945168852806,
0.056991904973983765,
-0.0917755737900734,
-0.010664395056664944,
-0.018344657495617867,
-0.08235283195972443,
-0.038904689252376556,
-0.06751681119203568,
0.023092377930879593,
0.018941592425107956,
0.003118101740255952,
-0.09043749421834946,
2.1633278243625125e-33,
-0.011445607990026474,
-0.0005994437960907817,
-0.0027900447603315115,
-0.07673632353544235,
0.02895953878760338,
-0.01278875395655632,
0.02861458994448185,
0.04513023421168327,
-0.03039427287876606,
0.02615855261683464,
-0.07292263954877853,
-0.050594210624694824,
-0.09769204258918762,
0.07226225733757019,
-0.007368014194071293,
-0.03974013403058052,
-0.05357692763209343,
-0.018892643973231316,
0.01896735280752182,
-0.07412884384393692,
0.018229058012366295,
-0.039110131561756134,
-0.0003371273342054337,
-0.02701191045343876,
0.054442040622234344,
-0.014375283382833004,
0.03588905185461044,
0.06255610287189484,
0.06241544336080551,
-0.013399522751569748,
0.007671970874071121,
0.04455839470028877,
0.0287918858230114,
0.012639768421649933,
0.0331982783973217,
0.016435831785202026,
-0.013380523771047592,
0.039354074746370316,
0.03453519567847252,
0.01038142666220665,
0.025482213124632835,
-0.011440778151154518,
-0.011633222922682762,
0.05098334327340126,
-0.04577016830444336,
-0.057717215269804,
0.03918268159031868,
0.00031523359939455986,
-0.046666089445352554,
-0.011418680660426617,
-0.008824756368994713,
0.053928811103105545,
-0.0040857745334506035,
-0.1080574169754982,
0.0061893342062830925,
0.0038718010764569044,
0.09251417219638824,
0.10400613397359848,
-0.04678511247038841,
0.07220543920993805,
-0.032677505165338516,
-0.006041505839675665,
0.02330360934138298,
0.061336059123277664,
0.02505406178534031,
0.12574046850204468,
-0.0400717668235302,
0.13693562150001526,
0.03578580543398857,
-0.09840769320726395,
0.08256305754184723,
0.04816332086920738,
-0.017842981964349747,
-0.03828723356127739,
0.06511423736810684,
0.005172814708203077,
0.02787131443619728,
-0.08775298297405243,
-0.03824152424931526,
0.06286274641752243,
0.003373662242665887,
-0.02897159941494465,
-0.08048646152019501,
0.05565304309129715,
0.011980375275015831,
-0.047786347568035126,
0.025853656232357025,
0.029397549107670784,
0.04530782997608185,
-0.012587890960276127,
-0.09093708544969559,
-0.05518978834152222,
0.011043641716241837,
0.02145671658217907,
0.00882041361182928,
-2.3574888180626376e-33,
0.05373317375779152,
-0.009305009618401527,
-0.008510848507285118,
0.05994383245706558,
0.006164512597024441,
0.0618758462369442,
-0.04732898622751236,
0.010557711124420166,
0.08920110762119293,
-0.005266934633255005,
-0.03289986401796341,
0.06491222232580185,
0.06295377016067505,
-0.041816215962171555,
0.03727104142308235,
-0.04822821170091629,
-0.04772030562162399,
-0.008084082044661045,
0.048436690121889114,
0.036150187253952026,
-0.05680275708436966,
0.04714842140674591,
-0.047916799783706665,
0.0798770934343338,
0.0036956211552023888,
0.022404702380299568,
-0.06211480498313904,
0.06027388945221901,
0.013146607205271721,
0.05713939294219017,
-0.054916977882385254,
-0.04054168239235878,
-0.012454883195459843,
-0.05090418457984924,
-0.008769741281867027,
0.04033326357603073,
0.06197350099682808,
-0.019382819533348083,
-0.04379025474190712,
0.04604611545801163,
0.024495424702763557,
-0.04758479446172714,
0.0030134215485304594,
0.054894354194402695,
0.04735410213470459,
0.010442515835165977,
-0.04169866070151329,
-0.03278670459985733,
-0.0842316672205925,
0.03078063763678074,
-0.0474235825240612,
0.04316733404994011,
-0.06090378016233444,
-0.025898553431034088,
0.024513715878129005,
0.012504294514656067,
-0.034286413341760635,
0.06305273622274399,
0.08078044652938843,
-0.02516498975455761,
0.007053803652524948,
0.005902346223592758,
-0.12181603908538818,
0.029438650235533714,
-0.04516720771789551,
0.053747501224279404,
-0.007294455077499151,
0.06321869045495987,
0.05722460150718689,
-0.04822588711977005,
-0.012930977158248425,
-0.013364051468670368,
0.02977994829416275,
0.011450152844190598,
0.020445315167307854,
-0.05975327640771866,
-0.011924488469958305,
-0.05109681189060211,
0.02398538403213024,
0.059657022356987,
0.05823151394724846,
0.015487011522054672,
0.05289190262556076,
0.028997283428907394,
0.04535442963242531,
0.023571113124489784,
-0.045091357082128525,
0.013969970867037773,
0.004517237655818462,
0.01170323695987463,
-0.0017273349221795797,
0.0390852615237236,
-0.0785544365644455,
0.02625121921300888,
-0.08659344911575317,
-4.5828691952465306e-8,
-0.05895842984318733,
0.03422606736421585,
0.055186692625284195,
-0.03504043072462082,
0.07220245897769928,
-0.06054762005805969,
0.026163365691900253,
-0.059398896992206573,
0.004803221672773361,
-0.01518731564283371,
0.10321798175573349,
0.030365321785211563,
-0.011675420217216015,
-0.14774033427238464,
-0.07690975815057755,
0.008043545298278332,
-0.004446920473128557,
-0.012068500742316246,
0.052965160459280014,
0.014557774178683758,
0.008169930428266525,
0.036668479442596436,
-0.032113030552864075,
-0.028795761987566948,
0.006901922170072794,
-0.0598844550549984,
-0.032673969864845276,
0.028315771371126175,
0.029563575983047485,
-0.05462676286697388,
0.04108240827918053,
-0.02921837382018566,
0.08534211665391922,
-0.027827858924865723,
-0.022452548146247864,
0.028670836240053177,
0.05867907032370567,
-0.07053599506616592,
0.052549876272678375,
-0.0016905806260183454,
-0.028871899470686913,
0.05880418419837952,
0.03279559686779976,
-0.04611777141690254,
-0.1302804946899414,
0.01641756109893322,
-0.027874276041984558,
-0.16046853363513947,
0.058906883001327515,
0.012171324342489243,
0.08413150906562805,
0.054981522262096405,
-0.02483864687383175,
0.05085939168930054,
0.08745956420898438,
-0.07138630002737045,
0.04002293944358826,
-0.09533150494098663,
-0.042446717619895935,
0.08919806778430939,
-0.03951987251639366,
0.03440805897116661,
0.009281660430133343,
0.04889339581131935
] |
Subsets and Splits