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
|
|---|---|---|---|---|---|---|---|---|---|
Let Your Features Tell The Differences: Understanding Graph Convolution By Feature Splitting | https://openreview.net/forum?id=I9omfcWfMp | [
"Yilun Zheng",
"Xiang Li",
"Sitao Luan",
"Xiaojiang Peng",
"Lihui Chen"
] | Poster | Graph Neural Networks (GNNs) have demonstrated strong capabilities in processing structured data. While traditional GNNs typically treat each feature dimension equally important during graph convolution, we raise an important question: **Is the graph convolution operation equally beneficial for each feature?** If not, the convolution operation on certain feature dimensions can possibly lead to harmful effects, even worse than convolution-free models. Therefore, it is required to distinguish convolution-favored and convolution-disfavored features. Traditional feature selection methods mainly focus on identifying informative features or reducing redundancy, but they are not suitable for structured data as they overlook graph structures. In graph community, some studies have investigated the performance of GNN with respect to node features using feature homophily metrics, which assess feature consistency across graph topology. Unfortunately, these metrics do not effectively align with GNN performance and cannot be reliably used for feature selection in GNNs. To address these limitations, we introduce a novel metric, Topological Feature Informativeness (TFI), to distinguish GNN-favored and GNN-disfavored features, where its effectiveness is validated through both theoretical analysis and empirical observations. Based on TFI, we propose a simple yet effective Graph Feature Selection (GFS) method, which processes GNN-favored and GNN-disfavored features with GNNs and non-GNN models separately. Compared to original GNNs, GFS significantly improves the extraction of useful topological information from each feature with comparable computational costs. Extensive experiments show that after applying GFS to $\textbf{8}$ baseline and state-of-the-art (SOTA) GNN architectures across $\textbf{10}$ datasets, $\textbf{90\%}$ of the GFS-augmented cases show significant performance boosts. Furthermore, our proposed TFI metric outperforms other feature selection methods for GFS. These results verify the effectiveness of both GFS and TFI. Additionally, we demonstrate that GFS's improvements are robust to hyperparameter tuning, highlighting its potential as a universally valid method for enhancing various GNN architectures. | Graph Neural Networks, Graph Homophily, Topological Feature Selection | We propose a new metric to identify GNN-favored and GNN-disfavored features and use topological feature selection to fuse these features into GNNs, which significantly improves GNNs performance without hyper-parameter tuning. | 1,933 | null | [
-0.030820710584521294,
-0.0724518820643425,
0.09225340932607651,
0.03065861575305462,
0.09321235865354538,
-0.022831248119473457,
-0.07363419979810715,
-0.047178562730550766,
-0.031386952847242355,
-0.048343680799007416,
-0.0015699105570092797,
-0.07235653698444366,
-0.016453485935926437,
0.10625846683979034,
0.027771711349487305,
-0.013083244673907757,
0.08695250749588013,
0.005958854220807552,
-0.027012845501303673,
-0.024680903181433678,
0.011335406452417374,
0.005306510254740715,
0.00036827163421548903,
0.01485510729253292,
0.08607003092765808,
0.04913216084241867,
0.0025763600133359432,
-0.005702931433916092,
0.04309747740626335,
0.00020469124137889594,
0.04177173599600792,
0.033523086458444595,
-0.06234438344836235,
-0.018061120063066483,
-0.054203666746616364,
0.040212541818618774,
0.012754122726619244,
0.02896069921553135,
0.024364285171031952,
-0.06137103959918022,
0.01394037064164877,
-0.021179411560297012,
0.020611003041267395,
0.06415215134620667,
0.07256877422332764,
0.11222046613693237,
-0.08827526122331619,
0.012414546683430672,
-0.011853024363517761,
-0.00005336381582310423,
-0.05630797892808914,
-0.014924590475857258,
-0.06607512384653091,
0.038548145443201065,
0.04492591693997383,
0.005331163760274649,
0.0034592801239341497,
-0.005581296980381012,
0.026926279067993164,
0.00857777614146471,
0.03152806684374809,
-0.03549329191446304,
-0.045921433717012405,
0.007631069980561733,
0.043593425303697586,
0.021738562732934952,
0.005089287646114826,
-0.03611280769109726,
0.027047740295529366,
0.043048612773418427,
0.06361314654350281,
0.10157989710569382,
-0.07409877330064774,
0.06249506026506424,
0.018129020929336548,
0.06124207377433777,
0.07588117569684982,
0.02180984802544117,
-0.026632385328412056,
-0.08303160965442657,
0.045711155980825424,
0.06713201850652695,
0.0110009228810668,
0.028501277789473534,
0.03549118712544441,
-0.035947300493717194,
-0.061095450073480606,
-0.054316699504852295,
-0.11892087757587433,
0.009151434525847435,
-0.07021818310022354,
0.03710522875189781,
0.004364740569144487,
-0.14114795625209808,
0.02765277586877346,
0.03504396975040436,
-0.03942618891596794,
-0.0602671355009079,
-0.028141304850578308,
0.054140061140060425,
-0.03175631910562515,
-0.07932157814502716,
0.055066872388124466,
-0.05597567930817604,
0.072502501308918,
0.02521275170147419,
-0.002262443071231246,
0.05579430237412453,
0.05629434436559677,
-0.04947999492287636,
0.011789851821959019,
0.06830000877380371,
0.004154742695391178,
0.03096414916217327,
0.011048303917050362,
0.026320945471525192,
-0.014860390685498714,
-0.0004407764063216746,
0.115360327064991,
0.10508690029382706,
-0.03822844475507736,
-0.002604059409350157,
0.01646163873374462,
-0.06798996031284332,
0.018338976427912712,
0.03595424070954323,
-0.045835886150598526,
8.225654697677394e-33,
0.0328172966837883,
-0.005820526275783777,
-0.01116093061864376,
-0.02629607543349266,
0.050945620983839035,
0.05600416660308838,
-0.11625005304813385,
0.031892575323581696,
0.004968611057847738,
0.02080206759274006,
-0.10384748131036758,
-0.004513786640018225,
-0.044568002223968506,
0.062181878834962845,
0.05679710954427719,
0.03142916038632393,
0.00494375079870224,
0.005803190171718597,
-0.0005347927799448371,
-0.11400750279426575,
0.0056819189339876175,
-0.0326196625828743,
0.030454019084572792,
0.09906021505594254,
0.005202350672334433,
-0.05538545548915863,
-0.04517102986574173,
-0.03831751272082329,
-0.07417416572570801,
-0.02373269386589527,
-0.07948075234889984,
0.08414386957883835,
0.03453519195318222,
0.05573105067014694,
-0.037597306072711945,
-0.013781175017356873,
-0.08153560012578964,
-0.10024359077215195,
-0.006998907309025526,
-0.03778919205069542,
-0.010317845270037651,
0.004902012180536985,
-0.06435848027467728,
-0.03291933238506317,
-0.016742724925279617,
0.03613542765378952,
-0.015026126988232136,
-0.12321460992097855,
-0.04731467366218567,
-0.06662927567958832,
0.07801299542188644,
-0.00679200841113925,
-0.02325069159269333,
0.04047189652919769,
-0.018439942970871925,
-0.03374192491173744,
0.08194266259670258,
0.07035408914089203,
0.047484830021858215,
0.03852984309196472,
-0.04509323835372925,
0.026444273069500923,
0.0016725995810702443,
-0.025096990168094635,
-0.04601991921663284,
-0.0025028043892234564,
-0.011879418976604939,
0.09537709504365921,
0.01820104941725731,
-0.025766227394342422,
-0.034114349633455276,
0.052406635135412216,
-0.04034062474966049,
-0.06396003812551498,
-0.009853956289589405,
-0.03091314435005188,
0.003484531305730343,
-0.11085552722215652,
-0.004340184852480888,
0.033733900636434555,
-0.06163860112428665,
-0.054130036383867264,
0.00013087572006043047,
-0.021113907918334007,
-0.01437996793538332,
0.022130226716399193,
0.09035447239875793,
-0.022477734833955765,
0.04428074136376381,
-0.05419856682419777,
-0.030244361609220505,
0.029249822720885277,
0.0029718982987105846,
0.026071825996041298,
0.025787197053432465,
-8.45360803155679e-33,
-0.038892365992069244,
0.10452142357826233,
0.0760081484913826,
0.07545920461416245,
0.013810748234391212,
-0.061759017407894135,
0.004685004707425833,
-0.06121078133583069,
-0.04996098205447197,
-0.026943663135170937,
0.026782751083374023,
-0.08951636403799057,
0.005541536025702953,
-0.04748924449086189,
0.039488714188337326,
0.005716197658330202,
-0.09519041329622269,
-0.05700629577040672,
0.005140559282153845,
0.10190500319004059,
0.019999176263809204,
0.02371561899781227,
-0.10738394409418106,
0.007982197217643261,
0.019734535366296768,
0.02740302123129368,
-0.06123858690261841,
-0.06767848134040833,
0.06107291206717491,
-0.018864870071411133,
-0.05976534262299538,
0.041342705488204956,
-0.07096320390701294,
-0.07054448127746582,
0.06079636886715889,
0.013041157275438309,
-0.008760063908994198,
0.003048537764698267,
-0.004059534054249525,
0.08431520313024521,
0.02099504880607128,
0.009645448997616768,
-0.07744322717189789,
0.024346662685275078,
-0.0007597532239742577,
0.030072517693042755,
-0.010903908871114254,
0.06722333282232285,
-0.087873674929142,
0.019159909337759018,
-0.0366198755800724,
0.05823006108403206,
-0.03602728247642517,
0.007721044588834047,
-0.01506813708692789,
-0.012094218283891678,
0.08553683757781982,
0.03537621349096298,
0.003005462931469083,
0.02154279313981533,
-0.07827433198690414,
-0.08870724588632584,
0.007861895486712456,
0.02760740928351879,
0.0060541508719325066,
-0.02553659863770008,
0.05310456454753876,
0.031062051653862,
-0.06602819263935089,
0.060017023235559464,
-0.053159430623054504,
-0.04578413814306259,
0.0016359667060896754,
-0.05758281052112579,
-0.006834208499640226,
-0.01826913096010685,
0.0304141603410244,
-0.0008770640706643462,
-0.0006088321097195148,
0.006260124500840902,
-0.006487847771495581,
0.01016481127589941,
-0.026273783296346664,
0.04853389039635658,
0.026892516762018204,
0.09726616740226746,
-0.00777775514870882,
0.07823791354894638,
0.053711242973804474,
0.022580761462450027,
0.004913932643830776,
0.04220281541347504,
-0.09948541224002838,
0.03153013065457344,
0.03340362012386322,
-4.9478206420872084e-8,
-0.12418422847986221,
0.009891429916024208,
-0.005584354512393475,
-0.02847978100180626,
0.023362373933196068,
-0.0512654073536396,
0.05751323327422142,
0.13265328109264374,
0.009978607296943665,
0.06948845088481903,
0.03659643977880478,
0.004476621281355619,
-0.08804936707019806,
-0.04433952644467354,
0.021305175498127937,
-0.017195872962474823,
-0.015448424033820629,
-0.039381902664899826,
0.021310074254870415,
0.02379102073609829,
0.010177268646657467,
-0.06379636377096176,
-0.008206218481063843,
0.014900293201208115,
-0.007426020689308643,
-0.13557271659374237,
-0.05273062363266945,
0.08590882271528244,
0.002254145685583353,
0.0009511932730674744,
-0.05901624262332916,
0.024484457448124886,
0.06525818258523941,
-0.02421071007847786,
0.07506367564201355,
0.0892469733953476,
0.030548522248864174,
-0.025433948263525963,
-0.04321569949388504,
0.05484448000788689,
0.03071088343858719,
0.10137563943862915,
-0.02760862372815609,
-0.015156296081840992,
-0.06761851161718369,
0.02986181527376175,
0.14226123690605164,
0.038474030792713165,
0.023224039003252983,
-0.0014871822204440832,
-0.05300169065594673,
-0.05625231936573982,
-0.01956799067556858,
-0.006030441261827946,
0.04413089156150818,
0.027849603444337845,
-0.07564246654510498,
-0.05245762690901756,
0.10012667626142502,
-0.02004997804760933,
-0.007394821383059025,
-0.06564854085445404,
-0.041342440992593765,
-0.09690270572900772
] |
Learning to Generate Diverse Pedestrian Movements from Web Videos with Noisy Labels | https://openreview.net/forum?id=DydCqKa6AH | [
"Zhizheng Liu",
"Joe Lin",
"Wayne Wu",
"Bolei Zhou"
] | Poster | Understanding and modeling pedestrian movements in the real world is crucial for applications like motion forecasting and scene simulation. Many factors influence pedestrian movements, such as scene context, individual characteristics, and goals, which are often ignored by the existing human generation methods. Web videos contain natural pedestrian behavior and rich motion context, but annotating them with pre-trained predictors leads to noisy labels. In this work, we propose learning diverse pedestrian movements from web videos. We first curate a large-scale dataset called CityWalkers that captures diverse real-world pedestrian movements in urban scenes. Then, based on CityWalkers, we propose a generative model called PedGen for diverse pedestrian movement generation. PedGen introduces automatic label filtering to remove the low-quality labels and a mask embedding to train with partial labels. It also contains a novel context encoder that lifts the 2D scene context to 3D and can incorporate various context factors in generating realistic pedestrian movements in urban scenes. Experiments show that PedGen outperforms existing baseline methods for pedestrian movement generation by learning from noisy labels and incorporating the context factors. In addition, PedGen achieves zero-shot generalization in both real-world and simulated environments. The code, model, and data are available at https://genforce.github.io/PedGen/. | Pedestrian Movement Analysis, Human Motion Dataset, Human Motion Generation | null | 1,930 | 2410.07500 | [
-0.0350240021944046,
-0.12201717495918274,
0.03636699542403221,
0.01678764447569847,
0.13023968040943146,
0.034796908497810364,
0.027396034449338913,
-0.08393078297376633,
-0.004345819354057312,
-0.11331172287464142,
0.0164254792034626,
-0.10376811772584915,
-0.009386763907968998,
0.06956816464662552,
0.021451594308018684,
0.0240179356187582,
0.1199367567896843,
0.13644056022167206,
-0.09649936109781265,
0.028686370700597763,
-0.03522161394357681,
-0.029146254062652588,
0.0588078498840332,
0.03078494407236576,
0.015966009348630905,
0.012226300314068794,
0.0464104562997818,
-0.027912503108382225,
0.06614848971366882,
-0.010799541138112545,
0.08597393333911896,
-0.02184215374290943,
-0.07293561846017838,
0.013153194449841976,
-0.10017314553260803,
0.06326425820589066,
-0.048543862998485565,
0.06915013492107391,
0.012481131590902805,
0.010175486095249653,
-0.007413820829242468,
0.015145916491746902,
0.010551952756941319,
0.03737231716513634,
0.024858931079506874,
0.06954062730073929,
0.02169637195765972,
-0.1078462228178978,
0.02970459870994091,
-0.010334458202123642,
-0.06012806296348572,
-0.04258939251303673,
0.054363783448934555,
0.026252979412674904,
-0.019876990467309952,
-0.04856732487678528,
0.04683641716837883,
-0.046632617712020874,
0.07289548218250275,
-0.007298843935132027,
0.016486341133713722,
-0.08428890258073807,
0.01886582188308239,
-0.045793481171131134,
-0.01854824274778366,
-0.060946669429540634,
0.03084430657327175,
-0.009165097028017044,
0.02352362871170044,
-0.0031150667928159237,
-0.021636608988046646,
0.003403006587177515,
0.03357037901878357,
0.06648379564285278,
0.03906923905014992,
0.02697226032614708,
0.00760974595323205,
-0.02740277722477913,
0.03692574054002762,
-0.021053630858659744,
0.033366475254297256,
-0.05375303328037262,
0.03404196351766586,
-0.0004339983279351145,
0.10783794522285461,
0.04391924664378166,
-0.008832563646137714,
0.10092981904745102,
-0.07809927314519882,
0.007505530025810003,
-0.12172059714794159,
0.05267404392361641,
0.042703427374362946,
-0.060109227895736694,
-0.046376075595617294,
0.0036115963011980057,
-0.05585075169801712,
-0.11076405644416809,
0.010446418076753616,
0.07002486288547516,
0.0039023319259285927,
0.011807562783360481,
0.03341590613126755,
0.007466578856110573,
0.056719787418842316,
-0.027937397360801697,
0.03701845929026604,
0.01963959075510502,
0.003570630680769682,
0.023133911192417145,
0.05253923684358597,
-0.03031259775161743,
0.0106856320053339,
0.03558516129851341,
0.016065020114183426,
0.007939331233501434,
-0.028984853997826576,
0.027897605672478676,
-0.008628087118268013,
0.09460987150669098,
-0.0626440942287445,
0.014219684526324272,
-0.06459996849298477,
0.04254671558737755,
-0.01848890446126461,
-0.07543274760246277,
-0.008098183199763298,
-6.542886963688279e-34,
0.06935888528823853,
0.00933845154941082,
0.0806962326169014,
0.08098474889993668,
0.06869132816791534,
-0.018677808344364166,
-0.062179818749427795,
-0.03986113518476486,
0.01697015017271042,
-0.03427775204181671,
0.052741989493370056,
-0.10296600311994553,
-0.0018618908943608403,
0.12064504623413086,
0.01705818437039852,
-0.03502616658806801,
-0.08571112900972366,
-0.0197660643607378,
-0.006133456714451313,
0.10610344260931015,
0.02455919235944748,
0.00541346799582243,
-0.0584913045167923,
-0.008083867840468884,
-0.007131786085665226,
-0.015445125289261341,
0.033982403576374054,
0.05150352045893669,
0.03834814205765724,
-0.02959045022726059,
-0.06563013792037964,
0.004015577491372824,
0.028831971809267998,
0.05060641095042229,
0.06579194217920303,
0.015422835014760494,
-0.05579565837979317,
-0.02477417327463627,
0.0005246325745247304,
0.021154196932911873,
-0.020509138703346252,
0.03249700739979744,
-0.02230127342045307,
-0.031028950586915016,
-0.05839886888861656,
-0.06150950491428375,
0.05103970691561699,
0.05656668171286583,
-0.07373003661632538,
0.05219980329275131,
0.05485311523079872,
0.01654200442135334,
-0.014449871145188808,
-0.005375216249376535,
-0.09129820764064789,
-0.030348392203450203,
0.06314961612224579,
0.000711760192643851,
-0.02109869010746479,
-0.08132895082235336,
-0.0050771962851285934,
0.058024823665618896,
-0.013226920738816261,
0.022768625989556313,
0.015734879299998283,
-0.029014185070991516,
0.04572174325585365,
0.10223566740751266,
0.061663538217544556,
0.01878325268626213,
0.005289361346513033,
0.04682379588484764,
0.037923842668533325,
-0.07073614746332169,
0.029140138998627663,
0.017803646624088287,
-0.009177811443805695,
-0.07808215916156769,
-0.02611265517771244,
0.04224890470504761,
-0.0469326376914978,
0.02257004752755165,
0.030454717576503754,
-0.012150513008236885,
0.022812914103269577,
-0.0603867769241333,
0.00909383874386549,
-0.04124993830919266,
-0.0016883577918633819,
-0.012036394327878952,
-0.0668649822473526,
-0.0720231756567955,
-0.10323579609394073,
0.01437217928469181,
-0.0028737233951687813,
6.4256520651803916e-34,
-0.0178108811378479,
0.08604808151721954,
0.020295310765504837,
-0.03444874659180641,
-0.025808855891227722,
-0.028790105134248734,
0.006593084894120693,
0.04392627999186516,
0.06475111097097397,
-0.016409585252404213,
-0.08836747705936432,
-0.021341504529118538,
0.060109104961156845,
-0.029772713780403137,
0.03681968152523041,
-0.029218284413218498,
0.0190482959151268,
0.02608766220510006,
-0.07195881754159927,
-0.0042579928413033485,
0.02201024629175663,
0.10446716099977493,
-0.07592158019542694,
-0.002132408320903778,
-0.0619766004383564,
0.017750803381204605,
-0.010689329355955124,
0.07736176252365112,
0.002843467053025961,
-0.014919874258339405,
0.034388888627290726,
-0.07452898472547531,
0.013194549828767776,
-0.05973151698708534,
-0.09725457429885864,
-0.004890811163932085,
0.013697149232029915,
0.019779136404395103,
-0.004501617979258299,
0.08504626899957657,
-0.00534872617572546,
-0.04659918695688248,
-0.023941805586218834,
0.021461058408021927,
-0.07574358582496643,
0.0093319620937109,
-0.03813806176185608,
-0.044484999030828476,
-0.05336681380867958,
0.05503029748797417,
-0.04179403558373451,
0.11067284643650055,
-0.1469697505235672,
0.0003450735821388662,
-0.004831099417060614,
-0.07919957488775253,
-0.037984419614076614,
-0.04248255863785744,
0.08479871600866318,
-0.06950972229242325,
-0.061154384166002274,
0.05782388150691986,
-0.05902811512351036,
0.05396849662065506,
-0.014571544714272022,
-0.03644413873553276,
-0.09622466564178467,
-0.012529812753200531,
-0.008373240008950233,
-0.03993256017565727,
0.05714894458651543,
-0.0016239779070019722,
0.0040621985681355,
0.05614224821329117,
-0.05294850468635559,
-0.06110437214374542,
0.05405997857451439,
0.030130065977573395,
0.046779565513134,
-0.03509463742375374,
0.06242617219686508,
-0.08847352117300034,
0.044260911643505096,
0.010121366940438747,
0.031129220500588417,
0.10633416473865509,
0.016767246648669243,
0.05179964751005173,
-0.028069019317626953,
0.004620387218892574,
-0.00046881544403731823,
0.06389907002449036,
-0.023043103516101837,
0.04220790043473244,
-0.05566222593188286,
-3.26256923699475e-8,
-0.10621768981218338,
0.10549253970384598,
-0.05261322483420372,
0.025780363008379936,
-0.06561022251844406,
0.017203710973262787,
0.022110896185040474,
0.05351486802101135,
0.009057838469743729,
0.037170134484767914,
0.06565754115581512,
0.03362511098384857,
-0.01777123101055622,
0.009594589471817017,
-0.02313704788684845,
0.0011934476206079125,
-0.0475844144821167,
-0.005445000249892473,
-0.005932483356446028,
0.09122904390096664,
-0.02872421219944954,
0.019730310887098312,
-0.029765821993350983,
-0.0207845289260149,
0.09525080770254135,
-0.002196576911956072,
0.0007909790147095919,
0.07425106316804886,
0.014777787029743195,
-0.029462631791830063,
-0.0029082871042191982,
0.026829374954104424,
-0.01975400187075138,
-0.05656563490629196,
-0.032484713941812515,
0.07210318744182587,
-0.0064949458464980125,
-0.021330218762159348,
0.042772866785526276,
-0.05900619924068451,
0.0036936160176992416,
0.04252591356635094,
-0.014627201482653618,
0.039870746433734894,
-0.011483648791909218,
0.05192643404006958,
0.01297834888100624,
-0.14732708036899567,
0.04462685063481331,
-0.007859766483306885,
-0.05201447382569313,
-0.07361860573291779,
0.028549468144774437,
0.07663534581661224,
0.10448936372995377,
-0.017421532422304153,
-0.08641748130321503,
-0.020495161414146423,
0.06444701552391052,
0.07457397133111954,
-0.01536140963435173,
0.059415094554424286,
-0.03531637415289879,
0.01225862093269825
] |
MIRAGE: Evaluating and Explaining Inductive Reasoning Process in Language Models | https://openreview.net/forum?id=tZCqSVncRf | [
"Jiachun Li",
"Pengfei Cao",
"Zhuoran Jin",
"Yubo Chen",
"Kang Liu",
"Jun Zhao"
] | Poster | Inductive reasoning is an essential capability for large language models (LLMs) to achieve higher intelligence, which requires the model to generalize rules from observed facts and then apply them to unseen examples. We present {\scshape Mirage}, a synthetic dataset that addresses the limitations of previous work, specifically the lack of comprehensive evaluation and flexible test data. In it, we evaluate LLMs' capabilities in both the inductive and deductive stages, allowing for flexible variation in input distribution, task scenario, and task difficulty to analyze the factors influencing LLMs' inductive reasoning. Based on these multi-faceted evaluations, we demonstrate that the LLM is a poor rule-based reasoner. In many cases, when conducting inductive reasoning, they do not rely on a correct rule to answer the unseen case. From the perspectives of different prompting methods, observation numbers, and task forms, models tend to consistently conduct correct deduction without correct inductive rules. Besides, we find that LLMs are good neighbor-based reasoners. In the inductive reasoning process, the model tends to focus on observed facts that are close to the current test example in feature space. By leveraging these similar examples, the model maintains strong inductive capabilities within a localized region, significantly improving its deductive performance. | inductive reasoning, large language model, model explanation | Our work introduces Mirage, a synthetic dataset that assesses LLMs' inductive reasoning, showing they often fail to apply rules but excel at using neighbor examples in feature space to improve performance | 1,915 | 2410.09542 | [
-0.04652542248368263,
-0.03627043217420578,
0.017965301871299744,
-0.016941942274570465,
0.12667697668075562,
-0.03137531876564026,
-0.032756052911281586,
0.03598697856068611,
-0.0023363849613815546,
-0.001764854067005217,
-0.06607606261968613,
-0.07816288620233536,
0.049107231199741364,
0.07733766734600067,
0.06003903970122337,
-0.00325103010982275,
0.06456027925014496,
-0.02615228109061718,
-0.10865694284439087,
-0.11366626620292664,
0.02145172655582428,
-0.04531162977218628,
0.01992037147283554,
0.011551727540791035,
-0.010588928125798702,
0.04067165032029152,
0.021145014092326164,
-0.052969254553318024,
0.011307569220662117,
0.03403892368078232,
-0.002389686182141304,
0.033512428402900696,
-0.051282353699207306,
-0.006816867273300886,
0.027550946921110153,
0.05504322797060013,
-0.041848067194223404,
0.08047249913215637,
-0.038112323731184006,
-0.015208051539957523,
-0.08014973998069763,
-0.06579754501581192,
0.008023268543183804,
-0.022776037454605103,
0.09535668790340424,
0.007712178863584995,
-0.05119887739419937,
0.012676376849412918,
-0.010314680635929108,
0.010609135031700134,
-0.08968561887741089,
-0.024590320885181427,
-0.04190312325954437,
0.03838203847408295,
0.032149385660886765,
-0.020098881796002388,
0.04788682237267494,
0.026182817295193672,
-0.016052518039941788,
-0.07649166882038116,
-0.05813547968864441,
-0.13198710978031158,
-0.062286198139190674,
-0.009658458642661572,
-0.06552782654762268,
-0.0064921132288873196,
0.004064165987074375,
0.035446178168058395,
-0.07028449326753616,
0.06317012757062912,
-0.01697564497590065,
0.1306639462709427,
-0.009679888375103474,
0.03855856880545616,
-0.01835562102496624,
0.06322784721851349,
0.028459135442972183,
0.0023232190869748592,
0.05061056837439537,
-0.06055207923054695,
-0.04563254863023758,
0.02330623008310795,
-0.03867724537849426,
0.009699104353785515,
0.11169109493494034,
0.009035772643983364,
0.013063096441328526,
-0.010665973648428917,
0.0070894923992455006,
0.07991824299097061,
0.011386608704924583,
-0.049840494990348816,
-0.0312601774930954,
0.06482214480638504,
0.06533503532409668,
0.06490052491426468,
-0.014039555564522743,
-0.023455694317817688,
-0.0453377291560173,
0.045522112399339676,
-0.0189717635512352,
0.09874089062213898,
0.042786870151758194,
-0.10501478612422943,
-0.010854902677237988,
-0.01965194195508957,
0.05169820040464401,
-0.002375513780862093,
0.056948453187942505,
-0.09809506684541702,
-0.010466654784977436,
0.033304523676633835,
0.037708137184381485,
0.017641523852944374,
-0.020013917237520218,
-0.051012810319662094,
0.03793777525424957,
0.000917004537768662,
-0.038539204746484756,
-0.06361132860183716,
0.015566768124699593,
0.05864640325307846,
0.05018216371536255,
0.08371935784816742,
0.0553399957716465,
-0.040351007133722305,
-0.057364922016859055,
-2.1326054989020547e-34,
0.09813683480024338,
0.03345920890569687,
-0.007122052367776632,
0.0740380808711052,
0.11138191819190979,
-0.024397218599915504,
-0.038735996931791306,
0.09254612028598785,
-0.02343159355223179,
0.05553063377737999,
-0.03196684271097183,
-0.02338622510433197,
-0.06063510477542877,
-0.019059140235185623,
0.017005277797579765,
0.0744960606098175,
-0.08087674528360367,
0.01055449154227972,
-0.10147478431463242,
-0.03360896185040474,
0.056345727294683456,
-0.06116682291030884,
0.012480178847908974,
-0.003619690425693989,
-0.030532686039805412,
0.019331486895680428,
0.06021902337670326,
-0.05410311371088028,
-0.04493110254406929,
0.009171247482299805,
-0.06710100919008255,
-0.10631994158029556,
0.03587907925248146,
0.06429148465394974,
-0.021720515564084053,
-0.02757865935564041,
-0.008139203302562237,
-0.052506592124700546,
0.07809071987867355,
0.009493948891758919,
0.008087052963674068,
0.00494516734033823,
-0.009342366829514503,
0.07933536916971207,
-0.047502871602773666,
-0.03660083934664726,
-0.07882086932659149,
0.0005182084278203547,
-0.013399708084762096,
-0.045995231717824936,
0.05780542641878128,
0.040272779762744904,
0.030100123956799507,
-0.06596562266349792,
-0.029180293902754784,
0.016040606424212456,
-0.009822497144341469,
0.009503846988081932,
-0.010744214057922363,
0.028256945312023163,
0.012897069565951824,
0.11113810539245605,
-0.05959799140691757,
0.12667717039585114,
-0.036295291036367416,
-0.030221622437238693,
-0.08540724962949753,
0.10118859261274338,
0.0920257717370987,
-0.037166208028793335,
-0.018788618966937065,
0.018153896555304527,
-0.0459011048078537,
-0.003864826401695609,
0.06271114200353622,
-0.03253304958343506,
0.05631573125720024,
-0.04835594445466995,
0.04618997126817703,
-0.014851762913167477,
0.04812108352780342,
0.010964246466755867,
-0.004249101504683495,
-0.010630943812429905,
-0.046132322400808334,
-0.05377748981118202,
-0.01129313837736845,
-0.08684208244085312,
-0.018374014645814896,
-0.04729517176747322,
-0.031213413923978806,
-0.0506565161049366,
0.014571404084563255,
-0.040055546909570694,
0.03772982954978943,
-1.0478729495344272e-33,
-0.05355677381157875,
0.021404555067420006,
-0.008121545426547527,
0.039275191724300385,
-0.047407738864421844,
-0.011925294995307922,
0.023832593113183975,
0.011513103730976582,
0.02170065976679325,
-0.05844280496239662,
-0.07869447767734528,
-0.02201301045715809,
0.010327550582587719,
0.0011317614698782563,
0.020901499316096306,
-0.050980377942323685,
-0.006607400719076395,
-0.01996666006743908,
0.008535997942090034,
0.09862995147705078,
-0.029180286452174187,
0.12670482695102692,
-0.165642648935318,
-0.004236399661749601,
-0.05257836729288101,
0.039323482662439346,
-0.046475328505039215,
0.06393813341856003,
0.029107453301548958,
0.03838914632797241,
-0.034022700041532516,
-0.03022075816988945,
0.009088891558349133,
-0.007518773898482323,
-0.06875145435333252,
0.008604485541582108,
0.018729109317064285,
-0.059209637343883514,
-0.05545886233448982,
0.11772479861974716,
0.042877063155174255,
-0.009127898141741753,
-0.042840708047151566,
0.0239173024892807,
-0.04631736874580383,
-0.008028889074921608,
-0.047181617468595505,
0.037535034120082855,
0.023493748158216476,
-0.00465426268056035,
0.007219373714178801,
-0.018110094591975212,
-0.006525735836476088,
0.06741290539503098,
-0.002706763334572315,
-0.04392750561237335,
-0.013451408594846725,
-0.03675585612654686,
0.0489557646214962,
-0.015725940465927124,
-0.022544344887137413,
0.008827745914459229,
0.05687612295150757,
-0.03967416658997536,
0.045246779918670654,
-0.011092976666986942,
-0.041317690163850784,
0.02063540741801262,
0.047075919806957245,
-0.04225020855665207,
0.037283897399902344,
-0.025798862800002098,
-0.0780334323644638,
0.005074198357760906,
0.06434616446495056,
-0.007095032837241888,
-0.028825554996728897,
-0.06723501533269882,
-0.0514075867831707,
-0.08513618260622025,
0.0006569065735675395,
-0.049302272498607635,
0.0028679948300123215,
0.02003900520503521,
0.0007576392381452024,
0.08322525769472122,
0.07632583379745483,
0.09076107293367386,
-0.01688433811068535,
0.053241051733493805,
-0.05650799348950386,
0.03230080008506775,
-0.04832049459218979,
0.10181712359189987,
-0.08087243139743805,
-4.2281872225657935e-8,
-0.06964891403913498,
0.054780181497335434,
-0.014636331237852573,
0.05063071846961975,
0.02952040359377861,
-0.0945039838552475,
-0.029672618955373764,
0.08265280723571777,
0.01820407621562481,
-0.1021002009510994,
0.03505121171474457,
0.008538871072232723,
-0.041283052414655685,
-0.017750060185790062,
0.050335679203271866,
0.08246297389268875,
0.012581784278154373,
0.02492048591375351,
-0.04154463857412338,
-0.00988817773759365,
0.046622149646282196,
0.0398373156785965,
-0.05724203959107399,
0.03443540632724762,
0.04787778481841087,
-0.08299912512302399,
-0.018447542563080788,
0.14473766088485718,
0.057471003383398056,
-0.030773859471082687,
0.02236591838300228,
-0.008237994275987148,
0.007493059150874615,
-0.028807666152715683,
0.06299394369125366,
0.08754008263349533,
0.011466274969279766,
-0.009799010120332241,
0.03654038533568382,
-0.006495368201285601,
0.043272800743579865,
0.09569767862558365,
-0.059136975556612015,
0.07270582020282745,
-0.002023224486038089,
-0.05416087806224823,
-0.03551829233765602,
-0.12905362248420715,
-0.04659529775381088,
0.003005140693858266,
-0.022056251764297485,
-0.041870489716529846,
-0.012916112318634987,
0.029607662931084633,
0.021716903895139694,
0.060890570282936096,
-0.020085537806153297,
0.01062336191534996,
-0.01608227752149105,
0.02746267057955265,
0.04140959680080414,
0.088188037276268,
0.029363881796598434,
0.07541340589523315
] |
One for all and all for one: Efficient computation of partial Wasserstein distances on the line | https://openreview.net/forum?id=kzEPsHbJDv | [
"Laetitia Chapel",
"Romain Tavenard"
] | Poster | Partial Wasserstein helps overcoming some of the limitations of Optimal Transport when the distributions at stake differ in mass, contain noise or outliers or exhibit mass mismatches across distribution modes.
We introduce PAWL, a novel algorithm designed to efficiently compute exact PArtial Wasserstein distances on the Line. PAWL not only solves the partial transportation problem for a specified amount of mass to be transported, but _for all_ admissible mass amounts. This flexibility is valuable for machine learning tasks where the level of noise is uncertain and needs to be determined through cross-validation, for example.
By achieving $O(n \log n)$ time complexity for the partial 1-Wasserstein problem on the line, it enables practical applications with large scale datasets.
Additionally, we introduce a novel slicing strategy tailored to Partial Wasserstein, which does not permit transporting mass between outliers or noisy data points. We demonstrate the advantages of PAWL in terms of computational efficiency and performance in downstream tasks, outperforming existing (sliced) Partial Optimal Transport techniques. | Optimal Transport, Partial Optimal Transport, Sliced Partial Optimal Transport | We introduce a novel algorithm designed to efficiently compute all exact PArtial Wasserstein distances on the Line with a complexity O(n log(n)) | 1,909 | null | [
-0.06586987525224686,
-0.06658640503883362,
0.07712603360414505,
0.05228147283196449,
0.06858912855386734,
-0.08721847832202911,
-0.010980168357491493,
-0.029776904731988907,
-0.025353558361530304,
-0.025947267189621925,
-0.014874404296278954,
-0.055114857852458954,
-0.031192781403660774,
0.06250524520874023,
-0.10713198781013489,
-0.006922516506165266,
0.10041707009077072,
0.05556817352771759,
-0.1093607172369957,
0.040799569338560104,
-0.0395902656018734,
0.03107638843357563,
-0.09303668141365051,
0.044930871576070786,
0.07985842972993851,
-0.03390548750758171,
0.04789697006344795,
-0.026078546419739723,
-0.04351373761892319,
-0.023283345624804497,
-0.003952525090426207,
-0.06430573761463165,
-0.02304486744105816,
-0.052027445286512375,
0.010522541590034962,
0.08942368626594543,
-0.02751227095723152,
0.06531494110822678,
0.024950586259365082,
-0.02077905833721161,
0.015079564414918423,
-0.011467487551271915,
0.06194945052266121,
0.07847056537866592,
-0.03646599501371384,
-0.02286398969590664,
-0.003270617453381419,
-0.04785619303584099,
0.09882792085409164,
0.044771187007427216,
-0.009581027552485466,
0.007455357350409031,
-0.03880771994590759,
0.11633165180683136,
-0.00041467550909146667,
-0.03090047463774681,
0.022760624065995216,
-0.02779274433851242,
-0.047144513577222824,
0.044715575873851776,
-0.01644020341336727,
-0.1368749439716339,
0.009846555069088936,
-0.08788180351257324,
-0.05710422992706299,
-0.051121845841407776,
-0.01838398352265358,
0.05499901622533798,
-0.01586596481502056,
0.05044367536902428,
0.008023061789572239,
0.12206292897462845,
-0.11899468302726746,
0.10517477244138718,
0.06444434076547623,
-0.0032503418624401093,
0.04166718199849129,
0.03031051717698574,
0.03884834051132202,
-0.04861060157418251,
-0.007956880144774914,
0.033998988568782806,
0.017775490880012512,
-0.049238741397857666,
0.09606817364692688,
0.005008900072425604,
-0.05133677273988724,
0.13391385972499847,
0.03641564026474953,
-0.03901803120970726,
0.04063103348016739,
-0.015055760741233826,
0.002613557269796729,
-0.0014750041300430894,
-0.06109592691063881,
-0.01783292181789875,
-0.028821418061852455,
0.024195468053221703,
0.058584533631801605,
0.1020006611943245,
-0.05159183219075203,
0.08412942290306091,
-0.01551075093448162,
-0.04175389185547829,
-0.027603160589933395,
-0.04106282442808151,
0.05573272332549095,
0.012259487994015217,
-0.009361138567328453,
-0.081842340528965,
0.07939474284648895,
-0.05366012454032898,
0.020018812268972397,
-0.02906947024166584,
0.010222353972494602,
-0.03783232718706131,
0.019017139449715614,
-0.02765575423836708,
-0.07338464260101318,
0.028354285284876823,
-0.09138917922973633,
-0.03110508807003498,
0.09691111743450165,
0.1111929640173912,
0.030886482447385788,
0.0007418618188239634,
0.06843866407871246,
3.758387149390976e-33,
-0.03477584570646286,
0.06127015873789787,
0.062345948070287704,
-0.007286181673407555,
0.001601383090019226,
-0.026825835928320885,
-0.010700227692723274,
-0.028630292043089867,
0.03413665294647217,
-0.013478224165737629,
-0.055914975702762604,
-0.08010833710432053,
-0.02440495602786541,
-0.011832654476165771,
0.033976275473833084,
-0.07014494389295578,
0.031155530363321304,
-0.0016061023343354464,
-0.0038322743494063616,
-0.030389249324798584,
0.04388664662837982,
-0.11741136759519577,
-0.02486717887222767,
-0.008276121690869331,
0.033245403319597244,
-0.03254036605358124,
-0.025331633165478706,
-0.03155058994889259,
0.07595279812812805,
0.022284455597400665,
-0.06938951462507248,
-0.034886982291936874,
0.010204496793448925,
0.04858876019716263,
0.03406112268567085,
-0.009686593897640705,
-0.02167530171573162,
-0.04298967123031616,
-0.016333406791090965,
-0.023110250011086464,
-0.004315359052270651,
0.06292102485895157,
-0.024274881929159164,
0.0050376965664327145,
-0.1062786877155304,
-0.028877858072519302,
0.03608778491616249,
-0.033223241567611694,
-0.042671240866184235,
-0.0917963832616806,
0.03390968590974808,
-0.049232423305511475,
0.08645441383123398,
-0.05589168146252632,
-0.06484018266201019,
-0.019672976806759834,
0.048618242144584656,
0.012582659721374512,
0.061498355120420456,
0.10953182727098465,
0.009399029426276684,
0.03856122866272926,
0.006508044432848692,
0.079131580889225,
0.055380672216415405,
0.030235009267926216,
0.010229961015284061,
0.013739914633333683,
0.003436704631894827,
-0.020900515839457512,
0.027750834822654724,
-0.009440572001039982,
0.045684974640607834,
-0.03234798461198807,
0.01164887472987175,
0.0072100404649972916,
0.026580877602100372,
0.005391033831983805,
0.014417833648622036,
-0.06686902046203613,
-0.0819954052567482,
0.04620108753442764,
0.013341267593204975,
-0.032473865896463394,
-0.057289205491542816,
-0.029694270342588425,
0.03464242443442345,
-0.07559729367494583,
-0.1302194595336914,
0.021102674305438995,
-0.06247542425990105,
0.037663087248802185,
-0.02231420762836933,
-0.008628832176327705,
0.014376402832567692,
-3.446394785635417e-33,
-0.0373355969786644,
0.05716494843363762,
0.07340465486049652,
-0.040937066078186035,
-0.042398154735565186,
0.036687519401311874,
-0.04010812193155289,
0.027140948921442032,
0.05782420560717583,
-0.004608260467648506,
-0.07874444872140884,
-0.006276726257055998,
0.055457886308431625,
-0.03228462487459183,
0.013988624326884747,
0.014229737222194672,
-0.021546058356761932,
-0.0003959597961511463,
-0.033092204481363297,
-0.03223460912704468,
0.06629744172096252,
0.023867979645729065,
-0.06487326323986053,
0.033748891204595566,
-0.04091836139559746,
0.04600038379430771,
0.024372827261686325,
0.06136776879429817,
-0.05767960473895073,
0.04620574042201042,
-0.046942088752985,
-0.026642316952347755,
-0.015268433839082718,
-0.05518611520528793,
-0.04802335426211357,
0.13821613788604736,
-0.0017138392431661487,
0.12274321168661118,
0.0014698597369715571,
0.09312708675861359,
-0.028414862230420113,
0.017533233389258385,
-0.07539132237434387,
0.012597217224538326,
-0.059777140617370605,
0.0020130148623138666,
-0.05787808075547218,
0.045519888401031494,
0.03862233832478523,
-0.03094819374382496,
-0.05157369002699852,
0.08231215924024582,
-0.08667344599962234,
0.09691528975963593,
0.010464690625667572,
0.031010696664452553,
-0.03210601583123207,
-0.023511294275522232,
0.03032771684229374,
-0.08114208281040192,
-0.09200029075145721,
-0.06832336634397507,
-0.03025546483695507,
0.009999912232160568,
0.012911279685795307,
-0.00751674547791481,
-0.030700137838721275,
-0.09941035509109497,
-0.01726258173584938,
0.04405885934829712,
-0.032505158334970474,
-0.05437102168798447,
0.01643608883023262,
0.03544415161013603,
-0.02317727729678154,
0.00032941243262030184,
0.10713548213243484,
0.003680165857076645,
0.028127053752541542,
0.08205730468034744,
-0.0017414238536730409,
-0.009552858769893646,
0.026831794530153275,
0.06113271787762642,
0.10225199162960052,
0.031009092926979065,
0.09871283173561096,
-0.04539439454674721,
0.006126207299530506,
-0.0491831973195076,
-0.04986942559480667,
0.06840980798006058,
0.026957480236887932,
0.008846613578498363,
-0.06902375817298889,
-4.087507221584019e-8,
-0.046500518918037415,
-0.05023068189620972,
0.04243721440434456,
-0.09318459033966064,
0.10333795100450516,
-0.007432641927152872,
0.04000410437583923,
0.14019934833049774,
-0.03778824210166931,
0.054621048271656036,
0.038010671734809875,
-0.022306889295578003,
-0.062132466584444046,
-0.0003096868167631328,
-0.00021538042346946895,
-0.008006470277905464,
-0.003398363711312413,
0.029351022094488144,
-0.06523429602384567,
0.0365862175822258,
-0.05120121315121651,
0.01266173180192709,
0.014167550019919872,
-0.040277738124132156,
0.06325184553861618,
-0.0962718203663826,
-0.0020952648483216763,
0.01567489095032215,
0.05635805428028107,
0.0031305572483688593,
0.015732131898403168,
-0.024119360372424126,
0.032304830849170685,
0.02917192131280899,
-0.00030974423862062395,
0.09653771668672562,
-0.03699564188718796,
0.06280836462974548,
-0.07616939395666122,
0.0626089796423912,
0.0226812195032835,
0.049172647297382355,
-0.03737404942512512,
0.026168247684836388,
0.024752141907811165,
0.04181761294603348,
-0.06065346300601959,
-0.0007543776300735772,
-0.00434260256588459,
0.02057688497006893,
0.0015344470739364624,
-0.010042642243206501,
0.05503682419657707,
0.022357730194926262,
0.027330458164215088,
0.01172199659049511,
-0.044524889439344406,
-0.04547997936606407,
0.026067154482007027,
0.020889410749077797,
0.04184577986598015,
0.0020404665265232325,
-0.017129631713032722,
0.009706670418381691
] |
Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models | https://openreview.net/forum?id=s20W12XTF8 | [
"Guobin Shen",
"Dongcheng Zhao",
"Yiting Dong",
"Xiang He",
"Yi Zeng"
] | Poster | As large language models (LLMs) become integral to various applications, ensuring both their safety and utility is paramount. Jailbreak attacks, which manipulate LLMs into generating harmful content, pose significant challenges to this balance. Existing defenses, such as prompt engineering and safety fine-tuning, often introduce computational overhead, increase inference latency, and lack runtime flexibility. Moreover, overly restrictive safety measures can degrade model utility by causing refusals of benign queries. In this paper, we introduce *Jailbreak Antidote*, a method that enables real-time adjustment of LLM safety preferences by manipulating a sparse subset of the model's internal states during inference. By shifting the model's hidden representations along a safety direction with varying strengths, we achieve flexible control over the safety-utility balance without additional token overhead or inference delays. Our analysis reveals that safety-related information in LLMs is sparsely distributed; adjusting approximately *5\%* of the internal state is as effective as modifying the entire state. Extensive experiments on nine LLMs (ranging from 2 billion to 72 billion parameters), evaluated against ten jailbreak attack methods and compared with six defense strategies, validate the effectiveness and efficiency of our approach. By directly manipulating internal states during reasoning, *Jailbreak Antidote* offers a lightweight, scalable solution that enhances LLM safety while preserving utility, opening new possibilities for real-time safety mechanisms in widely-deployed AI systems. | Large Language Models, Jailbreak Defense, Safety-Utility Balance, Internal State Manipulation, Sparse Representation Adjustment | null | 1,906 | 2410.02298 | [
-0.017279325053095818,
-0.047245968133211136,
-0.03783446550369263,
0.002616594545543194,
0.061300914734601974,
0.08255348354578018,
0.033507466316223145,
-0.0004402445338200778,
0.03998066484928131,
0.000402345642214641,
-0.023937759920954704,
-0.07167009264230728,
0.07272713631391525,
-0.054633669555187225,
0.04539012163877487,
0.027742700651288033,
0.14695000648498535,
0.0752221941947937,
-0.1337358057498932,
-0.009965701028704643,
0.012986782938241959,
0.03781421110033989,
0.019911330193281174,
-0.0052012186497449875,
0.008166979067027569,
-0.012949694879353046,
-0.0771791860461235,
-0.00371964811347425,
0.014974554069340229,
-0.015004869550466537,
-0.019792497158050537,
-0.014224461279809475,
0.03313998132944107,
-0.005744820460677147,
0.0030165533535182476,
-0.0498540960252285,
-0.052002809941768646,
0.0339478999376297,
0.003463500877842307,
-0.08877956122159958,
-0.07060647010803223,
0.03876136988401413,
-0.01856771856546402,
0.08889767527580261,
-0.016179654747247696,
-0.06437648087739944,
0.0009313467307947576,
0.04541591554880142,
0.004473975393921137,
-0.02729318104684353,
-0.02880098670721054,
0.0151154650375247,
0.056718047708272934,
0.031009884551167488,
-0.021635636687278748,
-0.05835651978850365,
0.01840846985578537,
0.10509229451417923,
-0.02255234681069851,
0.06923808157444,
-0.06160305440425873,
-0.058446869254112244,
-0.03443530574440956,
0.03469220921397209,
0.00018389264005236328,
0.04048145189881325,
0.01726442016661167,
0.050163716077804565,
0.012119038961827755,
0.05660947784781456,
-0.02219865471124649,
0.04979677125811577,
-0.04818863794207573,
-0.00031648960430175066,
-0.03658781200647354,
0.04932013154029846,
-0.007075455039739609,
-0.05846578627824783,
0.047510094940662384,
-0.03516608849167824,
0.013786735944449902,
-0.045959245413541794,
0.047169093042612076,
0.05433580279350281,
0.09255403280258179,
0.05775047466158867,
-0.016128161922097206,
0.048567622900009155,
0.08383947610855103,
0.03431330993771553,
0.008149971254169941,
0.00376334716565907,
0.06495923548936844,
0.02289585769176483,
0.024356426671147346,
0.03270004317164421,
-0.020764848217368126,
-0.09675677120685577,
-0.07869648188352585,
0.10736899822950363,
0.03594956919550896,
0.053473442792892456,
0.03877178207039833,
-0.029400639235973358,
0.0311881173402071,
-0.011790934950113297,
0.03660983964800835,
0.0213286355137825,
-0.041638389229774475,
0.011566074565052986,
0.05695488303899765,
0.06981060653924942,
0.05987918749451637,
0.03982476890087128,
-0.01352814119309187,
-0.02785056084394455,
0.014503095299005508,
-0.046564679592847824,
0.04737814888358116,
0.053849752992391586,
0.021859271451830864,
0.07106202840805054,
0.047541555017232895,
-0.02477033995091915,
-0.017906472086906433,
0.002905944362282753,
-0.07665617018938065,
1.598620521898478e-33,
0.027450943365693092,
0.03449561074376106,
-0.06311067193746567,
0.026453211903572083,
0.020945824682712555,
-0.04087936505675316,
0.010335354134440422,
0.07812951505184174,
-0.08548033982515335,
0.0374126061797142,
0.016412323340773582,
-0.0035895812325179577,
0.018989983946084976,
0.008452323265373707,
0.10770809650421143,
0.037726908922195435,
-0.04522732272744179,
0.02497493475675583,
0.012064495123922825,
-0.0682157352566719,
0.009321996010839939,
-0.07846799492835999,
-0.0036767025012522936,
-0.032422568649053574,
0.007060361560434103,
0.009317017160356045,
0.041085924953222275,
-0.03161429613828659,
-0.007034525275230408,
0.0404641330242157,
-0.16365869343280792,
-0.04039531946182251,
0.007252408191561699,
-0.02967572771012783,
0.0591777004301548,
-0.02834555320441723,
-0.01839079149067402,
-0.07731792330741882,
0.016440412029623985,
-0.031115708872675896,
-0.060255296528339386,
-0.021955573931336403,
0.032866522669792175,
0.002926790388301015,
-0.049420617520809174,
-0.041692234575748444,
-0.06892025470733643,
-0.004638110287487507,
-0.08971605449914932,
-0.03720797225832939,
0.015345807187259197,
0.05133992061018944,
-0.02306538075208664,
-0.03383402153849602,
-0.14176824688911438,
-0.019402949139475822,
-0.01661464013159275,
-0.004107472952455282,
0.012546570040285587,
0.13615742325782776,
-0.06139647588133812,
0.054694533348083496,
-0.012475443072617054,
0.019011689350008965,
0.049544043838977814,
0.01145089603960514,
-0.0872238501906395,
0.010120436549186707,
0.04863191023468971,
0.0016851129475980997,
-0.07523717731237411,
-0.03136426955461502,
0.006608191411942244,
0.009322197176516056,
-0.079999178647995,
0.004091953858733177,
0.10918256640434265,
-0.017348194494843483,
0.01050011906772852,
-0.049038540571928024,
0.06199263036251068,
-0.061707694083452225,
0.035490766167640686,
0.04869356006383896,
-0.027575064450502396,
-0.08601394295692444,
0.01188726257532835,
-0.07305095344781876,
-0.05821346491575241,
0.03330491483211517,
0.002693061251193285,
-0.05631459131836891,
0.01948462799191475,
0.015488183125853539,
-0.032200612127780914,
-2.133358412217306e-33,
-0.053011782467365265,
-0.08954928070306778,
-0.024725336581468582,
0.02433605305850506,
-0.0797286406159401,
-0.11394642293453217,
-0.05046171322464943,
0.03580315783619881,
0.005968725774437189,
-0.08856025338172913,
-0.1135917529463768,
-0.0038398848846554756,
0.08549301326274872,
-0.005274137016385794,
0.07346232980489731,
-0.01936485432088375,
-0.025603659451007843,
0.013656430877745152,
-0.005010008811950684,
0.10172300785779953,
-0.004450683481991291,
0.07035869359970093,
-0.10354429483413696,
0.04706459119915962,
-0.028001949191093445,
0.00574026582762599,
-0.0933375433087349,
0.029250191524624825,
0.030603474006056786,
0.03382071480154991,
-0.05567334592342377,
-0.012457717210054398,
-0.049870565533638,
-0.0449356883764267,
-0.09164357930421829,
0.04366142675280571,
0.04031863063573837,
-0.017207421362400055,
0.022554388269782066,
0.045248910784721375,
0.08231838047504425,
-0.003994325641542673,
-0.10056712478399277,
-0.0417415015399456,
0.0029637955594807863,
0.028692709282040596,
-0.08032513409852982,
-0.02807634323835373,
-0.0003069541126023978,
-0.04356406629085541,
0.003890251275151968,
-0.004907954018563032,
0.025029288604855537,
0.07610117644071579,
-0.03230981528759003,
-0.12796148657798767,
0.024840977042913437,
-0.08380834758281708,
0.032743725925683975,
-0.025134125724434853,
-0.03266175463795662,
0.008345117792487144,
-0.038128651678562164,
-0.09033475816249847,
0.06919475644826889,
-0.053872693330049515,
-0.028378279879689217,
-0.022651515901088715,
0.02640470489859581,
-0.05064253509044647,
0.06833070516586304,
-0.032359618693590164,
-0.014603689312934875,
0.022790636867284775,
-0.00811831932514906,
0.006272902246564627,
-0.026614198461174965,
-0.0922500491142273,
-0.057691991329193115,
0.05299707129597664,
0.0614967942237854,
0.08920802175998688,
0.033469557762145996,
0.02500947192311287,
0.02728438377380371,
0.07956606894731522,
0.045463256537914276,
0.07746172696352005,
0.03968813642859459,
0.05006427690386772,
-0.0492897629737854,
0.03874516859650612,
-0.03932517394423485,
0.10068278759717941,
-0.08275558054447174,
-4.308173018330308e-8,
0.006134413182735443,
0.03507078438997269,
0.020223120227456093,
0.05536016821861267,
0.04342640936374664,
0.02148769609630108,
-0.053572650998830795,
-0.03351733833551407,
0.025937387719750404,
-0.04715181514620781,
0.08030159771442413,
0.020990435034036636,
-0.032390691339969635,
-0.04678545519709587,
-0.06493018567562103,
0.09954523295164108,
-0.03611919656395912,
0.010586782358586788,
-0.04472291097044945,
-0.020312979817390442,
-0.050284117460250854,
0.07251617312431335,
0.016774965450167656,
-0.018592137843370438,
0.07098587602376938,
-0.03637738153338432,
-0.02397695556282997,
0.04723338410258293,
0.10346519947052002,
0.07066610455513,
-0.04771833494305611,
-0.009779004380106926,
-0.024998845532536507,
-0.014133807271718979,
-0.02104778215289116,
0.13132454454898834,
0.0005819697980768979,
-0.031018896028399467,
0.02008376270532608,
0.006511186249554157,
0.006553837563842535,
-0.012081428430974483,
-0.07512185722589493,
0.023228298872709274,
-0.00857364572584629,
-0.028811033815145493,
-0.018967684358358383,
-0.12606091797351837,
0.10669896006584167,
-0.03219025582075119,
0.03974040597677231,
-0.037363991141319275,
0.02172974869608879,
0.10248629003763199,
0.04108038172125816,
0.06905210018157959,
-0.018438879400491714,
-0.03083726391196251,
0.03090614639222622,
0.039329152554273605,
0.02811550162732601,
0.021501043811440468,
0.030663006007671356,
-0.014018511399626732
] |
Precise Parameter Localization for Textual Generation in Diffusion Models | https://openreview.net/forum?id=gdHtZlaaSo | [
"Łukasz Staniszewski",
"Bartosz Cywiński",
"Franziska Boenisch",
"Kamil Deja",
"Adam Dziedzic"
] | Poster | Novel diffusion models can synthesize photo-realistic images with integrated high-quality text. Surprisingly, we demonstrate through
attention activation patching that only less than $1$\% of diffusion models' parameters, all contained in attention layers, influence the generation of textual content within the images. Building on this observation, we improve textual generation efficiency and performance by targeting cross and joint attention layers of diffusion models. We introduce several applications that benefit from localizing the layers responsible for textual content generation. We first show that a LoRA-based fine-tuning solely of the localized layers enhances, even more, the general text-generation capabilities of large diffusion models while preserving the quality and diversity of the diffusion models' generations. Then, we demonstrate how we can use the localized layers to edit textual content in generated images. Finally, we extend this idea to the practical use case of preventing the generation of toxic text in a cost-free manner. In contrast to prior work, our localization approach is broadly applicable across various diffusion model architectures, including U-Net (e.g., SDXL and DeepFloyd IF) and transformer-based (e.g., Stable Diffusion 3), utilizing diverse text encoders (e.g., from CLIP to the large language models like T5). Project page available at https://t2i-text-loc.github.io/. | diffusion models, text edition, LoRA, localization, SD-XL, SD3, DeepFloyd IF | We introduce a novel method for editing text within images generated by diffusion models, which modifies very few parameters and leaves the other visual content intact. | 1,905 | 2502.09935 | [
0.018480896949768066,
-0.029975688084959984,
0.015049844048917294,
0.05799303576350212,
0.13846860826015472,
-0.029448892921209335,
-0.08206073939800262,
-0.05269787460565567,
0.12118959426879883,
-0.034582484513521194,
0.0038276570849120617,
0.032038480043411255,
0.04140369966626167,
0.02125183492898941,
-0.03858671337366104,
-0.0012394931400194764,
0.08119673281908035,
0.04753817245364189,
-0.13834995031356812,
-0.020531795918941498,
0.006855897605419159,
-0.007433733902871609,
0.08899195492267609,
-0.04230710119009018,
0.0344008170068264,
0.04137770086526871,
-0.042645763605833054,
-0.04544549435377121,
0.064680315554142,
-0.0422021746635437,
0.0688667744398117,
0.006857012864202261,
-0.026863010600209236,
0.060889799147844315,
-0.04713134095072746,
0.06528398394584656,
-0.0809391662478447,
-0.020473793148994446,
-0.005575249902904034,
-0.04273021221160889,
0.043362196534872055,
0.0067531513050198555,
-0.023936621844768524,
0.03972650691866875,
0.07300581783056259,
-0.04175945743918419,
0.01583644561469555,
-0.002554308157414198,
-0.003990668337792158,
0.018651805818080902,
-0.03614092618227005,
-0.04408987984061241,
-0.07124501466751099,
0.01899341493844986,
-0.013127905316650867,
0.010445197112858295,
0.07538451254367828,
0.01765914261341095,
0.029672896489501,
-0.028694618493318558,
-0.07335241883993149,
-0.006522971671074629,
-0.010210453532636166,
-0.024166306480765343,
0.05648098886013031,
0.021096421405673027,
0.024711621925234795,
0.018299706280231476,
-0.08476276695728302,
-0.022915145382285118,
-0.015484782867133617,
0.04047248885035515,
-0.03996367007493973,
0.04410296678543091,
0.01001883763819933,
0.038991063833236694,
0.025628555566072464,
-0.060798436403274536,
-0.013111259788274765,
-0.08686283975839615,
0.11269767582416534,
-0.045282550156116486,
0.12443361431360245,
-0.027798373252153397,
0.04489169642329216,
0.009028089232742786,
0.030933596193790436,
0.013003217056393623,
0.01939125917851925,
-0.004168116953223944,
-0.09353409707546234,
-0.017710700631141663,
-0.026736119762063026,
-0.011016030795872211,
-0.018860388547182083,
0.0359567366540432,
-0.030491920188069344,
-0.06483013927936554,
-0.0065165925770998,
0.07416984438896179,
0.06185270845890045,
-0.05378767102956772,
0.021057095378637314,
-0.008268313482403755,
0.06294111907482147,
-0.06156674399971962,
0.06927039474248886,
0.05398964881896973,
-0.03727927803993225,
-0.014007966965436935,
0.044204723089933395,
0.03358010575175285,
-0.02356104925274849,
-0.016935458406805992,
0.046537552028894424,
-0.11274521052837372,
0.07496169209480286,
-0.04753823205828667,
0.0007615005597472191,
-0.005062017124146223,
-0.0037608554121106863,
0.02021530643105507,
-0.10405288636684418,
0.03577470779418945,
-0.005520138423889875,
-0.08419229835271835,
-0.007800314575433731,
-2.884327478179436e-34,
0.02818317711353302,
0.009027642197906971,
0.023163214325904846,
0.0918617844581604,
0.09039421379566193,
-0.018272878602147102,
0.010991298593580723,
-0.06196646764874458,
-0.07715979218482971,
-0.1100875586271286,
-0.024101998656988144,
-0.03315577283501625,
-0.04611000418663025,
0.10562565177679062,
0.08819661289453506,
-0.042605187743902206,
-0.030521908774971962,
0.04061989486217499,
0.004899392835795879,
0.05362839624285698,
-0.051304955035448074,
-0.05047105625271797,
0.02941862680017948,
-0.05190378800034523,
-0.035080406814813614,
0.03016742318868637,
0.0904010534286499,
-0.08067280054092407,
-0.031421322375535965,
-0.021781057119369507,
-0.14442740380764008,
0.002424038015305996,
0.0640130564570427,
0.023508626967668533,
-0.015475421212613583,
-0.020892664790153503,
-0.06650084257125854,
-0.008417658507823944,
0.032991040498018265,
-0.0077484650537371635,
-0.0015970392851158977,
0.09087494760751724,
-0.018022440373897552,
0.03309538587927818,
-0.07217000424861908,
0.06227396801114082,
-0.02804453857243061,
0.04104813188314438,
-0.03446965292096138,
-0.07054245471954346,
0.06802316009998322,
0.005140851251780987,
0.009735739789903164,
-0.04182588681578636,
-0.0002072975185001269,
-0.015501447021961212,
0.10005176067352295,
-0.07928122580051422,
0.04883219301700592,
-0.0180903822183609,
0.11045219004154205,
0.047498222440481186,
-0.04548131301999092,
0.1015680655837059,
0.04767912998795509,
0.0011836012126877904,
0.03301578387618065,
0.059502068907022476,
0.003666936419904232,
-0.030571475625038147,
-0.04340888932347298,
-0.0078430799767375,
-0.016338879242539406,
-0.028119033202528954,
-0.02866266295313835,
-0.02284334786236286,
0.006560005247592926,
-0.015551934018731117,
0.05415147915482521,
0.021300408989191055,
-0.02258506417274475,
-0.048396624624729156,
-0.0319100059568882,
-0.06410395354032516,
-0.05121566727757454,
-0.017319360747933388,
-0.019431648775935173,
-0.09834457188844681,
-0.04159881919622421,
-0.012187061831355095,
0.02564762346446514,
-0.021257007494568825,
-0.030826466158032417,
-0.004806710872799158,
0.0026169000193476677,
6.033584291562849e-34,
-0.016713576391339302,
0.03660421445965767,
-0.028025317937135696,
0.09382741153240204,
-0.044549982994794846,
0.01080916915088892,
-0.022713204845786095,
0.029948504641652107,
0.07608435302972794,
-0.03368092328310013,
-0.04493856430053711,
-0.016942206770181656,
-0.059637296944856644,
-0.05138813704252243,
-0.015399961732327938,
-0.11578203737735748,
0.06855766475200653,
0.010123980231583118,
-0.08695182204246521,
0.02364826761186123,
-0.020194703713059425,
0.015269181691110134,
-0.04104483500123024,
0.06187717244029045,
0.018275689333677292,
0.03941226005554199,
-0.04703904315829277,
0.0564916767179966,
0.03837749734520912,
-0.07417386025190353,
-0.0663481131196022,
0.038099225610494614,
0.027875715866684914,
-0.014525460079312325,
-0.048586249351501465,
0.09123127162456512,
0.03805386647582054,
0.020781053230166435,
-0.006960076745599508,
0.05680748447775841,
0.053262412548065186,
-0.021026547998189926,
-0.0971318855881691,
0.07936496287584305,
-0.06140880659222603,
-0.016918212175369263,
-0.1002044603228569,
-0.018438266590237617,
-0.044197928160429,
0.04623570665717125,
0.035625189542770386,
0.009034310467541218,
-0.04431817680597305,
0.009590184316039085,
-0.06008594110608101,
-0.04118828475475311,
-0.11549129337072372,
-0.056325700134038925,
0.12403316050767899,
-0.05842175707221031,
-0.03492000326514244,
-0.04000895097851753,
-0.04014677554368973,
-0.12041443586349487,
-0.061709921807050705,
-0.05283705145120621,
-0.018751006573438644,
-0.00013811164535582066,
0.0017617738340049982,
0.07802064716815948,
0.08771757781505585,
-0.04390575736761093,
-0.0027411319315433502,
0.007834943011403084,
0.02953057736158371,
-0.05210549756884575,
0.05735427141189575,
0.027667421847581863,
0.02935810387134552,
-0.060114696621894836,
0.01096308697015047,
0.08431798219680786,
-0.015272588469088078,
0.038283757865428925,
0.08752183616161346,
0.019881341606378555,
0.012475511059165001,
0.032048776745796204,
0.01769532449543476,
0.015182683244347572,
-0.020042603835463524,
0.03380679339170456,
0.032623033970594406,
0.04232265055179596,
-0.0742630660533905,
-3.983947394203824e-8,
-0.06542894244194031,
0.009082276374101639,
-0.04831370338797569,
0.06910067051649094,
-0.03092363476753235,
-0.030129902064800262,
0.07234524190425873,
0.10141512006521225,
0.047727201133966446,
-0.04098972678184509,
0.019587824121117592,
-0.027998896315693855,
-0.04018460586667061,
-0.08632928878068924,
-0.03836909309029579,
0.1031051054596901,
0.049428846687078476,
0.06475275754928589,
0.02062624879181385,
-0.028343988582491875,
-0.04796956479549408,
0.10379135608673096,
0.02348151244223118,
0.0054671308025717735,
0.05908329784870148,
0.006915309466421604,
-0.021587004885077477,
0.016268139705061913,
0.07871200889348984,
-0.008784987963736057,
0.05845878645777702,
-0.02065291814506054,
-0.014688113704323769,
0.03659295290708542,
-0.023966429755091667,
0.08743113279342651,
0.014501960016787052,
0.005698848515748978,
-0.039526768028736115,
-0.03769654780626297,
0.026027115061879158,
0.06134575605392456,
0.004544658586382866,
0.0023985635489225388,
-0.025570034980773926,
-0.023076949641108513,
0.03465158864855766,
-0.05805542692542076,
0.040416471660137177,
-0.04105732589960098,
0.03110341727733612,
-0.07756416499614716,
-0.030986236408352852,
0.05976928398013115,
-0.014216632582247257,
-0.06354297697544098,
0.1011529490351677,
0.034247394651174545,
0.07524577528238297,
0.0873696431517601,
0.02771637961268425,
0.05032481253147125,
0.011536793783307076,
0.026433289051055908
] |
Lightning-Fast Image Inversion and Editing for Text-to-Image Diffusion Models | https://openreview.net/forum?id=t9l63huPRt | [
"Dvir Samuel",
"Barak Meiri",
"Haggai Maron",
"Yoad Tewel",
"Nir Darshan",
"Shai Avidan",
"Gal Chechik",
"Rami Ben-Ari"
] | Poster | Diffusion inversion is the problem of taking an image and a text prompt that describes it and finding a noise latent that would generate the exact same image.
Most current deterministic inversion techniques operate by approximately solving an implicit equation and may converge slowly or yield poor reconstructed images. We formulate the problem by finding the roots of an implicit equation and devlop a method to solve it efficiently. Our solution is based on Newton-Raphson (NR), a well-known technique in numerical analysis. We show that a vanilla application of NR is computationally infeasible while naively transforming it to a computationally tractable alternative tends to converge to out-of-distribution solutions, resulting in poor reconstruction and editing. We therefore derive an efficient guided formulation that fastly converges and provides high-quality reconstructions and editing. We showcase our method on real image editing with three popular open-sourced diffusion models: Stable Diffusion, SDXL-Turbo, and Flux with different deterministic schedulers. Our solution, **Guided Newton-Raphson Inversion**, inverts an image within 0.4 sec (on an A100 GPU) for few-step models (SDXL-Turbo and Flux.1),
opening the door for interactive image editing. We further show improved results in image interpolation and generation of rare objects. | Deterministic Image Inversion, Image Editing, Diffusion Models, Flow Matching, Image Generation | A new image inversion method for text-to-image diffusion models, that introduces real-time editing by employing Newton-Rahson numerical scheme | 1,900 | 2312.12540 | [
-0.08174151182174683,
-0.02333441935479641,
0.06771768629550934,
0.059078119695186615,
0.04917556047439575,
-0.09188075363636017,
-0.09428870677947998,
-0.07658792287111282,
0.07434578984975815,
-0.005449499934911728,
0.006758872885257006,
0.040338821709156036,
-0.011254370212554932,
0.04097386449575424,
-0.04025261104106903,
0.012339295819401741,
0.028018154203891754,
0.07017611712217331,
-0.07625115662813187,
-0.050550539046525955,
-0.07980573922395706,
0.009845880791544914,
-0.06397627294063568,
-0.05340482294559479,
0.10024426132440567,
0.02728201635181904,
0.0006204870878718793,
-0.05887695029377937,
-0.0016391545068472624,
-0.0069251395761966705,
0.04457550868391991,
-0.010120776481926441,
0.012875556014478207,
0.013459055684506893,
-0.05702010169625282,
0.07135779410600662,
-0.014634324237704277,
0.02307618036866188,
0.021949240937829018,
-0.0199859868735075,
0.03810950368642807,
0.02849084697663784,
0.022683538496494293,
0.00047857360914349556,
0.0813949853181839,
-0.029595952481031418,
0.012482292018830776,
-0.04097588360309601,
0.12052503973245621,
0.05200563371181488,
-0.005076548084616661,
-0.024303512647747993,
-0.06444884836673737,
0.0205659382045269,
0.030369017273187637,
0.02859327383339405,
0.07148176431655884,
-0.08888069540262222,
0.04152541980147362,
-0.015493717044591904,
0.01909535378217697,
0.01078728772699833,
-0.032484035938978195,
0.016159696504473686,
-0.016471263021230698,
0.015433789230883121,
0.026156073436141014,
-0.007266159635037184,
-0.006497654132544994,
0.06620597839355469,
-0.014333601109683514,
0.004391942638903856,
-0.031880512833595276,
-0.0118250772356987,
-0.043654754757881165,
-0.03494136780500412,
0.08273065090179443,
-0.03322339430451393,
0.014134185388684273,
-0.05725730583071709,
0.13110144436359406,
0.0017103699501603842,
0.06421975046396255,
-0.03127001225948334,
-0.018227795138955116,
0.013952785171568394,
0.018846163526177406,
0.08889363706111908,
0.03403497487306595,
0.020296305418014526,
-0.05960049107670784,
-0.008221236057579517,
-0.09457266330718994,
-0.0657266154885292,
0.004742044024169445,
-0.0006878150743432343,
0.0636003166437149,
-0.04516732320189476,
-0.004628188908100128,
0.13068801164627075,
0.004354182630777359,
-0.028719503432512283,
-0.026326250284910202,
-0.060204435139894485,
-0.03121713176369667,
-0.04888560622930527,
0.04741143807768822,
0.05061917379498482,
-0.055593349039554596,
-0.008929619565606117,
0.10865870863199234,
-0.01280313078314066,
-0.011359763331711292,
0.041528116911649704,
0.0940825343132019,
-0.10159022361040115,
0.045374978333711624,
-0.020096318796277046,
-0.027113312855362892,
0.049880508333444595,
-0.023767268285155296,
0.0015434626257047057,
-0.023897958919405937,
0.039545439183712006,
0.017212672159075737,
-0.05605646222829819,
-0.01402968168258667,
-1.1617735766833067e-33,
-0.0032069419976323843,
0.05570892617106438,
0.03558580204844475,
0.028741968795657158,
0.06670170277357101,
0.0009699304937385023,
0.022222371771931648,
-0.07884129881858826,
-0.012093295343220234,
0.012431330047547817,
0.010229786857962608,
-0.033526286482810974,
-0.04671528562903404,
0.06464964896440506,
0.03304333612322807,
-0.05957210063934326,
-0.00902816466987133,
0.059518877416849136,
-0.030529433861374855,
0.022098401561379433,
0.009678099304437637,
-0.0398239828646183,
-0.04704070836305618,
-0.07445891201496124,
-0.0932062417268753,
-0.003273583482950926,
0.07013975083827972,
-0.09930388629436493,
0.08317043632268906,
-0.022014372050762177,
-0.10051948577165604,
-0.002810945501551032,
0.021593594923615456,
0.021643856540322304,
-0.0020806570537388325,
0.00663105258718133,
0.05067313089966774,
-0.1177155151963234,
0.012697739526629448,
-0.00043631542939692736,
-0.019542057067155838,
0.10567230731248856,
0.004726348910480738,
-0.04351889714598656,
-0.08760672062635422,
0.07771874964237213,
-0.015354909002780914,
0.035452891141176224,
0.027293115854263306,
-0.0008059483952820301,
0.030228180810809135,
0.0033220520708709955,
-0.031109794974327087,
-0.0763293132185936,
-0.0015378087991848588,
-0.014922080561518669,
0.08184046298265457,
-0.06962639093399048,
0.07702300697565079,
0.04156685248017311,
0.06307174265384674,
0.05409926548600197,
-0.052162203937768936,
0.038268279284238815,
0.0027723682578653097,
0.020006194710731506,
-0.03451021388173103,
0.012042340822517872,
-0.016638118773698807,
-0.02123747020959854,
-0.10251777619123459,
-0.01288935262709856,
-0.05930529907345772,
-0.060369353741407394,
0.03646537661552429,
0.00794139876961708,
0.08485260605812073,
0.0033501239959150553,
0.04727243259549141,
-0.03386955335736275,
-0.11277949810028076,
-0.03840941563248634,
0.010486166924238205,
-0.06380797922611237,
-0.012104843743145466,
-0.03286357969045639,
-0.04815416410565376,
-0.013596558012068272,
-0.06306608021259308,
-0.06469636410474777,
-0.073063924908638,
0.023414544761180878,
-0.02078855037689209,
-0.006276577245444059,
0.022346829995512962,
9.532677866478072e-34,
-0.006032074801623821,
-0.007986434735357761,
-0.026359064504504204,
0.08118163049221039,
-0.018713563680648804,
-0.02698655240237713,
-0.002715452341362834,
0.029057450592517853,
0.10166692733764648,
-0.07838135212659836,
-0.06526753306388855,
0.0561642050743103,
-0.023663921281695366,
-0.021744444966316223,
0.025686077773571014,
-0.09596651047468185,
0.04094884172081947,
-0.021778207272291183,
-0.12786349654197693,
0.016878711059689522,
0.004435243550688028,
0.10214613378047943,
0.0714758038520813,
0.002848112490028143,
-0.05857577547430992,
0.08879078924655914,
-0.01572970114648342,
0.12578867375850677,
0.02214239537715912,
-0.019892117008566856,
-0.05857361853122711,
0.015535939484834671,
0.04928046464920044,
-0.03426032140851021,
-0.04514884203672409,
0.11392886936664581,
0.04349999129772186,
-0.006196308881044388,
-0.0154661750420928,
0.021884087473154068,
0.04643469303846359,
0.0016024105716496706,
-0.03080684505403042,
0.009939738549292088,
-0.010506987571716309,
0.03433425351977348,
-0.09122925996780396,
-0.024534380063414574,
0.06250590831041336,
0.08719414472579956,
-0.024055149406194687,
0.012904458679258823,
-0.0357818566262722,
0.09058182686567307,
-0.012837361544370651,
-0.07684686034917831,
-0.08855482935905457,
-0.05947599187493324,
0.09152255952358246,
-0.03769468143582344,
-0.09445490688085556,
-0.038634102791547775,
0.05312754958868027,
-0.059517309069633484,
-0.028450796380639076,
-0.033497586846351624,
-0.03639493137598038,
-0.021764986217021942,
-0.05205028876662254,
0.10748450458049774,
0.0002182347234338522,
-0.0235911775380373,
0.037374347448349,
0.029314158484339714,
0.02288886532187462,
-0.07678194344043732,
0.09433672577142715,
0.035222116857767105,
0.03147101402282715,
-0.0668165534734726,
0.027258016169071198,
0.09271596372127533,
0.010014847852289677,
0.03629546985030174,
0.02891160547733307,
0.008231885731220245,
0.024980371817946434,
-0.028868364170193672,
0.07464693486690521,
0.02835220657289028,
-0.043693557381629944,
0.04194868355989456,
0.11419913917779922,
-0.019285740330815315,
-0.02452050894498825,
-4.337432457646173e-8,
-0.031235285103321075,
0.05073145031929016,
0.042256880551576614,
-0.015416869893670082,
-0.0017426194390282035,
-0.03210017830133438,
0.08270211517810822,
-0.05631948634982109,
0.04497150331735611,
-0.11777196824550629,
0.08170325309038162,
-0.0303316880017519,
-0.02840346284210682,
-0.0833253562450409,
-0.10362651944160461,
0.08003463596105576,
-0.03382011502981186,
-0.05396677553653717,
-0.002554624807089567,
-0.009416958317160606,
-0.1215503141283989,
0.07224393635988235,
0.0033258176408708096,
-0.02065180242061615,
0.05695435032248497,
0.05848843604326248,
-0.043025288730859756,
-0.03532678633928299,
0.08168155699968338,
-0.01711130142211914,
0.02522309496998787,
0.008166114799678326,
0.05108474940061569,
0.07013870030641556,
-0.043171148747205734,
0.0023284044582396746,
0.04385681822896004,
0.05384770408272743,
-0.04778486117720604,
-0.09328269213438034,
-0.05259733274579048,
0.034614402800798416,
0.0027649959083646536,
-0.031672414392232895,
-0.02273287996649742,
-0.000046674031182192266,
0.07155876606702805,
-0.04020950570702553,
0.02182498574256897,
0.017840107902884483,
-0.008746604435145855,
-0.004330337978899479,
-0.0272351261228323,
0.0060389987193048,
0.04287852346897125,
-0.005618696101009846,
0.0003934735432267189,
0.026103373616933823,
0.059091683477163315,
0.048822805285453796,
0.006255933549255133,
0.019765207543969154,
-0.0008508116006851196,
0.01685452274978161
] |
Revisiting Convolution Architecture in the Realm of DNA Foundation Models | https://openreview.net/forum?id=B07dLVWLyD | [
"Yu Bo",
"Weian Mao",
"Yanjun Shao",
"Weiqiang Bai",
"Peng Ye",
"Xinzhu Ma",
"Junbo Zhao",
"Hao Chen",
"Chunhua Shen"
] | Poster | In recent years, A variety of methods based on Transformer and state space model (SSM) architectures have been proposed, advancing foundational DNA language models.
However, there is a lack of comparison between these recent approaches and the classical architecture—convolutional networks (CNNs)—on foundation model benchmarks.
This raises the question: are CNNs truly being surpassed by these recent approaches based on transformer and SSM architectures? In this paper, we develop a simple but well-designed CNN-based method, termed ConvNova. ConvNova identifies and proposes three effective designs: 1) dilated convolutions, 2) gated convolutions, and 3) a dual-branch framework for gating mechanisms.
Through extensive empirical experiments, we demonstrate that ConvNova significantly outperforms recent methods on more than half of the tasks across several foundation model benchmarks. For example, in histone-related tasks, ConvNova exceeds the second-best method by an average of 5.8\%, while generally utilizing fewer parameters and enabling faster computation. In addition, the experiments observed findings that may be related to biological characteristics. This indicates that CNNs are still a strong competitor compared to Transformers and SSMs. We anticipate that this work will spark renewed interest in CNN-based methods for DNA foundation models. | DNA modeling, foundation model, Genomic Language Model, Representation Learning | null | 1,899 | 2502.18538 | [
-0.057851456105709076,
-0.128147691488266,
0.059027887880802155,
-0.038658544421195984,
-0.019410109147429466,
-0.02896723710000515,
-0.11123048514127731,
-0.008550744503736496,
0.029426319524645805,
-0.08662404865026474,
-0.04622162878513336,
-0.018904229626059532,
-0.055764324963092804,
0.06006315350532532,
-0.09766875207424164,
-0.0018917019478976727,
0.002061259699985385,
0.07786544412374496,
-0.1432061791419983,
-0.08885139971971512,
-0.00570843368768692,
-0.00501561863347888,
0.09343720972537994,
-0.02099505625665188,
0.08506450057029724,
0.06989619880914688,
-0.05128060281276703,
-0.06225191801786423,
0.006196162663400173,
-0.03033757582306862,
0.07776546478271484,
0.010614363476634026,
-0.04092350974678993,
0.014870701357722282,
-0.09802501648664474,
0.06581883132457733,
0.007063365541398525,
-0.027071619406342506,
0.0342896468937397,
-0.08649609982967377,
0.031606681644916534,
0.01129097305238247,
0.07106811553239822,
0.11278001219034195,
0.0638534352183342,
0.061059847474098206,
0.040453240275382996,
-0.057032857090234756,
0.052255742251873016,
0.0056671565398573875,
-0.002331170719116926,
-0.011328322812914848,
0.03394566476345062,
0.0842062383890152,
-0.020252585411071777,
0.0443560853600502,
0.023513058200478554,
-0.017468124628067017,
0.021030887961387634,
0.036045122891664505,
0.04515299201011658,
-0.06149011477828026,
-0.019634777680039406,
0.026848839595913887,
-0.02014629729092121,
-0.05467800423502922,
0.06640719622373581,
0.014137819409370422,
-0.06049054116010666,
-0.04995756596326828,
0.049350179731845856,
0.06661150604486465,
0.003685715841129422,
0.050671275705099106,
0.03887128829956055,
0.04712718352675438,
0.05550021678209305,
0.012167165987193584,
0.09226588904857635,
-0.05467601865530014,
0.027978286147117615,
-0.01455833949148655,
0.07052235305309296,
0.027776191011071205,
0.12677831947803497,
0.032383039593696594,
-0.035269517451524734,
0.07802868634462357,
-0.007067106664180756,
0.07788195461034775,
-0.04106501489877701,
-0.07974764704704285,
0.038462962955236435,
-0.08433369547128677,
-0.05925588682293892,
0.009562871418893337,
-0.051802027970552444,
-0.054771989583969116,
-0.05368712171912193,
0.03153432905673981,
-0.05027274042367935,
-0.05061834305524826,
0.08332400023937225,
-0.013039159588515759,
0.039858028292655945,
-0.00897114910185337,
0.04230911657214165,
0.0484195314347744,
0.013492448255419731,
0.01601588912308216,
0.015451452694833279,
0.06655480712652206,
0.04557827115058899,
0.03703751787543297,
0.04508084058761597,
-0.009525828063488007,
0.0033741644583642483,
-0.03184555098414421,
0.11748712509870529,
-0.0004938017809763551,
-0.06561524420976639,
0.02301318198442459,
-0.08509064465761185,
-0.04972558096051216,
-0.02815335988998413,
0.004109014291316271,
-0.06999831646680832,
2.8434929166482815e-33,
-0.005568171851336956,
0.012528353370726109,
0.03437599539756775,
0.04987684264779091,
0.03536440432071686,
0.005318683572113514,
-0.013628006912767887,
0.018856318667531013,
-0.1268351525068283,
0.004301645327359438,
-0.036078713834285736,
-0.01526679191738367,
-0.03536707162857056,
0.07531307637691498,
-0.017471415922045708,
-0.00786442682147026,
-0.04157793149352074,
0.03003864362835884,
-0.040238939225673676,
-0.06436122208833694,
-0.04773986339569092,
-0.014241877943277359,
0.01809343695640564,
-0.00569571740925312,
0.03348260372877121,
-0.0589958056807518,
-0.02440425008535385,
-0.03910987451672554,
-0.044735584408044815,
-0.03554259240627289,
-0.13327109813690186,
0.02827860228717327,
-0.018231255933642387,
0.05781535059213638,
0.04938896372914314,
-0.0362546406686306,
0.04946274310350418,
-0.0417524091899395,
0.01568945311009884,
0.006926120724529028,
-0.032019373029470444,
0.027067754417657852,
-0.008492044173181057,
-0.05446711182594299,
-0.07381974160671234,
0.0055145868100225925,
0.04552929848432541,
-0.05663421377539635,
-0.021747710183262825,
0.01189638115465641,
0.09220888465642929,
0.00140758219640702,
-0.04114479199051857,
-0.04404943436384201,
0.0336468368768692,
0.031037887558341026,
0.021756431087851524,
0.017499282956123352,
0.007925273850560188,
0.04850102216005325,
-0.05845518782734871,
0.05184543877840042,
-0.04397327080368996,
0.05463868007063866,
0.08678851276636124,
-0.023750977590680122,
-0.06520875543355942,
0.05121774971485138,
0.02367469295859337,
0.036254607141017914,
-0.02768009342253208,
-0.03297446295619011,
-0.009164305403828621,
-0.06814524531364441,
0.007184620480984449,
0.015881240367889404,
0.01520575862377882,
-0.08680108934640884,
-0.08618015795946121,
0.032074347138404846,
0.0037596120964735746,
0.0016159644583240151,
-0.021420352160930634,
0.021668264642357826,
-0.06642802059650421,
0.006409119814634323,
0.025829046964645386,
-0.06669594347476959,
0.0018330110469833016,
-0.03083695098757744,
-0.018001459538936615,
-0.06505746394395828,
0.057388145476579666,
-0.04927922785282135,
0.01396851520985365,
-3.750885291380822e-33,
-0.026030980050563812,
0.0067774155177176,
0.003842657897621393,
0.05123964324593544,
-0.03791864588856697,
-0.07507186383008957,
0.04909243807196617,
-0.020861148834228516,
0.016753990203142166,
-0.07334081828594208,
0.02693580649793148,
-0.010360083542764187,
-0.006545128766447306,
0.015110733918845654,
0.04619011655449867,
-0.057786229997873306,
-0.03684304282069206,
0.008612086065113544,
0.07114367932081223,
0.027644546702504158,
0.0030393374618142843,
0.15756334364414215,
-0.08845631778240204,
0.018281832337379456,
-0.05735374614596367,
-0.021934499964118004,
-0.0852896124124527,
0.04933539777994156,
0.0538569912314415,
-0.06323674321174622,
-0.121578648686409,
-0.0020292685367166996,
-0.023828964680433273,
-0.011947046034038067,
-0.010622581467032433,
0.08579541742801666,
0.051285844296216965,
0.009669453836977482,
-0.0005374051979742944,
-0.06238735467195511,
0.03612734377384186,
0.030943574383854866,
-0.07492314279079437,
0.0357687845826149,
-0.03155110776424408,
0.10786107182502747,
-0.04365290328860283,
0.08661993592977524,
-0.08046426624059677,
0.031646519899368286,
-0.001804049126803875,
0.038336288183927536,
-0.11056482046842575,
0.007455312181264162,
-0.027586905285716057,
-0.07649809122085571,
0.019833698868751526,
-0.05196675285696983,
0.075624480843544,
-0.007766872178763151,
-0.07465667277574539,
-0.05484161153435707,
-0.02360231988132,
-0.0925009474158287,
-0.04299408942461014,
0.043139249086380005,
-0.03925904631614685,
0.06084826588630676,
0.02503209561109543,
0.04899203032255173,
-0.03744884580373764,
-0.03287530317902565,
0.08927837014198303,
0.07352744787931442,
-0.007175746839493513,
-0.02735542692244053,
0.02071545273065567,
-0.11063969880342484,
-0.021903740242123604,
-0.018673567101359367,
0.030847979709506035,
-0.058218903839588165,
-0.008750200271606445,
0.002500481903553009,
0.15094667673110962,
0.1144719272851944,
-0.02303704433143139,
0.03964655473828316,
0.05157260596752167,
-0.00528872711583972,
-0.05225995182991028,
0.04742312431335449,
0.01460350677371025,
0.07318998128175735,
-0.018838198855519295,
-4.8042785749657924e-8,
0.0286649651825428,
0.0336814746260643,
-0.018834145739674568,
-0.04036149010062218,
0.0001292249362450093,
-0.07393281161785126,
0.029499826952815056,
0.09672243893146515,
0.03602565824985504,
-0.04614110291004181,
0.009698635898530483,
0.04660182073712349,
-0.04987489432096481,
-0.04642375186085701,
-0.00918775424361229,
0.05128808319568634,
0.029906053096055984,
0.041914720088243484,
0.02059187740087509,
0.0070941331796348095,
0.029589006677269936,
0.04842018336057663,
-0.07041911035776138,
0.012062260881066322,
0.0160942692309618,
-0.023714452981948853,
0.016027668491005898,
0.09447437524795532,
0.045377444475889206,
0.00572201469913125,
0.05162836238741875,
0.02866504155099392,
0.02549739181995392,
-0.03640079125761986,
0.012205289676785469,
-0.0036523404996842146,
0.010352341458201408,
-0.0005975281819701195,
0.03146712854504585,
0.01252006459981203,
0.06079733744263649,
0.10580207407474518,
-0.09319496154785156,
0.014030334539711475,
-0.07716435194015503,
0.013679961673915386,
0.05242302641272545,
-0.06594938784837723,
0.0737246423959732,
-0.01179984025657177,
0.0433625802397728,
-0.0028792412485927343,
-0.031201135367155075,
0.025926582515239716,
0.012496127746999264,
0.025273501873016357,
0.011670819483697414,
-0.040849413722753525,
-0.0314754880964756,
0.05538049712777138,
0.021774668246507645,
0.027802716940641403,
-0.004565694835036993,
-0.08180610835552216
] |
Kronecker Mask and Interpretive Prompts are Language-Action Video Learners | https://openreview.net/forum?id=RUF7j1cJzK | [
"Yang JingYi",
"Zitong YU",
"Nixiuming",
"He Jia",
"Hui Li"
] | Poster | Contrastive language-image pretraining (CLIP) has significantly advanced image-based vision learning. A pressing topic subsequently arises: how can we effectively adapt CLIP to the video domain? Recent studies have focused on adjusting either the textual or visual branch of CLIP for action recognition. However, we argue that adaptations of both branches are crucial. In this paper, we propose a **C**ontrastive **L**anguage-**A**ction **V**ideo Learn**er** (**CLAVER**), designed to shift CLIP's focus from the alignment of static visual objects and concrete nouns to the alignment of dynamic action behaviors and abstract verbs. Specifically, we introduce a novel Kronecker mask attention for temporal modeling. Our tailored Kronecker mask offers three benefits 1) it expands the temporal receptive field for each token, 2) it serves as an effective spatiotemporal heterogeneity inductive bias, mitigating the issue of spatiotemporal homogenization, and 3) it can be seamlessly plugged into transformer-based models. Regarding the textual branch, we leverage large language models to generate diverse, sentence-level and semantically rich interpretive prompts of actions, which shift the model's focus towards the verb comprehension. Extensive experiments on various benchmarks and learning scenarios demonstrate the superiority and generality of our approach. The code will be available soon. | Action Recognition, Video Recognition, Spatiotemporal Modeling | null | 1,897 | 2502.03549 | [
0.00005481028347276151,
-0.1158708930015564,
0.009142827242612839,
-0.0726127102971077,
0.09994897991418839,
0.0828540250658989,
-0.0024970665108412504,
-0.0440739244222641,
0.08370883017778397,
-0.06163272261619568,
-0.02315806970000267,
-0.06479665637016296,
-0.062154896557331085,
0.11784298717975616,
-0.010214747861027718,
-0.01983925700187683,
0.01839928701519966,
0.14969022572040558,
-0.057147663086652756,
-0.034473467618227005,
0.03615689277648926,
-0.0518588125705719,
0.0021347738802433014,
-0.01414914708584547,
0.025639686733484268,
0.07291034609079361,
0.020738547667860985,
-0.07644280046224594,
0.02100997045636177,
-0.03305577114224434,
-0.003775994060561061,
-0.028436221182346344,
0.09757458418607712,
0.03465437889099121,
-0.10717161744832993,
0.08069848269224167,
-0.07797561585903168,
0.0035861055366694927,
-0.0361446850001812,
-0.04171795770525932,
-0.029002057388424873,
0.017901115119457245,
0.021859372034668922,
-0.00889788381755352,
0.018523726612329483,
0.008480207063257694,
0.03631964325904846,
-0.018127674236893654,
-0.03521274775266647,
-0.006358928047120571,
-0.03871522843837738,
-0.0922270119190216,
-0.004564798902720213,
0.035207394510507584,
-0.10595140606164932,
-0.014183119870722294,
-0.006247124634683132,
0.04166003316640854,
0.07353843003511429,
-0.036775123327970505,
0.012651712633669376,
-0.06497357785701752,
0.05488426983356476,
0.0027828312013298273,
-0.06487011164426804,
-0.008697174489498138,
0.013938113115727901,
-0.03845338150858879,
-0.045115672051906586,
0.040157631039619446,
-0.059683140367269516,
0.055341389030218124,
-0.008691159076988697,
0.021971382200717926,
-0.03976818174123764,
0.024478068575263023,
0.056444358080625534,
-0.07027079164981842,
0.061673663556575775,
-0.100075863301754,
0.10331279784440994,
0.0044684745371341705,
0.037968963384628296,
-0.030964523553848267,
0.12854823470115662,
0.005471975542604923,
-0.04929032176733017,
0.02802020125091076,
-0.018916670233011246,
-0.03936176747083664,
-0.058817923069000244,
-0.07970047742128372,
0.053728796541690826,
-0.024360766634345055,
0.05720636248588562,
0.000051106300816172734,
-0.053364347666502,
-0.0039316872134804726,
0.00929564330726862,
0.07825817167758942,
0.03133595734834671,
0.021081725135445595,
-0.03534059226512909,
-0.0049894689582288265,
-0.008306628093123436,
-0.02609679475426674,
0.03646368533372879,
-0.04854172468185425,
0.0020851308945566416,
-0.04748031869530678,
0.014006605371832848,
0.02184106595814228,
0.013726183213293552,
-0.005728248506784439,
0.031962800770998,
-0.006434231530874968,
-0.013791951350867748,
-0.06215715780854225,
0.0013607790460810065,
0.019297005608677864,
0.07173845916986465,
0.005742620676755905,
-0.046939779072999954,
0.053767167031764984,
0.0697140321135521,
-0.04926476255059242,
0.004548696801066399,
1.021734271440459e-33,
0.049433525651693344,
0.02121281996369362,
0.03287699073553085,
0.013572725467383862,
0.03297555074095726,
-0.031696517020463943,
0.004313380923122168,
-0.03134917840361595,
-0.013410774059593678,
-0.039683979004621506,
0.015804477035999298,
-0.006105051841586828,
-0.02817522920668125,
0.1007498949766159,
-0.029201265424489975,
-0.02102244272828102,
-0.02105911262333393,
0.0008073802455328405,
-0.018880313262343407,
0.03647241368889809,
-0.01623133197426796,
0.011707098223268986,
-0.011209994554519653,
-0.0009489969816058874,
-0.02051481604576111,
-0.04138148948550224,
0.07472701370716095,
-0.09961389750242233,
0.004440950229763985,
0.009042897261679173,
-0.11191481351852417,
-0.036853447556495667,
-0.04560903087258339,
0.031003989279270172,
0.046340085566043854,
-0.06563406437635422,
-0.00780663825571537,
-0.03614826127886772,
0.03089301474392414,
-0.06676503270864487,
-0.031284309923648834,
0.07225493341684341,
-0.07790202647447586,
-0.013491938821971416,
-0.028218388557434082,
-0.026801759377121925,
0.07893391698598862,
0.12524031102657318,
-0.0650881752371788,
0.011100775562226772,
0.04663807898759842,
-0.016245532780885696,
-0.02608095481991768,
-0.0909571722149849,
-0.0037263871636241674,
0.05249680206179619,
0.06866142153739929,
0.02025928720831871,
0.0051441253162920475,
-0.021090013906359673,
0.014150632545351982,
0.07194394618272781,
0.04763340950012207,
0.0682235136628151,
0.01625008136034012,
0.06405122578144073,
0.0056584542617201805,
0.07974308729171753,
0.047308992594480515,
0.024718554690480232,
-0.018522124737501144,
0.030454151332378387,
-0.04566952586174011,
-0.03361906111240387,
-0.002859964035451412,
0.015417926013469696,
0.07966084033250809,
-0.12160354107618332,
-0.03517673909664154,
0.05516515672206879,
-0.06018700450658798,
-0.092534139752388,
0.06224636361002922,
-0.08407030999660492,
-0.08451267331838608,
-0.04618314281105995,
0.080584816634655,
-0.11154362559318542,
0.015713458880782127,
-0.0466272346675396,
0.006649188231676817,
0.0072057219222188,
-0.01837698556482792,
0.04079611599445343,
0.08320318162441254,
-1.0553166756755132e-33,
0.047173161059617996,
0.05650842562317848,
-0.03809986636042595,
0.03612145408987999,
-0.0476769395172596,
-0.013921461999416351,
0.028458857908844948,
0.020345274358987808,
0.0038916082121431828,
-0.03722384199500084,
-0.07748265564441681,
-0.038178831338882446,
-0.10454787313938141,
-0.0250877533107996,
0.0479368232190609,
0.013657812029123306,
0.028254812583327293,
0.10218599438667297,
-0.043183792382478714,
-0.05687351152300835,
-0.028159189969301224,
0.06133226677775383,
-0.12648659944534302,
-0.0012670486466959119,
-0.10246005654335022,
0.08620923012495041,
0.0050693051889538765,
0.07526667416095734,
-0.029052119702100754,
0.01897885836660862,
-0.04258598014712334,
-0.03893902152776718,
-0.024270715191960335,
0.029103701934218407,
-0.06557578593492508,
0.07981695979833603,
0.02407258003950119,
-0.055560968816280365,
-0.009747241623699665,
0.04263285920023918,
0.0558193139731884,
-0.03902168571949005,
-0.0461561493575573,
0.036849893629550934,
-0.05860172212123871,
-0.03792784363031387,
-0.028512118384242058,
0.031443919986486435,
-0.026332853361964226,
-0.014335649088025093,
-0.042122095823287964,
0.03132309019565582,
-0.09276575595140457,
-0.025269141420722008,
-0.06820821017026901,
0.009443945251405239,
-0.02966351807117462,
-0.017036737874150276,
0.024943355470895767,
-0.02174307219684124,
0.019145628437399864,
0.006036906037479639,
-0.027072690427303314,
-0.05253587290644646,
0.044613271951675415,
-0.04939598590135574,
-0.037714842706918716,
-0.011331922374665737,
0.05073748528957367,
0.0020967177115380764,
0.08608003705739975,
-0.03180202841758728,
-0.05106702446937561,
0.004020552150905132,
0.03524814546108246,
-0.011942588724195957,
0.03851563483476639,
0.0015315324999392033,
-0.014188705012202263,
-0.009607408195734024,
-0.03947169706225395,
-0.05652464181184769,
0.012709720991551876,
0.06196226924657822,
0.01928364858031273,
0.15604980289936066,
-0.013134145177900791,
0.05973018705844879,
0.0554017499089241,
0.03626158833503723,
-0.05682714283466339,
0.05602056160569191,
0.00480258371680975,
0.0863199532032013,
-0.011001327075064182,
-4.459164060222065e-8,
-0.09271971136331558,
0.03855861723423004,
-0.04445790871977806,
0.0030044137965887785,
-0.015879441052675247,
-0.06978865712881088,
-0.015865866094827652,
0.018424557521939278,
0.02530236542224884,
-0.026817670091986656,
0.017782995477318764,
0.0006721293902955949,
0.05126778408885002,
-0.05195529758930206,
-0.04281137138605118,
0.058480288833379745,
0.04137579724192619,
0.047219615429639816,
-0.023244747892022133,
0.09749842435121536,
0.06506004929542542,
0.04695920646190643,
-0.01766112819314003,
-0.05058535933494568,
-0.07161501795053482,
-0.038249995559453964,
-0.034231480211019516,
0.12316633015871048,
0.06933590024709702,
0.019505983218550682,
0.04325889050960541,
0.09480606764554977,
0.016156744211912155,
-0.08107573539018631,
-0.022801514714956284,
0.06563921272754669,
0.024571647867560387,
-0.031156688928604126,
-0.004468151833862066,
0.016783954575657845,
0.023938916623592377,
-0.019671354442834854,
0.034139808267354965,
0.03209849074482918,
-0.08678644895553589,
0.044916536659002304,
-0.018300004303455353,
-0.1326049566268921,
0.04657954350113869,
-0.025351984426379204,
0.020637495443224907,
-0.018600763753056526,
-0.0205041766166687,
0.152075856924057,
0.09239688515663147,
-0.013085170648992062,
0.029134243726730347,
0.009404185228049755,
0.040632519870996475,
0.04084116220474243,
0.0013934321468695998,
0.06002743914723396,
-0.028842274099588394,
-0.0027132502291351557
] |
MME-RealWorld: Could Your Multimodal LLM Challenge High-Resolution Real-World Scenarios that are Difficult for Humans? | https://openreview.net/forum?id=k5VHHgsRbi | [
"YiFan Zhang",
"Huanyu Zhang",
"Haochen Tian",
"Chaoyou Fu",
"Shuangqing Zhang",
"Junfei Wu",
"Feng Li",
"Kun Wang",
"Qingsong Wen",
"Zhang Zhang",
"Liang Wang",
"Rong Jin"
] | Poster | Comprehensive evaluation of Multimodal Large Language Models (MLLMs) has recently garnered widespread attention in the research community. However, we observe that existing benchmarks present several common barriers that make it difficult to measure the significant challenges that models face in the real world, including: 1) small data scale leads to a large performance variance; 2) reliance on model-based annotations results in restricted data quality; 3) insufficient task difficulty, especially caused by the limited image resolution. To tackle these issues, we introduce MME-RealWorld. Specifically, we collect more than $300$ K images from public datasets and the Internet, filtering $13,366$ high-quality images for annotation. This involves the efforts of professional $25$ annotators and $7$ experts in MLLMs, contributing to $29,429$ question-answer pairs that cover $43$ subtasks across $5$ real-world scenarios, extremely challenging even for humans. As far as we know, **MME-RealWorld is the largest manually annotated benchmark to date, featuring the highest resolution and a targeted focus on real-world applications**. We further conduct a thorough evaluation involving $29$ prominent MLLMs, such as GPT-4o, Gemini 1.5 Pro, and Claude 3.5 Sonnet. Our results show that even the most advanced models struggle with our benchmarks, where none of them reach 60\% accuracy. The challenges of perceiving high-resolution images and understanding complex real-world scenarios remain urgent issues to be addressed. The data and evaluation code are released in our Project Page. | multimodal Large Language Models, benchmark, high-resolution images, real-world scenarios | We introduce MME-RealWorld, the largest manually annotated benchmark for evaluating Multimodal Large Language Models, featuring over 29,000 question-answer pairs and high-resolution images to address significant challenges in real-world scenarios. | 1,887 | null | [
-0.011806030757725239,
-0.0773201733827591,
0.018474746495485306,
0.017643922939896584,
0.11062540113925934,
-0.03821597248315811,
-0.031928665935993195,
0.014524633064866066,
-0.003477211808785796,
0.016827093437314034,
-0.07202786207199097,
-0.1346755027770996,
0.01453410368412733,
0.09070378541946411,
-0.01361869741231203,
-0.0277412086725235,
0.07362396270036697,
0.0321442186832428,
-0.11704972386360168,
-0.005459531210362911,
0.047404248267412186,
0.0027819774113595486,
0.02089857868850231,
-0.038350027054548264,
0.08785821497440338,
-0.03331021964550018,
-0.0015081041492521763,
-0.10003992915153503,
-0.01044553890824318,
-0.06218690797686577,
0.00207610079087317,
0.029304513707756996,
0.006689031608402729,
0.007589802145957947,
0.020959842950105667,
0.04652569070458412,
-0.02957952953875065,
-0.00868377648293972,
-0.01130523532629013,
-0.06637199223041534,
-0.027318574488162994,
-0.05298237502574921,
0.08133763074874878,
-0.06645986437797546,
0.07987692207098007,
-0.008775090798735619,
0.022429144009947777,
0.007170260418206453,
0.040133293718099594,
0.0080941803753376,
-0.16163219511508942,
-0.11534111201763153,
0.006273353472352028,
0.04842957481741905,
-0.03713775426149368,
-0.05843041464686394,
0.035905834287405014,
-0.0022328482009470463,
0.047366913408041,
0.022779248654842377,
-0.01737900823354721,
-0.08710700273513794,
-0.03588374704122543,
0.04059910401701927,
0.05638604611158371,
-0.04125766456127167,
0.014590045437216759,
-0.03126480430364609,
0.05952718108892441,
-0.02006249874830246,
-0.025789815932512283,
0.0246245376765728,
-0.029830876737833023,
0.060290105640888214,
0.05787855014204979,
0.03429170697927475,
0.003216278040781617,
-0.018784429877996445,
0.10210885852575302,
-0.047736234962940216,
0.09719870239496231,
-0.04053657129406929,
0.03834901005029678,
-0.021832656115293503,
0.09222899377346039,
-0.012963205575942993,
-0.0059651946648955345,
0.062315188348293304,
0.03337281942367554,
-0.0004993477487005293,
-0.07858441770076752,
-0.05411045253276825,
-0.020618697628378868,
0.044365208595991135,
0.009435772895812988,
-0.02017125114798546,
0.037921980023384094,
-0.04581639543175697,
-0.022870980203151703,
0.0925149992108345,
0.04648490622639656,
0.04903820902109146,
-0.009208220057189465,
-0.012097136117517948,
0.07957904040813446,
0.004012683872133493,
0.08717241883277893,
0.00927787460386753,
-0.022345006465911865,
-0.06843219697475433,
-0.01762741059064865,
-0.0008236235589720309,
0.03344890475273132,
-0.04103652387857437,
0.06608103960752487,
-0.07390986382961273,
0.0714692696928978,
-0.0008787119877524674,
0.04457360506057739,
-0.03096236102283001,
-0.0696663111448288,
0.005865040235221386,
0.026005703955888748,
-0.06834369152784348,
0.120059534907341,
0.05613205209374428,
-0.04791538789868355,
4.392368721671868e-33,
0.016479438170790672,
0.0625782385468483,
0.07686605304479599,
0.026947787031531334,
0.02803836390376091,
-0.06257235258817673,
-0.07034529745578766,
0.01748322695493698,
0.0027717372868210077,
-0.027440989390015602,
0.0317271463572979,
0.01595848612487316,
-0.062266021966934204,
0.07183774560689926,
0.06740141659975052,
0.004567896481603384,
-0.04745969921350479,
0.0060911355540156364,
-0.05061466246843338,
0.07676365226507187,
0.0225222148001194,
-0.069697804749012,
0.023444602265954018,
-0.007392617408186197,
0.023461010307073593,
0.021869104355573654,
0.07891832292079926,
-0.0738288164138794,
0.018045637756586075,
0.011908327229321003,
-0.11973390728235245,
-0.047282200306653976,
0.02975410781800747,
0.01867121458053589,
0.023135308176279068,
-0.030776599422097206,
-0.028526943176984787,
-0.0698554590344429,
-0.031290918588638306,
0.03470032289624214,
-0.05244896560907364,
0.031238840892910957,
-0.034917693585157394,
-0.010094241239130497,
-0.041839394718408585,
0.01294984482228756,
0.0035262713208794594,
0.042130082845687866,
-0.0831415206193924,
-0.009906156919896603,
0.031101927161216736,
0.05480637773871422,
-0.05002439022064209,
-0.03982453793287277,
-0.03330056369304657,
0.018175682052969933,
0.04316962882876396,
-0.01874048262834549,
0.053700704127550125,
0.023563481867313385,
-0.029760288074612617,
-0.005625480320304632,
0.052225176244974136,
0.10490714758634567,
0.04929583519697189,
-0.003319020150229335,
0.007739399094134569,
0.049689579755067825,
0.033682383596897125,
0.00582195445895195,
-0.041396141052246094,
0.0022443379275500774,
0.061385877430438995,
-0.09003214538097382,
0.05488777160644531,
-0.003138734959065914,
0.09869039803743362,
-0.047570426017045975,
0.026988815516233444,
0.054602429270744324,
-0.01749148964881897,
0.016444751992821693,
0.021013857796788216,
-0.059973664581775665,
-0.01031977403908968,
-0.04635288938879967,
-0.01025121845304966,
-0.10824960470199585,
-0.016839997842907906,
-0.03812756761908531,
-0.028129730373620987,
-0.014158819802105427,
-0.04822150990366936,
-0.008780235424637794,
-0.01935720629990101,
-1.9684439016366317e-33,
0.004376390483230352,
0.001996274571865797,
-0.025222059339284897,
0.05152638629078865,
0.0007479701889678836,
-0.07285070419311523,
-0.006516859866678715,
0.05357499048113823,
0.027625475078821182,
-0.055241163820028305,
-0.017300933599472046,
-0.039340540766716,
0.03032119944691658,
0.020758045837283134,
0.024915378540754318,
-0.10531778633594513,
0.031916458159685135,
-0.06023285165429115,
0.025397196412086487,
0.04563185200095177,
0.006633387878537178,
0.07792910933494568,
-0.08035746216773987,
0.06932766735553741,
-0.0627061054110527,
0.09748341888189316,
-0.08316484838724136,
-0.026689568534493446,
-0.04514123126864433,
-0.04229044169187546,
0.015879733487963676,
-0.03443186357617378,
-0.012367291375994682,
-0.05300791934132576,
-0.00020502996630966663,
0.06555131822824478,
0.07191842794418335,
-0.01807824894785881,
-0.01431224588304758,
0.12011095881462097,
0.08395878225564957,
-0.004185968544334173,
-0.15942983329296112,
-0.03802253678441048,
-0.040059421211481094,
-0.014552373439073563,
-0.10033652186393738,
-0.06838299334049225,
0.00020875905465800315,
0.05107630044221878,
-0.04430198296904564,
-0.0002763458760455251,
-0.06778262555599213,
-0.01655668392777443,
-0.013290089555084705,
-0.09994909167289734,
-0.1051291897892952,
-0.013883844017982483,
0.06551722437143326,
0.08362417668104172,
-0.07637288421392441,
-0.04647102952003479,
-0.05079791322350502,
0.02397015132009983,
-0.0099935382604599,
-0.021546460688114166,
-0.019464952871203423,
0.015638120472431183,
-0.0585283525288105,
0.03327746316790581,
-0.0026395712047815323,
-0.049893200397491455,
-0.018523775041103363,
0.04306412115693092,
-0.005832288879901171,
0.048606570810079575,
0.018600981682538986,
0.05203058943152428,
0.08640453219413757,
-0.04815177246928215,
0.045651473104953766,
-0.0651748925447464,
0.039598673582077026,
0.0550016388297081,
0.10493739694356918,
0.09259385615587234,
-0.0018106657080352306,
0.02566010132431984,
0.030029170215129852,
0.038153789937496185,
-0.015292219817638397,
0.005502880550920963,
-0.0078598503023386,
0.059154149144887924,
0.0009201224893331528,
-4.49524755197217e-8,
-0.009165408089756966,
0.05531885847449303,
-0.009667414240539074,
-0.04368332028388977,
0.005672084167599678,
-0.006659127771854401,
-0.08302484452724457,
0.11033100634813309,
0.007240550592541695,
0.08534033596515656,
0.11189030110836029,
-0.05342939868569374,
-0.04348655045032501,
0.02556123211979866,
0.008047266863286495,
0.07711231708526611,
-0.03964777663350105,
0.042142532765865326,
-0.025851791724562645,
-0.0066658491268754005,
0.048975106328725815,
0.05862302705645561,
0.0039977277629077435,
-0.07976311445236206,
0.021656904369592667,
-0.023405596613883972,
0.025155693292617798,
0.09155476838350296,
-0.01515381783246994,
0.01254668552428484,
-0.011129594407975674,
0.024972153827548027,
-0.02573285438120365,
-0.0062688784673810005,
0.016008328646421432,
0.0013689560582861304,
-0.018994497135281563,
0.020687228068709373,
-0.026989439502358437,
-0.013439108617603779,
0.0007185729336924851,
0.0917249396443367,
-0.041246186941862106,
-0.00909661129117012,
0.06606798619031906,
0.041952166706323624,
-0.06421516835689545,
-0.054296717047691345,
-0.004714000504463911,
-0.05253603681921959,
-0.06053541973233223,
-0.01907486841082573,
0.023963872343301773,
0.08596039563417435,
0.09425199031829834,
0.021713919937610626,
0.051963794976472855,
0.047714490443468094,
0.0237420741468668,
0.09798963367938995,
0.06410905718803406,
-0.05284614861011505,
-0.10630164295434952,
0.02679932489991188
] |
State Space Model Meets Transformer: A New Paradigm for 3D Object Detection | https://openreview.net/forum?id=Tisu1L0Jwt | [
"ChuXin Wang",
"Wenfei Yang",
"Xiang Liu",
"Tianzhu Zhang"
] | Poster | DETR-based methods, which use multi-layer transformer decoders to refine object queries iteratively, have shown promising performance in 3D indoor object detection. However, the scene point features in the transformer decoder remain fixed, leading to minimal contributions from later decoder layers, thereby limiting performance improvement. Recently, State Space Models (SSM) have shown efficient context modeling ability with linear complexity through iterative interactions between system states and inputs. Inspired by SSMs, we propose a new 3D object DEtection paradigm with an interactive STate space model (DEST). In the interactive SSM, we design a novel state-dependent SSM parameterization method that enables system states to effectively serve as queries in 3D indoor detection tasks. In addition, we introduce four key designs tailored to the characteristics of point cloud and SSM: The serialization and bidirectional scanning strategies enable bidirectional feature interaction among scene points within the SSM. The inter-state attention mechanism models the relationships between state points, while the gated feed-forward network enhances inter-channel correlations. To the best of our knowledge, this is the first method to model queries as system states and scene points as system inputs, which can simultaneously update scene point features and query features with linear complexity. Extensive experiments on two challenging datasets demonstrate the effectiveness of our DEST-based method. Our method improves the GroupFree baseline in terms of $\text{AP}_{50}$ on ScanNet V2 (+5.3) and SUN RGB-D (+3.2) datasets. Based on the VDETR baseline, Our method sets a new state-of-the-art on the ScanNetV2 and SUN RGB-D datasets. | Point Cloud; 3D Object Detection; State Space Model | This study introduces a novel paradigm, DEST, which significantly enhances 3D object detection performance on two challenging datasets through the proposed interactive state space model. | 1,881 | 2503.14493 | [
-0.016516583040356636,
-0.08827784657478333,
-0.003632995067164302,
0.027133630588650703,
0.04876967519521713,
-0.013069987297058105,
0.01586490124464035,
-0.0768745169043541,
0.036605510860681534,
-0.03843140974640846,
-0.0022679243702441454,
-0.014111964032053947,
0.036144305020570755,
0.08123742043972015,
0.03225282207131386,
-0.03354061767458916,
0.11521948873996735,
0.04044577479362488,
-0.01005857065320015,
0.05002889037132263,
0.05370549112558365,
0.021463755518198013,
-0.039806630462408066,
-0.011692296713590622,
0.024918628856539726,
0.025368154048919678,
0.07646310329437256,
-0.10187123715877533,
-0.018790462985634804,
0.014841238968074322,
0.03708968311548233,
-0.039718419313430786,
-0.040554311126470566,
0.05258160084486008,
-0.03908563777804375,
-0.09286521375179291,
0.05775519087910652,
-0.09480348974466324,
-0.06959673762321472,
-0.0544453002512455,
0.03532801941037178,
0.006210312247276306,
0.02235301397740841,
-0.010068017058074474,
0.030273258686065674,
0.0448959618806839,
0.0037632142193615437,
-0.03367903456091881,
0.08250579237937927,
-0.07229232043027878,
-0.00662480853497982,
0.03483325615525246,
0.028893746435642242,
0.11832739412784576,
-0.03833358734846115,
0.017255639657378197,
0.03996876999735832,
0.04775727540254593,
-0.007303937338292599,
-0.026599816977977753,
-0.05118824169039726,
-0.05878068879246712,
0.07404758036136627,
0.002710073022171855,
0.0011562794679775834,
-0.010090094059705734,
0.07512172311544418,
-0.0020599470008164644,
-0.02198054827749729,
-0.03996088728308678,
-0.037547752261161804,
0.053950611501932144,
0.0032068691216409206,
-0.021867599338293076,
0.03716270253062248,
-0.014676794409751892,
0.01097835786640644,
-0.061677463352680206,
0.010610713623464108,
-0.03372162953019142,
0.03583968058228493,
-0.008669666945934296,
0.007904591038823128,
0.03489266708493233,
0.05558299645781517,
0.017887737601995468,
-0.1076752170920372,
0.03181764855980873,
0.004772180691361427,
0.018603289499878883,
-0.08520683646202087,
-0.12045057117938995,
-0.07569897919893265,
-0.03620234876871109,
0.011089645326137543,
-0.035689227283000946,
-0.0028714805375784636,
-0.03659575805068016,
0.004918181337416172,
0.020887574180960655,
-0.024892179295420647,
0.005566195584833622,
0.07284118235111237,
-0.0839979499578476,
0.0779094398021698,
0.023189079016447067,
0.006304615642875433,
0.0412789024412632,
0.030080435797572136,
0.047437798231840134,
0.014642508700489998,
0.011632471345365047,
0.036933571100234985,
0.017924433574080467,
0.020089853554964066,
0.02282986417412758,
0.05207071453332901,
-0.06008319929242134,
0.04423327371478081,
-0.03253190964460373,
0.029907101765275,
-0.03529326245188713,
-0.03610967844724655,
0.011161982081830502,
0.0007417556480504572,
0.004196280147880316,
-0.07123300433158875,
-1.2815166862622614e-33,
-0.05553140491247177,
-0.001728778355754912,
0.0056382049806416035,
0.04698292165994644,
0.005644019227474928,
0.004671678878366947,
0.08699691295623779,
0.13209180533885956,
0.008527741767466068,
0.02384684421122074,
-0.07017470896244049,
-0.025007836520671844,
-0.0001874275621958077,
0.038850415498018265,
0.044185612350702286,
-0.02480359748005867,
-0.030660046264529228,
0.003763079410418868,
-0.010352076962590218,
0.048685021698474884,
0.015587318688631058,
-0.08942890167236328,
-0.0113833574578166,
-0.014338248409330845,
0.03448764607310295,
0.008800052106380463,
0.03461615368723869,
0.037822213023900986,
-0.013109921477735043,
-0.03588799759745598,
-0.04899825528264046,
0.040655795484781265,
-0.03698456659913063,
-0.015982020646333694,
0.05991789698600769,
-0.06820718944072723,
-0.032138027250766754,
0.006379564758390188,
-0.027795784175395966,
-0.06796717643737793,
0.03766763210296631,
-0.012432661838829517,
-0.06220991536974907,
-0.04652722552418709,
-0.06281393766403198,
-0.04609398543834686,
0.027701208367943764,
-0.04189838096499443,
-0.05350939556956291,
0.0676097720861435,
0.11372263729572296,
0.048916079103946686,
-0.1123967319726944,
-0.026703057810664177,
0.05301860719919205,
-0.05780378356575966,
0.039632365107536316,
0.03080420009791851,
0.06450152397155762,
0.05642582103610039,
-0.12080013006925583,
0.0005791631992906332,
0.008312340825796127,
0.06777428835630417,
0.006724183913320303,
-0.046991974115371704,
-0.02555040828883648,
0.11533357203006744,
0.012915374711155891,
0.055050548166036606,
-0.01783568225800991,
0.005267455242574215,
0.016845807433128357,
-0.015217836014926434,
0.07520151138305664,
-0.01901467703282833,
-0.027718879282474518,
0.019058382138609886,
-0.001536860945634544,
0.028897497802972794,
-0.02277289517223835,
-0.03900007903575897,
-0.030235428363084793,
0.07236076891422272,
-0.04735388234257698,
0.010356822982430458,
0.015611660666763783,
-0.08630038052797318,
-0.056503068655729294,
-0.023721791803836823,
0.015804937109351158,
-0.03451528027653694,
-0.03950438275933266,
0.01888732612133026,
-0.03237234055995941,
4.6523628173123145e-34,
0.08734768629074097,
-0.045041266828775406,
-0.032122474163770676,
-0.021349214017391205,
-0.04393337294459343,
-0.12463411688804626,
-0.027340294793248177,
0.048408977687358856,
0.03962322697043419,
-0.05747263506054878,
-0.033347442746162415,
-0.014276767149567604,
0.1012347862124443,
0.09685905277729034,
0.08380245417356491,
0.11775245517492294,
-0.04429777339100838,
-0.05118066444993019,
-0.053391601890325546,
-0.0049195438623428345,
0.04202314466238022,
0.09322912991046906,
-0.08204919099807739,
-0.05702035874128342,
-0.13913340866565704,
0.05549507588148117,
-0.025302017107605934,
0.030539413914084435,
0.07761043310165405,
0.027156557887792587,
-0.019054878503084183,
-0.015375119633972645,
0.05475848168134689,
-0.0012388918548822403,
-0.053434353321790695,
0.03527456149458885,
0.03768635168671608,
-0.028326239436864853,
-0.021721452474594116,
-0.008798542432487011,
0.05407913401722908,
-0.06924069672822952,
-0.023995814844965935,
0.11040818691253662,
-0.0016182682011276484,
0.012058636173605919,
-0.014990860596299171,
0.033712927252054214,
-0.07344396412372589,
0.05000373348593712,
-0.0730988159775734,
0.023941947147250175,
-0.09478192776441574,
0.02861529402434826,
-0.0034752769861370325,
-0.018354233354330063,
-0.07267091423273087,
0.029003070667386055,
0.09330806881189346,
-0.01331534143537283,
0.038589850068092346,
-0.025647753849625587,
-0.033100713044404984,
-0.03662073612213135,
-0.030776817351579666,
-0.04659464210271835,
-0.007993844337761402,
-0.06890927255153656,
0.01495886966586113,
0.07647546380758286,
0.021487008780241013,
0.010656802915036678,
0.00027489414787851274,
0.01112425234168768,
0.018514228984713554,
-0.02236936427652836,
0.04823967069387436,
-0.044682297855615616,
0.06731181591749191,
-0.03859555348753929,
-0.0786384642124176,
-0.032509613782167435,
-0.06301457434892654,
0.0644279271364212,
0.05381852760910988,
0.09524030983448029,
0.005638515576720238,
0.02940693497657776,
-0.03882458060979843,
-0.022716563194990158,
-0.07301610708236694,
0.07450146973133087,
-0.02509864792227745,
0.032650452107191086,
0.016225852072238922,
-4.210879112065413e-8,
-0.15723776817321777,
0.08706130087375641,
-0.011784878559410572,
-0.009537504985928535,
-0.05427594482898712,
-0.02021697722375393,
0.03142558038234711,
0.08676057308912277,
0.004498636815696955,
-0.034781429916620255,
0.0410153865814209,
-0.08256042748689651,
-0.013470472767949104,
-0.03984381631016731,
0.07449132204055786,
0.044671908020973206,
-0.014558074995875359,
-0.02892153337597847,
-0.0009086115169338882,
0.0481191910803318,
0.03223384544253349,
0.03364250436425209,
0.01701737940311432,
0.07352039963006973,
0.031154222786426544,
0.027162600308656693,
-0.016343530267477036,
0.10880665481090546,
0.0757640078663826,
-0.018036378547549248,
0.03694862127304077,
0.04104325920343399,
-0.022043883800506592,
0.05809721350669861,
-0.021102601662278175,
0.025421608239412308,
0.028436215594410896,
-0.05306367948651314,
0.08644309639930725,
-0.010828173719346523,
-0.010693219490349293,
-0.0039018287789076567,
-0.1334320455789566,
0.02555164135992527,
-0.07867544889450073,
0.08870913833379745,
0.11353889852762222,
-0.07034625113010406,
-0.024701246991753578,
-0.026184728369116783,
-0.00858920719474554,
-0.04747212678194046,
-0.06555881351232529,
0.0647515133023262,
-0.027054451406002045,
0.057639140635728836,
0.0628068745136261,
-0.06397642195224762,
0.0629052072763443,
0.09199836105108261,
-0.026268865913152695,
0.011313830502331257,
-0.12800371646881104,
0.04798617213964462
] |
What to align in multimodal contrastive learning? | https://openreview.net/forum?id=Pe3AxLq6Wf | [
"Benoit Dufumier",
"Javiera Castillo Navarro",
"Devis Tuia",
"Jean-Philippe Thiran"
] | Poster | Humans perceive the world through multisensory integration, blending the information of different modalities to adapt their behavior.
Contrastive learning offers an appealing solution for multimodal self-supervised learning. Indeed, by considering each modality as a different view of the same entity, it learns to align features of different modalities in a shared representation space. However, this approach is intrinsically limited as it only learns shared or redundant information between modalities, while multimodal interactions can arise in other ways. In this work, we introduce CoMM, a Contrastive Multimodal learning strategy that enables the communication between modalities in a single multimodal space. Instead of imposing cross- or intra- modality constraints, we propose to align multimodal representations by maximizing the mutual information between augmented versions of these multimodal features. Our theoretical analysis shows that shared, synergistic and unique terms of information naturally emerge from this formulation, allowing us to estimate multimodal interactions beyond redundancy. We test CoMM both in a controlled and in a series of real-world settings: in the former, we demonstrate that CoMM effectively captures redundant, unique and synergistic information between modalities. In the latter, CoMM learns complex multimodal interactions and achieves state-of-the-art results on seven multimodal tasks. | Multimodal representation learning, Self-supervised learning, Contrastive learning | We present CoMM, a new multimodal representation learning method that learns synergistic, unique and redundant information between modalities. | 1,878 | 2409.07402 | [
0.03350325673818588,
-0.17412030696868896,
0.019308950752019882,
0.03355911746621132,
0.10522978752851486,
-0.03577106446027756,
-0.009423358365893364,
-0.044286150485277176,
0.0576767735183239,
-0.05765942111611366,
-0.013775531202554703,
-0.11337669938802719,
0.016678912565112114,
0.07874266058206558,
0.017086397856473923,
0.012630286626517773,
0.048144709318876266,
0.1159820556640625,
-0.143462672829628,
-0.01928483322262764,
0.04782281070947647,
0.008078224025666714,
0.034515444189310074,
0.005106769036501646,
0.042650505900382996,
0.07849153876304626,
-0.0007086603436619043,
0.01841791346669197,
0.011594808660447598,
-0.072963185608387,
0.025069601833820343,
0.010759838856756687,
-0.024689650163054466,
0.047577228397130966,
-0.07072955369949341,
0.0953710526227951,
-0.05756789818406105,
0.047441642731428146,
-0.029363857582211494,
-0.09461115300655365,
-0.010353508405387402,
0.07597850263118744,
0.001759082661010325,
-0.009116778150200844,
-0.0216738972812891,
0.0687420591711998,
-0.029057208448648453,
-0.04217562824487686,
0.0112053994089365,
0.0015865596942603588,
-0.08445391803979874,
-0.014069718308746815,
-0.03054579719901085,
0.08347250521183014,
0.019500738009810448,
0.06717254966497421,
-0.017220137640833855,
0.014955651946365833,
0.023801781237125397,
-0.04215984418988228,
0.04562420770525932,
-0.07429250329732895,
0.0022894665598869324,
-0.005881504155695438,
0.06519868969917297,
-0.035307034850120544,
0.04759444668889046,
0.015752572566270828,
-0.0246767345815897,
0.02793477475643158,
-0.019804498180747032,
0.014277095906436443,
-0.019110525026917458,
0.011597712524235249,
-0.0019381477031856775,
0.05719780549407005,
0.055322837084531784,
-0.02424263209104538,
-0.01613154076039791,
-0.05309377238154411,
0.0541330948472023,
0.03442196547985077,
0.03083493746817112,
-0.021709084510803223,
0.13274919986724854,
0.00955208670347929,
-0.0769244059920311,
0.031089693307876587,
-0.029476085677742958,
0.05123600363731384,
-0.0567815899848938,
-0.020435409620404243,
-0.03197381645441055,
-0.03834123909473419,
0.04572899639606476,
-0.04252928867936134,
0.09889440983533859,
-0.06746168434619904,
0.009562310762703419,
0.10147521644830704,
0.002202050294727087,
-0.008556540124118328,
-0.0437457449734211,
0.03241599351167679,
0.06070757284760475,
-0.0634554922580719,
-0.03126164525747299,
-0.0612197071313858,
-0.005330053623765707,
-0.08883842080831528,
0.01555003970861435,
-0.007084559183567762,
-0.015804752707481384,
0.022232862189412117,
0.018459569662809372,
-0.0011957611422985792,
0.058824323117733,
0.016203967854380608,
0.11274014413356781,
-0.031874097883701324,
-0.001138350460678339,
-0.07190944999456406,
-0.0007051186985336244,
0.03534301370382309,
0.04383482411503792,
-0.0766739547252655,
-0.07233068346977234,
1.5767274906273897e-33,
0.05208628624677658,
0.015727899968624115,
0.04575215280056,
0.10535953938961029,
0.04413154348731041,
-0.016800252720713615,
-0.07733055204153061,
0.0205658208578825,
-0.018467269837856293,
0.0027124558109790087,
0.04660797119140625,
0.08189285546541214,
-0.0028523141518235207,
0.11216600239276886,
-0.06036334112286568,
0.022644490003585815,
-0.07416590303182602,
0.02879786305129528,
0.015241438522934914,
-0.004920466803014278,
-0.029363567009568214,
0.03259458765387535,
-0.059942811727523804,
0.0376422293484211,
-0.03236282989382744,
-0.036477454006671906,
0.04818158596754074,
-0.0875794067978859,
0.016562780365347862,
-0.017309078946709633,
-0.10609893500804901,
0.060105372220277786,
-0.011408628895878792,
0.004053631331771612,
0.06998755782842636,
-0.035039544105529785,
-0.04428818076848984,
-0.051431432366371155,
-0.022535428404808044,
-0.013650545850396156,
-0.02726687304675579,
0.004360162187367678,
-0.0532228946685791,
-0.0307313259691,
-0.04407637566328049,
-0.027496283873915672,
0.013883565552532673,
0.04643633961677551,
-0.07803516834974289,
-0.06879816204309464,
-0.026227308437228203,
0.015846945345401764,
0.008956481702625751,
-0.1186172366142273,
-0.03136131912469864,
-0.01741192489862442,
-0.014950308948755264,
-0.012333364225924015,
0.0366680771112442,
0.02990678697824478,
-0.06202118471264839,
0.07170415669679642,
0.022048112004995346,
0.05065047740936279,
0.0774168148636818,
0.08535534143447876,
-0.027879251167178154,
0.07793707400560379,
0.13029758632183075,
-0.02069987542927265,
-0.06982852518558502,
0.06660766154527664,
-0.042425487190485,
-0.071844682097435,
0.0236368365585804,
0.03654983267188072,
0.08127167820930481,
-0.06762385368347168,
0.049826640635728836,
0.06954889744520187,
-0.0024796214420348406,
0.006738904397934675,
0.02042338252067566,
-0.04640655592083931,
-0.060299478471279144,
-0.0009999473113566637,
0.03265690058469772,
-0.12838369607925415,
-0.050629064440727234,
-0.03290495276451111,
-0.04410521686077118,
-0.0055579617619514465,
0.008259085938334465,
-0.022223806008696556,
0.03770969435572624,
-2.7536181083332634e-33,
0.03028891235589981,
0.03182828426361084,
0.005908859893679619,
0.044859591871500015,
0.015528024174273014,
-0.024022404104471207,
0.010796905495226383,
0.01789027824997902,
-0.028374595567584038,
-0.0632859393954277,
0.01782871223986149,
-0.01312429178506136,
0.01198357529938221,
-0.020125307142734528,
-0.038363631814718246,
0.05897112190723419,
0.016307177022099495,
0.052915383130311966,
-0.0008288727840408683,
0.01186070591211319,
0.02987655997276306,
0.07864166796207428,
-0.00005126991163706407,
0.04696514457464218,
-0.09839536249637604,
0.05274854227900505,
-0.05589495971798897,
-0.036157961934804916,
-0.02920977771282196,
0.0055497088469564915,
0.006174110341817141,
-0.018492495641112328,
-0.003688487457111478,
-0.08630889654159546,
-0.039834100753068924,
0.04799136146903038,
-0.03662123903632164,
-0.02652679942548275,
-0.03784671425819397,
0.0818711370229721,
0.08672619611024857,
-0.025977585464715958,
-0.0679490715265274,
-0.025719204917550087,
0.016331534832715988,
-0.09996093809604645,
-0.09282046556472778,
-0.046245262026786804,
0.03439551964402199,
0.07997918874025345,
-0.03559419885277748,
0.017755478620529175,
-0.09772936254739761,
-0.09518018364906311,
-0.04566485062241554,
-0.10420737415552139,
0.039006058126688004,
0.016764681786298752,
0.10014571249485016,
0.030139228329062462,
-0.007153445854783058,
-0.07788760960102081,
-0.05075874924659729,
0.009758815169334412,
0.016537979245185852,
-0.023056073114275932,
0.021709628403186798,
-0.0868886336684227,
0.038768742233514786,
0.007982348091900349,
0.06632844358682632,
-0.013294375501573086,
0.013222593814134598,
-0.05613796412944794,
0.08038871735334396,
0.004772169049829245,
0.08225957304239273,
-0.019389456138014793,
-0.05107057839632034,
-0.019964775070548058,
0.02596968784928322,
-0.036288581788539886,
0.059768274426460266,
0.052206192165613174,
-0.021130776032805443,
0.11555521935224533,
-0.002664115745574236,
0.06354262679815292,
0.07945321500301361,
-0.018546845763921738,
-0.012688995338976383,
0.04220695048570633,
-0.0378933846950531,
-0.0008480896358378232,
-0.013514252379536629,
-4.057494606968248e-8,
-0.07074092328548431,
-0.020137464627623558,
0.017036370933055878,
-0.03662431240081787,
0.025114547461271286,
-0.027034835889935493,
0.005020123906433582,
0.011937522329390049,
-0.03311658278107643,
-0.005588208790868521,
0.061055976897478104,
-0.007058680988848209,
0.0007811319665051997,
-0.05138454958796501,
-0.027268745005130768,
0.0039686174131929874,
-0.0744042694568634,
-0.04135963320732117,
-0.017933769151568413,
0.06154020130634308,
0.04360008239746094,
0.00865687895566225,
0.03141779080033302,
-0.03292310982942581,
0.05913955345749855,
-0.048070862889289856,
-0.036468636244535446,
0.02280079573392868,
0.02503504417836666,
-0.03251664713025093,
-0.020862456411123276,
0.003387393429875374,
0.027982987463474274,
-0.058402519673109055,
-0.07051774859428406,
0.05727152153849602,
0.029168860986828804,
-0.008675809018313885,
-0.06310318410396576,
-0.061202116310596466,
-0.054048847407102585,
0.028970833867788315,
-0.03000584803521633,
0.07328034937381744,
0.07513084262609482,
0.04296804592013359,
0.003599172458052635,
-0.07928872853517532,
-0.0038121778052300215,
-0.0385419987142086,
-0.008279332891106606,
0.0068611386232078075,
0.08532638847827911,
0.09693710505962372,
0.09642933309078217,
-0.046192895621061325,
-0.022811971604824066,
0.09061409533023834,
0.0492636114358902,
0.06135570630431175,
0.02133161947131157,
0.027404718101024628,
-0.11687925457954407,
-0.017427723854780197
] |
Smoothing the Shift: Towards Stable Test-Time Adaptation under Complex Multimodal Noises | https://openreview.net/forum?id=rObkvzJxTG | [
"Zirun Guo",
"Tao Jin"
] | Poster | Test-Time Adaptation (TTA) aims to tackle distribution shifts using unlabeled test data without access to the source data. In the context of multimodal data, there are more complex noise patterns than unimodal data such as simultaneous corruptions for multiple modalities and missing modalities. Besides, in real-world applications, corruptions from different distribution shifts are always mixed. Existing TTA methods always fail in such multimodal scenario because the abrupt distribution shifts will destroy the prior knowledge from the source model, thus leading to performance degradation. To this end, we reveal a new challenge named *multimodal wild TTA*. To address this challenging problem, we propose two novel strategies: sample identification with interquartile range **S**moothing and **u**nimodal assistance, and **M**utual **i**nformation sharing (SuMi). SuMi smooths the adaptation process by interquartile range which avoids the abrupt distribution shifts. Then, SuMi fully utilizes the unimodal features to select low-entropy samples with rich multimodal information for optimization. Furthermore, mutual information sharing is introduced to align the information, reduce the discrepancies and enhance the information utilization across different modalities. Extensive experiments show the effectiveness and superiority over existing methods under the complex noise patterns in multimodal data. Code is available at https://github.com/zrguo/SuMi. | Test-Time Adaptation, Multimodal Noises, Transfer Learning | We reveal a new challenge named multimodal wild test-time adaptation. | 1,864 | 2503.02616 | [
-0.07989013940095901,
-0.12000441551208496,
0.056548863649368286,
0.013033853843808174,
0.08250652998685837,
-0.04496435448527336,
-0.07212255895137787,
-0.06810829043388367,
-0.010658561252057552,
-0.09235752373933792,
-0.035699114203453064,
-0.10520461946725845,
-0.013419164344668388,
0.012380574829876423,
-0.006517205853015184,
-0.08478786796331406,
0.017995165660977364,
0.12043119966983795,
-0.03950946778059006,
0.06123996153473854,
0.06936495006084442,
0.07125119864940643,
-0.0076972972601652145,
-0.00419970415532589,
0.09014907479286194,
0.020197909325361252,
-0.014494561590254307,
-0.03383666276931763,
0.06124605983495712,
-0.1235015019774437,
0.03177297115325928,
0.07474985718727112,
-0.029256561771035194,
0.007116695400327444,
0.033890821039676666,
-0.019092753529548645,
-0.06483037024736404,
0.07648301869630814,
0.0247910525649786,
0.00025111716240644455,
0.038532696664333344,
0.001795584917999804,
-0.001527645392343402,
-0.025976618751883507,
-0.035285793244838715,
-0.0072549255564808846,
0.0021657792385667562,
-0.036199167370796204,
0.041510406881570816,
0.032254479825496674,
-0.05108741298317909,
0.020809946581721306,
-0.0029236453119665384,
0.13383033871650696,
-0.02076338417828083,
-0.03720689192414284,
0.10001874715089798,
-0.04473740607500076,
0.05803156644105911,
0.03656954690814018,
-0.012186593376100063,
-0.02759569324553013,
0.02214706875383854,
-0.01061827689409256,
0.016784127801656723,
0.019812507554888725,
0.019799094647169113,
-0.008628368377685547,
0.03902726247906685,
-0.028797902166843414,
-0.07911822199821472,
0.08987262099981308,
-0.05128352716565132,
0.0566394180059433,
-0.012003542855381966,
0.08352766185998917,
0.04568808153271675,
-0.015920983627438545,
0.08300384134054184,
-0.0879732072353363,
0.071685791015625,
0.013648169115185738,
0.012932060286402702,
-0.07623620331287384,
0.13768288493156433,
0.040765050798654556,
-0.005290600936859846,
0.09290183335542679,
0.017414573580026627,
0.05114179104566574,
-0.06613005697727203,
0.0480002723634243,
-0.008731961250305176,
0.03739507868885994,
-0.004807584919035435,
0.010391440242528915,
0.09384050965309143,
-0.03469352424144745,
0.10074448585510254,
0.10420700907707214,
0.03397754952311516,
-0.019169094040989876,
-0.0713411197066307,
-0.005794022232294083,
0.020491626113653183,
-0.058638136833906174,
0.11524337530136108,
-0.07164722681045532,
0.0320592001080513,
0.014544427394866943,
0.02567923255264759,
0.0322004109621048,
0.03882911801338196,
-0.0034168020356446505,
-0.003999786917120218,
-0.02702464535832405,
-0.03095339797437191,
0.02738720178604126,
-0.05199562385678291,
0.003730131546035409,
-0.0027794658672064543,
-0.053733449429273605,
-0.030918002128601074,
0.0024673196021467447,
0.04078206419944763,
0.06354016810655594,
-0.03286682069301605,
2.492566894296542e-33,
0.049726180732250214,
0.0016976731130853295,
-0.02554858848452568,
0.026480985805392265,
0.0716608315706253,
-0.04863870143890381,
-0.0699496939778328,
0.03776304051280022,
-0.04876701533794403,
0.011397800408303738,
-0.0028827395290136337,
0.028442922979593277,
-0.11475996673107147,
0.03914472460746765,
-0.02409251593053341,
0.045481014996767044,
-0.013107185252010822,
0.02343633584678173,
0.017162851989269257,
0.03689907118678093,
0.056091342121362686,
-0.05902807414531708,
0.0064742909744381905,
-0.026070022955536842,
-0.010353409685194492,
0.011846070177853107,
0.007803838234394789,
-0.05717882886528969,
0.03334406018257141,
0.026152582839131355,
-0.14337745308876038,
-0.03712115436792374,
-0.02571907266974449,
-0.014324188232421875,
0.045236777514219284,
-0.08431907743215561,
0.003969061188399792,
-0.03388369828462601,
-0.11383950710296631,
0.03391661122441292,
-0.06992045789957047,
0.012284326367080212,
-0.05896075442433357,
-0.01431721169501543,
-0.03057604283094406,
-0.07140935212373734,
0.01156175509095192,
-0.009806831367313862,
-0.020683692768216133,
-0.08645841479301453,
-0.022833997383713722,
0.025246499106287956,
-0.02076120860874653,
-0.1573266088962555,
-0.05875032767653465,
-0.0204592514783144,
0.035595472902059555,
-0.03543027862906456,
0.06931640207767487,
0.03421064466238022,
-0.041748691350221634,
0.048461854457855225,
-0.00244144001044333,
0.08202871680259705,
0.029170051217079163,
0.0268973670899868,
-0.06776165217161179,
-0.00871227402240038,
0.08837045729160309,
0.019415322691202164,
-0.005259273108094931,
-0.02393648773431778,
-0.03707515448331833,
0.017006823793053627,
-0.012786534614861012,
-0.02829867973923683,
0.06407885998487473,
0.023314962163567543,
0.04312698543071747,
0.01639968901872635,
0.09080874174833298,
0.0020767017267644405,
-0.002741927281022072,
-0.05632005259394646,
-0.009043562225997448,
-0.02738008089363575,
-0.012971426360309124,
-0.07434193789958954,
-0.07034336030483246,
-0.08632355183362961,
-0.005770989693701267,
0.03538202866911888,
-0.04544619098305702,
-0.044185440987348557,
0.02363097481429577,
-2.999757625516827e-33,
0.01675829105079174,
0.04747466742992401,
0.004789569415152073,
0.031528692692518234,
0.013954679481685162,
-0.041761428117752075,
0.02354222722351551,
0.05781763046979904,
0.059716369956731796,
-0.037744760513305664,
-0.05792868509888649,
-0.08348382264375687,
0.04250417649745941,
-0.01415208075195551,
-0.004424591548740864,
-0.03114195354282856,
0.019010599702596664,
0.02138548716902733,
-0.03405803442001343,
0.008923506364226341,
0.022067466750741005,
0.04480001702904701,
-0.05748215690255165,
0.006353738252073526,
-0.08661361783742905,
0.04673687368631363,
-0.12417004257440567,
0.07242586463689804,
-0.051577452570199966,
-0.033803973346948624,
-0.0258783008903265,
0.04323979467153549,
-0.02165752649307251,
-0.03228121995925903,
0.024820052087306976,
0.07272109389305115,
-0.004673711955547333,
-0.0030454625375568867,
-0.026896977797150612,
0.12266606092453003,
0.10009625554084778,
0.08439750969409943,
-0.01179751567542553,
0.02239089645445347,
-0.011401083320379257,
0.020987708121538162,
-0.08021005243062973,
-0.018137330189347267,
0.02873242273926735,
0.06271052360534668,
0.027851637452840805,
-0.007584801875054836,
-0.04821902513504028,
0.02335912361741066,
0.012138829566538334,
-0.05320771783590317,
-0.022049523890018463,
-0.0412382148206234,
-0.02479373849928379,
0.04497739300131798,
-0.029039889574050903,
-0.005343866068869829,
-0.08066658675670624,
-0.06000005081295967,
0.0078011490404605865,
-0.023147327825427055,
0.026059869676828384,
-0.06235669553279877,
0.03271302208304405,
0.027728799730539322,
0.04485362023115158,
-0.03261328116059303,
-0.007173544727265835,
-0.04034433886408806,
0.06998808681964874,
-0.0045673782005906105,
-0.03160369396209717,
-0.033754054456949234,
0.009716046042740345,
-0.06373501569032669,
-0.019636614248156548,
-0.03516881540417671,
0.022263523191213608,
0.07772690057754517,
-0.02893795445561409,
0.2060619294643402,
0.05989908054471016,
0.05581312254071236,
0.09885678440332413,
-0.04403889551758766,
-0.014170520007610321,
0.07422033697366714,
0.006448362488299608,
-0.00019313629309181124,
0.015605672262609005,
-4.520071072988685e-8,
-0.008108050562441349,
0.008721787482500076,
-0.031109249219298363,
-0.019397415220737457,
0.0474306084215641,
-0.05811014398932457,
-0.05324646830558777,
-0.0208026934415102,
0.07894296944141388,
-0.05435976758599281,
0.08421864360570908,
-0.029320083558559418,
0.044917136430740356,
-0.02276964858174324,
-0.016625573858618736,
0.04590049386024475,
-0.01595313660800457,
0.06132661923766136,
-0.042693428695201874,
-0.028986357152462006,
-0.013720321469008923,
0.07984796166419983,
0.04112770035862923,
-0.020849915221333504,
-0.00031911410042084754,
-0.051269419491291046,
0.03136374056339264,
0.013674999587237835,
0.059257447719573975,
-0.054500892758369446,
-0.04335509613156319,
-0.011183664202690125,
-0.05123203247785568,
-0.0596628375351429,
-0.03753625974059105,
0.08847946673631668,
0.05171310156583786,
0.027777690440416336,
-0.03487207740545273,
0.027512118220329285,
-0.0011035731295123696,
0.014022226445376873,
-0.10221397876739502,
-0.006883909925818443,
-0.08289210498332977,
-0.04332151636481285,
-0.03418101370334625,
-0.07364340871572495,
0.008111138828098774,
-0.05050475522875786,
0.04983125254511833,
-0.03372028470039368,
0.015105622820556164,
0.06714096665382385,
0.07784537225961685,
-0.009480767883360386,
-0.02541038766503334,
0.026608256623148918,
0.041746970266103745,
0.006004270166158676,
0.058526281267404556,
-0.04615489020943642,
-0.07963072508573532,
-0.04583686217665672
] |
Improving Language Model Distillation through Hidden State Matching | https://openreview.net/forum?id=IcVSKhVpKu | [
"Sayantan Dasgupta",
"Trevor Cohn"
] | Poster | Hidden State Matching is shown to improve knowledge distillation of language models by encouraging similarity between a student and its teacher's hidden states since DistilBERT. This typically uses a cosine loss, which restricts the dimensionality of the student to the teacher's, severely limiting the compression ratio. We present an alternative technique using Centered Kernel Alignment (CKA) to match hidden states of different dimensionality, allowing for smaller students and higher compression ratios. We show the efficacy of our method using encoder--decoder (BART, mBART \& T5) and encoder-only (BERT) architectures across a range of tasks from classification to summarization and translation. Our technique is competitive with the current state-of-the-art distillation methods at comparable compression rates and does not require already pretrained student models. It can scale to students smaller than the current methods, is no slower in training and inference, and is considerably more flexible. The code is available on github. | Knowledge Distillation, Centered Kernel Alignment, BART, mBART, T5 | null | 1,863 | null | [
-0.11866841465234756,
-0.05956215038895607,
0.04987943917512894,
-0.040502917021512985,
0.045780621469020844,
-0.0032538429368287325,
-0.058421436697244644,
-0.015044709667563438,
0.0619606152176857,
0.005256597883999348,
-0.012770348228514194,
0.0222772266715765,
-0.015691764652729034,
0.023522481322288513,
-0.04538438096642494,
0.005342384800314903,
0.027533812448382378,
0.14659009873867035,
-0.039504244923591614,
-0.09923675656318665,
0.026956086978316307,
0.03243628889322281,
-0.03645961731672287,
0.0747445747256279,
0.1269443780183792,
0.04030477628111839,
0.053546831011772156,
-0.020784029737114906,
-0.008406969718635082,
-0.02804170921444893,
0.013582654297351837,
-0.028057213872671127,
-0.05722816288471222,
0.09044993668794632,
-0.08057443052530289,
0.07121660560369492,
-0.07123584300279617,
0.004311526194214821,
0.03576594963669777,
-0.029244713485240936,
-0.04929881542921066,
0.05653529241681099,
-0.02484987862408161,
0.038925014436244965,
0.03764363378286362,
0.013767292723059654,
-0.05728330835700035,
-0.03584206476807594,
0.0619162954390049,
-0.046343084424734116,
-0.07102561742067337,
0.05817243829369545,
0.0038042785599827766,
0.05821249634027481,
-0.02053893357515335,
-0.05906135216355324,
0.01354038342833519,
-0.015159711241722107,
-0.011971008963882923,
-0.044485077261924744,
-0.04675411060452461,
-0.0922241285443306,
0.00789071898907423,
0.01622150093317032,
-0.005894121248275042,
-0.023375026881694794,
0.03966545686125755,
0.07665012776851654,
-0.001159470179118216,
0.040464166551828384,
-0.03843967989087105,
0.0619841106235981,
0.02487294375896454,
0.08297830075025558,
0.022299760952591896,
0.03964821994304657,
0.028878584504127502,
0.03363411873579025,
0.04911414906382561,
-0.020085206255316734,
0.007686207070946693,
-0.009595583193004131,
0.08941337466239929,
-0.0318574421107769,
0.087868332862854,
0.00724735576659441,
0.04344996064901352,
-0.031622886657714844,
0.021888842806220055,
-0.0650525689125061,
-0.04057852551341057,
-0.04783739149570465,
0.06939160078763962,
-0.0382523387670517,
0.0507531613111496,
0.01322406530380249,
0.02170737087726593,
0.03998540714383125,
-0.05789465457201004,
0.060844480991363525,
0.024836938828229904,
0.1446591466665268,
0.01781836338341236,
-0.0964396744966507,
-0.04794929549098015,
-0.036079324781894684,
0.06895694136619568,
0.04206595942378044,
0.03483980521559715,
-0.07450088113546371,
0.04328550770878792,
0.02909287065267563,
-0.015763742849230766,
0.057650212198495865,
0.025786448270082474,
-0.01894553191959858,
0.09528641402721405,
-0.012100474908947945,
0.0346040204167366,
0.03537696227431297,
-0.008783008903265,
0.015095061622560024,
-0.07671397179365158,
-0.004279596731066704,
-0.06751459091901779,
-0.04185619577765465,
-0.061436641961336136,
2.1163822534105167e-33,
0.04718771204352379,
0.09113092720508575,
-0.0413784459233284,
0.04818776622414589,
-0.04779887944459915,
0.0006920245359651744,
0.023375242948532104,
0.04544608294963837,
-0.019484061747789383,
0.006459974218159914,
-0.03292320668697357,
0.005195466335862875,
-0.08393334597349167,
0.05413757264614105,
-0.025189600884914398,
0.013377828523516655,
-0.06509499996900558,
0.04970856383442879,
-0.00206860713660717,
-0.02751782536506653,
0.04904290661215782,
-0.01773751527070999,
-0.020308952778577805,
-0.05135119706392288,
-0.03056482970714569,
-0.010125355795025826,
0.03820715472102165,
-0.11372304707765579,
-0.0062130833975970745,
-0.013289143331348896,
-0.06668704003095627,
-0.04851003363728523,
0.006379662547260523,
-0.04143880680203438,
0.06389476358890533,
-0.0777510330080986,
0.015270000323653221,
-0.02546743117272854,
0.03025851584970951,
-0.0808834359049797,
-0.029629193246364594,
-0.0033326540142297745,
-0.03153888136148453,
-0.06457734107971191,
-0.03705539554357529,
-0.00445456150919199,
0.024507757276296616,
0.01030705776065588,
0.0001076387707144022,
-0.07162198424339294,
0.03549874573945999,
0.030858460813760757,
-0.1264888048171997,
-0.10152330994606018,
-0.008726146072149277,
-0.010411364957690239,
0.11895079910755157,
-0.0029404705855995417,
0.04889893904328346,
0.060507882386446,
-0.07507038116455078,
0.04358325153589249,
0.03562643751502037,
0.028014399111270905,
0.03458579257130623,
-0.052105970680713654,
-0.08258403837680817,
0.015908677130937576,
0.07857971638441086,
0.02946562133729458,
-0.0744374617934227,
-0.02259337343275547,
0.007266636472195387,
-0.015611611306667328,
0.016179731115698814,
-0.02966814674437046,
0.03742777183651924,
-0.11282438039779663,
-0.033432863652706146,
-0.02555394358932972,
0.049534834921360016,
-0.10071749985218048,
-0.01886008493602276,
-0.05182524770498276,
-0.10173312574625015,
-0.058893922716379166,
0.014184513129293919,
-0.0765179842710495,
-0.011287031695246696,
-0.003626860911026597,
0.06330699473619461,
-0.032763902097940445,
0.04544449970126152,
0.05179029330611229,
0.01448766328394413,
-9.699467657890292e-34,
0.03409132361412048,
0.08504720032215118,
-0.01160183921456337,
0.02609865553677082,
-0.00965424720197916,
0.027857167646288872,
0.05470966920256615,
0.09334580600261688,
-0.014987216331064701,
-0.0690852701663971,
0.01976892724633217,
-0.038874994963407516,
0.016367662698030472,
0.0031045617070049047,
0.021754248067736626,
0.05460038036108017,
0.039354290813207626,
0.08297894150018692,
-0.02203763648867607,
0.05201975628733635,
-0.023997878655791283,
0.11197074502706528,
-0.0971057116985321,
0.04047182574868202,
-0.004595952574163675,
0.04254141077399254,
-0.001308166654780507,
0.05021341145038605,
-0.014370430260896683,
-0.010669446550309658,
0.026663830503821373,
-0.014280453324317932,
-0.0073562366887927055,
-0.05001034587621689,
-0.155123770236969,
-0.025000374764204025,
-0.04828998073935509,
-0.0014164443127810955,
-0.022540371865034103,
0.09792426228523254,
0.03964388370513916,
-0.031614791601896286,
-0.042189113795757294,
-0.05103481933474541,
-0.04003618657588959,
0.006066698580980301,
-0.04577361419796944,
-0.02151481993496418,
0.02081666886806488,
0.042672768235206604,
0.04636741802096367,
0.06645867973566055,
-0.13170692324638367,
0.0064649926498532295,
0.03288792446255684,
-0.018358631059527397,
-0.07323166728019714,
-0.009989649057388306,
-0.039260778576135635,
0.058344367891550064,
-0.03177018463611603,
-0.03211196884512901,
0.11079787462949753,
-0.02683853544294834,
0.009175755083560944,
-0.07138845324516296,
0.0023033753968775272,
0.12833572924137115,
-0.007601418532431126,
0.06112898141145706,
0.039704352617263794,
-0.023227710276842117,
0.05074401944875717,
-0.00013826164649799466,
0.06489730626344681,
-0.022112861275672913,
0.04364222288131714,
-0.020319271832704544,
-0.04559692367911339,
0.020507019013166428,
-0.012831974774599075,
-0.04567582160234451,
0.007836407981812954,
0.02129167877137661,
0.07907716184854507,
0.08323957771062851,
0.05237210541963577,
0.018405454233288765,
0.006760501302778721,
0.08092082291841507,
-0.05761454999446869,
0.017024323344230652,
0.04041914641857147,
0.012013806961476803,
0.019887732341885567,
-4.400374109536642e-8,
-0.10974948853254318,
-0.0013789020013064146,
-0.06327387690544128,
0.0595623217523098,
-0.0764453187584877,
-0.10561716556549072,
0.0009581748163327575,
0.04299350827932358,
-0.056225962936878204,
0.031161488965153694,
0.029640555381774902,
-0.03072443976998329,
-0.0655026063323021,
-0.008403124287724495,
0.030196400359272957,
0.08506916463375092,
-0.013805029913783073,
0.059452082961797714,
0.04168710485100746,
-0.004685705062001944,
0.047251906245946884,
-0.01812722347676754,
-0.015961850062012672,
0.05166994035243988,
0.0406813770532608,
-0.04775703698396683,
0.013190430589020252,
0.1330580711364746,
0.06670957803726196,
0.04310055822134018,
-0.053076136857271194,
0.041574835777282715,
0.017814071848988533,
-0.07831207662820816,
0.019407426938414574,
0.04948293790221214,
-0.01893037185072899,
0.029048198834061623,
-0.0058723753318190575,
0.0709407851099968,
0.057084906846284866,
0.034642741084098816,
-0.10642881691455841,
0.04164733737707138,
0.012996850535273552,
0.04198620468378067,
-0.010670009069144726,
-0.06576210260391235,
0.03339880332350731,
0.012018844485282898,
0.08296915888786316,
-0.0498565137386322,
0.0014486652798950672,
0.021021058782935143,
0.06015780195593834,
0.03982505202293396,
-0.06273727118968964,
0.025945154950022697,
0.02002865821123123,
-0.03163399547338486,
0.002148853149265051,
0.014216319657862186,
-0.011506594717502594,
0.055917419493198395
] |
Long-Context Linear System Identification | https://openreview.net/forum?id=2TuUXtLGhT | [
"Oğuz Kaan Yüksel",
"Mathieu Even",
"Nicolas Flammarion"
] | Poster | This paper addresses the problem of long-context linear system identification, where the state $x_t$ of the system at time $t$ depends linearly on previous states $x_s$ over a fixed context window of length $p$. We establish a sample complexity bound that matches the _i.i.d._ parametric rate, up to logarithmic factors for a broad class of systems, extending previous work that considered only first-order dependencies. Our findings reveal a ``learning-without-mixing'' phenomenon, indicating that learning long-context linear autoregressive models is not hindered by slow mixing properties potentially associated with extended context windows. Additionally, we extend these results to _(i)_ shared low-rank feature representations, where rank-regularized estimators improve rates with respect to dimensionality, and _(ii)_ misspecified context lengths in strictly stable systems, where shorter contexts offer statistical advantages. | autoregressive, linear, statistics, low rank, mispecification | This paper provides sample complexity for long-context linear dynamical systems. | 1,861 | 2410.05690 | [
-0.018863987177610397,
-0.07447126507759094,
0.03994297608733177,
0.020458905026316643,
0.15722884237766266,
0.03674917668104172,
0.03743546083569527,
-0.02387838251888752,
0.02069641463458538,
-0.045819275081157684,
-0.006577844265848398,
0.008557459339499474,
0.012859804555773735,
0.033406052738428116,
-0.05084283649921417,
-0.01586967147886753,
0.06921997666358948,
0.022824563086032867,
-0.08987051248550415,
0.045359231531620026,
-0.020178576931357384,
0.017104312777519226,
-0.02449938654899597,
0.035528115928173065,
0.08719182759523392,
0.021457064896821976,
-0.04649777710437775,
-0.041236646473407745,
0.06867794692516327,
0.023075230419635773,
0.01536343153566122,
0.09534069895744324,
0.017323942855000496,
-0.08871166408061981,
-0.011629385873675346,
0.03469013050198555,
0.026294462382793427,
0.07467462122440338,
-0.024832287803292274,
0.016357501968741417,
0.0611482597887516,
0.04958489537239075,
0.02897532656788826,
0.0776471421122551,
0.0034959728363901377,
-0.0075418842025101185,
0.03826812654733658,
-0.05294138193130493,
-0.04414628446102142,
-0.0349532812833786,
-0.06199268996715546,
0.07422172278165817,
-0.061497174203395844,
0.05496882647275925,
-0.01705520786345005,
-0.04698556289076805,
0.025457017123699188,
0.0602438785135746,
-0.027743497863411903,
0.06706953048706055,
-0.07292380928993225,
-0.0516379214823246,
-0.02322680875658989,
-0.018666187301278114,
0.04008195921778679,
0.04372696578502655,
-0.0047696116380393505,
0.06635693460702896,
-0.02850319817662239,
0.005452737212181091,
-0.10777322202920914,
0.05821657180786133,
-0.08753398060798645,
0.021499257534742355,
0.01760062202811241,
-0.08000953495502472,
0.017432736232876778,
-0.017501799389719963,
0.030554847791790962,
-0.0713743343949318,
0.042301394045352936,
0.03737242519855499,
-0.026672745123505592,
-0.03875986859202385,
0.0820607915520668,
0.03357351943850517,
-0.04407389089465141,
0.006107653956860304,
0.03518672659993172,
0.0010969339637085795,
-0.011044006794691086,
-0.05578785389661789,
0.009198974817991257,
-0.026295127347111702,
0.1301841288805008,
0.013662084937095642,
0.0418354868888855,
-0.013429555110633373,
0.026353657245635986,
0.1292625069618225,
-0.030049579218029976,
0.010683615691959858,
-0.011591418646275997,
0.08973995596170425,
0.05828867107629776,
0.022323399782180786,
-0.0630754828453064,
0.062382083386182785,
-0.0029643632005900145,
-0.04911847785115242,
0.04300033673644066,
0.004420312121510506,
-0.03303488716483116,
0.060932718217372894,
-0.009905907325446606,
-0.1122216060757637,
0.06845960766077042,
-0.0876297578215599,
0.04323253408074379,
0.04207805171608925,
-0.0015813028439879417,
-0.06534843891859055,
0.016829343512654305,
-0.016142308712005615,
0.00845538079738617,
0.061554573476314545,
0.004150525666773319,
1.2001754159212361e-33,
-0.0073424335569143295,
0.044708363711833954,
0.02741854451596737,
0.004394045565277338,
-0.021108048036694527,
0.06476172804832458,
-0.047690365463495255,
-0.03479256480932236,
-0.014453364536166191,
0.07142540067434311,
0.022026952356100082,
0.07665185630321503,
-0.07025495916604996,
0.02828076295554638,
0.01651580072939396,
0.03493884578347206,
-0.019328035414218903,
0.08513934165239334,
0.003285019425675273,
-0.09133569151163101,
0.0038082939572632313,
-0.05140947178006172,
-0.06367749720811844,
-0.07676981389522552,
-0.0009850435890257359,
-0.07637286931276321,
-0.004540489055216312,
0.014011073857545853,
-0.03199851140379906,
-0.004606781993061304,
-0.07367363572120667,
0.024705257266759872,
-0.014643554575741291,
0.05528315529227257,
-0.009242658503353596,
-0.04769180715084076,
0.028908515349030495,
-0.02258136123418808,
0.012104924768209457,
-0.10811878740787506,
0.0058914972469210625,
0.007341593503952026,
0.008840953931212425,
-0.0036091164220124483,
-0.09802329540252686,
-0.17718510329723358,
-0.061138592660427094,
0.001890212413854897,
-0.08894415199756622,
-0.038553386926651,
0.021327363327145576,
-0.004179595969617367,
-0.029778027907013893,
-0.06261029094457626,
-0.04670491814613342,
-0.021479163318872452,
0.00206275237724185,
0.05103306099772453,
0.05217409506440163,
0.09954414516687393,
-0.028729142621159554,
0.003560800338163972,
0.00452865194529295,
0.03015071153640747,
0.07866932451725006,
0.05310771241784096,
-0.012515834532678127,
0.019146502017974854,
0.07527611404657364,
-0.05709075182676315,
0.02027835324406624,
-0.03991752490401268,
-0.052257735282182693,
0.0066728657111525536,
0.09814851731061935,
-0.05961000919342041,
0.005234888754785061,
0.017032859846949577,
-0.0731707438826561,
0.11092279851436615,
0.05490170419216156,
-0.037843648344278336,
0.04009731486439705,
-0.042774997651576996,
-0.07127077877521515,
-0.04727065563201904,
-0.04264015704393387,
-0.09144265949726105,
-0.06681744009256363,
-0.09364411979913712,
-0.0421365387737751,
-0.016917508095502853,
-0.020451171323657036,
0.054773472249507904,
0.009718900546431541,
-1.958778215666003e-33,
0.017606457695364952,
-0.05404431000351906,
0.04386935755610466,
0.007868655025959015,
0.035534728318452835,
-0.051039766520261765,
-0.032654598355293274,
0.017010396346449852,
-0.0987144187092781,
-0.0654081478714943,
0.018824860453605652,
-0.054769035428762436,
0.060616787523031235,
0.021459858864545822,
0.12436377257108688,
0.05240657925605774,
0.01156045962125063,
0.04072635620832443,
0.04161781072616577,
0.015119215473532677,
0.03703397884964943,
-0.01657438650727272,
-0.027074020355939865,
0.06027299165725708,
-0.0006872000521980226,
-0.05607156828045845,
-0.07300129532814026,
0.027234675362706184,
-0.07598498463630676,
-0.029838012531399727,
-0.03615754097700119,
-0.06422934681177139,
-0.04015988111495972,
0.02681632898747921,
-0.06822043657302856,
0.03609325736761093,
-0.031991444528102875,
-0.01361628994345665,
-0.048934608697891235,
0.10466655343770981,
0.060905564576387405,
-0.040063999593257904,
-0.06369771808385849,
0.03278079628944397,
-0.009878476150333881,
-0.03233344852924347,
-0.0362648069858551,
-0.001873475150205195,
0.09515658020973206,
-0.039481475949287415,
-0.025102579966187477,
0.08297071605920792,
-0.043512724339962006,
0.046637162566185,
-0.016368359327316284,
-0.007920916192233562,
-0.009697388857603073,
0.03255947679281235,
0.09202447533607483,
-0.07381996512413025,
0.028167108073830605,
-0.04374868422746658,
-0.027085985988378525,
-0.01677461713552475,
0.06964419782161713,
-0.05909828096628189,
0.04182128608226776,
-0.05718843266367912,
-0.003247317159548402,
0.04935756325721741,
0.01853790320456028,
-0.0170427393168211,
-0.09613978862762451,
0.008146494626998901,
-0.03487309068441391,
0.00790180079638958,
0.05220463499426842,
-0.06644845753908157,
-0.035527389496564865,
-0.031092049553990364,
0.0021505674812942743,
-0.01599281281232834,
0.06973988562822342,
-0.006965599954128265,
-0.06994123756885529,
0.032826367765665054,
0.06048145517706871,
0.04620221257209778,
0.02626105584204197,
0.013255470432341099,
0.016463737934827805,
0.019429974257946014,
-0.04901823773980141,
0.02325887605547905,
0.023927917703986168,
-4.5062108711135807e-8,
-0.033887527883052826,
0.05901803448796272,
-0.005774463061243296,
-0.008872193284332752,
0.02347346395254135,
0.07696707546710968,
-0.04914156347513199,
0.061474598944187164,
0.01948443613946438,
0.03682109713554382,
0.006309254560619593,
-0.012473173439502716,
0.029722057282924652,
-0.09318259358406067,
-0.02759716846048832,
0.04156290739774704,
-0.037788260728120804,
0.026355208829045296,
-0.000668409513309598,
0.037953753024339676,
0.09258070588111877,
0.05037896707653999,
0.05275145545601845,
-0.04350071772933006,
-0.04595880210399628,
-0.08752790093421936,
0.0152830695733428,
0.08454489707946777,
0.1195586696267128,
0.005218081176280975,
-0.039199426770210266,
0.0769878700375557,
-0.06227809190750122,
-0.08451228588819504,
-0.04199961572885513,
0.05815896764397621,
0.03266523778438568,
-0.04204564914107323,
-0.05438099429011345,
-0.0317784920334816,
-0.005532421171665192,
0.01722072809934616,
-0.08583378791809082,
0.0028414407279342413,
0.05660457909107208,
0.08310382813215256,
0.03127284348011017,
-0.11539178341627121,
0.03195716813206673,
0.004281891509890556,
0.13384923338890076,
-0.023293763399124146,
0.03140460327267647,
0.04657050222158432,
0.05486110597848892,
0.0008186521590687335,
0.023275699466466904,
-0.00810034479945898,
0.06372826546430588,
-0.01582927629351616,
-0.05574958771467209,
0.037878721952438354,
-0.05429502949118614,
0.03241194784641266
] |
Controllable Unlearning for Image-to-Image Generative Models via ϵ-Constrained Optimization | https://openreview.net/forum?id=9OJflnNu6C | [
"XiaoHua Feng",
"Yuyuan Li",
"Chaochao Chen",
"Li Zhang",
"Longfei Li",
"JUN ZHOU",
"Xiaolin Zheng"
] | Poster | While generative models have made significant advancements in recent years, they also raise concerns such as privacy breaches and biases. Machine unlearning has emerged as a viable solution, aiming to remove specific training data, e.g., containing private information and bias, from models. In this paper, we study the machine unlearning problem in Image-to-Image (I2I) generative models. Previous studies mainly treat it as a single objective optimization problem, offering a solitary solution, thereby neglecting the varied user expectations towards the trade-off between complete unlearning and model utility. To address this issue, we propose a controllable unlearning framework that uses a control coefficient $\epsilon$ to control the trade-off. We reformulate the I2I generative model unlearning problem into a $\epsilon$-constrained optimization problem and solve it with a gradient-based method to find optimal solutions for unlearning boundaries. These boundaries define the valid range for the control coefficient. Within this range, every yielded solution is theoretically guaranteed with Pareto optimality. We also analyze the convergence rate of our framework under various control functions. Extensive experiments on two benchmark datasets across three mainstream I2I models demonstrate the effectiveness of our controllable unlearning framework. | Machine unlearning, Generative model, Controllable | null | 1,859 | null | [
-0.08175860345363617,
-0.0525251142680645,
-0.008501652628183365,
0.06421737372875214,
0.04444246366620064,
0.029276465997099876,
-0.0017489228630438447,
-0.04338585212826729,
-0.01873067393898964,
0.019135039299726486,
0.017403410747647285,
-0.058157842606306076,
0.0888761505484581,
0.0035263230092823505,
-0.0075502474792301655,
0.012913778424263,
0.10077577829360962,
0.05982457101345062,
-0.10501419752836227,
-0.03750225901603699,
-0.00525343231856823,
0.015216672793030739,
0.03600544482469559,
-0.028317991644144058,
0.04372275620698929,
-0.03302910551428795,
0.05961208418011665,
-0.028436880558729172,
-0.02372674271464348,
-0.07967451959848404,
0.05633818730711937,
-0.03033769503235817,
-0.013814003206789494,
-0.0477028489112854,
-0.0028422889299690723,
0.06418070197105408,
-0.043120015412569046,
-0.011353720910847187,
0.0068843550980091095,
-0.041363734751939774,
0.007304621860384941,
0.04358348622918129,
-0.005121347960084677,
-0.03724813833832741,
0.02834511734545231,
-0.007800897117704153,
0.010671211406588554,
-0.0058888401836156845,
-0.02138056419789791,
-0.04941466078162193,
-0.01032493356615305,
-0.026491889730095863,
-0.050066232681274414,
0.007933961227536201,
0.06451425701379776,
-0.0330498181283474,
0.028902048245072365,
0.03146261349320412,
0.009721686132252216,
0.06117980182170868,
-0.08031297475099564,
-0.09837512671947479,
-0.0945572778582573,
0.005441791843622923,
0.0007850218680687249,
-0.012922108173370361,
0.03384222090244293,
0.014057982712984085,
0.0003518984012771398,
0.0058542778715491295,
0.012313568964600563,
0.053434863686561584,
-0.03834719955921173,
-0.03769619017839432,
0.053686149418354034,
0.02976449951529503,
0.04883693903684616,
-0.012150196358561516,
0.07152983546257019,
-0.06923041492700577,
0.04863879084587097,
0.01862824708223343,
0.06358327716588974,
0.012090151198208332,
0.06628625094890594,
0.0007844241335988045,
-0.03348338603973389,
0.08637672662734985,
0.14452554285526276,
0.052497394382953644,
-0.05026936158537865,
0.003122507594525814,
-0.03853979334235191,
0.03581145033240318,
0.05457289516925812,
-0.05248725041747093,
0.02575679123401642,
-0.06232581287622452,
-0.0378323495388031,
0.06770946085453033,
-0.005531069822609425,
0.031364556401968,
0.007896059192717075,
-0.041136495769023895,
0.09885162115097046,
-0.02327464148402214,
0.07564602792263031,
0.02461065910756588,
0.019732501357793808,
-0.03140867501497269,
0.07874827831983566,
0.015229511074721813,
-0.03381172940135002,
-0.021084928885102272,
-0.015001232735812664,
-0.008401276543736458,
0.027379000559449196,
0.009035402908921242,
-0.007770959287881851,
0.060327235609292984,
-0.06848500669002533,
-0.04573259502649307,
0.007085041608661413,
0.03989207372069359,
-0.044217485934495926,
-0.09603114426136017,
-0.10063694417476654,
1.4599530782561123e-34,
-0.0722619816660881,
-0.019278010353446007,
0.009905009530484676,
0.008981901220977306,
0.034145794808864594,
0.03512180224061012,
-0.04747253656387329,
0.06646765023469925,
-0.05954616516828537,
0.005800288636237383,
0.024347934871912003,
-0.03560115769505501,
-0.14426501095294952,
0.09688302129507065,
0.1400241255760193,
-0.036745451390743256,
-0.010396841913461685,
0.12767626345157623,
0.015762608498334885,
-0.03918164223432541,
0.08726651221513748,
-0.028492143377661705,
-0.02886327914893627,
-0.03566897287964821,
-0.021397443488240242,
-0.04684519022703171,
0.09337571263313293,
-0.05393228307366371,
0.040099259465932846,
0.005615826230496168,
-0.06218452751636505,
-0.04794393852353096,
0.07826189696788788,
0.048883117735385895,
0.020860940217971802,
0.031359266489744186,
-0.05793043226003647,
0.008355828002095222,
0.026793692260980606,
0.05506878346204758,
0.014863818883895874,
0.04083317518234253,
0.01476249285042286,
0.01762518845498562,
-0.10064487159252167,
-0.06727801263332367,
0.079686738550663,
0.035439059138298035,
-0.07867491990327835,
-0.059649743139743805,
-0.030307596549391747,
0.07308953255414963,
-0.009069056250154972,
-0.07056692987680435,
-0.04022051766514778,
-0.015390759333968163,
0.07230028510093689,
0.005690878257155418,
0.019746728241443634,
-0.007372159045189619,
0.005281566176563501,
0.08311785012483597,
0.03433844819664955,
0.05508587881922722,
0.013231880962848663,
0.020778821781277657,
-0.030867280438542366,
-0.019342388957738876,
0.024650920182466507,
0.007479316554963589,
-0.05739771947264671,
0.00614318810403347,
-0.0666220560669899,
-0.10767264664173126,
0.07084093987941742,
-0.006030047312378883,
0.05115559697151184,
0.017180265858769417,
0.01467309333384037,
-0.0020367950201034546,
0.01932075433433056,
0.07600931823253632,
-0.04630807787179947,
-0.025497086346149445,
-0.052324045449495316,
-0.0715731680393219,
0.004388989880681038,
0.016972839832305908,
-0.07816600054502487,
0.016147084534168243,
-0.018460551276803017,
-0.02367016300559044,
-0.08930902928113937,
0.016888508573174477,
0.030590837821364403,
-9.808334046810783e-34,
0.008907395415008068,
0.02302631549537182,
-0.024618158116936684,
-0.007603500969707966,
-0.010198974050581455,
-0.05245308205485344,
-0.011911239475011826,
0.027392501011490822,
-0.038759831339120865,
-0.06809628009796143,
-0.00798181165009737,
0.0186000969260931,
-0.01404270064085722,
-0.023488285019993782,
-0.04975154623389244,
0.01115063764154911,
-0.022317688912153244,
-0.03265778720378876,
-0.039823018014431,
0.0184637364000082,
-0.014596779830753803,
0.09330592304468155,
-0.05469246208667755,
0.04185750335454941,
-0.047837529331445694,
0.06695523113012314,
-0.08394766598939896,
0.10924891382455826,
0.03798365220427513,
0.024269986897706985,
-0.0721692219376564,
-0.028402773663401604,
-0.08154790103435516,
-0.03931383043527603,
-0.0973421186208725,
0.037570513784885406,
0.008185912854969501,
0.06396892666816711,
-0.029547015205025673,
0.17942245304584503,
-0.007458601612597704,
-0.001460365136153996,
-0.13146242499351501,
0.08860109746456146,
0.047175049781799316,
-0.07433991879224777,
0.0008271809201687574,
-0.05424081161618233,
0.07162772119045258,
0.05247028172016144,
-0.04505869746208191,
0.019713299348950386,
-0.027727734297513962,
0.03590456396341324,
-0.07852797210216522,
-0.12698617577552795,
-0.019445810467004776,
-0.036354485899209976,
0.08728278428316116,
-0.0014776689931750298,
-0.0166986845433712,
-0.06788108497858047,
-0.058043017983436584,
-0.006430451758205891,
-0.05956011638045311,
-0.03895983472466469,
0.0014775077579542994,
-0.04670688882470131,
0.055522412061691284,
0.03497317433357239,
-0.008206458762288094,
-0.04711554944515228,
0.055121999233961105,
0.028624091297388077,
-0.008287331089377403,
0.013316438533365726,
0.030947770923376083,
0.009059920907020569,
0.05154888331890106,
-0.08377803862094879,
0.06005798652768135,
0.01767745614051819,
0.0786847397685051,
-0.016795961186289787,
0.0597829632461071,
-0.029169054701924324,
0.06442547589540482,
-0.021808598190546036,
0.007223537657409906,
-0.004733765497803688,
-0.023012911900877953,
-0.0057172407396137714,
0.03762212023139,
0.014487131498754025,
-0.05400421842932701,
-4.372519057938007e-8,
-0.03735741600394249,
-0.041975587606430054,
0.1145356073975563,
-0.019664131104946136,
0.03285247087478638,
-0.06021425873041153,
-0.0004778090224135667,
0.008059087209403515,
0.025723114609718323,
-0.046738896518945694,
0.0752776712179184,
-0.019450675696134567,
-0.04654970020055771,
-0.03667636588215828,
-0.03882351517677307,
0.05527372658252716,
0.0456506684422493,
0.018995651975274086,
0.03919786959886551,
-0.006192583590745926,
-0.02070804126560688,
-0.03295794501900673,
-0.04072513431310654,
-0.05771063268184662,
0.04947354272007942,
-0.0384010411798954,
-0.07064071297645569,
-0.005743709858506918,
0.0841052234172821,
0.0810643807053566,
0.0035663044545799494,
-0.02380264736711979,
0.053606897592544556,
-0.005526217631995678,
-0.01266376581043005,
0.08959465473890305,
-0.002599215367808938,
-0.06968177855014801,
-0.05754464864730835,
-0.08346851170063019,
0.06338606774806976,
0.1299511045217514,
-0.027318544685840607,
0.0022736736573278904,
-0.010024315677583218,
0.022940734401345253,
0.12262040376663208,
-0.02761637605726719,
0.03610434755682945,
0.04919041693210602,
0.04812288656830788,
-0.017062652856111526,
-0.012998724356293678,
0.06063742935657501,
0.02900051511824131,
-0.048025671392679214,
0.06955398619174957,
-0.04022775590419769,
0.03155846148729324,
0.10427933931350708,
0.009218025021255016,
0.05318748205900192,
-0.09396513551473618,
-0.045748479664325714
] |
Random Is All You Need: Random Noise Injection on Feature Statistics for Generalizable Deep Image Denoising | https://openreview.net/forum?id=z8PcUSKXXN | [
"Zhengwei Yin",
"Hongjun Wang",
"Guixu Lin",
"Weihang Ran",
"Yinqiang Zheng"
] | Poster | Recent advancements in generalizable deep image denoising have catalyzed the development of robust noise-handling models. The current state-of-the-art, Masked Training (MT), constructs a masked swinir model which is trained exclusively on Gaussian noise ($\sigma$=15) but can achieve commendable denoising performance across various noise types (*i.e.* speckle noise, poisson noise). However, this method, while focusing on content reconstruction, often produces over-smoothed images and poses challenges in mask ratio optimization, complicating its integration with other methodologies. In response, this paper introduces RNINet, a novel architecture built on a streamlined encoder-decoder framework to enhance both efficiency and overall performance. Initially, we train a pure RNINet (only simple encoder-decoder) on individual noise types, observing that feature statistics such as mean and variance shift in response to different noise conditions. Leveraging these insights, we incorporate a noise injection block that injects random noise into feature statistics within our framework, significantly improving generalization across unseen noise types. Our framework not only simplifies the architectural complexity found in MT but also delivers superior performance. Comprehensive experimental evaluations demonstrate that our method outperforms MT in various unseen noise conditions in terms of denoising effectiveness and computational efficiency (lower MACs and GPU memory usage), achieving up to 10 times faster inference speeds and underscoring it's capability for large scale deployments. | Image Denoising, Low-Level Vision, Generalization Problem | null | 1,857 | null | [
-0.05740648880600929,
-0.1519349217414856,
0.09567254781723022,
0.042598482221364975,
0.031350940465927124,
-0.020060550421476364,
0.03602316975593567,
-0.1666286587715149,
-0.03124486654996872,
-0.08436734974384308,
-0.030972493812441826,
-0.005259951576590538,
0.01851591281592846,
0.04603637754917145,
-0.04257631674408913,
0.006016256287693977,
0.051233574748039246,
0.12515504658222198,
-0.08785552531480789,
-0.05354995280504227,
-0.020006736740469933,
0.02062905952334404,
0.02522270940244198,
-0.02553814835846424,
0.08304455876350403,
-0.08228500932455063,
-0.0046553839929401875,
-0.061898041516542435,
0.00873678270727396,
-0.003108031814917922,
0.09391729533672333,
-0.08542128652334213,
0.03198857605457306,
-0.05612267553806305,
0.004210487939417362,
-0.044738322496414185,
-0.059148404747247696,
0.023913532495498657,
0.04173934459686279,
0.0023542286362499,
-0.06103537604212761,
0.029575863853096962,
-0.0467480793595314,
-0.03645476698875427,
0.06453117728233337,
-0.00648342864587903,
0.028869401663541794,
-0.0730135589838028,
0.07294337451457977,
0.043926384299993515,
0.04817454516887665,
0.03822404518723488,
-0.08931718021631241,
0.052758775651454926,
0.060298047959804535,
-0.0009148415992967784,
-0.003203327301889658,
-0.0241810642182827,
-0.004188977647572756,
0.020940320566296577,
0.05272115394473076,
-0.05343184247612953,
0.043066900223493576,
0.001922065857797861,
-0.004774150904268026,
-0.010101749561727047,
0.02079916000366211,
0.014159232378005981,
0.06081477925181389,
-0.02945888601243496,
-0.0623142309486866,
0.05403588339686394,
-0.004277018830180168,
0.045003462582826614,
-0.03163774684071541,
0.05128812789916992,
0.01820155791938305,
-0.003720055567100644,
0.10506618767976761,
0.02222590148448944,
0.07287906110286713,
0.01369569543749094,
0.08324424922466278,
-0.01676037162542343,
0.10909867286682129,
-0.022057291120290756,
0.06161744147539139,
0.053052354604005814,
-0.0376935675740242,
0.0015047506894916296,
-0.063383087515831,
-0.07850194722414017,
-0.04719724878668785,
0.04062438756227493,
-0.02327701449394226,
-0.027887975797057152,
0.013016748242080212,
-0.0611729733645916,
-0.03723233565688133,
0.05874599516391754,
-0.02009548805654049,
-0.0526859425008297,
0.00202603405341506,
-0.049867551773786545,
0.045117609202861786,
-0.010529126040637493,
0.14277786016464233,
0.009267346933484077,
-0.007490326184779406,
0.007891405373811722,
0.00854271650314331,
0.06484775245189667,
-0.028203923255205154,
0.010679920203983784,
0.05288631096482277,
-0.00044628765317611396,
0.01353590376675129,
0.03737884759902954,
-0.005790218245238066,
0.028811929747462273,
-0.007705664727836847,
-0.01578657701611519,
-0.09628545492887497,
0.001315380330197513,
0.056726884096860886,
0.09442321211099625,
-0.007496407255530357,
4.2205584673556824e-33,
0.04616764932870865,
0.03698555752635002,
0.008361185900866985,
0.009574072435498238,
0.09354439377784729,
-0.015524286776781082,
-0.01847303844988346,
0.020464856177568436,
-0.006042678374797106,
0.07595159858465195,
-0.03516596555709839,
-0.07462757080793381,
-0.10487194359302521,
0.0666302740573883,
0.011745697818696499,
-0.0034998126793652773,
-0.07289593666791916,
-0.007643610239028931,
0.05071987211704254,
-0.046142157167196274,
-0.006639724597334862,
0.009605658240616322,
-0.013098174706101418,
-0.006523256190121174,
0.0016281971475109458,
-0.04612914100289345,
0.04078733175992966,
-0.06031162664294243,
0.0009896320989355445,
0.022361844778060913,
-0.12473751604557037,
0.026593487709760666,
0.04566564783453941,
0.07642199099063873,
-0.00994921661913395,
-0.06508153676986694,
0.067154660820961,
-0.026761962100863457,
-0.012122382409870625,
-0.012013617902994156,
0.03240666911005974,
0.07241086661815643,
-0.0816427692770958,
-0.04647434875369072,
-0.016575630754232407,
0.01735076680779457,
0.056994128972291946,
0.029255365952849388,
-0.04484066739678383,
-0.005136174149811268,
0.04512713477015495,
0.02206464111804962,
-0.09570706635713577,
-0.030303413048386574,
0.010396250523626804,
-0.009716729633510113,
0.029490582644939423,
-0.0159568190574646,
0.059090159833431244,
-0.004418230149894953,
-0.03231411427259445,
0.000007426234333252069,
0.01608722284436226,
0.00488588772714138,
0.0093440106138587,
0.03887643665075302,
0.015414788387715816,
0.03711004555225372,
-0.07264278829097748,
-0.08044638484716415,
-0.00782355759292841,
0.06395560503005981,
-0.028269518166780472,
-0.007168066222220659,
-0.03554974123835564,
0.0758809968829155,
0.08459248393774033,
-0.029068736359477043,
0.010239633731544018,
0.02719051204621792,
-0.047701746225357056,
0.024674449115991592,
-0.07202337682247162,
-0.040530938655138016,
-0.06402834504842758,
0.04591089114546776,
-0.008874046616256237,
-0.04167964681982994,
-0.023151051253080368,
-0.051225870847702026,
-0.0473414771258831,
0.02937568537890911,
0.021779818460345268,
-0.05126376077532768,
-0.03917614743113518,
-2.6716134183679928e-33,
-0.03720840811729431,
0.10592526942491531,
-0.07259640842676163,
0.04724385216832161,
-0.019150398671627045,
-0.07617468386888504,
-0.024046704173088074,
-0.012772232294082642,
0.059984706342220306,
-0.05370462313294411,
-0.06500408053398132,
-0.022850550711154938,
0.032606471329927444,
-0.07553572952747345,
-0.04273242875933647,
-0.057251252233982086,
-0.014985260553658009,
-0.006483140867203474,
-0.0802704319357872,
0.03015790320932865,
0.000670617853756994,
0.05474145710468292,
0.01133365835994482,
0.014399075880646706,
-0.053767889738082886,
0.10462727397680283,
-0.07022128254175186,
0.08217322081327438,
0.08752966672182083,
-0.044423412531614304,
-0.0633031353354454,
0.039223819971084595,
0.009797542355954647,
-0.06659743189811707,
-0.05463661625981331,
0.052481554448604584,
0.07160297781229019,
-0.0013297131517902017,
0.006086581386625767,
0.03235355019569397,
0.025816461071372032,
-0.040924765169620514,
0.007200068328529596,
0.059073902666568756,
0.028107335790991783,
-0.030361758545041084,
0.013710994273424149,
-0.047539953142404556,
0.08082883805036545,
0.021018575876951218,
-0.002431369386613369,
0.019472084939479828,
-0.1063743382692337,
0.08824747055768967,
-0.029710354283452034,
-0.027396541088819504,
-0.08759763836860657,
-0.056442756205797195,
0.05961287021636963,
0.11087111383676529,
-0.004535111133009195,
-0.01939522661268711,
-0.10741789638996124,
-0.046879712492227554,
-0.07547253370285034,
-0.008813311345875263,
-0.03356771171092987,
0.018859148025512695,
-0.022355226799845695,
0.058634135872125626,
-0.010578256100416183,
-0.07591728121042252,
0.03786582127213478,
0.062109872698783875,
-0.006006326526403427,
-0.0371975414454937,
0.00898926891386509,
-0.0821254700422287,
0.03168334439396858,
0.009098702110350132,
0.03349871188402176,
-0.031067771837115288,
-0.029686905443668365,
0.0197787843644619,
0.13213880360126495,
0.06360748410224915,
-0.054011158645153046,
-0.07778540998697281,
0.08415782451629639,
-0.025068702176213264,
-0.036631759256124496,
0.03288726136088371,
-0.02157624252140522,
-0.015895221382379532,
-0.024672778323292732,
-4.48771295680217e-8,
0.04084247350692749,
0.04803875833749771,
-0.026947665959596634,
-0.06488645821809769,
0.022343015298247337,
-0.08422140032052994,
0.02600695937871933,
0.06184915825724602,
0.02465003915131092,
-0.11307933926582336,
0.11881893128156662,
-0.03993239253759384,
-0.07424819469451904,
-0.002743673976510763,
0.05916197597980499,
0.03771866485476494,
-0.01667710766196251,
0.05362556502223015,
0.002648106310516596,
-0.05155019834637642,
-0.016713805496692657,
0.019302673637866974,
-0.027913933619856834,
-0.04585213586688042,
0.012211338616907597,
-0.03052198700606823,
0.07378560304641724,
0.07182158529758453,
0.0623348094522953,
0.05054127052426338,
0.0045532239601016045,
0.021557778120040894,
0.08738506585359573,
-0.006109504494816065,
-0.03141830116510391,
0.04884270578622818,
0.08283346891403198,
0.0393986813724041,
-0.02493593655526638,
-0.040318772196769714,
-0.015377718955278397,
-0.0016666839364916086,
0.0018875484820455313,
-0.03555784374475479,
-0.07910577952861786,
0.03796882554888725,
0.04855647310614586,
-0.08127202093601227,
0.04824244603514671,
0.04860639199614525,
0.08425024151802063,
-0.04604857414960861,
0.053669095039367676,
0.06134385988116264,
0.09586437046527863,
0.036073897033929825,
0.014543835073709488,
-0.017337054014205933,
0.02575244940817356,
0.06378567963838577,
-0.04264666512608528,
0.0007670575869269669,
-0.03658374026417732,
-0.030605817213654518
] |
Navigating Neural Space: Revisiting Concept Activation Vectors to Overcome Directional Divergence | https://openreview.net/forum?id=Q95MaWfF4e | [
"Frederik Pahde",
"Maximilian Dreyer",
"Moritz Weckbecker",
"Leander Weber",
"Christopher J. Anders",
"Thomas Wiegand",
"Wojciech Samek",
"Sebastian Lapuschkin"
] | Poster | With a growing interest in understanding neural network prediction strategies, Concept Activation Vectors (CAVs) have emerged as a popular tool for modeling human-understandable concepts in the latent space.
Commonly, CAVs are computed by leveraging linear classifiers optimizing the *separability* of latent representations of samples with and without a given concept. However, in this paper we show that such a separability-oriented computation leads to solutions, which may diverge from the actual goal of precisely modeling the concept direction.
This discrepancy can be attributed to the significant influence of distractor directions, i.e., signals unrelated to the concept, which are picked up by filters (i.e., weights) of linear models to optimize class-separability.
To address this, we introduce *pattern-based CAVs*, solely focussing on concept signals, thereby providing more accurate concept directions.
We evaluate various CAV methods in terms of their alignment with the true concept direction and their impact on CAV applications, including concept sensitivity testing and model correction for shortcut behavior caused by data artifacts.
We demonstrate the benefits of pattern-based CAVs using the Pediatric Bone Age, ISIC2019, and FunnyBirds datasets with VGG, ResNet, ReXNet, EfficientNet, and Vision Transformer as model architectures. | Explainable AI, Concept-based Explanations, Concept Activation Vectors | We introduce pattern-based CAVs, an alternative to widely used filter (e.g., SVM) CAVs, more robust to distractor patterns and thereby providing more accurate concept directions. | 1,856 | 2202.03482 | [
-0.03916603699326515,
-0.18020641803741455,
0.004779677838087082,
0.02285563386976719,
0.0615910105407238,
-0.018665703013539314,
0.0727563425898552,
-0.08035445958375931,
0.012689635157585144,
0.0027089782524853945,
-0.025984695181250572,
-0.05217451974749565,
0.014138801023364067,
0.03778303042054176,
0.039172928780317307,
-0.08365622162818909,
0.08068711310625076,
0.08029044419527054,
-0.10249122232198715,
-0.00728942733258009,
0.11829046905040741,
-0.009234994649887085,
-0.05027088150382042,
0.01350302156060934,
0.034653376787900925,
0.09051197022199631,
0.010136856697499752,
0.005154830869287252,
0.05174415931105614,
-0.02476145699620247,
0.05049440264701843,
0.026248833164572716,
-0.010299008339643478,
0.01770792528986931,
-0.1298944354057312,
0.04576313868165016,
-0.11020015925168991,
0.10766690224409103,
-0.06107800826430321,
0.0034486392978578806,
-0.013238824903964996,
-0.003276899689808488,
-0.001929865567944944,
0.10158149898052216,
0.05161263793706894,
0.021087301895022392,
-0.03681402653455734,
-0.12197349965572357,
0.045011114329099655,
0.004157173912972212,
-0.06322506815195084,
0.024449406191706657,
-0.023613175377249718,
0.06259499490261078,
-0.05836065858602524,
0.001995205646380782,
0.030809564515948296,
0.07483606040477753,
-0.007648744620382786,
-0.025116585195064545,
0.00477308826521039,
-0.08257027715444565,
0.054592881351709366,
-0.05438440665602684,
-0.02650446444749832,
-0.027972588315606117,
-0.0277659110724926,
0.024874906986951828,
0.037081532180309296,
0.04896457493305206,
-0.004766189027577639,
0.12410720437765121,
-0.048896849155426025,
0.015348480083048344,
0.024944597855210304,
0.08127757906913757,
0.027976833283901215,
0.008901289664208889,
0.027772463858127594,
-0.040846675634384155,
-0.00865787174552679,
0.011035979725420475,
-0.009473828598856926,
0.02773251384496689,
0.1395169496536255,
0.11615365743637085,
-0.015571693889796734,
0.02754628285765648,
-0.001303558237850666,
-0.09019105136394501,
-0.005492766387760639,
-0.06387601047754288,
0.062293604016304016,
-0.050646014511585236,
0.07878964394330978,
-0.0007318294956348836,
0.004844233859330416,
-0.07623070478439331,
-0.041394609957933426,
0.03597049042582512,
-0.03749890625476837,
0.05097286030650139,
-0.09155368059873581,
-0.04568801075220108,
0.05810980498790741,
-0.009825863875448704,
0.01984533667564392,
-0.002116213319823146,
0.04397675767540932,
-0.060438040643930435,
-0.04601406306028366,
0.05157484859228134,
-0.03854746371507645,
0.004173173103481531,
0.04176466912031174,
-0.007882165722548962,
0.05712030082941055,
-0.09776642173528671,
0.07420217245817184,
0.06472590565681458,
-0.06434585154056549,
0.04210599511861801,
-0.03460351750254631,
-0.02112266793847084,
0.050104204565286636,
-0.01976155675947666,
-0.08373118191957474,
1.892838860056627e-33,
-0.07742926478385925,
0.03126567229628563,
0.02355007641017437,
0.06240947172045708,
0.08782383054494858,
-0.024789845570921898,
-0.016431907191872597,
0.0823020413517952,
0.03257565200328827,
0.024724304676055908,
-0.03367961198091507,
0.03570114076137543,
-0.04214540123939514,
0.0727895051240921,
0.07793553173542023,
0.0387633815407753,
-0.070311538875103,
-0.045920222997665405,
-0.09936463087797165,
-0.0795149877667427,
0.002210640348494053,
0.010849828831851482,
0.013720611110329628,
0.011051759123802185,
-0.0054954891093075275,
0.03497292473912239,
0.041119351983070374,
-0.04867949336767197,
-0.08038020133972168,
-0.0123048210516572,
-0.11627325415611267,
0.054277628660202026,
0.058534953743219376,
-0.01871919445693493,
0.01470755971968174,
0.048163123428821564,
0.019597617909312248,
-0.029954174533486366,
0.04413389414548874,
0.005616222973912954,
-0.06466444581747055,
0.0029309922829270363,
-0.019827967509627342,
-0.023705635219812393,
-0.07364732772111893,
-0.07813574373722076,
0.018413353711366653,
0.04488939419388771,
0.05427410453557968,
0.024320565164089203,
0.017788322642445564,
0.02915126644074917,
-0.0014835235197097063,
-0.09091118723154068,
-0.06776785850524902,
0.024299852550029755,
-0.020112568512558937,
0.07879798859357834,
0.0074816253036260605,
0.014743996784090996,
-0.059082068502902985,
0.022995110601186752,
0.06132488325238228,
0.08691879361867905,
-0.01874227449297905,
0.00895568635314703,
-0.028389060869812965,
0.11239593476057053,
0.03965187072753906,
0.020619355142116547,
-0.025101441890001297,
0.06965763866901398,
-0.08024270832538605,
-0.11888159811496735,
0.06547439098358154,
0.03983241319656372,
-0.008075651712715626,
-0.11182719469070435,
-0.0060731712728738785,
-0.014754576608538628,
0.022431327030062675,
-0.02706753835082054,
0.047779835760593414,
-0.02749740332365036,
-0.06974244117736816,
0.006995893083512783,
0.11237060278654099,
-0.045140527188777924,
-0.07232137024402618,
0.0011419352376833558,
-0.054380618035793304,
0.014627719298005104,
0.022186068817973137,
0.002292538760229945,
-0.0356525294482708,
-3.4416770125267886e-33,
-0.009866267442703247,
0.03065423294901848,
-0.04748869314789772,
-0.028944144025444984,
0.023051468655467033,
-0.00384185160510242,
-0.01246198732405901,
-0.014672375284135342,
-0.07073964923620224,
-0.01158153172582388,
-0.04418430104851723,
-0.03640944883227348,
0.019809739664196968,
0.0145422937348485,
-0.007662005722522736,
0.01027312595397234,
-0.03091726079583168,
0.05921010300517082,
0.02422540821135044,
-0.011710730381309986,
0.03249600902199745,
0.04869783669710159,
-0.12318629771471024,
0.016192134469747543,
-0.03897733986377716,
-0.015199288725852966,
-0.013586720451712608,
0.06752927601337433,
-0.03373747691512108,
-0.019765036180615425,
-0.09141583740711212,
0.025334667414426804,
0.04052497446537018,
0.009894363582134247,
-0.019150473177433014,
0.04808926209807396,
-0.031174838542938232,
-0.03927512839436531,
0.006316476967185736,
0.07518800348043442,
0.002224376890808344,
-0.04802374169230461,
-0.07173427194356918,
-0.05101298540830612,
-0.03073345310986042,
-0.04173124581575394,
0.008539916947484016,
0.05324834585189819,
0.007850098423659801,
0.06078128516674042,
-0.008062890730798244,
0.007467597723007202,
-0.094328872859478,
0.04989266395568848,
-0.041345421224832535,
0.0840994194149971,
-0.012601812370121479,
0.007320475298911333,
0.06091796234250069,
0.04311217740178108,
0.011712933890521526,
0.03260687738656998,
-0.0034634980838745832,
-0.08072574436664581,
0.02643878199160099,
-0.04971730336546898,
-0.03781118616461754,
0.006536912638694048,
0.08465688675642014,
-0.06827054917812347,
-0.015832137316465378,
0.029717452824115753,
0.016714442521333694,
-0.04147423058748245,
-0.028973599895834923,
-0.037065308541059494,
0.07691986113786697,
-0.03621762618422508,
-0.014308634214103222,
-0.032249417155981064,
0.016068367287516594,
-0.04120723903179169,
0.007614981848746538,
0.03998212143778801,
-0.015207568183541298,
0.12328261882066727,
-0.03424820303916931,
0.06160321831703186,
0.05582081526517868,
-0.026719730347394943,
-0.048554085195064545,
0.012916303239762783,
-0.05108218640089035,
0.06037355214357376,
-0.061079323291778564,
-4.99100032413935e-8,
-0.07492177188396454,
0.028250351548194885,
0.014169087633490562,
0.05069887638092041,
-0.018117625266313553,
-0.06550680100917816,
-0.009144344367086887,
0.0726376548409462,
-0.06508985161781311,
-0.012242968194186687,
0.05238472297787666,
-0.024719107896089554,
-0.029796792194247246,
0.008331595920026302,
0.02950366772711277,
0.014021203853189945,
-0.005183847155421972,
0.01590750925242901,
0.0013433318817988038,
0.034291040152311325,
0.04540299251675606,
-0.04201480746269226,
0.020192323252558708,
0.04779782518744469,
0.0028000730089843273,
-0.055835966020822525,
-0.06530975550413132,
0.17268073558807373,
0.03941938653588295,
-0.027775214985013008,
0.06312022358179092,
0.04709320142865181,
-0.043457213789224625,
-0.02809235267341137,
0.00034387820051051676,
0.1013672947883606,
0.05707184597849846,
0.020547660067677498,
0.001139158965088427,
0.008355895057320595,
0.015919608995318413,
0.009345878846943378,
-0.013201532885432243,
0.04327904060482979,
0.01637241244316101,
0.04078865051269531,
-0.0013557245256379247,
-0.0971367359161377,
0.0247320719063282,
-0.0659395083785057,
-0.02494579553604126,
0.04679675027728081,
-0.012946865521371365,
0.0636303722858429,
0.014065813273191452,
-0.015798434615135193,
-0.04176366329193115,
-0.025503722950816154,
-0.05454983934760094,
-0.030676262453198433,
0.05336792767047882,
0.054997123777866364,
-0.0386136956512928,
0.0034698594827204943
] |
LLaVA-MoD: Making LLaVA Tiny via MoE-Knowledge Distillation | https://openreview.net/forum?id=uWtLOy35WD | [
"Fangxun Shu",
"Yue Liao",
"Lei Zhang",
"Le Zhuo",
"Chenning Xu",
"Guanghao Zhang",
"Haonan Shi",
"Long Chan",
"TaoZhong",
"Zhelun Yu",
"Wanggui He",
"Siming Fu",
"Haoyuan Li",
"Si Liu",
"Hongsheng Li",
"Hao Jiang"
] | Poster | We introduce LLaVA-MoD, a novel framework designed to enable the efficient training of small-scale Multimodal Language Models ($s$-MLLM) distilling knowledge from large-scale MLLM ($l$-MLLM). Our approach tackles two fundamental challenges in MLLM distillation. First, we optimize the network structure of $s$-MLLM by integrating a sparse Mixture of Experts (MoE) architecture into the language model, striking a balance between computational efficiency and model expressiveness. Second, we propose a progressive knowledge transfer strategy for comprehensive knowledge transfer. This strategy begins with mimic distillation, where we minimize the Kullback-Leibler (KL) divergence between output distributions to enable $s$-MLLM to emulate $s$-MLLM's understanding. Following this, we introduce preference distillation via Preference Optimization (PO), where the key lies in treating $l$-MLLM as the reference model. During this phase, the $s$-MLLM's ability to discriminate between superior and inferior examples is significantly enhanced beyond $l$-MLLM, leading to a better $s$-MLLM that surpasses $l$-MLLM, particularly in hallucination benchmarks.
Extensive experiments demonstrate that LLaVA-MoD surpasses existing works across various benchmarks while maintaining a minimal activated parameters and low computational costs. Remarkably, LLaVA-MoD-2B surpasses Qwen-VL-Chat-7B with an average gain of 8.8\%, using merely $0.3\%$ of the training data and 23\% trainable parameters. The results underscore LLaVA-MoD's ability to effectively distill comprehensive knowledge from its teacher model, paving the way for developing efficient MLLMs. | MLLM, MoE, Distillation | null | 1,853 | null | [
0.014784532599151134,
-0.11139390617609024,
0.09586983174085617,
0.0321219339966774,
0.0200115405023098,
-0.029115920886397362,
0.012573901563882828,
0.0015767557779327035,
0.011199679225683212,
-0.03004654124379158,
-0.03346264734864235,
0.007913121953606606,
0.045204125344753265,
0.009513413533568382,
-0.0055043636821210384,
0.033663030713796616,
0.08486558496952057,
0.12162280082702637,
-0.15024633705615997,
-0.020434586331248283,
0.016482608392834663,
-0.0364786721765995,
-0.058206453919410706,
-0.04417547583580017,
0.09506774693727493,
0.07433254271745682,
-0.029964419081807137,
-0.034580521285533905,
0.07482337206602097,
-0.08982398360967636,
0.021443743258714676,
0.014972923323512077,
-0.02323490008711815,
-0.01469702459871769,
-0.06717557460069656,
0.10319754481315613,
-0.08498752117156982,
0.08990347385406494,
0.041497185826301575,
-0.027170972898602486,
-0.024614529684185982,
0.013515608385205269,
0.00033599245944060385,
0.03794875368475914,
0.019946476444602013,
0.04803372919559479,
-0.022264840081334114,
0.04096359387040138,
0.03469032794237137,
0.011804777197539806,
-0.05509866401553154,
-0.0576988160610199,
-0.0027240687049925327,
0.05320471525192261,
0.044150952249765396,
-0.0214424729347229,
-0.003095954190939665,
-0.003007906023412943,
-0.06025223806500435,
-0.00286913407035172,
-0.0730915442109108,
-0.07918474823236465,
-0.05414549633860588,
-0.008032452315092087,
0.02578975446522236,
-0.05460347980260849,
-0.0139609444886446,
0.07992641627788544,
0.0072414083406329155,
0.01884346827864647,
0.011999936774373055,
0.04332754760980606,
-0.02328486554324627,
0.08281658589839935,
0.028883231803774834,
0.03935348242521286,
0.08233360201120377,
0.039192430675029755,
0.01571587473154068,
-0.011422703042626381,
0.060424480587244034,
0.033213526010513306,
-0.019683577120304108,
-0.058713313192129135,
0.08411025255918503,
-0.0019112530862912536,
-0.02455354854464531,
-0.003975153435021639,
0.09388379752635956,
-0.039409056305885315,
-0.07061676681041718,
0.02147054672241211,
0.0016195642529055476,
-0.047489989548921585,
0.01824387162923813,
0.04081137478351593,
0.03316100314259529,
-0.04262859374284744,
-0.061142537742853165,
0.14676348865032196,
0.03685469925403595,
0.0963376984000206,
0.01058483961969614,
-0.049681536853313446,
-0.0395221970975399,
-0.06483796238899231,
0.010653357952833176,
0.045861050486564636,
0.028896447271108627,
-0.09148604422807693,
-0.0035072893369942904,
0.09919571876525879,
0.032830290496349335,
0.0309936311095953,
0.03591397777199745,
-0.08656663447618484,
0.03971001133322716,
-0.014677041210234165,
0.0127339456230402,
0.02653188444674015,
-0.0538749173283577,
-0.016194036230444908,
0.012606779113411903,
0.010823264718055725,
0.01785176247358322,
0.07335520535707474,
-0.03920195996761322,
1.3713689948807208e-33,
0.07641046494245529,
0.12116728723049164,
0.01897497847676277,
0.040762271732091904,
0.014732410199940205,
-0.020928217098116875,
0.01797332614660263,
0.003182566026225686,
-0.08672347664833069,
-0.03665895760059357,
-0.0575186125934124,
0.024622095748782158,
-0.10533716529607773,
0.04987866058945656,
0.027181867510080338,
0.01804167777299881,
0.005112010519951582,
0.01646433211863041,
0.021887145936489105,
-0.10033652186393738,
0.03378237783908844,
0.0000027527289603312965,
-0.039466820657253265,
0.019786730408668518,
-0.05421367660164833,
0.00038447428960353136,
0.11653170734643936,
-0.05048119276762009,
-0.04410937801003456,
0.025005171075463295,
-0.08819234371185303,
-0.041092969477176666,
0.005003846250474453,
-0.004839194007217884,
-0.007650549523532391,
-0.08165348321199417,
-0.02572491019964218,
-0.04614335298538208,
0.024993859231472015,
-0.02751097083091736,
-0.02444862574338913,
-0.019414354115724564,
-0.052778251469135284,
-0.0015598211903125048,
-0.08149140328168869,
-0.006259117741137743,
0.0749860480427742,
0.0061990064568817616,
-0.05712215229868889,
-0.06901206821203232,
0.033427417278289795,
-0.02932782843708992,
-0.014683146961033344,
-0.030799418687820435,
-0.02514202892780304,
-0.0274640079587698,
0.002125565893948078,
0.05034976825118065,
0.0885152593255043,
0.016085583716630936,
-0.05107681825757027,
0.09904440492391586,
0.0032467269338667393,
0.08930469304323196,
0.09896864742040634,
0.020375989377498627,
-0.0757756307721138,
0.0257524773478508,
0.09804850816726685,
-0.040325067937374115,
-0.07043084502220154,
-0.004249710123986006,
-0.02608909085392952,
-0.055521611124277115,
0.023589126765727997,
-0.036897070705890656,
0.005029196850955486,
-0.05660516396164894,
0.12961365282535553,
0.02077893726527691,
0.029601113870739937,
0.004630991257727146,
-0.04993271455168724,
-0.061377499252557755,
-0.01977650821208954,
-0.004866325296461582,
0.06418987363576889,
-0.07834579795598984,
-0.007803982123732567,
-0.06663532555103302,
-0.034357257187366486,
-0.005909436848014593,
0.012171694077551365,
-0.007414058782160282,
0.03977179527282715,
-8.528200494956157e-34,
0.005742232780903578,
-0.015331457369029522,
0.01581505872309208,
0.13904118537902832,
0.0276862271130085,
-0.03423111513257027,
-0.005844800733029842,
0.04698171094059944,
-0.09271799772977829,
-0.02489824779331684,
-0.02942407689988613,
-0.10564864426851273,
0.003319523297250271,
-0.05501187592744827,
0.034656450152397156,
0.016055116429924965,
-0.017863301560282707,
-0.002655146876350045,
0.009906620718538761,
0.05677598714828491,
-0.02197486348450184,
0.03531073033809662,
-0.07698877155780792,
0.09241458028554916,
-0.04124554991722107,
0.018379559740424156,
-0.021030807867646217,
0.11330411583185196,
-0.020028239116072655,
0.029225023463368416,
-0.00021642856881953776,
-0.03258100897073746,
-0.0172218419611454,
-0.08553744107484818,
-0.06440142542123795,
0.044063810259103775,
-0.013143150135874748,
0.01604650728404522,
-0.020402824506163597,
0.09121650457382202,
-0.005572227761149406,
-0.0524374395608902,
-0.04649744927883148,
-0.0668780505657196,
-0.017473356798291206,
-0.03660498559474945,
-0.11207380890846252,
-0.09995391964912415,
0.07746368646621704,
0.0019203172996640205,
0.1229761466383934,
0.02981649525463581,
-0.07896038889884949,
0.0012264623073861003,
0.00605421606451273,
-0.05729672312736511,
0.05798126757144928,
0.030442414805293083,
0.016011780127882957,
0.02395687624812126,
-0.06450844556093216,
-0.12721982598304749,
-0.0013422874035313725,
-0.08208738267421722,
0.006325147580355406,
-0.03095836006104946,
0.039771534502506256,
0.04453844949603081,
0.01838180609047413,
-0.04258323460817337,
-0.03067627362906933,
-0.04434242472052574,
0.051873162388801575,
-0.019408978521823883,
0.039077598601579666,
0.0045212553814053535,
0.013797146268188953,
0.01013421081006527,
0.005367935169488192,
0.006149337161332369,
-0.035893283784389496,
-0.018152350559830666,
0.02353384718298912,
0.01871982216835022,
0.09774643182754517,
0.0414138026535511,
0.04848771542310715,
0.06402269750833511,
0.01504746824502945,
0.01950841024518013,
-0.026360882446169853,
0.009717138484120369,
0.08004197478294373,
0.036709047853946686,
0.002966740168631077,
-4.638253869870823e-8,
-0.06687796860933304,
-0.011757235042750835,
0.0047034695744514465,
0.06186152994632721,
0.040394384413957596,
-0.038778290152549744,
-0.05991844832897186,
0.06424055993556976,
-0.014780555851757526,
0.012732988223433495,
0.0841219425201416,
0.038940880447626114,
-0.03366590291261673,
-0.02084086462855339,
0.06002301350235939,
0.07979314774274826,
0.04868626594543457,
0.007784438785165548,
-0.013880130834877491,
-0.03016861155629158,
0.11239135265350342,
-0.034038834273815155,
0.03744366765022278,
-0.0412423349916935,
0.10336214303970337,
-0.10295438021421432,
-0.043244801461696625,
0.09409021586179733,
0.04886702820658684,
0.0013922700891271234,
-0.08043613284826279,
-0.025151489302515984,
-0.04306705296039581,
-0.02972176857292652,
0.014848302118480206,
0.08951742947101593,
0.029730232432484627,
-0.10011232644319534,
-0.030701568350195885,
-0.031192108988761902,
0.02245352789759636,
0.016863299533724785,
-0.06908757239580154,
0.042225491255521774,
-0.006718115881085396,
-0.01106566097587347,
-0.005026342812925577,
-0.04916875436902046,
0.0022435111459344625,
-0.007686130702495575,
0.034584566950798035,
-0.04823305457830429,
0.02075091376900673,
0.05494111031293869,
0.04516125097870827,
-0.020571205765008926,
-0.029122209176421165,
0.01610245555639267,
0.01698245294392109,
-0.00382341630756855,
0.06608038395643234,
0.036110784858465195,
-0.011701268143951893,
-0.022626925259828568
] |
I2VControl-Camera: Precise Video Camera Control with Adjustable Motion Strength | https://openreview.net/forum?id=AcAD4VEgCX | [
"Wanquan Feng",
"Jiawei Liu",
"Pengqi Tu",
"Tianhao Qi",
"Mingzhen Sun",
"Tianxiang Ma",
"Songtao Zhao",
"SiYu Zhou",
"Qian HE"
] | Poster | Video generation technologies are developing rapidly and have broad potential applications. Among these technologies, camera control is crucial for generating professional-quality videos that accurately meet user expectations. However, existing camera control methods still suffer from several limitations, including control precision and the neglect of the control for subject motion dynamics. In this work, we propose I2VControl-Camera, a novel camera control method that significantly enhances controllability while providing adjustability over the strength of subject motion. To improve control precision, we employ point trajectory in the camera coordinate system instead of only extrinsic matrix information as our control signal. To accurately control and adjust the strength of subject motion, we explicitly model the higher-order components of the video trajectory expansion, not merely the linear terms, and design an operator that effectively represents the motion strength. We use an adapter architecture that is independent of the base model structure. Experiments on static and dynamic scenes show that our framework outperformances previous methods both quantitatively and qualitatively. Project page: https://wanquanf.github.io/I2VControlCamera. | Video Generation, Camera Control | Precise Video Camera Control with Adjustable Motion Strength | 1,844 | 2411.06525 | [
-0.022503558546304703,
-0.08858609199523926,
-0.02884714864194393,
-0.06141338124871254,
0.03849026560783386,
-0.005437151528894901,
-0.008902892470359802,
0.04212849959731102,
0.04140552505850792,
0.008927402086555958,
0.04096135497093201,
-0.02535412460565567,
-0.047578033059835434,
0.08279445022344589,
0.008536874316632748,
-0.004257757216691971,
0.08058017492294312,
0.12069929391145706,
-0.05385740101337433,
0.04836400970816612,
0.033590029925107956,
-0.04623685032129288,
0.0911443829536438,
-0.03927358612418175,
0.0029003394301980734,
0.0010109099093824625,
-0.014142196625471115,
-0.023330386728048325,
0.030877646058797836,
-0.05769232660531998,
0.022563882172107697,
-0.018239755183458328,
-0.00968780554831028,
-0.011064833961427212,
-0.09475556761026382,
-0.04371579736471176,
0.062251828610897064,
-0.0008017360814847052,
-0.059629909694194794,
0.025732269510626793,
0.007174061611294746,
0.11075057089328766,
0.1169029101729393,
0.0016069246921688318,
0.08364670723676682,
0.10553843528032303,
0.09532161056995392,
-0.018721727654337883,
-0.027059154585003853,
0.05184633284807205,
-0.0874251052737236,
-0.020112257450819016,
0.11158481240272522,
-0.0437689907848835,
-0.004413638263940811,
-0.00012138109013903886,
-0.005577112548053265,
0.0029689643997699022,
0.06837008893489838,
-0.02559160441160202,
0.05804150179028511,
0.0047454978339374065,
0.016123270615935326,
0.009304918348789215,
-0.05711924657225609,
-0.06324741244316101,
0.09215070307254791,
0.001270800712518394,
0.015481386333703995,
0.059044644236564636,
-0.12444427609443665,
-0.0017764769727364182,
0.02279563806951046,
0.02871711365878582,
-0.03967025503516197,
0.0018109917873516679,
0.027926601469516754,
-0.0199227724224329,
0.02314702793955803,
-0.019894229248166084,
0.09713875502347946,
-0.07442492991685867,
-0.03257480263710022,
0.014179670251905918,
-0.037996258586645126,
-0.009003627113997936,
-0.002008328679949045,
0.06968490779399872,
0.04872830957174301,
0.04991777986288071,
-0.08998053520917892,
-0.08936893194913864,
0.018799209967255592,
-0.13765418529510498,
0.03971099108457565,
-0.03517909348011017,
0.007517264690250158,
-0.07648773491382599,
-0.010758092626929283,
0.01191389188170433,
0.03215066343545914,
-0.05872330442070961,
-0.040347084403038025,
0.042515821754932404,
0.0899011492729187,
0.0046047307550907135,
-0.00940792914479971,
0.03604111447930336,
0.016936250030994415,
0.04929528012871742,
0.0034493703860789537,
-0.003204519394785166,
-0.07690297067165375,
-0.0061214459128677845,
0.08595776557922363,
0.08134172111749649,
0.01799716427922249,
-0.0014136547688394785,
-0.0065999883227050304,
-0.006193122826516628,
0.01697286032140255,
-0.06706806272268295,
-0.05441298335790634,
0.03398079797625542,
0.004564689937978983,
-0.008927784860134125,
0.010085643269121647,
9.150789139254681e-34,
-0.013027418404817581,
0.0603807233273983,
0.037419382482767105,
0.1157442256808281,
0.03986794129014015,
-0.011644504964351654,
0.006395623087882996,
0.013436424545943737,
-0.026232963427901268,
-0.030715886503458023,
0.0675540417432785,
-0.08821509778499603,
-0.03107825107872486,
0.0744292289018631,
0.05131852626800537,
-0.02769182063639164,
0.0010739959543570876,
-0.03906155005097389,
0.017411628738045692,
0.08526710420846939,
0.060142166912555695,
0.009144200012087822,
-0.02832173742353916,
-0.012845903635025024,
-0.04708129167556763,
0.02735247276723385,
0.08193350583314896,
0.061889491975307465,
0.03445132449269295,
-0.057903025299310684,
0.03651442751288414,
-0.06294077634811401,
-0.017753880470991135,
-0.029419908300042152,
0.046270571649074554,
-0.024540679529309273,
-0.05332084000110626,
-0.06832024455070496,
0.06656160205602646,
0.03769630193710327,
-0.010990343056619167,
0.09269466996192932,
-0.0684894472360611,
-0.012863056734204292,
-0.04823185130953789,
-0.037945594638586044,
0.06873244792222977,
0.15822061896324158,
-0.10535858571529388,
0.00915401428937912,
0.032653190195560455,
0.033583592623472214,
-0.03420691564679146,
-0.07508821785449982,
-0.02773728035390377,
-0.020760107785463333,
0.0069202035665512085,
-0.004970648791640997,
-0.02792268432676792,
0.051739923655986786,
-0.0431031696498394,
-0.03797086700797081,
-0.007509495597332716,
-0.019798550754785538,
-0.005117189139127731,
-0.023365970700979233,
-0.0002896079677157104,
0.10034302622079849,
0.030863439664244652,
0.039639003574848175,
-0.01125440001487732,
0.004486718215048313,
0.005305144935846329,
-0.08346427977085114,
0.04161815717816353,
0.05281203240156174,
-0.06390900164842606,
-0.06207643449306488,
-0.04797119274735451,
0.07752089202404022,
-0.10059402883052826,
0.04879993945360184,
0.07229238003492355,
0.06407546997070312,
-0.015526480972766876,
-0.05108509212732315,
-0.032236091792583466,
-0.036193858832120895,
-0.07577715814113617,
-0.0907687097787857,
-0.003751829033717513,
-0.029494192451238632,
-0.001321844756603241,
0.013156656175851822,
0.0071356394328176975,
-4.339560556316138e-34,
0.05109754949808121,
0.01332078780978918,
-0.025444474071264267,
0.012392651289701462,
-0.047322534024715424,
-0.09540577232837677,
0.04618534445762634,
0.008898204192519188,
0.07454080134630203,
-0.05743538215756416,
-0.007146740797907114,
-0.06812256574630737,
-0.04620999097824097,
0.04994959384202957,
0.06440419703722,
-0.009537911973893642,
-0.07739391177892685,
-0.03709624335169792,
-0.046737655997276306,
-0.07440413534641266,
0.03215354308485985,
0.060830216854810715,
0.0685083195567131,
-0.023325812071561813,
0.01348878350108862,
0.04862763732671738,
0.022909823805093765,
0.1020636111497879,
0.05313261225819588,
-0.0590742789208889,
-0.039835818111896515,
-0.05425231531262398,
0.03628367558121681,
-0.0030287890695035458,
-0.0388525016605854,
0.0527864471077919,
-0.006564909126609564,
-0.06295771896839142,
-0.04343400523066521,
0.08259371668100357,
-0.009426693432033062,
0.01873915083706379,
-0.01902569830417633,
-0.0021025817841291428,
0.009247051551938057,
0.036096904426813126,
-0.008502734825015068,
0.07695426046848297,
-0.06055907905101776,
0.020530082285404205,
-0.13322865962982178,
0.04648842662572861,
-0.09695126116275787,
0.04716718941926956,
-0.010673366487026215,
-0.041250232607126236,
-0.05411190167069435,
0.0007608768646605313,
0.0897752121090889,
-0.0005747780669480562,
-0.0531090684235096,
0.03682342916727066,
-0.08043049275875092,
-0.005377566907554865,
0.01699453592300415,
0.04424876347184181,
-0.0162823349237442,
-0.03128466755151749,
0.0063428692519664764,
0.05919273570179939,
0.09629400819540024,
-0.049500852823257446,
0.06569739431142807,
-0.006997149437665939,
0.034466471523046494,
-0.0757962167263031,
0.02574453316628933,
-0.019023539498448372,
0.05358610302209854,
-0.09380722790956497,
-0.0198346097022295,
-0.1018841564655304,
0.07242904603481293,
0.04632287099957466,
0.00836099311709404,
0.09106089174747467,
-0.027040518820285797,
0.019752327352762222,
0.00894218310713768,
0.05711403116583824,
-0.004080246668308973,
0.043980855494737625,
0.0014095548540353775,
0.02297941781580448,
-0.059951286762952805,
-4.058339797552435e-8,
-0.15969975292682648,
0.009401140734553337,
0.022450512275099754,
0.035940393805503845,
-0.037586409598588943,
-0.061170436441898346,
0.04547545686364174,
0.03995591029524803,
0.06473314017057419,
-0.04204775020480156,
0.010237259790301323,
-0.012750576250255108,
0.046566519886255264,
-0.029154054820537567,
-0.009882364422082901,
0.02781233750283718,
0.01490089949220419,
0.01333653461188078,
-0.021960197016596794,
0.04192417487502098,
-0.0477975569665432,
-0.021782800555229187,
-0.01711120642721653,
-0.03736129775643349,
-0.0379076823592186,
-0.0029543235432356596,
-0.06882832199335098,
0.06028377264738083,
0.07284258306026459,
0.0012256484478712082,
0.019201045855879784,
0.012843585573136806,
0.05156290531158447,
0.005781867541372776,
-0.053093403577804565,
-0.013633216731250286,
-0.0749933272600174,
-0.04482610896229744,
-0.00923192035406828,
0.008759227581322193,
-0.032668907195329666,
0.0817563459277153,
0.00943795032799244,
0.07285166531801224,
0.03143228217959404,
0.08466634899377823,
0.04796844720840454,
-0.056406497955322266,
0.0029671124648302794,
-0.053225137293338776,
-0.05962828919291496,
0.04557105898857117,
-0.028616763651371002,
0.07231315970420837,
0.04003644734621048,
0.01935381256043911,
0.07785994559526443,
-0.0442887507379055,
-0.012739329598844051,
0.04314596578478813,
-0.04015075042843819,
0.02786394953727722,
-0.06937968730926514,
-0.024108344689011574
] |
BinaryDM: Accurate Weight Binarization for Efficient Diffusion Models | https://openreview.net/forum?id=YaeZwhXJ4k | [
"Xingyu Zheng",
"Xianglong Liu",
"Haotong Qin",
"Xudong Ma",
"Mingyuan Zhang",
"Haojie Hao",
"Jiakai Wang",
"Zixiang Zhao",
"Jinyang Guo",
"Michele Magno"
] | Poster | With the advancement of diffusion models (DMs) and the substantially increased computational requirements, quantization emerges as a practical solution to obtain compact and efficient low-bit DMs. However, the highly discrete representation leads to severe accuracy degradation, hindering the quantization of diffusion models to ultra-low bit-widths. This paper proposes a novel weight binarization approach for DMs, namely BinaryDM, pushing binarized DMs to be accurate and efficient by improving the representation and optimization. From the representation perspective, we present an Evolvable-Basis Binarizer (EBB) to enable a smooth evolution of DMs from full-precision to accurately binarized. EBB enhances information representation in the initial stage through the flexible combination of multiple binary bases and applies regularization to evolve into efficient single-basis binarization. The evolution only occurs in the head and tail of the DM architecture to retain the stability of training. From the optimization perspective, a Low-rank Representation Mimicking (LRM) is applied to assist the optimization of binarized DMs. The LRM mimics the representations of full-precision DMs in low-rank space, alleviating the direction ambiguity of the optimization process caused by fine-grained alignment. Comprehensive experiments demonstrate that BinaryDM achieves significant accuracy and efficiency gains compared to SOTA quantization methods of DMs under ultra-low bit-widths. With 1-bit weight and 4-bit activation (W1A4), BinaryDM achieves as low as 7.74 FID and saves the performance from collapse (baseline FID 10.87). As the first binarization method for diffusion models, W1A4 BinaryDM achieves impressive 15.2x OPs and 29.2x model size savings, showcasing its substantial potential for edge deployment. | Model Quantization, Model Compression, Generative Model, Diffusion Model | In this paper, we propose BinaryDM, which pushes the quantization of the diffusion model to the limit with binarized weight. | 1,841 | 2404.05662 | [
0.03132452443242073,
-0.03099728189408779,
-0.010500187054276466,
0.00840752013027668,
-0.015500102192163467,
-0.04361807554960251,
0.007445554714649916,
-0.041992634534835815,
-0.002965915249660611,
-0.019791798666119576,
-0.030647631734609604,
-0.03714699670672417,
-0.014562984928488731,
0.05067460611462593,
-0.027010804042220116,
-0.026203010231256485,
-0.01524263620376587,
0.0968601405620575,
-0.07026068866252899,
-0.008237474597990513,
0.04685759171843529,
-0.08516786992549896,
-0.059570085257291794,
-0.00319667742587626,
0.04088464751839638,
0.05083402246236801,
0.06535296887159348,
0.008304108865559101,
0.041982196271419525,
-0.09433333575725555,
0.05441991239786148,
0.00989446323364973,
0.09066399931907654,
-0.03276972845196724,
-0.10842546075582504,
0.038716379553079605,
0.047109413892030716,
0.0300612710416317,
0.037497151643037796,
-0.04076581075787544,
-0.018761400133371353,
0.013806678354740143,
0.04707540571689606,
0.05648869276046753,
0.03344541788101196,
-0.02509627677500248,
0.022769177332520485,
0.022830845788121223,
0.035494234412908554,
-0.062272533774375916,
0.14604754745960236,
0.024910954758524895,
-0.011437281966209412,
0.08910273015499115,
-0.07477820664644241,
-0.034630920737981796,
0.04715174436569214,
0.04167456924915314,
-0.021755870431661606,
-0.010073903016746044,
-0.09685143083333969,
0.03514860197901726,
0.06557715684175491,
-0.036910753697156906,
-0.012146273627877235,
0.0017037182115018368,
0.0010129761649295688,
-0.019272368401288986,
0.0035805348306894302,
0.015436813235282898,
-0.051723413169384,
-0.04483548551797867,
-0.12520794570446014,
0.04920801892876625,
-0.048127129673957825,
-0.022491713985800743,
0.0371599905192852,
0.016588769853115082,
0.06253542006015778,
0.03666432574391365,
-0.016939986497163773,
0.0043058875016868114,
0.09803327918052673,
-0.03275291621685028,
0.07681325823068619,
-0.0004497792397160083,
-0.0657070204615593,
0.025403859093785286,
0.004859234672039747,
-0.017239173874258995,
-0.03389699384570122,
0.08744931221008301,
-0.06500345468521118,
0.049131911247968674,
0.04079029709100723,
-0.011081055738031864,
-0.01075525302439928,
0.06873603910207748,
0.0345197357237339,
0.12080089002847672,
0.04116564244031906,
-0.03853170573711395,
-0.1056135818362236,
0.010007830336689949,
0.05097382888197899,
-0.02118537202477455,
0.05508125573396683,
0.005316246300935745,
0.022475184872746468,
-0.06798823177814484,
0.036802954971790314,
0.019948311150074005,
-0.04469749331474304,
0.07598892599344254,
-0.02178567834198475,
-0.024212338030338287,
-0.033775534480810165,
-0.005512450356036425,
0.015662996098399162,
-0.012548004277050495,
-0.05992411449551582,
-0.007066032849252224,
-0.0808933824300766,
0.03429804742336273,
0.013300925493240356,
-0.025926366448402405,
-0.0439213402569294,
1.793317467554181e-33,
-0.012743507511913776,
0.045120615512132645,
-0.017119180411100388,
-0.01705622300505638,
0.04203834757208824,
0.07193822413682938,
0.01778549887239933,
-0.0006966744549572468,
-0.03485226631164551,
-0.008392971940338612,
-0.06197095289826393,
0.02735459990799427,
-0.04817863926291466,
0.04204829782247543,
0.05075589939951897,
-0.042822420597076416,
-0.04128860682249069,
0.044429097324609756,
-0.023471619933843613,
-0.10411681234836578,
0.026134660467505455,
-0.040773987770080566,
-0.023135116323828697,
0.012665258720517159,
0.020183978602290154,
-0.006782757584005594,
0.07269273698329926,
-0.04068857803940773,
0.04418560862541199,
0.032087020576000214,
-0.13400453329086304,
-0.0895758718252182,
0.06960538029670715,
-0.08504507690668106,
0.0607643648982048,
-0.02205485664308071,
-0.01603185012936592,
0.018279055133461952,
0.006765907630324364,
-0.06342510879039764,
0.036576852202415466,
-0.00017783201474230736,
0.03137291967868805,
-0.09567301720380783,
-0.15965071320533752,
0.034998778253793716,
0.02946332097053528,
0.003497667144984007,
-0.03297547623515129,
-0.0702822208404541,
0.04507854953408241,
-0.05386853590607643,
-0.07192561030387878,
-0.002451084554195404,
0.031143294647336006,
-0.06880603730678558,
0.06890253722667694,
-0.004917391110211611,
-0.005452509969472885,
0.08166757225990295,
0.004972629249095917,
0.07890559732913971,
-0.015596133656799793,
0.0485401414334774,
0.01993708498775959,
-0.0685516893863678,
0.021536031737923622,
0.012360350228846073,
-0.03386952355504036,
0.020170601084828377,
-0.09309602528810501,
0.051945142447948456,
0.0824202299118042,
-0.023635899648070335,
0.027209702879190445,
-0.04184012860059738,
0.09833180159330368,
-0.017138322815299034,
-0.007096034940332174,
-0.0010505890240892768,
0.013154176063835621,
0.030473757535219193,
0.006235180422663689,
-0.05494141951203346,
-0.03509266674518585,
-0.020056888461112976,
-0.05780234560370445,
-0.044146209955215454,
-0.10977884382009506,
-0.063702292740345,
-0.021512767300009727,
-0.029556754976511,
-0.02538853883743286,
0.029695969074964523,
-0.04722430557012558,
-1.876981808606019e-33,
-0.03779897466301918,
0.06973101943731308,
0.03794167563319206,
0.06115198880434036,
-0.05796944350004196,
-0.03676602989435196,
0.05328040570020676,
-0.04504120722413063,
0.08803215622901917,
-0.03884490951895714,
0.004496993497014046,
-0.01600296050310135,
0.02512364089488983,
0.03823120892047882,
0.017998581752181053,
-0.025005316361784935,
-0.0891367644071579,
0.04068974032998085,
-0.019007688388228416,
0.01813407614827156,
0.009630980901420116,
0.06642789393663406,
-0.10222791135311127,
0.003839780343696475,
0.0030401975382119417,
0.05521898716688156,
-0.05146274343132973,
0.11169090867042542,
0.0763867199420929,
0.047213878482580185,
-0.03380010277032852,
0.0029257091227918863,
0.010071882046759129,
-0.046720799058675766,
-0.03293929994106293,
0.04392652213573456,
-0.007107021752744913,
0.029160723090171814,
0.055172596126794815,
0.0529780238866806,
-0.009395691566169262,
-0.045028574764728546,
-0.07145461440086365,
-0.03582130745053291,
0.07065681368112564,
0.007757833693176508,
-0.06243712827563286,
0.020430903881788254,
-0.00374883902259171,
-0.03596647456288338,
0.012473327107727528,
0.0670258179306984,
-0.0022154543548822403,
0.1354188621044159,
-0.039395056664943695,
0.001696065068244934,
-0.14516036212444305,
0.06777175515890121,
0.08522428572177887,
-0.031644269824028015,
-0.0182499997317791,
-0.016524557024240494,
0.09568378329277039,
-0.07274454087018967,
-0.08220009505748749,
-0.02179734595119953,
0.0009169051772914827,
-0.0012875585816800594,
-0.028406662866473198,
0.045704469084739685,
-0.028288746252655983,
-0.01876027137041092,
0.03468576446175575,
0.03222151845693588,
-0.023602129891514778,
-0.11841809004545212,
0.02844686061143875,
-0.05458230525255203,
0.002163772704079747,
0.03904324397444725,
-0.004405821207910776,
0.08049111813306808,
0.02739890292286873,
-0.010576915927231312,
0.029004070907831192,
0.04151540994644165,
0.0089833689853549,
0.024527598172426224,
0.024928459897637367,
0.02309405617415905,
0.000702574965544045,
0.062271226197481155,
0.07628130167722702,
0.04166289418935776,
-0.07772467285394669,
-4.720898161281184e-8,
-0.07301509380340576,
-0.00008580272697145119,
0.0007762148161418736,
-0.04941152408719063,
0.060562215745449066,
-0.0375775471329689,
0.03813149034976959,
0.019772924482822418,
0.05192882940173149,
-0.11743421107530594,
0.07916386425495148,
-0.06954947859048843,
-0.094879150390625,
-0.0843125507235527,
0.006624078843742609,
0.10986468940973282,
-0.03306851536035538,
-0.05365291237831116,
0.03649510443210602,
0.0026544041465967894,
-0.0852164700627327,
0.07503556460142136,
-0.004752913024276495,
-0.026697000488638878,
0.011690712533891201,
-0.055462002754211426,
-0.03286000341176987,
0.09843126684427261,
0.09383223205804825,
0.05627014860510826,
0.017341863363981247,
0.06770286709070206,
0.07508484274148941,
0.03845479339361191,
-0.005916053894907236,
0.07520630955696106,
-0.017054395750164986,
0.025167517364025116,
-0.03332100808620453,
-0.03336828574538231,
-0.0018171747215092182,
0.015174605883657932,
-0.0452490970492363,
0.02980201132595539,
0.016441481187939644,
-0.016063936054706573,
0.019362671300768852,
0.07575376331806183,
-0.007760846521705389,
-0.024032676592469215,
0.11152859777212143,
0.03518269583582878,
-0.07550687342882156,
0.04213233292102814,
-0.008587364107370377,
0.029727384448051453,
-0.04263049364089966,
-0.06607511639595032,
0.019134502857923508,
0.01564989611506462,
-0.021624810993671417,
-0.07315024733543396,
-0.04578128457069397,
0.05013071000576019
] |
LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs | https://openreview.net/forum?id=kQ5s9Yh0WI | [
"Yushi Bai",
"Jiajie Zhang",
"Xin Lv",
"Linzhi Zheng",
"Siqi Zhu",
"Lei Hou",
"Yuxiao Dong",
"Jie Tang",
"Juanzi Li"
] | Poster | Current long context large language models (LLMs) can process inputs up to 100,000 tokens, yet struggle to generate outputs exceeding even a modest length of 2,000 words. Through controlled experiments, we find that the model's effective generation length is inherently bounded by the sample it has seen during supervised fine-tuning (SFT). In other words, their output limitation is due to the scarcity of long-output examples in existing SFT datasets. To address this, we introduce AgentWrite, an agent-based pipeline that decomposes ultra-long generation tasks into subtasks, enabling off-the-shelf LLMs to generate coherent outputs exceeding 20,000 words. Leveraging AgentWrite, we construct LongWriter-6k, a dataset containing 6,000 SFT data with output lengths ranging from 2k to 32k words. By incorporating this dataset into model training, we successfully scale the output length of existing models to over 10,000 words while maintaining output quality. We also develop LongBench-Write, a comprehensive benchmark for evaluating ultra-long generation capabilities. Our 9B parameter model, further improved through DPO, achieves state-of-the-art performance on this benchmark, surpassing even much larger proprietary models. In general, our work demonstrates that existing long context LLM already possesses the potential for a larger output window--all you need is data with extended output during model alignment to unlock this capability. | long context, large language model, long-form generation | null | 1,838 | 2408.07055 | [
-0.02081509679555893,
-0.07228126376867294,
-0.0024173043202608824,
0.06595742702484131,
-0.00787072628736496,
-0.009070467203855515,
-0.10274983197450638,
-0.02765808254480362,
-0.020739907398819923,
-0.02943100407719612,
-0.03069678507745266,
-0.05759464576840401,
0.04203379154205322,
-0.057999350130558014,
-0.0064183552749454975,
0.03466539829969406,
0.06772411614656448,
0.03514382988214493,
-0.1280251443386078,
-0.04812643304467201,
0.08915334194898605,
0.0541088804602623,
0.029050473123788834,
-0.02770659513771534,
0.02373310551047325,
-0.016791798174381256,
-0.032418105751276016,
-0.08688706159591675,
0.033405132591724396,
-0.033436741679906845,
0.05316421762108803,
0.06964128464460373,
-0.006228994112461805,
0.03501091152429581,
0.0424748994410038,
0.06313701719045639,
-0.05955898016691208,
-0.022397741675376892,
0.03815164417028427,
-0.07604309916496277,
0.027371058240532875,
0.016841407865285873,
-0.005652106367051601,
0.05215839296579361,
0.021453948691487312,
-0.024781053885817528,
-0.016184287145733833,
-0.06059334799647331,
-0.013649718835949898,
0.052833136171102524,
-0.1033264696598053,
-0.022612761706113815,
-0.012303384952247143,
0.05677070841193199,
-0.05973608046770096,
-0.029746294021606445,
0.0029235321562737226,
0.055440712720155716,
-0.05040302500128746,
-0.04041264206171036,
-0.06819711625576019,
-0.06865062564611435,
-0.04211974889039993,
-0.0653761699795723,
-0.04557359591126442,
-0.03583161160349846,
-0.01752471923828125,
0.037848200649023056,
0.0024261828511953354,
0.02181243523955345,
-0.07801450788974762,
0.054687563329935074,
-0.08662155270576477,
0.04619750380516052,
-0.010173263028264046,
0.016378598287701607,
0.03397858142852783,
-0.018464650958776474,
0.0670793429017067,
-0.07529666274785995,
0.005033036693930626,
-0.030639225617051125,
0.06985969841480255,
0.0029479768127202988,
0.0648999810218811,
0.026436464861035347,
0.06795946508646011,
0.07234977185726166,
-0.0372038371860981,
0.035730041563510895,
-0.04936664551496506,
-0.08838853985071182,
0.1041686162352562,
-0.022023744881153107,
-0.04746536538004875,
0.0385611318051815,
-0.0031795212998986244,
-0.05209853872656822,
-0.031000182032585144,
0.07146114110946655,
0.005929823499172926,
0.08688751608133316,
-0.011212003417313099,
-0.004139183554798365,
-0.02920832298696041,
-0.034497056156396866,
0.04853549227118492,
0.0899956151843071,
-0.020634736865758896,
-0.11756069213151932,
0.07072604447603226,
0.053893595933914185,
0.0182727612555027,
-0.01109394058585167,
0.11493431031703949,
-0.02474234439432621,
-0.043469708412885666,
-0.029189689084887505,
-0.01518483180552721,
0.09689436107873917,
-0.018394848331809044,
0.021195029839873314,
-0.05692901462316513,
0.037362899631261826,
-0.029799049720168114,
0.010544355027377605,
0.027002248913049698,
2.0428446125200206e-33,
0.05386580526828766,
0.0792168453335762,
-0.0030838409438729286,
0.011166554875671864,
0.08377933502197266,
0.06488070636987686,
0.0433058999478817,
0.08729929476976395,
-0.0665651485323906,
0.012352196499705315,
0.004834479186683893,
-0.0318160206079483,
-0.03921185061335564,
0.059419404715299606,
0.08349509537220001,
-0.07463068515062332,
-0.0273041483014822,
0.08175322413444519,
-0.04240432754158974,
0.02204993925988674,
0.0477033406496048,
-0.0064870924688875675,
0.023492559790611267,
-0.049628254026174545,
0.026381706818938255,
-0.028999658301472664,
0.01707386039197445,
-0.04748550429940224,
-0.07409729808568954,
0.0036006676964461803,
-0.122029609978199,
-0.00794681254774332,
0.04931730777025223,
0.06687790155410767,
-0.0555705763399601,
-0.08104806393384933,
-0.02428428642451763,
-0.06417937576770782,
0.02018122933804989,
0.0005639942828565836,
-0.07151596248149872,
0.04711959511041641,
0.050721269100904465,
-0.026873890310525894,
-0.06854848563671112,
-0.04272566735744476,
-0.019133582711219788,
0.027073098346590996,
-0.002933469135314226,
-0.025982001796364784,
0.08833848685026169,
0.05319610983133316,
0.009013454429805279,
0.02391829527914524,
-0.004287184681743383,
-0.015077836811542511,
0.08992558717727661,
0.0033221361227333546,
0.0319109745323658,
0.08948107808828354,
0.0006161933997645974,
0.05162152647972107,
-0.03479742258787155,
0.1489635705947876,
0.026973940432071686,
0.015951350331306458,
0.043296538293361664,
0.11029175668954849,
0.035273849964141846,
-0.06425981223583221,
0.00296383211389184,
-0.026686860248446465,
0.02859090454876423,
-0.037065960466861725,
0.00950118899345398,
-0.042317669838666916,
0.05717569217085838,
-0.10218514502048492,
-0.02788732387125492,
0.011135516688227654,
0.041381534188985825,
0.038911089301109314,
0.007825088687241077,
-0.06780888140201569,
0.04974862560629845,
-0.08440237492322922,
-0.0784280076622963,
-0.10654131323099136,
0.014076445251703262,
-0.01493033953011036,
-0.09219149500131607,
-0.031254176050424576,
-0.017781255766749382,
-0.06699439883232117,
-0.04855218157172203,
-7.8351896371388e-34,
0.0224094707518816,
-0.047309260815382004,
-0.10164270550012589,
0.09796411544084549,
-0.03316723182797432,
-0.06827349960803986,
-0.0312761589884758,
0.0690399706363678,
-0.05304384231567383,
-0.03985151648521423,
-0.03980335220694542,
-0.05084524676203728,
0.038730818778276443,
-0.047888848930597305,
0.012784380465745926,
-0.09627673029899597,
0.030139077454805374,
-0.0027772688772529364,
0.06471004337072372,
-0.014587420038878918,
0.06248346343636513,
0.04911769554018974,
-0.009828047826886177,
0.04918227717280388,
0.027325259521603584,
0.008021976798772812,
-0.05953838676214218,
0.07803059369325638,
0.011273016221821308,
0.01806679740548134,
-0.006596083752810955,
0.010696041397750378,
0.019853295758366585,
-0.05272334814071655,
-0.05896426737308502,
0.030219651758670807,
0.04328303411602974,
0.05781614035367966,
0.01027096901088953,
0.09445449709892273,
0.0760449543595314,
-0.05580677464604378,
-0.022686010226607323,
0.03374579921364784,
-0.016028612852096558,
-0.011088765226304531,
-0.13871489465236664,
-0.0025683301500976086,
0.05839464068412781,
0.035103749483823776,
-0.009422976523637772,
0.04094648361206055,
-0.033189576119184494,
0.015202406793832779,
-0.058904264122247696,
-0.12715303897857666,
-0.05655824765563011,
-0.0732736587524414,
-0.034581419080495834,
-0.05672997608780861,
-0.06972955912351608,
0.0027042939327657223,
0.06543077528476715,
-0.04594675451517105,
0.07817108929157257,
-0.04468042775988579,
-0.011456727050244808,
0.013226335868239403,
0.05678671598434448,
-0.002927083056420088,
0.02162907086312771,
-0.056176651269197464,
-0.058989252895116806,
0.09261716157197952,
-0.04728524014353752,
0.02765505015850067,
-0.023820113390684128,
-0.08249226957559586,
0.0644148588180542,
0.0073959706351161,
-0.015203475952148438,
0.10079819709062576,
0.009136892855167389,
0.039139676839113235,
0.05254479497671127,
0.020345792174339294,
0.018952175974845886,
0.10006894916296005,
0.03713244944810867,
0.07219503819942474,
-0.007232083007693291,
0.0618559829890728,
0.03441273421049118,
-0.0018335465574637055,
-0.05361299589276314,
-3.6511185186327566e-8,
-0.003151039592921734,
0.033462394028902054,
-0.005620298907160759,
0.06735190004110336,
0.019780559465289116,
-0.0345030352473259,
0.024659082293510437,
0.0742528960108757,
0.057144004851579666,
-0.025408506393432617,
0.09966476261615753,
-0.058564070612192154,
-0.07891709357500076,
-0.011080124415457249,
0.04323209077119827,
0.010832002386450768,
0.025120608508586884,
-0.023150643333792686,
-0.03875920549035072,
-0.018000341951847076,
0.039386581629514694,
0.07050356268882751,
-0.017814017832279205,
0.017215466126799583,
0.03328033164143562,
-0.08544306457042694,
-0.0606025867164135,
0.09279630333185196,
0.026934292167425156,
-0.04015417397022247,
0.005713591817766428,
-0.03251812234520912,
-0.06005529314279556,
-0.03630353510379791,
0.018764186650514603,
0.036639925092458725,
-0.00894880574196577,
-0.007753332611173391,
0.029351692646741867,
-0.0008861047681421041,
0.023478977382183075,
0.0840405523777008,
-0.03358966484665871,
0.029163073748350143,
0.024102898314595222,
-0.06465024501085281,
-0.039708852767944336,
-0.09000983089208603,
0.05790559574961662,
-0.03506031259894371,
0.007806821260601282,
0.0071476236917078495,
0.02071051113307476,
0.006891855504363775,
0.04393630847334862,
0.065920390188694,
-0.02835121564567089,
0.022943295538425446,
0.05524206906557083,
0.025706520304083824,
0.008688647300004959,
0.00958728976547718,
-0.011407919228076935,
-0.003528430126607418
] |
GROOT-2: Weakly Supervised Multimodal Instruction Following Agents | https://openreview.net/forum?id=S9GyQUXzee | [
"Shaofei Cai",
"Bowei Zhang",
"Zihao Wang",
"Haowei Lin",
"Xiaojian Ma",
"Anji Liu",
"Yitao Liang"
] | Poster | Developing agents that can follow multimodal instructions remains a fundamental challenge in robotics and AI. Although large-scale pre-training on unlabeled datasets has enabled agents to learn diverse behaviors, these agents often struggle with following instructions. While augmenting the dataset with instruction labels can mitigate this issue, acquiring such high-quality annotations at scale is impractical.
To address this issue, we frame the problem as a semi-supervised learning task and introduce \agent, a multimodal instructable agent trained using a novel approach that combines weak supervision with latent variable models. Our method consists of two key components: constrained self-imitating, which utilizes large amounts of unlabeled demonstrations to enable the policy to learn diverse behaviors, and human intention alignment, which uses a smaller set of labeled demonstrations to ensure the latent space reflects human intentions. \agent’s effectiveness is validated across four diverse environments, ranging from video games to robotic manipulation, demonstrating its robust multimodal instruction-following capabilities. | Reinforcement Learning, Open-world Agent, Weakly Supervised Learning, Goal-Conditioned Policy | null | 1,830 | null | [
0.03230023756623268,
-0.10980362445116043,
0.035307563841342926,
-0.03565335273742676,
0.08327996730804443,
0.019451146945357323,
0.007894693873822689,
-0.051596254110336304,
-0.009736898355185986,
-0.030507616698741913,
0.019850844517350197,
-0.07175535708665848,
-0.0041812146082520485,
0.0802178829908371,
0.05290719494223595,
-0.015834860503673553,
0.11215938627719879,
0.04659762233495712,
-0.09383634477853775,
-0.033002402633428574,
0.07359232008457184,
0.06016312912106514,
0.07227295637130737,
0.01895175129175186,
-0.0019255065126344562,
0.07120546698570251,
0.0530833825469017,
-0.05296268314123154,
0.05448418855667114,
-0.04627055302262306,
0.04148136079311371,
-0.006344416178762913,
-0.043879102915525436,
0.027207504957914352,
-0.02571096457540989,
0.11986421048641205,
-0.09982180595397949,
0.051848024129867554,
-0.019357096403837204,
-0.04339545965194702,
-0.03380803018808365,
0.07309867441654205,
0.00968386884778738,
0.038413748145103455,
0.04238197207450867,
0.025093063712120056,
-0.03468623757362366,
-0.08072394877672195,
-0.009248290210962296,
0.03561302274465561,
-0.07057218998670578,
-0.05448676273226738,
0.09855250269174576,
-0.006025160662829876,
-0.014558840543031693,
-0.02163144014775753,
0.04243830591440201,
-0.007606555707752705,
0.09468737244606018,
0.01434010174125433,
0.04155882075428963,
-0.08787321299314499,
0.00429704412817955,
-0.06595611572265625,
-0.02061072178184986,
-0.06050936505198479,
0.0030228132382035255,
-0.03799668326973915,
-0.005529771558940411,
0.016528049483895302,
-0.04513387382030487,
-0.01742757484316826,
0.007239648140966892,
0.017767909914255142,
-0.017250362783670425,
0.0016314404783770442,
-0.02262013405561447,
-0.02325228974223137,
0.03600957244634628,
-0.06763013452291489,
-0.029988322407007217,
-0.054390959441661835,
0.03237144649028778,
0.06759315729141235,
0.05785580351948738,
-0.009452450089156628,
0.04588546231389046,
0.06792231649160385,
0.024720044806599617,
0.033239640295505524,
-0.037660904228687286,
-0.007393712643533945,
0.04665622487664223,
-0.014578278176486492,
0.005339674651622772,
0.023423904553055763,
-0.007146662566810846,
-0.16058778762817383,
-0.025889843702316284,
0.07834604382514954,
0.0034923250786960125,
0.08628502488136292,
-0.005945767275989056,
0.0544680655002594,
-0.012262294068932533,
-0.06144597753882408,
0.06796372681856155,
-0.002311732154339552,
0.01597015932202339,
-0.052204154431819916,
-0.0056236376985907555,
0.0028300879057496786,
0.02775077149271965,
0.0492677167057991,
0.10376248508691788,
-0.02598763071000576,
0.040662843734025955,
0.04357533156871796,
-0.007732035126537085,
0.03664516657590866,
0.011449780315160751,
-0.05240670219063759,
0.0072635384276509285,
-0.004286288749426603,
0.09360692650079727,
-0.019112413749098778,
-0.061793968081474304,
1.9914220611303227e-33,
0.1306658238172531,
0.05789098143577576,
0.06485777348279953,
0.055580344051122665,
0.02723107486963272,
-0.004480080679059029,
-0.0032656071707606316,
0.03515966236591339,
-0.0447552315890789,
-0.006550934631377459,
0.08399835228919983,
-0.03651915863156319,
-0.08859439194202423,
0.14083944261074066,
-0.011088203638792038,
0.0036801514215767384,
-0.03497105464339256,
0.03958655893802643,
-0.01048345398157835,
0.022095205262303352,
0.04023659974336624,
0.03659609332680702,
-0.02577057294547558,
-0.018395349383354187,
0.03118303045630455,
-0.04101811721920967,
0.022270459681749344,
-0.048105426132678986,
0.05866948515176773,
-0.014833960682153702,
-0.10056708008050919,
-0.024067038670182228,
0.023088732734322548,
0.07146135717630386,
0.006218243855983019,
-0.07467544823884964,
-0.008149161003530025,
-0.05297365412116051,
0.016680965200066566,
-0.0114077003672719,
0.02582922950387001,
-0.022312650457024574,
0.07097160071134567,
-0.09649742394685745,
-0.09794202446937561,
-0.07136368006467819,
0.037874821573495865,
0.03173667937517166,
-0.017410315573215485,
-0.022775128483772278,
0.020657094195485115,
0.07783272117376328,
-0.003882406046614051,
-0.13582098484039307,
-0.040244415402412415,
-0.04580998420715332,
0.00529986247420311,
0.020693711936473846,
-0.009100984781980515,
-0.019340550526976585,
0.004634269513189793,
0.06070084497332573,
0.006478298455476761,
0.09800295531749725,
0.03621711954474449,
-0.019172878935933113,
0.013297731988132,
0.10630474984645844,
0.1365739107131958,
0.014067232608795166,
-0.07879724353551865,
0.028856778517365456,
0.01016270276159048,
-0.10446774959564209,
0.0038445780519396067,
-0.008427745662629604,
0.010081973858177662,
-0.13416129350662231,
0.07353347539901733,
0.019029363989830017,
-0.024020833894610405,
-0.013161326758563519,
-0.0311676487326622,
0.01034434512257576,
-0.005738153122365475,
-0.056918803602457047,
-0.011659267358481884,
-0.07825802266597748,
0.0014484263956546783,
0.014858962036669254,
-0.05770769342780113,
-0.05174829810857773,
-0.09591442346572876,
0.03581010550260544,
-0.028240758925676346,
-1.9156283894412093e-33,
0.05503508821129799,
0.021302074193954468,
-0.01882372982800007,
0.032767318189144135,
0.019034579396247864,
-0.029452219605445862,
0.011023675091564655,
0.022865567356348038,
0.0004598943633027375,
-0.062462639063596725,
-0.04044293239712715,
-0.007706474978476763,
-0.045143917202949524,
-0.009574957191944122,
-0.003602410666644573,
-0.01163025014102459,
0.03415146470069885,
0.053236253559589386,
0.04492727294564247,
-0.06336645781993866,
-0.011286054737865925,
0.08906413614749908,
-0.010308947414159775,
0.008603660389780998,
-0.03318898752331734,
0.03326684981584549,
0.03823845088481903,
0.054945118725299835,
0.01867610216140747,
0.07035539299249649,
-0.006189539562910795,
-0.04482137784361839,
0.04977782443165779,
0.021903136745095253,
-0.06235064938664436,
0.03515297919511795,
-0.006176487542688847,
0.01414850540459156,
-0.06494322419166565,
0.013937795534729958,
0.06867486983537674,
-0.025915931910276413,
-0.06899725645780563,
0.007481497246772051,
-0.010912800207734108,
-0.043588556349277496,
-0.052873775362968445,
0.04143594950437546,
-0.00037137954495847225,
0.07168611139059067,
-0.07205256819725037,
0.006137475837022066,
-0.07861775904893875,
-0.06896239519119263,
-0.08541569113731384,
-0.10771409422159195,
0.030820520594716072,
-0.04478730261325836,
0.055234260857105255,
-0.01942312903702259,
-0.08810870349407196,
0.025625601410865784,
0.017042508348822594,
0.0033198886085301638,
0.018780168145895004,
-0.06773179769515991,
-0.05166760832071304,
0.02686566486954689,
0.03995579481124878,
-0.05942992493510246,
0.04931927099823952,
0.005316607654094696,
0.03214059770107269,
0.0082108024507761,
0.06656241416931152,
-0.014709165319800377,
0.030558889731764793,
-0.02802255190908909,
0.024280088022351265,
-0.07166420668363571,
0.07349511235952377,
-0.008983202278614044,
0.003977728076279163,
0.06088827922940254,
0.021189499646425247,
0.11507195234298706,
-0.06054530665278435,
0.08554071187973022,
0.028173713013529778,
0.05562002211809158,
0.008485944010317326,
0.0732296034693718,
0.03983069583773613,
0.05442328006029129,
-0.06987031549215317,
-3.8753654507672763e-8,
-0.11492444574832916,
0.03495802730321884,
0.08453231304883957,
-0.023743391036987305,
0.005345803685486317,
0.0032915030606091022,
0.013699444942176342,
0.01756569929420948,
0.011828280985355377,
-0.00045166819472797215,
0.06144862249493599,
-0.05192629620432854,
0.008794350549578667,
-0.03763597458600998,
-0.01245205756276846,
0.04770957678556442,
0.012226059101521969,
0.02610607258975506,
-0.0007864717626944184,
-0.009693346917629242,
-0.007140812464058399,
-0.04452710971236229,
0.00024646383826620877,
-0.038973379880189896,
0.00569043168798089,
-0.04149164631962776,
-0.0511733815073967,
0.0574844591319561,
-0.0030192530248314142,
0.0500619076192379,
0.015394669957458973,
-0.031002135947346687,
0.05128307268023491,
-0.03017730452120304,
-0.026073511689901352,
0.05584834888577461,
0.01591014303267002,
-0.07932765036821365,
0.001637589419260621,
-0.10804793238639832,
-0.035401783883571625,
0.06407288461923599,
-0.05948986858129501,
-0.0031313076615333557,
0.030789632350206375,
0.010503903962671757,
-0.06750655174255371,
-0.1718902438879013,
0.028363946825265884,
-0.05469118431210518,
-0.0294569693505764,
-0.0710611566901207,
0.039316706359386444,
0.10985192656517029,
0.10942766070365906,
0.060144394636154175,
-0.00046721435501240194,
-0.042583197355270386,
0.03567592799663544,
0.08886964619159698,
-0.009075654670596123,
0.03549620509147644,
-0.040341924875974655,
-0.016711287200450897
] |
Medium-Difficulty Samples Constitute Smoothed Decision Boundary for Knowledge Distillation on Pruned Datasets | https://openreview.net/forum?id=Rz4UkJziFe | [
"Yudong Chen",
"Xuwei Xu",
"Frank de Hoog",
"Jiajun Liu",
"Sen Wang"
] | Poster | This paper tackles a new problem of dataset pruning for Knowledge Distillation (KD), from a fresh perspective of Decision Boundary (DB) preservation and drifts. Existing dataset pruning methods generally assume that the post-pruning DB formed by the selected samples can be well-captured by future networks that use those samples for training. Therefore, they tend to preserve hard samples since hard samples are closer to the DB and better characterize the nuances in the distribution of the entire dataset. However, in KD, the limited learning capacity from the student network leads to imperfect preservation of the teacher's feature distribution, resulting in the drift of DB in the student space. Specifically, hard samples worsen such drifts as they are difficult for the student to learn, creating a situation where the student's DB can drift deeper into other classes and make incorrect classifications. Motivated by these findings, our method selects medium-difficulty samples for KD-based dataset pruning. We show that these samples constitute a smoothed version of the teacher's DB and are easier for the student to learn, obtaining a general feature distribution preservation for a class of samples and reasonable DB between different classes for the student. In addition, to reduce the distributional shift due to dataset pruning, we leverage the class-wise distributional information of the teacher's outputs to reshape the logits of the preserved samples. Experiments show that the proposed static pruning method can even perform better than the state-of-the-art dynamic pruning method which needs access to the entire dataset. In addition, our method halves the training times of KD and improves the student's accuracy by 0.4% on ImageNet with a 50% keep ratio. When the ratio further increases to 70%, our method achieves higher accuracy over the vanilla KD while reducing the training times by 30%. Code is available at https://github.com/chenyd7/MDSLR. | Knowledge distillation, dataset pruning, image recognition | This paper proposes a static dataset pruning method for efficient knowledge distillation. | 1,822 | null | [
-0.05693740397691727,
-0.03753848373889923,
0.06713435053825378,
-0.003085042117163539,
0.035359639674425125,
-0.03437710180878639,
0.033434946089982986,
-0.036885641515254974,
-0.06246327608823776,
0.031368058174848557,
0.01743641309440136,
0.022536270320415497,
0.05651883780956268,
0.0016284503508359194,
-0.009735074825584888,
-0.000488672056235373,
0.01673893630504608,
0.06017536669969559,
-0.08741399645805359,
0.018414458259940147,
-0.06203651800751686,
-0.002204792806878686,
-0.06792645901441574,
0.03174404054880142,
0.07458149641752243,
0.012203733436763287,
0.06722266227006912,
-0.020418481901288033,
0.031163400039076805,
-0.08768114447593689,
-0.045877426862716675,
0.020976487547159195,
-0.05100894719362259,
0.0042361910454928875,
-0.04094771295785904,
-0.025418369099497795,
-0.05297090485692024,
0.1215146854519844,
0.0007608539890497923,
0.02912563644349575,
0.005071063060313463,
0.01752272993326187,
-0.04279039055109024,
0.059635382145643234,
0.03294125944375992,
0.07354147732257843,
-0.05280506610870361,
-0.06483010202646255,
0.05018187686800957,
0.041595038026571274,
-0.00001905397220980376,
0.009410697035491467,
-0.035515718162059784,
0.04340752586722374,
0.0010334551334381104,
-0.0033072049263864756,
0.016665108501911163,
0.04412399232387543,
-0.011785897426307201,
0.004161010961979628,
-0.009556030854582787,
-0.12016595900058746,
-0.01146563608199358,
-0.011060020886361599,
-0.024281324818730354,
-0.08459252119064331,
0.06382466852664948,
0.08324677497148514,
0.06197226420044899,
-0.004365496337413788,
0.007730224635452032,
0.09328338503837585,
-0.10711612552404404,
0.06651309877634048,
0.054758984595537186,
0.002389770234003663,
0.0637807622551918,
0.0050038802437484264,
-0.010908049531280994,
-0.024014383554458618,
-0.038409315049648285,
0.030748410150408745,
0.00392748462036252,
-0.018360834568738937,
0.07078693807125092,
-0.04338261857628822,
0.0330374538898468,
-0.010635393671691418,
0.013189360499382019,
-0.024762367829680443,
-0.04216424375772476,
0.104803167283535,
0.009225839748978615,
0.004061625339090824,
0.014247431419789791,
0.05029140040278435,
-0.008771700784564018,
0.03571946173906326,
-0.012010902166366577,
0.09607449918985367,
-0.06833618879318237,
0.08509748429059982,
-0.011124206706881523,
-0.08780467510223389,
0.012242216616868973,
-0.08556750416755676,
0.03790442645549774,
-0.023629317060112953,
0.049515895545482635,
-0.08411700278520584,
0.006732050329446793,
0.017868323251605034,
-0.07960309088230133,
-0.016648991033434868,
0.03804479166865349,
0.011932953260838985,
0.03545917570590973,
-0.01258887629956007,
-0.07067669928073883,
0.01160438172519207,
0.008935927413403988,
0.02870703861117363,
-0.017061524093151093,
0.0005862461403012276,
0.03945222496986389,
0.04222753271460533,
0.006975598633289337,
2.875473526159348e-33,
0.08865169435739517,
0.023545267060399055,
-0.03392105549573898,
-0.06214625760912895,
0.02296523191034794,
-0.03495815768837929,
0.0023570843040943146,
0.0911431834101677,
0.013090194202959538,
0.067236989736557,
-0.041272759437561035,
-0.003766884794458747,
-0.04165210202336311,
-0.049104366451501846,
0.09572137892246246,
-0.033961374312639236,
-0.05031859874725342,
0.06252168864011765,
-0.04099390283226967,
-0.016988037154078484,
0.06561239063739777,
-0.03975718468427658,
0.04678896814584732,
-0.09987235814332962,
-0.04575454443693161,
-0.018648497760295868,
-0.006872162688523531,
0.07119634747505188,
0.00940043292939663,
-0.005536156706511974,
-0.03823218122124672,
-0.02017783187329769,
-0.019365329295396805,
0.019145933911204338,
-0.003077821107581258,
-0.04307927191257477,
0.016654087230563164,
-0.012516668997704983,
0.006400790996849537,
-0.01382245309650898,
-0.03010401502251625,
0.013011644594371319,
0.025839615613222122,
0.033290695399045944,
0.0068692294880747795,
0.009887369349598885,
0.057779461145401,
-0.04003803804516792,
-0.01673084869980812,
-0.015362218022346497,
-0.051010169088840485,
-0.04901941493153572,
0.036207906901836395,
-0.05061916261911392,
-0.06215862184762955,
0.059250134974718094,
0.039970315992832184,
-0.01864449866116047,
0.02641163021326065,
0.05474337190389633,
-0.0379900299012661,
0.0555727444589138,
-0.008495228365063667,
0.11645367741584778,
0.020957857370376587,
0.02922004461288452,
-0.058731190860271454,
0.049932047724723816,
0.08936671912670135,
-0.11361590772867203,
-0.04588599503040314,
-0.021855900064110756,
-0.05657816678285599,
-0.144830584526062,
0.08144326508045197,
-0.03841615468263626,
0.05059750750660896,
-0.012923655100166798,
-0.009216789156198502,
-0.06512784957885742,
0.034382693469524384,
-0.0786772146821022,
-0.07769641280174255,
-0.049925416707992554,
-0.06856103986501694,
0.0511135570704937,
0.05813140422105789,
-0.054949451237916946,
0.054294660687446594,
-0.022050784900784492,
-0.10022609680891037,
0.05052362009882927,
-0.05630768463015556,
-0.031710997223854065,
0.10682526975870132,
-2.0553733619188707e-33,
-0.041486989706754684,
0.05118333175778389,
-0.02707175351679325,
0.12654978036880493,
0.0687929317355156,
0.0037844597827643156,
-0.012342952191829681,
0.049026623368263245,
-0.03948202356696129,
-0.08439279347658157,
-0.05157088488340378,
-0.07921501249074936,
-0.03928564488887787,
-0.04912687465548515,
-0.03869577869772911,
0.05433318391442299,
-0.0866367295384407,
-0.00881104450672865,
0.0038224051240831614,
0.04216828942298889,
-0.03548557311296463,
0.03765140101313591,
-0.14065182209014893,
0.032964061945676804,
0.0041595385409891605,
0.013249137438833714,
-0.031529057770967484,
0.038128629326820374,
0.012845383957028389,
-0.00923664215952158,
-0.025402257218956947,
-0.06705126911401749,
-0.10889542102813721,
-0.04410203918814659,
-0.04548950120806694,
-0.0399167537689209,
0.0039620292373001575,
-0.008943784050643444,
-0.05523493513464928,
0.18587400019168854,
-0.002930279355496168,
0.008095073513686657,
-0.05011585354804993,
0.0028156323824077845,
0.02868773601949215,
0.026507100090384483,
0.025519894436001778,
-0.05859191343188286,
0.007470245007425547,
0.011551170609891415,
0.06012076511979103,
0.04122905433177948,
-0.005374238360673189,
0.0829143300652504,
0.07651637494564056,
0.016178954392671585,
-0.001902954769320786,
0.04798956215381622,
0.007881788536906242,
0.0679028108716011,
-0.02049035020172596,
-0.05234917998313904,
-0.022830292582511902,
0.035913728177547455,
0.01052278745919466,
-0.0842260792851448,
0.03932444378733635,
0.12035897374153137,
-0.07572716474533081,
0.06264390051364899,
-0.0432378351688385,
0.06033985689282417,
0.008202810771763325,
-0.07095862925052643,
0.01885077729821205,
0.010443338192999363,
0.010669641196727753,
-0.04274376481771469,
0.0001944464456755668,
0.04496577009558678,
0.030146263539791107,
-0.06625107675790787,
-0.007162627298384905,
-0.014149582013487816,
0.06739408522844315,
0.062202420085668564,
0.07311386615037918,
-0.0470472052693367,
0.07792958617210388,
-0.012888850644230843,
-0.03463288024067879,
0.01612834446132183,
0.016422253102064133,
0.037670742720365524,
-0.02808815985918045,
-4.465699987576954e-8,
-0.05067073553800583,
-0.0174523014575243,
-0.028381409123539925,
0.034046273678541183,
0.14893610775470734,
0.005917369853705168,
-0.05808987468481064,
0.1395794004201889,
-0.042900655418634415,
0.0037882979959249496,
0.0755421593785286,
-0.011326989158987999,
-0.037446364760398865,
0.009449261240661144,
0.05459609255194664,
0.07871793955564499,
-0.010742947459220886,
0.03340265154838562,
-0.011076124384999275,
0.0493948757648468,
0.05352087691426277,
-0.020214127376675606,
-0.010702368803322315,
-0.028024744242429733,
-0.01778254844248295,
-0.08185592293739319,
0.04498043656349182,
0.11830873787403107,
0.036990806460380554,
0.029882656410336494,
-0.0416860394179821,
-0.021446511149406433,
0.049832385033369064,
-0.05506542697548866,
0.06879277527332306,
0.08293846994638443,
0.012683514505624771,
0.021134868264198303,
-0.07105281203985214,
0.016167011111974716,
-0.06637579202651978,
-0.02977815270423889,
0.0004159103555139154,
0.028430970385670662,
-0.016870539635419846,
-0.034239958971738815,
-0.01013997197151184,
-0.03419705852866173,
0.08804060518741608,
-0.018658602610230446,
0.04425155371427536,
-0.05467601493000984,
0.0026375462766736746,
-0.03704303875565529,
0.06842060387134552,
0.051840294152498245,
-0.05818851292133331,
-0.027555014938116074,
-0.06479962170124054,
-0.04165998101234436,
-0.004369707778096199,
0.01369163766503334,
-0.08072500675916672,
0.09240195155143738
] |
Mixture Compressor for Mixture-of-Experts LLMs Gains More | https://openreview.net/forum?id=hheFYjOsWO | [
"Wei Huang",
"Yue Liao",
"Jianhui Liu",
"Ruifei He",
"Haoru Tan",
"Shiming Zhang",
"Hongsheng Li",
"Si Liu",
"XIAOJUAN QI"
] | Poster | Mixture-of-Experts large language models (MoE-LLMs) marks a significant step forward of language models, however, they encounter two critical challenges in practice: 1) expert parameters lead to considerable memory consumption and loading latency; and 2) the current activated experts are redundant, as many tokens may only require a single expert. Motivated by these issues, we investigate the MoE-LLMs and make two key observations: a) different experts exhibit varying behaviors on activation reconstruction error, routing scores, and activated frequencies, highlighting their differing importance, and b) not all tokens are equally important-- only a small subset is critical. Building on these insights, we propose MC, a training-free Mixture-Compressor for MoE-LLMs, which leverages the significance of both experts and tokens to achieve an extreme compression. First, to mitigate storage and loading overheads, we introduce Pre-Loading Mixed-Precision Quantization (PMQ), which formulates the adaptive bit-width allocation as a Linear Programming (LP) problem, where the objective function balances multi-factors reflecting the importance of each expert. Additionally, we develop Online Dynamic Pruning (ODP), which identifies important tokens to retain and dynamically select activated experts for other tokens during inference to optimize efficiency while maintaining performance. Our MC integrates static quantization and dynamic pruning to collaboratively achieve extreme compression for MoE-LLMs with less accuracy loss, ensuring an optimal trade-off between performance and efficiency Extensive experiments confirm the effectiveness of our approach. For instance, at 2.54 bits, MC compresses 76.6% of the model, with only a 3.8% average accuracy loss. During dynamic inference, we further reduce activated parameters by 15%, with a performance drop of less than 0.6%. Remarkably, MC even surpasses floating-point 13b dense LLMs with significantly smaller parameter sizes, suggesting that mixture compression in MoE-LLMs has the potential to outperform both comparable and larger dense LLMs. Our code is
available at https://github.com/Aaronhuang-778/MC-MoE | Mixture-of-Expert, LLM, Quantization, Pruning | A Mixture compression method, which achieves extreme compression for MoE-LLMs with less accuracy loss, ensuring a better trade-off between performance and efficiency. | 1,819 | 2410.06270 | [
-0.00655789952725172,
-0.05348155274987221,
-0.00033812978654168546,
0.07845385372638702,
0.0019162368262186646,
0.003840495366603136,
-0.007299881428480148,
-0.012660075910389423,
0.0044654356315732,
-0.06501410901546478,
-0.01603582501411438,
-0.010738352313637733,
0.012636415660381317,
-0.01143599022179842,
0.03260692209005356,
0.04517586901783943,
0.04564055800437927,
0.14824388921260834,
-0.12374793738126755,
-0.09882144629955292,
0.013074892573058605,
-0.01713494025170803,
-0.0042343842796981335,
-0.016505936160683632,
0.05491338670253754,
0.03769656643271446,
-0.024715092033147812,
-0.018008911982178688,
0.11142686009407043,
-0.02057359181344509,
0.03950078412890434,
0.01875203102827072,
0.09410041570663452,
0.006512683350592852,
-0.030205801129341125,
0.0704612210392952,
-0.022631831467151642,
0.009040185250341892,
0.008016448467969894,
-0.034717679023742676,
-0.04126409813761711,
0.005201296880841255,
0.004524923861026764,
0.052827708423137665,
-0.0023783256765455008,
0.045008327811956406,
0.028300108388066292,
0.05137254297733307,
0.0010553586762398481,
0.014229673892259598,
0.010648303665220737,
0.016252586618065834,
-0.059016112238168716,
0.06635026633739471,
-0.06556981801986694,
0.033871933817863464,
0.03212584927678108,
0.07785247266292572,
-0.032740768045186996,
0.03600055351853371,
-0.12848104536533356,
-0.08882369846105576,
0.011162333190441132,
0.02409988082945347,
-0.00704014440998435,
0.00029955466743558645,
0.06102460250258446,
0.03809770196676254,
0.016277378425002098,
0.06194788217544556,
-0.0826566144824028,
-0.016184424981474876,
-0.026374949142336845,
0.08728961646556854,
-0.0416305810213089,
0.07223793119192123,
0.06553041189908981,
-0.003667825600132346,
0.0199496541172266,
0.019888008013367653,
0.0171259306371212,
-0.022367380559444427,
0.03135424852371216,
-0.05919250473380089,
0.15037640929222107,
-0.04995843023061752,
-0.02411535568535328,
0.012280609458684921,
0.03243281692266464,
-0.015831205993890762,
-0.1086365357041359,
0.05796105042099953,
0.06629084795713425,
-0.019560791552066803,
0.06686250865459442,
0.046065445989370346,
0.0010596992215141654,
-0.017283249646425247,
-0.00379434903152287,
0.11267749965190887,
-0.02264597825706005,
0.04887723922729492,
0.026285432279109955,
-0.030637631192803383,
-0.06375449150800705,
-0.058575794100761414,
-0.006815624888986349,
0.07945124059915543,
0.06739023327827454,
0.011778786778450012,
0.041806917637586594,
0.05452057719230652,
0.03805201128125191,
-0.0725976750254631,
0.040477968752384186,
0.006944174412637949,
-0.033016953617334366,
-0.06260428577661514,
0.021624702960252762,
0.023692255839705467,
-0.10161583870649338,
-0.05265943333506584,
-0.030098991468548775,
0.020106149837374687,
0.033362336456775665,
-0.02094287797808647,
-0.0587192066013813,
-6.232530741132271e-36,
-0.0024937218986451626,
0.08447113633155823,
-0.0055647604167461395,
-0.006974568124860525,
0.02693880721926689,
0.0012241830118000507,
0.118505097925663,
0.04465791955590248,
-0.012879332527518272,
0.0033988580107688904,
-0.019594671204686165,
0.03845382481813431,
-0.07007981091737747,
0.06870336085557938,
0.0680660605430603,
-0.012059160508215427,
-0.04323628172278404,
0.07290852814912796,
-0.0551738478243351,
-0.09325382113456726,
0.036509815603494644,
-0.028859881684184074,
-0.022268174216151237,
-0.006980242673307657,
0.012702565640211105,
0.000043354979425203055,
0.1281319111585617,
-0.04093942046165466,
-0.015324504114687443,
-0.004212898667901754,
-0.11090768128633499,
-0.06135847419500351,
0.0029841475188732147,
-0.016303718090057373,
0.07835430651903152,
-0.060560572892427444,
-0.0004920607898384333,
-0.003596880007535219,
-0.03387650474905968,
0.011490285396575928,
-0.058005232363939285,
0.04500710964202881,
0.04125149920582771,
-0.07192103564739227,
-0.15869759023189545,
-0.08706838637590408,
-0.022171882912516594,
0.005872613284736872,
0.0005342708900570869,
-0.04578886181116104,
0.04990856721997261,
-0.024090565741062164,
-0.018779318779706955,
0.04646216332912445,
0.02393711917102337,
-0.04821250960230827,
0.017474552616477013,
0.02546544373035431,
0.029428793117403984,
0.05962147191166878,
-0.03071196749806404,
0.048824820667505264,
0.02107301540672779,
0.08946149796247482,
0.08149849623441696,
0.007394983898848295,
-0.014640342444181442,
0.10904005169868469,
0.005574634298682213,
0.007362307049334049,
-0.025630444288253784,
-0.01837901771068573,
0.06099492684006691,
-0.03153660520911217,
0.013952890411019325,
-0.02199721522629261,
0.012748192995786667,
-0.07427298277616501,
-0.05652100592851639,
0.04580298438668251,
0.039950717240571976,
0.04682781547307968,
-0.0040847561322152615,
-0.04730208218097687,
-0.043661825358867645,
-0.03501191362738609,
0.03466499596834183,
-0.06574146449565887,
-0.028543254360556602,
-0.008341457694768906,
-0.044933021068573,
-0.059621717780828476,
-0.04414580389857292,
0.0005923647549934685,
-0.06680059432983398,
2.9090183696764736e-34,
-0.07342850416898727,
0.027226613834500313,
-0.04174787923693657,
0.12055633962154388,
-0.051851529628038406,
-0.08360868692398071,
0.04289323836565018,
0.004380478989332914,
-0.025220515206456184,
-0.0685800239443779,
-0.07918550819158554,
-0.07155311852693558,
0.03958294540643692,
0.013953548856079578,
0.03303036466240883,
0.011326491832733154,
-0.1571844220161438,
0.058317817747592926,
0.041320908814668655,
0.08729693293571472,
0.011654884554445744,
-0.004019086714833975,
-0.07430721819400787,
0.0005959064583294094,
-0.005232236813753843,
0.00970727950334549,
-0.048144467175006866,
0.11297053098678589,
0.0017832436133176088,
0.010675863362848759,
-0.005883533973246813,
-0.051350876688957214,
-0.09407948702573776,
-0.04146421328186989,
-0.013185471296310425,
0.0260960403829813,
0.042335301637649536,
0.07394877821207047,
0.04185708984732628,
0.0723065733909607,
0.12749862670898438,
-0.08376733958721161,
-0.05910523980855942,
0.05617467686533928,
0.02002083510160446,
-0.0017421192023903131,
-0.04914800822734833,
-0.08086293935775757,
-0.0024608797393739223,
-0.024518469348549843,
0.002083278726786375,
0.06410593539476395,
-0.03466309234499931,
0.05475940182805061,
0.0053168716840445995,
-0.043800827115774155,
-0.050270333886146545,
-0.020114729180932045,
0.0484822653234005,
-0.005820562597364187,
0.012378228828310966,
-0.062457453459501266,
0.032673776149749756,
-0.10024099797010422,
-0.0058504072949290276,
-0.0571647584438324,
0.016478372737765312,
-0.00004277387779438868,
-0.07014193385839462,
0.043923504650592804,
-0.03363766521215439,
-0.023100832477211952,
0.03383563458919525,
0.03348470479249954,
-0.027660246938467026,
-0.0030084794852882624,
0.010026352480053902,
0.004592011217027903,
-0.007618545554578304,
0.008909960277378559,
0.026004554703831673,
0.06531692296266556,
-0.015937285497784615,
-0.009923120960593224,
0.030727259814739227,
0.08282831311225891,
0.039750516414642334,
0.11849331110715866,
0.012583269737660885,
-0.00006686731649097055,
-0.030221443623304367,
0.07954337447881699,
0.06263162195682526,
0.02119065448641777,
-0.06717042624950409,
-4.77275960975021e-8,
-0.03073115460574627,
0.028253240510821342,
-0.016724489629268646,
-0.0275164395570755,
0.08856306970119476,
-0.01088062021881342,
-0.08837997168302536,
0.06641735136508942,
-0.015208999626338482,
-0.026847153902053833,
0.1444992572069168,
-0.032014891505241394,
-0.10897012799978256,
-0.046501997858285904,
-0.03799964860081673,
0.057417791336774826,
-0.03503977134823799,
-0.006671297363936901,
-0.006286670919507742,
-0.04013160988688469,
-0.005137908738106489,
0.07469548285007477,
-0.0026343802455812693,
0.0047864606603980064,
0.04348776489496231,
-0.0636054277420044,
0.041888706386089325,
0.11396752297878265,
0.07290608435869217,
-0.013096763752400875,
-0.08106255531311035,
0.005055248737335205,
0.0037861703895032406,
-0.03538298234343529,
-0.016493739560246468,
0.060816600918769836,
0.024293115362524986,
-0.00821771752089262,
-0.029647961258888245,
-0.07729659229516983,
0.01051422581076622,
-0.020099205896258354,
-0.011086639016866684,
0.03471257910132408,
0.06268563121557236,
-0.03948839381337166,
-0.07152023911476135,
-0.0187059473246336,
0.004717034287750721,
0.015008771792054176,
-0.010672210715711117,
0.008252636529505253,
-0.03905649855732918,
0.04539477452635765,
0.05720161274075508,
-0.029981138184666634,
-0.02183135785162449,
-0.029596852138638496,
0.0302564799785614,
0.07087886333465576,
-0.008906745351850986,
-0.020506706088781357,
-0.05111237242817879,
-0.023192526772618294
] |
Methods for Convex (L0,L1)-Smooth Optimization: Clipping, Acceleration, and Adaptivity | https://openreview.net/forum?id=0wmfzWPAFu | [
"Eduard Gorbunov",
"Nazarii Tupitsa",
"Sayantan Choudhury",
"Alen Aliev",
"Peter Richtárik",
"Samuel Horváth",
"Martin Takáč"
] | Poster | Due to the non-smoothness of optimization problems in Machine Learning, generalized smoothness assumptions have been gaining a lot of attention in recent years. One of the most popular assumptions of this type is $(L_0,L_1)$-smoothness (Zhang et al., 2020). In this paper, we focus on the class of (strongly) convex $(L_0,L_1)$-smooth functions and derive new convergence guarantees for several existing methods. In particular, we derive improved convergence rates for Gradient Descent with (Smoothed) Gradient Clipping and for Gradient Descent with Polyak Stepsizes. In contrast to the existing results, our rates do not rely on the standard smoothness assumption and do not suffer from the exponential dependency on the initial distance to the solution. We also extend these results to the stochastic case under the over-parameterization assumption, propose a new accelerated method for convex $(L_0,L_1)$-smooth optimization, and derive new convergence rates for Adaptive Gradient Descent (Malitsky and Mishchenko, 2020). | generalized smoothness, first-order optimization, convex optimization, Polyak stepsizes, gradient clipping, adaptive optimization, acceleration | null | 1,810 | null | [
-0.10528815537691116,
-0.10817962139844894,
-0.0059502702206373215,
-0.012224319390952587,
0.03481712192296982,
-0.036869682371616364,
0.02123682200908661,
-0.03180307149887085,
-0.06350462138652802,
-0.003000547643750906,
0.026582736521959305,
0.06508905440568924,
-0.005920145660638809,
-0.031480737030506134,
-0.010486375540494919,
0.053041011095047,
0.0207396000623703,
0.06917033344507217,
-0.004888468887656927,
0.008187995292246342,
-0.01625453308224678,
0.032406289130449295,
-0.00952057633548975,
0.008118846453726292,
0.04394517093896866,
-0.08459156006574631,
-0.03212222456932068,
-0.0012756857322528958,
-0.010273865424096584,
0.05260055884718895,
0.055894337594509125,
-0.07794716954231262,
-0.03080029785633087,
-0.04544529691338539,
0.006438874639570713,
0.03442206606268883,
-0.02076713740825653,
0.07995152473449707,
0.002786502940580249,
0.060220394283533096,
-0.050844062119722366,
-0.011461198329925537,
0.020568616688251495,
-0.0814034715294838,
0.06983509659767151,
-0.018049687147140503,
-0.016818121075630188,
-0.0779135599732399,
-0.02280714176595211,
0.021470781415700912,
-0.004590524826198816,
-0.0759863331913948,
-0.08253373950719833,
0.013145730830729008,
-0.020155545324087143,
-0.015406540594995022,
-0.029403287917375565,
0.0718948021531105,
0.02642415650188923,
0.0101660480722785,
-0.03464221581816673,
-0.16030040383338928,
0.00692773750051856,
0.04759491980075836,
-0.053064651787281036,
-0.04511956870555878,
0.09436319768428802,
-0.03377010300755501,
-0.004535615909844637,
0.11301274597644806,
-0.05414492264389992,
0.05987309291958809,
-0.07218018919229507,
0.03342057392001152,
-0.03923271223902702,
0.004906265530735254,
0.06572280079126358,
0.04495744779706001,
-0.012922207824885845,
-0.08004745095968246,
0.04170701652765274,
0.029013538733124733,
0.03822674974799156,
-0.05376794934272766,
-0.01165471225976944,
0.036394547671079636,
-0.027649221941828728,
0.01025687251240015,
0.00955343060195446,
-0.032114434987306595,
0.020835040137171745,
0.004147056024521589,
-0.09786500781774521,
0.04781971126794815,
-0.012577868066728115,
-0.06946179270744324,
0.031714316457509995,
-0.04184991493821144,
-0.004737560637295246,
0.1029258444905281,
-0.05528987944126129,
0.05779113620519638,
-0.003806865541264415,
-0.019620180130004883,
0.07814782857894897,
-0.05805331468582153,
0.061006780713796616,
0.015072491019964218,
0.0669398233294487,
-0.023451313376426697,
0.0503898523747921,
0.04108457639813423,
0.025181742385029793,
0.01851661689579487,
-0.014662138186395168,
-0.07708127796649933,
0.06230011209845543,
-0.09413566440343857,
0.043834298849105835,
0.06378668546676636,
-0.06925053894519806,
0.010087932460010052,
0.0014948552707210183,
0.06005203351378441,
0.008248825557529926,
0.036594804376363754,
-0.029039660468697548,
9.246531317419233e-34,
-0.029500911012291908,
0.04475880041718483,
0.07538042217493057,
-0.09514688700437546,
0.037132713943719864,
-0.005667069926857948,
0.042893391102552414,
-0.03299974650144577,
-0.04525850713253021,
0.020711122080683708,
-0.01439495850354433,
-0.044942427426576614,
-0.08331431448459625,
0.02739756554365158,
0.021552111953496933,
-0.01901991292834282,
-0.05795915797352791,
0.07052882015705109,
0.009940329939126968,
-0.031769778579473495,
0.06932105123996735,
-0.07636303454637527,
-0.026582801714539528,
0.00921565294265747,
-0.030827531591057777,
-0.0374227873980999,
0.06805513054132462,
0.03564968705177307,
-0.0014937726082280278,
0.03345702216029167,
-0.08571162819862366,
0.006388560403138399,
-0.04131648316979408,
0.017540927976369858,
0.008666623383760452,
0.05425442010164261,
0.03842740133404732,
-0.05258344113826752,
-0.01441319938749075,
0.046194855123758316,
-0.025776132941246033,
0.0746694952249527,
0.051389701664447784,
-0.03928887099027634,
-0.09412514418363571,
-0.12109292298555374,
0.005704085808247328,
0.03872980177402496,
-0.03245456889271736,
-0.06593594700098038,
-0.05725343897938728,
0.029987521469593048,
-0.030271610245108604,
-0.030559243634343147,
-0.033465370535850525,
-0.020085876807570457,
0.05226918309926987,
0.03261129558086395,
0.035917311906814575,
0.053080569952726364,
-0.03037456050515175,
-0.022053074091672897,
-0.015206720679998398,
-0.020058007910847664,
-0.030096398666501045,
0.03161371126770973,
0.04108896106481552,
0.05336308479309082,
-0.011313118040561676,
-0.08878926187753677,
0.0055152480490505695,
-0.026732325553894043,
-0.021828563883900642,
-0.10043669492006302,
0.06272431463003159,
0.011390737257897854,
0.025498148053884506,
-0.11393123865127563,
-0.011988023295998573,
-0.014454295858740807,
-0.06617710739374161,
0.06396787613630295,
-0.017175979912281036,
-0.06834957748651505,
-0.037502992898225784,
-0.06253686547279358,
0.0650319755077362,
-0.08556630462408066,
-0.038736868649721146,
-0.007962441071867943,
-0.15329208970069885,
0.02832097001373768,
-0.010844244621694088,
0.021762175485491753,
0.00462622195482254,
-1.36627313503441e-33,
0.013236732222139835,
-0.008311454206705093,
-0.04322568699717522,
0.11061037331819534,
0.023229295387864113,
0.0033942919690161943,
-0.05791980400681496,
-0.06336744874715805,
-0.04557031765580177,
-0.10113528370857239,
0.003502901876345277,
0.016587985679507256,
0.029381239786744118,
-0.017399033531546593,
0.0075746257789433,
-0.004176746122539043,
-0.0006214335444383323,
0.12426791340112686,
-0.041326962411403656,
0.018093854188919067,
0.026527339592576027,
-0.005048657767474651,
-0.037254706025123596,
0.03822072595357895,
0.013843717984855175,
0.023289453238248825,
-0.0765702947974205,
0.15433241426944733,
0.032240405678749084,
-0.07390201091766357,
-0.08752526342868805,
0.011668129824101925,
-0.02219807729125023,
-0.01813204027712345,
0.0015337081858888268,
0.11235621571540833,
-0.009499471634626389,
0.046226296573877335,
0.065312460064888,
0.0824500098824501,
0.0783025398850441,
0.006554508116096258,
-0.015065985731780529,
0.03746994212269783,
0.09054064750671387,
-0.02637719362974167,
0.0056307511404156685,
-0.01893172413110733,
-0.015296642668545246,
0.0022122091613709927,
-0.02781594544649124,
0.0116820577532053,
-0.04571375995874405,
0.15313300490379333,
-0.05386403575539589,
0.025006724521517754,
-0.018842924386262894,
-0.04788997769355774,
-0.03421783074736595,
-0.008626595139503479,
-0.06593876332044601,
0.07244362682104111,
-0.09280016273260117,
0.13566306233406067,
0.048116765916347504,
-0.042520955204963684,
0.017965244129300117,
0.03669699653983116,
0.013523362576961517,
0.034790921956300735,
0.002044761786237359,
0.0215082336217165,
0.016718178987503052,
0.05156771466135979,
-0.02015542797744274,
-0.018461525440216064,
0.05859799683094025,
-0.03732273727655411,
-0.04715724289417267,
-0.0711081251502037,
0.12592190504074097,
0.03504418209195137,
0.061739351600408554,
-0.009855336509644985,
0.07229337096214294,
-0.0018251746660098433,
0.07175484299659729,
0.044200025498867035,
-0.022550104185938835,
-0.021117113530635834,
-0.019640536978840828,
-0.02181398682296276,
0.007479635067284107,
-0.025218414142727852,
-0.0818633958697319,
-4.2316170123513075e-8,
-0.04066167771816254,
-0.004511257633566856,
-0.0031375486869364977,
0.008665106259286404,
0.07151991128921509,
0.009016319178044796,
-0.026354333385825157,
0.02373003214597702,
-0.03211609274148941,
-0.10492843389511108,
0.03492841124534607,
-0.021689165383577347,
0.05655522271990776,
-0.017485303804278374,
0.013754910789430141,
0.07683670520782471,
-0.06962333619594574,
0.026802897453308105,
-0.005281771067529917,
0.05323592945933342,
-0.07259901612997055,
0.029948562383651733,
0.053279343992471695,
-0.0696810856461525,
0.02506076544523239,
-0.024424361065030098,
0.0016304892487823963,
0.05954236537218094,
0.14454638957977295,
0.0014961597044020891,
-0.11092802882194519,
0.08595531433820724,
0.005026275757700205,
0.012130366638302803,
0.03609529882669449,
0.10015023499727249,
0.03343698009848595,
0.02510068006813526,
-0.009325575083494186,
0.03197191655635834,
-0.011906682513654232,
0.04757353290915489,
-0.011150715872645378,
-0.04348613694310188,
0.024091331288218498,
-0.038873765617609024,
0.00144987425301224,
-0.06355924904346466,
-0.029278110712766647,
-0.010920992121100426,
0.05384194105863571,
0.0390206053853035,
0.04862949997186661,
0.08229434490203857,
0.05671723932027817,
-0.01826263964176178,
-0.030094293877482414,
-0.03289273381233215,
0.027481799945235252,
0.09594397246837616,
-0.009967630729079247,
0.003915960900485516,
-0.020803816616535187,
0.044348787516355515
] |
Minimax Optimal Reinforcement Learning with Quasi-Optimism | https://openreview.net/forum?id=i8LCUpKvAz | [
"Harin Lee",
"Min-hwan Oh"
] | Poster | In our quest for a reinforcement learning (RL) algorithm that is both practical and provably optimal, we introduce EQO (Exploration via Quasi-Optimism). Unlike existing minimax optimal approaches, EQO avoids reliance on empirical variances and employs a simple bonus term proportional to the inverse of the state-action visit count. Central to EQO is the concept of *quasi-optimism*, where estimated values need not be fully optimistic, allowing for a simpler yet effective exploration strategy. The algorithm achieves the sharpest known regret bound for tabular RL under the mildest assumptions, proving that fast convergence can be attained with a practical and computationally efficient approach. Empirical evaluations demonstrate that EQO consistently outperforms existing algorithms in both regret performance and computational efficiency, providing the best of both theoretical soundness and practical effectiveness. | Reinforcement Learning, Tabular Reinforcement Learning, Regret Analysis | We propose a simple and practical algorithm with the tightest minimax regret bound for tabular reinforcement learning. | 1,797 | 2503.00810 | [
-0.04151420667767525,
-0.0022403409238904715,
0.021290451288223267,
0.05141456052660942,
-0.031074149534106255,
0.008743786253035069,
-0.01961897499859333,
0.005208355840295553,
0.04126172140240669,
0.028661973774433136,
-0.057494211941957474,
-0.006139775738120079,
0.06381280720233917,
0.02128099463880062,
-0.06589113920927048,
0.0042257471941411495,
0.06355001032352448,
0.00006381455750670284,
0.042914796620607376,
-0.008976846002042294,
-0.012989573180675507,
-0.04949156567454338,
0.04452255368232727,
-0.01809253916144371,
0.0489780493080616,
-0.0006818410474807024,
-0.04566211625933647,
0.06342140585184097,
-0.015789475291967392,
0.010135984979569912,
0.01555490680038929,
0.00249291374348104,
0.017408888787031174,
-0.0889163538813591,
-0.0817081555724144,
0.026934459805488586,
-0.15553264319896698,
-0.07451209425926208,
-0.004184121731668711,
0.010086639784276485,
-0.029438519850373268,
0.026662996038794518,
0.04853530600667,
0.02907872572541237,
0.044545065611600876,
-0.03538745641708374,
0.028815744444727898,
-0.07786305993795395,
0.014632795006036758,
-0.010031797923147678,
-0.012026225216686726,
-0.01883513852953911,
-0.0005815663025714457,
-0.015291575342416763,
-0.0019145164405927062,
0.07581935077905655,
-0.008659781888127327,
-0.011279807426035404,
-0.004396934062242508,
-0.009136805310845375,
-0.016833888366818428,
-0.12211624532938004,
0.03929147124290466,
-0.032323043793439865,
-0.03223254904150963,
-0.037829238921403885,
-0.01235128566622734,
0.0235979612916708,
-0.014622039161622524,
0.060422543436288834,
0.005175822880119085,
0.04267677292227745,
0.019055839627981186,
0.0006399835110642016,
0.025994321331381798,
-0.04181060940027237,
0.054476696997880936,
0.005701795220375061,
-0.02250167354941368,
-0.016068290919065475,
-0.012703717686235905,
0.0037416121922433376,
-0.03932502493262291,
-0.0082700215280056,
0.029682420194149017,
-0.0002953545772470534,
-0.014669405296444893,
0.019137034192681313,
0.13168907165527344,
-0.0006682896055281162,
0.05474427342414856,
-0.04790014773607254,
-0.04611864313483238,
-0.001384612638503313,
-0.030618304386734962,
0.035807546228170395,
-0.026464873924851418,
-0.06450455635786057,
-0.07215806841850281,
0.06462307274341583,
0.07199805229902267,
0.10818649083375931,
-0.02813049778342247,
-0.03842729702591896,
-0.01113111525774002,
-0.08669467270374298,
0.04443088173866272,
0.00977596640586853,
0.04345432296395302,
-0.07210522145032883,
0.0016222471604123712,
-0.020486880093812943,
0.10555130988359451,
0.035307347774505615,
-0.0499989353120327,
0.053378183394670486,
0.0034414282999932766,
0.011225040070712566,
-0.003151070326566696,
0.08611626923084259,
0.062038492411375046,
-0.0744086280465126,
0.08900284022092819,
0.0691869929432869,
0.0016070069978013635,
-0.0004671814094763249,
-0.008411397226154804,
4.693503191045675e-33,
0.04966861754655838,
-0.03230281546711922,
0.09550812095403671,
-0.06982322782278061,
0.022154439240694046,
0.037833619862794876,
0.1031150370836258,
-0.027875518426299095,
-0.056931670755147934,
-0.059374723583459854,
0.03432717174291611,
0.061047859489917755,
-0.055562540888786316,
0.06144804507493973,
0.04015339910984039,
-0.04620957747101784,
-0.07699461281299591,
0.06820380687713623,
0.00710325688123703,
-0.04634007066488266,
0.010816585272550583,
-0.10835336148738861,
0.02924654632806778,
-0.07856730371713638,
0.08722008019685745,
0.04637110233306885,
-0.013951173052191734,
0.013272332958877087,
0.011264035478234291,
0.004990663845092058,
-0.025952089577913284,
0.05706239491701126,
-0.06313738971948624,
0.06082487106323242,
0.026250548660755157,
0.025321805849671364,
-0.037239961326122284,
0.0420120470225811,
0.022855734452605247,
0.03969552740454674,
-0.043091002851724625,
0.0030035884119570255,
0.037472475320100784,
-0.0028604462277144194,
-0.06741181761026382,
-0.08676645904779434,
0.037759166210889816,
-0.016544505953788757,
-0.07686357200145721,
-0.04651373624801636,
0.017237765714526176,
-0.029599878937005997,
-0.07055582106113434,
-0.044104546308517456,
-0.03961838781833649,
-0.0021364414133131504,
0.012357056140899658,
0.11973200738430023,
-0.04756907746195793,
0.015501761808991432,
-0.014069456607103348,
-0.03335344046354294,
-0.02504066936671734,
-0.0025594239123165607,
-0.09991025924682617,
0.16562463343143463,
-0.03192751854658127,
0.010274595580995083,
0.037171296775341034,
0.010888817720115185,
0.057089462876319885,
0.022683298215270042,
0.037250716239213943,
-0.11811327189207077,
0.12583906948566437,
-0.003697820007801056,
0.08899571001529694,
-0.012654699385166168,
0.07397111505270004,
-0.0861012190580368,
0.005048768129199743,
-0.016794584691524506,
-0.07676655054092407,
-0.005757991690188646,
-0.00014279555762186646,
-0.05838661640882492,
-0.0011267844820395112,
-0.04615605250000954,
-0.06324273347854614,
0.008651461452245712,
-0.10762717574834824,
0.000591086340136826,
0.03310174494981766,
0.04436541348695755,
0.0023014377802610397,
-4.73553886903108e-33,
0.07574417442083359,
-0.08461154252290726,
0.049680884927511215,
0.0046625505201518536,
0.006615240592509508,
-0.0013015330769121647,
-0.0093728331848979,
-0.03381378576159477,
0.01577494479715824,
-0.04377152398228645,
-0.025350186973810196,
0.08864358812570572,
0.0693357065320015,
0.05860895290970802,
-0.0024320620577782393,
-0.05215171352028847,
0.09038815647363663,
0.020777858793735504,
0.017857467755675316,
-0.029971661046147346,
0.0070693097077310085,
0.05850309133529663,
-0.09686801582574844,
0.10155293345451355,
-0.04170207679271698,
0.02407226525247097,
-0.018218088895082474,
0.0321468748152256,
-0.04482703283429146,
-0.019349565729498863,
-0.08100788295269012,
-0.062205970287323,
-0.027209842577576637,
-0.006315174046903849,
-0.012019863352179527,
0.09615408629179001,
0.07860975712537766,
0.013554377481341362,
-0.11016140133142471,
0.08007454127073288,
0.05090702697634697,
-0.03303471952676773,
0.04240361973643303,
-0.03238774836063385,
0.024908414110541344,
-0.01108388788998127,
-0.023594031110405922,
0.01603028178215027,
-0.03586635738611221,
0.060998328030109406,
-0.02604691870510578,
0.022063659504055977,
-0.07524831593036652,
0.03749389946460724,
-0.024968719109892845,
-0.048861827701330185,
-0.10264018177986145,
-0.035386040806770325,
0.057536058127880096,
0.012368029914796352,
-0.08011998981237411,
0.03332541137933731,
-0.05770120769739151,
0.0869777724146843,
-0.0637941062450409,
0.049192897975444794,
-0.03371672332286835,
0.032998617738485336,
0.023208143189549446,
-0.019126616418361664,
-0.0069713350385427475,
0.038314975798130035,
0.01544199138879776,
-0.022463155910372734,
0.03721504658460617,
-0.025804845616221428,
-0.010977228172123432,
0.03165165334939957,
-0.005196324549615383,
0.017207292839884758,
0.045971017330884933,
0.06324005126953125,
0.08605604618787766,
0.004187467508018017,
0.06185709312558174,
0.04534651339054108,
0.011889737099409103,
0.10673628747463226,
0.0018666678806766868,
-0.014142789877951145,
-0.02884303778409958,
-0.08824271708726883,
0.04487970098853111,
-0.017954660579562187,
-0.046698879450559616,
-4.3824336160014354e-8,
-0.0911463126540184,
0.020083976909518242,
0.10497764497995377,
-0.0021050693467259407,
0.15878307819366455,
0.003908971324563026,
0.04468483850359917,
0.011867964640259743,
-0.06965169310569763,
-0.035300176590681076,
0.08151426166296005,
0.028992217034101486,
0.03781777620315552,
-0.04689508303999901,
-0.030574265867471695,
0.05795145779848099,
0.011677050031721592,
-0.08720895648002625,
0.016378024592995644,
-0.0051217046566307545,
0.016096213832497597,
0.034872256219387054,
0.022832773625850677,
-0.0683184340596199,
0.031129777431488037,
-0.03584663197398186,
-0.044392313808202744,
0.04709884151816368,
0.02372392639517784,
-0.007391833700239658,
-0.05030630901455879,
-0.0538170300424099,
-0.0021128873340785503,
0.07585670799016953,
-0.019142327830195427,
0.031593967229127884,
0.07320648431777954,
-0.023100143298506737,
0.02607388235628605,
-0.0407751239836216,
-0.06404966861009598,
0.04664546251296997,
-0.044327352195978165,
-0.03156045824289322,
-0.06529783457517624,
0.03312210738658905,
-0.05097963660955429,
-0.09313534945249557,
0.014820015989243984,
-0.12885388731956482,
0.07307649403810501,
-0.03405600041151047,
-0.06162872910499573,
0.08221442252397537,
0.04038073867559433,
-0.0034505033399909735,
-0.015694160014390945,
-0.02906116284430027,
-0.03400708734989166,
0.11303717643022537,
0.0457642488181591,
-0.048031602054834366,
-0.05901946872472763,
0.021802587434649467
] |
Weakly-Supervised Affordance Grounding Guided by Part-Level Semantic Priors | https://openreview.net/forum?id=0823rvTIhs | [
"Peiran Xu",
"Yadong MU"
] | Poster | In this work, we focus on the task of weakly supervised affordance grounding, where a model is trained to identify affordance regions on objects using human-object interaction images and egocentric object images without dense labels.
Previous works are mostly built upon class activation maps, which are effective for semantic segmentation but may not be suitable for locating actions and functions. Leveraging recent advanced foundation models, we develop a supervised training pipeline based on pseudo labels. The pseudo labels are generated from an off-the-shelf part segmentation model, guided by a mapping from affordance to part names.
Furthermore, we introduce three key enhancements to the baseline model: a label refining stage, a fine-grained feature alignment process, and a lightweight reasoning module. These techniques harness the semantic knowledge of static objects embedded in off-the-shelf foundation models to improve affordance learning, effectively bridging the gap between objects and actions.
Extensive experiments demonstrate that the performance of the proposed model has achieved a breakthrough improvement over existing methods. | weakly supervised affordance grounding, foundation model, pseudo label | We propose a pseudo-label based method for weakly supervised affordance grounding, utilizing the semantic priors of vision foundation models. | 1,796 | null | [
0.06996262073516846,
-0.016705326735973358,
0.072999969124794,
0.0688520222902298,
0.11901715397834778,
-0.060049764811992645,
0.03219136223196983,
0.0031921209301799536,
-0.002159882802516222,
-0.06398842483758926,
-0.05139261856675148,
-0.05570586025714874,
-0.012386943213641644,
0.04697318747639656,
0.0335850790143013,
-0.017683889716863632,
0.11582312732934952,
0.045847661793231964,
-0.0898834764957428,
0.049719519913196564,
-0.05036957189440727,
0.024586528539657593,
0.11281352490186691,
-0.024305522441864014,
-0.009361582808196545,
0.051174066960811615,
0.01112180482596159,
-0.0050062742084264755,
0.03578357398509979,
-0.02116469107568264,
0.0819554254412651,
-0.02006162703037262,
0.02780219539999962,
0.03225789591670036,
-0.037649400532245636,
0.127073273062706,
0.04041200131177902,
0.005834739655256271,
-0.037265874445438385,
-0.07712243497371674,
0.020884810015559196,
-0.0698857307434082,
-0.024290133267641068,
0.04374588280916214,
0.06667274981737137,
0.0585368350148201,
0.031509704887866974,
-0.04630924388766289,
0.017967374995350838,
0.006534371059387922,
-0.0266305860131979,
-0.045530036091804504,
-0.02225871942937374,
0.012468408793210983,
-0.0582897886633873,
0.08828693628311157,
0.0961546003818512,
-0.08927514404058456,
0.04874826967716217,
0.0076172794215381145,
0.04595886543393135,
-0.015718642622232437,
-0.005992596503347158,
-0.018822282552719116,
0.012519842945039272,
-0.05696948617696762,
0.05705288425087929,
-0.030975256115198135,
-0.006468605250120163,
-0.02514476329088211,
0.05142521485686302,
0.044579096138477325,
0.05144847556948662,
-0.06605159491300583,
0.018535520881414413,
0.048158805817365646,
-0.02124439738690853,
-0.06332850456237793,
-0.010683497413992882,
-0.1108609065413475,
-0.011789741925895214,
-0.000947579275816679,
0.056694723665714264,
0.004230080172419548,
0.062428705394268036,
0.0344180203974247,
0.01718813367187977,
0.018774887546896935,
-0.06427852064371109,
0.0029960693791508675,
-0.08613324165344238,
-0.07375738024711609,
0.031149206683039665,
-0.008272210136055946,
-0.000692257599439472,
-0.0012584145879372954,
-0.020497096702456474,
-0.1362939327955246,
-0.04321333393454552,
0.08524502068758011,
-0.04183204472064972,
0.07687914371490479,
0.04708167538046837,
-0.010624735616147518,
0.027860993519425392,
-0.02572990581393242,
-0.037423014640808105,
0.057612095028162,
0.025637229904532433,
-0.049114327877759933,
0.061044659465551376,
-0.034007374197244644,
-0.019066160544753075,
0.032454028725624084,
-0.05759590491652489,
-0.09878390282392502,
-0.014320244081318378,
-0.11033548414707184,
0.07422912120819092,
-0.09426840394735336,
0.02643447555601597,
0.025080259889364243,
-0.00018744994304142892,
0.04329643398523331,
0.028115903958678246,
-0.11918321251869202,
-0.039719343185424805,
9.33774784233312e-34,
0.027241338044404984,
-0.039401840418577194,
0.08065151423215866,
-0.01915922202169895,
0.04379341006278992,
-0.027018658816814423,
0.002480283612385392,
0.008104607462882996,
-0.01999068446457386,
0.07092556357383728,
0.05951013043522835,
-0.040545664727687836,
-0.012933220714330673,
0.07919400930404663,
-0.018617473542690277,
-0.08339724689722061,
0.0035288052167743444,
0.024456707760691643,
-0.056371983140707016,
0.004253878258168697,
-0.06512735038995743,
0.0010654826182872057,
-0.02472512796521187,
0.004073452204465866,
0.022916387766599655,
-0.006866561248898506,
0.07749471813440323,
0.03364917263388634,
-0.010276229120790958,
0.0014027564320713282,
-0.053155869245529175,
-0.01301486510783434,
0.05330516770482063,
0.05672326683998108,
0.015100411139428616,
-0.023848680779337883,
-0.03231481835246086,
-0.015832828357815742,
0.05215026065707207,
0.01306210272014141,
0.005324395373463631,
0.054230254143476486,
0.024629976600408554,
-0.03252824395895004,
-0.05892939493060112,
0.01912742294371128,
0.02803655155003071,
0.08334646373987198,
-0.06346088647842407,
0.04756472632288933,
-0.0007185316644608974,
-0.026325387880206108,
0.023613575845956802,
-0.06663345545530319,
-0.17842978239059448,
-0.06466160714626312,
-0.00799468532204628,
0.05417204648256302,
-0.011792827397584915,
-0.08045509457588196,
-0.03075222298502922,
0.04185225069522858,
-0.07867297530174255,
0.029534298926591873,
-0.055918965488672256,
0.08480335772037506,
0.04521883651614189,
0.06713489443063736,
0.030824612826108932,
0.05090313032269478,
-0.07136444747447968,
0.07064571231603622,
0.05069880560040474,
-0.013303237035870552,
0.08462545275688171,
0.014043818227946758,
-0.004993127193301916,
-0.07840604335069656,
0.023300733417272568,
0.04076937586069107,
-0.0029441097285598516,
-0.048536572605371475,
-0.016290932893753052,
0.030215317383408546,
-0.050588496029376984,
0.018614843487739563,
0.05911749228835106,
-0.0831235721707344,
-0.027234533801674843,
-0.01347256451845169,
-0.04495604708790779,
-0.05773117393255234,
-0.07171367108821869,
0.03479362651705742,
0.00044990165042690933,
1.500737107305506e-34,
0.08090228587388992,
0.008683302439749241,
-0.031047364696860313,
0.048577576875686646,
-0.009152987971901894,
-0.07219817489385605,
-0.039676081389188766,
0.030249685049057007,
-0.03946773707866669,
-0.019857516512274742,
-0.016690313816070557,
0.009318104945123196,
-0.01956985704600811,
-0.047132205218076706,
0.0049989111721515656,
-0.0051578073762357235,
-0.035390544682741165,
0.040757711976766586,
0.021078942343592644,
0.08491362631320953,
0.03482241556048393,
0.08961866796016693,
-0.06917475908994675,
0.001953263534232974,
-0.08204872906208038,
0.057844724506139755,
0.0011286132503300905,
0.009993493556976318,
-0.00582491559907794,
0.011416844092309475,
-0.07550757378339767,
-0.06982540339231491,
0.04564129188656807,
-0.08952778577804565,
-0.13448156416416168,
0.005826272070407867,
-0.028969870880246162,
-0.06279979646205902,
-0.006468463223427534,
0.022164098918437958,
-0.004332749173045158,
-0.05817384645342827,
-0.0800585150718689,
0.03476228564977646,
-0.06088742986321449,
-0.04263338819146156,
-0.03937568888068199,
0.026149598881602287,
-0.09009522199630737,
0.04729818180203438,
-0.06432796269655228,
0.06308930367231369,
-0.0952286422252655,
-0.0031285856384783983,
-0.0191954318434,
-0.021260296925902367,
0.025853026658296585,
-0.020713532343506813,
0.09203244000673294,
-0.02087434194982052,
-0.04060734808444977,
0.05468441918492317,
0.058654408901929855,
0.0179928969591856,
0.05910775437951088,
-0.0017734509892761707,
-0.045668426901102066,
-0.03341924026608467,
0.01079168077558279,
0.03014422208070755,
0.0670829564332962,
0.024222591891884804,
0.042307015508413315,
0.0018404192524030805,
0.024369116872549057,
-0.012586935423314571,
0.0008765485254116356,
0.012268339283764362,
0.03638231009244919,
-0.03281496837735176,
0.07129202038049698,
-0.1106538400053978,
0.049920301884412766,
0.09023191034793854,
0.04037335142493248,
0.07904715836048126,
-0.001129256677813828,
0.05936979129910469,
-0.027098773047327995,
0.02448039874434471,
-0.06698518991470337,
0.06088464334607124,
0.037201810628175735,
0.11574887484312057,
-0.0334693044424057,
-3.634513490169411e-8,
-0.0982314795255661,
0.09016502648591995,
-0.007224665489047766,
0.02944222092628479,
0.010286052711308002,
-0.0036480226553976536,
0.03237651288509369,
0.1345527321100235,
-0.011521883308887482,
0.006889029406011105,
0.026434987783432007,
0.01447533443570137,
0.013369079679250717,
-0.10499956458806992,
-0.03130209073424339,
0.07146076112985611,
0.03890601918101311,
0.04775581881403923,
-0.0586123988032341,
0.07145045697689056,
-0.00008905045251594856,
-0.003219681093469262,
0.09825781732797623,
-0.03222273290157318,
0.0473940446972847,
-0.06765991449356079,
-0.003687706310302019,
0.11132502555847168,
0.06122375279664993,
-0.02509138360619545,
0.01277645118534565,
0.009410440921783447,
-0.022502252832055092,
0.041925907135009766,
0.04949246346950531,
0.04401106387376785,
0.0027305595576763153,
-0.05045001581311226,
-0.03656227886676788,
-0.01068549882620573,
-0.012238714843988419,
0.01403488777577877,
-0.017342032864689827,
0.04904903843998909,
0.05014518275856972,
0.05633789673447609,
-0.029248209670186043,
-0.03902866318821907,
0.0249966848641634,
0.01363281812518835,
-0.035042643547058105,
-0.0073246737010777,
-0.05206632614135742,
0.09282844513654709,
-0.04183203727006912,
-0.010160732083022594,
0.005637628026306629,
0.000963553786277771,
0.026730962097644806,
0.09850385040044785,
0.023762766271829605,
0.017373334616422653,
-0.04644044488668442,
-0.03134845942258835
] |
Safety Layers in Aligned Large Language Models: The Key to LLM Security | https://openreview.net/forum?id=kUH1yPMAn7 | [
"Shen Li",
"Liuyi Yao",
"Lan Zhang",
"Yaliang Li"
] | Poster | Aligned LLMs are secure, capable of recognizing and refusing to answer malicious questions. However, the role of internal parameters in maintaining such security is not well understood yet, further these models can be vulnerable to security degradation when subjected to fine-tuning attacks. To address these challenges, our work uncovers the mechanism behind security in aligned LLMs at the parameter level, identifying a small set of contiguous layers in the middle of the model that are crucial for distinguishing malicious queries from normal ones, referred to as "safety layers". We first confirm the existence of these safety layers by analyzing variations in input vectors within the model's internal layers. Additionally, we leverage the over-rejection phenomenon and parameters scaling analysis to precisely locate the safety layers. Building on these findings, we propose a novel fine-tuning approach, Safely Partial-Parameter Fine-Tuning (SPPFT), that fixes the gradient of the safety layers during fine-tuning to address the security degradation. Our experiments demonstrate that the proposed approach can significantly preserve LLM security while maintaining performance and reducing computational resources compared to full fine-tuning. | safety layers, LLM security, LLM alignment | null | 1,788 | 2408.17003 | [
-0.04940874129533768,
-0.04344554618000984,
-0.021343694999814034,
0.01304942648857832,
0.03838615119457245,
0.04114142060279846,
0.01986839808523655,
-0.011296081356704235,
0.04676782339811325,
-0.054528798907995224,
-0.05228080600500107,
-0.06195269152522087,
0.07973102480173111,
-0.0151484040543437,
0.009512635879218578,
0.008383069187402725,
0.05033498629927635,
0.11740788072347641,
-0.1027153879404068,
-0.003999680280685425,
-0.034449122846126556,
0.03954027220606804,
-0.029042163863778114,
-0.01652836613357067,
-0.06967296451330185,
-0.030368223786354065,
-0.04394993931055069,
-0.005834161303937435,
-0.0078117516823112965,
-0.03454384580254555,
0.01065381895750761,
0.00863343570381403,
-0.052717551589012146,
0.10463331639766693,
-0.01783987134695053,
-0.02132008597254753,
-0.031648945063352585,
-0.06353582441806793,
-0.026434356346726418,
-0.11091462522745132,
-0.0661073699593544,
-0.017229758203029633,
-0.030531005933880806,
0.06299600005149841,
-0.015996810048818588,
-0.053109489381313324,
-0.0033278046175837517,
0.018029896542429924,
-0.02675444632768631,
-0.03808460384607315,
-0.04987071454524994,
-0.012351161800324917,
-0.024570966139435768,
0.06029006466269493,
-0.012959768995642662,
-0.05370224639773369,
-0.012141696177423,
0.025898827239871025,
-0.02597132697701454,
0.005238729063421488,
-0.07712298631668091,
-0.06381335109472275,
-0.017693249508738518,
0.043358247727155685,
-0.010881726630032063,
-0.01735731028020382,
0.0015341913094744086,
0.039175692945718765,
0.012459538877010345,
0.09359699487686157,
-0.06263942271471024,
0.032340288162231445,
-0.03984398767352104,
0.0511101670563221,
0.04860867187380791,
0.10842996090650558,
-0.0561554916203022,
-0.04350454360246658,
0.0268460214138031,
-0.06768935918807983,
0.041367560625076294,
0.003060943214222789,
0.027449514716863632,
0.05218878760933876,
0.10715033859014511,
0.07107479870319366,
0.027569107711315155,
-0.03957315906882286,
0.14326786994934082,
0.04951896518468857,
-0.03347437083721161,
-0.04783652350306511,
0.04391849413514137,
0.07340074330568314,
0.0893625020980835,
0.017478318884968758,
-0.0006741294055245817,
-0.062173422425985336,
-0.024913791567087173,
0.068870909512043,
0.028744082897901535,
0.039310578256845474,
0.017852066084742546,
-0.023657767102122307,
0.02033657766878605,
-0.00044678730773739517,
-0.005798807833343744,
0.01904319040477276,
0.0029706954956054688,
-0.06874175369739532,
-0.009797674603760242,
0.05578935891389847,
0.05156835913658142,
-0.006026983726769686,
0.014864775352180004,
-0.04876923933625221,
0.03202611953020096,
-0.044570617377758026,
-0.02328973077237606,
0.031043291091918945,
0.0144088976085186,
0.03957859054207802,
0.0038877781480550766,
-0.013363662175834179,
-0.015566498041152954,
-0.008781001903116703,
-0.05885222554206848,
-1.615352398285487e-34,
0.031806640326976776,
0.074106864631176,
-0.0703183189034462,
0.019031651318073273,
-0.04227793589234352,
0.03192276880145073,
0.046109236776828766,
0.07412607222795486,
-0.05454428866505623,
0.04090206325054169,
-0.044773977249860764,
-0.02169581688940525,
0.031013676896691322,
0.00932866521179676,
0.08622535318136215,
0.07321055978536606,
0.03572729974985123,
0.0807851105928421,
-0.027699923142790794,
-0.037553396075963974,
0.0663263127207756,
-0.028802162036299706,
0.044341567903757095,
-0.04164319112896919,
0.036943353712558746,
0.006009511649608612,
-0.00338363298214972,
-0.0013914677547290921,
-0.06914962828159332,
0.031513772904872894,
-0.1459578424692154,
-0.06229459494352341,
0.007010983768850565,
0.04754788428544998,
0.06762130558490753,
-0.008318094536662102,
-0.015776097774505615,
-0.017652707174420357,
0.029911503195762634,
-0.014779223129153252,
0.01202109269797802,
0.03768734633922577,
0.025585852563381195,
0.04428928717970848,
-0.025487955659627914,
-0.007819881662726402,
-0.11898240447044373,
-0.008634439669549465,
-0.018535753712058067,
-0.09748616069555283,
0.02141285315155983,
0.028356317430734634,
0.015106362290680408,
0.003605310805141926,
-0.06602995842695236,
-0.01002204604446888,
0.007174069527536631,
0.04854021221399307,
-0.06272277981042862,
0.12147384136915207,
-0.060544632375240326,
0.03493993356823921,
-0.05271795019507408,
0.0014450820162892342,
0.05512566864490509,
-0.07178276032209396,
-0.0022935201413929462,
0.06567402929067612,
0.048614539206027985,
0.003993289079517126,
-0.03295058012008667,
-0.023761041462421417,
0.019781269133090973,
0.029516713693737984,
-0.11712583154439926,
-0.06274806708097458,
0.04196750745177269,
-0.011335390619933605,
0.09162138402462006,
-0.009522154927253723,
0.06732592731714249,
0.024938635528087616,
0.046872302889823914,
-0.0054217176511883736,
-0.11086099594831467,
-0.1215430349111557,
-0.030830681324005127,
-0.024025220423936844,
-0.03463607281446457,
0.03750939667224884,
-0.013631767593324184,
-0.04471311345696449,
0.062265150249004364,
-0.007863534614443779,
-0.053738318383693695,
-9.708342640239e-34,
-0.060383349657058716,
-0.09120719134807587,
-0.07124149054288864,
0.09504427015781403,
-0.1293942630290985,
-0.0890045315027237,
0.004632541444152594,
0.06918834149837494,
-0.03589717298746109,
-0.0367250069975853,
-0.033451106399297714,
-0.0224982388317585,
0.08677765727043152,
-0.05585664510726929,
0.07596906274557114,
0.016597941517829895,
-0.06580565869808197,
0.0381595715880394,
0.02323145978152752,
0.0552743636071682,
0.013149813748896122,
0.08458609879016876,
-0.1281578093767166,
0.08854597061872482,
-0.0033321769442409277,
0.013578692451119423,
-0.04645565524697304,
0.004642271902412176,
0.027838274836540222,
0.08438317477703094,
-0.046973083168268204,
0.04953264072537422,
-0.026386598125100136,
-0.030607648193836212,
-0.035156525671482086,
-0.008751948364078999,
0.045453328639268875,
-0.003157195867970586,
0.08199655264616013,
0.02303367853164673,
0.06804235279560089,
0.045046672224998474,
-0.06455297023057938,
-0.015806453302502632,
0.027086952701210976,
-0.036573320627212524,
-0.047404684126377106,
0.015544826164841652,
0.04770444333553314,
-0.07479076087474823,
-0.018900377675890923,
-0.050936657935380936,
0.0723392441868782,
-0.0024461334105581045,
-0.0008014650084078312,
-0.03838344290852547,
-0.06194597855210304,
-0.044940341264009476,
0.05635777860879898,
0.0065390197560191154,
-0.03434476628899574,
-0.031312935054302216,
0.0033136417623609304,
-0.07137338072061539,
0.04138937219977379,
0.004027757328003645,
-0.03019787184894085,
-0.03460012748837471,
-0.0013929045526310802,
-0.06367330253124237,
0.08994363248348236,
-0.07351541519165039,
-0.024348197504878044,
0.043921858072280884,
-0.019360139966011047,
-0.09739028662443161,
0.034126874059438705,
-0.0993645191192627,
0.0007450159173458815,
0.06253551691770554,
0.06073817238211632,
0.04629693925380707,
-0.03608064353466034,
0.09394372999668121,
0.04619750753045082,
0.01755833812057972,
0.02600858360528946,
0.06240391731262207,
0.01262580044567585,
-0.0003642821975518018,
-0.004792546387761831,
-0.005980513524264097,
-0.035396818071603775,
0.04726344719529152,
-0.03680722415447235,
-4.1083929147589515e-8,
0.002490529790520668,
-0.0028515418525785208,
-0.027167780324816704,
0.07257699966430664,
0.04592220112681389,
0.008274356834590435,
-0.09632662683725357,
0.025958934798836708,
0.044228844344615936,
-0.018575124442577362,
0.06835543364286423,
0.008821116760373116,
-0.0332353338599205,
-0.08742491900920868,
-0.01685856096446514,
0.07112488150596619,
-0.0854133740067482,
0.039362311363220215,
-0.08908556401729584,
-0.06161043792963028,
-0.04436509311199188,
0.1079750806093216,
-0.04659269005060196,
-0.0048169284127652645,
0.07561607658863068,
-0.027129704132676125,
0.01849512755870819,
0.06281349807977676,
0.069569893181324,
0.07787591964006424,
-0.035796862095594406,
-0.05940236523747444,
-0.0037098529282957315,
0.030619684606790543,
-0.019755924120545387,
0.18187230825424194,
-0.057052455842494965,
-0.011486929841339588,
0.010267382487654686,
-0.012232608161866665,
0.05345838889479637,
0.03076089359819889,
0.003871356137096882,
0.036845315247774124,
-0.013980763964354992,
-0.017360946163535118,
-0.006687844172120094,
-0.054828401654958725,
0.028154281899333,
-0.05636852607131004,
0.049226656556129456,
0.004567999858409166,
-0.00616706907749176,
0.11929178982973099,
-0.04364262893795967,
0.02480042725801468,
-0.04532892629504204,
-0.03046257607638836,
0.06327483803033829,
0.038054101169109344,
0.008373220451176167,
-0.0077683464623987675,
0.048070456832647324,
0.00622255215421319
] |
Episodic Novelty Through Temporal Distance | https://openreview.net/forum?id=I7DeajDEx7 | [
"Yuhua Jiang",
"Qihan Liu",
"Yiqin Yang",
"Xiaoteng Ma",
"Dianyu Zhong",
"Hao Hu",
"Jun Yang",
"Bin Liang",
"Bo XU",
"Chongjie Zhang",
"Qianchuan Zhao"
] | Poster | Exploration in sparse reward environments remains a significant challenge in reinforcement learning, particularly in Contextual Markov Decision Processes (CMDPs), where environments differ across episodes. Existing episodic intrinsic motivation methods for CMDPs primarily rely on count-based approaches, which are ineffective in large state spaces, or on similarity-based methods that lack appropriate metrics for state comparison. To address these shortcomings, we propose Episodic Novelty Through Temporal Distance (ETD), a novel approach that introduces temporal distance as a robust metric for state similarity and intrinsic reward computation. By employing contrastive learning, ETD accurately estimates temporal distances and derives intrinsic rewards based on the novelty of states within the current episode. Extensive experiments on various benchmark tasks demonstrate that ETD significantly outperforms state-of-the-art methods, highlighting its effectiveness in enhancing exploration in sparse reward CMDPs. | Reinforcement Learning | null | 1,766 | 2501.15418 | [
-0.07188744843006134,
-0.03696815297007561,
0.020191334187984467,
-0.026038020849227905,
0.04058228060603142,
0.07046947628259659,
-0.02183382213115692,
-0.006532804574817419,
0.04885018989443779,
0.010797109454870224,
-0.04721866920590401,
-0.09630466252565384,
0.051283251494169235,
0.06217319145798683,
0.031183283776044846,
-0.01000981219112873,
0.056745920330286026,
-0.03748396039009094,
0.01501717884093523,
-0.06497472524642944,
0.05138381943106651,
0.017779061570763588,
0.07482081651687622,
0.008625023998320103,
0.05169009789824486,
0.09232701361179352,
0.07063768804073334,
-0.027877869084477425,
0.049320999532938004,
0.003988074604421854,
0.06766238808631897,
-0.010370633564889431,
0.062464263290166855,
-0.10713650286197662,
-0.008420449681580067,
0.0957898423075676,
-0.11636123806238174,
0.03293117508292198,
-0.08632175624370575,
-0.04441118612885475,
0.039212923496961594,
0.04365995153784752,
0.02897188626229763,
0.0319652296602726,
0.01235185656696558,
-0.033618051558732986,
0.011636885814368725,
-0.06303044408559799,
0.04856986552476883,
-0.035235341638326645,
0.040448274463415146,
0.028365230187773705,
0.038455065339803696,
-0.02226020209491253,
-0.005368595477193594,
0.10761770606040955,
0.041776593774557114,
0.036452073603868484,
-0.06069152429699898,
-0.0648157075047493,
0.002394346287474036,
-0.1312338411808014,
-0.0015094251139089465,
-0.032305989414453506,
-0.00581963500007987,
-0.0050551313906908035,
-0.031739842146635056,
0.01775098405778408,
-0.009199634194374084,
-0.033576950430870056,
-0.028363440185785294,
0.07394254207611084,
-0.011676170863211155,
-0.00436614453792572,
0.02113574557006359,
0.047960564494132996,
0.013401160947978497,
-0.0943460538983345,
0.007357291877269745,
-0.0050872680731117725,
-0.05880998820066452,
-0.015542908571660519,
-0.008007948286831379,
-0.029826359823346138,
0.10826431214809418,
-0.007306578569114208,
0.06156498193740845,
-0.0117573831230402,
0.058379530906677246,
-0.00021665099484380335,
0.04694727808237076,
-0.024033067747950554,
-0.0226249098777771,
-0.031816136091947556,
-0.0473116859793663,
0.007563418243080378,
-0.029986685141921043,
-0.005996688734740019,
-0.04376321658492088,
0.08253216743469238,
0.02605481445789337,
0.060982633382081985,
-0.12318912148475647,
0.03724675253033638,
0.052757326513528824,
0.008373369462788105,
-0.0025268548633903265,
0.0036823940463364124,
0.05389770120382309,
-0.10610777884721756,
0.01579725183546543,
-0.07543298602104187,
0.10233070701360703,
0.08337501436471939,
-0.01677786558866501,
0.006862326059490442,
0.07091803103685379,
-0.05142499879002571,
0.009136717766523361,
0.0791609063744545,
0.009399880655109882,
-0.1248805969953537,
0.01858956180512905,
0.025737328454852104,
-0.035661257803440094,
-0.006740887649357319,
0.03908495977520943,
2.6655012150857015e-33,
0.018085811287164688,
-0.00679665757343173,
0.008403037674725056,
-0.004838088061660528,
0.025451676920056343,
0.017734838649630547,
0.03563700616359711,
0.006483030505478382,
-0.02871609479188919,
-0.02784769982099533,
0.0052959611639380455,
0.10967621952295303,
0.009582940489053726,
0.06489171087741852,
0.050510331988334656,
-0.020203175023198128,
-0.05421721190214157,
0.07527501881122589,
0.08726142346858978,
-0.09030891954898834,
0.018039165064692497,
-0.03980303928256035,
-0.015872785821557045,
-0.09421241283416748,
0.018314938992261887,
-0.040809113532304764,
0.010634141974151134,
-0.01439819298684597,
-0.024615440517663956,
-0.02205108292400837,
-0.024976717308163643,
0.06398548185825348,
-0.018932364881038666,
0.012108368799090385,
-0.024867627769708633,
0.07713364064693451,
-0.0023024573456496,
0.00012681672524195164,
0.018711596727371216,
-0.029726067557930946,
-0.05231146886944771,
-0.03777008131146431,
0.017338501289486885,
-0.0415833443403244,
-0.10573612153530121,
-0.024350207298994064,
0.049650244414806366,
-0.06754648685455322,
-0.07129945605993271,
0.016993969678878784,
0.07438058406114578,
0.011487232521176338,
0.02839989773929119,
-0.10328937321901321,
-0.061152201145887375,
-0.0383572056889534,
-0.016142765060067177,
-0.025888795033097267,
-0.010444384068250656,
0.021019916981458664,
0.08312106132507324,
0.020079180598258972,
0.05294228717684746,
0.018988924100995064,
0.003874981077387929,
0.13248498737812042,
0.04106941074132919,
0.045452456921339035,
0.0592726394534111,
0.06052849814295769,
0.0206189826130867,
0.03370368108153343,
0.059638917446136475,
-0.1545981466770172,
0.09287891536951065,
-0.04545264318585396,
0.0870298370718956,
-0.07732591032981873,
-0.022520452737808228,
-0.02782513201236725,
-0.06405974924564362,
-0.09037460386753082,
-0.07132937014102936,
-0.0179393719881773,
0.0036303680390119553,
0.018548035994172096,
-0.025234228000044823,
-0.06471186131238937,
-0.005991713143885136,
-0.03275277838110924,
-0.0003778414102271199,
-0.01782822236418724,
-0.014961112290620804,
0.05376031994819641,
0.028444698080420494,
-3.09163610068702e-33,
-0.009958576411008835,
-0.036627549678087234,
0.035202477127313614,
-0.04113387688994408,
0.07027328014373779,
-0.02110636979341507,
-0.07959620654582977,
0.017464399337768555,
-0.0222941841930151,
-0.02726096846163273,
-0.05966141074895859,
0.019086994230747223,
0.01460231188684702,
0.013689021579921246,
-0.02645217627286911,
0.00564171327278018,
-0.009930331259965897,
0.03879781439900398,
-0.004996138159185648,
0.03711778298020363,
0.09214001148939133,
0.04993518441915512,
-0.18228256702423096,
0.016104359179735184,
0.05187520384788513,
0.04258125647902489,
0.026818640530109406,
0.01586715504527092,
-0.02312115952372551,
-0.04472864419221878,
-0.05616046115756035,
-0.027543071657419205,
-0.07573764771223068,
-0.03648603707551956,
-0.04282093420624733,
0.10306322574615479,
0.08010131120681763,
-0.023871880024671555,
-0.11151442676782608,
0.09466536343097687,
-0.03751237317919731,
-0.031224971637129784,
-0.01879875361919403,
-0.017044341191649437,
-0.008286726661026478,
-0.002631834940984845,
-0.031231127679347992,
0.00430987449362874,
-0.018530219793319702,
0.00028717785608023405,
0.029771242290735245,
0.06257861852645874,
-0.09299644082784653,
0.02998504787683487,
-0.08046334981918335,
-0.026041319593787193,
-0.1147955060005188,
-0.0029685646295547485,
0.0606076754629612,
-0.015320862643420696,
-0.05741381272673607,
0.02683352120220661,
-0.07140912860631943,
0.08741634339094162,
-0.012609384022653103,
0.019228167831897736,
0.018064679577946663,
0.009195717051625252,
-0.047498930245637894,
-0.006362912245094776,
0.015213344246149063,
0.048982974141836166,
0.02173176035284996,
-0.06373666226863861,
0.08035695552825928,
0.04623263329267502,
0.006874537095427513,
-0.044446833431720734,
-0.05871852859854698,
-0.03211362287402153,
-0.012656793929636478,
0.025740351527929306,
0.013275078497827053,
-0.0019584964029490948,
-0.02729984000325203,
0.16406527161598206,
0.037438541650772095,
0.06940776109695435,
0.027596618980169296,
-0.0160111952573061,
-0.051000650972127914,
-0.025920094922184944,
0.031476620584726334,
0.026825647801160812,
-0.0364995002746582,
-3.780827739774395e-8,
-0.07575219124555588,
0.01133898738771677,
0.01959514431655407,
-0.0038556859362870455,
0.026313718408346176,
0.023511087521910667,
-0.04121715947985649,
0.00850155670195818,
-0.04567369073629379,
-0.09446533024311066,
0.0704706609249115,
0.014934827573597431,
0.02717665769159794,
0.013282367028295994,
-0.008149545639753342,
0.034282904118299484,
0.04423252493143082,
-0.022556817159056664,
0.0009572886629030108,
0.047557808458805084,
-0.013117627240717411,
0.03821273148059845,
0.04965287074446678,
-0.0656999722123146,
-0.04948526620864868,
-0.056330595165491104,
-0.006843778770416975,
0.08596950024366379,
0.11728251725435257,
-0.005652574356645346,
0.047523174434900284,
0.01037483662366867,
0.043506771326065063,
0.04198077321052551,
0.044052254408597946,
0.015791840851306915,
0.015187160111963749,
-0.042693011462688446,
-0.11756017059087753,
0.024718521162867546,
-0.06311272829771042,
-0.012439422309398651,
0.010944433510303497,
0.03557133302092552,
-0.04826850816607475,
0.04763546586036682,
-0.04849747195839882,
-0.07167766243219376,
0.004757347982376814,
-0.09027622640132904,
0.02882593311369419,
-0.026401294395327568,
-0.04588216915726662,
0.031241919845342636,
0.1151948794722557,
-0.007316297851502895,
-0.01180677954107523,
-0.06708667427301407,
0.0232009869068861,
0.04817082732915878,
0.04274442046880722,
-0.004663565196096897,
-0.055623557418584824,
-0.0014387492556124926
] |
ViDiT-Q: Efficient and Accurate Quantization of Diffusion Transformers for Image and Video Generation | https://openreview.net/forum?id=E1N1oxd63b | [
"Tianchen Zhao",
"Tongcheng Fang",
"Haofeng Huang",
"Rui Wan",
"Widyadewi Soedarmadji",
"Enshu Liu",
"Shiyao Li",
"Zinan Lin",
"Guohao Dai",
"Shengen Yan",
"Huazhong Yang",
"Xuefei Ning",
"Yu Wang"
] | Poster | Diffusion transformers have demonstrated remarkable performance in visual generation tasks, such as generating realistic images or videos based on textual instructions. However, larger model sizes and multi-frame processing for video generation lead to increased computational and memory costs, posing challenges for practical deployment on edge devices. Post-Training Quantization (PTQ) is an effective method for reducing memory costs and computational complexity.
When quantizing diffusion transformers, we find that existing quantization methods face challenges when applied to text-to-image and video tasks. To address these challenges, we begin by systematically analyzing the source of quantization error and conclude with the unique challenges posed by DiT quantization. Accordingly, we design an improved quantization scheme: ViDiT-Q (**V**ideo \& **I**mage **Di**ffusion **T**ransformer **Q**uantization), tailored specifically for DiT models. We validate the effectiveness of ViDiT-Q across a variety of text-to-image and video models, achieving W8A8 and W4A8 with negligible degradation in visual quality and metrics. Additionally, we implement efficient GPU kernels to achieve practical 2-2.5x memory optimization and a 1.4-1.7x end-to-end latency speedup. | video generation, low-bit quantization, diffusion model | We design the quantization scheme ViDiT-Q tailored for Video & Image Diffusion Transformer Quantization. It achieves W4A8 quantization with negligible performance loss, which brings 2x memory and 1.5x end2end latency speedup. | 1,762 | null | [
-0.03882381319999695,
0.007632690016180277,
0.007074758876115084,
-0.024493465200066566,
0.04270217567682266,
-0.04877590388059616,
-0.05598362907767296,
-0.03467538580298424,
0.07546607404947281,
-0.06762881577014923,
0.034873295575380325,
-0.023126939311623573,
-0.05327913910150528,
0.08248677849769592,
-0.0015345261199399829,
0.012752286158502102,
0.01088227704167366,
0.10961107164621353,
-0.14365266263484955,
-0.05567975714802742,
0.0442192405462265,
-0.043455563485622406,
-0.003861823584884405,
-0.059332262724637985,
0.061659201979637146,
0.09169895201921463,
-0.014190172776579857,
-0.00012558524031192064,
0.05724771320819855,
-0.03345394879579544,
0.051362428814172745,
-0.028828544542193413,
0.00646764412522316,
0.028045665472745895,
-0.05528264120221138,
0.04455801844596863,
-0.040035609155893326,
-0.028092676773667336,
-0.017017222940921783,
-0.059224698692560196,
-0.0042728944681584835,
0.06926147639751434,
0.006307168398052454,
-0.0018619309412315488,
0.07887173444032669,
0.029625210911035538,
-0.02763422392308712,
0.004768263082951307,
0.016503406688570976,
-0.037710145115852356,
0.002389319706708193,
0.03587917983531952,
-0.06695742160081863,
0.09357017278671265,
-0.06438253074884415,
0.014969561249017715,
0.0673794075846672,
-0.008723888546228409,
0.011280958540737629,
-0.0183430016040802,
-0.0683891624212265,
0.007694159634411335,
0.049311596900224686,
0.024163352325558662,
-0.0485110767185688,
-0.023706601932644844,
0.08483441919088364,
-0.019479570910334587,
-0.027815863490104675,
-0.03444775938987732,
-0.09332437068223953,
-0.007823373191058636,
-0.10516687482595444,
0.04401985555887222,
-0.016643602401018143,
0.06680242717266083,
0.042284395545721054,
-0.012957314029335976,
-0.03516203165054321,
-0.031096290796995163,
0.12570713460445404,
-0.03059745393693447,
0.08185891062021255,
0.03124932385981083,
0.06764351576566696,
0.007166191004216671,
-0.07617396861314774,
0.03864915668964386,
-0.03801456466317177,
0.0013585533015429974,
-0.10467136651277542,
0.02084621787071228,
-0.021746914833784103,
-0.0170828178524971,
0.03744411841034889,
-0.01011015847325325,
-0.02456335350871086,
-0.06304381787776947,
-0.02480596862733364,
0.06964579969644547,
0.04957512766122818,
-0.0809173434972763,
-0.04360136017203331,
0.011918474920094013,
0.03679749369621277,
-0.03412775322794914,
0.08295323699712753,
0.07222931832075119,
0.014954783022403717,
-0.024418743327260017,
0.06537172198295593,
0.010097840800881386,
0.0051649524830281734,
-0.05265628919005394,
0.055111829191446304,
0.002781196730211377,
-0.022453803569078445,
0.048531122505664825,
0.029201483353972435,
-0.015308665111660957,
0.002941876882687211,
-0.014844974502921104,
-0.09731791168451309,
0.02449723333120346,
-0.0005046537262387574,
-0.08711201697587967,
-0.04788518324494362,
9.700602744622805e-34,
-0.01993536204099655,
0.014043531380593777,
0.05132893845438957,
0.0814838856458664,
0.0854395404458046,
0.045599035918712616,
0.013103265315294266,
-0.04315871372818947,
-0.04946177080273628,
-0.060541074723005295,
-0.0030328279826790094,
-0.0834694653749466,
-0.08061813563108444,
0.051171720027923584,
0.083433136343956,
-0.058472804725170135,
-0.028522418811917305,
-0.00026137748500332236,
-0.03667186200618744,
0.024403605610132217,
0.10629884153604507,
-0.03681512922048569,
-0.004688517190515995,
0.021807847544550896,
-0.05936812609434128,
0.017882568761706352,
0.0697459951043129,
-0.034244805574417114,
0.04807710647583008,
-0.017783716320991516,
-0.17588382959365845,
-0.06171322613954544,
0.046135470271110535,
-0.07406549900770187,
0.047427233308553696,
0.016271954402327538,
-0.041264086961746216,
-0.039016880095005035,
-0.016605932265520096,
-0.028711000457406044,
-0.036956049501895905,
0.07960834354162216,
0.02103595808148384,
-0.07779382169246674,
-0.08622464537620544,
0.0432807058095932,
0.04059254750609398,
0.045122966170310974,
-0.0407671257853508,
-0.01870149001479149,
0.042299311608076096,
-0.017992811277508736,
-0.026697730645537376,
-0.03774105757474899,
0.03204571083188057,
-0.03387591242790222,
0.1420058012008667,
-0.013970939442515373,
0.01533487904816866,
0.054482027888298035,
0.043759215623140335,
-0.023566434159874916,
-0.07440079003572464,
0.09256158024072647,
0.00018800771795213223,
-0.05940420553088188,
0.0318477489054203,
0.03383598104119301,
-0.06235230714082718,
0.0749896988272667,
-0.0658612996339798,
0.017594724893569946,
0.037450168281793594,
-0.08090413361787796,
0.0934574082493782,
-0.0010837066220119596,
0.054589033126831055,
0.015172114595770836,
-0.028408989310264587,
0.003686085809022188,
-0.04496656358242035,
0.006786706857383251,
0.031825147569179535,
-0.08523347228765488,
0.01613497920334339,
-0.07414694130420685,
-0.037820685654878616,
-0.06458280235528946,
-0.03176155313849449,
-0.032785672694444656,
-0.02801240235567093,
-0.015231864526867867,
-0.013040791265666485,
0.012222995981574059,
-0.027784008532762527,
-9.369528434133554e-34,
-0.06724579632282257,
0.11839637905359268,
0.006918335799127817,
0.1033121645450592,
-0.044680751860141754,
-0.03968995809555054,
0.04255090653896332,
0.029329609125852585,
0.13473688066005707,
-0.12360405176877975,
-0.04127373918890953,
0.01132108736783266,
-0.0439281091094017,
-0.04173314571380615,
-0.03124256059527397,
0.019428489729762077,
0.05157442390918732,
-0.010520920157432556,
-0.09460243582725525,
0.0659133717417717,
-0.024460241198539734,
0.05426345020532608,
0.012127972207963467,
-0.05106700584292412,
-0.0156870037317276,
0.04525190219283104,
0.008362796157598495,
0.08437558263540268,
0.06957395374774933,
-0.05281476303935051,
0.029193777590990067,
-0.04531745985150337,
-0.02528347447514534,
-0.063141368329525,
-0.061195679008960724,
0.0233332347124815,
0.07787475734949112,
0.03646327927708626,
-0.03412124514579773,
0.08132706582546234,
0.05228196457028389,
-0.03794093802571297,
-0.10142244398593903,
0.06934616714715958,
-0.0065055047161877155,
0.023401034995913506,
-0.08923798054456711,
-0.040916889905929565,
0.000927371671423316,
0.06199638172984123,
0.025037622079253197,
0.06819066405296326,
-0.04329590126872063,
0.08988722413778305,
0.001494416268542409,
-0.09601886570453644,
-0.09843456000089645,
0.04976548254489899,
0.028859438374638557,
0.010415636003017426,
-0.004460128489881754,
-0.04345652833580971,
-0.03283538669347763,
-0.09958170354366302,
-0.03845774754881859,
-0.00888141430914402,
-0.05338871106505394,
0.03528818488121033,
0.0027244784869253635,
0.08143199980258942,
0.009757970459759235,
0.020098095759749413,
-0.00326737854629755,
-0.019339801743626595,
0.007444212678819895,
-0.05612654983997345,
-0.010372554883360863,
-0.007829408161342144,
0.05267654359340668,
0.015424199402332306,
-0.0024270217400044203,
0.031181514263153076,
0.021411107853055,
-0.011251729913055897,
0.08075570315122604,
0.05733450502157211,
0.025365129113197327,
0.032753244042396545,
0.008911934681236744,
0.07162415981292725,
-0.06527761369943619,
0.04142538458108902,
0.07194431126117706,
0.029766181483864784,
-0.05430639535188675,
-3.9439338905822297e-8,
-0.06159080192446709,
0.017799532040953636,
-0.06317205727100372,
-0.0555843785405159,
-0.007288709282875061,
-0.06592247635126114,
0.0587112195789814,
0.04259590432047844,
0.018248334527015686,
-0.0592639334499836,
0.07962863892316818,
-0.024336354807019234,
-0.11389710009098053,
-0.021829763427376747,
-0.04891657084226608,
0.10419958829879761,
0.027265308424830437,
0.01167621836066246,
0.051844820380210876,
0.01866867206990719,
-0.06556473672389984,
0.08377120643854141,
-0.01748635433614254,
-0.028072932735085487,
-0.0021485406905412674,
-0.0004063615342602134,
0.01060978602617979,
0.005687826778739691,
0.10501033067703247,
-0.015348177403211594,
0.01286622229963541,
-0.011073330417275429,
0.03079831786453724,
-0.00894775427877903,
-0.042197518050670624,
-0.007671143393963575,
-0.027713492512702942,
0.0027359803207218647,
-0.017241915687918663,
-0.033634185791015625,
0.012769775465130806,
0.023074442520737648,
-0.012147476896643639,
0.026615357026457787,
-0.01005612500011921,
0.03221844136714935,
0.052134279161691666,
-0.02019328624010086,
0.004731568042188883,
0.0033676261082291603,
0.0326043963432312,
0.02904440090060234,
-0.05114224553108215,
0.01077241450548172,
0.05399215221405029,
-0.03614608570933342,
0.12047161906957626,
-0.021815398707985878,
0.034754760563373566,
0.037684593349695206,
-0.004133503418415785,
-0.024611152708530426,
0.05218581482768059,
0.022426679730415344
] |
Do You Keep an Eye on What I Ask? Mitigating Multimodal Hallucination via Attention-Guided Ensemble Decoding | https://openreview.net/forum?id=ziw5bzg2NO | [
"Yeongjae Cho",
"Keonwoo Kim",
"Taebaek Hwang",
"Sungzoon Cho"
] | Poster | Recent advancements in Large Vision-Language Models (LVLMs) have significantly expanded their utility in tasks like image captioning and visual question answering. However, they still struggle with object hallucination, where models generate descriptions that inaccurately reflect the visual content by including nonexistent objects or misrepresenting existing ones. While previous methods, such as data augmentation and training-free approaches, strive to tackle this issue, they still encounter scalability challenges and often depend on additional external modules. In this work, we propose Ensemble Decoding (ED), a novel strategy that splits the input image into sub-images and combines logit distributions by assigning weights through the attention map. Furthermore, we introduce ED adaptive plausibility constraint to calibrate logit distribution and FastED, a variant designed for speed-critical applications. Extensive experiments across hallucination benchmarks demonstrate that our proposed method achieves state-of-the-art performance, validating the effectiveness of our approach. | Hallucination, Multimodal Hallucination, Large Vision-Language Model | We introduce Ensemble Decoding (ED), a method designed to mitigate object hallucination in Large Vision-Language Models by dividing an image into sub-images and combining logit distributions with attention-guided weights to improve accuracy. | 1,757 | null | [
-0.00019238398817833513,
-0.061102718114852905,
0.02825276181101799,
0.013443026691675186,
0.12133622914552689,
0.007864234037697315,
-0.06071104481816292,
-0.013852824456989765,
-0.00042752345325425267,
-0.04866820573806763,
-0.04601498320698738,
-0.09230752289295197,
0.057644300162792206,
0.023409567773342133,
0.005476055201143026,
-0.003325394121930003,
0.0801718533039093,
0.13220463693141937,
-0.10870857536792755,
-0.01875976473093033,
0.023159554228186607,
-0.05919823423027992,
0.07404012978076935,
0.008288074284791946,
0.094236359000206,
0.012249748222529888,
-0.017806079238653183,
-0.06448931992053986,
-0.052906978875398636,
-0.07676666229963303,
0.0356098972260952,
0.010908213444054127,
-0.07307697087526321,
0.03203453868627548,
-0.038679979741573334,
0.07784657180309296,
-0.09034689515829086,
0.005102983210235834,
0.02362457476556301,
-0.07062417268753052,
-0.054214704781770706,
0.03641640022397041,
0.007693178951740265,
0.002217161236330867,
0.08250339329242706,
0.010629087686538696,
0.011483345180749893,
-0.03691140189766884,
0.022606397047638893,
-0.046744830906391144,
0.00635778671130538,
0.01614326238632202,
0.016803078353405,
0.028308743610978127,
-0.07465347647666931,
0.061030421406030655,
0.031539008021354675,
-0.04349857568740845,
0.010294103063642979,
-0.0022940905764698982,
-0.022636253386735916,
-0.018317244946956635,
0.05475538969039917,
-0.0037827137857675552,
-0.00671106344088912,
-0.056800778955221176,
0.0584585927426815,
-0.0065010227262973785,
0.006081979256123304,
-0.003466642228886485,
0.002354573691263795,
0.03568192943930626,
0.01444385014474392,
0.0074735055677592754,
0.09391070902347565,
0.011553673073649406,
0.02137887477874756,
-0.04906221851706505,
0.1140330582857132,
-0.11054349690675735,
0.005959630478173494,
0.017222810536623,
0.0037663355469703674,
0.020446978509426117,
0.14612525701522827,
-0.01741141639649868,
-0.01350380014628172,
0.08227758854627609,
-0.024696841835975647,
0.006430237554013729,
-0.09811131656169891,
-0.11456981301307678,
0.06077752262353897,
-0.015700584277510643,
0.04289834201335907,
0.02875153161585331,
-0.00942996609956026,
-0.06953104585409164,
-0.04767749086022377,
0.06907099485397339,
-0.02476448006927967,
0.0032213591039180756,
0.02870705910027027,
-0.06899552792310715,
0.01769294962286949,
-0.061912935227155685,
0.1150616705417633,
0.06553675979375839,
-0.010995911434292793,
-0.03909463435411453,
0.04882843792438507,
0.029100684449076653,
-0.007200884632766247,
0.01824314147233963,
0.017054134979844093,
0.026592066511511803,
0.01780523732304573,
0.004914871416985989,
0.0044835349544882774,
0.0632510706782341,
-0.03659787401556969,
-0.0037032999098300934,
-0.0345553494989872,
-0.01276760920882225,
0.10424969345331192,
-0.0644240453839302,
-0.009918228723108768,
2.054972959155622e-33,
0.04720465466380119,
0.024891337379813194,
0.06546295434236526,
0.038709741085767746,
0.1212025061249733,
-0.02117070183157921,
-0.04599134996533394,
0.011546570807695389,
-0.016090262681245804,
-0.0032141925767064095,
-0.05804358422756195,
0.00346046919003129,
-0.06351690739393234,
0.10608970373868942,
0.012251312844455242,
0.028332997113466263,
-0.08401398360729218,
0.04743005335330963,
-0.05900062620639801,
0.009103902615606785,
0.010158644057810307,
-0.06482845544815063,
-0.02596529759466648,
-0.020359322428703308,
0.004385716747492552,
-0.04493263363838196,
0.055183399468660355,
-0.09840461611747742,
-0.029935620725154877,
0.019503269344568253,
-0.12854856252670288,
-0.013882460072636604,
-0.0227682963013649,
0.025567688047885895,
0.06347929686307907,
-0.07237996906042099,
-0.022558342665433884,
-0.0261390283703804,
-0.0363890640437603,
-0.007205393631011248,
0.011433073319494724,
0.06088930740952492,
-0.05398314818739891,
-0.01232889574021101,
-0.08676153421401978,
0.003856632625684142,
0.0035269802901893854,
0.06491375714540482,
-0.07138670980930328,
-0.03444922715425491,
0.06402416527271271,
0.03792619705200195,
-0.06082737445831299,
-0.05854298174381256,
-0.032466188073158264,
-0.07515265792608261,
0.06537042558193207,
0.021062934771180153,
0.06453633308410645,
-0.03309503197669983,
0.036551348865032196,
0.045232143253088,
-0.014637800864875317,
0.1418079435825348,
0.03291857987642288,
0.012799695134162903,
0.014586033299565315,
0.021601855754852295,
0.03303072229027748,
-0.05536118894815445,
-0.059181343764066696,
-0.0015117275761440396,
0.0030455528758466244,
0.014009211212396622,
0.03840496763586998,
0.0360538475215435,
0.059264931827783585,
-0.04094281047582626,
0.037748925387859344,
0.040579233318567276,
0.014661147259175777,
0.0001976355561055243,
-0.0038851231802254915,
-0.054591331630945206,
-0.01613333821296692,
-0.027781128883361816,
0.03174836188554764,
-0.08244725316762924,
-0.029423274099826813,
-0.052347250282764435,
-0.043504748493433,
-0.034341294318437576,
-0.04531049355864525,
-0.06803973019123077,
-0.0007359938463196158,
-6.754618751032817e-34,
0.02792615443468094,
0.01922312192618847,
0.04578801617026329,
-0.003911684732884169,
-0.008173177018761635,
-0.019087087363004684,
-0.01606179028749466,
0.03172709792852402,
0.0458657369017601,
-0.04679577797651291,
-0.04798895865678787,
-0.022823240607976913,
0.03496778383851051,
0.0013008747482672334,
0.03563844785094261,
-0.03255030885338783,
0.03933801129460335,
0.039884280413389206,
-0.04507826641201973,
0.018370963633060455,
-0.02223179116845131,
0.12298522889614105,
-0.031403928995132446,
-0.03644106164574623,
-0.03767562285065651,
0.0928468257188797,
-0.04569680616259575,
0.05657721683382988,
0.016892043873667717,
-0.055915866047143936,
0.00007040733908070251,
0.03623054549098015,
0.011781318113207817,
-0.04948611184954643,
-0.038427144289016724,
0.0541820265352726,
0.0268918015062809,
-0.027944443747401237,
-0.07702333480119705,
0.08981459587812424,
0.05446069687604904,
-0.07846367359161377,
-0.06021151691675186,
0.03876467049121857,
-0.05249718204140663,
-0.11717371642589569,
-0.049253832548856735,
-0.07772321999073029,
0.00704495282843709,
0.1295854151248932,
-0.029073307290673256,
-0.011625109240412712,
-0.09744741022586823,
0.025719737634062767,
-0.03880754113197327,
-0.033975835889577866,
-0.12101374566555023,
0.047210585325956345,
0.10260345041751862,
-0.010481120087206364,
-0.027253126725554466,
0.03715895488858223,
-0.01208152249455452,
0.00583897577598691,
-0.0036831216420978308,
-0.06261038780212402,
-0.0031775985844433308,
-0.049701739102602005,
0.057179078459739685,
0.027325181290507317,
0.07023902982473373,
-0.05331802740693092,
0.047545336186885834,
0.11067003011703491,
0.021279219537973404,
-0.00039015847141854465,
0.06451914459466934,
0.014874046668410301,
0.01777001842856407,
-0.10357411205768585,
-0.008181905373930931,
-0.06284446269273758,
0.032126836478710175,
0.10458225011825562,
0.10221979767084122,
0.006236399989575148,
0.03908756747841835,
0.03027004562318325,
0.07000528275966644,
-0.004808046389371157,
-0.018579546362161636,
-0.01027070451527834,
-0.0025776727125048637,
0.10896307229995728,
0.003256123047322035,
-4.387129948213442e-8,
-0.045649152249097824,
0.09108151495456696,
-0.08019638061523438,
0.0183370690792799,
0.039405617862939835,
-0.06052080914378166,
-0.07932932674884796,
0.07809959352016449,
0.014160008169710636,
-0.05364713817834854,
0.10586486756801605,
0.020312197506427765,
0.011419499292969704,
-0.08093299716711044,
-0.022953571751713753,
0.08926840871572495,
0.020857350900769234,
0.0421634167432785,
-0.024930158630013466,
-0.014658293686807156,
0.03359556198120117,
-0.022656256332993507,
0.029895741492509842,
-0.016388313844799995,
-0.02765577659010887,
0.00891026109457016,
-0.023030636832118034,
0.01582537777721882,
0.04322991892695427,
-0.0715814009308815,
0.028046784922480583,
0.04844171926379204,
0.032982680946588516,
-0.03189043328166008,
-0.00004274034290574491,
0.08319791406393051,
-0.018990175798535347,
0.022094042971730232,
-0.021916251629590988,
-0.02247009612619877,
0.06562940776348114,
-0.006770541425794363,
-0.00806045439094305,
0.021901775151491165,
0.012463097460567951,
-0.04389293119311333,
-0.05086667090654373,
-0.08360328525304794,
0.016602132469415665,
-0.048818059265613556,
-0.011927664279937744,
0.023848794400691986,
0.014428945258259773,
0.10753937810659409,
0.026464873924851418,
0.03047259710729122,
0.0889539048075676,
-0.06117691099643707,
0.05931811407208443,
0.043489016592502594,
-0.006909623276442289,
-0.033727485686540604,
-0.04231422021985054,
-0.006708073895424604
] |
Masked Temporal Interpolation Diffusion for Procedure Planning in Instructional Videos | https://openreview.net/forum?id=HnpDHiItd2 | [
"Yufan Zhou",
"Zhaobo Qi",
"Lingshuai Lin",
"Junqi Jing",
"Tingting Chai",
"Beichen Zhang",
"Shuhui Wang",
"Weigang Zhang"
] | Poster | In this paper, we address the challenge of procedure planning in instructional videos, aiming to generate coherent and task-aligned action sequences from start and end visual observations. Previous work has mainly relied on text-level supervision to bridge the gap between observed states and unobserved actions, but it struggles with capturing intricate temporal relationships among actions. Building on these efforts, we propose the Masked Temporal Interpolation Diffusion (MTID) model that introduces a latent space temporal interpolation module within the diffusion model. This module leverages a learnable interpolation matrix to generate intermediate latent features, thereby augmenting visual supervision with richer mid-state details. By integrating this enriched supervision into the model, we enable end-to-end training tailored to task-specific requirements, significantly enhancing the model's capacity to predict temporally coherent action sequences. Additionally, we introduce an action-aware mask projection mechanism to restrict the action generation space, combined with a task-adaptive masked proximity loss to prioritize more accurate reasoning results close to the given start and end states over those in intermediate steps. Simultaneously, it filters out task-irrelevant action predictions, leading to contextually aware action sequences. Experimental results across three widely used benchmark datasets demonstrate that our MTID achieves promising action planning performance on most metrics. | procedure planning, diffusion, U-Net, temporal logic interpolation, action prediction, mask | A masked temporal interpolation predictor for procedure planning in instructional videos | 1,751 | null | [
-0.028061024844646454,
-0.09136952459812164,
0.031792882829904556,
0.020496193319559097,
0.07530582696199417,
0.036982979625463486,
-0.06962763518095016,
-0.02963075414299965,
0.06677708029747009,
-0.04984418302774429,
-0.051360804587602615,
-0.033255770802497864,
-0.026569463312625885,
0.14053623378276825,
-0.040952168405056,
-0.004864427261054516,
0.025972995907068253,
0.012668387033045292,
-0.05942594259977341,
-0.05478888750076294,
0.09443262964487076,
0.0064738173969089985,
0.05789320915937424,
-0.005776236765086651,
0.0033482476137578487,
0.07672326266765594,
0.03640545532107353,
-0.069821797311306,
0.04951874166727066,
0.004621091298758984,
0.052406467497348785,
-0.08214205503463745,
-0.019968045875430107,
-0.025777414441108704,
-0.09308843314647675,
0.0785466879606247,
-0.06568096578121185,
0.06506580859422684,
-0.03420981019735336,
-0.05281728133559227,
0.01776226796209812,
0.027122821658849716,
0.047465283423662186,
0.076653391122818,
0.04743099585175514,
-0.02674979530274868,
-0.02788509614765644,
-0.06916831433773041,
-0.019593307748436928,
0.026668697595596313,
-0.06292608380317688,
-0.015524609945714474,
0.034860819578170776,
0.0028141988441348076,
-0.07702097296714783,
0.00717956805601716,
0.05211929976940155,
-0.001125057926401496,
0.04368339106440544,
-0.033770836889743805,
-0.014219770208001137,
-0.09401310235261917,
0.021953174844384193,
-0.06640484184026718,
-0.05295656993985176,
-0.06467562168836594,
0.06302359700202942,
0.02161085419356823,
-0.04552539065480232,
0.0599215030670166,
-0.04436817392706871,
0.027391603216528893,
-0.06321236491203308,
0.021128112450242043,
-0.018344981595873833,
0.057160135358572006,
0.006247785408049822,
-0.11564437299966812,
-0.010464553721249104,
-0.025541134178638458,
0.0652601346373558,
-0.008853179402649403,
0.030469119548797607,
0.015683209523558617,
0.10222534090280533,
0.0060762823559343815,
0.044466592371463776,
0.07554524391889572,
-0.031128110364079475,
-0.028119834139943123,
-0.04156169295310974,
-0.02351127751171589,
-0.003798720659688115,
0.0014568474143743515,
0.014418377541005611,
0.03442274034023285,
-0.01145266741514206,
-0.08147032558917999,
0.012495643459260464,
0.05152595043182373,
-0.004161531571298838,
0.06369257718324661,
-0.06714990735054016,
-0.009042512625455856,
-0.011217274703085423,
-0.04427134990692139,
0.06376340985298157,
0.024958061054348946,
-0.012038064189255238,
-0.06437195837497711,
0.04921920225024223,
-0.031068967655301094,
0.03861960396170616,
0.034953318536281586,
0.06787223368883133,
-0.007813933305442333,
0.01957557536661625,
0.02216915600001812,
0.023762693628668785,
0.048155784606933594,
0.05151059851050377,
-0.009193439036607742,
-0.02550036460161209,
-0.010131328366696835,
0.034210361540317535,
-0.046159178018569946,
-0.0035003856755793095,
-2.4523374368495725e-33,
0.07732018083333969,
-0.04599792882800102,
0.07493069022893906,
0.05572479963302612,
0.06510318070650101,
0.032169606536626816,
-0.009366246871650219,
0.03554028272628784,
-0.04885770007967949,
-0.03958994150161743,
0.010290197096765041,
-0.0756860226392746,
-0.07062558829784393,
0.09476285427808762,
-0.012262231670320034,
-0.01866007409989834,
-0.031001994386315346,
0.050525784492492676,
-0.012042304500937462,
0.05057269334793091,
-0.006519598886370659,
-0.021535927429795265,
-0.023747656494379044,
-0.06270132213830948,
-0.0006542436312884092,
-0.007695457898080349,
0.08895617723464966,
-0.07069000601768494,
0.0187432412058115,
-0.02934950590133667,
-0.12760332226753235,
-0.0050447117537260056,
-0.04291950911283493,
-0.0007371121901087463,
0.051402702927589417,
0.0075090802274644375,
0.011171369813382626,
-0.08001109957695007,
0.09621226042509079,
-0.02937990613281727,
0.03306913375854492,
0.03490929678082466,
0.005923505872488022,
0.001637563924305141,
-0.1001351848244667,
-0.06737813353538513,
0.028569137677550316,
0.07363764196634293,
-0.010606183670461178,
-0.0008054350037127733,
0.03105742111802101,
0.05567115172743797,
0.026143204420804977,
-0.17212313413619995,
-0.02572326548397541,
-0.009914319962263107,
0.08108928799629211,
-0.012404559180140495,
0.020794665440917015,
-0.008210303261876106,
0.022070175036787987,
0.036801718175411224,
-0.043831244111061096,
0.11454277485609055,
-0.0023322340566664934,
-0.012648259289562702,
-0.013344959355890751,
0.07690755277872086,
0.11527682095766068,
-0.035148799419403076,
-0.07021628320217133,
0.024051165208220482,
0.002846191171556711,
-0.0854712724685669,
0.04705316945910454,
0.012547468766570091,
0.04993409663438797,
-0.09487633407115936,
0.0034037858713418245,
0.06330061703920364,
0.023484239354729652,
-0.07996269315481186,
0.0018912854138761759,
-0.07522599399089813,
-0.02098969556391239,
-0.04327774420380592,
0.001845411490648985,
-0.04990541934967041,
0.01138367224484682,
-0.06007072329521179,
-0.023023469373583794,
-0.02881336584687233,
-0.0394866019487381,
0.07555167376995087,
0.0655488595366478,
2.0243713514548637e-33,
0.08421868830919266,
0.032096978276968,
-0.02042010799050331,
0.002154479967430234,
0.060415614396333694,
-0.028652772307395935,
0.03044731542468071,
-0.029067939147353172,
0.06786391884088516,
-0.06562136113643646,
-0.027523940429091454,
-0.06086132302880287,
-0.05044904723763466,
-0.02909746766090393,
0.03028821013867855,
0.006609782110899687,
0.030841177329421043,
0.039530783891677856,
0.01568443514406681,
-0.022252948954701424,
-0.02551150880753994,
0.032548364251852036,
-0.045442111790180206,
0.017520617693662643,
-0.05204203724861145,
0.0026803920045495033,
0.07020246982574463,
0.04436411336064339,
0.001001570955850184,
0.016541538760066032,
0.017120907083153725,
-0.06667677313089371,
0.026171399280428886,
-0.01039030123502016,
-0.12260272353887558,
0.05595485866069794,
0.018157148733735085,
-0.06265763938426971,
-0.06557706743478775,
0.07375074923038483,
0.0697210505604744,
-0.05730795860290527,
-0.019608914852142334,
0.056163109838962555,
-0.04065287858247757,
-0.023581434041261673,
-0.067233607172966,
0.0005317665636539459,
-0.02602440118789673,
0.035741399973630905,
-0.08604084700345993,
0.07916218042373657,
-0.12348584830760956,
-0.04996319115161896,
-0.008873815648257732,
-0.05532645434141159,
-0.05734412744641304,
0.030707985162734985,
0.07114469259977341,
-0.03664489462971687,
0.021387340500950813,
0.08048241585493088,
0.0114542031660676,
-0.08805807679891586,
0.07497500628232956,
0.017112310975790024,
-0.04660165682435036,
-0.025716368108987808,
0.04576392099261284,
-0.0022782981395721436,
0.030834421515464783,
-0.019013691693544388,
0.0016798656433820724,
-0.011801304295659065,
0.13296042382717133,
-0.03934997320175171,
0.013515972532331944,
-0.024261150509119034,
-0.01705356501042843,
-0.026340950280427933,
0.0057146986946463585,
-0.03870874270796776,
-0.030218617990612984,
0.04217715561389923,
0.012300354428589344,
0.12588679790496826,
0.016415076330304146,
0.1004047691822052,
-0.0067811752669513226,
0.03359946236014366,
-0.0512160062789917,
-0.0011391047155484557,
0.004423703998327255,
0.024654805660247803,
-0.05475616082549095,
-3.2367999835969385e-8,
-0.10891228914260864,
0.05671023204922676,
-0.006792124360799789,
-0.0003550792462192476,
0.010144021362066269,
-0.08550947159528732,
0.048124413937330246,
0.051584526896476746,
0.06334396451711655,
0.005338949151337147,
0.07904714345932007,
-0.03066099062561989,
0.03629818558692932,
-0.07623787224292755,
-0.007970881648361683,
0.04174692556262016,
0.036022547632455826,
0.02970394305884838,
0.014112046919763088,
0.029847990721464157,
-0.008480974473059177,
0.02106456644833088,
-0.02726757898926735,
-0.05546477064490318,
-0.024610942229628563,
-0.009038407355546951,
-0.008482039906084538,
0.1351800262928009,
0.10021115094423294,
0.04194585233926773,
0.04781801253557205,
-0.056703608483076096,
0.033700987696647644,
-0.04541732743382454,
-0.03442441672086716,
0.055898379534482956,
0.07910526543855667,
-0.03861512616276741,
0.03503139689564705,
-0.04610500857234001,
-0.017084302380681038,
0.023362120613455772,
0.06005866825580597,
0.034426845610141754,
-0.040912121534347534,
0.03928441181778908,
-0.07572018355131149,
-0.09873485565185547,
0.04418565705418587,
-0.037983786314725876,
-0.023266460746526718,
-0.061480045318603516,
-0.04246411472558975,
0.09544315189123154,
0.103945292532444,
0.05968845635652542,
0.06508833169937134,
-0.050858769565820694,
0.008887241594493389,
0.04056548327207565,
-0.01505985576659441,
0.043875738978385925,
0.016604607924818993,
0.06032445654273033
] |
BlueSuffix: Reinforced Blue Teaming for Vision-Language Models Against Jailbreak Attacks | https://openreview.net/forum?id=wwVGZRnAYG | [
"Yunhan Zhao",
"Xiang Zheng",
"Lin Luo",
"Yige Li",
"Xingjun Ma",
"Yu-Gang Jiang"
] | Poster | In this paper, we focus on black-box defense for VLMs against jailbreak attacks.
Existing black-box defense methods are either unimodal or bimodal. Unimodal methods enhance either the vision or language module of the VLM, while bimodal methods robustify the model through text-image representation realignment.
However, these methods suffer from two limitations: 1) they fail to fully exploit the cross-modal information, or 2) they degrade the model performance on benign inputs.
To address these limitations, we propose a novel blue-team method BlueSuffix that defends target VLMs against jailbreak attacks without compromising its performance under black-box setting. BlueSuffix includes three key components: 1) a visual purifier against jailbreak images, 2) a textual purifier against jailbreak texts, and 3) a blue-team suffix generator using reinforcement fine-tuning for enhancing cross-modal robustness. We empirically show on four VLMs (LLaVA, MiniGPT-4, InstructionBLIP, and Gemini) and four safety benchmarks (Harmful Instruction, AdvBench, MM-SafetyBench, and RedTeam-2K) that BlueSuffix outperforms the baseline defenses by a significant margin.
Our BlueSuffix opens up a promising direction for defending VLMs against jailbreak attacks. Code is available at https://github.com/Vinsonzyh/BlueSuffix. | Jailbreak Defense, Blue-Teaming, Large Vision-Language Model | In this work, we focus on black-box defense against VLM jailbreaks and propose a novel defense framework dubbed BlueSuffix to achieve multimodal robustness by training a blue-team suffix generator using reinforcement learning (RL). | 1,748 | 2410.20971 | [
-0.03606405854225159,
-0.03674207627773285,
-0.007725014816969633,
-0.01904694177210331,
0.06268354505300522,
0.07158603519201279,
0.004237501882016659,
-0.0322900265455246,
0.01062870491296053,
-0.044621679931879044,
0.027060307562351227,
-0.046408120542764664,
0.050142738968133926,
-0.025622397661209106,
0.03310937061905861,
0.05816425383090973,
0.07577697187662125,
0.09934793412685394,
-0.11214902251958847,
0.0283846203237772,
0.049926698207855225,
-0.012216858565807343,
-0.025835510343313217,
-0.016822071745991707,
0.015993446111679077,
0.09434280544519424,
-0.10962933301925659,
-0.04985315352678299,
0.03564171493053436,
-0.08086532354354858,
-0.051960114389657974,
-0.01600591093301773,
0.03471143916249275,
0.043632157146930695,
-0.046128857880830765,
-0.028225375339388847,
-0.05083472654223442,
0.04952094703912735,
-0.016817927360534668,
-0.09247094392776489,
-0.12857945263385773,
0.08346302062273026,
-0.04310249909758568,
0.017931949347257614,
0.015973202884197235,
-0.04639329016208649,
0.030401332303881645,
0.009983142837882042,
0.010808517225086689,
-0.049667954444885254,
-0.03557094186544418,
0.00660948408767581,
0.0654265284538269,
0.059295956045389175,
-0.0320134200155735,
0.008948705159127712,
0.007969196885824203,
0.015749450773000717,
0.09711822867393494,
0.0606113076210022,
-0.010103782638907433,
0.06383366137742996,
-0.013605191372334957,
0.018432630226016045,
0.023167235776782036,
0.009420681744813919,
0.07433310151100159,
0.0012310094898566604,
-0.004989026114344597,
0.0036967184860259295,
-0.0450405515730381,
0.0017903411062434316,
-0.001724093221127987,
-0.022339796647429466,
-0.00899223517626524,
0.11416417360305786,
0.025922494009137154,
-0.041157424449920654,
-0.0023029777221381664,
-0.14616362750530243,
0.06938497722148895,
-0.06790236383676529,
0.03993622213602066,
0.053192924708127975,
0.08910653740167618,
0.08249576389789581,
-0.04265918955206871,
0.0444890595972538,
-0.00044148662709631026,
0.03893473371863365,
-0.051163941621780396,
0.07233434915542603,
0.07215991616249084,
-0.05428376421332359,
0.039202503859996796,
-0.020876340568065643,
-0.023884860798716545,
-0.0974370613694191,
-0.02351894974708557,
0.08056167513132095,
0.018592555075883865,
-0.018404003232717514,
0.0532773993909359,
-0.05357217788696289,
0.08587680757045746,
-0.029172906652092934,
0.10218165069818497,
-0.024260742589831352,
0.02405429631471634,
-0.025250254198908806,
0.06052040681242943,
0.03003844805061817,
0.008393554948270321,
0.024474887177348137,
-0.0028126451652497053,
-0.007959610782563686,
-0.034154586493968964,
0.04857638105750084,
0.04930618405342102,
0.006293568294495344,
0.07580442726612091,
-0.001187482732348144,
-0.013793033547699451,
-0.0468968041241169,
0.00319266808219254,
-0.005019020289182663,
-0.09020145982503891,
2.692847989131628e-33,
0.06320352852344513,
0.04864245280623436,
-0.0640115812420845,
0.018772834911942482,
0.07513761520385742,
-0.08433610945940018,
-0.01266409270465374,
0.04058651998639107,
-0.079982690513134,
-0.018940167501568794,
-0.025517122820019722,
-0.0275834109634161,
0.0009187579853460193,
0.1041807159781456,
0.13167785108089447,
-0.05614331364631653,
-0.08299865573644638,
-0.033265370875597,
-0.03125157952308655,
0.039134398102760315,
-0.017887938767671585,
-0.06773622334003448,
-0.028961703181266785,
-0.0250957403331995,
0.0011633803369477391,
0.01524911168962717,
0.07624915987253189,
-0.025083720684051514,
0.0005407971912063658,
0.03327948600053787,
-0.11259479075670242,
0.026695189997553825,
0.023882197216153145,
0.011805785819888115,
0.06403878331184387,
-0.08036738634109497,
0.012147900648415089,
-0.060514651238918304,
-0.08192869275808334,
0.008124245330691338,
-0.019909949973225594,
-0.00042316538747400045,
-0.14738506078720093,
-0.07048064470291138,
0.06618531793355942,
-0.023858118802309036,
-0.020952362567186356,
0.05435111001133919,
-0.06939002871513367,
-0.04887132719159126,
-0.04283206909894943,
0.04560504108667374,
-0.016733411699533463,
-0.0629105195403099,
-0.05207882076501846,
-0.018221620470285416,
0.02077493444085121,
0.05467035993933678,
0.042656779289245605,
0.11537071317434311,
-0.06044120714068413,
0.08268597722053528,
-0.026998406276106834,
0.05818162485957146,
-0.027550673112273216,
-0.020489109680056572,
-0.02995634451508522,
-0.014566022902727127,
-0.03702370449900627,
-0.05345385894179344,
-0.0919937789440155,
0.018535055220127106,
0.016598133370280266,
0.024905048310756683,
-0.05646701902151108,
-0.024370752274990082,
0.02539806067943573,
0.023136364296078682,
-0.011506513692438602,
-0.052916985005140305,
0.0013792829122394323,
-0.012303242459893227,
0.06743163615465164,
-0.023704687133431435,
-0.023289348930120468,
-0.07491324096918106,
0.0077226422727108,
-0.0807342529296875,
-0.07024533301591873,
-0.013211363926529884,
-0.011406647972762585,
-0.06730187684297562,
-0.04023674875497818,
-0.0004141349927522242,
-0.04301268607378006,
-2.588700108003735e-33,
-0.022663429379463196,
-0.03284040465950966,
0.007900667376816273,
0.07179605215787888,
-0.01316024549305439,
-0.04870562255382538,
0.04550889506936073,
0.029938505962491035,
0.03226279094815254,
-0.05193690210580826,
-0.049358975142240524,
0.07863765209913254,
0.01579633168876171,
0.0277248602360487,
0.05289410427212715,
-0.05563497170805931,
0.010007827542722225,
-0.00344100478105247,
-0.05992349982261658,
0.04306314140558243,
0.02506907843053341,
0.0986846461892128,
-0.008085091598331928,
0.04064948111772537,
-0.0014176196418702602,
0.04377231001853943,
-0.06257480382919312,
0.051913224160671234,
0.028384989127516747,
-0.003262650454416871,
0.06897349655628204,
0.05101299285888672,
0.0048418203368783,
-0.073404960334301,
-0.00045071361819282174,
0.09364589303731918,
0.0479310005903244,
-0.041282910853624344,
-0.020114969462156296,
0.08293457329273224,
0.08680453896522522,
-0.0030197277665138245,
-0.09889651089906693,
-0.07704520225524902,
0.02769373171031475,
-0.01985028013586998,
-0.08125180751085281,
-0.030195433646440506,
-0.029168998822569847,
-0.04107380658388138,
0.034579407423734665,
0.006879068445414305,
-0.02760470286011696,
0.01646394655108452,
-0.0465676486492157,
-0.1308557093143463,
0.020122135058045387,
-0.002374098403379321,
0.06111988425254822,
0.03394161909818649,
0.03022109530866146,
-0.02127946726977825,
-0.03467801585793495,
-0.04775559529662132,
0.03407122567296028,
-0.0586957111954689,
-0.018901435658335686,
0.00007393223495455459,
0.04019453376531601,
-0.034179676324129105,
0.030633367598056793,
-0.07507491111755371,
-0.025383837521076202,
0.10504045337438583,
-0.010053150355815887,
-0.0007240630220621824,
-0.03619521111249924,
-0.07135819643735886,
-0.04779210686683655,
0.04670478031039238,
0.028471680358052254,
0.016754543408751488,
0.04487938433885574,
0.08654586225748062,
0.0385611467063427,
0.05673736333847046,
0.015336282551288605,
0.0515604242682457,
0.03783608600497246,
-0.04226551577448845,
-0.0597064271569252,
0.016611618921160698,
0.003113378770649433,
0.1670503169298172,
-0.027815459296107292,
-4.1030883579651345e-8,
0.036851733922958374,
0.009860159829258919,
0.009801886044442654,
0.01500787679105997,
0.012603919953107834,
0.028365176171064377,
-0.04804801195859909,
-0.08010253310203552,
0.07039625197649002,
-0.041052866727113724,
0.03830419108271599,
-0.07359566539525986,
-0.0523093044757843,
-0.03451443091034889,
-0.030390899628400803,
0.1005704402923584,
-0.06482662260532379,
-0.00004088273271918297,
-0.014691854827105999,
-0.03389899432659149,
-0.04415289685130119,
0.015315614640712738,
0.009389408864080906,
-0.005911393091082573,
-0.0028783113230019808,
-0.06620637327432632,
-0.03659097105264664,
0.017113331705331802,
0.08220025897026062,
0.03780484199523926,
0.006561639253050089,
0.058797359466552734,
0.024862714111804962,
-0.030494442209601402,
-0.01305415853857994,
0.07551247626543045,
-0.02298974245786667,
0.019228586927056313,
0.06619236618280411,
0.06238465756177902,
0.007666776422411203,
0.006943661719560623,
-0.049278952181339264,
-0.01645592413842678,
0.009964203462004662,
-0.05127643793821335,
-0.04454386606812477,
-0.10243293642997742,
0.00737339211627841,
-0.0605914406478405,
0.04110806807875633,
-0.04409124702215195,
-0.012843174859881401,
0.1122625544667244,
0.06584297120571136,
-0.008733054623007774,
0.04353703185915947,
-0.04206225648522377,
0.07908561825752258,
0.07672081142663956,
0.007659211754798889,
0.023562287911772728,
0.02020883932709694,
-0.03225928544998169
] |
Active Learning for Continual Learning: Keeping the Past Alive in the Present | https://openreview.net/forum?id=mnLmmtW7HO | [
"Jaehyun Park",
"Dongmin Park",
"Jae-Gil Lee"
] | Poster | *Continual learning (CL)* enables deep neural networks to adapt to ever-changing data distributions. In practice, there may be scenarios where annotation is costly, leading to *active continual learning (ACL)*, which performs *active learning (AL)* for the CL scenarios when reducing the labeling cost by selecting the most informative subset is preferable. However, conventional AL strategies are not suitable for ACL, as they focus solely on learning the new knowledge, leading to *catastrophic forgetting* of previously learned tasks. Therefore, ACL requires a new AL strategy that can balance the prevention of catastrophic forgetting and the ability to quickly learn new tasks. In this paper, we propose **AccuACL**, **Accu**mulated informativeness-based **A**ctive **C**ontinual **L**earning, by the novel use of the Fisher information matrix as a criterion for sample selection, derived from a theoretical analysis of the Fisher-optimality preservation properties within the framework of ACL, while also addressing the scalability issue of Fisher information-based AL. Extensive experiments demonstrate that AccuACL significantly outperforms AL baselines across various CL algorithms, increasing the average accuracy and forgetting by 23.8% and 17.0%, respectively, on average. | active learning, continual learning, Fisher information | Propose a novel active continual learning method based on the accumulative informativeness to avoid catastrophic forgetting | 1,747 | 2501.14278 | [
-0.06270507723093033,
-0.06775858998298645,
-0.013022138737142086,
0.08425245434045792,
0.0767768993973732,
0.07601150125265121,
0.041639186441898346,
-0.05814540386199951,
0.025476763024926186,
-0.009054243564605713,
-0.030684994533658028,
0.03595000505447388,
-0.041099585592746735,
0.020886855199933052,
-0.09568081796169281,
0.07993080466985703,
0.014839604496955872,
0.06989195197820663,
-0.0759127289056778,
-0.021004579961299896,
0.013070441782474518,
0.07287237048149109,
0.0676591768860817,
0.05149615928530693,
-0.014892060309648514,
-0.012878320179879665,
-0.043312713503837585,
-0.061048876494169235,
0.0456424281001091,
-0.03734210133552551,
0.016343573108315468,
0.042727891355752945,
-0.001064213807694614,
-0.023357555270195007,
-0.043986134231090546,
-0.010313042439520359,
-0.10831016302108765,
0.09026166796684265,
0.05304975435137749,
0.008233658969402313,
0.0425078384578228,
-0.0015201341593638062,
0.005741986446082592,
0.013775423169136047,
0.02002337947487831,
0.012387155555188656,
0.007481125183403492,
-0.06758525222539902,
0.05185340344905853,
0.11674798280000687,
-0.030538959428668022,
-0.03218008577823639,
-0.002309916540980339,
0.012403110042214394,
0.010107574053108692,
0.010814601555466652,
0.002234441926702857,
-0.033478762954473495,
-0.08757612854242325,
0.03801221773028374,
0.010531256906688213,
-0.060115884989500046,
-0.017275094985961914,
-0.007649564184248447,
-0.07891274988651276,
0.010289684869349003,
0.05389485880732536,
-0.02317321114242077,
0.014578436501324177,
0.06095035374164581,
-0.030614560469985008,
0.1120988056063652,
-0.04730752855539322,
0.01932772807776928,
0.09273058921098709,
0.019886093214154243,
0.06958264857530594,
-0.08943359553813934,
0.07459814846515656,
0.02488482929766178,
0.030971551313996315,
0.04655390977859497,
0.029177749529480934,
-0.08095859736204147,
0.13237477838993073,
-0.024993058294057846,
0.05368524417281151,
0.013118485920131207,
0.04788250848650932,
-0.06631627678871155,
-0.03330301493406296,
-0.020346740260720253,
0.017712725326418877,
-0.04457420855760574,
0.04973451420664787,
-0.038192421197891235,
-0.0005241985782049596,
-0.04321270436048508,
-0.01607891544699669,
0.05250895395874977,
-0.014625787734985352,
0.04832050949335098,
-0.060712967067956924,
0.009303698316216469,
-0.020505517721176147,
-0.09219539910554886,
0.026188695803284645,
0.11026250571012497,
0.07441990822553635,
-0.0198834128677845,
0.05065450072288513,
-0.02599511481821537,
0.0327976830303669,
-0.001205922570079565,
-0.004250719211995602,
-0.005610578693449497,
-0.015214288607239723,
0.027867086231708527,
-0.02434244193136692,
0.07123639434576035,
-0.01890348084270954,
-0.00007908044062787667,
-0.01171854604035616,
0.028160689398646355,
0.0066542173735797405,
-0.06916116923093796,
0.0005122604197822511,
-4.561912659613867e-34,
0.0018857750110328197,
-0.030377928167581558,
0.008315721526741982,
0.025402218103408813,
0.007331995293498039,
-0.01979486458003521,
-0.027500154450535774,
0.09266828745603561,
0.01298580877482891,
-0.041155412793159485,
0.010925446636974812,
0.036781951785087585,
-0.1012304499745369,
0.09156116098165512,
0.036896198987960815,
-0.041840191930532455,
0.007203788496553898,
0.008367227390408516,
0.012913154438138008,
-0.1124214306473732,
0.007834156043827534,
0.006343642715364695,
0.024436617270112038,
-0.04008355364203453,
0.01902177929878235,
0.011159384623169899,
0.06502676755189896,
0.05251592770218849,
-0.08049813657999039,
0.006879173219203949,
-0.11722514033317566,
0.0011429963633418083,
-0.0117581682279706,
-0.010289343073964119,
-0.001576022943481803,
-0.01951764151453972,
-0.03643161058425903,
0.007537897676229477,
0.09075930714607239,
-0.01664947345852852,
0.028197146952152252,
0.11628223955631256,
0.018837198615074158,
-0.022592972964048386,
-0.061216749250888824,
-0.12352630496025085,
0.05758070573210716,
0.035684652626514435,
-0.04972890019416809,
-0.058927468955516815,
0.02035028301179409,
-0.05855712294578552,
-0.0868736207485199,
-0.0487961508333683,
-0.07244345545768738,
-0.03282839432358742,
-0.053218960762023926,
0.06883349269628525,
-0.009947200305759907,
0.03792142495512962,
0.03912106901407242,
0.03242507576942444,
0.01945139281451702,
0.1004447191953659,
-0.028881506994366646,
0.15923941135406494,
0.004809792619198561,
-0.012593347579240799,
0.049990370869636536,
-0.05191950500011444,
0.0723818987607956,
-0.0187369417399168,
0.01927211508154869,
-0.059718117117881775,
0.045860253274440765,
-0.04652527719736099,
0.05219997838139534,
-0.10218191146850586,
0.018349668011069298,
0.05572791397571564,
0.0034276547376066446,
-0.04547666013240814,
-0.028787706047296524,
0.012278653681278229,
-0.00878470204770565,
-0.0265424232929945,
0.08877601474523544,
-0.02943359687924385,
0.009588862769305706,
0.03331628814339638,
-0.018686562776565552,
0.020098567008972168,
-0.04622018337249756,
-0.04961138963699341,
0.020487874746322632,
-1.987792860075403e-34,
-0.01898369938135147,
0.027539018541574478,
-0.06253105401992798,
0.053020503371953964,
-0.030417341738939285,
-0.005506641697138548,
-0.06867049634456635,
-0.002699370961636305,
-0.05346117913722992,
-0.11201803386211395,
0.005766953341662884,
-0.019328648224473,
-0.04682284966111183,
-0.01485509891062975,
-0.025962121784687042,
-0.0024028951302170753,
-0.08693148195743561,
-0.006732129957526922,
-0.0773121565580368,
0.055444397032260895,
0.009191840887069702,
0.016118459403514862,
-0.04465477913618088,
-0.005004106089472771,
0.05079637095332146,
0.055706705898046494,
-0.05133630707859993,
0.09271502494812012,
0.028182314708828926,
-0.005545064806938171,
-0.08618190139532089,
-0.0695820301771164,
-0.026002157479524612,
-0.06539804488420486,
-0.0494195930659771,
0.06512922048568726,
0.11374000459909439,
0.044490400701761246,
-0.09263576567173004,
0.115179643034935,
0.13234031200408936,
-0.03661523014307022,
-0.065467968583107,
-0.0019362468738108873,
0.05187457427382469,
-0.09412340819835663,
-0.013088636100292206,
0.004451422020792961,
0.007518211379647255,
0.04080939665436745,
0.04706389456987381,
-0.004861693829298019,
-0.07196152210235596,
-0.010052797384560108,
-0.07102619856595993,
0.032515183091163635,
-0.06623643636703491,
-0.02647480182349682,
-0.010579993948340416,
-0.021068869158625603,
-0.03410739079117775,
0.015454117208719254,
-0.0068605560809373856,
0.0400148443877697,
-0.0033831261098384857,
0.01875254511833191,
0.03442799672484398,
0.0724969282746315,
-0.04058298468589783,
-0.0052643464878201485,
0.025293421000242233,
0.08799552172422409,
-0.029618024826049805,
-0.0764787569642067,
-0.006749190390110016,
0.027040889486670494,
-0.044463351368904114,
-0.03595399856567383,
-0.03855123743414879,
-0.024090616032481194,
-0.06347756832838058,
-0.04347528889775276,
-0.011786427348852158,
0.07199376821517944,
0.02708793617784977,
0.0940699502825737,
0.03206891193985939,
0.020650360733270645,
0.06380576640367508,
0.001683077891357243,
0.013190153986215591,
0.008681079372763634,
-0.02822658233344555,
-0.029611552134156227,
-0.05450328439474106,
-4.228040850762227e-8,
-0.042708687484264374,
0.08004236966371536,
0.015894634649157524,
-0.013844492845237255,
0.10887235403060913,
-0.06438781321048737,
-0.03240988403558731,
0.0908283218741417,
-0.031560204923152924,
-0.008514313958585262,
0.1018659695982933,
0.06454121321439743,
0.02131691761314869,
-0.06951136142015457,
-0.02617994137108326,
0.06164630874991417,
0.07908429205417633,
-0.049854815006256104,
0.009192348457872868,
-0.005004441365599632,
-0.0029119679238647223,
-0.019279994070529938,
0.004684009123593569,
-0.016943037509918213,
-0.053305041044950485,
-0.07772034406661987,
0.058632992208004,
0.07280261069536209,
0.060873594135046005,
-0.018099378794431686,
-0.0012533217668533325,
-0.015063775703310966,
0.02393888682126999,
-0.0012074168771505356,
0.0424458310008049,
-0.004621368367224932,
0.07038209587335587,
-0.11545336246490479,
-0.0722552016377449,
-0.025523709133267403,
-0.030902190133929253,
0.07522975653409958,
-0.03598907217383385,
0.04711655154824257,
0.027957459911704063,
-0.009711234830319881,
-0.050912704318761826,
-0.06440681219100952,
0.019684400409460068,
-0.04921533539891243,
0.01534783374518156,
0.03247606381773949,
-0.001071212813258171,
0.13303425908088684,
0.09193790704011917,
-0.03962315246462822,
0.029435554519295692,
0.0021163697820156813,
-0.02997070737183094,
0.05026945844292641,
0.035520490258932114,
0.0068485611118376255,
-0.06954532116651535,
0.002004633191972971
] |
Group Downsampling with Equivariant Anti-aliasing | https://openreview.net/forum?id=sOte83GogU | [
"Md Ashiqur Rahman",
"Raymond A. Yeh"
] | Poster | Downsampling layers are crucial building blocks in CNN architectures, which help to increase the receptive field for learning high-level features and reduce the amount of memory/computation in the model. In this work, we study the generalization of the uniform downsampling layer for group equivariant architectures, e.g., $G$-CNNs. That is, we aim to downsample signals (feature maps) on general finite groups *with* anti-aliasing. This involves the following: **(a)** Given a finite group and a downsampling rate, we present an algorithm to form a suitable choice of subgroup. **(b)** Given a group and a subgroup, we study the notion of bandlimited-ness and propose how to perform anti-aliasing. Notably, our method generalizes the notion of downsampling based on classical sampling theory. When the signal is on a cyclic group, i.e., periodic, our method recovers the standard downsampling of an ideal low-pass filter followed by a subsampling operation. Finally, we conducted experiments on image classification tasks demonstrating that the proposed downsampling operation improves accuracy, better preserves equivariance, and reduces model size when incorporated into $G$-equivariant networks | equivariance, downsampling, signal processing | null | 1,744 | null | [
-0.09222906827926636,
-0.00577184185385704,
0.1049724668264389,
-0.021435966715216637,
0.01636091247200966,
-0.04363827034831047,
0.04140245541930199,
-0.06433252990245819,
0.002467166632413864,
-0.10599055141210556,
-0.005093772429972887,
-0.015516783110797405,
0.036330122500658035,
-0.014139398001134396,
-0.02886592224240303,
-0.037623900920152664,
0.03576294705271721,
0.14483381807804108,
-0.07703591138124466,
-0.05589509382843971,
0.03206843137741089,
-0.039201077073812485,
0.058263447135686874,
0.006670582573860884,
0.053092822432518005,
-0.0224912129342556,
-0.02200581133365631,
-0.04043549299240112,
0.014881761744618416,
-0.04585311934351921,
-0.03602450713515282,
0.11671195924282074,
0.017871864140033722,
0.030136507004499435,
-0.07859151810407639,
0.09458810091018677,
-0.006226957309991121,
0.04494820535182953,
0.05029497668147087,
0.03377588465809822,
0.015613759867846966,
0.04878022521734238,
0.010013960301876068,
0.04340018332004547,
0.0547187365591526,
0.07452883571386337,
0.0961105152964592,
-0.035722944885492325,
-0.03386891260743141,
0.06844493001699448,
0.07391003519296646,
0.05348559468984604,
-0.09834552556276321,
0.1680714190006256,
-0.05801757052540779,
-0.07749311625957489,
-0.05279175564646721,
-0.013625022023916245,
-0.0492427758872509,
-0.019836336374282837,
0.017235342413187027,
-0.06514190882444382,
-0.012894186191260815,
0.017868924885988235,
-0.11490880697965622,
0.003040200099349022,
0.06921052187681198,
-0.0065900785848498344,
0.07029219716787338,
0.03117096610367298,
0.07977447658777237,
0.039930060505867004,
-0.016116037964820862,
0.041060443967580795,
0.01269187405705452,
0.043063994497060776,
-0.015152014791965485,
0.022241409868001938,
0.04682471603155136,
-0.10294373333454132,
0.08288192749023438,
0.014730633236467838,
0.022398296743631363,
0.00890602171421051,
0.09775169938802719,
-0.0035132216289639473,
-0.0661499872803688,
-0.01616710238158703,
-0.012764031998813152,
-0.015002240426838398,
-0.11654294282197952,
0.011147093027830124,
0.023598451167345047,
-0.025465743616223335,
-0.04071221500635147,
-0.02235141582787037,
0.08928333222866058,
-0.048518192023038864,
-0.00876842811703682,
0.14308080077171326,
-0.002896316582337022,
-0.016178954392671585,
-0.017174169421195984,
0.01714690402150154,
0.02560112066566944,
-0.07000897824764252,
0.08348491787910461,
0.00961976870894432,
0.008085321635007858,
-0.0753435343503952,
-0.032036151736974716,
0.053246352821588516,
0.01882847584784031,
-0.023409903049468994,
0.01778971590101719,
-0.014528253115713596,
0.05457785353064537,
0.030172094702720642,
0.04309979826211929,
-0.009951000101864338,
-0.04355955868959427,
-0.0023806521203368902,
-0.017236346378922462,
0.0480608195066452,
0.06730090081691742,
0.07243695855140686,
-0.1312447488307953,
-6.973810335730268e-34,
0.024261265993118286,
0.014814252965152264,
0.05916958674788475,
-0.032622870057821274,
0.040647029876708984,
0.0253281332552433,
-0.04288298636674881,
0.062107980251312256,
-0.03634044900536537,
0.06498844176530838,
-0.05351657420396805,
-0.04508976638317108,
-0.02137143351137638,
0.10522469133138657,
0.021619293838739395,
-0.0064400010742247105,
-0.046359047293663025,
-0.02644071727991104,
0.024059677496552467,
-0.08141711354255676,
-0.07011356204748154,
-0.013065299950540066,
-0.06538306176662445,
-0.012499858625233173,
-0.06913676857948303,
-0.01428135670721531,
-0.0628919005393982,
-0.016080079600214958,
-0.0040903519839048386,
0.017132669687271118,
-0.0601884126663208,
0.041289590299129486,
-0.04151901975274086,
-0.0014144498854875565,
-0.005346810910850763,
-0.01063696201890707,
0.013235275633633137,
0.011822100728750229,
0.001997375162318349,
-0.007066653575748205,
-0.015323416329920292,
0.017155954614281654,
0.0034825140610337257,
-0.0858134925365448,
0.00682024797424674,
-0.08568230271339417,
0.0493226982653141,
0.00716280797496438,
-0.05728646367788315,
0.028995860368013382,
0.04092591628432274,
-0.07053611427545547,
-0.05359388515353203,
-0.04103735461831093,
-0.05454093590378761,
0.02131345123052597,
0.031843990087509155,
0.0754292830824852,
0.006159489043056965,
0.07319986820220947,
-0.0031289842445403337,
0.0072940681129693985,
-0.042881935834884644,
0.02945154905319214,
-0.06224208325147629,
-0.01961437612771988,
0.007294254843145609,
0.04738154262304306,
0.005662402603775263,
0.03520874306559563,
-0.0010525332763791084,
0.006424139719456434,
-0.060599155724048615,
-0.020278526470065117,
-0.008159291930496693,
0.026217568665742874,
0.04146026074886322,
-0.0010928019182756543,
-0.028609251603484154,
0.09157150983810425,
0.009848265908658504,
-0.062230728566646576,
0.0028807083144783974,
-0.06381294131278992,
-0.07372649759054184,
0.031867820769548416,
0.050224173814058304,
-0.11993440985679626,
0.011022582650184631,
-0.05850343406200409,
-0.15545125305652618,
0.033678267151117325,
0.03440750017762184,
0.0038803413044661283,
0.00954134576022625,
-4.519751442505214e-34,
0.040607430040836334,
0.06327684223651886,
-0.018776381388306618,
0.035027265548706055,
-0.022450830787420273,
-0.08181456476449966,
0.05409512296319008,
0.011280166916549206,
0.0034205156844109297,
0.00757264718413353,
-0.022967645898461342,
-0.0015059056458994746,
0.04230089113116264,
0.017222072929143906,
-0.04150594398379326,
-0.020905939862132072,
-0.043974701315164566,
-0.021320443600416183,
0.05407897010445595,
0.057628605514764786,
-0.03640705719590187,
0.045992039144039154,
-0.012020964175462723,
0.03179776668548584,
-0.05500362068414688,
0.058427199721336365,
-0.06619695574045181,
0.03760943561792374,
0.00961924996227026,
-0.014353353530168533,
-0.05609171465039253,
-0.07193324714899063,
-0.050341684371232986,
-0.07728295773267746,
0.00453718239441514,
0.026477815583348274,
-0.039260830730199814,
0.043942105025053024,
-0.043832842260599136,
0.01707892119884491,
0.02722933515906334,
-0.02699930965900421,
0.042637791484594345,
0.03532124310731888,
0.07197552174329758,
-0.06323900818824768,
0.01830771379172802,
0.031806930899620056,
-0.055768124759197235,
0.051490284502506256,
-0.04627364128828049,
0.04060230776667595,
-0.019463498145341873,
0.09305383265018463,
-0.012391764670610428,
-0.00550859933719039,
0.012989060953259468,
-0.041288215667009354,
0.0776313990354538,
0.057357966899871826,
-0.04162054881453514,
-0.08326306939125061,
-0.10604246705770493,
-0.047062139958143234,
-0.023602601140737534,
-0.020368680357933044,
-0.07539767771959305,
0.02125312015414238,
0.05308493599295616,
0.05894192308187485,
-0.017549218609929085,
-0.11343572288751602,
-0.028888754546642303,
-0.04554907605051994,
-0.06519749015569687,
-0.08002260327339172,
0.043171945959329605,
0.008870248682796955,
-0.02377910166978836,
-0.003756702644750476,
0.016541825607419014,
-0.037933800369501114,
-0.0557205006480217,
0.11241253465414047,
0.006749750580638647,
0.040259528905153275,
-0.06031586974859238,
0.020562821999192238,
0.001467798720113933,
0.022070670500397682,
-0.013232363387942314,
-0.0022648025769740343,
0.06278285384178162,
0.12014954537153244,
-0.020243145525455475,
-4.628128635886242e-8,
-0.05002956464886665,
0.09680803120136261,
-0.010901943780481815,
-0.009558035992085934,
0.052952468395233154,
-0.0952172800898552,
-0.021282803267240524,
0.05957712605595589,
0.005056781228631735,
-0.08258815854787827,
0.09919334948062897,
-0.030336305499076843,
-0.07367757707834244,
0.03125875070691109,
0.029333187267184258,
0.05191436782479286,
0.00761299068108201,
0.05410376563668251,
0.0068264915607869625,
0.03441363573074341,
-0.005478496663272381,
-0.044875022023916245,
0.03409024327993393,
-0.025878995656967163,
0.021427199244499207,
-0.08018128573894501,
0.04998409375548363,
-0.004411690402776003,
0.0071835205890238285,
-0.03487978130578995,
-0.09131789952516556,
0.0512097030878067,
0.021824510768055916,
0.027123695239424706,
0.08188337087631226,
0.011618376709520817,
-0.048097241669893265,
0.04164295643568039,
0.04140602424740791,
-0.006854059174656868,
-0.021793989464640617,
0.05232595279812813,
-0.008856379427015781,
-0.0010337774874642491,
-0.0036355294287204742,
-0.006608662661164999,
0.14730054140090942,
-0.03041660226881504,
0.042797766625881195,
0.06052543222904205,
0.050691571086645126,
0.035675209015607834,
0.016518529504537582,
0.059304624795913696,
-0.014375000260770321,
-0.0013381866738200188,
-0.010503954254090786,
-0.043257251381874084,
0.022350896149873734,
-0.0203405749052763,
-0.050447892397642136,
0.08052213490009308,
-0.10824299603700638,
-0.04786863923072815
] |
Visually Guided Decoding: Gradient-Free Hard Prompt Inversion with Language Models | https://openreview.net/forum?id=mQ55y4s5hj | [
"Donghoon Kim",
"Minji Bae",
"Kyuhong Shim",
"Byonghyo Shim"
] | Poster | Text-to-image generative models like DALL-E and Stable Diffusion have revolutionized visual content creation across various applications, including advertising, personalized media, and design prototyping.
However, crafting effective textual prompts to guide these models remains challenging, often requiring extensive trial and error.
Existing prompt inversion approaches, such as soft and hard prompt techniques, are not so effective due to the limited interpretability and incoherent prompt generation.
To address these issues, we propose Visually Guided Decoding (VGD), a gradient-free approach that leverages large language models (LLMs) and CLIP-based guidance to generate coherent and semantically aligned prompts.
In essence, VGD utilizes the robust text generation capabilities of LLMs to produce human-readable prompts.
Further, by employing CLIP scores to ensure alignment with user-specified visual concepts, VGD enhances the interpretability, generalization, and flexibility of prompt generation without the need for additional training.
Our experiments demonstrate that VGD outperforms existing prompt inversion techniques in generating understandable and contextually relevant prompts, facilitating more intuitive and controllable interactions with text-to-image models. | text-to-image, inversion, gradient free hard prompt inversion, language model guidance on latent diffusion model | Visually Guided Decoding (VGD) improves prompt generation for text-to-image models by using language models and CLIP guidance to create coherent, human-readable prompts aligned with visual concepts. | 1,743 | null | [
0.020704366266727448,
-0.08737150579690933,
0.07577475905418396,
0.02453332021832466,
0.06965915113687515,
-0.014275401830673218,
-0.056695133447647095,
0.010349920950829983,
0.030218342319130898,
-0.044659703969955444,
-0.05313391610980034,
-0.028461012989282608,
0.055176883935928345,
0.020824167877435684,
0.024300865828990936,
-0.015094042755663395,
0.08052176237106323,
0.05724084749817848,
-0.050314877182245255,
-0.0280484389513731,
0.006750940810889006,
0.010953709483146667,
0.023707440122961998,
-0.016742832958698273,
0.00841501820832491,
0.027070092037320137,
-0.018533052876591682,
-0.01418973132967949,
0.07092242687940598,
-0.027105215936899185,
0.016357768326997757,
-0.04174202308058739,
0.04956534132361412,
0.07691722363233566,
-0.04413967207074165,
0.09679180383682251,
-0.0014788112603127956,
0.05438460037112236,
-0.034102506935596466,
-0.07593460381031036,
-0.02537902072072029,
-0.025322193279862404,
0.027214158326387405,
0.024154983460903168,
0.0771869644522667,
-0.1173413097858429,
-0.05740908905863762,
-0.017811113968491554,
-0.0015215191524475813,
0.031040746718645096,
-0.03662806749343872,
-0.023874366655945778,
0.0057102167047560215,
0.010716604068875313,
-0.06633319705724716,
0.07043100893497467,
0.07800159603357315,
-0.0050818114541471004,
-0.003513387404382229,
0.012796413153409958,
-0.0818396732211113,
-0.041357941925525665,
-0.02401188388466835,
0.04662690311670303,
0.011894981376826763,
-0.051968321204185486,
0.04124075174331665,
0.03520017862319946,
-0.008830057457089424,
0.008941330015659332,
-0.06416980922222137,
-0.017839767038822174,
-0.052943672984838486,
0.027613824233412743,
-0.006474214140325785,
0.0900510922074318,
-0.019850892946124077,
-0.08357039839029312,
0.029400216415524483,
-0.062143538147211075,
0.04795413091778755,
0.024705249816179276,
0.07187879085540771,
0.03605257347226143,
0.0081512201577425,
0.018759749829769135,
0.018266664817929268,
0.04633387178182602,
0.02095011994242668,
0.028091171756386757,
-0.06300671398639679,
-0.08352004736661911,
0.02311825193464756,
0.03361988812685013,
0.0002809495781548321,
0.07597672194242477,
0.04359082877635956,
-0.16862785816192627,
-0.07406988739967346,
0.02549947425723076,
0.0018669308628886938,
0.04103763401508331,
0.011993452906608582,
-0.1115594282746315,
-0.02077539637684822,
-0.06907311081886292,
0.07322721183300018,
0.04551953822374344,
-0.0031913965940475464,
-0.049856871366500854,
0.06856373697519302,
0.04263406991958618,
0.026477551087737083,
-0.0022224660497158766,
0.014312060549855232,
-0.06070083752274513,
-0.0032656011171638966,
-0.10750996321439743,
0.05932173877954483,
-0.009634586051106453,
0.020272018387913704,
0.010154837742447853,
-0.11198006570339203,
0.03282159939408302,
0.06973721086978912,
-0.08253036439418793,
0.01561201922595501,
4.719444055480919e-34,
0.03710376098752022,
0.06019445136189461,
0.08931940793991089,
0.1508154720067978,
0.06750918924808502,
0.005058087408542633,
-0.04484260454773903,
-0.015818243846297264,
-0.10006365925073624,
0.007862918078899384,
0.02098676562309265,
-0.06434588134288788,
-0.05247023329138756,
0.15946064889431,
0.009900834411382675,
-0.024492720142006874,
-0.02348140813410282,
0.05363369733095169,
-0.07464305311441422,
-0.008848197758197784,
-0.06053280830383301,
0.01015520840883255,
0.017389845103025436,
-0.023124676197767258,
0.05504487454891205,
-0.00819284375756979,
0.06076018512248993,
-0.040220364928245544,
-0.032919660210609436,
0.006619549356400967,
-0.1331855207681656,
-0.07974404096603394,
0.057436078786849976,
-0.023533128201961517,
-0.055102258920669556,
-0.02435397170484066,
-0.005253506358712912,
-0.08744628727436066,
0.03688349202275276,
0.03862403705716133,
0.021373186260461807,
0.014345748350024223,
0.0064899311400949955,
0.052662502974271774,
-0.10083150118589401,
0.027902813628315926,
0.004690950270742178,
0.06511961668729782,
-0.09024512022733688,
0.019742881879210472,
0.009101759642362595,
0.027395175769925117,
-0.06232732534408569,
-0.05538216233253479,
0.012013915926218033,
-0.04167107865214348,
0.06650712341070175,
-0.022065764293074608,
0.00644896412268281,
-0.034570373594760895,
0.04449555650353432,
0.10024601221084595,
0.007150867488235235,
0.061389632523059845,
0.0009120245813392103,
-0.014504814520478249,
0.00618418212980032,
0.013535678386688232,
-0.001532856491394341,
0.022686010226607323,
-0.1122758537530899,
0.005175391212105751,
-0.03091365471482277,
-0.03898086026310921,
-0.008011776022613049,
0.010063296183943748,
0.04806151241064072,
-0.045049697160720825,
0.10199904441833496,
0.03237301483750343,
-0.022552041336894035,
-0.05561424046754837,
-0.0007258190889842808,
-0.034394118934869766,
-0.027780167758464813,
-0.031252194195985794,
0.06613235175609589,
-0.07125365734100342,
-0.060093410313129425,
-0.06945744901895523,
-0.01451422180980444,
-0.008897602558135986,
-0.03179150074720383,
0.007173607591539621,
0.004530461970716715,
-1.1434763642642847e-34,
0.010906439274549484,
0.06279722601175308,
-0.04084579274058342,
0.005075708497315645,
-0.014336848631501198,
-0.027935082092881203,
0.022152256220579147,
0.06567751616239548,
0.02140372060239315,
-0.08989452570676804,
-0.04592999815940857,
0.048515211790800095,
0.004511925391852856,
0.026277286931872368,
0.0026795233134180307,
0.008358651772141457,
0.045861534774303436,
0.01640091836452484,
-0.011648612096905708,
0.021730123087763786,
-0.0034988196566700935,
0.08138567954301834,
-0.08084964007139206,
0.009687229059636593,
-0.045917268842458725,
0.03872634842991829,
-0.029003998264670372,
0.05030715838074684,
0.005216539837419987,
-0.01195746473968029,
-0.02630448527634144,
0.03270060941576958,
0.032043587416410446,
-0.007497052662074566,
-0.062187083065509796,
0.06547200679779053,
0.03948154300451279,
-0.04454808309674263,
-0.031981632113456726,
0.10112603008747101,
0.02696307934820652,
-0.0574677512049675,
-0.03628525882959366,
0.0020481725223362446,
-0.07474850118160248,
-0.05719329044222832,
0.0009651068830862641,
-0.07305291295051575,
0.06129664182662964,
0.10372140258550644,
-0.05902924761176109,
0.05186546593904495,
-0.06633330136537552,
0.0028204184491187334,
-0.10205315053462982,
-0.1155039444565773,
-0.09255895018577576,
0.007837408222258091,
0.09525822848081589,
-0.0904267355799675,
-0.009853644296526909,
0.0001712315424811095,
0.0659824013710022,
-0.09502999484539032,
-0.01471028570085764,
-0.059750065207481384,
-0.002938448218628764,
-0.04151926934719086,
0.018861394375562668,
-0.031173227354884148,
0.030953995883464813,
-0.07361669838428497,
0.01646139845252037,
0.023920834064483643,
0.08405865728855133,
0.010588680393993855,
0.02299623191356659,
-0.016406310722231865,
-0.028762083500623703,
-0.07650463283061981,
0.06682360172271729,
0.041424646973609924,
-0.008289664052426815,
0.08804158121347427,
0.07895996421575546,
0.044449612498283386,
-0.010876321233808994,
0.1231786236166954,
0.022723805159330368,
0.06071871891617775,
-0.02899867668747902,
0.060900282114744186,
0.048890434205532074,
0.10969648510217667,
-0.03725656121969223,
-3.9273214014201585e-8,
-0.05887530371546745,
-0.003217266872525215,
0.0017643228638917208,
0.05097167193889618,
-0.03063635714352131,
-0.03675823658704758,
0.009281197562813759,
-0.0351683646440506,
0.05482063069939613,
-0.08658480644226074,
-0.013841923326253891,
0.008568543009459972,
-0.03793181851506233,
-0.04588906094431877,
-0.029152048751711845,
0.10007765144109726,
-0.09834098815917969,
0.04990905895829201,
-0.014746285043656826,
-0.00730101577937603,
0.01593305915594101,
0.05451356992125511,
-0.03966072201728821,
-0.007134233135730028,
0.03715556859970093,
0.027226611971855164,
-0.04932476580142975,
0.04034706950187683,
0.029003065079450607,
-0.05437612161040306,
0.04457765445113182,
0.008626104332506657,
0.025000175461173058,
0.021819159388542175,
-0.024349242448806763,
0.07970109581947327,
-0.017696380615234375,
0.02520117349922657,
-0.007714441046118736,
-0.0016834910493344069,
0.01688862219452858,
0.04701455682516098,
0.016031891107559204,
0.022331995889544487,
-0.034638434648513794,
-0.0008471423643641174,
0.07811558246612549,
-0.08412926644086838,
0.03406606242060661,
-0.0363655611872673,
-0.04689841717481613,
-0.09584712237119675,
0.0359797403216362,
0.02897169254720211,
0.009937589056789875,
0.004334277473390102,
0.03464178740978241,
0.056489262729883194,
0.05944415554404259,
0.09604553133249283,
-0.01630871184170246,
0.0871170312166214,
-0.014105381444096565,
-0.010347944684326649
] |
BP-Modified Local Loss for Efficient Training of Deep Neural Networks | https://openreview.net/forum?id=MtW30ql5Oj | [
"REN Lianhai",
"Qianxiao Li"
] | Poster | The training of large models is memory-constrained, one direction to relieve
this is training using local loss, like GIM, LoCo, and Forward-Forward
algorithms. However, the local loss methods often face the issue of slow or
non-convergence. In this paper, we propose a novel BP-modified local loss
method that uses the true Backward Propagation (BP) gradient to modify the
local loss gradient to improve the performance of local loss training. We
use the stochastic modified equation to analyze our method and show that
modified offset decreases the bias between the BP gradient and local loss
gradient, but introduces additional variance, which results in a
bias-variance balance. Numerical experiments on full-tuning and LoKr tuning
on the ResNet-50 model and LoRA tuning on the ViT-b16 model on CIFAR-100
datasets show 20.5\% test top-1 accuracy improvement for the Forward-Forward
algorithm, 18.6\% improvement for LoCo algorithm and achieve only on average
7.7\% of test accuracy loss compared to the BP algorithm, with up to 75\%
memory savings. | deep learning optimization, local loss training, bias-variance balance | We proposed a novel method that periodically compute the BP gradient and use it to modify the local loss gradient. This method improve the performance of the original local loss methods with negligible additional memory usage. | 1,740 | null | [
-0.037188176065683365,
-0.03993872180581093,
0.048447925597429276,
0.043612778186798096,
0.007187632378190756,
0.02438906580209732,
-0.0308403130620718,
-0.016908153891563416,
0.005702354945242405,
-0.0769365206360817,
-0.037608277052640915,
0.06958671659231186,
0.04061324894428253,
-0.057877250015735626,
-0.07123901695013046,
0.025506436824798584,
0.09281960129737854,
0.13874876499176025,
-0.0830722227692604,
-0.04074689745903015,
-0.06582827866077423,
0.056277062743902206,
0.025872090831398964,
0.07435378432273865,
0.06221776455640793,
-0.04127182066440582,
-0.035457443445920944,
-0.019486013799905777,
0.021815100684762,
-0.05621457099914551,
0.004638591781258583,
-0.01331462524831295,
-0.047024939209222794,
0.014572483487427235,
-0.0948859453201294,
0.05566203221678734,
-0.07536821812391281,
0.02279084362089634,
-0.01080634817481041,
0.04431834816932678,
0.015478291548788548,
0.044661737978458405,
0.05905120447278023,
-0.00815193634480238,
0.11841166019439697,
0.031163135543465614,
0.023982714861631393,
-0.047737836837768555,
0.04318918287754059,
0.043236248195171356,
-0.00991345476359129,
-0.017583560198545456,
-0.07052813470363617,
0.06012055277824402,
-0.05234939232468605,
0.005745110101997852,
0.052650850266218185,
0.05020097643136978,
-0.04692956432700157,
0.10605984181165695,
0.036043085157871246,
-0.03511327877640724,
-0.06320330500602722,
-0.08423458784818649,
-0.09582435339689255,
-0.09003236889839172,
0.006691254209727049,
0.004082261584699154,
0.024152159690856934,
-0.04373869672417641,
0.043902620673179626,
0.08827656507492065,
-0.03620092570781708,
0.06147713214159012,
-0.00477533508092165,
0.04403770714998245,
0.1609273999929428,
0.02580125629901886,
0.014997100457549095,
-0.028510892763733864,
-0.008894911967217922,
0.003305977676063776,
0.07341280579566956,
-0.07751604914665222,
0.09446048736572266,
-0.031752221286296844,
-0.004252466838806868,
-0.01787078194320202,
0.054482631385326385,
-0.07471321523189545,
0.03357989713549614,
0.014777706936001778,
-0.05020515248179436,
-0.002995295450091362,
0.04759394749999046,
0.004506751429289579,
-0.01180583331733942,
-0.048967164009809494,
-0.006287682335823774,
0.10911072790622711,
0.026556072756648064,
0.018021756783127785,
0.008620919659733772,
0.016435712575912476,
0.08217398822307587,
-0.012128823436796665,
0.0993780642747879,
0.021105268970131874,
0.04726315662264824,
-0.013370699249207973,
0.03179221972823143,
0.02771306410431862,
0.003540873760357499,
0.018604589626193047,
-0.0005158589337952435,
0.01806751824915409,
-0.0397457592189312,
-0.03990968316793442,
-0.03513066843152046,
0.062393248081207275,
-0.17244189977645874,
-0.019511424005031586,
-0.01770792342722416,
0.017526041716337204,
0.009705056436359882,
-0.07378986477851868,
-0.06310964375734329,
8.763800786929637e-34,
0.013449164107441902,
-0.04059525579214096,
0.011968015693128109,
-0.05409960821270943,
0.03594852611422539,
-0.011729631572961807,
0.044135864824056625,
0.006455579772591591,
-0.06638789921998978,
0.018170403316617012,
-0.021137919276952744,
-0.026268362998962402,
-0.05403188616037369,
0.05252635478973389,
0.054700493812561035,
-0.0715826004743576,
-0.006794900167733431,
-0.03160255029797554,
0.031004806980490685,
0.020559722557663918,
0.07489129900932312,
-0.06752050668001175,
-0.02177879586815834,
-0.06973516941070557,
-0.02887536585330963,
0.013151266612112522,
0.051463112235069275,
0.01107849832624197,
-0.07933956384658813,
0.0472957082092762,
-0.052941445261240005,
-0.037455737590789795,
-0.017568761482834816,
-0.044991638511419296,
0.03166037052869797,
-0.02137436531484127,
-0.048466648906469345,
0.048143789172172546,
0.008628202602267265,
0.014094464480876923,
0.011565983295440674,
0.08484683930873871,
-0.04053919389843941,
-0.06415192782878876,
-0.08598685264587402,
-0.09367319941520691,
0.07409850507974625,
0.03593943640589714,
-0.008080034516751766,
-0.04185733199119568,
-0.040085598826408386,
-0.0016610577004030347,
-0.08299481123685837,
0.022904494777321815,
0.013980889692902565,
-0.0017194250831380486,
0.07452894002199173,
0.1067148745059967,
0.0340474508702755,
0.05803374573588371,
0.04131285473704338,
0.061220910400152206,
-0.0356832854449749,
0.14214253425598145,
0.017780354246497154,
-0.04236236959695816,
0.020906049758195877,
0.03539799526333809,
-0.0007219842518679798,
-0.024782337248325348,
0.002632123650982976,
-0.003315946087241173,
0.020878911018371582,
0.004510974045842886,
-0.0034350028727203608,
-0.047144316136837006,
0.0133571932092309,
-0.051200829446315765,
0.0029176888056099415,
-0.07849506288766861,
-0.012695923447608948,
0.11699479818344116,
-0.07205744087696075,
0.04449187219142914,
-0.04491270333528519,
0.008727260865271091,
0.0411541685461998,
-0.10831166058778763,
-0.019688842818140984,
0.012196211144328117,
-0.06861098110675812,
-0.0680122971534729,
0.041684120893478394,
-0.046686671674251556,
-0.0731697678565979,
-3.897931920487013e-34,
-0.05883710831403732,
0.057970575988292694,
-0.06408901512622833,
0.05451671779155731,
-0.03341900184750557,
-0.033991023898124695,
0.05844118446111679,
0.038657527416944504,
-0.014662256464362144,
0.013811861164867878,
-0.0013872500276193023,
-0.030719058588147163,
-0.015529433265328407,
0.08281264454126358,
-0.0009285727282986045,
-0.010564533062279224,
-0.03564898297190666,
-0.0034871376119554043,
0.010960988700389862,
-0.05542547628283501,
-0.0129245575517416,
0.09743843227624893,
-0.07128755748271942,
0.02199549227952957,
-0.051585689187049866,
-0.012270706705749035,
-0.017066363245248795,
0.12461011111736298,
-0.03763285651803017,
-0.08273863047361374,
-0.06402262300252914,
0.06323356926441193,
0.003703424008563161,
0.03807516023516655,
-0.04154378920793533,
0.08536958694458008,
0.013492879457771778,
0.02435985393822193,
0.06715402752161026,
0.03133227303624153,
0.07563045620918274,
-0.005038774106651545,
0.0002928679168689996,
-0.027815241366624832,
0.0037491980474442244,
0.012354012578725815,
-0.11359602957963943,
-0.026458704844117165,
-0.0035707084462046623,
0.023097794502973557,
0.02419472299516201,
-0.07655612379312515,
-0.0341087207198143,
0.06595119088888168,
-0.053599849343299866,
-0.04613981023430824,
0.016179557889699936,
0.0321943461894989,
0.01269458793103695,
0.015844659879803658,
-0.03199081867933273,
-0.09225543588399887,
0.0026229151990264654,
-0.024772554636001587,
-0.05489721894264221,
0.08024518191814423,
0.015532554127275944,
0.0352909155189991,
0.10247960686683655,
0.014619356021285057,
0.021939195692539215,
0.0975494235754013,
0.00040488175000064075,
0.025361504405736923,
-0.0613071583211422,
-0.01212610024958849,
-0.005723764654248953,
-0.00648837024345994,
0.027598707005381584,
-0.012478780932724476,
0.07786229997873306,
-0.01743849739432335,
-0.041118551045656204,
-0.010499291121959686,
0.10357939451932907,
0.0626634955406189,
0.01247329916805029,
0.08378245681524277,
0.0074987346306443214,
-0.024336939677596092,
-0.008609789423644543,
0.0040058717131614685,
0.04284914582967758,
-0.08727876096963882,
-0.038888756185770035,
-4.84720921178905e-8,
-0.09795500338077545,
0.04034343361854553,
-0.007785586174577475,
0.14221543073654175,
-0.009982522577047348,
-0.03307513892650604,
0.05574825033545494,
0.07074647396802902,
-0.0035312012769281864,
-0.026004739105701447,
0.0768689215183258,
-0.01506873406469822,
-0.01917230151593685,
-0.002482681768015027,
0.004850656725466251,
0.05101853236556053,
0.0399903804063797,
0.0015492235543206334,
0.032546449452638626,
0.04105091094970703,
0.000169460239703767,
0.05353005230426788,
0.06285735219717026,
-0.012105504982173443,
0.052197881042957306,
-0.08120319247245789,
0.003649249905720353,
0.07192548364400864,
0.09367011487483978,
-0.07396314293146133,
0.0010394033743068576,
0.006466142367571592,
-0.025852680206298828,
0.04368695244193077,
-0.005410917103290558,
0.03211585432291031,
-0.001400350476615131,
0.020383741706609726,
-0.03273797035217285,
0.05014713108539581,
0.025288157165050507,
-0.014363644644618034,
-0.04848182201385498,
0.0009575082804076374,
-0.04101526364684105,
0.014542772434651852,
0.028743138536810875,
-0.07256311178207397,
-0.049061428755521774,
-0.06990079581737518,
0.08361036330461502,
0.00047803117195144296,
0.0015560234896838665,
0.02644192799925804,
0.06156374141573906,
-0.011141881346702576,
-0.011250543408095837,
-0.06396113336086273,
-0.023405248299241066,
0.08406274765729904,
-0.10756882280111313,
0.01570122316479683,
-0.06780031323432922,
-0.02220200002193451
] |
The Crucial Role of Samplers in Online Direct Preference Optimization | https://openreview.net/forum?id=F6z3utfcYw | [
"Ruizhe Shi",
"Runlong Zhou",
"Simon Shaolei Du"
] | Poster | Direct Preference Optimization (DPO) has emerged as a stable, scalable, and efficient solution for language model alignment.
Despite its empirical success, the optimization properties, particularly the impact of samplers on its convergence rates, remain under-explored. In this paper, we provide a rigorous analysis of DPO's convergence rates with different sampling strategies under the exact gradient setting, revealing a surprising separation: uniform sampling achieves $\textbf{linear}$ convergence, while our proposed online sampler achieves $\textbf{quadratic}$ convergence. We further adapt the sampler to practical settings by incorporating posterior distributions and logit mixing, demonstrating improvements over previous methods. For example, it outperforms vanilla DPO by over $7.4$% on Safe-RLHF dataset. Our results not only offer insights into the theoretical understanding of DPO but also pave the way for further algorithm designs. | direct preference optimization, online DPO, tabular softmax policy | We study convergence rates of (online) DPO from optimization perspective, and show the impact of samplers through a theoretical separation and empirical experiments. | 1,735 | 2409.19605 | [
-0.07589994370937347,
-0.11100170016288757,
0.015574714168906212,
-0.040046174079179764,
0.05201181024312973,
0.010577314533293247,
-0.004453565459698439,
0.039818622171878815,
0.0398729182779789,
-0.0365331806242466,
-0.009194627404212952,
0.003264909377321601,
0.0014042977709323168,
-0.03698324039578438,
0.08523939549922943,
-0.050509121268987656,
0.08510766923427582,
0.05982545018196106,
-0.0413578636944294,
-0.04212472215294838,
-0.03256402164697647,
0.07699066400527954,
0.009353459812700748,
-0.044205520302057266,
0.07218698412179947,
-0.02992534451186657,
0.04709092900156975,
0.010225842706859112,
0.04804868996143341,
-0.011843658052384853,
0.06302309036254883,
0.03829061612486839,
0.021497545763850212,
-0.05152878537774086,
-0.03540296107530594,
-0.010921494103968143,
-0.11930627375841141,
-0.07022810727357864,
-0.004922166466712952,
0.022784531116485596,
-0.0469319149851799,
0.024671869352459908,
0.009459379129111767,
-0.021056171506643295,
0.029180968180298805,
0.03208453953266144,
-0.01784190535545349,
0.06713135540485382,
0.022547191008925438,
0.04752856865525246,
-0.07830333709716797,
-0.0028627440333366394,
0.0070242262445390224,
0.03019939363002777,
-0.027622757479548454,
0.0243056770414114,
0.025059636682271957,
0.04470811039209366,
0.04128403961658478,
-0.017010867595672607,
-0.09615912288427353,
-0.11504777520895004,
-0.0005880146054551005,
0.019512224942445755,
-0.009915304370224476,
-0.025959959253668785,
0.038039352744817734,
0.044590339064598083,
0.03621075674891472,
0.0887007936835289,
-0.09607646614313126,
0.14234091341495514,
-0.026238439604640007,
0.029044289141893387,
-0.006291612051427364,
0.06829266995191574,
0.007833042182028294,
-0.03946572169661522,
0.038884975016117096,
-0.08693215250968933,
0.011299630627036095,
0.022687776014208794,
0.05939296633005142,
0.04989071935415268,
0.06300871074199677,
-0.019686905667185783,
-0.015089361928403378,
-0.05062459036707878,
0.07703333348035812,
-0.03253527730703354,
-0.048828091472387314,
-0.0006406711763702333,
0.005810903385281563,
-0.021425168961286545,
-0.04791142791509628,
0.03288824483752251,
-0.0030134597327560186,
-0.01278773508965969,
-0.006169317755848169,
0.08411753922700882,
0.03822891414165497,
0.0911715030670166,
-0.024546075612306595,
-0.052123893052339554,
-0.07186879217624664,
-0.11976409703493118,
0.022804364562034607,
0.05919927358627319,
0.07762640714645386,
-0.06551699340343475,
0.04057955741882324,
0.036270368844270706,
0.00010590728925308213,
0.034508515149354935,
-0.035405371338129044,
-0.06175093352794647,
0.05274783447384834,
-0.11855597048997879,
0.02727087028324604,
-0.00003604095400078222,
-0.06873814016580582,
-0.014414509758353233,
-0.004420030862092972,
-0.03907995671033859,
0.03722938522696495,
0.016641421243548393,
0.007970858365297318,
3.5274910388726926e-33,
0.030054058879613876,
0.060842521488666534,
0.04047943279147148,
-0.051777347922325134,
-0.0008652429678477347,
-0.00014224016922526062,
0.04451455920934677,
0.03408908471465111,
-0.10144324600696564,
0.008579729124903679,
0.026470692828297615,
-0.056218471378088,
-0.053132541477680206,
0.07093328982591629,
0.04927772656083107,
-0.06976842135190964,
0.00793012697249651,
0.11288142204284668,
-0.05431806668639183,
-0.004671494010835886,
0.09076827019453049,
0.08232233673334122,
0.0017543004360049963,
-0.031065214425325394,
-0.020382508635520935,
0.023806612938642502,
-0.010197199881076813,
-0.016345759853720665,
-0.051302120089530945,
0.005666801705956459,
-0.08495310693979263,
-0.06659598648548126,
-0.00298744416795671,
0.04922565817832947,
0.051544830203056335,
0.00201717228628695,
-0.038668740540742874,
-0.050050970166921616,
-0.029158424586057663,
-0.006744610145688057,
-0.006093035917729139,
0.07851644605398178,
0.07994838058948517,
-0.01938444934785366,
-0.08460375666618347,
-0.022830316796898842,
0.017232680693268776,
-0.0033528234343975782,
0.06391303986310959,
-0.10740296542644501,
-0.002063908614218235,
0.023637384176254272,
-0.008868577890098095,
0.08445370942354202,
0.005618325434625149,
-0.001368030090816319,
-0.032115112990140915,
0.04284895583987236,
0.05817194655537605,
0.047315776348114014,
-0.0292830653488636,
0.0331517793238163,
-0.02544355019927025,
0.023620886728167534,
0.05633870139718056,
0.011631687171757221,
0.025935139507055283,
-0.08054645359516144,
0.09732655435800552,
-0.01763574406504631,
0.05375880375504494,
0.019581850618124008,
-0.06145801022648811,
-0.04216337576508522,
0.007892969995737076,
0.0340597927570343,
0.03512899950146675,
-0.025487631559371948,
0.04782538115978241,
0.027111703529953957,
0.008971246890723705,
0.02498646080493927,
-0.010987025685608387,
-0.07127200812101364,
-0.05116310715675354,
-0.08637651801109314,
0.0702928677201271,
-0.02868260256946087,
-0.03852732107043266,
-0.0005369396531023085,
-0.01092291809618473,
0.015034638345241547,
-0.042951758950948715,
-0.02306469716131687,
0.05871887132525444,
-3.244881625783875e-33,
-0.08266609162092209,
-0.030471989884972572,
0.10960641503334045,
0.10613049566745758,
-0.04795277863740921,
-0.06573311239480972,
-0.03193129226565361,
-0.04233751446008682,
0.03267493098974228,
-0.07418742030858994,
-0.0477900467813015,
-0.047305457293987274,
0.021832648664712906,
-0.028038837015628815,
0.06938386708498001,
0.04548688977956772,
-0.03155698999762535,
0.06129009649157524,
0.02887081913650036,
-0.028617404401302338,
0.05669626221060753,
0.02880829945206642,
-0.07150396704673767,
0.05775289237499237,
-0.005702633876353502,
-0.0008919022511690855,
-0.001073588035069406,
0.1012943908572197,
-0.016738586127758026,
-0.06584520637989044,
-0.0816967785358429,
-0.005296613089740276,
-0.13852307200431824,
-0.0031907178927212954,
-0.06591183692216873,
0.046087637543678284,
-0.07284646481275558,
0.09344544261693954,
0.023826900869607925,
0.15356434881687164,
0.010048303753137589,
-0.031255390495061874,
-0.07073115557432175,
-0.04243861138820648,
0.00700167752802372,
-0.03237007558345795,
-0.0823655053973198,
-0.01624927669763565,
0.02187330089509487,
-0.009110496379435062,
0.010703355073928833,
0.09777765721082687,
-0.019025078043341637,
0.07315225899219513,
-0.03724559769034386,
-0.09320148080587387,
0.0004573932383209467,
-0.027482282370328903,
0.061500184237957,
0.021112166345119476,
-0.009194329380989075,
0.04980335012078285,
-0.02843168005347252,
0.0030605425126850605,
0.021913912147283554,
-0.10837017744779587,
0.03274446725845337,
-0.007249640766531229,
0.0005578990094363689,
-0.06364671885967255,
-0.006840445566922426,
-0.10881518572568893,
0.02491428144276142,
0.04243793711066246,
-0.0453307181596756,
-0.03239939734339714,
0.027447164058685303,
-0.0065090288408100605,
0.0031588103156536818,
-0.03166252374649048,
0.0130047257989645,
0.09146545827388763,
-0.013461299240589142,
0.018459167331457138,
0.056321099400520325,
0.09196338802576065,
0.029031936079263687,
0.049616407603025436,
-0.0022155572660267353,
-0.04468412697315216,
-0.0036913950461894274,
-0.024669237434864044,
0.09513309597969055,
0.044283170253038406,
-0.09994790703058243,
-5.059085950165354e-8,
-0.0720507875084877,
-0.02270941436290741,
0.02600654773414135,
0.13371235132217407,
0.025899698957800865,
-0.035708874464035034,
0.031915415078401566,
0.06336865574121475,
-0.06773781031370163,
-0.02726714313030243,
0.053014516830444336,
0.04554951190948486,
-0.024023614823818207,
-0.08498487621545792,
-0.02187802642583847,
0.02899775840342045,
-0.023111458867788315,
0.030227700248360634,
-0.03070373460650444,
-0.005830247886478901,
0.03145872429013252,
0.05649423599243164,
0.03426793962717056,
-0.032082051038742065,
0.043110087513923645,
0.023409120738506317,
0.041086047887802124,
0.0645231232047081,
0.11556995660066605,
0.028631776571273804,
-0.03650563582777977,
0.01836976408958435,
-0.04025573283433914,
-0.027231309562921524,
0.021149013191461563,
0.03356827050447464,
-0.024499522522091866,
-0.02116404101252556,
-0.0491311252117157,
-0.07461969554424286,
0.0341823510825634,
0.022963713854551315,
-0.06860555708408356,
-0.018862107768654823,
0.03807644546031952,
0.013457121327519417,
-0.02206818200647831,
-0.06845937669277191,
0.059138935059309006,
-0.06300384551286697,
-0.05122337490320206,
0.0059170834720134735,
0.0171872079372406,
0.035608019679784775,
0.0533057376742363,
-0.04635528847575188,
-0.015678269788622856,
-0.036045148968696594,
0.056145261973142624,
0.07021435350179672,
0.07753623276948929,
0.004951080773025751,
-0.10952721536159515,
0.01158086210489273
] |
Near-Optimal Policy Identification in Robust Constrained Markov Decision Processes via Epigraph Form | https://openreview.net/forum?id=G5sPv4KSjR | [
"Toshinori Kitamura",
"Tadashi Kozuno",
"Wataru Kumagai",
"Kenta Hoshino",
"Yohei Hosoe",
"Kazumi Kasaura",
"Masashi Hamaya",
"Paavo Parmas",
"Yutaka Matsuo"
] | Poster | Designing a safe policy for uncertain environments is crucial in real-world control systems. However, this challenge remains inadequately addressed within the Markov decision process (MDP) framework. This paper presents the first algorithm guaranteed to identify a near-optimal policy in a robust constrained MDP (RCMDP), where an optimal policy minimizes cumulative cost while satisfying constraints in the worst-case scenario across a set of environments. We first prove that the conventional policy gradient approach to the Lagrangian max-min formulation can become trapped in suboptimal solutions. This occurs when its inner minimization encounters a sum of conflicting gradients from the objective and constraint functions. To address this, we leverage the epigraph form of the RCMDP problem, which resolves the conflict by selecting a single gradient from either the objective or the constraints. Building on the epigraph form, we propose a bisection search algorithm with a policy gradient subroutine and prove that it identifies an $\varepsilon$-optimal policy in an RCMDP with $\widetilde{\mathcal{O}}(\varepsilon^{-4})$ robust policy evaluations. | Markov Decision Process, Constrained Optimization, Robust Optimization | We propose the first algorithm to identify a near-optimal policy in a robust constrained MDP. | 1,733 | 2408.16286 | [
-0.03317617252469063,
0.022462842985987663,
0.0007726636831648648,
-0.07320500910282135,
0.08136142790317535,
0.012988096103072166,
-0.0020899572409689426,
0.013169018551707268,
0.0051155127584934235,
0.07163065671920776,
-0.10868999361991882,
-0.018759911879897118,
0.10151547193527222,
0.01658514514565468,
0.0005475171492435038,
0.030480077490210533,
0.0541628859937191,
-0.01692967489361763,
-0.021418770775198936,
0.02952129952609539,
0.005887045990675688,
0.04176907241344452,
0.05568036437034607,
0.0011604779865592718,
-0.022394418716430664,
-0.005575109273195267,
0.046704504638910294,
0.02398388646543026,
-0.01042749173939228,
0.020950889214873314,
0.041203465312719345,
-0.06980631500482559,
0.01846296712756157,
-0.10928597301244736,
0.03866283968091011,
0.10879265516996384,
-0.10604855418205261,
0.0017749386606737971,
-0.015794826671481133,
0.006243655923753977,
-0.008523814380168915,
0.0752190500497818,
-0.04667362570762634,
-0.02139834314584732,
0.013811920769512653,
-0.004793898202478886,
0.012038400396704674,
-0.0032979047391563654,
-0.010399281047284603,
-0.08292823284864426,
0.015579087659716606,
-0.005941019393503666,
-0.044786423444747925,
-0.014816677197813988,
0.008677723817527294,
0.008389776572585106,
0.04903561994433403,
-0.02208572067320347,
0.025582360103726387,
0.00583392009139061,
-0.02202746830880642,
-0.14156369864940643,
-0.01986360177397728,
-0.01262033823877573,
-0.0047752526588737965,
-0.02163158543407917,
0.008118387311697006,
0.011555843986570835,
0.04257053881883621,
0.11343269795179367,
-0.07099128514528275,
0.04741119220852852,
-0.0031900035683065653,
0.013886213302612305,
0.050634127110242844,
-0.01229812577366829,
0.09351971000432968,
-0.0004923518863506615,
0.018316201865673065,
-0.02380276285111904,
0.005498002283275127,
0.046869345009326935,
-0.01916304975748062,
0.013307128101587296,
0.09046954661607742,
-0.07812757790088654,
-0.0181986503303051,
-0.027787331491708755,
0.23793719708919525,
-0.010411448776721954,
-0.00639580562710762,
-0.0038662366569042206,
-0.010354800149798393,
0.00921093113720417,
0.05821569636464119,
-0.01816040463745594,
0.02888909913599491,
-0.015106095001101494,
-0.0018022808944806457,
0.03934001550078392,
0.014830024912953377,
-0.01887447014451027,
-0.013903646729886532,
0.007178983185440302,
0.061708685010671616,
-0.054320078343153,
-0.050223104655742645,
0.00048752521979622543,
0.021257484331727028,
-0.07605632394552231,
0.04962983354926109,
-0.03974727541208267,
0.1005876287817955,
0.007510388735681772,
-0.0472748726606369,
-0.00479881139472127,
0.03450453653931618,
-0.06570488959550858,
0.03810909762978554,
0.08201319724321365,
-0.10177526623010635,
-0.07882700115442276,
0.01941991224884987,
0.03160248324275017,
0.029763150960206985,
0.0407196581363678,
-0.004780750256031752,
1.0832614824724207e-33,
0.010788331739604473,
-0.022560622543096542,
0.027396822348237038,
-0.08194836974143982,
0.021815631538629532,
0.03759755194187164,
0.0004993550828658044,
-0.0178738534450531,
-0.05413500964641571,
0.016507845371961594,
0.056521885097026825,
-0.027060650289058685,
-0.058627769351005554,
0.01803121343255043,
0.10646896809339523,
-0.04486647993326187,
-0.0539361946284771,
0.07445146888494492,
-0.014455697499215603,
-0.08214084059000015,
0.008745585568249226,
-0.07479481399059296,
-0.03698104992508888,
-0.06120298430323601,
0.04754779115319252,
0.02764222025871277,
0.06409915536642075,
-0.009558841586112976,
0.03355078771710396,
0.0050587523728609085,
-0.06495928764343262,
0.011263584718108177,
-0.02436755783855915,
0.06986834853887558,
0.03196684271097183,
0.08427281677722931,
-0.09600818902254105,
-0.033392567187547684,
-0.0411003939807415,
-0.022546079009771347,
-0.028220999985933304,
0.026907140389084816,
0.037971481680870056,
0.02265869453549385,
-0.0488080233335495,
-0.09157709032297134,
-0.012744705192744732,
0.011637017130851746,
-0.02236957661807537,
-0.04942699894309044,
-0.055905893445014954,
0.03325358033180237,
0.006156043615192175,
-0.09062182158231735,
-0.10545103251934052,
-0.07033206522464752,
0.0011598264100030065,
0.024125970900058746,
-0.007262905593961477,
0.02961880899965763,
0.02704232558608055,
-0.013346362859010696,
0.0864654928445816,
0.042950812727212906,
0.029406530782580376,
0.07068484276533127,
-0.10849887132644653,
0.005360924173146486,
0.07518842071294785,
-0.030738821253180504,
-0.035205963999032974,
0.0010155894560739398,
0.12752941250801086,
-0.09655346721410751,
0.06762005388736725,
-0.002535489620640874,
0.055536575615406036,
0.007542802952229977,
-0.0008816676563583314,
-0.00014676820137538016,
-0.06569085270166397,
0.02957908622920513,
-0.04696672037243843,
0.00838952511548996,
-0.000936182274017483,
-0.04477548599243164,
-0.02346568927168846,
-0.018291642889380455,
-0.03288362920284271,
0.02511240728199482,
-0.11227583885192871,
-0.024484001100063324,
-0.00476082181558013,
0.06036605313420296,
0.011250771582126617,
-2.4320809141200198e-33,
0.063186876475811,
-0.03864424303174019,
0.06578239798545837,
-0.017400266602635384,
-0.012944389134645462,
-0.06617164611816406,
-0.03520474210381508,
-0.03979673236608505,
-0.02419605478644371,
-0.0364309661090374,
-0.10153805464506149,
-0.023488515987992287,
0.05290048569440842,
-0.006735589355230331,
-0.04280297830700874,
-0.037990085780620575,
0.023201778531074524,
0.0249515101313591,
0.029658464714884758,
0.05109240114688873,
0.043024659156799316,
0.029063396155834198,
-0.07595506310462952,
0.07373710721731186,
0.03435783088207245,
0.041532061994075775,
-0.02968847006559372,
0.017279790714383125,
-0.03178912773728371,
0.02085958607494831,
-0.11741191148757935,
0.0016620990354567766,
-0.09500506520271301,
0.05831960588693619,
-0.048440299928188324,
0.022428827360272408,
0.030699122697114944,
-0.010681414976716042,
-0.02766815572977066,
0.10263534635305405,
0.009114954620599747,
-0.0640137568116188,
-0.0035441196523606777,
0.026442864909768105,
0.045532818883657455,
-0.04978169500827789,
0.0021588497329503298,
0.04318654537200928,
-0.02018601819872856,
-0.00887448899447918,
-0.0949842557311058,
0.012849850580096245,
-0.08267175406217575,
0.0715196281671524,
-0.06789328902959824,
-0.08382809907197952,
-0.032201047986745834,
0.031242087483406067,
0.05737443268299103,
-0.007758618798106909,
-0.0018825428560376167,
0.06288284063339233,
-0.09263209998607635,
0.08985090255737305,
0.013237692415714264,
0.07974247634410858,
-0.02855564095079899,
-0.06155946105718613,
0.08495262265205383,
-0.014793620444834232,
0.034071557223796844,
0.02103487402200699,
0.05976342782378197,
0.00918765552341938,
0.05395517498254776,
0.07542768120765686,
0.052816856652498245,
-0.042369991540908813,
-0.04011999070644379,
0.014509494416415691,
0.019514737650752068,
0.0730997696518898,
0.047207556664943695,
0.10432150959968567,
-0.05764466151595116,
0.040867164731025696,
-0.019913967698812485,
0.003046808997169137,
-0.014990035444498062,
-0.03200985863804817,
-0.05982998013496399,
-0.10465121269226074,
0.04712230712175369,
-0.03188909590244293,
-0.06684452295303345,
-4.679053589029536e-8,
-0.07000094652175903,
0.029399849474430084,
0.05999886244535446,
-0.06002076342701912,
0.06543973833322525,
0.0068540023639798164,
-0.019607430323958397,
-0.09438474476337433,
0.025013074278831482,
-0.005743118934333324,
0.0657276064157486,
0.03301938995718956,
0.04734699800610542,
-0.0590205192565918,
-0.11456328630447388,
0.02758934162557125,
-0.01671474799513817,
0.019589724019169807,
0.0034493906423449516,
0.02074711211025715,
-0.08375253528356552,
0.0071159047074615955,
0.009968613274395466,
0.03646086901426315,
0.07533276826143265,
0.018227780237793922,
-0.034979794174432755,
0.05193445086479187,
0.0836711972951889,
0.05255598947405815,
-0.06362799555063248,
-0.015314430929720402,
0.09012667834758759,
0.10313732922077179,
0.015006250701844692,
0.06802495568990707,
-0.01725797913968563,
0.010201281867921352,
-0.06125517934560776,
-0.07852794229984283,
-0.011788930743932724,
0.05994652956724167,
-0.042916033416986465,
-0.018598398193717003,
-0.00086387456394732,
0.0889323279261589,
-0.03767162188887596,
-0.04640699550509453,
0.04242328926920891,
-0.03440948575735092,
0.013754241168498993,
-0.02007164992392063,
0.04434395581483841,
0.03948190435767174,
0.013012158684432507,
-0.00332260993309319,
-0.0034784884192049503,
-0.111786387860775,
0.023440588265657425,
0.05932197719812393,
-0.03137895464897156,
0.05319572985172272,
-0.039340484887361526,
-0.053627148270606995
] |
Scaling Autonomous Agents via Automatic Reward Modeling And Planning | https://openreview.net/forum?id=womU9cEwcO | [
"Zhenfang Chen",
"Delin Chen",
"Rui Sun",
"Wenjun Liu",
"Chuang Gan"
] | Poster | Large language models (LLMs) have demonstrated remarkable capabilities across a range of text-generation tasks. However, LLMs still struggle with problems requiring multi-step decision-making and environmental feedback, such as online shopping, scientific reasoning, and mathematical problem-solving. Unlike pure text data, collecting large-scale decision-making data is challenging. Moreover, many powerful LLMs are only accessible through APIs, which hinders their fine-tuning for agent tasks due to cost and complexity. To address LLM agents' limitations, we propose a framework that can automatically learn a reward model from the environment without human annotations. This model can be used to evaluate the action trajectories of LLM agents and provide heuristics for task planning. Specifically, our approach involves employing one LLM-based agent to navigate an environment randomly, generating diverse action trajectories. Subsequently, a separate LLM is leveraged to assign a task intent and synthesize a negative response alongside the correct response for each trajectory. These triplets (task intent, positive response, and negative response) are then utilized as training data to optimize a reward model capable of scoring action trajectories. This reward model can be integrated with LLM-based agents and various planning algorithms to enhance task-solving performance. The effectiveness and generalizability of our framework are demonstrated through evaluations conducted on different agent benchmarks. In conclusion, our proposed framework represents a significant advancement in enhancing LLM agents' decision-making capabilities. By automating the learning of reward models, we overcome the challenges of data scarcity and API limitations, potentially revolutionizing the application of LLMs in complex and interactive environments. This research paves the way for more sophisticated AI agents capable of tackling a wide range of real-world problems requiring multi-step decision-making. | agents, large language models, planning | null | 1,719 | 2502.12130 | [
-0.010054259561002254,
-0.04630659520626068,
0.0010539565701037645,
0.012326156720519066,
0.011284153908491135,
0.019421406090259552,
0.012955035082995892,
0.0046969931572675705,
0.0416075699031353,
0.018923528492450714,
-0.042701371014118195,
-0.12687364220619202,
0.07523278892040253,
0.11441678553819656,
0.046884581446647644,
0.018258094787597656,
0.08847156912088394,
0.027636557817459106,
-0.058106809854507446,
-0.039039529860019684,
0.03858347237110138,
0.06003459542989731,
0.0627753883600235,
0.011405027471482754,
-0.03657655790448189,
-0.0044590481556952,
0.020966293290257454,
-0.06556549668312073,
0.0021311044692993164,
-0.006955035496503115,
0.0792669802904129,
-0.01428129244595766,
0.005554220639169216,
0.0025914297439157963,
-0.04365919530391693,
0.05550093576312065,
-0.14723995327949524,
-0.04409090057015419,
-0.03903831169009209,
-0.03838968649506569,
-0.02635692059993744,
-0.028930500149726868,
-0.00659957667812705,
0.02764788083732128,
0.10424554347991943,
-0.042774926871061325,
-0.04031796380877495,
-0.030886417254805565,
0.02276388742029667,
0.052053727209568024,
-0.03605441004037857,
-0.05779663845896721,
0.018305940553545952,
-0.006715272553265095,
-0.08981899172067642,
0.02999001555144787,
0.01918594166636467,
0.007578520104289055,
0.02716073952615261,
-0.11664414405822754,
0.008593182079494,
-0.1564202457666397,
-0.025398705154657364,
-0.03804194927215576,
-0.04767898470163345,
-0.039698097854852676,
-0.019418863579630852,
0.037680838257074356,
-0.024704746901988983,
0.0033883445430547,
-0.02316047064960003,
0.00628163805231452,
0.06557139754295349,
0.009643147699534893,
0.028709776699543,
0.04096609726548195,
-0.06648749858140945,
-0.03231365233659744,
0.035770125687122345,
-0.014111557975411415,
-0.04854883253574371,
-0.03481287136673927,
0.013113757595419884,
0.08375231921672821,
0.12032215297222137,
0.0058541083708405495,
0.05771108344197273,
0.08923671394586563,
0.08113020658493042,
0.030282607302069664,
-0.04728994145989418,
-0.06665296852588654,
0.07099775969982147,
-0.01955268904566765,
-0.05175343155860901,
0.09113958477973938,
-0.09624506533145905,
-0.12334900349378586,
-0.03553817793726921,
0.08253028243780136,
0.04529561474919319,
0.11452878266572952,
-0.025227965787053108,
-0.015889404341578484,
0.02282349020242691,
0.03499539941549301,
0.006084409076720476,
0.016069823876023293,
0.05753730610013008,
-0.11853964626789093,
-0.034804586321115494,
0.005039536394178867,
0.07091625034809113,
0.06404992938041687,
0.07899844646453857,
0.007119067944586277,
0.06374671310186386,
0.015261502005159855,
0.01625862717628479,
0.0756901279091835,
0.01694442331790924,
-0.03433288261294365,
0.04145193472504616,
-0.004533800762146711,
0.02143382653594017,
0.018462883308529854,
0.011074656620621681,
-1.5050642810901004e-34,
0.06602707505226135,
0.05072944611310959,
0.09365715831518173,
0.05950421467423439,
0.08822578936815262,
0.021266527473926544,
0.03540867194533348,
0.0367688313126564,
-0.03007005713880062,
-0.04753001406788826,
-0.00235777092166245,
0.029889173805713654,
-0.015983987599611282,
0.10866067558526993,
0.07282513380050659,
-0.030908191576600075,
-0.04414728656411171,
0.0232289656996727,
0.0022891180124133825,
-0.05185539275407791,
0.029758071526885033,
-0.016119565814733505,
0.052568674087524414,
-0.05612197890877724,
0.0740119218826294,
-0.0005828767898492515,
0.06115876883268356,
-0.03530514985322952,
-0.05554746463894844,
-0.036393456161022186,
-0.055976033210754395,
-0.023532144725322723,
0.02245595119893551,
0.04005208984017372,
0.004212164785712957,
-0.0010645646834746003,
-0.07198032736778259,
-0.009807913564145565,
0.028970681130886078,
-0.03763454407453537,
-0.030397741124033928,
0.04725755378603935,
0.03451187536120415,
-0.007031232118606567,
-0.07763449102640152,
-0.0326043963432312,
0.023270253092050552,
0.030365213751792908,
-0.030259672552347183,
-0.009022761136293411,
0.08136791735887527,
0.026509378105401993,
0.021404094994068146,
-0.07438167929649353,
-0.024789292365312576,
-0.0013693086802959442,
-0.01817554235458374,
0.03858150914311409,
0.0353214293718338,
-0.017731735482811928,
0.01655237376689911,
-0.05642084404826164,
0.013939856551587582,
0.09772992134094238,
0.053068943321704865,
0.060370296239852905,
-0.01844220981001854,
0.09910176694393158,
0.06406929343938828,
-0.004027656279504299,
0.044964633882045746,
0.06366565823554993,
0.04882798716425896,
-0.06306899338960648,
0.010253618471324444,
0.006109150592237711,
0.0607287622988224,
-0.167068213224411,
0.012230074033141136,
-0.013699859380722046,
-0.038104839622974396,
-0.054723743349313736,
0.00012371923367027193,
-0.059466663748025894,
0.005369895603507757,
-0.056982945650815964,
0.0073345014825463295,
-0.1314363330602646,
-0.006144373212009668,
0.001778669422492385,
-0.08900640159845352,
0.013684691861271858,
-0.07488662749528885,
0.019456496462225914,
0.0262178722769022,
-1.8947017430961917e-34,
-0.03273748606443405,
-0.0924198180437088,
-0.06922312080860138,
0.017640043050050735,
0.014723599888384342,
-0.038163963705301285,
-0.07316633313894272,
0.004573134705424309,
-0.0018678285414353013,
-0.024676719680428505,
-0.10984312742948532,
0.051135193556547165,
0.049347806721925735,
-0.024236083030700684,
0.019255897030234337,
-0.0370657853782177,
0.04221994802355766,
0.03588029742240906,
0.02876720391213894,
0.019531434401869774,
-0.051030565053224564,
0.1285993605852127,
-0.1492307037115097,
0.007455501239746809,
0.0052223508246243,
0.03487122803926468,
0.03042881190776825,
0.014176984317600727,
-0.01386263594031334,
0.08265594393014908,
-0.06120985746383667,
-0.0018675217870622873,
-0.0019125074613839388,
0.001162833534181118,
-0.04732861742377281,
0.05514874309301376,
0.04615996032953262,
-0.013372965157032013,
-0.04884663596749306,
0.08060936629772186,
0.06317637115716934,
-0.08014500886201859,
0.0031972499564290047,
0.008491150103509426,
-0.025772055611014366,
0.015860652551054955,
-0.12759514153003693,
0.046786051243543625,
-0.004036119673401117,
0.006903830915689468,
-0.03361247107386589,
0.0071786209009587765,
-0.10108046233654022,
-0.020234864205121994,
-0.04872884973883629,
-0.05512721836566925,
-0.03561881557106972,
-0.06488639861345291,
0.036626558750867844,
-0.05835971608757973,
-0.08827833831310272,
0.03431776538491249,
0.050709813833236694,
-0.008610526099801064,
-0.06002374738454819,
-0.03594100847840309,
-0.012614022940397263,
0.012132201343774796,
-0.02583005465567112,
0.009555268101394176,
0.05693807825446129,
0.007010837085545063,
0.02161337248980999,
-0.01996038667857647,
0.04048548638820648,
-0.04878726974129677,
-0.014398491941392422,
-0.002881871536374092,
-0.00879472866654396,
-0.10230791568756104,
0.03423648700118065,
0.05633570998907089,
0.02791137620806694,
0.04335799068212509,
0.02064470760524273,
0.00615747319534421,
0.010444924235343933,
0.103333979845047,
-0.0034562996588647366,
0.0558101087808609,
0.006379608530551195,
0.049412187188863754,
-0.006462853401899338,
0.06052907556295395,
-0.044556356966495514,
-3.9092306280963385e-8,
-0.041701480746269226,
0.04905863478779793,
-0.0014785878593102098,
0.07867060601711273,
-0.02287355624139309,
-0.0006930245435796678,
-0.030353711917996407,
0.052448004484176636,
0.027359284460544586,
0.00623950082808733,
0.0789678618311882,
0.0010243573924526572,
-0.025451580062508583,
-0.02689608745276928,
-0.026794344186782837,
0.05252312496304512,
0.060619890689849854,
-0.03683006018400192,
-0.033885158598423004,
-0.024516131728887558,
0.03989707678556442,
0.038032807409763336,
-0.003953521139919758,
-0.009190461598336697,
0.033982958644628525,
-0.05114661902189255,
-0.04667456075549126,
0.07067271322011948,
0.07225780934095383,
-0.019401177763938904,
0.04675730690360069,
-0.03870781883597374,
-0.02272786945104599,
0.020208729431033134,
0.048289667814970016,
0.00314989872276783,
0.013075456954538822,
-0.07163022458553314,
0.01016342081129551,
-0.06490772217512131,
0.013229092583060265,
0.08630963414907455,
0.0038605062291026115,
-0.006079288199543953,
-0.011331329122185707,
0.006379696074873209,
-0.0954933613538742,
-0.11982191354036331,
0.08756465464830399,
-0.054408665746450424,
-0.05662029981613159,
-0.049371387809515,
-0.011555791832506657,
0.039766035974025726,
0.07509846985340118,
0.023902172222733498,
-0.027774082496762276,
-0.0361064188182354,
0.051181234419345856,
0.03800265118479729,
0.036495331674814224,
-0.011873464100062847,
-0.06983137875795364,
-0.0368570052087307
] |
Generalized Video Moment Retrieval | https://openreview.net/forum?id=qdOIkeZ5e4 | [
"You Qin",
"Qilong Wu",
"Yicong Li",
"Wei Ji",
"Li Li",
"Pengcheng Cai",
"Lina Wei",
"Roger Zimmermann"
] | Poster | In this paper, we introduce the Generalized Video Moment Retrieval (GVMR) framework, which extends traditional Video Moment Retrieval (VMR) to handle a wider range of query types. Unlike conventional VMR systems, which are often limited to simple, single-target queries, GVMR accommodates both non-target and multi-target queries. To support this expanded task, we present the NExT-VMR dataset, derived from the YFCC100M collection, featuring diverse query scenarios to enable more robust model evaluation.
Additionally, we propose BCANet, a transformer-based model incorporating the novel Boundary-aware Cross Attention (BCA) module. The BCA module enhances boundary detection and uses cross-attention to achieve a comprehensive understanding of video content in relation to queries. BCANet accurately predicts temporal video segments based on natural language descriptions, outperforming traditional models in both accuracy and adaptability. Our results demonstrate the potential of the GVMR framework, the NExT-VMR dataset, and BCANet to advance VMR systems, setting a new standard for future multimedia information retrieval research. | video moment retrieval, diverse query types, model versatility | null | 1,718 | null | [
0.014528290368616581,
-0.10559050738811493,
-0.01733439415693283,
-0.02464951202273369,
0.11868493258953094,
0.06617875397205353,
-0.021934233605861664,
0.01135076954960823,
0.06513924896717072,
-0.13646262884140015,
-0.018349697813391685,
-0.06883260607719421,
-0.03533288091421127,
0.07090052962303162,
0.030128171667456627,
-0.01721087656915188,
0.06828051805496216,
0.03711720556020737,
-0.07495283335447311,
-0.014821747317910194,
0.04389193281531334,
0.0025274022482335567,
0.0845881849527359,
-0.017646871507167816,
-0.036538369953632355,
-0.006967619992792606,
-0.005967350210994482,
0.006100906524807215,
0.03342701867222786,
-0.020023662596940994,
0.0543985553085804,
0.015063956379890442,
0.010008740238845348,
0.11761987209320068,
-0.0728297308087349,
-0.015780555084347725,
-0.12481524795293808,
-0.0176978949457407,
-0.030757630243897438,
-0.07885468006134033,
-0.02238992042839527,
0.007627684623003006,
0.022361528128385544,
-0.0025398724246770144,
0.057540684938430786,
0.04862510412931442,
-0.03299996256828308,
-0.04753752797842026,
0.07085832208395004,
0.005983317736536264,
-0.04439606890082359,
-0.04955332726240158,
-0.02916881814599037,
0.0707811713218689,
-0.024486219510436058,
0.05115370452404022,
-0.04359274357557297,
-0.03799087181687355,
0.056166090071201324,
-0.014422700740396976,
-0.013964885845780373,
-0.08046708256006241,
0.059896815568208694,
-0.018552729859948158,
0.012691670097410679,
-0.05926474928855896,
0.08072515577077866,
-0.002413664013147354,
-0.033987466245889664,
-0.08379165083169937,
-0.120784692466259,
0.014383966103196144,
-0.0038690129294991493,
0.05334920436143875,
0.009610062465071678,
0.06531666219234467,
-0.01417594775557518,
-0.010039880871772766,
0.019136125221848488,
-0.043807901442050934,
0.010482324287295341,
-0.014923187904059887,
0.061363693326711655,
0.023395471274852753,
0.17460200190544128,
-0.03373012691736221,
0.014911293983459473,
0.02083088643848896,
-0.047811657190322876,
0.029522283002734184,
-0.0773383155465126,
-0.08484674990177155,
0.08893806487321854,
0.0027513671666383743,
0.012840857729315758,
0.005412123631685972,
-0.03164763003587723,
-0.04422378912568092,
0.0006326989969238639,
-0.006667715962976217,
0.06229540333151817,
0.009577696211636066,
0.027561377733945847,
-0.03978155925869942,
0.07420995831489563,
-0.009501866064965725,
0.038313958793878555,
0.053652435541152954,
0.0488746203482151,
-0.07224799692630768,
0.00022385804913938046,
0.050413213670253754,
-0.0003221321676392108,
-0.11842559278011322,
0.05435575544834137,
-0.0465579479932785,
-0.03683701902627945,
-0.01596870832145214,
-0.0055281417444348335,
-0.0017473059706389904,
0.03780170530080795,
0.026476923376321793,
-0.07188256084918976,
-0.05097253620624542,
0.05383080989122391,
-0.04857311770319939,
0.012409769929945469,
1.3826466693154105e-33,
-0.005448747891932726,
-0.07294998317956924,
0.029436133801937103,
0.030363647267222404,
0.022645827382802963,
0.04694775491952896,
0.020739777013659477,
0.07795227319002151,
-0.013763505965471268,
-0.025096766650676727,
0.008769397623836994,
-0.039228655397892,
-0.020291047170758247,
-0.006437688134610653,
-0.03073650784790516,
-0.03888653963804245,
-0.11543804407119751,
-0.07618852704763412,
0.03996626287698746,
0.06795485317707062,
-0.009077928960323334,
0.02823529951274395,
0.12245851010084152,
-0.008083333261311054,
-0.0053187040612101555,
-0.012055053375661373,
0.03113844431936741,
-0.08058501034975052,
0.003021996235474944,
-0.006576926913112402,
-0.1274397075176239,
0.00241556903347373,
-0.01396536361426115,
-0.010157433338463306,
0.05915823206305504,
-0.05826708674430847,
-0.055029209703207016,
-0.008640257641673088,
0.015214604325592518,
-0.007584638427942991,
-0.055367544293403625,
0.06404720991849899,
-0.029079530388116837,
-0.03451307490468025,
-0.0926431193947792,
-0.038745176047086716,
0.0564529187977314,
0.06406961381435394,
-0.007071869447827339,
0.04605729877948761,
0.07700637727975845,
-0.002816561609506607,
-0.05662483721971512,
-0.08349044620990753,
0.025851134210824966,
0.05539324879646301,
0.09877926856279373,
0.10673398524522781,
-0.03271937370300293,
0.015912575647234917,
-0.014236359857022762,
0.040337249636650085,
0.03284851834177971,
0.03483591973781586,
-0.005018950905650854,
-0.024347631260752678,
0.013476739637553692,
0.14197713136672974,
-0.006155754439532757,
0.06796468794345856,
0.04799970984458923,
0.015808312222361565,
0.07013995945453644,
-0.045216623693704605,
0.021896377205848694,
0.02041471004486084,
0.0044138734228909016,
-0.051628392189741135,
0.06282408535480499,
0.006981252692639828,
0.007846485823392868,
0.002401888370513916,
0.0544787161052227,
-0.04609714448451996,
-0.0537494532763958,
-0.00144265650305897,
0.06702963262796402,
-0.1082594245672226,
-0.027152758091688156,
-0.07558479905128479,
-0.0185224749147892,
0.03033328428864479,
0.00046218460192903876,
0.05268409475684166,
-0.029138192534446716,
8.476835868805661e-35,
0.07236255705356598,
0.07889164239168167,
-0.01339894998818636,
0.011567428708076477,
0.01421347726136446,
-0.04155375808477402,
0.047338031232357025,
0.016030477359890938,
-0.03243667259812355,
-0.03009123168885708,
0.03153666481375694,
-0.08577752113342285,
-0.008047560229897499,
-0.0021442093420773745,
0.003016649279743433,
0.054123010486364365,
0.009334743022918701,
-0.04619617015123367,
-0.13242994248867035,
-0.004694669507443905,
-0.016466591507196426,
0.036105699837207794,
-0.06803515553474426,
-0.01604112982749939,
-0.06933552771806717,
0.014581725932657719,
-0.059386033564805984,
0.01276474166661501,
0.04033719375729561,
-0.0652751699090004,
-0.028794705867767334,
-0.003949345555156469,
0.0049902345053851604,
-0.08065836876630783,
-0.046437639743089676,
-0.019861437380313873,
0.06275827437639236,
-0.09300651401281357,
0.013156946748495102,
0.06910215318202972,
0.06807763874530792,
0.024074457585811615,
-0.05666137486696243,
0.05071890726685524,
-0.02705942839384079,
0.03266249597072601,
-0.09581347554922104,
0.053850531578063965,
0.004142663441598415,
0.05424752086400986,
-0.07200168073177338,
-0.002430835971608758,
-0.06674302369356155,
-0.003559016389772296,
-0.007500969804823399,
-0.05936051160097122,
-0.0793246403336525,
-0.02849486842751503,
0.04872886463999748,
-0.030246540904045105,
0.012418526224792004,
0.025945276021957397,
0.015232491306960583,
-0.11703994125127792,
0.04696618765592575,
0.02951216511428356,
0.02414556033909321,
-0.05950043722987175,
-0.024242175742983818,
0.00981126632541418,
0.07070427387952805,
-0.017981793731451035,
-0.01320237573236227,
-0.0401555560529232,
0.0080623934045434,
0.04963387921452522,
0.033976394683122635,
-0.07679472118616104,
0.019037233665585518,
-0.06789655238389969,
-0.047094523906707764,
-0.05017392709851265,
0.030869998037815094,
0.06196067854762077,
0.038571905344724655,
0.1142706573009491,
0.08634822815656662,
0.0001247827021870762,
0.05190258100628853,
-0.02195015177130699,
-0.03540338948369026,
0.042945560067892075,
0.02977478876709938,
0.07666891813278198,
0.03828895092010498,
-3.893490685413781e-8,
-0.1340440958738327,
0.002507431199774146,
-0.10096757113933563,
-0.011424126103520393,
-0.007730494253337383,
-0.03980007395148277,
0.005779041443020105,
0.061065103858709335,
0.042055997997522354,
0.00393122062087059,
0.06689511239528656,
-0.04278780147433281,
-0.025780167430639267,
0.06160394847393036,
-0.006501633208245039,
0.009962499141693115,
0.05099852755665779,
0.05037969350814819,
0.028322070837020874,
-0.03861232474446297,
-0.007697029504925013,
0.05305027216672897,
0.08027669787406921,
-0.03785741329193115,
0.027830103412270546,
0.025435930117964745,
-0.06068974360823631,
0.06769829243421555,
0.072644904255867,
-0.06860468536615372,
0.06143910437822342,
0.013544336892664433,
-0.005619824398308992,
-0.08390545845031738,
0.03874052315950394,
0.04894408956170082,
0.019224464893341064,
0.008854449726641178,
0.005318866111338139,
0.023881906643509865,
0.032643187791109085,
-0.0005412230966612697,
0.010166932828724384,
0.0741792619228363,
-0.03501254692673683,
0.032689131796360016,
0.05907302722334862,
-0.03840109705924988,
0.04190758988261223,
-0.003714117221534252,
0.08520225435495377,
-0.06950023025274277,
-0.029470598325133324,
0.07622253894805908,
0.07466080784797668,
-0.0014842289965599775,
0.06681910157203674,
0.004845462739467621,
0.059072207659482956,
-0.015141275711357594,
-0.0007971597369760275,
0.01641886867582798,
-0.0788184404373169,
0.04274178668856621
] |
A Training-Free Sub-quadratic Cost Transformer Model Serving Framework with Hierarchically Pruned Attention | https://openreview.net/forum?id=PTcMzQgKmn | [
"Heejun Lee",
"Geon Park",
"Youngwan Lee",
"Jaduk Suh",
"Jina Kim",
"Wonyong Jeong",
"Bumsik Kim",
"Hyemin Lee",
"Myeongjae Jeon",
"Sung Ju Hwang"
] | Poster | In modern large language models (LLMs), increasing the context length is crucial for improving comprehension and coherence in long-context, multi-modal, and retrieval-augmented language generation.
While many recent transformer models attempt to extend their context length over a million tokens, they remain impractical due to the quadratic time and space complexities.
Although recent works on linear and sparse attention mechanisms can achieve this goal, their real-world applicability is often limited by the need to re-train from scratch and significantly worse performance. In response, we propose a novel approach, Hierarchically Pruned Attention (HiP), which reduces the time complexity of the attention mechanism to $O(T \log T)$ and the space complexity to $O(T)$, where $T$ is the sequence length.
We notice a pattern in the attention scores of pretrained LLMs where tokens close together tend to have similar scores, which we call "attention locality". Based on this observation, we utilize a novel tree-search-like algorithm that estimates the top-$k$ key tokens for a given query on the fly, which is mathematically guaranteed to have better performance than random attention pruning. In addition to improving the time complexity of the attention mechanism, we further optimize GPU memory usage by implementing KV cache offloading, which stores only $O(\log T)$ tokens on the GPU while maintaining similar decoding throughput. Experiments on benchmarks show that HiP, with its training-free nature, significantly reduces both prefill and decoding latencies, as well as memory usage, while maintaining high-quality generation with minimal degradation.
HiP enables pretrained LLMs to scale up to millions of tokens on commodity GPUs, potentially unlocking long-context LLM applications previously deemed infeasible. | Efficient Attention Mechanism, Long-context LLM Decoding, KV Cache Offloading | To response of quadratic scaling (time & GPU memory) problem of LLM decoding framework, we are proposing sub-quadratic time & GPU memory cost HiP attention. | 1,716 | 2406.09827 | [
-0.01524206344038248,
-0.046547580510377884,
0.021663973107933998,
0.03508707135915756,
0.06301353126764297,
0.027103520929813385,
-0.027475573122501373,
0.05831423029303551,
0.0958138182759285,
-0.03700796514749527,
0.016273153945803642,
-0.0070419530384242535,
0.04477990046143532,
0.034506093710660934,
0.004549291450530291,
0.028490770608186722,
0.052485391497612,
0.07227414846420288,
-0.07751365751028061,
-0.08474428206682205,
0.07246014475822449,
-0.011597279459238052,
0.06294766068458557,
-0.038730546832084656,
0.019956115633249283,
-0.0007785478373989463,
-0.04171377420425415,
-0.09501013159751892,
0.011782552115619183,
0.0253099724650383,
0.0158135537058115,
0.022789139300584793,
0.03259984776377678,
0.04178136587142944,
-0.08048424869775772,
0.07831375300884247,
-0.0547381266951561,
-0.009807085618376732,
0.014515778981149197,
-0.036920785903930664,
0.02643693797290325,
0.0380997508764267,
-0.03037181682884693,
0.0386541485786438,
0.08250033855438232,
-0.026928460225462914,
0.06094809249043465,
-0.0265959482640028,
-0.0190106388181448,
-0.04244484379887581,
-0.03506111353635788,
-0.019382141530513763,
-0.033120106905698776,
0.0814778134226799,
-0.032156385481357574,
0.04587521776556969,
-0.015126774087548256,
0.044886790215969086,
-0.0036506676115095615,
-0.007195944897830486,
-0.07080430537462234,
-0.1050734594464302,
-0.05166172981262207,
-0.08338658511638641,
-0.03025214374065399,
-0.021960722282528877,
0.07403486967086792,
-0.01719341054558754,
-0.062215715646743774,
0.01822328381240368,
0.00866522453725338,
0.08065493404865265,
-0.03654075041413307,
0.06728070974349976,
0.07269951701164246,
0.024974660947918892,
0.07372114062309265,
-0.019833043217658997,
0.080734983086586,
-0.03597647696733475,
0.020100487396121025,
-0.0336776077747345,
0.04577340930700302,
-0.02014065347611904,
0.1326017826795578,
0.014164344407618046,
0.05252942070364952,
0.03458937630057335,
-0.02320660464465618,
0.03136181831359863,
-0.07251680642366409,
-0.12798695266246796,
0.007074493914842606,
-0.04488169774413109,
0.028679760172963142,
0.011362443678081036,
0.02722381427884102,
-0.01850990578532219,
-0.0469311885535717,
0.06993410736322403,
0.006293443962931633,
0.055439677089452744,
0.03300373628735542,
-0.020406177267432213,
-0.02873423881828785,
-0.00742519274353981,
-0.03466052561998367,
0.07858618348836899,
-0.002318669343367219,
-0.04703119769692421,
0.015413965098559856,
0.04132690280675888,
0.034270018339157104,
0.03508305549621582,
0.048159193247556686,
0.012547858990728855,
0.02178596891462803,
-0.05768921971321106,
0.016133779659867287,
0.05447752773761749,
0.04356196150183678,
0.030491629615426064,
-0.040022145956754684,
0.03993677720427513,
0.03143390640616417,
-0.003657754510641098,
-0.0303974412381649,
-1.4386983266421962e-34,
0.022014008834958076,
0.11834953725337982,
0.057232968509197235,
-0.012161231599748135,
0.0709095299243927,
-0.020327646285295486,
0.05094994604587555,
0.0718623474240303,
0.0027252952568233013,
0.0013625684659928083,
0.004694825503975153,
0.019737455993890762,
-0.03650354966521263,
0.04789083078503609,
0.021800275892019272,
-0.014548063278198242,
-0.0235797930508852,
0.0355282723903656,
-0.054872073233127594,
-0.023728495463728905,
-0.0240408256649971,
-0.012762792408466339,
0.0002442934492137283,
-0.06212310492992401,
0.02560962736606598,
-0.08616150915622711,
0.03640583157539368,
-0.14551308751106262,
-0.05735540762543678,
-0.013672521337866783,
-0.17157112061977386,
0.03280620276927948,
-0.06340162456035614,
-0.003566785715520382,
0.03466341644525528,
-0.11600889265537262,
0.011267666704952717,
-0.07457060366868973,
0.009816432371735573,
-0.01067547220736742,
-0.030083339661359787,
0.048818666487932205,
0.026597918942570686,
-0.024086903780698776,
-0.05814455822110176,
-0.08776606619358063,
0.007733403239399195,
0.03467536345124245,
-0.01879970356822014,
0.0018394228536635637,
0.06542765349149704,
0.05104812607169151,
-0.06970701366662979,
-0.007279578596353531,
0.014352106489241123,
0.013010498136281967,
0.04990693926811218,
0.06325419992208481,
0.02999843657016754,
0.0869167298078537,
0.07426584511995316,
-0.005015737842768431,
0.028783626854419708,
0.12963984906673431,
0.05516154691576958,
0.0032667953055351973,
-0.03178960457444191,
0.058013785630464554,
0.07049553841352463,
0.011845901608467102,
-0.03232593834400177,
-0.011975692585110664,
-0.013960057869553566,
-0.06485339254140854,
0.06064005568623543,
-0.00023183740267995745,
0.13202446699142456,
-0.08080189675092697,
0.009374922141432762,
-0.00784243457019329,
0.005261698737740517,
-0.04003007337450981,
0.03933718800544739,
-0.07263309508562088,
-0.03434264659881592,
-0.07343675941228867,
0.05716106668114662,
-0.09511490166187286,
-0.03875098004937172,
-0.1060740053653717,
-0.030802011489868164,
0.018073860555887222,
-0.06749631464481354,
-0.04836085066199303,
-0.0008964881999418139,
1.1908094810100672e-33,
0.036562249064445496,
-0.06764435023069382,
-0.005041932221502066,
0.0015281917294487357,
-0.01699863001704216,
-0.05127127841114998,
-0.02527410164475441,
0.05139993503689766,
-0.09258558601140976,
-0.012684992514550686,
-0.05083911120891571,
-0.05161469429731369,
0.09277213364839554,
-0.02848825603723526,
0.08254485577344894,
0.04377944394946098,
0.004296279512345791,
0.05475800484418869,
-0.029457038268446922,
0.032480187714099884,
0.023771075531840324,
0.08081679046154022,
-0.02876153402030468,
0.027018850669264793,
-0.08413930237293243,
-0.013743402436375618,
-0.07909312844276428,
0.03029952012002468,
-0.02688281610608101,
-0.010466713458299637,
-0.07980012148618698,
0.006431251298636198,
0.008669600822031498,
-0.06867411732673645,
-0.11275948584079742,
0.027136370539665222,
0.017429152503609657,
-0.006569857709109783,
-0.04278033226728439,
0.11362769454717636,
0.08535655587911606,
-0.0719001516699791,
-0.03388345241546631,
0.050523802638053894,
-0.06314460188150406,
-0.018536431714892387,
-0.08548340201377869,
-0.041639335453510284,
0.01499238982796669,
0.04788467288017273,
0.018198728561401367,
0.04338619112968445,
-0.05932828038930893,
-0.01672401651740074,
-0.03992796689271927,
-0.08750047534704208,
-0.07272804528474808,
-0.032244972884655,
0.07817869633436203,
-0.08443550020456314,
0.0064980220049619675,
0.04583468660712242,
-0.01605396345257759,
-0.06291288882493973,
0.05254627391695976,
-0.07078947871923447,
0.023694122210144997,
-0.0034105265513062477,
0.013877781108021736,
-0.005301039200276136,
0.011601715348660946,
0.002634823089465499,
-0.002740734489634633,
-0.00752802100032568,
-0.011909140273928642,
0.07386267185211182,
0.0466448999941349,
-0.05114198103547096,
-0.03234676644206047,
-0.07927249372005463,
-0.02446606568992138,
0.03849245607852936,
-0.018801163882017136,
-0.003800158156082034,
0.008755188435316086,
0.07010450959205627,
0.13142997026443481,
0.06835637241601944,
0.04348469898104668,
0.0072660623118281364,
-0.01097175758332014,
0.03173380345106125,
0.038745101541280746,
0.07234182953834534,
-0.01476671639829874,
-4.6122689667527084e-8,
-0.0684615895152092,
0.040276192128658295,
-0.09363643079996109,
0.056678447872400284,
0.02155156247317791,
-0.09245333820581436,
-0.012724323198199272,
0.1598110795021057,
-0.07427340745925903,
0.019816290587186813,
0.020741654559969902,
0.010122782550752163,
-0.010588466189801693,
-0.017590930685400963,
-0.033235885202884674,
0.04504001885652542,
-0.001997606595978141,
-0.0019067019456997514,
-0.022840512916445732,
-0.010455623269081116,
0.01941090263426304,
0.09854545444250107,
0.043094586580991745,
0.05619091913104057,
-0.01100607868283987,
-0.032351475208997726,
0.02635197713971138,
0.0695396363735199,
0.09309440851211548,
-0.029218723997473717,
-0.014960548840463161,
0.006750890985131264,
-0.05759008228778839,
-0.0685562938451767,
-0.04531038925051689,
0.0494290366768837,
0.03617230802774429,
0.009349213913083076,
-0.039893604815006256,
0.0005190991214476526,
0.05030210316181183,
0.0029525028076022863,
-0.01640423573553562,
0.06427706778049469,
0.02644016034901142,
0.004249220713973045,
-0.07555320858955383,
-0.09106244891881943,
0.06952854990959167,
-0.08916571736335754,
0.06112191081047058,
-0.01181331928819418,
-0.017336832359433174,
-0.0016635141801089048,
0.012017780914902687,
0.047602664679288864,
0.0032566888257861137,
0.01614878512918949,
0.03792349621653557,
0.015871547162532806,
0.057390082627534866,
0.02057897485792637,
-0.032520171254873276,
0.022655388340353966
] |
GameGen-X: Interactive Open-world Game Video Generation | https://openreview.net/forum?id=8VG8tpPZhe | [
"Haoxuan Che",
"Xuanhua He",
"Quande Liu",
"Cheng Jin",
"Hao Chen"
] | Poster | We introduce GameGen-$\mathbb{X}$, the first diffusion transformer model specifically designed for both generating and interactively controlling open-world game videos.
This model facilitates high-quality, open-domain generation by approximating various game elements, such as innovative characters, dynamic environments, complex actions, and diverse events.
Additionally, it provides interactive controllability, predicting and altering future content based on the current clip, thus allowing for gameplay simulation.
To realize this vision, we first collected and built an Open-World Video Game Dataset (OGameData) from scratch.
It is the first and largest dataset for open-world game video generation and control, which comprises over one million diverse gameplay video clips with informative captions.
GameGen-$\mathbb{X}$ undergoes a two-stage training process, consisting of pre-training and instruction tuning.
Firstly, the model was pre-trained via text-to-video generation and video continuation, enabling long-sequence open-domain game video generation with improved fidelity and coherence.
Further, to achieve interactive controllability, we designed InstructNet to incorporate game-related multi-modal control signal experts.
This allows the model to adjust latent representations based on user inputs, advancing the integration of character interaction and scene content control in video generation.
During instruction tuning, only the InstructNet is updated while the pre-trained foundation model is frozen, enabling the integration of interactive controllability without loss of diversity and quality of generated content.
GameGen-$\mathbb{X}$ contributes to advancements in open-world game design using generative models.
It demonstrates the potential of generative models to serve as auxiliary tools to traditional rendering techniques, demonstrating the potential for merging creative generation with interactive capabilities.
The project will be available at https://github.com/GameGen-X/GameGen-X. | Open-world Game Video Generation, Interactive Control, Diffusion Transformers | We introduce GameGen-$\mathbb{X}$, the first diffusion transformer model tailored for the generation and controllable interaction of open-world game videos, unifying multi-modal game-related control signals. | 1,714 | null | [
-0.0052308859303593636,
-0.09869079291820526,
0.0025745900347828865,
-0.09096454083919525,
0.07135705649852753,
0.018965279683470726,
-0.013724852353334427,
-0.019198553636670113,
0.05040379613637924,
-0.007616148330271244,
-0.02392716519534588,
-0.08942443132400513,
-0.016205761581659317,
0.001501211547292769,
0.004509812220931053,
0.013041230849921703,
0.05221329256892204,
0.040994200855493546,
-0.09932278841733932,
-0.05331342667341232,
0.08777254819869995,
0.03920293226838112,
0.06751053780317307,
-0.038448579609394073,
-0.01943807490170002,
0.0478990338742733,
0.02751055359840393,
0.020759332925081253,
0.03609625622630119,
-0.0626918375492096,
0.10617207735776901,
-0.0011490160832181573,
-0.005420597270131111,
-0.002048990223556757,
-0.04972527548670769,
0.03812803700566292,
-0.05953945592045784,
-0.030660076066851616,
-0.03904521465301514,
-0.06034751608967781,
0.0007759789004921913,
0.040221069008111954,
0.046983540058135986,
0.05838850885629654,
0.01390653382986784,
0.02208852767944336,
-0.05694125220179558,
-0.08870858699083328,
0.015749521553516388,
0.09053070098161697,
-0.07174549251794815,
-0.07851890474557877,
0.04156571626663208,
0.011956600472331047,
0.039942316710948944,
0.04993320629000664,
0.04678322374820709,
0.011555257253348827,
0.04799087345600128,
-0.03375820443034172,
-0.016283296048641205,
-0.08738858997821808,
0.006945325527340174,
-0.0046540251933038235,
-0.061067547649145126,
-0.07928366214036942,
0.1270740032196045,
0.002017405815422535,
-0.02267543040215969,
-0.07354595512151718,
-0.04813122749328613,
0.04036260396242142,
-0.023411309346556664,
-0.00008609151700511575,
-0.03672690689563751,
0.05509863421320915,
0.02364252880215645,
-0.07139256596565247,
-0.019984213635325432,
-0.032036133110523224,
0.08971286565065384,
-0.02450830303132534,
0.07371893525123596,
0.03040575608611107,
-0.006406686268746853,
0.00880315713584423,
0.0038754944689571857,
0.04832526296377182,
-0.0009175670566037297,
0.06138733774423599,
-0.14300937950611115,
0.04080846905708313,
0.12725397944450378,
-0.008125302381813526,
-0.019948359578847885,
0.043185245245695114,
-0.0025543582160025835,
-0.13910405337810516,
0.010730411857366562,
0.07499658316373825,
-0.035489629954099655,
0.026030447334051132,
0.05028419569134712,
0.06911167502403259,
0.04806378856301308,
-0.03799442946910858,
0.08831138908863068,
0.08536574989557266,
-0.07282816618680954,
-0.01108362153172493,
0.02968304604291916,
0.011314702220261097,
-0.020585665479302406,
-0.01497567631304264,
0.03832484036684036,
0.004432164132595062,
-0.0198721494525671,
0.02820686250925064,
-0.007215287536382675,
0.049579229205846786,
0.05778864771127701,
-0.012436988763511181,
-0.04648425802588463,
0.07844261825084686,
0.037834301590919495,
-0.04026006907224655,
-0.009485076181590557,
2.0387057703791722e-34,
0.06224538013339043,
-0.005646924488246441,
0.10877472162246704,
0.06898152828216553,
0.06394078582525253,
-0.010363385081291199,
0.06143172085285187,
0.01762801967561245,
-0.06834127753973007,
-0.034915659576654434,
-0.025879010558128357,
0.017214354127645493,
-0.11844534426927567,
0.09903183579444885,
0.052833400666713715,
-0.0321408249437809,
-0.08961883187294006,
0.024921411648392677,
0.0030739225912839174,
0.10440082848072052,
0.05217723548412323,
0.014895141124725342,
0.023050492629408836,
-0.04129158332943916,
-0.03595024719834328,
0.011575485579669476,
0.023946696892380714,
-0.045086197555065155,
0.04248179867863655,
-0.01879529282450676,
-0.10082289576530457,
-0.0866236612200737,
-0.027569372206926346,
-0.01754515804350376,
0.07251943647861481,
-0.004859749227762222,
-0.0261155404150486,
-0.07055626064538956,
-0.0214915219694376,
0.0005786384572274983,
-0.0011613527312874794,
0.01735035702586174,
-0.054785653948783875,
0.008812317624688148,
-0.09186524897813797,
-0.058143459260463715,
0.04237998276948929,
0.0017393429297953844,
-0.11968667060136795,
0.0027308908756822348,
0.011945588514208794,
0.04428091272711754,
-0.007234477903693914,
-0.0770534873008728,
-0.03180425614118576,
-0.05001647397875786,
0.06382071226835251,
-0.04832649230957031,
0.0014051685575395823,
0.05029986426234245,
-0.02714962139725685,
0.06011458486318588,
-0.008568358607590199,
0.061190854758024216,
0.0015290979063138366,
0.04042626917362213,
0.05290214344859123,
0.033225227147340775,
0.06214197725057602,
-0.01764446683228016,
-0.038168903440237045,
-0.008559529669582844,
0.026465965434908867,
-0.028105488047003746,
0.05604036897420883,
0.03614949807524681,
0.005952327977865934,
-0.08087440580129623,
-0.01656394638121128,
0.03441481292247772,
0.03925040364265442,
-0.018672281876206398,
-0.055314186960458755,
-0.01663331873714924,
-0.018462281674146652,
-0.07083828002214432,
0.0022144478280097246,
-0.024648888036608696,
-0.060968074947595596,
-0.06467451900243759,
-0.07431501150131226,
-0.028981026262044907,
-0.002203888725489378,
-0.00895683467388153,
-0.009377664886415005,
2.473288901737427e-35,
0.009388635866343975,
0.018366781994700432,
-0.06104926019906998,
0.07044748216867447,
0.02325580082833767,
-0.028240403160452843,
-0.026865357533097267,
0.04339011758565903,
0.0332646518945694,
-0.051390502601861954,
-0.02948744036257267,
0.011130628176033497,
-0.02320607379078865,
-0.045376986265182495,
-0.010055373422801495,
-0.08241237699985504,
0.032081689685583115,
-0.03074108436703682,
-0.08090656995773315,
0.004948440473526716,
0.02905305102467537,
0.10041319578886032,
-0.0690770298242569,
-0.03695124015212059,
-0.04385533183813095,
0.030231526121497154,
0.004892575088888407,
0.0540199913084507,
0.049145907163619995,
0.07043494284152985,
0.011372857727110386,
0.004203736782073975,
0.050565123558044434,
0.01474536769092083,
-0.11736859381198883,
0.03359530121088028,
0.1094379797577858,
-0.004772852174937725,
-0.05982599034905434,
0.0655653104186058,
0.021240776404738426,
-0.07411991059780121,
-0.11191340535879135,
0.07241622358560562,
-0.041419968008995056,
0.06355774402618408,
-0.053782958537340164,
-0.01457089651376009,
-0.0016169344307854772,
0.05157826468348503,
-0.03800816088914871,
0.10234599560499191,
-0.06971433013677597,
-0.09369616210460663,
0.013161033391952515,
-0.14111268520355225,
-0.09177414327859879,
-0.02633325383067131,
0.057662080973386765,
-0.014803260564804077,
0.0247625894844532,
0.01339165773242712,
-0.02513592317700386,
-0.02028587833046913,
0.03462347388267517,
-0.017579039558768272,
-0.024459263309836388,
-0.004598128609359264,
0.010372993536293507,
-0.010873552411794662,
-0.011164854280650616,
0.02612088806927204,
0.022787228226661682,
-0.01349687296897173,
-0.009462416172027588,
-0.034610338509082794,
-0.012500098906457424,
-0.015744179487228394,
0.07230708748102188,
0.0014122269349172711,
0.015625767409801483,
-0.007068559527397156,
0.09366074949502945,
0.0380248948931694,
0.07733987271785736,
0.0798790231347084,
0.04852893203496933,
0.10525272786617279,
-0.012825495563447475,
0.03955145552754402,
-0.008032258599996567,
0.09765567630529404,
-0.005494432989507914,
0.09372779726982117,
-0.04007050767540932,
-4.014580667899281e-8,
-0.08693573623895645,
0.01934284344315529,
0.027269165962934494,
0.03800266981124878,
-0.038698893040418625,
-0.004138248972594738,
0.03376064822077751,
-0.02134489081799984,
0.07630883902311325,
-0.03572049364447594,
0.04741411283612251,
0.05946763604879379,
0.01810789480805397,
-0.017365358769893646,
-0.04711102321743965,
0.04110727459192276,
0.007511148229241371,
0.009767095558345318,
-0.0032011496368795633,
0.09658296406269073,
-0.008547675795853138,
0.05126384645700455,
-0.021558038890361786,
-0.11333317309617996,
0.010632725432515144,
-0.04865843057632446,
-0.056290268898010254,
0.023130979388952255,
0.05464370176196098,
0.02463286556303501,
0.0551762729883194,
-0.02071700431406498,
0.02721024677157402,
-0.06775055080652237,
-0.010341158136725426,
0.03284275904297829,
0.056300584226846695,
-0.008761544711887836,
0.04264429211616516,
-0.06449437141418457,
-0.026464184746146202,
0.03808244690299034,
-0.03265620395541191,
-0.018564373254776,
-0.02417386882007122,
0.028238600119948387,
0.0007761954329907894,
-0.10521342605352402,
0.025668129324913025,
-0.030350985005497932,
-0.055669672787189484,
0.01693231426179409,
-0.04899879917502403,
0.06364330649375916,
0.12946303188800812,
0.046337202191352844,
-0.022724663838744164,
0.03283929079771042,
0.08919092267751694,
0.06206933408975601,
-0.0245109461247921,
-0.0002278108149766922,
-0.01988527551293373,
0.0034310142509639263
] |
StochSync: Stochastic Diffusion Synchronization for Image Generation in Arbitrary Spaces | https://openreview.net/forum?id=XPNprvlxuQ | [
"Kyeongmin Yeo",
"Jaihoon Kim",
"Minhyuk Sung"
] | Poster | We propose a zero-shot method for generating images in arbitrary spaces (e.g., a sphere for 360◦ panoramas and a mesh surface for texture) using a pretrained image diffusion model. The zero-shot generation of various visual content using a pretrained image diffusion model has been explored mainly in two directions. First, Diffusion Synchronization–performing reverse diffusion processes jointly across different projected spaces while synchronizing them in the target space–generates high-quality outputs when enough conditioning is provided, but it struggles in its absence. Second, Score Distillation Sampling–gradually updating the target space data through gradient descent–results in better coherence but often lacks detail. In this paper, we reveal for the first time the interconnection between these two methods while highlighting their differences. To this end, we propose StochSync, a novel approach that combines the strengths of both, enabling effective performance with weak conditioning. Our experiments demonstrate that StochSync provides the best performance in 360◦ panorama generation (where image conditioning is not given), outperforming previous finetuning-based methods, and also delivers comparable results in 3D mesh texturing (where depth conditioning is provided) with previous methods. | Diffusion Models, Synchronization, Score Distillation, Panorama, Texturing | We propose a zero-shot state-of-the-art method for generating images in arbitrary spaces (e.g., a sphere for 360-degree panoramas and a mesh surface for texture) using a pretrained image diffusion model. | 1,709 | 2501.15445 | [
0.03153001889586449,
-0.05519025772809982,
0.04843745008111,
0.010954473167657852,
0.07159457355737686,
-0.04775708168745041,
-0.0617942251265049,
-0.13377760350704193,
0.06559152156114578,
-0.04735695943236351,
-0.02564791589975357,
-0.009835539385676384,
0.05946490913629532,
0.003237704513594508,
-0.025831123813986778,
-0.044760409742593765,
0.06583505123853683,
0.09522926062345505,
-0.07743147760629654,
-0.035298317670822144,
-0.09116581827402115,
-0.07823308557271957,
-0.024603206664323807,
-0.08776664733886719,
0.08036847412586212,
0.016005201265215874,
0.02087279222905636,
-0.08999605476856232,
0.05567795783281326,
-0.02425229176878929,
0.04380929470062256,
-0.06345689296722412,
-0.0775124803185463,
0.0033147698268294334,
-0.04885562136769295,
0.07202628999948502,
-0.027200311422348022,
0.028875630348920822,
-0.017940226942300797,
0.01954621821641922,
-0.01744154654443264,
0.043705571442842484,
-0.016025414690375328,
-0.0028670192696154118,
0.0618145614862442,
0.03503849357366562,
-0.011790627613663673,
0.015801120549440384,
-0.008788734674453735,
0.002132965484634042,
-0.06799585372209549,
-0.031273115426301956,
-0.14174994826316833,
0.009320446290075779,
0.004574844613671303,
0.08539655804634094,
0.07744649052619934,
0.057118009775877,
0.014183368533849716,
-0.061023224145174026,
0.010196330025792122,
-0.055758148431777954,
-0.013111972250044346,
0.024040963500738144,
0.05223224684596062,
-0.05971016734838486,
0.07306741178035736,
0.02631993219256401,
0.0486077181994915,
0.011060252785682678,
-0.02746407873928547,
0.04757040739059448,
0.003370145335793495,
0.05001828819513321,
-0.02244422398507595,
0.05220818147063255,
0.024772027507424355,
-0.004433444235473871,
-0.014798193238675594,
-0.045700520277023315,
0.09267088770866394,
0.06102251634001732,
0.04635800048708916,
-0.0017220152076333761,
0.006522968411445618,
0.021586855873465538,
-0.03955672308802605,
0.06001155078411102,
-0.026479782536625862,
-0.03697262331843376,
-0.01254678051918745,
0.017326584085822105,
-0.1408078968524933,
-0.035952430218458176,
-0.022230321541428566,
0.007814574055373669,
0.00925663486123085,
-0.03641461953520775,
-0.022515883669257164,
0.07528849691152573,
0.010336101055145264,
-0.03408089280128479,
0.04936602711677551,
0.01860876753926277,
0.060739461332559586,
-0.049491964280605316,
0.09411963075399399,
0.05987447127699852,
-0.017998013645410538,
0.0032949631568044424,
0.05226118862628937,
0.03611157834529877,
-0.033117663115262985,
0.01767108403146267,
-0.010037246160209179,
-0.030672181397676468,
0.009517040103673935,
-0.02338825725018978,
0.06179838627576828,
-0.007721032481640577,
0.04569130390882492,
-0.014764140360057354,
-0.06661643087863922,
0.013263971544802189,
-0.012256929650902748,
-0.06323089450597763,
-0.003600016236305237,
2.160593330628912e-33,
-0.0018127041403204203,
-0.013029680587351322,
0.09260232746601105,
0.030073588714003563,
0.07189823687076569,
0.04812655225396156,
0.01895553432404995,
-0.014389428310096264,
0.012345739640295506,
0.018997156992554665,
-0.02387618087232113,
-0.06142299249768257,
-0.07237416505813599,
0.09990650415420532,
0.06023992970585823,
-0.019186031073331833,
-0.02515440620481968,
0.02135438658297062,
0.007609909400343895,
0.09920545667409897,
-0.00319833867251873,
-0.025070521980524063,
-0.0648232027888298,
-0.04914863035082817,
-0.056898750364780426,
0.037071000784635544,
0.024530507624149323,
-0.009727071970701218,
-0.03198259696364403,
0.026036839932203293,
-0.07501032948493958,
-0.01724734902381897,
-0.041431501507759094,
0.003201450454071164,
0.018119513988494873,
-0.05168852210044861,
-0.09017551690340042,
-0.0606306716799736,
-0.023816341534256935,
-0.03495171666145325,
-0.027832860127091408,
0.09402863681316376,
-0.02623235061764717,
-0.018856829032301903,
-0.08137257397174835,
-0.02008017897605896,
0.008339219726622105,
0.02932862378656864,
-0.05266128480434418,
-0.02103990688920021,
0.08806176483631134,
0.0290317852050066,
-0.05776036158204079,
-0.057127490639686584,
-0.11461853235960007,
-0.01496042963117361,
0.07830368727445602,
-0.09320691972970963,
0.028724586591124535,
0.06505876779556274,
0.04078362137079239,
-0.019200963899493217,
-0.08986132591962814,
0.03610214963555336,
0.059539537876844406,
-0.008625581860542297,
0.04280277341604233,
0.047929797321558,
-0.026940612122416496,
-0.00225479481741786,
-0.03922506421804428,
0.028379397466778755,
-0.07881142944097519,
-0.10013119131326675,
0.0713827908039093,
-0.013028242625296116,
0.05436291918158531,
0.02998621203005314,
0.024403216317296028,
0.008259463123977184,
-0.05186868831515312,
0.03211964666843414,
-0.07872938364744186,
-0.10059352964162827,
-0.07419191300868988,
0.030589114874601364,
-0.03011886030435562,
-0.03646088391542435,
-0.04516603797674179,
-0.03628050908446312,
0.014049422927200794,
-0.0002609959337860346,
0.03212611749768257,
0.028046390041708946,
-0.03546902537345886,
-2.86428759538532e-33,
0.01704149879515171,
0.032401908189058304,
-0.011723062954843044,
0.07447231560945511,
0.005657556466758251,
0.05903139337897301,
0.023646749556064606,
-0.03760991618037224,
0.010668886825442314,
-0.01966080255806446,
-0.03230152279138565,
-0.04757406562566757,
-0.09701766818761826,
0.014149433001875877,
-0.04876753315329552,
-0.01296664122492075,
0.050449565052986145,
0.0027515904512256384,
-0.07208113372325897,
0.07369819283485413,
0.03269782289862633,
0.12391480803489685,
0.06378665566444397,
0.017585206776857376,
-0.07168460637331009,
0.07813341170549393,
0.01573987491428852,
0.09254511445760727,
-0.006983393803238869,
-0.008726497180759907,
0.018783293664455414,
0.0072082336992025375,
0.06401395797729492,
-0.060763075947761536,
-0.08643697202205658,
0.06977228820323944,
0.02627265267074108,
0.003632567822933197,
-0.034144867211580276,
0.0612221322953701,
0.055409595370292664,
-0.06899445503950119,
-0.06913870573043823,
0.08006608486175537,
0.003974603489041328,
-0.016276059672236443,
-0.10061675310134888,
0.01752312108874321,
-0.008395971730351448,
0.07349130511283875,
-0.04575945809483528,
0.04035818576812744,
-0.12977632880210876,
0.04670380428433418,
0.027522867545485497,
-0.025355327874422073,
-0.14230987429618835,
0.023916389793157578,
0.08545149117708206,
0.03104417212307453,
-0.07265037298202515,
-0.032206322997808456,
-0.03958691284060478,
-0.03274611756205559,
0.030788607895374298,
0.010404564440250397,
-0.07452195882797241,
0.0077837067656219006,
0.04350942745804787,
0.15509945154190063,
-0.027239413931965828,
-0.011002679355442524,
-0.013022046536207199,
0.010277093388140202,
-0.0130090843886137,
-0.05779116228222847,
0.07617749273777008,
-0.02261681668460369,
0.028159912675619125,
-0.024908334016799927,
-0.028917131945490837,
0.008494080975651741,
0.016087207943201065,
0.08167149871587753,
0.08364938944578171,
0.02623259834945202,
0.0496608130633831,
-0.05073043331503868,
0.03809072822332382,
0.02980283461511135,
0.006958527024835348,
-0.016282249242067337,
0.024175012484192848,
0.04114164784550667,
-0.036042045801877975,
-4.5846171303765004e-8,
-0.024431396275758743,
0.041049692779779434,
0.052966807037591934,
0.012756755575537682,
-0.0658196285367012,
-0.09907218813896179,
0.05192481726408005,
0.04729850962758064,
-0.004786883946508169,
-0.0683891773223877,
0.029981879517436028,
-0.08251805603504181,
-0.06194315105676651,
-0.08675162494182587,
-0.02075699344277382,
0.0644211620092392,
-0.035676706582307816,
0.01488745491951704,
0.008616775274276733,
0.018199559301137924,
-0.04770737886428833,
0.02872861921787262,
0.026697834953665733,
0.0023037083446979523,
0.03751815855503082,
-0.0014137213584035635,
0.04098854959011078,
0.05567748472094536,
0.1161414384841919,
-0.1109902635216713,
0.006956757977604866,
-0.005633551627397537,
-0.0002529459889046848,
0.01960531249642372,
-0.007342081982642412,
0.013850321061909199,
-0.026894623413681984,
-0.000480330316349864,
-0.011239883489906788,
-0.1196979284286499,
-0.0670841634273529,
0.09341564774513245,
0.015191645361483097,
-0.0009684342658147216,
-0.0004000109329354018,
0.04863695055246353,
0.10981428623199463,
-0.038706786930561066,
0.023015201091766357,
-0.00136856222525239,
-0.012953328900039196,
-0.052816424518823624,
-0.0055411760695278645,
0.04641071707010269,
0.06539492309093475,
-0.033057183027267456,
0.027474820613861084,
-0.01716764084994793,
0.07284145802259445,
0.08533868938684464,
0.006456047296524048,
0.012588080018758774,
-0.027277179062366486,
-0.02388611249625683
] |
FlickerFusion: Intra-trajectory Domain Generalizing Multi-agent Reinforcement Learning | https://openreview.net/forum?id=MRYyOaNxh3 | [
"Woosung Koh",
"Wonbeen Oh",
"Siyeol Kim",
"Suhin Shin",
"Hyeongjin Kim",
"Jaein Jang",
"Junghyun Lee",
"Se-Young Yun"
] | Poster | Multi-agent reinforcement learning has demonstrated significant potential in addressing complex cooperative tasks across various real-world applications. However, existing MARL approaches often rely on the restrictive assumption that the number of entities (e.g., agents, obstacles) remains constant between training and inference. This overlooks scenarios where entities are dynamically removed or $\textit{added}$ $\textit{during}$ the inference trajectory—a common occurrence in real-world environments like search and rescue missions and dynamic combat situations. In this paper, we tackle the challenge of intra-trajectory dynamic entity composition under zero-shot out-of-domain (OOD) generalization, where such dynamic changes cannot be anticipated beforehand. Our empirical studies reveal that existing MARL methods suffer $\textit{significant}$ performance degradation and increased uncertainty in these scenarios. In response, we propose FlickerFusion, a novel OOD generalization method that acts as a $\textit{universally}$ applicable augmentation technique for MARL backbone methods. FlickerFusion stochastically drops out parts of the observation space, emulating being in-domain when inferenced OOD. The results show that FlickerFusion not only achieves superior inference rewards but also $\textit{uniquely}$ reduces uncertainty vis-à-vis the backbone, compared to existing methods. Benchmarks, implementations, and model weights are organized and open-sourced at $\texttt{\href{flickerfusion305.github.io}{\textbf{flickerfusion305.github.io}}}$, accompanied by ample demo video renderings. | Domain Generalization, Multi-agent Reinforcement Learning, Reliability, Safety | We present a novel, simple yet effective method for reliable and safe domain generalization in multi-agent reinforcement learning | 1,688 | null | [
-0.019906142726540565,
-0.08395250141620636,
0.03732921555638313,
0.06271474063396454,
0.05561819300055504,
-0.013998202048242092,
0.004641331732273102,
-0.06396400183439255,
0.0970422774553299,
0.014667419716715813,
-0.02735714800655842,
-0.06022314727306366,
0.01940258964896202,
0.037055082619190216,
0.047480519860982895,
0.012620815075933933,
0.09496895968914032,
0.06478597968816757,
-0.10519056022167206,
-0.03076735883951187,
0.051693569868803024,
-0.004524571355432272,
0.0411309152841568,
0.05300731584429741,
-0.04587119445204735,
-0.05602714791893959,
-0.0329974889755249,
-0.05619526654481888,
0.016361573711037636,
-0.022617805749177933,
0.07402478158473969,
0.03608281537890434,
-0.09201370179653168,
-0.05451644957065582,
-0.02213156968355179,
0.06865141540765762,
-0.06547772139310837,
0.05619097501039505,
-0.08246149867773056,
-0.019715990871191025,
0.006259788293391466,
0.006059253588318825,
0.056930456310510635,
-0.03211862966418266,
0.054343994706869125,
0.014875197783112526,
0.0095222108066082,
-0.0919046625494957,
0.03858761116862297,
0.014639953151345253,
-0.08429109305143356,
-0.04365510493516922,
-0.011964213103055954,
0.06657502055168152,
0.031224941834807396,
0.050624191761016846,
0.0095386803150177,
0.02151767536997795,
0.0015490916557610035,
-0.06522760540246964,
-0.037547517567873,
-0.0895753875374794,
0.04295943304896355,
-0.013105965219438076,
-0.08220530301332474,
-0.006394105032086372,
0.02362576313316822,
0.030329464003443718,
0.0036052772775292397,
0.05763481184840202,
-0.0260308925062418,
0.026008225977420807,
0.026001516729593277,
0.04001679643988609,
0.061589717864990234,
0.07598180323839188,
-0.03195364028215408,
-0.03222513571381569,
0.06487525999546051,
-0.04519319906830788,
0.03781405836343765,
0.015115510672330856,
-0.025409633293747902,
0.028835885226726532,
0.1011277288198471,
0.05123315006494522,
-0.000431167398346588,
0.03454649820923805,
0.08266062289476395,
0.06236483156681061,
-0.0584275908768177,
-0.011631856672465801,
0.10893683135509491,
0.00501297228038311,
-0.039006076753139496,
0.007708757184445858,
-0.00877959281206131,
-0.09554047137498856,
-0.023548435419797897,
0.08443085849285126,
-0.009056969545781612,
0.06973812729120255,
0.002648990834131837,
0.007733508944511414,
-0.025833329185843468,
0.002056850353255868,
0.030365772545337677,
0.021781668066978455,
0.0014053210616111755,
-0.08888435363769531,
0.02910110168159008,
0.030168073251843452,
0.04163634777069092,
-0.03874678164720535,
-0.0567520335316658,
0.003212795127183199,
-0.03665304556488991,
-0.06693880259990692,
-0.015819193795323372,
0.04417302459478378,
0.023951362818479538,
-0.061085425317287445,
0.031029045581817627,
0.039483532309532166,
0.09486208856105804,
0.007335198577493429,
0.01620878092944622,
1.6008210840574157e-33,
0.129344180226326,
0.049458980560302734,
0.009333567693829536,
0.05125484988093376,
-0.0076903109438717365,
-0.0006855707033537328,
0.07090920954942703,
0.02092699520289898,
-0.02712378278374672,
-0.013155318796634674,
0.0028282359708100557,
0.007261272519826889,
-0.009466915391385555,
0.016734512522816658,
0.06908737868070602,
-0.006165803410112858,
-0.04782534018158913,
0.06617259979248047,
-0.011104261502623558,
-0.026898063719272614,
0.05326053127646446,
0.0010924473172053695,
-0.08394105732440948,
-0.019811663776636124,
0.06124350428581238,
0.06898195296525955,
-0.0022199160885065794,
-0.020895862951874733,
-0.03293079510331154,
0.03025771677494049,
-0.023776398971676826,
0.018451238051056862,
-0.03424596041440964,
0.08979935199022293,
0.02050819993019104,
-0.02133861742913723,
-0.10897603631019592,
-0.07863301783800125,
-0.01509413868188858,
0.04879814386367798,
-0.08370022475719452,
-0.035331450402736664,
-0.01704462245106697,
-0.0027132448740303516,
-0.08097061514854431,
-0.10365834832191467,
-0.01405969075858593,
0.05125724896788597,
-0.08388060331344604,
-0.02158222533762455,
0.03308788686990738,
0.06489656120538712,
0.036837562918663025,
-0.044473983347415924,
-0.0644894540309906,
0.031195666640996933,
-0.03351735323667526,
-0.014627043157815933,
0.005027647130191326,
-0.006474312394857407,
-0.02174411714076996,
-0.000813762191683054,
-0.05806821957230568,
0.08718311041593552,
0.03928211331367493,
0.08507243543863297,
-0.03206397965550423,
0.01879889890551567,
0.08264753967523575,
-0.059427760541439056,
0.05376381427049637,
0.062139686197042465,
0.008906925097107887,
-0.08453094959259033,
0.1010250449180603,
-0.035361550748348236,
0.05338817089796066,
-0.07070290297269821,
0.014128103852272034,
0.027389584109187126,
-0.01837899349629879,
-0.020416826009750366,
-0.049145881086587906,
0.01672622375190258,
0.029545163735747337,
-0.01790955290198326,
0.057249538600444794,
-0.07210259884595871,
-0.07996907085180283,
0.006532168947160244,
-0.06498643755912781,
-0.02764294110238552,
-0.05160382390022278,
0.02535666711628437,
0.06780720502138138,
-1.6780686953217023e-33,
-0.005644548684358597,
-0.018407324329018593,
-0.01899849995970726,
0.01978539489209652,
0.003173425095155835,
-0.004809959791600704,
-0.022500835359096527,
-0.08078920841217041,
0.030250735580921173,
-0.09703674167394638,
-0.11195480823516846,
-0.006182684563100338,
-0.01627444475889206,
-0.01165811438113451,
-0.027233915403485298,
-0.02877337485551834,
-0.040998466312885284,
-0.05677815154194832,
0.04066626355051994,
0.0728890597820282,
0.043205276131629944,
0.022785309702157974,
-0.04728061333298683,
-0.0033404415007680655,
0.10376505553722382,
0.03787270933389664,
0.028204357251524925,
0.06810826808214188,
0.008374584838747978,
0.02317240834236145,
-0.010448443703353405,
-0.034327637404203415,
-0.024649979546666145,
-0.022121837362647057,
-0.0549674890935421,
0.09859439730644226,
0.01176577527076006,
0.0075239147990942,
-0.07188514620065689,
0.05689888074994087,
0.02515227347612381,
-0.08629828691482544,
-0.08726682513952255,
0.020917803049087524,
0.0008040005923248827,
0.020360657945275307,
-0.0957050547003746,
0.060211364179849625,
0.07556583732366562,
0.015637382864952087,
-0.08885036408901215,
0.04521358013153076,
-0.13935737311840057,
0.017606690526008606,
-0.035584744065999985,
-0.039574034512043,
0.02352898195385933,
-0.011651730164885521,
0.02079714648425579,
-0.02371833473443985,
-0.0346604585647583,
0.036816731095314026,
0.0041917734779417515,
0.07821856439113617,
0.01214032992720604,
0.03587815538048744,
-0.0006394220399670303,
0.05010605975985527,
-0.03161310777068138,
-0.059958573430776596,
0.0472395159304142,
-0.02008582279086113,
0.029184866696596146,
-0.043385572731494904,
0.07430307567119598,
-0.020066160708665848,
-0.0018964122282341123,
-0.06325782090425491,
-0.026820063591003418,
0.0029418806079775095,
0.04129274934530258,
-0.0055980258621275425,
0.07583729177713394,
0.07803939282894135,
-0.01189508754760027,
0.07783477753400803,
0.036496903747320175,
0.14435909688472748,
0.019525688141584396,
0.0003847293264698237,
-0.040767621248960495,
0.024886539205908775,
0.0011187728960067034,
0.013063337653875351,
-0.07808580994606018,
-4.89367870670776e-8,
-0.10313091427087784,
0.0970773696899414,
0.04177642613649368,
0.04252900928258896,
0.0381312258541584,
-0.0334768183529377,
-0.025413651019334793,
0.036756400018930435,
-0.008140848949551582,
-0.008238510228693485,
0.07738025486469269,
0.016664886847138405,
0.0006353993085213006,
-0.047165267169475555,
-0.03294640779495239,
0.02739410102367401,
-0.0357106551527977,
-0.07471687346696854,
-0.06340790539979935,
-0.0016533300513401628,
-0.032290980219841,
-0.010790417902171612,
-0.02513749524950981,
-0.0826776996254921,
-0.008088009431958199,
-0.05160653218626976,
-0.035863976925611496,
-0.002050899202004075,
0.10640392452478409,
-0.02195327915251255,
0.0060264733619987965,
0.02323487028479576,
0.06685029715299606,
0.024157974869012833,
-0.002546564443036914,
0.09064488112926483,
0.056002989411354065,
-0.009896369650959969,
-0.004810409154742956,
-0.07463391125202179,
-0.04686524346470833,
0.07421227544546127,
-0.008757494390010834,
-0.00914349127560854,
0.02395280823111534,
-0.003217423800379038,
-0.028729509562253952,
-0.1224982887506485,
0.046059463173151016,
-0.06950687617063522,
-0.0047500901855528355,
-0.021753618493676186,
-0.07305154204368591,
0.12152834236621857,
0.11081549525260925,
-0.018166434019804,
0.07779362052679062,
-0.11593257635831833,
0.07176513224840164,
0.06271275132894516,
-0.013290989212691784,
-0.07129016518592834,
-0.056162815541028976,
0.01454560924321413
] |
DiTTo-TTS: Diffusion Transformers for Scalable Text-to-Speech without Domain-Specific Factors | https://openreview.net/forum?id=hQvX9MBowC | [
"Keon Lee",
"Dong Won Kim",
"Jaehyeon Kim",
"Seungjun Chung",
"Jaewoong Cho"
] | Poster | Large-scale latent diffusion models (LDMs) excel in content generation across various modalities, but their reliance on phonemes and durations in text-to-speech (TTS) limits scalability and access from other fields. While recent studies show potential in removing these domain-specific factors, performance remains suboptimal. In this work, we introduce DiTTo-TTS, a Diffusion Transformer (DiT)-based TTS model, to investigate whether LDM-based TTS can achieve state-of-the-art performance without domain-specific factors. Through rigorous analysis and empirical exploration, we find that (1) DiT with minimal modifications outperforms U-Net, (2) variable-length modeling with a speech length predictor significantly improves results over fixed-length approaches, and (3) conditions like semantic alignment in speech latent representations are key to further enhancement. By scaling our training data to 82K hours and the model size to 790M parameters, we achieve superior or comparable zero-shot performance to state-of-the-art TTS models in naturalness, intelligibility, and speaker similarity, all without relying on domain-specific factors. Speech samples are available at https://ditto-tts.github.io. | speech generation, speech synthesis, text-to-speech, tts, zero-shot, non-autoregressive, latent diffusion, transformer | null | 1,686 | null | [
-0.07366059720516205,
-0.09579223394393921,
0.03323345631361008,
0.04207870736718178,
0.043562598526477814,
-0.03857611492276192,
-0.07825536280870438,
-0.03295923396945,
0.03076121024787426,
-0.0474429614841938,
-0.002038366161286831,
-0.0439734049141407,
-0.040712933987379074,
0.04372601956129074,
0.011540363542735577,
-0.04655394330620766,
0.08190500736236572,
0.08976190537214279,
-0.1380351185798645,
-0.04100550711154938,
0.028404269367456436,
0.11280210316181183,
-0.0005948655889369547,
0.01887504570186138,
0.07032911479473114,
0.06605645269155502,
-0.01088002510368824,
-0.06449764221906662,
0.030073605477809906,
-0.030640332028269768,
0.02688024565577507,
0.030401261523365974,
0.0180604737251997,
0.017461778596043587,
-0.0929030254483223,
0.03743475303053856,
-0.029512008652091026,
-0.017036549746990204,
-0.006987154949456453,
-0.020197663456201553,
0.07416936010122299,
0.023962773382663727,
0.01939862221479416,
0.08831759542226791,
0.05841981619596481,
-0.0887068659067154,
-0.07967216521501541,
0.013255243189632893,
0.002418409800156951,
0.04895244538784027,
0.008858281187713146,
0.03741438314318657,
0.0010443481151014566,
0.15053033828735352,
-0.11660627275705338,
0.023042375221848488,
0.0650503858923912,
0.08705226331949234,
-0.01291321124881506,
-0.01546794269233942,
-0.09053272008895874,
-0.06212206929922104,
-0.0059892358258366585,
-0.07612533122301102,
-0.03834535554051399,
-0.02201942540705204,
-0.02729598432779312,
0.056895822286605835,
-0.09434613585472107,
0.04884736239910126,
-0.05927742272615433,
0.05138440430164337,
-0.10924944281578064,
0.09639281034469604,
0.021491335704922676,
0.04726249352097511,
0.03851921111345291,
-0.021054888144135475,
0.016146644949913025,
0.010961337946355343,
0.08562137186527252,
-0.008435823954641819,
0.11670485883951187,
-0.027307869866490364,
0.04577787220478058,
-0.002329835668206215,
0.004093326162546873,
0.009289535693824291,
-0.03803550824522972,
-0.00940448697656393,
-0.06284069269895554,
-0.049943044781684875,
0.04759720712900162,
0.016172992065548897,
0.006904061883687973,
0.0409943163394928,
-0.03463900461792946,
-0.023930303752422333,
-0.019116539508104324,
0.061188213527202606,
-0.010805011726915836,
0.039107874035835266,
-0.036737147718667984,
-0.03454267606139183,
-0.06929437816143036,
-0.05050083249807358,
0.04900998994708061,
0.03527412563562393,
0.002486777724698186,
-0.000978657160885632,
0.05295519530773163,
0.031756993383169174,
0.043139271438121796,
-0.04113539308309555,
0.08335740864276886,
-0.09699609875679016,
-0.031206559389829636,
-0.06569844484329224,
0.09791643172502518,
0.008293328806757927,
-0.03420286253094673,
0.03949859365820885,
-0.08962249010801315,
0.008560565300285816,
-0.009267223067581654,
-0.013059082441031933,
-0.008201465010643005,
1.166934823402882e-33,
-0.03223998472094536,
0.06696753203868866,
0.016929591074585915,
0.08179210126399994,
0.04389137402176857,
-0.02169322781264782,
-0.016009941697120667,
-0.007406007032841444,
-0.04223392531275749,
-0.05006813257932663,
-0.039983995258808136,
0.05987449362874031,
-0.08573733270168304,
0.009604411199688911,
0.06252425163984299,
0.003877198789268732,
-0.009340822696685791,
0.053966689854860306,
0.004068080801516771,
0.004184715449810028,
0.053991105407476425,
0.0013001799816265702,
-0.02729291096329689,
-0.020099055022001266,
0.06942936778068542,
-0.03097604773938656,
0.041666436940431595,
-0.12144547700881958,
-0.017754923552274704,
0.000016967089322861284,
-0.18887631595134735,
-0.003964876756072044,
0.007863905280828476,
-0.004375775344669819,
0.03864012286067009,
-0.05120839551091194,
0.014027269557118416,
-0.06286151707172394,
-0.01701340079307556,
-0.09470918774604797,
-0.05746349319815636,
0.038808416575193405,
-0.01935647241771221,
-0.01810571178793907,
-0.14490067958831787,
-0.021952319890260696,
0.048244986683130264,
0.03244513273239136,
-0.017157379537820816,
-0.056375086307525635,
0.056973159313201904,
0.015578375197947025,
-0.03999807685613632,
-0.024065960198640823,
0.025945110246539116,
-0.017014259472489357,
0.060460165143013,
-0.04132267087697983,
0.039037127047777176,
0.040054917335510254,
0.040889039635658264,
0.044622309505939484,
-0.002582447836175561,
0.04301752150058746,
0.013051305897533894,
-0.041322872042655945,
-0.06333009153604507,
0.07535149157047272,
0.003456811187788844,
-0.030773747712373734,
-0.02326897531747818,
-0.02201518416404724,
0.012603667564690113,
-0.023332232609391212,
0.0367039330303669,
-0.003646689699962735,
0.07823247462511063,
-0.06062736362218857,
0.004197276197373867,
0.014918042346835136,
0.0014034660998731852,
-0.021036120131611824,
-0.0036590592935681343,
-0.06461502611637115,
0.0012168593239039183,
-0.03604355826973915,
0.0075449212454259396,
-0.09793869405984879,
-0.0431087464094162,
-0.0355970561504364,
-0.04382750019431114,
0.036724962294101715,
-0.018536629155278206,
0.054220616817474365,
0.018315937370061874,
-2.0362153753943598e-33,
-0.0684867575764656,
0.027379406616091728,
-0.045598145574331284,
0.11110197752714157,
0.013179387897253036,
-0.054189350455999374,
0.031179876998066902,
0.11236404627561569,
0.08013323694467545,
-0.06863757222890854,
-0.017750179395079613,
-0.025496317073702812,
0.029797837138175964,
-0.021263008937239647,
0.07010548561811447,
0.011728573590517044,
0.04650096595287323,
0.04638141393661499,
-0.020804882049560547,
0.014486046507954597,
0.04050124064087868,
0.035722196102142334,
-0.09404660016298294,
0.03946558013558388,
-0.0374787263572216,
0.004515660461038351,
-0.07738500833511353,
0.0383608378469944,
0.012229849584400654,
-0.03282908350229263,
-0.05956827104091644,
0.06040382385253906,
-0.02355138771235943,
-0.033246397972106934,
-0.06508326530456543,
0.07622598856687546,
0.011664124205708504,
0.013497773557901382,
0.04219801723957062,
0.052881333976984024,
0.07441329210996628,
-0.03608306497335434,
-0.02457990124821663,
0.005945268087089062,
0.02504591830074787,
-0.022647053003311157,
-0.0753740668296814,
0.014920629560947418,
0.06506508588790894,
0.037329237908124924,
0.061269693076610565,
0.05721019208431244,
-0.021775241941213608,
0.07461772114038467,
-0.03856455907225609,
-0.0939895361661911,
-0.05940239131450653,
-0.06486298143863678,
0.029077542945742607,
-0.024658381938934326,
-0.011712233535945415,
-0.02555219642817974,
0.045425571501255035,
-0.1297525316476822,
-0.03210790082812309,
-0.046945761889219284,
0.008716669864952564,
0.02643650583922863,
0.04448694735765457,
0.0012580081820487976,
0.010730731301009655,
-0.028323126956820488,
0.0043445248156785965,
0.03617434948682785,
-0.026690073311328888,
-0.0588763989508152,
-0.0010794568806886673,
-0.0657409131526947,
-0.04310186579823494,
-0.02151091769337654,
0.01291281171143055,
0.08080493658781052,
-0.0024553039111196995,
-0.0827828198671341,
0.06786484271287918,
0.08172354847192764,
0.04513891786336899,
0.011326881125569344,
0.01308211125433445,
0.0857420265674591,
-0.04225844517350197,
0.021970275789499283,
-0.0061408900655806065,
0.07835494726896286,
-0.09024091809988022,
-4.472444103953421e-8,
-0.06927870213985443,
0.05462926998734474,
-0.040785059332847595,
0.006790896877646446,
-0.04519368335604668,
-0.0954718291759491,
0.044055648148059845,
0.06025690957903862,
0.07110461592674255,
0.04477908834815025,
0.09812311083078384,
-0.046895191073417664,
-0.022822365164756775,
0.01664685644209385,
0.004084368702024221,
0.09410517662763596,
0.018379153683781624,
-0.02640293724834919,
-0.0030632144771516323,
-0.06915656477212906,
0.02217857912182808,
0.12754859030246735,
0.009624502621591091,
-0.013432689011096954,
0.060104236006736755,
-0.031774234026670456,
0.0010323977330699563,
0.07177843898534775,
0.05415186658501625,
0.005573808681219816,
-0.016146982088685036,
0.027270805090665817,
-0.0953291729092598,
-0.044586747884750366,
-0.030001243576407433,
0.041456710547208786,
0.03179306536912918,
-0.028674466535449028,
0.006874078884720802,
0.015289940871298313,
0.011787711642682552,
0.04190584644675255,
-0.10389532148838043,
0.05810363590717316,
-0.004370007198303938,
-0.025402165949344635,
0.0076685273088514805,
-0.03840197995305061,
0.011177420616149902,
-0.03662543371319771,
0.044911742210388184,
-0.023297276347875595,
-0.01971772126853466,
-0.0011579619022086263,
0.10321520268917084,
0.02328408695757389,
0.03164166212081909,
0.028155246749520302,
0.052331309765577316,
0.0017734499415382743,
0.0004302756569813937,
-0.00799199566245079,
0.022115053609013557,
0.04328440502285957
] |
Intelligent Go-Explore: Standing on the Shoulders of Giant Foundation Models | https://openreview.net/forum?id=apErWGzCAA | [
"Cong Lu",
"Shengran Hu",
"Jeff Clune"
] | Poster | Go-Explore is a powerful family of algorithms designed to solve hard-exploration problems built on the principle of archiving discovered states, and iteratively returning to and exploring from the most promising states. This approach has led to superhuman performance across a wide variety of challenging problems including Atari games and robotic control, but requires manually designing heuristics to guide exploration (i.e., determine which states to save and explore from, and what actions to consider next), which is time-consuming and infeasible in general. To resolve this, we propose Intelligent Go-Explore (IGE) which greatly extends the scope of the original Go-Explore by replacing these handcrafted heuristics with the intelligence and internalized human notions of interestingness captured by giant pretrained foundation models (FMs). This provides IGE with a human-like ability to instinctively identify how interesting or promising any new state is (e.g., discovering new objects, locations, or behaviors), even in complex environments where heuristics are hard to define. Moreover, IGE offers the exciting opportunity to recognize and capitalize on serendipitous discoveries---states encountered during exploration that are valuable in terms of exploration, yet where what makes them interesting was not anticipated by the human user. We evaluate our algorithm on a diverse range of language and vision-based tasks that require search and exploration. Across these tasks, IGE strongly exceeds classic reinforcement learning and graph search baselines, and also succeeds where prior state-of-the-art FM agents like Reflexion completely fail. Overall, Intelligent Go-Explore combines the tremendous strengths of FMs and the powerful Go-Explore algorithm, opening up a new frontier of research into creating more generally capable agents with impressive exploration capabilities. All our code is open-sourced at: https://github.com/conglu1997/intelligent-go-explore. | Exploration, Large Language Models, LLM agents, Open-endedness | We propose a new FM agent for hard exploration tasks based on the classic Go-Explore algorithm augmented with foundation model intelligent selection. | 1,679 | null | [
0.027140097692608833,
-0.06563752889633179,
0.03033176064491272,
0.019090931862592697,
0.03632890805602074,
-0.016685733571648598,
0.05186816304922104,
-0.0151676582172513,
-0.012219028547406197,
0.005557150114327669,
-0.04672250896692276,
-0.02250772900879383,
-0.009311874397099018,
0.09753086417913437,
-0.0008473703637719154,
-0.05241275578737259,
0.06177613139152527,
-0.04594691842794418,
-0.03207765892148018,
-0.021268334239721298,
0.0602714829146862,
-0.012901361100375652,
0.011691785417497158,
-0.039939869195222855,
-0.07565917819738388,
0.06102181598544121,
-0.028326671570539474,
-0.12188030034303665,
0.033362749963998795,
-0.03492587059736252,
0.05892263352870941,
0.02543642558157444,
0.014034025371074677,
-0.047536879777908325,
-0.011163240298628807,
0.050006888806819916,
-0.06321460753679276,
-0.01697429083287716,
0.025599956512451172,
-0.08858203142881393,
-0.051798224449157715,
-0.009650187566876411,
0.05046942085027695,
0.032182566821575165,
0.008544844575226307,
0.005288481246680021,
-0.02152322791516781,
-0.05865630507469177,
0.041231919080019,
-0.01812868006527424,
-0.10523127019405365,
-0.07757759094238281,
-0.012658683583140373,
-0.06484195590019226,
0.09765370190143585,
0.10189811885356903,
0.024494599550962448,
-0.049826040863990784,
-0.0041015357710421085,
-0.05260734632611275,
0.01685297302901745,
-0.039304789155721664,
0.02011682838201523,
-0.009082290343940258,
0.02462976798415184,
0.01850060001015663,
0.027906807139515877,
0.017341433092951775,
0.04665642976760864,
0.004230978433042765,
0.04378078132867813,
0.01954936794936657,
0.013836978003382683,
0.016066517680883408,
0.1286810338497162,
0.055684708058834076,
-0.041989006102085114,
0.0359465591609478,
0.009563775733113289,
-0.00928129255771637,
-0.006685326807200909,
-0.05272489786148071,
-0.025352636352181435,
0.05568942800164223,
-0.00907759740948677,
-0.012850572355091572,
-0.06251650303602219,
-0.009374464862048626,
0.14803628623485565,
0.01687675341963768,
-0.006042645312845707,
-0.1316913366317749,
-0.06666819006204605,
-0.027314873412251472,
0.018466876819729805,
-0.004080090671777725,
-0.018606087192893028,
-0.06548462808132172,
-0.023684047162532806,
0.03194703534245491,
0.024265188723802567,
0.057397257536649704,
0.0603492334485054,
0.017654510214924812,
0.04024665802717209,
-0.0036621224135160446,
0.029267583042383194,
0.010675721801817417,
0.06832843273878098,
0.003219140926375985,
-0.04777822643518448,
0.025917967781424522,
0.11033882945775986,
0.015828685835003853,
-0.013633757829666138,
0.01465508621186018,
-0.004403992090374231,
0.08504349738359451,
0.008932948112487793,
0.052723806351423264,
0.10339607298374176,
-0.0186245609074831,
0.0657450258731842,
0.04525022208690643,
0.022201936691999435,
0.004152209963649511,
-0.1395527571439743,
1.3746983989584326e-33,
0.016536828130483627,
-0.006755598355084658,
0.07516680657863617,
0.07975605130195618,
-0.015473793260753155,
-0.05708765238523483,
0.060763753950595856,
-0.025615200400352478,
-0.02447972074151039,
0.02158447541296482,
-0.01199323870241642,
0.07290738075971603,
-0.03968189284205437,
0.09240548312664032,
0.05415709316730499,
-0.04595724120736122,
-0.0911267101764679,
-0.008440949022769928,
-0.01719188317656517,
-0.07009770721197128,
0.10454116761684418,
-0.0015115292044356465,
0.05222618579864502,
-0.10110639035701752,
0.09294605255126953,
0.05185546353459358,
-0.008091741241514683,
-0.08032453060150146,
-0.017861036583781242,
0.02358013577759266,
-0.07670549303293228,
0.020351825281977654,
-0.08798971772193909,
0.0661800354719162,
-0.0029662405140697956,
0.04319719597697258,
-0.027741458266973495,
-0.06367159634828568,
0.012715059332549572,
-0.026951342821121216,
0.0006756018265150487,
-0.015014690347015858,
-0.057201191782951355,
0.02015833556652069,
-0.044485580176115036,
-0.05377792567014694,
0.05510919541120529,
0.00016344494360964745,
-0.0977088212966919,
-0.04038883373141289,
-0.05472869053483009,
-0.02699696086347103,
-0.021062027662992477,
-0.07699117064476013,
-0.07115048170089722,
-0.0029028018470853567,
0.026524126529693604,
0.031763821840286255,
0.0494103841483593,
0.02437077835202217,
0.01597815565764904,
0.027105387300252914,
-0.0004875360755249858,
0.05477065220475197,
0.022978391498327255,
0.07382490485906601,
-0.011807717382907867,
0.035696059465408325,
0.03700682520866394,
0.07371488958597183,
-0.013024027459323406,
-0.012686961330473423,
0.01250655110925436,
-0.04871704429388046,
0.04737203195691109,
-0.08038327842950821,
0.03447743505239487,
-0.05577104538679123,
-0.01716776378452778,
-0.077123261988163,
-0.009351671673357487,
-0.0426376573741436,
-0.13305650651454926,
0.04199446365237236,
0.026564517989754677,
-0.056552786380052567,
0.022525159642100334,
-0.08617467433214188,
-0.07312650978565216,
-0.016582032665610313,
-0.050823874771595,
-0.03954079747200012,
0.039422787725925446,
0.03290611132979393,
-0.046556275337934494,
-2.7575615245382875e-33,
-0.018389003351330757,
-0.1816360354423523,
-0.0030039947014302015,
0.030558045953512192,
0.009129121899604797,
-0.07195664942264557,
-0.02932136319577694,
-0.009556765668094158,
-0.04281317815184593,
-0.044026076793670654,
-0.11488205939531326,
0.0606689527630806,
0.06583084911108017,
-0.02868805266916752,
-0.01803736574947834,
0.0011179987341165543,
0.029544927179813385,
-0.031176742166280746,
0.055373284965753555,
0.0698988288640976,
-0.010683577507734299,
0.03788060322403908,
-0.1358930766582489,
-0.028272151947021484,
0.04025793820619583,
0.09984491020441055,
0.02299732156097889,
-0.002374005503952503,
-0.006604330148547888,
0.04908938333392143,
-0.058215461671352386,
-0.04186888784170151,
-0.023725958541035652,
-0.02086063288152218,
-0.006957782432436943,
0.12553177773952484,
0.059115808457136154,
-0.008260572329163551,
-0.03975970298051834,
0.07789099961519241,
-0.01678319275379181,
-0.0231564212590456,
0.012221226468682289,
0.03664796054363251,
0.0652676671743393,
0.052814923226833344,
-0.11480482667684555,
0.11013519018888474,
-0.012841988354921341,
-0.003335637738928199,
0.030413974076509476,
0.056716352701187134,
-0.08646133542060852,
-0.08234300464391708,
-0.0372176319360733,
0.028539972379803658,
-0.00518167857080698,
-0.009572112932801247,
0.041105806827545166,
0.049485936760902405,
-0.06688050925731659,
-0.05198210850358009,
-0.04911402612924576,
0.01923709362745285,
-0.07684770971536636,
-0.032886963337659836,
-0.02888142131268978,
0.036867428570985794,
-0.08356358110904694,
-0.025658797472715378,
-0.0481124147772789,
-0.017879901453852654,
-0.005105594638735056,
-0.028269516304135323,
0.04592811316251755,
-0.019253559410572052,
-0.01236304547637701,
-0.010617789812386036,
0.06903063505887985,
-0.04433515667915344,
-0.03322196379303932,
0.02563437633216381,
0.06336919218301773,
0.004059819038957357,
0.0908663421869278,
0.052568234503269196,
-0.03221418336033821,
0.1291961669921875,
-0.032201480120420456,
-0.03810283541679382,
-0.03271815925836563,
0.036295801401138306,
-0.013309497386217117,
0.05948108434677124,
-0.0711374282836914,
-4.757468730076653e-8,
-0.026648348197340965,
0.03290460258722305,
0.07279987633228302,
0.013031155802309513,
0.10203384608030319,
0.018895354121923447,
-0.06411182135343552,
0.034574612975120544,
-0.08112704753875732,
-0.055217064917087555,
0.029840609058737755,
-0.04321089759469032,
0.023348698392510414,
0.06749392300844193,
0.07064998149871826,
0.02427629753947258,
0.022869961336255074,
-0.00898191798478365,
-0.07313273102045059,
0.020906630903482437,
0.04576370492577553,
-0.01619897037744522,
0.039960745722055435,
-0.04014335945248604,
0.025321707129478455,
-0.04382903501391411,
-0.030641023069620132,
-0.04271100088953972,
0.03059852309525013,
0.04759860038757324,
-0.02942018397152424,
0.01512982975691557,
-0.003747830167412758,
0.04611743241548538,
0.027795758098363876,
0.021766025573015213,
-0.01490854099392891,
-0.05173712596297264,
0.028979536145925522,
-0.005406857002526522,
-0.0024675321765244007,
0.0995132252573967,
-0.007915575988590717,
0.03578192740678787,
-0.03082713671028614,
0.040055859833955765,
-0.021319737657904625,
-0.08347048610448837,
0.010127800516784191,
0.04612109437584877,
-0.02525504305958748,
-0.03384512662887573,
-0.03647210821509361,
0.061732713133096695,
0.10374733060598373,
0.03804353252053261,
-0.019297489896416664,
-0.00961688905954361,
0.019912678748369217,
0.08236882090568542,
0.05008821561932564,
-0.023488132283091545,
-0.056111983954906464,
-0.000035478980862535536
] |
Online Reward-Weighted Fine-Tuning of Flow Matching with Wasserstein Regularization | https://openreview.net/forum?id=2IoFFexvuw | [
"Jiajun Fan",
"Shuaike Shen",
"Chaoran Cheng",
"Yuxin Chen",
"Chumeng Liang",
"Ge Liu"
] | Poster | Recent advancements in reinforcement learning (RL) have achieved great success in fine-tuning diffusion-based generative models. However, fine-tuning continuous flow-based generative models to align with arbitrary user-defined reward functions remains challenging, particularly due to issues such as policy collapse from overoptimization and the prohibitively high computational cost of likelihoods in continuous-time flows. In this paper, we propose an easy-to-use and theoretically sound RL fine-tuning method, which we term Online Reward-Weighted Conditional Flow Matching with Wasserstein-2 Regularization (ORW-CFM-W2). Our method integrates RL into the flow matching framework to fine-tune generative models with arbitrary reward functions, without relying on gradients of rewards or filtered datasets. By introducing an online reward-weighting mechanism, our approach guides the model to prioritize high-reward regions in the data manifold. To prevent policy collapse and maintain diversity, we incorporate Wasserstein-2 (W2) distance regularization into our method and derive a tractable upper bound for it in flow matching, effectively balancing exploration and exploitation of policy optimization. We provide theoretical analyses to demonstrate the convergence properties and induced data distributions of our method, establishing connections with traditional RL algorithms featuring Kullback-Leibler (KL) regularization and offering a more comprehensive understanding of the underlying mechanisms and learning behavior of our approach. Extensive experiments on tasks including target image generation, image compression, and text-image alignment demonstrate the effectiveness of our method, where our method achieves optimal policy convergence while allowing controllable trade-offs between reward maximization and diversity preservation. | Flow Matching, Reinforcement Learning, Wasserstein Regularization, Exploration-Exploitation Trade-off, Fine-Tuning, Generative Model | We propose a novel and theoretically sound method for fine-tuning flow matching generative models using a reward-weighted mechanism and Wasserstein-2 regularization to optimize user-defined rewards while preventing overoptimization. | 1,675 | 2502.06061 | [
-0.06093406677246094,
-0.05767839401960373,
0.055326834321022034,
-0.01875663921236992,
0.058794718235731125,
0.07523557543754578,
-0.021096516400575638,
-0.048917707055807114,
0.0038843201473355293,
-0.023808779194951057,
-0.06275229901075363,
-0.08698367327451706,
0.019900759682059288,
0.057527851313352585,
-0.042064230889081955,
0.03601551428437233,
0.11104985326528549,
0.03888680413365364,
-0.12219777703285217,
-0.015522187575697899,
0.028759948909282684,
0.03557296097278595,
0.023486265912652016,
0.016206946223974228,
-0.0006546674412675202,
0.02204323746263981,
0.05646767094731331,
-0.01554899476468563,
0.046943869441747665,
0.010351786389946938,
0.0018726343987509608,
-0.011075842194259167,
-0.024850232526659966,
-0.023819176480174065,
-0.11021002382040024,
0.10508114099502563,
-0.14288972318172455,
0.02324678562581539,
-0.044238876551389694,
-0.0007992906612344086,
-0.026667581871151924,
0.05638493597507477,
0.03402850031852722,
0.05950779467821121,
-0.012267728336155415,
0.03538814187049866,
-0.03251156583428383,
-0.024555673822760582,
0.018817950040102005,
0.022053450345993042,
-0.0009749988093972206,
-0.04827837646007538,
-0.02623612806200981,
0.07456512749195099,
0.032966405153274536,
0.016701145097613335,
0.0684424340724945,
-0.008271168917417526,
0.025636838749051094,
-0.018014216795563698,
-0.01987641677260399,
-0.18863946199417114,
-0.007264528889209032,
-0.0374104306101799,
-0.047299064695835114,
-0.0919414684176445,
-0.03936370089650154,
0.08852221071720123,
-0.008591361343860626,
0.024975838139653206,
0.022723939269781113,
0.0664234533905983,
-0.0005967701436020434,
0.004309855867177248,
0.08361842483282089,
0.06693152338266373,
-0.003649743739515543,
-0.05676871910691261,
0.036421969532966614,
0.003368502017110586,
0.018758410587906837,
-0.006097814533859491,
0.052164558321237564,
-0.0025790801737457514,
0.059404052793979645,
0.015043449588119984,
-0.04027009382843971,
0.019221261143684387,
0.14379730820655823,
-0.0002131959336111322,
-0.023272914811968803,
0.026818012818694115,
-0.023882608860731125,
-0.04141185060143471,
-0.02203514240682125,
0.021826326847076416,
-0.006851718295365572,
-0.047385841608047485,
-0.008567468263208866,
0.07937902212142944,
0.0014110654592514038,
0.06462980061769485,
-0.0405266210436821,
0.015541616827249527,
0.0400572307407856,
-0.042596686631441116,
0.026461852714419365,
0.115334153175354,
0.01750076562166214,
-0.11867271363735199,
0.03355278819799423,
-0.022584347054362297,
0.12846943736076355,
0.03179924562573433,
-0.08077745139598846,
-0.014811580069363117,
0.059115953743457794,
-0.036714114248752594,
-0.07189390808343887,
0.09778019785881042,
0.009898406453430653,
-0.025246182456612587,
0.004560874309390783,
-0.004031807649880648,
0.024028541520237923,
-0.016921209171414375,
0.01426995825022459,
1.6279658888884169e-34,
0.05948294326663017,
0.031593356281518936,
-0.021925581619143486,
0.046351853758096695,
0.02422850765287876,
0.03061973862349987,
0.03798401728272438,
-0.001050052116625011,
-0.06875526160001755,
-0.05056758224964142,
-0.0245690755546093,
0.033592045307159424,
-0.13750283420085907,
0.13264229893684387,
-0.012431745417416096,
-0.09266243129968643,
-0.023181622847914696,
0.031115787103772163,
0.05567659065127373,
-0.05048481374979019,
0.03232817351818085,
-0.04345713183283806,
-0.04911701753735542,
-0.010990330949425697,
0.045707788318395615,
-0.02843516506254673,
0.04251817241311073,
-0.022050466388463974,
0.056425370275974274,
-0.03164869174361229,
-0.044861603528261185,
-0.05194418877363205,
-0.0021807835437357426,
0.02793307788670063,
0.0028810298535972834,
-0.009105188772082329,
-0.04165005311369896,
-0.0039061226416379213,
-0.004112797789275646,
-0.03567900508642197,
-0.001091792481020093,
0.051645390689373016,
-0.05069899931550026,
-0.020923515781760216,
-0.09770650416612625,
-0.1365124136209488,
0.05124721676111221,
0.014151430688798428,
-0.05234847590327263,
-0.0581347681581974,
0.06715185195207596,
0.04176020994782448,
0.047445185482501984,
-0.061630189418792725,
-0.11023764312267303,
-0.017723096534609795,
0.05279743671417236,
0.057578328996896744,
0.011017273180186749,
0.04880739003419876,
0.034474246203899384,
0.09557453542947769,
0.004042331129312515,
0.05165679752826691,
0.03206123784184456,
0.03752787411212921,
-0.021778147667646408,
0.004233002196997404,
0.01970895752310753,
0.029209429398179054,
0.02414311282336712,
0.008691045455634594,
0.032977938652038574,
-0.08534359931945801,
0.05668531358242035,
0.015086786821484566,
0.051008742302656174,
-0.016202371567487717,
0.02871515415608883,
-0.040061093866825104,
-0.08661627769470215,
0.020490938797593117,
-0.03004308044910431,
0.019220076501369476,
-0.012568583711981773,
-0.044865258038043976,
0.03185271844267845,
-0.12122282385826111,
-0.04010007902979851,
-0.003960902336984873,
-0.013769716024398804,
0.015056335367262363,
-0.02341054193675518,
0.01291949674487114,
0.06340127438306808,
2.893488298509936e-34,
-0.008335416205227375,
0.009282535873353481,
0.04873492196202278,
-0.0041027492843568325,
0.020076192915439606,
0.010484461672604084,
-0.042097125202417374,
-0.003183211199939251,
0.0392657108604908,
-0.06740060448646545,
0.019679661840200424,
0.017189376056194305,
0.01554189808666706,
-0.02457820624113083,
0.007887599989771843,
-0.038144782185554504,
-0.02065078541636467,
0.044629890471696854,
-0.037199363112449646,
0.03310364857316017,
0.08151950687170029,
0.10416396707296371,
-0.15058015286922455,
0.04187735170125961,
0.004938123747706413,
0.04890548810362816,
0.015087291598320007,
0.03308846056461334,
0.031352389603853226,
0.0826917216181755,
-0.012207753956317902,
-0.09150884300470352,
-0.07643328607082367,
-0.019031543284654617,
-0.059382032603025436,
0.09248115122318268,
0.0306493379175663,
0.05352708324790001,
-0.09140704572200775,
0.16572773456573486,
-0.024958856403827667,
-0.03171559423208237,
-0.10475730895996094,
0.059054113924503326,
-0.05446070805191994,
-0.03240788355469704,
-0.06117922440171242,
0.012595550157129765,
-0.01709846593439579,
-0.02834472618997097,
-0.09139750152826309,
0.07814671099185944,
-0.10246603190898895,
0.07502623647451401,
-0.05609731748700142,
-0.046505142003297806,
-0.01983995921909809,
-0.06624401360750198,
0.014625309035182,
-0.05335584655404091,
-0.052731532603502274,
-0.03267435356974602,
-0.010807895101606846,
-0.005552988499403,
-0.01758311316370964,
-0.03214456886053085,
0.007274342700839043,
-0.02177882008254528,
-0.007644854485988617,
0.057135164737701416,
0.0034166716504842043,
0.010040500201284885,
0.08426286280155182,
0.03855044022202492,
0.05497734621167183,
-0.027884604409337044,
0.04810594394803047,
0.04296306520700455,
0.05608413368463516,
0.04392770305275917,
-0.015006473287940025,
0.06564000248908997,
0.07647457718849182,
0.03154590353369713,
0.0980386883020401,
0.01168882381170988,
0.00034403230529278517,
0.062421780079603195,
-0.011012760922312737,
-0.023800745606422424,
-0.020707400515675545,
-0.0655297115445137,
0.046503107994794846,
-0.014959965832531452,
-0.06508872658014297,
-4.1070634893003444e-8,
-0.09754392504692078,
-0.023723077028989792,
0.04075134918093681,
-0.004106173291802406,
-0.008903635665774345,
-0.01228663232177496,
0.02179493010044098,
0.08019963651895523,
-0.034644320607185364,
0.01703478768467903,
0.08309660851955414,
0.0519179068505764,
-0.014901177957654,
-0.13357266783714294,
-0.022054698318243027,
0.034834809601306915,
0.030627574771642685,
0.05529483035206795,
0.004300088621675968,
0.02497892826795578,
-0.011653810739517212,
-0.00018322867981623858,
-0.01963934861123562,
-0.08045867085456848,
0.025301583111286163,
-0.04691905155777931,
0.01840401068329811,
0.03506789356470108,
0.04832146689295769,
0.005997725762426853,
0.016763916239142418,
-0.007084907963871956,
0.06141728535294533,
-0.002660302445292473,
-0.009583068080246449,
0.0348227433860302,
-0.04930054396390915,
0.023505622521042824,
-0.06061425805091858,
0.03002404421567917,
0.034584563225507736,
0.08461277931928635,
-0.023028472438454628,
-0.007898323237895966,
0.04985072836279869,
0.05745268240571022,
-0.00978744775056839,
-0.01882483810186386,
0.061804965138435364,
-0.01924217864871025,
0.027512462809681892,
-0.05505308508872986,
0.05201977491378784,
0.05996362119913101,
0.06843359023332596,
-0.041785430163145065,
-0.01654372550547123,
-0.015589803457260132,
0.0025176475755870342,
0.041698310524225235,
-0.027249176055192947,
-0.0035116064827889204,
-0.03829043358564377,
-0.02808964252471924
] |
GenEx: Generating an Explorable World | https://openreview.net/forum?id=8NlUL0Cv1L | [
"TaiMing Lu",
"Tianmin Shu",
"Alan Yuille",
"Daniel Khashabi",
"Jieneng Chen"
] | Poster | Planning with partial observation is a central challenge in embodied AI. A majority of prior works have tackled this challenge by developing agents that physically explore their environment to update their beliefs about the world state. However, humans can imagine unseen parts of the world through a mental exploration and revise their beliefs with imagined observations. Such updated beliefs can allow them to make more informed decisions at the current step, without having to physically explore the world first. To achieve this human-like ability, we introduce the **Generative World Explorer (Genex)**, a video generation model that allows an agent to mentally explore a large-scale 3D world (e.g., urban scenes) and acquire imagined observations to update its belief. This updated belief will then help the agent to make a more informed decision at the current step. To train Genex, we create a synthetic urban scene dataset, Genex-DB. Our experimental results demonstrate that (1) Genex can generate high-quality and consistent observations during long-horizon mental exploration of large 3D scenes and (2) the beliefs updated with the generated observations can inform an existing decision-making model (e.g., an LLM agent) to make better plans. | Generative Models, Video Generation, Embodied AI | We propose Genex to allow the agent for imaginatively exploration in a physical world, and acquire imagined observations to update its belief. | 1,658 | 2412.09624 | [
0.03217364847660065,
-0.1234198659658432,
0.08583353459835052,
0.0015466812765225768,
0.0626760944724083,
-0.019507985562086105,
-0.012469477951526642,
-0.060731224715709686,
0.03073333203792572,
0.0009811613708734512,
-0.0868317186832428,
-0.07402018457651138,
0.0037563072983175516,
0.030030015856027603,
0.02268734760582447,
-0.004293649923056364,
0.09042973816394806,
0.04996200278401375,
-0.0667058601975441,
0.007462401408702135,
-0.0035749205853790045,
-0.011112857609987259,
0.04773776978254318,
-0.014186380431056023,
-0.0014970280462875962,
0.0328628346323967,
0.10166212916374207,
-0.03950902447104454,
0.07489117234945297,
-0.02497396059334278,
0.12435927242040634,
0.02944205515086651,
-0.01609232649207115,
-0.05938100069761276,
-0.0287723857909441,
0.07866270840167999,
-0.10910559445619583,
0.04249083623290062,
-0.030788352712988853,
-0.08127765357494354,
0.019434304907917976,
0.02304379642009735,
0.012030608020722866,
0.02710914984345436,
0.025300947949290276,
0.021768759936094284,
-0.01296325959265232,
-0.06143530458211899,
0.0757967084646225,
-0.03326835855841637,
-0.13721470534801483,
-0.08551231026649475,
0.03091799095273018,
-0.09145189821720123,
-0.04067861661314964,
0.08158033341169357,
0.07656341046094894,
-0.021365055814385414,
0.04292608052492142,
-0.04255211353302002,
0.01550345215946436,
-0.06911592185497284,
0.024090809747576714,
-0.04006682708859444,
0.041503481566905975,
-0.049094248563051224,
0.009026940912008286,
0.02545180730521679,
0.04081277921795845,
-0.08165235817432404,
-0.0011950283078476787,
0.04946427792310715,
0.007250844035297632,
-0.05644018575549126,
0.014606117270886898,
0.0031909295357763767,
-0.004254082217812538,
-0.094288170337677,
0.07622847706079483,
-0.02235923707485199,
0.025394894182682037,
0.00027039815904572606,
0.0045905099250376225,
0.05679077282547951,
0.024774132296442986,
0.04898824915289879,
0.026727372780442238,
0.10819970071315765,
-0.028758365660905838,
0.05971239134669304,
-0.0759032592177391,
-0.0940079540014267,
-0.025943700224161148,
0.019021516665816307,
0.01117723435163498,
0.05324375256896019,
-0.011540423147380352,
-0.13531947135925293,
0.03814316913485527,
0.025788018479943275,
0.008219663985073566,
0.02275908924639225,
0.02085348777472973,
0.04136573523283005,
0.02124851755797863,
-0.035482436418533325,
0.03513244166970253,
0.007809827104210854,
-0.018880752846598625,
-0.0038332471158355474,
-0.045478276908397675,
-0.0203265193849802,
0.004886940121650696,
0.034761302173137665,
-0.04249328747391701,
0.03383928909897804,
0.01576734147965908,
0.08689939975738525,
-0.0335579551756382,
0.008433273993432522,
0.051958680152893066,
-0.038342058658599854,
0.06453375518321991,
0.05484948679804802,
0.07551049441099167,
-0.051326390355825424,
0.0033266881946474314,
-8.95730736088414e-34,
-0.0005279196775518358,
-0.047277823090553284,
0.03994985297322273,
0.0893857479095459,
0.08848504722118378,
0.012160907499492168,
0.02377631887793541,
-0.0013077518669888377,
-0.05672651156783104,
0.0026541014667600393,
-0.042695168405771255,
-0.04462919384241104,
-0.044244397431612015,
0.15351668000221252,
0.04919823259115219,
-0.038699474185705185,
-0.09014958143234253,
0.038349296897649765,
-0.034595638513565063,
0.058537229895591736,
0.025716982781887054,
-0.0060137007385492325,
0.01899683102965355,
-0.042396627366542816,
0.013811648823320866,
-0.00016121685621328652,
0.035491760820150375,
0.025185732170939445,
-0.004140051547437906,
-0.02984032779932022,
-0.05696551501750946,
0.04740331321954727,
-0.05199562385678291,
0.05695226043462753,
-0.005429222714155912,
0.016683658584952354,
-0.04303284361958504,
-0.07217683643102646,
0.018573401495814323,
0.042035892605781555,
0.01830374076962471,
0.08265592902898788,
0.00003010579530382529,
-0.006307108793407679,
-0.0738297626376152,
-0.007536858320236206,
0.05261664837598801,
0.020573653280735016,
-0.09748376905918121,
0.06641308218240738,
0.0583033449947834,
0.05007420480251312,
0.000226662028580904,
-0.1416119784116745,
-0.06896127760410309,
-0.020535355433821678,
0.051321908831596375,
-0.017727138474583626,
0.047698840498924255,
-0.007841335609555244,
-0.010615966282784939,
0.03637170046567917,
0.007758117746561766,
0.11131651699542999,
0.0037545773666352034,
0.04660111293196678,
0.03917740657925606,
0.08152526617050171,
0.05107869207859039,
0.0650588795542717,
-0.018557647243142128,
0.048430535942316055,
0.03730357065796852,
-0.07092230767011642,
0.03821033611893654,
-0.03978104516863823,
-0.011470956727862358,
-0.03913867846131325,
-0.04843749850988388,
0.0039010888431221247,
0.006258342415094376,
0.009012452326714993,
-0.10725244134664536,
0.014383090659976006,
0.08668561279773712,
-0.007703253999352455,
-0.03993745148181915,
-0.0409233532845974,
-0.03824419900774956,
-0.03566177561879158,
-0.05549188330769539,
-0.0871874988079071,
-0.02740723080933094,
-0.021242523565888405,
-0.06545137614011765,
6.330225801122623e-34,
-0.013063923455774784,
-0.05035380274057388,
0.01641257107257843,
0.017485523596405983,
-0.038064856082201004,
-0.09054653346538544,
-0.042862050235271454,
-0.012668781913816929,
0.006148494314402342,
-0.06165594235062599,
-0.07256460189819336,
0.03895057365298271,
0.05007382482290268,
0.0021667066030204296,
-0.02000376768410206,
-0.05992840230464935,
0.016812285408377647,
-0.05078865960240364,
-0.08216822147369385,
-0.01794619858264923,
0.037339724600315094,
0.050849538296461105,
-0.13305558264255524,
-0.029635434970259666,
-0.044462062418460846,
0.05105556175112724,
-0.003611375577747822,
0.11964966356754303,
0.011271147057414055,
-0.019747843965888023,
-0.051920197904109955,
-0.03248760104179382,
-0.023057619109749794,
-0.008735226467251778,
-0.03662429004907608,
0.07622814923524857,
0.11266018450260162,
-0.013852080330252647,
-0.06936345249414444,
0.06101493909955025,
-0.00724569009616971,
-0.0610918290913105,
-0.04353055730462074,
0.044189453125,
-0.044206466525793076,
0.022489136084914207,
-0.0615101121366024,
0.047353778034448624,
-0.02230837568640709,
0.050323158502578735,
-0.03633301705121994,
0.11222665011882782,
-0.07681765407323837,
-0.04906155541539192,
0.027127835899591446,
-0.08500465005636215,
-0.0467386431992054,
0.019987476989626884,
0.059929076582193375,
-0.004757414106279612,
-0.049600135535001755,
-0.04207433760166168,
0.0016516715986654162,
0.04849052429199219,
-0.045326679944992065,
0.015917722135782242,
-0.0768652930855751,
0.02049250528216362,
0.03552430123090744,
-0.007100848015397787,
0.022347502410411835,
0.01687620021402836,
-0.034570008516311646,
0.011136461049318314,
0.08116179704666138,
-0.028514422476291656,
-0.02522917091846466,
0.03203464671969414,
0.10334393382072449,
-0.018927766010165215,
0.02305610105395317,
-0.04442717134952545,
0.04994841665029526,
0.04569147899746895,
0.10904348641633987,
0.09351510554552078,
-0.051201917231082916,
0.0687536895275116,
0.013142595998942852,
0.003386900294572115,
-0.054823316633701324,
0.06047132983803749,
-0.05104865878820419,
0.03138529881834984,
-0.08133286982774734,
-4.150026455818079e-8,
-0.06257819384336472,
-0.02674766071140766,
0.06797826290130615,
0.0016807550564408302,
0.023312536999583244,
-0.005634886212646961,
-0.015395053662359715,
0.02140943333506584,
-0.02769872359931469,
-0.02768169716000557,
0.08903323858976364,
-0.014980176463723183,
0.007523065898567438,
0.05571626499295235,
0.045082271099090576,
0.006545087322592735,
0.010060391388833523,
-0.014411070384085178,
-0.01010729931294918,
0.03193547949194908,
0.008043107576668262,
0.058316633105278015,
0.010013687424361706,
0.0019419296877458692,
0.09021764248609543,
-0.034605931490659714,
-0.06432059407234192,
0.0293912123888731,
-0.004672505427151918,
0.03925739601254463,
0.027609942480921745,
-0.020856300368905067,
0.01719762571156025,
0.023249134421348572,
0.004598343279212713,
-0.02996687777340412,
0.0007352429674938321,
0.005178438499569893,
0.019905660301446915,
-0.05628805607557297,
0.031794194132089615,
0.046484507620334625,
0.0002639969752635807,
0.014684784226119518,
-0.01926119439303875,
0.01229966338723898,
-0.015936685726046562,
-0.13692761957645416,
0.05117619037628174,
-0.028568768873810768,
-0.061962101608514786,
-0.011323446407914162,
-0.03073418326675892,
0.09270209819078445,
0.12092674523591995,
0.05134706199169159,
-0.04716867953538895,
-0.04764069244265556,
0.04277528449892998,
0.07188276946544647,
-0.01601981744170189,
0.0054614972323179245,
-0.1103464812040329,
-0.06242926046252251
] |
Gradient-Free Generation for Hard-Constrained Systems | https://openreview.net/forum?id=teE4pl9ftK | [
"Chaoran Cheng",
"Boran Han",
"Danielle C. Maddix",
"Abdul Fatir Ansari",
"Andrew Stuart",
"Michael W. Mahoney",
"Bernie Wang"
] | Poster | Generative models that satisfy hard constraints are critical in many scientific and engineering applications, where physical laws or system requirements must be strictly respected. Many existing constrained generative models, especially those developed for computer vision, rely heavily on gradient information, which is often sparse or computationally expensive in some fields, e.g., partial differential equations (PDEs).
In this work, we introduce a novel framework for adapting pre-trained, unconstrained flow-matching models to satisfy constraints exactly in a zero-shot manner without requiring expensive gradient computations or fine-tuning.
Our framework, *ECI sampling*, alternates between extrapolation (E), correction (C), and interpolation (I) stages during each iterative sampling step of flow matching sampling to ensure accurate integration of constraint information while preserving the validity of the generation.
We demonstrate the effectiveness of our approach across various PDE systems, showing that ECI-guided generation strictly adheres to physical constraints and accurately captures complex distribution shifts induced by these constraints.
Empirical results demonstrate that our framework consistently outperforms baseline approaches in various zero-shot constrained generation tasks and also achieves competitive results in the regression tasks without additional fine-tuning. | Flow Matching, Generative Model, Constrained Generation, Partial Differential Equations, Conservation Laws | We propose ECI sampling, a gradient-free approach for guiding pre-trained generative models for hard-constrained generation. | 1,656 | 2412.01786 | [
-0.11341838538646698,
-0.008933166973292828,
0.07829786837100983,
0.0335359051823616,
0.06437772512435913,
-0.010189230553805828,
-0.0976441279053688,
-0.046969935297966,
0.017008693888783455,
-0.06528280675411224,
-0.052090276032686234,
-0.16594114899635315,
0.017281167209148407,
-0.021668031811714172,
-0.009861284866929054,
-0.017221296206116676,
0.013927343301475048,
0.07922863215208054,
-0.09422904998064041,
-0.030346186831593513,
-0.03265855088829994,
0.03490053117275238,
-0.023704588413238525,
-0.0240629892796278,
0.02050730027258396,
-0.030445551499724388,
0.05077243223786354,
-0.018015103414654732,
-0.00005886996950721368,
-0.009599021635949612,
0.052853502333164215,
0.0031011048704385757,
-0.12177672982215881,
-0.0020229960791766644,
-0.009001491591334343,
0.06494656205177307,
-0.026797207072377205,
0.03193242847919464,
0.010816873982548714,
-0.018572019413113594,
0.0017339583719149232,
0.08396587520837784,
0.04028219357132912,
0.018690424039959908,
0.09821822494268417,
-0.015141760930418968,
0.008660933934152126,
-0.04822991415858269,
-0.022875087335705757,
-0.014803888276219368,
-0.04828552529215813,
-0.015177452005445957,
-0.0026054191403090954,
0.015599512495100498,
0.0242796391248703,
0.019396068528294563,
0.13887780904769897,
-0.010052002966403961,
0.05707898363471031,
-0.010931886732578278,
-0.022872094064950943,
-0.11141800135374069,
-0.04703286662697792,
0.00003844335151370615,
-0.069283626973629,
-0.0668865218758583,
0.01709480583667755,
0.05410246178507805,
0.03294997289776802,
0.05718856304883957,
-0.07607313245534897,
0.05395270138978958,
-0.043300382792949677,
0.03125937655568123,
0.015531362034380436,
0.0626799687743187,
0.058216556906700134,
-0.0061034467071294785,
0.0449608750641346,
-0.05470456928014755,
0.03349750488996506,
0.001202797400765121,
-0.002527134260162711,
-0.007487239316105843,
0.05072769895195961,
-0.008854559622704983,
-0.028932876884937286,
0.06235148757696152,
0.11167319864034653,
0.046732570976018906,
-0.04332134127616882,
-0.029972439631819725,
-0.08559678494930267,
0.005207565147429705,
0.02820560708642006,
-0.008481680415570736,
0.058918632566928864,
-0.0846809521317482,
-0.01590512879192829,
0.09138173609972,
0.008252087980508804,
0.05664140358567238,
0.045115888118743896,
0.016585757955908775,
0.02322704903781414,
-0.025835147127509117,
0.06465732306241989,
0.08807890117168427,
-0.0362272672355175,
-0.038370806723833084,
0.09861119091510773,
-0.039296604692935944,
0.05327298864722252,
-0.01358359307050705,
-0.03243410959839821,
-0.05356698855757713,
-0.029997600242495537,
-0.0989522635936737,
-0.028179766610264778,
0.032199110835790634,
-0.07572107762098312,
-0.0168340764939785,
-0.04371412470936775,
0.06172329932451248,
0.014653721824288368,
-0.021016009151935577,
-0.040226686745882034,
3.0565532421979885e-34,
0.011494780890643597,
0.01216974202543497,
0.05251511558890343,
0.06340271234512329,
0.06779037415981293,
0.03132554143667221,
-0.003906309138983488,
0.05402294173836708,
-0.053893305361270905,
0.04972083866596222,
-0.06180574372410774,
-0.02468865178525448,
-0.1362174153327942,
0.08650505542755127,
0.026796096935868263,
-0.03218628466129303,
-0.04141106456518173,
0.032653212547302246,
-0.02579626999795437,
0.08908947557210922,
0.023924265056848526,
-0.10649166256189346,
-0.08195386826992035,
-0.04160575196146965,
0.012018629349768162,
0.05203631892800331,
0.01648341305553913,
0.02229592762887478,
0.011601662263274193,
-0.020018968731164932,
-0.06861855089664459,
0.01713143289089203,
0.0503772608935833,
0.018258443102240562,
0.045778028666973114,
-0.004191781394183636,
0.06791675835847855,
-0.04227573797106743,
-0.03230384364724159,
0.0038344634231179953,
0.02134649083018303,
0.06623127311468124,
0.05641213059425354,
0.029657861217856407,
-0.1024489626288414,
-0.0551934689283371,
0.036963991820812225,
0.07913249731063843,
-0.047592923045158386,
-0.018242742866277695,
-0.052727505564689636,
0.07584433257579803,
0.006908307317644358,
-0.08400919288396835,
-0.055735453963279724,
-0.018055683001875877,
0.019279299303889275,
-0.05278772488236427,
0.010216805152595043,
0.051602963358163834,
-0.030141208320856094,
0.09165053814649582,
-0.04032314196228981,
0.035711515694856644,
0.027502723038196564,
0.04712100699543953,
-0.048538852483034134,
0.05310254544019699,
0.07638265192508698,
-0.01424887403845787,
-0.05757547914981842,
-0.02620338276028633,
-0.024397598579525948,
-0.07054309546947479,
0.13196498155593872,
0.06337478756904602,
0.05192376673221588,
0.011782640591263771,
0.004548497963696718,
0.045349378138780594,
-0.06289903074502945,
0.07567575573921204,
0.0006156841409392655,
0.023476529866456985,
0.00893213041126728,
-0.06833568960428238,
-0.024275856092572212,
-0.002206997713074088,
-0.043901409953832626,
-0.10636604577302933,
-0.04404491186141968,
-0.0817873403429985,
-0.057160813361406326,
0.037789128720760345,
0.0206308513879776,
-1.2862944279498809e-33,
-0.006651629228144884,
0.029967768117785454,
0.020865976810455322,
0.0002119355631293729,
0.03055562824010849,
-0.022205984219908714,
-0.02019023336470127,
0.00123431661631912,
0.08372848480939865,
-0.07732392102479935,
-0.05100471153855324,
-0.02274436689913273,
-0.0062698740512132645,
-0.0212765634059906,
0.009403405711054802,
-0.010120661929249763,
-0.14670243859291077,
-0.02868659980595112,
0.013771522790193558,
0.029664674773812294,
0.010683950036764145,
0.06958434730768204,
-0.0625324472784996,
0.03371676057577133,
-0.044278062880039215,
0.0732397735118866,
-0.04428768903017044,
0.006733639165759087,
-0.031332194805145264,
0.011295796371996403,
-0.002721878234297037,
-0.02442198619246483,
-0.04985319823026657,
-0.001406337833032012,
-0.07669636607170105,
0.08611064404249191,
0.04020947962999344,
0.06499816477298737,
0.004992296919226646,
0.09281501919031143,
0.026306694373488426,
-0.019223762676119804,
-0.07018911838531494,
0.10304217040538788,
0.017399044707417488,
0.02593170665204525,
-0.004414249677211046,
-0.053317196667194366,
0.026329951360821724,
0.07171802967786789,
-0.09299860894680023,
0.0471876785159111,
-0.06849373131990433,
0.0015680453507229686,
-0.03295149654150009,
-0.12965509295463562,
-0.09404623508453369,
-0.0033866027370095253,
0.12039587646722794,
-0.03468088433146477,
-0.06161290034651756,
-0.04704515263438225,
-0.03978120535612106,
-0.04774864763021469,
-0.0031118979677557945,
-0.028809377923607826,
-0.04081307351589203,
-0.01747032068669796,
0.08217078447341919,
0.0413188710808754,
0.03342730551958084,
-0.018354356288909912,
0.03519603610038757,
0.014825001358985901,
0.044278621673583984,
-0.07264301925897598,
0.04424974322319031,
-0.04582945629954338,
0.08258374035358429,
-0.046598851680755615,
0.07191315293312073,
0.018761927261948586,
0.05247020348906517,
0.06405071914196014,
0.02758563868701458,
0.003211184171959758,
-0.01672431454062462,
-0.010261607356369495,
0.03011668287217617,
-0.0010512826265767217,
-0.04246237874031067,
0.00854089017957449,
0.03759404644370079,
0.019856108352541924,
-0.09682395309209824,
-4.311223378294926e-8,
0.002363031031563878,
0.07058260589838028,
0.02326393499970436,
-0.0025666782166808844,
0.02219224162399769,
-0.00306499726139009,
0.01309226918965578,
0.024453971534967422,
0.03445883467793465,
-0.08436308801174164,
0.07001309841871262,
-0.016641100868582726,
0.01824425533413887,
-0.0598122663795948,
-0.029915116727352142,
0.08452096581459045,
0.003390721743926406,
0.027514249086380005,
-0.03537103906273842,
0.026832658797502518,
-0.10121642798185349,
0.0672316774725914,
-0.06050776317715645,
-0.014100173488259315,
0.0715392604470253,
-0.1040637195110321,
0.009409227408468723,
0.008581757545471191,
0.11938286572694778,
-0.018492236733436584,
-0.03182714432477951,
-0.03586003929376602,
0.04298068583011627,
0.03212391957640648,
0.02925783582031727,
0.0486443005502224,
-0.03295944258570671,
0.004240985028445721,
-0.025827135890722275,
-0.10227140784263611,
-0.056486114859580994,
0.08305119723081589,
-0.01692158356308937,
-0.011631614528596401,
0.013239637017250061,
0.06927772611379623,
0.008612186647951603,
-0.03870546817779541,
0.016916004940867424,
0.04873042181134224,
0.0037697507068514824,
-0.014929449185729027,
0.007770774886012077,
-0.002748775063082576,
-0.016181431710720062,
-0.008777988143265247,
0.018996557220816612,
-0.02871892601251602,
0.05458337068557739,
0.025468146428465843,
-0.0236396174877882,
0.060492660850286484,
0.00493223499506712,
-0.03010527603328228
] |
UniDrive: Towards Universal Driving Perception Across Camera Configurations | https://openreview.net/forum?id=jVDPq9EdzT | [
"Ye Li",
"Wenzhao Zheng",
"Xiaonan Huang",
"Kurt Keutzer"
] | Poster | Vision-centric autonomous driving has demonstrated excellent performance with economical sensors. As the fundamental step, 3D perception aims to infer 3D information from 2D images based on 3D-2D projection. This makes driving perception models susceptible to sensor configuration (e.g., camera intrinsics and extrinsics) variations. However, generalizing across camera configurations is important for deploying autonomous driving models on different car models. In this paper, we present UniDrive, a novel framework for vision-centric autonomous driving to achieve universal perception across camera configurations. We deploy a set of unified virtual cameras and propose a ground-aware projection method to effectively transform the original images into these unified virtual views. We further propose a virtual configuration optimization method by minimizing the expected projection error between original cameras and virtual cameras. The proposed virtual camera projection can be applied to existing 3D perception methods as a plug-and-play module to mitigate the challenges posed by camera parameter variability, resulting in more adaptable and reliable driving perception models. To evaluate the effectiveness of our framework, we collect a dataset on CARLA by driving the same routes while only modifying the camera configurations. Experimental results demonstrate that our method trained on one specific camera configuration can generalize to varying configurations with minor performance degradation. | Autonomous Driving, 3D Detection, Sensor Configuration | null | 1,655 | 2410.13864 | [
-0.0033042875584214926,
-0.10806538909673691,
-0.029907165095210075,
-0.004821841139346361,
0.05611952766776085,
0.003063468262553215,
0.006141464691609144,
0.0015947123756632209,
-0.007434530649334192,
-0.03615150228142738,
-0.03700108453631401,
-0.10110300779342651,
0.013595396652817726,
0.07288546860218048,
0.014739838428795338,
-0.04745297506451607,
0.12027767300605774,
0.12670208513736725,
-0.09091290086507797,
0.05192023888230324,
0.011059689335525036,
-0.013607586733996868,
0.05010203272104263,
-0.007359488867223263,
0.023347090929746628,
0.028534021228551865,
0.048949845135211945,
-0.033224139362573624,
-0.05724012851715088,
-0.04380740597844124,
0.029968751594424248,
-0.04564564675092697,
-0.04087080433964729,
-0.002234424464404583,
0.007672908715903759,
-0.0902232751250267,
-0.035086892545223236,
-0.030064888298511505,
-0.08518000692129135,
-0.07329051196575165,
-0.03739014267921448,
0.0482022799551487,
-0.008378891274333,
0.0012046955525875092,
0.07163669914007187,
0.10029066354036331,
0.0853002667427063,
-0.007522709667682648,
0.055000897496938705,
-0.03331075608730316,
0.012447097338736057,
-0.055230263620615005,
0.027724694460630417,
-0.010331615805625916,
-0.0468885637819767,
0.06251540035009384,
-0.0118525680154562,
0.010082281194627285,
0.010210281237959862,
0.05488669499754906,
0.0686788558959961,
-0.012404198758304119,
0.03154204785823822,
0.036106593906879425,
-0.051806166768074036,
-0.001452998723834753,
0.051047831773757935,
-0.05368317291140556,
-0.0027248321566730738,
0.00972438883036375,
-0.053707223385572433,
0.09581179916858673,
0.05218387395143509,
-0.12256398797035217,
0.020163564011454582,
0.0029499223455786705,
0.00011201472079847008,
-0.0007610365282744169,
-0.028779104351997375,
-0.09237982332706451,
0.07881388813257217,
0.08976906538009644,
0.002557894214987755,
0.1086910143494606,
0.061593133956193924,
0.0004412206762935966,
-0.028628185391426086,
0.06080485135316849,
0.024380367249250412,
-0.00765012064948678,
-0.12366906553506851,
-0.09295804053544998,
-0.05579458549618721,
-0.044972918927669525,
0.0375191830098629,
-0.03145000711083412,
0.0225590281188488,
-0.1220901757478714,
0.048637814819812775,
0.02264559082686901,
0.004259409848600626,
0.006670084316283464,
0.04627864807844162,
-0.02115885727107525,
0.06227602809667587,
0.0705617368221283,
0.03437640145421028,
0.043122582137584686,
-0.026515498757362366,
-0.01663452945649624,
0.08171825855970383,
-0.019574930891394615,
0.04909660667181015,
-0.02722926251590252,
-0.02366630546748638,
0.03335941582918167,
-0.0004109296714887023,
0.05735979974269867,
0.031215397641062737,
0.030996961519122124,
-0.06599719077348709,
-0.10045529156923294,
0.05116096884012222,
0.06354554742574692,
0.07092920690774918,
-0.0109212351962924,
0.030001400038599968,
1.3042816955693696e-33,
-0.07488154619932175,
0.057216111570596695,
0.07574012875556946,
0.02211274392902851,
0.062032300978899,
0.014607836492359638,
-0.02980361133813858,
0.06332851201295853,
-0.03568897768855095,
-0.025881962850689888,
0.035724930465221405,
-0.04956089332699776,
-0.0442052036523819,
0.06424082815647125,
0.05189764127135277,
0.03954122215509415,
-0.04199323430657387,
0.026286402717232704,
-0.010646692477166653,
0.03968915343284607,
0.06833131611347198,
0.005497063975781202,
-0.0014951329212635756,
-0.023598920553922653,
0.014628749340772629,
0.022463640198111534,
0.018632296472787857,
0.06079067289829254,
0.008779733441770077,
-0.03236910700798035,
-0.06203962489962578,
0.04106712341308594,
-0.026570020243525505,
0.06168831139802933,
-0.028899282217025757,
0.021470410749316216,
-0.06652017682790756,
-0.019217422232031822,
-0.025768732652068138,
0.06232661381363869,
0.020189618691802025,
0.1062774658203125,
-0.037110429257154465,
-0.04713968560099602,
-0.07396719604730606,
-0.02474653534591198,
-0.003026222111657262,
0.13845916092395782,
-0.08490096777677536,
-0.0002542977163102478,
0.016544632613658905,
-0.02257128246128559,
-0.0023298272863030434,
-0.07778371870517731,
-0.08108986914157867,
-0.09209135174751282,
-0.004151396453380585,
0.009419774636626244,
-0.002040460240095854,
-0.06665030866861343,
0.006516536232084036,
-0.027777763083577156,
-0.044892631471157074,
-0.014072375372052193,
-0.05470024794340134,
-0.006279152352362871,
0.027605753391981125,
0.05394104868173599,
0.015472633764147758,
0.07665515691041946,
0.0017849461873993278,
0.006765614729374647,
-0.047188300639390945,
-0.08499249070882797,
0.09417987614870071,
0.03313833102583885,
-0.04197720065712929,
-0.0718647688627243,
0.02402486838400364,
0.060544636100530624,
-0.09073719382286072,
0.11186980456113815,
0.017569908872246742,
-0.02866542711853981,
0.030408024787902832,
-0.07671220600605011,
-0.02346990630030632,
-0.04656654968857765,
-0.04551425576210022,
-0.013240023516118526,
-0.024340029805898666,
-0.04340849444270134,
-0.03886353224515915,
-0.020489122718572617,
0.018263911828398705,
-8.491093444378559e-34,
0.029909148812294006,
-0.016341792419552803,
-0.012326464056968689,
0.01391767617315054,
0.007937068119645119,
-0.09019175916910172,
0.01808813214302063,
0.033756695687770844,
0.05233197659254074,
-0.04724493250250816,
-0.07919738441705704,
-0.010964062996208668,
0.0712512880563736,
0.057807743549346924,
-0.01670239493250847,
0.0025992710143327713,
-0.0022279187105596066,
-0.018644634634256363,
-0.04126689210534096,
-0.006617062725126743,
0.010609542019665241,
0.08438210189342499,
-0.00903476681560278,
0.011114567518234253,
-0.04129263386130333,
0.06085975095629692,
-0.02244282141327858,
0.0789569616317749,
-0.04117751866579056,
-0.02476470172405243,
0.012335828505456448,
-0.027266593649983406,
-0.020209958776831627,
-0.07990727573633194,
-0.0629630759358406,
0.06780952215194702,
-0.009634300135076046,
-0.01567080430686474,
-0.10624166578054428,
0.03582046553492546,
-0.058354027569293976,
-0.07001575827598572,
-0.032459888607263565,
0.03690386563539505,
0.0375080481171608,
-0.02076195552945137,
0.02646728977560997,
0.04356203228235245,
-0.0007528952555730939,
0.045607130974531174,
-0.04868848994374275,
0.07538694143295288,
-0.07621176540851593,
0.026323359459638596,
0.025752563029527664,
-0.047327183187007904,
-0.05687680095434189,
0.007672356907278299,
0.04321398586034775,
0.006704447325319052,
0.009753390215337276,
-0.010965605266392231,
-0.11730087548494339,
0.07408875972032547,
-0.018626924604177475,
-0.01872931607067585,
-0.07292209565639496,
0.015038429759442806,
0.0133920107036829,
0.050211843103170395,
-0.019801829010248184,
0.005603146273642778,
-0.05342325195670128,
-0.02318703383207321,
-0.0004808623343706131,
-0.038870636373758316,
0.05419096350669861,
0.03770630434155464,
0.1261623501777649,
0.006673119496554136,
0.0651768147945404,
-0.07103821635246277,
0.08911562711000443,
0.1255064457654953,
0.019226448610424995,
0.0619015172123909,
-0.0067645469680428505,
0.029741838574409485,
0.03052613139152527,
0.08853078633546829,
0.018141228705644608,
-0.00890970230102539,
-0.07183724641799927,
0.0777880996465683,
-0.08411014825105667,
-3.5236773499036644e-8,
-0.08627589792013168,
0.06121399998664856,
0.0528828464448452,
-0.016529006883502007,
-0.08475837111473083,
-0.05232341215014458,
0.037204500287771225,
0.0711701363325119,
-0.06874684989452362,
-0.03994441032409668,
0.01635013520717621,
0.008833669126033783,
0.047694139182567596,
0.021250197663903236,
0.013435648754239082,
0.08590144664049149,
0.002435174770653248,
0.03645722568035126,
-0.04411914944648743,
0.05662301555275917,
-0.027664026245474815,
0.044014763087034225,
-0.05002877116203308,
0.0006965987267903984,
0.08205115795135498,
-0.07219059020280838,
-0.04767521843314171,
0.004103260580450296,
0.07495313137769699,
-0.01863604597747326,
-0.036857228726148605,
-0.018697639927268028,
0.014119061641395092,
0.023996254429221153,
-0.030536087229847908,
-0.01471345778554678,
-0.039412710815668106,
0.0697559267282486,
0.053312335163354874,
-0.10452111065387726,
-0.015407035127282143,
0.05717591196298599,
0.003921004012227058,
0.019693473353981972,
-0.012225588783621788,
0.0534609779715538,
0.043433427810668945,
-0.02396375872194767,
-0.056335579603910446,
-0.001051939558237791,
-0.06406017392873764,
0.007319926284253597,
-0.05129044130444527,
0.126095250248909,
0.053465161472558975,
-0.007396823726594448,
0.04007214307785034,
-0.08723584562540054,
0.001797788543626666,
0.07989147305488586,
-0.02443196438252926,
0.0007958016940392554,
-0.029477354139089584,
0.0026221261359751225
] |
ImageFolder: Autoregressive Image Generation with Folded Tokens | https://openreview.net/forum?id=QE1LFzXQPL | [
"Xiang Li",
"Kai Qiu",
"Hao Chen",
"Jason Kuen",
"Jiuxiang Gu",
"Bhiksha Raj",
"Zhe Lin"
] | Poster | Image tokenizers are crucial for visual generative models, \eg, diffusion models (DMs) and autoregressive (AR) models, as they construct the latent representation for modeling. Increasing token length is a common approach to improve image reconstruction quality. However, tokenizers with longer token lengths are not guaranteed to achieve better generation quality. There exists a trade-off between reconstruction and generation quality regarding token length. In this paper, we investigate the impact of token length on both image reconstruction and generation and provide a flexible solution to the tradeoff. We propose \textbf{ImageFolder}, a semantic tokenizer that provides spatially aligned image tokens that can be folded during autoregressive modeling to improve both efficiency and quality. To enhance the representative capability without increasing token length, we leverage dual-branch product quantization to capture different contexts of images. Specifically, semantic regularization is introduced in one branch to encourage compacted semantic information while another branch is designed to capture pixel-level details. Extensive experiments demonstrate the superior quality of image generation and shorter token length with ImageFolder tokenizer. | Semantic tokenizer, Autoregressive generation, Product quantization | null | 1,652 | 2410.01756 | [
-0.017994914203882217,
-0.06307564675807953,
0.05624283477663994,
0.050088342279195786,
0.04496362432837486,
-0.020627779886126518,
-0.07152730226516724,
0.016614306718111038,
0.07564431428909302,
-0.059336207807064056,
0.04862675815820694,
-0.029493780806660652,
0.034663595259189606,
0.06792743504047394,
0.016427071765065193,
0.045596227049827576,
-0.019204087555408478,
0.14815011620521545,
-0.16534824669361115,
-0.04109833389520645,
0.03048243373632431,
-0.03158211335539818,
0.01821923814713955,
-0.06467496603727341,
0.02013610675930977,
0.07784346491098404,
0.03185122460126877,
-0.05887628346681595,
0.07015681266784668,
-0.03480452671647072,
0.05695055052638054,
-0.016196230426430702,
-0.00047439782065339386,
0.02671823650598526,
0.03576605021953583,
0.11071170121431351,
0.028298422694206238,
0.059263698756694794,
0.039016660302877426,
-0.04566463455557823,
0.05002488195896149,
0.025744393467903137,
0.014841951429843903,
0.011965098790824413,
0.03349184989929199,
0.036751385778188705,
0.006047482136636972,
-0.04000645503401756,
0.0023285660427063704,
0.0027683512307703495,
-0.009341692551970482,
0.024661952629685402,
-0.08417101204395294,
0.005813951604068279,
-0.025520185008645058,
0.03786366805434227,
0.031035467982292175,
0.002064632950350642,
0.019162338227033615,
0.031141769140958786,
-0.0306989923119545,
-0.012081528082489967,
0.021038157865405083,
-0.0270784180611372,
-0.013022814877331257,
-0.0595635287463665,
0.038706034421920776,
0.011405033059418201,
0.01804831624031067,
0.031048661097884178,
-0.08062922954559326,
0.016520723700523376,
-0.1246592327952385,
0.03259345889091492,
-0.03249797970056534,
0.04124753177165985,
0.0452210009098053,
-0.053842056542634964,
0.06938350945711136,
-0.06173349544405937,
0.07284656167030334,
0.000703292025718838,
0.07237117737531662,
0.07502279430627823,
0.05789874121546745,
0.03051873669028282,
-0.0015507728094235063,
0.03663447126746178,
-0.002289745258167386,
-0.0031336452811956406,
-0.08314434438943863,
-0.0414205826818943,
-0.006088447291404009,
0.0349234901368618,
0.0012406065361574292,
-0.04294746369123459,
0.04234830290079117,
-0.05386633798480034,
0.0002505065640434623,
0.08911623060703278,
0.06305564194917679,
0.021027710288763046,
0.015611891634762287,
-0.00942561961710453,
0.06452759355306625,
-0.06057814508676529,
0.04682152718305588,
0.016258591786026955,
0.050730276852846146,
0.025806130841374397,
0.05498575419187546,
0.03084915690124035,
-0.04515324532985687,
0.022901849821209908,
0.044500682502985,
-0.04494627192616463,
-0.046204037964344025,
-0.02670828439295292,
0.03891509026288986,
-0.005640182178467512,
-0.016779348254203796,
0.028940372169017792,
-0.04772530496120453,
0.02045745775103569,
-0.015279356390237808,
-0.06786300987005234,
-0.0034333535004407167,
3.808235089363549e-33,
-0.06628958880901337,
0.0775752067565918,
0.03968653082847595,
0.035133808851242065,
0.09857872128486633,
0.0003436729602981359,
-0.06231626123189926,
-0.026181770488619804,
-0.018994973972439766,
-0.014469892717897892,
-0.03211800754070282,
-0.039944250136613846,
-0.08433755487203598,
0.12578332424163818,
0.03709428384900093,
-0.05932756885886192,
-0.09470544010400772,
0.052607350051403046,
-0.03098595328629017,
-0.01930549368262291,
-0.055007148534059525,
-0.003810703055933118,
-0.012004602700471878,
0.03884506970643997,
-0.033798109740018845,
-0.08047458529472351,
0.12005802989006042,
-0.020217621698975563,
0.015196584165096283,
-0.023309219628572464,
-0.15551218390464783,
-0.038739822804927826,
0.06907989084720612,
0.00009636914182920009,
-0.02690332569181919,
-0.017945919185876846,
0.015168249607086182,
-0.05667883902788162,
-0.05260441079735756,
0.03352399170398712,
-0.0424799770116806,
0.0848529040813446,
0.011288546957075596,
-0.03684934973716736,
-0.05705739185214043,
-0.05714988335967064,
0.03934038057923317,
0.02357548102736473,
-0.05656903237104416,
-0.0447796992957592,
0.08222588151693344,
-0.0035388562828302383,
0.03338852524757385,
-0.05066889151930809,
-0.007437012158334255,
-0.06975521892309189,
0.040948133915662766,
-0.06535270810127258,
0.018249887973070145,
0.02421765774488449,
0.0017930413596332073,
0.06386864185333252,
0.04574707895517349,
0.06609579175710678,
-0.0032977787777781487,
-0.025233782827854156,
0.09209064394235611,
0.005741553846746683,
-0.030338335782289505,
0.04274832084774971,
-0.0523594431579113,
0.07172095775604248,
-0.04716445133090019,
-0.09086547791957855,
0.018039114773273468,
-0.025698458775877953,
0.06568581610918045,
-0.01125693041831255,
-0.0017282288754358888,
0.060051433742046356,
-0.06594771146774292,
-0.02136225253343582,
0.02335374988615513,
-0.08655428141355515,
-0.018879210576415062,
-0.07116806507110596,
0.011150218546390533,
-0.11651738733053207,
-0.022396404296159744,
-0.07447251677513123,
-0.025769054889678955,
-0.03589295968413353,
-0.04469309374690056,
0.054729945957660675,
-0.016325950622558594,
-3.0312801595730574e-33,
0.027327649295330048,
0.08334638923406601,
0.011932549998164177,
0.02782621420919895,
-0.061098452657461166,
-0.06202100217342377,
0.04388394579291344,
0.05439872294664383,
-0.046751607209444046,
-0.10954739153385162,
-0.0734313428401947,
0.040594734251499176,
0.0012695787008851767,
0.004315927159041166,
-0.040157053619623184,
-0.04336307942867279,
0.04230017960071564,
-0.045368634164333344,
-0.06577348709106445,
0.04867027327418327,
0.05562865734100342,
0.05062246322631836,
-0.023607784882187843,
0.038678623735904694,
-0.08407894521951675,
0.10245606303215027,
-0.012250881642103195,
0.02254173904657364,
0.03023706190288067,
-0.026620380580425262,
0.05688200145959854,
-0.08038347214460373,
0.04590568691492081,
-0.06788258999586105,
-0.07128925621509552,
-0.018670113757252693,
0.03865908086299896,
0.025935720652341843,
0.0013003282947465777,
0.02880208194255829,
0.020401984453201294,
-0.06130991876125336,
-0.04215499013662338,
0.07923226803541183,
-0.03563227504491806,
-0.042877428233623505,
0.013000286184251308,
-0.010116231627762318,
0.015125608071684837,
0.07407587021589279,
-0.03802690654993057,
0.04995317384600639,
-0.0903969556093216,
0.07276710867881775,
-0.04516349360346794,
-0.07348232716321945,
-0.09756279736757278,
-0.030538242310285568,
0.023982562124729156,
0.005723280366510153,
-0.02540726400911808,
-0.10886252671480179,
-0.005555001553148031,
-0.09101741015911102,
-0.018582135438919067,
-0.05071872100234032,
-0.048446837812662125,
0.01119177509099245,
-0.052394237369298935,
0.06810696423053741,
0.04906117916107178,
-0.005949108861386776,
-0.023023376241326332,
0.060342926532030106,
0.06275087594985962,
-0.027992170304059982,
0.03278374299407005,
0.023962730541825294,
0.09028357267379761,
-0.038746900856494904,
-0.0507596880197525,
-0.0013016237644478679,
0.01417601015418768,
0.04657987877726555,
0.0037953790742903948,
-0.0016237640520557761,
0.002893674187362194,
0.03646428510546684,
0.06068676337599754,
0.05656132847070694,
0.020517738536000252,
0.008861654438078403,
0.03932968154549599,
0.07253327965736389,
-0.027545101940631866,
-4.491656113714271e-8,
-0.06570826470851898,
0.08999618142843246,
0.05100477859377861,
-0.025653531774878502,
-0.07298610359430313,
-0.0912492573261261,
0.01823912188410759,
0.03164629638195038,
0.024403950199484825,
-0.01261490024626255,
0.03681240230798721,
-0.048890892416238785,
-0.1535090059041977,
-0.04651881381869316,
-0.019383497536182404,
0.05618596822023392,
0.009693103842437267,
0.0452079139649868,
0.03352215886116028,
0.006003604270517826,
-0.10409678518772125,
0.06201518699526787,
0.0030072855297476053,
-0.04886253923177719,
0.03272946551442146,
-0.035204239189624786,
-0.017681529745459557,
0.07934452593326569,
0.06405491381883621,
-0.08131866157054901,
0.011203267611563206,
0.10086216777563095,
-0.0019837594591081142,
-0.014933406375348568,
-0.09396746754646301,
0.03747699409723282,
0.01086680218577385,
-0.025180315598845482,
-0.04226117581129074,
-0.08807239681482315,
-0.002730383537709713,
0.035239335149526596,
-0.004867046605795622,
0.023890884593129158,
0.045293089002370834,
0.011085456237196922,
0.10932622849941254,
0.004081456921994686,
0.0380428284406662,
-0.0021750887390226126,
-0.02420561946928501,
-0.03963004797697067,
-0.0005747329560108483,
0.02893099933862686,
0.003324956865981221,
0.02966465801000595,
0.028594624251127243,
0.03709621727466583,
0.11668236553668976,
0.0646878108382225,
-0.023140236735343933,
-0.020116696134209633,
-0.05623495206236839,
0.010586529970169067
] |
InstaSHAP: Interpretable Additive Models Explain Shapley Values Instantly | https://openreview.net/forum?id=ky7vVlBQBY | [
"James Enouen",
"Yan Liu"
] | Poster | In recent years, the Shapley value and SHAP explanations have emerged as one
of the most dominant paradigms for providing post-hoc explanations of blackbox models. Despite their well-founded theoretical properties, many recent works
have focused on the limitations in both their computational efficiency and their
representation power. The underlying connection with additive models, however,
is left critically under-emphasized in the current literature. In this work, we find
that a variational perspective linking GAM models and SHAP explanations is able
to provide deep insights into nearly all recent developments. In light of this connection, we borrow in the other direction to develop a new method to train interpretable GAM models which are automatically purified to compute the Shapley
value in a single forward pass. Finally, we provide theoretical results showing the
limited representation power of GAM models is the same Achilles’ heel existing
in SHAP and discuss the implications for SHAP’s modern usage in CV and NLP. | additive models, GAM, SHAP, Shapley | null | 1,649 | 2502.14177 | [
-0.05583060532808304,
-0.017971383407711983,
0.019217537716031075,
0.019429754465818405,
0.05856652930378914,
0.05063159391283989,
0.03704666346311569,
-0.07023701071739197,
0.06033222749829292,
-0.055472783744335175,
-0.007143467199057341,
-0.030436109751462936,
0.04689226672053337,
0.012236474081873894,
0.017845652997493744,
-0.02329210564494133,
0.08637029677629471,
0.03986498340964317,
-0.13107502460479736,
-0.046889737248420715,
0.008315043523907661,
-0.014375001192092896,
-0.025310352444648743,
0.030480656772851944,
0.03950967267155647,
-0.060693785548210144,
0.01923529990017414,
0.022887758910655975,
0.0235176682472229,
-0.03059905767440796,
-0.008719285950064659,
0.038789380341768265,
-0.053135570138692856,
0.04840868338942528,
-0.08595000952482224,
0.05871107056736946,
-0.03176155686378479,
0.02830202504992485,
-0.04591124504804611,
-0.032805345952510834,
-0.022885818034410477,
0.02056785859167576,
-0.040867775678634644,
0.004642491228878498,
0.019585464149713516,
-0.034855131059885025,
0.04853859543800354,
-0.07893610000610352,
-0.06395995616912842,
-0.04586721584200859,
-0.059595394879579544,
0.03749614208936691,
-0.053573399782180786,
0.02602395974099636,
0.007734193466603756,
-0.03353721648454666,
-0.051308322697877884,
-0.021929191425442696,
-0.041911911219358444,
-0.07756305485963821,
-0.09103856980800629,
-0.040878769010305405,
-0.028257127851247787,
0.0010737741831690073,
0.06971195340156555,
0.006304480135440826,
0.03264489397406578,
0.02359180524945259,
-0.060734644532203674,
0.07566533237695694,
-0.03656016290187836,
0.049549102783203125,
-0.136891171336174,
-0.030107302591204643,
-0.003203350119292736,
0.08607957512140274,
0.034648943692445755,
0.0375884547829628,
0.049765996634960175,
0.0002384054969297722,
-0.08352911472320557,
-0.05626162886619568,
0.046188049018383026,
0.053162749856710434,
0.059969622641801834,
-0.020911751314997673,
-0.000543045112863183,
0.037945229560136795,
0.0048302519135177135,
0.0766533613204956,
-0.009509248659014702,
0.01017451286315918,
-0.010873024351894855,
-0.0006678429781459272,
0.09187984466552734,
0.024596204981207848,
-0.041663605719804764,
-0.07092613726854324,
0.015872448682785034,
0.09945189952850342,
-0.04010220244526863,
0.09821675717830658,
0.06479553878307343,
-0.06328070908784866,
0.07789447158575058,
-0.051893554627895355,
0.029961181804537773,
-0.00716044008731842,
-0.00778130954131484,
-0.028256533667445183,
0.05415207892656326,
0.01083845179527998,
-0.0012413443764671683,
-0.03706207871437073,
-0.013501672074198723,
0.04118131846189499,
0.0024508165661245584,
-0.07175583392381668,
0.0031555346213281155,
-0.0033153905533254147,
0.03976193815469742,
0.019800076261162758,
0.04849620908498764,
0.09355577826499939,
-0.07247567176818848,
-0.052727680653333664,
-0.11010017991065979,
1.5747521091050314e-33,
0.04178394749760628,
0.034389689564704895,
0.06760626286268234,
-0.027638332918286324,
0.06859292834997177,
0.026493888348340988,
-0.09775242954492569,
0.025153273716568947,
-0.018385423347353935,
0.006124863866716623,
-0.04902678355574608,
0.10276110470294952,
-0.004274986684322357,
0.018108122050762177,
0.0654168576002121,
0.008497473783791065,
-0.048593077808618546,
0.027777569368481636,
-0.042717210948467255,
-0.022943785414099693,
0.0548570342361927,
-0.03216143697500229,
0.018511302769184113,
-0.05985954776406288,
0.013474796898663044,
0.050221092998981476,
0.03422418236732483,
-0.01828429289162159,
-0.08987978845834732,
0.0243530236184597,
-0.05551993101835251,
0.015053349547088146,
-0.01095514465123415,
0.014109382405877113,
0.02450123056769371,
-0.011401963420212269,
-0.0230094064027071,
-0.06564300507307053,
0.07331036031246185,
-0.08445262908935547,
0.0456748865544796,
-0.012627639807760715,
0.012570700608193874,
-0.03262504190206528,
-0.134257972240448,
0.001158117433078587,
-0.002089728368446231,
0.006618825253099203,
-0.04409055411815643,
-0.020815899595618248,
0.03229425847530365,
0.042808897793293,
-0.09181609749794006,
-0.01843852549791336,
-0.06362945586442947,
0.02614370360970497,
-0.030334999784827232,
-0.00034704976133070886,
-0.018242565914988518,
0.08462641388177872,
0.010461297817528248,
0.09880278259515762,
0.06425090879201889,
0.005676747765392065,
-0.038609106093645096,
0.032212723046541214,
-0.07360035926103592,
0.057917993515729904,
0.09449462592601776,
0.030739791691303253,
-0.06261280924081802,
0.03210531175136566,
-0.05713159590959549,
0.020004913210868835,
0.013596284203231335,
-0.07044285535812378,
0.018720025196671486,
-0.10681968927383423,
0.04916580021381378,
0.05577527731657028,
-0.06217174604535103,
0.007746207993477583,
-0.005054530221968889,
-0.013026142492890358,
-0.05316883325576782,
-0.047564782202243805,
0.011623053811490536,
-0.04341863840818405,
-0.05129105970263481,
-0.10592897981405258,
-0.07310869544744492,
0.010846822522580624,
0.022823020815849304,
-0.03080230951309204,
-0.04085506498813629,
-3.402522765068837e-33,
-0.04329179599881172,
-0.049366146326065063,
-0.06560453772544861,
0.08514609932899475,
-0.04167461395263672,
-0.09282919764518738,
0.008883821777999401,
0.003747932380065322,
0.029439467936754227,
0.03526739403605461,
0.011046316474676132,
0.01913374289870262,
0.0752670168876648,
0.029428327456116676,
0.04822715371847153,
-0.06143426522612572,
-0.0641830563545227,
-0.06074589490890503,
0.06657776236534119,
0.05341286584734917,
0.0624203085899353,
0.07341825217008591,
-0.07523632049560547,
-0.008200128562748432,
-0.03267209231853485,
0.07116164267063141,
-0.05471379682421684,
0.009456801228225231,
0.0057879905216395855,
0.03892458230257034,
-0.0801018476486206,
0.02388399839401245,
-0.09868433326482773,
-0.08513152599334717,
-0.08407726138830185,
-0.002299845451489091,
0.04309234395623207,
0.045016080141067505,
-0.01671796664595604,
0.062398761510849,
0.07987583428621292,
-0.039305489510297775,
-0.03758217766880989,
0.038391415029764175,
0.004380892962217331,
0.012836418114602566,
0.02521469071507454,
0.07793585956096649,
0.015779055655002594,
-0.03787776082754135,
0.023031705990433693,
0.00547243095934391,
-0.032647013664245605,
0.10961882025003433,
-0.057431772351264954,
-0.014543426223099232,
-0.00836928840726614,
-0.05175124108791351,
0.0027657162863761187,
0.011259859427809715,
-0.08532265573740005,
0.012587965466082096,
0.0031357610132545233,
-0.008501030504703522,
0.0030550335068255663,
0.0034640906378626823,
0.029915712773799896,
-0.09076172858476639,
0.03694823756814003,
-0.0982452780008316,
0.009457752108573914,
-0.05435355007648468,
-0.05358510836958885,
0.051787689328193665,
0.020775554701685905,
-0.0008578327251598239,
-0.009054651483893394,
-0.0008996743126772344,
0.052873533219099045,
-0.02649618871510029,
0.07797833532094955,
0.026793792843818665,
0.025749288499355316,
0.036891575902700424,
0.06087789684534073,
-0.12197642028331757,
-0.01959414593875408,
0.061197709292173386,
0.01524399220943451,
0.010724972933530807,
-0.06853614002466202,
0.09782165288925171,
-0.06981859356164932,
0.12776415050029755,
-0.10404198616743088,
-3.980846230433599e-8,
-0.0622439980506897,
0.0868367925286293,
-0.015996374189853668,
0.0714709535241127,
0.09273192286491394,
-0.05863562226295471,
-0.043036967515945435,
-0.012162616476416588,
-0.02626841329038143,
-0.024213148280978203,
0.06939594447612762,
0.025998026132583618,
-0.03577122837305069,
-0.061151083558797836,
0.002960559679195285,
0.12097709625959396,
0.02634105272591114,
0.014887901954352856,
-0.02056787721812725,
0.025652598589658737,
0.003039322094991803,
0.05825750529766083,
-0.01489868201315403,
0.01295559573918581,
0.07414573431015015,
-0.008722237311303616,
-0.0779539942741394,
0.08190341293811798,
0.0356772281229496,
0.09160418063402176,
0.01572292484343052,
0.051707923412323,
0.046598535031080246,
-0.004819776397198439,
-0.007672780193388462,
0.08986587822437286,
-0.023900099098682404,
-0.009700805880129337,
0.047905731946229935,
-0.05105016753077507,
-0.022196004167199135,
0.022801553830504417,
-0.08936619013547897,
-0.019230332225561142,
0.034079983830451965,
-0.000344409781973809,
0.006452560890465975,
-0.11466523259878159,
-0.04972568526864052,
0.09016103297472,
-0.006727423518896103,
0.004153057932853699,
0.08037100732326508,
0.05034634843468666,
-0.0009378056274726987,
-0.013492796570062637,
0.015049233101308346,
-0.08839356899261475,
0.020000556483864784,
0.03656548261642456,
0.03093927912414074,
-0.005918759386986494,
-0.014108576811850071,
0.002503616502508521
] |
Watermark Anything With Localized Messages | https://openreview.net/forum?id=IkZVDzdC8M | [
"Tom Sander",
"Pierre Fernandez",
"Alain Oliviero Durmus",
"Teddy Furon",
"Matthijs Douze"
] | Poster | Image watermarking methods are not tailored to handle small watermarked areas.
This restricts applications in real-world scenarios where parts of the image may come from different sources or have been edited.
We introduce a deep-learning model for localized image watermarking, dubbed the Watermark Anything Model (WAM).
The WAM embedder imperceptibly modifies the input image, while the extractor segments the received image into watermarked and non-watermarked areas and recovers one or several hidden messages from the areas found to be watermarked.
The models are jointly trained at low resolution and without perceptual constraints, then post-trained for imperceptibility and multiple watermarks.
Experiments show that WAM is competitive with state-of-the art methods in terms of imperceptibility and robustness, especially against inpainting and splicing, even on high-resolution images.
Moreover, it offers new capabilities: WAM can locate watermarked areas in spliced images and extract distinct 32-bit messages with less than 1 bit error from multiple small regions -- no larger than 10\% of the image surface -- even for small $256\times 256$ images.
Training and inference code and model weights are available at https://github.com/facebookresearch/watermark-anything. | Image Watermarking; Segmentation | Watermark Anything Models (WAM), the first deep-learning approach for localized image watermarking that can handle small watermarked areas, inpainting, splicing, edited images and multiple watermarks in a single image. | 1,647 | 2411.07231 | [
-0.042441606521606445,
-0.003903077682480216,
0.07959260791540146,
0.05123169720172882,
0.16782081127166748,
-0.017076408490538597,
0.003631507745012641,
-0.0916827917098999,
0.033545155078172684,
-0.06514297425746918,
-0.0025658193044364452,
-0.0009320419630967081,
0.10094691812992096,
0.06727512180805206,
-0.08286703377962112,
-0.0007331609376706183,
0.013281144201755524,
0.07384584099054337,
-0.1090482547879219,
-0.06111922860145569,
0.008760900236666203,
-0.08867322653532028,
0.049566600471735,
-0.031775522977113724,
0.05264945700764656,
0.043254051357507706,
0.00586259551346302,
-0.010388323105871677,
-0.019213499501347542,
-0.031219417229294777,
0.07130680978298187,
-0.08118192851543427,
-0.0027717468328773975,
0.005354367662221193,
0.0017455255147069693,
0.08372049033641815,
-0.023053424432873726,
0.03624418005347252,
0.07221686840057373,
-0.05314388871192932,
0.014584055170416832,
-0.020221086218953133,
-0.04536975547671318,
0.022072825580835342,
0.06143900379538536,
0.012681501917541027,
0.10079021006822586,
0.056937504559755325,
-0.04825473576784134,
0.023568350821733475,
0.010358116589486599,
-0.04088183119893074,
-0.08348975330591202,
0.0791713073849678,
-0.02997940219938755,
-0.0914410874247551,
-0.03400197997689247,
0.015054065734148026,
-0.02276708371937275,
0.020041346549987793,
0.013863896951079369,
-0.04044490307569504,
0.023692533373832703,
-0.018542176112532616,
0.0389552116394043,
-0.03584451973438263,
-0.014227412641048431,
0.003353566164150834,
-0.016507798805832863,
-0.04295452684164047,
0.04651021584868431,
0.05866406857967377,
-0.030243558809161186,
0.06265965104103088,
-0.006740529555827379,
0.07491908967494965,
0.00002671464972081594,
0.050824981182813644,
0.021474642679095268,
-0.016907285898923874,
0.044061772525310516,
-0.02931503765285015,
0.06349904090166092,
0.0008974561933428049,
0.14007219672203064,
-0.04556139186024666,
-0.038793448358774185,
0.04084164649248123,
0.021025074645876884,
-0.0351746529340744,
-0.10224974155426025,
-0.0359402671456337,
0.040433119982481,
-0.00901914481073618,
0.0460190586745739,
0.01706916280090809,
-0.014903025701642036,
-0.004030583892017603,
-0.049130458384752274,
0.05605652183294296,
-0.05273160710930824,
-0.031466029584407806,
0.010802600532770157,
-0.07579858601093292,
0.16062289476394653,
0.07114328444004059,
0.0002832658647093922,
0.019584646448493004,
0.07304642349481583,
-0.031009208410978317,
0.015077966265380383,
0.07329359650611877,
-0.006170020438730717,
-0.011415638960897923,
0.046867139637470245,
-0.04175738990306854,
-0.04059389606118202,
-0.03349043056368828,
-0.0301029235124588,
-0.02538718283176422,
-0.03900489956140518,
-0.030990131199359894,
-0.03537142649292946,
0.07349648326635361,
-0.05094130337238312,
-0.021388376131653786,
-0.059602558612823486,
4.117431613274074e-33,
0.06549046188592911,
-0.007114004343748093,
-0.02288910001516342,
-0.048369042575359344,
0.08374513685703278,
0.023026268929243088,
-0.028209934011101723,
0.029915237799286842,
-0.018014907836914062,
-0.04394840821623802,
0.00874012429267168,
-0.09073761850595474,
-0.02246382087469101,
0.11242105066776276,
0.045532241463661194,
0.026812372729182243,
-0.03275512531399727,
-0.0456402450799942,
0.03508242219686508,
0.0032737795263528824,
-0.026056557893753052,
-0.07315987348556519,
-0.06594306230545044,
-0.007798910140991211,
0.01532582938671112,
0.0012980205938220024,
0.10085482150316238,
-0.0355599969625473,
0.030366279184818268,
-0.010602342896163464,
-0.06533057242631912,
-0.021974697709083557,
0.08207803219556808,
0.021877586841583252,
0.004182495176792145,
-0.010240413248538971,
0.037703968584537506,
-0.00855123158544302,
0.013692209497094154,
0.05024800822138786,
0.020153988152742386,
0.0267091803252697,
-0.05084478482604027,
-0.010035242885351181,
-0.07216048985719681,
0.004855250474065542,
0.01911018230021,
0.03604360297322273,
-0.08132535219192505,
-0.009487377479672432,
0.03367605805397034,
0.04886895418167114,
-0.010195841081440449,
-0.0682007297873497,
-0.1073484793305397,
-0.0223776176571846,
0.03544698655605316,
-0.08235333859920502,
0.017545251175761223,
0.0008794231107458472,
-0.02153736725449562,
0.051754578948020935,
0.024419071152806282,
0.07880184054374695,
-0.026965806260704994,
-0.05916402488946915,
0.08358938246965408,
0.03282582387328148,
-0.03181939199566841,
-0.05101833492517471,
-0.03335443511605263,
0.017713962122797966,
0.04141898453235626,
0.029080364853143692,
-0.07817663252353668,
-0.03687126561999321,
0.06260937452316284,
-0.005876263603568077,
-0.02231275476515293,
0.05319109559059143,
-0.025845762342214584,
-0.028314175084233284,
0.0591265894472599,
-0.030412035062909126,
-0.08112156391143799,
-0.058558739721775055,
0.02781769447028637,
-0.10383633524179459,
-0.041028738021850586,
-0.0060205920599401,
-0.030003905296325684,
-0.01871701329946518,
0.039340756833553314,
-0.06396893411874771,
0.013348544016480446,
-2.3824247266634247e-33,
-0.014837764203548431,
0.016001710668206215,
-0.048565663397312164,
0.03244183212518692,
-0.006723035592585802,
0.01285172812640667,
0.03219899162650108,
0.0642695501446724,
0.02626456320285797,
-0.052940621972084045,
-0.07110965251922607,
0.0057963114231824875,
0.0007526719127781689,
-0.07436083257198334,
-0.07928845286369324,
-0.01607007533311844,
-0.007237995974719524,
0.10064155608415604,
-0.025390464812517166,
-0.013822280801832676,
0.07359012216329575,
0.033524513244628906,
-0.0075839911587536335,
0.12460887432098389,
-0.05874089524149895,
0.07028212398290634,
0.029492048546671867,
-0.020583871752023697,
-0.04560745507478714,
0.009839894250035286,
-0.04242993891239166,
0.03912920132279396,
0.05816809460520744,
-0.0412895642220974,
-0.03787284344434738,
0.03795001283288002,
0.03527157008647919,
0.015263549983501434,
-0.04438944160938263,
0.015112779103219509,
0.07282737642526627,
-0.017103422433137894,
-0.08855193108320236,
0.048608776181936264,
-0.04686456173658371,
-0.003353897947818041,
-0.027169518172740936,
0.01758441887795925,
0.0476827397942543,
0.0942068099975586,
-0.03686312586069107,
-0.06896614283323288,
-0.06310897320508957,
0.01646769419312477,
-0.04619472101330757,
-0.024325106292963028,
-0.044953104108572006,
0.03500043600797653,
0.06558354198932648,
0.02180328406393528,
-0.014277903363108635,
-0.05052131041884422,
-0.03753483667969704,
-0.06299814581871033,
0.02231728285551071,
-0.009766974486410618,
-0.0434844009578228,
0.044060152024030685,
-0.04164677858352661,
-0.003928706515580416,
0.05136708915233612,
-0.09635081887245178,
0.002464899094775319,
0.06308496743440628,
0.05904946103692055,
0.03436026722192764,
-0.04577629268169403,
0.022467084228992462,
-0.04860060289502144,
-0.014843071810901165,
0.03497824817895889,
-0.06195495277643204,
-0.03697856888175011,
0.09075307101011276,
0.15810352563858032,
0.09113267809152603,
-0.013643038459122181,
-0.06833646446466446,
-0.012554274871945381,
0.028569066897034645,
-0.05670132488012314,
0.1055520623922348,
-0.025208333507180214,
0.028248710557818413,
0.02639087475836277,
-3.914429314022527e-8,
-0.035817600786685944,
-0.007802152540534735,
-0.029051844030618668,
-0.08373672515153885,
-0.03613714873790741,
0.022959642112255096,
-0.014227627776563168,
0.10276384651660919,
0.04295836016535759,
-0.09487665444612503,
0.06787645816802979,
-0.07496102154254913,
-0.13180328905582428,
-0.12389474362134933,
-0.06968116015195847,
-0.005658803507685661,
0.046492233872413635,
0.04006116837263107,
-0.03356979787349701,
-0.024429092183709145,
-0.02875519171357155,
0.061084337532520294,
0.0009095965069718659,
-0.06154116615653038,
0.03360613435506821,
-0.03897222504019737,
0.0826057493686676,
0.08473124355077744,
0.0007305871113203466,
0.04810997098684311,
-0.017231915146112442,
-0.02975231222808361,
0.07149349898099899,
0.03776190057396889,
0.03705412894487381,
0.06539907306432724,
0.06957309693098068,
0.017401838675141335,
-0.054885827004909515,
-0.0029005000833421946,
0.030865315347909927,
-0.0005400626105256379,
0.006250724196434021,
-0.03944004327058792,
0.006965354550629854,
0.016447873786091805,
0.04676806926727295,
-0.0837925374507904,
0.025400305166840553,
0.03835609555244446,
0.029757875949144363,
0.028216972947120667,
-0.008143441751599312,
0.10092417150735855,
0.02511107549071312,
-0.068553127348423,
0.051025837659835815,
-0.023042449727654457,
0.043240658938884735,
0.08165248483419418,
0.06137261912226677,
0.021952485665678978,
0.023536603897809982,
0.02731228619813919
] |
M^3PC: Test-time Model Predictive Control using Pretrained Masked Trajectory Model | https://openreview.net/forum?id=inOwd7hZC1 | [
"Kehan Wen",
"Yutong Hu",
"Yao Mu",
"Lei Ke"
] | Poster | Recent work in Offline Reinforcement Learning (RL) has shown that a unified transformer trained under a masked auto-encoding objective can effectively capture the relationships between different modalities (e.g., states, actions, rewards) within given trajectory datasets. However, this information has not been fully exploited during the inference phase, where the agent needs to generate an optimal policy instead of just reconstructing masked components from unmasked. Given that a pretrained trajectory model can act as both a Policy Model and a World Model with appropriate mask patterns, we propose using Model Predictive Control (MPC) at test time to leverage the model's own predictive capacity to guide its action selection. Empirical results on D4RL and RoboMimic show that our inference-phase MPC significantly improves the decision-making performance of a pretrained trajectory model without any additional parameter training. Furthermore, our framework can be adapted to Offline to Online (O2O) RL and Goal Reaching RL, resulting in more substantial performance gains when an additional online interaction budget is given, and better generalization capabilities when different task targets are specified. Code is available: \href{https://github.com/wkh923/m3pc}{\texttt{https://github.com/wkh923/m3pc}}. | Offline-to-Online Reinforcement Learning, Model-based Reinforcement Learning, Masked Autoencoding, Robot Learning | Enhance Transformer for RL by employing the Model itself for test-time MPC, achieving better performance in offline RL and offline-to-online RL for both simulated and real-world robotic tasks, with additional goal-reaching capabilities. | 1,642 | null | [
-0.08497340977191925,
-0.09312564134597778,
-0.02073555625975132,
0.13843397796154022,
0.06389331817626953,
0.04807836562395096,
-0.036503732204437256,
-0.03520454838871956,
-0.012990597635507584,
0.0475926548242569,
-0.05739610642194748,
-0.055137258023023605,
-0.00968535989522934,
0.10970862954854965,
0.05113039165735245,
-0.003291651140898466,
0.12245526164770126,
0.04840073361992836,
-0.02523374743759632,
-0.03329906985163689,
0.11329535394906998,
0.02780584804713726,
0.08335717767477036,
0.04222307726740837,
0.01439051516354084,
-0.03745158016681671,
0.018905334174633026,
-0.037487518042325974,
-0.04048861190676689,
-0.026118053123354912,
0.06381667405366898,
-0.07107848674058914,
0.027573419734835625,
-0.08632969856262207,
-0.0733993723988533,
-0.0031932347919791937,
-0.07890401035547256,
0.016510652378201485,
-0.006063207518309355,
-0.030344322323799133,
0.03474918007850647,
-0.01360597088932991,
0.04221714287996292,
0.035322386771440506,
0.019714053720235825,
0.04059859737753868,
0.013660346157848835,
-0.019623633474111557,
-0.04298969730734825,
-0.021356303244829178,
0.0012822834542021155,
-0.04630748927593231,
0.014512785710394382,
-0.00230111600831151,
-0.0858258605003357,
0.006276270840317011,
0.07146306335926056,
0.023466046899557114,
0.04232108220458031,
-0.06858781725168228,
-0.10012032836675644,
-0.11320517212152481,
0.00431405333802104,
-0.045034684240818024,
-0.1367284208536148,
0.017718778923153877,
0.00995871052145958,
0.009381860494613647,
0.03500870242714882,
0.0008699868922121823,
-0.037580955773591995,
0.037699826061725616,
-0.0155975092202425,
-0.030352983623743057,
0.07428667694330215,
0.012141411192715168,
0.05239061266183853,
-0.011135163716971874,
0.04296676442027092,
-0.022822756320238113,
0.0614967942237854,
0.008973168209195137,
-0.020153958350419998,
0.01550288312137127,
0.0077551016584038734,
0.033386360853910446,
-0.024138499051332474,
0.08346331119537354,
0.1015036404132843,
-0.007584215607494116,
-0.029900597408413887,
0.030405675992369652,
0.022847022861242294,
0.0029383525252342224,
-0.04223823547363281,
-0.024216344580054283,
0.012460555881261826,
-0.05067240446805954,
-0.049454208463430405,
0.06722371280193329,
0.025758715346455574,
0.04667925834655762,
-0.02234645187854767,
0.019120508804917336,
0.0039363885298371315,
-0.03695286810398102,
0.03860676288604736,
0.03613675385713577,
0.05723806843161583,
0.014966321177780628,
0.04436040669679642,
0.03651910275220871,
0.03286067768931389,
-0.02971935085952282,
0.023080214858055115,
0.09584417939186096,
-0.008328519761562347,
-0.018155574798583984,
-0.048381637781858444,
0.06604151427745819,
-0.013448934070765972,
-0.1050930768251419,
0.07701873034238815,
0.0029238243587315083,
0.05726538598537445,
-0.04119572788476944,
-0.07049085944890976,
2.1920763754217947e-33,
0.003512610914185643,
-0.029603775590658188,
0.034406810998916626,
0.028526509180665016,
0.009838989935815334,
0.017180148512125015,
0.08063555508852005,
-0.0005009084125049412,
0.020956723019480705,
-0.02510053664445877,
0.014357561245560646,
-0.022590018808841705,
-0.05603354424238205,
0.14553958177566528,
0.0072885616682469845,
0.03493642434477806,
-0.030247611925005913,
0.06200633943080902,
0.01746664196252823,
-0.03715723007917404,
0.08564050495624542,
0.007404252886772156,
-0.01886972039937973,
-0.08511002361774445,
-0.0024063424207270145,
-0.031499505043029785,
-0.00165202422067523,
-0.06694656610488892,
0.010771268978714943,
0.01386337075382471,
-0.022870458662509918,
-0.0005637929425574839,
-0.03344682231545448,
0.04257526993751526,
0.025516822934150696,
0.013412108644843102,
-0.07557535916566849,
-0.11694914847612381,
-0.016902979463338852,
0.07527612149715424,
-0.05116156116127968,
-0.02626536600291729,
-0.010951395146548748,
-0.01661837100982666,
-0.07849825173616409,
-0.1755094975233078,
0.01849197968840599,
0.018335560336709023,
-0.12575073540210724,
0.052422963082790375,
0.06850812584161758,
0.044499803334474564,
-0.05038939043879509,
-0.11568433046340942,
-0.0396173819899559,
0.01583804003894329,
-0.0053392560221254826,
0.020181698724627495,
0.02523311786353588,
0.02916531451046467,
-0.013107671402394772,
-0.015700707212090492,
0.030701888725161552,
0.048508014529943466,
-0.019946428015828133,
0.027597110718488693,
-0.030713394284248352,
0.04699642211198807,
0.009526235982775688,
-0.00971216894686222,
0.010387733578681946,
-0.03181000053882599,
0.021749181672930717,
-0.05487705022096634,
0.03818020597100258,
-0.018011977896094322,
0.1253865510225296,
-0.02742590755224228,
-0.015728218480944633,
-0.004998994991183281,
0.00887664221227169,
0.016816651448607445,
0.010505855083465576,
-0.03231259807944298,
0.04045730456709862,
-0.08233591914176941,
0.03231244161725044,
-0.06428682059049606,
-0.03801592066884041,
-0.04389573633670807,
-0.09493909031152725,
0.008678188547492027,
-0.05773526430130005,
0.033033933490514755,
0.04134823754429817,
-2.1006199761919437e-33,
0.04402272775769234,
0.01072621438652277,
0.02132718451321125,
0.046348873525857925,
-0.047504451125860214,
-0.0727338120341301,
-0.004801856819540262,
-0.012944932095706463,
0.008336436003446579,
-0.0481235571205616,
-0.05014448240399361,
-0.05637379363179207,
0.03987254947423935,
-0.005179582163691521,
0.039549682289361954,
-0.07071647047996521,
0.011285747401416302,
-0.005278661381453276,
0.04026690870523453,
0.003464038949459791,
0.020402023568749428,
-0.008728643879294395,
-0.08810396492481232,
0.01972039043903351,
-0.08756802976131439,
-0.011617662385106087,
-0.012630137614905834,
0.08850832283496857,
0.022781401872634888,
0.05573714151978493,
-0.0631619468331337,
0.05281350761651993,
0.03248089924454689,
-0.011839260347187519,
-0.09066886454820633,
0.11531791090965271,
-0.008844356052577496,
0.0628795176744461,
-0.05110543966293335,
0.14658430218696594,
0.0033585031051188707,
-0.021984757855534554,
-0.07088852673768997,
0.07259850203990936,
-0.05850669741630554,
-0.012189943343400955,
-0.03617819398641586,
0.04681514576077461,
0.09621448069810867,
0.009613938629627228,
-0.042179711163043976,
0.04370609298348427,
-0.08282453566789627,
-0.013228204101324081,
-0.03208420053124428,
-0.04952152073383331,
-0.010182252153754234,
-0.028958041220903397,
0.058372750878334045,
-0.05691736936569214,
0.009251340292394161,
0.0513051375746727,
-0.01495265867561102,
-0.017179761081933975,
0.04471208155155182,
-0.0439176931977272,
0.002072376199066639,
0.028083819895982742,
0.06691119074821472,
0.003909608349204063,
-0.0483921617269516,
-0.032502420246601105,
0.0721692144870758,
-0.016415920108556747,
0.07968367636203766,
-0.06666794419288635,
-0.03920375555753708,
0.027706436812877655,
0.002276652492582798,
0.033812351524829865,
-0.02203848771750927,
0.021299926564097404,
0.007442887872457504,
-0.014959841035306454,
0.021502969786524773,
0.07914148271083832,
-0.03691995143890381,
0.13416078686714172,
0.02159198746085167,
-0.044664379209280014,
-0.04841148108243942,
0.03345652297139168,
0.02931603603065014,
0.05583425611257553,
0.021059714257717133,
-4.924663699057419e-8,
-0.015861039981245995,
0.09819188714027405,
0.06622408330440521,
-0.009721285663545132,
-0.017475059255957603,
-0.00828165840357542,
0.03209304437041283,
-0.0039028015453368425,
-0.002720523625612259,
0.06328073889017105,
0.11572599411010742,
0.00004722154699265957,
0.022864574566483498,
-0.06873694807291031,
-0.006364718079566956,
0.05421385541558266,
0.004411227069795132,
-0.012551111169159412,
0.00799840409308672,
0.03379136323928833,
0.01132705807685852,
-0.019592804834246635,
-0.028906546533107758,
-0.05413413420319557,
0.08136504888534546,
-0.015775606036186218,
-0.03568987920880318,
0.11136788129806519,
0.03301502391695976,
0.09552957117557526,
0.03576714172959328,
-0.07882076501846313,
0.0523211695253849,
-0.044904373586177826,
-0.023265236988663673,
0.023061424493789673,
0.049064092338085175,
-0.0698976069688797,
0.007986540906131268,
0.02887094020843506,
-0.01677503064274788,
0.03968598693609238,
-0.04412401095032692,
-0.018641790375113487,
-0.0351516492664814,
0.0016661949921399355,
-0.07883494347333908,
-0.08246450871229172,
0.010240260511636734,
-0.06514710932970047,
-0.002613473916426301,
-0.05260424688458443,
-0.03265341371297836,
0.07626070827245712,
0.13883104920387268,
0.03983857482671738,
0.02009098418056965,
-0.06073751300573349,
0.03659585118293762,
0.07577811181545258,
-0.009008718654513359,
-0.05498994514346123,
-0.11818784475326538,
-0.011159182526171207
] |
TEASER: Token Enhanced Spatial Modeling for Expressions Reconstruction | https://openreview.net/forum?id=pTeOOKnjGM | [
"Yunfei Liu",
"Lei Zhu",
"Lijian Lin",
"Ye Zhu",
"Ailing Zhang",
"Yu Li"
] | Poster | 3D facial reconstruction from a single in-the-wild image is a crucial task in human-centered computer vision tasks. While existing methods can recover accurate facial shapes, there remains significant space for improvement in fine-grained expression capture. Current approaches struggle with irregular mouth shapes, exaggerated expressions, and asymmetrical facial movements. We present TEASER (Token EnhAnced Spatial modeling for Expressions Reconstruction), which addresses these challenges and enhances 3D facial geometry performance. TEASER tackles two main limitations of existing methods: insufficient photometric loss for self-reconstruction and inaccurate localization of subtle expressions. We introduce a multi-scale tokenizer to extract facial appearance information. Combined with a neural renderer, these tokens provide precise geometric guidance for expression reconstruction. Furthermore, TEASER incorporates a pose-dependent landmark loss to further improve geometric performance. Our approach not only significantly enhances expression reconstruction quality but also offers interpretable tokens suitable for various downstream applications, such as photorealistic facial video driving, expression transfer, and identity swapping. Quantitative and qualitative experimental results across multiple datasets demonstrate that TEASER achieves state-of-the-art performance in precise expression reconstruction. | Expression reconstruction, Hybrid parameters | TEASER reconstructs precise 3D facial expression and generates high-fidelity face image through estimating hybrid parameters for 3D facial reconstruction. | 1,633 | 2502.10982 | [
-0.10962767153978348,
-0.008780425414443016,
0.08833087980747223,
0.017545660957694054,
0.032156072556972504,
0.05422905460000038,
-0.04513182118535042,
-0.03806878626346588,
0.031937528401613235,
-0.06369956582784653,
-0.061434388160705566,
-0.09332538396120071,
-0.021658530458807945,
0.059464167803525925,
0.07866188883781433,
-0.0046219127252697945,
0.005538115743547678,
0.16225086152553558,
-0.03091140277683735,
0.0014779448974877596,
0.04865793511271477,
-0.029286349192261696,
0.028234099969267845,
0.007394419517368078,
0.07912179082632065,
0.0030386364087462425,
0.01958315446972847,
-0.08771570026874542,
0.047453831881284714,
-0.015370327979326248,
-0.025313401594758034,
-0.053516242653131485,
-0.0020997102838009596,
0.0008763612131588161,
-0.06163893640041351,
0.06568976491689682,
-0.0030392571352422237,
0.04121873155236244,
0.021234797313809395,
-0.04627752676606178,
0.012087048962712288,
0.04626118764281273,
-0.008131696842610836,
0.04804433509707451,
0.05774639546871185,
0.05579543858766556,
0.10484249144792557,
-0.04913373664021492,
0.009478544816374779,
-0.062116559594869614,
-0.009443553164601326,
-0.06348948925733566,
0.02193196304142475,
-0.003530403831973672,
-0.0266755148768425,
0.045545659959316254,
0.043135423213243484,
-0.08786790817975998,
0.04426424950361252,
0.014787761494517326,
0.03178587555885315,
-0.03722720220685005,
0.058072444051504135,
0.03843473643064499,
-0.03819986432790756,
-0.05269467085599899,
0.029963217675685883,
-0.061208147555589676,
-0.01987156644463539,
0.03154712915420532,
-0.0720057263970375,
-0.03792773187160492,
-0.04400378093123436,
-0.019758226349949837,
0.021607467904686928,
0.030969809740781784,
0.08016032725572586,
-0.0005081749404780567,
0.03282884880900383,
0.007243165746331215,
0.032811593264341354,
-0.014090030454099178,
0.026768550276756287,
-0.008434929884970188,
-0.022295793518424034,
-0.037185732275247574,
-0.06293982267379761,
-0.027605688199400902,
-0.03670042008161545,
0.09396656602621078,
-0.08961973339319229,
-0.007988319732248783,
0.0004858903703279793,
-0.06556037068367004,
-0.004857741761952639,
-0.05902383103966713,
-0.030447350814938545,
-0.04884794354438782,
0.04925013333559036,
0.09800096601247787,
0.003846763866022229,
0.018114963546395302,
0.03445691242814064,
-0.05415260046720505,
0.062135305255651474,
-0.02594953030347824,
0.01886369101703167,
-0.01357768103480339,
0.0801909789443016,
0.004488693550229073,
0.03724333643913269,
-0.028020771220326424,
0.002293089171871543,
0.04110139608383179,
0.08195526897907257,
-0.020672427490353584,
-0.056284140795469284,
-0.03599398210644722,
0.04959028586745262,
0.016551857814192772,
0.0067785996943712234,
-0.023259466513991356,
-0.02418735809624195,
0.0022657853551208973,
-0.002840649802237749,
0.08626994490623474,
-0.03231373801827431,
1.7155750466365952e-33,
0.020880384370684624,
0.07518567144870758,
0.0501369945704937,
0.008421938866376877,
0.006193954963237047,
0.011875477619469166,
-0.0462125688791275,
-0.0450197272002697,
-0.046849966049194336,
0.020394379273056984,
-0.035592060536146164,
-0.06167539581656456,
-0.04460425302386284,
0.16350118815898895,
-0.036275964230298996,
0.008601815439760685,
-0.01796342059969902,
0.02227284573018551,
-0.031425945460796356,
0.04622282460331917,
-0.027013659477233887,
-0.04573052003979683,
-0.045811716467142105,
-0.011544708162546158,
-0.04031115397810936,
0.01052854023873806,
0.05324351042509079,
-0.035772740840911865,
-0.03777877613902092,
-0.03797855228185654,
-0.07601653039455414,
0.03678650036454201,
0.045145273208618164,
-0.021045846864581108,
0.09361916780471802,
-0.05341818183660507,
0.08311393111944199,
-0.11061561107635498,
-0.010762719437479973,
0.03559773415327072,
-0.03377332538366318,
0.05812383070588112,
-0.03924969956278801,
-0.0282878577709198,
-0.08154775947332382,
0.010835233144462109,
0.020907839760184288,
0.06863950192928314,
-0.02810276299715042,
-0.007211956195533276,
0.0796269103884697,
0.11149615794420242,
0.0017477410146966577,
-0.014850527048110962,
-0.02492506057024002,
-0.017781732603907585,
-0.057177573442459106,
-0.06006350740790367,
0.05494881793856621,
0.020881595090031624,
-0.023208439350128174,
0.0025570394936949015,
0.01997310109436512,
0.013006781227886677,
-0.0461575873196125,
0.045164577662944794,
-0.0006542095215991139,
0.048153236508369446,
0.02524205856025219,
0.07291288673877716,
0.022068217396736145,
0.0654660165309906,
0.0032496298663318157,
-0.06704656779766083,
0.023464705795049667,
0.01403321884572506,
0.10588137805461884,
-0.04990791529417038,
-0.02205326035618782,
0.05492695793509483,
-0.04620097205042839,
0.0005888615269213915,
-0.030143337324261665,
-0.0017389549175277352,
0.01595783233642578,
-0.06854893267154694,
0.04964454472064972,
-0.12169481068849564,
0.015096725896000862,
-0.016839956864714622,
-0.04963141307234764,
-0.054403576999902725,
-0.02020350657403469,
0.02861277014017105,
-0.06541988998651505,
-7.783234623542379e-34,
0.013035311363637447,
0.04515165090560913,
-0.09119784832000732,
-0.0956907868385315,
-0.018779737874865532,
0.0032373920548707247,
0.03509875386953354,
0.03204207494854927,
0.021595215424895287,
-0.04597455635666847,
-0.07434500008821487,
-0.056768544018268585,
0.01691489852964878,
-0.07655360549688339,
0.005082509480416775,
0.0026008407585322857,
0.011932898312807083,
-0.014515739865601063,
-0.05287918075919151,
0.0035525846760720015,
0.06019977107644081,
0.06400420516729355,
-0.022035561501979828,
-0.01042187213897705,
-0.10283824801445007,
0.09280894696712494,
-0.05475512892007828,
0.06125611439347267,
-0.03088129311800003,
0.008741019293665886,
0.022433355450630188,
-0.017313694581389427,
0.03627832233905792,
-0.0075638173148036,
-0.05984257161617279,
-0.013470705598592758,
0.004663066007196903,
-0.02822914905846119,
-0.03607119247317314,
-0.009564010426402092,
0.055824700742959976,
-0.057433996349573135,
-0.01960425078868866,
0.08354321867227554,
0.011142555624246597,
0.02260318212211132,
-0.034950632601976395,
-0.050088509917259216,
-0.053306691348552704,
0.12900696694850922,
-0.062264055013656616,
0.047576744109392166,
-0.10599114745855331,
0.044881753623485565,
-0.0878467708826065,
-0.09622123837471008,
-0.024081915616989136,
-0.032225798815488815,
0.05131224915385246,
-0.02712629921734333,
-0.05331139639019966,
-0.054722730070352554,
-0.04070669412612915,
0.02830064482986927,
-0.01943610981106758,
0.07247993350028992,
0.045399587601423264,
0.008031709119677544,
0.04033764824271202,
0.042488016188144684,
0.055970579385757446,
0.015526080504059792,
0.018192734569311142,
0.03217041864991188,
0.03202180936932564,
-0.0037479561287909746,
0.02050761878490448,
0.011568870395421982,
-0.015422175638377666,
0.0034757249522954226,
-0.01331364270299673,
-0.07196483016014099,
0.04950243979692459,
0.08183341473340988,
0.11097843945026398,
0.08830836415290833,
-0.010679543018341064,
0.08575134724378586,
-0.035599905997514725,
0.08871149271726608,
-0.09282314777374268,
0.11613238602876663,
-0.04351094365119934,
0.007402508053928614,
0.04810808598995209,
-3.764935385675017e-8,
-0.11069031059741974,
0.05645812675356865,
0.009906142950057983,
-0.006093089003115892,
-0.03901594132184982,
0.027539225295186043,
0.01756441779434681,
-0.027675453573465347,
-0.06532249599695206,
-0.07751499861478806,
0.01383234653621912,
-0.0316569060087204,
-0.002484608208760619,
-0.08154917508363724,
-0.015513363294303417,
0.02183905802667141,
0.0125903794541955,
0.05401962250471115,
-0.047241874039173126,
0.013434645719826221,
-0.13414163887500763,
0.0957546979188919,
-0.015321439132094383,
-0.07355471700429916,
0.005629838444292545,
-0.022801518440246582,
0.005916103255003691,
0.12631875276565552,
-0.017573652788996696,
-0.021867936477065086,
-0.010087727569043636,
0.07440780848264694,
0.044908713549375534,
-0.021121729165315628,
-0.019128786399960518,
-0.007972346618771553,
0.01397702470421791,
-0.027668975293636322,
0.007465248927474022,
0.010220413096249104,
-0.05012006312608719,
0.04940730333328247,
0.03597673401236534,
0.023887036368250847,
-0.0257241353392601,
0.04711131006479263,
0.06154035031795502,
-0.07146835327148438,
0.06600672751665115,
0.02503373473882675,
-0.01955466717481613,
-0.04038671776652336,
-0.04603362828493118,
0.15890568494796753,
0.07432293891906738,
0.03076820820569992,
-0.07212112098932266,
0.04818761348724365,
0.003738504834473133,
0.1468995213508606,
0.03039117343723774,
-0.008128325454890728,
-0.03915991634130478,
-0.050450701266527176
] |
SurFhead: Affine Rig Blending for Geometrically Accurate 2D Gaussian Surfel Head Avatars | https://openreview.net/forum?id=1x1gGg49jr | [
"Jaeseong Lee",
"Taewoong Kang",
"Marcel Buehler",
"Min-Jung Kim",
"Sungwon Hwang",
"Junha Hyung",
"Hyojin Jang",
"Jaegul Choo"
] | Poster | Recent advancements in head avatar rendering using Gaussian primitives have achieved significantly high-fidelity results. Although precise head geometry is crucial for applications like mesh reconstruction and relighting, current methods struggle to capture intricate geometric details and render unseen poses due to their reliance on similarity transformations, which cannot handle stretch and shear transforms essential for detailed deformations of geometry. To address this, we propose SurFhead, a novel method that reconstructs riggable head geometry from RGB videos using 2D Gaussian surfels, which offer well-defined geometric properties, such as precise depth from fixed ray intersections and normals derived from their surface orientation, making them advantageous over 3D counterparts. SurFhead ensures high-fidelity rendering of both normals and images, even in extreme poses, by leveraging classical mesh-based deformation transfer and affine transformation interpolation. SurFhead introduces precise geometric deformation and blends surfels through polar decomposition of transformations, including those affecting normals. Our key contribution lies in bridging classical graphics techniques, such as mesh-based deformation, with modern Gaussian primitives, achieving state-of-the-art geometry reconstruction and rendering quality. Unlike previous avatar rendering approaches, SurFhead enables efficient reconstruction driven by Gaussian primitives while preserving high-fidelity geometry. | dynamic head avatars, rigging, inverse-graphics | null | 1,632 | 2410.11682 | [
-0.05630940571427345,
0.01945350132882595,
0.04142846539616585,
-0.05639614164829254,
0.06284041702747345,
-0.048889413475990295,
-0.004969940986484289,
-0.05805433541536331,
0.008722781203687191,
-0.03305555135011673,
0.007170556113123894,
-0.05930735170841217,
0.00882240291684866,
0.05358678475022316,
-0.016497306525707245,
-0.04045352339744568,
-0.028929846361279488,
0.12973332405090332,
-0.06294242292642593,
0.08617038279771805,
0.03273271396756172,
-0.0023016284685581923,
-0.051046498119831085,
-0.039015911519527435,
0.0392945297062397,
0.03912271559238434,
0.07463501393795013,
-0.07255399972200394,
0.0008672574767842889,
-0.03912428021430969,
0.039545588195323944,
-0.0013685033190995455,
-0.05788711458444595,
-0.054576095193624496,
-0.09364723414182663,
0.0038516882341355085,
0.00785741861909628,
0.0544191338121891,
-0.06617588549852371,
-0.03134861961007118,
-0.05654444172978401,
0.05870860442519188,
-0.009015138261020184,
0.017647897824645042,
0.06848915666341782,
0.07395171374082565,
-0.0032728398218750954,
-0.007265699096024036,
0.05356137827038765,
0.008391554467380047,
-0.01710711233317852,
-0.0887066051363945,
0.016885729506611824,
-0.001228123321197927,
-0.01238106656819582,
0.07919476181268692,
-0.0005051666521467268,
-0.14996302127838135,
0.042034924030303955,
0.049094390124082565,
0.025691591203212738,
0.010364688001573086,
0.07464323192834854,
0.04016565904021263,
-0.025078045204281807,
-0.08113421499729156,
0.008915025740861893,
-0.07930898666381836,
-0.035815801471471786,
0.05146908015012741,
-0.1781504899263382,
-0.017056070268154144,
-0.040703725069761276,
-0.06672333180904388,
-0.01678672432899475,
0.03651462867856026,
0.006805104669183493,
-0.011384410783648491,
-0.02609947882592678,
-0.021026866510510445,
0.02367943897843361,
0.0022193724289536476,
0.08732300996780396,
0.09031340479850769,
-0.0401516817510128,
0.06282918155193329,
-0.019710354506969452,
0.06657671183347702,
-0.029300928115844727,
0.059844695031642914,
-0.038493383675813675,
0.05498648062348366,
-0.020381301641464233,
-0.042778074741363525,
0.08770143985748291,
-0.06070605292916298,
-0.06523255258798599,
-0.05371752381324768,
0.023755693808197975,
0.058907799422740936,
0.03314654529094696,
-0.06208014115691185,
0.05002591013908386,
-0.04593290016055107,
0.029682455584406853,
0.008711189031600952,
0.0024666166864335537,
0.03394458442926407,
0.0213500764220953,
-0.005846299231052399,
0.037528734654188156,
-0.05467210337519646,
-0.028040844947099686,
0.007835227064788342,
-0.010467460379004478,
0.07520575821399689,
-0.013787017203867435,
0.006020031403750181,
-0.0582427978515625,
-0.01765109971165657,
-0.02097274363040924,
0.05238481983542442,
-0.04613903537392616,
-0.02081458643078804,
0.014481496065855026,
-0.02338593266904354,
-0.04911366477608681,
1.2765938444907122e-33,
0.05388599634170532,
0.06421119719743729,
-0.008125494234263897,
0.06686636805534363,
-0.010024859569966793,
0.015545169822871685,
-0.027716606855392456,
-0.0035511895548552275,
-0.020799102261662483,
0.03448421508073807,
-0.041659899055957794,
-0.004596389830112457,
-0.030463287606835365,
0.058276351541280746,
-0.011654190719127655,
0.04839537665247917,
-0.03952718526124954,
-0.005410598590970039,
-0.027737418189644814,
0.03567524999380112,
0.04016967490315437,
0.019552743062376976,
-0.004832576494663954,
-0.07107783108949661,
-0.08973526209592819,
0.044933218508958817,
-0.042120080441236496,
-0.008895454928278923,
0.05874047055840492,
-0.0036665936931967735,
-0.020075524225831032,
-0.0005589534412138164,
-0.024069013074040413,
-0.075410857796669,
-0.015805279836058617,
-0.011988130398094654,
0.07728427648544312,
-0.09785984456539154,
-0.07867639511823654,
0.045636750757694244,
0.019804716110229492,
0.06918545812368393,
-0.019292205572128296,
0.013456327840685844,
-0.10675927996635437,
0.03252555802464485,
0.08678820729255676,
0.09539765119552612,
-0.03680689260363579,
0.05206966772675514,
0.04993880167603493,
0.02863987348973751,
-0.004359063226729631,
-0.041325099766254425,
-0.05632884055376053,
-0.017900677397847176,
0.013170595280826092,
-0.019960859790444374,
-0.02656267210841179,
0.045438576489686966,
0.01684654876589775,
0.0028603188693523407,
-0.1029137372970581,
-0.014537300914525986,
-0.08788595348596573,
-0.009371069259941578,
0.14175155758857727,
0.0742998793721199,
-0.02885093353688717,
0.08736960589885712,
-0.07157496362924576,
-0.007643916644155979,
0.009946162812411785,
-0.021114984527230263,
-0.031705740839242935,
-0.0030786674469709396,
0.026320841163396835,
-0.034616682678461075,
0.039840616285800934,
0.07535676658153534,
-0.05273834988474846,
0.11017252504825592,
0.035580530762672424,
-0.06270553916692734,
-0.07843383401632309,
-0.13401751220226288,
0.04298902302980423,
0.0037520381156355143,
0.013896367512643337,
-0.03190254420042038,
-0.07846267521381378,
-0.03700946643948555,
0.005514486692845821,
-0.03272589296102524,
-0.004416094161570072,
-1.534716975567932e-33,
0.022523969411849976,
-0.055799733847379684,
-0.08116710186004639,
0.06494388729333878,
0.04143336042761803,
-0.034877710044384,
0.06855268776416779,
-0.02392740361392498,
-0.003870112355798483,
-0.1274581402540207,
-0.006666745524853468,
0.005284403916448355,
-0.017743170261383057,
0.0428360216319561,
0.05025303363800049,
-0.05504676699638367,
-0.05689831078052521,
-0.037804052233695984,
-0.15434233844280243,
-0.09698329865932465,
0.03835960477590561,
0.04353233799338341,
0.07327151298522949,
-0.021340299397706985,
-0.0413961224257946,
0.12620045244693756,
-0.024637972936034203,
0.03409167751669884,
-0.03794737905263901,
0.05772533267736435,
-0.029198646545410156,
-0.012173762544989586,
0.009598778560757637,
-0.03263993561267853,
-0.04547499865293503,
0.039954427629709244,
-0.010823584161698818,
0.05979471653699875,
0.004465820733457804,
-0.013937129639089108,
-0.0365268811583519,
-0.01372590847313404,
0.03783006966114044,
0.015121995471417904,
0.017579900100827217,
0.03834345564246178,
-0.01758771389722824,
0.01183994859457016,
-0.013048898428678513,
0.05281984806060791,
-0.08436800539493561,
0.04710422828793526,
0.011088082566857338,
0.02091873064637184,
-0.017171917483210564,
-0.0673501268029213,
-0.14209359884262085,
-0.031619396060705185,
0.041744545102119446,
-0.009998061694204807,
-0.0016942290822044015,
-0.02974526397883892,
-0.11518886685371399,
0.038150690495967865,
-0.025920182466506958,
0.031102126464247704,
0.051349978893995285,
-0.06814204901456833,
-0.031149595975875854,
0.08799555897712708,
-0.014740879647433758,
-0.03931655362248421,
0.0015614511212334037,
0.0354556143283844,
0.03267866000533104,
-0.045007359236478806,
0.08233164995908737,
0.017811516299843788,
0.05959225073456764,
-0.016186801716685295,
0.06485184282064438,
0.026615673676133156,
0.0701465904712677,
0.06561360508203506,
0.1026655063033104,
0.08138012140989304,
-0.05906493216753006,
-0.016919612884521484,
0.03624643757939339,
0.05295804888010025,
-0.007425559684634209,
0.02510596439242363,
-0.05704478546977043,
0.07099974900484085,
-0.003274560207501054,
-3.6863735175529655e-8,
-0.014073433354496956,
0.024897094815969467,
0.030184345319867134,
-0.013520682230591774,
-0.05986647307872772,
0.0018254101742058992,
0.013532276265323162,
-0.024288179352879524,
-0.0042765079997479916,
-0.08376620709896088,
-0.02131970226764679,
-0.08159352838993073,
0.0035171369090676308,
-0.023055529221892357,
0.010541792958974838,
0.07836925983428955,
-0.02461015060544014,
0.04888107255101204,
-0.01936321146786213,
-0.02655981294810772,
-0.13803265988826752,
0.05349040403962135,
0.0055016339756548405,
0.023992720991373062,
-0.03590795397758484,
0.01837475597858429,
-0.035993605852127075,
0.019357986748218536,
0.12107899785041809,
0.020855160430073738,
-0.011287348344922066,
0.03554876521229744,
0.0259202029556036,
-0.05451647564768791,
-0.002112551359459758,
-0.007691757287830114,
-0.019665664061903954,
-0.012435641139745712,
-0.015594149008393288,
-0.0032017026096582413,
0.004841147921979427,
0.11321725696325302,
0.08532274514436722,
-0.004810291808098555,
0.06675945222377777,
0.09307249635457993,
0.11240789294242859,
0.01814253255724907,
-0.017454048618674278,
0.013766796328127384,
-0.010689799673855305,
0.009740984998643398,
0.0014575672103092074,
0.1193121075630188,
0.02538730576634407,
-0.006278147455304861,
-0.001250205677933991,
-0.01770402118563652,
0.05651437118649483,
0.0025262904819101095,
-0.023112857714295387,
-0.03117099590599537,
-0.001240034936927259,
0.005958708003163338
] |
BALROG: Benchmarking Agentic LLM and VLM Reasoning On Games | https://openreview.net/forum?id=fp6t3F669F | [
"Davide Paglieri",
"Bartłomiej Cupiał",
"Samuel Coward",
"Ulyana Piterbarg",
"Maciej Wolczyk",
"Akbir Khan",
"Eduardo Pignatelli",
"Łukasz Kuciński",
"Lerrel Pinto",
"Rob Fergus",
"Jakob Nicolaus Foerster",
"Jack Parker-Holder",
"Tim Rocktäschel"
] | Poster | Large Language Models (LLMs) and Vision Language Models (VLMs) possess extensive knowledge and exhibit promising reasoning abilities, however, they still struggle to perform well in complex, dynamic environments. Real-world tasks require handling intricate interactions, advanced spatial reasoning, long-term planning, and continuous exploration of new strategies—areas in which we lack effective methodologies for comprehensively evaluating these capabilities. To address this gap, we introduce BALROG, a novel benchmark designed to assess the agentic capabilities of LLMs and VLMs through a diverse set of challenging games. Our benchmark incorporates a range of existing reinforcement learning environments with varying levels of difficulty, including tasks that are solvable by non-expert humans in seconds to extremely challenging ones that may take years to master (e.g., the NetHack Learning Environment).
We devise fine-grained metrics to measure performance and conduct an extensive evaluation of several popular open-source and closed-source LLMs and VLMs. Our findings indicate that while current models achieve partial success in the easier games, they struggle significantly with more challenging tasks. Notably, we observe severe deficiencies in vision-based decision-making, as several models perform worse when visual representations of the environments are provided. We release BALROG as an open and user-friendly benchmark to facilitate future research and development in the agentic community. Code and Leaderboard at balrogai.com | LLM, VLM, Agents, Benchmark, RL, Reasoning, Games | null | 1,630 | 2411.13543 | [
0.03679225221276283,
-0.08291783183813095,
0.005547659937292337,
0.014860782772302628,
0.07992003113031387,
0.0052719758823513985,
-0.02571171522140503,
0.0003695542982313782,
-0.02385522797703743,
0.027579892426729202,
-0.07730895280838013,
-0.11894012242555618,
-0.0066972230561077595,
0.10078711062669754,
0.02422081120312214,
0.029232751578092575,
0.08064568042755127,
0.04243455454707146,
-0.053354356437921524,
-0.03612411767244339,
0.01768498308956623,
0.0011235761921852827,
0.0643477588891983,
-0.030981404706835747,
0.029315339401364326,
0.03121289610862732,
0.007941988296806812,
-0.033072639256715775,
-0.0016367853386327624,
-0.024681981652975082,
0.02891370840370655,
0.006971324328333139,
-0.03655394911766052,
0.005374656990170479,
-0.009894209913909435,
0.11294145882129669,
-0.060886722058057785,
0.0051477225497365,
-0.014008223079144955,
-0.086814284324646,
-0.07190341502428055,
-0.030277244746685028,
0.03409121558070183,
-0.019100770354270935,
0.07073459774255753,
0.050529658794403076,
0.005985328461974859,
-0.008215123787522316,
0.013292081654071808,
-0.028197601437568665,
-0.1009405255317688,
-0.06969773024320602,
0.03423253074288368,
-0.024590253829956055,
-0.051776885986328125,
0.046093061566352844,
0.03381728380918503,
-0.024337496608495712,
0.0428847074508667,
-0.08240973204374313,
-0.014554444700479507,
-0.1155630424618721,
-0.01306199748069048,
0.03983520716428757,
-0.08979354053735733,
-0.0035581830888986588,
0.029374534264206886,
0.025476302951574326,
-0.019358407706022263,
0.02088843286037445,
-0.033791836351156235,
0.024890461936593056,
-0.04264145717024803,
0.01275231596082449,
0.014602293260395527,
0.08416774868965149,
0.017994755879044533,
-0.08054099977016449,
0.06308721750974655,
-0.07195615768432617,
0.012167614884674549,
-0.04014650359749794,
-0.04301605373620987,
0.0708027109503746,
0.11935070902109146,
-0.005935838911682367,
-0.016046183183789253,
0.05013897269964218,
0.0586218386888504,
-0.022928666323423386,
0.006812414154410362,
-0.03603711351752281,
0.04285695403814316,
0.007397122215479612,
0.06403262913227081,
0.060144249349832535,
-0.017387498170137405,
-0.14215494692325592,
-0.08556258678436279,
0.10659728199243546,
-0.052957627922296524,
0.07370428740978241,
-0.007917812094092369,
-0.02333785407245159,
0.05314129963517189,
0.03585269674658775,
0.05623399466276169,
0.016609933227300644,
-0.0012440390419214964,
-0.0777464509010315,
0.03904041275382042,
-0.009679372422397137,
0.07730521261692047,
0.1001107320189476,
0.02675655484199524,
0.038459841161966324,
0.02214537002146244,
-0.018535099923610687,
-0.008601236157119274,
0.03352956473827362,
0.031250908970832825,
-0.014336238615214825,
0.03382839262485504,
-0.004700124729424715,
0.07552409172058105,
0.019451485946774483,
0.03785904496908188,
4.285242615403571e-34,
0.08150942623615265,
0.06039823219180107,
0.07499951124191284,
0.06117025017738342,
0.07630114257335663,
-0.045854486525058746,
0.04980839788913727,
0.004500228445976973,
-0.052605535835027695,
0.0708603635430336,
-0.05722054839134216,
-0.00008666042413096875,
-0.07357551157474518,
0.08382770419120789,
0.09155146777629852,
-0.011112122796475887,
-0.045109380036592484,
0.014249402098357677,
-0.07565026730298996,
-0.020011398941278458,
0.006089786533266306,
-0.05025382712483406,
0.019673319533467293,
-0.08527687937021255,
0.00485899718478322,
0.07849282771348953,
0.0521325021982193,
-0.060076724737882614,
-0.014418108388781548,
-0.030842721462249756,
-0.0762951448559761,
0.03504856303334236,
-0.025980401784181595,
0.1235220730304718,
0.02071245200932026,
0.0017073708586394787,
-0.06330692768096924,
-0.06359170377254486,
0.0187599565833807,
0.04133569821715355,
-0.005922601092606783,
0.04374431446194649,
-0.0006414674571715295,
-0.020796649158000946,
-0.02996697835624218,
-0.03957999870181084,
0.006754652597010136,
0.0043060798197984695,
-0.09041593968868256,
0.02201087772846222,
0.042533911764621735,
0.07549279928207397,
-0.05595779046416283,
-0.04952825978398323,
-0.005658676847815514,
0.023692848160862923,
0.0159694105386734,
0.032947149127721786,
0.0042549846693873405,
0.020508743822574615,
0.05350911617279053,
0.006934712640941143,
-0.027794014662504196,
0.09680770337581635,
0.0005543247098103166,
0.060217004269361496,
-0.011370264925062656,
0.10333878546953201,
0.04707494378089905,
-0.06593843549489975,
0.011980751529335976,
-0.0008317207684740424,
0.04849693551659584,
-0.07764570415019989,
0.08753858506679535,
-0.022766778245568275,
0.08078255504369736,
-0.09011910855770111,
0.05622313171625137,
-0.01858050376176834,
0.044168826192617416,
0.01975853554904461,
-0.058024417608976364,
-0.04778677225112915,
-0.020663421601057053,
-0.06243359297513962,
0.003820309415459633,
-0.11869070678949356,
-0.03570059314370155,
-0.02562435157597065,
-0.08177414536476135,
-0.04768052324652672,
-0.06612860411405563,
-0.013712872751057148,
0.011201038025319576,
-8.418248610471079e-34,
-0.03368442505598068,
-0.08951343595981598,
-0.024737151339650154,
0.07163269072771072,
-0.04744444042444229,
-0.07216420769691467,
-0.017184089869260788,
0.010280543938279152,
0.023545091971755028,
-0.09350533783435822,
-0.06356234848499298,
0.06263790279626846,
0.011356072500348091,
0.017144164070487022,
-0.0063425772823393345,
-0.10092137008905411,
0.0023784711956977844,
-0.03239288181066513,
0.021112406626343727,
0.026604270562529564,
0.03350190445780754,
0.11138711124658585,
-0.09296616166830063,
-0.02531050145626068,
-0.04788675904273987,
0.1043226569890976,
-0.01588570512831211,
-0.023717470467090607,
-0.023749375715851784,
0.04610229656100273,
0.042690251022577286,
-0.007982957176864147,
-0.024680081754922867,
-0.045806072652339935,
-0.024158157408237457,
0.07809147238731384,
0.06430031359195709,
0.003797944402322173,
-0.10900170356035233,
0.11373540759086609,
0.04136655107140541,
-0.08276963233947754,
-0.08150960505008698,
-0.059160564094781876,
-0.013447599485516548,
0.04445214197039604,
-0.03784129396080971,
-0.058679401874542236,
-0.03584380820393562,
0.014337941072881222,
-0.004053111653774977,
0.03336118534207344,
-0.11125390976667404,
-0.0055611636489629745,
-0.018576940521597862,
-0.046670641750097275,
-0.059704627841711044,
-0.015571595169603825,
0.0444912351667881,
-0.028512418270111084,
-0.06663402169942856,
0.017329584807157516,
-0.052604883909225464,
0.016132298856973648,
0.008521096780896187,
0.07831171154975891,
-0.0704028382897377,
-0.02277407795190811,
0.05131416395306587,
-0.011242141015827656,
0.014416199177503586,
-0.022928956896066666,
-0.03378421813249588,
0.00092536979354918,
0.0918726846575737,
0.04806658625602722,
-0.035315416753292084,
0.010929038748145103,
0.029515564441680908,
-0.05541139096021652,
0.03648511320352554,
-0.05123673379421234,
0.02596270479261875,
0.07424848526716232,
0.004288898315280676,
0.053111933171749115,
-0.025906892493367195,
0.04959111660718918,
0.014657439664006233,
-0.0074834576807916164,
-0.001200695987790823,
0.010351909324526787,
0.008461492136120796,
0.07644909620285034,
-0.06707324832677841,
-4.051305069197042e-8,
-0.06042852997779846,
0.05676780268549919,
0.007961622439324856,
-0.03841712698340416,
0.027686424553394318,
-0.028975365683436394,
-0.0679275244474411,
0.042490530759096146,
-0.033223602920770645,
0.029803326353430748,
0.0843236893415451,
0.02618112415075302,
0.022547326982021332,
-0.023574814200401306,
0.008143505081534386,
0.08234063535928726,
-0.005499028135091066,
-0.021841177716851234,
-0.03343390300869942,
0.005227907560765743,
0.024543067440390587,
0.030460184440016747,
-0.022354744374752045,
-0.06840559840202332,
-0.02989203855395317,
-0.0737060084939003,
-0.09344951808452606,
0.06385718286037445,
0.06998679786920547,
0.023645712062716484,
0.06216447055339813,
0.018881816416978836,
0.09410563111305237,
-0.005598770454525948,
0.04606340825557709,
0.03658387437462807,
-0.048805780708789825,
-0.06321119517087936,
0.028658801689743996,
-0.02035623788833618,
-0.004571069963276386,
0.08267591893672943,
-0.0031980411149561405,
0.02351280301809311,
-0.024878334254026413,
0.047551535069942474,
-0.05058389902114868,
-0.08582597225904465,
0.0021797448862344027,
-0.08702541142702103,
-0.05545216053724289,
0.01056636031717062,
-0.05111703649163246,
0.06623595952987671,
0.05783848464488983,
0.043902624398469925,
0.010592620819807053,
-0.05531344562768936,
0.04062190279364586,
0.06150127202272415,
-0.003729758784174919,
0.020219406113028526,
-0.07990303635597229,
0.033049412071704865
] |
Ensembles of Low-Rank Expert Adapters | https://openreview.net/forum?id=l0gZS0sAlf | [
"Yinghao Li",
"Vianne R. Gao",
"Chao Zhang",
"MohamadAli Torkamani"
] | Poster | The training and fine-tuning of large language models (LLMs) often involve diverse textual data from multiple sources, which poses challenges due to conflicting gradient directions, hindering optimization and specialization. These challenges can undermine model generalization across tasks, resulting in reduced downstream performance. Recent research suggests that fine-tuning LLMs on carefully selected, task-specific subsets of data can match or even surpass the performance of using the entire dataset. Building on these insights, we propose the Ensembles of Low-Rank Expert Adapters (ELREA) framework to improve the model's capability to handle diverse tasks. ELREA clusters the training instructions based on their gradient directions, representing different areas of expertise and thereby reducing conflicts during optimization. Expert adapters are then trained on these clusters, utilizing the low-rank adaptation (LoRA) technique to ensure training efficiency and model scalability. During inference, ELREA combines predictions from the most relevant expert adapters based on the input data's gradient similarity to the training clusters, ensuring optimal adapter selection for each task. Experiments show that our method outperforms baseline LoRA adapters trained on the full dataset and other ensemble approaches with similar training and inference complexity across a range of domain-specific tasks. | Language Model, LoRA, MoE, Ensembles | null | 1,623 | 2502.00089 | [
-0.013846250250935555,
-0.12275680899620056,
0.012207423336803913,
0.07669195532798767,
0.10919562727212906,
0.016709910705685616,
-0.03460538759827614,
0.005510367453098297,
-0.01906919851899147,
-0.07598894089460373,
-0.11066865921020508,
-0.10966464132070541,
0.07483150064945221,
0.0148172527551651,
-0.04636567085981369,
0.049687325954437256,
0.09546162933111191,
0.04410251975059509,
-0.08738037943840027,
-0.059668298810720444,
-0.00037357903784140944,
0.01753407157957554,
0.07983693480491638,
0.04555899277329445,
0.01646219566464424,
-0.007361382246017456,
-0.008023486472666264,
0.004691282752901316,
-0.06275651603937149,
-0.043553296476602554,
0.026597948744893074,
0.02661222778260708,
-0.014025777578353882,
0.0027817089576274157,
-0.061933211982250214,
0.06474669277667999,
-0.08618750423192978,
0.07810831815004349,
0.06710012257099152,
-0.016709033399820328,
0.0007675731903873384,
0.030871864408254623,
0.017203735187649727,
-0.021572085097432137,
0.06099918857216835,
0.018705788999795914,
-0.03127617388963699,
-0.04541504383087158,
-0.03750675544142723,
0.06771298497915268,
-0.06999021023511887,
-0.058965981006622314,
-0.02618410438299179,
0.032305292785167694,
-0.06632549315690994,
0.0295275766402483,
0.046705037355422974,
0.025402147322893143,
0.06402482092380524,
-0.040494490414857864,
-0.045049916952848434,
-0.13472211360931396,
0.017192604020237923,
-0.03817131370306015,
0.024967607110738754,
-0.07286205887794495,
0.049265794456005096,
0.05305427685379982,
-0.09073330461978912,
-0.031824372708797455,
-0.059977319091558456,
0.07702488452196121,
-0.02686082012951374,
0.07802069187164307,
0.06107962504029274,
0.007020885590463877,
0.06881017982959747,
0.010749706998467445,
0.05242379382252693,
-0.006079480517655611,
-0.024102240800857544,
0.07377207279205322,
0.00380677985958755,
-0.011829539202153683,
0.11288788914680481,
0.0016072953585535288,
-0.03161277249455452,
0.012047499418258667,
0.06040558964014053,
-0.02278093434870243,
-0.04821900650858879,
-0.06138217821717262,
0.07340405881404877,
0.005371025297790766,
0.001816643518395722,
0.08574595302343369,
0.010388804599642754,
-0.023074602708220482,
-0.022799775004386902,
0.06033067777752876,
-0.002397461561486125,
0.016040144488215446,
0.011385434307157993,
-0.034527506679296494,
-0.038656458258628845,
-0.050251640379428864,
0.022174635902047157,
0.03762291744351387,
-0.0021374898497015238,
-0.09763389825820923,
0.030860891565680504,
0.07683206349611282,
0.024676302447915077,
0.0560125932097435,
-0.008122575469315052,
0.08319760859012604,
0.027641745284199715,
-0.019894447177648544,
-0.030275734141469002,
0.12255541980266571,
-0.050369780510663986,
0.030747799202799797,
0.05923235043883324,
-0.01087161060422659,
0.047827981412410736,
0.10404990613460541,
-0.03391624987125397,
2.841031541680255e-33,
0.09338261932134628,
0.009791187010705471,
0.04550538212060928,
0.06005697324872017,
0.05186636373400688,
-0.060206297785043716,
0.04643654450774193,
0.057752687484025955,
-0.05838758870959282,
-0.057598359882831573,
-0.05386286973953247,
0.03984807804226875,
0.02146008051931858,
0.04057835787534714,
0.03655180707573891,
0.052434924989938736,
-0.0632861778140068,
0.046448323875665665,
-0.08703013509511948,
-0.04401553422212601,
0.0953509584069252,
-0.03275882452726364,
0.02534075267612934,
-0.0590522326529026,
-0.010529785417020321,
-0.0007469761767424643,
0.07943695783615112,
-0.016342854127287865,
-0.033505670726299286,
-0.011206853203475475,
-0.10370822250843048,
-0.0622905008494854,
0.051374562084674835,
0.027662189677357674,
0.07821812480688095,
-0.03768458217382431,
-0.008687099441885948,
-0.035553429275751114,
0.007653637323528528,
0.007876244373619556,
-0.03463565558195114,
0.03517867997288704,
0.02773648500442505,
-0.0011321317870169878,
-0.0434877909719944,
-0.07023590058088303,
-0.04220147058367729,
0.02439194731414318,
-0.026630334556102753,
-0.02178659290075302,
0.02497255429625511,
-0.0425780825316906,
0.021275032311677933,
0.007681792601943016,
0.004251750186085701,
0.027200087904930115,
0.017054682597517967,
0.008685484528541565,
0.045537132769823074,
-0.019345080479979515,
-0.013830401934683323,
0.049385081976652145,
-0.01668623834848404,
0.10056214034557343,
0.06396530568599701,
-0.0031835585832595825,
-0.020067330449819565,
0.10251829773187637,
0.057306453585624695,
-0.027220411226153374,
0.02198953367769718,
-0.006327740848064423,
0.01305823028087616,
-0.006587247829884291,
0.013342498801648617,
-0.013867279514670372,
0.012893607839941978,
-0.04837179556488991,
0.030647333711385727,
0.01806465908885002,
0.009103029035031796,
0.03013119287788868,
0.031226707622408867,
-0.10645240545272827,
-0.06938039511442184,
-0.09205754846334457,
0.044303666800260544,
-0.0335417203605175,
0.03503034636378288,
0.021090371534228325,
-0.09959625452756882,
-0.05025619640946388,
-0.027825355529785156,
-0.09089689701795578,
0.020306624472141266,
-1.2433493948815194e-33,
-0.042924489825963974,
-0.03926078975200653,
-0.008018993772566319,
0.0654098391532898,
0.03499628230929375,
-0.014797697775065899,
0.010698959231376648,
0.028121372684836388,
-0.07212799787521362,
-0.0654962882399559,
-0.04166288301348686,
-0.0399194061756134,
0.07482850551605225,
0.009992904029786587,
0.08284463733434677,
-0.008337324485182762,
-0.07796746492385864,
0.07951150089502335,
0.019436316564679146,
0.03958315774798393,
0.006140637211501598,
0.03823122754693031,
-0.022438114508986473,
0.016179082915186882,
-0.019031863659620285,
-0.01782887428998947,
-0.040275003761053085,
0.03531194478273392,
-0.05812502279877663,
-0.014222999103367329,
-0.05694236978888512,
0.010362247005105019,
-0.061559949070215225,
-0.019414497539401054,
-0.02412273921072483,
0.1000075489282608,
-0.00019544023962225765,
-0.05484607443213463,
-0.0175927747040987,
0.10636088997125626,
-0.01441661361604929,
-0.03931664302945137,
-0.060861971229314804,
-0.007235902827233076,
-0.04761992394924164,
-0.034440744668245316,
-0.10753443092107773,
-0.08598233014345169,
0.0279970895498991,
-0.02392462082207203,
-0.05878816172480583,
-0.0351305790245533,
-0.04361065849661827,
0.043155815452337265,
-0.04455601051449776,
-0.08687669783830643,
0.019997911527752876,
-0.02854747883975506,
0.03277307376265526,
-0.027746813371777534,
0.03470851853489876,
0.005884993821382523,
0.03921152278780937,
-0.03828056529164314,
0.05007283762097359,
-0.007706521078944206,
-0.011357567273080349,
-0.001982081215828657,
0.011786837130784988,
0.06051787734031677,
0.04073426499962807,
-0.08020102232694626,
0.04442600533366203,
-0.0033738117199391127,
-0.03519991412758827,
-0.0012658433988690376,
0.007218025624752045,
-0.016909338533878326,
-0.05049579590559006,
-0.053212057799100876,
0.0520235151052475,
-0.012875520624220371,
0.005582970101386309,
0.03582748770713806,
0.06877166777849197,
0.11928564310073853,
0.060305915772914886,
0.1335727721452713,
0.029035767540335655,
-0.014152709394693375,
0.022273864597082138,
-0.057661909610033035,
0.0159283634275198,
0.007165444549173117,
-0.06656327843666077,
-4.2679200618067625e-8,
-0.05903352424502373,
0.04827986657619476,
-0.09165927022695541,
0.0949094146490097,
-0.056916557252407074,
-0.03569459170103073,
-0.11731129884719849,
0.05137466639280319,
0.02266530692577362,
0.028829438611865044,
0.06372727453708649,
0.0015963861951604486,
0.01291386317461729,
-0.04370948672294617,
0.11128513514995575,
0.0298521276563406,
0.007654047105461359,
0.09824495762586594,
-0.03559359535574913,
0.003363455878570676,
0.07191374897956848,
0.053875938057899475,
0.059075213968753815,
0.012444190680980682,
0.10624279081821442,
-0.06922034174203873,
-0.04131481051445007,
0.0857616513967514,
0.0970030277967453,
-0.055589400231838226,
-0.048479411751031876,
-0.024463919922709465,
-0.021935738623142242,
-0.051744911819696426,
0.02915903553366661,
0.07865700870752335,
0.017832618206739426,
-0.015798553824424744,
-0.04706163331866264,
0.03787830471992493,
0.00883823074400425,
0.016660435125231743,
-0.06493944674730301,
0.025147495791316032,
0.02945253625512123,
-0.025811176747083664,
-0.05974666029214859,
-0.07077593356370926,
0.03981668874621391,
-0.01997937634587288,
0.00021370319882407784,
-0.06398936361074448,
0.007637423928827047,
0.08684944361448288,
0.054787904024124146,
0.004592223558574915,
-0.03718525916337967,
-0.04338300973176956,
0.037694819271564484,
-0.02255501225590706,
0.06046614795923233,
0.016162604093551636,
-0.04008049517869949,
0.026719698682427406
] |
Does Refusal Training in LLMs Generalize to the Past Tense? | https://openreview.net/forum?id=aJUuere4fM | [
"Maksym Andriushchenko",
"Nicolas Flammarion"
] | Poster | Refusal training is widely used to prevent LLMs from generating harmful, undesirable, or illegal outputs. We reveal a curious generalization gap in the current refusal training approaches: simply reformulating a harmful request in the past tense (e.g., *"How to make a Molotov cocktail?"* to *"How did people make a Molotov cocktail?"*) is often sufficient to jailbreak many state-of-the-art LLMs. We systematically evaluate this method on Llama-3 8B, Claude-3.5 Sonnet, GPT-3.5 Turbo, Gemma-2 9B, Phi-3-Mini, GPT-4o-mini, GPT-4o, o1-mini, o1-preview, and R2D2 models using GPT-3.5 Turbo as a reformulation model. For example, the success rate of this simple attack on GPT-4o increases from 1\% using direct requests to 88\% using 20 past-tense reformulation attempts on harmful requests from JailbreakBench with GPT-4 as a jailbreak judge. Interestingly, we also find that reformulations in the future tense are less effective, suggesting that refusal guardrails tend to consider past historical questions more benign than hypothetical future questions. Moreover, our experiments on fine-tuning GPT-3.5 Turbo show that defending against past reformulations is feasible when past tense examples are explicitly included in the fine-tuning data. Overall, our findings highlight that the widely used alignment techniques---such as SFT, RLHF, and adversarial training---employed to align the studied models can be brittle and do not always generalize as intended. We provide code and jailbreak artifacts at https://github.com/tml-epfl/llm-past-tense. | Jailbreaking, adversarial attacks, adversarial robustness, AI safety | Frontier LLMs refuse harmful prompts in the present tense, but not in the past tense | 1,622 | 2407.11969 | [
-0.09295683354139328,
0.03318518400192261,
0.025816848501563072,
-0.019647864624857903,
0.04669683426618576,
0.024709142744541168,
-0.01583411730825901,
-0.034908831119537354,
0.050412606447935104,
-0.043088626116514206,
-0.05042654648423195,
0.07474333792924881,
0.026509586721658707,
0.022668616846203804,
-0.00018559934687800705,
-0.028684237971901894,
0.10506094992160797,
0.06340017914772034,
-0.0089686568826437,
0.04278647527098656,
0.0651518926024437,
-0.020520927384495735,
0.07483845204114914,
-0.019148685038089752,
0.035652004182338715,
-0.028799111023545265,
-0.01450126338750124,
-0.0006064441986382008,
0.05724473297595978,
-0.041503552347421646,
0.006851518526673317,
0.06801092624664307,
-0.038551539182662964,
-0.0012398402905091643,
-0.0666995570063591,
-0.06781193614006042,
-0.0032056295312941074,
0.011131870560348034,
-0.0624031163752079,
-0.027604391798377037,
-0.02368162013590336,
0.049072135239839554,
0.010198383592069149,
0.016907140612602234,
0.013540114276111126,
-0.07669525593519211,
0.02072964981198311,
-0.055189430713653564,
-0.09895960986614227,
-0.04394370689988136,
-0.03544241935014725,
0.031891558319330215,
0.01932354085147381,
-0.01296311616897583,
0.021948983892798424,
-0.06284841150045395,
0.026736656203866005,
0.15247200429439545,
-0.008275379426777363,
0.031087679788470268,
-0.05642276629805565,
-0.04453623294830322,
-0.10514466464519501,
-0.027600755915045738,
0.027591103687882423,
0.07004231214523315,
-0.0018944480689242482,
0.05535304918885231,
0.06070956215262413,
0.05751020088791847,
-0.029173828661441803,
-0.036215171217918396,
-0.09282993525266647,
-0.023084886372089386,
-0.10180854797363281,
0.02504260651767254,
0.009215260855853558,
0.04765768349170685,
0.01292414404451847,
-0.07715938240289688,
0.056058626621961594,
-0.05611887201666832,
0.01951909437775612,
-0.00799937080591917,
-0.0005160720320418477,
-0.025257455185055733,
-0.04300970584154129,
0.018509306013584137,
0.12257549911737442,
0.021463172510266304,
0.005047392565757036,
-0.003960501868277788,
0.022485947236418724,
0.04544299468398094,
0.034289032220840454,
-0.01683971658349037,
0.01003021840006113,
0.009993946179747581,
-0.011313320137560368,
0.13232769072055817,
0.06825413554906845,
0.09034398198127747,
-0.059523943811655045,
-0.06460075080394745,
0.05393120273947716,
-0.040348540991544724,
-0.004152367357164621,
0.008908151648938656,
0.02521115355193615,
-0.0023898868821561337,
-0.04354622960090637,
0.08276647329330444,
0.03898830711841583,
0.012344066053628922,
-0.0185535978525877,
-0.030201604589819908,
-0.008271495811641216,
0.014682784676551819,
-0.03203156962990761,
0.08624596893787384,
-0.002388274995610118,
0.020456470549106598,
0.045419979840517044,
-0.014609423466026783,
0.020687518641352654,
-0.06735583394765854,
-0.11175830662250519,
3.035621958160415e-33,
0.07994075864553452,
-0.004915349185466766,
-0.04609328880906105,
0.013776223175227642,
0.02741747908294201,
0.07096424698829651,
0.00041271018562838435,
0.029736850410699844,
0.08222437649965286,
0.025637729093432426,
-0.012742165476083755,
-0.026439955458045006,
-0.00773982098326087,
0.006755440961569548,
0.10814757645130157,
0.028127508237957954,
-0.07130177319049835,
-0.012756261974573135,
-0.008610236458480358,
0.005781632382422686,
0.05351492390036583,
0.027484089136123657,
0.020366238430142403,
-0.05626027658581734,
-0.02569727972149849,
0.05202551558613777,
0.0009961259784176946,
0.008178925141692162,
0.06117771938443184,
0.04103906452655792,
-0.05466008558869362,
-0.0357804149389267,
-0.024221496656537056,
-0.006542737130075693,
0.07968698441982269,
0.041672829538583755,
-0.002173934830352664,
-0.05625288188457489,
0.019891923293471336,
-0.06601646542549133,
-0.07041116058826447,
-0.0358692891895771,
-0.04148244857788086,
-0.00989647675305605,
-0.037759825587272644,
-0.0887906402349472,
0.008689328096807003,
-0.00874515250325203,
-0.02559942938387394,
0.00022932638239581138,
-0.009207364171743393,
0.05717376992106438,
-0.028252271935343742,
-0.008313492871820927,
-0.11276166141033173,
0.052977703511714935,
-0.014485683292150497,
0.018978329375386238,
-0.023894280195236206,
0.03523628041148186,
0.07073861360549927,
0.010234108194708824,
-0.031086279079318047,
-0.0012587031815201044,
-0.07013637572526932,
0.02369525097310543,
-0.10573925822973251,
-0.00759916240349412,
-0.016509823501110077,
-0.014855346642434597,
-0.02840389311313629,
-0.012206889688968658,
-0.06198636814951897,
0.010901299305260181,
0.0809355080127716,
-0.13111664354801178,
0.1382819414138794,
0.033131226897239685,
0.06824745982885361,
-0.10940392315387726,
0.04200653359293938,
-0.054618582129478455,
-0.048996906727552414,
-0.004848691169172525,
-0.015710514038801193,
-0.051811814308166504,
0.04569455236196518,
-0.11413729935884476,
-0.010983653366565704,
0.04331325367093086,
-0.062255389988422394,
-0.09410993754863739,
0.011519900523126125,
0.05758769065141678,
-0.0004052532895002514,
-3.927432787930821e-33,
-0.05197538062930107,
-0.0027191932313144207,
-0.04634108021855354,
0.07619026303291321,
0.01566031202673912,
-0.07990695536136627,
0.008109962567687035,
0.01782919280230999,
0.016302935779094696,
-0.028451062738895416,
0.021717017516493797,
-0.032140735536813736,
0.07668085396289825,
-0.004599972162395716,
-0.018313588574528694,
-0.06954172998666763,
-0.02978876419365406,
-0.015311824157834053,
-0.006422749720513821,
-0.021258583292365074,
0.010016757994890213,
0.07607758045196533,
-0.07507726550102234,
0.05139746889472008,
-0.019041037186980247,
0.0060062101110816,
0.01420007273554802,
0.040217090398073196,
0.04188617691397667,
0.000919068232178688,
0.02501053363084793,
0.027576616033911705,
-0.0720592737197876,
0.06614875793457031,
-0.03444048762321472,
0.06013987958431244,
0.1334504634141922,
0.09135397523641586,
-0.06737657636404037,
0.028422122821211815,
0.061267461627721786,
-0.028665266931056976,
-0.0316154882311821,
0.0756012499332428,
0.032321035861968994,
0.06020841747522354,
-0.05155046656727791,
-0.006472376640886068,
0.08404886722564697,
-0.043108999729156494,
0.06049134209752083,
-0.03732331469655037,
0.01078940462321043,
0.07080886512994766,
-0.02101638726890087,
-0.09657357633113861,
0.007861941121518612,
-0.07531342655420303,
-0.059276118874549866,
0.06172596290707588,
-0.04666934162378311,
0.029056737199425697,
0.006453035864979029,
-0.06917472183704376,
0.02748396061360836,
-0.09267935901880264,
-0.015571723692119122,
0.07254950702190399,
0.03971012681722641,
-0.002992855617776513,
0.055088259279727936,
-0.026404138654470444,
-0.04332052543759346,
-0.01039181649684906,
0.05603490397334099,
-0.011632945388555527,
-0.10458534210920334,
-0.07593595236539841,
-0.03515532612800598,
-0.09275877475738525,
0.025620372965931892,
0.06660512834787369,
0.01915670558810234,
0.03404121473431587,
0.04360445588827133,
0.05713801831007004,
0.09051574021577835,
0.1059776321053505,
0.018954887986183167,
0.017186490818858147,
-0.03017433173954487,
0.030842524021863937,
0.026483846828341484,
0.0986630842089653,
-0.03886772692203522,
-5.8071449871022196e-8,
0.029299451038241386,
-0.039947107434272766,
-0.0031451282557100058,
0.06513132154941559,
0.05534135550260544,
0.013550074771046638,
-0.08824319392442703,
0.012261291034519672,
-0.043396130204200745,
-0.07282450795173645,
0.09719055145978928,
0.002891112584620714,
-0.0949842557311058,
-0.10199590772390366,
-0.009344539605081081,
0.08393241465091705,
0.009961794130504131,
-0.05613184720277786,
-0.06855803728103638,
-0.04061692953109741,
-0.029987214133143425,
-0.00426281476393342,
0.007185169495642185,
-0.007585506420582533,
-0.033989276736974716,
-0.03312056139111519,
-0.02380346693098545,
0.09867662936449051,
0.009136877954006195,
0.08542977273464203,
-0.021569447591900826,
0.017560752108693123,
-0.016096144914627075,
0.014188798144459724,
-0.10729902237653732,
0.02209833264350891,
-0.016218770295381546,
-0.03806304186582565,
0.07616914808750153,
-0.007666922174394131,
-0.028821058571338654,
-0.022581778466701508,
-0.004865079186856747,
0.05144849792122841,
-0.02820405177772045,
0.005819806829094887,
-0.08172589540481567,
-0.13480573892593384,
-0.0025355624966323376,
-0.04912913218140602,
0.05585944652557373,
-0.002563753165304661,
-0.047073207795619965,
0.05541900917887688,
0.059704460203647614,
0.03566763550043106,
-0.039700306951999664,
-0.05880679190158844,
-0.016204437240958214,
-0.01199757307767868,
0.04969382286071777,
0.012901699170470238,
-0.011359351687133312,
0.014333835802972317
] |
Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks | https://openreview.net/forum?id=hXA8wqRdyV | [
"Maksym Andriushchenko",
"Francesco Croce",
"Nicolas Flammarion"
] | Poster | We show that even the most recent safety-aligned LLMs are not robust to simple *adaptive* jailbreaking attacks. First, we demonstrate how to successfully leverage access to *logprobs* for jailbreaking: we initially design an adversarial prompt template (sometimes adapted to the target LLM), and then we apply random search on a suffix to maximize a target logprob (e.g., of the token *``Sure''*), potentially with multiple restarts. In this way, we achieve 100\% attack success rate---according to GPT-4 as a judge---on Vicuna-13B, Mistral-7B, Phi-3-Mini, Nemotron-4-340B, Llama-2-Chat-7B/13B/70B, Llama-3-Instruct-8B, Gemma-7B, GPT-3.5, GPT-4o, and R2D2 from HarmBench that was adversarially trained against the GCG attack. We also show how to jailbreak *all* Claude models---that do not expose logprobs---via either a transfer or prefilling attack with a *100\% success rate*. In addition, we show how to use random search on a restricted set of tokens for finding trojan strings in poisoned models---a task that shares many similarities with jailbreaking---which is the algorithm that brought us the *first place* in a recent trojan detection competition. The common theme behind these attacks is that *adaptivity* is crucial: different models are vulnerable to different prompting templates (e.g., R2D2 is very sensitive to in-context learning prompts), some models have unique vulnerabilities based on their APIs (e.g., prefilling for Claude), and in some settings, it is crucial to restrict the token search space based on prior knowledge (e.g., for trojan detection). For reproducibility purposes, we provide the code, logs, and jailbreak artifacts in the JailbreakBench format at https://github.com/tml-epfl/llm-adaptive-attacks. | Jailbreaking, adversarial attacks, adversarial robustness, AI safety | We show how to jailbreak basically all leading frontier LLMs with 100% success rate. | 1,621 | 2404.02151 | [
-0.07875142991542816,
-0.00822744332253933,
-0.018373824656009674,
-0.06615141779184341,
0.034221649169921875,
-0.0023001611698418856,
0.036423128098249435,
-0.012273615226149559,
-0.018937652930617332,
0.04255623742938042,
-0.005081479903310537,
-0.01824670098721981,
0.04994088038802147,
-0.04731369391083717,
0.026095259934663773,
0.025219697505235672,
0.08379997313022614,
0.07605845481157303,
-0.06617102026939392,
0.06603053957223892,
-0.0346376933157444,
0.0009393172804266214,
0.04383968189358711,
0.01006778422743082,
-0.04801901802420616,
-0.045315902680158615,
-0.07281729578971863,
-0.04111925885081291,
0.03420593589544296,
-0.07802107185125351,
0.04937388375401497,
-0.022778309881687164,
-0.012488596141338348,
0.02913651429116726,
-0.04219558462500572,
-0.020216291770339012,
-0.032585762441158295,
0.012386840768158436,
-0.012686442583799362,
-0.07035697996616364,
-0.024134324863553047,
0.07055884599685669,
-0.03834155574440956,
0.029350459575653076,
-0.01790180802345276,
-0.06390520185232162,
0.0029494522605091333,
0.010168456472456455,
-0.0546126514673233,
-0.05170392617583275,
-0.006114176940172911,
-0.00853126123547554,
0.031072942540049553,
0.055513035506010056,
-0.02645576372742653,
-0.019425826147198677,
0.005973123479634523,
0.04297367110848427,
0.07348518818616867,
0.045454297214746475,
0.0022908616811037064,
0.04082466661930084,
-0.04552597552537918,
-0.0005068392492830753,
0.014079970307648182,
-0.0020468414295464754,
0.05333876982331276,
0.09503036737442017,
0.04577326774597168,
0.08387628942728043,
0.034999921917915344,
-0.03623184934258461,
-0.05969639867544174,
-0.018697086721658707,
-0.030042530968785286,
0.09102346748113632,
-0.03402693197131157,
0.022149384021759033,
0.017994606867432594,
-0.064647376537323,
0.04440835490822792,
-0.05839730426669121,
0.06429959088563919,
0.022837674245238304,
0.04244126006960869,
-0.027950983494520187,
-0.062900111079216,
-0.0006342125125229359,
0.16223371028900146,
0.03057832643389702,
0.03120485506951809,
0.05644170194864273,
0.014320731163024902,
0.06941598653793335,
0.059407319873571396,
-0.040420155972242355,
0.019030248746275902,
-0.03145375847816467,
-0.039527587592601776,
0.11651960760354996,
0.04349572956562042,
0.040102891623973846,
-0.026519324630498886,
-0.04838024452328682,
0.07014455646276474,
0.020359158515930176,
0.01023096963763237,
-0.035721324384212494,
0.026431020349264145,
-0.01904613897204399,
0.03934277594089508,
0.055462777614593506,
0.05545611307024956,
0.03414063900709152,
-0.05019649118185043,
-0.025937309488654137,
-0.1020476371049881,
0.023625852540135384,
-0.007908791303634644,
0.048857104033231735,
0.03185508772730827,
0.022129429504275322,
0.026955261826515198,
-0.03607417643070221,
0.02578185312449932,
0.025317901745438576,
-0.09218603372573853,
1.8302718875371644e-33,
0.02702871710062027,
0.008010456338524818,
-0.07048212736845016,
0.009631329216063023,
0.06253346055746078,
0.026659782975912094,
-0.023127976804971695,
0.04474060609936714,
-0.059154342859983444,
0.0281244944781065,
-0.07551831007003784,
-0.04036721959710121,
-0.008124390617012978,
-0.012200660072267056,
0.12723073363304138,
0.02501085214316845,
-0.02974109910428524,
-0.03940078988671303,
-0.030948581174016,
-0.029425978660583496,
0.003939754329621792,
0.004848782904446125,
-0.009312872774899006,
-0.07942336797714233,
0.0430179201066494,
0.061059243977069855,
0.012721133418381214,
-0.0222008153796196,
0.009033042937517166,
0.0761055201292038,
-0.08509357273578644,
-0.03163367882370949,
-0.023139795288443565,
0.0725138932466507,
0.0354502908885479,
0.011656133458018303,
-0.05200863257050514,
-0.08993958681821823,
-0.03912079706788063,
-0.04018533602356911,
-0.07904580235481262,
-0.027427541092038155,
-0.017614087089896202,
-0.060436468571424484,
-0.016230283305048943,
-0.05935949832201004,
-0.09092304855585098,
0.027145713567733765,
-0.03625829145312309,
-0.02615244872868061,
-0.03411059081554413,
0.015108051709830761,
-0.09062036126852036,
-0.010467885062098503,
-0.10362650454044342,
0.001572187407873571,
-0.06343349814414978,
0.06824339181184769,
0.016132477670907974,
0.15341012179851532,
-0.029298193752765656,
0.014600886031985283,
-0.013708888553082943,
0.0053702169097959995,
-0.024975797161459923,
0.0038441424258053303,
-0.04329788684844971,
0.004766293801367283,
0.008589756675064564,
0.04656646400690079,
-0.058269649744033813,
-0.004184330347925425,
0.05889798700809479,
0.04841625690460205,
-0.03332299366593361,
-0.02836172841489315,
0.10399390012025833,
0.039446309208869934,
-0.03315448388457298,
-0.04692600294947624,
0.009850683622062206,
0.022456999868154526,
-0.010346158407628536,
0.05623863264918327,
-0.04695310816168785,
-0.09635589271783829,
-0.009404679760336876,
-0.12090612202882767,
-0.10265203565359116,
0.07855547219514847,
0.0012920608278363943,
-0.05194256454706192,
0.0369771346449852,
0.067714624106884,
-0.08137984573841095,
-1.7856699584471205e-33,
-0.07663989812135696,
-0.04569404944777489,
0.011788835749030113,
0.02327808551490307,
-0.03425320237874985,
-0.07953611016273499,
-0.034652527421712875,
-0.0024495585821568966,
-0.06748411059379578,
-0.007244722452014685,
-0.06622759252786636,
0.020230412483215332,
0.06432686746120453,
0.00017191721417475492,
0.03070235252380371,
-0.03400442376732826,
0.006355643272399902,
0.004584762267768383,
-0.021035784855484962,
0.017273658886551857,
0.08764731884002686,
0.09537802636623383,
-0.12629477679729462,
0.07832592725753784,
0.06119354069232941,
0.02765479125082493,
-0.02780812978744507,
0.08041627705097198,
0.06402454525232315,
0.04818333312869072,
0.018472332507371902,
-0.005480608437210321,
-0.04713869467377663,
0.00755871320143342,
-0.03540601208806038,
0.05933337286114693,
0.10355820506811142,
0.03177737072110176,
-0.008373587392270565,
0.030766522511839867,
0.08625983446836472,
0.011431066319346428,
-0.028803635388612747,
-0.009291065856814384,
0.025546163320541382,
0.053503721952438354,
-0.020023275166749954,
0.013252219185233116,
0.006824224255979061,
-0.06042087823152542,
0.05370897427201271,
-0.055142082273960114,
0.01914266124367714,
0.07819299399852753,
-0.047603789716959,
-0.068865567445755,
0.04888509213924408,
-0.03727946802973747,
-0.06627263873815536,
0.03304099664092064,
0.030113128945231438,
-0.022985413670539856,
0.046907875686883926,
0.005319611635059118,
0.02567564696073532,
-0.06829793751239777,
-0.058045804500579834,
0.05082405358552933,
0.005635133944451809,
-0.022484343498945236,
0.07322344928979874,
-0.004888369236141443,
-0.056130487471818924,
0.02336076833307743,
-0.005907610524445772,
0.04794719070196152,
-0.09524573385715485,
-0.14250999689102173,
-0.04877013340592384,
-0.041914135217666626,
-0.007208355236798525,
0.04352765530347824,
0.016434533521533012,
0.05796922743320465,
0.05593903735280037,
0.01651979610323906,
0.1160263940691948,
0.08740422129631042,
0.05898699164390564,
-0.00019557378254830837,
-0.050759121775627136,
0.024436818435788155,
0.017323847860097885,
0.14002355933189392,
-0.016830112785100937,
-5.263093072471747e-8,
0.017351621761918068,
-0.060375191271305084,
-0.022863928228616714,
0.054655522108078,
0.042556487023830414,
0.06372804939746857,
-0.08742963522672653,
-0.00011305369844194502,
-0.04405811429023743,
-0.052017148584127426,
0.10504675656557083,
-0.0058675846084952354,
-0.07929178327322006,
-0.05788211524486542,
-0.00020783988293260336,
0.0448833629488945,
-0.07348018884658813,
-0.022135909646749496,
-0.07628300040960312,
-0.027158407494425774,
-0.033282723277807236,
0.06646265834569931,
0.02491014078259468,
-0.015997283160686493,
-0.0067854938097298145,
-0.06849711388349533,
-0.021294042468070984,
0.04125412926077843,
0.0022662703413516283,
0.08346415311098099,
-0.06689301878213882,
-0.02938844822347164,
-0.02184254489839077,
0.020864132791757584,
-0.051118675619363785,
0.15166842937469482,
-0.03959732875227928,
-0.0639730915427208,
0.05355329066514969,
0.0008142957231029868,
-0.02383362129330635,
-0.03876421973109245,
0.01686941646039486,
0.03136935457587242,
-0.025231322273612022,
-0.06558318436145782,
-0.02900104410946369,
-0.0867641270160675,
0.06868339329957962,
-0.013526416383683681,
0.04936191439628601,
-0.03680676221847534,
-0.01995355449616909,
0.05892763286828995,
0.08883361518383026,
0.010595516301691532,
-0.06894563138484955,
-0.03963775560259819,
0.042112719267606735,
0.046360358595848083,
0.04667402431368828,
-0.04250335693359375,
0.022364147007465363,
-0.024592561647295952
] |
Laplace Sample Information: Data Informativeness Through a Bayesian Lens | https://openreview.net/forum?id=qO6dk9KfIp | [
"Johannes Kaiser",
"Kristian Schwethelm",
"Daniel Rueckert",
"Georgios Kaissis"
] | Poster | Accurately estimating the informativeness of individual samples in a dataset is an important objective in deep learning, as it can guide sample selection, which can improve model efficiency and accuracy by removing redundant or potentially harmful samples.
We propose $\text{\textit{Laplace Sample Information}}$ ($\mathsf{LSI}$) measure of sample informativeness grounded in information theory widely applicable across model architectures and learning settings.
$\mathsf{LSI}$ leverages a Bayesian approximation to the weight posterior and the KL divergence to measure the change in the parameter distribution induced by a sample of interest from the dataset.
We experimentally show that $\mathsf{LSI}$ is effective in ordering the data with respect to typicality, detecting mislabeled samples, measuring class-wise informativeness, and assessing dataset difficulty.
We demonstrate these capabilities of $\mathsf{LSI}$ on image and text data in supervised and unsupervised settings.
Moreover, we show that $\mathsf{LSI}$ can be computed efficiently through probes and transfers well to the training of large models. | Sample informativeness, Sample Information, Sample Difficulty, Long-tailed distribution, Leave-one-out retraining, KL Divergence | This paper establishes a novel measure of sample informativeness based on the KL divergence applied on a quasi Bayesian approximaiton of the model parameters. | 1,607 | null | [
-0.07107439637184143,
-0.05392645671963692,
0.0511796697974205,
0.029539544135332108,
0.1218520775437355,
-0.006164314690977335,
0.023205846548080444,
-0.005261648911982775,
0.034218087792396545,
-0.030091816559433937,
0.009644330479204655,
-0.057583920657634735,
0.024580899626016617,
-0.00003145251321257092,
-0.03528435900807381,
-0.03452266752719879,
0.10284823179244995,
0.09450925141572952,
-0.16429315507411957,
0.006209812592715025,
0.1057087630033493,
0.059655070304870605,
0.056180644780397415,
-0.035300374031066895,
0.06446023285388947,
0.013160227797925472,
0.07294256240129471,
-0.022700216621160507,
0.04348248615860939,
-0.06677332520484924,
0.03066839650273323,
0.010943264700472355,
-0.07532314211130142,
0.009622160345315933,
-0.018951542675495148,
-0.0002696310402825475,
-0.022616123780608177,
0.05063951388001442,
0.04529080539941788,
0.022989924997091293,
0.002245721872895956,
0.011293822899460793,
0.08709011971950531,
0.019236762076616287,
0.060022983700037,
-0.009775592014193535,
-0.0339338444173336,
-0.04108680039644241,
0.0381612665951252,
0.06472337990999222,
-0.00187750905752182,
-0.03895251080393791,
-0.03231528028845787,
0.04599916189908981,
-0.11412166804075241,
0.027458369731903076,
0.04426737502217293,
0.012574497610330582,
0.006168256979435682,
0.017775610089302063,
-0.04938594996929169,
-0.08646851778030396,
0.01523769274353981,
-0.01087427418678999,
-0.008156304247677326,
-0.02669854275882244,
0.024258382618427277,
0.01241917721927166,
-0.0043989187106490135,
0.025870803743600845,
-0.007021802477538586,
0.09452386200428009,
-0.04223208501935005,
0.07874834537506104,
0.06067695841193199,
0.005233061499893665,
0.01760040409862995,
0.008065837435424328,
0.043111030012369156,
-0.058966148644685745,
0.050688669085502625,
0.06590675562620163,
-0.05031048506498337,
0.020860793069005013,
0.14344671368598938,
-0.05115194991230965,
-0.046155184507369995,
0.04768797382712364,
0.0401502288877964,
0.0011982751311734319,
-0.05753565579652786,
-0.01064466591924429,
-0.0682099312543869,
-0.010190559551119804,
0.0263673085719347,
0.023986339569091797,
-0.029553093016147614,
-0.08524858206510544,
-0.017366476356983185,
0.07847969233989716,
-0.024524573236703873,
0.11847828328609467,
0.006723197642713785,
0.020909879356622696,
-0.022604789584875107,
-0.08769173175096512,
0.03215470165014267,
0.048616405576467514,
0.06473778188228607,
-0.12177476286888123,
0.02673770673573017,
0.038730449974536896,
0.004540174268186092,
-0.06280975788831711,
0.004366180393844843,
-0.03834444656968117,
0.043986424803733826,
0.004306360613554716,
-0.08753010630607605,
0.025150451809167862,
-0.06025872379541397,
0.007422679103910923,
-0.025253625586628914,
-0.027787785977125168,
0.02165294624865055,
0.011105604469776154,
-0.062207069247961044,
3.724064551056888e-33,
0.08910194039344788,
0.004435598384588957,
0.049361251294612885,
0.06886600703001022,
0.014612102881073952,
-0.005089289508759975,
-0.09660277515649796,
-0.029812900349497795,
-0.061970025300979614,
0.03665779531002045,
-0.05274653434753418,
0.01603374257683754,
-0.0907258689403534,
0.0013809348456561565,
0.08961091190576553,
0.041549116373062134,
-0.05548212304711342,
0.12463473528623581,
-0.024925652891397476,
-0.03193412348628044,
0.03046935237944126,
0.01938093639910221,
0.03549860417842865,
-0.11651501804590225,
-0.0005132297519594431,
0.052590262144804,
0.07025734335184097,
0.005777518730610609,
-0.024120334535837173,
0.013468269258737564,
-0.1013968288898468,
-0.014928205870091915,
0.07552830874919891,
-0.012398675084114075,
0.01650342158973217,
-0.010230118408799171,
-0.021909333765506744,
-0.05895578861236572,
0.08937308192253113,
-0.017500881105661392,
0.05222489312291145,
0.043591637164354324,
0.06265661120414734,
0.02015627734363079,
-0.10539425164461136,
-0.09265367686748505,
0.009959924034774303,
-0.062429118901491165,
0.064615398645401,
-0.07224564254283905,
0.0217985101044178,
-0.06774963438510895,
0.03050624392926693,
-0.00756035465747118,
0.009590535424649715,
0.03847905248403549,
-0.023004200309515,
0.010755940340459347,
0.011671331711113453,
0.000829417142085731,
0.0038149331230670214,
0.11261379718780518,
0.0028151297010481358,
0.12389423698186874,
0.016198011115193367,
0.017547471448779106,
-0.003356935689225793,
0.07127400487661362,
0.03088923916220665,
-0.03457498922944069,
-0.05288365110754967,
0.013592477887868881,
-0.0812988206744194,
-0.03849457949399948,
0.07916010916233063,
0.044251564890146255,
0.049775492399930954,
-0.05154610797762871,
0.005616139620542526,
0.063662089407444,
0.05003215745091438,
-0.06101643294095993,
0.04303060472011566,
0.011572015471756458,
-0.01916828751564026,
0.0474177710711956,
0.058117929846048355,
-0.07661762833595276,
0.013430599123239517,
-0.03699193522334099,
-0.026633093133568764,
-0.021826116368174553,
-0.047832150012254715,
-0.056292951107025146,
0.002923708874732256,
-2.266976671404283e-33,
-0.059906404465436935,
-0.009430401027202606,
0.02031700313091278,
0.0488717257976532,
0.005532324779778719,
-0.055752139538526535,
-0.00528904190286994,
-0.0062461425550282,
-0.027008797973394394,
-0.08060801029205322,
-0.055371347814798355,
-0.04545941203832626,
-0.05952548608183861,
-0.0853361189365387,
0.04164822772145271,
0.07237719744443893,
-0.02353062480688095,
-0.03423818200826645,
-0.04835697263479233,
-0.046134255826473236,
-0.003961183596402407,
0.07202152162790298,
-0.009971060790121555,
-0.005107814446091652,
-0.056178826838731766,
0.045507293194532394,
-0.03474101796746254,
0.10272256284952164,
-0.0013842041371390224,
-0.08998561650514603,
-0.03727688640356064,
-0.03129902482032776,
-0.0511687733232975,
-0.019859271124005318,
-0.07529293745756149,
0.05610570311546326,
0.0846351906657219,
0.012312660925090313,
-0.04108801484107971,
0.08268104493618011,
0.04812456667423248,
0.04109722003340721,
-0.09549732506275177,
-0.026746880263090134,
-0.02118845097720623,
-0.0802522599697113,
-0.018946807831525803,
-0.07089614868164062,
0.03901618346571922,
0.0019092891598120332,
-0.050013843923807144,
0.05047481134533882,
-0.05356011912226677,
0.06478238105773926,
-0.033479709178209305,
-0.07290598005056381,
-0.06376728415489197,
0.019516371190547943,
0.041159626096487045,
0.07093994319438934,
-0.10748907923698425,
0.02325710840523243,
-0.08456260710954666,
-0.022066639736294746,
-0.033235158771276474,
-0.03492408245801926,
-0.047546740621328354,
-0.02893916890025139,
0.02821170724928379,
-0.02058880776166916,
0.04316557198762894,
-0.08021102100610733,
0.04910004511475563,
-0.018372852355241776,
0.01795918680727482,
0.0005208244547247887,
0.05480450019240379,
0.055336542427539825,
0.022208569571375847,
-0.03749266639351845,
0.10253545641899109,
-0.02665567956864834,
-0.030619323253631592,
0.023782020434737206,
0.06925968080759048,
0.06580081582069397,
0.026910647749900818,
0.00405059615150094,
0.018665211275219917,
-0.0327792726457119,
-0.07350344955921173,
-0.014973215758800507,
0.008663593791425228,
0.03266378864645958,
-0.06581613421440125,
-4.889601612489969e-8,
-0.00393107021227479,
0.061688587069511414,
0.03596342355012894,
-0.023267701268196106,
0.03541228175163269,
-0.018536701798439026,
-0.06690678745508194,
0.1391855776309967,
-0.019713055342435837,
-0.0483388677239418,
0.07657279819250107,
-0.009112445637583733,
-0.008301914669573307,
-0.06912445276975632,
-0.005602928809821606,
0.11068758368492126,
0.031019551679491997,
0.031400009989738464,
0.006061900407075882,
-0.02063150890171528,
0.014335955493152142,
0.04105135798454285,
0.061511557549238205,
-0.03267764300107956,
0.031919535249471664,
-0.05471168830990791,
0.02941145747900009,
0.06683110445737839,
0.02534368820488453,
0.039568349719047546,
0.004426558967679739,
-0.018822865560650826,
-0.03900957852602005,
0.05485924705862999,
0.025294765830039978,
0.07828542590141296,
-0.004572708625346422,
-0.04163849726319313,
0.006681295111775398,
-0.05777117237448692,
-0.04553278535604477,
0.10536112636327744,
-0.061823997646570206,
0.06901096552610397,
0.03015950880944729,
0.007579382508993149,
0.03318013250827789,
0.01271718181669712,
-0.011695874854922295,
-0.02243613637983799,
-0.03932330384850502,
-0.04206530749797821,
-0.019556736573576927,
0.07671970874071121,
-0.015139716677367687,
-0.011753600090742111,
0.028200339525938034,
-0.03989844769239426,
-0.0011498898966237903,
0.042121317237615585,
0.03312801197171211,
0.03576141595840454,
-0.06027745455503464,
0.021570676937699318
] |
Structural-Entropy-Based Sample Selection for Efficient and Effective Learning | https://openreview.net/forum?id=xUMI52rrW7 | [
"Tianchi Xie",
"Jiangning Zhu",
"Guozu Ma",
"Minzhi Lin",
"Wei Chen",
"Weikai Yang",
"Shixia Liu"
] | Poster | Sample selection improves the efficiency and effectiveness of machine learning models by providing informative and representative samples. Typically, samples can be modeled as a sample graph, where nodes are samples and edges represent their similarities. Most existing methods are based on local information, such as the training difficulty of samples, thereby overlooking global information, such as connectivity patterns. This oversight can result in suboptimal selection because global information is crucial for ensuring that the selected samples well represent the structural properties of the graph. To address this issue, we employ structural entropy to quantify global information and losslessly decompose it from the whole graph to individual nodes using the Shapley value. Based on the decomposition, we present $\textbf{S}$tructural-$\textbf{E}$ntropy-based sample $\textbf{S}$election ($\textbf{SES}$), a method that integrates both global and local information to select informative and representative samples. SES begins by constructing a $k$NN-graph among samples based on their similarities. It then measures sample importance by combining structural entropy (global metric) with training difficulty (local metric). Finally, SES applies importance-biased blue noise sampling to select a set of diverse and representative samples. Comprehensive experiments on three learning scenarios --- supervised learning, active learning, and continual learning --- clearly demonstrate the effectiveness of our method. | Sample selection, graph, structural entropy, blue noise sampling | null | 1,606 | 2410.02268 | [
0.001474500517360866,
-0.03195694833993912,
0.06308232992887497,
-0.0078243063762784,
0.13406787812709808,
-0.006008600816130638,
0.020004261285066605,
-0.057173460721969604,
0.023030072450637817,
-0.03266887739300728,
-0.06802831590175629,
-0.03570655733346939,
0.029838550835847855,
0.04578576982021332,
-0.0010478346375748515,
-0.010938758961856365,
0.0751735046505928,
0.0827033594250679,
-0.04540964961051941,
-0.016123129054903984,
0.03460988402366638,
-0.0037819352000951767,
0.0207466222345829,
-0.004082054365426302,
0.03382906690239906,
-0.0772152990102768,
0.056772083044052124,
0.07921876013278961,
0.042135510593652725,
-0.03913644701242447,
0.01728433184325695,
-0.014584601856768131,
0.034903064370155334,
-0.046739451587200165,
-0.06873731315135956,
0.02297971583902836,
-0.08136610686779022,
0.06546531617641449,
-0.032091181725263596,
0.03675893321633339,
0.007206536829471588,
0.028283637017011642,
0.0009828031761571765,
0.0408715084195137,
0.03406721353530884,
0.060274623334407806,
-0.06328378617763519,
-0.024124151095747948,
0.0075169513002038,
0.11979444324970245,
-0.027361011132597923,
0.027197474613785744,
0.002252016682177782,
-0.02624453417956829,
0.008380419574677944,
-0.011878163553774357,
-0.021351661533117294,
-0.02754274383187294,
-0.08060791343450546,
-0.04746890813112259,
0.014842947013676167,
-0.10070319473743439,
-0.08115188032388687,
-0.02306385710835457,
0.09899240732192993,
0.03694123029708862,
0.05529862269759178,
0.030902646481990814,
-0.0423000305891037,
0.03924846276640892,
-0.007829615846276283,
-0.007101332303136587,
-0.10595517605543137,
0.12287925183773041,
0.021663321182131767,
0.09933203458786011,
0.03173139691352844,
0.025677338242530823,
0.009426268748939037,
-0.026216307654976845,
-0.0019059233600273728,
0.10402121394872665,
-0.012416132725775242,
0.04369691014289856,
0.10060764104127884,
-0.011596575379371643,
0.03380364924669266,
-0.018503036350011826,
0.031083326786756516,
0.04003966599702835,
-0.09066005051136017,
0.084687240421772,
0.02860904484987259,
-0.06209953874349594,
0.048381127417087555,
0.02736577019095421,
-0.0014055733336135745,
0.015319814905524254,
0.04349399358034134,
0.08271097391843796,
-0.03991754725575447,
0.040697429329156876,
0.01054338552057743,
-0.004378422629088163,
0.016518713906407356,
-0.029137790203094482,
-0.07816523313522339,
0.043716251850128174,
0.06837864965200424,
-0.0835300013422966,
-0.013009416870772839,
0.05222437158226967,
-0.05664254352450371,
0.016889875754714012,
0.05635324865579605,
0.01184720080345869,
0.053853839635849,
0.014487450942397118,
0.006056465208530426,
0.08616066724061966,
-0.0050962516106665134,
0.045170046389102936,
-0.013873385265469551,
-0.014197707176208496,
-0.029220500960946083,
-0.03237806260585785,
-0.039785608649253845,
3.765738397187431e-33,
0.08003721386194229,
-0.007196194026619196,
0.031500011682510376,
-0.03703102841973305,
0.02134430967271328,
0.023095373064279556,
-0.09061762690544128,
-0.009621602483093739,
-0.01890331506729126,
0.07500255852937698,
-0.11326023191213608,
0.04411110281944275,
0.07550835609436035,
0.02642754279077053,
0.05820254981517792,
0.06594876945018768,
-0.0459139347076416,
0.05604879930615425,
-0.02651127241551876,
-0.06426969915628433,
0.02958986908197403,
0.025410180911421776,
0.07287012040615082,
-0.045187316834926605,
-0.00333427800796926,
-0.02744121290743351,
-0.006656744051724672,
0.06109309196472168,
-0.043931957334280014,
-0.011511809192597866,
-0.023852132260799408,
-0.09282191842794418,
-0.003237001830711961,
-0.05229499191045761,
0.040112949907779694,
0.02521413192152977,
-0.019895577803254128,
-0.0027454430237412453,
-0.0007474790909327567,
0.006419496610760689,
0.03938959538936615,
0.018952442333102226,
0.07923460751771927,
-0.010067879222333431,
-0.09373560547828674,
-0.0013847275404259562,
-0.039856504648923874,
-0.07697848230600357,
0.02630833350121975,
-0.05720288306474686,
0.032909639179706573,
-0.09054720401763916,
0.03896849974989891,
-0.04181687906384468,
-0.012722976505756378,
0.07071252167224884,
0.0201105959713459,
0.06269454956054688,
0.030424322932958603,
0.01471435371786356,
-0.05892249941825867,
0.03187951073050499,
0.05410554260015488,
0.05686625465750694,
0.05696432292461395,
0.010613562539219856,
-0.032723210752010345,
-0.007908117026090622,
0.08330234885215759,
-0.0950034111738205,
-0.006871722638607025,
0.04386728256940842,
-0.039267901331186295,
-0.08716731518507004,
0.05350816622376442,
-0.01500934548676014,
0.06568170338869095,
-0.06462542712688446,
-0.00025800272123888135,
0.07452359795570374,
-0.10834204405546188,
-0.031425122171640396,
0.03423215448856354,
-0.1299384981393814,
-0.023422090336680412,
0.024738343432545662,
0.05578146502375603,
-0.07189257442951202,
-0.06330456584692001,
-0.04349835216999054,
-0.10043396800756454,
0.01566547527909279,
-0.047706250101327896,
-0.07230622321367264,
-0.009157855994999409,
-4.071085911782982e-33,
-0.05376888066530228,
-0.015977846458554268,
0.11456604301929474,
0.0413871668279171,
-0.0674603134393692,
-0.02066633850336075,
0.01670442894101143,
-0.04068945720791817,
-0.09513386338949203,
-0.04570815712213516,
-0.002186576835811138,
-0.045050639659166336,
0.14508424699306488,
-0.08208447694778442,
0.00998737383633852,
0.15886180102825165,
-0.11569125205278397,
0.041990552097558975,
-0.002834795508533716,
0.045368652790784836,
0.022816386073827744,
0.0528186559677124,
-0.08408301323652267,
0.023789918050169945,
0.02303461916744709,
0.005331346299499273,
-0.006525971461087465,
0.037316761910915375,
0.036922961473464966,
-0.06777340918779373,
-0.07497801631689072,
-0.015351061709225178,
-0.08165595680475235,
-0.03676183894276619,
-0.04990768060088158,
0.09329965710639954,
0.0006709484732709825,
-0.022753076627850533,
-0.05274166166782379,
0.11940525472164154,
-0.059176232665777206,
0.07491513341665268,
-0.129206582903862,
-0.026558134704828262,
-0.009699838235974312,
-0.03527916222810745,
-0.07576853781938553,
0.020541636273264885,
-0.021382465958595276,
0.028408968821167946,
0.019709821790456772,
0.033289145678281784,
-0.06487657129764557,
0.10408847033977509,
-0.04469708353281021,
-0.01905727945268154,
-0.03196495771408081,
0.06458456069231033,
-0.01027190312743187,
0.05409814044833183,
-0.024057181552052498,
-0.010098252445459366,
0.0000844114474602975,
0.013651259243488312,
0.020597901195287704,
-0.06932125985622406,
0.03502722829580307,
-0.010622896254062653,
-0.018457243219017982,
-0.0147738391533494,
-0.00880062859505415,
0.021476738154888153,
-0.013363855890929699,
-0.0424378402531147,
0.021109580993652344,
-0.013597597368061543,
0.07669412344694138,
0.09662096202373505,
-0.056606367230415344,
-0.02935524843633175,
0.04046214744448662,
0.04573354125022888,
-0.021107254549860954,
0.019707132130861282,
0.08381983637809753,
0.026045704260468483,
0.05393073707818985,
0.03424303978681564,
0.03522471338510513,
-0.014298005029559135,
-0.034936338663101196,
0.021514276042580605,
-0.0019195053027942777,
0.06084073707461357,
-0.05385889112949371,
-5.1745072227049604e-8,
-0.03175511956214905,
0.0283059049397707,
0.005938512738794088,
-0.010937884449958801,
0.05906042456626892,
0.02316945232450962,
-0.03071163222193718,
0.09652254730463028,
-0.06960231810808182,
-0.012304251082241535,
0.0833049863576889,
0.01368633285164833,
-0.07002775371074677,
-0.02757962793111801,
-0.013334139250218868,
0.010984855704009533,
0.0037899476010352373,
0.04855557903647423,
-0.011007039807736874,
0.0023095677606761456,
0.004004622343927622,
-0.01981477066874504,
-0.00812230072915554,
0.061818238347768784,
0.08775092661380768,
-0.07003634423017502,
-0.061658866703510284,
-0.019754020497202873,
0.027337756007909775,
0.04585232958197594,
-0.06974012404680252,
-0.005709158722311258,
0.07389407604932785,
-0.042844124138355255,
0.020389005541801453,
0.06737153232097626,
-0.0358293242752552,
-0.059473056346178055,
-0.036637816578149796,
-0.03223888948559761,
-0.016005614772439003,
0.02568839117884636,
-0.05090143531560898,
-0.006747453473508358,
0.012889457866549492,
-0.04359647259116173,
0.0046074544079601765,
0.024680152535438538,
0.008488797582685947,
-0.012736376374959946,
0.015340032987296581,
-0.0884641483426094,
-0.01818087324500084,
0.005237473174929619,
0.05790416896343231,
-0.034746237099170685,
-0.007567563559859991,
-0.020076192915439606,
0.08122577518224716,
-0.003338748821988702,
0.055214740335941315,
-0.020385773852467537,
-0.053481996059417725,
-0.07111500948667526
] |
Stable Segment Anything Model | https://openreview.net/forum?id=ooxj2Audlq | [
"Qi Fan",
"Xin Tao",
"Lei Ke",
"Mingqiao Ye",
"Di ZHANG",
"Pengfei Wan",
"Yu-Wing Tai",
"Chi-Keung Tang"
] | Poster | The Segment Anything Model (SAM) achieves remarkable promptable segmentation given high-quality prompts which, however, often require good skills to specify. To make SAM robust to casual prompts, this paper presents the first comprehensive analysis on SAM’s segmentation stability across a diverse spectrum of prompt qualities, notably imprecise bounding boxes and insufficient points. Our key finding reveals that given such low-quality prompts, SAM’s mask decoder tends to activate image features that are biased towards the background or confined to specific object parts. To mitigate this issue, our key idea consists of calibrating solely SAM’s mask attention by adjusting the sampling locations and amplitudes of image features, while the original SAM model architecture and weights remain unchanged. Consequently, our deformable sampling plugin (DSP) enables SAM to adaptively shift attention to the prompted target regions in a data-driven manner. During inference, dynamic routing plugin (DRP) is proposed that toggles SAM between the deformable and regular grid sampling modes, conditioned on the input prompt quality. Thus, our solution, termed Stable-SAM, offers several advantages: 1) improved SAM’s segmentation stability across a wide range of prompt qualities, while 2) retaining SAM’s powerful promptable segmentation efficiency and generality, with 3) minimal learnable parameters (0.08 M) and fast adaptation. Extensive experiments validate the effectiveness and advantages of our approach, underscoring Stable-SAM as a more robust solution for segmenting anything. Codes are at https://github.com/fanq15/Stable-SAM. | Segment Anything Model, Interactive Segmentation, Segmentation Stability, Deformable Feature Sampling | Our Stable-SAM analyzes and improves SAM’s segmentation stability across a wide range of prompt qualities. | 1,604 | 2311.15776 | [
0.02298477105796337,
-0.04371405392885208,
0.003312602872028947,
0.01670285314321518,
0.12427656352519989,
-0.04220161214470863,
0.026815246790647507,
-0.07005029171705246,
0.01517500914633274,
-0.03798646479845047,
-0.09870589524507523,
0.006311976350843906,
-0.022562021389603615,
0.01422838307917118,
0.03715173527598381,
-0.05364541709423065,
0.1024656817317009,
0.08479409664869308,
-0.07575765997171402,
-0.02099171094596386,
0.07302358746528625,
-0.04587444290518761,
0.03348139673471451,
-0.036199796944856644,
-0.02077634446322918,
0.02601100318133831,
-0.006178760901093483,
-0.08917629718780518,
0.061444442719221115,
-0.03294964134693146,
0.006944464985281229,
-0.0277929175645113,
0.05944596230983734,
0.04157230257987976,
-0.05429007112979889,
-0.05024764686822891,
-0.00009370849875267595,
0.03208179771900177,
-0.0645417645573616,
-0.03213205188512802,
-0.09193821251392365,
0.02263982966542244,
-0.04661751165986061,
-0.03704303130507469,
0.0202135369181633,
-0.017055673524737358,
0.053855881094932556,
0.006081472150981426,
-0.0014040933456271887,
0.012227687053382397,
0.00117008900269866,
-0.05136474594473839,
-0.020021073520183563,
0.0439712293446064,
-0.027450164780020714,
0.05781257152557373,
0.009911183267831802,
0.011929430067539215,
0.030534004792571068,
0.030875975266098976,
-0.05917014926671982,
-0.04932897165417671,
0.06721988320350647,
0.034052446484565735,
-0.03595859557390213,
-0.008379284292459488,
0.007022956386208534,
-0.048851870000362396,
-0.011669781990349293,
0.012723795138299465,
-0.07249732315540314,
0.024068884551525116,
-0.039575256407260895,
0.04572976380586624,
0.00804974790662527,
0.04280543327331543,
-0.01032273005694151,
-0.06837490200996399,
0.06228663772344589,
-0.0024358383379876614,
0.046204984188079834,
-0.033824075013399124,
-0.00879492424428463,
0.013073884882032871,
0.05858628824353218,
0.01572168804705143,
-0.03625579550862312,
0.0386740006506443,
-0.014887582510709763,
0.006530020851641893,
-0.10127002745866776,
-0.018958061933517456,
0.025201987475156784,
0.02804604172706604,
-0.04929342120885849,
-0.023771969601511955,
-0.057458288967609406,
-0.13476529717445374,
0.02489226683974266,
0.07443612813949585,
-0.03843379020690918,
0.020577717572450638,
0.06391929090023041,
-0.044664494693279266,
0.05981487035751343,
0.023458506911993027,
0.07127992808818817,
0.042314693331718445,
0.02555953525006771,
0.014693794772028923,
0.038533154875040054,
0.01721377484500408,
0.05024910345673561,
0.010587453842163086,
-0.014687209390103817,
-0.026282886043190956,
-0.0016130090225487947,
-0.05081026256084442,
-0.03150780498981476,
-0.005917090456932783,
-0.0027486493345350027,
-0.007512327283620834,
-0.08297097682952881,
0.06492545455694199,
0.11344248801469803,
-0.06924726814031601,
-0.04818115383386612,
4.0733979622342556e-33,
0.04036210849881172,
0.023390840739011765,
-0.01426693145185709,
0.0027115389239042997,
0.04659930616617203,
0.020135264843702316,
0.05850423499941826,
0.025491731241345406,
-0.007690489757806063,
0.00820873212069273,
0.007184057030826807,
-0.08203285932540894,
-0.025133566930890083,
0.07694101333618164,
0.09995893388986588,
-0.03148636594414711,
0.03841492906212807,
0.01699761301279068,
-0.05020871385931969,
-0.012169756926596165,
-0.01193806529045105,
-0.06678370386362076,
-0.04552697762846947,
-0.01617901213467121,
0.13753168284893036,
0.05682129040360451,
0.03459514304995537,
-0.002403863240033388,
0.002502662595361471,
0.01008018758147955,
-0.09740103781223297,
-0.020764440298080444,
0.04780377820134163,
0.00718603515997529,
-0.026408985257148743,
-0.03854546323418617,
-0.015084181912243366,
-0.031616486608982086,
-0.010513132438063622,
0.03554590046405792,
0.04837026074528694,
0.02184911258518696,
0.05234020948410034,
0.02996154874563217,
-0.0826847180724144,
-0.009840217418968678,
0.022747183218598366,
0.12907563149929047,
-0.14163246750831604,
0.05691651254892349,
0.04241309314966202,
0.005012720823287964,
-0.02239920385181904,
-0.06327827274799347,
-0.08436707407236099,
-0.08409496396780014,
0.02459004335105419,
0.04548315703868866,
0.01389295980334282,
0.07360099256038666,
-0.017898045480251312,
0.007269741967320442,
-0.05957493185997009,
0.020520061254501343,
-0.008657949976623058,
-0.04635308310389519,
-0.004353792406618595,
-0.008009656332433224,
-0.04490181431174278,
0.021703919395804405,
-0.08194202184677124,
0.004876743536442518,
-0.018671516329050064,
0.01682145707309246,
-0.0492248460650444,
-0.02549773082137108,
-0.054012127220630646,
0.040067464113235474,
0.016733728349208832,
0.059331201016902924,
-0.043609049171209335,
0.04580524191260338,
-0.04483906552195549,
-0.011280542239546776,
0.00021401082631200552,
-0.05568016692996025,
0.08145134896039963,
-0.10028857737779617,
-0.06935624778270721,
-0.04036504775285721,
-0.018216704949736595,
0.011083129793405533,
-0.009502321481704712,
0.04456038028001785,
0.03194664046168327,
-2.5496687364396426e-33,
0.055695097893476486,
0.05978090688586235,
-0.06233253329992294,
0.14705036580562592,
-0.07101216167211533,
-0.030944721773266792,
0.031220875680446625,
0.0316331721842289,
0.010504561476409435,
-0.13690270483493805,
-0.09613772481679916,
-0.017922960221767426,
0.030492058023810387,
-0.02475040778517723,
0.06058983877301216,
-0.0034526169765740633,
-0.07584843784570694,
0.00391077995300293,
-0.046776801347732544,
0.04361077770590782,
-0.014903003349900246,
0.06515220552682877,
-0.08644180744886398,
-0.04995957389473915,
-0.03625257685780525,
0.059013716876506805,
-0.06087731197476387,
0.07595814019441605,
0.025097405537962914,
0.01755618117749691,
-0.03589623048901558,
0.014115508645772934,
-0.028528952971100807,
-0.03949460759758949,
-0.0012412345968186855,
0.058476753532886505,
0.016220858320593834,
0.0414804182946682,
0.05155937373638153,
0.023222284391522408,
0.04992375150322914,
-0.044501032680273056,
-0.006166006438434124,
0.09472212940454483,
-0.006872647907584906,
-0.09306707978248596,
0.042645685374736786,
0.028699828311800957,
-0.06693227589130402,
0.023524636402726173,
-0.1287965029478073,
0.06895265728235245,
-0.06804048269987106,
0.04662461206316948,
-0.017337284982204437,
0.025195052847266197,
-0.16393691301345825,
0.015274700708687305,
0.09225098043680191,
-0.01696251891553402,
-0.014010845683515072,
-0.006815686356276274,
-0.01699376106262207,
0.0031178945209831,
0.07055342942476273,
-0.04550445079803467,
-0.006082364823669195,
-0.0664401650428772,
0.03823022171854973,
-0.01494961604475975,
0.0526210255920887,
-0.09370838850736618,
0.08533329516649246,
0.08428773283958435,
0.027848005294799805,
0.021635111421346664,
0.009585436433553696,
-0.03682470694184303,
0.03035692684352398,
0.029091181233525276,
-0.02420113794505596,
0.049742572009563446,
-0.05544165149331093,
0.09139375388622284,
0.08474437147378922,
0.11625569313764572,
-0.005358476657420397,
0.029920045286417007,
0.02300436422228813,
0.04590495303273201,
-0.023075632750988007,
0.035204995423555374,
0.032321613281965256,
0.0713125467300415,
-0.03726956248283386,
-5.147208526068425e-8,
-0.07948844134807587,
0.03497045114636421,
-0.035419806838035583,
0.02831306867301464,
0.031552478671073914,
-0.01995226740837097,
0.03061777912080288,
0.004166021477431059,
-0.005297488532960415,
-0.06639572978019714,
0.08378412574529648,
-0.016808077692985535,
-0.07892958074808121,
-0.04000642150640488,
0.0009647326660342515,
0.09007401764392853,
-0.03354145959019661,
0.03947172313928604,
-0.03742831200361252,
0.031144332140684128,
-0.0801902487874031,
-0.002941940212622285,
-0.0499994121491909,
-0.022909345105290413,
0.012234518304467201,
-0.017595496028661728,
-0.001422510715201497,
0.06500893086194992,
0.05413695052266121,
0.007379371207207441,
-0.003431518329307437,
0.008779194205999374,
0.058949828147888184,
0.07883773744106293,
0.0389467291533947,
0.05685802549123764,
-0.05196510627865791,
0.014512700028717518,
0.031177081167697906,
-0.05634826049208641,
0.08021923899650574,
-0.025367608293890953,
0.0004393641138449311,
-0.015600962564349174,
-0.018929852172732353,
0.07532243430614471,
0.08850998431444168,
-0.06980134546756744,
0.009784257970750332,
-0.033486101776361465,
0.05060051754117012,
-0.07475656270980835,
-0.032633669674396515,
0.0279306098818779,
0.04404214024543762,
-0.04566517099738121,
0.1025519073009491,
-0.058741360902786255,
0.0706636980175972,
0.09105651080608368,
0.0029604402370750904,
0.0747009739279747,
-0.11440721154212952,
-0.03358478099107742
] |
ProtoSnap: Prototype Alignment For Cuneiform Signs | https://openreview.net/forum?id=XHTirKsQV6 | [
"Rachel Mikulinsky",
"Morris Alper",
"Shai Gordin",
"Enrique Jiménez",
"Yoram Cohen",
"Hadar Averbuch-Elor"
] | Poster | The cuneiform writing system served as the medium for transmitting knowledge
in the ancient Near East for a period of over three thousand years. Cuneiform
signs have a complex internal structure which is the subject of expert paleographic
analysis, as variations in sign shapes bear witness to historical developments and
transmission of writing and culture over time. However, prior automated techniques
mostly treat sign types as categorical and do not explicitly model their highly varied
internal configurations. In this work, we present an unsupervised approach for
recovering the fine-grained internal configuration of cuneiform signs by leveraging
powerful generative models and the appearance and structure of prototype font
images as priors. Our approach, ProtoSnap, enforces structural consistency on
matches found with deep image features to estimate the diverse configurations
of cuneiform characters, snapping a skeleton-based template to photographed
cuneiform signs. We provide a new benchmark of expert annotations and evaluate
our method on this task. Our evaluation shows that our approach succeeds in
aligning prototype skeletons to a wide variety of cuneiform signs. Moreover, we
show that conditioning on structures produced by our method allows for generating
synthetic data with correct structural configurations, significantly boosting the
performance of cuneiform sign recognition beyond existing techniques, in particular
over rare signs. Our code, data, and trained models are available at the project page:
https://tau-vailab.github.io/ProtoSnap/ | Machine learning for social sciences, Ancient character recognition, generative models | An unsupervised approach for recovering the fine-grained internal configuration of cuneiform signs using diffusion-based generative models. | 1,598 | 2502.00129 | [
-0.09231667965650558,
0.04084624722599983,
0.12160616368055344,
-0.041211310774087906,
-0.010091105476021767,
-0.07228905707597733,
-0.11866088956594467,
-0.014452761970460415,
-0.021691037341952324,
-0.03774741664528847,
-0.044854581356048584,
-0.09802375733852386,
0.05725311115384102,
0.03998027741909027,
-0.0512494221329689,
-0.0472613088786602,
-0.036506738513708115,
0.05664413049817085,
0.008692778646945953,
0.0008391840965487063,
-0.06841202825307846,
0.02228078804910183,
0.04945893958210945,
-0.024805938825011253,
0.11271097511053085,
0.03948502615094185,
0.02660907246172428,
-0.021778961643576622,
0.01864672638475895,
-0.04210018366575241,
-0.053367480635643005,
0.02294398657977581,
0.009470486082136631,
0.047659777104854584,
0.0263537410646677,
0.12816132605075836,
-0.01141876820474863,
0.04010998085141182,
0.0013531933072954416,
-0.03500291332602501,
0.027410127222537994,
0.015465769916772842,
0.00046369165647774935,
0.08426189422607422,
0.07454650849103928,
0.10533100366592407,
0.020943639799952507,
0.07031738758087158,
-0.03216853365302086,
0.006793901789933443,
-0.017498740926384926,
-0.07080072164535522,
-0.024937661364674568,
-0.00090859952615574,
-0.01248127967119217,
0.010191041976213455,
0.06469171494245529,
-0.11906618624925613,
0.029776019975543022,
-0.026314986869692802,
0.03732552379369736,
-0.0022908253595232964,
-0.055393461138010025,
-0.032209474593400955,
-0.024984177201986313,
0.009002861566841602,
0.09685607254505157,
0.038167744874954224,
0.017210358753800392,
0.015196295455098152,
0.004343988839536905,
0.08051914721727371,
-0.016142522916197777,
0.04618556797504425,
-0.006434855051338673,
0.08023597300052643,
-0.02282131090760231,
-0.06751055270433426,
-0.03385029733181,
-0.11594186723232269,
-0.008386054076254368,
0.05716710910201073,
0.0438196137547493,
0.03223097324371338,
0.054783012717962265,
0.01545524038374424,
-0.04119408130645752,
0.023458564653992653,
0.03624885156750679,
-0.011643745936453342,
-0.05125952884554863,
-0.06905633956193924,
0.03215222805738449,
-0.04282323271036148,
-0.023158850148320198,
0.05452268198132515,
0.10568059235811234,
0.006526453420519829,
0.03141327574849129,
0.08126822113990784,
-0.005599142983555794,
0.002566152485087514,
0.11982952803373337,
-0.061991509050130844,
0.002380738267675042,
-0.057600364089012146,
-0.005095674656331539,
-0.07872773706912994,
0.022452060133218765,
-0.07748325169086456,
0.05868913605809212,
-0.05000893399119377,
-0.03592483699321747,
0.018692078068852425,
0.025173846632242203,
-0.04254981130361557,
-0.050215695053339005,
-0.03007803112268448,
0.02741255611181259,
-0.023082448169589043,
-0.038361866027116776,
0.0065887924283742905,
-0.015210826881229877,
0.05476711690425873,
-0.012245992198586464,
0.05986759811639786,
-0.038817793130874634,
1.4156736524697012e-33,
0.06680943816900253,
0.1280772089958191,
-0.03188425302505493,
0.07490263134241104,
0.09055250138044357,
-0.02118326723575592,
-0.12541469931602478,
-0.022970687597990036,
-0.03853496164083481,
-0.03142310678958893,
0.007296729367226362,
-0.009088465943932533,
-0.07063465565443039,
0.12493076175451279,
0.012418409809470177,
0.03285355865955353,
-0.007650131359696388,
-0.005081630777567625,
-0.07591375708580017,
-0.013244169764220715,
-0.08326105028390884,
0.01664656214416027,
-0.027293777093291283,
-0.06403354555368423,
-0.04508393630385399,
0.019945407286286354,
-0.00020270059758331627,
-0.04126105457544327,
-0.077896848320961,
-0.038735587149858475,
-0.015141415409743786,
-0.054800912737846375,
0.058589160442352295,
0.020630070939660072,
0.009886003099381924,
-0.08976253867149353,
0.07574036717414856,
-0.11728500574827194,
0.012300748378038406,
0.08537052571773529,
0.08292213082313538,
0.026555778458714485,
-0.03349422663450241,
0.01648138090968132,
-0.049653299152851105,
0.04766550660133362,
0.03538290038704872,
0.028451744467020035,
0.013884945772588253,
-0.06757145375013351,
0.0349336639046669,
0.07085376977920532,
-0.04371177405118942,
-0.04159106686711311,
-0.0861145630478859,
-0.019156647846102715,
-0.002481597475707531,
-0.09556152671575546,
-0.003294697031378746,
0.0071759093552827835,
0.058132149279117584,
0.07788364589214325,
0.055125754326581955,
0.05956263095140457,
-0.02884138748049736,
-0.0015436672838404775,
-0.04707365855574608,
0.06861545890569687,
0.047100260853767395,
0.03501246124505997,
-0.06216960772871971,
0.015467818826436996,
-0.009514755569398403,
-0.022055799141526222,
0.014141745865345001,
0.07922990620136261,
0.01927190274000168,
-0.023274190723896027,
-0.052402351051568985,
0.019420797005295753,
-0.06446655839681625,
0.025662332773208618,
-0.045532308518886566,
-0.08393857628107071,
-0.018218111246824265,
-0.07121911644935608,
0.09003542363643646,
-0.05989043787121773,
-0.024386417120695114,
-0.018074171617627144,
-0.0666266605257988,
0.0052559454925358295,
-0.01736067794263363,
-0.03961900994181633,
-0.00966912042349577,
-2.6612405990769477e-33,
-0.02511233650147915,
0.04524829238653183,
-0.055017974227666855,
-0.011768479831516743,
-0.11872544884681702,
-0.06094919145107269,
-0.06239521503448486,
0.04591253772377968,
-0.015390568412840366,
-0.07644104957580566,
0.024783533066511154,
0.02907746098935604,
0.012271175161004066,
-0.08073548227548599,
-0.028599616140127182,
-0.05097866430878639,
0.017303943634033203,
0.08171135187149048,
-0.008952643722295761,
-0.0003472824755590409,
-0.03066300041973591,
-0.011792061850428581,
-0.14512847363948822,
0.03697174787521362,
0.002463716547936201,
0.0714646726846695,
0.03229285404086113,
-0.0155984191223979,
-0.0771464854478836,
0.012234627269208431,
-0.04735705256462097,
-0.03628595545887947,
0.053598273545503616,
0.023800451308488846,
-0.05290978401899338,
0.005723940674215555,
0.0464940071105957,
-0.01550600491464138,
0.011963432654738426,
0.05132492259144783,
0.06327778846025467,
0.06817909330129623,
-0.040568120777606964,
0.027171041816473007,
-0.056728068739175797,
-0.03492499887943268,
-0.05872468650341034,
0.041215404868125916,
-0.0011445160489529371,
0.07781407237052917,
0.06756912171840668,
0.03910510241985321,
0.010629515163600445,
-0.11409331113100052,
-0.04164087027311325,
-0.01637721061706543,
-0.045338042080402374,
-0.049646902829408646,
0.03914165124297142,
0.03842751681804657,
-0.07879854738712311,
-0.06913790106773376,
-0.00285376631654799,
-0.09544656425714493,
0.03698773682117462,
-0.07699250429868698,
-0.06396484375,
0.018454784527420998,
-0.02527850866317749,
0.007367309648543596,
0.01487452257424593,
-0.05701833963394165,
-0.029490739107131958,
0.06951796263456345,
0.025671418756246567,
-0.020085584372282028,
0.021054143086075783,
0.01717495173215866,
0.023614907637238503,
-0.029739463701844215,
-0.026808498427271843,
-0.08445523679256439,
0.013636508025228977,
0.0860506221652031,
0.0704956129193306,
0.1343182772397995,
-0.07059264928102493,
0.06607240438461304,
0.07788955420255661,
0.03886716067790985,
-0.0098333228379488,
0.0684613585472107,
0.02024734392762184,
0.10311180353164673,
0.040235575288534164,
-3.962076178254392e-8,
-0.009313397109508514,
-0.006082446780055761,
-0.011090770363807678,
-0.029887648299336433,
-0.003277034731581807,
-0.03272182121872902,
0.019957121461629868,
-0.039419904351234436,
-0.04175971448421478,
-0.050233758985996246,
0.019489403814077377,
0.020322149619460106,
-0.07852604985237122,
-0.04797770082950592,
0.009468058124184608,
0.06308092176914215,
0.03706037998199463,
0.005912579596042633,
-0.09491587430238724,
0.027874192222952843,
0.01974935084581375,
0.08501362055540085,
-0.03268514201045036,
0.009242784231901169,
-0.04211156815290451,
-0.024904221296310425,
0.010410350747406483,
0.08836781233549118,
-0.014168541878461838,
0.020895924419164658,
-0.009045375511050224,
0.01531605888158083,
0.01083117350935936,
-0.039102815091609955,
0.04655079543590546,
0.041248954832553864,
0.013404353521764278,
-0.026911133900284767,
0.02288130484521389,
-0.003806719556450844,
0.08415936678647995,
0.013812843710184097,
-0.011483430862426758,
-0.001680568908341229,
0.038354191929101944,
-0.0017863494576886296,
0.106299988925457,
-0.017605362460017204,
0.024442991241812706,
0.00491075124591589,
-0.006420187186449766,
0.03204013779759407,
0.035909924656152725,
0.06877732276916504,
-0.019930828362703323,
-0.0313880555331707,
-0.03002934157848358,
0.08852510899305344,
0.01658863201737404,
0.026031535118818283,
0.06678985804319382,
-0.0013258338440209627,
-0.011334427632391453,
-0.005013157147914171
] |
Learning Fine-Grained Representations through Textual Token Disentanglement in Composed Video Retrieval | https://openreview.net/forum?id=wGa2plE8ka | [
"Yue WU",
"Zhaobo Qi",
"Yiling Wu",
"Junshu Sun",
"Yaowei Wang",
"Shuhui Wang"
] | Poster | With the explosive growth of video data, finding videos that meet detailed requirements in large datasets has become a challenge. To address this, the composed video retrieval task has been introduced, enabling users to retrieve videos using complex queries that involve both visual and textual information. However, the inherent heterogeneity between the modalities poses significant challenges. Textual data are highly abstract, while video content contains substantial redundancy. The modality gap in information representation makes existing methods struggle with the modality fusion and alignment required for fine-grained composed retrieval. To overcome these challenges, we first introduce FineCVR-1M, a fine-grained composed video retrieval dataset containing 1,010,071 video-text triplets with detailed textual descriptions. This dataset is constructed through an automated process that identifies key concept changes between video pairs to generate textual descriptions for both static and action concepts. For fine-grained retrieval methods, the key challenge lies in understanding the detailed requirements. Text description serves as clear expressions of intent, but it requires models to distinguish subtle differences in the description of video semantics. Therefore, we propose a textual Feature Disentanglement and Cross-modal Alignment framework (FDCA) that disentangles features at both the sentence and token levels. At the sequence level, we separate text features into retained and injected features. At the token level, an Auxiliary Token Disentangling mechanism is proposed to disentangle texts into retained, injected, and excluded tokens. The disentanglement at both levels extracts fine-grained features, which are aligned and fused with the reference video to extract global representations for video retrieval. Experiments on FineCVR-1M dataset demonstrate the superior performance of FDCA. Our code and dataset are available at: https://may2333.github.io/FineCVR/. | Composed Video Retrieval; Fine-grained Representation; Feature Disentanglement | null | 1,597 | null | [
0.011941136792302132,
-0.10136402398347855,
0.042098142206668854,
-0.026717865839600563,
0.08082891255617142,
0.021405281499028206,
-0.0024259849451482296,
0.015535864047706127,
0.04002625122666359,
-0.10715726763010025,
-0.0015033417148515582,
-0.05926726013422012,
-0.00669837323948741,
0.09489626437425613,
-0.0016863540513440967,
-0.019789675250649452,
0.1310049146413803,
0.07101227343082428,
-0.1258520632982254,
-0.006594799924641848,
0.09897623211145401,
-0.0017870094161480665,
0.07824436575174332,
0.006785419769585133,
0.021577168256044388,
0.024493630975484848,
0.010081936605274677,
0.03415074571967125,
0.0638241171836853,
-0.04496772214770317,
0.018740452826023102,
0.04426341876387596,
-0.04147328436374664,
0.0941784456372261,
-0.05340774729847908,
0.042518842965364456,
-0.08507293462753296,
0.0011380688520148396,
-0.05674869194626808,
-0.06806715577840805,
0.020579122006893158,
0.04829167202115059,
0.01610472798347473,
0.021750902757048607,
0.02770979516208172,
0.04944857954978943,
-0.03552770987153053,
-0.061343688517808914,
0.06169970706105232,
0.01604599691927433,
-0.10006088018417358,
0.059243980795145035,
-0.015544673427939415,
0.06808095425367355,
-0.01388438232243061,
0.000021801553884870373,
0.013944044709205627,
-0.007369876839220524,
0.02497735433280468,
-0.03965819254517555,
0.03050864487886429,
-0.10405611991882324,
0.03981870040297508,
-0.012891622260212898,
0.04565322399139404,
-0.07242311537265778,
0.0325351282954216,
0.012559415772557259,
0.01107404101639986,
-0.04477759078145027,
-0.07473466545343399,
0.04030647501349449,
-0.017183711752295494,
0.0714372768998146,
-0.05866517871618271,
0.06533550471067429,
-0.0009274680051021278,
-0.0068992022424936295,
-0.012949858792126179,
-0.033296745270490646,
0.041594237089157104,
-0.020411895588040352,
0.10834552347660065,
0.053391359746456146,
0.09958135336637497,
-0.01709403097629547,
-0.022354405373334885,
-0.023163696750998497,
-0.027401508763432503,
0.008997228927910328,
-0.1413845717906952,
0.023032234981656075,
0.08443636447191238,
-0.024237368255853653,
0.07785581797361374,
0.05610528588294983,
-0.0087787676602602,
-0.0624869242310524,
-0.003531214315444231,
0.05942630395293236,
0.05373368784785271,
0.04482810199260712,
0.00565901817753911,
-0.05982706695795059,
0.0390959195792675,
-0.00012321354006417096,
0.017250264063477516,
0.02039632759988308,
0.03274618461728096,
-0.07997909933328629,
0.03185100480914116,
0.029236920177936554,
-0.029672062024474144,
-0.04273219406604767,
0.022508591413497925,
0.022622402757406235,
-0.011617053300142288,
-0.01188051700592041,
-0.016788875684142113,
0.005414068698883057,
0.024020962417125702,
-0.027773860841989517,
-0.026534445583820343,
-0.06098141893744469,
-0.03306722268462181,
-0.04526906833052635,
0.028215033933520317,
2.07691833665952e-33,
0.04071659594774246,
0.007542535196989775,
0.02937508188188076,
0.10680001229047775,
0.04670426622033119,
-0.013887641951441765,
-0.029392842203378677,
0.0798279270529747,
-0.069658562541008,
0.006256067659705877,
-0.004342297557741404,
-0.04915447160601616,
-0.004469703417271376,
0.032818954437971115,
-0.05443401634693146,
-0.06350953131914139,
-0.11222941428422928,
-0.05588332191109657,
0.010112258605659008,
0.06413030624389648,
-0.003922720439732075,
0.030954474583268166,
0.02647446095943451,
0.029854342341423035,
0.010691128671169281,
-0.07551068812608719,
0.037726979702711105,
-0.11121813207864761,
-0.03366679698228836,
-0.02129831537604332,
-0.10381390154361725,
-0.04012518376111984,
0.044096291065216064,
-0.03506554663181305,
0.09761922806501389,
-0.015173734165728092,
-0.0942155197262764,
-0.043278954923152924,
0.007922657765448093,
-0.023475952446460724,
-0.012107765302062035,
-0.017392273992300034,
-0.04058993607759476,
-0.08417275547981262,
-0.04949420690536499,
-0.038604605942964554,
0.015860050916671753,
0.04106668755412102,
-0.07538679242134094,
0.040242064744234085,
0.08211185038089752,
0.05574825406074524,
0.031345319002866745,
-0.10949638485908508,
0.014943813905119896,
0.06145286187529564,
0.029101045802235603,
0.011961815878748894,
0.02898181602358818,
0.015836074948310852,
-0.008441741578280926,
0.035755787044763565,
-0.015140911564230919,
0.06070864573121071,
-0.011189237236976624,
0.02858065627515316,
0.054244980216026306,
0.11478185653686523,
0.06123681738972664,
0.05325911194086075,
-0.039452679455280304,
0.018000204116106033,
0.06404752284288406,
-0.07136436551809311,
0.03494169935584068,
0.026002870872616768,
0.007780096028000116,
-0.09668996185064316,
-0.012129501439630985,
-0.007928736507892609,
0.00895783957093954,
-0.06196165084838867,
0.07564306259155273,
-0.03063497319817543,
-0.11527470499277115,
-0.023202447220683098,
0.047576360404491425,
-0.07620943337678909,
0.037184689193964005,
-0.03923052176833153,
0.011711020953953266,
0.04744509235024452,
-0.05470774322748184,
-0.006501886993646622,
0.06509272009134293,
-3.367368412646046e-34,
0.04873603582382202,
0.032867856323719025,
-0.04756300151348114,
-0.03252676501870155,
0.03877229988574982,
-0.011676370166242123,
0.006715015042573214,
-0.018950263038277626,
0.05830001458525658,
-0.03903132304549217,
-0.015420799143612385,
-0.07079071551561356,
-0.08578883111476898,
-0.0741814374923706,
-0.04911959916353226,
0.04677823185920715,
-0.03926209732890129,
0.01999109797179699,
-0.030361050739884377,
0.0632554143667221,
0.0379582941532135,
0.02803088165819645,
-0.11014130711555481,
-0.008389418944716454,
-0.03398806229233742,
0.030069958418607712,
-0.003092873375862837,
0.0019894265569746494,
0.012656789273023605,
-0.014513122849166393,
0.03931798040866852,
-0.04696660861372948,
-0.031937118619680405,
-0.012240924872457981,
-0.10332103818655014,
-0.05227772146463394,
0.04705948382616043,
-0.049037106335163116,
-0.04725855588912964,
0.06985393911600113,
0.05611572787165642,
0.013393903151154518,
-0.07189933210611343,
0.05355735495686531,
-0.03261851146817207,
-0.04676163196563721,
-0.053097713738679886,
0.011056244373321533,
0.029779426753520966,
0.013933947309851646,
-0.015595757402479649,
-0.011295064352452755,
-0.06918879598379135,
0.05438985675573349,
0.04299263656139374,
-0.041104983538389206,
-0.054099973291158676,
0.04235810041427612,
0.09898801892995834,
-0.009990665130317211,
-0.025231851264834404,
-0.014295379631221294,
-0.04349832236766815,
-0.02083251252770424,
0.09122945368289948,
-0.032977230846881866,
-0.030808856710791588,
-0.049468085169792175,
-0.04425927624106407,
-0.012328209355473518,
0.03579915687441826,
-0.07354370504617691,
0.007390468847006559,
-0.036986254155635834,
0.03054404817521572,
-0.002625600667670369,
0.04586274176836014,
-0.017347224056720734,
-0.0191949475556612,
-0.026484480127692223,
-0.008891910314559937,
-0.057080045342445374,
0.021274052560329437,
0.10026384145021439,
0.03195654973387718,
0.11412422358989716,
0.009026320651173592,
0.10518241673707962,
-0.0033935762476176023,
0.02446426823735237,
-0.030155247077345848,
0.03856394812464714,
0.006540525238960981,
0.06204287335276604,
0.034038245677948,
-3.5607033765927554e-8,
-0.09992715716362,
0.04121091216802597,
-0.17671097815036774,
0.016566703096032143,
-0.031100912019610405,
0.017706865444779396,
0.01228280644863844,
0.09734994173049927,
-0.00006988807581365108,
-0.028059976175427437,
0.04694407060742378,
-0.009768085554242134,
-0.05443844199180603,
-0.046970583498477936,
-0.04576913267374039,
0.04538006708025932,
0.004436826799064875,
0.04199652001261711,
0.0054389177821576595,
0.09318161755800247,
0.034850675612688065,
0.04148266464471817,
0.059314604848623276,
0.028619272634387016,
-0.03895501792430878,
0.028205957263708115,
-0.028323141857981682,
0.09211844205856323,
0.09031336009502411,
0.031237266957759857,
-0.01409598346799612,
0.015524045564234257,
0.050294551998376846,
-0.13414610922336578,
-0.04635707288980484,
0.04386156052350998,
-0.01725182868540287,
-0.016042210161685944,
0.027896692976355553,
-0.031768906861543655,
0.012168943881988525,
0.06649564206600189,
0.0688430517911911,
0.03979521617293358,
0.0009120216709561646,
0.01706375926733017,
0.04932321980595589,
-0.05371994152665138,
0.07617589831352234,
0.04073823243379593,
-0.0428653359413147,
-0.016111232340335846,
-0.0122469337657094,
0.1366608887910843,
0.022168589755892754,
0.02817692793905735,
-0.0343407578766346,
0.00940013024955988,
0.08143037557601929,
-0.021904265508055687,
-0.0023206372279673815,
0.021804841235280037,
-0.019446691498160362,
-0.015505858696997166
] |
Salvage: Shapley-distribution Approximation Learning Via Attribution Guided Exploration for Explainable Image Classification | https://openreview.net/forum?id=WBUVagRgsd | [
"Mehdi Naouar",
"Hanne Raum",
"Jens Rahnfeld",
"Yannick Vogt",
"Joschka Boedecker",
"Gabriel Kalweit",
"Maria Kalweit"
] | Poster | The integration of deep learning into critical vision application areas has given rise to a necessity for techniques that can explain the rationale behind predictions. In this paper, we address this need by introducing Salvage, a novel removal-based explainability method for image classification. Our approach involves training an explainer model that learns the prediction distribution of the classifier on masked images. We first introduce the concept of Shapley-distributions, which offers a more accurate approximation of classification probability distributions than existing methods. Furthermore, we address the issue of unbalanced important and unimportant features. In such settings, naive uniform sampling of feature subsets often results in a highly unbalanced ratio of samples with high and low prediction likelihoods, which can hinder effective learning. To mitigate this, we propose an informed sampling strategy that leverages approximated feature importance scores, thereby reducing imbalance and facilitating the estimation of underrepresented features. After incorporating these two principles into our method, we conducted an extensive analysis on the ImageNette, MURA, WBC, and Pet datasets. The results show that Salvage outperforms various baseline explainability methods, including attention-, gradient-, and removal-based approaches, both qualitatively and quantitatively. Furthermore, we demonstrate that our explainer model can serve as a fully explainable classifier without a major decrease in classification performance, paving the way for fully explainable image classification. | Explainability, XAI, feature attribution | null | 1,596 | null | [
-0.07285112887620926,
-0.05447106435894966,
0.09475163370370865,
0.042539678514003754,
0.13246463239192963,
-0.012598459608852863,
-0.010403674095869064,
-0.059629280120134354,
-0.009098764508962631,
-0.052473194897174835,
-0.056574709713459015,
-0.06436536461114883,
-0.010519634932279587,
0.0412435457110405,
-0.06305363029241562,
-0.01725899428129196,
0.08333434909582138,
0.07869338989257812,
-0.13499368727207184,
0.021855460479855537,
0.04769578576087952,
0.011594352312386036,
0.10578111559152603,
0.03308979049324989,
0.04186425730586052,
-0.035806551575660706,
0.014947260729968548,
0.006291569210588932,
0.00039498068508692086,
-0.0135652432218194,
-0.0030839897226542234,
0.008353723213076591,
-0.027421897277235985,
-0.040622808039188385,
-0.036705177277326584,
0.03175397589802742,
-0.09269268065690994,
0.07903659343719482,
0.010071279481053352,
-0.026213038712739944,
0.0085730766877532,
0.041487812995910645,
-0.04286053031682968,
-0.0009138412424363196,
-0.003246667794883251,
-0.011678150855004787,
0.012093888595700264,
-0.09310837090015411,
0.06030946224927902,
-0.028545675799250603,
0.0027551092207431793,
0.00814325176179409,
-0.09217788279056549,
0.03587121516466141,
-0.030643552541732788,
-0.07367286086082458,
-0.02017761766910553,
0.01277479063719511,
0.015567760914564133,
-0.032317761331796646,
-0.016831079497933388,
-0.03399360179901123,
0.05865148827433586,
0.021427791565656662,
0.006541159935295582,
-0.014471599832177162,
0.031578611582517624,
0.03658987581729889,
0.0043558464385569096,
-0.008726853877305984,
-0.0009965719655156136,
0.06299649178981781,
-0.06682274490594864,
0.01922428235411644,
-0.002530446508899331,
0.11665021628141403,
0.02148817852139473,
0.05047588422894478,
0.016957834362983704,
-0.031092924997210503,
-0.01647544465959072,
-0.010165820829570293,
0.028068898245692253,
0.03033178113400936,
0.1065831333398819,
0.007546020671725273,
-0.027410581707954407,
0.04398539662361145,
-0.022328054532408714,
0.04491959512233734,
-0.07372862845659256,
-0.054085735231637955,
-0.03780461475253105,
-0.048906195908784866,
0.06834726780653,
0.016901690512895584,
-0.022410891950130463,
-0.11894559115171432,
-0.0703304260969162,
0.10011887550354004,
-0.09885747730731964,
0.0055563743226230145,
0.06154679134488106,
-0.1098499447107315,
0.08237055689096451,
-0.07149457931518555,
0.07008814066648483,
-0.0023476609494537115,
0.07625633478164673,
-0.05078773573040962,
0.056428536772727966,
0.009315351955592632,
-0.01939927041530609,
0.006603993475437164,
0.05682508274912834,
0.05401887372136116,
0.03426646068692207,
0.00856875628232956,
-0.03978772088885307,
0.0003272309259045869,
-0.010685006156563759,
0.01585237868130207,
-0.010436083190143108,
-0.01996597833931446,
0.0031130530405789614,
-0.008843881078064442,
-0.07851734757423401,
9.347769850028841e-34,
0.06059769168496132,
-0.06126905605196953,
0.03891168534755707,
0.03227119520306587,
0.09818825870752335,
-0.023218469694256783,
-0.01531265303492546,
0.05486788600683212,
-0.02155155874788761,
0.0912715271115303,
-0.02536352537572384,
0.03453197330236435,
-0.0022587215062230825,
0.08959544450044632,
-0.0109503623098135,
0.02743939496576786,
-0.000979033880867064,
0.05756274610757828,
-0.06686059385538101,
-0.07180069386959076,
-0.030537500977516174,
-0.0369892418384552,
-0.04732314869761467,
-0.08596403151750565,
-0.016753140836954117,
-0.04729854688048363,
-0.03128794953227043,
-0.029822086915373802,
-0.030727019533514977,
-0.015768034383654594,
-0.10830876231193542,
0.02055947668850422,
0.01923965848982334,
-0.02607184648513794,
-0.052415233105421066,
-0.046792201697826385,
-0.04299098625779152,
-0.04142417013645172,
0.0595666877925396,
0.02603609301149845,
0.058054983615875244,
-0.013173055835068226,
-0.04828542843461037,
-0.018398426473140717,
-0.017342668026685715,
-0.0164212454110384,
0.04036429896950722,
-0.06102001294493675,
-0.07481439411640167,
0.011564532294869423,
0.06817090511322021,
0.004913384560495615,
-0.023608604446053505,
-0.0672907680273056,
-0.09243971854448318,
-0.004206527955830097,
0.07956226170063019,
0.012759468518197536,
-0.04014845937490463,
0.03762708604335785,
0.0007601709803566337,
0.0650826096534729,
-0.017741698771715164,
0.03253650292754173,
-0.01683753915131092,
-0.006823310162872076,
-0.03013491816818714,
0.10023857653141022,
-0.01386087853461504,
0.0412009172141552,
-0.016996702179312706,
0.057222142815589905,
-0.055457498878240585,
-0.06480418145656586,
0.04692642763257027,
0.05621238052845001,
0.053954463452100754,
-0.038859620690345764,
0.022451894357800484,
0.07340037077665329,
0.02779712900519371,
0.004776597023010254,
0.01423528604209423,
-0.01244440022855997,
-0.08736690878868103,
0.03835485130548477,
0.11504700034856796,
-0.046602699905633926,
-0.04500184208154678,
-0.05488275736570358,
-0.09233145415782928,
-0.0017902690451592207,
-0.023045435547828674,
-0.04578559473156929,
-0.0056948293931782246,
-1.720489531377994e-33,
-0.05831676721572876,
0.06289007514715195,
-0.008563973009586334,
-0.02549097314476967,
-0.07515281438827515,
-0.012713647447526455,
-0.046945951879024506,
0.023354440927505493,
-0.001252911053597927,
-0.06460411101579666,
-0.022893328219652176,
-0.04562605172395706,
0.056134916841983795,
-0.06871236860752106,
-0.057154446840286255,
0.008626840077340603,
-0.06421027332544327,
0.019357269629836082,
-0.008070100098848343,
0.0654066652059555,
0.025773558765649796,
0.10835669934749603,
-0.061989445239305496,
0.007051934953778982,
-0.11068437993526459,
0.0639597550034523,
-0.0038960531819611788,
0.07149883359670639,
0.046184055507183075,
-0.045013103634119034,
-0.1021089181303978,
-0.012080381624400616,
-0.037387873977422714,
-0.0017541488632559776,
-0.056616976857185364,
0.0381317101418972,
0.03559951111674309,
0.012717137113213539,
-0.08824887126684189,
0.11135511100292206,
0.07327345758676529,
-0.028720872476696968,
-0.1118445172905922,
0.03938905894756317,
-0.052977304905653,
-0.06464816629886627,
0.09595376253128052,
-0.0077768475748598576,
0.055125005543231964,
0.07127701491117477,
-0.059674590826034546,
-0.008889585733413696,
-0.05931031331419945,
0.1515560895204544,
-0.04105406627058983,
-0.011687755584716797,
-0.03157772868871689,
0.039746373891830444,
0.050189290195703506,
0.044154781848192215,
-0.040456537157297134,
0.02179989404976368,
-0.05027800425887108,
0.008843408897519112,
-0.021159959957003593,
-0.0001794844720279798,
-0.0459638312458992,
-0.008446341380476952,
0.03583072870969772,
-0.007045229896903038,
0.015844015404582024,
0.003446246962994337,
0.009795029647648335,
-0.015966203063726425,
0.03486410900950432,
0.04881640896201134,
0.04942331090569496,
-0.017741609364748,
0.020962567999958992,
-0.022985907271504402,
0.08064848184585571,
-0.05341460555791855,
-0.019047163426876068,
0.04925250634551048,
0.09738073498010635,
0.054672129452228546,
-0.008766526356339455,
-0.014329150319099426,
0.039960525929927826,
-0.04259456694126129,
-0.0771610215306282,
0.004986172541975975,
-0.025681784376502037,
0.10715146362781525,
-0.02253880351781845,
-4.272186160392266e-8,
-0.026014914736151695,
0.08766385912895203,
-0.018737424165010452,
-0.0012834218796342611,
0.09277679026126862,
-0.03151721507310867,
-0.07183001935482025,
0.08263177424669266,
-0.06527794897556305,
-0.020701013505458832,
0.06613446027040482,
-0.0617460273206234,
-0.055866971611976624,
-0.04432028532028198,
0.00876846257597208,
0.10286060720682144,
0.03485853970050812,
0.07655061781406403,
0.018149787560105324,
0.03562978655099869,
-0.0542602501809597,
-0.03210042417049408,
0.08377678692340851,
-0.05726868659257889,
0.002569072414189577,
-0.06791958957910538,
0.008087004534900188,
0.06349779665470123,
0.017356276512145996,
0.0622008815407753,
-0.02880033291876316,
-0.010885958559811115,
0.06965948641300201,
-0.041105419397354126,
0.06788615882396698,
0.10387887060642242,
0.014849669300019741,
-0.043336980044841766,
-0.023524196818470955,
0.01454464253038168,
0.00933229923248291,
0.054730337113142014,
0.006533147767186165,
-0.0030490742065012455,
0.014900144189596176,
-0.011136972345411777,
0.06676226854324341,
-0.03556648641824722,
0.022790830582380295,
-0.015749827027320862,
-0.007089504972100258,
0.01716363988816738,
0.03132591024041176,
0.0919092446565628,
0.008703235536813736,
0.030983733013272285,
-0.02017926052212715,
-0.007175468374043703,
-0.024877235293388367,
0.1071314662694931,
0.06569579243659973,
0.06869438290596008,
-0.025116242468357086,
-0.017649341374635696
] |
Fine-tuning can Help Detect Pretraining Data from Large Language Models | https://openreview.net/forum?id=X8dzvdkQwO | [
"Hengxiang Zhang",
"Songxin Zhang",
"Bingyi Jing",
"Hongxin Wei"
] | Poster | In the era of large language models (LLMs), detecting pretraining data has been increasingly important due to concerns about fair evaluation and ethical risks. Current methods differentiate members and non-members by designing scoring functions, like Perplexity and Min-k%. However, the diversity and complexity of training data magnifies the difficulty of distinguishing, leading to suboptimal performance in detecting pretraining data. In this paper, we first explore the benefits of unseen data, which can be easily collected after the release of the LLM. We find that the perplexities of LLMs shift differently for members and non-members, after fine-tuning with a small amount of previously unseen data. In light of this, we introduce a novel and effective method termed Fine-tuned Score Deviation (FSD), which improves the performance of current scoring functions for pretraining data detection. In particular, we propose to measure the deviation distance of current scores after fine-tuning on a small amount of unseen data within the same domain. In effect, using a few unseen data can largely decrease the scores of all non-members, leading to a larger deviation distance than members. Extensive experiments demonstrate the effectiveness of our method, significantly improving the AUC score on common benchmark datasets across various models. | Large language models, Fine-tuning, Pretraining data detection | We propose Fine-tuned Score Deviation (FSD), a novel and effective approach that improves the detection capabilities of current methods for pretraining data detection. | 1,584 | 2410.10880 | [
-0.0043514566496014595,
-0.05617853254079819,
-0.015287407673895359,
0.057861827313899994,
0.08706200867891312,
0.07540260255336761,
0.0005839939112775028,
-0.03917920961976051,
0.075116366147995,
-0.05039486289024353,
0.005826544016599655,
-0.06456287205219269,
0.06400116533041,
-0.016962286084890366,
-0.04999622702598572,
-0.009584891609847546,
0.0696553885936737,
0.05388043075799942,
-0.12129524350166321,
-0.09237190335988998,
0.009006095118820667,
0.038060031831264496,
0.024437522515654564,
0.003505873726680875,
0.05319496616721153,
-0.11209583282470703,
-0.0035230389330536127,
0.004309691023081541,
0.01615288481116295,
-0.008612662553787231,
-0.029054952785372734,
0.03178573399782181,
0.09287258982658386,
0.059328239411115646,
-0.057828500866889954,
-0.012931241653859615,
-0.0016620884416624904,
0.020478595048189163,
0.004825315438210964,
-0.0017082395497709513,
-0.009906421415507793,
0.009783181361854076,
0.019465815275907516,
0.05836983770132065,
-0.009052108973264694,
0.023641489446163177,
-0.008862544782459736,
-0.042005036026239395,
-0.08484620600938797,
0.01181740127503872,
-0.07416626065969467,
-0.03003394603729248,
0.00394004862755537,
0.07195287942886353,
-0.03989715501666069,
0.0009652297012507915,
0.013400393538177013,
0.017551714554429054,
-0.09137635678052902,
-0.021975811570882797,
-0.09626125544309616,
-0.12631551921367645,
-0.031997933983802795,
-0.010367030277848244,
-0.05290872976183891,
0.011411079205572605,
-0.010810920037329197,
0.035367898643016815,
-0.0010257158428430557,
0.019292254000902176,
0.019258612766861916,
0.1154203861951828,
0.020930267870426178,
0.09691470116376877,
0.021423600614070892,
0.09321121126413345,
-0.03461683914065361,
0.0346987247467041,
0.0834738239645958,
-0.06242724508047104,
-0.030605502426624298,
-0.011705764569342136,
0.04414865002036095,
-0.007604255806654692,
0.13832826912403107,
0.04237673804163933,
0.10188469290733337,
-0.024468692019581795,
-0.006259835325181484,
0.05742117390036583,
-0.05537789314985275,
-0.007029096595942974,
0.037502471357584,
-0.008310345001518726,
-0.008875668980181217,
0.08477136492729187,
-0.028384478762745857,
-0.05545885115861893,
0.008102355524897575,
0.08442804217338562,
-0.029524102807044983,
0.14467652142047882,
-0.014813517220318317,
-0.031138962134718895,
0.005956417415291071,
0.003398084081709385,
0.055965665727853775,
-0.031392112374305725,
-0.0009788786992430687,
-0.05317138507962227,
0.04131617024540901,
0.09613245725631714,
-0.010842240415513515,
-0.0009011884103529155,
0.062359001487493515,
0.005841340869665146,
0.029039479792118073,
-0.023166455328464508,
0.0038879106286913157,
0.07091592997312546,
-0.020660247653722763,
0.023049907758831978,
0.023799145594239235,
0.02185417339205742,
-0.0028405427001416683,
0.02415776439011097,
-0.0700325146317482,
1.6171303334911057e-33,
0.07422375679016113,
0.025690842419862747,
-0.01345494482666254,
-0.01197685208171606,
-0.02773878537118435,
-0.06783240288496017,
-0.013356025330722332,
0.030496012419462204,
-0.04689766466617584,
0.05894618481397629,
-0.01826121099293232,
0.019094979390501976,
0.0280649121850729,
-0.011527666822075844,
0.04171598330140114,
0.032072871923446655,
-0.00897286832332611,
-0.00007931296568131074,
-0.03336234390735626,
-0.012576914392411709,
0.17284707725048065,
-0.05374554917216301,
0.04484657570719719,
-0.061625659465789795,
0.0047141979448497295,
0.0024002855643630028,
0.013446439057588577,
-0.10201263427734375,
-0.026440037414431572,
0.030357694253325462,
-0.09175970405340195,
-0.04159712418913841,
0.07830122858285904,
0.017659587785601616,
0.010108203627169132,
-0.03205689787864685,
-0.03775927051901817,
0.01731840893626213,
0.04327351227402687,
0.0010359352454543114,
0.012735839001834393,
0.008843193762004375,
0.0366269052028656,
-0.06780952960252762,
-0.06596199423074722,
-0.021961180493235588,
-0.04180801659822464,
-0.017734568566083908,
0.015053607523441315,
0.024635806679725647,
0.04866563901305199,
0.022228652611374855,
0.023265540599822998,
0.0026186814066022635,
-0.0686175674200058,
0.02117256633937359,
0.0359148271381855,
-0.021241741254925728,
-0.038499582558870316,
-0.007683995645493269,
-0.03905998542904854,
0.012912175618112087,
0.0012669276911765337,
0.07231390476226807,
0.006768558174371719,
-0.03534352406859398,
-0.05107580125331879,
0.004836434964090586,
0.02604854665696621,
-0.061135660856962204,
-0.008450616151094437,
-0.05020296946167946,
-0.05063886567950249,
-0.05617726966738701,
-0.03908175602555275,
0.004438767209649086,
0.07808767259120941,
-0.05020062252879143,
0.06748753786087036,
-0.0427289642393589,
0.07531126588582993,
-0.04752827808260918,
-0.01659899204969406,
-0.108978271484375,
-0.055637504905462265,
-0.06562477350234985,
0.031169764697551727,
-0.05140335112810135,
-0.0259313452988863,
-0.003634051885455847,
0.012465614825487137,
0.017220068722963333,
-0.05334828794002533,
-0.028608517721295357,
-0.015287802554666996,
-2.1030886979996628e-33,
-0.06095097213983536,
0.024593474343419075,
-0.012718341313302517,
0.039837028831243515,
-0.0024011798668652773,
-0.04953964427113533,
0.03807266429066658,
0.0972956046462059,
0.005490686744451523,
-0.01759675145149231,
-0.03610514476895332,
-0.1101345643401146,
0.043879762291908264,
-0.017333202064037323,
0.04108617082238197,
0.012658519670367241,
-0.0015562172047793865,
0.13351301848888397,
0.031020527705550194,
0.06973279267549515,
0.016579436138272285,
0.11972934752702713,
-0.11798233538866043,
0.03599586337804794,
-0.014988476410508156,
0.020832320675253868,
-0.029681958258152008,
0.05730345472693443,
-0.06227128580212593,
0.006538666784763336,
-0.05568849667906761,
0.010713660158216953,
-0.042461346834897995,
-0.03969244658946991,
-0.06779900938272476,
-0.04995192959904671,
0.03718261793255806,
-0.03760436177253723,
0.008614505641162395,
0.14344468712806702,
0.07688898593187332,
0.03501793369650841,
-0.15146900713443756,
-0.03617604449391365,
-0.03048660419881344,
-0.049560096114873886,
-0.09186394512653351,
-0.028131522238254547,
0.0514550618827343,
-0.011000261642038822,
-0.020286254584789276,
0.004559747874736786,
0.016962818801403046,
0.02019287645816803,
-0.03179195895791054,
-0.06693634390830994,
-0.04716034606099129,
-0.04670058563351631,
0.035822268575429916,
0.059302251785993576,
-0.0414639450609684,
0.01793425902724266,
-0.006629667244851589,
-0.031994275748729706,
0.028641076758503914,
-0.06641239672899246,
-0.031396377831697464,
0.030886197462677956,
-0.015247081406414509,
0.02927899919450283,
-0.0074983034282922745,
-0.011796129867434502,
-0.06903613358736038,
-0.04002860561013222,
-0.04186714068055153,
0.018277766183018684,
-0.00648794649168849,
-0.044271234422922134,
-0.05351538583636284,
-0.030338358134031296,
0.0469699501991272,
-0.038562435656785965,
-0.0020119454711675644,
0.0425054170191288,
0.06660380214452744,
0.10026626288890839,
0.0722501277923584,
0.03669526427984238,
0.025489402934908867,
0.029689989984035492,
-0.06476131081581116,
0.037477172911167145,
-0.010436448268592358,
0.013887127861380577,
-0.07352807372808456,
-4.810227594020944e-8,
-0.054642606526613235,
-0.013217669911682606,
-0.07801930606365204,
0.13400980830192566,
0.0005613702232949436,
0.006987022235989571,
-0.13442957401275635,
0.08986840397119522,
-0.04941285029053688,
-0.02094653993844986,
0.08133480697870255,
0.014518212527036667,
-0.08638250082731247,
-0.041319046169519424,
-0.0042371139861643314,
0.06600941717624664,
-0.06477289646863937,
0.09825628995895386,
-0.024380607530474663,
-0.02499575912952423,
0.014098330400884151,
0.09490571916103363,
0.05289885401725769,
-0.012476840056478977,
-0.001645980286411941,
-0.04238518327474594,
-0.011877068318426609,
0.053040578961372375,
0.062204811722040176,
-0.029403990134596825,
0.004159354604780674,
0.006012303754687309,
0.013567479327321053,
-0.007005949970334768,
0.017541131004691124,
0.1544409990310669,
-0.0559733584523201,
-0.005148634780198336,
-0.034707196056842804,
0.08227267116308212,
0.02448776736855507,
0.07688067108392715,
-0.04288288578391075,
0.03383863344788551,
0.00046204557293094695,
-0.07336894422769547,
-0.03669564053416252,
-0.05228529870510101,
0.036618564277887344,
-0.07450520992279053,
0.020274672657251358,
-0.026250971481204033,
-0.041426125913858414,
0.04537871107459068,
0.06693051010370255,
0.0629582554101944,
-0.018823275342583656,
0.044775497168302536,
0.014342116191983223,
-0.014311341568827629,
0.04966903105378151,
-0.03899103030562401,
-0.012425665743649006,
-0.027947448194026947
] |
Uncertainty and Influence aware Reward Model Refinement for Reinforcement Learning from Human Feedback | https://openreview.net/forum?id=iamWnRpMuQ | [
"Zexu Sun",
"Yiju Guo",
"Yankai Lin",
"Xu Chen",
"Qi Qi",
"Xing Tang",
"xiuqiang He",
"Ji-Rong Wen"
] | Poster | Reinforcement Learning from Human Feedback (RLHF) has emerged as a standard and effective approach for training large language models (LLMs) with human preferences. In this framework, a learned reward model approximates human preferences and guides policy optimization, making it crucial to develop an accurate reward model. However, without the ``true'' reward function, challenges arise when the reward model is an imperfect proxy for human preference. Since the policy optimization continuously shifts the human preference training dataset's distribution. The fixed reward model suffers from this problem of off-distribution, especially the on policy methods. While collecting new preference data can mitigate this issue, it is costly and challenging to optimize. Thus, reusing the policy interaction samples becomes a possible way to further refine the reward model. To tackle these challenges, we introduce a novel method \textbf{U}ncertainty-\textbf{G}radient based \textbf{D}ata \textbf{A}ugmentation (\textbf{UGDA} for short) to enhance reward modeling by leveraging policy samples to maintain on-distribution performance. Specifically, UGDA selects interaction samples based on the uncertainty of the reward ensembles and the gradient based influence of policy optimization. After the reward relabeling of selected samples, we use supervised learning to refine the reward ensembles, then get the retrained policy. Extensive experiments demonstrate that by leveraging UGDA to select a few samples without the costly human preference data collection, we can improve the ability of the policy and surpass the state-of-the-art methods. | Reward Modeling, Large Language Model, Data Augmentation | null | 1,579 | null | [
-0.08915666490793228,
-0.08081097900867462,
0.01065860316157341,
0.01913858950138092,
0.008220560848712921,
0.06750841438770294,
0.037318333983421326,
-0.0074610537849366665,
-0.036294788122177124,
0.04949026554822922,
-0.014572786167263985,
-0.05276330187916756,
0.06688243895769119,
0.03160715103149414,
0.028285548090934753,
-0.014183830469846725,
0.13048909604549408,
0.042798902839422226,
-0.06967958062887192,
-0.04515650123357773,
0.030890898779034615,
0.028757838532328606,
0.023868577554821968,
-0.020145876333117485,
0.009602670557796955,
-0.053205911070108414,
0.03753626346588135,
-0.03042013943195343,
0.04943644255399704,
-0.011065375991165638,
0.09617108851671219,
-0.03494878113269806,
0.03365914523601532,
-0.08741418272256851,
-0.0878390520811081,
0.04771732911467552,
-0.15174657106399536,
-0.05901525542140007,
-0.004196458030492067,
-0.0065864273346960545,
-0.020700624212622643,
0.01697128638625145,
0.02343066968023777,
0.07127345353364944,
0.034993916749954224,
0.012656702660024166,
0.009629381820559502,
-0.006022927816957235,
0.0417725034058094,
0.07100702822208405,
0.017073238268494606,
-0.0216146782040596,
0.0651584342122078,
-0.003512265160679817,
0.0015117423608899117,
0.015883686020970345,
0.020177653059363365,
0.04416435584425926,
-0.002310920739546418,
-0.08876834809780121,
-0.06777897477149963,
-0.17764922976493835,
-0.06037531793117523,
0.013588136993348598,
-0.06133939325809479,
-0.07615363597869873,
-0.015046900138258934,
0.06717066466808319,
-0.0013679071562364697,
-0.000033400396205252036,
-0.008040741086006165,
0.07786010205745697,
-0.02072068303823471,
-0.017030352726578712,
0.032061617821455,
0.03407464176416397,
0.003227740293368697,
-0.036582931876182556,
0.07996075600385666,
-0.03296313062310219,
-0.046453800052404404,
0.005740968976169825,
0.06436984241008759,
-0.03441120684146881,
0.10077526420354843,
-0.04746011644601822,
0.031153222545981407,
0.04210975393652916,
0.07834655046463013,
0.01594880037009716,
-0.010326961055397987,
0.03943422809243202,
0.018915414810180664,
-0.0051833330653607845,
-0.09848708659410477,
0.08737712353467941,
-0.02370811626315117,
-0.06216924637556076,
-0.08712045848369598,
0.10108175128698349,
0.05852899327874184,
0.08075246959924698,
-0.08623351156711578,
0.007928446866571903,
0.00024829612812027335,
-0.04931033402681351,
-0.05694897472858429,
0.08892416954040527,
0.03662538900971413,
-0.07419215142726898,
-0.013497734442353249,
0.027996281161904335,
0.00043818241101689637,
0.031515467911958694,
0.015139862895011902,
-0.019607553258538246,
0.0286723505705595,
-0.08548680692911148,
-0.03841989487409592,
0.022404378280043602,
-0.023705249652266502,
-0.03086923249065876,
0.05499687418341637,
0.026494892314076424,
-0.006379758473485708,
0.019533399492502213,
-0.021851252764463425,
-7.345460782664485e-34,
0.04164857044816017,
0.03276176378130913,
0.00287231826223433,
-0.011110840365290642,
0.031375590711832047,
0.026688171550631523,
0.04745907709002495,
0.05224975198507309,
-0.06323117017745972,
-0.057536326348781586,
0.023519637063145638,
0.038853127509355545,
-0.0551765114068985,
0.12271524220705032,
0.03684267774224281,
-0.004178917035460472,
-0.03176291659474373,
0.03163245692849159,
0.014354821294546127,
-0.03421662002801895,
0.048764217644929886,
0.004056095145642757,
0.07794591784477234,
-0.07919164001941681,
0.05029832571744919,
-0.00022047903621569276,
0.07177063077688217,
-0.0023630622308701277,
-0.046662408858537674,
0.01025441475212574,
-0.015911266207695007,
-0.04561930149793625,
0.06442713737487793,
-0.013364814221858978,
-0.03516102582216263,
0.0071336738765239716,
-0.027730057016015053,
-0.029850022867321968,
-0.012225952930748463,
-0.056909363716840744,
0.04236257076263428,
0.01951461099088192,
0.042315524071455,
0.00883528497070074,
-0.06850763410329819,
-0.056763846427202225,
0.0696331188082695,
0.004534706473350525,
-0.027510439977049828,
-0.02957071363925934,
0.06579059362411499,
0.05355053022503853,
0.044566892087459564,
-0.05812515690922737,
-0.05467042326927185,
-0.012647872790694237,
-0.028457781299948692,
0.045045942068099976,
-0.0022848935332149267,
-0.02389690652489662,
0.006079426500946283,
0.014063247479498386,
0.03625551983714104,
0.10170701891183853,
0.03025699220597744,
0.060435693711042404,
-0.06432951241731644,
0.011420190334320068,
0.023594507947564125,
-0.013639970682561398,
-0.011517097242176533,
0.0024746109265834093,
0.00038337442674674094,
-0.04995433986186981,
0.03227302432060242,
-0.02387419529259205,
0.04744929447770119,
-0.035449136048555374,
0.05161070451140404,
-0.04959092289209366,
-0.03725080192089081,
-0.03889758884906769,
-0.04102378338575363,
-0.07380840182304382,
-0.007792164571583271,
-0.055325429886579514,
0.04687203839421272,
-0.10814617574214935,
0.01815950497984886,
-0.01977686583995819,
-0.0338476300239563,
-0.02617637999355793,
-0.014591162092983723,
0.06925899535417557,
0.06978651881217957,
-1.9904269316939547e-34,
-0.03402251377701759,
-0.00231828261166811,
0.011130335740745068,
0.029716068878769875,
-0.012327293865382671,
-0.013951314613223076,
-0.012225301936268806,
0.07388582080602646,
-0.027857748791575432,
-0.09418400377035141,
-0.05131097882986069,
-0.005287246312946081,
0.022396229207515717,
0.0003201176878064871,
0.031387750059366226,
-0.030146969482302666,
-0.006543335970491171,
0.03605636954307556,
0.0214591883122921,
0.020902464166283607,
0.022750409319996834,
0.07375995814800262,
-0.10949292033910751,
0.08357718586921692,
-0.013502380810678005,
0.0031391552183777094,
0.011906829662621021,
0.03641168773174286,
-0.00823307503014803,
0.021394697949290276,
-0.06316933035850525,
-0.03155435994267464,
-0.10324148088693619,
-0.023555733263492584,
-0.06446671485900879,
0.03070301003754139,
0.03933871537446976,
0.002889093244448304,
-0.0542578361928463,
0.17188894748687744,
0.033603083342313766,
-0.06204519048333168,
-0.0718158483505249,
-0.01688127964735031,
-0.009220851585268974,
-0.007577352225780487,
-0.11495018750429153,
-0.04298661649227142,
-0.020784620195627213,
-0.060965221375226974,
-0.030098699033260345,
-0.014888697303831577,
-0.03911014646291733,
0.09200136363506317,
-0.060095977038145065,
-0.034522414207458496,
-0.034783173352479935,
-0.03447609394788742,
0.12319104373455048,
-0.03488992527127266,
-0.10121176391839981,
0.04076863080263138,
-0.00825948640704155,
-0.004892783705145121,
-0.044179074466228485,
-0.021358923986554146,
0.0091256657615304,
0.023281829431653023,
0.02312101051211357,
0.01230443175882101,
0.05126146227121353,
0.022378290072083473,
0.02971160225570202,
-0.04136420786380768,
0.03056591935455799,
0.009673581458628178,
0.04483919218182564,
-0.045342281460762024,
-0.01483877282589674,
-0.029359277337789536,
0.056769777089357376,
0.05150623247027397,
0.07367298752069473,
0.020533183589577675,
0.05428925156593323,
0.09840256720781326,
0.053454261273145676,
0.12646757066249847,
-0.008895570412278175,
0.033972933888435364,
-0.056433700025081635,
0.018171129748225212,
0.05269390717148781,
0.053307272493839264,
-0.012596981599926949,
-4.467558767373703e-8,
-0.10107661038637161,
0.06948389858007431,
0.08153758198022842,
0.07373086363077164,
0.026065539568662643,
-0.009983299300074577,
-0.04900623485445976,
-0.002229786477982998,
-0.04982150346040726,
-0.06271488219499588,
0.05050110071897507,
0.0989326611161232,
0.005089962854981422,
-0.11485116183757782,
-0.05846407637000084,
0.08470073342323303,
-0.014614689163863659,
0.039107318967580795,
0.006290396209806204,
0.018111394718289375,
0.059205543249845505,
0.03634721040725708,
0.051204126328229904,
-0.06144197657704353,
0.038253430277109146,
-0.053944990038871765,
0.006755521986633539,
0.09504558145999908,
0.056448038667440414,
0.017114222049713135,
-0.0050012352876365185,
-0.07074906677007675,
0.02968418598175049,
0.009579932317137718,
0.08268257975578308,
0.05334784835577011,
0.03851831331849098,
-0.11668842285871506,
0.0174074899405241,
-0.04436641186475754,
0.05156371369957924,
0.06564147025346756,
0.020772784948349,
-0.0076666539534926414,
-0.0046775261871516705,
0.013560332357883453,
-0.04800761863589287,
-0.05417663976550102,
0.03177812695503235,
-0.05442265048623085,
-0.034523963928222656,
-0.049964603036642075,
-0.09551125764846802,
0.03672312572598457,
0.08012490719556808,
-0.05069452524185181,
-0.06671638041734695,
-0.05244189873337746,
0.021207045763731003,
-0.002733697183430195,
0.05252263322472572,
0.013969061896204948,
-0.11271754652261734,
-0.00429391348734498
] |
A Simple Framework for Open-Vocabulary Zero-Shot Segmentation | https://openreview.net/forum?id=QzPKSUUcud | [
"Thomas Stegmüller",
"Tim Lebailly",
"Nikola Đukić",
"Behzad Bozorgtabar",
"Tinne Tuytelaars",
"Jean-Philippe Thiran"
] | Poster | Zero-shot classification capabilities naturally arise in models trained within a vision-language contrastive framework. Despite their classification prowess, these models struggle in dense tasks like zero-shot open-vocabulary segmentation. This deficiency is often attributed to the absence of localization cues in captions and the intertwined nature of the learning process, which encompasses both image/text representation learning and cross-modality alignment. To tackle these issues, we propose SimZSS, a $\textbf{Sim}$ple framework for open-vocabulary $\textbf{Z}$ero-$\textbf{S}$hot $\textbf{S}$egmentation. The method is founded on two key principles: i) leveraging frozen vision-only models that exhibit spatial awareness while exclusively aligning the text encoder and ii) exploiting the discrete nature of text and linguistic knowledge to pinpoint local concepts within captions. By capitalizing on the quality of the visual representations, our method requires only image-caption pair datasets and adapts to both small curated and large-scale noisy datasets. When trained on COCO Captions across 8 GPUs, SimZSS achieves state-of-the-art results on 7 out of 8 benchmark datasets in less than 15 minutes. Our code and pretrained models are publicly available at https://github.com/tileb1/simzss. | Vision-language models, Zero-shot segmentation | We densely align vision and language for open-vocabulary zero-shot segmentation. | 1,573 | 2406.16085 | [
0.059075530618429184,
-0.08765105903148651,
-0.0024989349767565727,
0.028603509068489075,
0.11568812280893326,
-0.0048267170786857605,
0.017037533223628998,
-0.019733482971787453,
0.04302392154932022,
-0.017404841259121895,
0.01190347783267498,
-0.06995023041963577,
0.03743898123502731,
0.07323743402957916,
-0.009608384221792221,
-0.007285017520189285,
0.08854783326387405,
0.1075541228055954,
-0.10341815650463104,
-0.013556456193327904,
0.0894434005022049,
-0.0019955977331846952,
0.12113472074270248,
0.001830951776355505,
0.029014116153120995,
-0.020056819543242455,
-0.013833719305694103,
-0.012605799362063408,
0.026386085897684097,
-0.05680316686630249,
0.03089950978755951,
-0.008419656194746494,
-0.024621957913041115,
0.03827895596623421,
-0.009895938448607922,
0.06015380844473839,
0.06020838022232056,
-0.030783098191022873,
-0.048646412789821625,
-0.024408388882875443,
0.003950204700231552,
0.05858316645026207,
-0.01494942232966423,
0.04830971360206604,
0.11364416778087616,
0.03788295388221741,
0.024190817028284073,
-0.028121400624513626,
0.00473834527656436,
-0.027602700516581535,
-0.07856880873441696,
-0.062075383961200714,
-0.07399852573871613,
0.04006034880876541,
-0.014810455963015556,
0.04546007141470909,
0.0323084332048893,
-0.06416318565607071,
0.06973210722208023,
-0.06523317098617554,
-0.023908942937850952,
-0.1171157956123352,
0.0021350893657654524,
0.031531739979982376,
-0.01937892660498619,
-0.0423407144844532,
0.00631606625393033,
0.00928976945579052,
-0.022617157548666,
-0.01844489946961403,
-0.016470640897750854,
0.05092305317521095,
0.011597245931625366,
0.030682245269417763,
-0.038962893187999725,
0.06173935905098915,
0.023097572848200798,
-0.049283888190984726,
0.08595026284456253,
-0.04959740862250328,
0.08351971954107285,
-0.05660531297326088,
0.057237621396780014,
-0.04764498025178909,
0.03368925303220749,
0.01686178520321846,
-0.007294626906514168,
0.026062145829200745,
-0.00008560103742638603,
0.0054521141573786736,
-0.06006892770528793,
-0.037224557250738144,
0.025283310562372208,
-0.008436070755124092,
0.01261197216808796,
-0.02828369103372097,
-0.025896724313497543,
-0.05704566091299057,
-0.016049008816480637,
0.08030160516500473,
0.04377908632159233,
0.036997806280851364,
0.0516909584403038,
-0.013434448279440403,
0.05336261913180351,
0.040382690727710724,
0.06195810064673424,
-0.024542422965168953,
-0.008152919821441174,
-0.06969691067934036,
0.029416397213935852,
-0.015142432413995266,
-0.002536306856200099,
-0.04069891571998596,
-0.040419839322566986,
-0.03511684760451317,
-0.04000397026538849,
-0.04504197835922241,
0.009236999787390232,
0.020938269793987274,
-0.08082111179828644,
0.05498649552464485,
-0.07464342564344406,
0.03691142797470093,
0.033580146729946136,
-0.03933671861886978,
0.012738890014588833,
1.7073030561556603e-33,
0.06334121525287628,
0.058186598122119904,
0.09423376619815826,
0.0015629865229129791,
0.07127410918474197,
-0.015357949770987034,
0.013010865077376366,
0.012704231776297092,
-0.06458473950624466,
0.02949744649231434,
0.0039838021621108055,
-0.004311054479330778,
-0.03006257861852646,
0.1065613329410553,
0.038975413888692856,
-0.04456876218318939,
-0.014159717597067356,
-0.019835691899061203,
-0.036837466061115265,
0.1097426563501358,
-0.02204311266541481,
0.0021010118070989847,
-0.008625122718513012,
-0.021998774260282516,
-0.0035699743311852217,
0.00585022009909153,
0.05195191502571106,
-0.10721223056316376,
-0.0704881101846695,
0.028781969100236893,
-0.11327601969242096,
0.049548037350177765,
0.03023865446448326,
0.04802161455154419,
0.039815157651901245,
-0.06116948649287224,
-0.03495548665523529,
-0.04117204248905182,
0.012845107354223728,
-0.06316983699798584,
-0.028633223846554756,
0.04195886477828026,
-0.08484160155057907,
-0.035603947937488556,
-0.05183373764157295,
-0.05990709364414215,
0.01077846996486187,
0.08864811807870865,
-0.022193700075149536,
0.015999775379896164,
0.049413297325372696,
0.018195923417806625,
-0.06132395192980766,
-0.013681289739906788,
-0.10945525020360947,
0.012952863238751888,
0.004141387064009905,
-0.02975473366677761,
-0.007861174643039703,
-0.04690801724791527,
0.0014566731406375766,
-0.0023615341633558273,
-0.03300061449408531,
0.044154100120067596,
0.03486566245555878,
0.0631546601653099,
-0.06039625033736229,
-0.010967552661895752,
0.05101071670651436,
-0.0076440973207354546,
-0.04468880966305733,
0.0029500755481421947,
-0.04306977987289429,
-0.023129917681217194,
0.04955282434821129,
0.06551756709814072,
0.04392370954155922,
-0.06069059297442436,
0.06054520234465599,
0.10394168645143509,
-0.02501104399561882,
-0.019207295030355453,
-0.024830836802721024,
-0.034701552242040634,
-0.11029843240976334,
-0.07012046873569489,
0.04730862379074097,
-0.033473316580057144,
0.0015247598057612777,
0.027749253436923027,
-0.03642285615205765,
-0.052559882402420044,
-0.02065768837928772,
0.016191426664590836,
0.02180406078696251,
-5.829502860225753e-34,
0.02724957838654518,
0.040574394166469574,
-0.035724129527807236,
0.06416846066713333,
-0.021963777020573616,
-0.012943783774971962,
0.030430080369114876,
0.03929781913757324,
0.054836027324199677,
-0.07187645882368088,
0.004692322574555874,
-0.02477412484586239,
-0.013669424690306187,
-0.06502766162157059,
0.06864510476589203,
0.01632789522409439,
0.021104639396071434,
0.06960729509592056,
0.005677155684679747,
0.12743140757083893,
0.03771481662988663,
0.047853052616119385,
-0.03709058463573456,
0.037548817694187164,
-0.0015829431358724833,
0.09162970632314682,
-0.019014917314052582,
0.009338864125311375,
-0.021048642694950104,
-0.009645059704780579,
0.028252236545085907,
0.028250962495803833,
-0.02464267425239086,
-0.03385042026638985,
-0.0857413113117218,
0.04131689295172691,
0.06192731857299805,
-0.022965917363762856,
0.03860248252749443,
0.048948176205158234,
0.10066826641559601,
-0.04196930304169655,
-0.08114521950483322,
0.0461529977619648,
-0.04531647637486458,
-0.08073551207780838,
-0.10851004719734192,
-0.03849807009100914,
-0.028180113062262535,
0.056926123797893524,
-0.11122570186853409,
0.050418559461832047,
-0.11713877320289612,
0.036068353801965714,
0.00295265787281096,
-0.11049489676952362,
-0.06768379360437393,
-0.04697757959365845,
0.032179899513721466,
0.011830704286694527,
0.012389939278364182,
0.014632337726652622,
-0.05717809125781059,
0.016005560755729675,
0.02994287945330143,
-0.06894838809967041,
-0.052152011543512344,
-0.001730073126964271,
0.004371160175651312,
-0.03364018723368645,
0.08814453333616257,
-0.03342505544424057,
0.011632040143013,
-0.03897877410054207,
-0.06757405400276184,
0.048470959067344666,
0.02705540880560875,
-0.020654266700148582,
-0.0203414149582386,
-0.07067853212356567,
0.06941729038953781,
-0.04229746758937836,
0.027181794866919518,
0.1023140624165535,
0.13085882365703583,
0.10238683223724365,
0.06268917769193649,
-0.0042946795001626015,
0.08485856652259827,
0.004715784918516874,
-0.07411586493253708,
-0.07780977338552475,
0.04536309465765953,
0.08506758511066437,
-0.06242012232542038,
-4.220962779299953e-8,
-0.09682003408670425,
-0.004508335608989,
-0.02874734438955784,
0.020116470754146576,
-0.06267030537128448,
-0.06802259385585785,
-0.0598047710955143,
0.07209930568933487,
-0.026001276448369026,
-0.020481111481785774,
0.11892136931419373,
0.016077710315585136,
0.005187198054045439,
-0.05233757197856903,
-0.01574706844985485,
0.12811748683452606,
0.04616933315992355,
0.057809822261333466,
0.024138446897268295,
0.09019618481397629,
-0.04975426197052002,
0.01431132759898901,
0.03739571571350098,
-0.02613215707242489,
0.0010036035673692822,
-0.04457845538854599,
-0.029344284906983376,
0.07616615295410156,
0.11535997688770294,
-0.05802752077579498,
0.04203063249588013,
0.0351720005273819,
0.037156201899051666,
-0.0009057199931703508,
0.0013457650784403086,
0.07145220041275024,
0.02422221563756466,
0.011142782866954803,
-0.04032307490706444,
-0.053642429411411285,
-0.009128449484705925,
0.024227548390626907,
0.053736455738544464,
-0.03362933546304703,
-0.003807647153735161,
0.03537863492965698,
0.004578867461532354,
-0.06466520577669144,
0.06132758408784866,
-0.021845703944563866,
-0.023664899170398712,
0.05905640870332718,
-0.052192896604537964,
0.08590462803840637,
0.03220955282449722,
-0.021068325266242027,
0.015832502394914627,
0.006934689357876778,
0.048767343163490295,
0.052320580929517746,
0.00997644942253828,
-0.004153203684836626,
-0.11348092555999756,
-0.03321729227900505
] |
Understanding the Generalization of In-Context Learning in Transformers: An Empirical Study | https://openreview.net/forum?id=yOhNLIqTEF | [
"Xingxuan Zhang",
"Haoran Wang",
"Jiansheng Li",
"Yuan Xue",
"Shikai Guan",
"Renzhe Xu",
"Hao Zou",
"Han Yu",
"Peng Cui"
] | Poster | Large language models (LLMs) like GPT-4 and LLaMA-3 utilize the powerful in-context learning (ICL) capability of Transformer architecture to learn on the fly from limited examples. While ICL underpins many LLM applications, its full potential remains hindered by a limited understanding of its generalization boundaries and vulnerabilities. We present a systematic investigation of transformers' generalization capability with ICL relative to training data coverage by defining a task-centric framework along three dimensions: inter-problem, intra-problem, and intra-task generalization. Through extensive simulation and real-world experiments, encompassing tasks such as function fitting, API calling, and translation, we find that transformers lack inter-problem generalization with ICL, but excel in intra-task and intra-problem generalization. When the training data includes a greater variety of mixed tasks, it significantly enhances the generalization ability of ICL on unseen tasks and even on known simple tasks. This guides us in designing training data to maximize the diversity of tasks covered and to combine different tasks whenever possible, rather than solely focusing on the target task for testing. | generalization, in-context learning, transformer | null | 1,554 | 2503.15579 | [
-0.08111998438835144,
0.004670246038585901,
0.025469332933425903,
0.0008871914469636977,
0.06047610193490982,
-0.059824537485837936,
0.038956835865974426,
0.024384332820773125,
-0.003992915153503418,
-0.04971315339207649,
-0.07652951776981354,
-0.08846297860145569,
0.07134732604026794,
0.07921546697616577,
0.09560386836528778,
-0.050743766129016876,
0.06266457587480545,
0.04696457460522652,
-0.16768547892570496,
-0.06435412168502808,
0.055089645087718964,
0.05538443475961685,
0.028634527698159218,
0.014642560854554176,
-0.0071359071880578995,
0.0014487062580883503,
0.021884655579924583,
-0.0813584253191948,
0.010440992191433907,
0.01837354153394699,
0.015862256288528442,
0.061150435358285904,
-0.049049150198698044,
0.07064623385667801,
-0.010097919963300228,
0.09005837887525558,
-0.1209322139620781,
0.031262509524822235,
0.04615669324994087,
-0.03368480131030083,
-0.005819404032081366,
0.013663791120052338,
0.02361268922686577,
0.011504213325679302,
0.04444023221731186,
-0.03586360439658165,
0.009401840157806873,
-0.08270371705293655,
-0.044817570596933365,
-0.000594750395976007,
-0.05173351243138313,
-0.04312385246157646,
-0.04419313743710518,
0.08081697672605515,
-0.021167177706956863,
-0.00009123136987909675,
0.04962409660220146,
0.011354787275195122,
-0.025588274002075195,
0.014429274946451187,
-0.053689371794462204,
-0.11002956330776215,
-0.06108935922384262,
-0.006305148359388113,
-0.07474297285079956,
-0.0030418571550399065,
0.026721393689513206,
0.0595276840031147,
-0.06440667062997818,
-0.052773892879486084,
-0.06412369757890701,
0.06834115833044052,
-0.05165723338723183,
0.0881090760231018,
0.09967220574617386,
0.022602522745728493,
0.03178349509835243,
-0.05361853539943695,
0.025170721113681793,
-0.06414235383272171,
0.06691611558198929,
0.0689840316772461,
0.03154098615050316,
0.05867821350693703,
0.1566331833600998,
0.023941077291965485,
0.006904181558638811,
0.02374693751335144,
-0.014482591301202774,
0.0300630871206522,
-0.05654563754796982,
-0.10751491785049438,
0.04442528635263443,
0.028576107695698738,
0.04475504904985428,
-0.01112257782369852,
0.043512918055057526,
-0.08633756637573242,
-0.023153584450483322,
0.09101453423500061,
-0.012694349512457848,
0.015745636075735092,
0.028963712975382805,
-0.022632917389273643,
-0.0785536840558052,
-0.02309711091220379,
-0.03033342771232128,
0.015961622819304466,
-0.017461320385336876,
-0.10463671386241913,
0.008469417691230774,
0.03838750347495079,
0.03146748244762421,
-0.04474346712231636,
0.006864635273814201,
-0.036390867084264755,
0.04049123451113701,
-0.0483882874250412,
-0.020728761330246925,
0.11386449635028839,
-0.03055143728852272,
0.0358610637485981,
0.01689464971423149,
0.05771150067448616,
-0.02742910012602806,
0.0009924430632963777,
-0.016398873180150986,
2.8994837329588625e-33,
0.028790369629859924,
0.015007725916802883,
0.02675684541463852,
0.08382568508386612,
0.0422821044921875,
-0.02075444906949997,
0.06555324792861938,
0.04923205450177193,
0.040575724095106125,
0.03829612582921982,
0.03615787997841835,
0.07214435189962387,
-0.013493938371539116,
0.026990259066224098,
0.006050141993910074,
0.07445025444030762,
-0.06764299422502518,
0.06152809038758278,
-0.06867144256830215,
0.010810299776494503,
0.04494135081768036,
-0.026869261637330055,
0.04218008741736412,
-0.053426045924425125,
0.024458473548293114,
0.013955986127257347,
0.06978822499513626,
-0.06443376839160919,
0.011903394013643265,
0.001577554503455758,
-0.09917140007019043,
-0.005173885263502598,
-0.005310318432748318,
0.10369319468736649,
0.02033647708594799,
-0.0010265713790431619,
0.016851460561156273,
-0.05248050019145012,
-0.03310832008719444,
0.00394844077527523,
-0.04205898195505142,
0.013620101846754551,
0.003988164942711592,
0.0028818349819630384,
-0.013102267868816853,
-0.042976316064596176,
-0.021001098677515984,
-0.026692621409893036,
-0.02750086411833763,
-0.034894078969955444,
0.017006102949380875,
0.03286029025912285,
-0.025758378207683563,
-0.045346375554800034,
0.06428245455026627,
0.05468467250466347,
0.023614730685949326,
0.010570473037660122,
0.049534425139427185,
0.024213653057813644,
-0.0016650928882882,
-0.000943133607506752,
-0.020555732771754265,
0.09984904527664185,
0.058252643793821335,
0.020445145666599274,
-0.039561741054058075,
0.026156097650527954,
0.06665005534887314,
-0.0507613867521286,
-0.005565730854868889,
-0.015807053074240685,
-0.013755274936556816,
0.002723140176385641,
0.08176379650831223,
-0.020251644775271416,
0.05792506784200668,
-0.06196809560060501,
-0.01765703223645687,
0.031357813626527786,
0.04461254924535751,
-0.005142739042639732,
0.06895164400339127,
-0.051124829798936844,
-0.07283646613359451,
-0.03040885180234909,
0.033519912511110306,
-0.10931458324193954,
-0.05327196419239044,
-0.06587780267000198,
-0.06544823199510574,
0.018411841243505478,
-0.0894327312707901,
0.007207056041806936,
0.007466101553291082,
-2.5715580942908844e-33,
-0.03253279626369476,
0.0004198912647552788,
-0.06745308637619019,
0.02963355928659439,
-0.011725611053407192,
-0.05220611020922661,
0.04278068244457245,
0.01009630598127842,
-0.029536472633481026,
0.017193350940942764,
-0.02625715173780918,
-0.047231368720531464,
-0.007971465587615967,
-0.02120758593082428,
0.01732398197054863,
-0.026186781004071236,
-0.04316214844584465,
0.01862737163901329,
0.03250735253095627,
0.10915391147136688,
0.01446934137493372,
0.09934177994728088,
-0.10966013371944427,
0.008035368286073208,
-0.06361120194196701,
-0.019624970853328705,
-0.10116846114397049,
0.02917814441025257,
0.06184826418757439,
0.013152576051652431,
-0.011040098033845425,
-0.009137621149420738,
-0.013640017248690128,
-0.022422699257731438,
-0.07343199849128723,
0.040444497019052505,
0.03900480642914772,
0.009346581064164639,
-0.035727132111787796,
0.07556606084108353,
0.019363733008503914,
-0.05603477731347084,
-0.025043487548828125,
0.040507152676582336,
-0.022013720124959946,
-0.009290044195950031,
-0.07060565799474716,
-0.07376561313867569,
0.016541441902518272,
0.017360320314764977,
-0.03714464232325554,
0.010873327031731606,
-0.08834878355264664,
-0.06015286594629288,
-0.009217077866196632,
-0.07628186792135239,
0.05376477539539337,
-0.07613278925418854,
0.0007822211482562125,
-0.020564379170536995,
-0.007471688557416201,
-0.025726940482854843,
0.10557118058204651,
-0.062003958970308304,
-0.006541331764310598,
-0.03804147243499756,
-0.042120955884456635,
-0.019652780145406723,
0.045779597014188766,
0.04469757899641991,
-0.016050836071372032,
-0.046813566237688065,
-0.003824579296633601,
-0.07343938946723938,
-0.00670252600684762,
-0.02552386000752449,
-0.008129688911139965,
-0.07369861006736755,
-0.005741712171584368,
-0.016466043889522552,
-0.003961540758609772,
-0.04350949078798294,
0.01791016012430191,
0.06946281343698502,
0.02134488709270954,
0.07225941121578217,
0.09241726994514465,
0.1228552833199501,
0.048182256519794464,
-0.03397432342171669,
-0.07175395637750626,
0.06173567473888397,
-0.03983617201447487,
0.052991628646850586,
-0.059421151876449585,
-4.558642530128054e-8,
-0.05747947469353676,
0.07608388364315033,
-0.050846535712480545,
0.0213614534586668,
-0.001612590509466827,
0.022611843422055244,
-0.0764436200261116,
0.10622281581163406,
0.020141134038567543,
0.039917051792144775,
-0.0056883832439780235,
-0.026661790907382965,
-0.024338150396943092,
-0.013544121757149696,
-0.0036552275996655226,
0.07004682719707489,
0.01635664887726307,
0.04055728390812874,
-0.0017506367294117808,
-0.010512685403227806,
0.034992944449186325,
0.08344609290361404,
-0.016317617148160934,
-0.06775444000959396,
0.03771053999662399,
-0.06198177486658096,
-0.02088875137269497,
0.09565061330795288,
0.06542780250310898,
0.022406326606869698,
-0.04728812351822853,
-0.006601092871278524,
-0.0350247360765934,
-0.025592627003788948,
-0.0035624706652015448,
0.11227724701166153,
0.09255257248878479,
-0.04463207721710205,
0.02955387532711029,
-0.03700565546751022,
0.023465421050786972,
0.06438694894313812,
-0.050151973962783813,
0.06814734637737274,
0.011497299186885357,
-0.00993899442255497,
-0.047780752182006836,
-0.11852529644966125,
0.027523575350642204,
-0.02245880290865898,
0.05349986255168915,
-0.0012409505434334278,
-0.0458780936896801,
0.07970570027828217,
0.09625109285116196,
0.049993012100458145,
-0.023667648434638977,
-0.109341099858284,
0.03576250746846199,
0.10110367089509964,
-0.020244350656867027,
0.0554596409201622,
-0.0763477087020874,
0.029397299513220787
] |
Progressive Mixed-Precision Decoding for Efficient LLM Inference | https://openreview.net/forum?id=OVxmpus9NA | [
"Hao Mark Chen",
"Fuwen Tan",
"Alexandros Kouris",
"Royson Lee",
"Hongxiang Fan",
"Stylianos Venieris"
] | Poster | In spite of the great potential of large language models (LLMs) across various tasks, their deployment on resource-constrained devices remains challenging due to their excessive computational and memory demands. Quantization has emerged as an effective solution by storing weights in reduced precision. However, utilizing low precisions (i.e.~2/3-bit) to substantially alleviate the memory-boundedness of LLM decoding, still suffers from prohibitive performance drop. In this work,
we argue that existing approaches fail to explore the diversity in computational patterns, redundancy, and sensitivity to approximations of the different phases of LLM inference, resorting to a uniform quantization policy throughout.
Instead, we propose a novel phase-aware method that selectively allocates precision during different phases of LLM inference, achieving both strong context extraction during prefill and efficient memory bandwidth utilization during decoding. To further address the memory-boundedness of the decoding phase, we introduce Progressive Mixed-Precision Decoding (PMPD), a technique that enables the gradual lowering of precision deeper in the generated sequence, together with a spectrum of precision-switching schedulers that dynamically drive the precision-lowering decisions in either task-adaptive or prompt-adaptive manner.
Extensive evaluation across diverse language tasks shows that when targeting Nvidia GPUs, PMPD achieves 1.4$-$12.2$\times$ speedup in matrix-vector multiplications over fp16 models, while when targeting an LLM-optimized NPU, our approach delivers a throughput gain of 3.8$-$8.0$\times$ over fp16 models and up to 1.54$\times$ over uniform quantization approaches while preserving the output quality. | LLM Quantization, Efficient LLM Inference | We propose a novel decoding scheme that adapts the precision throughout stages in LLM inference. | 1,551 | 2410.13461 | [
0.02584870532155037,
-0.05415985360741615,
-0.007258427329361439,
0.04937983304262161,
0.08262301981449127,
0.008491321466863155,
-0.0706019476056099,
-0.056924089789390564,
0.06035980209708214,
-0.05289805680513382,
-0.05744427815079689,
-0.03497174009680748,
-0.008711226284503937,
0.009222920052707195,
0.049885399639606476,
-0.011447692289948463,
0.061840105801820755,
0.1308910995721817,
-0.07671286910772324,
-0.04237115755677223,
0.07876946032047272,
0.0008160052238963544,
0.02907492406666279,
-0.018874458968639374,
0.08831043541431427,
-0.008401242084801197,
-0.037331681698560715,
0.009880796074867249,
0.10395585745573044,
0.02645360492169857,
0.04967648535966873,
0.11110544204711914,
-0.024035820737481117,
0.0344146229326725,
-0.024519210681319237,
0.06374483555555344,
-0.027845432981848717,
-0.019511789083480835,
0.013252126052975655,
-0.024316608905792236,
0.012458877637982368,
0.006027508527040482,
0.034611914306879044,
0.040930427610874176,
0.01914030872285366,
0.029719045385718346,
0.03303945064544678,
0.013065622188150883,
-0.020158415660262108,
-0.004870248958468437,
0.056894220411777496,
0.04612428694963455,
-0.05513354763388634,
0.01963595487177372,
-0.07620958983898163,
-0.015188012272119522,
0.06629098206758499,
0.061359819024801254,
-0.03895469754934311,
0.02451344206929207,
-0.11334896832704544,
-0.14104098081588745,
0.028225472196936607,
0.02440694533288479,
-0.04041599482297897,
-0.01197455357760191,
0.04853758215904236,
0.07262876629829407,
-0.0015353712951764464,
0.03445320203900337,
-0.0331348180770874,
0.009671464562416077,
-0.05234965682029724,
0.09575489163398743,
-0.02417636103928089,
0.03847301006317139,
-0.012533777393400669,
-0.05197978392243385,
0.08416096866130829,
-0.0043125152587890625,
-0.025131475180387497,
-0.014721704646945,
0.05271909013390541,
-0.024657631292939186,
0.14757591485977173,
-0.011488725431263447,
-0.02170213870704174,
0.039052899926900864,
-0.01986399106681347,
0.04453813657164574,
-0.08307255059480667,
-0.052907757461071014,
0.026428991928696632,
0.05679502338171005,
0.017024101689457893,
0.046831369400024414,
0.011331828311085701,
-0.035774726420640945,
-0.018177146092057228,
0.08061004430055618,
0.07003027200698853,
0.07098913937807083,
-0.023188745602965355,
-0.06305817514657974,
-0.05785601958632469,
-0.08110719174146652,
0.05074248090386391,
0.09685808420181274,
-0.040817029774188995,
-0.04595731198787689,
0.08548249304294586,
0.08339182287454605,
0.056010760366916656,
0.026609865948557854,
0.03796696290373802,
-0.020277148112654686,
0.021540731191635132,
-0.08107928186655045,
0.0464487187564373,
0.0325838141143322,
-0.052334461361169815,
-0.011689320206642151,
-0.04867425188422203,
-0.0014003454707562923,
0.009813958778977394,
-0.029335174709558487,
-0.05391259863972664,
7.898043687419253e-34,
0.038179561495780945,
0.0051151467487216,
0.022656802088022232,
0.0584227629005909,
0.0040214466862380505,
0.06593597680330276,
0.0407041497528553,
0.0751619040966034,
-0.0542210228741169,
0.046069104224443436,
0.022948339581489563,
-0.03600494563579559,
-0.062041230499744415,
0.0027376350481063128,
0.04412528872489929,
-0.002921218518167734,
-0.04258709028363228,
0.04228874668478966,
-0.020019177347421646,
-0.0467071458697319,
0.05763191729784012,
-0.018127193674445152,
-0.01671847142279148,
-0.05774049460887909,
0.0006818577530793846,
0.04999377578496933,
0.0807720273733139,
-0.040107205510139465,
-0.021160895004868507,
0.027068613097071648,
-0.11311503499746323,
-0.04830959811806679,
0.0010755021357908845,
0.013626699335873127,
0.047787152230739594,
-0.061478141695261,
0.004035762045532465,
-0.017722342163324356,
-0.018766213208436966,
-0.03809281438589096,
-0.00015909731155261397,
0.018410947173833847,
0.0553467757999897,
-0.0405760332942009,
-0.10356397181749344,
-0.11062370985746384,
-0.040266551077365875,
0.00980796106159687,
-0.06360086053609848,
-0.00429760804399848,
0.08424503356218338,
0.023015078157186508,
-0.05443871021270752,
-0.03961101919412613,
0.048496510833501816,
-0.07488253712654114,
-0.010724425315856934,
-0.00487864064052701,
0.020528724417090416,
0.07653886079788208,
-0.005141282454133034,
0.02366468869149685,
-0.036585405468940735,
0.07542192190885544,
0.08163107186555862,
0.053152091801166534,
-0.08952463418245316,
0.07902584224939346,
-0.025810495018959045,
-0.0030357488431036472,
-0.010256445966660976,
-0.014375782571732998,
0.07433950901031494,
-0.06491793692111969,
0.034982822835445404,
0.019060827791690826,
0.05179286375641823,
-0.0498185008764267,
-0.02143154852092266,
0.029691392555832863,
0.08866472542285919,
0.038955289870500565,
-0.02425300143659115,
0.011980137787759304,
-0.08917398750782013,
-0.06993621587753296,
0.008446648716926575,
-0.07190421968698502,
-0.07238462567329407,
-0.006692436523735523,
-0.052257388830184937,
-0.05472518876194954,
0.013968522660434246,
-0.02644704468548298,
-0.0772164911031723,
-1.5207830512423685e-33,
-0.042511943727731705,
0.008599039167165756,
-0.01944326050579548,
0.08448819071054459,
-0.023832377046346664,
-0.10957261919975281,
0.020400041714310646,
-0.016332034021615982,
-0.016214190050959587,
-0.11739177256822586,
-0.02867615967988968,
-0.03985299915075302,
0.0305539071559906,
0.020959723740816116,
0.06712165474891663,
-0.02875458635389805,
-0.00955204013735056,
0.03764672204852104,
0.027516378089785576,
0.08296925574541092,
-0.0034561739303171635,
0.05596923828125,
-0.030223052948713303,
0.0074480758048594,
0.03226565197110176,
0.03245912492275238,
-0.07737679034471512,
0.03712834045290947,
-0.007586168125271797,
-0.015103902667760849,
-0.07097163051366806,
-0.03982223570346832,
-0.025449777022004128,
-0.05354047194123268,
-0.07760559767484665,
0.015456032007932663,
0.026739442721009254,
0.048382557928562164,
-0.011563559994101524,
0.04192725569009781,
0.11712796241044998,
-0.10404126346111298,
-0.06483722478151321,
0.01811673492193222,
0.0290600024163723,
-0.04733845219016075,
-0.11162007600069046,
0.004056878387928009,
0.057760030031204224,
-0.018864810466766357,
-0.014026188291609287,
0.018121901899576187,
-0.0032640809658914804,
0.05945364013314247,
0.043045204132795334,
-0.08434077352285385,
-0.08144500851631165,
-0.05832265689969063,
-0.005722423084080219,
-0.03892732784152031,
-0.05917089059948921,
-0.013984833844006062,
-0.031648196280002594,
-0.07768440991640091,
0.016009796410799026,
-0.04579579830169678,
0.0103665916249156,
-0.08320444822311401,
-0.007881203666329384,
-0.023487171158194542,
0.002631212817505002,
-0.03227579966187477,
-0.038732655346393585,
0.024197937920689583,
-0.022069819271564484,
-0.022657768800854683,
0.0037029068917036057,
-0.043219659477472305,
-0.013074774295091629,
0.056330520659685135,
0.02636733092367649,
0.0341898612678051,
-0.0220248531550169,
0.02778191864490509,
0.046186383813619614,
0.016997674480080605,
-0.054865095764398575,
0.09082019329071045,
0.01873661018908024,
0.01950753666460514,
-0.0738598108291626,
0.08517047762870789,
-0.002718319185078144,
0.009361843578517437,
-0.05003951117396355,
-4.3772566016286873e-8,
-0.04073784500360489,
0.037631552666425705,
0.00685507757589221,
0.00943917129188776,
0.030497416853904724,
-0.0620994046330452,
-0.06724441796541214,
0.03590499982237816,
0.06825138628482819,
0.002313624368980527,
0.1146712601184845,
-0.05286519601941109,
-0.05643985792994499,
-0.0726080909371376,
0.00863701943308115,
0.08731526136398315,
0.0415494367480278,
-0.03736129403114319,
0.0051241363398730755,
-0.05815579369664192,
-0.01330147311091423,
0.036957647651433945,
0.001597363967448473,
-0.022276954725384712,
0.05331386253237724,
-0.0809822827577591,
0.029004674404859543,
0.14982233941555023,
0.11726094782352448,
-0.030651142820715904,
-0.0592893622815609,
-0.0038590566255152225,
-0.05659791827201843,
0.021013295277953148,
-0.047549065202474594,
0.051720984280109406,
0.0032840785570442677,
0.03821887448430061,
0.08054500073194504,
-0.09185947477817535,
0.03911344334483147,
-0.012535973452031612,
-0.08449063450098038,
0.07895781844854355,
0.024339335039258003,
-0.027044273912906647,
0.01381338108330965,
-0.003491661511361599,
-0.011275884695351124,
-0.02317711152136326,
-0.021046707406640053,
0.05070488899946213,
-0.00411220034584403,
0.05594278872013092,
0.014352318830788136,
0.05013299360871315,
0.011755620129406452,
-0.06818856298923492,
-0.03956114873290062,
0.041019219905138016,
0.010100471787154675,
0.05001920834183693,
-0.12787295877933502,
-0.01893274299800396
] |
Manifold Constraint Reduces Exposure Bias in Accelerated Diffusion Sampling | https://openreview.net/forum?id=5xmXUwDxep | [
"Yuzhe YAO",
"Jun Chen",
"Zeyi Huang",
"Haonan Lin",
"Mengmeng Wang",
"Guang Dai",
"Jingdong Wang"
] | Poster | Diffusion models have demonstrated significant potential for generating high-quality images, audio, and videos. However, their iterative inference process entails substantial computational costs, limiting practical applications. Recently, researchers have introduced accelerated sampling methods that enable diffusion models to generate samples with far fewer timesteps than those used during training. Nonetheless, as the number of sampling steps decreases, the prediction errors significantly degrade the quality of generated outputs. Additionally, the exposure bias in diffusion models further amplifies these errors. To address these challenges, we leverage a manifold hypothesis to explore the exposure bias problem in depth. Based on this geometric perspective, we propose a manifold constraint that effectively reduces exposure bias during accelerated sampling of diffusion models. Notably, our method involves no additional training and requires only minimal hyperparameter tuning. Extensive experiments demonstrate the effectiveness of our approach, achieving a FID score of 15.60 with 10-step SDXL on MS-COCO, surpassing the baseline by a reduction of 2.57 in FID. | Diffusion Models, Exposure Bias | null | 1,547 | null | [
0.0029439101926982403,
-0.09638851881027222,
0.05243546888232231,
0.04273313283920288,
0.0948375016450882,
-0.018481602892279625,
-0.0540195032954216,
-0.07543259114027023,
0.027896076440811157,
-0.026938114315271378,
-0.06926015764474869,
-0.03341468423604965,
-0.008524803444743156,
0.05672145262360573,
-0.0268836859613657,
-0.06304047256708145,
0.07742895931005478,
0.06821141391992569,
-0.07814137637615204,
-0.04232260212302208,
-0.026595626026391983,
-0.0070014880038797855,
0.029213277623057365,
-0.01446288637816906,
0.04639261215925217,
-0.0669226422905922,
0.06413436681032181,
-0.06514821201562881,
0.033037252724170685,
-0.04072576388716698,
0.08902550488710403,
-0.0013675399823114276,
0.02445952221751213,
-0.04528238996863365,
-0.014553134329617023,
-0.00467356713488698,
-0.03598629683256149,
0.014334818348288536,
-0.002689850050956011,
-0.005356919951736927,
0.016608387231826782,
0.07702352851629257,
0.040453482419252396,
0.042762212455272675,
0.01682954840362072,
-0.00300421635620296,
0.009569002315402031,
-0.02843690849840641,
0.009633656591176987,
0.08024396747350693,
-0.027586983516812325,
-0.01106429472565651,
-0.06585510820150375,
0.019256291911005974,
-0.07274165749549866,
0.008466345258057117,
0.0617038831114769,
0.007435888051986694,
0.026981886476278305,
-0.020168492570519447,
-0.06057937443256378,
-0.1221405565738678,
-0.058352407068014145,
0.03011021390557289,
-0.007152029313147068,
-0.016669901087880135,
0.03261279687285423,
0.049150947481393814,
-0.0038965395651757717,
0.09473324567079544,
-0.051063310354948044,
0.07473811507225037,
-0.10629254579544067,
-0.018026620149612427,
0.024455362930893898,
0.0262437891215086,
0.04103953018784523,
-0.046511393040418625,
0.02421039529144764,
-0.046799104660749435,
0.09139024466276169,
-0.015395115129649639,
0.09770811349153519,
-0.047265902161598206,
0.01152343861758709,
-0.010983040556311607,
0.004173688590526581,
0.0450517013669014,
-0.031224209815263748,
-0.0002901925181504339,
-0.03657982870936394,
-0.0014556498499587178,
-0.12036309391260147,
0.018438277766108513,
-0.004165002144873142,
0.013641434721648693,
-0.01842997968196869,
-0.04733423516154289,
0.03143152967095375,
0.08446668088436127,
0.0005142324371263385,
-0.07277929782867432,
-0.06539286673069,
0.03696517273783684,
0.0664680227637291,
-0.07735345512628555,
0.09823273867368698,
0.050589531660079956,
-0.015795189887285233,
-0.024458257481455803,
0.07427854090929031,
0.048508383333683014,
0.04348012059926987,
0.023076385259628296,
0.04551912471652031,
-0.030627796426415443,
0.026695221662521362,
-0.028438713401556015,
0.015201115980744362,
0.03410269320011139,
-0.022341901436448097,
-0.013916733674705029,
-0.07485827058553696,
-0.0035216519609093666,
0.0030515340622514486,
-0.09205132722854614,
-0.02443729154765606,
1.1061835304779591e-33,
-0.012566880322992802,
-0.0005255286814644933,
0.09218090027570724,
-0.01383020170032978,
0.05844854190945625,
0.00689309136942029,
0.03688323125243187,
-0.03502444177865982,
-0.014338064007461071,
0.005833304487168789,
-0.045072685927152634,
-0.04787837713956833,
-0.07522435486316681,
0.08363315463066101,
0.02422258071601391,
0.005622854456305504,
-0.01678125001490116,
0.12388501316308975,
-0.05671318620443344,
0.03917772322893143,
0.028245680034160614,
-0.07942458242177963,
-0.048254404217004776,
0.015483889728784561,
-0.031930238008499146,
0.04969199746847153,
0.028784971684217453,
-0.005512821022421122,
0.03125709667801857,
0.01567496918141842,
-0.11988871544599533,
-0.03525197505950928,
-0.025608686730265617,
0.02787643112242222,
0.05762345716357231,
0.060596976429224014,
-0.0022079937625676394,
0.025568069890141487,
-0.04177282750606537,
-0.00010668535105651245,
-0.044249027967453,
0.07634809613227844,
0.07410600036382675,
-0.02958374284207821,
-0.18528595566749573,
-0.048152223229408264,
-0.0017030263552442193,
-0.007477148901671171,
-0.06538800150156021,
0.02769497036933899,
0.010601652786135674,
-0.03909839317202568,
0.049676407128572464,
-0.050027791410684586,
-0.07948513329029083,
-0.019619883969426155,
0.07072239369153976,
-0.06041870266199112,
0.022685999050736427,
0.045123763382434845,
0.04464380070567131,
-0.012135051190853119,
-0.06075958535075188,
0.022444719448685646,
0.007884724996984005,
0.030570384114980698,
0.010071977972984314,
0.01171465776860714,
0.01696307212114334,
0.027722850441932678,
-0.04665829613804817,
0.049935948103666306,
-0.06863337010145187,
-0.10624120384454727,
0.054616719484329224,
0.01786188967525959,
0.041853610426187515,
-0.004286519251763821,
0.034370072185993195,
0.061956267803907394,
-0.08673902601003647,
-0.03598227724432945,
-0.04689594730734825,
-0.10181652754545212,
-0.04440775513648987,
0.010686502791941166,
-0.028945161029696465,
-0.0629054456949234,
-0.01044743973761797,
-0.0593661367893219,
-0.04356160759925842,
-0.0073865377344191074,
-0.005737560801208019,
0.016791928559541702,
0.004052357282489538,
-7.517486115712672e-34,
0.05068112164735794,
0.07511026412248611,
0.06094040349125862,
0.0567111037671566,
-0.044329073280096054,
0.01902889646589756,
0.03585220128297806,
0.044146761298179626,
0.06900770217180252,
-0.06116541847586632,
-0.02434571087360382,
-0.02740342915058136,
0.003267622319981456,
-0.015658387914299965,
-0.010710152797400951,
-0.0008209100342355669,
-0.019005857408046722,
-0.029335791245102882,
-0.061076052486896515,
-0.007838727906346321,
-0.016011016443371773,
0.053735990077257156,
0.046912431716918945,
0.017934400588274002,
-0.04851462319493294,
0.06722361594438553,
-0.04749196395277977,
0.07659934461116791,
0.033605411648750305,
-0.0665886402130127,
-0.07763092964887619,
-0.04352997615933418,
-0.024712134152650833,
-0.0729418471455574,
-0.07287713140249252,
0.10968294739723206,
0.08492889255285263,
0.04079429805278778,
-0.05431394279003143,
0.09674520045518875,
0.019663572311401367,
-0.004973390139639378,
-0.06046563759446144,
0.0688101053237915,
0.02959093637764454,
-0.03403780981898308,
-0.041186098009347916,
-0.012290986254811287,
0.020281387493014336,
0.02812296338379383,
-0.027932235971093178,
0.07183101028203964,
-0.006249550264328718,
0.08716999739408493,
-0.019786562770605087,
-0.11266109347343445,
-0.10710342228412628,
-0.003671196522191167,
0.12241104990243912,
-0.008251124992966652,
-0.041901737451553345,
0.005725557450205088,
-0.08335509151220322,
-0.05347553268074989,
-0.04273884370923042,
-0.011216901242733002,
-0.0727100819349289,
-0.029820794239640236,
0.01893775910139084,
0.11979395896196365,
-0.011923017911612988,
0.0011394231114536524,
-0.004952951334416866,
0.0033654258586466312,
-0.02489321678876877,
-0.06642181426286697,
0.022216031327843666,
-0.045420438051223755,
0.008590941317379475,
-0.04711431637406349,
0.06007326766848564,
0.018521979451179504,
0.04093948006629944,
0.01885075494647026,
0.10558314621448517,
0.0557076521217823,
0.06874828040599823,
-0.05499689653515816,
0.03231291100382805,
0.049988675862550735,
-0.044078849256038666,
-0.009257691912353039,
0.07678624242544174,
0.05959378927946091,
-0.0943571925163269,
-4.184718704891566e-8,
-0.0781727284193039,
0.0623946413397789,
-0.01871424727141857,
0.003526763990521431,
-0.002054742071777582,
-0.038081083446741104,
0.06261494755744934,
0.100599005818367,
0.01832459308207035,
-0.05111449956893921,
0.10015863925218582,
-0.023876799270510674,
0.000054902546253288165,
0.011323142796754837,
-0.012728379108011723,
0.05509018898010254,
0.01928020641207695,
0.03302587941288948,
0.005273162852972746,
-0.0028876212891191244,
-0.05229710042476654,
0.06281235069036484,
0.07857732474803925,
-0.05333271622657776,
0.07936310023069382,
0.011982721276581287,
0.08186914026737213,
0.006026757415384054,
0.1482265442609787,
-0.053925465792417526,
-0.07340819388628006,
0.01202335674315691,
-0.03991038352251053,
0.004230889491736889,
0.0006701177917420864,
0.05754384398460388,
-0.036294106394052505,
-0.033869095146656036,
-0.06049896404147148,
-0.025660037994384766,
-0.09802760928869247,
0.0849882960319519,
0.009040302596986294,
0.011230950243771076,
0.013623630627989769,
0.058631010353565216,
0.08984270691871643,
0.008948247879743576,
-0.011899502016603947,
-0.003912410233169794,
-0.019813844934105873,
-0.026621712371706963,
0.00046832679072394967,
0.049111027270555496,
0.06984744966030121,
-0.005991664249449968,
0.042235005646944046,
0.021585145965218544,
0.025942092761397362,
0.04111122339963913,
0.0356501005589962,
-0.020863132551312447,
-0.04582063853740692,
-0.03088160790503025
] |
Gaussian Head & Shoulders: High Fidelity Neural Upper Body Avatars with Anchor Gaussian Guided Texture Warping | https://openreview.net/forum?id=HtbqsbNw9c | [
"Tianhao Walter Wu",
"Jing Yang",
"Zhilin Guo",
"Jingyi Wan",
"Fangcheng Zhong",
"Cengiz Oztireli"
] | Poster | The ability to reconstruct realistic and controllable upper body avatars from casual monocular videos is critical for various applications in communication and entertainment. By equipping the most recent 3D Gaussian Splatting representation with head 3D morphable models (3DMM), existing methods manage to create head avatars with high fidelity. However, most existing methods only reconstruct a head without the body, substantially limiting their application scenarios. We found that naively applying Gaussians to model the clothed chest and shoulders tends to result in blurry reconstruction and noisy floaters under novel poses. This is because of the fundamental limitation of Gaussians and point clouds -- each Gaussian or point can only have a single directional radiance without spatial variance, therefore an unnecessarily large number of them is required to represent complicated spatially varying texture, even for simple geometry. In contrast, we propose to model the body part with a neural texture that consists of coarse and pose-dependent fine colors. To properly render the body texture for each view and pose without accurate geometry nor UV mapping, we optimize another sparse set of Gaussians as anchors that constrain the neural warping field that maps image plane coordinates to the texture space. We demonstrate that Gaussian Head & Shoulders can fit the high-frequency details on the clothed upper body with high fidelity and potentially improve the accuracy and fidelity of the head region. We evaluate our method with casual phone-captured and internet videos and show our method archives superior reconstruction quality and robustness in both self and cross reenactment tasks. To fully utilize the efficient rendering speed of Gaussian splatting, we additionally propose an accelerated inference method of our trained model without Multi-Layer Perceptron (MLP) queries and reach a stable rendering speed of around 130 FPS for any subjects. | Neural Radiance Field, Gaussian Splatting, Neural Head Avatar | null | 1,545 | 2405.12069 | [
-0.0446607731282711,
-0.048740047961473465,
0.09321274608373642,
-0.05031999945640564,
0.038426801562309265,
-0.044112760573625565,
0.06335963308811188,
-0.08245086669921875,
-0.05018755421042442,
-0.06181369721889496,
0.023101815953850746,
-0.05870697274804115,
0.05094799026846886,
0.0740736871957779,
0.018383188173174858,
-0.06315187364816666,
0.12140011787414551,
0.12887808680534363,
-0.039450231939554214,
0.10606406629085541,
0.011905625462532043,
-0.01795515976846218,
-0.02575848437845707,
-0.003793632611632347,
0.00841055903583765,
-0.0674162432551384,
0.0558442622423172,
-0.0983448177576065,
-0.0016416042344644666,
-0.030472762882709503,
0.021660424768924713,
-0.08905386924743652,
-0.06203074753284454,
-0.01656273566186428,
-0.10394436120986938,
0.0595511756837368,
0.053247395902872086,
0.02651289664208889,
-0.04381798952817917,
-0.0007174965576268733,
0.01360396295785904,
0.04793845862150192,
-0.06920686364173889,
0.07336745411157608,
0.17588409781455994,
0.11353526264429092,
0.07316412776708603,
-0.034803181886672974,
0.0019071146380156279,
0.004839797504246235,
-0.05307690054178238,
-0.07430041581392288,
0.04701203107833862,
0.04169056937098503,
0.005979208741337061,
0.04256238043308258,
-0.037789855152368546,
-0.10535838454961777,
-0.010389791801571846,
0.043432895094156265,
0.01530405879020691,
0.02341299131512642,
0.036212384700775146,
0.015533801168203354,
0.032691530883312225,
-0.05343719571828842,
0.04202704131603241,
-0.012712902389466763,
-0.06552394479513168,
0.05960264801979065,
-0.07008830457925797,
-0.01954890601336956,
-0.0748172476887703,
-0.028629766777157784,
-0.011884728446602821,
0.043485675007104874,
0.030121885240077972,
-0.03409599885344505,
0.0033020502887666225,
0.022074973210692406,
0.002017190447077155,
-0.022331569343805313,
0.027074050158262253,
0.0003898911236319691,
-0.06175689771771431,
0.05574281886219978,
-0.03882782161235809,
0.10796316713094711,
-0.11144878715276718,
0.023395629599690437,
-0.05823853984475136,
0.047031696885824203,
0.0201355442404747,
-0.10621704161167145,
0.012853281572461128,
0.015687942504882812,
-0.03330593183636665,
-0.047072991728782654,
0.04597163572907448,
0.047779470682144165,
0.024632683023810387,
-0.11211274564266205,
0.10624191164970398,
0.029400484636425972,
0.03865458816289902,
-0.027014797553420067,
0.05347483232617378,
0.02208505943417549,
0.007216492667794228,
0.015375334769487381,
-0.0027736055199056864,
-0.041948821395635605,
-0.03568459674715996,
0.01985113136470318,
0.015595076605677605,
0.04149916395545006,
-0.006328053306788206,
0.025842171162366867,
-0.03853895515203476,
-0.03965848311781883,
0.041554972529411316,
0.009768882766366005,
0.021981775760650635,
-0.026502268388867378,
-0.012679855339229107,
-0.05876784399151802,
-0.08064156025648117,
4.7768666790120994e-33,
0.03536287322640419,
0.06413955241441727,
0.07743582874536514,
0.0586935430765152,
-0.00668490631505847,
0.0531914122402668,
-0.01748831942677498,
-0.026170240715146065,
-0.0009717422071844339,
0.03517008200287819,
-0.03265807777643204,
-0.017129145562648773,
-0.050705283880233765,
0.12588593363761902,
-0.01807478256523609,
0.029963376000523567,
-0.055922918021678925,
0.027561131864786148,
-0.014567560516297817,
0.07149410247802734,
0.02537107840180397,
-0.014416749589145184,
-0.009351694956421852,
-0.023365767672657967,
-0.08821890503168106,
0.005353418178856373,
-0.012897883541882038,
-0.036211516708135605,
-0.019406568259000778,
-0.01750282011926174,
-0.03482162952423096,
0.019660137593746185,
0.0003926908830180764,
-0.0065195607021451,
0.017149142920970917,
-0.046578265726566315,
0.05767042562365532,
-0.058083027601242065,
-0.032991550862789154,
0.09102042019367218,
0.05572523549199104,
0.058768775314092636,
0.01223005075007677,
0.011725841090083122,
-0.07689392566680908,
0.06092201918363571,
0.14055205881595612,
0.0649048238992691,
-0.045748163014650345,
0.03633953630924225,
0.05502717196941376,
0.010795566253364086,
-0.027042265981435776,
-0.07483906298875809,
-0.04902999848127365,
0.023916922509670258,
-0.029196415096521378,
-0.03085368499159813,
0.01076118927448988,
-0.02772018499672413,
0.000028652804758166894,
0.0015691914595663548,
0.008793824352324009,
-0.05068979412317276,
-0.050678551197052,
-0.031931567937135696,
0.043037738651037216,
0.004600983578711748,
0.008726410567760468,
0.02184334583580494,
-0.021067338064312935,
0.04341627284884453,
0.028122715651988983,
-0.009186528623104095,
0.04170556366443634,
0.01804220676422119,
0.0685294046998024,
-0.04014553874731064,
-0.02537369355559349,
0.011803199537098408,
-0.017106937244534492,
0.0817553699016571,
-0.015922317281365395,
-0.027456358075141907,
-0.06579259783029556,
-0.060094188898801804,
-0.0461425818502903,
0.022011183202266693,
-0.05388454720377922,
-0.05205962061882019,
-0.029962489381432533,
-0.030406635254621506,
0.0018540330929681659,
-0.08968353271484375,
-0.05972912162542343,
-2.5433583521566424e-33,
0.06496194750070572,
-0.01354307122528553,
-0.028568299487233162,
0.03377218917012215,
0.05102001875638962,
-0.09003531187772751,
0.06042157858610153,
0.04635749012231827,
0.015910319983959198,
-0.025217056274414062,
-0.00036982871824875474,
-0.024522921070456505,
0.009297464042901993,
0.04271814599633217,
0.05997452884912491,
0.01592389866709709,
-0.01243782788515091,
-0.030455484986305237,
-0.05139504745602608,
-0.06669075042009354,
0.013383572921156883,
0.05895065516233444,
0.03204512968659401,
-0.025854866951704025,
-0.08956354111433029,
0.14213700592517853,
-0.052393075078725815,
0.06989754736423492,
0.0107507249340415,
0.012837197631597519,
-0.035288140177726746,
-0.018651602789759636,
0.010251917876303196,
-0.03577415645122528,
-0.03626742586493492,
0.05046272277832031,
0.04813173785805702,
0.06451171636581421,
-0.03176598623394966,
-0.03619975969195366,
0.01896573044359684,
-0.08398017287254333,
0.029928721487522125,
0.017080962657928467,
0.00926682073622942,
-0.01865832321345806,
-0.05390286073088646,
-0.05082333832979202,
-0.02604524791240692,
0.10526943951845169,
-0.09640360623598099,
0.04018332436680794,
-0.06254469603300095,
-0.007406964432448149,
-0.02988203801214695,
-0.07365141063928604,
-0.12111897766590118,
0.004686689004302025,
0.09760091453790665,
0.006560288369655609,
-0.05224546045064926,
-0.031422436237335205,
-0.1089668944478035,
-0.003067929996177554,
-0.03608257696032524,
0.09349822252988815,
0.02472119778394699,
-0.033435970544815063,
-0.028472574427723885,
0.05555642396211624,
-0.013124463148415089,
-0.06309543550014496,
0.051005199551582336,
0.08802422881126404,
0.023260124027729034,
-0.023459943011403084,
0.020252879709005356,
0.018668444827198982,
0.03241611272096634,
0.030850350856781006,
0.03798086941242218,
-0.06144983321428299,
-0.03486542031168938,
0.06750132143497467,
0.08676853775978088,
0.11284632980823517,
-0.05085412412881851,
-0.034138549119234085,
-0.032485492527484894,
-0.028276002034544945,
-0.027552597224712372,
0.048188816756010056,
-0.011291757225990295,
0.09062913805246353,
0.01320687122642994,
-4.253997332170911e-8,
-0.0403585322201252,
0.03285035490989685,
0.05810502544045448,
-0.05397895723581314,
-0.06524422019720078,
0.020387837663292885,
0.005359672009944916,
-0.03565211594104767,
0.016944631934165955,
-0.08456984162330627,
0.016017219051718712,
-0.08903445303440094,
0.03204350546002388,
0.007884575054049492,
0.04258868843317032,
0.06575644761323929,
-0.0822172462940216,
0.04924947768449783,
-0.04861418157815933,
-0.02934339828789234,
-0.07235262542963028,
0.0318119153380394,
-0.024528957903385162,
0.02051595225930214,
0.03917498141527176,
0.009947669692337513,
-0.06386692821979523,
0.1073160171508789,
0.015138550661504269,
0.05113891884684563,
0.008115116506814957,
0.027867473661899567,
0.014229814521968365,
-0.02385920099914074,
-0.014959150925278664,
-0.043243180960416794,
-0.022410301491618156,
-0.019418349489569664,
-0.02411571703851223,
0.023598404601216316,
-0.01461485680192709,
0.04708779230713844,
0.06613125652074814,
0.015912985429167747,
0.09269794821739197,
0.027780717238783836,
0.076557956635952,
-0.08335884660482407,
0.07611972093582153,
0.023724092170596123,
-0.010095280595123768,
-0.014365305192768574,
-0.021671948954463005,
0.13929642736911774,
0.03060167469084263,
-0.011825473047792912,
0.03827100247144699,
-0.008214076980948448,
0.02400698885321617,
0.05636784806847572,
-0.009740200825035572,
0.00042950600618496537,
-0.04478735104203224,
0.045699700713157654
] |
Bootstrapping Language-Guided Navigation Learning with Self-Refining Data Flywheel | https://openreview.net/forum?id=OUuhwVsk9Z | [
"Zun Wang",
"Jialu Li",
"Yicong Hong",
"Songze Li",
"Kunchang Li",
"Shoubin Yu",
"Yi Wang",
"Yu Qiao",
"Yali Wang",
"Mohit Bansal",
"Limin Wang"
] | Poster | Creating high-quality data for training robust language-instructed agents is a long-lasting challenge in embodied AI. In this paper, we introduce a Self-Refining Data Flywheel (SRDF) that generates high-quality and large-scale navigational instruction-trajectory pairs by iteratively refining the data pool through the collaboration between two models, the instruction generator and the navigator, without any human-in-the-loop annotation.
Specifically, SRDF starts with using a base generator to create an initial data pool for training a base navigator, followed by applying the trained navigator to filter the data pool. This leads to higher-fidelity data to train a better generator, which can, in turn, produce higher-quality data for training the next-round navigator. Such a flywheel establishes a data self-refining process, yielding a continuously improved and highly effective dataset for large-scale language-guided navigation learning. Our experiments demonstrate that after several flywheel rounds, the navigator elevates the performance boundary from 70\% to 78\% SPL on the classic R2R test set, surpassing human performance (76\%) for the first time.
Meanwhile, this process results in a superior generator, evidenced by a SPICE increase from 23.5 to 26.2, better than all previous VLN instruction generation methods. Finally, we demonstrate the scalability of our method through increasing environment and instruction diversity, and
the generalization ability of our pre-trained navigator across various downstream navigation tasks, surpassing state-of-the-art methods by a large margin in all cases. | vision-and-language navigation, data flywheel, dataset curation | null | 1,544 | 2412.08467 | [
0.026557963341474533,
-0.06508147716522217,
0.044317279011011124,
0.019629042595624924,
-0.005371810868382454,
-0.039686329662799835,
0.016850603744387627,
-0.04786668345332146,
-0.023503392934799194,
-0.07504818588495255,
-0.030362024903297424,
0.0036475087981671095,
0.020559027791023254,
0.01721922867000103,
0.0011912524932995439,
-0.006174629088491201,
0.08127527683973312,
0.08667530864477158,
-0.07387889921665192,
-0.048298805952072144,
0.04586312919855118,
0.06506767868995667,
0.010927936062216759,
0.019049298018217087,
-0.008014134131371975,
0.044875238090753555,
-0.03166874498128891,
-0.10412478446960449,
0.06158570572733879,
-0.07512955367565155,
0.041933342814445496,
0.037932563573122025,
-0.08947640657424927,
-0.013454095460474491,
-0.01847890205681324,
0.02488616108894348,
-0.03478028625249863,
0.025364689528942108,
0.005657997447997332,
-0.05494381859898567,
0.006376918870955706,
0.012544565834105015,
0.060995958745479584,
0.04198497533798218,
0.05068788304924965,
0.025281088426709175,
0.003109174082055688,
-0.06501100212335587,
0.05473727732896805,
0.07918663322925568,
-0.13937775790691376,
-0.04319611191749573,
0.09191974252462387,
-0.04133787006139755,
-0.021692883223295212,
0.06975365430116653,
0.05066704750061035,
0.06630441546440125,
-0.013664268888533115,
0.013538618572056293,
-0.01648668572306633,
-0.06487450748682022,
0.011377695947885513,
-0.005800630431622267,
-0.03983204439282417,
-0.09152358025312424,
0.03842351213097572,
0.03168169781565666,
0.036987658590078354,
0.011641527526080608,
-0.00424891896545887,
0.01499464362859726,
0.04813838377594948,
0.04229319468140602,
0.0431813970208168,
0.045188192278146744,
-0.026777930557727814,
-0.04576429724693298,
0.06531528383493423,
-0.10159525275230408,
-0.04040570557117462,
-0.04424256831407547,
0.03252497315406799,
0.02110379748046398,
0.0898440033197403,
0.011939828284084797,
0.09461797028779984,
0.016954518854618073,
0.0003182307118549943,
0.04734534025192261,
-0.03903549537062645,
-0.0336323156952858,
0.0440777949988842,
-0.04203631356358528,
-0.031833186745643616,
0.06436067819595337,
-0.02563132904469967,
-0.09753915667533875,
-0.03842676803469658,
0.06347718089818954,
-0.013479596003890038,
0.08147254586219788,
0.011810410767793655,
0.022729728370904922,
-0.06715501099824905,
-0.04177510738372803,
0.09073318541049957,
0.03442375734448433,
-0.010377720929682255,
-0.02729148231446743,
0.03611395135521889,
0.022649094462394714,
0.007377581670880318,
-0.007817935198545456,
0.008937680162489414,
-0.006485086865723133,
-0.08671655505895615,
0.017047807574272156,
0.02800147980451584,
0.07563649117946625,
-0.02881803922355175,
-0.003848499385640025,
0.03369082137942314,
0.054156169295310974,
0.06982864439487457,
-0.07931333780288696,
-0.04411567375063896,
-1.0561705621187902e-33,
0.041991326957941055,
0.04997550696134567,
0.039063550531864166,
0.018658341839909554,
0.04455919563770294,
-0.035157933831214905,
0.015572216361761093,
0.026199521496891975,
-0.02271288074553013,
0.0638473629951477,
0.01934940740466118,
-0.010114526376128197,
-0.055791761726140976,
0.07477331906557083,
0.02942045032978058,
-0.009173005819320679,
-0.05784411355853081,
0.018385842442512512,
-0.11533752828836441,
-0.012012551538646221,
0.12905019521713257,
-0.0033974372781813145,
0.03086545132100582,
-0.11294322460889816,
0.06825271248817444,
0.03381836414337158,
0.008552786894142628,
-0.0060026319697499275,
-0.01844293251633644,
0.026250598952174187,
-0.06140860915184021,
-0.013192878104746342,
-0.06711320579051971,
0.04393022507429123,
0.02291066385805607,
-0.06675122678279877,
0.021602928638458252,
-0.07473379373550415,
0.02764296904206276,
0.0895649790763855,
0.021737312898039818,
0.022572781890630722,
0.06498336791992188,
-0.015259064733982086,
-0.11126435548067093,
-0.0748995840549469,
0.04475734010338783,
0.05261782556772232,
-0.02678774669766426,
-0.02286442555487156,
-0.026009272783994675,
0.014723699539899826,
-0.0050587160512804985,
-0.07014912366867065,
-0.007659709546715021,
-0.0452444925904274,
0.02920568361878395,
0.03554859384894371,
0.017479874193668365,
0.0737442895770073,
-0.02459602989256382,
0.02894734963774681,
-0.042449627071619034,
0.13164761662483215,
0.07724644988775253,
0.04506933316588402,
0.03147067874670029,
0.0695924162864685,
0.08202262967824936,
-0.08313612639904022,
-0.054020702838897705,
-0.050791677087545395,
0.01895860768854618,
0.002119804732501507,
0.04991390183568001,
-0.05724557861685753,
0.05075909197330475,
-0.11076700687408447,
0.04853341355919838,
-0.042516835033893585,
0.01600487530231476,
-0.03792116418480873,
-0.041543588042259216,
0.0021417909301817417,
0.00020376825705170631,
-0.06808240711688995,
-0.0219382606446743,
-0.07782584428787231,
0.006815882865339518,
0.024755289778113365,
-0.09026990085840225,
-0.0769132673740387,
-0.04612341895699501,
-0.04270341992378235,
-0.02428281679749489,
1.1750666729166797e-33,
0.011039892211556435,
0.017144974321126938,
-0.008178560994565487,
0.1060008704662323,
-0.05168404430150986,
-0.03641744703054428,
0.08413243293762207,
-0.010124980472028255,
-0.017785480245947838,
-0.06920678913593292,
-0.09574270248413086,
-0.0019800965674221516,
0.04596227779984474,
0.03036988340318203,
0.021485593169927597,
-0.029690925031900406,
0.037532247602939606,
0.01914389804005623,
0.04281086102128029,
-0.025397883728146553,
0.02219821885228157,
0.08401969075202942,
-0.058039791882038116,
0.03213931992650032,
-0.016218189150094986,
0.0033578616566956043,
0.003552837297320366,
0.0833091288805008,
0.04369691014289856,
0.019793100655078888,
0.049616724252700806,
0.01649067923426628,
0.07083319127559662,
-0.03600630164146423,
-0.0645340159535408,
0.040511440485715866,
0.06481605768203735,
-0.014043235220015049,
-0.04436863213777542,
0.06882815808057785,
0.07605750858783722,
-0.06321031600236893,
-0.0700560137629509,
-0.04347378760576248,
-0.0020358948968350887,
-0.01991538517177105,
-0.13248632848262787,
0.036961253732442856,
-0.02266012690961361,
0.026084046810865402,
-0.0439164824783802,
0.021153230220079422,
-0.07262861728668213,
-0.08179149031639099,
-0.08028066903352737,
-0.1088523268699646,
0.027833517640829086,
-0.01754588447511196,
0.07768725603818893,
-0.03433319181203842,
-0.047066472470760345,
-0.0011266780784353614,
0.019399741664528847,
-0.011665936559438705,
0.002730516018345952,
-0.05216324329376221,
-0.028824327513575554,
0.013383317738771439,
0.004122868645936251,
0.007521715015172958,
-0.007832275703549385,
-0.029725564643740654,
0.04888714849948883,
0.022908974438905716,
-0.01885480433702469,
0.027209287509322166,
0.0623922161757946,
-0.04985205456614494,
0.05823829025030136,
-0.05730152130126953,
0.018459966406226158,
-0.003680964233353734,
0.01872534677386284,
0.0538652241230011,
0.0713503509759903,
0.17376522719860077,
-0.06589823216199875,
0.07004975527524948,
-0.001242752419784665,
0.03181163966655731,
0.020117057487368584,
0.018133217468857765,
-0.019851576536893845,
0.0707540214061737,
-0.0732223242521286,
-4.2651386422676296e-8,
-0.09257660061120987,
0.07200218737125397,
0.06026732921600342,
0.09995632618665695,
0.04292028024792671,
-0.0014210558729246259,
0.01615118980407715,
0.06994164735078812,
-0.019290655851364136,
-0.040277495980262756,
0.02773183397948742,
-0.011793861165642738,
0.005128246732056141,
-0.06867610663175583,
0.08119489997625351,
0.08158312737941742,
0.0023125046864151955,
0.011008569970726967,
-0.033023614436388016,
-0.007754811085760593,
0.04062961786985397,
0.010942772962152958,
-0.021535612642765045,
-0.03979703038930893,
0.01268573384732008,
-0.0201667919754982,
-0.02293003723025322,
0.055181339383125305,
0.02974856086075306,
-0.050963692367076874,
0.030559930950403214,
-0.040392931550741196,
0.02017875947058201,
0.005331697873771191,
0.004319678992033005,
0.041361913084983826,
-0.00019094104936812073,
-0.008751604706048965,
-0.007862350903451443,
-0.03550061210989952,
-0.0007083488744683564,
0.12367208302021027,
-0.02457459643483162,
-0.008364981040358543,
-0.04156816378235817,
-0.01639997959136963,
-0.07922187447547913,
-0.13091114163398743,
0.012882188893854618,
-0.10618318617343903,
-0.0153230931609869,
-0.02291601337492466,
-0.04671884700655937,
0.06514959037303925,
0.10653097927570343,
0.04169386625289917,
-0.04871854931116104,
-0.03762541711330414,
-0.007456418126821518,
0.08985477685928345,
-0.01751180924475193,
0.02212538756430149,
-0.06896325945854187,
-0.03976535424590111
] |
Hydra-SGG: Hybrid Relation Assignment for One-stage Scene Graph Generation | https://openreview.net/forum?id=tpD1rs25Uu | [
"Minghan Chen",
"Guikun Chen",
"Wenguan Wang",
"Yi Yang"
] | Poster | DETR introduces a simplified one-stage framework for scene graph generation (SGG) but faces challenges of sparse supervision and false negative samples. The former occurs because each image typically contains fewer than 10 relation annotations, while DETR-based SGG models employ over 100 relation queries. Each ground truth relation is assigned to only one query during training. The latter arises when one ground truth relation may have multiple queries with similar matching scores, leading to suboptimally matched queries being treated as negative samples. To address these, we propose Hydra-SGG, a one-stage SGG method featuring a Hybrid Relation Assignment. This approach combines a One-to-One Relation Assignment with an IoU-based One-to-Many Relation Assignment, increasing positive training samples and mitigating sparse supervision. In addition, we empirically demonstrate that removing self-attention between relation queries leads to duplicate predictions, which actually benefits the proposed One-to-Many Relation Assignment. With this insight, we introduce Hydra Branch, an auxiliary decoder without self-attention layers, to further enhance One-to-Many Relation Assignment by promoting different queries to make the same relation prediction. Hydra-SGG achieves state-of-the-art performance on multiple datasets, including VG150 (16.0 mR@50), Open Images V6 (50.1 weighted score), and GQA (12.7 mR@50). Our code and pre-trained models will be released on Hydra-SGG. | Scene Graph Generation, Visual Relation Detection | null | 1,542 | null | [
-0.051295436918735504,
-0.05666171386837959,
0.07569268345832825,
0.057218510657548904,
0.04464418813586235,
0.010457731783390045,
-0.005628187209367752,
-0.05525633692741394,
0.04326246678829193,
-0.09352923184633255,
-0.025567948818206787,
-0.08159882575273514,
0.05533156543970108,
0.11041712760925293,
0.008462347090244293,
0.010073870420455933,
0.05662127584218979,
-0.018705477938055992,
-0.06916636973619461,
-0.06419812887907028,
0.008284264244139194,
-0.011619058437645435,
0.05265920236706734,
-0.00033966818591579795,
0.023707501590251923,
-0.044636957347393036,
0.009424715302884579,
-0.020417872816324234,
0.06793300062417984,
-0.02247537672519684,
0.10016214847564697,
0.07525550574064255,
-0.046550702303647995,
0.054114148020744324,
-0.01277995202690363,
0.08224358409643173,
-0.019194720312952995,
0.03251916542649269,
-0.014578673988580704,
-0.030113929882645607,
0.023930681869387627,
-0.002350772498175502,
0.05201182886958122,
0.022260533645749092,
0.11078937351703644,
0.050554368644952774,
-0.018237123265862465,
-0.03320101648569107,
0.05120990425348282,
0.014188040047883987,
-0.04839463159441948,
0.045795053243637085,
-0.05029856786131859,
-0.01694212295114994,
0.03718710318207741,
0.04730464145541191,
0.027095086872577667,
-0.035827893763780594,
-0.02391764335334301,
-0.019910966977477074,
-0.0019938151817768812,
-0.06196710839867592,
0.011319927871227264,
-0.024075299501419067,
0.01077644806355238,
-0.026054201647639275,
0.039335962384939194,
0.05591108649969101,
-0.015841806307435036,
0.05308094993233681,
-0.009860104881227016,
0.0018398306565359235,
-0.030636580660939217,
0.08462615311145782,
-0.019307304173707962,
0.1342436820268631,
0.012137976475059986,
-0.060008659958839417,
0.05642994865775108,
-0.04532026872038841,
-0.024460840970277786,
-0.03403182700276375,
0.018565364181995392,
0.0008358585764653981,
0.07160932570695877,
-0.038151998072862625,
0.029158417135477066,
0.001712755300104618,
-0.013215180486440659,
0.051733147352933884,
-0.07479659467935562,
-0.07183283567428589,
0.0693691298365593,
0.0050353086553514,
-0.036342885345220566,
0.08413867652416229,
-0.03328762203454971,
-0.07998084276914597,
-0.07326389849185944,
0.05432058870792389,
-0.004566041752696037,
0.08609481155872345,
0.010466186329722404,
0.0016868340317159891,
0.003149560187011957,
0.014781513251364231,
0.021097876131534576,
0.023138726130127907,
0.01604975387454033,
-0.08211155980825424,
0.017488466575741768,
0.07579691708087921,
-0.0032199767883867025,
-0.01859685778617859,
0.02185339480638504,
0.02828480303287506,
0.06177716702222824,
-0.0034789959900081158,
-0.005637827329337597,
0.03614242002367973,
-0.050224341452121735,
-0.004727750550955534,
-0.025959650054574013,
0.041241198778152466,
-0.0056894319131970406,
-0.08029820770025253,
-0.06718713045120239,
2.794417129527405e-33,
0.12491198629140854,
0.030201466754078865,
0.09766821563243866,
-0.062012042850255966,
0.01652967743575573,
-0.0075926389545202255,
-0.02400670014321804,
0.13709664344787598,
-0.0890207588672638,
0.08420555293560028,
0.011091064661741257,
-0.04011176899075508,
-0.04314480721950531,
0.054495055228471756,
-0.020046420395374298,
0.03971093147993088,
-0.04679516330361366,
-0.020412517711520195,
-0.03043125942349434,
-0.0177720095962286,
-0.08040449023246765,
0.03724261745810509,
-0.046115729957818985,
0.008068358525633812,
0.04935666173696518,
0.04477039724588394,
0.07779613137245178,
0.011498866602778435,
-0.02566605992615223,
-0.046170424669981,
-0.1482314169406891,
0.01843094825744629,
0.07581011205911636,
0.034627385437488556,
0.021812736988067627,
-0.07313832640647888,
-0.05500509962439537,
-0.0598486065864563,
0.012913359329104424,
0.018349142745137215,
-0.013960877433419228,
0.04334934428334236,
0.01692136377096176,
-0.10013937205076218,
-0.06452634930610657,
-0.0035219816491007805,
-0.010350110940635204,
0.004937989637255669,
-0.025374233722686768,
0.035895369946956635,
0.038808103650808334,
0.03718515858054161,
-0.09342027455568314,
0.004045142326503992,
-0.03219551593065262,
-0.06313782930374146,
0.025995248928666115,
0.005414470564574003,
-0.008274495601654053,
0.005411832593381405,
0.023870553821325302,
-0.044937241822481155,
-0.07213754951953888,
0.07546302676200867,
0.03476724401116371,
0.07264797389507294,
0.02298901230096817,
0.034010957926511765,
0.09318246692419052,
-0.03669280931353569,
-0.05013642460107803,
0.029378069564700127,
-0.053619399666786194,
-0.11067662388086319,
0.08248841017484665,
0.002617713762447238,
0.04625147953629494,
-0.09924137592315674,
0.005795606877654791,
0.044484663754701614,
-0.07471927255392075,
-0.03145919740200043,
-0.0019276428502053022,
0.02368014305830002,
-0.08871803432703018,
0.02237551286816597,
0.08077236264944077,
-0.02836461551487446,
0.03128398582339287,
-0.041725970804691315,
-0.017735060304403305,
-0.07368719577789307,
-0.05960419774055481,
-0.036226313561201096,
0.028377195820212364,
-1.926408407322219e-33,
0.0683528333902359,
0.041805438697338104,
-0.04360229894518852,
-0.06553831696510315,
0.030055467039346695,
-0.0860995426774025,
0.03783871978521347,
-0.029249222949147224,
-0.0033859126269817352,
0.003243149956688285,
-0.022737232968211174,
-0.06833715736865997,
0.08201300352811813,
-0.06940891593694687,
-0.00463495310395956,
0.03535551577806473,
-0.015851177275180817,
-0.04232574999332428,
-0.10695844888687134,
0.06801804900169373,
-0.02873382531106472,
0.16400638222694397,
-0.02829093486070633,
0.041757479310035706,
-0.04440317675471306,
0.044620316475629807,
0.038660477846860886,
0.04163578897714615,
-0.004188053775578737,
-0.015691710636019707,
0.006908849347382784,
0.021622639149427414,
-0.005574103444814682,
-0.10092295706272125,
-0.04242447763681412,
0.022214028984308243,
0.03898092731833458,
-0.028546003624796867,
-0.0459800660610199,
0.07735110819339752,
0.0579458512365818,
-0.013192841783165932,
-0.04312555491924286,
0.059771399945020676,
-0.0802120789885521,
-0.007927639409899712,
-0.10046102106571198,
-0.0024254745803773403,
-0.018164895474910736,
0.06838015466928482,
-0.08462276309728622,
0.03046872839331627,
-0.08176719397306442,
0.014715894125401974,
-0.024872785434126854,
-0.13087601959705353,
-0.01635763980448246,
0.07078748196363449,
0.07424832135438919,
-0.004532976541668177,
0.005508812610059977,
0.009980732575058937,
-0.03378130495548248,
-0.0930693969130516,
-0.015732653439044952,
-0.04139547795057297,
-0.03073188103735447,
-0.011651664040982723,
0.016111649572849274,
0.051362790167331696,
0.011406992562115192,
0.0056219277903437614,
0.03048551268875599,
0.026083679869771004,
0.05166247487068176,
0.01350057777017355,
-0.017741205170750618,
-0.057300642132759094,
0.05055776983499527,
-0.08657816797494888,
-0.027248762547969818,
0.027925794944167137,
0.01931135728955269,
0.10260695964097977,
0.0330331027507782,
0.021847639232873917,
0.03086506389081478,
0.08023148030042648,
0.05216873064637184,
0.023776350542902946,
-0.07913349568843842,
-0.03802021965384483,
-0.03964479640126228,
0.06348351389169693,
-0.04656358063220978,
-4.285512389401447e-8,
-0.11371951550245285,
0.024781513959169388,
-0.019284706562757492,
0.009685383178293705,
-0.04621642455458641,
-0.02901303395628929,
0.016571274027228355,
0.09959250688552856,
-0.048782702535390854,
0.03054102696478367,
0.07455925643444061,
0.0017576554091647267,
-0.08320972323417664,
-0.04387291148304939,
-0.003050424624234438,
0.01835346780717373,
0.033485446125268936,
-0.031183328479528427,
-0.002992575289681554,
-0.01666703261435032,
-0.0032561845146119595,
-0.0007044204394333065,
-0.0184207521378994,
0.01783042959868908,
0.05695851519703865,
-0.043826326727867126,
0.026246540248394012,
0.08724084496498108,
0.1080707535147667,
-0.018384922295808792,
0.0347084142267704,
-0.07931427657604218,
0.06903018057346344,
-0.058387577533721924,
0.049345169216394424,
0.06704533100128174,
0.02422216534614563,
-0.024048393592238426,
0.0011498212115839124,
-0.011450313031673431,
0.014155981130897999,
0.04267185553908348,
-0.05004711449146271,
0.06836827844381332,
-0.015283011831343174,
0.05783715099096298,
0.02528473548591137,
-0.11974148452281952,
0.03827051445841789,
-0.05639398470520973,
-0.058081384748220444,
-0.09013921022415161,
-0.03498030826449394,
0.037302255630493164,
-0.02270331233739853,
0.05169463902711868,
0.026596587151288986,
-0.022756323218345642,
0.050097908824682236,
0.036956313997507095,
0.03495097905397415,
0.015943599864840508,
-0.030233150348067284,
0.008223135024309158
] |
SANER: Annotation-free Societal Attribute Neutralizer for Debiasing CLIP | https://openreview.net/forum?id=x5hXkSMOd1 | [
"Yusuke Hirota",
"Min-Hung Chen",
"Chien-Yi Wang",
"Yuta Nakashima",
"Yu-Chiang Frank Wang",
"Ryo Hachiuma"
] | Poster | Large-scale vision-language models, such as CLIP, are known to contain societal bias regarding protected attributes (e.g., gender, age). This paper aims to address the problems of societal bias in CLIP. Although previous studies have proposed to debias societal bias through adversarial learning or test-time projecting, our comprehensive study of these works identifies two critical limitations: 1) loss of attribute information when it is explicitly disclosed in the input and 2) use of the attribute annotations during debiasing process. To mitigate societal bias in CLIP and overcome these limitations simultaneously, we introduce a simple-yet-effective debiasing method called SANER (societal attribute neutralizer) that eliminates attribute information from CLIP text features only of attribute-neutral descriptions. Experimental results show that SANER, which does not require attribute annotations and preserves original information for attribute-specific descriptions, demonstrates superior debiasing ability than the existing methods. | Societal bias, CLIP, Debiasing, Fairness | We propose SANER, a debiasing method for CLIP that removes societal bias without requiring attribute annotations, while preserving attribute-specific information. | 1,541 | 2408.10202 | [
-0.01617857627570629,
0.046913087368011475,
0.00008219473966164514,
0.036257196217775345,
0.17278167605400085,
-0.03297955542802811,
0.012430421076714993,
-0.09292297810316086,
0.017752286046743393,
-0.043236978352069855,
-0.006579841952770948,
-0.055201731622219086,
0.0423608273267746,
0.01732875034213066,
0.010455434210598469,
0.02291657216846943,
0.06567540019750595,
0.07142484188079834,
-0.07905980199575424,
-0.024467740207910538,
0.008921568281948566,
0.011734823696315289,
0.03960222750902176,
0.027305157855153084,
0.0057638175785541534,
-0.07013355195522308,
-0.025390464812517166,
-0.02299005724489689,
-0.05487167835235596,
0.037428222596645355,
0.06157217547297478,
-0.01386368740350008,
0.07898780703544617,
0.05910646542906761,
-0.0598943866789341,
0.00676309410482645,
-0.019605198875069618,
0.0925833061337471,
-0.02657320722937584,
-0.009843559935688972,
-0.07419399172067642,
-0.05751112475991249,
-0.02485993318259716,
-0.02429562248289585,
0.04517827183008194,
0.015597927384078503,
-0.002876353682950139,
-0.03268204629421234,
-0.0438838005065918,
0.032731372863054276,
-0.06365795433521271,
-0.020280340686440468,
0.030000677332282066,
0.019702976569533348,
-0.01156346034258604,
-0.025803035125136375,
0.01863466016948223,
0.05496681109070778,
-0.0544673353433609,
0.04467760771512985,
-0.03337467089295387,
-0.056012652814388275,
-0.03321196883916855,
0.0514383539557457,
-0.016965068876743317,
-0.04295963793992996,
0.01794588379561901,
0.04864278808236122,
-0.08544538170099258,
0.004936307668685913,
0.0670819878578186,
0.04679062217473984,
0.013139638118445873,
0.045663584023714066,
0.06503713876008987,
0.09039349108934402,
-0.0055502657778561115,
-0.044068511575460434,
0.03438252955675125,
-0.11262261867523193,
0.028753837570548058,
-0.013084455393254757,
0.09975465387105942,
0.09246943145990372,
0.05905323103070259,
0.004060659557580948,
-0.04293055087327957,
0.012632207944989204,
-0.019475875422358513,
0.043678127229213715,
-0.10375934839248657,
-0.05478254333138466,
0.160732701420784,
0.042034659534692764,
0.0721864104270935,
-0.015745315700769424,
-0.0496869795024395,
-0.07744977623224258,
-0.0387570858001709,
0.07947903871536255,
-0.04551243036985397,
0.03303411602973938,
-0.022832978516817093,
-0.05561104416847229,
0.002842338988557458,
-0.019157933071255684,
0.07833798974752426,
-0.008605285547673702,
-0.020675819367170334,
0.034426189959049225,
0.043503109365701675,
0.03533947095274925,
-0.00890833605080843,
-0.0649799033999443,
0.08251354098320007,
-0.05730871483683586,
0.05225834995508194,
0.0079072006046772,
0.011335372924804688,
0.002667295513674617,
-0.017514996230602264,
-0.004140257369726896,
-0.0007103019161149859,
0.035718366503715515,
0.08680717647075653,
0.028714550659060478,
-0.051592543721199036,
9.938610963305548e-34,
0.008488595485687256,
0.04461827129125595,
-0.0038631216157227755,
0.013536907732486725,
-0.02490745298564434,
-0.010094108060002327,
-0.04977799579501152,
0.01912141777575016,
-0.04046685993671417,
-0.045031677931547165,
0.07872211188077927,
-0.010577118955552578,
-0.04127432405948639,
0.02955089695751667,
0.0432351678609848,
0.02607530914247036,
-0.045859094709157944,
0.10468389093875885,
-0.031081119552254677,
0.039597705006599426,
0.021787505596876144,
-0.003460718784481287,
-0.023497071117162704,
-0.011893883347511292,
-0.0031591150909662247,
-0.02842666395008564,
-0.0007509617717005312,
-0.032885726541280746,
-0.058705080300569534,
0.035775311291217804,
-0.09793909639120102,
-0.02783946879208088,
0.07977854460477829,
-0.025081535801291466,
-0.006577661260962486,
-0.06516778469085693,
0.005488220602273941,
0.0322004109621048,
0.04179200530052185,
0.031596165150403976,
0.03362784907221794,
0.02333747036755085,
0.03292021155357361,
-0.0038283474277704954,
-0.007854224182665348,
0.01274978555738926,
0.017046524211764336,
0.023544838652014732,
-0.08111844211816788,
0.06067891791462898,
0.0348186120390892,
0.060177892446517944,
0.0001882111100712791,
-0.10550125688314438,
-0.07579729706048965,
0.029845938086509705,
0.005964311771094799,
-0.036776378750801086,
-0.022996218875050545,
-0.10810071229934692,
-0.01232105866074562,
0.06739285588264465,
0.01632332056760788,
-0.06767301261425018,
-0.013033844530582428,
0.013804902322590351,
0.046338096261024475,
-0.006903355475515127,
-0.005591756198555231,
0.03499900549650192,
-0.023826178163290024,
0.061317551881074905,
-0.1074814423918724,
0.042798686772584915,
-0.009686271660029888,
0.04523856192827225,
0.049240127205848694,
-0.0030887050088495016,
0.09934358298778534,
-0.015595399774610996,
0.041555847972631454,
-0.03927842155098915,
0.05047467723488808,
-0.07840366661548615,
-0.061276912689208984,
-0.06123604252934456,
0.09348292648792267,
-0.05888362601399422,
0.02850002609193325,
0.01710658147931099,
-0.0012735555646941066,
-0.048432908952236176,
-0.12162744998931885,
-0.01197826862335205,
-0.012834334746003151,
-2.1738969879443434e-33,
-0.06828507035970688,
-0.03414561599493027,
-0.05185040831565857,
0.007762810681015253,
-0.01606251485645771,
-0.04951094835996628,
-0.014724760316312313,
0.004943554755300283,
0.027340251952409744,
-0.03641105070710182,
-0.01582191325724125,
-0.08346229791641235,
-0.008582182228565216,
-0.008365578949451447,
0.025039559230208397,
-0.06657097488641739,
-0.03134582191705704,
0.04210548847913742,
-0.034102100878953934,
0.010592830367386341,
0.018579525873064995,
0.09583692252635956,
-0.054679516702890396,
0.05791879817843437,
-0.04455234110355377,
-0.011950310319662094,
0.020637335255742073,
0.009346239268779755,
0.08601825684309006,
-0.019501080736517906,
-0.07172085344791412,
-0.022549506276845932,
-0.01649242267012596,
-0.04338015615940094,
-0.07195350527763367,
0.01547173596918583,
0.010547331534326077,
-0.053676705807447433,
0.0028786398470401764,
0.07347802817821503,
0.012348109856247902,
-0.022671349346637726,
-0.17724162340164185,
-0.021073060110211372,
-0.11163874715566635,
-0.0949307456612587,
-0.006701924372464418,
-0.005178677383810282,
-0.00496615469455719,
-0.001753235119394958,
-0.06092345342040062,
-0.06510207802057266,
-0.0497971810400486,
0.051249753683805466,
-0.07826519757509232,
-0.09665541350841522,
0.004942193627357483,
0.026615755632519722,
0.026097306981682777,
0.08889096975326538,
-0.08937214314937592,
0.04674902185797691,
-0.07679889351129532,
0.04077744856476784,
-0.019342396408319473,
-0.07167000323534012,
-0.016425523906946182,
0.009546909481287003,
-0.030655544251203537,
-0.042717721313238144,
0.14000239968299866,
-0.032658327370882034,
-0.05151194706559181,
0.017468560487031937,
-0.040702663362026215,
0.004001261666417122,
0.03293510153889656,
0.08007123321294785,
-0.059913717210292816,
-0.024041475728154182,
0.0794953927397728,
-0.05816337838768959,
0.07523730397224426,
0.017822833731770515,
0.08043547719717026,
0.04747258871793747,
-0.03370567783713341,
0.04968954250216484,
-0.054810985922813416,
-0.010019620880484581,
-0.048528704792261124,
0.006966283544898033,
-0.06735199689865112,
0.09035476297140121,
-0.06544129550457001,
-3.7997075708062766e-8,
-0.049258846789598465,
0.03749743476510048,
-0.011948912404477596,
0.12596793472766876,
0.020661108195781708,
-0.0016864145873114467,
-0.05807081237435341,
-0.01747862622141838,
-0.00017086842854041606,
-0.00903567485511303,
0.03712763264775276,
0.045599594712257385,
-0.030926022678613663,
-0.015566926449537277,
-0.002667854307219386,
0.05253861844539642,
0.011618468910455704,
0.07816678285598755,
0.005536896642297506,
0.04582831263542175,
0.09371370822191238,
0.01625118963420391,
-0.04242076352238655,
-0.08322449773550034,
0.02407750114798546,
-0.05496571213006973,
-0.09587136656045914,
0.03192129358649254,
0.0788475051522255,
0.019242001697421074,
-0.01613149419426918,
0.04351697117090225,
-0.05903980880975723,
0.011375373229384422,
0.07269950956106186,
0.05879887565970421,
-0.041672658175230026,
-0.013075546361505985,
0.04566686227917671,
0.057724662125110626,
0.0809086263179779,
0.039702776819467545,
-0.010263808071613312,
0.009924988262355328,
0.035536013543605804,
0.015420814044773579,
0.015347921289503574,
-0.04452546313405037,
0.013945582322776318,
-0.029634667560458183,
-0.01594688929617405,
-0.0734221339225769,
0.04405266419053078,
0.08001614362001419,
0.009699435904622078,
-0.07499728351831436,
0.034561142325401306,
0.07254872471094131,
0.037348672747612,
0.1158338412642479,
0.09481441229581833,
-0.0029595789965242147,
-0.028830626979470253,
-0.03741836920380592
] |
Dynamical Diffusion: Learning Temporal Dynamics with Diffusion Models | https://openreview.net/forum?id=c5JZEPyFUE | [
"Xingzhuo Guo",
"Yu Zhang",
"Baixu Chen",
"Haoran Xu",
"Jianmin Wang",
"Mingsheng Long"
] | Poster | Diffusion models have emerged as powerful generative frameworks by progressively adding noise to data through a forward process and then reversing this process to generate realistic samples. While these models have achieved strong performance across various tasks and modalities, their application to temporal predictive learning remains underexplored. Existing approaches treat predictive learning as a conditional generation problem, but often fail to fully exploit the temporal dynamics inherent in the data, leading to challenges in generating temporally coherent sequences. To address this, we introduce Dynamical Diffusion (DyDiff), a theoretically sound framework that incorporates temporally aware forward and reverse processes. Dynamical Diffusion explicitly models temporal transitions at each diffusion step, establishing dependencies on preceding states to better capture temporal dynamics. Through the reparameterization trick, Dynamical Diffusion achieves efficient training and inference similar to any standard diffusion model. Extensive experiments across scientific spatiotemporal forecasting, video prediction, and time series forecasting demonstrate that Dynamical Diffusion consistently improves performance in temporal predictive tasks, filling a crucial gap in existing methodologies. Code is available at this repository: https://github.com/thuml/dynamical-diffusion. | Diffusion Model, Generative Model, Prediction Learning, Dynamics | We propose Dynamical Diffusion, which incorporates temporally aware forward and reverse processes and improves performance in general temporal predictive tasks. | 1,535 | 2503.00951 | [
-0.06279676407575607,
-0.14088523387908936,
0.040257398039102554,
0.07831764221191406,
0.0568687841296196,
-0.024291805922985077,
-0.07864956557750702,
-0.10642784088850021,
0.042546484619379044,
-0.036046963185071945,
-0.07936366647481918,
-0.0709972009062767,
-0.03713241219520569,
0.08802423626184464,
-0.06470569968223572,
-0.051119811832904816,
-0.00716429203748703,
0.02469996176660061,
-0.03139004111289978,
-0.07143855094909668,
0.011927193030714989,
0.03731045499444008,
-0.010154902935028076,
0.047153353691101074,
0.09691523760557175,
0.0528903529047966,
0.05944939702749252,
-0.03979158774018288,
0.0334918238222599,
-0.016268916428089142,
0.09477826952934265,
-0.023675667122006416,
-0.04768925905227661,
-0.02876877784729004,
-0.09571060538291931,
0.013487406075000763,
-0.08010131120681763,
0.04153492674231529,
-0.016412800177931786,
-0.01109327469021082,
0.004562538582831621,
-0.009196351282298565,
-0.003519064048305154,
0.02704457938671112,
-0.004152812063694,
0.008047747425734997,
0.0051626237109303474,
-0.06175420433282852,
-0.025982504710555077,
0.06273218989372253,
-0.06516328454017639,
-0.05223686248064041,
-0.004014519974589348,
0.046744782477617264,
-0.04349596053361893,
0.08736741542816162,
0.04737790301442146,
0.02995149791240692,
0.015784798189997673,
-0.04553719237446785,
-0.07616081088781357,
-0.10993334650993347,
-0.025299979373812675,
-0.055962178856134415,
-0.05581333860754967,
-0.01186157763004303,
-0.008778522722423077,
0.10587428510189056,
0.004666825290769339,
0.014315526932477951,
-0.017076823860406876,
0.09123104065656662,
-0.07614371180534363,
-0.009829549118876457,
0.032621629536151886,
0.02092226967215538,
0.07366059720516205,
-0.06080467998981476,
0.03859331086277962,
-0.0598207451403141,
0.06409996747970581,
0.021949803456664085,
0.024611536413431168,
-0.02639932744204998,
0.07224547117948532,
0.018787134438753128,
0.06285040825605392,
0.05635543167591095,
-0.013794681057333946,
-0.04993734136223793,
-0.02854265458881855,
0.009406941011548042,
-0.028604578226804733,
-0.03976045548915863,
0.04491221904754639,
0.05681031569838524,
0.02925146371126175,
-0.0722832903265953,
0.049818262457847595,
0.05568096414208412,
-0.01573893241584301,
0.027282599359750748,
-0.02543419413268566,
0.03084884025156498,
0.052518006414175034,
-0.0869731456041336,
0.03168989345431328,
0.012711262330412865,
-0.014001612551510334,
-0.013729754835367203,
0.05256692320108414,
0.05991620197892189,
0.07921923696994781,
-0.011875510215759277,
0.01784420758485794,
-0.036186475306749344,
-0.008097744546830654,
0.0025017878506332636,
-0.01925434172153473,
0.046374257653951645,
0.012364729307591915,
-0.023546654731035233,
-0.05343613028526306,
0.037700895220041275,
0.06454988569021225,
-0.09785687923431396,
0.015637818723917007,
1.6766761018579625e-33,
0.02325078472495079,
-0.1080232560634613,
0.032016247510910034,
0.04752560704946518,
0.017697621136903763,
-0.036870744079351425,
-0.04867364466190338,
-0.030592992901802063,
0.04434254392981529,
-0.07261357456445694,
-0.021090708673000336,
0.021829145029187202,
-0.0873001292347908,
0.12019641697406769,
0.011950665153563023,
-0.047838591039180756,
0.0063778553158044815,
0.0752064660191536,
0.014796537347137928,
-0.00118824967648834,
0.008763828314840794,
0.008073131553828716,
-0.04915311187505722,
-0.05974264815449715,
0.004829530604183674,
0.039327554404735565,
0.07495427876710892,
-0.030849823728203773,
0.03793966397643089,
0.020518362522125244,
-0.11791577935218811,
-0.03350202366709709,
-0.07959946990013123,
0.015002521686255932,
0.05908223241567612,
0.04693405702710152,
-0.06686078011989594,
0.009319808334112167,
0.022015640512108803,
-0.02314198762178421,
0.0232147928327322,
0.02711416780948639,
-0.03860940411686897,
-0.01838599145412445,
-0.09547217190265656,
-0.04078265279531479,
0.02009863592684269,
0.007998625747859478,
-0.06306516379117966,
-0.00849438738077879,
0.04202713072299957,
-0.05885518714785576,
-0.013155018910765648,
-0.11852632462978363,
-0.0833253413438797,
0.021686336025595665,
0.06532863527536392,
-0.06120993196964264,
0.022203663364052773,
-0.0105133056640625,
0.008901143446564674,
0.06557289510965347,
-0.019458336755633354,
-0.001567617990076542,
-0.014746484346687794,
0.02492626942694187,
-0.005991937592625618,
0.10999485850334167,
0.03980318456888199,
-0.014710860326886177,
0.043738000094890594,
0.028154511004686356,
0.05019424483180046,
-0.11255130171775818,
0.09557950496673584,
0.04543602094054222,
0.08560914546251297,
-0.03097464330494404,
0.002977729542180896,
0.022289928048849106,
-0.058712512254714966,
-0.03982822597026825,
-0.013063643127679825,
-0.02762858010828495,
0.04227723181247711,
-0.0162420142441988,
0.05930893123149872,
-0.06744369864463806,
-0.09163744002580643,
-0.0635906383395195,
-0.0992501750588417,
-0.03768881410360336,
0.03794422373175621,
0.06469853967428207,
0.0011178271379321814,
-3.640604452148506e-33,
0.018671199679374695,
0.06703920662403107,
0.014313423074781895,
0.04161205515265465,
0.0141206756234169,
-0.003277347655966878,
-0.03933989629149437,
0.0672123059630394,
0.06907269358634949,
-0.09424415230751038,
-0.05080400034785271,
-0.04921486973762512,
-0.003243205603212118,
-0.01115344651043415,
0.002676059026271105,
-0.021910928189754486,
0.05544715374708176,
0.009600338526070118,
-0.05048263072967529,
0.008280714973807335,
-0.06739155948162079,
0.031875867396593094,
-0.05539665371179581,
0.008787058293819427,
-0.04632316529750824,
-0.007696524728089571,
0.03673146292567253,
0.11874665319919586,
-0.06203526258468628,
-0.00866646133363247,
-0.06410083174705505,
-0.024852264672517776,
0.05770853906869888,
-0.029910823330283165,
-0.12430384755134583,
0.09707539528608322,
0.022373225539922714,
-0.03827715665102005,
-0.04055410623550415,
0.04100268334150314,
0.04559710994362831,
-0.01460322830826044,
-0.0693877562880516,
0.03669412434101105,
-0.0433356910943985,
-0.07453687489032745,
-0.0321054607629776,
0.09911559522151947,
0.03530808165669441,
0.036522164940834045,
0.07394400238990784,
0.11955629289150238,
-0.11134739965200424,
-0.021171212196350098,
-0.017580805346369743,
-0.01780901476740837,
-0.007130540907382965,
-0.04068813472986221,
0.04031284525990486,
0.03633245453238487,
-0.06190900132060051,
-0.028002599254250526,
0.049036651849746704,
-0.041583333164453506,
-0.03864740952849388,
-0.045703038573265076,
-0.0732511430978775,
-0.05397488549351692,
0.0853385180234909,
0.017813345417380333,
0.022741181775927544,
0.059531070291996,
-0.02605677954852581,
-0.009341070428490639,
0.06089453771710396,
-0.07680250704288483,
0.01078860368579626,
-0.03659236058592796,
-0.012305366806685925,
0.0014050041791051626,
-0.01924903318285942,
0.01823676936328411,
0.0031944492366164923,
-0.017023596912622452,
0.014519089832901955,
0.017626281827688217,
0.03516409173607826,
0.023071233183145523,
0.04089543968439102,
0.033113494515419006,
-0.022381030023097992,
-0.043857935816049576,
0.007674104999750853,
0.06556284427642822,
-0.011650578118860722,
-4.0255478950257384e-8,
-0.06564465910196304,
0.08710943907499313,
0.08345115184783936,
0.046441804617643356,
0.054200299084186554,
-0.00964384712278843,
0.11412405967712402,
0.045904118567705154,
0.04170224815607071,
-0.06492995470762253,
0.13814979791641235,
-0.03575287014245987,
-0.021901588886976242,
-0.02558841183781624,
0.00427942955866456,
0.10410942137241364,
0.045325879007577896,
-0.022745594382286072,
-0.035046253353357315,
0.013242360204458237,
0.015552508644759655,
0.006960784550756216,
-0.014630748890340328,
-0.032803669571876526,
0.07972465455532074,
-0.005609703715890646,
-0.05842221528291702,
0.02290947735309601,
0.016245635226368904,
0.011479462496936321,
0.020627152174711227,
0.012048672884702682,
-0.02072528377175331,
-0.04667489230632782,
-0.06880529969930649,
0.05716583877801895,
0.04115801304578781,
-0.030070306733250618,
-0.00840483233332634,
-0.030519526451826096,
-0.027688968926668167,
0.07025831937789917,
-0.04980086535215378,
0.0318051241338253,
-0.08011049777269363,
-0.0022572705056518316,
0.04197488725185394,
-0.06632943451404572,
0.03625619783997536,
0.02859727293252945,
0.017887162044644356,
-0.03711460158228874,
0.04944480210542679,
-0.00015336842625401914,
0.1560613363981247,
0.02687493897974491,
0.004853453487157822,
-0.019872432574629784,
0.017483854666352272,
0.05322703719139099,
0.026515213772654533,
0.020361199975013733,
-0.0052648852579295635,
0.041520893573760986
] |
A General Framework for Off-Policy Learning with Partially-Observed Reward | https://openreview.net/forum?id=mUbYof5MKp | [
"Rikiya Takehi",
"Masahiro Asami",
"Kosuke Kawakami",
"Yuta Saito"
] | Poster | Off-policy learning (OPL) in contextual bandits aims to learn a decision-making policy that maximizes the target rewards by using only historical interaction data collected under previously developed policies. Unfortunately, when rewards are only partially observed, the effectiveness of OPL degrades severely. Well-known examples of such partial rewards include explicit ratings in content recommendations, conversion signals on e-commerce platforms that are partial due to delay, and the issue of censoring in medical problems. One possible solution to deal with such partial rewards is to use secondary rewards, such as dwelling time, clicks, and medical indicators, which are more densely observed. However, relying solely on such secondary rewards can also lead to poor policy learning since they may not align with the target reward. Thus, this work studies a new and general problem of OPL where the goal is to learn a policy that maximizes the expected target reward by leveraging densely observed secondary rewards as supplemental data. We then propose a new method called Hybrid Policy Optimization for Partially-Observed Reward (HyPeR), which effectively uses the secondary rewards in addition to the partially observed target reward to achieve effective OPL despite the challenging scenario. We also discuss a case where we aim to optimize not only the expected target reward but also the expected secondary rewards to some extent; counter-intuitively, we will show that leveraging the two objectives is in fact advantageous also for the optimization of only the target reward. Along with statistical analysis of our proposed methods, empirical evaluations on both synthetic and real-world data show that HyPeR outperforms existing methods in various scenarios. | off-policy learning, partially-observed rewards, contextual bandits | We introduce HyPeR, a method that enhances off-policy learning in contextual bandits with partially-observed rewards by leveraging secondary rewards to optimize policies effectively. | 1,531 | null | [
-0.09173131734132767,
0.015664106234908104,
-0.014499080367386341,
0.015674464404582977,
0.11407563835382462,
0.035471878945827484,
0.09949172288179398,
0.00958207342773676,
0.010939544066786766,
0.002355072181671858,
-0.016839494928717613,
-0.017120856791734695,
0.03949781507253647,
0.03562719747424126,
0.010571763850748539,
0.038330622017383575,
0.12968304753303528,
0.03657977283000946,
-0.06870206445455551,
-0.034452006220817566,
-0.0028874557465314865,
0.02440645918250084,
0.08414701372385025,
0.000929598172660917,
-0.06716998666524887,
-0.025332417339086533,
-0.026254933327436447,
-0.012265649624168873,
0.014600666239857674,
0.018909147009253502,
0.0696883350610733,
-0.03160792961716652,
0.0318882055580616,
-0.1048317477107048,
-0.07354290783405304,
-0.0021408796310424805,
-0.11866416782140732,
-0.06106569990515709,
-0.0469859316945076,
-0.01574033685028553,
-0.01726742461323738,
0.04578502103686333,
-0.05818231776356697,
0.0527942031621933,
0.03199400752782822,
-0.06875475496053696,
0.010239094495773315,
-0.019470462575554848,
0.03207376226782799,
0.04252743721008301,
-0.026571311056613922,
-0.014464092440903187,
0.006119299214333296,
-0.011955922469496727,
0.0029070039745420218,
0.036613523960113525,
-0.03429434075951576,
0.023302778601646423,
-0.026641227304935455,
-0.02381259575486183,
-0.04554485157132149,
-0.1343870609998703,
-0.06640687584877014,
-0.0457184724509716,
-0.08129734545946121,
-0.039282314479351044,
-0.018169550225138664,
0.030354011803865433,
-0.051402680575847626,
0.05615435540676117,
-0.03682798147201538,
0.033399030566215515,
-0.03223254904150963,
0.014592204242944717,
0.04267224296927452,
0.0013794173719361424,
-0.022761952131986618,
-0.053163204342126846,
0.02373787946999073,
-0.029893742874264717,
-0.03528297692537308,
-0.061850547790527344,
0.003015530062839389,
-0.025531161576509476,
0.059599533677101135,
-0.06423956900835037,
0.00009091043466469273,
0.03230755776166916,
0.07738116383552551,
-0.040554728358983994,
0.006888674572110176,
-0.0078100054524838924,
0.022018110379576683,
0.019414404407143593,
0.02179723232984543,
-0.0016570331063121557,
-0.019337520003318787,
-0.06857281178236008,
-0.03879609331488609,
0.11257454007863998,
-0.006561055313795805,
0.05516459792852402,
-0.049194689840078354,
-0.042044755071401596,
-0.004645985085517168,
-0.07558535784482956,
-0.03807559236884117,
0.08301985263824463,
0.07064962387084961,
-0.04949694126844406,
-0.004052823409438133,
0.04835397005081177,
0.09295325726270676,
0.016431128606200218,
-0.0008598474087193608,
0.03489066660404205,
0.012292275205254555,
-0.035979244858026505,
0.050579674541950226,
0.09823013097047806,
0.06436183303594589,
-0.05612039938569069,
0.057635463774204254,
-0.012607776559889317,
-0.00162235286552459,
0.0049650054425001144,
0.016085634008049965,
1.032502258871472e-33,
0.06661831587553024,
-0.005321155767887831,
-0.010427284985780716,
-0.06348825246095657,
0.00341984536498785,
0.04239431023597717,
0.05709937959909439,
-0.029340576380491257,
0.0055011301301419735,
-0.0717509314417839,
0.02482721023261547,
0.0517837256193161,
-0.05194925144314766,
0.09468385577201843,
0.046260833740234375,
-0.0534062497317791,
-0.026760466396808624,
0.14039106667041779,
0.0881069228053093,
-0.051060523837804794,
0.04910069331526756,
-0.02601403370499611,
-0.006392325274646282,
-0.04432740435004234,
0.033811431378126144,
-0.01783466897904873,
0.0009550159447826445,
0.0021373569034039974,
0.02907324768602848,
0.01374143548309803,
-0.03313713148236275,
-0.007529921364039183,
-0.012370079755783081,
0.027409378439188004,
-0.00755592156201601,
0.04629438742995262,
-0.04953920468688011,
-0.006852874998003244,
-0.0128181092441082,
-0.03558171167969704,
-0.008363673463463783,
0.006599099840968847,
0.04652996361255646,
0.040969233959913254,
-0.10875482112169266,
-0.09243658185005188,
0.01394085306674242,
-0.0633607730269432,
-0.06421763449907303,
-0.03305543586611748,
-0.003036626847460866,
0.00117149343714118,
0.03441512584686279,
-0.12356528639793396,
-0.1346091628074646,
0.034837737679481506,
-0.07456079870462418,
-0.017531326040625572,
-0.017480887472629547,
-0.027171969413757324,
0.07985751330852509,
-0.028816018253564835,
0.03291676193475723,
0.034977760165929794,
-0.08196402341127396,
0.11115926504135132,
0.001584865851327777,
-0.007328005973249674,
0.03089301288127899,
-0.013870054855942726,
0.014329072088003159,
0.07348804175853729,
0.0751757100224495,
-0.11355520784854889,
0.09688504785299301,
0.008087619207799435,
0.054039258509874344,
-0.006408147979527712,
0.0016512369038537145,
-0.06641023606061935,
0.053524501621723175,
-0.05370336398482323,
-0.002548465272411704,
-0.04283451288938522,
0.030467255041003227,
0.006519863847643137,
-0.039744798094034195,
-0.07951389253139496,
-0.06817109137773514,
-0.022586477920413017,
-0.11917108297348022,
0.04819776490330696,
-0.09976249188184738,
0.05672939121723175,
0.07275932282209396,
-2.3052353391037022e-33,
-0.014348240569233894,
-0.0011216988787055016,
0.07541806995868683,
-0.04897515848278999,
-0.004135288298130035,
0.045156292617321014,
-0.049089059233665466,
-0.07684212923049927,
0.010136477649211884,
-0.04439680278301239,
-0.035774409770965576,
-0.019930919632315636,
0.03621542453765869,
-0.03277972340583801,
-0.06376239657402039,
0.012377695180475712,
0.001957035157829523,
0.09109631925821304,
-0.005400869529694319,
0.008305156603455544,
0.04497088864445686,
0.06939335912466049,
-0.14780348539352417,
0.06587670743465424,
-0.012456214055418968,
-0.001321026822552085,
0.004843731410801411,
0.042536407709121704,
0.04782811552286148,
0.01050086785107851,
-0.04061213508248329,
-0.05965127423405647,
-0.040487807244062424,
-0.02483941987156868,
-0.09871901571750641,
0.06468293070793152,
0.04219624772667885,
-0.023804601281881332,
-0.10453491657972336,
0.15475960075855255,
0.09180588275194168,
-0.04043576493859291,
-0.028285805135965347,
0.011713221669197083,
0.04072896018624306,
-0.01775333099067211,
-0.048797607421875,
0.00921565294265747,
0.015796754509210587,
0.0000647181659587659,
-0.02466895617544651,
0.03747139498591423,
-0.05026005953550339,
0.043075475841760635,
-0.09143741428852081,
-0.06311561167240143,
-0.03485254943370819,
-0.047929797321558,
0.03756532073020935,
0.011258977465331554,
-0.00720475846901536,
0.034311871975660324,
-0.050560541450977325,
0.039185501635074615,
0.01881542056798935,
-0.004215757828205824,
0.04191293194890022,
0.015032032504677773,
0.007948708720505238,
0.01926976814866066,
0.03128783777356148,
0.06679733097553253,
0.03677983582019806,
-0.08984563499689102,
0.07814164459705353,
0.04602184519171715,
0.0011382014490664005,
0.002528620418161154,
-0.044660262763500214,
-0.0276933740824461,
0.018300915136933327,
-0.01246221549808979,
0.08399941772222519,
0.006089037749916315,
0.010525153949856758,
0.07671220600605011,
0.0026545703876763582,
0.05500230938196182,
0.005047030746936798,
-0.04725054278969765,
-0.024952441453933716,
-0.02833537384867668,
0.005176837556064129,
0.04055175930261612,
-0.019664591178297997,
-4.87551261585395e-8,
-0.017754562199115753,
0.034433528780937195,
0.0286705419421196,
0.041051723062992096,
0.06514500081539154,
0.0679803118109703,
-0.06454651057720184,
0.04195055738091469,
-0.02604273520410061,
-0.053835559636354446,
0.029800638556480408,
0.03579418361186981,
-0.02299881912767887,
-0.11593047529459,
-0.06528424471616745,
0.04143428057432175,
0.012853089720010757,
-0.025034423917531967,
0.01962725818157196,
0.048168595880270004,
0.04905010759830475,
0.0010251746280118823,
0.03822937607765198,
-0.06500382721424103,
0.010747285559773445,
-0.041154127568006516,
0.021723216399550438,
0.0829147920012474,
0.11523807048797607,
0.04581884294748306,
-0.021694140508770943,
-0.0072228312492370605,
-0.005947084631770849,
-0.02512325905263424,
0.04670155420899391,
-0.005828148685395718,
0.04046374559402466,
-0.08102400600910187,
-0.06339971721172333,
0.006903782486915588,
0.041703421622514725,
0.005427305120974779,
-0.000018957471183966845,
0.012826229445636272,
0.03459900617599487,
0.033677615225315094,
-0.05817234143614769,
-0.07551556825637817,
0.08864463865756989,
-0.03829333186149597,
0.07129482179880142,
-0.0734628215432167,
0.03840915113687515,
0.05565401911735535,
0.0900178849697113,
-0.028962386772036552,
-0.003472458105534315,
-0.020259292796254158,
-0.00024797258083708584,
0.10589465498924255,
0.09727361053228378,
-0.0028057650197297335,
-0.08342809230089188,
0.07244693487882614
] |
Not-So-Optimal Transport Flows for 3D Point Cloud Generation | https://openreview.net/forum?id=62Ff8LDAJZ | [
"Ka-Hei Hui",
"Chao Liu",
"Xiaohui Zeng",
"Chi-Wing Fu",
"Arash Vahdat"
] | Poster | Learning generative models of 3D point clouds is one of the fundamental problems in 3D generative learning. One of the key properties of point clouds is their permutation invariance, i.e., changing the order of points in a point cloud does not change the shape they represent. In this paper, we analyze the recently proposed equivariant OT flows that learn permutation invariant generative models for point-based molecular data and we show that these models scale poorly on large point clouds. Also, we observe learning (equivariant) OT flows is generally challenging since straightening flow trajectories makes the learned flow model complex at the beginning of the trajectory. To remedy these, we propose not-so-optimal transport flow models that obtain an approximate OT by an offline OT precomputation, enabling an efficient construction of OT pairs for training. During training, we can additionally construct a hybrid coupling by combining our approximate OT and independent coupling to make the target flow models easier to learn. In an extensive empirical study, we show that our proposed model outperforms prior diffusion- and flow -based approaches on a wide range of unconditional generation and shape completion on the ShapeNet benchmark. | Generative models, 3D point cloud generation, flow matching, optimal transport flows | null | 1,530 | 2502.12456 | [
-0.04600317031145096,
-0.10760493576526642,
0.1252993792295456,
0.00507836788892746,
-0.009234912693500519,
0.00984710082411766,
0.003844020189717412,
-0.09969673305749893,
0.10001061111688614,
-0.01471469085663557,
0.0010703245643526316,
-0.006975814234465361,
-0.03544865548610687,
0.06732307374477386,
-0.004953798837959766,
0.031240282580256462,
0.03577783331274986,
0.08605537563562393,
-0.039698224514722824,
0.02598506212234497,
-0.024204764515161514,
0.08407047390937805,
0.02120594121515751,
0.08429832011461258,
0.06100613251328468,
0.052222415804862976,
0.09470338374376297,
-0.013924618251621723,
0.004381031263619661,
-0.05221562087535858,
0.017025120556354523,
0.006809728220105171,
-0.11972662061452866,
0.002172478474676609,
-0.0011080298572778702,
0.024936607107520103,
-0.06160940229892731,
0.025075256824493408,
-0.026108762249350548,
-0.006920620799064636,
0.07329390943050385,
0.010117022320628166,
0.040164314210414886,
0.01612948626279831,
0.09641709178686142,
0.007458061911165714,
-0.008987627923488617,
-0.11659587174654007,
0.033751774579286575,
0.011457063257694244,
0.006700693164020777,
-0.04708091914653778,
-0.083401158452034,
0.015998266637325287,
-0.010195017792284489,
0.014040245674550533,
-0.009857210330665112,
-0.04036463424563408,
0.047568388283252716,
-0.04298132285475731,
0.003360823495313525,
-0.0913940817117691,
-0.0012271825689822435,
-0.006991541478782892,
-0.05663447454571724,
-0.07689887285232544,
0.05562189966440201,
0.0656491219997406,
0.030684709548950195,
-0.02042825147509575,
-0.023305552080273628,
0.08456789702177048,
-0.03937919810414314,
-0.030627837404608727,
0.05309760197997093,
0.06790869683027267,
0.055289071053266525,
-0.03190945088863373,
-0.0553942434489727,
-0.03218524903059006,
-0.004650905728340149,
0.06952986866235733,
0.05661562457680702,
0.041809700429439545,
0.000047286612243624404,
0.04698849096894264,
-0.037760864943265915,
0.10451274365186691,
0.016914930194616318,
-0.04487293213605881,
-0.05180021747946739,
0.016493458300828934,
-0.03404867649078369,
-0.07410046458244324,
-0.0009695732733234763,
0.04397988319396973,
0.014591838233172894,
-0.05831494554877281,
0.037962496280670166,
0.04995357245206833,
0.01685057394206524,
0.061093028634786606,
0.04746786877512932,
0.017469102516770363,
0.10769226402044296,
-0.009544171392917633,
0.03355738893151283,
0.019076908007264137,
-0.015597843565046787,
-0.027161477133631706,
0.013979893177747726,
-0.05097988247871399,
0.06696034222841263,
0.00742317782714963,
-0.048716865479946136,
0.011358799412846565,
-0.0709795281291008,
-0.01742604561150074,
0.00823045801371336,
0.10964735597372055,
-0.05483296513557434,
-0.05970832332968712,
-0.014140978455543518,
0.05753793939948082,
-0.051347874104976654,
-0.0305873341858387,
-0.04502978175878525,
2.945109719666564e-34,
0.05214162915945053,
0.05414822697639465,
0.08540971577167511,
0.06447786837816238,
0.08369115740060806,
0.03649740666151047,
-0.022648582234978676,
-0.07013130933046341,
-0.010366939939558506,
-0.04778796434402466,
-0.142872154712677,
-0.07172312587499619,
-0.07705722749233246,
0.04234273359179497,
0.0033159193117171526,
-0.08173998445272446,
-0.04204178601503372,
-0.0072050816379487514,
-0.0229620523750782,
0.05348581448197365,
0.030571285635232925,
-0.0518047958612442,
-0.039266686886548996,
-0.06331221759319305,
-0.05031811445951462,
0.02536124177277088,
-0.011399761773645878,
0.05614178255200386,
-0.02652418613433838,
-0.011651961132884026,
-0.02698836848139763,
-0.04739033803343773,
-0.025339698418974876,
0.04023153334856033,
0.016661645844578743,
-0.0361371673643589,
-0.03351723775267601,
-0.04394664242863655,
0.015485876239836216,
-0.04872722551226616,
0.030148066580295563,
0.03791343420743942,
-0.06043129786849022,
-0.006255872081965208,
-0.1250673532485962,
-0.027543742209672928,
0.08150999248027802,
0.052995309233665466,
-0.008562063798308372,
0.0056777941063046455,
0.0525294728577137,
0.019463589414954185,
-0.04925282672047615,
-0.07626532018184662,
-0.051875486969947815,
-0.014845882542431355,
0.01417408138513565,
-0.011611994355916977,
0.03860585764050484,
0.049500204622745514,
-0.018918586894869804,
0.04488581418991089,
-0.01994803547859192,
-0.031100677326321602,
0.005466741509735584,
-0.008214627392590046,
-0.002232956001535058,
0.05792626738548279,
0.08305197954177856,
-0.009026856161653996,
-0.003986341413110495,
0.03259752690792084,
-0.03957947716116905,
-0.052209217101335526,
0.1052246242761612,
-0.028054939582943916,
-0.006215703673660755,
-0.08534892648458481,
0.07888004928827286,
0.025298122316598892,
-0.11435361951589584,
0.017896991223096848,
-0.016535652801394463,
-0.055405449122190475,
-0.004189129453152418,
-0.03951970860362053,
0.0009183199726976454,
-0.05848487466573715,
-0.037305478006601334,
-0.02618374489247799,
-0.02335352823138237,
-0.10081169754266739,
-0.031055781990289688,
-0.017457321286201477,
-0.01673230528831482,
-7.725440709100231e-34,
-0.015692846849560738,
0.009934539906680584,
-0.02398289367556572,
0.0653386190533638,
0.03515094146132469,
-0.030351385474205017,
-0.008425150997936726,
0.02176312357187271,
0.02312413416802883,
-0.0689079686999321,
-0.04093490540981293,
0.05270499363541603,
0.03656187281012535,
-0.03252400457859039,
0.11054480820894241,
0.013961168937385082,
-0.026638394221663475,
-0.06437960267066956,
-0.00960654579102993,
-0.0020867816638201475,
-0.029103444889187813,
-0.021385693922638893,
-0.06887004524469376,
-0.012033691629767418,
-0.0947214886546135,
0.11891208589076996,
-0.04147873446345329,
0.040067512542009354,
0.028137564659118652,
0.10277536511421204,
-0.004469504579901695,
0.00851553212851286,
-0.0024056565016508102,
0.020916877314448357,
-0.09944501519203186,
0.09414409846067429,
0.052030254155397415,
0.0890139639377594,
-0.06460012495517731,
0.05445364862680435,
0.031586818397045135,
-0.08595944941043854,
-0.07421219348907471,
0.07199230790138245,
0.02986972965300083,
0.03760455176234245,
0.012015122920274734,
-0.019618377089500427,
0.03223133459687233,
0.026905518025159836,
-0.052467525005340576,
0.060914047062397,
-0.0410793162882328,
-0.012700052931904793,
-0.011621840298175812,
-0.053563691675662994,
0.0011857611825689673,
-0.044673699885606766,
0.08801878243684769,
-0.09222105145454407,
-0.040502339601516724,
-0.07781536877155304,
0.03733472898602486,
0.039392415434122086,
-0.057730525732040405,
-0.01972350664436817,
-0.027225825935602188,
-0.03379325568675995,
-0.007536662742495537,
0.028936689719557762,
-0.036213312298059464,
-0.006810069549828768,
0.02633138746023178,
-0.0724145770072937,
0.04110719636082649,
-0.0977313444018364,
0.041398510336875916,
-0.009908623993396759,
0.04867373779416084,
-0.050111979246139526,
0.011741861701011658,
0.06863842159509659,
0.011023668572306633,
0.11812547594308853,
0.08304541558027267,
0.015053438022732735,
0.030332062393426895,
0.012566967867314816,
-0.0543350987136364,
0.050820834934711456,
0.001168487942777574,
0.06929093599319458,
0.02854926511645317,
0.04455674812197685,
-0.03823172301054001,
-3.816186833205393e-8,
-0.04928746074438095,
0.0008816052577458322,
0.08960816264152527,
-0.0030798937659710646,
0.027931708842515945,
0.05978886038064957,
0.017019540071487427,
0.15467026829719543,
0.024156801402568817,
0.004042306449264288,
0.040428489446640015,
-0.05707047879695892,
-0.02447362057864666,
-0.07394236326217651,
0.04295055568218231,
0.1280796378850937,
0.03159021586179733,
0.0032132414635270834,
-0.030325140804052353,
0.0005742417415603995,
-0.07669713348150253,
0.07868945598602295,
-0.019856208935379982,
0.016391513869166374,
0.06101275607943535,
-0.048277001827955246,
0.017521552741527557,
0.05558348819613457,
0.09392829984426498,
-0.056315045803785324,
-0.05069831386208534,
-0.03083326853811741,
0.019916051998734474,
0.03888446092605591,
-0.02870343066751957,
0.0067695751786231995,
0.03559482470154762,
-0.018426550552248955,
-0.045681729912757874,
-0.009760649874806404,
-0.0633135586977005,
0.09086043387651443,
-0.006749346852302551,
-0.04184528812766075,
0.041918449103832245,
0.11464205384254456,
0.03269284591078758,
-0.009835784323513508,
-0.043859731405973434,
0.047624457627534866,
-0.01210431195795536,
-0.00836983323097229,
-0.018301360309123993,
0.040505602955818176,
0.000778737070504576,
0.038413070142269135,
-0.06784002482891083,
-0.05231460556387901,
0.06334468722343445,
0.005118252709507942,
-0.09559608995914459,
0.0036144067998975515,
-0.014456086792051792,
-0.021853573620319366
] |
Self-Boosting Large Language Models with Synthetic Preference Data | https://openreview.net/forum?id=7visV100Ms | [
"Qingxiu Dong",
"Li Dong",
"Xingxing Zhang",
"Zhifang Sui",
"Furu Wei"
] | Poster | Through alignment with human preferences, Large Language Models (LLMs) have advanced significantly in generating honest, harmless, and helpful responses. However, collecting high-quality preference data is a resource-intensive and creativity-demanding process, especially for the continual improvement of LLMs. We introduce SynPO, a self-boosting paradigm that leverages synthetic preference data for model alignment. SynPO employs an iterative mechanism wherein a self-prompt generator creates diverse prompts, and a response improver refines model responses progressively. This approach trains LLMs to autonomously learn the generative rewards for their own outputs and eliminates the need for large-scale annotation of prompts and human preferences. After four SynPO iterations, Llama3-8B and Mistral-7B show significant enhancements in instruction-following abilities, achieving over 22.1% win rate improvements on AlpacaEval 2.0 and ArenaHard. Simultaneously, SynPO improves the general performance of LLMs on various tasks, validated by a 3.2 to 5.0 average score increase on the well-recognized Open LLM leaderboard. | preference optimization, synthetic data, LLM alignment | This paper introduces a self-boosting mechanism for LLMs, enabling them to self-synthesize preference data for iterative improvement. | 1,529 | 2410.06961 | [
-0.05472996085882187,
-0.08646436780691147,
0.022817270830273628,
-0.0226999931037426,
-0.008244454860687256,
0.02253965102136135,
0.0162751954048872,
0.036187827587127686,
0.04960440844297409,
0.00035962858237326145,
-0.0582185797393322,
-0.07210195809602737,
0.036307983100414276,
-0.017210477963089943,
0.060583144426345825,
-0.01241238322108984,
0.08910900354385376,
0.027951737865805626,
-0.03042498044669628,
-0.0516861230134964,
0.0004034063604194671,
0.003266999963670969,
0.04228435084223747,
0.047017037868499756,
0.029454845935106277,
0.036543041467666626,
0.013139803893864155,
-0.006127619184553623,
0.05029487609863281,
-0.07415222376585007,
0.01243321318179369,
-0.011766788549721241,
0.0569801889359951,
-0.007334918715059757,
-0.06939154863357544,
0.031793370842933655,
-0.10553848743438721,
0.005564375780522823,
-0.03121805004775524,
-0.06357373297214508,
-0.054517053067684174,
-0.0248199924826622,
0.03368327394127846,
0.001854328322224319,
0.06577494740486145,
-0.010244956240057945,
-0.016265930607914925,
-0.04204715043306351,
0.025369910523295403,
0.051072169095277786,
-0.08247867971658707,
-0.04968807101249695,
0.025253742933273315,
0.04046565666794777,
-0.03293759748339653,
0.05775875970721245,
0.07234625518321991,
0.06970139592885971,
0.02017483115196228,
-0.031095508486032486,
-0.07047227025032043,
-0.09192261099815369,
-0.016353121027350426,
0.015959622338414192,
-0.01640041545033455,
-0.06379479914903641,
0.009288000874221325,
0.1020151749253273,
-0.0323210284113884,
0.013384737074375153,
0.016672082245349884,
0.06617545336484909,
0.04953021928668022,
0.06995386630296707,
0.04146983474493027,
0.12156639993190765,
-0.01916603557765484,
-0.01708686351776123,
0.08143388479948044,
-0.035110145807266235,
-0.05405290424823761,
0.011854608543217182,
-0.0035088651347905397,
0.0357666052877903,
0.0953824445605278,
-0.005644007585942745,
0.038537245243787766,
0.07039111107587814,
0.06419343501329422,
0.032244857400655746,
-0.049970511347055435,
-0.036726322025060654,
0.021631594747304916,
-0.014169631525874138,
-0.019193381071090698,
0.13182923197746277,
0.012807483784854412,
-0.12726493179798126,
-0.04652916640043259,
0.06612430512905121,
0.042111415416002274,
0.12115232646465302,
-0.05214490368962288,
-0.0661642849445343,
-0.014516542665660381,
-0.025989297777414322,
0.03301246836781502,
0.05070862919092178,
0.04743041470646858,
-0.1181972324848175,
0.006250788923352957,
0.04775919392704964,
0.011174913495779037,
0.002953950548544526,
0.020355824381113052,
-0.03135085105895996,
0.034605395048856735,
-0.0401887372136116,
0.005374398548156023,
-0.019462645053863525,
-0.0022297706454992294,
-0.0375913642346859,
0.0012007879558950663,
0.034793417900800705,
0.0037531687412410975,
-0.04948107525706291,
-0.06648755073547363,
2.3009603967575477e-33,
0.08567671477794647,
0.08729109913110733,
0.03302706032991409,
0.023590227589011192,
0.06213691085577011,
-0.039863526821136475,
-0.003045917022973299,
0.03870364651083946,
-0.09262016415596008,
-0.018429985269904137,
-0.018303068354725838,
-0.02068176493048668,
-0.021474815905094147,
0.11152144521474838,
0.05626131221652031,
-0.029977859929203987,
-0.05683625489473343,
0.02653980813920498,
-0.07674671709537506,
-0.044940654188394547,
0.07345198839902878,
0.02347436733543873,
0.03332234174013138,
-0.05636946111917496,
0.07693073153495789,
0.021896518766880035,
0.07236012071371078,
-0.03417859226465225,
-0.09871584922075272,
0.0060117230750620365,
-0.04041880741715431,
-0.0698133036494255,
-0.013091561384499073,
-0.010075976140797138,
0.01534595899283886,
-0.04823894426226616,
-0.03185110166668892,
-0.04988305643200874,
0.03499568998813629,
0.03084205836057663,
0.03824824467301369,
0.0115424869582057,
0.025428518652915955,
0.05628475919365883,
-0.06275461614131927,
-0.012440665625035763,
0.008223227225244045,
0.023864584043622017,
-0.005040042567998171,
-0.05055377259850502,
0.022696534171700478,
0.04808333143591881,
-0.014030312187969685,
0.0017104157013818622,
-0.08646181225776672,
0.0030973542015999556,
0.008464295417070389,
0.07857772707939148,
0.0438651517033577,
-0.011113411746919155,
-0.0065560066141188145,
0.05128767713904381,
-0.015162999741733074,
0.038999468088150024,
0.014764495193958282,
-0.015657279640436172,
0.049214232712984085,
0.019499357789754868,
0.0639529675245285,
-0.00404904643073678,
-0.005191517062485218,
-0.010323096066713333,
-0.07563859969377518,
-0.014282507821917534,
-0.0008259128080680966,
0.001556705217808485,
0.04729034751653671,
-0.08082865923643112,
0.06063089147210121,
-0.05142732337117195,
0.013936164788901806,
-0.030729865655303,
-0.026057401672005653,
-0.058980248868465424,
-0.004615102428942919,
-0.09104982763528824,
0.0847993940114975,
-0.07577250897884369,
-0.009790144860744476,
0.014188732951879501,
0.02479429729282856,
-0.009470775723457336,
0.00616840086877346,
0.037116385996341705,
-0.04632564261555672,
-6.982420831850041e-34,
-0.010412737727165222,
-0.019504360854625702,
-0.0006632177392020822,
0.08957276493310928,
-0.06542447954416275,
-0.03545279800891876,
0.0013992423191666603,
0.053602807223796844,
-0.054793983697891235,
-0.057078346610069275,
-0.07162564992904663,
0.03267534077167511,
0.06766500324010849,
-0.01729133538901806,
0.03108939342200756,
-0.006392965093255043,
0.009424641728401184,
0.037783194333314896,
-0.015486184507608414,
-0.010233183391392231,
-0.014916897751390934,
0.1101890504360199,
-0.10491233319044113,
0.02550439164042473,
-0.04933611676096916,
0.06069101393222809,
-0.018336882814764977,
0.06077009066939354,
0.011697516776621342,
0.019545534625649452,
-0.02092740871012211,
0.04507724568247795,
-0.06743468344211578,
-0.049796465784311295,
-0.04679115489125252,
0.02291778288781643,
-0.008695408701896667,
-0.06114821881055832,
-0.010945947840809822,
0.1514832228422165,
0.053454939275979996,
-0.01947508007287979,
-0.0742734894156456,
-0.003992691170424223,
-0.039878085255622864,
-0.04001687094569206,
-0.09780727326869965,
-0.01866762526333332,
0.02238967828452587,
0.012532180175185204,
-0.0193465743213892,
0.014708247035741806,
-0.03889323025941849,
0.03594033047556877,
-0.0036985883489251137,
-0.13577492535114288,
-0.038376085460186005,
-0.09414912760257721,
0.05356789380311966,
-0.005888607352972031,
-0.06442133337259293,
0.021246332675218582,
0.11583361774682999,
-0.08526197075843811,
0.0016598952934145927,
-0.04281841963529587,
-0.01963786594569683,
0.0003671544836834073,
-0.003077892819419503,
-0.03984838351607323,
0.054814137518405914,
-0.033152054995298386,
0.002930445596575737,
-0.0018779520178213716,
0.0014024173142388463,
-0.01963479444384575,
0.007436606101691723,
-0.00847843661904335,
-0.002729143714532256,
-0.06833872199058533,
0.027210460975766182,
0.06066103279590607,
0.034497007727622986,
0.030180515721440315,
-0.0027728229761123657,
0.08354306221008301,
0.06960072368383408,
0.18311724066734314,
-0.02815435826778412,
0.047795671969652176,
0.002105436287820339,
0.04149549826979637,
0.012698610313236713,
0.11539636552333832,
-0.07349592447280884,
-4.372486017700794e-8,
-0.09332873672246933,
-0.02091812901198864,
0.03269760310649872,
0.18330450356006622,
-0.01844174414873123,
-0.04218463972210884,
-0.03904221951961517,
0.07148002833127975,
0.023931363597512245,
-0.041004907339811325,
0.03229917958378792,
0.006642220076173544,
-0.03107554465532303,
-0.06723962724208832,
0.047901641577482224,
0.08927684277296066,
-0.00034166104160249233,
0.037784043699502945,
-0.052547331899404526,
-0.08048713952302933,
0.029981819912791252,
0.06157398223876953,
0.015449250116944313,
-0.041193559765815735,
0.06045620143413544,
-0.026721026748418808,
-0.05804681032896042,
0.06990666687488556,
0.023516550660133362,
0.017184726893901825,
-0.008227257058024406,
-0.03469816967844963,
-0.03230380266904831,
-0.04349759966135025,
0.061634745448827744,
0.09177794307470322,
0.011454093270003796,
-0.02045387774705887,
0.02757825329899788,
-0.05340731889009476,
0.053505703806877136,
0.0584549643099308,
-0.056043293327093124,
0.043213699012994766,
-0.06254357099533081,
-0.04868929088115692,
-0.04115601256489754,
-0.12193453311920166,
-0.01670241169631481,
-0.09740441292524338,
0.015380755998194218,
-0.07306418567895889,
-0.008218235336244106,
0.02769809402525425,
0.026058977469801903,
0.02218768373131752,
-0.06708244234323502,
-0.010863936506211758,
0.04123615473508835,
0.04734066501259804,
0.06763793528079987,
-0.030274199321866035,
-0.02736392617225647,
0.01251307874917984
] |
Unifying Unsupervised Graph-Level Anomaly Detection and Out-of-Distribution Detection: A Benchmark | https://openreview.net/forum?id=g90RNzs8wX | [
"Yili Wang",
"Yixin Liu",
"Xu Shen",
"Chenyu Li",
"Rui Miao",
"Kaize Ding",
"Ying Wang",
"Shirui Pan",
"Xin Wang"
] | Poster | To build safe and reliable graph machine learning systems, unsupervised graph-level anomaly detection (GLAD) and unsupervised graph-level out-of-distribution (OOD) detection (GLOD) have received significant attention in recent years. Though these two lines of research share the same objective, they have been studied independently in the community due to distinct evaluation setups, creating a gap that hinders the application and evaluation of methods from one to the other. To bridge the gap, in this work, we present a Unified Benchmark for unsupervised Graph-level OOD and anomaly Detection (UB-GOLD), a comprehensive evaluation framework that unifies GLAD and GLOD under the concept of generalized graph-level OOD detection. Our benchmark encompasses 35 datasets spanning four practical anomaly and OOD detection scenarios, facilitating the comparison of 18 representative GLAD/GLOD methods. We conduct multi-dimensional analyses to explore the effectiveness, generalizability, robustness, and efficiency of existing methods, shedding light on their strengths and limitations. Furthermore, we provide an open-source codebase of UB-GOLD to foster reproducible research and outline potential directions for future investigations based on our insights. | Graph out-of-distribution detection; graph anomaly detection; benchmark | null | 1,522 | 2406.15523 | [
-0.0949326753616333,
-0.054313044995069504,
0.039010800421237946,
0.08728612959384918,
0.13367226719856262,
-0.0684981569647789,
-0.0008841757080517709,
-0.00017700815806165338,
-0.05627338960766792,
0.019430002197623253,
0.00565336225554347,
-0.039613883942365646,
0.02223680354654789,
0.055995866656303406,
-0.025689104571938515,
0.028408192098140717,
0.08661722391843796,
-0.01168951764702797,
-0.03236432746052742,
-0.024839261546730995,
0.018843602389097214,
0.04485105350613594,
-0.010043139569461346,
0.016973678022623062,
0.07694173604249954,
0.007295085582882166,
-0.020380396395921707,
-0.035400744527578354,
-0.016731347888708115,
-0.024949342012405396,
0.07750973105430603,
0.03208378329873085,
-0.04342372342944145,
-0.03127510845661163,
-0.002039481420069933,
-0.028175918385386467,
0.04933205619454384,
-0.01528100948780775,
0.02281954139471054,
0.014819320291280746,
0.07689359784126282,
-0.04235495254397392,
0.03752734512090683,
0.06670933961868286,
-0.023657819256186485,
-0.021119186654686928,
-0.0978618860244751,
-0.00503129418939352,
0.0007695902022533119,
-0.033475082367658615,
0.030101323500275612,
-0.07515806704759598,
-0.026155002415180206,
0.05694570019841194,
0.01125575602054596,
-0.06381222605705261,
-0.002386638894677162,
-0.08750400692224503,
0.061123497784137726,
0.005978910252451897,
-0.010920745320618153,
-0.03642281889915466,
0.023998457938432693,
-0.06498317420482635,
-0.03318347409367561,
0.0329744927585125,
0.02778032422065735,
0.023062685504555702,
-0.0024066201876848936,
0.05076511204242706,
0.06407702714204788,
0.09151536226272583,
-0.09267990291118622,
0.07989777624607086,
0.07063106447458267,
0.08179271966218948,
-0.005048088729381561,
0.03471997380256653,
0.022735515609383583,
-0.04133212938904762,
0.042551252990961075,
0.0627606213092804,
0.03418666496872902,
-0.01709754765033722,
0.0652725026011467,
0.02045552060008049,
-0.045460913330316544,
0.024519270285964012,
-0.017410799860954285,
0.057804495096206665,
-0.08500455319881439,
0.04075523093342781,
0.03345615789294243,
-0.0200862567871809,
0.0337119959294796,
-0.015807606279850006,
0.03192995861172676,
-0.027429696172475815,
0.014490779489278793,
0.059372931718826294,
-0.057826925069093704,
0.02437918819487095,
0.010160795412957668,
-0.02846520021557808,
0.047023702412843704,
0.015975765883922577,
-0.022624319419264793,
0.05592460557818413,
0.0641196221113205,
-0.05960528552532196,
0.009346961975097656,
0.04418694227933884,
0.06459829211235046,
-0.010719547979533672,
-0.016419600695371628,
0.01834745705127716,
0.01810431480407715,
0.025479763746261597,
-0.042628899216651917,
0.0029199584387242794,
-0.015817707404494286,
-0.015907013788819313,
-0.000040195904148276895,
0.029113108292222023,
0.0030504767782986164,
0.0551459938287735,
-0.062193308025598526,
2.008286731644776e-33,
0.09141906350851059,
0.05610736086964607,
-0.007767406292259693,
-0.04830542951822281,
0.039606958627700806,
-0.014534033834934235,
-0.13348543643951416,
0.013423839583992958,
0.033822499215602875,
0.0986296758055687,
-0.07139153778553009,
0.03443737328052521,
0.027662016451358795,
0.032345518469810486,
0.06827087700366974,
0.008862718008458614,
0.05701009929180145,
0.018873369321227074,
-0.02788015827536583,
-0.09432585537433624,
0.011579887941479683,
-0.07502865046262741,
-0.01193289365619421,
0.03573029115796089,
0.023517418652772903,
0.08275102823972702,
-0.014095498248934746,
-0.008814726024866104,
0.08376579731702805,
0.024094529449939728,
-0.03597987815737724,
0.07694073766469955,
0.006789532955735922,
0.02967144176363945,
-0.017655417323112488,
-0.11000554263591766,
-0.06815586239099503,
-0.054660484194755554,
-0.031247183680534363,
-0.11016833037137985,
-0.06671905517578125,
-0.03328298032283783,
-0.039966825395822525,
-0.04847097024321556,
0.006932523567229509,
-0.003874616464599967,
0.02240087278187275,
-0.03916621208190918,
-0.0051796347834169865,
-0.10123428702354431,
-0.04315190389752388,
-0.010382250882685184,
0.011487147770822048,
0.0980708971619606,
-0.047021932899951935,
0.014485117048025131,
0.03310703858733177,
0.04704710841178894,
-0.003038778668269515,
-0.0008085803128778934,
0.05948907136917114,
0.04324919730424881,
0.0067277210764586926,
0.0175937507301569,
-0.05749433860182762,
-0.0466122105717659,
-0.010334549471735954,
0.07264290004968643,
-0.05233342945575714,
0.09199647605419159,
0.02842140942811966,
0.0675247386097908,
-0.0678805485367775,
-0.00046426744665950537,
0.001900930074043572,
0.009267960675060749,
0.05405297502875328,
-0.0032705587800592184,
-0.041856549680233,
-0.0024754281621426344,
-0.03098430670797825,
-0.06595730781555176,
0.05199117586016655,
-0.021239548921585083,
-0.07208477705717087,
-0.003909469116479158,
-0.000256055238423869,
0.033028606325387955,
-0.051226768642663956,
0.04738161340355873,
-0.03155289590358734,
-0.038501545786857605,
-0.029957978054881096,
-0.0326571986079216,
0.026759710162878036,
-7.790259120643503e-34,
-0.0024113371036946774,
0.11616228520870209,
-0.016155768185853958,
-0.026738910004496574,
-0.009169131517410278,
-0.08372533321380615,
-0.0014828495914116502,
0.08999885618686676,
-0.03365042060613632,
0.007684009615331888,
-0.06959094107151031,
-0.027024900540709496,
0.021043099462985992,
-0.03232787922024727,
0.06424388289451599,
0.0478244349360466,
-0.0937902182340622,
0.05906989797949791,
-0.06514544039964676,
0.015439478680491447,
-0.0277915820479393,
0.013099709525704384,
-0.06301550567150116,
-0.05163516849279404,
0.08810770511627197,
0.0789891853928566,
-0.09722431004047394,
0.01964161917567253,
-0.035524558275938034,
0.004673444200307131,
-0.05393353849649429,
0.0768299251794815,
-0.008757050149142742,
-0.051307398825883865,
-0.011552492156624794,
0.0192704014480114,
0.05011827498674393,
-0.061470333486795425,
0.05759327858686447,
0.004114682320505381,
-0.045920226722955704,
0.04486067593097687,
-0.03369465842843056,
0.03426935896277428,
-0.0017850426957011223,
0.08388176560401917,
-0.0011582713341340423,
0.08130662888288498,
-0.04459107294678688,
-0.015225176699459553,
-0.1010449230670929,
0.06882281601428986,
-0.0001859522599261254,
0.06377366930246353,
-0.06466840207576752,
-0.024549247696995735,
0.04347311332821846,
0.033208414912223816,
-0.05748303234577179,
-0.006582704838365316,
-0.03847508132457733,
-0.01992231421172619,
-0.023769162595272064,
0.022973371669650078,
-0.0471838153898716,
-0.02295163832604885,
0.05328679457306862,
-0.009890223853290081,
-0.018245775252580643,
0.04370995983481407,
0.0009456055122427642,
-0.024309437721967697,
-0.052849724888801575,
-0.030909646302461624,
0.020166903734207153,
0.02720714919269085,
-0.004353070165961981,
-0.006477283779531717,
-0.03208881616592407,
-0.021266279742121696,
0.01138348039239645,
-0.020759426057338715,
-0.004429235588759184,
0.09468308091163635,
0.05405362322926521,
0.09939379245042801,
0.03958316892385483,
-0.013000189326703548,
0.03156772255897522,
0.05379645898938179,
-0.06060316413640976,
0.004348055459558964,
-0.1134500727057457,
0.04908367618918419,
0.06510892510414124,
-3.702652406900597e-8,
-0.1253027617931366,
-0.0009812447242438793,
-0.018992271274328232,
-0.07347274571657181,
0.10504012554883957,
0.043818265199661255,
-0.015288935974240303,
0.1660013347864151,
-0.04889863729476929,
0.07527483254671097,
0.04039672762155533,
-0.09274809807538986,
-0.05090458691120148,
-0.0911085233092308,
-0.004176585003733635,
-0.015972428023815155,
0.04435248672962189,
0.03606400638818741,
-0.017331702634692192,
0.04445711895823479,
-0.0005379138165153563,
0.01535913348197937,
0.07462958991527557,
-0.001875169575214386,
-0.049072857946157455,
-0.1465669870376587,
0.029295658692717552,
0.013575808145105839,
0.04855251684784889,
-0.04709934815764427,
-0.051638443022966385,
-0.0019012471893802285,
0.07465051114559174,
0.009955410845577717,
-0.009194759652018547,
0.18742473423480988,
0.0009454145911149681,
-0.04442862048745155,
-0.005201705265790224,
0.028826696798205376,
-0.0004973724135197699,
0.0702626034617424,
-0.05609363690018654,
0.041861068457365036,
-0.06806610524654388,
-0.010285106487572193,
-0.018161330372095108,
-0.0027844805736094713,
0.036822158843278885,
-0.0964033231139183,
-0.03134377673268318,
-0.1034979298710823,
-0.007231052033603191,
0.0583922415971756,
-0.0071600438095629215,
0.03937757387757301,
-0.050731319934129715,
-0.07182590663433075,
0.04971081018447876,
0.07235883921384811,
0.07372435927391052,
-0.0525054968893528,
-0.09937263280153275,
-0.006079569458961487
] |
Building Interactable Replicas of Complex Articulated Objects via Gaussian Splatting | https://openreview.net/forum?id=ix2yRWarPn | [
"Yu Liu",
"Baoxiong Jia",
"Ruijie Lu",
"Junfeng Ni",
"Song-Chun Zhu",
"Siyuan Huang"
] | Poster | Building interactable replicas of articulated objects is a key challenge in computer vision. Existing methods often fail to effectively integrate information across different object states, limiting the accuracy of part-mesh reconstruction and part dynamics modeling, particularly for complex multi-part articulated objects. We introduce ArtGS, a novel approach that leverages 3D Gaussians as a flexible and efficient representation to address these issues. Our method incorporates canonical Gaussians with coarse-to-fine initialization and updates for aligning articulated part information across different object states, and employs a skinning-inspired part dynamics modeling module to improve both part-mesh reconstruction and articulation learning. Extensive experiments on both synthetic and real-world datasets, including a new benchmark for complex multi-part objects, demonstrate that ArtGS achieves state-of-the-art performance in joint parameter estimation and part mesh reconstruction. Our approach significantly improves reconstruction quality and efficiency, especially for multi-part articulated objects. Additionally, we provide comprehensive analyses of our design choices, validating the effectiveness of each component to highlight potential areas for future improvement. | articulated object modeling | null | 1,521 | 2502.19459 | [
0.002041909145191312,
-0.038027651607990265,
0.05328352749347687,
0.007944577373564243,
0.09092365205287933,
-0.03196680545806885,
-0.009778664447367191,
-0.07231124490499496,
0.0032039599027484655,
-0.05990305170416832,
-0.021360736340284348,
-0.020081182941794395,
-0.027236085385084152,
0.1258377879858017,
0.004413875285536051,
-0.07851459830999374,
0.08334711194038391,
0.13633993268013,
-0.10793378949165344,
0.09294574707746506,
-0.006827345117926598,
0.0417112335562706,
0.029015520587563515,
-0.030864253640174866,
0.022602643817663193,
-0.01668483205139637,
0.08073949068784714,
-0.0688743069767952,
0.0069532678462564945,
-0.05602192506194115,
-0.010145300067961216,
-0.06324496120214462,
-0.04481300711631775,
-0.012600393034517765,
-0.06677263230085373,
0.05466636270284653,
0.009916780516505241,
0.044603075832128525,
-0.008197428658604622,
-0.026625625789165497,
0.001798707526177168,
0.10413674265146255,
-0.07008033245801926,
0.011832267977297306,
0.10341984033584595,
0.11655706912279129,
-0.0005202462780289352,
-0.08069337159395218,
0.046051401644945145,
0.047593314200639725,
-0.05437880754470825,
-0.011470076628029346,
0.0612332820892334,
0.05762975662946701,
0.003842678852379322,
0.056888315826654434,
-0.006947445683181286,
-0.11252529919147491,
0.0323048010468483,
-0.0004467352118808776,
0.04231773316860199,
-0.00400485098361969,
0.03217913955450058,
0.0064612505957484245,
-0.08386718481779099,
-0.10584793239831924,
0.04657112807035446,
-0.017498008906841278,
-0.06187712773680687,
0.08725176751613617,
-0.04227173700928688,
0.01199037954211235,
0.015559196472167969,
0.025383763015270233,
0.015914898365736008,
0.06195516884326935,
0.02946176566183567,
-0.02220543660223484,
0.029774660244584084,
-0.015169107355177402,
-0.009578411467373371,
-0.04000065103173256,
0.013667672872543335,
-0.0236839409917593,
-0.010098596103489399,
-0.004537591710686684,
-0.041369516402482986,
0.1376716047525406,
-0.04136695712804794,
0.008850759826600552,
-0.0788343995809555,
0.03548712283372879,
0.03828165680170059,
-0.05563151463866234,
0.013740744441747665,
0.022092295810580254,
0.037141405045986176,
-0.005803161300718784,
0.03611880540847778,
0.06637518852949142,
-0.026991022750735283,
-0.024824174121022224,
0.11996974796056747,
-0.0038523986004292965,
-0.018505165353417397,
-0.05611288920044899,
-0.013310256414115429,
-0.0009031299268826842,
0.012863731011748314,
-0.002898162929341197,
0.06196986883878708,
-0.05462963879108429,
-0.00509619852527976,
0.029922697693109512,
0.020725548267364502,
-0.02351556345820427,
-0.03689171373844147,
-0.012880818918347359,
-0.01168009266257286,
-0.004423130303621292,
0.0244845412671566,
-0.015935758128762245,
0.009578453376889229,
0.050921883434057236,
-0.022310804575681686,
-0.07343608140945435,
-0.060603946447372437,
7.1115746002267935e-34,
0.041605934500694275,
0.04657401144504547,
0.049903709441423416,
0.06036381423473358,
-0.001502490253187716,
0.03442735970020294,
-0.01927376165986061,
-0.026587164029479027,
-0.034837909042835236,
0.041665446013212204,
-0.07750546187162399,
-0.044743746519088745,
-0.03925056383013725,
0.107907235622406,
-0.005041802767664194,
0.01932854764163494,
-0.08757499605417252,
0.02016873098909855,
-0.08062442392110825,
0.11293930560350418,
0.043944716453552246,
-0.021992802619934082,
-0.049044784158468246,
-0.006273901090025902,
-0.030315609648823738,
0.02426932007074356,
0.01726674474775791,
-0.034092798829078674,
-0.019048091024160385,
-0.012935202568769455,
-0.012424439191818237,
0.04083907604217529,
-0.01803155429661274,
0.0251537524163723,
0.0688922330737114,
-0.0839894562959671,
0.059034932404756546,
-0.08318451046943665,
-0.0725092664361,
-0.026778634637594223,
0.04771140217781067,
0.04779540374875069,
0.02608645334839821,
-0.00006148979446152225,
-0.10847541689872742,
0.002085715066641569,
0.023848645389080048,
0.0902634859085083,
-0.021601416170597076,
0.014044879004359245,
0.031769558787345886,
0.06564213335514069,
0.02705635502934456,
-0.050722766667604446,
-0.09995577484369278,
-0.010838297195732594,
-0.008904576301574707,
-0.05908204987645149,
0.007502192631363869,
-0.03770396113395691,
-0.04027711972594261,
0.005035336595028639,
-0.012249560095369816,
-0.023255744948983192,
-0.07762151211500168,
0.05043718218803406,
0.042921118438243866,
0.047105077654123306,
0.05454765632748604,
0.0027677163016051054,
-0.05711215361952782,
-0.03959285095334053,
0.013879897072911263,
0.013927674852311611,
0.03434370085597038,
0.026969434693455696,
0.03520921617746353,
-0.009082495234906673,
-0.03719587251543999,
0.03307290002703667,
-0.0789705142378807,
0.031622689217329025,
0.023167721927165985,
-0.036752793937921524,
-0.02769087441265583,
-0.11982438713312149,
0.013158422894775867,
-0.004301621578633785,
-0.04545538127422333,
0.0209749024361372,
-0.052875861525535583,
-0.01945664919912815,
-0.07362008839845657,
0.047365788370370865,
-0.03262929990887642,
-2.8355986455630444e-34,
0.04051806405186653,
0.005963263101875782,
0.012376274913549423,
0.03929682448506355,
0.02963530272245407,
-0.08976138383150101,
-0.03441191092133522,
0.019883496686816216,
0.03639833629131317,
-0.04474160447716713,
0.026525581255555153,
-0.055742811411619186,
0.003691786201670766,
0.021576862782239914,
0.06066082790493965,
0.029009448364377022,
-0.04580605775117874,
0.0020357510074973106,
0.00649957824498415,
-0.053297270089387894,
0.00982407946139574,
-0.011031487956643105,
-0.007071011699736118,
-0.05659681186079979,
-0.0768800899386406,
0.12145784497261047,
-0.01925177127122879,
-0.04069402441382408,
0.0002661769976839423,
0.05794845521450043,
-0.019689172506332397,
-0.09659692645072937,
-0.03834838047623634,
0.04270919784903526,
-0.08231697976589203,
0.054957449436187744,
0.0012913751415908337,
0.005667096935212612,
0.0959257036447525,
-0.0170388575643301,
0.0605781227350235,
-0.033446166664361954,
-0.05799652263522148,
0.10044006258249283,
0.008918912149965763,
-0.020876822993159294,
0.006104346364736557,
0.02189173549413681,
-0.05854285508394241,
0.08303192257881165,
-0.07879285514354706,
0.10120503604412079,
-0.03932425379753113,
-0.0729127749800682,
0.017940588295459747,
-0.09064343571662903,
-0.002798964735120535,
-0.07430371642112732,
0.053217530250549316,
0.010996845550835133,
-0.05485108867287636,
0.010735807940363884,
-0.019212400540709496,
0.01422825176268816,
0.004170373547822237,
0.08595146983861923,
0.02416398376226425,
-0.0721265971660614,
-0.016719941049814224,
0.03235235810279846,
0.005917682778090239,
-0.010051408782601357,
0.010001027956604958,
0.033630020916461945,
0.0464196614921093,
-0.06137636676430702,
0.011807834729552269,
0.003839496523141861,
0.08352872729301453,
0.0015151109546422958,
0.04459938034415245,
-0.0939420759677887,
0.026607926934957504,
0.050477851182222366,
0.10097737610340118,
0.10374753177165985,
-0.08676855266094208,
0.06569680571556091,
-0.02633626013994217,
0.03609862178564072,
-0.014896566048264503,
0.10394197702407837,
0.03393428027629852,
0.035702645778656006,
-0.011611361987888813,
-3.551356542175199e-8,
-0.05373753234744072,
0.1058320552110672,
0.0004757909628096968,
-0.07215363532304764,
-0.017443887889385223,
0.03921717032790184,
-0.006627124734222889,
0.04523774981498718,
-0.04700056463479996,
-0.008403048850595951,
0.06893976777791977,
-0.12081476300954819,
0.010078403167426586,
-0.06892508268356323,
-0.02009524032473564,
0.076516292989254,
-0.03830708935856819,
0.016051674261689186,
-0.041917793452739716,
-0.03208807483315468,
-0.0523662231862545,
0.02716170810163021,
-0.0006802162388339639,
0.044634751975536346,
0.02502492628991604,
-0.06757410615682602,
-0.07385089248418808,
0.07451843470335007,
0.03265558183193207,
0.04269983246922493,
-0.03829909488558769,
0.02433718368411064,
-0.027905529364943504,
0.02448815107345581,
-0.0033986461348831654,
0.015949081629514694,
-0.06236422061920166,
-0.03373783454298973,
0.032940302044153214,
-0.059157516807317734,
0.034787580370903015,
0.07025373727083206,
0.013877548277378082,
0.030088938772678375,
0.10382772982120514,
0.11004369705915451,
-0.01858319155871868,
-0.08951853960752487,
0.03145331144332886,
0.035117655992507935,
-0.0012274443870410323,
-0.021081743761897087,
-0.0022674466017633677,
0.07553227990865707,
0.015014264732599258,
0.010442814789712429,
0.033985015004873276,
-0.03653959557414055,
0.1517072468996048,
0.021267535164952278,
-0.0581657849252224,
-0.019571205601096153,
-0.08040530979633331,
0.030625564977526665
] |
MACPO: Weak-to-Strong Alignment via Multi-Agent Contrastive Preference Optimization | https://openreview.net/forum?id=x1Okv4kbVR | [
"Yougang Lyu",
"Lingyong Yan",
"Zihan Wang",
"Dawei Yin",
"Pengjie Ren",
"Maarten de Rijke",
"Zhaochun Ren"
] | Poster | As large language models (LLMs) are rapidly advancing and achieving near-human capabilities on specific tasks, aligning them with human values is becoming more urgent. In scenarios where LLMs outperform humans, we face a weak-to-strong alignment problem where we need to effectively align strong student LLMs through weak supervision generated by weak teachers. Existing alignment methods mainly focus on strong-to-weak alignment and self-alignment settings, and it is impractical to adapt them to the much harder weak-to-strong alignment setting. To fill this gap, we propose a multi-agent contrastive preference optimization (MACPO) framework. MACPO facilitates weak teachers and strong students to learn from each other by iteratively reinforcing unfamiliar positive behaviors while penalizing familiar negative ones. To get this, we devise a mutual positive behavior augmentation strategy to encourage weak teachers and strong students to learn from each other's positive behavior and further provide higher quality positive behavior for the next iteration. Additionally, we propose a hard negative behavior construction strategy to induce weak teachers and strong students to generate familiar negative behavior by fine-tuning on negative behavioral data. Experimental results on the HH-RLHF and PKU-SafeRLHF datasets, evaluated using both automatic metrics and human judgments, demonstrate that MACPO simultaneously improves the alignment performance of strong students and weak teachers. Moreover, as the number of weak teachers increases, MACPO achieves better weak-to-strong alignment performance through more iteration optimization rounds. | weak-to-strong alignment, preference optimization | We propose a multi-agent contrastive preference optimization (MACPO) framework to facilitate weak teachers and strong students learn from each other to improve weak-to-strong alignment performance. | 1,511 | 2410.07672 | [
-0.028000641614198685,
-0.06718941032886505,
-0.02460353448987007,
0.00003721713073900901,
-0.036756426095962524,
-0.014953557401895523,
-0.03166653960943222,
-0.017083842307329178,
0.045711345970630646,
-0.021994899958372116,
0.0413268581032753,
-0.08090338110923767,
0.06622058898210526,
0.09143862873315811,
0.018290480598807335,
0.037307970225811005,
0.16567650437355042,
0.07847750186920166,
-0.07408652454614639,
-0.009968998841941357,
-0.01990421488881111,
0.012986818328499794,
-0.02302571013569832,
-0.05352206155657768,
-0.08316672593355179,
0.03833399713039398,
-0.02978234365582466,
-0.09032423794269562,
0.03217986598610878,
-0.06561969220638275,
0.04202926903963089,
-0.05236002802848816,
0.02063015103340149,
0.054233018308877945,
-0.01535737980157137,
0.09153120219707489,
-0.12922146916389465,
0.07845670729875565,
-0.04730676859617233,
-0.005975859239697456,
-0.07761015743017197,
0.05951780080795288,
-0.0174430962651968,
-0.05931767821311951,
-0.015308824367821217,
-0.02460625395178795,
-0.008906601928174496,
-0.037002578377723694,
0.056199684739112854,
0.005938740912824869,
-0.08032282441854477,
-0.004619833547621965,
0.03866611793637276,
0.03181719034910202,
0.01635557785630226,
0.08864567428827286,
0.010051432996988297,
0.06454962491989136,
0.0021313661709427834,
-0.06859996914863586,
-0.008548911660909653,
-0.07447298616170883,
0.02124812826514244,
-0.039486825466156006,
0.006882977206259966,
-0.0040913415141403675,
-0.07971392571926117,
0.05518248677253723,
-0.08877985924482346,
0.10250911861658096,
-0.04577537253499031,
0.028770383447408676,
0.0762258991599083,
0.01628926210105419,
0.01089514885097742,
0.1110677495598793,
0.028197938576340675,
-0.012450280599296093,
0.0778181403875351,
-0.025740645825862885,
-0.04418204724788666,
-0.022898701950907707,
0.01020820066332817,
0.043507885187864304,
0.14320024847984314,
0.009981957264244556,
-0.02624446712434292,
0.008265127427875996,
-0.005024375859647989,
0.04606551304459572,
-0.060150571167469025,
0.008410763926804066,
0.045013368129730225,
0.04772069305181503,
0.05232435464859009,
0.056400034576654434,
-0.010576997883617878,
-0.011107798665761948,
-0.05303589627146721,
0.08533565700054169,
-0.042916905134916306,
0.12731324136257172,
-0.04027951508760452,
0.002317646984010935,
-0.03739756718277931,
-0.11542561650276184,
-0.011861426755785942,
-0.05081430822610855,
0.06594963371753693,
-0.10687916725873947,
0.013900656253099442,
0.022352764382958412,
0.00776858301833272,
0.05086008831858635,
0.053556572645902634,
-0.0363176129758358,
0.0845322534441948,
0.02968977950513363,
-0.018204933032393456,
-0.022806743159890175,
0.04310796782374382,
-0.04755805432796478,
-0.01095343567430973,
0.001969423610717058,
0.07902716100215912,
0.026695827022194862,
-0.0017795228632166982,
1.0554359699850224e-33,
0.043652698397636414,
0.09929781407117844,
0.04687263071537018,
0.017333494499325752,
0.03671351820230484,
-0.04286382719874382,
-0.03567991778254509,
0.01665201224386692,
-0.10653907805681229,
0.020116420462727547,
0.06138288974761963,
0.04410083591938019,
-0.05830622836947441,
0.1031455546617508,
-0.003042826894670725,
-0.06547105312347412,
-0.03864793851971626,
-0.020990051329135895,
-0.08900562673807144,
-0.01628280244767666,
0.03355204686522484,
0.03807522729039192,
0.041988808661699295,
-0.08053891360759735,
-0.005835241172462702,
-0.034142665565013885,
0.07717622816562653,
-0.0695643499493599,
0.012627732940018177,
-0.02251562289893627,
-0.10806245356798172,
0.01933247782289982,
-0.008640605956315994,
0.05002758279442787,
-0.017335860058665276,
-0.03441297262907028,
-0.04907739907503128,
-0.056457437574863434,
0.011784921400249004,
0.05604294687509537,
0.030555687844753265,
0.07837612926959991,
0.060995087027549744,
-0.0820632129907608,
0.0007319679716601968,
-0.05624710023403168,
0.05226544290781021,
-0.009442397393286228,
0.015494490042328835,
-0.08481419831514359,
0.016616737470030785,
0.027881579473614693,
-0.032254643738269806,
-0.048695869743824005,
-0.06837575882673264,
0.01937769167125225,
0.004631356801837683,
0.0890093669295311,
-0.015403942205011845,
-0.028677381575107574,
-0.0356266051530838,
-0.02967732958495617,
-0.008956974372267723,
0.03157858923077583,
0.0676732212305069,
-0.028812259435653687,
0.009063757956027985,
0.04565151035785675,
0.07686493545770645,
-0.05731424689292908,
0.03822294622659683,
0.02242422290146351,
-0.04622572660446167,
-0.10725510865449905,
0.005933593958616257,
0.0155298737809062,
0.07770267128944397,
-0.06951930373907089,
0.07459557056427002,
-0.04483192786574364,
0.014669973403215408,
-0.019434113055467606,
0.008482430130243301,
-0.07891164720058441,
-0.08483204245567322,
-0.043185923248529434,
0.013310609385371208,
-0.07835602015256882,
0.004350865259766579,
0.01018955372273922,
-0.028409400954842567,
-0.017643820494413376,
-0.0582817867398262,
0.06247168779373169,
0.039848923683166504,
-1.638251028239505e-33,
0.020127685740590096,
-0.011175991035997868,
-0.012312577106058598,
0.038929283618927,
-0.005849630106240511,
-0.016292955726385117,
0.04131723940372467,
0.005068398080766201,
-0.062115542590618134,
-0.039287809282541275,
-0.012173141352832317,
-0.03567042201757431,
0.038065582513809204,
-0.04552903026342392,
-0.025254718959331512,
-0.01562550850212574,
0.045927803963422775,
0.10706323385238647,
0.017895832657814026,
-0.09625140577554703,
0.025891508907079697,
0.1321653574705124,
-0.08401563763618469,
0.1145012378692627,
0.03226900473237038,
0.01310601457953453,
0.03887023776769638,
0.010785599239170551,
-0.0066559938713908195,
-0.017720747739076614,
-0.019921647384762764,
0.045383989810943604,
-0.002925779204815626,
-0.01941663585603237,
-0.04865587502717972,
0.08619125187397003,
-0.08392618596553802,
-0.003073985455557704,
-0.042744267731904984,
0.1038888618350029,
0.02677297778427601,
-0.04122321680188179,
-0.038632575422525406,
-0.0513700507581234,
-0.009664243087172508,
0.03895463049411774,
-0.06780826300382614,
0.0015154771972447634,
0.037101611495018005,
-0.0160796120762825,
-0.032235100865364075,
-0.05535978078842163,
-0.06171996146440506,
0.006629314739257097,
-0.01587963104248047,
-0.053465068340301514,
0.06219661235809326,
-0.008914620615541935,
0.0541636198759079,
0.00016675602819304913,
-0.059264395385980606,
0.010543886572122574,
0.06797327846288681,
-0.0339212529361248,
-0.06212272495031357,
-0.03279056400060654,
0.01290172804147005,
0.026744898408651352,
0.028874078765511513,
-0.050495877861976624,
0.04080791026353836,
0.008960183709859848,
0.018350090831518173,
-0.018267694860696793,
0.03391652554273605,
-0.05660846456885338,
-0.008607996627688408,
-0.03385215997695923,
0.0046766409650444984,
-0.05933033302426338,
0.02204129658639431,
0.0706961378455162,
-0.032560352236032486,
-0.007425086107105017,
0.03378503769636154,
0.05204523727297783,
0.03596322610974312,
0.16486500203609467,
0.035415712743997574,
0.005570916458964348,
-0.06356658041477203,
0.025325465947389603,
0.0025820501614362,
-0.028423326089978218,
-0.14122018218040466,
-4.251320007142567e-8,
-0.07851564884185791,
-0.05620469152927399,
0.027778057381510735,
0.07619357854127884,
-0.048357922583818436,
-0.09127320349216461,
0.0046536363661289215,
0.01812354475259781,
0.030326800420880318,
0.02610270120203495,
0.03833369165658951,
-0.003620557254180312,
-0.024344785138964653,
-0.11897102743387222,
0.01638573408126831,
0.08725157380104065,
-0.0054396940395236015,
0.016123948618769646,
0.025574233382940292,
-0.014423995278775692,
0.003948526922613382,
0.013048543594777584,
-0.0021292208693921566,
0.04251442849636078,
0.0329410545527935,
-0.04672085866332054,
-0.03569144010543823,
0.05024753138422966,
0.04923442006111145,
0.010990658774971962,
0.022878194227814674,
0.002262621419504285,
0.04480966180562973,
-0.058224473148584366,
-0.04540379345417023,
0.0846397876739502,
-0.04083634167909622,
-0.011884981766343117,
0.035831671208143234,
-0.049385543912649155,
-0.012158003635704517,
0.06650710105895996,
0.0180833637714386,
0.008004499599337578,
-0.04162513092160225,
0.016127893701195717,
-0.08665628731250763,
-0.05026299133896828,
0.03765295445919037,
-0.05970141291618347,
0.02205725386738777,
-0.042591530829668045,
0.036241594702005386,
0.027551190927624702,
0.03807470202445984,
0.015192507766187191,
-0.0029477833304554224,
-0.030839819461107254,
0.06102391704916954,
0.0033002698328346014,
0.06295423954725266,
-0.00428660586476326,
-0.0496700294315815,
0.02662394754588604
] |
EDiT: A Local-SGD-Based Efficient Distributed Training Method for Large Language Models | https://openreview.net/forum?id=xtlMtbVfWu | [
"Jialiang Cheng",
"Ning Gao",
"Yun Yue",
"Zhiling Ye",
"Jiadi Jiang",
"Jian Sha"
] | Poster | Distributed training methods are crucial for large language models (LLMs). However, existing distributed training methods often suffer from communication bottlenecks, stragglers, and limited elasticity, particularly in heterogeneous or large-scale environments. Local SGD methods have been proposed to address these issues, but their effectiveness remains limited to small-scale training due to additional memory overhead and lack of concerns on efficiency and stability. To tackle these issues, we propose EDiT, an innovative Efficient Distributed Training method that combines a tailored Local SGD approach with model sharding techniques to enhance large-scale training efficiency. EDiT performs layer-wise parameter synchronization during forward pass, reducing communication and memory overhead and enabling overlap. Besides, EDiT employs a pseudo gradient penalty strategy to suppress loss spikes, which ensures training stability and improves performance. Additionally, we introduce A-EDiT, a fully asynchronous variant of EDiT that accommodates heterogeneous clusters. Building on EDiT/A-EDiT, we conduct a series of experiments to validate large-scale asynchronous training for LLMs, accompanied by comprehensive analyses. Experimental results demonstrate the superior performance of EDiT/A-EDiT, establishing them as robust solutions for distributed LLM training in diverse computational ecosystems. The code is available at Atorch codebase: https://github.com/intelligent-machine-learning/atorch/tree/main/atorch/local_sgd. | Distributed Training, Large Language Models, Local SGD, Training Acceleration | We propose a novel Local SGD-based distributed training method for training LLMs effectively and efficiently, and we provide a large-scale verification of asynchronous pre-training for LLMs. | 1,502 | 2412.07210 | [
-0.006073253694921732,
-0.08046150207519531,
0.01193767786026001,
0.06428235024213791,
0.048839714378118515,
0.03664924576878548,
-0.14987796545028687,
-0.06371542811393738,
0.03241918608546257,
-0.06762197613716125,
-0.025804413482546806,
0.030839281156659126,
0.0022154406178742647,
-0.028068559244275093,
-0.026686208322644234,
-0.04560524970293045,
0.06213301792740822,
0.10383431613445282,
-0.1377321481704712,
-0.058258652687072754,
-0.041045691817998886,
-0.014482984319329262,
0.05577639117836952,
0.051578547805547714,
0.04236115887761116,
-0.016769269481301308,
-0.061350394040346146,
-0.04011889919638634,
0.005457496270537376,
-0.029073823243379593,
0.04373622685670853,
0.013385836966335773,
-0.04617886617779732,
0.06804706901311874,
-0.09468833357095718,
0.10865073651075363,
-0.029282014816999435,
0.03586018830537796,
-0.006110232323408127,
-0.03728647902607918,
0.009798280894756317,
-0.041161034256219864,
0.004418346099555492,
-0.0030980880837887526,
0.055473461747169495,
-0.025164656341075897,
-0.016295772045850754,
0.0023824272211641073,
-0.043337490409612656,
0.014755071140825748,
0.007239397149533033,
-0.06275507062673569,
-0.06232129782438278,
0.08200965076684952,
-0.05121304839849472,
0.04906454309821129,
0.08365480601787567,
0.02115010842680931,
0.00621337816119194,
-0.02492828294634819,
0.006682601757347584,
-0.12965163588523865,
-0.03221343457698822,
0.007693467196077108,
-0.08877439051866531,
-0.014308095909655094,
0.08787558972835541,
0.08101391792297363,
-0.018389178439974785,
-0.028660917654633522,
-0.00048545829486101866,
0.06543295830488205,
-0.062438856810331345,
0.08889570832252502,
-0.01339317113161087,
0.058835219591856,
0.04232924431562424,
-0.0123876528814435,
0.04725511744618416,
-0.02159828320145607,
0.01676151342689991,
0.02712220512330532,
0.03418627381324768,
-0.09856205433607101,
0.1242065280675888,
-0.006382476072758436,
0.04956689476966858,
0.028640443459153175,
0.035970211029052734,
0.01486822497099638,
0.02037470042705536,
0.04743015393614769,
-0.023460835218429565,
-0.03190029412508011,
-0.02677902765572071,
0.005212179385125637,
-0.009604311548173428,
0.027888000011444092,
-0.03853403776884079,
0.10982170701026917,
0.032308634370565414,
0.040659788995981216,
0.0007869249093346298,
0.015330296009778976,
0.05738167464733124,
-0.020495586097240448,
0.054452504962682724,
0.006216349545866251,
0.013155288062989712,
-0.0666474774479866,
0.058127108961343765,
0.07601429522037506,
-0.03520655632019043,
0.0283893421292305,
0.013650079257786274,
0.036718957126140594,
0.022207438945770264,
-0.06245837360620499,
-0.024731386452913284,
0.0834188237786293,
-0.04732128232717514,
-0.016741802915930748,
0.04014662280678749,
-0.01631556637585163,
0.003431527176871896,
-0.00851344596594572,
-0.08265034854412079,
7.754826231160866e-34,
0.09501315653324127,
0.028030777350068092,
0.017781298607587814,
-0.07001659274101257,
0.05814575031399727,
0.003365442156791687,
0.019453980028629303,
0.04891518875956535,
-0.0849774181842804,
-0.07952890545129776,
0.003464524634182453,
-0.010015305131673813,
0.034818511456251144,
0.042571067810058594,
0.062295008450746536,
0.005240365397185087,
0.018504813313484192,
0.06365223973989487,
0.007679251953959465,
-0.046509191393852234,
0.07594747841358185,
-0.0841359868645668,
0.011299096047878265,
-0.03853997960686684,
-0.019721033051609993,
0.001983395777642727,
0.08302466571331024,
-0.013872888870537281,
0.01178274117410183,
0.06921756267547607,
-0.08371879160404205,
-0.06410788744688034,
0.01430747751146555,
0.01694064401090145,
0.04993613064289093,
-0.029327884316444397,
-0.06687292456626892,
-0.07361308485269547,
0.052813392132520676,
-0.06967709958553314,
0.005641620140522718,
0.0012709815055131912,
0.026366939768195152,
-0.07517639547586441,
-0.1476656049489975,
-0.07329240441322327,
-0.016084887087345123,
0.038094017654657364,
-0.04513044282793999,
-0.061223048716783524,
0.055212900042533875,
0.020355025306344032,
-0.06350350379943848,
-0.0010909391567111015,
-0.03338851407170296,
0.0399949811398983,
0.08991403132677078,
0.019979378208518028,
-0.016610916703939438,
0.10214467346668243,
0.025176145136356354,
0.03150094300508499,
-0.06540787220001221,
0.12080474197864532,
0.05958491191267967,
-0.027956485748291016,
-0.056632984429597855,
0.11036945879459381,
0.06661610305309296,
-0.08242928981781006,
-0.03261403739452362,
-0.017631707713007927,
-0.00031518982723355293,
0.04043450579047203,
0.013733137398958206,
-0.061599139124155045,
0.015215575695037842,
-0.04297025129199028,
-0.0026357679162174463,
0.01797392964363098,
0.0034581308718770742,
0.018374014645814896,
-0.0808992013335228,
-0.026429153978824615,
-0.029652094468474388,
-0.05469872057437897,
-0.01467711478471756,
-0.021894702687859535,
-0.02462846413254738,
0.02071140892803669,
-0.10879988223314285,
-0.03454018756747246,
0.04576276242733002,
-0.058126069605350494,
0.006259677931666374,
-1.4564049909031743e-33,
-0.014039603061974049,
-0.020194068551063538,
-0.0660051703453064,
0.14355164766311646,
0.024039726704359055,
-0.022867409512400627,
0.0028913114219903946,
0.062359519302845,
-0.08855818212032318,
-0.05110704153776169,
-0.056285351514816284,
-0.02843194082379341,
0.028802422806620598,
0.012380349449813366,
0.03100794181227684,
-0.0267681535333395,
-0.0007777201244607568,
0.02271251752972603,
0.06113043799996376,
0.04805753752589226,
-0.024567818269133568,
0.01571553386747837,
-0.0016040618065744638,
0.08565988391637802,
0.007113760802894831,
-0.0056653572246432304,
-0.05556783452630043,
0.02935551293194294,
-0.026041613891720772,
0.003131520701572299,
-0.025585727766156197,
0.04046313837170601,
-0.03405527397990227,
-0.06873006373643875,
-0.05965980887413025,
0.05920933932065964,
0.05409757047891617,
-0.02398465760052204,
0.010341614484786987,
0.07996730506420135,
0.13791432976722717,
-0.019941460341215134,
-0.028179168701171875,
0.023792287334799767,
0.008096655830740929,
0.009558555670082569,
-0.17287200689315796,
0.02372954785823822,
-0.0794643834233284,
0.021208908408880234,
-0.07383676618337631,
0.0005298691685311496,
-0.028891844674944878,
0.03027280420064926,
-0.009289758279919624,
-0.07383137196302414,
-0.0399589017033577,
0.0005071207415312529,
-0.007416246924549341,
-0.04000645875930786,
-0.0414709597826004,
-0.05378920957446098,
0.02120610512793064,
-0.08126536011695862,
0.06207830831408501,
-0.02437891624867916,
0.009768841788172722,
0.00040013936813920736,
0.037503838539123535,
0.022616948932409286,
0.052174732089042664,
0.06347564607858658,
-0.06513627618551254,
0.026045046746730804,
-0.015214160084724426,
-0.03915882855653763,
0.004714860115200281,
-0.08790341019630432,
0.0065086353570222855,
-0.0036197728477418423,
0.005280032753944397,
0.04123035445809364,
0.0018917947309091687,
0.025051183998584747,
0.06584770977497101,
0.0625791922211647,
0.046753063797950745,
0.11918829381465912,
-0.0019282340072095394,
-0.03282882273197174,
-0.019875267520546913,
0.06207900494337082,
-0.008083977736532688,
-0.009876156225800514,
-0.06844732910394669,
-3.9470684498610353e-8,
-0.03135387599468231,
0.06553006917238235,
-0.006518984213471413,
0.07851530611515045,
0.01765701361000538,
-0.038421645760536194,
-0.013009404763579369,
0.015039218589663506,
0.0654754564166069,
0.002448573475703597,
0.03836017847061157,
-0.04198189079761505,
-0.00013473333092406392,
-0.08759158849716187,
-0.04701412841677666,
0.051780909299850464,
-0.04333571344614029,
0.06364934146404266,
-0.05051759257912636,
0.011700582690536976,
0.05358395352959633,
0.0725213885307312,
0.0229067113250494,
0.020265568047761917,
0.05812107026576996,
-0.05934976786375046,
0.03065609186887741,
0.09859535098075867,
0.06395872682332993,
-0.07207721471786499,
-0.05118214339017868,
-0.05935380235314369,
-0.024212442338466644,
-0.013190673664212227,
0.007621937431395054,
0.06342887133359909,
0.03035476617515087,
0.006923530250787735,
0.06676913052797318,
0.014238997362554073,
0.05731958523392677,
-0.0004697985132224858,
-0.028217991814017296,
0.012017258442938328,
-0.012871062383055687,
-0.0034880591556429863,
-0.046427905559539795,
-0.06870777159929276,
-0.01477017067372799,
-0.038305848836898804,
0.01290676649659872,
-0.013296583667397499,
-0.004863121546804905,
0.08333425223827362,
0.029690522700548172,
0.02510812133550644,
-0.01930214650928974,
-0.09216421842575073,
0.05933638662099838,
0.04774169996380806,
0.0020277530420571566,
-0.04389271140098572,
-0.10581688582897186,
0.016094503924250603
] |
IDA-VLM: Towards Movie Understanding via ID-Aware Large Vision-Language Model | https://openreview.net/forum?id=N5YTixK4F1 | [
"Yatai Ji",
"Shilong Zhang",
"Jie Wu",
"Peize Sun",
"Weifeng Chen",
"Xuefeng Xiao",
"Sidi Yang",
"Yujiu Yang",
"Ping Luo"
] | Poster | The rapid advancement of Large Vision-Language models (LVLMs) has demonstrated a spectrum of emergent capabilities. Nevertheless, current models only focus on the visual content of a single scenario, while their ability to associate instances across different scenes has not yet been explored, which is essential for understanding complex visual content, such as movies with multiple characters and intricate plots. Towards movie understanding, a critical initial step for LVLMs is to unleash the potential of character identities memory and recognition across multiple visual scenarios. To achieve the goal, we propose visual instruction tuning with ID reference and develop an ID-Aware Large Vision-Language Model, IDA-VLM. Furthermore, our research introduces a novel benchmark MM-ID, to examine LVLMs on instance IDs memory and recognition across four dimensions: matching, location, question-answering, and captioning. Our findings highlight the limitations of existing LVLMs in recognizing and associating instance identities with ID reference. This paper paves the way for future artificial intelligence systems to possess multi-identity visual inputs, thereby facilitating the comprehension of complex visual narratives like movies. | large vision-language model; ID recognition | We propose an ID-aware LVLM, IDA-VLM to recognize instance IDs across diverse scenes, towards understanding complex visual inputs, such as movies. We construct a new benchmark, MM-ID, to examine LVLMs on instance IDs recognition. | 1,496 | null | [
0.00017092324560508132,
-0.08093998581171036,
-0.04616708680987358,
-0.04085047170519829,
0.06507512927055359,
0.04561229795217514,
0.009223638102412224,
0.015495858155190945,
0.07448837906122208,
-0.08883707970380783,
-0.04614556208252907,
-0.09247620403766632,
0.038317956030368805,
0.07213904708623886,
0.008294230327010155,
0.010984143242239952,
0.05143257975578308,
0.07242406159639359,
-0.09393958002328873,
-0.02050686627626419,
0.055551912635564804,
-0.020823892205953598,
0.06306733936071396,
-0.01531164813786745,
0.0006108260131441057,
-0.012048470787703991,
0.016501320526003838,
-0.021610649302601814,
-0.020187878981232643,
-0.051773253828287125,
0.05771303549408913,
0.02802208997309208,
-0.0035468824207782745,
0.05714983120560646,
0.00046262843534350395,
0.058689914643764496,
-0.06909042596817017,
0.04572611302137375,
0.005706984084099531,
-0.07459163665771484,
-0.031780730932950974,
-0.012691821902990341,
0.0672997459769249,
0.013144364580512047,
0.11531729996204376,
0.007529648020863533,
-0.04750703647732735,
-0.01659725233912468,
0.04718821495771408,
-0.0036946770269423723,
-0.09249512851238251,
0.01343473233282566,
-0.05419164523482323,
0.015843408182263374,
-0.03746306151151657,
0.023895524442195892,
0.06550222635269165,
-0.02450036257505417,
0.00353722064755857,
0.027504902333021164,
-0.01093264576047659,
-0.059951502829790115,
0.10667888820171356,
0.02619343250989914,
-0.07359902560710907,
-0.032198987901210785,
0.021557405591011047,
-0.01772455871105194,
-0.04723421111702919,
-0.028208985924720764,
-0.05778972804546356,
0.05967986211180687,
0.0512181855738163,
0.023116163909435272,
0.00675947405397892,
0.048115022480487823,
-0.015703510493040085,
-0.11287831515073776,
0.019782064482569695,
-0.1023319885134697,
0.0020926096476614475,
0.01882885955274105,
0.04033401608467102,
0.012318305671215057,
0.10682593286037445,
0.024151716381311417,
-0.012337991036474705,
0.00821707583963871,
-0.012913376092910767,
0.04123081639409065,
-0.07835762947797775,
-0.1177978664636612,
0.07841216772794724,
0.01081047672778368,
0.18255175650119781,
0.022070536389946938,
0.02647356688976288,
-0.05412141978740692,
-0.004010114353150129,
-0.014856614172458649,
-0.004186598584055901,
0.05848665535449982,
0.020628796890378,
-0.01714225299656391,
0.007023164536803961,
0.007421995513141155,
0.11708204448223114,
0.006622523535043001,
-0.017810750752687454,
-0.1306285709142685,
0.046743832528591156,
0.010416196659207344,
0.024990959092974663,
0.04731656610965729,
0.02190285362303257,
-0.00013762862363364547,
0.047889698296785355,
-0.008608534932136536,
0.026371048763394356,
0.01980317384004593,
0.001974932849407196,
0.017996758222579956,
-0.01689862087368965,
0.0343986451625824,
0.0012602994684129953,
0.02750272862613201,
0.033716484904289246,
-1.2429865528361992e-34,
0.021602241322398186,
0.05127730220556259,
0.004790435545146465,
0.06523638963699341,
0.04249170050024986,
-0.06137299910187721,
-0.07749103754758835,
0.10265415161848068,
-0.08371124416589737,
0.051142409443855286,
-0.0298972986638546,
-0.0023554919753223658,
-0.06409477442502975,
0.09829267114400864,
-0.020882312208414078,
0.06333806365728378,
-0.040430545806884766,
0.06920095533132553,
-0.07499068975448608,
-0.023298999294638634,
-0.012886480428278446,
-0.07639137655496597,
0.002280069049447775,
-0.003513180883601308,
0.05719775706529617,
0.026660947129130363,
0.06062578037381172,
-0.04971716180443764,
-0.009480330161750317,
-0.009611744433641434,
-0.08624578267335892,
0.019468534737825394,
0.0018601084593683481,
-0.019990671426057816,
0.0650639533996582,
-0.025892700999975204,
-0.0334489680826664,
-0.036242302507162094,
0.02819032035768032,
0.015687888488173485,
-0.03794260695576668,
0.03661491721868515,
-0.06513206660747528,
-0.022935450077056885,
-0.10193779319524765,
0.001564661506563425,
0.020353907719254494,
0.06411252170801163,
-0.11755101382732391,
0.002249516313895583,
0.03406019136309624,
0.0552370548248291,
-0.09159697592258453,
-0.07070613652467728,
-0.00859733484685421,
-0.047901760786771774,
-0.002792592626065016,
-0.0008317785686813295,
0.033816032111644745,
-0.007438646163791418,
-0.026701247319579124,
0.08060607314109802,
-0.06726079434156418,
0.11643881350755692,
0.0029750762041658163,
0.04815223813056946,
0.05939026549458504,
0.06723393499851227,
0.050143495202064514,
0.042894478887319565,
-0.011057175695896149,
-0.04510433226823807,
0.01663907617330551,
-0.0501696839928627,
-0.005054340232163668,
-0.0007523966487497091,
0.021007537841796875,
-0.08091626316308975,
-0.0026896705385297537,
0.0857597291469574,
0.003633328713476658,
-0.01946881040930748,
-0.02968033216893673,
-0.07534172385931015,
-0.017060941085219383,
-0.04524980112910271,
0.0683460608124733,
-0.10452711582183838,
0.005736027844250202,
-0.06951508671045303,
0.0022457565646618605,
-0.03434426337480545,
-0.031733546406030655,
-0.07085540145635605,
-0.002290964126586914,
-3.712675617918603e-34,
0.06644897162914276,
-0.02896825782954693,
-0.028245355933904648,
-0.02674659714102745,
0.017762724310159683,
-0.09915009140968323,
0.01025236677378416,
-0.02780444547533989,
-0.01900484226644039,
-0.08308692276477814,
-0.03131920471787453,
-0.020142437890172005,
-0.023792689666152,
0.0012679554056376219,
0.046976834535598755,
-0.02031303010880947,
-0.008961351588368416,
0.01824818179011345,
-0.004971087910234928,
0.07965067774057388,
0.01452973298728466,
0.04975791648030281,
-0.038104865700006485,
-0.04319935664534569,
-0.01987912692129612,
0.0603938102722168,
-0.04846286401152611,
-0.00216285209171474,
-0.03160066530108452,
-0.035822294652462006,
-0.0009501855820417404,
0.01214705128222704,
-0.01821262575685978,
-0.01445435918867588,
-0.05958767980337143,
0.07428344339132309,
0.09805856645107269,
-0.09090933203697205,
-0.031831156462430954,
0.11612243205308914,
0.09120568633079529,
-0.081505686044693,
-0.0617109015583992,
0.016708258539438248,
-0.03685477375984192,
-0.03721514716744423,
-0.010639762505888939,
0.03509048372507095,
-0.0009217566694132984,
0.027308817952871323,
-0.07958167791366577,
0.005493421573191881,
-0.0294532161206007,
-0.05271128937602043,
-0.04000149294734001,
-0.07993778586387634,
-0.05523446202278137,
0.03149505332112312,
-0.0021690784487873316,
-0.02161855809390545,
0.012694772332906723,
0.0054796794429421425,
-0.024002471938729286,
0.01505284383893013,
-0.006447490304708481,
-0.03338859975337982,
-0.05127299576997757,
-0.030558351427316666,
-0.030713070183992386,
-0.023940477520227432,
0.026369450613856316,
-0.10130586475133896,
-0.0305455531924963,
-0.015754714608192444,
-0.00832885317504406,
-0.053357869386672974,
-0.03681008890271187,
-0.009335815906524658,
0.0020355472806841135,
-0.014666241593658924,
0.00619642436504364,
-0.043570056557655334,
0.10097873955965042,
0.10488715767860413,
0.051279712468385696,
0.09523718059062958,
0.0038588084280490875,
0.08478256314992905,
0.023552099242806435,
0.030379049479961395,
-0.06363634020090103,
0.024456366896629333,
0.018959924578666687,
0.04001769796013832,
-0.0682946965098381,
-3.750922417111724e-8,
-0.08886222541332245,
0.09435306489467621,
0.016701163724064827,
-0.010538509115576744,
-0.0077264513820409775,
-0.03738477826118469,
-0.14955537021160126,
0.11727290600538254,
0.04357133433222771,
0.021903080865740776,
0.017349814996123314,
0.0035556075163185596,
0.011223292909562588,
-0.06002786010503769,
0.023619091138243675,
0.05739140138030052,
-0.023938102647662163,
-0.010095691308379173,
0.035898495465517044,
0.03425668179988861,
0.05526813864707947,
0.015177854336798191,
-0.002840121742337942,
0.02427070029079914,
-0.03486380726099014,
-0.043208204209804535,
-0.017337854951620102,
0.05187477543950081,
0.05785488337278366,
0.00991915725171566,
0.027990451082587242,
0.03965895622968674,
0.022830545902252197,
-0.055773284286260605,
0.09617441892623901,
-0.028418658301234245,
0.040821563452482224,
-0.007033337838947773,
0.031842492520809174,
-0.006147414445877075,
0.0562957301735878,
0.031388334929943085,
-0.0069960192777216434,
0.052436210215091705,
0.05053136497735977,
0.05022186040878296,
-0.04390466585755348,
-0.10897009074687958,
-0.010230808518826962,
0.0003084628260694444,
-0.05689409375190735,
-0.0631742849946022,
0.0030176618602126837,
0.18031664192676544,
0.04895959421992302,
-0.012204891070723534,
0.0423983559012413,
-0.003241712460294366,
0.08902469277381897,
0.0571257509291172,
0.003605517791584134,
0.06845927983522415,
-0.003750524017959833,
-0.05854974687099457
] |
Gaussian-Based Instance-Adaptive Intensity Modeling for Point-Supervised Facial Expression Spotting | https://openreview.net/forum?id=daD6uGMeLs | [
"Yicheng Deng",
"Hideaki Hayashi",
"Hajime Nagahara"
] | Poster | Point-supervised facial expression spotting (P-FES) aims to localize facial expression instances in untrimmed videos, requiring only a single timestamp label for each instance during training. To address label sparsity, hard pseudo-labeling is often employed to propagate point labels to unlabeled frames; however, this approach can lead to confusion when distinguishing between neutral and expression frames with various intensities, which can negatively impact model performance. In this paper, we propose a two-branch framework for P-FES that incorporates a Gaussian-based instance-adaptive Intensity Modeling (GIM) module for soft pseudo-labeling. GIM models the expression intensity distribution for each instance. Specifically, we detect the pseudo-apex frame around each point label, estimate the duration, and construct a Gaussian distribution for each expression instance. We then assign soft pseudo-labels to pseudo-expression frames as intensity values based on the Gaussian distribution. Additionally, we introduce an Intensity-Aware Contrastive (IAC) loss to enhance discriminative feature learning and suppress neutral noise by contrasting neutral frames with expression frames of various intensities. Extensive experiments on the SAMM-LV and CAS(ME)$^2$ datasets demonstrate the effectiveness of our proposed framework. Code is available at https://github.com/KinopioIsAllIn/GIM. | micro-expression spotting, semi-supervised learning, soft pseudo-labeling | An instance-adaptive Gaussian-based soft pseudo-labeling method for point-supervised facial expression spotting, which models the expression intensity distribution at the instance level. | 1,490 | null | [
-0.04554111883044243,
-0.13155725598335266,
-0.0346812903881073,
0.05015770345926285,
0.10311967134475708,
0.05223759263753891,
0.06664899736642838,
-0.0654473528265953,
0.007615873124450445,
-0.11191828548908234,
-0.0531766377389431,
-0.10248434543609619,
0.02428419515490532,
0.13492557406425476,
0.04036824405193329,
-0.01566934771835804,
0.07639160007238388,
0.05934518203139305,
0.018091505393385887,
-0.00873518269509077,
0.04378780350089073,
0.039457306265830994,
0.019541552290320396,
0.036122698336839676,
0.05421889200806618,
0.02101209945976734,
0.0436098538339138,
0.012258732691407204,
0.054690781980752945,
-0.0110079450532794,
0.002097110962495208,
-0.08060309290885925,
0.0838819071650505,
0.052206218242645264,
-0.023405777290463448,
-0.01694590225815773,
-0.036303892731666565,
0.07892706245183945,
0.004994731862097979,
0.025336317718029022,
0.0376904122531414,
-0.021375911310315132,
0.004117698408663273,
-0.03610501438379288,
0.06590315699577332,
0.06466519832611084,
0.06436313688755035,
-0.06982307136058807,
0.0005142160225659609,
-0.0011361412471160293,
0.011215941980481148,
-0.05045520141720772,
-0.05730527639389038,
0.0024487669579684734,
-0.051431793719530106,
0.005418302491307259,
0.056206416338682175,
-0.09482964873313904,
0.05285797268152237,
0.056051939725875854,
-0.029533294960856438,
-0.053571976721286774,
0.08099053055047989,
0.013203279115259647,
-0.08013464510440826,
-0.010151170194149017,
0.06052245944738388,
0.028463104739785194,
-0.031319741159677505,
0.025560980662703514,
-0.09928824007511139,
0.02742510847747326,
-0.023017294704914093,
-0.022543761879205704,
-0.07929545640945435,
0.008071819320321083,
-0.004768643528223038,
-0.0003211138828191906,
0.034165725111961365,
-0.002493140287697315,
0.03602897748351097,
-0.054636429995298386,
0.08046459406614304,
-0.04214530065655708,
0.06360480934381485,
-0.07024698704481125,
0.005806971341371536,
0.026052745059132576,
-0.09958247095346451,
0.10919758677482605,
-0.05734803155064583,
-0.024051319807767868,
0.016041968017816544,
-0.01719658263027668,
0.049235545098781586,
-0.030738292261958122,
-0.08085829764604568,
-0.005323600489646196,
0.020558902993798256,
0.058430444449186325,
-0.02441919967532158,
0.006511664483696222,
-0.026961183175444603,
-0.044909220188856125,
0.03489290177822113,
-0.01738912984728813,
-0.014147150330245495,
0.0726260095834732,
0.07797333598136902,
0.007274662144482136,
0.039458245038986206,
0.02731146663427353,
0.01030977163463831,
-0.02620510570704937,
0.02733718976378441,
0.034107889980077744,
0.001007791724987328,
-0.06312651932239532,
-0.018096426501870155,
-0.0039598471485078335,
0.014049447141587734,
-0.006191601045429707,
-0.05558126047253609,
0.014344463124871254,
0.054911401122808456,
0.03703607618808746,
-0.07260191440582275,
2.6678371427659762e-33,
0.05093495547771454,
0.026859896257519722,
-0.005457104649394751,
-0.01645657792687416,
-0.01877661794424057,
0.04441552609205246,
-0.0991034209728241,
-0.035761792212724686,
0.004560545086860657,
0.05431583896279335,
0.031693655997514725,
0.04724357649683952,
-0.04442518949508667,
-0.008451947011053562,
-0.011282223276793957,
0.01790745183825493,
0.011050720699131489,
-0.03646209090948105,
0.020974650979042053,
0.008979341015219688,
-0.051941435784101486,
-0.024942928925156593,
-0.007462894544005394,
-0.01413081306964159,
0.00319303129799664,
0.054537370800971985,
0.06918124854564667,
-0.027908235788345337,
-0.004630291368812323,
-0.03569643571972847,
-0.06329850107431412,
-0.04083072766661644,
0.036346033215522766,
-0.02610683999955654,
0.07155781239271164,
-0.14212746918201447,
0.0359300933778286,
-0.05513232946395874,
-0.005924944765865803,
-0.02912805788218975,
-0.09039416164159775,
0.07391496747732162,
-0.0030065865721553564,
-0.07667401432991028,
-0.029029984027147293,
0.03294491022825241,
0.02598670683801174,
0.07250463217496872,
-0.03142290934920311,
0.06372877955436707,
0.06287630647420883,
-0.010281992144882679,
0.029512792825698853,
-0.022896619513630867,
-0.060627102851867676,
0.07460827380418777,
-0.0014727733796462417,
0.02501348964869976,
-0.0029955548234283924,
-0.0020823541563004255,
0.001657035551033914,
0.03907889127731323,
0.025151720270514488,
-0.001334065105766058,
-0.0444454699754715,
0.06091347709298134,
-0.02397477626800537,
0.08157593011856079,
0.014678528532385826,
0.03186000883579254,
0.030370185151696205,
0.045620594173669815,
-0.0001454922603443265,
-0.06708545237779617,
0.0435740202665329,
-0.05979016795754433,
0.07712450623512268,
-0.031105300411581993,
-0.05497632920742035,
0.0705871731042862,
-0.05236463248729706,
-0.014290411956608295,
0.011620846576988697,
-0.0017645780462771654,
-0.03887530043721199,
-0.018215833231806755,
0.005736959166824818,
-0.039360277354717255,
-0.01652233488857746,
-0.06033403053879738,
0.015618150122463703,
-0.013823151588439941,
-0.07358042150735855,
0.04687865823507309,
-0.05329114571213722,
-1.371243547592969e-33,
0.047233741730451584,
0.02895905263721943,
-0.08573070168495178,
0.014396891929209232,
0.07987649738788605,
-0.04471477493643761,
0.037678781896829605,
0.04806569218635559,
0.05365420877933502,
-0.03917752206325531,
0.03501816466450691,
-0.08851645141839981,
-0.04527239128947258,
-0.024900056421756744,
0.025393996387720108,
0.002292392309755087,
-0.06283847242593765,
0.0009447354823350906,
-0.0642831027507782,
0.03979497402906418,
0.008807916194200516,
0.09308313578367233,
0.02224881388247013,
-0.04838409274816513,
-0.05595811828970909,
0.049532823264598846,
-0.08629431575536728,
0.049242742359638214,
0.04022864252328873,
0.004527698270976543,
-0.01631046272814274,
0.02296825312077999,
-0.04038527235388756,
-0.06568143516778946,
-0.08757619559764862,
0.0037118929903954268,
0.027157438918948174,
-0.016021760180592537,
-0.04327578842639923,
0.04322769120335579,
0.0678950771689415,
0.010389719158411026,
-0.08055475354194641,
0.06785336881875992,
-0.0097357127815485,
0.023201297968626022,
-0.006341950036585331,
-0.052838925272226334,
-0.04182938113808632,
0.010459118522703648,
-0.15392500162124634,
0.0030963041353970766,
-0.09864728897809982,
0.10877684503793716,
-0.10241205990314484,
-0.12603487074375153,
-0.09060485661029816,
0.018320705741643906,
-0.021283671259880066,
-0.023302242159843445,
-0.008600397035479546,
-0.03646421059966087,
-0.04145188629627228,
-0.0048181661404669285,
-0.0007834372809156775,
0.04258357360959053,
0.027931347489356995,
-0.041416142135858536,
0.0024767389986664057,
0.05434754118323326,
0.0843575969338417,
0.04681024327874184,
-0.06811931729316711,
0.03898381069302559,
-0.04729732871055603,
0.03408927097916603,
-0.01992732845246792,
-0.05869090557098389,
-0.014819152653217316,
-0.04104944318532944,
0.07306056469678879,
-0.03978367894887924,
-0.033646754920482635,
0.05168057233095169,
0.024692343547940254,
0.11409805715084076,
0.01914108172059059,
0.04012181609869003,
0.030608557164669037,
0.0004403939237818122,
-0.055905137211084366,
0.05650968477129936,
0.02151295728981495,
0.008980042301118374,
-0.10178571939468384,
-4.1728597466317296e-8,
-0.1400837004184723,
0.013283072970807552,
-0.03384441137313843,
-0.06248798593878746,
0.03662724792957306,
-0.06080076843500137,
-0.016688495874404907,
0.052434369921684265,
0.024766480550169945,
-0.06739707291126251,
0.04380734637379646,
-0.049073103815317154,
0.009496045298874378,
-0.016861697658896446,
-0.006101110018789768,
0.10773974657058716,
0.04004577919840813,
0.07556517422199249,
0.0474184975028038,
0.05613402649760246,
-0.10305868089199066,
0.03758731856942177,
-0.0009163016802631319,
-0.004216430708765984,
-0.028833327814936638,
-0.024631641805171967,
0.056444648653268814,
0.06809035688638687,
0.10015570372343063,
-0.0019736867398023605,
0.023115739226341248,
0.07450604438781738,
0.021839652210474014,
0.0035513401962816715,
0.00731586292386055,
0.00421670638024807,
-0.01592164859175682,
-0.018354766070842743,
-0.018981069326400757,
0.05250781774520874,
-0.09114929288625717,
0.0028662672266364098,
-0.009664583019912243,
0.010679814964532852,
0.011531221680343151,
0.11946161836385727,
0.03638595715165138,
-0.107420414686203,
0.058878980576992035,
-0.026601694524288177,
0.06716720759868622,
-0.0020061079412698746,
-0.05309908464550972,
0.07319719344377518,
0.09111081063747406,
0.010879569686949253,
0.06046868860721588,
-0.0012566337827593088,
0.07287436723709106,
0.10731630027294159,
-0.051154933869838715,
0.03293326497077942,
-0.009732802398502827,
-0.051941417157649994
] |
Learning Spatial-Semantic Features for Robust Video Object Segmentation | https://openreview.net/forum?id=EM93t94zEi | [
"Xin Li",
"Deshui Miao",
"Zhenyu He",
"Yaowei Wang",
"Huchuan Lu",
"Ming-Hsuan Yang"
] | Poster | Tracking and segmenting multiple similar objects with distinct or complex parts in long-term videos is particularly challenging due to the ambiguity in identifying target components and the confusion caused by occlusion, background clutter, and changes in appearance or environment over time. In this paper, we propose a robust video object segmentation framework that learns spatial-semantic features and discriminative object queries to address the above issues. Specifically, we construct a spatial-semantic block comprising a semantic embedding component and a spatial dependency modeling part for associating global semantic features and local spatial features, providing a comprehensive target representation. In addition, we develop a masked cross-attention module to generate object queries that focus on the most discriminative parts of target objects during query propagation, alleviating noise accumulation to ensure effective long-term query propagation. The experimental results show that the proposed method sets new state-of-the-art performance on multiple data sets, including the DAVIS2017 test (\textbf{87.8\%}), YoutubeVOS 2019 (\textbf{88.1\%}), MOSE val (\textbf{74.0\%}), and LVOS test (\textbf{73.0\%}), which demonstrate the effectiveness and generalization capacity of the proposed method. We will make all the source code and trained models publicly available. | Video Object Segmentation, Spatial-Semantic Feature, Long-Term, Discriminative Object Queries | null | 1,489 | 2407.07760 | [
0.04979467764496803,
-0.14105968177318573,
0.06233970448374748,
0.05856623873114586,
0.14251799881458282,
0.014778699725866318,
0.05439328774809837,
-0.09515619277954102,
-0.00780078861862421,
-0.0988079383969307,
-0.08665253221988678,
-0.08807462453842163,
-0.00647145276889205,
0.07606726139783859,
0.004751559346914291,
-0.0387035608291626,
0.12195714563131332,
0.09652164578437805,
-0.09388764202594757,
0.03537587448954582,
0.06104006618261337,
-0.06823915243148804,
0.03283756598830223,
0.027176672592759132,
0.026965266093611717,
0.03123720921576023,
0.013981789350509644,
-0.03068956360220909,
0.04390094801783562,
0.008846210315823555,
0.05002330243587494,
-0.00854372326284647,
-0.01445848774164915,
0.08855760842561722,
-0.04880938678979874,
-0.031181538477540016,
-0.06005335971713066,
-0.00017055396165233105,
-0.02139117754995823,
-0.017417101189494133,
0.0030157268047332764,
0.05778367817401886,
-0.015818621963262558,
-0.031068257987499237,
-0.0020148595795035362,
-0.03152341768145561,
0.02021303027868271,
-0.036707304418087006,
0.03376813977956772,
-0.003892624517902732,
-0.059187475591897964,
-0.08208943903446198,
-0.05540592223405838,
0.0074556623585522175,
-0.0413578636944294,
-0.006787784863263369,
0.01881876215338707,
0.0035746588837355375,
0.03897767886519432,
0.01924000307917595,
-0.01502313558012247,
-0.08753584325313568,
0.12285618484020233,
-0.0028665976133197546,
0.02049950323998928,
-0.10370170325040817,
-0.03729739412665367,
-0.04236353188753128,
0.0079544298350811,
0.014613130129873753,
-0.06843064725399017,
0.05137117952108383,
0.012644506059587002,
0.03554481267929077,
0.021966829895973206,
0.10908626765012741,
-0.029146088287234306,
-0.06890761107206345,
0.037925757467746735,
-0.07341746985912323,
0.0629488006234169,
-0.02123013138771057,
-0.00015718593203928322,
-0.031872790306806564,
0.19051221013069153,
-0.05277615040540695,
0.056031256914138794,
0.022781237959861755,
-0.07258304953575134,
-0.02654828503727913,
-0.04527154192328453,
-0.033676858991384506,
0.09550930559635162,
0.0001241173449670896,
0.03723137825727463,
-0.06399302184581757,
-0.05163269117474556,
-0.047588251531124115,
0.021750742569565773,
0.022022172808647156,
0.002496464177966118,
0.04094541817903519,
0.05476315692067146,
0.020646097138524055,
0.025234811007976532,
0.00114784250035882,
0.01754228211939335,
0.021787650883197784,
0.04158080741763115,
-0.047328051179647446,
0.018167028203606606,
0.04650379344820976,
0.004405360668897629,
-0.0254437904804945,
-0.051270026713609695,
0.0010579159716144204,
-0.02365919016301632,
-0.0020816754549741745,
0.019755328074097633,
0.0018923087045550346,
0.012399840168654919,
0.0026959923561662436,
-0.05424284562468529,
-0.0356866717338562,
0.0725393146276474,
-0.02699112519621849,
-0.010327158495783806,
-8.99887933849057e-35,
0.06673812866210938,
-0.0939335823059082,
0.024854378774762154,
-0.021054252982139587,
0.0323774479329586,
0.015732893720269203,
0.021902838721871376,
0.04765576124191284,
-0.02980688400566578,
0.026008283719420433,
0.02229965664446354,
-0.0904310941696167,
-0.033462394028902054,
0.027507590129971504,
0.007981842383742332,
-0.009998421184718609,
-0.0408044271171093,
-0.06670457124710083,
-0.05448590964078903,
0.11313597112894058,
0.002206148114055395,
0.0049658180214464664,
-0.01769900508224964,
0.024407831951975822,
0.05735594034194946,
0.006868528667837381,
0.025466354563832283,
-0.09879622608423233,
0.007899069227278233,
0.02895628660917282,
-0.08436006307601929,
0.03723858296871185,
-0.032558441162109375,
0.09557552635669708,
0.06286069750785828,
-0.0558936670422554,
-0.06661898642778397,
-0.031469911336898804,
0.06092822179198265,
-0.026643332093954086,
-0.03618254512548447,
0.033307842910289764,
-0.05094233527779579,
-0.0446401908993721,
-0.03695100173354149,
-0.04273753613233566,
0.05119079723954201,
0.05109461769461632,
-0.05025975778698921,
0.07900295406579971,
0.04461061209440231,
0.030999628826975822,
-0.016038522124290466,
-0.0639374703168869,
-0.03544595092535019,
-0.012978626415133476,
0.05995422601699829,
0.048591967672109604,
0.024078749120235443,
-0.05548163130879402,
-0.005485814530402422,
0.015312324278056622,
-0.018506590276956558,
0.03834938257932663,
-0.03863474726676941,
0.002024353016167879,
0.08538118004798889,
0.08116736263036728,
0.03270268440246582,
0.09121344983577728,
0.00016094806778710335,
0.086287721991539,
-0.030105652287602425,
-0.04440432786941528,
0.04082752764225006,
-0.0002909544564317912,
0.011132806539535522,
-0.029803413897752762,
0.005352600943297148,
0.01763368770480156,
-0.012702930718660355,
-0.031096840277314186,
0.025670114904642105,
-0.049764543771743774,
-0.06356393545866013,
-0.012126591056585312,
0.0776456743478775,
-0.08847109973430634,
-0.019721901044249535,
-0.000005900040832784725,
0.03951327130198479,
-0.02138061262667179,
-0.031188342720270157,
0.004107312764972448,
0.02156188152730465,
6.99040684659594e-34,
0.03775423392653465,
0.022367235273122787,
0.05450534448027611,
-0.017105599865317345,
-0.06535787135362625,
-0.04139268398284912,
-0.06344969570636749,
0.011543254368007183,
-0.017720716074109077,
-0.043795663863420486,
-0.09042516350746155,
-0.09311315417289734,
-0.02126098982989788,
0.007188857067376375,
-0.023318512365221977,
0.053074467927217484,
0.025656357407569885,
0.03229178488254547,
-0.07416921854019165,
-0.0016683946596458554,
-0.045306190848350525,
0.04065684974193573,
-0.05829356610774994,
-0.014367845840752125,
-0.09609099477529526,
0.0521155446767807,
-0.0007075674366205931,
0.0039825174026191235,
-0.0046860044822096825,
-0.04591861739754677,
-0.05316833406686783,
-0.022270672023296356,
0.005666144657880068,
-0.07396462559700012,
-0.1376648247241974,
0.014043844304978848,
0.03747484087944031,
-0.04587826132774353,
-0.04580388590693474,
0.019961584359407425,
0.01750560663640499,
-0.05896386504173279,
-0.00041193433571606874,
0.05712330713868141,
-0.08850028365850449,
-0.042271628975868225,
-0.046670690178871155,
0.1283160299062729,
-0.0384773425757885,
0.017030108720064163,
-0.0703018307685852,
0.019148515537381172,
-0.09903548657894135,
0.03472568839788437,
-0.007284713443368673,
0.008417647331953049,
-0.05664398893713951,
0.05251318961381912,
0.0605023056268692,
-0.04825998470187187,
-0.0027290915604680777,
0.05741307884454727,
-0.020128794014453888,
0.0013724567834287882,
0.035188253968954086,
0.0362267792224884,
-0.027456335723400116,
0.026485096663236618,
0.055183544754981995,
-0.01627066731452942,
0.08033991605043411,
-0.01092837005853653,
0.040353383868932724,
0.03880348056554794,
0.023839004337787628,
0.06353697180747986,
0.02540724165737629,
-0.04124808311462402,
0.09040010720491409,
0.04470810666680336,
0.027510082349181175,
-0.1137690395116806,
0.018524430692195892,
0.05799197405576706,
0.07807967811822891,
0.10939483344554901,
0.059734780341386795,
-0.007577129639685154,
-0.04240146279335022,
0.0033629140816628933,
-0.07591063529253006,
0.0017885962733998895,
-0.03168436512351036,
0.050082091242074966,
-0.001701099332422018,
-3.888899158255299e-8,
-0.09217483550310135,
0.04051817208528519,
-0.11082718521356583,
-0.0055196527391672134,
-0.018518196418881416,
-0.03707611560821533,
0.025580385699868202,
0.1376921683549881,
-0.016310404986143112,
0.016050396487116814,
0.06024280935525894,
-0.013251405209302902,
0.03852612525224686,
0.011928222142159939,
-0.000823529320769012,
0.022791573777794838,
0.002301062922924757,
0.04122364521026611,
-0.019804697483778,
0.05997199937701225,
-0.08370907604694366,
0.012467456050217152,
0.06337753683328629,
0.018997445702552795,
-0.005786405876278877,
0.0005793255404569209,
0.0335717536509037,
0.15149416029453278,
0.11769718676805496,
-0.025461537763476372,
-0.03332086279988289,
0.01101701520383358,
0.03605850040912628,
-0.03835413604974747,
0.04574473574757576,
0.05696214735507965,
-0.008431266993284225,
-0.028243491426110268,
-0.024043738842010498,
-0.009201117791235447,
0.06027623265981674,
-0.0319681242108345,
0.05374760180711746,
0.036170583218336105,
0.010884230025112629,
0.09494119882583618,
0.010342034511268139,
-0.043668098747730255,
0.03742824122309685,
-0.06109102442860603,
-0.001616169116459787,
-0.02249481901526451,
-0.06119121238589287,
0.12091421335935593,
0.027311010286211967,
-0.026230910792946815,
0.007937674410641193,
-0.0069496347568929195,
0.025387080386281013,
0.029408125206828117,
-0.03958143666386604,
0.0364944189786911,
-0.0193773340433836,
0.027589229866862297
] |
Boosting the visual interpretability of CLIP via adversarial fine-tuning | https://openreview.net/forum?id=khuIvzxPRp | [
"Shizhan Gong",
"Haoyu LEI",
"Qi Dou",
"Farzan Farnia"
] | Poster | CLIP has achieved great success in visual representation learning and is becoming an important plug-in component for many large multi-modal models like LLaVA and DALL-E. However, the lack of interpretability caused by the intricate image encoder architecture and training process restricts its wider use in high-stake decision making applications. In this work, we propose an unsupervised adversarial fine-tuning (AFT) with norm-regularization to enhance the visual interpretability of CLIP. We provide theoretical analysis showing that AFT has implicit regularization that enforces the image encoder to encode the input features sparsely, directing the network's focus towards meaningful features. Evaluations by both feature attribution techniques and network dissection offer convincing evidence that the visual interpretability of CLIP has significant improvements. With AFT, the image encoder prioritizes pertinent input features, and the neuron within the encoder exhibits better alignment with human-understandable concepts. Moreover, these effects are generalizable to out-of-distribution datasets and can be transferred to downstream tasks. Additionally, AFT enhances the visual interpretability of derived large vision-language models that incorporate the pre-trained CLIP an integral component. The code of this paper is available at [the CLIP_AFT GitHub repository](https://github.com/peterant330/CLIP_AFT). | interpretability, vision-language models, CLIP, adversarial fine-tuning | null | 1,472 | null | [
-0.030780304223299026,
-0.0970117524266243,
0.03744199499487877,
0.037117019295692444,
0.1255332976579666,
0.014551911503076553,
-0.030205117538571358,
-0.08011157065629959,
-0.003505360335111618,
-0.060568708926439285,
-0.046002697199583054,
-0.06659776717424393,
-0.014767663553357124,
0.05202522501349449,
-0.09132051467895508,
-0.014380767941474915,
0.11738526076078415,
0.11629585176706314,
-0.15404367446899414,
-0.05684097856283188,
0.05546798184514046,
-0.05162627622485161,
0.02412775531411171,
-0.005883651785552502,
0.07211607694625854,
0.01634981483221054,
0.0056891608983278275,
-0.0592733658850193,
0.03365398570895195,
-0.08904575556516647,
0.031160840764641762,
-0.03402317315340042,
0.005747917573899031,
0.04959433153271675,
-0.1494288295507431,
0.08704397082328796,
-0.0175812765955925,
0.07549142837524414,
0.04380181059241295,
-0.048514485359191895,
0.03817562386393547,
0.007897157222032547,
-0.022446447983384132,
-0.008526070043444633,
0.012590417638421059,
-0.05065589025616646,
0.022237537428736687,
-0.07943222671747208,
0.04571918770670891,
0.021662071347236633,
0.008787658996880054,
-0.01292397826910019,
-0.03479410335421562,
0.07889828085899353,
0.02153688110411167,
0.014178923331201077,
0.06698065251111984,
0.047619886696338654,
0.026251597329974174,
-0.005090823397040367,
0.026808204129338264,
-0.03145430609583855,
0.002357473596930504,
0.009681391529738903,
0.04252266883850098,
-0.08148781210184097,
0.0581534244120121,
0.0303192101418972,
-0.03773588314652443,
-0.047179821878671646,
0.03311151638627052,
0.07954639196395874,
-0.06457868963479996,
-0.0033517596311867237,
0.03578382357954979,
0.09323540329933167,
0.012726062908768654,
0.012031812220811844,
0.05803675577044487,
-0.04733547195792198,
0.059324588626623154,
0.02611847035586834,
0.04798722267150879,
0.047342222183942795,
0.09523569792509079,
0.06660868972539902,
-0.07139544188976288,
0.034717004746198654,
0.045789480209350586,
0.03376924246549606,
-0.06020159646868706,
-0.03399842977523804,
-0.021391410380601883,
0.019832663238048553,
0.06411165744066238,
0.048348840326070786,
-0.006664082873612642,
-0.08996643871068954,
-0.07922524213790894,
0.07884301990270615,
-0.0008629689691588283,
0.05957826226949692,
0.008447393774986267,
-0.008604596368968487,
0.07938714325428009,
-0.018243441358208656,
0.11353524774312973,
0.00984279252588749,
0.036761898547410965,
-0.016846606507897377,
0.03593084216117859,
0.05207649618387222,
0.04169447720050812,
0.0217302106320858,
0.008354201912879944,
-0.012528919614851475,
-0.0455368272960186,
-0.04376890882849693,
0.016621002927422523,
-0.024443626403808594,
-0.04460732266306877,
0.01489156112074852,
-0.057935118675231934,
0.03854158893227577,
0.10255865007638931,
-0.03362264484167099,
-0.022769911214709282,
1.9905327261855287e-33,
0.0018364995485171676,
0.025623798370361328,
0.02248547598719597,
-0.010563363321125507,
0.067515067756176,
-0.04711388424038887,
-0.07106315344572067,
0.03759011998772621,
-0.0713818371295929,
0.04388390854001045,
-0.04009850323200226,
-0.00795156043022871,
-0.06425484269857407,
0.14014992117881775,
0.045690104365348816,
-0.03935809060931206,
0.015261068940162659,
0.029195114970207214,
-0.02524886466562748,
-0.01606721431016922,
-0.015426439233124256,
-0.04861504212021828,
-0.050490617752075195,
-0.023328717797994614,
0.014698199927806854,
-0.05167635530233383,
0.06793820858001709,
-0.0338812991976738,
-0.024628896266222,
0.014989625662565231,
-0.09360980987548828,
0.04661482572555542,
0.08842869102954865,
0.001818174379877746,
-0.04675062745809555,
-0.0868239477276802,
-0.02859274297952652,
-0.0015897760167717934,
0.011416144669055939,
0.04346845671534538,
0.048946280032396317,
0.05627613887190819,
-0.06271976232528687,
0.03496772050857544,
-0.03273019194602966,
-0.022351155057549477,
0.01793104223906994,
0.05775586515665054,
-0.14930106699466705,
-0.005194178316742182,
0.06133287772536278,
0.023091938346624374,
0.0073016611859202385,
-0.1094895750284195,
-0.09490364789962769,
-0.01311838161200285,
0.07016822695732117,
0.0024030939675867558,
0.024965912103652954,
0.006434318609535694,
-0.03506976738572121,
0.06789353489875793,
-0.015360296703875065,
0.09337447583675385,
-0.011099196970462799,
-0.0035278357099741697,
-0.00796002522110939,
0.06347782909870148,
-0.003191425232216716,
0.013641607947647572,
-0.11828649789094925,
0.10671161860227585,
-0.042580947279930115,
-0.056327033787965775,
0.00695525947958231,
0.021887660026550293,
0.021327098831534386,
-0.07089237123727798,
0.06451208889484406,
0.050002455711364746,
-0.008341650478541851,
0.0037390743382275105,
0.00937661062926054,
0.01206094678491354,
-0.05641308054327965,
-0.043149251490831375,
0.044691234827041626,
-0.06745597720146179,
0.0012488432694226503,
0.00963966827839613,
0.0011590411886572838,
-0.006625349633395672,
-0.026201874017715454,
-0.04776158928871155,
0.04117581993341446,
-1.1218016948441862e-34,
-0.010784157551825047,
0.09123218804597855,
-0.04200458899140358,
0.040583085268735886,
-0.0004904704983346164,
0.027088353410363197,
-0.029589621350169182,
0.05120132490992546,
-0.033446457237005234,
-0.05508873611688614,
-0.03512139618396759,
-0.032795511186122894,
0.008234592154622078,
-0.01731841266155243,
0.007547799963504076,
-0.028790252283215523,
-0.05864382162690163,
0.06959617137908936,
-0.08089510351419449,
-0.017198195680975914,
0.07785842567682266,
0.04857999458909035,
-0.01648726686835289,
0.0632135272026062,
-0.04835682734847069,
0.05774706229567528,
-0.04913927987217903,
0.031101129949092865,
-0.018312616273760796,
0.006361660081893206,
-0.026909906417131424,
0.030561860650777817,
-0.018317649140954018,
-0.06119588762521744,
-0.05093924328684807,
0.06378698348999023,
0.023965410888195038,
0.0052940272726118565,
-0.048584915697574615,
0.04864075779914856,
0.03400598093867302,
0.007003836799412966,
-0.09416461735963821,
0.04032287746667862,
-0.029600854963064194,
-0.0816740095615387,
0.03542441874742508,
-0.03311225399374962,
0.02911899983882904,
0.05716630071401596,
-0.04020508751273155,
0.005885713268071413,
-0.02890883944928646,
0.06794899702072144,
-0.01841421239078045,
-0.04970763623714447,
-0.011817162856459618,
0.0278756245970726,
0.14155705273151398,
-0.005070639308542013,
-0.04037481173872948,
-0.0342576839029789,
-0.04252519831061363,
-0.03186163678765297,
-0.02016413025557995,
-0.03446799889206886,
-0.01106411311775446,
0.012680526822805405,
0.043712418526411057,
-0.005851015914231539,
0.015197972767055035,
-0.08291269838809967,
0.00609164871275425,
0.04284510761499405,
0.04687231034040451,
-0.045576129108667374,
0.029085105285048485,
0.021737482398748398,
0.003349314909428358,
0.034673988819122314,
0.017177987843751907,
-0.05122021958231926,
-0.008114115335047245,
0.04162195324897766,
0.07696396857500076,
0.03550543263554573,
0.005574468988925219,
0.002601881278678775,
0.012893564067780972,
0.02783135138452053,
-0.007378758862614632,
0.01891685090959072,
-0.0912758931517601,
0.06394701451063156,
0.022645141929388046,
-4.073300630125232e-8,
-0.08154158294200897,
0.010138402692973614,
-0.0069288220256567,
0.018820375204086304,
0.010589157231152058,
-0.07734785228967667,
-0.030466478317975998,
0.08451370894908905,
-0.00749171432107687,
-0.02092447504401207,
0.047659482806921005,
-0.00836711935698986,
-0.09120295941829681,
-0.055081598460674286,
0.014930672012269497,
0.061587635427713394,
0.05601099506020546,
0.10149289667606354,
0.005345266778022051,
0.061018191277980804,
0.012159406207501888,
-0.03853199630975723,
0.011149143800139427,
-0.059751491993665695,
0.04337632656097412,
-0.08417456597089767,
-0.08995227515697479,
0.09630998224020004,
0.0714704617857933,
0.034653227776288986,
0.030880693346261978,
0.02900788187980652,
0.05647337809205055,
-0.04665115848183632,
-0.0783592164516449,
0.06948859244585037,
-0.024060195311903954,
-0.035806626081466675,
0.015327020548284054,
0.04595481976866722,
0.03615817427635193,
0.06251634657382965,
0.04141458496451378,
-0.002959917765110731,
-0.0804109051823616,
0.004018791019916534,
0.09599422663450241,
-0.04853823781013489,
0.017377929762005806,
-0.07777513563632965,
0.0060408045537769794,
-0.06991784274578094,
0.07113400101661682,
0.09727805852890015,
-0.0033124482724815607,
-0.06653590500354767,
-0.0055870735086500645,
-0.033022794872522354,
0.0334123931825161,
0.06714969128370285,
0.0327809602022171,
0.03017401322722435,
-0.06850405782461166,
-0.052173011004924774
] |
DiffSplat: Repurposing Image Diffusion Models for Scalable Gaussian Splat Generation | https://openreview.net/forum?id=eajZpoQkGK | [
"Chenguo Lin",
"Panwang Pan",
"Bangbang Yang",
"Zeming Li",
"Yadong MU"
] | Poster | Recent advancements in 3D content generation from text or a single image struggle with limited high-quality 3D datasets and inconsistency from 2D multi-view generation. We introduce DiffSplat, a novel 3D generative framework that natively generates 3D Gaussian splats by taming large-scale text-to-image diffusion models. It differs from previous 3D generative models by effectively utilizing web-scale 2D priors while maintaining 3D consistency in a unified model. To bootstrap the training, a lightweight reconstruction model is proposed to instantly produce multi-view Gaussian splat grids for scalable dataset curation. In conjunction with the regular diffusion loss on these grids, a 3D rendering loss is introduced to facilitate 3D coherence across arbitrary views. The compatibility with image diffusion models enables seamless adaptions of numerous techniques for image generation to the 3D realm. Extensive experiments reveal the superiority of DiffSplat in text- and image-conditioned generation tasks and downstream applications. Thorough ablation studies validate the efficacy of each critical design choice and provide insights into the underlying mechanism. | 3D Generation, Diffusion models, 3D Gaussian Splatting | We introduce a novel 3D generative framework that directly generate 3D Gaussian splats by taming large text-to-image diffusion models, effectively utilizing 2D priors and maintaining 3D consistency in a unified model. | 1,469 | 2501.16764 | [
-0.04093708470463753,
-0.10149631649255753,
0.00995201151818037,
0.029554324224591255,
0.08473515510559082,
-0.04974500462412834,
-0.09974639117717743,
-0.10386432707309723,
0.002069012960419059,
-0.06786900013685226,
-0.01078152097761631,
-0.0005343675147742033,
0.007860759273171425,
0.05378863960504532,
-0.05570894479751587,
-0.02939450554549694,
0.07054127752780914,
0.07783029228448868,
-0.10789783298969269,
0.01873885467648506,
-0.018356777727603912,
-0.06277666240930557,
-0.0060410816222429276,
-0.02107737958431244,
0.09749212861061096,
0.09207642078399658,
0.05097422003746033,
-0.03456950932741165,
0.05792054906487465,
-0.01681472547352314,
0.03929676488041878,
-0.0129297636449337,
-0.07395357638597488,
0.018420835956931114,
-0.01431456208229065,
0.07322436571121216,
-0.021777361631393433,
0.053819816559553146,
0.021293241530656815,
0.0021071583032608032,
0.0644570142030716,
0.07962733507156372,
-0.02153508923947811,
0.030634401366114616,
0.06502838432788849,
-0.014479639939963818,
-0.04298332706093788,
-0.06231016293168068,
0.039128512144088745,
0.0022860458120703697,
-0.06769195944070816,
-0.04364631325006485,
-0.029209788888692856,
0.015115645714104176,
0.020771455019712448,
0.021777482703328133,
0.006850480567663908,
-0.03422115370631218,
0.03302868828177452,
-0.017933672294020653,
-0.03585785999894142,
-0.0299636609852314,
0.024113493040204048,
-0.019682398065924644,
0.01646140031516552,
-0.029733654111623764,
0.04992435500025749,
-0.031541645526885986,
-0.05997518450021744,
0.05570182949304581,
-0.07808952778577805,
0.03800129145383835,
-0.0028932567220181227,
-0.003925074357539415,
-0.03723619505763054,
0.03534158319234848,
-0.016044344753026962,
-0.08421959728002548,
-0.04287916421890259,
-0.025538615882396698,
0.061094339936971664,
0.03672304004430771,
0.08810316771268845,
0.0383651927113533,
-0.013044700026512146,
-0.022391309961676598,
0.026960616931319237,
0.09397336095571518,
-0.041915036737918854,
0.02542555145919323,
-0.04992588609457016,
0.04754609614610672,
-0.011859433725476265,
-0.014776925556361675,
-0.09853487461805344,
0.019729232415556908,
-0.00416934210807085,
-0.0669865533709526,
0.024357322603464127,
0.031485866755247116,
0.03757821023464203,
-0.018296608701348305,
0.06810927391052246,
0.024206437170505524,
0.05739954113960266,
-0.06547758728265762,
0.08628231287002563,
0.04638918861746788,
0.005108832847326994,
0.029488375410437584,
0.06488177925348282,
0.04571021720767021,
0.010234949178993702,
-0.007288064807653427,
-0.0035589218605309725,
-0.0007870877743698657,
-0.03413998335599899,
-0.03358747437596321,
0.04027131572365761,
0.005755294114351273,
0.012413762509822845,
-0.005741138011217117,
-0.04909723624587059,
-0.015722863376140594,
0.0028860762249678373,
-0.027622060850262642,
-0.006593276280909777,
2.9222356097900202e-33,
0.0045146034099161625,
-0.0017837523482739925,
0.016428042203187943,
0.09637431055307388,
0.10610179603099823,
0.04373547434806824,
-0.034240759909152985,
-0.056391600519418716,
-0.03694373741745949,
-0.07825270295143127,
-0.045895058661699295,
-0.02892918884754181,
-0.055083297193050385,
0.1385718435049057,
0.012882813811302185,
-0.02391514740884304,
-0.034137580543756485,
0.014065963216125965,
-0.012939409352838993,
0.10377282649278641,
0.013905507512390614,
-0.04443037509918213,
-0.01775292307138443,
-0.009226600639522076,
-0.06659425795078278,
0.012209239415824413,
0.03416347876191139,
-0.02872779406607151,
-0.016054688021540642,
-0.02843800000846386,
-0.05831728130578995,
0.008679388090968132,
0.017945820465683937,
0.012675420381128788,
0.004539333749562502,
-0.0572257936000824,
0.005549673456698656,
-0.0760425478219986,
-0.04433559253811836,
-0.012920142151415348,
-0.05703010782599449,
0.05553539842367172,
0.020619213581085205,
-0.049463242292404175,
-0.10115470737218857,
0.07886803895235062,
0.09000294655561447,
0.0361294150352478,
-0.03335491195321083,
-0.06281416863203049,
0.10408949851989746,
0.017281845211982727,
0.009964874014258385,
-0.05953339487314224,
-0.04388892278075218,
-0.06988465040922165,
0.06322610378265381,
-0.05050906538963318,
0.12168619781732559,
0.02668480947613716,
0.011407660320401192,
0.04368160292506218,
-0.024096496403217316,
0.08729922026395798,
0.009904942475259304,
-0.021670302376151085,
0.007181250955909491,
0.03690426051616669,
0.017147215083241463,
0.00590870063751936,
-0.01732356660068035,
-0.021631130948662758,
-0.013379590585827827,
-0.050650522112846375,
0.01980532892048359,
-0.018342502415180206,
0.04157193377614021,
-0.02349783107638359,
0.06462496519088745,
0.018478211015462875,
-0.06835883855819702,
-0.03727810084819794,
0.0391501747071743,
-0.1275825947523117,
-0.051622431725263596,
-0.04593954607844353,
-0.030885327607393265,
-0.058725133538246155,
-0.028147201985120773,
-0.06304316222667694,
-0.0801238939166069,
-0.06052250787615776,
0.005588939879089594,
-0.06334229558706284,
0.023382654413580894,
-2.4360485748960296e-33,
0.02483527548611164,
0.03613514453172684,
-0.05310545116662979,
0.06156950816512108,
-0.014914615079760551,
0.041139982640743256,
0.0007345885969698429,
0.05578014254570007,
0.09146182984113693,
-0.10641676187515259,
-0.03265104815363884,
0.027383141219615936,
0.005049446132034063,
0.008490582928061485,
-0.017238248139619827,
-0.028994718566536903,
0.05121854320168495,
-0.0725378543138504,
-0.10737679153680801,
-0.028898291289806366,
0.026349011808633804,
0.0028263614512979984,
-0.04169173911213875,
0.03699332848191261,
-0.010467855259776115,
0.11281958967447281,
-0.006348986644297838,
0.031979579478502274,
-0.019193042069673538,
-0.0012381954584270716,
-0.020817263051867485,
0.023886844515800476,
0.05732607841491699,
-0.07047663629055023,
-0.1114935353398323,
0.06076177954673767,
0.07547392696142197,
0.025234172120690346,
-0.072936050593853,
0.014874589629471302,
-0.022162562236189842,
-0.022239835932850838,
-0.057576656341552734,
0.1133541390299797,
-0.010589344426989555,
0.027029240503907204,
-0.023616697639226913,
0.004004695452749729,
0.03974993899464607,
0.13362374901771545,
-0.054658129811286926,
0.11370866000652313,
-0.04444868490099907,
0.04508117586374283,
0.008382095955312252,
-0.17123320698738098,
-0.1289278268814087,
-0.0527026429772377,
0.0591617152094841,
-0.032560933381319046,
-0.028271008282899857,
-0.06458155810832977,
-0.021512065082788467,
-0.05662141367793083,
-0.014146383851766586,
0.008927534334361553,
-0.060959041118621826,
-0.0040150536224246025,
0.009492055512964725,
0.116886205971241,
-0.037669338285923004,
0.021246660500764847,
-0.005971184000372887,
-0.014148370362818241,
0.04011993855237961,
-0.025681225582957268,
-0.0004300894506741315,
0.04238732531666756,
0.05114171281456947,
0.04327722638845444,
0.09241019934415817,
0.054580897092819214,
0.020047899335622787,
0.08971510827541351,
0.08748592436313629,
0.010281515307724476,
0.015071133151650429,
0.009915636852383614,
0.027834750711917877,
0.0279028732329607,
-0.000005281964604364475,
-0.038228362798690796,
0.01575005054473877,
0.057459279894828796,
-0.027297837659716606,
-3.940651183143018e-8,
-0.07012943178415298,
0.051431167870759964,
-0.0025326362811028957,
-0.023413727059960365,
-0.02186385914683342,
-0.05223676189780235,
0.024625612422823906,
0.05998746305704117,
0.001419421168975532,
-0.11030823737382889,
0.03159201890230179,
-0.10745575278997421,
-0.03904837742447853,
-0.04000610485672951,
0.07196744531393051,
0.10070306807756424,
0.02498319372534752,
-0.030591482296586037,
0.01838872767984867,
0.013086033053696156,
-0.07990778982639313,
0.034540168941020966,
-0.006800079718232155,
0.07827507704496384,
0.03885046765208244,
0.07098241150379181,
-0.04994768276810646,
0.01647261343896389,
0.11028210073709488,
-0.04566073417663574,
0.018429888412356377,
-0.02231587842106819,
0.02163008041679859,
0.006756340153515339,
-0.046339355409145355,
0.028491809964179993,
-0.00991015788167715,
0.0021943359170109034,
0.028391005471348763,
-0.11723918467760086,
-0.009748945012688637,
0.04473350942134857,
0.030377143993973732,
-0.005841145757585764,
-0.023806264623999596,
0.04287976399064064,
0.02943728119134903,
-0.0654873326420784,
0.049956440925598145,
0.03018389642238617,
-0.006854998879134655,
-0.06988094002008438,
0.00318856630474329,
0.08809219300746918,
0.07362671196460724,
-0.062356263399124146,
0.03208000585436821,
0.0212794728577137,
0.08949511498212814,
0.032752010971307755,
0.0010613800259307027,
0.010693298652768135,
0.0007162510883063078,
0.029323311522603035
] |
Scaling Offline Model-Based RL via Jointly-Optimized World-Action Model Pretraining | https://openreview.net/forum?id=T1OvCSFaum | [
"Jie Cheng",
"Ruixi Qiao",
"YINGWEI MA",
"Binhua Li",
"Gang Xiong",
"Qinghai Miao",
"Yongbin Li",
"Yisheng Lv"
] | Poster | A significant aspiration of offline reinforcement learning (RL) is to develop a generalist agent with high capabilities from large and heterogeneous datasets. However, prior approaches that scale offline RL either rely heavily on expert trajectories or struggle to generalize to diverse unseen tasks. Inspired by the excellent generalization of world model in conditional video generation, we explore the potential of image observation-based world model for scaling offline RL and enhancing generalization on novel tasks. In this paper, we introduce JOWA: Jointly-Optimized World-Action model, an offline model-based RL agent pretrained on multiple Atari games with 6 billion tokens data to learn general-purpose representation and decision-making ability. Our method jointly optimizes a world-action model through a shared transformer backbone, which stabilize temporal difference learning with large models during pretraining. Moreover, we propose a provably efficient and parallelizable planning algorithm to compensate for the Q-value estimation error and thus search out better policies. Experimental results indicate that our largest agent, with 150 million parameters, achieves 78.9% human-level performance on pretrained games using only 10% subsampled offline data, outperforming existing state-of-the-art large-scale offline RL baselines by 31.6% on averange. Furthermore, JOWA scales favorably with model capacity and can sample-efficiently transfer to novel games using only 5k offline fine-tuning data (approximately 4 trajectories) per game, demonstrating superior generalization. | reinforcement learning, offline reinforcement learning, world model | We scale offline model-based RL through a jointly-optimized world-action model pretrained across multiple games, which achieves sample-efficient transfer to novel games. | 1,467 | 2410.00564 | [
-0.047209758311510086,
-0.07128782570362091,
0.034133218228816986,
0.011640745215117931,
0.06197246164083481,
0.07480242848396301,
-0.038200993090867996,
-0.010173080489039421,
0.02336922660470009,
0.00032653106609359384,
-0.03787276893854141,
0.002291555982083082,
0.016767820343375206,
0.1451554298400879,
-0.020970655605196953,
0.03936631977558136,
0.025997783988714218,
0.049283064901828766,
-0.16262377798557281,
-0.07872902601957321,
-0.005431136582046747,
0.010248207487165928,
0.04343877360224724,
-0.021791618317365646,
0.037152741104364395,
0.015639998018741608,
0.04156455770134926,
-0.051507532596588135,
0.005333921406418085,
-0.03372330963611603,
0.05481642857193947,
-0.0578867569565773,
-0.0135642746463418,
-0.08882087469100952,
-0.057030607014894485,
0.08580116927623749,
-0.13751552999019623,
-0.022063899785280228,
-0.009034094400703907,
-0.06499462574720383,
0.024860091507434845,
0.02410830743610859,
-0.0006981014739722013,
0.002065759152173996,
0.05038346350193024,
0.020559784024953842,
-0.027409445494413376,
-0.029984906315803528,
0.0359843410551548,
-0.021392930299043655,
-0.03650698810815811,
-0.03568660840392113,
0.003127699252218008,
-0.0001819365716073662,
0.0002169259823858738,
0.058011848479509354,
0.05597444623708725,
-0.031418394297361374,
0.03126337379217148,
-0.051688309758901596,
-0.04805273562669754,
-0.08003582060337067,
-0.0016075829043984413,
-0.04202906787395477,
-0.06721042841672897,
-0.03964594006538391,
0.0704401507973671,
0.03324022516608238,
0.03142496570944786,
-0.009748673997819424,
-0.0922621563076973,
0.05364445969462395,
0.06560405343770981,
-0.010275539942085743,
0.02013939805328846,
-0.07132843881845474,
0.04983137547969818,
-0.009465289302170277,
0.042816564440727234,
-0.040359172970056534,
0.02341417595744133,
-0.036326728761196136,
-0.030279189348220825,
0.020575596019625664,
0.11697471886873245,
0.03146146982908249,
-0.04200929030776024,
0.056876495480537415,
0.07403691112995148,
-0.04144085571169853,
-0.06267306953668594,
0.019586335867643356,
0.05170949175953865,
-0.017874347046017647,
-0.021314388141036034,
0.0032625223975628614,
-0.048430684953927994,
-0.12176162004470825,
-0.026141509413719177,
0.0817677304148674,
0.03126705437898636,
0.02273678593337536,
0.07913637161254883,
0.05429038032889366,
0.06774428486824036,
-0.0167617779225111,
0.048769205808639526,
0.08406911045312881,
0.011808170937001705,
-0.07188232988119125,
0.007530217990279198,
-0.012582528404891491,
0.03566132113337517,
0.05114464834332466,
0.01323674526065588,
0.024099206551909447,
0.060926567763090134,
0.05784560739994049,
-0.09512989223003387,
0.0369006022810936,
0.07075457274913788,
-0.07158465683460236,
0.015947291627526283,
0.020597267895936966,
0.01655888743698597,
-0.02139934152364731,
-0.04753434285521507,
1.7404345481045788e-33,
0.06544815003871918,
0.012844929471611977,
0.13171085715293884,
0.06272776424884796,
0.0739256963133812,
-0.011055119335651398,
0.040352534502744675,
0.0066633280366659164,
-0.045750390738248825,
-0.00788328517228365,
0.0031785538885742426,
-0.010719693265855312,
-0.07593371719121933,
0.14174358546733856,
0.049206625670194626,
0.013880553655326366,
-0.09656728804111481,
0.01830383576452732,
0.03526681289076805,
0.016208278015255928,
0.0697147324681282,
-0.058846697211265564,
-0.017584210261702538,
-0.04829694330692291,
-0.02542608603835106,
-0.029078977182507515,
0.06391441822052002,
-0.050300922244787216,
0.0019760411232709885,
-0.03458762168884277,
-0.030398618429899216,
-0.03828093409538269,
-0.019868912175297737,
0.0732194036245346,
0.0054580275900661945,
-0.003420746885240078,
-0.10077137500047684,
-0.043959345668554306,
-0.053109098225831985,
0.014112098142504692,
0.05494195967912674,
0.024426382035017014,
-0.019397510215640068,
0.007416427135467529,
-0.07589150965213776,
-0.11381670832633972,
0.05453132465481758,
0.07936302572488785,
-0.06613410264253616,
0.026742076501250267,
0.05323924869298935,
0.01012403704226017,
-0.008108912035822868,
-0.10011225193738937,
-0.03621828928589821,
0.02287496067583561,
0.04161883518099785,
0.045637428760528564,
-0.013915854506194592,
0.003150872653350234,
-0.010356503538787365,
0.012589114718139172,
0.005114289000630379,
0.13158151507377625,
0.0018299135845154524,
0.03746819868683815,
0.008162268437445164,
0.07502530515193939,
0.011210169643163681,
0.05356983095407486,
0.005453657358884811,
0.05679383873939514,
0.10384254157543182,
-0.07176072895526886,
0.08296455442905426,
-0.027259495109319687,
0.035315606743097305,
-0.06910295784473419,
0.0065890224650502205,
0.02991827391088009,
-0.028078574687242508,
-0.015397795476019382,
-0.019408194348216057,
-0.032296277582645416,
0.032863594591617584,
-0.10305368900299072,
0.02648010291159153,
-0.04517956078052521,
-0.008651956915855408,
-0.011341704986989498,
-0.08849991858005524,
-0.02104298211634159,
-0.06598871946334839,
0.0029394349548965693,
0.013811551034450531,
-8.048795327782442e-34,
0.02508966438472271,
-0.0577419139444828,
0.038269732147455215,
0.033597420901060104,
-0.003235551994293928,
-0.08192142844200134,
-0.03189795836806297,
0.002386577660217881,
-0.010205878876149654,
-0.034663449972867966,
-0.0621732734143734,
0.037742335349321365,
0.023456035181879997,
-0.01197034865617752,
0.01200074888765812,
-0.04393036291003227,
0.02541678585112095,
-0.026508403941988945,
0.018809616565704346,
0.009938260540366173,
0.01376026775687933,
0.02650962583720684,
-0.051890116184949875,
0.0386396162211895,
-0.021442508324980736,
0.027957754209637642,
0.009976863861083984,
0.07429024577140808,
0.006985102314502001,
0.06384259462356567,
-0.04175892844796181,
-0.07333550602197647,
0.003510971087962389,
-0.04712148383259773,
-0.09635826200246811,
0.11249933391809464,
0.027750186622142792,
0.053438883274793625,
-0.09034329652786255,
0.12899444997310638,
0.01706058345735073,
-0.09087581932544708,
-0.09369079768657684,
0.030948521569371223,
-0.023789362981915474,
0.029536457732319832,
-0.061296675354242325,
-0.001921782037243247,
-0.05520041659474373,
0.005335293710231781,
-0.05466633290052414,
0.07489238679409027,
-0.08086439222097397,
-0.02708510495722294,
-0.04572088643908501,
-0.06205756217241287,
-0.03568173199892044,
0.008927215822041035,
0.04731608182191849,
-0.0034641099628061056,
0.00466674892231822,
-0.05357961729168892,
-0.08474317938089371,
0.001144113135524094,
-0.039882492274045944,
-0.042033080011606216,
0.03904065862298012,
0.040731657296419144,
0.021380145102739334,
0.003950723446905613,
-0.03175223991274834,
-0.02680799551308155,
0.03632503002882004,
-0.02263493463397026,
0.037313640117645264,
-0.05821680277585983,
-0.006742381490767002,
-0.02019539289176464,
0.07248839735984802,
-0.021923048421740532,
0.011369132436811924,
-0.02920105867087841,
0.04272995889186859,
0.053626898676157,
0.05094105005264282,
0.07305750250816345,
0.021265292540192604,
0.15522199869155884,
0.039513688534498215,
-0.00480625731870532,
-0.02690732292830944,
0.05599913373589516,
-0.006098890211433172,
0.0005518298712559044,
-0.04857730120420456,
-3.973421769387642e-8,
-0.06578046828508377,
0.03932882100343704,
0.0857551321387291,
0.006487131118774414,
0.0026160716079175472,
-0.00947930570691824,
0.05169534310698509,
0.003418148960918188,
0.03293055668473244,
0.041955623775720596,
0.040274541825056076,
0.019985156133770943,
0.009782915003597736,
-0.05934140831232071,
-0.04638567566871643,
0.02267518639564514,
0.016130171716213226,
0.0025544464588165283,
0.0040467362850904465,
0.01785391755402088,
0.02981901355087757,
0.014843028038740158,
-0.019957028329372406,
-0.08086805045604706,
0.0703294426202774,
-0.06788080930709839,
-0.01092427410185337,
0.07070150226354599,
0.05245312303304672,
-0.040589358657598495,
0.041237398982048035,
-0.03379906713962555,
0.06676793843507767,
-0.029410453513264656,
-0.00726183969527483,
0.0020309099927544594,
0.08603847026824951,
-0.005878985393792391,
0.01675586588680744,
-0.06885333359241486,
-0.004272776655852795,
0.06857013702392578,
-0.035263899713754654,
-0.0039492277428507805,
-0.0009458382264710963,
0.04910779371857643,
-0.09228023886680603,
-0.1034182533621788,
0.07185081392526627,
-0.002464079996570945,
0.027582410722970963,
-0.03854447603225708,
0.0008315686718560755,
0.056297965347766876,
0.11809257417917252,
-0.012370198965072632,
0.0587126649916172,
-0.05368364229798317,
0.09103859215974808,
0.047136735171079636,
0.02133934013545513,
-0.04462004080414772,
-0.12531223893165588,
-0.009278506971895695
] |
Deep Incomplete Multi-view Learning via Cyclic Permutation of VAEs | https://openreview.net/forum?id=s4MwstmB8o | [
"Xin Gao",
"Jian Pu"
] | Poster | Multi-View Representation Learning (MVRL) aims to derive a unified representation from multi-view data by leveraging shared and complementary information across views. However, when views are irregularly missing, the incomplete data can lead to representations that lack sufficiency and consistency. To address this, we propose Multi-View Permutation of Variational Auto-Encoders (MVP), which excavates invariant relationships between views in incomplete data. MVP establishes inter-view correspondences in the latent space of Variational Auto-Encoders, enabling the inference of missing views and the aggregation of more sufficient information. To derive a valid Evidence Lower Bound (ELBO) for learning, we apply permutations to randomly reorder variables for cross-view generation and then partition them by views to maintain invariant meanings under permutations. Additionally, we enhance consistency by introducing an informational prior with cyclic permutations of posteriors, which turns the regularization term into a similarity measure across distributions. We demonstrate the effectiveness of our approach on seven diverse datasets with varying missing ratios, achieving superior performance in multi-view clustering and generation tasks. | Multi-View Learning, Representation Learning, Multimodal VAEs, Generative Models | This paper proposes the Multi-View Permutation of VAEs (MVP), designed to learn more sufficient and consistent representations from incomplete multi-view data. | 1,466 | 2502.11037 | [
-0.018266752362251282,
-0.16271086037158966,
0.011898898519575596,
-0.04173927754163742,
0.04445816949009895,
0.0753592997789383,
-0.04861234501004219,
-0.09245054423809052,
0.06760214269161224,
-0.05648523196578026,
-0.001631511957384646,
-0.0475478358566761,
0.025256209075450897,
0.017172429710626602,
-0.08339335769414902,
0.028263330459594727,
0.07186152786016464,
0.06296069920063019,
-0.13971737027168274,
-0.04193923622369766,
-0.028929051011800766,
0.06589221209287643,
-0.007232486270368099,
0.027217086404561996,
0.0916205570101738,
0.07769183069467545,
0.007000325713306665,
-0.01905149035155773,
0.025315094739198685,
-0.03929635509848595,
0.06129808351397514,
0.03860754519701004,
-0.06597013771533966,
0.00740190502256155,
-0.02093956619501114,
0.05917138606309891,
-0.14020749926567078,
0.05568002536892891,
0.03696592152118683,
0.0178066398948431,
-0.004479167517274618,
0.09101644903421402,
-0.00532573601230979,
0.042585231363773346,
-0.013938496820628643,
-0.02354259230196476,
-0.029957303777337074,
-0.11769087612628937,
0.018066884949803352,
0.0101435212418437,
-0.0051369816064834595,
-0.008252251893281937,
-0.0524631030857563,
0.03389781713485718,
-0.003543619532138109,
-0.06189052760601044,
-0.02537807822227478,
-0.012218169867992401,
-0.013103500008583069,
0.002612535608932376,
0.028758546337485313,
-0.06552814692258835,
0.023394688963890076,
0.009135845117270947,
-0.05376242846250534,
-0.04344233125448227,
0.048819445073604584,
0.011771618388593197,
-0.03928527235984802,
0.018563034012913704,
-0.03320976719260216,
0.05773736536502838,
-0.06263046711683273,
0.05904131755232811,
0.002349252812564373,
0.08659329265356064,
0.06643085181713104,
-0.11199919134378433,
0.03654232621192932,
-0.07366831600666046,
-0.041212890297174454,
0.05280261114239693,
-0.008987666107714176,
0.03461256995797157,
0.16146790981292725,
-0.03554527834057808,
0.02152438461780548,
0.03373965248465538,
-0.02195095457136631,
0.02619423344731331,
-0.09457802027463913,
-0.021766306832432747,
-0.009587246924638748,
-0.008161783218383789,
0.012076414190232754,
0.047446511685848236,
0.04965731501579285,
-0.08165380358695984,
0.014406000263988972,
0.08597668260335922,
-0.013513307087123394,
0.07784978300333023,
0.021965138614177704,
-0.028270108625292778,
-0.0015677778283134103,
-0.04638022929430008,
0.029464727267622948,
0.06909109652042389,
-0.04318510740995407,
-0.07771074771881104,
0.0054261269979178905,
0.03154578432440758,
0.030969474464654922,
0.061088766902685165,
-0.017791742458939552,
-0.013502698391675949,
0.052512217313051224,
-0.02240240015089512,
0.07800070196390152,
0.04158470779657364,
-0.0489942841231823,
-0.06397009640932083,
0.03595530986785889,
0.060775939375162125,
-0.005350202322006226,
-0.016504978761076927,
0.01868169568479061,
8.52141281864075e-35,
-0.03485497832298279,
0.0022986535914242268,
0.0917930155992508,
0.021164346486330032,
0.06088845059275627,
0.01980539597570896,
-0.06475889682769775,
0.032886140048503876,
-0.010900218039751053,
0.0094930874183774,
0.015922723338007927,
0.0018779914826154709,
-0.00867376197129488,
0.05579268932342529,
-0.03550220653414726,
-0.009057625196874142,
-0.05026363953948021,
0.000479280948638916,
-0.06383746862411499,
-0.054111625999212265,
0.004913134966045618,
-0.021025096997618675,
-0.059813205152750015,
-0.027298014611005783,
-0.09257318824529648,
0.018952544778585434,
0.09546007215976715,
-0.04103565961122513,
-0.0618320070207119,
-0.012432430870831013,
-0.08264487236738205,
0.031171265989542007,
0.0783655121922493,
0.028566861525177956,
-0.011354182846844196,
-0.02980789542198181,
0.08575933426618576,
-0.005573939997702837,
0.035093337297439575,
-0.019637618213891983,
0.03929179534316063,
0.02710193395614624,
-0.030215639621019363,
-0.09739260375499725,
-0.03205176070332527,
-0.08297447860240936,
0.07533080130815506,
0.06826584786176682,
-0.02184627205133438,
-0.0376015305519104,
0.05310701206326485,
-0.02638721466064453,
0.008528240956366062,
-0.08951425552368164,
-0.026718780398368835,
-0.015793051570653915,
-0.03363063186407089,
0.00011570577771635726,
0.03746955841779709,
0.028429903090000153,
-0.014847866259515285,
0.09273487329483032,
-0.004971239250153303,
0.04739954322576523,
-0.013773142360150814,
0.06231140345335007,
0.030217766761779785,
0.056205764412879944,
0.07007448375225067,
-0.018338939175009727,
-0.027210963889956474,
-0.0014898956287652254,
-0.030110610648989677,
-0.0977330207824707,
0.06814023107290268,
0.05874846503138542,
-0.03129933774471283,
-0.07369323819875717,
0.02561083808541298,
0.0740288570523262,
0.0038739440497010946,
-0.018934734165668488,
0.06579463183879852,
0.022425759583711624,
-0.0416884683072567,
-0.04992496594786644,
0.039777111262083054,
-0.051937997341156006,
0.026686113327741623,
-0.018847420811653137,
0.0060180905275046825,
-0.08062703162431717,
-0.010331087745726109,
-0.06719677895307541,
0.04977647215127945,
-8.318820155490745e-34,
-0.025135910138487816,
0.027455313131213188,
0.029591267928481102,
0.013481882400810719,
0.03686726093292236,
-0.09282278269529343,
-0.03818764537572861,
0.032670069485902786,
0.011469799093902111,
-0.10304371267557144,
-0.041967250406742096,
-0.038818471133708954,
0.015702906996011734,
-0.03137485310435295,
0.026481565088033676,
0.05083860456943512,
0.012624519877135754,
0.0036301969084888697,
-0.03640415519475937,
0.015844659879803658,
0.08379893004894257,
0.040808357298374176,
-0.03196321055293083,
0.031962424516677856,
-0.0016537107294425368,
0.09840322285890579,
0.0286010280251503,
0.01787942461669445,
0.027378419414162636,
0.010548697784543037,
0.058552246540784836,
-0.08951728045940399,
0.018247032538056374,
-0.061122506856918335,
-0.10057378560304642,
0.054866187274456024,
-0.019511623308062553,
0.05505335330963135,
-0.10029733180999756,
0.10523249953985214,
0.03206576406955719,
-0.05031537637114525,
-0.004782502073794603,
0.0330970361828804,
-0.03567007556557655,
-0.007920104078948498,
-0.00036123301833868027,
-0.0156955998390913,
0.04517153650522232,
0.056331437081098557,
-0.06809815019369125,
0.06611599028110504,
-0.0871162861585617,
0.03329486399888992,
-0.07399021834135056,
-0.05875576287508011,
0.013848728500306606,
-0.00619913125410676,
0.07390569150447845,
0.003820596495643258,
0.004445584490895271,
-0.03818429633975029,
0.026834260672330856,
-0.0792880654335022,
-0.04031556099653244,
-0.0708182230591774,
-0.03166951611638069,
-0.026733452454209328,
0.04924831539392471,
-0.0003537514712661505,
0.03715599328279495,
-0.0069237640127539635,
-0.053173165768384933,
-0.08591601997613907,
0.11306537687778473,
0.01804519258439541,
0.06803064793348312,
0.023153509944677353,
0.046147409826517105,
-0.045847631990909576,
0.015390579588711262,
-0.025509679690003395,
0.01564788445830345,
0.06334417313337326,
0.0380038321018219,
0.03066033311188221,
-0.02973535656929016,
-0.007070235908031464,
0.05436486750841141,
-0.02707493118941784,
-0.03769546374678612,
0.011796257458627224,
0.01524446438997984,
0.04913719370961189,
-0.028799552470445633,
-3.6547813664356e-8,
-0.0472901351749897,
0.04883100837469101,
0.03247341886162758,
-0.015376809053122997,
0.003551498055458069,
-0.06039094924926758,
-0.02763570100069046,
0.10687437653541565,
-0.03971623629331589,
0.018968945369124413,
0.10340509563684464,
-0.013430198654532433,
-0.02652634121477604,
-0.03568825125694275,
-0.010237375274300575,
0.055531829595565796,
0.08661369979381561,
0.04424770921468735,
0.008961694315075874,
0.013690494932234287,
-0.01587630994617939,
0.0061550443060696125,
0.046708881855010986,
-0.08227789402008057,
0.055279504507780075,
-0.0511520616710186,
-0.06856190413236618,
0.015141184441745281,
0.09632638841867447,
-0.0794588029384613,
0.01593647338449955,
0.033709026873111725,
0.013277999125421047,
-0.06764011085033417,
-0.06917427480220795,
0.10146781802177429,
0.01820405013859272,
-0.02817522920668125,
-0.04024704918265343,
-0.06873904168605804,
-0.008968327194452286,
-0.026297729462385178,
0.05569576099514961,
0.06154073774814606,
0.027948696166276932,
0.025761736556887627,
0.034589897841215134,
-0.04118145629763603,
0.00917245913296938,
-0.06287156045436859,
-0.01087622344493866,
-0.08055754005908966,
0.06895145773887634,
0.06532454490661621,
0.051849380135536194,
0.02186858467757702,
0.04628250375390053,
-0.051162321120500565,
0.06221884861588478,
-0.007553932256996632,
0.004014541860669851,
0.08358870446681976,
-0.14927895367145538,
0.010556835681200027
] |
X-NeMo: Expressive Neural Motion Reenactment via Disentangled Latent Attention | https://openreview.net/forum?id=ML8FH4s5Ts | [
"XIAOCHEN ZHAO",
"Hongyi Xu",
"Guoxian Song",
"You Xie",
"Chenxu Zhang",
"Xiu Li",
"Linjie Luo",
"Jinli Suo",
"Yebin Liu"
] | Poster | We propose X-NeMo, a novel zero-shot diffusion-based portrait animation pipeline that animates a static portrait using facial movements from a driving video of a different individual. Our work first identifies the root causes of the limitations in prior approaches, such as identity leakage and difficulty in capturing subtle and extreme expressions. To address these challenges, we introduce a fully end-to-end training framework that distills a 1D identity-agnostic latent motion descriptor from driving image, effectively controlling motion through cross-attention during image generation. Our implicit motion descriptor captures expressive facial motion in fine detail, learned end-to-end from a diverse video dataset without reliance on any pre-trained motion detectors. We further disentangle motion latents from identity cues with enhanced expressiveness by supervising their learning with a dual GAN decoder, alongside spatial and color augmentations. By embedding the driving motion into a 1D latent vector and controlling motion via cross-attention instead of additive spatial guidance, our design effectively eliminates the transmission of spatial-aligned structural clues from the driving condition to the diffusion backbone, substantially mitigating identity leakage. Extensive experiments demonstrate that X-NeMo surpasses state-of-the-art baselines, producing highly expressive animations with superior identity resemblance. Our code and models will be available for research. | Portrait Animation, Head Avatar, Conditional Video Generation | null | 1,465 | null | [
-0.04409731552004814,
-0.09372448176145554,
0.05298638716340065,
0.03024025447666645,
0.08216758817434311,
-0.016303980723023415,
-0.027069803327322006,
-0.08703429996967316,
0.015533579513430595,
-0.0691511332988739,
-0.01161150075495243,
-0.12141887843608856,
-0.07332760095596313,
0.0529228039085865,
-0.015119998715817928,
0.03153081610798836,
0.07712846994400024,
0.11394423991441727,
-0.03150993585586548,
0.03285844624042511,
0.025252465158700943,
-0.031899403780698776,
0.010461276397109032,
0.013540059328079224,
0.03674495592713356,
0.011521318927407265,
0.03732440620660782,
-0.056895989924669266,
0.051400456577539444,
-0.06287698447704315,
0.05823512375354767,
-0.062106505036354065,
-0.05825784429907799,
0.06057792901992798,
-0.10273460298776627,
0.07762476801872253,
-0.0054585617035627365,
0.051293179392814636,
-0.02538793720304966,
-0.07461296766996384,
0.03739330917596817,
0.049407534301280975,
-0.010027234442532063,
0.041004180908203125,
0.04111466184258461,
-0.006604488473385572,
0.019217586144804955,
-0.05161011964082718,
0.03463289514183998,
-0.015321462415158749,
0.004764447454363108,
-0.02924191579222679,
0.03729595988988876,
0.016329094767570496,
0.016771551221609116,
0.032360754907131195,
0.052736420184373856,
-0.03538237884640694,
-0.028399638831615448,
0.02941068261861801,
0.014065741561353207,
-0.07441195100545883,
0.05424020066857338,
-0.03729063645005226,
-0.008558440022170544,
-0.05181160941720009,
0.04100882261991501,
-0.026765400543808937,
-0.042869891971349716,
0.04229167476296425,
-0.007600376848131418,
-0.03938615694642067,
-0.018976977095007896,
-0.04225806146860123,
0.022108493372797966,
0.015674155205488205,
0.031317710876464844,
-0.022557033225893974,
0.041027624160051346,
-0.025544840842485428,
0.061898037791252136,
-0.012424652464687824,
0.0361185260117054,
0.018452556803822517,
0.06176294386386871,
0.0006732743349857628,
-0.0541476272046566,
0.04510265961289406,
-0.048275720328092575,
0.07491257041692734,
-0.11173368245363235,
0.021926801651716232,
0.0239703506231308,
-0.08079663664102554,
0.10897158831357956,
-0.02366659790277481,
-0.015966249629855156,
-0.03442304581403732,
0.005672720726579428,
0.03151513636112213,
0.006010186858475208,
0.05448678508400917,
-0.039127375930547714,
0.05201730504631996,
0.030315851792693138,
-0.06769511848688126,
0.05996057391166687,
0.05836392939090729,
0.0071317777037620544,
-0.00942172296345234,
0.05642359331250191,
-0.050650253891944885,
0.0485706701874733,
0.008845370262861252,
0.050741445273160934,
-0.002746076323091984,
-0.03134487569332123,
0.045744314789772034,
0.022353656589984894,
-0.014353909529745579,
0.024989333003759384,
-0.001591315376572311,
-0.08474848419427872,
-0.00699000945314765,
0.0579758919775486,
-0.026432501152157784,
-0.04467545449733734,
-4.49688210807447e-34,
0.03164653107523918,
0.0034064240753650665,
0.029924295842647552,
0.09236510097980499,
0.022877424955368042,
-0.021329868584871292,
-0.04684319347143173,
0.0073076277039945126,
-0.11951305717229843,
0.008172831498086452,
0.03296772763133049,
-0.026346873492002487,
-0.09505171328783035,
0.1130666509270668,
-0.04635583609342575,
-0.017069756984710693,
-0.02003660425543785,
-0.024236157536506653,
0.04305536299943924,
0.05253173038363457,
-0.009429735131561756,
-0.012315800413489342,
-0.07940046489238739,
-0.04349055141210556,
-0.05660955235362053,
-0.006832735612988472,
0.047860294580459595,
-0.06566867232322693,
-0.005180508829653263,
-0.02938048541545868,
-0.1709529459476471,
0.014639613218605518,
0.031202921643853188,
0.027085598558187485,
0.05347713455557823,
-0.03352265805006027,
-0.012515916489064693,
-0.0679459199309349,
0.06561306118965149,
0.015266570262610912,
0.011475483886897564,
0.05913924798369408,
-0.06117086485028267,
-0.04166479781270027,
-0.10244984179735184,
-0.03613081946969032,
0.04760530963540077,
0.1183917447924614,
-0.03553665801882744,
0.046348653733730316,
0.04264778271317482,
0.024368878453969955,
-0.00004811316466657445,
-0.06056392565369606,
-0.05977702885866165,
-0.05216565728187561,
0.04082466661930084,
0.020947417244315147,
0.002922152867540717,
-0.00965924933552742,
0.002896314486861229,
0.004152128007262945,
-0.04440367594361305,
0.005522474646568298,
0.006201157812029123,
0.008699459955096245,
-0.04360945150256157,
0.09096477180719376,
0.08981131762266159,
0.06436812877655029,
-0.040814291685819626,
0.018790584057569504,
-0.012756964191794395,
-0.08122226595878601,
0.07710783183574677,
0.05451751500368118,
0.05829096585512161,
-0.02811436541378498,
0.04277864471077919,
0.09050870686769485,
0.026312045753002167,
0.004734222777187824,
-0.05905956029891968,
-0.00117898837197572,
-0.05815926939249039,
-0.03716307505965233,
0.011639105156064034,
-0.01289363857358694,
-0.025968901813030243,
-0.006938682869076729,
0.00956135243177414,
-0.09538926184177399,
-0.031004728749394417,
0.0345391221344471,
-0.013406624086201191,
1.395604565987807e-33,
0.031518012285232544,
0.05021039396524429,
-0.07471898943185806,
-0.024923846125602722,
0.004166676662862301,
0.03970455750823021,
0.08388447016477585,
-0.03134777396917343,
0.10218453407287598,
-0.038777969777584076,
-0.005611100234091282,
-0.04215114191174507,
0.028314201161265373,
-0.029423044994473457,
0.02709461748600006,
0.03406299650669098,
0.014080645516514778,
0.0023797699250280857,
-0.059474244713783264,
-0.024855906143784523,
0.005118117202073336,
0.13528303802013397,
-0.03295277804136276,
0.00031202062382362783,
-0.05105844512581825,
0.10665366053581238,
-0.0506599135696888,
0.07728878408670425,
-0.009083515964448452,
-0.011828242801129818,
0.02770141139626503,
0.0032026804983615875,
0.03660339117050171,
-0.03520309552550316,
-0.13296939432621002,
0.030121158808469772,
0.022503238171339035,
0.010336726903915405,
-0.015185052528977394,
0.04962616786360741,
0.05378991365432739,
-0.05212423950433731,
-0.0010880479821935296,
0.10738031566143036,
0.037070609629154205,
0.006632660515606403,
-0.04001859202980995,
-0.0243669580668211,
-0.03298244997859001,
0.12911657989025116,
-0.07325128465890884,
0.05071387439966202,
-0.10761020332574844,
0.023119375109672546,
-0.08680856227874756,
-0.07957570999860764,
-0.013587402179837227,
0.0515742301940918,
0.06929615139961243,
0.035244107246398926,
0.025175930932164192,
0.06751439720392227,
-0.05344405397772789,
-0.03968430683016777,
-0.07937823235988617,
-0.0019215630600228906,
-0.0013052785070613027,
-0.03426217660307884,
0.03867039456963539,
-0.02147395722568035,
0.06393324583768845,
0.023964937776327133,
0.0074326274916529655,
-0.04429232329130173,
-0.0005114405648782849,
-0.04470646381378174,
0.05207604542374611,
-0.00453943433240056,
0.004519266076385975,
-0.04953693225979805,
0.014105403795838356,
-0.05559327453374863,
0.03497663885354996,
0.01749754697084427,
0.06675361841917038,
0.05765072628855705,
0.02878512255847454,
0.009242619387805462,
-0.04170915484428406,
0.08105258643627167,
-0.08588476479053497,
0.060216911137104034,
-0.07447163015604019,
0.006033312063664198,
-0.010137801989912987,
-3.488640487603334e-8,
-0.14214609563350677,
0.038578279316425323,
0.016073396429419518,
-0.026059839874505997,
-0.030744729563593864,
-0.03761877119541168,
0.04778232052922249,
0.026593005284667015,
0.04056462273001671,
-0.019358953461050987,
0.09758391231298447,
-0.02239527739584446,
0.031050601974129677,
-0.07910379022359848,
-0.029374901205301285,
0.05526542663574219,
0.0028462328482419252,
0.01814141497015953,
-0.014235042035579681,
-0.000051108763727825135,
-0.05276624485850334,
0.016137341037392616,
-0.0456167533993721,
-0.11866196244955063,
-0.012868099845945835,
-0.06879124045372009,
0.015804344788193703,
0.039298221468925476,
0.06706809997558594,
-0.04215878248214722,
-0.07825218886137009,
0.02411309815943241,
0.09052443504333496,
-0.07268961519002914,
-0.021548040211200714,
0.06139034405350685,
-0.001772447139956057,
-0.007465404458343983,
0.007234529592096806,
-0.025892533361911774,
-0.028667517006397247,
0.08346573263406754,
0.027208387851715088,
0.031483057886362076,
-0.07958613336086273,
0.057446688413619995,
-0.008806339465081692,
-0.08737583458423615,
0.06956673413515091,
0.04145658761262894,
0.05285242572426796,
-0.08231895416975021,
0.04804177209734917,
0.12715965509414673,
0.08442524820566177,
0.013262290507555008,
-0.00600637448951602,
0.027047481387853622,
-0.00030967846396379173,
0.11360423266887665,
-0.02387789636850357,
0.015675047412514687,
0.030461685732007027,
-0.019818883389234543
] |
On Discriminative Probabilistic Modeling for Self-Supervised Representation Learning | https://openreview.net/forum?id=s15HrqCqbr | [
"Bokun Wang",
"Yunwen Lei",
"Yiming Ying",
"Tianbao Yang"
] | Poster | We study the discriminative probabilistic modeling on a continuous domain for the data prediction task of (multimodal) self-supervised representation learning. To address the challenge of computing the integral in the partition function for each anchor data, we leverage the multiple importance sampling (MIS) technique for robust Monte Carlo integration, which can recover InfoNCE-based contrastive loss as a special case. Within this probabilistic modeling framework, we conduct generalization error analysis to reveal the limitation of current InfoNCE-based contrastive loss for self-supervised representation learning and derive insights for developing better approaches by reducing the error of Monte Carlo integration. To this end, we propose a novel non-parametric method for approximating the sum of conditional probability densities required by MIS through convex optimization, yielding a new contrastive objective for self-supervised representation learning. Moreover, we design an efficient algorithm for solving the proposed objective. We empirically compare our algorithm to representative baselines on the contrastive image-language pretraining task. Experimental results on the CC3M and CC12M datasets demonstrate the superior overall performance of our algorithm. Our code is available at https://github.com/bokun-wang/NUCLR. | Discriminative Probabilistic Modeling; Self-Supervised Representation Learning; Multiple Importance Sampling | null | 1,457 | 2410.09156 | [
0.0011525494046509266,
-0.10922059416770935,
0.002385955536738038,
0.00906461663544178,
0.08229438215494156,
-0.017795255407691002,
0.0620139054954052,
0.0007418301538564265,
-0.01268461998552084,
-0.05676603317260742,
-0.06115878000855446,
-0.0871635228395462,
0.01842590980231762,
0.09089571237564087,
-0.03929411247372627,
-0.009534253738820553,
0.11217613518238068,
0.11705754697322845,
-0.13083647191524506,
-0.07288437336683273,
-0.01825440116226673,
0.03723125532269478,
0.05944210663437843,
-0.01697547733783722,
0.08168227225542068,
-0.044068560004234314,
0.08628561347723007,
-0.0020457259379327297,
-0.00501134991645813,
-0.018163910135626793,
0.024209387600421906,
-0.050729699432849884,
0.04017595574259758,
-0.012010253965854645,
-0.028424635529518127,
0.005375871900469065,
-0.08657317608594894,
0.03458492457866669,
0.027670826762914658,
-0.008267547003924847,
-0.0029112957417964935,
0.07561337947845459,
0.061228711158037186,
0.012736725620925426,
0.1016281247138977,
0.024759279564023018,
-0.005920226685702801,
-0.06155044212937355,
0.054424386471509933,
0.02547544054687023,
0.0036527710035443306,
-0.008267705328762531,
-0.05251041427254677,
0.037491608411073685,
-0.049806900322437286,
0.04179579019546509,
0.04013692960143089,
-0.02797938883304596,
0.0042847758159041405,
0.04933284968137741,
-0.00438981968909502,
-0.112618587911129,
0.02172687277197838,
-0.008091664873063564,
0.01989893801510334,
-0.0011073112254962325,
0.07483934611082077,
0.05351352319121361,
-0.023382775485515594,
0.026648491621017456,
-0.04022322595119476,
0.036901287734508514,
-0.030872665345668793,
0.039179008454084396,
0.03348039463162422,
0.04531006142497063,
0.05820031836628914,
-0.01737716794013977,
0.026955129578709602,
-0.10797343403100967,
0.004421544261276722,
0.07002705335617065,
0.017936760559678078,
0.023002544417977333,
0.13392828404903412,
0.06774861365556717,
-0.009429600089788437,
0.10884769260883331,
0.012298024259507656,
0.03449975326657295,
-0.04831942543387413,
0.004681563936173916,
-0.06879344582557678,
-0.013863908126950264,
-0.012144666165113449,
-0.004871335346251726,
0.07925507426261902,
-0.10052622109651566,
0.009547887369990349,
0.07238513231277466,
-0.012872617691755295,
0.0012017152039334178,
0.05293017253279686,
-0.001017957809381187,
0.03965414687991142,
-0.07540769129991531,
0.04188904911279678,
0.07013340294361115,
0.03327903151512146,
-0.06096614897251129,
0.029021544381976128,
-0.0647335797548294,
-0.006719756871461868,
0.0001820861507439986,
-0.012945865280926228,
0.005110647063702345,
0.07632099092006683,
-0.001307200058363378,
0.04039044678211212,
0.024217763915657997,
-0.08087410777807236,
0.005647363606840372,
0.013578562997281551,
0.023854071274399757,
-0.022545190528035164,
-0.08250924944877625,
-0.02652403898537159,
1.6484361357764033e-33,
0.02525603584945202,
-0.024027125909924507,
0.07772000133991241,
0.017720414325594902,
0.03406427800655365,
-0.05671120807528496,
-0.011737636290490627,
-0.07140491157770157,
-0.027685029432177544,
-0.009598915465176105,
0.030776171013712883,
0.01653933897614479,
-0.07925117015838623,
0.06661958992481232,
0.03593059629201889,
0.07149932533502579,
-0.0880032405257225,
0.04043470695614815,
0.003448139177635312,
-0.04584828019142151,
0.034016191959381104,
0.09088166803121567,
-0.01996852084994316,
-0.016164520755410194,
0.025444436818361282,
0.03240734711289406,
0.06284452974796295,
-0.07463830709457397,
-0.016513533890247345,
0.008551230654120445,
-0.13023029267787933,
0.04285094514489174,
0.010838919319212437,
-0.06410171091556549,
0.12432295083999634,
-0.057057786732912064,
-0.000991386128589511,
-0.012170340865850449,
-0.02742893621325493,
-0.027727050706744194,
-0.028505804017186165,
0.018184222280979156,
0.05424226075410843,
-0.0581434927880764,
-0.1162634789943695,
-0.12287481874227524,
0.07302991300821304,
0.01895095966756344,
-0.011782653629779816,
-0.0560019426047802,
0.04119395464658737,
-0.010526604019105434,
-0.008625848218798637,
-0.028494108468294144,
-0.06487569957971573,
0.0001718694402370602,
-0.007117925677448511,
0.05370405688881874,
0.02956540510058403,
-0.018841397017240524,
-0.04163103178143501,
0.045328788459300995,
0.004788623191416264,
0.0857819989323616,
-0.0008653877303004265,
0.04584013670682907,
-0.04560133069753647,
0.06476465612649918,
0.09367965161800385,
-0.031601544469594955,
-0.014375143684446812,
-0.014096079394221306,
-0.006956565193831921,
-0.13416939973831177,
0.07180622220039368,
0.05860326439142227,
0.03372066840529442,
-0.020371973514556885,
0.0016641983529552817,
0.018882133066654205,
0.004764813929796219,
-0.027209093794226646,
-0.029785165563225746,
-0.0768471360206604,
-0.015485151670873165,
0.010129706934094429,
0.06901194900274277,
-0.08198080211877823,
-0.042488936334848404,
-0.02008415386080742,
-0.02731652744114399,
0.029477184638381004,
0.0036467292811721563,
0.0059145670384168625,
-0.007298730313777924,
-1.361368935868581e-33,
-0.05193959176540375,
0.06575844436883926,
0.02900090254843235,
-0.006940438412129879,
0.0044588129967451096,
-0.062353454530239105,
0.018900500610470772,
0.03834928199648857,
0.024456825107336044,
-0.08784932643175125,
-0.04540586099028587,
-0.05785243958234787,
0.0008325971430167556,
-0.002081750426441431,
-0.03394946828484535,
0.07183893024921417,
-0.07225816696882248,
0.11590412259101868,
-0.06250856071710587,
0.043577633798122406,
0.012852808460593224,
0.07281998544931412,
-0.011144793592393398,
0.025969021022319794,
-0.03854413703083992,
0.0717848390340805,
-0.02390252612531185,
0.007299512159079313,
0.011004549451172352,
-0.06888949126005173,
-0.029299234971404076,
0.0471423864364624,
-0.028223052620887756,
-0.03202308341860771,
-0.10966536402702332,
0.0003991665143985301,
0.010165154933929443,
-0.03219585865736008,
-0.0510898195207119,
0.07720078527927399,
0.04075700417160988,
-0.009177083149552345,
-0.078497976064682,
-0.08176373690366745,
-0.027243873104453087,
-0.05304251238703728,
-0.028190376237034798,
0.02133023925125599,
0.05203484743833542,
0.050712455064058304,
-0.06987966597080231,
0.01723339408636093,
-0.07358629256486893,
0.049452412873506546,
-0.02055259235203266,
-0.09680145233869553,
0.0254269577562809,
-0.02831171452999115,
0.08595351129770279,
0.05732288956642151,
-0.055112384259700775,
-0.039700254797935486,
0.04697738215327263,
-0.04807044193148613,
-0.02680107392370701,
-0.08131323009729385,
-0.06077676638960838,
-0.0685347393155098,
0.02781427465379238,
0.020203500986099243,
0.03270895406603813,
0.006630698684602976,
0.04778162017464638,
-0.04419691115617752,
-0.017899103462696075,
-0.044883470982313156,
0.09050947427749634,
0.012307562865316868,
-0.025534238666296005,
0.02078840136528015,
0.10543929785490036,
-0.047618210315704346,
0.04213316738605499,
0.023490557447075844,
0.05994178354740143,
0.08210180699825287,
0.032627128064632416,
0.008546832017600536,
0.07791908830404282,
-0.03260347619652748,
-0.05075144022703171,
0.046418074518442154,
0.035064730793237686,
0.036454565823078156,
-0.11524880677461624,
-4.3584904574345273e-8,
-0.0756801888346672,
0.01729397289454937,
-0.005840897094458342,
-0.00013736719847656786,
0.01009447779506445,
-0.0339994877576828,
-0.04528597742319107,
0.07547280937433243,
-0.008262417279183865,
-0.013463137671351433,
0.1459544152021408,
-0.048988718539476395,
-0.02922392077744007,
-0.09799350798130035,
-0.04367709532380104,
0.05650525540113449,
0.06275075674057007,
0.02285950630903244,
0.03671683743596077,
-0.0365336574614048,
-0.03459789231419563,
-0.043887827545404434,
0.019872700795531273,
-0.023254426196217537,
0.011362911202013493,
-0.029602983966469765,
-0.017660265788435936,
0.015885530039668083,
0.05027568340301514,
0.011226780712604523,
-0.03366118296980858,
0.003441855311393738,
0.008844339288771152,
0.0026122620329260826,
0.010579532012343407,
0.04836951941251755,
0.05018705502152443,
-0.06288076192140579,
-0.03635725751519203,
-0.011697250418365002,
0.014891750179231167,
0.12064878642559052,
-0.060788460075855255,
0.047786254435777664,
0.06985321640968323,
0.018924489617347717,
0.03602780029177666,
-0.016547152772545815,
-0.0023777163587510586,
-0.019543085247278214,
0.03923249617218971,
-0.04503173753619194,
0.008749821223318577,
0.058048415929079056,
0.08814739435911179,
0.016711626201868057,
0.02869337983429432,
0.04851652309298515,
0.011227058246731758,
0.07148311287164688,
0.0001570228923810646,
0.028979511931538582,
-0.1007758155465126,
0.008659303188323975
] |
ThermalGaussian: Thermal 3D Gaussian Splatting | https://openreview.net/forum?id=ybFRoGxZjs | [
"Rongfeng Lu",
"Hangyu Chen",
"Zunjie Zhu",
"Yuhang Qin",
"Ming Lu",
"Le zhang",
"Chenggang Yan",
"anke xue"
] | Poster | Thermography is especially valuable for the military and other users of surveillance cameras. Some recent methods based on Neural Radiance Fields (NeRF) are proposed to reconstruct the thermal scenes in 3D from a set of thermal and RGB images. However, unlike NeRF, 3D Gaussian splatting (3DGS) prevails due to its rapid training and real-time rendering. In this work, we propose ThermalGaussian, the first thermal 3DGS approach capable of rendering high-quality images in RGB and thermal modalities. We first calibrate the RGB camera and the thermal camera to ensure that both modalities are accurately aligned. Subsequently, we use the registered images to learn the multimodal 3D Gaussians. To prevent the overfitting of any single modality, we introduce several multimodal regularization constraints. We also develop smoothing constraints tailored to the physical characteristics of the thermal modality.
Besides, we contribute a real-world dataset named RGBT-Scenes, captured by a hand-hold thermal-infrared camera, facilitating future research on thermal scene reconstruction. We conduct comprehensive experiments to show that ThermalGaussian achieves photorealistic rendering of thermal images and improves the rendering quality of RGB images. With the proposed multimodal regularization constraints, we also reduced the model's storage cost by 90\%. Our project page is at https://thermalgaussian.github.io/. | 3D reconstruction; Thermal fild reconstruction; 3D Computer Vision; Machine learning approaches; | The use of infrared thermal imaging generates new perspective images and realistic 3D reconstructions, improving both thermal and color image quality while significantly reducing memory requirements. | 1,451 | 2409.07200 | [
-0.061634235084056854,
-0.09069856256246567,
-0.005650540813803673,
0.02856050617992878,
0.1031932681798935,
-0.021527647972106934,
0.030866576358675957,
-0.09179886430501938,
-0.023426657542586327,
-0.04360232874751091,
0.015900487080216408,
-0.08307848870754242,
0.02387859858572483,
0.048656363040208817,
0.006687135901302099,
-0.07135213911533356,
0.05680657550692558,
0.09649970382452011,
-0.08858036249876022,
-0.004999068565666676,
0.015353466384112835,
-0.02385217882692814,
-0.02890404872596264,
-0.022043604403734207,
0.08572419732809067,
-0.017285283654928207,
0.0699981227517128,
-0.03025575913488865,
-0.0558057501912117,
0.06464847922325134,
0.006963985040783882,
-0.03322509676218033,
-0.04863061383366585,
0.041634682565927505,
0.016708848997950554,
0.004798894748091698,
-0.021707331761717796,
-0.02309104986488819,
-0.09453170001506805,
0.02265123836696148,
-0.017424145713448524,
0.0133783845230937,
0.01316017098724842,
0.009401648305356503,
0.05633831024169922,
0.09597944468259811,
0.035641588270664215,
-0.11687721312046051,
0.0584559440612793,
-0.038931164890527725,
0.004312707576900721,
0.004989169538021088,
0.006597302854061127,
0.07511284202337265,
-0.004821443464607,
0.07465946674346924,
-0.02948511578142643,
-0.06651974469423294,
0.01951698213815689,
-0.013940439559519291,
0.03906285762786865,
-0.011260513216257095,
-0.07362976670265198,
-0.015527743846178055,
0.012560427188873291,
0.007118221838027239,
0.04151899367570877,
-0.04617886617779732,
0.050852593034505844,
0.02471093088388443,
-0.10259350389242172,
0.09980689734220505,
-0.016709551215171814,
-0.05916552245616913,
-0.050802700221538544,
0.02502579241991043,
0.037749748677015305,
-0.02884458750486374,
0.006635376717895269,
-0.06359995156526566,
0.12068276107311249,
-0.006114083342254162,
0.07618461549282074,
0.04253121837973595,
0.03557687625288963,
-0.001957961358129978,
0.028899330645799637,
0.11762293428182602,
-0.015522396191954613,
0.07419205456972122,
-0.0530753917992115,
0.04885297268629074,
-0.08833172172307968,
-0.05276283621788025,
-0.08139180392026901,
-0.056713566184043884,
0.04884801432490349,
-0.06187175586819649,
0.0953415259718895,
-0.021724572405219078,
0.00599544495344162,
-0.06829147785902023,
-0.01235760934650898,
0.09195272624492645,
0.07345213741064072,
0.00153048534411937,
0.0434272326529026,
-0.026141991838812828,
0.05615480989217758,
-0.014129753224551678,
-0.010074238292872906,
0.016394473612308502,
0.00027841448900289834,
-0.024844523519277573,
0.021949663758277893,
0.032862622290849686,
-0.05824742466211319,
0.05402016639709473,
-0.04376007243990898,
0.00862948689609766,
-0.03583686426281929,
0.03302093967795372,
-0.027143992483615875,
0.015041009522974491,
0.0022157528437674046,
-0.03309463709592819,
0.0019846586510539055,
3.5173524000968504e-33,
0.023385215550661087,
0.08694162964820862,
-0.05710102990269661,
0.031804125756025314,
0.049925681203603745,
0.06621473282575607,
-0.06699740141630173,
0.0065688882023096085,
-0.014125099405646324,
0.024262359365820885,
0.046687766909599304,
0.0023244679905474186,
-0.04285654425621033,
0.05257034674286842,
0.019791828468441963,
0.04860929399728775,
-0.06631484627723694,
0.009792613796889782,
-0.04700970649719238,
0.1063515692949295,
-0.053466010838747025,
0.011007425375282764,
0.09640008956193924,
0.033480189740657806,
-0.03293728083372116,
0.0006710575544275343,
0.045970216393470764,
0.05906721577048302,
0.025836769491434097,
-0.00808323360979557,
-0.0028076879680156708,
-0.02375839464366436,
-0.0484846830368042,
-0.01126960851252079,
0.0022161337547004223,
-0.05152006447315216,
0.011581730097532272,
-0.0628189891576767,
-0.0414741151034832,
0.04670782387256622,
0.016118252649903297,
0.11417863517999649,
-0.001881109899841249,
0.03347716107964516,
-0.05011489987373352,
0.05236111581325531,
0.07153141498565674,
0.08528728783130646,
-0.03172319382429123,
-0.012318314053118229,
-0.014958180487155914,
-0.03355376794934273,
-0.04825014993548393,
-0.10052578151226044,
-0.06347773969173431,
-0.03560784086585045,
0.05250263586640358,
0.0007134054321795702,
0.08161042630672455,
-0.054583415389060974,
-0.030738573521375656,
-0.0008179965079762042,
-0.027575574815273285,
0.00959052424877882,
-0.04962052404880524,
0.042953163385391235,
0.014158505015075207,
0.0696474090218544,
0.017518704757094383,
-0.03683660179376602,
-0.053931254893541336,
0.013380554504692554,
0.01663140393793583,
-0.030065549537539482,
0.03597177937626839,
0.0556207075715065,
0.06884396076202393,
0.009938386268913746,
-0.014876189641654491,
0.0256356168538332,
-0.13996171951293945,
0.04211539402604103,
0.05083790794014931,
-0.07333642244338989,
-0.05532871559262276,
-0.0920434519648552,
-0.05649472400546074,
-0.013061197474598885,
-0.036234788596630096,
-0.07282483577728271,
-0.11250501871109009,
-0.07971043884754181,
-0.03424261882901192,
0.0036051974166184664,
-0.08284585177898407,
-1.0177667024999453e-33,
0.058050308376550674,
0.0007593391346745193,
-0.09942595660686493,
0.0883835107088089,
-0.03806373104453087,
-0.069676473736763,
0.009352917782962322,
0.0571138858795166,
0.06372363865375519,
-0.05651388689875603,
0.0670585110783577,
-0.015676427632570267,
0.009934922680258751,
-0.02230241522192955,
0.017042098566889763,
-0.04783029854297638,
-0.02972797490656376,
0.001080437214113772,
-0.06825634837150574,
0.008973261341452599,
-0.0024471725337207317,
0.0280643068253994,
0.0014731441624462605,
0.023453189060091972,
-0.09640909731388092,
0.10625782608985901,
-0.03940100222826004,
0.010456555522978306,
-0.027278004214167595,
-0.01722436584532261,
0.02320760302245617,
-0.006096047814935446,
0.005428055766969919,
-0.08658241480588913,
-0.0181217510253191,
0.03069486655294895,
0.08329373598098755,
0.014217585325241089,
-0.003249873872846365,
0.013280368410050869,
0.0840458795428276,
0.027756236493587494,
-0.0489443875849247,
0.01912643015384674,
-0.017135556787252426,
-0.009416352026164532,
0.011836349964141846,
-0.07232218235731125,
0.04202200099825859,
0.11042580753564835,
0.003663998330011964,
-0.02875288762152195,
-0.07623506337404251,
0.06371074914932251,
0.023795323446393013,
-0.11960624903440475,
-0.1293540596961975,
0.0069185541942715645,
0.1298636794090271,
0.005123088601976633,
0.007835106924176216,
-0.03179286792874336,
-0.01762664131820202,
0.019669000059366226,
-0.021354777738451958,
0.08125461637973785,
-0.02237735688686371,
0.030259471386671066,
0.060877956449985504,
0.08791258186101913,
-0.0020700180903077126,
-0.022574441507458687,
0.033911436796188354,
0.005192611366510391,
-0.014367382042109966,
-0.027522139251232147,
0.03003409504890442,
0.0017145400634035468,
0.0768737643957138,
0.047705646604299545,
0.04423094540834427,
-0.037508390843868256,
0.045911919325590134,
0.1208447590470314,
0.11866694688796997,
0.06243668496608734,
-0.014703569933772087,
-0.00467018922790885,
0.051123280078172684,
-0.03798146918416023,
0.0012033312814310193,
0.006757297553122044,
0.05991419404745102,
0.05339210852980614,
0.023443205282092094,
-3.9178164712438956e-8,
-0.013154675252735615,
0.026358716189861298,
0.05478541925549507,
-0.00962802954018116,
0.008334750309586525,
-0.07222867012023926,
0.009005566127598286,
-0.02408336102962494,
0.0038436483591794968,
-0.056569911539554596,
0.045475952327251434,
-0.08114786446094513,
0.016341112554073334,
-0.06342039257287979,
-0.01383749395608902,
0.04871588572859764,
-0.017011644318699837,
0.01200411282479763,
0.004895315505564213,
0.04490387439727783,
-0.054963406175374985,
0.025012638419866562,
0.05999910086393356,
0.011742416769266129,
0.04164432734251022,
0.09063491225242615,
-0.0551665723323822,
-0.01948780193924904,
0.11654502153396606,
-0.02625960297882557,
-0.0272329393774271,
-0.061765167862176895,
0.07830282300710678,
-0.045227840542793274,
-0.05899539217352867,
-0.0026191335637122393,
0.009276467375457287,
0.01669665053486824,
0.03956291824579239,
-0.05049455538392067,
-0.04313003271818161,
-0.0008939630351960659,
0.0165864210575819,
0.04477376118302345,
-0.0016876856097951531,
-0.014609022997319698,
0.02761339209973812,
-0.07996512949466705,
-0.021755846217274666,
0.05138189345598221,
-0.002938122022897005,
-0.02319825068116188,
0.007058247923851013,
0.09605805575847626,
0.02242175303399563,
-0.063018299639225,
0.027500003576278687,
0.01125613134354353,
0.038897473365068436,
0.07718083262443542,
-0.04509539157152176,
-0.032856639474630356,
-0.12990117073059082,
0.009928738698363304
] |
Zero-shot forecasting of chaotic systems | https://openreview.net/forum?id=TqYjhJrp9m | [
"Yuanzhao Zhang",
"William Gilpin"
] | Poster | Time-series forecasting is a challenging problem that traditionally requires specialized models custom-trained for the specific task at hand. Recently, inspired by the success of large language models, foundation models pre-trained on vast amounts of time-series data from diverse domains have emerged as a promising candidate for general-purpose time-series forecasting. The defining characteristic of these foundation models is their ability to perform zero-shot learning, that is, forecasting a new system from limited context data without explicit re-training or fine-tuning. Here, we evaluate whether the zero-shot learning paradigm extends to the challenging task of forecasting chaotic systems. Across 135 distinct chaotic dynamical systems and $10^8$ timepoints, we find that foundation models produce competitive forecasts compared to custom-trained models (including NBEATS, TiDE, etc.), particularly when training data is limited. Interestingly, even after point forecasts fail, large foundation models are able to preserve the geometric and statistical properties of the chaotic attractors.
We attribute this success to foundation models' ability to perform in-context learning and identify context parroting as a simple mechanism used by these models to capture the long-term behavior of chaotic dynamical systems. Our results highlight the potential of foundation models as a tool for probing nonlinear and complex systems. | chaos, nonlinear dynamics, forecasting, physics, scientific machine learning | Large language models zero-shot forecast chaotic dynamics | 1,445 | 2409.15771 | [
-0.03338339179754257,
-0.06282927095890045,
0.0808890089392662,
0.03918648511171341,
0.05465264245867729,
0.018944012001156807,
-0.030341623350977898,
-0.06851474195718765,
0.11676941066980362,
-0.08267784863710403,
-0.08155176043510437,
-0.07591351121664047,
0.017045585438609123,
0.060562726110219955,
-0.07810008525848389,
-0.05276983231306076,
0.005257383454591036,
-0.01498743612319231,
-0.007033424451947212,
-0.015611574053764343,
-0.03292243927717209,
0.0006895464030094445,
0.032521430402994156,
0.03508440777659416,
0.02633013390004635,
-0.01566190831363201,
-0.07246366143226624,
0.005465073045343161,
0.014181005768477917,
0.011856945231556892,
0.08853437751531601,
0.0052469652146101,
-0.05265155807137489,
-0.03747856989502907,
-0.03719543665647507,
0.02631317637860775,
-0.0013001665938645601,
0.04619818925857544,
-0.006910474505275488,
0.03541404753923416,
0.06441190093755722,
0.010171391069889069,
0.04833373427391052,
0.04617636650800705,
0.05418545752763748,
0.05473237857222557,
0.0246112197637558,
-0.09231457859277725,
0.025865938514471054,
-0.013664603233337402,
-0.16021603345870972,
-0.021885797381401062,
-0.025900159031152725,
0.031185245141386986,
0.04106004908680916,
0.06726902723312378,
0.04085391387343407,
0.046973634511232376,
0.005184788256883621,
-0.08369757980108261,
-0.06537648290395737,
-0.08547075092792511,
-0.04254087060689926,
0.010984673164784908,
-0.017497694119811058,
0.05181717872619629,
-0.06242453306913376,
0.10588331520557404,
-0.029791751876473427,
-0.004618790931999683,
-0.027581769973039627,
0.0623239167034626,
0.005494318902492523,
0.051132358610630035,
0.03705323860049248,
0.06914126873016357,
0.009623964317142963,
-0.04744476452469826,
0.0021566334180533886,
-0.040935542434453964,
0.027272041887044907,
-0.026146167889237404,
-0.01226510014384985,
-0.06111535802483559,
0.07103265076875687,
0.05666324123740196,
0.04139464721083641,
0.027296317741274834,
0.07134424149990082,
-0.013069245964288712,
-0.015379643067717552,
-0.05208173394203186,
0.04367305338382721,
-0.02235415391623974,
0.057066891342401505,
0.11100837588310242,
-0.00342277018353343,
-0.05136644467711449,
0.002019238192588091,
0.08525191992521286,
-0.027394846081733704,
-0.016841258853673935,
0.05912931263446808,
-0.011085855774581432,
0.045256324112415314,
0.03633052855730057,
-0.017955360934138298,
0.02647053264081478,
0.011733578518033028,
-0.05948200449347496,
0.03918788582086563,
0.057644035667181015,
0.03341050446033478,
0.018366435542702675,
0.037730541080236435,
-0.039633117616176605,
-0.050010696053504944,
-0.02857087552547455,
-0.04055942967534065,
0.1300307810306549,
0.05804193764925003,
-0.005403576884418726,
-0.06337427347898483,
-0.027963021770119667,
-0.002015516860410571,
-0.018709475174546242,
-0.05823738873004913,
3.695699596865058e-34,
0.05868498608469963,
0.01356612890958786,
0.066785529255867,
-0.004627212416380644,
0.04778607189655304,
-0.033135779201984406,
-0.016817910596728325,
0.02234015241265297,
0.020061340183019638,
0.03937283530831337,
0.009589306078851223,
0.06432180106639862,
-0.0807432010769844,
0.06708525121212006,
0.0213333610445261,
-0.009646772406995296,
-0.0034543380606919527,
-0.030544884502887726,
0.005382702685892582,
-0.03154829889535904,
0.0025274751242250204,
-0.005407419521361589,
-0.045978646725416183,
-0.1005578339099884,
0.04925655946135521,
0.03379351273179054,
0.08035355806350708,
-0.047508303076028824,
-0.05031199753284454,
0.00844280980527401,
-0.10891488939523697,
0.028071340173482895,
-0.05995063856244087,
-0.00011836783960461617,
0.03444461524486542,
-0.05082573741674423,
-0.058147359639406204,
-0.00748120853677392,
-0.01248873583972454,
-0.03278157114982605,
-0.011186155490577221,
-0.003435494378209114,
-0.05905796214938164,
-0.007181267254054546,
-0.07745377719402313,
-0.09804712235927582,
0.01798199489712715,
0.04228733107447624,
-0.0672069862484932,
-0.038096025586128235,
0.08189664781093597,
-0.03194820135831833,
-0.048413924872875214,
-0.09688849747180939,
-0.04404227063059807,
0.003608913626521826,
0.03771566227078438,
-0.06495746970176697,
-0.02286348305642605,
0.03607659041881561,
-0.0024009861517697573,
-0.034541063010692596,
0.012039813213050365,
-0.007355847861617804,
0.0044318074360489845,
0.05163121223449707,
0.03518922999501228,
0.007740861736238003,
0.08037325739860535,
-0.07031764835119247,
0.01316132489591837,
-0.013588872738182545,
-0.030890140682458878,
-0.021327393129467964,
0.08399428427219391,
-0.0066447788849473,
0.10291634500026703,
-0.05162583664059639,
0.010361556895077229,
0.08294440805912018,
0.04806550219655037,
-0.06473460048437119,
0.017715344205498695,
-0.0001614236389286816,
-0.04236435517668724,
-0.027067705988883972,
0.05774115398526192,
-0.06338976323604584,
-0.09547753632068634,
-0.08431902527809143,
-0.11635464429855347,
-0.08145023882389069,
-0.0007767541101202369,
0.013957139104604721,
-0.042122114449739456,
-2.3998387571158953e-33,
-0.012898766435682774,
-0.03896922245621681,
-0.08752357214689255,
0.01170336827635765,
-0.0014560003764927387,
0.020152173936367035,
0.020291855558753014,
0.064752958714962,
-0.0074958219192922115,
-0.042859748005867004,
-0.02648855373263359,
0.00412070844322443,
0.12051212042570114,
-0.015121026895940304,
0.010840468108654022,
-0.056890521198511124,
0.04510440304875374,
-0.0033752263989299536,
0.0020271537359803915,
0.005905900616198778,
0.0009415840613655746,
0.056267593055963516,
-0.09660252928733826,
0.06359874457120895,
0.033251821994781494,
0.0017731214175000787,
-0.02477952279150486,
0.109864242374897,
-0.053663741797208786,
0.03045189008116722,
-0.059663042426109314,
-0.0753999650478363,
0.06953137367963791,
-0.04467162489891052,
-0.05270049348473549,
0.11159113049507141,
0.055439531803131104,
-0.057006340473890305,
-0.031045984476804733,
0.0749095156788826,
0.09368015825748444,
0.008549408987164497,
-0.002849976299330592,
0.007838723249733448,
-0.0434555858373642,
-0.029839487746357918,
-0.06359373778104782,
0.03770281374454498,
0.05024790018796921,
0.03142790496349335,
-0.031112493947148323,
0.04615652561187744,
-0.08535409718751907,
-0.012202988378703594,
-0.03456695005297661,
-0.023932529613375664,
0.020404569804668427,
-0.05226705223321915,
0.07660851627588272,
0.012897924520075321,
-0.04940002039074898,
0.01463677641004324,
-0.013219346292316914,
-0.033467650413513184,
0.0027221760246902704,
0.005524664185941219,
-0.061973556876182556,
-0.04865697771310806,
-0.018099183216691017,
-0.01884900964796543,
0.06085394695401192,
0.03924676030874252,
-0.014921066351234913,
-0.013067950494587421,
-0.07291413098573685,
0.013221051543951035,
-0.013518416322767735,
-0.02532837726175785,
-0.03704942762851715,
-0.024869000539183617,
0.0021566920913755894,
-0.01778459921479225,
0.06403729319572449,
0.06161102280020714,
0.08092445135116577,
0.09997401386499405,
0.10590492933988571,
-0.012155739590525627,
0.05644150450825691,
-0.0038965591229498386,
-0.006873926147818565,
0.0040303245186805725,
-0.04999219626188278,
0.11124879866838455,
-0.005351062398403883,
-4.170937373260131e-8,
-0.0968109518289566,
0.034748658537864685,
0.06651357561349869,
0.06460676342248917,
0.027895426377654076,
0.0032932835165411234,
0.04840081185102463,
0.0012052521342411637,
-0.015294884331524372,
-0.03526679426431656,
0.11532856523990631,
-0.072101891040802,
-0.061499908566474915,
-0.03238455578684807,
0.013810886070132256,
0.09310023486614227,
0.017733585089445114,
0.03906669840216637,
-0.057589393109083176,
-0.006419745273888111,
0.09283699095249176,
0.05141810327768326,
0.013723047450184822,
-0.048077840358018875,
0.04299788549542427,
-0.09499780833721161,
-0.07224761694669724,
0.07320356369018555,
0.08205638080835342,
-0.010627193376421928,
-0.01959368959069252,
-0.011571301147341728,
-0.026904549449682236,
-0.08709922432899475,
-0.12713153660297394,
0.02580198086798191,
0.015000716783106327,
-0.01903565414249897,
0.014941837638616562,
-0.03493543714284897,
-0.017933575436472893,
0.08316825330257416,
-0.05691501498222351,
-0.00479182368144393,
-0.053407203406095505,
-0.012051752768456936,
0.028756869956851006,
-0.12592114508152008,
0.06095322594046593,
0.0400237999856472,
0.09279536455869675,
-0.024085136130452156,
0.014602788724005222,
0.054607123136520386,
0.09120821952819824,
-0.003669444937258959,
0.007039347197860479,
-0.01128541212528944,
-0.028924476355314255,
0.0628981962800026,
-0.03329647332429886,
-0.008348995819687843,
-0.03719606250524521,
0.0632583498954773
] |
MIA-DPO: Multi-Image Augmented Direct Preference Optimization For Large Vision-Language Models | https://openreview.net/forum?id=f7WBRSuf9l | [
"Ziyu Liu",
"Yuhang Zang",
"Xiaoyi Dong",
"Pan Zhang",
"Yuhang Cao",
"Haodong Duan",
"Conghui He",
"Yuanjun Xiong",
"Dahua Lin",
"Jiaqi Wang"
] | Poster | Visual preference alignment involves training Large Vision-Language Models (LVLMs) to predict human preferences between visual inputs. This is typically achieved by using labeled datasets of chosen/rejected pairs and employing optimization algorithms like direct preference optimization (DPO).
Existing visual alignment methods, primarily designed for single-image scenarios, struggle to effectively handle the complexity of multi-image tasks due to the scarcity of diverse training data and the high cost of annotating chosen/rejected pairs.
We present Multi-Image Augmented Direct Preference Optimization (MIA-DPO), a visual preference alignment approach that effectively handles multi-image inputs.
MIA-DPO mitigates the scarcity of diverse multi-image training data by extending single-image data with unrelated images arranged in grid collages or pic-in-pic formats, significantly reducing the costs associated with multi-image data annotations.
Our observation reveals that attention values of LVLMs vary considerably across different images. We use attention values to identify and filter out rejected responses the model may have mistakenly focused on.
Our attention-aware selection for constructing the chosen/rejected pairs without relying on (i) human annotation, (ii) extra data, and (iii) external models or APIs.
MIA-DPO is compatible with various architectures and outperforms existing methods on five multi-image benchmarks, achieving an average performance boost of 3.0% on LLaVA-v1.5 and 4.3% on the recent InternLM-XC2.5.
Moreover, MIA-DPO has a minimal effect on the model's ability to understand single images. | Large Vision Language Models | null | 1,438 | null | [
-0.02374524064362049,
-0.09171149879693985,
0.004479453433305025,
0.010554767213761806,
0.1254291832447052,
0.03129219636321068,
-0.04353366047143936,
0.02407350018620491,
0.07772716134786606,
0.0004991439054720104,
-0.009434716776013374,
-0.10336270183324814,
0.027842899784445763,
0.0704488605260849,
0.010201312601566315,
-0.0038836393505334854,
0.04277579486370087,
0.02281634509563446,
-0.1076681837439537,
-0.002628626534715295,
-0.021858802065253258,
-0.06877337396144867,
0.06233677640557289,
-0.019102076068520546,
-0.015949944034218788,
0.016447529196739197,
0.028077250346541405,
-0.0003188690752722323,
-0.058039531111717224,
-0.07499104738235474,
-0.004937529098242521,
-0.025071101263165474,
0.011542808264493942,
0.06455989927053452,
-0.05862007662653923,
0.07235591858625412,
-0.12591056525707245,
0.009197271429002285,
-0.011303763836622238,
-0.05186286196112633,
-0.0469735711812973,
-0.0010244171135127544,
0.0227329283952713,
-0.03829875960946083,
0.06693193316459656,
0.017311273142695427,
-0.013159191235899925,
-0.06169680505990982,
0.04320457577705383,
0.012868300080299377,
-0.07765735685825348,
-0.037761762738227844,
-0.006076531019061804,
0.044221632182598114,
-0.0007629863102920353,
0.014076695777475834,
0.03834511339664459,
0.0032239998690783978,
0.03893104940652847,
-0.0012775373179465532,
-0.02117558754980564,
-0.0615369975566864,
-0.015025459229946136,
0.028184043243527412,
-0.0012906212359666824,
-0.020837431773543358,
0.056367065757513046,
-0.059701334685087204,
-0.055322758853435516,
-0.027485709637403488,
-0.04148806631565094,
0.10324206203222275,
0.035675182938575745,
0.039688341319561005,
0.06946463882923126,
0.0858379676938057,
0.03795601800084114,
-0.057357750833034515,
0.07164832949638367,
-0.0777704119682312,
0.0034214702900499105,
0.05916101485490799,
0.01735026389360428,
-0.01402277685701847,
0.1621280014514923,
0.016629667952656746,
-0.05352043733000755,
0.05519356578588486,
0.06392531841993332,
-0.028210192918777466,
-0.1172466054558754,
-0.07335341721773148,
0.02743706665933132,
0.02315906621515751,
0.010174211114645004,
0.1032705157995224,
-0.02287326380610466,
-0.018947690725326538,
-0.09812640398740768,
-0.000810490339063108,
0.044384993612766266,
0.06261207908391953,
-0.01569269225001335,
-0.022884830832481384,
0.07961094379425049,
-0.03545913100242615,
0.0550936721265316,
0.03000470995903015,
0.05870526283979416,
-0.1268942803144455,
0.014955726452171803,
-0.02374793030321598,
-0.032783180475234985,
0.044157419353723526,
-0.05725035071372986,
-0.007429815363138914,
0.03214908391237259,
-0.039759282022714615,
0.02098209783434868,
0.005265905521810055,
-0.048382122069597244,
-0.05615127831697464,
-0.007467235438525677,
-0.04776978865265846,
0.03947978839278221,
-0.016637738794088364,
-0.03468450903892517,
2.89972066853895e-33,
-0.017287321388721466,
0.04566885530948639,
0.09961406141519547,
0.00403217226266861,
0.057123832404613495,
-0.0830855593085289,
-0.047736652195453644,
0.04583613574504852,
-0.07090907543897629,
-0.018510572612285614,
-0.03175920248031616,
-0.019181406125426292,
-0.06581895798444748,
0.16382543742656708,
-0.018717268481850624,
-0.040206097066402435,
-0.008521643467247486,
0.04792937636375427,
-0.05261066555976868,
0.05857145041227341,
0.017367897555232048,
0.05066884309053421,
-0.01799086108803749,
-0.027278810739517212,
0.010082562454044819,
-0.05639536306262016,
0.0448937863111496,
-0.05956023558974266,
-0.037091221660375595,
0.005902243312448263,
-0.11422597616910934,
0.03766338527202606,
0.06302326917648315,
0.048965934664011,
0.02341853827238083,
-0.07683996856212616,
-0.06619421392679214,
-0.023766333237290382,
0.021408557891845703,
0.058430735021829605,
0.05921642854809761,
0.09334252029657364,
0.002493933541700244,
0.04366209730505943,
-0.012400424107909203,
-0.010157602839171886,
-0.002497847657650709,
0.0497998371720314,
-0.040863871574401855,
-0.06021532416343689,
0.052720628678798676,
0.03883155807852745,
-0.0376177579164505,
-0.026819828897714615,
-0.07399642467498779,
-0.023554641753435135,
0.047657132148742676,
0.02010190300643444,
0.0495322048664093,
-0.07527446001768112,
-0.051225870847702026,
0.029036762192845345,
-0.04932493716478348,
0.07621181756258011,
0.05225600302219391,
-0.017962291836738586,
0.07003125548362732,
0.028552759438753128,
0.05833149328827858,
0.02721850760281086,
-0.016202328726649284,
0.018626699224114418,
-0.04243287444114685,
-0.07644159346818924,
0.028257958590984344,
0.018761996179819107,
0.11951366066932678,
-0.07867912203073502,
0.08091478049755096,
0.006580010987818241,
-0.005430896300822496,
0.009832512587308884,
-0.013693644665181637,
-0.04405589774250984,
-0.036531802266836166,
-0.06570032984018326,
0.07250027358531952,
-0.09725259244441986,
-0.0041601466946303844,
-0.04474826157093048,
0.007058366667479277,
0.04709959402680397,
-0.05416988953948021,
-0.06438944488763809,
0.05199919641017914,
-7.6464130504139145e-34,
0.039368320256471634,
0.051263611763715744,
0.04104827716946602,
0.025178207084536552,
-0.018145393580198288,
-0.022618647664785385,
0.02685278281569481,
-0.0009600527700968087,
0.01838035136461258,
-0.07980078458786011,
0.01567430980503559,
-0.0766678974032402,
0.018590621650218964,
-0.06709109246730804,
-0.02239481545984745,
0.0062178559601306915,
-0.004799127113074064,
0.02708904817700386,
-0.038202591240406036,
0.012076309882104397,
0.03832652047276497,
0.08189206570386887,
-0.024160146713256836,
0.04811347275972366,
-0.1198817566037178,
0.08160942047834396,
0.016399545595049858,
-0.0188676156103611,
-0.010800212621688843,
-0.03740271180868149,
-0.04096108302474022,
0.023007448762655258,
-0.02278631366789341,
-0.04725639894604683,
0.0016766076441854239,
0.05829589441418648,
-0.018006937578320503,
-0.0204415675252676,
-0.033472105860710144,
0.1381700038909912,
0.023839198052883148,
-0.06324644386768341,
-0.09078237414360046,
0.02585705555975437,
-0.04660075530409813,
-0.032823704183101654,
-0.02750982716679573,
-0.07519441097974777,
0.015685535967350006,
0.09417691081762314,
-0.05246993154287338,
0.02978828363120556,
-0.06370691955089569,
0.027777526527643204,
0.002827355405315757,
-0.13640660047531128,
-0.030893810093402863,
0.02431401051580906,
0.09121251851320267,
-0.030992576852440834,
0.009885582141578197,
0.025323646143078804,
-0.0023782288189977407,
-0.01527561154216528,
0.01729828119277954,
-0.004961058031767607,
-0.01503636036068201,
-0.062962606549263,
-0.013521688990294933,
-0.015729226171970367,
0.05470570549368858,
-0.09381222724914551,
0.019720550626516342,
0.03886927291750908,
-0.009749691002070904,
-0.01845928095281124,
0.041022658348083496,
-0.0016087241237983108,
0.034810494631528854,
-0.007318796124309301,
0.016226070001721382,
-0.008091306313872337,
-0.02282930538058281,
0.07637190073728561,
0.024748774245381355,
0.12052667886018753,
0.06811772286891937,
0.04327145963907242,
0.053351595997810364,
-0.027402717620134354,
-0.04572847858071327,
0.024281077086925507,
0.06478646397590637,
0.01339590735733509,
-0.060945186764001846,
-3.9920788452718625e-8,
-0.059313010424375534,
0.012635942548513412,
0.02659095823764801,
0.0336240790784359,
-0.03991365060210228,
-0.07918386906385422,
-0.01373635046184063,
0.1303236484527588,
-0.003575074952095747,
0.019602948799729347,
0.027275921776890755,
0.10945479571819305,
-0.016443906351923943,
-0.0514371432363987,
0.007507655303925276,
0.03614388033747673,
0.01713501662015915,
0.055862654000520706,
0.004469934850931168,
0.06616892665624619,
-0.027270197868347168,
0.022069130092859268,
0.041895460337400436,
-0.026971125975251198,
0.043579280376434326,
-0.053797412663698196,
-0.03701654449105263,
0.06480257958173752,
0.06393574178218842,
-0.04158366844058037,
-0.01265067420899868,
-0.05216111242771149,
0.032333992421627045,
-0.04834840074181557,
-0.0223644208163023,
0.030266957357525826,
-0.011512266471982002,
-0.02484029345214367,
-0.0049785650335252285,
-0.03876015916466713,
0.06807318329811096,
0.033040791749954224,
0.02009553834795952,
0.017369652166962624,
0.021336296573281288,
0.019283387809991837,
0.023442985489964485,
-0.0409943126142025,
-0.010095314122736454,
-0.04188774153590202,
-0.05670507624745369,
-0.04465994983911514,
0.029889531433582306,
0.11945081502199173,
0.0095461905002594,
-0.08514243364334106,
0.008867144584655762,
-0.03137231990695,
0.13220006227493286,
0.04660854488611221,
0.05755789205431938,
-0.012266105972230434,
-0.08967946469783783,
0.006787245161831379
] |
Causal Information Prioritization for Efficient Reinforcement Learning | https://openreview.net/forum?id=nDj45w5wam | [
"Hongye Cao",
"Fan Feng",
"Tianpei Yang",
"Jing Huo",
"Yang Gao"
] | Poster | Current Reinforcement Learning (RL) methods often suffer from sample-inefficiency, resulting from blind exploration strategies that neglect causal relationships among states, actions, and rewards. Although recent causal approaches aim to address this problem, they lack grounded modeling of reward-guided causal understanding of states and actions for goal-orientation, thus impairing learning efficiency. To tackle this issue, we propose a novel method named Causal Information Prioritization (CIP) that improves sample efficiency by leveraging factored MDPs to infer causal relationships between different dimensions of states and actions with respect to rewards, enabling the prioritization of causal information. Specifically, CIP identifies and leverages causal relationships between states and rewards to execute counterfactual data augmentation to prioritize high-impact state features under the causal understanding of the environments. Moreover, CIP integrates a causality-aware empowerment learning objective, which significantly enhances the agent's execution of reward-guided actions for more efficient exploration in complex environments.
To fully assess the effectiveness of CIP, we conduct extensive experiments across $39$ tasks in $5$ diverse continuous control environments, encompassing both locomotion and manipulation skills learning with pixel-based and sparse reward settings. Experimental results demonstrate that CIP consistently outperforms existing RL methods across a wide range of scenarios. | causality, reinforcement learning, empowerment, sample efficiency | To address limitations of blind exploration and poor sample efficiency, we introduce CIP, a novel efficient RL framework that prioritizes causal information through the lens of reward feedback. | 1,433 | 2502.10097 | [
-0.01564158871769905,
-0.0340384878218174,
0.04659842327237129,
0.054789841175079346,
0.05914393812417984,
0.03228131681680679,
-0.017458394169807434,
-0.056453343480825424,
-0.00475422153249383,
0.04228725656867027,
-0.052655357867479324,
-0.04417553171515465,
0.051537852734327316,
0.09496350586414337,
0.004660779610276222,
0.017082028090953827,
0.1045195609331131,
0.07014063000679016,
-0.10821494460105896,
-0.041987448930740356,
0.01579505018889904,
-0.0542789101600647,
0.07231456786394119,
0.0100468210875988,
-0.027178049087524414,
-0.013340545818209648,
0.0012098628794774413,
-0.050782348960638046,
0.031911399215459824,
-0.013165286742150784,
0.06332582980394363,
-0.028422711417078972,
-0.05712311714887619,
-0.12457439303398132,
-0.08113348484039307,
0.09037501364946365,
-0.05126294493675232,
-0.008698003366589546,
-0.04606151953339577,
-0.04706476256251335,
-0.012444696389138699,
0.03483106568455696,
0.00028185336850583553,
0.008681580424308777,
0.02054343745112419,
0.042856939136981964,
0.019091608002781868,
-0.019487978890538216,
0.019388621672987938,
-0.04079101234674454,
-0.022251049056649208,
-0.01332456711679697,
0.0137712387368083,
-0.029011374339461327,
0.02991054579615593,
0.08205267041921616,
0.053098082542419434,
-0.03698055073618889,
-0.018233273178339005,
-0.03713913634419441,
-0.060850098729133606,
-0.07011983543634415,
-0.06823810189962387,
-0.011289959773421288,
-0.04121244698762894,
0.006990640424191952,
0.006741881836205721,
0.06754843145608902,
-0.027937354519963264,
0.0011043534614145756,
-0.010289444588124752,
0.022412080317735672,
0.030202850699424744,
-0.026706239208579063,
0.06672883778810501,
0.013294242322444916,
-0.03249244764447212,
-0.056942299008369446,
-0.01770106889307499,
-0.020640887320041656,
0.011550364084541798,
0.03994086757302284,
-0.06726991385221481,
0.07403688877820969,
0.11630262434482574,
-0.07135164737701416,
0.04217565432190895,
0.07177209854125977,
0.07594935595989227,
0.036960434168577194,
0.01213541068136692,
0.03931652009487152,
0.0032975610811263323,
-0.00559406029060483,
0.0121273472905159,
0.019067391753196716,
-0.04569530114531517,
-0.06236093118786812,
-0.08229709416627884,
0.024368684738874435,
0.05943255126476288,
0.03554927930235863,
-0.04145555943250656,
0.022373680025339127,
-0.036071475595235825,
-0.1009824126958847,
0.03064189851284027,
0.060763414949178696,
0.05078904703259468,
-0.03651103004813194,
0.010666276328265667,
-0.0381317138671875,
0.020526990294456482,
0.06814919412136078,
-0.009283699095249176,
0.020931443199515343,
0.008567821234464645,
0.0003048329381272197,
0.05507379025220871,
-0.005205563735216856,
0.018154805526137352,
-0.08365394920110703,
0.10939573496580124,
0.015434539876878262,
0.012364408932626247,
-0.08225933462381363,
-0.061851616948843,
1.082740316031724e-33,
0.071953684091568,
0.0019296326208859682,
0.0677434578537941,
-0.00829885620623827,
0.04579903930425644,
0.01742604933679104,
0.020663656294345856,
0.04258657246828079,
-0.025592423975467682,
0.03480348736047745,
-0.001639092923142016,
0.03655160218477249,
-0.09031841158866882,
0.1030949279665947,
0.04952646419405937,
-0.02774251624941826,
-0.07521414756774902,
0.11086541414260864,
-0.047799620777368546,
-0.05652445927262306,
0.05507109686732292,
-0.043888021260499954,
-0.04221121966838837,
-0.027536455541849136,
0.0760955885052681,
-0.0007106224074959755,
-0.0096869096159935,
0.012837777845561504,
-0.0440070703625679,
0.00891353003680706,
-0.08354122936725616,
0.006698355078697205,
-0.023447606712579727,
0.02148009091615677,
-0.02359108440577984,
0.014307012781500816,
-0.05495056137442589,
-0.0284286979585886,
-0.01575085148215294,
0.07281739264726639,
-0.010266296565532684,
-0.048474375158548355,
0.0606565847992897,
-0.016484957188367844,
-0.12617717683315277,
-0.12220930308103561,
0.021831005811691284,
-0.007088321726769209,
-0.04831027239561081,
0.0010962423402816057,
0.03210127353668213,
0.03450750932097435,
0.03932012990117073,
-0.13284587860107422,
-0.03870612010359764,
-0.006801339332014322,
0.030192825943231583,
-0.010664430446922779,
-0.022741539403796196,
0.008691112510859966,
-0.001649018027819693,
-0.0036145742051303387,
-0.08242753893136978,
0.08626410365104675,
-0.02271668054163456,
0.07995114475488663,
-0.0036072947550565004,
0.05478140339255333,
0.07273025065660477,
0.015235218219459057,
-0.06565632671117783,
0.04149044677615166,
-0.004210923332720995,
-0.10090664774179459,
0.13675935566425323,
-0.002963518025353551,
-0.006357716396450996,
-0.054035965353250504,
0.050568677484989166,
-0.004866461735218763,
-0.017970893532037735,
-0.030665751546621323,
-0.053065575659275055,
0.011054527945816517,
0.0038487082347273827,
-0.03904661163687706,
-0.023126283660531044,
-0.06589743494987488,
-0.008756948634982109,
-0.0393470823764801,
0.004116615280508995,
-0.013566194102168083,
-0.03501327708363533,
0.04503440484404564,
-0.018408726900815964,
-2.594640027895234e-33,
0.03644212707877159,
-0.04878012463450432,
0.06972230225801468,
-0.04004910960793495,
0.030043771490454674,
-0.01711212284862995,
0.0107169384136796,
-0.12453757226467133,
0.05267937108874321,
-0.07198740541934967,
-0.07458098977804184,
0.024999067187309265,
0.012958303093910217,
0.034222982823848724,
-0.03616319224238396,
0.029363732784986496,
-0.0022944645024836063,
0.047022800892591476,
0.00039529401692561805,
-0.012117871083319187,
0.0054239374585449696,
0.07433900982141495,
-0.024494636803865433,
-0.007332511246204376,
-0.04609338566660881,
0.004183352459222078,
0.050968464463949203,
0.05278142914175987,
0.07293370366096497,
0.08957688510417938,
-0.028205404058098793,
-0.04820084944367409,
-0.042052820324897766,
-0.055156413465738297,
-0.07810982316732407,
0.09819896519184113,
0.07351411879062653,
0.011776413768529892,
-0.1027628555893898,
0.09802599251270294,
0.08995387703180313,
-0.0059393178671598434,
-0.042802587151527405,
0.03857448324561119,
0.013302134349942207,
0.008968556299805641,
-0.06005621701478958,
0.055358946323394775,
-0.0036900839768350124,
0.07828454673290253,
-0.028638921678066254,
0.061390221118927,
-0.047056447714567184,
0.02795909158885479,
-0.06659600883722305,
-0.0988558828830719,
-0.0007855559815652668,
-0.008148262277245522,
0.08399201929569244,
-0.002009534277021885,
-0.0655360221862793,
0.03643614426255226,
-0.05606910213828087,
0.03575828671455383,
-0.02768677845597267,
0.008544955402612686,
0.03181742876768112,
0.035966020077466965,
0.03361980989575386,
0.011026119813323021,
-0.006028680596500635,
0.0506562739610672,
0.02656431868672371,
-0.12850165367126465,
0.08062844723463058,
0.03419361263513565,
0.016205156221985817,
0.02035980112850666,
-0.06066424027085304,
-0.04208631440997124,
0.08745691180229187,
-0.008440610952675343,
0.020256832242012024,
-0.03769221156835556,
-0.06434141844511032,
0.11723340302705765,
0.003958446439355612,
0.09433015435934067,
0.061513155698776245,
-0.024644462391734123,
-0.07543714344501495,
-0.0021995550487190485,
-0.015154854394495487,
0.07322466373443604,
-0.059698086231946945,
-4.3188400411509065e-8,
-0.05638522282242775,
0.07061737775802612,
0.04975472763180733,
-0.004068916197866201,
0.02885797992348671,
-0.0037235147319734097,
-0.06415627151727676,
0.0371171273291111,
0.004983507562428713,
-0.054873548448085785,
0.08255773037672043,
0.007521703839302063,
0.0449758879840374,
-0.07187188416719437,
-0.026080217212438583,
0.01818736456334591,
0.03053961507976055,
-0.04082091152667999,
0.002885946072638035,
0.0021590297110378742,
0.07424082607030869,
-0.029834017157554626,
0.00009787812450667843,
-0.06953510642051697,
0.02361489273607731,
-0.077438123524189,
-0.05111149325966835,
0.07147060334682465,
0.06923405081033707,
0.005218991544097662,
0.07258797436952591,
-0.0396563895046711,
0.009779747575521469,
0.013330851681530476,
-0.015957597643136978,
0.03945126011967659,
0.04802987724542618,
-0.09036841243505478,
-0.00890685897320509,
-0.079593226313591,
-0.03842747211456299,
0.10922440886497498,
-0.0019061530474573374,
0.034781329333782196,
-0.009325765073299408,
0.0274747796356678,
-0.07480393350124359,
-0.08145864307880402,
0.008185427635908127,
-0.04217128828167915,
-0.03443574160337448,
-0.014474672265350819,
-0.011473986320197582,
0.08267977833747864,
0.13036538660526276,
0.012260563671588898,
0.02619451843202114,
-0.050585053861141205,
0.018271399661898613,
0.06823179125785828,
0.08608453720808029,
-0.00012056627019774169,
-0.06517606973648071,
-0.002594877267256379
] |
Towards Empowerment Gain through Causal Structure Learning in Model-Based Reinforcement Learning | https://openreview.net/forum?id=vgXI1Ws0ma | [
"Hongye Cao",
"Fan Feng",
"Meng Fang",
"Shaokang Dong",
"Tianpei Yang",
"Jing Huo",
"Yang Gao"
] | Poster | In Model-Based Reinforcement Learning (MBRL), incorporating causal structures into dynamics models provides agents with a structured understanding of the environments, enabling efficient decision.
Empowerment as an intrinsic motivation enhances the ability of agents to actively control their environments by maximizing the mutual information between future states and actions.
We posit that empowerment coupled with causal understanding can improve controllability, while enhanced empowerment gain can further facilitate causal reasoning in MBRL.
To improve learning efficiency and controllability, we propose a novel framework, Empowerment through Causal Learning (ECL), where an agent with the awareness of causal dynamics models achieves empowerment-driven exploration and optimizes its causal structure for task learning.
Specifically, ECL operates by first training a causal dynamics model of the environment based on collected data. We then maximize empowerment under the causal structure for exploration, simultaneously using data gathered through exploration to update causal dynamics model to be more controllable than dense dynamics model without causal structure. In downstream task learning, an intrinsic curiosity reward is included to balance the causality, mitigating overfitting.
Importantly, ECL is method-agnostic and is capable of integrating various causal discovery methods.
We evaluate ECL combined with $3$ causal discovery methods across $6$ environments including pixel-based tasks, demonstrating its superior performance compared to other causal MBRL methods, in terms of causal discovery, sample efficiency, and asymptotic performance. | Causal RL, MBRL, Empowerment, Intrinsic Motivation | We propose a framework, Empowerment through Causal Learning , where an agent with the awareness of causal models achieves empowerment-driven exploration and utilize its structured causal perception and control for task learning. | 1,429 | null | [
-0.02066078782081604,
-0.0209562536329031,
0.025265181437134743,
0.08103811740875244,
0.04107963293790817,
-0.00822775810956955,
0.030065303668379784,
-0.013692273758351803,
-0.012795745395123959,
0.01609284244477749,
-0.06403502821922302,
-0.057100482285022736,
0.06326522678136826,
0.08109142631292343,
0.03146154806017876,
0.03048356994986534,
0.10345654934644699,
0.049843817949295044,
-0.12371926754713058,
-0.05860540270805359,
-0.029766034334897995,
-0.07079869508743286,
0.03526006266474724,
0.04825209453701973,
-0.022727366536855698,
0.018453095108270645,
-0.007162983529269695,
-0.03712799400091171,
0.05159291625022888,
-0.015742285177111626,
0.05476389825344086,
-0.06303737312555313,
-0.0521078035235405,
-0.09885542839765549,
-0.0467020645737648,
0.05440773814916611,
-0.015857022255659103,
0.012849017046391964,
-0.040005121380090714,
-0.022449713200330734,
-0.016070447862148285,
-0.007283948827534914,
0.0034021532628685236,
-0.012711937539279461,
0.003600167576223612,
0.034963276237249374,
0.02978602610528469,
-0.024698596447706223,
-0.00576047133654356,
-0.019956035539507866,
-0.033663034439086914,
-0.030999960377812386,
0.036236658692359924,
-0.058258190751075745,
0.08198484033346176,
0.021671241149306297,
0.004503064788877964,
-0.017611879855394363,
-0.026267539709806442,
-0.046217672526836395,
-0.09468843787908554,
-0.07258348912000656,
-0.06350438296794891,
-0.004148328211158514,
-0.05639245733618736,
0.02656317688524723,
0.008026842027902603,
0.12609194219112396,
-0.030634496361017227,
0.01898930035531521,
0.042379915714263916,
0.02501870132982731,
0.02821047231554985,
-0.008983032777905464,
0.06939040124416351,
-0.0002263872738694772,
-0.025308744981884956,
-0.029749037697911263,
0.04590402543544769,
-0.0034626945853233337,
0.0022964330855757,
0.09158988296985626,
-0.0738375261425972,
0.035887885838747025,
0.06002894788980484,
-0.06264466047286987,
0.05974286049604416,
0.045148998498916626,
0.08115331083536148,
0.05810660123825073,
0.04335182532668114,
-0.028063081204891205,
0.031064914539456367,
-0.007611830253154039,
0.04968325421214104,
0.048329904675483704,
-0.058748308569192886,
-0.07641129195690155,
-0.03963744640350342,
0.038635578006505966,
0.016573144122958183,
0.06536222994327545,
-0.024118678644299507,
0.037043265998363495,
-0.023506078869104385,
-0.10326046496629715,
-0.0023892184253782034,
0.07542400807142258,
0.06250165402889252,
-0.08605842292308807,
-0.017640210688114166,
-0.039147622883319855,
0.07088802754878998,
0.02217414602637291,
-0.032111331820487976,
0.01483478955924511,
-0.009043688885867596,
0.0015675583854317665,
0.019191930070519447,
-0.04789959266781807,
0.03614092990756035,
-0.0807853564620018,
0.046396706253290176,
0.048206690698862076,
0.049412861466407776,
-0.08497895300388336,
-0.029643923044204712,
5.269295772579949e-34,
0.08323147147893906,
0.021971285343170166,
0.07196425646543503,
0.06054706126451492,
0.03636758401989937,
0.03364210203289986,
0.019907962530851364,
0.044512588530778885,
-0.0753796324133873,
0.00583159364759922,
-0.0031413526739925146,
0.014715496450662613,
-0.0871727466583252,
0.03182169049978256,
0.043484073132276535,
-0.025100097060203552,
-0.07520169764757156,
0.08625011891126633,
-0.021855376660823822,
-0.08502981066703796,
0.08056986331939697,
-0.05776385962963104,
-0.021868949756026268,
-0.03951815143227577,
0.031148655340075493,
-0.009272541850805283,
0.02081829123198986,
0.03476125746965408,
-0.03559187054634094,
0.029100967571139336,
-0.010493987239897251,
-0.021951748058199883,
-0.0250723697245121,
0.06670304387807846,
-0.04780859500169754,
0.026277028024196625,
-0.10151825845241547,
-0.01826365664601326,
0.01861836388707161,
0.045573633164167404,
0.03513578698039055,
-0.04440908506512642,
0.045634303241968155,
-0.005696665495634079,
-0.1271882951259613,
-0.0697462409734726,
0.045220594853162766,
-0.0037829603534191847,
-0.09931767731904984,
-0.02753021940588951,
0.030751436948776245,
-0.01133730262517929,
0.039219725877046585,
-0.13213060796260834,
-0.0069044423289597034,
0.049478981643915176,
0.0005026329308748245,
0.021520301699638367,
-0.0632433071732521,
-0.024751286953687668,
-0.0271489005535841,
-0.0047165872529149055,
-0.0340566448867321,
0.06868010014295578,
0.03137100487947464,
0.07756970077753067,
-0.02562973089516163,
0.06252506375312805,
0.05053756386041641,
0.0197634045034647,
-0.05053066089749336,
0.03785760700702667,
-0.015971343964338303,
-0.02798117883503437,
0.06851282715797424,
-0.023852292448282242,
0.010352999903261662,
-0.07875428348779678,
0.021103840321302414,
0.058790694922208786,
-0.07193069159984589,
-0.06821788847446442,
-0.005216796416789293,
0.058788008987903595,
0.01672692596912384,
-0.036852847784757614,
-0.014727744273841381,
-0.06975721567869186,
0.016702715307474136,
-0.04194074124097824,
-0.0445999950170517,
-0.011584209278225899,
0.004714994225651026,
0.026792367920279503,
-0.04893847554922104,
-1.612825636444196e-33,
0.05282922089099884,
-0.08854381740093231,
0.03478866070508957,
-0.06515955179929733,
0.020680896937847137,
-0.015146777033805847,
0.012831523083150387,
-0.07599323242902756,
0.015282820910215378,
-0.04127141460776329,
-0.03596866503357887,
-0.0077018775045871735,
0.025310907512903214,
-0.03143847733736038,
-0.0407852940261364,
0.012178773991763592,
-0.014577222988009453,
0.004810826852917671,
0.006633330602198839,
-0.027933962643146515,
-0.11148574203252792,
0.04929075390100479,
-0.011548434384167194,
-0.015955079346895218,
0.053996775299310684,
0.010758206248283386,
-0.00268569472245872,
0.14598354697227478,
0.11481909453868866,
0.10333652049303055,
-0.09372638165950775,
-0.05362425372004509,
-0.027662647888064384,
-0.05119883641600609,
-0.09545696526765823,
0.12760911881923676,
0.09046926349401474,
0.00010611562902340665,
-0.05984031781554222,
0.07368146628141403,
-0.01718679443001747,
-0.030581116676330566,
-0.04403344914317131,
0.04784867912530899,
0.025997450575232506,
0.0347757451236248,
-0.013490456156432629,
0.08448197692632675,
0.025905674323439598,
0.05556875094771385,
-0.011888979002833366,
0.03492813557386398,
0.013227433897554874,
0.07383368909358978,
-0.05250571668148041,
-0.08884858340024948,
0.04032722860574722,
-0.015552998520433903,
0.09814085066318512,
-0.02131032757461071,
-0.0885269045829773,
0.018028611317276955,
0.01309911534190178,
0.08411099761724472,
-0.05912226065993309,
-0.028031067922711372,
-0.016550080850720406,
-0.002330140443518758,
0.013050192967057228,
0.0013979581417515874,
0.02395891584455967,
0.006144838407635689,
-0.02270827442407608,
-0.08858540654182434,
0.09771032631397247,
0.014255650341510773,
-0.019612643867731094,
-0.006328875664621592,
-0.09753502905368805,
-0.1192585825920105,
0.07088568806648254,
0.0177440382540226,
0.03462371975183487,
0.0008677422883920372,
-0.026582961902022362,
0.06665100157260895,
0.035192400217056274,
0.12516982853412628,
0.0355692021548748,
0.028940193355083466,
-0.12162469327449799,
-0.0038744728080928326,
-0.003935696091502905,
0.10554345697164536,
-0.09365225583314896,
-3.7741386904599494e-8,
-0.05286828801035881,
0.011478828266263008,
0.05739060416817665,
-0.012574366293847561,
0.10167288780212402,
0.020354604348540306,
-0.018025770783424377,
0.04687628149986267,
0.025512032210826874,
-0.0506424680352211,
0.03236347809433937,
-0.009602660313248634,
-0.00020777566533070058,
-0.02637319639325142,
-0.024630937725305557,
0.020598754286766052,
0.037882838398218155,
-0.03349674865603447,
-0.020779570564627647,
0.017983492463827133,
0.06132066994905472,
-0.04819338023662567,
-0.022751519456505775,
-0.015656793490052223,
0.07472975552082062,
-0.10906597971916199,
-0.07631136476993561,
0.04142520949244499,
0.062127672135829926,
0.0034862381871789694,
0.009642558172345161,
0.015427105128765106,
0.007856952026486397,
0.018911154940724373,
0.00452193571254611,
0.08455553650856018,
0.03922330588102341,
-0.08633369952440262,
-0.07986810058355331,
-0.05917347967624664,
-0.007898250594735146,
0.11983218789100647,
-0.02335513010621071,
0.009832928888499737,
0.021250421181321144,
0.01314628031104803,
-0.03356734290719032,
-0.06433165073394775,
-0.023118335753679276,
-0.005422407295554876,
0.012994477525353432,
-0.008161014877259731,
0.005968662444502115,
0.028925277292728424,
0.07546212524175644,
0.0432225838303566,
0.02587245963513851,
-0.018855348229408264,
0.026079237461090088,
0.08321388810873032,
0.04557902738451958,
0.018984321504831314,
-0.019863933324813843,
-0.028258752077817917
] |
Subsets and Splits