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
|
|---|---|---|---|---|---|---|---|---|---|
SmartRAG: Jointly Learn RAG-Related Tasks From the Environment Feedback | https://openreview.net/forum?id=OCd3cffulp | [
"Jingsheng Gao",
"Linxu Li",
"Ke Ji",
"Weiyuan Li",
"Yixin Lian",
"yuzhuo fu",
"Bin Dai"
] | Poster | RAG systems consist of multiple modules to work together. However, these modules are usually separately trained. We argue that a system like RAG that incorporates multiple modules should be jointly optimized to achieve optimal performance. To demonstrate this, we design a specific pipeline called SmartRAG that includes a policy network and a retriever. The policy network can serve as 1) a decision maker that decides when to retrieve, 2) a query rewriter to generate a query most suited to the retriever and 3) an answer generator that produces the final response with/without the observations. We then propose to jointly optimize the whole system using a reinforcement learning algorithm, with the reward designed to encourage the system to achieve the highest performance with minimal retrieval cost. When jointly optimized, each module can be aware of how other modules are working and thus find the best way to work together as a complete system. Empirical results demonstrate that the jointly optimized system can achieve better performance than separately optimized counterparts. | Retrieval-augmented Generation, Language Models, Reinforcement Learning | We introduce SmartRAG, a joint framework to enable an LM learn when to retrieve, what to retrieve and how to answer. | 2,389 | 2410.18141 | [
-0.08468986302614212,
0.048428844660520554,
0.022618677467107773,
0.04360567778348923,
-0.02442367747426033,
0.006719327066093683,
0.022220486775040627,
0.0002697007148526609,
-0.03107096441090107,
0.02365186996757984,
-0.05220429599285126,
-0.002096570562571287,
0.08580046147108078,
0.004550394602119923,
0.0007570028537884355,
0.0508965402841568,
0.10576952248811722,
0.08807597309350967,
-0.06488233804702759,
-0.06134718284010887,
-0.040853749960660934,
0.016804957762360573,
0.06270278990268707,
-0.03868532553315163,
-0.011948519386351109,
0.03750036656856537,
-0.033932168036699295,
-0.025838326662778854,
-0.0007602506084367633,
0.028224147856235504,
0.06482729315757751,
-0.055462419986724854,
-0.04267096519470215,
-0.0012768800370395184,
-0.07152356952428818,
0.04870963469147682,
-0.12460342049598694,
0.0028140104841440916,
-0.007751253433525562,
-0.10667991638183594,
-0.0331614650785923,
0.006768071558326483,
-0.0433223620057106,
-0.0037859862204641104,
0.04138804227113724,
0.0005843606777489185,
-0.045961685478687286,
-0.06274782121181488,
0.01564272865653038,
0.005542222876101732,
-0.08778403699398041,
-0.07459092885255814,
0.044801950454711914,
0.03269673138856888,
0.011554287746548653,
0.08965038508176804,
0.004250241909176111,
0.010127916932106018,
-0.05788774788379669,
-0.0058729387819767,
-0.07487616688013077,
-0.12189418822526932,
-0.06233604624867439,
-0.017454789951443672,
0.03545110672712326,
-0.05918708071112633,
0.025410007685422897,
0.06515283137559891,
-0.011146243661642075,
-0.05840657278895378,
-0.048671744763851166,
0.059800006449222565,
-0.03612840175628662,
0.054428450763225555,
0.01575389876961708,
0.041507214307785034,
-0.027056152001023293,
-0.043321263045072556,
0.0005873646587133408,
-0.0415368527173996,
-0.13034884631633759,
-0.005568154621869326,
-0.03656909614801407,
0.02005045861005783,
0.10987892001867294,
-0.06775045394897461,
0.006134110037237406,
0.03329523652791977,
0.06852694600820541,
-0.026645109057426453,
-0.016656504943966866,
0.015239275060594082,
0.09645577520132065,
-0.059399839490652084,
-0.033063583076000214,
0.07387813180685043,
-0.012728756293654442,
-0.03409067541360855,
-0.022133901715278625,
0.07985252141952515,
0.022761661559343338,
0.06433787941932678,
0.04872256517410278,
-0.043706733733415604,
0.018359407782554626,
-0.025552554056048393,
0.013393327593803406,
0.06468081474304199,
0.05136949196457863,
-0.06816452741622925,
-0.01967901550233364,
0.007448732852935791,
0.03228446841239929,
-0.058653537184000015,
-0.05142676457762718,
0.0459430031478405,
-0.012264924123883247,
-0.0003201517974957824,
-0.01473053079098463,
0.11456497013568878,
0.025329364463686943,
-0.03522111847996712,
0.030900580808520317,
0.011760544963181019,
0.10330606251955032,
-0.0652083158493042,
0.010179136879742146,
4.459829240439605e-34,
0.06100326403975487,
-0.014442338608205318,
-0.020491236820816994,
-0.021751768887043,
0.02553955279290676,
0.024113448336720467,
0.08753195405006409,
0.006721074692904949,
-0.009880749508738518,
-0.0808098241686821,
-0.01677023246884346,
0.02406942844390869,
0.024738870561122894,
0.05411380156874657,
0.028629528358578682,
-0.024058904498815536,
-0.0661359578371048,
0.08147835731506348,
0.010957423597574234,
0.0262870155274868,
0.01213301345705986,
-0.012517592869699001,
0.04630101099610329,
-0.037119198590517044,
0.03767646476626396,
-0.005131954792886972,
0.10359848290681839,
0.03544313833117485,
-0.03992116078734398,
0.007936772890388966,
0.02009125053882599,
-0.033045973628759384,
-0.05928020551800728,
0.0349167138338089,
0.03960098698735237,
0.031047504395246506,
-0.024007512256503105,
0.03333381190896034,
-0.0355556420981884,
0.01791972480714321,
-0.02882012166082859,
-0.005894005298614502,
0.014786493964493275,
-0.0350712314248085,
-0.05704157426953316,
-0.06372541189193726,
0.015064938925206661,
0.010050875134766102,
-0.014580263756215572,
0.022848058491945267,
0.06860791146755219,
0.01717028021812439,
0.04980338364839554,
-0.013424612581729889,
-0.012312467209994793,
0.0187879279255867,
0.008737020194530487,
0.0672137662768364,
-0.01904986798763275,
0.0668511688709259,
-0.017840584740042686,
0.008961888961493969,
-0.040663640946149826,
0.09427006542682648,
0.02217605523765087,
0.04444115236401558,
0.08410271257162094,
0.08089158684015274,
0.11142666637897491,
-0.004561666399240494,
0.0034413449466228485,
0.006561244837939739,
0.046540629118680954,
-0.0180466640740633,
0.07063177227973938,
0.014062979258596897,
-0.004626650828868151,
-0.06915809214115143,
-0.02415918931365013,
-0.05573263019323349,
0.030606044456362724,
0.010783269070088863,
-0.03898894414305687,
0.010394319891929626,
0.028864063322544098,
-0.052299581468105316,
-0.04912147670984268,
-0.053892530500888824,
-0.052303656935691833,
0.030136052519083023,
-0.0688859298825264,
0.004581535700708628,
-0.04803909733891487,
0.04252924770116806,
0.03669370710849762,
-6.507953659304741e-34,
0.08994372189044952,
-0.034093227237463,
0.009195375256240368,
0.038817279040813446,
0.048283517360687256,
0.04829958453774452,
-0.030266953632235527,
-0.06738073378801346,
-0.10869159549474716,
0.04994941130280495,
-0.030809713527560234,
0.003073625499382615,
0.047784727066755295,
-0.02438187599182129,
-0.0010576499626040459,
0.06454301625490189,
-0.034706443548202515,
-0.050102490931749344,
0.024649329483509064,
0.04473093897104263,
-0.05976073816418648,
0.14103060960769653,
-0.0657314881682396,
0.05305510759353638,
-0.02520754002034664,
0.03387901186943054,
-0.025149958208203316,
0.05658627301454544,
0.02191377989947796,
0.0227501280605793,
-0.06213689222931862,
-0.1171388328075409,
-0.04880272597074509,
-0.08677688986063004,
0.01857200264930725,
0.05183647200465202,
0.039739299565553665,
0.022835157811641693,
-0.09088283032178879,
0.1255340278148651,
0.07375223189592361,
-0.10802190005779266,
-0.04871561750769615,
0.03788698464632034,
-0.03769936040043831,
-0.03758285567164421,
-0.10551469773054123,
0.0028663810808211565,
-0.13046830892562866,
0.01769641786813736,
0.008382156491279602,
-0.0007805718341842294,
-0.09445875883102417,
-0.07216878980398178,
-0.0005922720301896334,
-0.08850988000631332,
0.03459611535072327,
0.026470623910427094,
0.03798062354326248,
0.01982148177921772,
0.0024028264451771975,
0.0038097843062132597,
0.01598351262509823,
0.06280462443828583,
-0.0025228869635611773,
0.002320303115993738,
0.026586567983031273,
-0.02057281881570816,
-0.03876360505819321,
0.011551844887435436,
0.0032752344850450754,
0.010800172574818134,
0.0647650882601738,
-0.01721935346722603,
0.09823315590620041,
0.031154559925198555,
0.004773093853145838,
-0.06776151806116104,
-0.002027801936492324,
-0.001974077196791768,
-0.04961076378822327,
0.030889291316270828,
0.0734575092792511,
0.03607131168246269,
-0.01473793014883995,
0.003499516984447837,
-0.043713394552469254,
0.06064600497484207,
0.05614842474460602,
-0.04362743720412254,
-0.007136891130357981,
0.0077689290046691895,
-0.0004020865308120847,
0.033750712871551514,
-0.029269183054566383,
-4.879571946503347e-8,
-0.007468086201697588,
0.06393148005008698,
0.016521982848644257,
0.08792801201343536,
0.031026756390929222,
0.013071168214082718,
-0.021359732374548912,
0.12688027322292328,
-0.03324279934167862,
0.027434589341282845,
0.15657001733779907,
-0.018040647730231285,
-0.020438972860574722,
-0.03681912273168564,
0.02710675075650215,
-0.03447439894080162,
0.0007795002893544734,
-0.022901158779859543,
-0.0849187821149826,
-0.06013554334640503,
0.10410855710506439,
0.014253396540880203,
-0.019085731357336044,
0.014100456610321999,
0.029708566144108772,
-0.04912499710917473,
-0.01468003448098898,
0.08167465031147003,
0.06616818904876709,
-0.008641407825052738,
-0.028172820806503296,
-0.04372255504131317,
-0.032955948263406754,
-0.059421319514513016,
0.04784790426492691,
0.03590855002403259,
0.004110542591661215,
-0.11385449767112732,
0.02748873457312584,
-0.029381778091192245,
-0.022418413311243057,
0.09177696704864502,
-0.024208584800362587,
0.030890746042132378,
0.01715676859021187,
-0.0610152967274189,
-0.06250084936618805,
-0.12001676857471466,
0.058150820434093475,
-0.08480532467365265,
0.0036286974791437387,
-0.13176919519901276,
-0.009792950935661793,
0.029718322679400444,
0.02068895660340786,
-0.019130820408463478,
0.042993996292352676,
-0.09321107715368271,
0.03266705945134163,
0.08915002644062042,
0.056012194603681564,
0.00533338775858283,
-0.07300662249326706,
0.02714165858924389
] |
Self-Supervised Diffusion Models for Electron-Aware Molecular Representation Learning | https://openreview.net/forum?id=UQ0RqfhgCk | [
"Gyoung S. Na",
"Chanyoung Park"
] | Poster | Physical properties derived from electronic distributions are essential information that determines molecular properties. However, the electron-level information is not accessible in most real-world complex molecules due to the extensive computational costs of determining uncertain electronic distributions. For this reason, existing methods for molecular property prediction have remained in regression models on simplified atom-level molecular descriptors, such as atomic structures and fingerprints. This paper proposes an efficient knowledge transfer method for electron-aware molecular representation learning. To this end, we devised a self-supervised diffusion method that estimates the electron-level information of real-world complex molecules without expensive quantum mechanical calculations. The proposed method achieved state-of-the-art prediction accuracy in the tasks of predicting molecular properties on extensive real-world molecular datasets. | Representation learning;Generative models;Molecular science;Scientific applications | null | 2,388 | null | [
-0.05101677402853966,
-0.052011869847774506,
0.009169962257146835,
0.08822935074567795,
0.0026619737036526203,
-0.0477001890540123,
0.005882325582206249,
0.017691142857074738,
0.005445480812340975,
-0.012311464175581932,
-0.0004473457229323685,
-0.07223666459321976,
-0.032716505229473114,
0.1122659221291542,
0.004075167700648308,
-0.038443516939878464,
0.031597256660461426,
0.014029699377715588,
-0.06462731212377548,
-0.046430572867393494,
0.002084473380818963,
0.03143363445997238,
0.05417977273464203,
0.012321261689066887,
0.09474727511405945,
0.09812222421169281,
0.10052403062582016,
0.015262613072991371,
-0.011295652016997337,
-0.03181089833378792,
0.10692542791366577,
-0.03168332949280739,
-0.044012296944856644,
-0.04157788306474686,
-0.05115858465433121,
0.008479353971779346,
-0.07495246827602386,
0.04243047535419464,
-0.0007697351393289864,
-0.06813211739063263,
0.04687731713056564,
0.02435494400560856,
0.010633011348545551,
0.051325298845767975,
0.048413004726171494,
0.023319996893405914,
0.010138597339391708,
0.031842369586229324,
0.038132015615701675,
-0.014470888301730156,
0.05291371792554855,
-0.024656422436237335,
-0.008024611510336399,
0.013354008086025715,
-0.012356671504676342,
0.0325472317636013,
0.04730401188135147,
-0.03901335969567299,
-0.036302488297224045,
0.035578906536102295,
0.02624276466667652,
-0.059823308140039444,
-0.002833038102835417,
0.04428354650735855,
-0.0038414469454437494,
0.056361787021160126,
0.10604395717382431,
0.0414716973900795,
-0.027036812156438828,
-0.014096039347350597,
-0.030852578580379486,
0.06634002923965454,
-0.08702512830495834,
0.04058120399713516,
0.09617923945188522,
-0.015074646100401878,
0.023691678419709206,
0.033660538494586945,
0.01635112427175045,
-0.03236810490489006,
-0.07627733051776886,
-0.02837243862450123,
0.07419635355472565,
0.05941732972860336,
0.11425859481096268,
0.04062610864639282,
0.02054535411298275,
0.08827613294124603,
-0.07480188459157944,
-0.08539502322673798,
-0.016777383163571358,
-0.042222581803798676,
-0.03347303345799446,
-0.06829371303319931,
-0.07929334789514542,
0.04457097500562668,
0.08247482776641846,
-0.024852728471159935,
0.0018540454329922795,
0.020224573090672493,
-0.03316112607717514,
0.031032679602503777,
-0.026225578039884567,
-0.019892385229468346,
0.0318492129445076,
-0.018009047955274582,
0.07773874700069427,
0.04269694164395332,
-0.027394354343414307,
-0.033636923879384995,
0.019261648878455162,
0.036316853016614914,
-0.025978179648518562,
-0.033316660672426224,
0.03248291090130806,
0.002673523500561714,
0.01711679808795452,
0.017872396856546402,
0.034579548984766006,
0.03118146024644375,
-0.0823838859796524,
-0.06525937467813492,
-0.002795830834656954,
0.05432900786399841,
0.017140842974185944,
-0.0022366801276803017,
-0.13552048802375793,
-1.1880815994321903e-33,
-0.009091820567846298,
0.01323654968291521,
0.05346471071243286,
-0.0022746468894183636,
0.005694893188774586,
-0.035665739327669144,
-0.021275624632835388,
-0.019378039985895157,
-0.033719751983881,
-0.04442520812153816,
-0.036844853311777115,
0.012023933231830597,
-0.003981024492532015,
-0.008284133858978748,
0.009889351204037666,
0.06604205071926117,
-0.04200693964958191,
0.0329902358353138,
0.027259184047579765,
-0.0843309760093689,
0.034192778170108795,
-0.013716211542487144,
0.0038288612850010395,
-0.03827446326613426,
-0.049705274403095245,
0.04879980534315109,
0.012063886970281601,
0.018588513135910034,
-0.047198377549648285,
-0.017404690384864807,
-0.036649543792009354,
0.055695004761219025,
-0.040219247341156006,
-0.0607227087020874,
0.07818878442049026,
-0.0015880282735452056,
-0.060851696878671646,
-0.02667805552482605,
0.06288261711597443,
-0.08389198035001755,
0.020654428750276566,
-0.014048013836145401,
-0.01394101232290268,
-0.022911401465535164,
-0.09538539499044418,
0.04210418835282326,
0.01719398982822895,
-0.08210176229476929,
-0.011196349747478962,
-0.02046756073832512,
-0.001333158928900957,
-0.017805390059947968,
-0.04649372026324272,
-0.06250987201929092,
-0.023643162101507187,
0.012975160963833332,
0.02133197709918022,
0.02697864919900894,
0.05185825005173683,
0.002501034177839756,
0.01164308562874794,
0.08132926374673843,
-0.015429981052875519,
0.04186957702040672,
0.006090028211474419,
0.026216143742203712,
-0.04086044430732727,
0.02137228287756443,
0.003490505740046501,
0.04830063879489899,
0.018913855776190758,
0.08412550389766693,
0.017911242321133614,
-0.10431437939405441,
0.114277683198452,
0.023358773440122604,
0.000042434505303390324,
-0.13875114917755127,
0.04548810422420502,
-0.003216358134523034,
0.04538580775260925,
-0.04613513872027397,
-0.008888829499483109,
-0.10850203037261963,
-0.05339472368359566,
-0.08276842534542084,
-0.006487619131803513,
-0.09792280942201614,
-0.10975225269794464,
-0.011062700301408768,
0.016973203048110008,
-0.09768520295619965,
0.0014411720912903547,
-0.042404551059007645,
-0.03658200800418854,
-1.1509119170462126e-33,
-0.05450967699289322,
0.06801612675189972,
0.06873659789562225,
0.01902427151799202,
0.014019823633134365,
-0.03733988106250763,
-0.08587928116321564,
0.05919695645570755,
0.09193062037229538,
0.01611975021660328,
0.06112020090222359,
0.035057567059993744,
-0.05200432240962982,
-0.05541088432073593,
-0.004755404777824879,
0.07280492782592773,
-0.024610061198472977,
-0.036879368126392365,
0.08722824603319168,
-0.027698244899511337,
-0.07555601000785828,
-0.01609206758439541,
-0.000971770437899977,
0.0761512964963913,
-0.0395374521613121,
-0.0015312428586184978,
0.00480924965813756,
0.004861464258283377,
0.03036242350935936,
-0.0022686654701828957,
-0.05652960389852524,
-0.02319747395813465,
-0.03109786845743656,
-0.012992572039365768,
-0.09994258731603622,
-0.0011656198184937239,
-0.030885756015777588,
0.004363093990832567,
0.009882539510726929,
0.02028747648000717,
-0.048428650945425034,
-0.040172405540943146,
-0.08372504264116287,
-0.02492213249206543,
0.066508449614048,
-0.04451313987374306,
-0.04362574592232704,
0.03462977707386017,
0.03894837945699692,
0.02059069648385048,
0.027706939727067947,
0.04172489047050476,
-0.003237152239307761,
-0.01828620955348015,
-0.03246000409126282,
0.042302343994379044,
-0.03034333325922489,
0.027411557734012604,
0.10600165277719498,
0.020527858287096024,
-0.07129087299108505,
-0.08105076104402542,
0.07446052134037018,
0.04109633341431618,
-0.10975177586078644,
-0.13872288167476654,
-0.0876941978931427,
0.08288650214672089,
0.04347654804587364,
-0.007024361751973629,
-0.06309354305267334,
0.013583733700215816,
0.04924899712204933,
-0.026302199810743332,
-0.00014103685680311173,
-0.12069384008646011,
0.06610604375600815,
-0.11225423216819763,
-0.05826989561319351,
0.03195125237107277,
0.056418608874082565,
0.051825132220983505,
0.006307350471615791,
-0.0024775248020887375,
0.07532630860805511,
0.03205827996134758,
0.06110022962093353,
0.008358393795788288,
0.025354871526360512,
-0.0139613663777709,
0.013640936464071274,
0.06432264298200607,
0.008790090680122375,
0.030818410217761993,
-0.052705585956573486,
-3.310806562240032e-8,
-0.017933523282408714,
-0.035585708916187286,
0.11317670345306396,
-0.02017742209136486,
0.08362443000078201,
0.08799077570438385,
0.04350656270980835,
0.06410551816225052,
-0.027332980185747147,
0.02133980579674244,
0.07854948192834854,
-0.00982846412807703,
-0.05065235495567322,
0.0010617680381983519,
0.0033521594014018774,
0.06664279848337173,
-0.010591557249426842,
-0.03938105329871178,
-0.00406627869233489,
0.03331771492958069,
-0.0012620290508493781,
-0.01109625119715929,
0.031903501600027084,
-0.019079238176345825,
0.08896752446889877,
-0.07939186692237854,
-0.02270566299557686,
0.006191573105752468,
-0.016293268650770187,
0.030781473964452744,
-0.07981634140014648,
-0.03267081454396248,
0.11778320372104645,
0.09929762035608292,
0.04796396195888519,
0.07458523660898209,
0.03073209896683693,
-0.02326619066298008,
-0.09377774596214294,
0.011900026351213455,
-0.036666013300418854,
0.06239275261759758,
-0.0665557011961937,
0.06939228624105453,
0.03927994892001152,
0.035103969275951385,
0.07622841000556946,
0.030274974182248116,
0.021334443241357803,
-0.005165048409253359,
0.09559200704097748,
-0.033400584012269974,
-0.038570139557123184,
-0.0005767603870481253,
0.018068745732307434,
0.034349385648965836,
-0.14360694587230682,
0.017412472516298294,
0.08834237605333328,
-0.020577680319547653,
0.05450824275612831,
-0.04301217570900917,
-0.021967241540551186,
0.050149738788604736
] |
HShare: Fast LLM Decoding by Hierarchical Key-Value Sharing | https://openreview.net/forum?id=Tb5PY5vwp6 | [
"Huaijin Wu",
"Lianqiang Li",
"Hantao Huang",
"Tu Yi",
"Jihang Zhang",
"Minghui Yu",
"Junchi Yan"
] | Poster | The frequent retrieval of Key-Value (KV) cache data has emerged as a significant factor contributing to the inefficiency of the inference process in large language models. Previous research has demonstrated that a small subset of critical KV cache tokens largely influences attention outcomes, leading to methods that either employ fixed sparsity patterns or dynamically select critical tokens based on the query. While dynamic sparse patterns have proven to be more effective, they introduce significant computational overhead, as critical tokens must be reselected for each self-attention computation. In this paper, we reveal substantial similarities in KV cache token criticality across neighboring queries, layers, and heads. Motivated by this insight, we propose HShare, a hierarchical KV sharing framework. HShare facilitates the sharing of critical KV cache token indices across layers, heads, and queries, which significantly reduces the computational overhead associated with query-aware dynamic token sparsity. In addition, we introduce a greedy algorithm that dynamically determines the optimal layer-level and head-level sharing configuration for the decoding phase. We evaluate the effectiveness and efficiency of HShare across various tasks using three models: LLaMA2-7b, LLaMA3-70b, and Mistral-7b. Experimental results demonstrate that HShare achieves competitive accuracy with different sharing ratios, while delivering up to an $8.6\times$ speedup in self-attention operations and a $2.7\times$ improvement in end-to-end throughput compared with FlashAttention2 and GPT-fast respectively. The source code is publicly available at ~\url{https://github.com/wuhuaijin/HShare}. | Large Language Model, Decode, Key-Value Sharing, Critical Token, Hierarchical | We propose a hierarchical critical KV cache indices sharing framework across layer, head, and query levels. | 2,383 | null | [
0.005578263662755489,
-0.032453373074531555,
0.005728057119995356,
0.040337029844522476,
0.02060193009674549,
0.01562024001032114,
0.018902717158198357,
0.008960503153502941,
0.07390911132097244,
-0.03797709569334984,
-0.02638525888323784,
-0.04835742712020874,
0.0517403669655323,
-0.05333775654435158,
0.0386631041765213,
-0.035148557275533676,
0.08543112874031067,
0.05243554338812828,
-0.10197095572948456,
-0.12898661196231842,
0.022258378565311432,
-0.008263491094112396,
0.021291742101311684,
-0.023115629330277443,
0.05060653015971184,
0.04406851530075073,
-0.04519885033369064,
0.0029197908006608486,
0.027858028188347816,
0.022573688998818398,
0.02075536549091339,
0.010059511289000511,
-0.00035371334524825215,
0.11381955444812775,
-0.04248497262597084,
0.04519364610314369,
-0.07874120026826859,
-0.028912357985973358,
-0.02750227227807045,
-0.05678839236497879,
-0.00697404844686389,
0.03169190511107445,
-0.01930702105164528,
0.07080122828483582,
0.09963269531726837,
0.06021546199917793,
0.030940532684326172,
0.009288622997701168,
-0.015885550528764725,
-0.02722684107720852,
-0.009345182217657566,
-0.02637758105993271,
0.008958930149674416,
0.07566988468170166,
-0.03879990056157112,
0.019056454300880432,
-0.036670129746198654,
0.02772652730345726,
-0.036926448345184326,
0.03311574459075928,
-0.07483046501874924,
-0.10175587236881256,
0.030334843322634697,
-0.0077178399078547955,
-0.011593316681683064,
-0.0546865276992321,
0.10108480602502823,
0.03507933393120766,
0.01877722144126892,
0.013691538944840431,
0.006270845420658588,
0.07624120265245438,
-0.026893625035881996,
0.04350240156054497,
0.005058400798588991,
0.08118999004364014,
-0.034520749002695084,
-0.09375187009572983,
0.05804181098937988,
-0.1041453629732132,
-0.015223057009279728,
0.005839597433805466,
0.023302266374230385,
0.02529175765812397,
0.1261424720287323,
-0.03157981112599373,
0.07459311187267303,
-0.02032696269452572,
0.013124014250934124,
0.03770783171057701,
-0.043324925005435944,
-0.10045792907476425,
0.06705988198518753,
-0.037448782473802567,
-0.0024165366776287556,
0.03561671823263168,
0.01573936454951763,
0.02252226136624813,
-0.01587195321917534,
0.04995295777916908,
0.025650961324572563,
0.06036192551255226,
-0.010505138896405697,
-0.05018007010221481,
0.00019570926087908447,
0.019395386800169945,
0.01076100766658783,
0.07601776719093323,
-0.007256008684635162,
-0.06408100575208664,
-0.042539555579423904,
0.09630652517080307,
0.04052925482392311,
-0.0446164533495903,
0.002733005443587899,
0.038818225264549255,
0.05808127298951149,
-0.07135274261236191,
0.048519302159547806,
0.016541795805096626,
0.03749837726354599,
-0.041957225650548935,
-0.002246126299723983,
-0.04091349616646767,
-0.038755495101213455,
-0.04338844120502472,
-0.06658787280321121,
9.957176427208847e-34,
0.054660651832818985,
0.04576254263520241,
-0.023633796721696854,
0.006989798974245787,
-0.04009426757693291,
-0.03020070679485798,
0.0618739053606987,
0.07116812467575073,
-0.09309138357639313,
-0.0057596019469201565,
-0.021540287882089615,
-0.00314462766982615,
-0.007904852740466595,
0.016597764566540718,
0.022142935544252396,
0.015061083249747753,
0.011112590320408344,
0.021571552380919456,
-0.03438638150691986,
-0.07596339285373688,
0.048930611461400986,
0.008768761530518532,
0.04426911100745201,
-0.058787811547517776,
0.05707727000117302,
0.02723546326160431,
0.06972698122262955,
-0.09382646530866623,
-0.029096851125359535,
0.03595670312643051,
-0.11307159066200256,
-0.03193136304616928,
0.00253916229121387,
0.005040501710027456,
-0.001712665893137455,
0.007075809873640537,
0.05052872747182846,
-0.024770542979240417,
0.03575606271624565,
-0.04979300871491432,
0.012328043580055237,
0.01907118409872055,
0.016294129192829132,
-0.07955525815486908,
-0.1169738918542862,
-0.04201345890760422,
-0.057081498205661774,
-0.002436540788039565,
0.006543789524585009,
-0.012373536825180054,
0.07818230241537094,
-0.017887407913804054,
-0.03151722997426987,
0.024895597249269485,
0.026312310248613358,
-0.02206716686487198,
0.03964444622397423,
0.041412778198719025,
0.05160115659236908,
0.16116344928741455,
-0.025263383984565735,
0.040114875882864,
-0.005979300942271948,
0.12959858775138855,
0.06638098508119583,
0.04171914979815483,
0.0026824313681572676,
0.09672161936759949,
0.027139704674482346,
0.004935539793223143,
0.004849825520068407,
0.01863575540482998,
0.0086599700152874,
-0.14861665666103363,
0.01738758757710457,
0.00812183041125536,
-0.009659928269684315,
-0.059025660157203674,
0.010865220800042152,
-0.02886182628571987,
0.009966399520635605,
-0.044006768614053726,
-0.001526897307485342,
0.03200168162584305,
-0.08329784870147705,
-0.04291396960616112,
0.03852434083819389,
-0.04746805503964424,
-0.09025529026985168,
0.01002689078450203,
-0.0074911159463226795,
-0.023296445608139038,
0.08802149444818497,
-0.08301999419927597,
-0.03567855805158615,
-6.174464075266535e-34,
-0.026038289070129395,
-0.040985483676195145,
-0.01930973120033741,
0.07808185368776321,
-0.036869559437036514,
-0.07070329040288925,
-0.03909663110971451,
-0.03544479236006737,
-0.08055882155895233,
-0.039539799094200134,
-0.09398265928030014,
-0.08382391184568405,
0.07965753227472305,
0.006967519409954548,
0.07111137360334396,
0.029488276690244675,
-0.021784497424960136,
0.0121100889518857,
0.0022745809983462095,
0.11130944639444351,
-0.06988104432821274,
0.08420481532812119,
-0.08976351469755173,
0.05042432248592377,
-0.024340378120541573,
0.03602980822324753,
-0.05418383330106735,
0.0348188541829586,
0.03663896024227142,
-0.025059841573238373,
-0.1025780737400055,
0.01105730701237917,
-0.04223836585879326,
-0.07993198186159134,
-0.05291781574487686,
0.05890258774161339,
-0.01513830665498972,
0.024726372212171555,
-0.06457560509443283,
0.0980972945690155,
0.06536947190761566,
-0.06339311599731445,
-0.017990637570619583,
0.0030727405101060867,
-0.013927964493632317,
-0.015450704842805862,
-0.13217855989933014,
0.027334023267030716,
-0.010443746112287045,
0.05633871257305145,
0.05532418191432953,
-0.018603160977363586,
-0.04834818094968796,
0.026007479056715965,
-0.02772713638842106,
-0.024277038872241974,
-0.11640553921461105,
0.023135023191571236,
0.011610728688538074,
-0.04984426125884056,
0.01985219120979309,
-0.05428292974829674,
-0.01925572194159031,
-0.023394810035824776,
0.024419547989964485,
-0.07437287271022797,
0.002989282365888357,
-0.08589203655719757,
-0.01864156499505043,
-0.07281143218278885,
0.05035687983036041,
-0.08468828350305557,
0.007928113453090191,
-0.03038349188864231,
0.01694902405142784,
0.038611121475696564,
0.010835481807589531,
-0.062340471893548965,
-0.0011361644137650728,
0.04757573455572128,
-0.03270379826426506,
0.07131722569465637,
-0.005543363746255636,
0.03353194147348404,
0.02001805789768696,
0.050680551677942276,
0.020630961284041405,
0.096927210688591,
0.008923075161874294,
-0.029014326632022858,
0.007551905699074268,
0.02230626903474331,
-0.023028839379549026,
0.014405466616153717,
-0.008616435341536999,
-4.1268954475981445e-8,
-0.05816267058253288,
-0.001095559448003769,
-0.02202828973531723,
0.06663840264081955,
0.03448040783405304,
-0.06953952461481094,
-0.003959899768233299,
0.13346633315086365,
-0.019346322864294052,
0.043227020651102066,
0.09727539122104645,
-0.04400501772761345,
-0.05323527753353119,
-0.0724325031042099,
0.004333461169153452,
0.0667237713932991,
0.006763726472854614,
0.01618294045329094,
-0.049293819814920425,
-0.05868549644947052,
-0.047380231320858,
0.10701056569814682,
0.019818106666207314,
-0.02195894718170166,
-0.001222469611093402,
-0.0065130083821713924,
0.025800202041864395,
0.09017089754343033,
0.11635216325521469,
-0.06313502788543701,
0.004393258132040501,
0.01942519098520279,
-0.02280486933887005,
-0.041540563106536865,
0.04811811447143555,
0.062276240438222885,
0.017771108075976372,
0.0052513726986944675,
-0.010627953335642815,
0.010079861618578434,
0.04783109948039055,
-0.0183748509734869,
-0.02926790527999401,
0.07778695970773697,
-0.002927051391452551,
0.02854122221469879,
-0.010013559833168983,
0.0005802190280519426,
0.044125910848379135,
-0.05963985621929169,
-0.010860173963010311,
-0.0802406370639801,
-0.05040545389056206,
0.04114179313182831,
-0.031793661415576935,
0.07836001366376877,
0.01911962404847145,
-0.003038109978660941,
0.08043481409549713,
0.009333810769021511,
0.10684174299240112,
0.008202430792152882,
-0.10719283670186996,
0.036991413682699203
] |
Do Vision & Language Decoders use Images and Text equally? How Self-consistent are their Explanations? | https://openreview.net/forum?id=lCasyP21Bf | [
"Letitia Parcalabescu",
"Anette Frank"
] | Poster | Vision and language model (VLM) decoders are currently the best-performing architectures on multimodal tasks. Next to answers, they are able to produce natural language explanations, either in post-hoc or CoT settings. However, it is not clear to what extent they are using the input vision and text modalities when generating answers or explanations.
In this work, we investigate if VLMs rely on their input modalities differently when they produce explanations as opposed to answers.
We also evaluate the self-consistency of VLM decoders in both post-hoc and CoT explanation settings, by extending existing unimodal tests and measures to VLM decoders. We find that most tested VLMs are less self-consistent than LLMs. Text contributions in all tested VL decoders are more important than image contributions in all examined tasks. However, when comparing explanation generation to answer generation, the contributions of images are significantly stronger for generating explanations compared to answers. This difference is even larger in CoT compared to post-hoc explanations.
Lastly, we provide an up-to-date benchmarking of state-of-the-art VL decoders on the VALSE benchmark, which before was restricted to VL encoders. We find that the tested VL decoders still struggle with most phenomena tested by VALSE. We will make our code publicly available. | interpretability, multimodality, natural language explanations, vision and language, benchmarking | We (1) measure how much VLMs use text and images when generating predictions or explanations, (2) evaluate their self-consistency in post-hoc and CoT explanation, (3) provide an update of state-of-the-art VL decoder accuracy on the VALSE benchmark. | 2,382 | 2404.18624 | [
-0.040052350610494614,
-0.05226607620716095,
0.013881820254027843,
0.0032291635870933533,
0.1497194766998291,
-0.04693635180592537,
-0.04352293536067009,
-0.017145682126283646,
0.09998967498540878,
-0.034235622733831406,
-0.04935197904706001,
-0.10948941856622696,
-0.02844052202999592,
0.03999580442905426,
0.012446562759578228,
-0.06416343152523041,
0.08891409635543823,
0.026940923184156418,
-0.09316407889127731,
-0.03997065871953964,
0.0525757372379303,
0.01367401797324419,
0.06950151175260544,
0.005323787685483694,
0.062336474657058716,
0.06644687056541443,
-0.05003712698817253,
-0.04372139647603035,
0.05353067070245743,
-0.04305659979581833,
-0.014687793329358101,
0.02852657623589039,
-0.0030649020336568356,
0.09083760529756546,
-0.06807586550712585,
0.07990909367799759,
-0.03812786564230919,
0.018913019448518753,
-0.02977268025279045,
-0.09471413493156433,
-0.12943662703037262,
0.03861597925424576,
0.02203756384551525,
-0.03500053659081459,
0.06659342348575592,
-0.008909087628126144,
0.012614086270332336,
-0.03177126124501228,
-0.062481727451086044,
-0.045560434460639954,
-0.08623798936605453,
0.001407940173521638,
-0.00757059920579195,
-0.033298105001449585,
-0.07364939898252487,
0.018101908266544342,
0.050742339342832565,
0.027957508340477943,
0.008396645076572895,
-0.024941280484199524,
-0.02114662528038025,
-0.04080037400126457,
-0.05134262889623642,
0.05430551990866661,
0.025286279618740082,
0.02206825278699398,
0.06945458054542542,
-0.040511541068553925,
-0.03443799912929535,
-0.06282968819141388,
-0.05108025670051575,
0.05249349772930145,
0.04301946610212326,
0.0046853311359882355,
0.0010463406797498465,
0.0997571051120758,
0.013438512571156025,
-0.0132130216807127,
0.037879496812820435,
-0.08055680990219116,
0.03713102266192436,
0.013158782385289669,
0.027246467769145966,
0.013146297074854374,
0.10729794204235077,
0.024703219532966614,
-0.00004961872036801651,
0.043394118547439575,
-0.032729823142290115,
0.017322178930044174,
-0.06145213171839714,
-0.08211489021778107,
0.0016213201452046633,
-0.0012266378616914153,
0.09768685698509216,
0.02313573844730854,
0.018396982923150063,
-0.06318283081054688,
-0.020489277318120003,
0.05796092748641968,
0.024513933807611465,
0.0611458458006382,
-0.012575571425259113,
-0.06050378456711769,
0.04004629701375961,
0.012649335898458958,
0.11374857276678085,
0.0304864514619112,
0.034283481538295746,
-0.06389319896697998,
0.002474686363711953,
0.04362177476286888,
0.022145558148622513,
0.02624119073152542,
0.020628733560442924,
-0.05352329835295677,
0.03687406703829765,
0.0006139326724223793,
0.04834272339940071,
-0.016349921002984047,
-0.015194038860499859,
-0.012562695890665054,
0.016148686408996582,
0.0005334098823368549,
0.0510292649269104,
-0.03485039621591568,
-0.020606473088264465,
2.6857358809671687e-33,
0.023381061851978302,
0.057484161108732224,
0.04182931408286095,
0.0830891951918602,
0.04997783899307251,
-0.020050175487995148,
-0.02701195515692234,
0.022579150274395943,
-0.0056893834844231606,
-0.007579585071653128,
-0.014905164018273354,
0.002486845478415489,
-0.053064581006765366,
0.0988917276263237,
0.05283511430025101,
0.03352820500731468,
-0.10662394762039185,
0.02618889883160591,
-0.0842328816652298,
-0.020751429721713066,
0.011950656771659851,
-0.042811036109924316,
-0.011833843775093555,
-0.05100097507238388,
0.03039957769215107,
-0.003163212211802602,
0.06045994907617569,
-0.09186318516731262,
-0.09516925364732742,
-0.004656861536204815,
-0.10485322773456573,
0.004918694030493498,
0.07601232081651688,
0.03229610249400139,
0.06348814070224762,
-0.046158451586961746,
0.014473408460617065,
-0.022502252832055092,
-0.010550334118306637,
-0.018151821568608284,
0.01692395657300949,
0.040045469999313354,
-0.05276120454072952,
-0.03543971851468086,
-0.035650093108415604,
-0.014626404270529747,
-0.03027678281068802,
0.05535893514752388,
-0.0991952046751976,
-0.027282698079943657,
0.04007690027356148,
0.10281325876712799,
-0.0378187857568264,
-0.06571409851312637,
0.061347853392362595,
-0.040465857833623886,
0.06374936550855637,
0.024774834513664246,
0.006498541682958603,
0.05631108209490776,
0.04552819952368736,
0.05501912534236908,
-0.0162078607827425,
0.09370432049036026,
0.04189562797546387,
0.04121391847729683,
-0.020184950903058052,
0.05551225692033768,
0.02097976766526699,
0.004082697909325361,
-0.0660681277513504,
0.06272470951080322,
-0.08364886045455933,
-0.03176184743642807,
0.023727349936962128,
-0.022678235545754433,
-0.017671650275588036,
-0.08313678950071335,
0.05403954163193703,
0.05453740432858467,
0.023010263219475746,
-0.015216611325740814,
-0.01934949867427349,
-0.06337360292673111,
-0.07231462001800537,
-0.05129994824528694,
0.042977459728717804,
-0.11415161192417145,
-0.029444366693496704,
-0.029246816411614418,
0.0212115366011858,
-0.06552643328905106,
0.0358252227306366,
-0.07135666906833649,
0.03930075094103813,
-2.6521529257191375e-33,
-0.04566310718655586,
-0.00550480792298913,
-0.0485098622739315,
0.0716501921415329,
-0.03020947426557541,
-0.029251381754875183,
0.03344936668872833,
0.004567862022668123,
0.06521142274141312,
-0.05900120362639427,
-0.03247620165348053,
0.015789182856678963,
-0.031670935451984406,
0.07931523770093918,
-0.00020377659529913217,
-0.06409980356693268,
0.029845593497157097,
-0.05516905337572098,
0.03449279069900513,
0.028409915044903755,
0.028544113039970398,
0.11056708544492722,
-0.057891733944416046,
-0.04031972587108612,
-0.023243112489581108,
0.10835354775190353,
-0.07268260419368744,
0.06553920358419418,
-0.02887113206088543,
-0.04884706810116768,
0.02040220983326435,
0.05747396498918533,
0.01307069230824709,
-0.04433722794055939,
0.009024509228765965,
0.042542871087789536,
-0.0077002691105008125,
-0.026474909856915474,
-0.02868858352303505,
0.1092238575220108,
0.07470645755529404,
-0.0041067637503147125,
-0.06246315315365791,
-0.00990386214107275,
-0.029827572405338287,
-0.05963430926203728,
-0.052739232778549194,
-0.059971511363983154,
0.021157292649149895,
0.045268911868333817,
0.012080871500074863,
-0.017021168023347855,
-0.11463422328233719,
-0.013574151322245598,
-0.031857483088970184,
-0.06843362003564835,
-0.07839386910200119,
-0.011854523792862892,
0.03658687695860863,
-0.02816847898066044,
-0.04649524390697479,
0.02264879085123539,
-0.02053925022482872,
0.01996159739792347,
0.028044268488883972,
-0.0072378190234303474,
-0.009325756691396236,
-0.05210795998573303,
0.07606936991214752,
-0.03381824120879173,
0.09564080089330673,
-0.10112447291612625,
-0.020178116858005524,
0.035736922174692154,
0.061865393072366714,
-0.01854485459625721,
0.03321939334273338,
-0.005540794692933559,
0.0067632379941642284,
-0.10222209990024567,
-0.03805394470691681,
-0.041579533368349075,
0.03600896894931793,
0.10557091236114502,
0.003602019278332591,
0.06795217841863632,
-0.024960318580269814,
0.007673600222915411,
-0.0015979225281625986,
0.11410292237997055,
-0.02265962027013302,
0.022390050813555717,
0.003421576227992773,
0.10522740334272385,
0.015367437154054642,
-3.927340230802656e-8,
-0.036068394780159,
0.03782806545495987,
-0.05603666603565216,
0.030683733522892,
-0.005381940398365259,
-0.0032169825863093138,
-0.036184605211019516,
0.008637980557978153,
-0.024416496977210045,
-0.057097651064395905,
0.05725841596722603,
-0.014921792782843113,
-0.06043887138366699,
-0.028240898624062538,
0.003911269828677177,
0.08346312493085861,
-0.03819599747657776,
0.014421981759369373,
-0.01499050110578537,
-0.05083228647708893,
0.06891066581010818,
0.0266940388828516,
0.004044585861265659,
-0.006429093424230814,
-0.0369776114821434,
-0.013780863955616951,
-0.06091879680752754,
0.08913848549127579,
0.02605140209197998,
-0.010890188626945019,
0.05717421695590019,
0.036440812051296234,
-0.011429836973547935,
-0.035338666290044785,
0.020115487277507782,
0.026640407741069794,
-0.004171530716121197,
-0.04571804776787758,
0.0633072704076767,
-0.0033540441654622555,
0.04753866791725159,
0.04028867185115814,
-0.06242881715297699,
0.015198418870568275,
0.041321419179439545,
-0.00042191060492768884,
-0.015829451382160187,
-0.12721852958202362,
-0.049883898347616196,
-0.12242542207241058,
-0.06522516161203384,
-0.04344220086932182,
-0.045274604111909866,
0.15363575518131256,
0.05859896168112755,
-0.022051380947232246,
0.04932112246751785,
-0.005679680500179529,
0.03157271817326546,
0.09269780665636063,
0.028617290779948235,
0.10360544174909592,
-0.02786303497850895,
-0.04139541834592819
] |
Generalizability of Neural Networks Minimizing Empirical Risk Based on Expressive Power | https://openreview.net/forum?id=8wAL9ywQNB | [
"Lijia Yu",
"Yibo Miao",
"Yifan Zhu",
"Xiao-Shan Gao",
"Lijun Zhang"
] | Poster | The primary objective of learning methods is generalization. Classic generalization bounds, based on VC-dimension or Rademacher complexity, are uniformly applicable to all networks in the hypothesis space. On the other hand, algorithm-dependent generalization bounds, like stability bounds, address more practical scenarios and provide generalization conditions for neural networks trained using SGD. However, these bounds often rely on strict assumptions, such as the NTK hypothesis or convexity of the empirical loss, which are typically not met by neural networks. In order to establish generalizability under less stringent assumptions, this paper investigates generalizability of neural networks that minimize the empirical risk. A lower bound for population accuracy is established based on the expressiveness of these networks, which indicates that with adequately large training sample and network sizes, these networks can generalize effectively. Additionally, we provide a lower bound necessary for generalization, demonstrating that, for certain data distributions, the quantity of data required to ensure generalization exceeds the network size needed to represent that distribution. Finally, we provide theoretical insights into several phenomena in deep learning, including robust overfitting, importance of over-parameterization networks, and effects of loss functions. | generalization bound, expressive power | We establish the generalization bound based on the expressive power for the network which minimize the Empirical Risk. | 2,381 | null | [
-0.06754963845014572,
-0.05228380113840103,
0.06143728271126747,
0.06184088811278343,
0.06488429009914398,
0.04398597776889801,
0.010976182296872139,
-0.005921318661421537,
-0.02020825818181038,
-0.043816350400447845,
-0.05173007771372795,
-0.06496916711330414,
0.019148239865899086,
-0.03144100680947304,
-0.10366010665893555,
-0.06311871111392975,
0.05349864065647125,
0.1117648184299469,
-0.14246144890785217,
0.051508497446775436,
0.03732930123806,
0.045843593776226044,
0.002613757038488984,
-0.006904403679072857,
0.010085304267704487,
-0.07276882976293564,
0.023179229348897934,
-0.05729980394244194,
-0.009080744348466396,
0.06077783554792404,
0.06829724460840225,
-0.06571225076913834,
0.007316496688872576,
0.01349819265305996,
0.01368678081780672,
0.01587638258934021,
-0.0794457197189331,
0.10028746724128723,
0.002263744128867984,
0.055150583386421204,
0.03494231030344963,
0.008831393904983997,
0.003428802592679858,
0.014943907968699932,
0.08255323022603989,
0.050969842821359634,
0.03082110360264778,
-0.09070234000682831,
-0.07223910093307495,
0.04372773319482803,
0.0015499948058277369,
-0.012479661963880062,
-0.06454326212406158,
0.026679545640945435,
-0.019431941211223602,
-0.07504621148109436,
-0.02674454264342785,
-0.009209848009049892,
-0.06455840915441513,
0.1009056493639946,
0.05736420303583145,
-0.12752452492713928,
-0.03956116735935211,
-0.04501420259475708,
0.036285851150751114,
-0.0028896296862512827,
0.05293796956539154,
0.04759621620178223,
-0.005397358443588018,
0.08618012815713882,
0.016312267631292343,
0.058597661554813385,
-0.10104216635227203,
0.09420466423034668,
0.03860705345869064,
0.02490989863872528,
0.05794393643736839,
0.005451807286590338,
0.031134841963648796,
-0.08288495242595673,
0.03229027986526489,
0.035659994930028915,
0.0202702097594738,
-0.003211294999346137,
0.10229340940713882,
0.05449015274643898,
0.008712907321751118,
0.02308000810444355,
0.03242355212569237,
-0.01176574919372797,
-0.03792763501405716,
0.014600110240280628,
0.013935531489551067,
0.03856077045202255,
0.0737699568271637,
-0.02707289159297943,
-0.0038774150889366865,
-0.12337663769721985,
-0.029155218973755836,
0.08129273355007172,
-0.034406181424856186,
0.020040908828377724,
0.07354456186294556,
-0.012111185118556023,
0.05639798194169998,
-0.048741552978754044,
0.010343200527131557,
-0.029034728184342384,
0.051770132035017014,
-0.061608366668224335,
0.058276981115341187,
0.08233561366796494,
-0.0031227641738951206,
-0.006913656834512949,
0.02205916866660118,
-0.02269081026315689,
-0.02107851207256317,
0.03684217110276222,
-0.040549106895923615,
0.04599597305059433,
-0.08003925532102585,
0.04273569956421852,
0.016812078654766083,
0.036806780844926834,
0.004176410846412182,
0.027613207697868347,
-0.05722527205944061,
4.4540766649602686e-33,
0.00916173867881298,
-0.02766123041510582,
0.022797517478466034,
0.026375770568847656,
0.04251238331198692,
-0.018009504303336143,
-0.02196716144680977,
-0.012478630058467388,
0.05259055271744728,
0.03729062154889107,
-0.011968155391514301,
0.004137062001973391,
-0.05781152471899986,
-0.006640682928264141,
0.06895279884338379,
0.12636896967887878,
-0.012156144715845585,
0.06900709867477417,
0.028844783082604408,
-0.1021571084856987,
0.05690612643957138,
-0.12012957781553268,
0.0022372559178620577,
-0.0208999365568161,
-0.024295032024383545,
-0.027885260060429573,
-0.02843284420669079,
0.016037052497267723,
-0.052336569875478745,
-0.030235445126891136,
-0.06952133029699326,
0.0021881391294300556,
0.08485855162143707,
-0.021039213985204697,
0.04701031744480133,
-0.005342209246009588,
-0.006431384477764368,
-0.01988084986805916,
-0.0046750521287322044,
0.053798556327819824,
-0.06087201088666916,
0.02640722505748272,
0.04554338380694389,
0.05257740244269371,
-0.01758882775902748,
-0.08831559121608734,
-0.016573844477534294,
-0.005539841018617153,
-0.08094838261604309,
-0.06796211749315262,
-0.031646113842725754,
-0.00646653026342392,
-0.026778634637594223,
-0.021300196647644043,
0.04121518135070801,
0.03634939342737198,
0.01296123955398798,
0.08916657418012619,
0.03054109215736389,
0.03403439745306969,
-0.03086533024907112,
0.026863906532526016,
0.04851704090833664,
0.04616330564022064,
-0.01731850579380989,
0.030281372368335724,
-0.011616158299148083,
0.021899884566664696,
0.05677768215537071,
-0.0289740152657032,
-0.01098792813718319,
-0.003448254195973277,
-0.06759050488471985,
-0.07367748767137527,
-0.0077693872153759,
0.0650172010064125,
0.05399271845817566,
-0.028128525242209435,
-0.010416821576654911,
0.06413322687149048,
-0.04328109323978424,
0.05866430327296257,
-0.004661085549741983,
-0.018167903646826744,
-0.051687583327293396,
-0.06267224997282028,
0.0372500978410244,
-0.014063620008528233,
0.03155415877699852,
-0.011385270394384861,
-0.10569234937429428,
0.013402728363871574,
-0.07244453579187393,
-0.05869675800204277,
-0.08168939501047134,
-4.165345130355575e-33,
-0.13949145376682281,
0.08283883333206177,
-0.07995006442070007,
0.06394127011299133,
-0.0019083680817857385,
-0.03176850825548172,
-0.04691573977470398,
-0.011708891950547695,
-0.07823293656110764,
-0.05776757374405861,
0.0019478211179375648,
-0.038198843598365784,
0.07454535365104675,
-0.02140389196574688,
-0.004853199236094952,
0.04076629504561424,
-0.009364637546241283,
0.06294409930706024,
0.057836394757032394,
-0.030775411054491997,
0.027720117941498756,
0.045085594058036804,
-0.08027683943510056,
0.05945491045713425,
0.029134593904018402,
-0.033519428223371506,
-0.1488114297389984,
0.11206737160682678,
0.00926445797085762,
-0.04294925928115845,
-0.019875530153512955,
-0.04312615096569061,
-0.0809207335114479,
0.005474851932376623,
0.030353359878063202,
0.036654140800237656,
0.049074575304985046,
0.043873585760593414,
-0.05829869210720062,
0.091925710439682,
0.0023836842738091946,
-0.03966396674513817,
-0.06175106018781662,
0.0011557984398677945,
0.025005070492625237,
-0.012660425156354904,
0.026725929230451584,
-0.018640141934156418,
0.08233992755413055,
-0.0020877334754914045,
0.038750749081373215,
-0.02140769362449646,
0.00189692503772676,
0.07908803969621658,
-0.044087935239076614,
0.0030844821594655514,
0.058255281299352646,
-0.004380057565867901,
0.08932861685752869,
0.045256227254867554,
-0.07979393005371094,
-0.0650494173169136,
-0.01701505109667778,
0.01591322384774685,
-0.07443449646234512,
-0.027678120881319046,
-0.02654425799846649,
0.05007162317633629,
0.03919747844338417,
0.00775115704163909,
0.007061162032186985,
-0.0005315892049111426,
0.02435818314552307,
0.03170132637023926,
-0.11957259476184845,
0.026987431570887566,
0.0951928123831749,
0.0006994539289735258,
-0.007290987763553858,
-0.005236592143774033,
0.07454881072044373,
0.01611492969095707,
0.07287232577800751,
-0.01989503763616085,
0.08297257125377655,
0.08791998773813248,
0.09858320653438568,
0.023864535614848137,
0.017024226486682892,
-0.015740979462862015,
-0.04002081975340843,
0.05530208721756935,
-0.09806361794471741,
-0.039286453276872635,
-0.09174525737762451,
-4.576357781616025e-8,
-0.015438364818692207,
0.06081748753786087,
0.03531188890337944,
-0.026853471994400024,
0.12969687581062317,
-0.038316674530506134,
-0.028741931542754173,
0.052016403526067734,
-0.015490733087062836,
-0.00543213402852416,
0.039223190397024155,
0.04657158628106117,
0.004135063383728266,
-0.032663971185684204,
-0.024137111380696297,
0.061079513281583786,
-0.11455867439508438,
-0.02843000739812851,
0.022593483328819275,
0.01983426697552204,
-0.04313008114695549,
0.005315584596246481,
0.00558055192232132,
-0.07666239142417908,
0.05950377136468887,
-0.11852557212114334,
-0.0008177465060725808,
0.036348696798086166,
-0.01577371172606945,
0.020613711327314377,
0.034820809960365295,
0.01355078537017107,
0.04373294860124588,
0.00451726233586669,
0.04032943770289421,
0.088150255382061,
-0.023448994383215904,
0.017390146851539612,
-0.018006127327680588,
-0.008178707212209702,
-0.03895964100956917,
0.0713055282831192,
0.014743798412382603,
0.021708158776164055,
-0.0038425824604928493,
-0.04055730625987053,
0.032834772020578384,
-0.0841992050409317,
0.011112446896731853,
-0.009563927538692951,
0.05330771207809448,
-0.03177647665143013,
0.033248208463191986,
0.030798200517892838,
0.03899227827787399,
-0.021816091611981392,
-0.03268636390566826,
-0.06804239749908447,
-0.06581683456897736,
0.04950608313083649,
-0.032878268510103226,
0.03518228977918625,
-0.049801915884017944,
-0.03779803588986397
] |
Limits of Deep Learning: Sequence Modeling through the Lens of Complexity Theory | https://openreview.net/forum?id=DhdqML3FdM | [
"Nikola Zubic",
"Federico Soldà",
"Aurelio Sulser",
"Davide Scaramuzza"
] | Poster | Despite their successes, deep learning models struggle with tasks requiring complex reasoning and function composition. We present a theoretical and empirical investigation into the limitations of Structured State Space Models (SSMs) and Transformers in such tasks. We prove that one-layer SSMs cannot efficiently perform function composition over large domains without impractically large state sizes, and even with Chain-of-Thought prompting, they require a number of steps that scale unfavorably with the complexity of the function composition. Also, the language of a finite-precision SSM is within the class of regular languages. Our experiments corroborate these theoretical findings. Evaluating models on tasks including various function composition settings, multi-digit multiplication, dynamic programming, and Einstein's puzzle, we find significant performance degradation even with advanced prompting techniques. Models often resort to shortcuts, leading to compounding errors. These findings highlight fundamental barriers within current deep learning architectures rooted in their computational capacities. We underscore the need for innovative solutions to transcend these constraints and achieve reliable multi-step reasoning and compositional task-solving, which is critical for advancing toward general artificial intelligence. | theory, complexity theory, state space models, deep learning architectures, logic in computer science | We theoretically prove the shortcomings of current deep learning models with a specific focus on State Space Models. | 2,378 | 2405.16674 | [
-0.046237263828516006,
-0.055505264550447464,
0.07463442534208298,
0.02944357693195343,
-0.03953593596816063,
-0.02034633979201317,
-0.08285333961248398,
-0.013583915308117867,
0.0058258892968297005,
-0.06712528318166733,
-0.11244293302297592,
-0.09569632261991501,
-0.057592540979385376,
0.07119634002447128,
-0.010114762000739574,
-0.02133498340845108,
0.02016402967274189,
0.052403371781110764,
-0.1100914478302002,
-0.06564868241548538,
0.04087534174323082,
0.013194911181926727,
-0.012648320756852627,
0.021149711683392525,
-0.01178804598748684,
0.004669160582125187,
-0.025261661037802696,
-0.11860351264476776,
0.010678515769541264,
0.011116103269159794,
0.10974041372537613,
-0.025128012523055077,
0.06418605893850327,
0.052559394389390945,
-0.01505982968956232,
0.10958138853311539,
-0.08161863684654236,
-0.01314566284418106,
-0.02847393788397312,
-0.05412047728896141,
-0.019735975190997124,
-0.031012699007987976,
0.0717664435505867,
0.026784194633364677,
0.06602991372346878,
-0.03180409595370293,
0.0633629858493805,
-0.05028130114078522,
0.013117684051394463,
0.026033179834485054,
-0.09470951557159424,
0.04942674562335014,
-0.042141836136579514,
0.08801566064357758,
0.01946704089641571,
0.004946182947605848,
0.055488038808107376,
0.0531671866774559,
-0.09857078641653061,
-0.07778953015804291,
-0.07047580182552338,
-0.1343797743320465,
0.02145766094326973,
-0.03221854567527771,
0.004532263148576021,
0.05685093253850937,
-0.013667484745383263,
0.01572106033563614,
-0.04427358880639076,
0.09603298455476761,
0.025215595960617065,
0.07013954967260361,
-0.038913823664188385,
0.021978888660669327,
0.09340999275445938,
0.014757664874196053,
0.039240095764398575,
-0.022523336112499237,
0.04650988429784775,
-0.026409050449728966,
0.03215917944908142,
0.025440866127610207,
-0.012699619866907597,
0.04027801379561424,
0.12634892761707306,
-0.032410651445388794,
-0.022109348326921463,
0.03224090859293938,
0.04988062009215355,
-0.0025607110001146793,
-0.023333679884672165,
-0.08491889387369156,
0.03637159988284111,
-0.04708274453878403,
0.08334264159202576,
0.0787230134010315,
-0.0336700864136219,
-0.07616985589265823,
-0.04346635937690735,
0.07567921280860901,
-0.04504995420575142,
0.061703018844127655,
0.05664912983775139,
-0.06301672011613846,
-0.029173782095313072,
0.02819080464541912,
0.01895168609917164,
0.06804566830396652,
-0.010444192215800285,
-0.11495015025138855,
0.058975182473659515,
0.05306921899318695,
0.08172350376844406,
0.051825813949108124,
-0.017196329310536385,
-0.04957537725567818,
0.028264811262488365,
-0.08163291960954666,
0.021231358870863914,
0.0812302976846695,
-0.040133051574230194,
0.021440505981445312,
-0.00745015824213624,
0.06431613117456436,
-0.02435889095067978,
0.006189892068505287,
-0.00406135618686676,
2.0075667413548975e-33,
-0.023434987291693687,
0.015113651752471924,
0.053942106664180756,
0.04998249188065529,
0.05174095183610916,
0.01082775928080082,
0.040314871817827225,
0.04944419115781784,
-0.0703265368938446,
0.08726997673511505,
-0.046029187738895416,
0.03499668464064598,
-0.04685826972126961,
0.06041942909359932,
0.04810993745923042,
-0.008415954187512398,
0.017036598175764084,
0.02207425981760025,
0.010646759532392025,
-0.09921475499868393,
0.07241777330636978,
-0.025652268901467323,
-0.024296468123793602,
-0.05214850977063179,
0.010233098641037941,
0.04606059566140175,
0.05383619666099548,
-0.027771297842264175,
-0.05417228490114212,
-0.028658099472522736,
-0.1538044959306717,
0.030283773317933083,
-0.033206649124622345,
0.07225986570119858,
0.05991693213582039,
-0.0047251880168914795,
0.0012242720695212483,
-0.04803716763854027,
0.0396776981651783,
-0.05317838490009308,
0.002763736294582486,
0.026243148371577263,
-0.020291341468691826,
0.05183378607034683,
-0.06159759685397148,
-0.031619004905223846,
0.03443584218621254,
-0.008928523398935795,
-0.06613563001155853,
0.022235391661524773,
0.04435810074210167,
0.02640136331319809,
-0.029303234070539474,
-0.09509483724832535,
0.07837360352277756,
-0.009761822409927845,
0.03891350328922272,
0.04913100600242615,
0.06902596354484558,
0.07460083812475204,
-0.009773979894816875,
0.061332497745752335,
-0.051174767315387726,
0.0954069197177887,
-0.007560299709439278,
0.053998954594135284,
-0.04054964706301689,
0.11110050231218338,
0.07325642555952072,
-0.02375771850347519,
0.03121132403612137,
0.0389346107840538,
-0.041714005172252655,
-0.04984091594815254,
0.08644269406795502,
0.0009031550143845379,
0.08111277222633362,
-0.2048414945602417,
-0.061953138560056686,
0.042807064950466156,
0.0000022828701276011998,
0.02437734045088291,
-0.026150736957788467,
0.05315272510051727,
-0.021407784894108772,
-0.06339258700609207,
0.03769141808152199,
-0.04561927169561386,
-0.023652026429772377,
-0.02167915739119053,
-0.056575533002614975,
-0.02827967330813408,
0.03969625383615494,
-0.025310050696134567,
-0.03444533050060272,
-2.4905696559260982e-33,
-0.06940003484487534,
-0.06830348819494247,
-0.05405309051275253,
0.04729711636900902,
-0.02146010845899582,
-0.07958412170410156,
-0.07819195836782455,
-0.024393200874328613,
-0.014184429310262203,
-0.03326227143406868,
-0.008529057726264,
-0.045701462775468826,
0.05893027409911156,
0.002665528329089284,
0.018987052142620087,
0.004874030128121376,
-0.03327201306819916,
0.014659433625638485,
0.026208432391285896,
0.01654684729874134,
0.027409985661506653,
0.11336322128772736,
-0.07594011723995209,
0.03284379467368126,
-0.07666851580142975,
0.03385661914944649,
-0.010546277277171612,
0.026009129360318184,
0.027283800765872,
0.06490598618984222,
-0.07969538867473602,
-0.02691120095551014,
-0.027950525283813477,
0.013217455707490444,
-0.014681022614240646,
0.04865053668618202,
0.02898545004427433,
0.00676949555054307,
-0.025882331654429436,
-0.03774503991007805,
0.03737732023000717,
-0.08434831351041794,
-0.047726988792419434,
0.010360351763665676,
-0.021597569808363914,
0.020428059622645378,
-0.022529102861881256,
0.018841654062271118,
-0.030089164152741432,
0.03467480465769768,
-0.02919270657002926,
-0.009711108170449734,
-0.08953766524791718,
-0.024258142337203026,
-0.06615965068340302,
0.013861129991710186,
-0.03431517258286476,
-0.004038933664560318,
0.07844063639640808,
-0.010090994648635387,
-0.06177898496389389,
-0.040027547627687454,
0.05941000208258629,
-0.07527299970388412,
0.014837565831840038,
-0.015893060714006424,
-0.04360683262348175,
0.02567729726433754,
0.04701092466711998,
0.061195701360702515,
0.03799694404006004,
-0.0500485859811306,
0.023688942193984985,
0.04918264225125313,
-0.019082162529230118,
0.0022517568431794643,
0.02127286233007908,
-0.022919440641999245,
0.010463682003319263,
-0.001255058217793703,
0.07341516762971878,
0.007311474531888962,
0.021341875195503235,
0.022674651816487312,
0.010127520188689232,
0.06667356193065643,
0.025940977036952972,
0.10063357651233673,
-0.0020618850830942392,
0.049282219260931015,
-0.05163612961769104,
0.05781841650605202,
-0.016092713922262192,
0.020644577220082283,
-0.07799143344163895,
-4.455877444797807e-8,
-0.048870861530303955,
0.06590893119573593,
0.038613803684711456,
-0.027105258777737617,
0.048766978085041046,
-0.030512209981679916,
0.0005070003680884838,
0.045962635427713394,
-0.024770699441432953,
-0.012531401589512825,
0.06895913183689117,
0.007588261738419533,
-0.02826778031885624,
-0.055744532495737076,
-0.03420371562242508,
0.0952361598610878,
0.03217393532395363,
-0.03687422722578049,
-0.03342144936323166,
0.011134822852909565,
0.0012037912383675575,
0.041493188589811325,
-0.008524846285581589,
-0.017813192680478096,
-0.04265976324677467,
-0.09057416766881943,
-0.0065867784433066845,
0.0698111355304718,
0.07443655282258987,
-0.019371336326003075,
0.03202889859676361,
0.00645210687071085,
0.004745453596115112,
0.021945146843791008,
-0.04316474497318268,
0.021224480122327805,
0.026429129764437675,
0.012658881954848766,
0.03388595208525658,
-0.08021131902933121,
-0.01658470556139946,
0.09679200500249863,
-0.033299654722213745,
0.0359807051718235,
-0.042635560035705566,
0.008060697466135025,
-0.04598088935017586,
-0.06855631619691849,
-0.02443617954850197,
0.007122966926544905,
0.02227931283414364,
0.038144052028656006,
-0.06013154983520508,
0.08619092404842377,
0.029541704803705215,
0.03951885178685188,
0.0079859159886837,
-0.051669489592313766,
-0.032783143222332,
0.10264943540096283,
-0.009028094820678234,
0.035873349756002426,
-0.07509797811508179,
-0.0138449901714921
] |
It Helps to Take a Second Opinion: Teaching Smaller LLMs To Deliberate Mutually via Selective Rationale Optimisation | https://openreview.net/forum?id=NHxwxc3ql6 | [
"Sohan Patnaik",
"Milan Aggarwal",
"Sumit Bhatia",
"Balaji Krishnamurthy"
] | Poster | Very large language models (LLMs) such as GPT-4 have shown the ability to handle complex tasks by generating and self-refining step-by-step rationales. Smaller language models (SLMs), typically with < 13B parameters, have been improved by using the data generated from very-large LMs through knowledge distillation. However, various practical constraints such as API costs, copyright, legal and ethical policies restrict using large (often opaque) models to train smaller models for commercial use. Limited success has been achieved at improving the ability of an SLM to explore the space of possible rationales and evaluate them by itself through self-deliberation. To address this, we propose COALITION, a trainable framework that facilitates interaction between two variants of the same SLM and trains them to generate and refine rationales optimized for the end-task. The variants exhibit different behaviors to produce a set of diverse candidate rationales during the generation and refinement steps. The model is then trained via Selective Rationale Optimization (SRO) to prefer generating rationale candidates that maximize the likelihood of producing the ground-truth answer. During inference, COALITION employs a controller to select the suitable variant for generating and refining the rationales. On five different datasets covering mathematical problems, commonsense reasoning, and natural language inference, COALITION outperforms several baselines by up to 5%. Our ablation studies reveal that cross-communication between the two variants performs better than using the single model to self-refine the rationales. We also demonstrate the applicability of COALITION for LMs of varying scales (4B to 14B parameters) and model families (Mistral, Llama, Qwen, Phi). We release the code for this work here. | Multi-LLM Deliberation, Smaller LLMs, Rationale Generation, Rationale Refinement, Selective Rationale Optimisation, Trainable, Task-Guided Rationale Selection | A trainable framework that facilitates interaction between two distinct Variants of the same LM to preferentially Generate and Refine better rationale choices guided by the end-task. | 2,371 | 2503.02463 | [
-0.00857226736843586,
-0.05840924754738808,
-0.0082766218110919,
-0.0025882450863718987,
0.04549055173993111,
-0.008898280560970306,
-0.01759343221783638,
-0.01562732458114624,
0.052819959819316864,
-0.011924516409635544,
-0.07443757355213165,
-0.058790285140275955,
0.06910642981529236,
0.050017185509204865,
0.07054636627435684,
-0.005432699806988239,
0.0712074413895607,
0.05026758462190628,
-0.11529078334569931,
-0.07698461413383484,
0.01676885224878788,
0.0010404677595943213,
-0.004988108295947313,
-0.002596590667963028,
0.012966026552021503,
-0.04107552766799927,
0.02922705188393593,
-0.050538528710603714,
0.061500873416662216,
-0.04331628605723381,
0.07480219751596451,
-0.013952227309346199,
0.034495819360017776,
0.027513517066836357,
0.0025853957049548626,
0.07843456417322159,
-0.07558949291706085,
-0.026743832975625992,
0.0452130064368248,
-0.022683629766106606,
-0.04367037117481232,
-0.01846056804060936,
-0.04804295301437378,
-0.005902460776269436,
0.07518515735864639,
-0.04258111119270325,
-0.024876510724425316,
0.02243105135858059,
-0.05449065938591957,
-0.01815946027636528,
-0.08591008186340332,
-0.054938748478889465,
-0.051281340420246124,
0.013312752358615398,
-0.006028352305293083,
-0.016620494425296783,
0.0022676405496895313,
0.036660224199295044,
-0.060570910573005676,
-0.08100791275501251,
-0.06813900917768478,
-0.0822322741150856,
-0.12112017720937729,
-0.01897399127483368,
-0.03620939701795578,
-0.01559118926525116,
0.010250450111925602,
0.12635290622711182,
-0.06792734563350677,
0.0217934288084507,
0.03818436712026596,
0.030855301767587662,
-0.026905516162514687,
0.06488978862762451,
0.013725603930652142,
0.07307813316583633,
0.02008097991347313,
0.03967056795954704,
0.006134623661637306,
-0.04395706206560135,
-0.022770624607801437,
0.002538449829444289,
0.009588174521923065,
-0.00031046196818351746,
0.08339693397283554,
-0.0027409647591412067,
0.039102427661418915,
0.08005046099424362,
0.11055221408605576,
0.04536822810769081,
-0.06762859970331192,
-0.010177361778914928,
0.03845766559243202,
-0.014428071677684784,
0.07229554653167725,
0.03205394372344017,
0.026506923139095306,
-0.0949268713593483,
0.006150819361209869,
0.11802785098552704,
-0.004007149953395128,
0.12598325312137604,
0.008659477345645428,
-0.10488293319940567,
0.0008473795605823398,
-0.0330713652074337,
-0.017027469351887703,
0.07115162163972855,
0.04254427179694176,
-0.02982804737985134,
0.004408855456858873,
0.08403890579938889,
0.04253252223134041,
-0.0035665675532072783,
0.07392153143882751,
0.0657276138663292,
0.026675011962652206,
-0.0008850607555359602,
0.015295124612748623,
0.05808178707957268,
-0.018395068123936653,
0.04905425012111664,
0.04860932007431984,
0.00830728281289339,
0.052248917520046234,
-0.03323956951498985,
-0.08036673069000244,
1.7009708150245745e-33,
0.07202974706888199,
0.04277218505740166,
0.02614104188978672,
0.03311128541827202,
0.0655340924859047,
0.023385636508464813,
-0.012440353631973267,
0.045632485300302505,
-0.028207438066601753,
0.02888484299182892,
-0.008025587536394596,
0.0791405737400055,
0.0068044718354940414,
0.006872976198792458,
0.08207403868436813,
0.027536893263459206,
-0.09271140396595001,
0.04427611082792282,
-0.08929436653852463,
-0.05817895382642746,
-0.009900519624352455,
-0.03750153258442879,
0.053095705807209015,
-0.03798889368772507,
-0.019732462242245674,
-0.020892491564154625,
0.08057201653718948,
-0.050629619508981705,
0.012502368539571762,
0.03796504810452461,
-0.10024674236774445,
-0.04123656824231148,
0.000792631006333977,
0.06378639489412308,
0.011793586425483227,
0.01803700067102909,
-0.03586604446172714,
-0.09352613240480423,
0.06812146306037903,
-0.003410578006878495,
0.009973892942070961,
-0.004110436886548996,
0.06338036060333252,
0.04039144515991211,
-0.030148370191454887,
-0.02699429728090763,
0.00627663591876626,
0.004274416249245405,
-0.07537633925676346,
-0.014427082613110542,
0.037918586283922195,
0.061373356729745865,
-0.01579044573009014,
-0.040343403816223145,
-0.03026045300066471,
0.013656195253133774,
-0.03337765112519264,
0.016803767532110214,
-0.04412868618965149,
0.0016843023477122188,
-0.006805169861763716,
0.03304195776581764,
-0.022659845650196075,
0.10194291919469833,
0.010148385539650917,
0.04519479721784592,
-0.0909247025847435,
0.0515095517039299,
0.0924261212348938,
-0.025975139811635017,
-0.04088914394378662,
-0.05327688530087471,
-0.013379697687923908,
-0.07552266120910645,
0.00492267869412899,
-0.05494672805070877,
0.08591226488351822,
-0.028286995366215706,
0.041492365300655365,
-0.05309631675481796,
0.056398238986730576,
0.04877443239092827,
-0.012152907438576221,
-0.04967620223760605,
-0.04326394200325012,
-0.05578142777085304,
0.048785921186208725,
-0.033537160605192184,
-0.06093868613243103,
-0.021139292046427727,
-0.1500018984079361,
-0.019978247582912445,
0.007938777096569538,
-0.03603653609752655,
0.011599179357290268,
-2.0294300179252305e-33,
-0.030472680926322937,
-0.033802177757024765,
-0.05864066258072853,
0.05525854229927063,
-0.007711272221058607,
-0.03896874189376831,
0.010718650184571743,
-0.022346004843711853,
0.00584459537640214,
-0.014184744097292423,
-0.06376542150974274,
-0.07362819463014603,
0.10480265319347382,
0.008656855672597885,
-0.04865402728319168,
-0.07513200491666794,
0.016152972355484962,
-0.032159224152565,
0.06657551229000092,
0.11242785304784775,
-0.026524009183049202,
0.09112267941236496,
-0.06208590045571327,
0.08583337813615799,
0.020961642265319824,
0.01734536699950695,
-0.07651472836732864,
0.0008324338705278933,
0.05642499029636383,
0.0778246745467186,
-0.0429864227771759,
-0.022896450012922287,
-0.02091490477323532,
-0.12074707448482513,
-0.07460848987102509,
0.04787768796086311,
0.03722085431218147,
0.04280969873070717,
-0.021893152967095375,
0.06094804033637047,
-0.008834371343255043,
-0.12137596309185028,
-0.026793649420142174,
0.023326043039560318,
-0.0323195680975914,
-0.07474595308303833,
-0.06669571995735168,
-0.10243760049343109,
0.004792159888893366,
0.037884268909692764,
0.022005965933203697,
0.006081925705075264,
-0.042753223329782486,
0.026564225554466248,
-0.040592070668935776,
-0.085147425532341,
-0.016670653596520424,
-0.037620559334754944,
0.007865386083722115,
0.016592564061284065,
0.0026144215371459723,
0.0408494807779789,
0.04134593904018402,
-0.06329303234815598,
0.011614098213613033,
-0.055215023458004,
-0.06543216854333878,
0.06920932233333588,
-0.023631606251001358,
-0.008205600082874298,
-0.05232836678624153,
-0.017174355685710907,
0.017127584666013718,
0.017605608329176903,
-0.04079052060842514,
-0.0062643554992973804,
-0.05819063261151314,
-0.017779912799596786,
-0.002661032835021615,
-0.048465047031641006,
0.11367896944284439,
0.01846407912671566,
0.03646305203437805,
0.09861395508050919,
-0.00006616996688535437,
-0.056483376771211624,
0.06656944751739502,
0.11970412731170654,
-0.009884039871394634,
0.05454400181770325,
-0.036080483347177505,
0.0013981085503473878,
0.07229006290435791,
0.11062656342983246,
-0.06474350392818451,
-4.879710857608188e-8,
0.004300897009670734,
0.02424585074186325,
0.0030376939103007317,
0.085578553378582,
0.0005237130098976195,
-0.020588597282767296,
-0.06691131740808487,
0.020021023228764534,
-0.02517908252775669,
0.0170455239713192,
0.016097933053970337,
0.03583047538995743,
-0.07406672835350037,
-0.00824502669274807,
-0.012813358567655087,
0.07198192924261093,
-0.04858317971229553,
-0.014593160711228848,
-0.026802437379956245,
-0.06273767352104187,
0.061512112617492676,
0.005887278355658054,
-0.03438323736190796,
0.0038692618254572153,
0.09289200603961945,
-0.046958934515714645,
-0.03056672029197216,
0.08541812002658844,
0.03562164306640625,
0.06941596418619156,
-0.024226944893598557,
-0.04461004212498665,
-0.060316383838653564,
0.015170413069427013,
0.022113967686891556,
0.017759157344698906,
-0.059224605560302734,
0.013416162692010403,
0.0707818791270256,
0.012759560719132423,
0.008754274807870388,
0.10224767029285431,
-0.03416488692164421,
0.03769063577055931,
0.003425460308790207,
-0.04764685407280922,
-0.1158241480588913,
-0.11731686443090439,
0.04556252434849739,
0.033289387822151184,
0.012691419571638107,
-0.03572617471218109,
-0.0037047064397484064,
0.02716124802827835,
0.03676975518465042,
0.0848141461610794,
-0.023487448692321777,
-0.012439664453268051,
0.008086810819804668,
0.03943897783756256,
0.031654492020606995,
0.06473828107118607,
0.01762501709163189,
-0.008981151506304741
] |
Denoising Levy Probabilistic Models | https://openreview.net/forum?id=SYmUS6qRub | [
"Dario Shariatian",
"Umut Simsekli",
"Alain Oliviero Durmus"
] | Poster | Investigating noise distributions beyond Gaussian in diffusion generative models remains an open challenge. The Gaussian case has been a large success experimentally and theoretically, admitting a unified stochastic differential equation (SDE) framework, encompassing score-based and denoising formulations. Recent studies have investigated the potential of \emph{heavy-tailed} noise distributions to mitigate mode collapse and effectively manage datasets exhibiting class imbalance, heavy tails, or prominent outliers.
Very recently, Yoon et al.\ (NeurIPS 2023), presented the Levy-Ito model (LIM), directly extending the SDE-based framework to a class of heavy-tailed SDEs, where the injected noise followed an $\alpha$-stable distribution -- a rich class of heavy-tailed distributions.
Despite its theoretical elegance and performance improvements, LIM relies on highly involved mathematical techniques, which may limit its accessibility and hinder its broader adoption and further development.
In this study, we take a step back, and instead of starting from the SDE formulation, we extend the denoising diffusion probabilistic model (DDPM) by directly replacing the Gaussian noise with $\alpha$-stable noise.
By using only elementary proof techniques, we show that the proposed approach, \emph{denoising L\'{e}vy probabilistic model} (DLPM) algorithmically boils down to running vanilla DDPM with minor modifications, hence allowing the use of existing implementations with minimal changes.
Remarkably, as opposed to the Gaussian case, DLPM and LIM yield different training algorithms and different backward processes, leading to distinct sampling algorithms.
This fundamental difference translates favorably for the performance of DLPM in various aspects: our experiments show that DLPM achieves better coverage of the tails of the data distribution, better generation of unbalanced datasets, and improved computation times requiring significantly smaller number of backward steps. | diffusion, generative model, deep learning, machine learning, heavy-tail | The paper introduces the Denoising Lévy Probabilistic Model (DLPM), which replaces Gaussian noise in denoising diffusion probabilistic models (DDPM) with heavy-tailed α-stable noise to improve performance, showing better tail coverage | 2,369 | 2407.18609 | [
-0.049354009330272675,
-0.09000088274478912,
0.09692792594432831,
0.037457820028066635,
0.03687762841582298,
-0.05317620187997818,
0.033877160400152206,
0.01427063625305891,
0.020932670682668686,
0.013497572392225266,
0.015779800713062286,
0.034978676587343216,
0.014497531577944756,
-0.06548048555850983,
0.01784507744014263,
-0.035853311419487,
0.04450266808271408,
-0.002076541306450963,
-0.08152973651885986,
0.03151559457182884,
0.0036663906648755074,
0.06355907022953033,
-0.08273270726203918,
0.02408922277390957,
0.07700929045677185,
-0.05794226750731468,
0.02867763675749302,
-0.06595124304294586,
-0.03131073713302612,
-0.11150310933589935,
0.11228035390377045,
-0.0010523280361667275,
-0.03385104238986969,
0.013136867433786392,
0.008796625770628452,
-0.023535603657364845,
0.007407759316265583,
0.04416686296463013,
0.025950562208890915,
0.014248614199459553,
0.04290945827960968,
0.005494208540767431,
-0.017381789162755013,
-0.00919275637716055,
-0.022642452269792557,
-0.05884166061878204,
0.014226531609892845,
-0.11202259361743927,
0.039888907223939896,
0.042672161012887955,
0.07947087287902832,
0.022650113329291344,
-0.04176976904273033,
0.03853887319564819,
0.056493598967790604,
0.03473683074116707,
0.07722868770360947,
0.006619925610721111,
0.015610824339091778,
0.03211859613656998,
-0.016051318496465683,
-0.04231136292219162,
-0.014520972035825253,
-0.0150590930134058,
-0.016949210315942764,
0.01949358731508255,
0.0421784482896328,
0.030301112681627274,
0.01348042581230402,
0.06641535460948944,
-0.008428659290075302,
-0.0004275586106814444,
-0.11577942222356796,
0.06051303818821907,
-0.01156669482588768,
-0.019158946350216866,
-0.005209111608564854,
-0.05186466872692108,
0.03332376852631569,
-0.02755715884268284,
0.02314525656402111,
0.03162193298339844,
0.05249197036027908,
-0.11354854702949524,
0.05644790828227997,
-0.022343425080180168,
0.03242914378643036,
0.024442745372653008,
0.035324014723300934,
-0.012331773526966572,
-0.008666764944791794,
0.10989991575479507,
-0.051269322633743286,
0.02555726282298565,
0.010312282480299473,
0.07531954348087311,
0.014081792905926704,
-0.05854129046201706,
0.06047547236084938,
0.08788443356752396,
-0.0048995958641171455,
0.07520119845867157,
-0.020776400342583656,
-0.03521515429019928,
0.012898207642138004,
-0.13765673339366913,
0.08628898859024048,
-0.03451157361268997,
-0.0419132299721241,
-0.03485747426748276,
-0.00706665962934494,
0.024370158091187477,
0.036068592220544815,
0.036428049206733704,
-0.04914906620979309,
-0.054311178624629974,
0.021713607013225555,
0.008485672995448112,
-0.044383704662323,
-0.024282181635499,
0.024019232019782066,
0.0271463505923748,
-0.07475675642490387,
-0.029310429468750954,
0.036989036947488785,
0.10451556742191315,
-0.01941687799990177,
2.176853723578646e-33,
0.01595393940806389,
-0.040888357907533646,
-0.0716223269701004,
-0.02561553753912449,
0.07352128624916077,
-0.012449950911104679,
-0.009739136323332787,
-0.006024216767400503,
-0.06396988034248352,
0.031038885936141014,
0.014451668597757816,
-0.01280438993126154,
-0.10604260116815567,
-0.007990533486008644,
0.04553372785449028,
-0.03517427295446396,
-0.06397625058889389,
0.037879373878240585,
0.0419052392244339,
-0.067152239382267,
0.06187012046575546,
-0.04059366136789322,
-0.033774737268686295,
-0.0166071318089962,
-0.01988411508500576,
0.009286506101489067,
0.041110433638095856,
-0.06592380255460739,
0.07972560077905655,
0.05083354935050011,
-0.11254210025072098,
0.06125691533088684,
-0.01022660918533802,
-0.04911186546087265,
0.010548367165029049,
-0.013000836595892906,
-0.052282799035310745,
-0.002241323236376047,
-0.06008770689368248,
-0.06540743261575699,
-0.016844341531395912,
0.05246790871024132,
-0.08294645696878433,
-0.01199227012693882,
-0.12828555703163147,
0.010391085408627987,
0.09292564541101456,
-0.1157115176320076,
-0.01250616554170847,
-0.12241076678037643,
-0.0225606020539999,
-0.04997336491942406,
-0.0019607185386121273,
0.020721403881907463,
0.02316868305206299,
0.06545180082321167,
0.0644555315375328,
-0.0784686729311943,
-0.022626370191574097,
-0.04173056036233902,
-0.03282889723777771,
0.03632119670510292,
0.016011733561754227,
-0.008850003592669964,
0.11598126590251923,
0.04566524177789688,
-0.046421051025390625,
0.05906054750084877,
-0.00660359114408493,
-0.031169883906841278,
0.09140586107969284,
-0.051654111593961716,
-0.05339569225907326,
-0.08898245543241501,
-0.0239237193018198,
-0.04779300093650818,
0.05517714098095894,
0.005787705071270466,
0.01982085593044758,
-0.021957334131002426,
-0.023529114201664925,
-0.08455222845077515,
-0.02116117812693119,
-0.021479591727256775,
0.02954656071960926,
0.017976826056838036,
-0.025500144809484482,
-0.0437333844602108,
-0.07285963743925095,
-0.12682068347930908,
-0.03423004597425461,
-0.023433208465576172,
0.00989712867885828,
-0.029751194640994072,
-0.09374994784593582,
-2.9378265219727623e-33,
-0.06630188971757889,
0.11204668134450912,
-0.043867845088243484,
0.04004478454589844,
-0.013504277914762497,
-0.02023960091173649,
-0.04873652383685112,
0.05551711469888687,
0.03881952166557312,
-0.039843495935201645,
-0.0870603546500206,
-0.013959335163235664,
-0.054616037756204605,
0.014852037653326988,
0.032896991819143295,
-0.110637366771698,
0.058348510414361954,
0.030865076929330826,
-0.06676008552312851,
-0.019882304593920708,
-0.0039230892434716225,
-0.0026249000802636147,
0.02654554694890976,
0.05533546209335327,
0.07696348428726196,
0.0008919340907596052,
-0.09399338066577911,
0.1265363097190857,
0.0005229869857430458,
-0.05751122906804085,
-0.05407760664820671,
0.004638406448066235,
0.05190371721982956,
-0.03817633539438248,
-0.041666753590106964,
0.10150404274463654,
0.027419310063123703,
0.03817496821284294,
-0.005319342482835054,
0.029353514313697815,
0.03134510666131973,
0.022001774981617928,
0.01942398212850094,
-0.04528927057981491,
0.09608352184295654,
0.0014996269019320607,
0.006044439971446991,
0.051204703748226166,
0.058348413556814194,
0.009525150060653687,
0.007526380475610495,
0.057088349014520645,
-0.03458369895815849,
0.18077035248279572,
-0.053616929799318314,
-0.02954496443271637,
0.05189993605017662,
-0.07985783368349075,
-0.018115568906068802,
-0.003094669431447983,
-0.07196163386106491,
0.022778693586587906,
-0.006090699695050716,
0.020455384626984596,
-0.1286482959985733,
-0.015311026945710182,
-0.054461270570755005,
-0.015850836411118507,
0.023659303784370422,
-0.012196500785648823,
-0.020598722621798515,
-0.047773249447345734,
0.006660809274762869,
0.01812516525387764,
-0.026705561205744743,
-0.056213825941085815,
0.0546397902071476,
-0.0002539555716793984,
0.06744703650474548,
-0.01690756343305111,
0.006561733782291412,
0.061057642102241516,
0.005211476236581802,
-0.0038736083079129457,
0.01965673267841339,
-0.042783550918102264,
-0.03700867295265198,
-0.09812591224908829,
0.056179918348789215,
0.014262499287724495,
-0.03228834643959999,
0.0033627189695835114,
0.0037828744389116764,
0.09100273996591568,
-0.038672082126140594,
-4.8499078530994666e-8,
-0.0033246325328946114,
0.043848298490047455,
0.020195933058857918,
-0.015397386625409126,
0.14972786605358124,
0.00961849931627512,
-0.014509888365864754,
0.01306274626404047,
0.008700993843376637,
-0.08883447200059891,
0.046247448772192,
-0.06762039661407471,
-0.06120934337377548,
-0.031121036037802696,
0.021596847102046013,
0.06955406069755554,
0.025074927136301994,
0.06821896135807037,
-0.01463727280497551,
-0.05103942006826401,
0.0053423927165567875,
0.020066354423761368,
-0.007991072721779346,
0.02612937055528164,
0.029544388875365257,
0.06383094191551208,
-0.02684158831834793,
0.009149957448244095,
0.02979457750916481,
0.052492521703243256,
0.00805850699543953,
0.079918771982193,
0.042269155383110046,
0.04276556521654129,
0.037584759294986725,
0.11113256961107254,
0.04385404661297798,
0.00396761717274785,
-0.00740768201649189,
0.004260527901351452,
0.01059700921177864,
0.016636641696095467,
-0.00923159345984459,
0.0026657511480152607,
-0.02544989436864853,
0.039723630994558334,
0.0205481369048357,
-0.011337187141180038,
0.1068151444196701,
0.026455648243427277,
-0.010619920678436756,
-0.01370889600366354,
0.044796451926231384,
0.04122117906808853,
-0.021441280841827393,
-0.00981219857931137,
0.02238278090953827,
-0.027039334177970886,
-0.04728972539305687,
-0.03068169206380844,
0.11497792601585388,
-0.02632504142820835,
-0.012884815223515034,
0.10793454945087433
] |
MMIU: Multimodal Multi-image Understanding for Evaluating Large Vision-Language Models | https://openreview.net/forum?id=WsgEWL8i0K | [
"Fanqing Meng",
"Jin Wang",
"Chuanhao Li",
"Quanfeng Lu",
"Hao Tian",
"Tianshuo Yang",
"Jiaqi Liao",
"Xizhou Zhu",
"Jifeng Dai",
"Yu Qiao",
"Ping Luo",
"Kaipeng Zhang",
"Wenqi Shao"
] | Poster | The capability to process multiple images is crucial for Large Vision-Language Models (LVLMs) to develop a more thorough and nuanced understanding of a scene. Recent multi-image LVLMs have begun to address this need. However, their evaluation has not kept pace with their development. To fill this gap, we introduce the Multimodal Multi-image Understanding (MMIU) benchmark, a comprehensive evaluation suite designed to assess LVLMs across a wide range of multi-image tasks. MMIU encompasses 7 types of multi-image relationships, 52 tasks, 77K images, and 11K meticulously curated multiple-choice questions, making it the most extensive benchmark of its kind. Our evaluation of nearly 30 popular LVLMs, including both open-source and proprietary models, reveals significant challenges in multi-image comprehension, particularly in tasks involving spatial understanding. Even the most advanced models, such as GPT-4o, achieve only 55.7\% accuracy on MMIU. Through multi-faceted analytical experiments, we identify key performance gaps and limitations, providing valuable insights for future model and data improvements. We aim for MMIU to advance the frontier of LVLM research and development. We release the data and code at https://github.com/MMIUBenchmark/MMIU. | Multi-image Understanding, Benchmark, LVLM, Evaluation | null | 2,353 | 2408.02718 | [
0.06730753183364868,
-0.1251644492149353,
0.005828604567795992,
-0.02757078967988491,
0.13225562870502472,
-0.03982122987508774,
-0.03835878521203995,
-0.014503843151032925,
0.017809560522437096,
-0.027248382568359375,
-0.044083304703235626,
-0.12055767327547073,
0.014690357260406017,
0.12866996228694916,
-0.014384991489350796,
0.0014523891732096672,
-0.00674299243837595,
0.04230742156505585,
-0.12381316721439362,
-0.00238471245393157,
0.01154859084635973,
-0.018273748457431793,
0.09200771898031235,
-0.07191068679094315,
0.09444048255681992,
0.04497310519218445,
-0.005348242353647947,
-0.07652595639228821,
-0.03692861646413803,
-0.04170187562704086,
0.020302098244428635,
0.018530428409576416,
0.018574660643935204,
0.06421144306659698,
0.01023197639733553,
0.07500869035720825,
-0.07465119659900665,
0.015983637422323227,
-0.010499835945665836,
-0.022372392937541008,
-0.01971193216741085,
0.01780213601887226,
0.05448850616812706,
-0.025978008285164833,
0.08681629598140717,
-0.024120237678289413,
-0.009945138357579708,
-0.08335331827402115,
0.03974490612745285,
0.0023553059436380863,
-0.07481159269809723,
-0.05088178440928459,
-0.042451877146959305,
0.05992693826556206,
-0.028983304277062416,
0.06559403240680695,
0.004842092748731375,
-0.03591557592153549,
0.0086124949157238,
-0.024068528786301613,
-0.00014435115735977888,
-0.06951914727687836,
-0.032784149050712585,
0.046262748539447784,
0.005238733720034361,
-0.03622344136238098,
-0.008852490223944187,
-0.06373295933008194,
-0.027126992121338844,
-0.04173249006271362,
-0.038988545536994934,
0.06623130291700363,
0.001612071180716157,
-0.002160491421818733,
0.009341398254036903,
0.053242217749357224,
0.03794112429022789,
-0.03792818263173103,
0.04806322604417801,
-0.1032964214682579,
0.08981560915708542,
0.011021068319678307,
0.0091826431453228,
-0.03314634785056114,
0.13407067954540253,
-0.0021807877346873283,
-0.013480489142239094,
0.06778974086046219,
0.00760635593906045,
0.0288862232118845,
-0.06451509892940521,
-0.067924864590168,
-0.01132370624691248,
0.011641004122793674,
0.09055879712104797,
0.012875689193606377,
0.03397790715098381,
-0.10523021221160889,
-0.07086501270532608,
0.044121984392404556,
0.03658604249358177,
0.028142137452960014,
-0.01662440225481987,
-0.04032793268561363,
0.020232688635587692,
-0.014431586489081383,
0.07314417511224747,
0.025550542399287224,
-0.0013569764560088515,
-0.05616844817996025,
0.04154054820537567,
0.002469497499987483,
-0.0009261762606911361,
0.030777353793382645,
-0.024881400167942047,
0.01140577532351017,
0.054712891578674316,
0.02194645255804062,
0.0703590139746666,
0.012809700332581997,
-0.07111408561468124,
-0.037427060306072235,
0.009045998565852642,
-0.03970520943403244,
0.03452882170677185,
0.003607469843700528,
-0.03712227940559387,
1.1015086445460366e-33,
-0.007023130543529987,
0.00262317038141191,
0.058533184230327606,
0.053443241864442825,
0.08674847334623337,
-0.018600638955831528,
-0.05411451682448387,
0.03875824809074402,
-0.07802601158618927,
0.026372047141194344,
-0.0078081353567540646,
0.01998816430568695,
-0.07798031717538834,
0.12347032129764557,
0.0028787213377654552,
-0.00621592765673995,
-0.0653555616736412,
0.03013128601014614,
-0.047035958617925644,
0.08595050871372223,
0.03514236584305763,
-0.013056750409305096,
-0.03285013139247894,
0.01972435973584652,
-0.0056091854348778725,
0.018649589270353317,
0.07231426984071732,
-0.06297408789396286,
-0.027581196278333664,
0.004523986019194126,
-0.16964396834373474,
0.002842704998329282,
-0.008014551363885403,
0.030304204672574997,
0.026092758402228355,
-0.04621157422661781,
-0.04517456889152527,
-0.052787479013204575,
0.010878048837184906,
-0.03142142668366432,
-0.06489808857440948,
0.0075224959291517735,
-0.00413114670664072,
-0.013363901525735855,
-0.09184285253286362,
0.022987065836787224,
-0.03167864680290222,
0.06766819208860397,
-0.07204051315784454,
0.05115731433033943,
0.041421208530664444,
0.042416710406541824,
-0.06716787070035934,
-0.056181713938713074,
-0.04216981679201126,
-0.007820356637239456,
0.03335270285606384,
0.012262126430869102,
0.036253463476896286,
-0.011829948984086514,
0.04259635880589485,
0.05248241871595383,
-0.03169587627053261,
0.10959003865718842,
0.03594233840703964,
0.03239727392792702,
-0.04223332554101944,
0.062375105917453766,
0.041209395974874496,
0.07381314039230347,
-0.057830892503261566,
0.019556181505322456,
0.06642455607652664,
-0.10206522792577744,
0.035480987280607224,
0.04757564514875412,
0.04737372323870659,
-0.07502100616693497,
0.0657033920288086,
0.03967816382646561,
-0.010879169218242168,
0.0169560257345438,
0.021467136219143867,
-0.07493025809526443,
-0.0699581652879715,
-0.013859559781849384,
0.026528548449277878,
-0.12935951352119446,
-0.01746424287557602,
-0.07229150831699371,
-0.013990991748869419,
0.005881036166101694,
-0.024027803912758827,
-0.04505405202507973,
-0.02652066759765148,
-8.322383372741675e-34,
0.028557978570461273,
0.013156137429177761,
0.013604676350951195,
0.041250359266996384,
-0.037050094455480576,
-0.0711313858628273,
0.020016135647892952,
0.01727013662457466,
0.021997200325131416,
-0.06906639039516449,
0.015292562544345856,
0.03110475465655327,
-0.03270788490772247,
0.022711163386702538,
0.0667991191148758,
-0.019164668396115303,
0.03754002973437309,
-0.043693818151950836,
-0.0020681514870375395,
0.043529633432626724,
0.0032260415609925985,
0.106608547270298,
-0.030162090435624123,
0.023039773106575012,
-0.13685137033462524,
0.12278509140014648,
-0.04847939312458038,
-0.05880722030997276,
-0.08914289623498917,
-0.045154836028814316,
0.00670497165992856,
-0.001307030557654798,
0.04084562510251999,
-0.018950290977954865,
-0.00349695747718215,
0.060759007930755615,
0.06192518770694733,
-0.09811024367809296,
-0.055818647146224976,
0.08018392324447632,
0.07027214020490646,
-0.05020061135292053,
-0.03707059845328331,
0.0042287916876375675,
-0.05346887931227684,
-0.01685934327542782,
-0.01593606546521187,
-0.015314250253140926,
0.02133983187377453,
0.10781895369291306,
-0.0482780821621418,
0.007372700143605471,
-0.11248779296875,
0.03162369132041931,
0.020944926887750626,
-0.08886951208114624,
-0.08276718109846115,
-0.02725440077483654,
0.03888098895549774,
-0.03335387632250786,
-0.01982996053993702,
-0.021778453141450882,
-0.06844314187765121,
0.03178877755999565,
0.01980399340391159,
-0.002432030625641346,
-0.04251554235816002,
-0.06509378552436829,
0.021723462268710136,
0.05223160609602928,
-0.03470080345869064,
-0.06207820773124695,
0.004553961101919413,
-0.006174503359943628,
0.011828785762190819,
0.005704485811293125,
0.03279683738946915,
-0.009458714164793491,
0.0468778982758522,
-0.03367336094379425,
0.028676357120275497,
-0.0811774805188179,
0.010809047147631645,
0.10241471976041794,
0.08721940219402313,
0.050401993095874786,
-0.012317296117544174,
0.030633671209216118,
0.09623308479785919,
0.06503751128911972,
-0.013817469589412212,
0.025147106498479843,
0.030047385022044182,
0.027694398537278175,
-0.03975699469447136,
-3.8746389208199616e-8,
-0.05035204067826271,
0.058106038719415665,
-0.0445542149245739,
-0.0342959500849247,
-0.02599853090941906,
-0.07731202989816666,
-0.05184167996048927,
0.10194575786590576,
0.006915417965501547,
0.022364266216754913,
0.06276224553585052,
-0.04146040976047516,
-0.0468897670507431,
0.002346162451431155,
0.04854618385434151,
0.05055882781744003,
0.005688659381121397,
0.08081303536891937,
-0.00041507030255161226,
0.029786191880702972,
0.027538815513253212,
0.023980041965842247,
0.032330840826034546,
-0.007845412008464336,
-0.025638889521360397,
-0.00461276900023222,
-0.051543235778808594,
0.11675196886062622,
0.06800961494445801,
-0.0647006407380104,
0.05736513435840607,
0.02534341998398304,
0.0009388833423145115,
-0.03724830970168114,
-0.01683744229376316,
0.0034794609528034925,
0.024987023323774338,
-0.004894906189292669,
0.035083312541246414,
-0.015921298414468765,
0.007725963369011879,
0.02010757476091385,
-0.026212722063064575,
0.06842141598463058,
0.006058902014046907,
0.037737954407930374,
-0.037610914558172226,
-0.03748540207743645,
-0.0187394917011261,
-0.001944876043125987,
-0.0838075652718544,
0.0271424762904644,
-0.011506116017699242,
0.1416693478822708,
0.05295410379767418,
-0.06929606944322586,
0.017262352630496025,
0.04012301564216614,
0.06752432882785797,
0.06498084962368011,
0.004004189744591713,
0.0007659734692424536,
-0.09937833249568939,
-0.007882836274802685
] |
From Layers to States: A State Space Model Perspective to Deep Neural Network Layer Dynamics | https://openreview.net/forum?id=msD4DHZzFg | [
"Qinshuo Liu",
"Weiqin Zhao",
"Wei Huang",
"Yanwen Fang",
"Lequan Yu",
"Guodong Li"
] | Poster | The depth of neural networks is a critical factor for their capability, with deeper models often demonstrating superior performance. Motivated by this, significant efforts have been made to enhance layer aggregation - reusing information from previous layers to better extract features at the current layer, to improve the representational power of deep neural networks. However, previous works have primarily addressed this problem from a discrete-state perspective which is not suitable as the number of network layers grows. This paper novelly treats the outputs from layers as states of a continuous process and considers leveraging the state space model (SSM) to design the aggregation of layers in very deep neural networks. Moreover, inspired by its advancements in modeling long sequences, the Selective State Space Models (S6) is employed to design a new module called Selective State Space Model Layer Aggregation (S6LA). This module aims to combine traditional CNN or transformer architectures within a sequential framework, enhancing the representational capabilities of state-of-the-art vision networks. Extensive experiments show that S6LA delivers substantial improvements in both image classification and detection tasks, highlighting the potential of integrating SSMs with contemporary deep learning techniques. | deep neural network, sequential model, state space model, statistical model | null | 2,351 | 2502.10463 | [
-0.11439797282218933,
-0.07774746417999268,
0.04688340052962303,
0.014751358889043331,
0.028399283066391945,
0.08832380175590515,
-0.08615385740995407,
-0.08040154725313187,
0.05207420140504837,
-0.12281329184770584,
-0.020923199132084846,
-0.01661287248134613,
-0.02037808671593666,
0.06274651736021042,
-0.08702506870031357,
0.017559783533215523,
0.048679858446121216,
0.0921192541718483,
-0.07866289466619492,
-0.0735076516866684,
0.07098942995071411,
-0.028102543205022812,
-0.007640598341822624,
-0.0038822146598249674,
0.06590630859136581,
0.08765346556901932,
0.008629821240901947,
-0.04631898179650307,
0.00342725683003664,
-0.038946498185396194,
0.035666573792696,
-0.04065285250544548,
-0.010940426960587502,
0.04704141616821289,
-0.04351004585623741,
0.025248926132917404,
-0.06600919365882874,
-0.0025888760574162006,
-0.0268548596650362,
-0.08907123655080795,
0.01639498397707939,
-0.015811428427696228,
0.05877818539738655,
0.008256616070866585,
0.049758587032556534,
0.00868793111294508,
0.014108429662883282,
-0.08211466670036316,
0.03661404177546501,
-0.01563805527985096,
-0.011572486720979214,
0.0556778647005558,
-0.03331577777862549,
0.13212816417217255,
-0.00801119301468134,
0.02323697693645954,
0.019001420587301254,
-0.03325655311346054,
-0.10410245507955551,
-0.03571844473481178,
-0.01863185502588749,
-0.019767718389630318,
0.057462457567453384,
-0.007155224680900574,
0.02009674906730652,
0.048371944576501846,
0.0158405639231205,
0.07476737350225449,
-0.005879600998014212,
-0.02700410783290863,
-0.015296101570129395,
0.10610691457986832,
0.034094713628292084,
0.004939267877489328,
0.030129538848996162,
-0.03456494212150574,
0.06164361536502838,
-0.035566024482250214,
0.11240355670452118,
-0.05263326317071915,
0.009317170828580856,
0.015020893886685371,
0.018338464200496674,
0.025020349770784378,
0.07878688722848892,
-0.008916052058339119,
-0.10378404706716537,
0.025710465386509895,
-0.04988645389676094,
-0.015898549929261208,
-0.06095787510275841,
-0.06788371503353119,
0.024451829493045807,
-0.09484465420246124,
0.08905787020921707,
-0.015991363674402237,
-0.0485110841691494,
-0.08140057325363159,
-0.04216959327459335,
0.02765440009534359,
-0.021264657378196716,
-0.02638905495405197,
0.07842525839805603,
-0.02931341342628002,
0.07826101779937744,
-0.01291117537766695,
-0.025462515652179718,
0.06254547834396362,
0.013915374875068665,
-0.07007210701704025,
0.04825476184487343,
0.023128895089030266,
0.0490676574409008,
0.012274390086531639,
0.025662431493401527,
0.003058340633288026,
-0.021223818883299828,
-0.01648322120308876,
0.004379223100841045,
0.024333849549293518,
-0.03439977392554283,
-0.004505596123635769,
-0.05886833742260933,
0.03517886996269226,
0.031792111694812775,
0.033252157270908356,
-0.052270494401454926,
2.163689105204398e-33,
-0.028403660282492638,
-0.07398966699838638,
0.03425605222582817,
0.03190367668867111,
0.013178536668419838,
0.004035392310470343,
0.03904706612229347,
0.03268947824835777,
-0.06862576305866241,
0.06305819004774094,
-0.11038605123758316,
0.054657403379678726,
-0.01707088202238083,
0.09796255081892014,
0.070672407746315,
-0.06498521566390991,
0.026884090155363083,
0.012899324297904968,
0.05623035505414009,
-0.07202181965112686,
-0.015188449062407017,
0.014045880176126957,
0.028440028429031372,
-0.010137490928173065,
-0.006529431324452162,
-0.002731130225583911,
0.07011386752128601,
0.03114381805062294,
-0.07975102961063385,
-0.009458002634346485,
-0.0775890052318573,
0.0747159868478775,
-0.015573051758110523,
-0.008960864506661892,
0.10600604861974716,
0.03308689594268799,
0.0022180357482284307,
0.006445505656301975,
0.05727075785398483,
-0.06467444449663162,
0.05517907068133354,
0.015848761424422264,
-0.09839427471160889,
0.036865200847387314,
-0.08186967670917511,
-0.00930931605398655,
0.004659677855670452,
-0.002872057259082794,
-0.0711655542254448,
0.012341597117483616,
0.0894247367978096,
-0.03752006217837334,
-0.04307207465171814,
-0.046344123780727386,
0.020032649859786034,
-0.022683776915073395,
0.04651416465640068,
0.09662359207868576,
0.016924843192100525,
0.03295623138546944,
-0.050403304398059845,
0.0067810299806296825,
-0.05160708352923393,
0.10764911025762558,
0.05846675857901573,
0.013544179499149323,
0.025855842977762222,
0.14576540887355804,
-0.016538729891180992,
-0.018065769225358963,
-0.00022646332217846066,
0.07008080184459686,
0.023960109800100327,
-0.02459828183054924,
0.07271799445152283,
-0.04427722468972206,
0.05875574052333832,
-0.06892900168895721,
-0.09154403954744339,
0.06380832940340042,
0.04181905835866928,
-0.012934105470776558,
-0.01619460992515087,
0.023634033277630806,
-0.021965254098176956,
0.05362898111343384,
0.005192853510379791,
-0.0067084976471960545,
0.02502746880054474,
0.010618473403155804,
-0.022215059027075768,
-0.035839833319187164,
0.05852833390235901,
0.02326085790991783,
-0.02089935913681984,
-1.7495573026219213e-33,
0.04093707725405693,
0.0726657435297966,
-0.021291829645633698,
0.003173856995999813,
-0.0036147632636129856,
-0.08466992527246475,
0.04140550270676613,
0.0697089433670044,
-0.03653382509946823,
-0.027638081461191177,
0.015682348981499672,
-0.013181470334529877,
0.04762427136301994,
-0.01640930585563183,
0.0444558709859848,
0.013940630480647087,
-0.06498014181852341,
-0.07529211044311523,
0.03188283368945122,
-0.011056855320930481,
-0.012539876624941826,
0.05225261300802231,
-0.07943670451641083,
0.05107266828417778,
-0.09274523705244064,
0.057410527020692825,
-0.061185676604509354,
0.017323201522231102,
0.06652259081602097,
-0.0004672383365686983,
-0.0023161068093031645,
-0.06137340888381004,
-0.055675555020570755,
-0.038745272904634476,
-0.028666634112596512,
0.026277318596839905,
0.046717576682567596,
-0.04592370614409447,
-0.024587485939264297,
0.032124023884534836,
0.005807244684547186,
-0.07964003831148148,
-0.01696695201098919,
0.058929115533828735,
0.006312128156423569,
-0.06439992785453796,
0.02972009778022766,
0.02343672513961792,
-0.09403444826602936,
0.00022237970551941544,
-0.08597083389759064,
0.007680319715291262,
-0.14246812462806702,
0.038231223821640015,
-0.029634887352585793,
0.009958307258784771,
0.003643685020506382,
0.005186005029827356,
0.08030789345502853,
-0.0007988752913661301,
0.016114331781864166,
-0.08464853465557098,
-0.03229886665940285,
-0.09479708969593048,
-0.04076391085982323,
-0.007574727758765221,
-0.021803822368383408,
-0.010489121079444885,
0.02437964268028736,
0.03245849162340164,
0.0027808002196252346,
0.00858897902071476,
0.03656897321343422,
0.0313180647790432,
0.011529381386935711,
-0.09851758927106857,
0.04936271160840988,
-0.015609437599778175,
0.09205301105976105,
-0.01792995259165764,
-0.06024707481265068,
-0.04414443299174309,
-0.03976377844810486,
0.039152055978775024,
0.09158175438642502,
0.08764921128749847,
0.0017770479898899794,
0.0081251822412014,
0.05234375596046448,
-0.010920203290879726,
-0.045954588800668716,
0.03803640976548195,
0.007785860449075699,
0.021471217274665833,
-0.02977566421031952,
-4.3404732252838585e-8,
-0.08210784196853638,
0.05029885470867157,
0.05335982143878937,
-0.06981581449508667,
0.06180446594953537,
-0.028318360447883606,
0.062210813164711,
0.08323263376951218,
0.02245941385626793,
-0.032413262873888016,
0.07647959887981415,
-0.021275147795677185,
-0.04980773106217384,
-0.08069039136171341,
0.00360098690725863,
0.03299630060791969,
0.060397662222385406,
-0.008944576606154442,
0.03217611089348793,
-0.023542365059256554,
0.015100590884685516,
-0.012278025969862938,
0.029228227213025093,
0.04972006753087044,
0.01211667712777853,
-0.0794144868850708,
-0.06774498522281647,
0.05526842921972275,
0.02999887429177761,
-0.012338425032794476,
0.010075533762574196,
0.025873886421322823,
0.02563420496881008,
-0.006558467634022236,
0.026685774326324463,
0.09557691216468811,
0.016662560403347015,
0.014203907921910286,
0.07395104318857193,
-0.025553269311785698,
-0.04236377775669098,
0.02466754987835884,
-0.044489786028862,
0.047997791320085526,
-0.07721472531557083,
-0.015009867027401924,
0.03187182545661926,
-0.10136839747428894,
0.009973166510462761,
0.01322353258728981,
0.01734333485364914,
-0.04570813849568367,
-0.019248846918344498,
0.13203871250152588,
0.027270032092928886,
0.010197075083851814,
0.09653229266405106,
-0.11690796911716461,
0.021091273054480553,
0.08999568223953247,
-0.007225682493299246,
-0.029691940173506737,
-0.09688949584960938,
-0.004022047854959965
] |
VAE-Var: Variational Autoencoder-Enhanced Variational Methods for Data Assimilation in Meteorology | https://openreview.net/forum?id=utz99dx2RN | [
"Yi Xiao",
"Qilong Jia",
"Kun Chen",
"LEI BAI",
"Wei Xue"
] | Poster | Data assimilation (DA) is an essential statistical technique for generating accurate estimates of a physical system's states by combining prior model predictions with observational data, especially in the realm of weather forecasting. Effectively modeling the prior distribution while adapting to diverse observational sources presents significant challenges for both traditional and neural network-based DA algorithms. This paper introduces VAE-Var, a novel neural network-based data assimilation algorithm aimed at 1) enhancing accuracy by capturing the non-Gaussian characteristics of the conditional background distribution $p(\mathbf{x}|\mathbf{x}_b)$, and 2) efficiently assimilating real-world observational data. VAE-Var utilizes a variational autoencoder to learn the background error distribution, with its decoder creating a variational cost function to optimize the analysis states. The advantages of VAE-Var include: 1) it maintains the framework of traditional variational assimilation, enabling it to accommodate various observation operators, particularly irregular observations; 2) it lessens the dependence on expert knowledge for constructing the background distribution, allowing for improved modeling of non-Gaussian structures; and 3) experimental results indicate that, when applied to the FengWu weather forecasting model, VAE-Var outperforms DiffDA and two traditional algorithms (interpolation and 3DVar) in terms of assimilation accuracy in sparse observational contexts, and is capable of assimilating real-world GDAS prepbufr observations over a year. | Data assimilation, Variational Autoencoder, Weather Forecasting | null | 2,350 | null | [
-0.050290048122406006,
-0.07701702415943146,
0.10646165907382965,
0.04507054015994072,
0.0329899899661541,
-0.02311311662197113,
0.016623463481664658,
-0.023439833894371986,
-0.005915953312069178,
-0.039395470172166824,
-0.06796760112047195,
-0.037562862038612366,
0.010881869122385979,
-0.014206822961568832,
-0.0779319703578949,
-0.02042263373732567,
0.03059326857328415,
0.03569606691598892,
-0.09219765663146973,
-0.07543309777975082,
-0.047040119767189026,
0.1289399117231369,
-0.11629239469766617,
0.040828052908182144,
0.11830977350473404,
0.05479620397090912,
-0.01620246283710003,
-0.002515987493097782,
0.0299576073884964,
0.054082535207271576,
0.03433007001876831,
-0.010591622442007065,
-0.0484074167907238,
0.056136779487133026,
-0.06638440489768982,
0.003371796803548932,
-0.10231011360883713,
0.04763086512684822,
-0.08377785980701447,
0.014112592674791813,
0.06372241675853729,
0.023056553676724434,
-0.013040319085121155,
-0.018844831734895706,
0.03365570306777954,
0.0032412095461040735,
-0.013360395096242428,
-0.04225840047001839,
0.07992859184741974,
0.06403840333223343,
0.037710368633270264,
-0.007009283639490604,
-0.03828404098749161,
-0.015349742025136948,
-0.014010868035256863,
0.0475715771317482,
-0.011759168468415737,
0.018315471708774567,
-0.011788126081228256,
-0.0026240667793899775,
-0.05939379334449768,
0.03678654134273529,
-0.035377975553274155,
0.02453087829053402,
0.04875459522008896,
0.022829921916127205,
0.0038481266237795353,
0.07177747040987015,
0.026073148474097252,
-0.08839961886405945,
-0.05401434749364853,
0.14362135529518127,
-0.03784246742725372,
0.0012809962499886751,
-0.030211826786398888,
-0.03559955209493637,
0.09712773561477661,
-0.029409822076559067,
0.10425744950771332,
-0.028599059209227562,
0.03723214939236641,
0.0269571952521801,
-0.0005417889915406704,
0.012449356727302074,
0.04988523945212364,
0.04552692547440529,
0.01606057398021221,
0.058981042355298996,
0.031959570944309235,
-0.03807150572538376,
-0.03867299482226372,
-0.05583907291293144,
0.030643466860055923,
0.017053142189979553,
0.020401449874043465,
0.10490065813064575,
0.025927580893039703,
-0.04831390082836151,
0.05947306752204895,
0.0759759396314621,
-0.02587946690618992,
0.011170847341418266,
0.06119370087981224,
0.026881519705057144,
0.03991936147212982,
-0.12689001858234406,
0.0367773100733757,
0.044612471014261246,
-0.02396313287317753,
-0.03255300596356392,
-0.0886959582567215,
0.06608185172080994,
0.03564383462071419,
-0.061346642673015594,
0.005349284037947655,
0.019402677193284035,
0.02745048888027668,
-0.0375480055809021,
-0.08837208151817322,
-0.0203595831990242,
-0.031346485018730164,
-0.04124673455953598,
0.01617581769824028,
0.10676953941583633,
0.07554636150598526,
-0.00033300204086117446,
-0.021335938945412636,
2.2734196661437354e-33,
-0.018208101391792297,
0.013418258167803288,
0.05104346573352814,
-0.010808776132762432,
0.00270565296523273,
0.04509427025914192,
-0.025957180187106133,
-0.05762322619557381,
0.012954949401319027,
0.019558314234018326,
-0.025861987844109535,
0.024942738935351372,
-0.0662468820810318,
0.07709228247404099,
0.053105004131793976,
0.0036973317619413137,
-0.013510447926819324,
-0.014522611163556576,
0.02098538540303707,
0.013252309523522854,
0.005217103753238916,
-0.06704061478376389,
-0.0387411043047905,
0.03151632845401764,
-0.045471206307411194,
-0.003698777174577117,
0.12725509703159332,
0.07358311861753464,
-0.03542330861091614,
-0.018784476444125175,
0.0012901979498565197,
-0.05641314014792442,
0.005114187020808458,
0.013058781623840332,
0.02946079522371292,
-0.09364775568246841,
0.01599951460957527,
-0.030832763761281967,
-0.010536313988268375,
0.0014997374964877963,
0.08634436875581741,
0.04122098162770271,
-0.03123297728598118,
-0.022886693477630615,
0.0007752105593681335,
-0.09514536708593369,
0.04806174337863922,
-0.03500799462199211,
0.04553666710853577,
-0.04769144579768181,
-0.019173868000507355,
-0.019802263006567955,
-0.0008473785128444433,
-0.05508987233042717,
-0.009873776696622372,
0.0317523255944252,
0.0591900497674942,
-0.025203155353665352,
0.06320688128471375,
-0.026834899559617043,
-0.04278198257088661,
-0.005438136402517557,
0.035367708653211594,
0.020202143117785454,
0.08362161368131638,
0.043311312794685364,
0.004278266802430153,
0.09892125427722931,
0.010024121962487698,
-0.04704717546701431,
0.008867420256137848,
-0.006004972383379936,
0.020876605063676834,
-0.041734788566827774,
0.0460180938243866,
-0.0041485801339149475,
-0.008453880436718464,
-0.033724360167980194,
-0.026203284040093422,
-0.010673132725059986,
-0.044824738055467606,
0.05115866661071777,
-0.01261014025658369,
0.08115431666374207,
-0.04898151382803917,
-0.002798045752570033,
0.0009431237122043967,
0.026380613446235657,
-0.04105573520064354,
-0.008851222693920135,
-0.025922656059265137,
-0.02274734526872635,
0.011289924383163452,
-0.055191926658153534,
-0.02859940007328987,
-1.676622837270191e-33,
-0.03168821707367897,
0.07050806283950806,
-0.020379601046442986,
0.03052903711795807,
-0.014318154193460941,
-0.026641903445124626,
-0.010281108319759369,
0.050835613161325455,
0.017153672873973846,
-0.06236351281404495,
-0.05946962535381317,
0.008015349507331848,
0.05963899567723274,
0.014123042114078999,
0.005152328405529261,
0.021481962874531746,
-0.04437795281410217,
0.06766989827156067,
0.004197134170681238,
-0.02037827856838703,
-0.04903832823038101,
-0.013317681849002838,
-0.002461969153955579,
-0.046824533492326736,
-0.039539072662591934,
-0.04517745599150658,
0.006298482418060303,
0.1262299120426178,
-0.0630037933588028,
0.027025755494832993,
-0.050788719207048416,
0.011112747713923454,
0.0005569958011619747,
-0.010890522040426731,
-0.06721153855323792,
0.08555059880018234,
0.0995086133480072,
-0.020665086805820465,
-0.010676481761038303,
-0.010435839183628559,
-0.10471329092979431,
0.0017818192718550563,
0.04028535634279251,
-0.10319890826940536,
-0.01108138170093298,
0.08276734501123428,
0.034564025700092316,
0.019291838631033897,
0.03077714517712593,
-0.02237754315137863,
0.019511176273226738,
0.007312633562833071,
-0.07819662988185883,
0.07045357674360275,
0.0024337300565093756,
-0.009651054628193378,
0.02230404131114483,
-0.05142729729413986,
0.05784411355853081,
-0.04022587090730667,
-0.029491353780031204,
-0.10896113514900208,
0.08740980178117752,
-0.09369532763957977,
-0.09620972722768784,
-0.005534963682293892,
0.007027872838079929,
0.08618418872356415,
0.033143769949674606,
0.03492745757102966,
0.04370206966996193,
-0.0408259741961956,
0.04926585406064987,
0.04628366976976395,
-0.0024536103010177612,
-0.0928836390376091,
0.048969049006700516,
0.011841517873108387,
0.025402173399925232,
-0.0177606213837862,
-0.045584771782159805,
-0.04092145711183548,
0.012452374212443829,
-0.029755661264061928,
0.054531607776880264,
0.03485926240682602,
-0.025512337684631348,
-0.0510486364364624,
0.06258663535118103,
0.0591067299246788,
-0.0890936404466629,
0.06390893459320068,
0.015870166942477226,
0.012756063602864742,
-0.09183194488286972,
-4.5330040165936225e-8,
-0.03875485807657242,
0.03848627209663391,
0.04606339707970619,
-0.01831117644906044,
0.0353982038795948,
-0.04538150876760483,
-0.009512792341411114,
-0.05723821744322777,
0.05269968882203102,
-0.08135003596544266,
0.044728025794029236,
-0.051266860216856,
-0.033404260873794556,
-0.06047471985220909,
0.02270374447107315,
0.03390979766845703,
0.02488880604505539,
0.020289240404963493,
-0.009800434112548828,
0.025712674483656883,
0.04887588694691658,
0.041932445019483566,
-0.0653916522860527,
-0.049496278166770935,
0.201985701918602,
-0.04866502061486244,
-0.02228613942861557,
0.07002726197242737,
0.06876040250062943,
-0.001466773566789925,
-0.04733043909072876,
0.06149044632911682,
-0.010474681854248047,
-0.06996116787195206,
-0.045917313545942307,
0.036381155252456665,
0.01980004832148552,
0.0176712553948164,
-0.07632918655872345,
-0.028114309534430504,
0.032245419919490814,
-0.01923193596303463,
-0.005742168519645929,
0.0042505161836743355,
0.02445162460207939,
0.038681577891111374,
0.07265717536211014,
-0.11708777397871017,
0.08870761841535568,
0.0028005188796669245,
0.0963607206940651,
-0.06198941171169281,
0.009918137453496456,
0.05728284269571304,
0.0923115611076355,
-0.03780519589781761,
0.020359564572572708,
-0.102574922144413,
0.05989145115017891,
0.011921425350010395,
-0.08096206188201904,
0.09049540013074875,
-0.1534460335969925,
-0.010242586024105549
] |
Towards Domain Adaptive Neural Contextual Bandits | https://openreview.net/forum?id=LNkMWCEssX | [
"Ziyan Wang",
"Xiaoming Huo",
"Hao Wang"
] | Poster | Contextual bandit algorithms are essential for solving real-world decision making problems. In practice, collecting a contextual bandit's feedback from different domains may involve different costs. For example, measuring drug reaction from mice (as a source domain) and humans (as a target domain). Unfortunately, adapting a contextual bandit algorithm from a source domain to a target domain with distribution shift still remains a major challenge and largely unexplored. In this paper, we introduce the first general domain adaptation method for contextual bandits. Our approach learns a bandit model for the target domain by collecting feedback from the source domain. Our theoretical analysis shows that our algorithm maintains a sub-linear regret bound even adapting across domains. Empirical results show that our approach outperforms the state-of-the-art contextual bandit algorithms on real-world datasets. Code will soon be available at https://github.com/Wang-ML-Lab/DABand. | Domain Adaptation, Deep Learning, Adversarial Learning | We introduce the first general domain adaptation method for contextual bandits. | 2,344 | 2406.09564 | [
-0.13353843986988068,
-0.07757582515478134,
-0.03520428389310837,
0.019447026774287224,
0.10362562537193298,
-0.06340057402849197,
0.07542620599269867,
0.03642070293426514,
0.03038722090423107,
-0.03171633556485176,
-0.05851733684539795,
-0.03341543301939964,
0.10727904736995697,
-0.0379011407494545,
0.027002835646271706,
-0.015199985355138779,
0.12208624184131622,
0.0630340650677681,
-0.0012809049803763628,
-0.012084625661373138,
-0.028548551723361015,
0.03686752915382385,
0.11042105406522751,
-0.031414881348609924,
0.01838565059006214,
-0.026962822303175926,
-0.00890736561268568,
-0.05496389418840408,
0.0359211340546608,
0.037758804857730865,
0.0748715028166771,
0.047572020441293716,
-0.022953597828745842,
-0.07348892092704773,
-0.04462669417262077,
0.03422543779015541,
-0.16924582421779633,
0.03255815431475639,
0.05993896722793579,
0.04777459800243378,
-0.0089087700471282,
0.03205806016921997,
0.012841255404055119,
0.015906263142824173,
0.013923087157309055,
-0.06838883459568024,
-0.04335290938615799,
0.003175550140440464,
0.11172541230916977,
0.018210837617516518,
-0.0763934925198555,
-0.04000566154718399,
-0.027638182044029236,
-0.00159748294390738,
-0.06049667298793793,
0.03604989871382713,
0.05089391767978668,
0.07722832262516022,
0.008086646907031536,
0.0058278911747038364,
0.028269175440073013,
-0.14430314302444458,
-0.06432108581066132,
-0.01277005486190319,
-0.027181338518857956,
-0.015874309465289116,
-0.018858235329389572,
0.10466594249010086,
-0.002969516906887293,
0.009930586442351341,
0.025509968400001526,
0.0600726455450058,
-0.02695482037961483,
0.021546941250562668,
0.0946701318025589,
0.024909760802984238,
0.018772898241877556,
-0.0043857344426214695,
0.06484191864728928,
-0.04944771155714989,
-0.020547162741422653,
-0.035581301897764206,
0.05925017595291138,
-0.010393347591161728,
0.11085017770528793,
-0.018960880115628242,
0.04794560745358467,
0.009364728815853596,
0.08184371143579483,
-0.037638939917087555,
0.007188461720943451,
-0.04682661592960358,
-0.02414887584745884,
0.005446741823107004,
-0.04923601076006889,
0.05507565289735794,
-0.009346868842840195,
-0.03158186003565788,
-0.02815626747906208,
0.0795239582657814,
-0.010569959878921509,
0.037132419645786285,
-0.031853172928094864,
-0.12254989892244339,
0.0067275515757501125,
-0.03748834878206253,
0.02817162685096264,
0.1020115464925766,
0.06559975445270538,
-0.06445086747407913,
0.0011363443918526173,
0.0860997810959816,
0.060210391879081726,
0.00387467909604311,
0.006391788832843304,
-0.07329441606998444,
0.03134031221270561,
-0.07006119936704636,
-0.06070815771818161,
0.06168746203184128,
-0.02245219051837921,
0.019350383430719376,
-0.008797847665846348,
-0.06204385310411453,
-0.020381499081850052,
-0.015427475795149803,
0.01631130464375019,
1.1998444407930827e-33,
0.035620227456092834,
-0.0058587887324392796,
0.02552097477018833,
-0.06616479158401489,
0.023099975660443306,
0.010138481855392456,
0.06849994510412216,
-0.06593017280101776,
-0.02927876077592373,
0.010451427660882473,
0.019121503457427025,
0.08041208982467651,
-0.09562627971172333,
0.06436988711357117,
0.05798233300447464,
-0.03338965028524399,
-0.04136928170919418,
0.012291845865547657,
0.12115591764450073,
-0.028557896614074707,
-0.012941695749759674,
0.011430667713284492,
-0.047094814479351044,
-0.0462518110871315,
0.03499538078904152,
-0.008301980793476105,
-0.042704444378614426,
0.015934685245156288,
-0.002460240386426449,
0.014708567410707474,
-0.10942365229129791,
0.05571715533733368,
-0.0562838539481163,
0.04053241014480591,
0.034791234880685806,
-0.0033229342661798,
-0.04252803698182106,
-0.014131958596408367,
-0.10349280387163162,
-0.004138711374253035,
-0.05884408950805664,
0.043474677950143814,
0.06398285925388336,
0.01642581634223461,
-0.08330199867486954,
-0.07462480664253235,
-0.06634580343961716,
-0.03912581875920296,
-0.01525105256587267,
0.004121322184801102,
0.02539520151913166,
0.02013370580971241,
0.08870300650596619,
-0.01807383820414543,
-0.049556147307157516,
0.0026298111770302057,
-0.045374684035778046,
0.007964732125401497,
0.009142298251390457,
0.033922042697668076,
0.08113045245409012,
-0.011579145677387714,
0.009039217606186867,
0.08575394749641418,
-0.03931359946727753,
0.028929123654961586,
-0.04334791377186775,
-0.017835630103945732,
-0.008037634193897247,
0.01299041323363781,
0.019971899688243866,
0.06946470588445663,
0.007690880913287401,
0.021206818521022797,
0.08648081123828888,
-0.03692187741398811,
0.029318593442440033,
-0.0023436499759554863,
-0.019609613344073296,
-0.03453307971358299,
0.015797892585396767,
-0.03305598348379135,
-0.058915574103593826,
0.014874184504151344,
-0.01754125952720642,
0.050893377512693405,
0.013617204502224922,
-0.14934895932674408,
-0.021555624902248383,
-0.061138030141592026,
-0.09945496916770935,
-0.03973471373319626,
-0.06859257817268372,
0.008503187447786331,
0.08091451972723007,
-1.4114113834220166e-33,
-0.0035150444600731134,
-0.016951505094766617,
0.09614890068769455,
0.01612675189971924,
0.05575726926326752,
0.03178488835692406,
-0.011640592478215694,
-0.006567367818206549,
0.02179720625281334,
-0.026341188699007034,
-0.10975009948015213,
0.008976278826594353,
-0.007785114925354719,
-0.006682582665234804,
-0.025165535509586334,
-0.0675656795501709,
0.01608477346599102,
0.03532722592353821,
-0.05655675381422043,
-0.04426806792616844,
0.031154023483395576,
0.07245027273893356,
-0.09665059298276901,
0.017809445038437843,
-0.016413560137152672,
0.0032150670886039734,
-0.029318001121282578,
0.12345758825540543,
-0.02985120192170143,
-0.049519721418619156,
-0.05867395177483559,
0.01042516715824604,
-0.024839915335178375,
0.016608871519565582,
-0.04747668653726578,
0.04259292036294937,
0.09052116423845291,
0.003462243126705289,
-0.11003521084785461,
0.07151425629854202,
0.05343877151608467,
-0.024769600480794907,
0.018375538289546967,
0.023288486525416374,
0.0511297732591629,
0.031082749366760254,
-0.07480236142873764,
-0.01406196877360344,
0.030730409547686577,
-0.011589989066123962,
-0.011724790558218956,
0.06772290170192719,
-0.06307830661535263,
0.007531004026532173,
-0.04294339567422867,
-0.10162214934825897,
0.01608513481914997,
-0.04534586891531944,
0.10864812880754471,
0.007928186096251011,
-0.07822437584400177,
0.016310133039951324,
-0.03146767616271973,
0.052631936967372894,
0.0029292243998497725,
0.002356251934543252,
0.029793042689561844,
0.06516718864440918,
0.01064287405461073,
0.023593515157699585,
0.04739956930279732,
0.029917828738689423,
0.011368883773684502,
-0.026609109714627266,
0.008018309250473976,
0.025457583367824554,
0.03194081038236618,
-0.00787552073597908,
0.01154960598796606,
-0.038857027888298035,
0.10553324222564697,
-0.03392606973648071,
0.08419002592563629,
-0.002055333461612463,
0.049707893282175064,
0.11521575599908829,
0.045738570392131805,
0.04360184073448181,
-0.015506355091929436,
-0.06004074588418007,
-0.010092431679368019,
-0.08283861726522446,
0.02081426978111267,
-0.005451285745948553,
-0.030625829473137856,
-3.923510405456909e-8,
-0.05034121125936508,
0.07175810635089874,
0.028659839183092117,
0.06875676661729813,
0.027011442929506302,
0.06101429834961891,
-0.0631600022315979,
0.025398574769496918,
-0.05900830775499344,
0.02316790260374546,
0.03063962049782276,
0.0485345683991909,
0.023332390934228897,
-0.0036117038689553738,
-0.07439982891082764,
0.039296939969062805,
0.009417856112122536,
0.011858630925416946,
-0.0006112545379437506,
-0.04083312675356865,
0.022179456427693367,
0.05133134871721268,
0.10152466595172882,
-0.06814279407262802,
0.08355263620615005,
-0.0966009795665741,
-0.012373385950922966,
0.08434191346168518,
0.08085610717535019,
0.010912055149674416,
-0.0344427190721035,
-0.0007736787665635347,
-0.03943553939461708,
-0.0019491243874654174,
0.0023960149846971035,
-0.037911925464868546,
0.016210755333304405,
-0.0625225156545639,
0.002869305433705449,
-0.04955406114459038,
-0.03638260066509247,
0.016883915290236473,
-0.00970893632620573,
-0.006547203287482262,
-0.025272773578763008,
0.0428650788962841,
0.05932262912392616,
-0.03257670998573303,
0.06014518067240715,
-0.09177248179912567,
0.015619519166648388,
-0.03606612980365753,
-0.0006989458343014121,
-0.01309931930154562,
0.07534882426261902,
-0.015720317140221596,
-0.06104641780257225,
-0.04948049783706665,
0.025339990854263306,
0.046496085822582245,
-0.04872657358646393,
-0.02229832299053669,
-0.07081468403339386,
0.03195762634277344
] |
Phidias: A Generative Model for Creating 3D Content from Text, Image, and 3D Conditions with Reference-Augmented Diffusion | https://openreview.net/forum?id=TEkoMEjf7E | [
"Zhenwei Wang",
"Tengfei Wang",
"Zexin He",
"Gerhard Petrus Hancke",
"Ziwei Liu",
"Rynson W. H. Lau"
] | Poster | Generative 3D modeling has made significant advances recently, but it remains constrained by its inherently ill-posed nature, leading to challenges in quality and controllability. Inspired by the real-world workflow that designers typically refer to existing 3D models when creating new ones, we propose Phidias, a novel generative model that uses diffusion for reference-augmented 3D generation. Given an image, our method leverages a retrieved or user-provided 3D reference model to guide the generation process, thereby enhancing the generation quality, generalization ability, and controllability. Phidias integrates three key components: 1) meta-ControlNet to dynamically modulate the conditioning strength, 2) dynamic reference routing to mitigate misalignment between the input image and 3D reference, and 3) self-reference augmentations to enable self-supervised training with a progressive curriculum. Collectively, these designs result in significant generative improvements over existing methods. Phidias forms a unified framework for 3D generation using text, image, and 3D conditions, offering versatile applications. | 3D generation, retrieval-augmented generation, multi-view diffusion | A 3D diffusion model with RAG, supporting reference-augmented 3D generation from text, image, and 3D conditions. | 2,341 | 2409.11406 | [
-0.04470088332891464,
-0.08933722972869873,
-0.023526372388005257,
0.049357812851667404,
-0.005638448987156153,
0.020418601110577583,
-0.034540820866823196,
-0.01754121482372284,
0.009236617013812065,
-0.029274264350533485,
-0.030336417257785797,
-0.022116132080554962,
0.025306550785899162,
0.017499098554253578,
-0.009833957068622112,
0.004527647513896227,
0.0878213495016098,
0.06855712831020355,
-0.08676637709140778,
-0.0037341804709285498,
0.04570852592587471,
0.004782271105796099,
0.028746241703629494,
0.013491522520780563,
0.00461047887802124,
0.07683534175157547,
0.01446675043553114,
-0.07117290794849396,
0.06994941830635071,
-0.08412384241819382,
0.01732654497027397,
0.024045897647738457,
-0.0739220604300499,
-0.023848995566368103,
-0.07261195778846741,
0.02564244158565998,
-0.016098108142614365,
0.10300394892692566,
-0.06571342051029205,
-0.04742755740880966,
-0.01507024746388197,
0.0593249648809433,
0.0017829202115535736,
-0.002626243978738785,
0.05342564731836319,
0.031112341210246086,
-0.03604591265320778,
-0.07296153157949448,
0.010499277152121067,
0.04903509467840195,
-0.0815185084939003,
-0.08929818123579025,
-0.032997976988554,
-0.0013060602359473705,
0.02406449057161808,
0.12700237333774567,
0.09325999766588211,
-0.05309966951608658,
-0.004908282775431871,
-0.00860433280467987,
-0.03656398877501488,
-0.010368931107223034,
-0.009091117419302464,
-0.05002487078309059,
0.014333182945847511,
-0.032229695469141006,
0.005835469346493483,
0.03051299974322319,
-0.041237056255340576,
-0.027547283098101616,
-0.0011004318948835135,
0.027133047580718994,
-0.007987722754478455,
-0.03321034088730812,
0.0589430071413517,
0.07020381838083267,
0.004056830890476704,
-0.06366639584302902,
0.026245782151818275,
-0.046367812901735306,
0.031626056879758835,
0.03358818590641022,
0.040404271334409714,
0.027690859511494637,
-0.03579959273338318,
0.0669671893119812,
-0.00391683354973793,
0.09637169539928436,
0.019730405882000923,
0.013446533121168613,
-0.023362861946225166,
-0.01573816128075123,
-0.03076082281768322,
-0.06211702898144722,
0.0043189674615859985,
-0.008548145182430744,
0.027549389749765396,
-0.1598660945892334,
-0.006108159199357033,
0.028114615008234978,
0.036537908017635345,
0.006744409445673227,
0.007117651402950287,
0.009899325668811798,
0.051487360149621964,
-0.031152743846178055,
0.08563201129436493,
0.013062508776783943,
-0.005805459804832935,
0.00859297625720501,
0.033346157521009445,
-0.0507824681699276,
-0.037945326417684555,
-0.0402667336165905,
-0.0013305554166436195,
-0.0016114027239382267,
-0.012496918439865112,
-0.05099404603242874,
0.05915118008852005,
-0.006902709603309631,
-0.018893882632255554,
0.010713905096054077,
-0.033406030386686325,
0.04190140590071678,
-0.038995642215013504,
-0.10328291356563568,
-0.02801930159330368,
2.216256845231186e-34,
0.035783711820840836,
0.043418027460575104,
0.01224013976752758,
0.12105949968099594,
0.060291118919849396,
-0.0008486144943162799,
-0.03025616705417633,
-0.001441202824935317,
-0.037658706307411194,
-0.1121528372168541,
-0.007212195545434952,
-0.03327174112200737,
-0.10112686455249786,
0.18696704506874084,
0.07856348901987076,
-0.012203660793602467,
-0.09667257964611053,
0.06785409897565842,
-0.022520458325743675,
0.05830464884638786,
0.00016021952615119517,
-0.06305322051048279,
-0.018024880439043045,
0.01695439964532852,
0.016223367303609848,
0.02865273505449295,
0.038391634821891785,
0.06417284160852432,
-0.01469478290528059,
-0.023485664278268814,
-0.05084068700671196,
-0.0037169388961046934,
0.022302569821476936,
-0.020100736990571022,
0.009208325296640396,
0.021481415256857872,
-0.03686662018299103,
-0.029697874560952187,
0.01734844595193863,
0.04353659227490425,
0.012907731346786022,
0.01953227072954178,
0.046568356454372406,
0.019662607461214066,
-0.10708601772785187,
0.013215247541666031,
0.06612135469913483,
0.07886584103107452,
-0.13821451365947723,
-0.02913340926170349,
0.0460544154047966,
0.03971480578184128,
-0.009359840303659439,
-0.05788262188434601,
-0.06998424977064133,
-0.06287381052970886,
0.03693578764796257,
0.013381928205490112,
0.054069411009550095,
0.01711718924343586,
0.03843868896365166,
0.07368357479572296,
-0.039400745183229446,
0.08609693497419357,
0.04871746525168419,
0.021433357149362564,
0.05246734991669655,
0.08181900531053543,
0.07297180593013763,
-0.026603683829307556,
-0.04019082337617874,
-0.02743474952876568,
-0.053904686123132706,
-0.028360864147543907,
0.046645063906908035,
-0.040838297456502914,
0.017430467531085014,
-0.08043726533651352,
0.003071469720453024,
0.018994035199284554,
-0.088857501745224,
0.007194157689809799,
-0.027454568073153496,
-0.039868053048849106,
-0.056980960071086884,
-0.08688293397426605,
0.014483808539807796,
-0.07706673443317413,
-0.016930000856518745,
-0.05829974263906479,
-0.01237897016108036,
-0.03236421197652817,
-0.018381919711828232,
0.07964503765106201,
0.002323887310922146,
-5.786419155558193e-34,
0.02947986125946045,
0.03532366082072258,
-0.031991664320230484,
0.019496776163578033,
0.008683969266712666,
0.005860366392880678,
0.009820139966905117,
0.05911429971456528,
0.04343383386731148,
-0.09860694408416748,
-0.009549714624881744,
0.05739076808094978,
-0.030131688341498375,
-0.04012987017631531,
-0.02821396104991436,
-0.03536965325474739,
-0.019658606499433517,
-0.04319965839385986,
-0.03957000747323036,
0.016874870285391808,
0.031252048909664154,
0.0498339980840683,
-0.0699683353304863,
0.020415080711245537,
-0.10144242644309998,
0.06568349152803421,
-0.015980027616024017,
0.0197772067040205,
0.04812662675976753,
0.07042834907770157,
-0.013807741925120354,
0.007107127923518419,
0.07354041188955307,
-0.04878756403923035,
-0.06780093163251877,
0.05864601954817772,
0.07176999002695084,
-0.008611179888248444,
-0.04132426902651787,
0.045690227299928665,
0.028847169131040573,
-0.014468117617070675,
-0.06335338205099106,
0.06691715121269226,
-0.021949609741568565,
0.017467370256781578,
-0.0007221776759251952,
0.033531539142131805,
0.04906812682747841,
0.10784287750720978,
-0.031074414029717445,
0.02334943599998951,
-0.0836922824382782,
-0.0047952947206795216,
-0.00391066400334239,
-0.1612301468849182,
-0.08216400444507599,
-0.06556791812181473,
0.15013988316059113,
0.002711867680773139,
-0.07862859964370728,
-0.08308406919240952,
0.018241945654153824,
-0.06624045222997665,
-0.05132259055972099,
-0.01629214733839035,
-0.057753268629312515,
-0.035450100898742676,
0.04809744656085968,
0.09268542379140854,
0.027964992448687553,
0.07894469797611237,
0.054699111729860306,
-0.02693132311105728,
0.03827128931879997,
-0.12167433649301529,
0.01874716766178608,
0.032431088387966156,
0.004948582965880632,
-0.064949631690979,
0.0529099702835083,
0.014229955151677132,
0.02389843948185444,
0.11040698736906052,
0.07781393826007843,
0.01273437775671482,
-0.012766758911311626,
0.05389280617237091,
-0.024814490228891373,
0.044139374047517776,
-0.02714209258556366,
0.05907652527093887,
0.008503067307174206,
0.08955537527799606,
-0.06738541275262833,
-4.423640831419107e-8,
0.009743096306920052,
0.052419573068618774,
0.06298749893903732,
0.06576187163591385,
-0.06307797878980637,
-0.011471563950181007,
0.05360310524702072,
0.03043184243142605,
0.03691812977194786,
-0.03313848376274109,
0.012134755961596966,
-0.061425644904375076,
-0.023882213979959488,
-0.049115102738142014,
0.060186296701431274,
0.07784099876880646,
0.014310508966445923,
0.07130873203277588,
0.019882215186953545,
-0.04077785462141037,
-0.045747265219688416,
-0.0005814394098706543,
-0.0028852992691099644,
-0.024725865572690964,
0.06758002191781998,
0.014011328108608723,
-0.01990344189107418,
-0.013446260243654251,
0.054376620799303055,
-0.019262200221419334,
0.07292047888040543,
0.012815559282898903,
0.028010761365294456,
0.013144840486347675,
-0.03528057038784027,
0.013700461015105247,
-0.007937222719192505,
-0.05179887264966965,
0.014956674538552761,
-0.09281733632087708,
-0.024904388934373856,
0.15435004234313965,
-0.0419180728495121,
-0.01617874577641487,
0.02637191116809845,
-0.012777351774275303,
0.044767726212739944,
-0.10511987656354904,
-0.00003832842776319012,
0.005445118993520737,
-0.08619209378957748,
-0.038638416677713394,
0.03987373784184456,
0.06902637332677841,
0.07274416089057922,
0.023346412926912308,
0.019075309857726097,
-0.001060068723745644,
0.037291038781404495,
0.06787481158971786,
0.009278534911572933,
0.008791090920567513,
-0.023907266557216644,
0.02399815432727337
] |
FLIP: Flow-Centric Generative Planning as General-Purpose Manipulation World Model | https://openreview.net/forum?id=B2N0nCVC91 | [
"Chongkai Gao",
"Haozhuo Zhang",
"Zhixuan Xu",
"Cai Zhehao",
"Lin Shao"
] | Poster | We aim to develop a model-based planning framework for world models that can be scaled with increasing model and data budgets for general-purpose manipulation tasks with only language and vision inputs. To this end, we present FLow-CentrIc generative Planning (FLIP), a model-based planning algorithm on visual space that features three key modules: 1) a multi-modal flow generation model as the general-purpose action proposal module; 2) a flow-conditioned video generation model as the dynamics module; and 3) a vision-language representation learning model as the value module. Given an initial image and language instruction as the goal, FLIP can progressively search for long-horizon flow and video plans that maximize the discounted return to accomplish the task. FLIP is able to synthesize long-horizon plans across objects, robots, and tasks with image flows as the general action representation, and the dense flow information also provides rich guidance for long-horizon video generation. In addition, the synthesized flow and video plans can guide the training of low-level control policies for robot execution. Experiments on diverse benchmarks demonstrate that FLIP can improve both the success rates and quality of long-horizon video plan synthesis and has the interactive world model property, opening up wider applications for future works. Video demos are on our website: https://nus-lins-lab.github.io/flipweb/. | World Model, Long-Horizon Planning, Robot Manipulation, Flow Generation | We train an interactive world-model that enables model-based planning on the image flow and video spaces for diverse manipualtion tasks with only language-annotated video datasets. | 2,338 | 2412.08261 | [
0.006141391582787037,
-0.10232972353696823,
0.053944822400808334,
-0.03035541996359825,
0.015286358073353767,
0.026471873745322227,
-0.048542048782110214,
0.0005325247766450047,
0.04818561673164368,
-0.017574187368154526,
0.0011847740970551968,
-0.05553218349814415,
0.004003541544079781,
0.05010423809289932,
0.00999717228114605,
0.019355377182364464,
0.06715167313814163,
0.08195388317108154,
-0.07415178418159485,
-0.03473657742142677,
0.03730139508843422,
-0.01592596434056759,
0.10702258348464966,
-0.02440311200916767,
-0.009502320550382137,
0.10270044207572937,
0.05827653408050537,
-0.08000212907791138,
0.019297974184155464,
-0.04032258689403534,
0.058600012212991714,
0.008572286926209927,
-0.040050797164440155,
-0.04115626960992813,
-0.05294535309076309,
0.1272803395986557,
-0.06440692394971848,
-0.003263816935941577,
-0.058326903730630875,
-0.07027027755975723,
-0.040831275284290314,
0.028403786942362785,
0.08812245726585388,
-0.02720331773161888,
0.04850466176867485,
0.05284906178712845,
-0.01366178784519434,
-0.057213861495256424,
0.009415913373231888,
0.021631626412272453,
-0.07115841656923294,
-0.07750079035758972,
0.01528783980756998,
-0.03217357024550438,
-0.03210247680544853,
0.04100216180086136,
0.06872130930423737,
-0.06201672554016113,
0.07714978605508804,
-0.023211795836687088,
-0.026041075587272644,
-0.0782928615808487,
0.023060426115989685,
-0.05827683210372925,
-0.07327686250209808,
-0.08433221280574799,
0.03999731317162514,
-0.028392309322953224,
-0.013475066050887108,
-0.013194643892347813,
-0.0068398527801036835,
0.007426070049405098,
0.043408770114183426,
0.011439134366810322,
-0.029056748375296593,
-0.021090468391776085,
0.054086003452539444,
-0.04838033765554428,
0.013161900453269482,
-0.03822711110115051,
0.08311443030834198,
-0.001553631853312254,
0.0002981688012368977,
0.04158683121204376,
0.04548034816980362,
-0.017431989312171936,
-0.011699584312736988,
0.11842607706785202,
0.0615997277200222,
0.04976986348628998,
-0.03977107256650925,
-0.00851301196962595,
0.004576748237013817,
-0.07904936373233795,
0.031076043844223022,
-0.017789237201213837,
-0.040148332715034485,
-0.16782575845718384,
-0.03953952342271805,
-0.008209583349525928,
0.001452979282476008,
-0.005192901007831097,
0.03144605830311775,
-0.011997667141258717,
0.05710108205676079,
-0.040319330990314484,
0.006536749191582203,
0.052243806421756744,
-0.024070721119642258,
-0.033610522747039795,
0.0497853122651577,
-0.04606843367218971,
0.10228808969259262,
0.017184631898999214,
0.013275135308504105,
0.022816821932792664,
-0.05339481681585312,
0.0065397098660469055,
-0.052006933838129044,
0.09147913008928299,
0.013384096324443817,
-0.00927791278809309,
-0.00798794999718666,
0.0068629733286798,
0.03646780923008919,
-0.1068943440914154,
0.03282802924513817,
-1.20610780530205e-33,
0.05777357891201973,
0.03527294099330902,
0.03737001121044159,
0.056196216493844986,
0.07357674837112427,
0.03717362880706787,
0.06440529227256775,
0.0477852001786232,
-0.04664774611592293,
-0.0233774296939373,
-0.0018212534487247467,
-0.07281024008989334,
-0.1100834533572197,
0.16350771486759186,
0.01488055381923914,
-0.024681156501173973,
0.006097343750298023,
0.03463512286543846,
-0.050846733152866364,
0.05685725435614586,
0.02662591077387333,
-0.001389855518937111,
0.002361971652135253,
-0.02378002181649208,
0.011799359694123268,
-0.0018684100359678268,
0.019680911675095558,
-0.018079521134495735,
-0.01698698103427887,
-0.03804950788617134,
-0.10595934092998505,
0.01682133413851261,
-0.01852680742740631,
0.05203109234571457,
0.036340996623039246,
0.034909091889858246,
-0.057567957788705826,
-0.0572674386203289,
-0.020141873508691788,
0.010157657787203789,
0.011305613443255424,
0.0734371617436409,
-0.060909464955329895,
-0.011667252518236637,
-0.06323812156915665,
-0.04444557800889015,
0.0499374195933342,
0.06940653920173645,
-0.09966525435447693,
-0.026273095980286598,
0.06508800387382507,
0.05782666429877281,
-0.02673896960914135,
-0.07226204127073288,
-0.011716105975210667,
-0.044017378240823746,
0.05515972524881363,
0.05994698405265808,
-0.013360356912016869,
0.06084785237908363,
0.04564717411994934,
0.059106018394231796,
-0.07532770186662674,
0.09098993986845016,
-0.023524228483438492,
0.007552932482212782,
0.05373547971248627,
0.058677807450294495,
0.10473130643367767,
-0.04752085730433464,
-0.04563023895025253,
0.042033448815345764,
0.011270921677350998,
-0.03943323716521263,
-0.0008848732686601579,
0.020909812301397324,
0.0417647510766983,
-0.08697014302015305,
0.00969270896166563,
0.07133830338716507,
-0.06908685714006424,
-0.008062665350735188,
0.0346563383936882,
0.010495406575500965,
0.08157501369714737,
-0.031261805444955826,
0.044002339243888855,
-0.011989934369921684,
-0.041836511343717575,
-0.06250942498445511,
-0.07233081758022308,
-0.06446520984172821,
-0.0396648570895195,
0.03310294821858406,
0.011338223703205585,
4.635115651956363e-34,
0.026766348630189896,
-0.013183921575546265,
-0.02606937848031521,
0.04872932657599449,
-0.01956184208393097,
-0.07206639647483826,
0.008571150712668896,
-0.022291207686066628,
0.023386364802718163,
-0.1037599965929985,
-0.08616992831230164,
-0.0031803755555301905,
0.0005265430081635714,
-0.00755976652726531,
0.050402481108903885,
-0.01646026223897934,
0.03365794196724892,
-0.046319857239723206,
-0.01496972143650055,
-0.005061606410890818,
-0.013867517001926899,
0.11941064894199371,
-0.05186609923839569,
-0.015448641031980515,
-0.09274640679359436,
0.059717606753110886,
0.000142076110932976,
0.023885272443294525,
0.04809503257274628,
0.06764943152666092,
-0.006682237610220909,
-0.05291331559419632,
0.03075724095106125,
0.03990849852561951,
-0.06455644965171814,
0.02046116627752781,
0.024368053302168846,
-0.004215162247419357,
-0.05106603354215622,
0.12128125876188278,
0.05925385653972626,
-0.06802186369895935,
-0.03626561537384987,
0.04881250113248825,
-0.039805520325899124,
-0.024392927065491676,
-0.03561394661664963,
0.0055645969696342945,
-0.015680918470025063,
-0.002523497212678194,
-0.10246875882148743,
0.07193047553300858,
-0.09097538888454437,
-0.06648720800876617,
-0.03278970345854759,
-0.08864227682352066,
-0.0244965311139822,
-0.027651214972138405,
0.13921910524368286,
-0.04218028113245964,
-0.02397298254072666,
0.026399526745080948,
0.014999572187662125,
0.0077425516210496426,
-0.007045958656817675,
0.05700894817709923,
-0.0527636781334877,
-0.059594955295324326,
0.03509870171546936,
-0.007925890386104584,
0.029154270887374878,
0.023988746106624603,
0.05337340757250786,
0.008006177842617035,
0.0652240663766861,
0.005273542366921902,
0.015947137027978897,
0.02645675279200077,
0.054969098418951035,
-0.006336679682135582,
-0.00186539173591882,
0.01122602540999651,
0.02941492386162281,
0.11819964647293091,
0.027541697025299072,
0.06427975744009018,
-0.05128201097249985,
0.019422508776187897,
0.004956417251378298,
0.04852767288684845,
-0.012939944863319397,
-0.0265590138733387,
0.015855057165026665,
0.06032440811395645,
-0.046615611761808395,
-3.764348832646647e-8,
-0.0886220708489418,
0.016435598954558372,
0.06261932104825974,
0.0013374047121033072,
0.017027441412210464,
0.03288089483976364,
0.041211094707250595,
0.000319336773827672,
0.04898858442902565,
-0.018803635612130165,
0.058094266802072525,
0.03714917600154877,
0.014213595539331436,
-0.08064369857311249,
-0.033496588468551636,
0.026403650641441345,
0.03423916548490524,
-0.009825482033193111,
0.03510570898652077,
0.033308807760477066,
0.026685556396842003,
-0.05595319718122482,
0.005740915425121784,
-0.010372176766395569,
-0.019575200974941254,
-0.03754705935716629,
-0.055447157472372055,
0.03895099461078644,
0.07417234778404236,
-0.046950213611125946,
0.02072839066386223,
-0.04363417997956276,
0.09679901599884033,
0.02311711013317108,
0.019125789403915405,
-0.030775735154747963,
-0.02991189807653427,
-0.03864853456616402,
0.021615197882056236,
-0.017301831394433975,
0.0451781265437603,
0.11596885323524475,
-0.027149250730872154,
-0.014235730282962322,
-0.04623057320713997,
0.06606006622314453,
-0.015855394303798676,
-0.16576501727104187,
-0.008394372649490833,
-0.03208465874195099,
-0.047118041664361954,
0.0012357020750641823,
-0.022684084251523018,
0.1264697015285492,
0.10405254364013672,
0.06561258435249329,
0.04062262549996376,
-0.10215383023023605,
-0.002594390884041786,
0.049700114876031876,
-0.06271819770336151,
0.030027206987142563,
-0.029194053262472153,
-0.03490552678704262
] |
Causal Graph Transformer for Treatment Effect Estimation Under Unknown Interference | https://openreview.net/forum?id=foQ4AeEGG7 | [
"Anpeng Wu",
"Haiyi Qiu",
"Zhengming Chen",
"Zijian Li",
"Ruoxuan Xiong",
"Fei Wu",
"Kun Zhang"
] | Poster | Networked interference, also known as the peer effect in social science and spillover effect in economics, has drawn increasing interest across various domains. This phenomenon arises when a unit’s treatment and outcome are influenced by the actions of its peers, posing significant challenges to causal inference, particularly in treatment assignment and effect estimation in real applications, due to the violation of the SUTVA assumption. While extensive graph models have been developed to identify treatment effects, these models often rely on structural assumptions about networked interference, assuming it to be identical to the social network, which can lead to misspecification issues in real applications. To address these challenges, we propose an Interference-Agnostic Causal Graph Transformer (CauGramer), which aggregates peers information via $L$-order Graph Transformer and employs cross-attention to infer aggregation function for learning interference representations. By integrating confounder balancing and minimax moment constraints, CauGramer fully incorporates peer information, enabling robust treatment effect estimation. Extensive experiments on two widely-used benchmarks demonstrate the effectiveness and superiority of CauGramer. The code is available at https://github.com/anpwu/CauGramer. | Causal Graph Transformer, Networked Interference, Unknown Interference Graph, Peer Effects, Treatment Effects Estimation | null | 2,331 | null | [
-0.11051659286022186,
-0.0024944005999714136,
0.07266269624233246,
0.0331808440387249,
0.08396446704864502,
-0.015365111641585827,
-0.05719136819243431,
0.028274456039071083,
0.01934283971786499,
0.03915439918637276,
-0.027559371665120125,
0.07772833108901978,
0.04481892287731171,
-0.017840571701526642,
-0.04093167558312416,
0.04894084855914116,
0.08666722476482391,
-0.000506920157931745,
-0.06178397312760353,
-0.09056083858013153,
0.03206830099225044,
-0.045966729521751404,
-0.07163165509700775,
-0.007852504029870033,
0.06463130563497543,
-0.029548775404691696,
-0.047464609146118164,
-0.06762398034334183,
0.01489472109824419,
-0.029126249253749847,
0.014954163692891598,
0.0330018512904644,
-0.13161636888980865,
-0.042017072439193726,
-0.1145743653178215,
-0.014217483811080456,
-0.0181192047894001,
0.028356807306408882,
0.011931709945201874,
0.009132077917456627,
0.07748337090015411,
0.024751557037234306,
-0.009757966734468937,
0.025317370891571045,
0.039644863456487656,
-0.01119745522737503,
0.03482746332883835,
0.025914039462804794,
-0.022902842611074448,
0.0056204162538051605,
-0.02627534046769142,
0.000022211701434571296,
0.002266277326270938,
0.06636852771043777,
0.06896605342626572,
-0.027331385761499405,
0.04289491847157478,
-0.04427936300635338,
-0.03122185915708542,
0.09198343753814697,
-0.06930441409349442,
-0.04773061349987984,
-0.03621901199221611,
0.002271846402436495,
-0.004404248669743538,
0.07198129594326019,
0.003577784402295947,
0.10122231394052505,
-0.07122179120779037,
0.04541920870542526,
0.04451501742005348,
0.009204541333019733,
-0.05596916005015373,
-0.005890095140784979,
0.11104832589626312,
0.01157602109014988,
0.031049927696585655,
0.04648977145552635,
0.055575188249349594,
-0.02731247805058956,
0.04563485458493233,
-0.03842181712388992,
-0.03780549392104149,
-0.027322646230459213,
0.09086061269044876,
-0.06567108631134033,
0.019299479201436043,
-0.009459234774112701,
0.03303280845284462,
0.08995746821165085,
-0.05250523239374161,
0.06094026938080788,
0.040093716233968735,
0.009617405012249947,
0.03074057400226593,
0.0062324595637619495,
-0.01543193869292736,
-0.02651762031018734,
0.030962977558374405,
0.1210286095738411,
-0.037197209894657135,
-0.0083995396271348,
0.039039526134729385,
-0.04382956773042679,
-0.05561918392777443,
-0.05713486298918724,
-0.030791867524385452,
0.07646104693412781,
0.032730959355831146,
-0.004022622015327215,
-0.030181515961885452,
0.014289367944002151,
0.027288440614938736,
0.03797634318470955,
-0.013335730880498886,
-0.06373319029808044,
0.13656960427761078,
0.024011757224798203,
0.03394709900021553,
-0.04694728925824165,
0.028737464919686317,
0.03812668099999428,
0.011360734701156616,
-0.026182159781455994,
0.015675218775868416,
0.062282104045152664,
-0.10492967069149017,
1.9353818373228086e-33,
-0.0007193614146672189,
-0.01350396778434515,
0.035173241049051285,
-0.04221238195896149,
0.04691832885146141,
-0.001009213156066835,
-0.08088765293359756,
-0.0043639796786010265,
-0.03448567911982536,
0.023075643926858902,
-0.04070097208023071,
-0.009272757917642593,
0.010020711459219456,
0.01394870225340128,
0.008018134161829948,
0.014758693985641003,
0.01830339804291725,
0.07681009918451309,
0.016444239765405655,
-0.09675074368715286,
0.04726474732160568,
-0.14076246321201324,
-0.02909252978861332,
0.0350671149790287,
-0.006895282305777073,
-0.0296105295419693,
-0.022952010855078697,
-0.027087125927209854,
0.09604911506175995,
-0.006647614762187004,
-0.03359081968665123,
0.0685337707400322,
0.0977492555975914,
0.02662338875234127,
-0.0008679360616952181,
-0.08531912416219711,
-0.07012984901666641,
-0.0705387145280838,
-0.023655561730265617,
-0.06500090658664703,
0.008365346118807793,
-0.0054898327216506,
0.04018386825919151,
-0.0338391438126564,
-0.03985312581062317,
-0.08013506978750229,
-0.0013426255900412798,
-0.022017475217580795,
-0.002832641825079918,
-0.018160147592425346,
-0.0166009534150362,
0.02932021953165531,
-0.02625856176018715,
-0.028569310903549194,
0.006403821054846048,
0.08194931596517563,
0.019281914457678795,
0.05506479740142822,
-0.04571664705872536,
-0.005068841390311718,
0.021827617660164833,
-0.003187487367540598,
0.011969932354986668,
0.010030853562057018,
0.053583983331918716,
0.058680444955825806,
-0.044922102242708206,
-0.02461312711238861,
0.031937070190906525,
-0.03553389385342598,
-0.07772216945886612,
0.07925733923912048,
-0.02745250053703785,
0.0004855841980315745,
-0.027150068432092667,
-0.047270193696022034,
-0.055339619517326355,
-0.01172350998967886,
0.016414498910307884,
0.03167978301644325,
-0.00404389388859272,
-0.1042623296380043,
-0.009627426974475384,
-0.0063978396356105804,
-0.08216913044452667,
0.008455979637801647,
0.011308031156659126,
0.017124058678746223,
0.0018858719849959016,
-0.05862250179052353,
-0.003887951374053955,
0.007109023630619049,
-0.11194844543933868,
0.06995798647403717,
0.026389090344309807,
-2.9719542453840027e-33,
-0.08633684366941452,
0.05037275701761246,
0.1435948610305786,
-0.051992326974868774,
0.04748748242855072,
-0.05138784646987915,
0.04727175459265709,
-0.08014657348394394,
0.08913090825080872,
0.021382177248597145,
0.002961956663057208,
-0.0388440266251564,
-0.0009123096824623644,
-0.04862956330180168,
0.031186064705252647,
0.01614290103316307,
0.01677754521369934,
0.03794265165925026,
-0.041211530566215515,
-0.05368383973836899,
0.04087056219577789,
0.052573513239622116,
0.041675128042697906,
0.04653334617614746,
0.07479796558618546,
-0.03336236998438835,
-0.03839724138379097,
0.010900476016104221,
0.0777297168970108,
0.040583088994026184,
-0.05174041539430618,
-0.03529014438390732,
-0.03330066800117493,
-0.0603313110768795,
-0.01752511039376259,
0.07676846534013748,
0.02004411816596985,
-0.02842152677476406,
-0.06814152747392654,
0.0012766646686941385,
0.05912899225950241,
-0.02139587514102459,
-0.18687891960144043,
0.025194641202688217,
0.01961117424070835,
0.020281316712498665,
-0.059877969324588776,
0.04394718259572983,
-0.060560595244169235,
-0.0028691666666418314,
-0.04982226714491844,
0.10403073579072952,
0.05291895195841789,
0.050634849816560745,
-0.034749213606119156,
-0.14809660613536835,
0.09727855026721954,
0.0018961412133648992,
0.02215588092803955,
-0.028788918629288673,
0.00811437051743269,
-0.02344377152621746,
-0.09978540986776352,
0.013216736726462841,
-0.04213200509548187,
-0.013093032874166965,
0.05936921015381813,
-0.0049993135035037994,
0.04572329297661781,
0.06079369783401489,
0.06148432940244675,
0.04920850694179535,
-0.02276952937245369,
-0.09870634227991104,
0.053647883236408234,
0.06553371250629425,
-0.003358497517183423,
-0.003447190625593066,
-0.12058231234550476,
0.04971802979707718,
-0.01967368647456169,
0.010546475648880005,
0.09586130827665329,
0.0014798810007050633,
-0.016042131930589676,
0.015654277056455612,
0.08882327377796173,
0.07540597766637802,
0.07581987977027893,
-0.0009603427024558187,
-0.12343311309814453,
0.0004381103499326855,
-0.09467896819114685,
0.01602177321910858,
-0.03973330557346344,
-4.691313293392341e-8,
-0.059848397970199585,
0.016628023236989975,
-0.014321590773761272,
0.05185019597411156,
0.012631765566766262,
-0.015498502179980278,
-0.03643489629030228,
0.06644337624311447,
-0.07064253836870193,
0.07746897637844086,
0.0134538309648633,
0.023617103695869446,
0.016006365418434143,
-0.06066397204995155,
-0.0014800142962485552,
-0.03920434042811394,
0.01998687908053398,
0.053488437086343765,
0.004717940464615822,
0.0036608828231692314,
-0.020534684881567955,
-0.027789654210209846,
0.01513561513274908,
-0.008132962509989738,
0.059221699833869934,
-0.09897023439407349,
-0.03495486080646515,
0.004666258580982685,
0.0023910447489470243,
0.021485060453414917,
-0.03137756884098053,
0.006104424595832825,
0.03825945779681206,
-0.06182357668876648,
-0.021645884960889816,
0.09156549721956253,
0.032728686928749084,
-0.023739466443657875,
-0.02033465914428234,
0.06515634059906006,
0.020076334476470947,
0.07409012317657471,
-0.0192938894033432,
0.06899376213550568,
0.07138432562351227,
0.03655855730175972,
-0.00527997687458992,
-0.02078515850007534,
0.04050645977258682,
-0.06073514372110367,
0.09884743392467499,
0.002652808092534542,
-0.06049501150846481,
-0.05621114373207092,
0.04606568440794945,
0.005433374550193548,
0.04098215699195862,
-0.007790057919919491,
0.011174062266945839,
-0.0029660214204341173,
0.06702902913093567,
-0.009217246435582638,
-0.0984373465180397,
-0.02919442020356655
] |
Can One Modality Model Synergize Training of Other Modality Models? | https://openreview.net/forum?id=5BXWhVbHAK | [
"Jae-Jun Lee",
"Sung Whan Yoon"
] | Poster | Learning with multiple modalities has recently demonstrated significant gains in many domains by maximizing the shared information across modalities. However, the current approaches strongly rely on high-quality paired datasets, which allow co-training from the paired labels from different modalities. In this context, we raise a pivotal question: Can a model with one modality synergize the training of other models with the different modalities, even without the paired multimodal labels? Our answer is 'Yes'. As a figurative description, we argue that a writer, i.e., a language model, can promote the training of a painter, i.e., a visual model, even without the paired ground truth of text and image. We theoretically argue that a superior representation can be achieved by the synergy between two different modalities without paired supervision. As proofs of concept, we broadly confirm the considerable performance gains from the synergy among visual, language, and audio models. From a theoretical viewpoint, we first establish a mathematical foundation of the synergy between two different modality models, where each one is trained with its own modality. From a practical viewpoint, our work aims to broaden the scope of multimodal learning to encompass the synergistic usage of single-modality models, relieving a strong limitation of paired supervision. The code is available at https://github.com/johnjaejunlee95/synergistic-multimodal. | Multimodal learning, Representation learning, learning theory | We conducted theoratical and empirical frameworks that one model training can be promoted by the other model across modalities without exactly paired labels. | 2,325 | null | [
-0.0066149672493338585,
-0.13663949072360992,
0.0232554841786623,
0.015712911263108253,
0.07523757964372635,
0.04127216339111328,
0.024593405425548553,
-0.04153968393802643,
0.06668289750814438,
-0.11437058448791504,
-0.019211899489164352,
-0.018226593732833862,
0.0585278682410717,
0.05787931755185127,
0.058943524956703186,
0.012855934910476208,
0.060273125767707825,
0.10793112963438034,
-0.10427984595298767,
-0.01676834002137184,
0.04254266992211342,
-0.0070578232407569885,
-0.008109386079013348,
-0.035149913281202316,
0.08990170806646347,
-0.01768244244158268,
-0.030528521165251732,
-0.037392470985651016,
0.07379314303398132,
-0.0230405293405056,
-0.01248721219599247,
0.05131084844470024,
-0.08516811579465866,
0.019831599667668343,
-0.029969993978738785,
0.08382346481084824,
-0.00567011209204793,
0.06854497641324997,
0.05025626718997955,
-0.050527166575193405,
-0.029856106266379356,
0.08333131670951843,
-0.04179760068655014,
-0.0001737519196467474,
0.006428637076169252,
-0.019901009276509285,
0.000004532378625299316,
0.015288072638213634,
-0.08971603959798813,
-0.002223926829174161,
-0.10060951113700867,
-0.039964742958545685,
-0.013879031874239445,
-0.005175815429538488,
-0.023883674293756485,
0.03668389841914177,
0.04780828207731247,
0.008308205753564835,
0.02723078243434429,
0.012243998236954212,
-0.04335818439722061,
-0.039149146527051926,
-0.07350115478038788,
-0.024326391518115997,
0.09419217705726624,
-0.10031349956989288,
0.018865689635276794,
0.03100549802184105,
-0.05998867005109787,
0.03830132633447647,
0.02877746894955635,
0.02229318581521511,
0.002827961463481188,
-0.04401465505361557,
0.027212172746658325,
0.024535374715924263,
0.023777367547154427,
0.005729546770453453,
0.0010910077253356576,
-0.10195987671613693,
0.049998458474874496,
0.0233705285936594,
0.04792381450533867,
-0.03918658569455147,
0.06732368469238281,
-0.031432848423719406,
-0.030125148594379425,
-0.016963239759206772,
-0.036106303334236145,
0.00010740722063928843,
-0.0932997614145279,
0.009400852955877781,
0.006319439969956875,
-0.015818821266293526,
0.03100396879017353,
-0.003872697940096259,
0.060293421149253845,
-0.01767229288816452,
0.03757729381322861,
0.09716867655515671,
0.006949870381504297,
0.059900615364313126,
-0.024854285642504692,
-0.019932817667722702,
0.11020472645759583,
-0.08073082566261292,
0.013265901245176792,
0.029506759718060493,
-0.017452888190746307,
-0.07280585914850235,
0.07949647307395935,
0.041633591055870056,
-0.01556396298110485,
0.008108607493340969,
0.04531685635447502,
0.00896409247070551,
-0.006197141949087381,
-0.01085597462952137,
0.04836910963058472,
-0.03982241079211235,
-0.08678758889436722,
-0.0010136902565136552,
0.03944717347621918,
0.007749142590910196,
0.013753916136920452,
-0.08897869288921356,
-0.08076376467943192,
2.645808011289582e-33,
0.15256917476654053,
0.01698855310678482,
0.07470810413360596,
0.04692170023918152,
0.08738869428634644,
-0.006128600798547268,
-0.07409451901912689,
-0.04240155220031738,
0.016032647341489792,
0.015697939321398735,
-0.006692505907267332,
0.021796397864818573,
-0.003527274588122964,
0.1253420114517212,
0.002178482012823224,
0.04533848911523819,
-0.05298598110675812,
-0.005390954203903675,
-0.018803320825099945,
0.04790369048714638,
0.08405901491641998,
0.01586313359439373,
-0.010421410202980042,
-0.02065707929432392,
-0.06726748496294022,
-0.027563028037548065,
0.03360750153660774,
-0.06914171576499939,
-0.049934979528188705,
0.012712664902210236,
-0.15253551304340363,
0.05547504499554634,
0.059919700026512146,
0.020385148003697395,
0.0231744647026062,
-0.0739453062415123,
-0.000594650860875845,
-0.058454688638448715,
-0.006141305901110172,
-0.04364193603396416,
-0.000013540244253817946,
0.02634715661406517,
0.003563410835340619,
-0.05884290114045143,
-0.05025358498096466,
0.015028774738311768,
-0.03338330611586571,
0.03701064735651016,
-0.12968027591705322,
-0.0383431650698185,
-0.01930088736116886,
0.04817871376872063,
-0.023957662284374237,
-0.03388388454914093,
-0.02640051394701004,
0.010740441270172596,
0.004658123943954706,
-0.007626570761203766,
0.05965862795710564,
-0.0405099019408226,
-0.010778876952826977,
0.044098611921072006,
-0.004036432597786188,
0.06954674422740936,
0.07351868599653244,
0.08788613975048065,
-0.061744965612888336,
-0.01908668503165245,
0.08193759620189667,
-0.04201402887701988,
-0.1044653132557869,
0.02454010210931301,
-0.10109291970729828,
-0.04791263863444328,
0.020314496010541916,
-0.06253143399953842,
0.03282963111996651,
-0.0910521075129509,
0.04013126716017723,
0.10404231399297714,
-0.02273072861135006,
0.024415798485279083,
0.008355237543582916,
-0.08289924263954163,
-0.036481402814388275,
0.008430184796452522,
-0.000032939929951680824,
-0.09175866842269897,
0.011138010770082474,
-0.00377401290461421,
-0.07929866015911102,
0.013374928385019302,
-0.033253222703933716,
-0.08859854191541672,
-0.01496752817183733,
-3.2023676685344795e-33,
-0.013371770270168781,
0.05596411973237991,
-0.0260449405759573,
0.04839757829904556,
0.005625375546514988,
-0.0405726358294487,
0.07360272109508514,
0.02216888777911663,
-0.05266018956899643,
-0.01417544949799776,
0.025288067758083344,
-0.04562678933143616,
0.037012435495853424,
-0.07843779027462006,
-0.023709628731012344,
-0.007901710458099842,
0.026771748438477516,
0.049677252769470215,
0.013009008951485157,
0.011624989099800587,
0.05408502370119095,
0.01624150015413761,
-0.014244341291487217,
0.0384361632168293,
-0.04721878841519356,
0.04155591130256653,
-0.11826347559690475,
0.047813333570957184,
-0.05836637318134308,
0.009806672111153603,
0.050121381878852844,
-0.030451063066720963,
0.013576329685747623,
-0.07195720821619034,
-0.05123748257756233,
0.033566270023584366,
0.04153075069189072,
-0.017385484650731087,
-0.0478702187538147,
0.1241866797208786,
0.04000023752450943,
-0.05508914962410927,
-0.022931979969143867,
0.038041602820158005,
0.014887858182191849,
-0.10632310807704926,
-0.08699247241020203,
-0.08738623559474945,
0.005638805218040943,
-0.014268426224589348,
0.018010644242167473,
-0.03598715364933014,
-0.07964690774679184,
-0.02230823040008545,
-0.055981166660785675,
-0.1689791977405548,
0.03451848775148392,
-0.03298182040452957,
0.09542292356491089,
0.05997615307569504,
0.002047061687335372,
-0.02563738450407982,
-0.0847715362906456,
-0.005292458459734917,
0.07919085770845413,
-0.03842560201883316,
-0.02030278369784355,
-0.008647959679365158,
0.0989685207605362,
0.018631670624017715,
0.036642417311668396,
-0.005363455042243004,
0.04484555125236511,
0.032013699412345886,
0.01001986674964428,
0.030175907537341118,
0.036770667880773544,
-0.012699284590780735,
-0.04998468607664108,
-0.06421627104282379,
-0.019462918862700462,
-0.04917740449309349,
0.01279321126639843,
0.040593553334474564,
0.0791894942522049,
0.11390049010515213,
-0.013136770576238632,
0.016912177205085754,
0.026740750297904015,
0.007714972831308842,
0.003113001585006714,
-0.003470488591119647,
0.03597142919898033,
0.00925037544220686,
-0.013766958378255367,
-4.56631354950332e-8,
-0.07653389871120453,
0.0457763709127903,
-0.01743491366505623,
0.0067636966705322266,
-0.00513681722804904,
-0.02682602033019066,
0.024827133864164352,
0.03465428203344345,
-0.022208265960216522,
0.018291452899575233,
0.06276974827051163,
-0.00600776169449091,
-0.02948882058262825,
-0.023991145193576813,
0.003570272121578455,
0.018156295642256737,
0.05778081715106964,
0.043805286288261414,
-0.004796336870640516,
-0.012568132020533085,
0.0832008570432663,
0.003404384944587946,
0.022827470675110817,
0.0066246227361261845,
0.045694321393966675,
-0.029152104631066322,
-0.016341961920261383,
0.0334462895989418,
-0.01172131672501564,
0.05680105835199356,
-0.06257284432649612,
0.026595985516905785,
-0.008962125517427921,
-0.04397345334291458,
-0.022428365424275398,
0.0065003568306565285,
-0.035319749265909195,
-0.0557975135743618,
-0.06379084289073944,
-0.01911981962621212,
-0.0021051662042737007,
0.07559142261743546,
-0.083075150847435,
0.058783158659935,
0.16045428812503815,
0.02278951369225979,
0.05118541046977043,
-0.0920548364520073,
-0.016357947140932083,
0.03178437054157257,
-0.03597145527601242,
0.0397174209356308,
0.0628666877746582,
0.05305450037121773,
0.04697783663868904,
-0.02684357389807701,
-0.021995633840560913,
0.07404577732086182,
0.010277598164975643,
0.01995908096432686,
0.04201354831457138,
0.028232328593730927,
-0.053935401141643524,
-0.06722946465015411
] |
Aligned Datasets Improve Detection of Latent Diffusion-Generated Images | https://openreview.net/forum?id=doBkiqESYq | [
"Anirudh Sundara Rajan",
"Utkarsh Ojha",
"Jedidiah Schloesser",
"Yong Jae Lee"
] | Poster | As latent diffusion models (LDMs) democratize image generation capabilities, there is a growing need to detect fake images. A good detector should focus on the generative model’s fingerprints while ignoring image properties such as semantic content, resolution, file format, etc. Fake image detectors are usually built in a data-driven way, where a model is trained to separate real from fake images. Existing works primarily investigate network architecture choices and training recipes. In this work, we argue that in addition to these algorithmic choices, we also require a well-aligned dataset of real/fake images to train a robust detector. For the family of LDMs, we propose a very simple way to achieve this: we reconstruct all the real images using the LDM's autoencoder, without any denoising operation. We then train a model to separate these real images from their reconstructions. The fakes created this way are extremely similar to the real ones in almost every aspect (e.g., size, aspect ratio, semantic content), which forces the model to look for the LDM decoder's artifacts. We empirically show that this way of creating aligned real/fake datasets, which also sidesteps the computationally expensive denoising process, helps in building a detector that focuses less on spurious correlations, something that a very popular existing method is susceptible to. Finally, to demonstrate the effectivenss of dataset alignment, we build a detector using images that are not natural objects, and present promising results. Overall, our work identifies the subtle but significant issues that arise when training a fake image detector and proposes a simple and inexpensive solution to address these problems. | Image Forensics, Latent Diffusion | null | 2,319 | 2410.11835 | [
-0.05162247270345688,
-0.058743853121995926,
-0.008779003284871578,
0.022834716364741325,
0.09180425852537155,
-0.060655251145362854,
-0.053010955452919006,
-0.13777747750282288,
0.015423454344272614,
-0.0498921275138855,
0.085835762321949,
-0.03202003985643387,
0.04844605550169945,
0.05282272398471832,
-0.04317346587777138,
-0.06291943788528442,
0.048520416021347046,
0.10096661746501923,
-0.07463077455759048,
-0.03568563237786293,
-0.059717658907175064,
-0.003059664275497198,
0.005896936636418104,
-0.04055093973875046,
0.004538940265774727,
0.062227945774793625,
0.04690336436033249,
-0.05553572624921799,
-0.0037233722396194935,
-0.09903822094202042,
0.13887181878089905,
0.010328799486160278,
-0.04610435292124748,
0.0420236811041832,
0.0037578956689685583,
0.03955059498548508,
-0.022333640605211258,
-0.009584789164364338,
0.04938302934169769,
-0.06542560458183289,
0.039386022835969925,
0.03708992898464203,
0.007110820151865482,
0.010513616725802422,
0.001887885620817542,
0.017151519656181335,
0.006035325117409229,
0.0028897838201373816,
0.017118148505687714,
-0.01923183724284172,
-0.030331742018461227,
0.017003152519464493,
-0.02382780984044075,
0.051019337028265,
-0.00276764458976686,
-0.0023974210489541292,
0.05588236078619957,
0.049822330474853516,
-0.013449607416987419,
0.07791601866483688,
0.0025410798843950033,
-0.0199822299182415,
-0.03399825096130371,
-0.014505399391055107,
-0.02013053558766842,
0.050162479281425476,
0.03509760648012161,
0.03157241642475128,
0.04429956153035164,
-0.026782164350152016,
0.044567693024873734,
0.0629609003663063,
-0.08706888556480408,
-0.0006554144201800227,
-0.039003875106573105,
0.0670023038983345,
-0.0030096909031271935,
-0.046034447848796844,
0.04132497310638428,
-0.07723502814769745,
0.021309539675712585,
-0.0053081815131008625,
0.10425098240375519,
0.02073783427476883,
0.07601732015609741,
-0.014294802211225033,
-0.021217044442892075,
0.05423302203416824,
-0.018178148195147514,
0.005088812205940485,
-0.11301958560943604,
0.0013548628194257617,
-0.061109207570552826,
0.006176571361720562,
0.07946904003620148,
0.05881621316075325,
0.042230140417814255,
-0.016612356528639793,
0.03230138495564461,
0.09184549748897552,
-0.05440249666571617,
0.012299957685172558,
-0.009056075476109982,
-0.04583032801747322,
0.09052493423223495,
-0.024283235892653465,
0.03462423384189606,
0.021452954038977623,
0.08128051459789276,
-0.026794008910655975,
0.037748731672763824,
0.037847425788640976,
-0.0071472665295004845,
-0.046080660074949265,
0.10745622962713242,
-0.09044540673494339,
0.05264057591557503,
0.00041584644350223243,
-0.029586181044578552,
-0.060252025723457336,
-0.03406081721186638,
0.044225357472896576,
-0.05519137531518936,
-0.05662743002176285,
0.08640356361865997,
-0.07466123253107071,
-0.05393461138010025,
2.200072492072271e-33,
0.00565005000680685,
0.00909309834241867,
0.034206733107566833,
0.03542269393801689,
0.07402663677930832,
-0.0400216318666935,
-0.0458565391600132,
-0.024422483518719673,
0.028725774958729744,
-0.01773793064057827,
-0.0601750873029232,
0.010691351257264614,
-0.08776308596134186,
0.11416196823120117,
0.034038905054330826,
-0.0002261723711853847,
-0.08761020004749298,
0.004858340136706829,
-0.006510051898658276,
0.05789060890674591,
0.017649993300437927,
-0.026732739061117172,
0.020879317075014114,
-0.034853655844926834,
-0.011763461865484715,
-0.04346678406000137,
0.05844789370894432,
-0.024954628199338913,
0.0035657293628901243,
-0.025555390864610672,
-0.13421332836151123,
-0.03438130021095276,
0.03810729458928108,
0.06280143558979034,
0.017111610621213913,
-0.00023947286535985768,
-0.04885633662343025,
0.010269957594573498,
-0.045494839549064636,
-0.016239436343312263,
0.05626768246293068,
0.053107116371393204,
-0.03130895271897316,
-0.015317833982408047,
-0.0381094329059124,
0.05701981484889984,
-0.008433211594820023,
-0.018751680850982666,
0.01569974608719349,
-0.02369949035346508,
0.10727650672197342,
-0.006550462916493416,
0.019363241270184517,
-0.07540718466043472,
-0.06359898298978806,
-0.002293251920491457,
0.06025101989507675,
-0.04716473072767258,
0.07914827764034271,
0.017522510141134262,
0.07531014829874039,
-0.03494292497634888,
-0.015910092741250992,
0.0505918487906456,
0.029430853202939034,
-0.04771589860320091,
0.032407138496637344,
-0.012427053414285183,
-0.0053104218095541,
0.0007968447171151638,
-0.027521679177880287,
0.008755143731832504,
-0.05261749029159546,
-0.05606598034501076,
0.06995497643947601,
0.006059583276510239,
0.06463297456502914,
-0.0014444685075432062,
0.00922947283834219,
0.03847063332796097,
-0.00235435226932168,
-0.05755272135138512,
0.05536877363920212,
-0.08865419775247574,
-0.07033563405275345,
0.0014636327978223562,
-0.03171145170927048,
-0.09636042267084122,
-0.039521750062704086,
0.002590194810181856,
-0.0227699875831604,
-0.05657435581088066,
-0.001322206691838801,
0.013362974859774113,
0.005151346325874329,
-1.528054402157166e-33,
-0.013220982626080513,
0.06851251423358917,
0.03278084471821785,
0.05586215853691101,
-0.05790271982550621,
-0.03474966436624527,
0.008070626296103,
0.1016589105129242,
0.026038506999611855,
-0.0448378250002861,
-0.01749965362250805,
-0.017541805282235146,
-0.011811868287622929,
-0.05017834156751633,
-0.04728531464934349,
-0.0660092681646347,
0.042031727731227875,
-0.030152251943945885,
-0.028465090319514275,
0.029713677242398262,
0.02798917330801487,
0.12177671492099762,
-0.07624680548906326,
0.06323482096195221,
0.006518355570733547,
0.09079132229089737,
-0.0010112590389326215,
0.05502006784081459,
0.025706414133310318,
-0.01460061501711607,
0.036255452781915665,
0.0057661565952003,
0.008101308718323708,
-0.0981636792421341,
-0.07220280170440674,
0.028073608875274658,
0.03834836184978485,
0.06341732293367386,
0.006055380683392286,
0.08830970525741577,
0.027699559926986694,
-0.00974654033780098,
-0.17099007964134216,
0.0695638433098793,
-0.01996787078678608,
-0.05199046432971954,
-0.02614409290254116,
0.03378136083483696,
0.008951960131525993,
0.02823709510266781,
-0.000567810726352036,
0.05333828181028366,
-0.04661761596798897,
0.02622387744486332,
-0.0030534008983522654,
-0.045037563890218735,
-0.0926884338259697,
0.007663177326321602,
0.0863577276468277,
0.05742054805159569,
-0.044525619596242905,
-0.09971101582050323,
-0.022797290235757828,
-0.08635802567005157,
-0.04749540239572525,
-0.06786851584911346,
-0.057266317307949066,
0.0071925874799489975,
-0.05194836109876633,
0.05869905650615692,
0.08348154276609421,
0.011708003468811512,
-0.039431728422641754,
0.07905713468790054,
0.045047760009765625,
-0.11387813091278076,
0.041568510234355927,
0.008456677198410034,
0.08785142004489899,
-0.0023814563173800707,
0.014054570347070694,
-0.002122002886608243,
0.04279107600450516,
0.03829546272754669,
0.0847494974732399,
0.04262762516736984,
-0.046237796545028687,
-0.0995279848575592,
0.02596116065979004,
0.06487932801246643,
-0.0009729508310556412,
0.001471243565902114,
-0.025624634698033333,
0.03836403414607048,
-0.04716366156935692,
-4.442035361762464e-8,
-0.06999702006578445,
0.034322340041399,
0.010728929191827774,
-0.04085031524300575,
-0.036010418087244034,
-0.037556249648332596,
0.019528057426214218,
0.08719370514154434,
0.019267573952674866,
-0.08213888108730316,
0.11197993904352188,
-0.03755275160074234,
-0.10103029757738113,
-0.04183818772435188,
-0.022743768990039825,
0.0658966675400734,
-0.015535017475485802,
-0.003931326791644096,
-0.026824628934264183,
0.0307298731058836,
-0.01046485174447298,
0.041464194655418396,
0.015213731676340103,
-0.002882772358134389,
0.09836482256650925,
-0.014617315493524075,
-0.025905821472406387,
0.03755287081003189,
-0.006285740062594414,
0.07360412180423737,
0.005881020333617926,
-0.007017030380666256,
0.05857022479176521,
-0.028481416404247284,
-0.011888504028320312,
0.1566067636013031,
0.0525277815759182,
-0.018345089629292488,
-0.04516976326704025,
-0.10050532221794128,
-0.08058629184961319,
0.008702865801751614,
-0.028674112632870674,
-0.03399849683046341,
-0.02072611078619957,
-0.046007853001356125,
0.06275114417076111,
-0.012331975623965263,
0.021359657868742943,
-0.016732752323150635,
0.03003232553601265,
-0.044456638395786285,
0.010842925868928432,
0.04927734285593033,
-0.007405469659715891,
-0.0724087730050087,
0.023681193590164185,
-0.013283700682222843,
0.09627986699342728,
0.09617474675178528,
0.03256066143512726,
0.0013940074713900685,
0.06209646537899971,
0.014859289862215519
] |
RocketEval: Efficient automated LLM evaluation via grading checklist | https://openreview.net/forum?id=zJjzNj6QUe | [
"Tianjun Wei",
"Wei Wen",
"Ruizhi Qiao",
"Xing Sun",
"Jianghong Ma"
] | Poster | Evaluating large language models (LLMs) in diverse and challenging scenarios is essential to align them with human preferences. To mitigate the prohibitive costs associated with human evaluations, utilizing a powerful LLM as a judge has emerged as a favored approach. Nevertheless, this methodology encounters several challenges, including substantial expenses, concerns regarding privacy and security, and reproducibility. In this paper, we propose a straightforward, replicable, and accurate automated evaluation method by leveraging a lightweight LLM as the judge, named RocketEval. Initially, we identify that the performance disparity between lightweight and powerful LLMs in evaluation tasks primarily stems from their ability to conduct comprehensive analyses, which is not easily enhanced through techniques such as chain-of-thought reasoning. By reframing the evaluation task as a multi-faceted Q\&A using an instance-specific checklist, we demonstrate that the limited judgment accuracy of lightweight LLMs is largely attributes to high uncertainty and positional bias. To address these challenges, we introduce an automated evaluation process grounded in checklist grading, which is designed to accommodate a variety of scenarios and questions. This process encompasses the creation of checklists, the grading of these checklists by lightweight LLMs, and the reweighting of checklist items to align with the supervised annotations. Our experiments carried out on the automated evaluation benchmarks, MT-Bench and WildBench datasets, reveal that RocketEval, when using $\textit{Gemma-2-2B}$ as the judge, achieves a high correlation (0.965) with human preferences, which is comparable to $\textit{GPT-4o}$. Moreover, RocketEval provides a cost reduction exceeding 50-fold for large-scale evaluation and comparison scenarios. Our code is available at https://github.com/Joinn99/RocketEval-ICLR. | automated evaluation, large language models, natural language processing | A simple, replicable,interpretable, and accurate automated evaluation method that uses lightweight LLMs as judges to efficiently assess various scenarios and questions. | 2,318 | 2503.05142 | [
-0.024932822212576866,
0.019388454034924507,
-0.039890140295028687,
-0.005280264653265476,
0.06731842458248138,
-0.005836085416376591,
0.04634895548224449,
0.011510001495480537,
0.01653064414858818,
0.001676458166912198,
-0.0686788260936737,
-0.11336681991815567,
0.05601081997156143,
0.047726552933454514,
0.02885730005800724,
0.0037278435193002224,
0.06957653909921646,
-0.008395703509449959,
-0.07154172658920288,
-0.07665296643972397,
0.00747154513373971,
0.045552026480436325,
0.033478979021310806,
0.02903997339308262,
-0.007249540649354458,
-0.06752732396125793,
-0.042409371584653854,
-0.005214033182710409,
-0.029193909838795662,
0.0023793294094502926,
0.014163568615913391,
0.013561065308749676,
0.0006736145005561411,
0.07575759291648865,
0.001689980854280293,
0.018730364739894867,
-0.04778444021940231,
-0.008486264385282993,
-0.03830977901816368,
-0.048161581158638,
-0.1052108108997345,
-0.09406311064958572,
-0.012985083274543285,
0.021419497206807137,
0.0037996461614966393,
-0.04056656360626221,
-0.0516527034342289,
-0.022069431841373444,
-0.05958384647965431,
0.016972916200757027,
-0.07720215618610382,
-0.04142523929476738,
0.00686713308095932,
-0.028535308316349983,
-0.058058734983205795,
0.04123200103640556,
0.03876730427145958,
0.028644200414419174,
-0.010807552374899387,
-0.06784939765930176,
-0.14465372264385223,
-0.08655229210853577,
-0.025969309732317924,
0.07580769062042236,
-0.042390380054712296,
-0.011925758793950081,
-0.0030049318447709084,
0.038474321365356445,
-0.009453617967665195,
0.024261031299829483,
-0.037860214710235596,
0.06805349141359329,
0.03486273065209389,
0.08823473006486893,
-0.029076410457491875,
0.0517890639603138,
-0.036385562270879745,
-0.03829972818493843,
0.041848719120025635,
-0.03948507830500603,
-0.026132861152291298,
-0.027796601876616478,
-0.038255903869867325,
0.03088146634399891,
0.09937442094087601,
0.018299872055649757,
0.06067032366991043,
0.041648413985967636,
0.028241291642189026,
0.1091582253575325,
0.0026850132271647453,
-0.03879771754145622,
0.09439698606729507,
0.003290478140115738,
0.10473953187465668,
0.03891485184431076,
-0.05036109685897827,
-0.10238859802484512,
-0.02049279399216175,
0.054712165147066116,
-0.012207023799419403,
0.10585237294435501,
-0.020794766023755074,
-0.11808186024427414,
-0.037539999932050705,
0.015075305476784706,
0.035885605961084366,
0.021293338388204575,
0.005431026220321655,
-0.069656603038311,
-0.008896042592823505,
0.044326405972242355,
0.05287797376513481,
-0.020581545308232307,
-0.0028004441410303116,
0.003591805463656783,
0.051679834723472595,
0.0033540227450430393,
0.03563368320465088,
0.030776947736740112,
0.014091975055634975,
0.018768854439258575,
0.10927196592092514,
-0.029981223866343498,
0.04432698339223862,
-0.023868752643465996,
-0.05677979066967964,
-2.1736789245587162e-34,
0.05502905324101448,
0.007763634901493788,
-0.011018814519047737,
0.050025574862957,
-0.01656840182840824,
-0.03607950359582901,
-0.0007178314845077693,
0.1222568228840828,
-0.12756894528865814,
0.03420805186033249,
0.015760133042931557,
0.039417020976543427,
0.0042289793491363525,
0.043668512254953384,
0.11622036248445511,
0.07449209690093994,
-0.011204428970813751,
0.1148386150598526,
-0.06924767047166824,
-0.04437899589538574,
0.08149560540914536,
-0.04934665188193321,
0.0731830820441246,
-0.01796405203640461,
0.018746647983789444,
0.033255044370889664,
0.07138650119304657,
-0.0017612904775887728,
0.0017437500646337867,
0.014142543077468872,
-0.12119826674461365,
-0.06963353604078293,
0.013001898303627968,
0.028130320832133293,
0.06584533303976059,
0.021381232887506485,
-0.0390566810965538,
-0.0022389048244804144,
0.08548258244991302,
-0.012206128798425198,
0.005716462153941393,
-0.03255297243595123,
0.04768884554505348,
0.02977490797638893,
-0.048479098826646805,
0.037290047854185104,
-0.056179266422986984,
0.02263697050511837,
0.03628257289528847,
0.0036276662722229958,
0.047752853482961655,
0.002871616743505001,
-0.01837640255689621,
0.04012790694832802,
-0.05032094195485115,
-0.03614840656518936,
0.014937976375222206,
0.04301588237285614,
-0.010730010457336903,
0.04544985294342041,
0.006356440018862486,
0.14731848239898682,
-0.06084465980529785,
0.003153854748234153,
-0.012881211936473846,
-0.05151138827204704,
-0.07651793956756592,
0.03689483925700188,
0.08861535042524338,
-0.026626089587807655,
-0.010623099282383919,
0.009829775430262089,
0.04256622493267059,
-0.018602954223752022,
0.018098540604114532,
-0.010157059878110886,
0.016915511339902878,
-0.032738931477069855,
0.04777922481298447,
-0.056788600981235504,
0.09616172313690186,
0.0038819361943751574,
-0.0055124107748270035,
-0.017863038927316666,
-0.04269452765583992,
-0.11934832483530045,
0.007943662814795971,
-0.032283712178468704,
-0.022161582484841347,
0.014527742750942707,
-0.06470154225826263,
-0.06570593267679214,
-0.05692833662033081,
-0.01176964957267046,
-0.014603142626583576,
-1.850290556331668e-33,
-0.06537115573883057,
-0.06394777446985245,
-0.010976800695061684,
0.08915308117866516,
-0.04211907833814621,
-0.08309389650821686,
-0.018874071538448334,
-0.006036016158759594,
-0.015844030305743217,
-0.03814932703971863,
-0.10674508661031723,
-0.014447539113461971,
0.0449262410402298,
-0.03127126395702362,
0.06430216878652573,
-0.023181041702628136,
-0.09258194267749786,
-0.03593052923679352,
-0.004727228078991175,
0.09547708928585052,
0.021340737119317055,
0.14817112684249878,
-0.09598161280155182,
0.048666927963495255,
-0.07307607680559158,
0.025967273861169815,
-0.0006717892829328775,
-0.02557508274912834,
0.032313719391822815,
0.05007113143801689,
-0.008296811021864414,
-0.015110383741557598,
-0.07103043794631958,
-0.006450374610722065,
-0.05143635347485542,
-0.018969161435961723,
0.11174000054597855,
-0.006025881040841341,
0.009993048384785652,
0.06994982808828354,
0.008304314687848091,
-0.027770448476076126,
-0.06823684275150299,
-0.03192666545510292,
-0.04616639390587807,
0.013191197998821735,
-0.09402827173471451,
-0.0005355314933694899,
0.017247991636395454,
0.007179837208241224,
-0.04267694801092148,
-0.02330307476222515,
-0.007719175890088081,
0.038404431194067,
-0.0365636944770813,
-0.0954468697309494,
-0.03841147571802139,
-0.07291720807552338,
0.009751920588314533,
0.04698290303349495,
-0.06153811514377594,
-0.01323010865598917,
0.054679084569215775,
-0.054765116423368454,
0.03416237235069275,
-0.05890573933720589,
-0.01256483793258667,
0.0015442835865542293,
-0.061041805893182755,
0.032529089599847794,
0.008731787092983723,
-0.04564496874809265,
-0.042635153979063034,
-0.022912299260497093,
0.03915486857295036,
0.005199202336370945,
0.048424240201711655,
-0.0434885174036026,
-0.021526232361793518,
-0.03816467896103859,
0.09914188086986542,
-0.019010452553629875,
0.025852860882878304,
-0.016223520040512085,
0.008685489185154438,
0.008858680725097656,
0.04178699478507042,
0.10616215318441391,
0.006499331910163164,
0.048120882362127304,
-0.030274245887994766,
-0.0024310287553817034,
-0.0007297095144167542,
0.037648797035217285,
-0.08504100143909454,
-4.389517371805596e-8,
-0.02186504565179348,
-0.02190617471933365,
0.015653349459171295,
0.12863753736019135,
0.017443455755710602,
-0.04248099401593208,
-0.08510904014110565,
0.0576363205909729,
0.00962760578840971,
-0.02738124504685402,
0.05542617291212082,
-0.0174175426363945,
-0.04565538465976715,
-0.030643543228507042,
0.025659210979938507,
0.0791400894522667,
-0.04150281846523285,
0.07063445448875427,
-0.04336100444197655,
-0.011185680516064167,
-0.017603173851966858,
0.03506989777088165,
-0.06912297755479813,
0.020070698112249374,
0.0027401770930737257,
-0.022866984829306602,
-0.03827553614974022,
0.08327537775039673,
0.04797673597931862,
0.02675998955965042,
0.014884021133184433,
0.009130630642175674,
-0.05950634181499481,
-0.008314627222716808,
0.04468818008899689,
0.073202945291996,
-0.05082128196954727,
0.021958881989121437,
0.0691375583410263,
0.05462520197033882,
0.04417455568909645,
0.035303253680467606,
-0.0746796652674675,
0.06447548419237137,
0.002659908728674054,
-0.04967866465449333,
-0.14391647279262543,
-0.07596976310014725,
-0.025524204596877098,
-0.046446215361356735,
-0.01316926721483469,
-0.08669405430555344,
-0.03779218718409538,
0.059348348528146744,
0.004274172242730856,
0.025686534121632576,
0.005852825008332729,
-0.0250317994505167,
0.046925220638513565,
0.07083945721387863,
0.12278302758932114,
0.017448708415031433,
-0.0019150788430124521,
0.028438935056328773
] |
PolaFormer: Polarity-aware Linear Attention for Vision Transformers | https://openreview.net/forum?id=kN6MFmKUSK | [
"Weikang Meng",
"Yadan Luo",
"Xin Li",
"Dongmei Jiang",
"Zheng Zhang"
] | Poster | Linear attention has emerged as a promising alternative to softmax-based attention, leveraging kernelized feature maps to reduce complexity from quadratic to linear in sequence length. However, the non-negative constraint on feature maps and the relaxed exponential function used in approximation lead to significant information loss compared to the original query-key dot products, resulting in less discriminative attention maps with higher entropy. To address the missing interactions driven by negative values in query-key pairs, we propose a polarity-aware linear attention mechanism that explicitly models both same-signed and opposite-signed query-key interactions, ensuring comprehensive coverage of relational information. Furthermore, to restore the spiky properties of attention maps, we provide a theoretical analysis proving the existence of a class of element-wise functions (with positive first and second derivatives) that can reduce entropy in the attention distribution. For simplicity, and recognizing the distinct contributions of each dimension, we employ a learnable power function for rescaling, allowing strong and weak attention signals to be effectively separated. Extensive experiments demonstrate that the proposed PolaFormer improves performance on various vision tasks, enhancing both expressiveness and efficiency by up to 4.6%. | Linear attention, polarity-aware attention, efficient transformer | null | 2,313 | 2501.15061 | [
0.003743414068594575,
-0.0461159385740757,
-0.014097574166953564,
0.014657184481620789,
0.05608292296528816,
0.027032215148210526,
0.052884332835674286,
-0.05318022519350052,
0.05713403970003128,
0.007627922110259533,
0.003759093349799514,
-0.005783848464488983,
0.06649898737668991,
0.058396823704242706,
0.0016631323378533125,
0.05950639769434929,
0.03213244304060936,
0.09065821021795273,
-0.07288647443056107,
0.028217392042279243,
0.031560808420181274,
-0.04561825096607208,
0.03845555707812309,
-0.022194042801856995,
-0.014373404905200005,
0.016548991203308105,
-0.006882247049361467,
-0.041138164699077606,
0.005819379352033138,
-0.005466898437589407,
-0.0060711936093866825,
0.0058148629032075405,
0.01863790489733219,
0.052695710211992264,
-0.07920225709676743,
0.10286103188991547,
-0.02667897380888462,
0.021875498816370964,
-0.06960891932249069,
-0.06792951375246048,
0.02256036549806595,
0.0038105023559182882,
-0.02770099602639675,
0.06692413985729218,
0.0651804730296135,
0.03174547106027603,
0.048039134591817856,
-0.05559053644537926,
0.03347660228610039,
-0.059340812265872955,
-0.08663307130336761,
0.026201996952295303,
-0.07162031531333923,
0.04225807264447212,
0.026329101994633675,
0.058369167149066925,
-0.054042551666498184,
-0.0049798996187746525,
0.0013250433839857578,
-0.03039577789604664,
-0.03391115367412567,
-0.09888825565576553,
0.04500347003340721,
-0.06544600427150726,
0.0051453597843647,
-0.05866293981671333,
0.043337058275938034,
-0.05267928168177605,
-0.07107848674058914,
0.05363059788942337,
-0.0555034838616848,
0.05498947575688362,
-0.043165240436792374,
-0.05912163853645325,
0.0528845489025116,
0.1392112374305725,
0.028869861736893654,
-0.05503742769360542,
0.018359068781137466,
-0.017923012375831604,
-0.029941603541374207,
-0.05209985747933388,
0.03225107863545418,
0.03911377117037773,
0.18249543011188507,
-0.006044114474207163,
0.02574281394481659,
-0.04665866494178772,
0.006318016909062862,
0.08082938194274902,
-0.10943767428398132,
-0.06075616925954819,
0.04418643191456795,
-0.06566649675369263,
0.036057375371456146,
-0.009491152130067348,
0.047702305018901825,
-0.04299716651439667,
-0.017250478267669678,
0.08511790633201599,
-0.030172429978847504,
-0.01571466773748398,
-0.024154027923941612,
-0.10992121696472168,
0.016231441870331764,
-0.03380880877375603,
-0.02642187662422657,
0.02198759652674198,
0.045380983501672745,
-0.08769666403532028,
0.0048960670828819275,
0.021846167743206024,
0.005221557337790728,
0.04219513386487961,
0.010992837138473988,
0.018286531791090965,
-0.02235395647585392,
0.006343412678688765,
0.09053933620452881,
0.010429611429572105,
0.030756784602999687,
0.039725884795188904,
-0.09918498247861862,
0.04336416348814964,
0.022698841989040375,
-0.0423206128180027,
-0.02525254897773266,
-2.307653826895023e-34,
0.024957166984677315,
0.09103501588106155,
0.045873645693063736,
-0.018860915675759315,
0.042200200259685516,
0.042085207998752594,
-0.031900838017463684,
0.022787509486079216,
-0.04845840856432915,
0.050477854907512665,
0.00030476655228994787,
0.046410977840423584,
-0.006559101399034262,
0.09616274386644363,
-0.001921599730849266,
-0.0031850074883550406,
-0.007454903796315193,
-0.001448746188543737,
-0.04491894319653511,
-0.03596369922161102,
-0.05997896194458008,
-0.009878197684884071,
0.014348060823976994,
-0.007405622396618128,
0.015033489093184471,
-0.03531280905008316,
-0.0014715752331539989,
-0.04261983931064606,
-0.07182847708463669,
-0.012185698375105858,
-0.1319420337677002,
0.0714593380689621,
-0.017428135499358177,
0.013863270170986652,
0.05394330620765686,
-0.05703642591834068,
-0.004159898031502962,
-0.06714823096990585,
0.04940805211663246,
-0.010707387700676918,
0.0051579889841377735,
0.0681978389620781,
-0.031978242099285126,
-0.022944288328289986,
-0.10095753520727158,
-0.010783267207443714,
0.08176662027835846,
0.05096106976270676,
-0.023138441145420074,
-0.020587589591741562,
0.018284671008586884,
-0.01859953999519348,
-0.02868231013417244,
-0.05177097022533417,
-0.014189488254487514,
-0.018916796892881393,
0.013493778184056282,
0.09844909608364105,
-0.02366263046860695,
0.04038233309984207,
-0.05717764422297478,
0.016679037362337112,
0.06355930119752884,
-0.00954487081617117,
-0.005504928063601255,
0.01838887482881546,
-0.01599162444472313,
0.0818072110414505,
0.08883649855852127,
-0.014983383007347584,
0.051405906677246094,
0.05741783604025841,
-0.04667019471526146,
-0.060732536017894745,
0.053710322827100754,
0.028735142201185226,
0.0695168599486351,
-0.1268860399723053,
0.036889586597681046,
0.0527065135538578,
-0.036809492856264114,
-0.014572820626199245,
-0.014123638160526752,
-0.035120900720357895,
-0.08818260580301285,
-0.0520441010594368,
0.02774805761873722,
-0.0421176441013813,
-0.07962821424007416,
-0.04624153673648834,
-0.04631209373474121,
-0.032919805496931076,
-0.013228850439190865,
0.014591983519494534,
-0.03677499294281006,
-6.0485989360072e-34,
-0.03177042305469513,
-0.014575932174921036,
-0.05397224426269531,
-0.014404146932065487,
-0.09029856324195862,
-0.011905103921890259,
-0.0024365808349102736,
-0.018318137153983116,
-0.012267469428479671,
-0.011953129433095455,
0.03548252210021019,
-0.0688653364777565,
0.04972545802593231,
-0.08181726187467575,
0.02497502602636814,
0.08073247224092484,
-0.04731864854693413,
0.04508484527468681,
-0.03674701228737831,
0.026583969593048096,
-0.0046464079059660435,
0.16026633977890015,
-0.0464228130877018,
0.10293601453304291,
-0.049439676105976105,
0.07294086366891861,
-0.03930668532848358,
0.014251223765313625,
-0.04569690302014351,
-0.051287878304719925,
-0.06929782778024673,
-0.01545007899403572,
-0.03007393516600132,
-0.02935258112847805,
-0.0786874070763588,
0.009781782515347004,
-0.02663091942667961,
-0.02620573341846466,
-0.009878021664917469,
0.13100048899650574,
0.048118267208337784,
-0.032968875020742416,
0.03723135218024254,
0.049376048147678375,
-0.006495197769254446,
-0.08092589676380157,
0.011253383941948414,
-0.01954157091677189,
-0.014786278828978539,
0.10019853711128235,
-0.036139629781246185,
0.03441963344812393,
-0.07379604130983353,
0.00974325742572546,
-0.12461383640766144,
-0.04422057792544365,
-0.10902964323759079,
0.03643647953867912,
0.10984890162944794,
-0.024336135014891624,
-0.042665936052799225,
-0.002641636412590742,
0.0056950789876282215,
0.020665735006332397,
-0.007377196568995714,
-0.017224984243512154,
0.04177335277199745,
-0.0760716125369072,
0.07682044059038162,
-0.03699367120862007,
0.06911422312259674,
0.020582135766744614,
0.040605492889881134,
0.02925930917263031,
0.0056911855936050415,
-0.023062093183398247,
0.02896985411643982,
-0.04220593720674515,
-0.05820857360959053,
-0.02901545912027359,
0.0023794902954250574,
0.0610199011862278,
0.03937651962041855,
0.03878498077392578,
0.025304747745394707,
0.11548417061567307,
0.07263810932636261,
0.030537158250808716,
0.043002527207136154,
-0.04770727455615997,
0.003248967230319977,
0.04969492554664612,
-0.05566025525331497,
0.09996992349624634,
0.005209796596318483,
-4.365378458714986e-8,
-0.10763611644506454,
0.07689505815505981,
-0.05882878601551056,
0.008958070538938046,
0.061931729316711426,
-0.1113874614238739,
0.005054954439401627,
0.13960464298725128,
-0.0841134637594223,
-0.021872444078326225,
0.10633733868598938,
0.025855187326669693,
0.0032964583951979876,
-0.07419199496507645,
0.04754815995693207,
0.003708621021360159,
0.023543035611510277,
0.015777267515659332,
-0.0019765403121709824,
-0.037786874920129776,
-0.020636487752199173,
-0.03345328941941261,
-0.002580279251560569,
0.03094620071351528,
-0.00537289772182703,
-0.04569633677601814,
-0.010299292393028736,
0.05844442546367645,
0.07906271517276764,
-0.018440013751387596,
0.02837563119828701,
0.0028568513225764036,
0.02471359819173813,
-0.008021626621484756,
-0.023930024355649948,
0.02217220328748226,
-0.0057884324342012405,
-0.01444641500711441,
-0.007559844292700291,
0.03226536139845848,
0.011852378956973553,
0.06201513111591339,
0.01521669328212738,
0.051707733422517776,
-0.0728193148970604,
0.038981664925813675,
0.012210417538881302,
-0.0964045450091362,
0.06553635746240616,
-0.026821685954928398,
0.10676895081996918,
-0.003453013254329562,
-0.007766700349748135,
0.03228483349084854,
0.00046963075874373317,
0.012170610949397087,
0.03570379316806793,
0.03814045339822769,
0.007387974299490452,
0.018309524282813072,
0.0036588364746421576,
0.030125092715024948,
-0.10183127969503403,
-0.04211962968111038
] |
Learning Gain Map for Inverse Tone Mapping | https://openreview.net/forum?id=GtHRhpgpzB | [
"Yinuo Liao",
"Yuanshen Guan",
"Ruikang Xu",
"Jiacheng Li",
"Shida Sun",
"Zhiwei Xiong"
] | Poster | For a more compatible and consistent high dynamic range (HDR) viewing experience, a new image format with a double-layer structure has been developed recently, which incorporates an auxiliary Gain Map (GM) within a standard dynamic range (SDR) image for adaptive HDR display. This new format motivates us to introduce a new task termed Gain Map-based Inverse Tone Mapping (GM-ITM), which focuses on learning the corresponding GM of an SDR image instead of directly estimating its HDR counterpart, thereby enabling a more effective up-conversion by leveraging the advantages of GM. The main challenge in this task, however, is to accurately estimate regional intensity variation with the fluctuating peak value. To this end, we propose a dual-branch network named GMNet, consisting of a Local Contrast Restoration (LCR) branch and a Global Luminance Estimation (GLE) branch to capture pixel-wise and image-wise information for GM estimation. Moreover, to facilitate the future research of the GM-ITM task, we build both synthetic and real-world datasets for comprehensive evaluations: synthetic SDR-GM pairs are generated from existing HDR resources, and real-world SDR-GM pairs are captured by mobile devices. Extensive experiments on these datasets demonstrate the superiority of our proposed GMNet over existing HDR-related methods both quantitatively and qualitatively. The codes and datasets are available at https://github.com/qtlark/GMNet. | Computational Photography, Inverse Tone Mapping, Gain Map | null | 2,310 | null | [
0.036057695746421814,
-0.12356801331043243,
0.05934799835085869,
-0.018952012062072754,
0.030294138938188553,
0.004727586638182402,
-0.04510404169559479,
-0.007469541393220425,
-0.04710599035024643,
-0.04174361377954483,
-0.07985654473304749,
0.008724536746740341,
0.04350178688764572,
-0.015632981434464455,
-0.043275125324726105,
0.029155436903238297,
0.1088775023818016,
0.057090774178504944,
-0.007437050808221102,
-0.06856755912303925,
0.016805250197649002,
-0.002488080644980073,
-0.01929793693125248,
-0.03745260089635849,
0.08872421085834503,
-0.08110496401786804,
0.03233867511153221,
0.06461451947689056,
-0.034507717937231064,
-0.013333846814930439,
0.06908038258552551,
-0.010338659398257732,
0.12025021761655807,
-0.005922486074268818,
-0.052176378667354584,
0.0038212027866393328,
0.06575432419776917,
-0.008448701351881027,
-0.0002941782004199922,
-0.01753266714513302,
0.021476805210113525,
0.030136141926050186,
0.04823709651827812,
-0.09421681612730026,
0.022179948166012764,
0.03089224174618721,
0.02839546836912632,
-0.030207067728042603,
-0.0036141660530120134,
0.03080696053802967,
0.08298598974943161,
0.02168196253478527,
-0.06658398360013962,
0.03512609004974365,
0.008773698471486568,
0.17705892026424408,
-0.05048735439777374,
-0.04224179685115814,
0.0328679122030735,
0.02317642793059349,
-0.099864162504673,
0.025122975930571556,
-0.025122661143541336,
-0.02835523709654808,
-0.0081633310765028,
0.006182903423905373,
0.03296925872564316,
-0.031210962682962418,
-0.0042886859737336636,
-0.06503933668136597,
-0.06889653950929642,
-0.003438942600041628,
-0.027750197798013687,
-0.061082713305950165,
-0.029270993545651436,
0.008135374635457993,
0.01782624050974846,
-0.025984853506088257,
0.059741098433732986,
-0.048593487590551376,
0.05407296493649483,
0.021556571125984192,
0.053570717573165894,
-0.0027698948979377747,
0.06014826521277428,
0.03980330005288124,
-0.01674792356789112,
0.06796106696128845,
0.012600955553352833,
0.09880826622247696,
-0.08764682710170746,
-0.019235478714108467,
-0.06132669374346733,
0.04683881625533104,
0.008537010289728642,
-0.015974126756191254,
0.058746807277202606,
-0.11669321358203888,
0.09643859416246414,
0.07788635045289993,
0.03826269879937172,
-0.07583490014076233,
0.009026742540299892,
-0.05453847348690033,
-0.0194980651140213,
-0.03576939180493355,
0.08921121060848236,
0.0811343789100647,
0.03827187418937683,
-0.006346425972878933,
0.05603202432394028,
0.018971117213368416,
0.014111117459833622,
-0.04205605387687683,
0.0232858844101429,
0.009369439445436,
0.0024405128788203,
-0.026760263368487358,
-0.010509127750992775,
-0.037330642342567444,
-0.033300407230854034,
-0.07762955874204636,
-0.010164405219256878,
0.0839788168668747,
0.033810805529356,
-0.06663652509450912,
-0.05666632950305939,
4.3345547756121214e-35,
-0.016008514910936356,
0.021113188937306404,
0.09961070865392685,
0.029754849150776863,
0.004812794271856546,
0.03213493153452873,
-0.03240140527486801,
-0.0003327435115352273,
-0.024525407701730728,
0.003495044307783246,
-0.056594546884298325,
0.10187683254480362,
-0.049249209463596344,
0.060941845178604126,
-0.05474039912223816,
0.025689030066132545,
0.0150837367400527,
-0.0126681262627244,
-0.03338121995329857,
0.05289921909570694,
-0.08897386491298676,
-0.026904381811618805,
0.03944472223520279,
-0.052441149950027466,
0.009620130993425846,
0.007953488267958164,
0.05871140956878662,
0.032675281167030334,
0.04218272492289543,
-0.01551427785307169,
-0.050094448029994965,
-0.03036651574075222,
0.02017579972743988,
0.012411815114319324,
-0.028069471940398216,
-0.05942627787590027,
-0.005185897462069988,
0.014493207447230816,
-0.07006283104419708,
0.01867707632482052,
-0.03437299281358719,
0.10099092125892639,
-0.0024372930638492107,
-0.018551599234342575,
-0.05931750312447548,
0.01735459454357624,
0.04049073904752731,
-0.014916487969458103,
-0.08276175707578659,
0.08126907795667648,
0.033613886684179306,
-0.05375085026025772,
-0.040951766073703766,
-0.09248600155115128,
-0.03900117054581642,
0.020135076716542244,
0.03809808939695358,
0.04406850412487984,
0.06846826523542404,
0.09327447414398193,
0.028166944161057472,
0.05837792158126831,
0.022678567096590996,
0.058541927486658096,
-0.046992722898721695,
0.028848933055996895,
0.0362822487950325,
-0.010798654519021511,
-0.09618787467479706,
0.015297060832381248,
0.054030630737543106,
-0.010396766476333141,
0.08760116249322891,
0.011956760659813881,
0.04007880389690399,
0.032041456550359726,
0.006155843380838633,
-0.027624398469924927,
0.049307335168123245,
0.05965340510010719,
-0.031775668263435364,
0.020557008683681488,
-0.019352944567799568,
-0.030949877575039864,
-0.05991709604859352,
0.06918459385633469,
-0.03133828938007355,
-0.10758299380540848,
-0.017579270526766777,
-0.043116774410009384,
-0.08701776713132858,
-0.08058197796344757,
-0.02956567332148552,
0.005318985786288977,
0.015413661487400532,
-4.432126603846066e-34,
0.10397903621196747,
0.061356525868177414,
-0.07099991291761398,
0.02764389105141163,
-0.01345417182892561,
-0.00940299965441227,
0.03693224489688873,
0.06271370500326157,
0.008527512662112713,
-0.06040806323289871,
0.10655388236045837,
-0.0076738763600587845,
0.009869021363556385,
0.024265222251415253,
0.022535501047968864,
-0.03774544969201088,
-0.006166167091578245,
0.0070868125185370445,
-0.10584414005279541,
0.004925820976495743,
0.019070589914917946,
0.05097987502813339,
0.008651639334857464,
0.04294588044285774,
-0.10577154904603958,
0.04117611050605774,
-0.04313034936785698,
0.10993999242782593,
0.01500346977263689,
-0.023489978164434433,
-0.04893884062767029,
0.006659530568867922,
0.05386398360133171,
-0.05351254716515541,
-0.024552183225750923,
0.07982580363750458,
0.0805196464061737,
-0.08576971292495728,
-0.09804777801036835,
0.03477099537849426,
0.02033662050962448,
0.008723258040845394,
0.03474767506122589,
0.001909213955514133,
-0.00319660478271544,
0.028780702501535416,
0.05124307796359062,
0.011233857832849026,
0.01249862089753151,
0.01696886122226715,
0.006197370123118162,
-0.052061401307582855,
-0.06787970662117004,
0.018012674525380135,
-0.00430330028757453,
-0.05485326051712036,
-0.013943302445113659,
0.03006332740187645,
0.06336618959903717,
0.024587146937847137,
-0.043522775173187256,
-0.07796096801757812,
-0.10284940898418427,
-0.03304673358798027,
-0.062468621879816055,
0.0011683824704959989,
0.0074805887416005135,
-0.09461389482021332,
-0.003168009454384446,
0.09809058159589767,
-0.0076838345266878605,
-0.0617428682744503,
0.03074668161571026,
-0.017363084480166435,
0.01953515224158764,
-0.048942264169454575,
-0.043954744935035706,
-0.023635530844330788,
-0.019251801073551178,
-0.01802455075085163,
0.035034697502851486,
-0.02948940172791481,
0.02210240624845028,
0.10320886969566345,
0.08227719366550446,
0.033279046416282654,
-0.043834034353494644,
-0.020253004506230354,
0.06747176498174667,
-0.014991563744843006,
-0.04762016236782074,
0.06322712451219559,
-0.11285455524921417,
-0.048565398901700974,
-0.08887924253940582,
-4.326174263269422e-8,
-0.0637703463435173,
0.017642823979258537,
-0.004113057628273964,
-0.006677554454654455,
-0.05785921588540077,
-0.05688191205263138,
0.1016659364104271,
-0.00817461870610714,
0.03027360700070858,
-0.07355313748121262,
0.09658824652433395,
-0.0895901471376419,
0.025225123390555382,
-0.036373429000377655,
-0.026531947776675224,
0.05144480988383293,
0.031578972935676575,
0.041346777230501175,
0.0006888399366289377,
0.02569815143942833,
-0.03450331091880798,
0.09026860445737839,
0.08129587769508362,
-0.03051850199699402,
0.07105004042387009,
-0.0525701530277729,
-0.06508869677782059,
0.041609253734350204,
0.06460701674222946,
-0.06125522032380104,
0.035336051136255264,
0.03267212212085724,
0.015031466260552406,
-0.019753146916627884,
-0.06362418085336685,
0.009784527122974396,
-0.010220066644251347,
0.05990942195057869,
0.005566265434026718,
0.019583558663725853,
-0.09392523020505905,
0.014610864222049713,
-0.025286713615059853,
-0.003444527741521597,
-0.020936431363224983,
0.015087162144482136,
0.06555888056755066,
0.028359411284327507,
-0.010316192172467709,
-0.02434617467224598,
0.04827316105365753,
0.014356652274727821,
0.000008085658919299021,
0.02470691129565239,
0.11045324802398682,
-0.08793903887271881,
0.04547091946005821,
-0.021324248984456062,
0.016123641282320023,
0.11082711070775986,
-0.018183022737503052,
-0.03341804817318916,
-0.11874300986528397,
-0.008622365072369576
] |
SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression | https://openreview.net/forum?id=LNYIUouhdt | [
"Xin Wang",
"Yu Zheng",
"Zhongwei Wan",
"Mi Zhang"
] | Poster | The advancements in Large Language Models (LLMs) have been hindered by
their substantial sizes, which necessitates LLM compression methods for practical
deployment. Singular Value Decomposition (SVD) offers a promising solution for
LLM compression. However, state-of-the-art SVD-based LLM compression meth-
ods have two key limitations: truncating smaller singular values may lead to higher
compression loss, and the lack of update on the compressed weights after SVD
truncation. In this work, we propose SVD-LLM, a SVD-based post-training LLM
compression method that addresses the limitations of existing methods. SVD-LLM
incorporates a truncation-aware data whitening technique to ensure a direct map-
ping between singular values and compression loss. Moreover, SVD-LLM adopts
a parameter update with sequential low-rank approximation to compensate for
the accuracy degradation after SVD compression. We evaluate SVD-LLM on 10
datasets and seven models from three different LLM families at three different
scales. Our results demonstrate the superiority of SVD-LLM over state-of-the-arts,
especially at high model compression ratios. | Large Language Model; Post-training Model Compression | We design a SVD-based post-training method to compress large language models | 2,301 | null | [
0.010539783164858818,
0.025338971987366676,
0.017147092148661613,
-0.0310169979929924,
0.08288506418466568,
-0.0003900790470652282,
-0.03348398953676224,
-0.013503444381058216,
0.020577669143676758,
-0.008575982414186,
-0.036935001611709595,
0.04844951629638672,
-0.03708997741341591,
0.006420105695724487,
-0.06374487280845642,
-0.009525254368782043,
-0.013215644285082817,
0.11137929558753967,
-0.044610559940338135,
-0.06422386318445206,
-0.00016252398199867457,
0.046444911509752274,
0.014030995778739452,
0.040713831782341,
0.13665826618671417,
0.02894633822143078,
0.005876483395695686,
-0.030753090977668762,
0.043630633503198624,
-0.015466040931642056,
0.04022558405995369,
-0.01418924331665039,
0.03245333209633827,
0.029319336637854576,
-0.02993694506585598,
0.04931727424263954,
0.05172581970691681,
0.03922130540013313,
-0.07795555144548416,
0.0005274637951515615,
0.03671691194176674,
0.043046772480010986,
0.03492589294910431,
0.03609064966440201,
-0.015276990830898285,
0.047953736037015915,
0.003301067277789116,
-0.06463319808244705,
-0.03723396360874176,
0.02782117761671543,
-0.018119949847459793,
0.029814910143613815,
-0.05462033674120903,
0.09361076354980469,
-0.11705534160137177,
-0.04863370582461357,
0.019264107570052147,
0.03430062532424927,
-0.03728362172842026,
-0.0010171569883823395,
-0.08762996643781662,
-0.08041781187057495,
0.022341595962643623,
-0.005525949411094189,
0.038223206996917725,
-0.029599128291010857,
0.07936722785234451,
0.07974839955568314,
-0.08183716237545013,
0.054776035249233246,
-0.017313983291387558,
0.006340888328850269,
0.010461683385074139,
0.12256218492984772,
-0.10335821658372879,
0.08912590891122818,
0.03876784071326256,
-0.03389373794198036,
0.08229491114616394,
0.005878685973584652,
0.024582313373684883,
0.013098853640258312,
0.026246095076203346,
-0.07104771584272385,
0.03766172006726265,
0.0019125319086015224,
0.04394296929240227,
0.003281635232269764,
0.011615175753831863,
0.07283668220043182,
-0.05159269645810127,
-0.014288809150457382,
0.045460294932127,
-0.016231000423431396,
-0.022680290043354034,
0.015275033190846443,
0.02506670169532299,
0.03069695457816124,
-0.007044930476695299,
0.06650127470493317,
0.017804265022277832,
0.07553566247224808,
0.01656939648091793,
-0.025120941922068596,
-0.007876022718846798,
-0.04588288441300392,
0.010677090846002102,
0.03597842529416084,
0.05817648023366928,
-0.021071624010801315,
0.03340250998735428,
0.014390179887413979,
0.0252789705991745,
-0.020069342106580734,
0.05036765709519386,
0.007774355821311474,
0.006547801662236452,
-0.03589823469519615,
0.035710159689188004,
0.03805520385503769,
-0.07152057439088821,
-0.07174556702375412,
-0.018180593848228455,
0.06628746539354324,
0.0052298931404948235,
0.010732422582805157,
-0.06568478047847748,
2.4215405868839825e-33,
0.04043462499976158,
0.06050873175263405,
-0.0401763841509819,
0.01175639033317566,
0.032778557389974594,
-0.03365293890237808,
-0.018959080800414085,
0.030699996277689934,
-0.011978103779256344,
0.02468392252922058,
-0.0786605253815651,
-0.08568970859050751,
0.0059792762622237206,
0.0655321478843689,
0.007505068089812994,
0.008139854297041893,
0.06557166576385498,
0.02481946535408497,
-0.08848078548908234,
-0.06836363673210144,
0.1175055280327797,
-0.008411267772316933,
0.04193078726530075,
-0.06271765381097794,
0.08223434537649155,
-0.0048344810493290424,
0.042206015437841415,
-0.002022974891588092,
-0.05990719050168991,
-0.007596300449222326,
-0.04600094258785248,
-0.008069895207881927,
0.0028552296571433544,
-0.0012614004081115127,
0.0699368566274643,
-0.030336681753396988,
-0.03190295770764351,
-0.007799035403877497,
-0.012661426328122616,
-0.05305301398038864,
0.0010211810003966093,
-0.019834809005260468,
-0.06485889852046967,
-0.03425706550478935,
-0.06242693215608597,
-0.06453322619199753,
0.05180775374174118,
-0.03860354423522949,
-0.034581102430820465,
-0.05456440523266792,
0.07217168807983398,
-0.0011161069851368666,
-0.06145578995347023,
0.045735910534858704,
-0.05586680397391319,
-0.05206583812832832,
0.061272479593753815,
-0.13440868258476257,
0.020967887714505196,
0.021555300801992416,
-0.048314936459064484,
0.05363504961133003,
-0.024082327261567116,
0.04353048652410507,
0.048403024673461914,
-0.0661170706152916,
0.03852405399084091,
0.04408048465847969,
0.0006528662634082139,
-0.04009988158941269,
-0.03557958826422691,
-0.08062303066253662,
0.040741823613643646,
-0.01216433197259903,
-0.02018575184047222,
-0.02413397841155529,
0.04792541638016701,
-0.08565659075975418,
0.02813797816634178,
0.006226797122508287,
-0.0006603689398616552,
0.007261923048645258,
0.01738058030605316,
-0.12078309804201126,
0.025872940197587013,
-0.07086525112390518,
0.04184601455926895,
-0.0635761171579361,
0.012159043923020363,
-0.07488314807415009,
-0.07139571011066437,
-0.04076290875673294,
0.017286846414208412,
-0.033534131944179535,
-0.02117125317454338,
-2.5726683854394345e-33,
-0.09795839339494705,
0.013406259939074516,
-0.08342593908309937,
0.08793147653341293,
-0.05747658386826515,
-0.05672954022884369,
0.036815159022808075,
0.1070355623960495,
-0.06979882717132568,
-0.08567941188812256,
-0.040942106395959854,
-0.13339558243751526,
-0.03460581600666046,
0.057510219514369965,
0.10204140841960907,
0.0021853067446500063,
-0.019126400351524353,
-0.001345236087217927,
-0.055497124791145325,
0.044285207986831665,
-0.07541146874427795,
0.07467847317457199,
-0.042743366211652756,
0.07356035709381104,
0.03157544881105423,
-0.00010925989772658795,
0.05995825305581093,
0.04834870621562004,
-0.022254424169659615,
-0.0234075877815485,
-0.06947049498558044,
-0.002605783985927701,
-0.0032747196964919567,
-0.006389849353581667,
-0.08935490995645523,
0.026782160624861717,
0.04611494019627571,
0.016890432685613632,
0.008833756670355797,
0.11695566773414612,
0.08590072393417358,
-0.027667516842484474,
-0.0910753533244133,
0.054168716073036194,
-0.019192831590771675,
-0.05673239007592201,
-0.11602745205163956,
0.024288883432745934,
0.032622311264276505,
-0.025837760418653488,
0.020786520093679428,
0.02747480943799019,
-0.038064561784267426,
0.04840000718832016,
0.05446431040763855,
-0.12141678482294083,
-0.13025622069835663,
-0.006914228200912476,
-0.0867253914475441,
-0.03794914484024048,
-0.03512023016810417,
-0.04311910644173622,
0.037194158881902695,
-0.0173847284168005,
0.07395030558109283,
-0.03564104437828064,
0.0049039279110729694,
0.02879106067121029,
-0.06280088424682617,
-0.004009941127151251,
-0.02974054589867592,
0.004218213725835085,
0.010917684063315392,
0.0582125298678875,
-0.006715932860970497,
-0.09471755474805832,
-0.0035997594241052866,
0.026199443265795708,
0.02884862571954727,
-0.02366272546350956,
0.03187815845012665,
0.03287860006093979,
0.010006827302277088,
0.10095270723104477,
0.055331990122795105,
0.13046739995479584,
0.004148497711867094,
0.07290183752775192,
-0.0001939296635100618,
0.058490779250860214,
-0.030537616461515427,
0.07662086933851242,
0.022760262712836266,
0.050727732479572296,
0.012678046710789204,
-4.028434830161132e-8,
-0.08013691008090973,
-0.018102683126926422,
-0.06859064102172852,
0.03379913046956062,
0.039175257086753845,
-0.009825635701417923,
-0.03103453293442726,
0.05340932682156563,
0.05778295174241066,
-0.022066565230488777,
-0.03371860459446907,
0.011896605603396893,
-0.10887973010540009,
-0.04960460588335991,
-0.02698388695716858,
0.032375916838645935,
-0.04513954371213913,
0.03344271704554558,
0.0026209258940070868,
0.054363664239645004,
-0.007834915071725845,
0.0779675543308258,
-0.028355158865451813,
0.05042601004242897,
0.06326398253440857,
-0.051411647349596024,
-0.0035079584922641516,
0.1103547140955925,
0.11066687107086182,
0.0019892100244760513,
-0.057427987456321716,
0.012204659171402454,
-0.03432100638747215,
-0.0539344847202301,
0.058807890862226486,
0.06606924533843994,
0.09449870884418488,
0.0030204036738723516,
-0.024268532171845436,
0.04876559227705002,
0.025575043633580208,
0.06217744201421738,
-0.05403442308306694,
0.02033877745270729,
-0.016829026862978935,
0.002992736641317606,
-0.028766881674528122,
-0.07157847285270691,
-0.05306418240070343,
-0.029298853129148483,
0.011948849074542522,
0.005875005852431059,
-0.02139194682240486,
0.07291135936975479,
0.002437891438603401,
-0.06943488121032715,
0.0015131453983485699,
0.011039890348911285,
0.03683951869606972,
0.0786895677447319,
0.04546027258038521,
-0.09435177594423294,
-0.032178621739149094,
0.0095198480412364
] |
CG-Bench: Clue-grounded Question Answering Benchmark for Long Video Understanding | https://openreview.net/forum?id=le4IoZZHy1 | [
"Guo Chen",
"Yicheng Liu",
"Yifei Huang",
"Baoqi Pei",
"Jilan Xu",
"Yuping He",
"Tong Lu",
"Yali Wang",
"Limin Wang"
] | Poster | The existing video understanding benchmarks for multimodal large language models (MLLMs) mainly focus on short videos. The few benchmarks for long video understanding often rely on multiple-choice questions (MCQs). Due to the limitations of MCQ evaluations and the advanced reasoning abilities of MLLMs, models can often answer correctly by combining short video insights with elimination, without truly understanding the content. To bridge this gap, we introduce CG-Bench, a benchmark for clue-grounded question answering in long videos. CG-Bench emphasizes the model's ability to retrieve relevant clues, enhancing evaluation credibility. It includes 1,219 manually curated videos organized into 14 primary, 171 secondary, and 638 tertiary categories, making it the largest benchmark for long video analysis. The dataset features 12,129 QA pairs in three question types: perception, reasoning, and hallucination. To address the limitations of MCQ-based evaluation, we develop two novel clue-based methods: clue-grounded white box and black box evaluations, assessing whether models generate answers based on accurate video understanding. We evaluated multiple closed-source and open-source MLLMs on CG-Bench. The results show that current models struggle significantly with long videos compared to short ones, and there is a notable gap between open-source and commercial models. We hope CG-Bench will drive the development of more reliable and capable MLLMs for long video comprehension. | Long Video Understanding | null | 2,298 | null | [
0.0039660329930484295,
-0.13694912195205688,
-0.02216312661767006,
0.01719255931675434,
0.1224617287516594,
0.05214580520987511,
-0.055768296122550964,
0.03773866593837738,
0.06426455080509186,
-0.057987064123153687,
-0.07191059738397598,
-0.09173473715782166,
-0.03696174547076225,
0.07560963183641434,
-0.037769630551338196,
0.000650322821456939,
0.033210255205631256,
0.023511173203587532,
-0.1157028004527092,
-0.03501567989587784,
0.03310910612344742,
-0.009179106913506985,
0.09364711493253708,
-0.029597966000437737,
0.010590330697596073,
-0.012698291800916195,
-0.0030157286673784256,
-0.023692339658737183,
-0.009114754386246204,
-0.03636520355939865,
0.050689637660980225,
0.03323773667216301,
-0.0011299130273982882,
0.0733930766582489,
0.014482209458947182,
0.07520866394042969,
-0.06523171812295914,
-0.009724202565848827,
-0.014645719900727272,
-0.050585050135850906,
0.00016526180843356997,
0.017493730410933495,
0.09765022248029709,
0.0002530332712922245,
0.0741824060678482,
-0.013819134794175625,
-0.0542396605014801,
-0.06905384361743927,
0.017114045098423958,
-0.0345769077539444,
-0.07750201970338821,
0.0007492100703530014,
-0.026885122060775757,
0.07252627611160278,
-0.05487089604139328,
0.017883354797959328,
-0.038385163992643356,
-0.009304573759436607,
0.03818322345614433,
-0.007416741456836462,
0.008608557283878326,
-0.12416873872280121,
-0.025542421266436577,
0.019934605807065964,
0.0148171903565526,
-0.04002797231078148,
0.010052379220724106,
-0.02222711779177189,
-0.026222703978419304,
-0.0008789487183094025,
-0.06328268349170685,
0.0472201369702816,
0.006633665878325701,
0.07192621380090714,
-0.010825078934431076,
0.043812885880470276,
0.023500127717852592,
-0.03893130645155907,
0.020882971584796906,
-0.11228667199611664,
0.07130742818117142,
-0.028763314709067345,
0.06586854159832001,
0.029069995507597923,
0.11363564431667328,
0.013085783459246159,
0.021581074222922325,
0.07101292163133621,
-0.10975063592195511,
0.028177062049508095,
-0.09687539935112,
-0.05076293647289276,
0.04378679394721985,
-0.032047074288129807,
0.08396925032138824,
0.04797365888953209,
0.007654912304133177,
-0.11551083624362946,
-0.0611601360142231,
0.040676772594451904,
0.04917219281196594,
0.0735621303319931,
0.010888095945119858,
-0.050298236310482025,
-0.004775808658450842,
-0.04208981990814209,
0.06090475246310234,
0.059475723654031754,
0.02975195273756981,
-0.07840524613857269,
0.0012428711634129286,
0.09244052320718765,
0.04595809057354927,
-0.024204570800065994,
0.03089510276913643,
0.03277077153325081,
0.03970855101943016,
0.05220961570739746,
-0.006806349381804466,
0.001258817152120173,
0.024147039279341698,
-0.013220957480370998,
0.00975112896412611,
-0.06909161061048508,
0.06496040523052216,
-0.009842540137469769,
0.002749995794147253,
8.823693142458994e-34,
0.04292850196361542,
0.005609284155070782,
0.09634243696928024,
0.0033531170338392258,
0.08257730305194855,
-0.00014062330592423677,
-0.06209458410739899,
0.09990343451499939,
-0.06426616758108139,
0.012180442921817303,
-0.006182579789310694,
-0.02845458686351776,
-0.11933853477239609,
0.057493750005960464,
-0.01583392545580864,
0.0659422054886818,
-0.14868368208408356,
-0.014943339861929417,
-0.06571263819932938,
0.01324678398668766,
0.05355142429471016,
-0.0016355186235159636,
0.042481064796447754,
0.055099908262491226,
0.013410492800176144,
-0.01998981647193432,
0.08794308453798294,
-0.09663476794958115,
-0.07613974064588547,
-0.01229908037930727,
-0.1691053956747055,
-0.045856621116399765,
-0.01901470310986042,
0.02028927020728588,
0.06871414184570312,
-0.0493672713637352,
-0.038273852318525314,
-0.06335415691137314,
0.012230364605784416,
-0.024355625733733177,
-0.03018610179424286,
-0.019834978505969048,
-0.05615509673953056,
0.04492080956697464,
-0.06335946917533875,
-0.03843769058585167,
0.008845504373311996,
-0.002835022285580635,
-0.061685845255851746,
0.030665718019008636,
0.06289064884185791,
0.025190910324454308,
-0.013391697779297829,
-0.054003436118364334,
-0.024772541597485542,
0.005126649513840675,
0.03646288067102432,
0.021830029785633087,
0.033373430371284485,
-0.002852807054296136,
0.039734821766614914,
0.04193739593029022,
-0.030254974961280823,
0.06290140748023987,
-0.00927222054451704,
0.019061194732785225,
-0.03420965373516083,
0.11498592793941498,
0.0737030878663063,
0.05812587961554527,
0.014280456118285656,
0.018632160499691963,
0.029222916811704636,
-0.054849773645401,
0.03288674354553223,
0.06121999770402908,
-0.0124592250213027,
-0.0719502717256546,
0.0603635273873806,
0.03607981279492378,
0.029483238235116005,
-0.01902618817985058,
0.015249970369040966,
-0.04823025315999985,
-0.0505734458565712,
-0.069290392100811,
0.0538368821144104,
-0.10014989227056503,
0.014918682165443897,
-0.04823184385895729,
-0.044599875807762146,
-0.017009112983942032,
0.006143194157630205,
-0.031176432967185974,
0.04204456880688667,
6.38039920609358e-34,
0.025093117728829384,
-0.022141270339488983,
-0.055600110441446304,
0.0315944142639637,
-0.001703395857475698,
-0.018778767436742783,
-0.015689032152295113,
0.0262664332985878,
0.04779569432139397,
-0.052904173731803894,
0.006243007257580757,
-0.059945378452539444,
-0.030245071277022362,
0.006860974244773388,
0.0326874777674675,
-0.0611654557287693,
-0.04842798411846161,
-0.02454596571624279,
-0.026659946888685226,
0.09630083292722702,
-0.0056846607476472855,
0.08218039572238922,
-0.06004878878593445,
-0.008164677768945694,
-0.05524073541164398,
0.0878395214676857,
-0.0351022370159626,
0.032449886202812195,
-0.026471052318811417,
-0.026670748367905617,
0.02273022197186947,
-0.003948705270886421,
-0.00889561790972948,
-0.014849083498120308,
-0.07417023926973343,
0.035194262862205505,
0.08742481470108032,
-0.05079048499464989,
-0.041846614331007004,
0.1407654583454132,
0.08177321404218674,
-0.018448026850819588,
-0.08941788226366043,
-0.035886138677597046,
-0.039104677736759186,
-0.029324937611818314,
-0.06865435838699341,
-0.007875385694205761,
-0.025500545278191566,
0.041667815297842026,
-0.007174608297646046,
0.02574671059846878,
-0.0876753032207489,
0.045638829469680786,
0.046077221632003784,
-0.07774795591831207,
-0.11989055573940277,
0.04628388211131096,
0.0618998259305954,
0.05808333307504654,
-0.038534026592969894,
0.003077967558056116,
-0.0020706187933683395,
-0.032390471547842026,
0.06892341375350952,
-0.0414673313498497,
-0.003228442743420601,
0.0021769015584141016,
-0.003174270037561655,
-0.008917059749364853,
-0.018915722146630287,
-0.020760372281074524,
0.011285021901130676,
0.02178136818110943,
0.03968296945095062,
0.025841010734438896,
-0.014211495406925678,
-0.0198952779173851,
0.051754213869571686,
-0.05754610523581505,
-0.0045247687958180904,
-0.09617368131875992,
0.03647075220942497,
0.07633545249700546,
0.05804915353655815,
0.055956389755010605,
0.0605778731405735,
0.08820775151252747,
-0.013908487744629383,
0.10022234916687012,
-0.008047712035477161,
0.015814341604709625,
0.03240009769797325,
0.06603379547595978,
-0.04390968754887581,
-3.4149991279264214e-8,
-0.0805700495839119,
0.057273782789707184,
-0.057620782405138016,
0.0025589410215616226,
-0.012663532048463821,
-0.019406253471970558,
-0.04377404600381851,
0.08100991696119308,
0.05368548631668091,
-0.0305264201015234,
0.03345813229680061,
-0.0005995461833663285,
-0.003654534462839365,
0.006189247593283653,
-0.05874310061335564,
0.09346844255924225,
-0.02693413756787777,
0.02247384935617447,
0.009816640987992287,
0.025435740128159523,
0.09806659072637558,
0.01660623587667942,
0.021771471947431564,
0.0041289483197033405,
-0.06313793361186981,
0.0022803046740591526,
-0.003831346519291401,
0.1049850657582283,
0.061710748821496964,
0.003343048971146345,
-0.006116995122283697,
0.012815439142286777,
0.009778440929949284,
-0.11519765853881836,
0.05064592882990837,
0.037542521953582764,
0.01885226182639599,
-0.037737179547548294,
0.04808655381202698,
0.041975561529397964,
0.026045598089694977,
0.0202559195458889,
-0.019434938207268715,
0.014910843223333359,
-0.016689661890268326,
0.03370684012770653,
-0.07376870512962341,
-0.03954842686653137,
0.00780699634924531,
0.0029059473890811205,
-0.04571943357586861,
-0.0651017352938652,
-0.023292506113648415,
0.10784830898046494,
0.062229372560977936,
-0.01874658279120922,
0.008025000803172588,
0.06606865674257278,
0.004242499824613333,
0.037855714559555054,
0.024893363937735558,
-0.003427274990826845,
-0.005494784563779831,
0.039175018668174744
] |
Parameter and Memory Efficient Pretraining via Low-rank Riemannian Optimization | https://openreview.net/forum?id=i0zzO7Hslk | [
"Zhanfeng Mo",
"Long-Kai Huang",
"Sinno Jialin Pan"
] | Poster | Pretraining large language models often requires significant computational resources and memory due to their vast parameter amount. An effective approach to enhance parameter efficiency in both training and inference is to parameterize each full-size weight as the product of two trainable low-rank factors. While low-rank fine-tuning has achieved great success, low-rank pretraining remains challenging as it requires learning extensive knowledge from scratch under the restrictive low-rank parameterization. During standard low-rank pretraining, separately optimizing the low-rank factors introduces redundant information from the full gradient, which hinders the learning process. To achieve efficient yet effective low-rank pretraining, we propose a **Lo**w-rank **R**iemannian **O**ptimizer (**LORO**). At each LORO update step, the low-rank factor pairs are jointly updated to ensure their full-size product moves along the steepest descent direction on the low-rank manifold, without the need to compute any memory-intensive full-size matrices or gradients. Hence, our LORO finds low-rank models that achieve high performance comparable to full-size pretrained models, while significantly reducing memory usage and accelerating both training and inference. A LLaMA 1B model pretrained with LORO achieves a perplexity score of 2\% better than the full-size baseline, with a 54\% reduction in model memory, a $\times1.8$ speedup in training, and a $\times2.2$ speedup in inference. The code is available on https://github.com/mzf666/LORO-main. | Neural Network Optimization, Parameter Efficient Pretraining, Low-rank Optimization | A Riemannian optimizer for pretraining fully low-rank parameterized language models. | 2,294 | null | [
0.0008682764018885791,
-0.09172594547271729,
0.012370191514492035,
0.04233372583985329,
0.03792605921626091,
0.07850995659828186,
-0.006732751615345478,
0.012446585111320019,
-0.025670602917671204,
-0.05811143293976784,
0.009082376956939697,
0.019139043986797333,
0.0031491403933614492,
0.03389323875308037,
0.027935774996876717,
0.060529835522174835,
0.04750509187579155,
0.13450977206230164,
-0.08133043348789215,
-0.06864817440509796,
0.019203152507543564,
-0.01394791528582573,
0.04157581552863121,
0.012714040465652943,
0.08432870358228683,
-0.031214548274874687,
0.0038935160264372826,
-0.01296444796025753,
0.031610943377017975,
-0.04914451017975807,
0.027763668447732925,
0.015204256400465965,
0.0514339953660965,
0.009896277450025082,
-0.058940544724464417,
0.06847114115953445,
-0.016065463423728943,
-0.02125551365315914,
-0.00236356514506042,
0.027528928592801094,
0.03493032231926918,
-0.0015886849723756313,
0.010338790714740753,
0.0698980987071991,
0.07674852013587952,
0.02830473519861698,
0.027771255001425743,
0.009218723513185978,
0.020299529656767845,
-0.008657308295369148,
-0.08749677240848541,
-0.007417682092636824,
-0.04343501478433609,
0.07061711698770523,
-0.05075858533382416,
-0.05819837376475334,
0.011288131587207317,
-0.0035814079456031322,
-0.011318482458591461,
-0.0019791284576058388,
-0.08876257389783859,
-0.1114904060959816,
0.008083434775471687,
-0.05537603050470352,
-0.044606804847717285,
-0.03619996830821037,
0.03843535855412483,
0.005214109551161528,
-0.017310813069343567,
0.08710998296737671,
-0.02947472222149372,
0.05751205235719681,
-0.027332445606589317,
0.04753878340125084,
-0.006391457747668028,
0.052844129502773285,
0.05714119225740433,
0.015956971794366837,
0.031378429383039474,
-0.01108572632074356,
0.029192054644227028,
0.0348302498459816,
0.0331190824508667,
-0.10318838804960251,
0.10356097668409348,
0.04134982079267502,
0.05127647891640663,
-0.02529590018093586,
0.062449727207422256,
-0.04705748334527016,
0.028551094233989716,
-0.06013205274939537,
-0.010103060863912106,
0.01953776739537716,
-0.0290296021848917,
0.017240820452570915,
0.019070565700531006,
-0.03698573261499405,
-0.021331384778022766,
0.11801706254482269,
0.05118110775947571,
-0.005870779510587454,
-0.020790183916687965,
0.06208482012152672,
-0.04486023634672165,
-0.049541451036930084,
0.05518645420670509,
0.09792209416627884,
-0.0387222059071064,
-0.08058004081249237,
0.0411933995783329,
0.021283118054270744,
0.000915073265787214,
0.05534006655216217,
-0.006830333266407251,
0.01915615238249302,
0.024406300857663155,
-0.06638667732477188,
0.024227268993854523,
0.06439806520938873,
-0.05176301673054695,
-0.012550479732453823,
0.024166440591216087,
0.04943850263953209,
-0.04622654616832733,
-0.004600724671036005,
-0.07974119484424591,
2.977313764939783e-33,
0.09933102130889893,
0.1024668961763382,
0.03455432504415512,
-0.029715178534388542,
-0.0678606778383255,
-0.0000032690390980860684,
0.035531092435121536,
0.016029387712478638,
-0.08551151305437088,
0.0211472325026989,
-0.043682146817445755,
-0.02336318977177143,
-0.06665382534265518,
0.07962604612112045,
-0.01575351320207119,
-0.030651774257421494,
0.004235582426190376,
0.034899916499853134,
0.001027859398163855,
-0.05872231721878052,
0.08990482240915298,
-0.004729677923023701,
-0.03246765211224556,
-0.02553218975663185,
0.0016980348154902458,
0.010066001676023006,
0.054847318679094315,
-0.1330159306526184,
-0.09977246075868607,
0.011948988772928715,
-0.11021676659584045,
-0.11416704952716827,
-0.0045461105182766914,
0.01934867724776268,
-0.02087954618036747,
-0.05558789521455765,
-0.06939234584569931,
-0.012772835791110992,
0.042342230677604675,
-0.10647675395011902,
-0.048664819449186325,
0.039993949234485626,
0.030729658901691437,
-0.026813773438334465,
-0.11592162400484085,
-0.09330674260854721,
-0.006690460257232189,
0.052744895219802856,
-0.03396227955818176,
-0.05576396733522415,
-0.020153667777776718,
0.0031704858411103487,
-0.020426884293556213,
0.000574967241846025,
-0.06807997077703476,
-0.05320543423295021,
0.0843682587146759,
0.005328884813934565,
0.014042406342923641,
0.060979172587394714,
-0.02274399995803833,
0.05879925191402435,
-0.04149915650486946,
0.05499850586056709,
-0.01149542722851038,
-0.06628381460905075,
0.008405433036386967,
0.05431780964136124,
0.0922359973192215,
0.03281721845269203,
-0.046327926218509674,
-0.051976144313812256,
0.02610107511281967,
-0.03407394140958786,
0.0634613186120987,
0.044161442667245865,
0.06252647936344147,
-0.0963868722319603,
0.042271289974451065,
0.021473202854394913,
-0.020662905648350716,
0.02315220609307289,
0.04632604122161865,
-0.06938133388757706,
-0.039506398141384125,
-0.0817062258720398,
0.02597694657742977,
-0.07226508855819702,
0.0001329364749835804,
-0.013233799487352371,
-0.008193251676857471,
0.007734087761491537,
-0.03076695092022419,
-0.03862793371081352,
-0.014805776998400688,
-1.130282909544243e-33,
0.058138128370046616,
-0.02818983979523182,
0.02650136686861515,
0.06398816406726837,
-0.044218726456165314,
-0.05340505391359329,
0.0025962654035538435,
0.05221047252416611,
-0.08730188757181168,
-0.011096329428255558,
0.024156786501407623,
-0.058784566819667816,
0.05901367962360382,
0.0074026454240083694,
0.13447797298431396,
-0.014467934146523476,
-0.03307729959487915,
0.06618735939264297,
0.05073051154613495,
0.03661100193858147,
-0.04233379662036896,
0.060332246124744415,
-0.037676453590393066,
0.04790162295103073,
-0.07464827597141266,
0.012888869270682335,
0.03351358696818352,
0.040463339537382126,
-0.06847614049911499,
0.03938170149922371,
-0.06272467225790024,
-0.006733990274369717,
-0.04104931280016899,
0.028233222663402557,
-0.054269108921289444,
0.06864012777805328,
0.024124225601553917,
-0.023623883724212646,
-0.021927090361714363,
0.0978047251701355,
0.07504729926586151,
-0.03654047101736069,
-0.07709108293056488,
-0.04311395436525345,
-0.029244927689433098,
-0.02138315699994564,
-0.10110639780759811,
-0.020274290814995766,
0.016463374719023705,
0.01081897597759962,
-0.041031330823898315,
0.04794720187783241,
0.02457137033343315,
-0.0028786014299839735,
-0.05598181113600731,
-0.11114592850208282,
-0.0796276181936264,
-0.035891253501176834,
0.055803555995225906,
-0.04410826042294502,
0.018259761855006218,
-0.01994754560291767,
-0.014988841488957405,
-0.08940224349498749,
-0.0025795195251703262,
-0.08233686536550522,
-0.004846244119107723,
0.029043078422546387,
-0.0012775711948052049,
0.04648704454302788,
-0.044952306896448135,
0.02921764925122261,
0.00953166838735342,
-0.002978944918140769,
-0.06699898093938828,
-0.0038061123341321945,
0.007864290848374367,
0.009342960081994534,
-0.00912785716354847,
-0.05882567539811134,
0.04189164936542511,
0.04280747473239899,
-0.010509161278605461,
0.05794079229235649,
0.04369045048952103,
0.13716496527194977,
0.06506475806236267,
0.09339294582605362,
0.05769307166337967,
0.004067907575517893,
-0.0034024203196167946,
-0.01232137717306614,
0.06972207874059677,
0.037255678325891495,
-0.06611926853656769,
-4.117062957220696e-8,
-0.09900688380002975,
0.02918328531086445,
0.0018302022945135832,
0.0885433480143547,
0.0149167375639081,
-0.04180966317653656,
-0.027604008093476295,
0.10083413869142532,
0.0010079607600346208,
0.030528122559189796,
0.03443610668182373,
0.022985396906733513,
-0.03547070920467377,
-0.07885854691267014,
-0.0032618825789541006,
0.08629768341779709,
-0.02853557839989662,
0.048334814608097076,
0.0056111812591552734,
-0.04001345485448837,
0.05506560951471329,
0.07789518684148788,
0.06193246319890022,
0.03519001975655556,
0.08938533067703247,
-0.10065379738807678,
0.04777100682258606,
0.08773251622915268,
0.14205600321292877,
-0.014228131622076035,
-0.04254239425063133,
-0.05505384877324104,
-0.02753724530339241,
-0.008959878236055374,
0.03944583982229233,
0.0810241624712944,
0.004166468046605587,
0.017446497455239296,
-0.05796642601490021,
-0.04516804963350296,
-0.021796492859721184,
0.031245684251189232,
-0.0607527457177639,
0.03806736320257187,
0.03791668266057968,
0.01922999694943428,
-0.08210747689008713,
-0.07517743855714798,
0.046790961176157,
-0.06088530272245407,
0.04932655766606331,
-0.006088294554501772,
0.030527858063578606,
0.07665771245956421,
0.00592505419626832,
0.04484807699918747,
-0.013085701502859592,
0.0018008638871833682,
0.0141919469460845,
-0.010840599425137043,
0.042482879012823105,
-0.018092507496476173,
0.0012105288915336132,
0.024457179009914398
] |
Latent Radiance Fields with 3D-aware 2D Representations | https://openreview.net/forum?id=vL9t9tpKli | [
"Chaoyi Zhou",
"Xi Liu",
"Feng Luo",
"Siyu Huang"
] | Poster | Latent 3D reconstruction has shown great promise in empowering 3D semantic understanding and 3D generation by distilling 2D features into the 3D space. However, existing approaches struggle with the domain gap between 2D feature space and 3D representations, resulting in degraded rendering performance. To address this challenge, we propose a novel framework that integrates 3D awareness into the 2D latent space. The framework consists of three stages: (1) a correspondence-aware autoencoding method that enhances the 3D consistency of 2D latent representations, (2) a latent radiance field (LRF) that lifts these 3D-aware 2D representations into 3D space, and (3) a VAE-Radiance Field (VAE-RF) alignment strategy that improves image decoding from the rendered 2D representations. Extensive experiments demonstrate that our method outperforms the state-of-the-art latent 3D reconstruction approaches in terms of synthesis performance and cross-dataset generalizability across diverse indoor and outdoor scenes. To our knowledge, this is the first work showing the radiance field representations constructed from 2D latent representations can yield photorealistic 3D reconstruction performance. | 3D Gaussian Splatting, 3D-aware Representation | To our knowledge, this is the first work demonstrating that radiance field representations in the latent space can achieve decent 3D reconstruction performance across various settings including indoor and unbounded outdoor scenes. | 2,290 | 2502.09613 | [
-0.035163044929504395,
-0.0935719758272171,
-0.014672587625682354,
-0.0026224025059491396,
0.0870962142944336,
-0.011639855802059174,
-0.087346151471138,
-0.08606912940740585,
0.0375371091067791,
0.006148537155240774,
-0.05122825875878334,
-0.06633792072534561,
-0.036141328513622284,
0.07903338223695755,
0.0056107258424162865,
0.027699731290340424,
0.04802825674414635,
0.1378854513168335,
-0.07103823870420456,
0.012871225364506245,
-0.004320215433835983,
0.005255056545138359,
0.004530295729637146,
-0.0360899455845356,
0.1263582408428192,
0.034978386014699936,
0.10611076653003693,
-0.013510181568562984,
-0.010865881107747555,
-0.004674088675528765,
0.020537598058581352,
0.006394765805453062,
-0.04859023168683052,
0.0029721802566200495,
-0.03768815100193024,
0.037467386573553085,
-0.018114719539880753,
-0.014149921014904976,
-0.10863450914621353,
-0.03716530650854111,
0.005303392186760902,
0.02566971257328987,
-0.05338070169091225,
0.02212393470108509,
0.0038584715221077204,
0.03269495442509651,
0.02842763438820839,
-0.10883501917123795,
0.06244106963276863,
-0.02881636656820774,
0.016706936061382294,
0.01383267343044281,
-0.036430153995752335,
0.035173673182725906,
-0.04677724465727806,
0.09317467361688614,
0.03335881978273392,
-0.051555901765823364,
0.012898806482553482,
0.017583612352609634,
0.030870042741298676,
0.010218165814876556,
0.0485522523522377,
-0.01818045601248741,
0.02637266181409359,
-0.07956697046756744,
-0.02189447544515133,
-0.03588176146149635,
-0.011749157682061195,
-0.06385830044746399,
-0.08575242012739182,
0.05220963805913925,
-0.019724305719137192,
-0.08803629875183105,
0.03680472448468208,
0.08990373462438583,
0.009560944512486458,
-0.06296882778406143,
0.015387979336082935,
-0.05809924378991127,
0.04844911769032478,
0.0710039734840393,
0.03989066928625107,
0.12281402945518494,
0.044190507382154465,
0.03208177536725998,
0.007659438066184521,
0.12491723150014877,
-0.042726606130599976,
-0.015815693885087967,
-0.09763295203447342,
-0.07615771144628525,
-0.06928227841854095,
-0.018323875963687897,
-0.0025857535656541586,
-0.028967995196580887,
0.11439718306064606,
-0.08954571932554245,
0.005669577047228813,
-0.01339293085038662,
0.021658897399902344,
0.02806946076452732,
0.031247403472661972,
0.041763316839933395,
0.08999001979827881,
-0.08151751756668091,
0.08342744410037994,
0.03392627090215683,
-0.006096040364354849,
-0.02680928260087967,
0.03935005143284798,
0.01675277203321457,
-0.005679903086274862,
0.01891765184700489,
0.03649498522281647,
0.07231062650680542,
-0.0054782661609351635,
-0.0033448119647800922,
0.043737780302762985,
0.005241988226771355,
-0.0528675802052021,
0.01571924425661564,
-0.04738205671310425,
-0.0014190145302563906,
0.07378977537155151,
0.006082635372877121,
0.045761868357658386,
1.7782626909980093e-33,
-0.02361365407705307,
0.17333529889583588,
0.03698968142271042,
0.05739103630185127,
0.0285427775233984,
-0.015463019721210003,
-0.04915279150009155,
0.014839316718280315,
0.0054636369459331036,
0.01291930116713047,
-0.04169729724526405,
-0.030853714793920517,
-0.003317075315862894,
0.09779749810695648,
-0.02488049305975437,
-0.07853496819734573,
-0.06738951057195663,
-0.010879449546337128,
-0.013242848217487335,
0.0698339194059372,
0.02395891211926937,
0.025410769507288933,
0.0217730849981308,
0.051858097314834595,
-0.06611145287752151,
-0.04940241202712059,
-0.007263858802616596,
0.029287002980709076,
-0.08273477107286453,
-0.020662743598222733,
-0.0843719094991684,
0.04289644956588745,
0.038518276065588,
-0.023258881643414497,
-0.029614293947815895,
-0.03360309451818466,
0.001754478900693357,
-0.08044574409723282,
-0.01660584844648838,
-0.017101094126701355,
0.018547583371400833,
0.1108090803027153,
-0.00014986352471169084,
-0.0772809088230133,
-0.08520270884037018,
-0.02825990878045559,
0.0662999227643013,
0.04228068143129349,
-0.03917618840932846,
0.015614937990903854,
0.06985285878181458,
0.026576068252325058,
-0.006511230021715164,
-0.05923876911401749,
-0.032397594302892685,
-0.08689091354608536,
0.04993653669953346,
-0.04294446483254433,
0.09260621666908264,
-0.020763276144862175,
-0.022732488811016083,
0.007647537160664797,
-0.03597956523299217,
0.0474703386425972,
-0.03337131813168526,
-0.008728362619876862,
0.03197282925248146,
0.041390206664800644,
0.0058532836847007275,
0.07101037353277206,
-0.030550336465239525,
0.002960383892059326,
0.00025275954976677895,
-0.06426439434289932,
-0.0006468855426646769,
-0.012699558399617672,
0.02383721061050892,
-0.09235968440771103,
0.049842823296785355,
0.030953343957662582,
-0.10620438307523727,
0.023953484371304512,
-0.028766710311174393,
-0.04274304211139679,
-0.055324096232652664,
-0.099838487803936,
0.026562411338090897,
-0.09096908569335938,
-0.025408565998077393,
-0.05590284615755081,
-0.046511210501194,
-0.06142599135637283,
0.016566313803195953,
-0.03253905847668648,
0.0011646264465525746,
-1.4306614809491739e-33,
0.05802517384290695,
-0.016923455521464348,
-0.06421931833028793,
0.0031830482184886932,
0.024401230737566948,
-0.04652949422597885,
-0.02273442968726158,
0.03534099459648132,
0.04343588650226593,
-0.045851580798625946,
0.00032920687226578593,
0.04857468977570534,
0.006218580529093742,
-0.04079866036772728,
0.03215527534484863,
-0.0011144491145387292,
0.023941311985254288,
-0.007300959900021553,
-0.006281913723796606,
-0.010055018588900566,
0.03250489383935928,
0.007230890449136496,
-0.02502208948135376,
0.032288599759340286,
-0.06309302151203156,
0.09229225665330887,
0.07781249284744263,
-0.03442744165658951,
-0.05108325555920601,
0.009460319764912128,
0.014741807244718075,
0.005317962262779474,
0.018922924995422363,
-0.007762691471725702,
-0.06044954061508179,
0.03514197841286659,
0.03302411362528801,
0.02798376977443695,
-0.06707465648651123,
-0.008787612430751324,
0.02790968306362629,
-0.05088372528553009,
-0.07567422837018967,
0.0503724105656147,
-0.047666244208812714,
-0.016320446506142616,
0.00864406954497099,
-0.014913897030055523,
0.02302497811615467,
0.1038406565785408,
-0.010258934460580349,
0.05179963260889053,
-0.047975096851587296,
0.0006784643628634512,
0.011604406870901585,
-0.047303706407547,
-0.07503336668014526,
-0.014287690632045269,
0.11287600547075272,
0.02015097625553608,
-0.0015083433827385306,
-0.03349672630429268,
-0.027104942128062248,
-0.009276436641812325,
-0.011870297603309155,
-0.0018526598578318954,
-0.055176712572574615,
0.057962000370025635,
0.06698457896709442,
0.0798310860991478,
0.049915339797735214,
-0.0018733764300122857,
0.008811955340206623,
0.029379626736044884,
0.046193331480026245,
-0.1303035467863083,
0.028164545074105263,
0.029255887493491173,
0.029042204841971397,
-0.04869046062231064,
0.03319476917386055,
-0.020411387085914612,
0.030291421338915825,
0.12811414897441864,
0.04253353923559189,
0.03485514596104622,
-0.012940543703734875,
0.01106962189078331,
-0.044992800801992416,
-0.01519771572202444,
-0.022588439285755157,
0.07729180157184601,
-0.05398682504892349,
0.04689093306660652,
-0.020144086331129074,
-3.823348393439119e-8,
-0.09738800674676895,
0.02398715727031231,
0.04926420748233795,
-0.022959060966968536,
-0.10845888406038284,
-0.07803632318973541,
0.01689734309911728,
0.11815052479505539,
0.014092743396759033,
-0.03396731615066528,
0.052744705229997635,
-0.06337099522352219,
0.01714632660150528,
-0.07001736015081406,
0.07069605588912964,
0.13136811554431915,
-0.018303537741303444,
0.006105019710958004,
-0.006285023409873247,
0.08353304862976074,
-0.041090041399002075,
0.0872991681098938,
-0.007005091290920973,
-0.04909950867295265,
0.030393242835998535,
0.04072875902056694,
-0.06428363174200058,
0.06411995738744736,
0.11030729860067368,
0.0043479143641889095,
0.011207593604922295,
-0.004340701270848513,
0.054530978202819824,
-0.04792087525129318,
-0.08322960138320923,
0.016333771869540215,
0.03223352134227753,
-0.07703401148319244,
-0.011022270657122135,
-0.08673487603664398,
-0.05606360733509064,
0.08921673148870468,
-0.013622094877064228,
0.0076360576786100864,
0.025304671376943588,
0.03377595171332359,
0.06252572685480118,
0.015752948820590973,
-0.011351272463798523,
-0.015408828854560852,
-0.06461985409259796,
-0.009930875152349472,
-0.0007735005929134786,
0.06682520359754562,
0.11163227260112762,
0.015287691727280617,
-0.014360488392412663,
-0.004287717863917351,
0.06512225419282913,
-0.012937323190271854,
-0.02695711888372898,
0.019783375784754753,
-0.03249680623412132,
0.02293982356786728
] |
Memory Efficient Transformer Adapter for Dense Predictions | https://openreview.net/forum?id=vJkktqyU8B | [
"Dong Zhang",
"Rui Yan",
"Pingcheng Dong",
"Kwang-Ting Cheng"
] | Poster | While current Vision Transformer (ViT) adapter methods have shown promising accuracy, their inference speed is implicitly hindered by inefficient memory access operations, e.g., standard normalization and frequent reshaping. In this work, we propose META, a simple and fast ViT adapter that can improve the model's memory efficiency and decrease memory time consumption by reducing the inefficient memory access operations. Our method features a memory-efficient adapter block that enables the common sharing of layer normalization between the self-attention and feed-forward network layers, thereby reducing the model's reliance on normalization operations. Within the proposed block, the cross-shaped self-attention is employed to reduce the model's frequent reshaping operations. Moreover, we augment the adapter block with a lightweight convolutional branch that can enhance local inductive biases, particularly beneficial for the dense prediction tasks, e.g., object detection, instance segmentation, and semantic segmentation. The adapter block is finally formulated in a cascaded manner to compute diverse head features, thereby enriching the variety of feature representations. Empirically, extensive evaluations on multiple representative datasets validate that META substantially enhances the predicted quality, while achieving a new state-of-the-art accuracy-efficiency trade-off. Theoretically, we demonstrate that META exhibits superior generalization capability and stronger adaptability. | Vision Transformer, Vision Transformer, Transformer | In this paper, we propose META, a straightforward and high-speed ViT adapter that enhances the model's memory efficiency and reduces memory access time by minimizing inefficient memory access operations. | 2,289 | 2502.01962 | [
-0.001595266512595117,
-0.038114938884973526,
0.0033944302704185247,
0.05153767392039299,
0.10570641607046127,
0.005917207803577185,
-0.060510024428367615,
0.016248567029833794,
-0.021699801087379456,
-0.06458190828561783,
-0.03092484548687935,
-0.08435952663421631,
-0.025807103142142296,
0.04964752122759819,
-0.016577960923314095,
0.008079075254499912,
0.05547766387462616,
0.14104676246643066,
-0.10749439895153046,
-0.0714213103055954,
0.04089469462633133,
0.02338826097548008,
0.007482754997909069,
0.029419586062431335,
0.0604899600148201,
0.02686639502644539,
-0.0006676081102341413,
-0.08884871751070023,
-0.0027274456806480885,
-0.059770599007606506,
0.03395938500761986,
-0.02430654875934124,
-0.06604330986738205,
0.10170966386795044,
-0.08116313815116882,
-0.02341378852725029,
-0.061577849090099335,
0.039659999310970306,
-0.026470746845006943,
-0.06561393290758133,
0.04369305074214935,
-0.0025836252607405186,
-0.027138877660036087,
0.015247330069541931,
0.05604604631662369,
0.1019781231880188,
0.0131139000877738,
-0.05655387416481972,
0.08279658108949661,
-0.035682179033756256,
-0.007361536845564842,
-0.004911364056169987,
-0.052977874875068665,
0.07317014783620834,
-0.025758646428585052,
0.06284978240728378,
0.06509564816951752,
-0.020473528653383255,
-0.0030995735432952642,
0.03297942876815796,
-0.01685093343257904,
-0.045109815895557404,
0.0752597525715828,
-0.01046381052583456,
-0.08928357809782028,
-0.04772615432739258,
0.0611497201025486,
0.002572922268882394,
-0.039837323129177094,
-0.04332125186920166,
-0.030845435336232185,
0.11102540791034698,
-0.019427670165896416,
-0.0322372131049633,
0.09093916416168213,
0.027712367475032806,
0.04094095155596733,
-0.031521402299404144,
0.04854116216301918,
-0.05728178471326828,
-0.002261039102450013,
0.03272240608930588,
0.07304145395755768,
-0.01830470934510231,
0.13693499565124512,
0.025954261422157288,
-0.0551321879029274,
0.05710070952773094,
-0.06103603169322014,
0.010373485274612904,
-0.06684709340333939,
-0.06709963828325272,
0.006814202759414911,
-0.0637279599905014,
0.05271931737661362,
-0.004672188777476549,
-0.05116274580359459,
-0.09051250666379929,
-0.0773015096783638,
0.06029055267572403,
0.005903169047087431,
-0.0032550266478210688,
0.07183510065078735,
0.024530483409762383,
0.05104593560099602,
-0.004679875914007425,
0.05222371593117714,
0.058915771543979645,
0.007477228529751301,
0.00034584288368932903,
0.07198375463485718,
-0.00833873637020588,
0.006672527175396681,
-0.06206708028912544,
0.04520568624138832,
0.015565590001642704,
-0.031748007982969284,
-0.005213289521634579,
0.0249777864664793,
0.010698649100959301,
0.003931245766580105,
0.031630080193281174,
-0.03828268125653267,
0.0058594150468707085,
0.05895920842885971,
-0.010250910185277462,
-0.040596771985292435,
1.0496022119270822e-33,
0.0031437885481864214,
0.010366563685238361,
0.018333004787564278,
0.025423724204301834,
0.018289171159267426,
0.01104991789907217,
0.0076266927644610405,
0.0829041451215744,
-0.011236452497541904,
-0.018133345991373062,
-0.03378092870116234,
-0.031060606241226196,
-0.02360561676323414,
0.10302940756082535,
-0.005893811583518982,
-0.007816619239747524,
-0.04060708358883858,
0.0397031307220459,
-0.04492238909006119,
-0.03934057056903839,
0.007149512879550457,
-0.031091241165995598,
0.0035191280767321587,
-0.008698111400008202,
-0.03082079254090786,
-0.028563417494297028,
0.042930714786052704,
-0.010757846757769585,
-0.006366037763655186,
-0.015723131597042084,
-0.07591785490512848,
0.07313552498817444,
0.020349351689219475,
-0.031948696821928024,
0.05003746971487999,
-0.058865103870630264,
0.011612559668719769,
-0.05786063149571419,
0.042021531611680984,
0.011756699532270432,
0.05934406816959381,
0.0922723114490509,
-0.015922361984848976,
-0.0440361313521862,
-0.07416151463985443,
-0.03048725239932537,
0.021900156512856483,
0.08781255781650543,
-0.07956205308437347,
0.049501895904541016,
0.09342236071825027,
-0.011813506484031677,
-0.11547806113958359,
-0.06167195737361908,
0.006750970147550106,
0.006514453794807196,
0.08406352996826172,
0.11689795553684235,
0.10005751252174377,
0.04755382612347603,
-0.05523128807544708,
0.03449459746479988,
-0.041204698383808136,
0.1256483793258667,
-0.015426570549607277,
0.002070480491966009,
0.07465074956417084,
0.13647520542144775,
-0.04926273971796036,
0.026221973821520805,
-0.006091167218983173,
0.03760988637804985,
-0.04766366630792618,
-0.057858906686306,
0.0012310468591749668,
0.005664865951985121,
0.006846527103334665,
-0.06829466670751572,
-0.042321451008319855,
0.021930841729044914,
0.016973717138171196,
0.06643228977918625,
-0.03371919319033623,
0.006612234283238649,
-0.030429044738411903,
0.011232447810471058,
0.07639022916555405,
-0.0680614486336708,
0.01622915454208851,
-0.010107980109751225,
0.015917334705591202,
-0.03641045093536377,
0.037158120423555374,
-0.04223153367638588,
-0.06884735822677612,
3.2172761511452716e-34,
0.041041769087314606,
0.04591669142246246,
-0.07097429782152176,
0.05122683197259903,
-0.00983743742108345,
-0.08576897531747818,
0.02488737925887108,
0.024912705644965172,
-0.05864253640174866,
-0.09000080823898315,
-0.018532073125243187,
-0.03761434182524681,
0.04111357778310776,
-0.00210716319270432,
0.08007866144180298,
-0.022935615852475166,
-0.033592745661735535,
0.03188653290271759,
-0.10615786164999008,
-0.02020818553864956,
-0.04900623485445976,
0.09485170990228653,
-0.00682904152199626,
-0.008429760113358498,
-0.10629624873399734,
0.02837403304874897,
-0.10669226199388504,
0.04844430461525917,
0.044182442128658295,
-0.05040024593472481,
-0.042398300021886826,
0.03985028341412544,
0.0011752670397982001,
-0.08023517578840256,
-0.04209674522280693,
0.09293191879987717,
0.0006130149704404175,
-0.015506234019994736,
0.029619719833135605,
0.03737364336848259,
0.01883763074874878,
-0.016053268685936928,
-0.02642885595560074,
-0.0008740609046071768,
-0.08266669511795044,
-0.03523578494787216,
-0.036055099219083786,
-0.01495310291647911,
-0.04559732973575592,
0.07267394661903381,
-0.10287173837423325,
0.009006626904010773,
-0.11827957630157471,
0.028110140934586525,
-0.08063098788261414,
0.024633686989545822,
-0.03934045881032944,
0.011519405990839005,
0.10610610246658325,
-0.057414885610342026,
0.00893288291990757,
-0.03179463744163513,
-0.013621130958199501,
-0.06841316819190979,
-0.003665725002065301,
-0.03578674793243408,
-0.028367575258016586,
-0.0869607925415039,
0.05757629871368408,
0.06400695443153381,
0.023294230923056602,
0.01257103681564331,
0.02431759610772133,
-0.014515232294797897,
-0.035468872636556625,
-0.06926325708627701,
0.08590402454137802,
-0.023755662143230438,
0.032019272446632385,
0.044271670281887054,
-0.041241392493247986,
-0.01501494087278843,
0.017412278801202774,
0.005275193601846695,
0.04564220458269119,
0.03626378998160362,
0.025148510932922363,
0.02811211533844471,
0.031039966270327568,
-0.019907761365175247,
-0.10106959193944931,
0.019404951483011246,
-0.017860649153590202,
0.09362029284238815,
-0.023912571370601654,
-4.064180814111751e-8,
-0.06918434053659439,
0.059721704572439194,
0.004911160562187433,
0.0349622406065464,
-0.013286661356687546,
-0.10166534781455994,
0.012801408767700195,
0.12316416203975677,
-0.013782280497252941,
0.023845497518777847,
0.0785728245973587,
-0.06291418522596359,
-0.05554472282528877,
-0.004313815385103226,
0.06278090924024582,
0.08923137933015823,
0.04915991052985191,
0.04736529290676117,
-0.00015788889140821993,
-0.061382897198200226,
-0.05086331069469452,
0.03536541759967804,
0.03562932834029198,
-0.04253983870148659,
0.029129069298505783,
-0.028799161314964294,
-0.014814712107181549,
0.10800018906593323,
0.0415513701736927,
-0.03579859435558319,
-0.005650739651173353,
0.06644074618816376,
-0.00556078739464283,
-0.05641317367553711,
0.03995168209075928,
0.04909704625606537,
0.021035758778452873,
0.0966053307056427,
0.024708816781640053,
-0.00383857567794621,
0.04313115030527115,
0.009856749325990677,
-0.013557557947933674,
0.0590113140642643,
-0.05621686950325966,
0.003617687616497278,
0.03487364947795868,
-0.03724273666739464,
-0.012064383365213871,
-0.037926074117422104,
0.05813406780362129,
0.007819456048309803,
0.028067823499441147,
0.09296127408742905,
-0.01338485162705183,
0.0058912429958581924,
0.06927205622196198,
-0.05135311931371689,
-0.006415456533432007,
0.06009896099567413,
-0.022935837507247925,
0.02826933190226555,
-0.050650786608457565,
0.002728014951571822
] |
Policy Optimization under Imperfect Human Interactions with Agent-Gated Shared Autonomy | https://openreview.net/forum?id=LfekK1E0QE | [
"Zhenghai Xue",
"Bo An",
"Shuicheng YAN"
] | Poster | We introduce AGSA, an Agent-Gated Shared Autonomy framework that learns from high-level human feedback to tackle the challenges of reward-free training, safe exploration, and imperfect low-level human control. Recent human-in-the loop learning methods enable human participants to intervene a learning agent’s control and provide online demonstrations. Nonetheless, these methods rely heavily on perfect human interactions, including accurate human-monitored intervention decisions and near-optimal human demonstrations. AGSA employs a dedicated gating agent to determine when to switch control, thereby reducing the need of constant human monitoring. To obtain a precise and foreseeable gating agent, AGSA trains a long-term gating value function from human evaluative feedback on the gating agent’s intervention requests and preference feedback on pairs of human intervention trajectories. Instead of relying on potentially suboptimal human demonstrations, the learning agent is trained using control-switching signals from the gating agent. We provide theoretical insights on performance bounds that respectively describe the ability of the two agents. Experiments are conducted with both simulated and real human participants at different skill levels in challenging continuous control environments. Comparative results highlight that AGSA achieves significant improvements over previous human-in-the-loop learning methods in terms of training safety, policy performance, and user-friendliness. | Reinforcement Learning, Human-in-the-loop Learning, Imperfect Human Interaction, Human Feedback | A new agent-gated shared autonomy framework that safely and efficiently trains RL policies from imperfect human interactions. | 2,283 | null | [
-0.03717004880309105,
-0.08946556597948074,
0.003348294412717223,
-0.010582772083580494,
0.022887196391820908,
-0.0405573844909668,
0.055431198328733444,
0.014832916669547558,
0.02164021134376526,
0.0036490161437541246,
0.008257061243057251,
-0.014648468233644962,
-0.01856275461614132,
0.016719549894332886,
0.008954852819442749,
-0.008580601774156094,
0.1290639489889145,
0.034832343459129333,
-0.10736079514026642,
-0.0032747096847742796,
0.009117623791098595,
-0.05285799503326416,
0.09448407590389252,
-0.027848243713378906,
-0.06448973715305328,
-0.04742932319641113,
-0.028442909941077232,
-0.0578722320497036,
0.04283332824707031,
-0.019049638882279396,
0.039158593863248825,
-0.11705298721790314,
0.05115178972482681,
-0.0639117956161499,
-0.0843314453959465,
0.029296308755874634,
-0.07295340299606323,
-0.03147660568356514,
-0.0018511939560994506,
-0.05087032914161682,
-0.026530291885137558,
-0.04493763670325279,
0.008646983653306961,
-0.03381665050983429,
0.008909592404961586,
0.1005876287817955,
-0.01841677539050579,
-0.08617199212312698,
0.008123883046209812,
0.005861691199243069,
0.0056412843987345695,
-0.010197232477366924,
0.05484993755817413,
-0.07809751480817795,
0.08640045672655106,
0.05540825054049492,
0.016527846455574036,
0.0012773280031979084,
0.04068855941295624,
-0.01632806472480297,
-0.013577016070485115,
-0.0826946273446083,
-0.06137537956237793,
-0.001961811212822795,
-0.07589464634656906,
-0.007500333245843649,
0.01650487259030342,
-0.01446225680410862,
0.01763194426894188,
-0.004907683003693819,
-0.011418290436267853,
-0.0325651578605175,
0.011040354147553444,
-0.02194034494459629,
0.06923192739486694,
0.010322350077331066,
-0.04917514696717262,
-0.02310849539935589,
0.038648318499326706,
-0.043034836649894714,
0.034684088081121445,
-0.021889489144086838,
0.022045673802495003,
0.05107082799077034,
0.028682641685009003,
-0.008828517980873585,
0.023594431579113007,
0.12111200392246246,
0.10444853454828262,
0.013607429340481758,
-0.04808894917368889,
0.02985163778066635,
-0.025974420830607414,
-0.03413033112883568,
-0.030805131420493126,
0.011174802668392658,
-0.05453409254550934,
-0.09099353104829788,
-0.06292333453893661,
0.08065931499004364,
-0.01871856302022934,
0.0432412289083004,
-0.0023916857317090034,
0.04461424797773361,
-0.014076627790927887,
-0.027352940291166306,
0.0572429858148098,
0.03884408250451088,
0.0611317902803421,
-0.010258818976581097,
-0.004854193888604641,
-0.04294756054878235,
0.022221386432647705,
0.06936196982860565,
0.002870900323614478,
0.012259194627404213,
-0.030492449179291725,
0.006849679630249739,
-0.00003158973413519561,
0.010879242792725563,
0.051613740622997284,
-0.08620484918355942,
0.06041650101542473,
0.02084490843117237,
0.07916285842657089,
-0.012238727882504463,
-0.033827342092990875,
1.1172947071581145e-33,
0.06505494564771652,
0.012017453089356422,
0.0833931639790535,
0.012501707300543785,
-0.00015460829308722168,
0.04937669634819031,
0.017975565046072006,
-0.04182172566652298,
-0.02247929759323597,
0.027107177302241325,
0.05374101549386978,
-0.022304663434624672,
-0.07674932479858398,
0.056596290320158005,
0.020849181339144707,
0.018574248999357224,
-0.02297171764075756,
-0.02506987936794758,
-0.03108610026538372,
-0.08807028084993362,
0.12238068878650665,
-0.08171279728412628,
0.025316154584288597,
-0.04733504727482796,
0.05866188183426857,
-0.03731514886021614,
-0.006999506149441004,
0.022119887173175812,
0.03949490189552307,
-0.011903430335223675,
-0.03614846616983414,
0.07115058600902557,
-0.04852776974439621,
0.0833863839507103,
-0.034480899572372437,
-0.0039693196304142475,
-0.0005730846896767616,
-0.06150626763701439,
-0.007282866630703211,
0.08823911100625992,
-0.016201211139559746,
-0.0010787276551127434,
0.10651808232069016,
-0.008104333654046059,
-0.06134796515107155,
-0.07674208283424377,
0.08011031895875931,
0.03489936143159866,
-0.07308536767959595,
0.018062308430671692,
-0.05777059495449066,
0.05644078925251961,
0.008241478353738785,
-0.05341436713933945,
0.020985016599297523,
0.07047777622938156,
0.04040704295039177,
0.05357533320784569,
-0.06853596866130829,
0.025587957352399826,
0.02143925242125988,
0.04242931306362152,
0.02873152494430542,
0.06461159139871597,
-0.024029115214943886,
0.10376712679862976,
-0.058207083493471146,
0.07257818430662155,
0.03497283533215523,
0.026975145563483238,
-0.04056068882346153,
0.05193179473280907,
-0.005881804041564465,
-0.0710189938545227,
0.05914495140314102,
-0.05945684760808945,
0.0386931374669075,
-0.06539854407310486,
-0.002235308289527893,
-0.029735997319221497,
0.042107731103897095,
-0.004954006522893906,
-0.03425896167755127,
-0.00394565612077713,
-0.03727758303284645,
-0.09669242054224014,
-0.06190148741006851,
-0.03268799930810928,
-0.05095662549138069,
0.026865657418966293,
-0.07899430394172668,
0.005534923169761896,
-0.05903392285108566,
0.08764708042144775,
-0.007627442013472319,
-1.9446851404255978e-33,
0.04381275922060013,
-0.07326482236385345,
-0.009397954680025578,
0.037533074617385864,
0.03561476618051529,
-0.0011255714343860745,
0.03512359410524368,
-0.0635673999786377,
0.0072129773907363415,
-0.027792854234576225,
-0.08949419856071472,
0.05722088739275932,
0.06084540858864784,
0.03328331559896469,
-0.015982089564204216,
-0.05277535691857338,
-0.027429090812802315,
0.00664018327370286,
0.07828240841627121,
-0.009153700433671474,
-0.010696029290556908,
0.08798787742853165,
-0.0540478341281414,
0.021395668387413025,
0.002911651972681284,
0.009278792887926102,
0.022472653537988663,
0.114844910800457,
0.042894937098026276,
0.0446518175303936,
-0.03022860735654831,
-0.010960584506392479,
-0.05784159153699875,
-0.02266933210194111,
-0.05096465349197388,
0.08503521978855133,
0.08515603840351105,
0.044138193130493164,
-0.09053188562393188,
0.07847988605499268,
-0.010253948159515858,
-0.059828586876392365,
-0.006838569883257151,
0.009456299245357513,
0.023764120414853096,
0.014634737744927406,
-0.04958457127213478,
0.00893871858716011,
-0.08525978028774261,
-0.012502733618021011,
-0.04751230776309967,
0.020162401720881462,
-0.09498250484466553,
-0.026020856574177742,
-0.09046809375286102,
-0.029362380504608154,
0.060773998498916626,
-0.024635424837470055,
0.049035586416721344,
-0.050772834569215775,
-0.03691773861646652,
0.03262653574347496,
-0.042463961988687515,
0.0826490968465805,
-0.018355241045355797,
0.04885245859622955,
0.0027783983387053013,
0.07350082695484161,
0.05473402142524719,
0.010971193201839924,
0.01875370740890503,
-0.04274553060531616,
0.0013516810722649097,
-0.09005247056484222,
0.0783407986164093,
0.03849875181913376,
-0.002841461915522814,
-0.04799436032772064,
0.0038970138411968946,
-0.06997179239988327,
0.03791380301117897,
0.030190901830792427,
0.02757136896252632,
0.03390996530652046,
-0.03355863317847252,
0.10813334584236145,
-0.0701303705573082,
0.14393198490142822,
-0.03433341905474663,
-0.02414204739034176,
-0.007635235786437988,
0.024310225620865822,
0.04554180055856705,
0.023954693228006363,
-0.08153875172138214,
-4.078165360965613e-8,
-0.1021721139550209,
0.058372244238853455,
0.08329610526561737,
0.013751671649515629,
0.04944824054837227,
0.0024723305832594633,
0.031136319041252136,
-0.06031019613146782,
-0.0384281724691391,
-0.011332792229950428,
-0.00835964921861887,
0.057982221245765686,
0.074273020029068,
-0.030201613903045654,
-0.002304776106029749,
0.015755483880639076,
-0.027743089944124222,
0.01170526072382927,
0.009361385367810726,
0.02829347550868988,
0.03251201659440994,
-0.027289923280477524,
-0.00807071104645729,
-0.07516535371541977,
-0.04495873674750328,
-0.04641388729214668,
-0.03021278604865074,
0.07086227834224701,
-0.005304573103785515,
0.04033607989549637,
0.035231102257966995,
-0.027862828224897385,
0.06764461100101471,
-0.00798860751092434,
0.03623747080564499,
0.06324883550405502,
-0.02921285293996334,
-0.07764779776334763,
0.014128739945590496,
-0.05540063604712486,
-0.0037947427481412888,
0.08868929743766785,
-0.009458553977310658,
-0.017812857404351234,
-0.04159000888466835,
0.021577028557658195,
-0.05384108051657677,
-0.11887777596712112,
0.07926320284605026,
-0.028769787400960922,
0.03451637551188469,
-0.05835842713713646,
-0.06581664830446243,
0.0984700545668602,
0.1400085687637329,
-0.015425358898937702,
0.029401518404483795,
-0.08451678603887558,
0.06877153366804123,
0.14224962890148163,
-0.0030158443842083216,
0.007957417517900467,
-0.1171257272362709,
-0.028105691075325012
] |
Multiview Equivariance Improves 3D Correspondence Understanding with Minimal Feature Finetuning | https://openreview.net/forum?id=CNO4rbSV6v | [
"Yang You",
"Yixin Li",
"Congyue Deng",
"Yue Wang",
"Leonidas Guibas"
] | Poster | Vision foundation models, particularly the ViT family, have revolutionized image understanding by providing rich semantic features. However, despite their success in 2D comprehension, their abilities on grasping 3D spatial relationships are still unclear.
In this work, we evaluate and enhance the 3D awareness of ViT-based models. We begin by systematically assessing their ability to learn 3D equivariant features, specifically examining the consistency of semantic embeddings across different viewpoints. Our findings indicate that improved 3D equivariance leads to better performance on various downstream tasks, including pose estimation, tracking, and semantic transfer. Building on this insight, we propose a simple yet effective finetuning strategy based on 3D correspondences, which significantly enhances the 3D understanding of existing vision models. Remarkably, even finetuning on a single object for just one iteration results in substantial performance gains. Code is available on https://github.com/qq456cvb/3DCorrEnhance. | Vision Foundation Models; 3D Representation Learning; Fine-tuning; 3D Equivariance | This work evaluates and improves the 3D awareness of ViT-based models by enhancing their 3D equivariance through a simple finetuning strategy using 3D correspondences. | 2,280 | 2411.19458 | [
0.023553289473056793,
-0.1417255401611328,
0.03450627252459526,
-0.0612151101231575,
0.05333869159221649,
-0.020448103547096252,
0.012028125114738941,
-0.01879013516008854,
-0.007450425531715155,
-0.006593081168830395,
-0.004564825911074877,
-0.08957074582576752,
-0.003046720754355192,
0.09648904204368591,
-0.016766488552093506,
-0.015387308783829212,
0.09608685970306396,
0.14741554856300354,
-0.08925112336874008,
0.06833086907863617,
-0.016079796478152275,
0.006596554070711136,
0.08486687391996384,
-0.014942944049835205,
0.06864644587039948,
0.017688671126961708,
0.030404919758439064,
-0.04847794398665428,
0.02005806379020214,
-0.07845578342676163,
-0.033877186477184296,
0.020651422441005707,
-0.08171942085027695,
0.02470904402434826,
-0.05548524111509323,
0.0463925376534462,
0.00641263835132122,
0.015407687053084373,
-0.06999224424362183,
-0.06846754252910614,
-0.03558987379074097,
0.08081857115030289,
-0.01432302501052618,
0.025873268023133278,
0.05991477146744728,
0.0923398807644844,
0.05094609037041664,
-0.046562716364860535,
0.05744069442152977,
0.010667839087545872,
-0.05170382559299469,
-0.06519512832164764,
-0.05880368500947952,
-0.005151348654180765,
-0.025866512209177017,
0.0446646474301815,
0.03580409288406372,
-0.08304446935653687,
0.010546581819653511,
-0.009362811222672462,
0.04971594363451004,
0.004812261089682579,
0.03316957876086235,
0.06073412671685219,
-0.012737991288304329,
-0.0894305482506752,
-0.01082019042223692,
-0.0001854889269452542,
-0.04242783784866333,
-0.0278768390417099,
-0.04705701395869255,
0.033275995403528214,
0.011135151609778404,
-0.052932482212781906,
0.05375703051686287,
0.047731876373291016,
0.019183943048119545,
-0.08116097748279572,
-0.014681782573461533,
-0.045651305466890335,
0.0017863658722490072,
0.06189658120274544,
0.03286254405975342,
0.04850836843252182,
0.07024654746055603,
0.02213258482515812,
-0.0841732770204544,
0.048098742961883545,
0.0012389294570311904,
-0.01663193665444851,
0.01336287148296833,
-0.07011280953884125,
0.0008571564103476703,
-0.027118761092424393,
0.03315737843513489,
-0.040862683206796646,
0.020756226032972336,
-0.11584693193435669,
-0.04833091050386429,
0.029865548014640808,
-0.010471610352396965,
0.06515128910541534,
0.032694023102521896,
0.00169934646692127,
0.03851766884326935,
0.010312718339264393,
0.05056958645582199,
-0.0404893159866333,
-0.021443670615553856,
-0.06629350036382675,
0.02928921952843666,
-0.0594838485121727,
-0.022981347516179085,
0.027238884940743446,
-0.028695184737443924,
0.06282807886600494,
-0.05795948952436447,
0.02471376582980156,
0.05315687879920006,
-0.019870372489094734,
-0.06623996794223785,
-0.00295029254630208,
0.027735939249396324,
0.0006518354639410973,
0.0039267512038350105,
0.012030583806335926,
0.010922899469733238,
1.6292472925251833e-33,
0.048409171402454376,
0.07914285361766815,
0.040108658373355865,
0.06232789531350136,
0.039379119873046875,
0.03119775280356407,
0.009304528124630451,
0.07395189255475998,
-0.0794844701886177,
0.03455903381109238,
0.006786684971302748,
-0.0727967917919159,
0.0061154356226325035,
0.13539905846118927,
-0.04444967582821846,
-0.06468331068754196,
-0.08733213692903519,
0.027717525139451027,
-0.07986762374639511,
0.13418829441070557,
0.04532748460769653,
-0.024426160380244255,
-0.03713645786046982,
-0.013707909733057022,
-0.018372559919953346,
0.010709264315664768,
-0.008761436678469181,
0.04860354959964752,
-0.020747710019350052,
-0.00003542801641742699,
-0.06265722960233688,
0.06501366943120956,
-0.04142065718770027,
0.015800893306732178,
-0.008102691732347012,
-0.055632319301366806,
0.0287189744412899,
-0.05000431090593338,
0.0362028032541275,
-0.036059021949768066,
0.025315377861261368,
0.05924298241734505,
-0.057552460581064224,
-0.07447319477796555,
-0.059177301824092865,
0.00656700786203146,
0.026074381545186043,
0.06941687315702438,
-0.08035030215978622,
0.005817691329866648,
0.05696120858192444,
0.00926277507096529,
-0.08558844029903412,
-0.06430227309465408,
-0.12124520540237427,
-0.0033593776170164347,
-0.011612705886363983,
0.03503935784101486,
0.021154623478651047,
-0.0390176922082901,
-0.03927356004714966,
0.026462508365511894,
-0.052891165018081665,
0.04279148578643799,
-0.00048662995686754584,
0.023614665493369102,
0.04859161376953125,
0.059826191514730453,
0.061530884355306625,
0.0858924388885498,
-0.05383671820163727,
-0.007210013456642628,
-0.01643127202987671,
-0.09250114113092422,
0.03236028552055359,
-0.04693357273936272,
-0.056351516395807266,
-0.0696663185954094,
0.07667946070432663,
0.07336528599262238,
-0.026179974898695946,
0.04399847984313965,
0.04826159402728081,
-0.02499687299132347,
-0.04538736119866371,
-0.08461432158946991,
0.02562819980084896,
-0.10612209141254425,
0.07211482524871826,
0.018914010375738144,
-0.011004281230270863,
-0.05726487562060356,
-0.08909443765878677,
-0.03543834760785103,
-0.007769865449517965,
-1.7960886954862446e-33,
0.06299462169408798,
-0.007415724918246269,
0.0306276623159647,
0.05619626119732857,
-0.026459312066435814,
-0.10443682223558426,
-0.009462371468544006,
0.017817165702581406,
0.0009551165858283639,
-0.07253293693065643,
-0.00510740838944912,
0.01268607098609209,
-0.04323700815439224,
-0.006049812771379948,
0.058503251522779465,
0.0005274810828268528,
-0.04646224528551102,
0.03802774101495743,
0.007028879597783089,
-0.005451765842735767,
0.05838789418339729,
-0.040148064494132996,
-0.02208651788532734,
-0.021974151954054832,
-0.09280619770288467,
0.11661433428525925,
-0.040284570306539536,
-0.05673935264348984,
-0.05800490826368332,
-0.0007575644413009286,
0.00429693516343832,
-0.05064011365175247,
0.0063357180915772915,
-0.0006739943637512624,
-0.04902731999754906,
0.03015025332570076,
-0.04066582769155502,
-0.006557896267622709,
-0.013191644102334976,
0.03300206735730171,
0.029315952211618423,
-0.019884655252099037,
-0.01044060755521059,
0.03696959838271141,
-0.033054549247026443,
0.016846461221575737,
-0.007596705108880997,
0.04061660170555115,
0.02253137342631817,
0.06829264760017395,
-0.06778080016374588,
0.06004327908158302,
-0.0637173131108284,
-0.001075399573892355,
0.0026765717193484306,
-0.02029549330472946,
-0.034906916320323944,
-0.08014606684446335,
0.07844655960798264,
-0.030674561858177185,
-0.009661594405770302,
0.009136350825428963,
-0.08207673579454422,
0.06956267356872559,
0.01999085396528244,
0.010162930004298687,
-0.04082963243126869,
-0.026935912668704987,
0.0934421718120575,
0.025658275932073593,
-0.014918623492121696,
-0.0439644381403923,
0.01894645392894745,
-0.10093723237514496,
0.029023462906479836,
-0.08227140456438065,
0.01859411969780922,
-0.022703617811203003,
0.05246753245592117,
-0.03516436740756035,
0.048329610377550125,
-0.0857551246881485,
0.051764875650405884,
0.15390513837337494,
0.07665081322193146,
0.0796859934926033,
-0.0019212211482226849,
-0.0027452006470412016,
0.017448127269744873,
0.03286658972501755,
-0.008384307846426964,
0.04738270118832588,
0.019952310249209404,
0.05671500787138939,
-0.0018200369086116552,
-3.712623453111519e-8,
-0.09731867164373398,
0.08985857665538788,
0.015607224777340889,
0.010948927141726017,
-0.06344043463468552,
-0.043994974344968796,
-0.004957254976034164,
0.15994644165039062,
-0.05937551334500313,
0.029566116631031036,
-0.0018111130921170115,
-0.06831765919923782,
-0.02300918847322464,
0.021901067346334457,
0.07101152092218399,
0.0584985576570034,
0.027147769927978516,
0.06529287993907928,
-0.013026162981987,
0.03441518545150757,
0.010304559022188187,
0.040632959455251694,
0.07532446831464767,
-0.013669145293533802,
-0.003707160707563162,
-0.005719441920518875,
-0.0727100670337677,
0.05927227810025215,
0.09671910852193832,
0.008000850677490234,
-0.004619680345058441,
0.01787659525871277,
-0.010539446957409382,
-0.031726498156785965,
-0.010927506722509861,
0.0008839059737510979,
-0.029652101919054985,
-0.049980245530605316,
0.03873453289270401,
-0.0649271085858345,
0.012530229985713959,
0.1145576611161232,
0.03444744274020195,
0.03609484061598778,
0.02690621092915535,
0.0621902234852314,
0.028798533603549004,
-0.06388197094202042,
-0.010530687868595123,
-0.04373679682612419,
-0.05397597700357437,
0.02629796601831913,
0.007157715503126383,
0.1334458291530609,
0.03905687853693962,
0.045171208679676056,
-0.012233532033860683,
-0.04936547577381134,
0.0450221449136734,
0.06904702633619308,
-0.03303716331720352,
-0.009648374281823635,
-0.05642680078744888,
-0.04178478941321373
] |
Scale-aware Recognition in Satellite Images under Resource Constraints | https://openreview.net/forum?id=QIxFo9mFwR | [
"Shreelekha Revankar",
"Cheng Perng Phoo",
"Utkarsh Mall",
"Bharath Hariharan",
"Kavita Bala"
] | Poster | Recognition of features in satellite imagery (forests, swimming pools, etc.) depends strongly on the spatial scale of the concept and therefore the resolution of the images. This poses two challenges:
Which resolution is best suited for recognizing a given concept, and where and when should the costlier higher-resolution (HR) imagery be acquired?
We present a novel scheme to address these challenges by introducing three components: (1) A technique to distill knowledge from models trained on HR imagery to recognition models that operate on imagery of lower resolution (LR), (2) a sampling strategy for HR imagery based on model disagreement, and (3) an LLM-based approach for inferring concept "scale". With these components we present a system to efficiently perform scale-aware recognition in satellite imagery, improving accuracy over single-scale inference while following budget constraints. **Our novel approach offers up to a 26.3\% improvement over entirely HR baselines, using 76.3 \% fewer HR images.** Resources are available at https://www.cs.cornell.edu/~revankar/scale_aware. | Satellite Imagery, Resolution | A system for scale-aware recognition in satellite imagery under resource constraints | 2,275 | 2411.00210 | [
0.044915344566106796,
-0.03806193172931671,
0.03761157765984535,
0.0716281533241272,
0.09999571740627289,
-0.035601649433374405,
-0.002016685204580426,
-0.039625056087970734,
0.022564474493265152,
0.03892611339688301,
-0.07409856468439102,
-0.03921518474817276,
0.05809730291366577,
0.10255105048418045,
0.008391862735152245,
-0.0009316516225226223,
0.027259236201643944,
0.11969371885061264,
-0.053675759583711624,
-0.035212717950344086,
-0.008039981126785278,
-0.012835964560508728,
0.025311553850769997,
-0.032119713723659515,
0.05192083492875099,
-0.04731426760554314,
0.00955750048160553,
0.00636587617918849,
-0.031990282237529755,
-0.016159072518348694,
0.03829196095466614,
0.08408360183238983,
0.036235492676496506,
0.002163661178201437,
-0.0013461826601997018,
0.0641104206442833,
-0.02589460276067257,
-0.03735123574733734,
-0.021997004747390747,
0.017245344817638397,
0.03770247846841812,
-0.03653234988451004,
0.003018296090885997,
-0.04267755150794983,
0.019217517226934433,
0.09612464904785156,
-0.02021816000342369,
0.0068745082244277,
-0.005775679834187031,
-0.0010372195392847061,
-0.05004385858774185,
-0.08836894482374191,
-0.1233452707529068,
0.018294574692845345,
-0.09262209385633469,
-0.05204959586262703,
0.04057082533836365,
-0.03445972874760628,
-0.003779294900596142,
-0.0021559924352914095,
-0.0433778315782547,
-0.08601333200931549,
-0.06412778049707413,
0.03455745056271553,
0.006678534671664238,
-0.018407711759209633,
-0.022373104467988014,
-0.03475723788142204,
0.0376637727022171,
-0.038906510919332504,
0.00738936010748148,
0.06524886190891266,
0.010051055811345577,
0.010027086362242699,
-0.020760703831911087,
-0.05902574956417084,
-0.010027768090367317,
0.05417794734239578,
0.07208048552274704,
-0.0630144402384758,
-0.047873854637145996,
0.02768147736787796,
-0.030337665230035782,
0.03893881291151047,
0.1474827229976654,
-0.006374118383973837,
0.003217042190954089,
0.06619051098823547,
0.020621541887521744,
-0.05149727314710617,
-0.002845525508746505,
-0.04220208153128624,
-0.003756157821044326,
0.005730756092816591,
0.02183341607451439,
0.013724802061915398,
0.03250123932957649,
-0.07278550416231155,
-0.05266505852341652,
0.05316513404250145,
-0.015782557427883148,
0.04636310786008835,
0.030993064865469933,
-0.034081410616636276,
0.057466089725494385,
0.01345509197562933,
0.05900413915514946,
0.07390844076871872,
0.0996350422501564,
-0.09992676973342896,
0.01575232669711113,
0.018369002267718315,
0.013035130687057972,
0.05283843353390694,
-0.00928592775017023,
-0.029208138585090637,
0.05132708325982094,
-0.028967401012778282,
-0.004555652849376202,
-0.01070061232894659,
-0.0177073385566473,
-0.01918003335595131,
0.03575900197029114,
-0.05876952037215233,
0.06616269797086716,
0.026177626103162766,
-0.09946320205926895,
-2.643322667257058e-34,
0.02756197564303875,
0.07521013170480728,
0.07734283804893494,
-0.010286648757755756,
0.03771361708641052,
-0.008419683203101158,
-0.06719420105218887,
0.026454558596014977,
-0.0016889682738110423,
0.08813846856355667,
-0.006460360251367092,
0.01273900456726551,
-0.01673305779695511,
0.07989788800477982,
0.0508023202419281,
0.021490847691893578,
-0.014017018489539623,
0.08821513503789902,
0.023433445021510124,
0.0165873933583498,
0.03670576587319374,
-0.04859071224927902,
0.01911843754351139,
-0.02677169255912304,
-0.029651856049895287,
-0.04893713444471359,
0.0164901465177536,
-0.0007212276686914265,
-0.029870258644223213,
-0.026756033301353455,
0.03987029567360878,
-0.033780138939619064,
0.018074557185173035,
0.07397781312465668,
-0.0018428955227136612,
-0.09896146506071091,
-0.024188144132494926,
0.00978810153901577,
-0.0012448475463315845,
0.007892432622611523,
-0.006464520003646612,
0.03636888042092323,
-0.01496855914592743,
-0.05712851136922836,
-0.0076623521745204926,
-0.007290302775800228,
0.0977846309542656,
-0.04715333133935928,
-0.0023597534745931625,
0.013715704903006554,
0.06185193732380867,
0.030040403828024864,
-0.05156369507312775,
-0.10507920384407043,
-0.03340163081884384,
-0.011950149200856686,
0.029981739819049835,
0.025320352986454964,
0.0011350405402481556,
0.004142351448535919,
0.009281052276492119,
-0.11026273667812347,
-0.05168486759066582,
0.06735250353813171,
0.05259725823998451,
-0.02530425600707531,
0.04675429314374924,
0.06846526265144348,
-0.04053271934390068,
0.009216181002557278,
0.027688225731253624,
0.041374459862709045,
0.032521456480026245,
-0.05944502726197243,
-0.004297985695302486,
-0.030478758737444878,
0.06301961839199066,
-0.02495625428855419,
-0.017708152532577515,
0.058485545217990875,
-0.013510061427950859,
0.0067062522284686565,
0.06846924871206284,
-0.011226669885218143,
-0.06170918792486191,
-0.0036041876301169395,
0.023304900154471397,
-0.07622071355581284,
-0.020542502403259277,
-0.009120245464146137,
-0.07888781279325485,
-0.01633918285369873,
-0.08019375056028366,
-0.04355984181165695,
0.026707936078310013,
-2.0913045508036617e-34,
0.02362116239964962,
-0.0028965515084564686,
0.06196262314915657,
0.006147739011794329,
-0.02753700502216816,
-0.012530808337032795,
-0.013269150629639626,
0.0599977932870388,
-0.06398723274469376,
-0.0956692174077034,
-0.040648505091667175,
0.018817007541656494,
-0.04463207349181175,
-0.1182311549782753,
-0.04548222944140434,
-0.03391837701201439,
0.014027866534888744,
-0.048194173723459244,
-0.020024854689836502,
0.05464937537908554,
0.03164459764957428,
0.06624181568622589,
-0.023247385397553444,
0.019325481727719307,
-0.04575125128030777,
0.044006846845149994,
-0.0037029508966952562,
0.024614613503217697,
-0.03606943413615227,
-0.014006558805704117,
0.00721095222979784,
-0.09112317860126495,
0.044278208166360855,
-0.03068252094089985,
-0.058399006724357605,
0.015688002109527588,
0.022270483896136284,
-0.053856413811445236,
-0.026673752814531326,
0.12032879889011383,
0.043948277831077576,
-0.0638582855463028,
-0.05224063619971275,
-0.09742644429206848,
0.0039552729576826096,
-0.1540844589471817,
-0.037221938371658325,
0.04367350786924362,
-0.013740705326199532,
-0.023601729422807693,
-0.024145327508449554,
-0.041544005274772644,
-0.07387391477823257,
0.03465303033590317,
0.03815559297800064,
0.009491348639130592,
-0.043513305485248566,
0.018272796645760536,
0.09052407741546631,
-0.04195589944720268,
-0.004093298688530922,
-0.03982570022344589,
-0.0880701094865799,
0.021441156044602394,
-0.059710726141929626,
-0.013511091470718384,
-0.02287456765770912,
0.004877457395195961,
-0.01968982070684433,
0.10038445889949799,
-0.03879174217581749,
-0.08832567930221558,
-0.008449564687907696,
-0.0427069254219532,
-0.09724429994821548,
-0.03188111260533333,
0.0453760139644146,
-0.022009486332535744,
0.053361501544713974,
-0.05856362730264664,
0.012525605037808418,
-0.07222595065832138,
-0.004210768733173609,
0.10438927263021469,
0.11731166392564774,
0.12661302089691162,
0.00927935540676117,
-0.03249603137373924,
0.06383317708969116,
0.064051553606987,
-0.08201008290052414,
0.029537387192249298,
0.010098104365170002,
0.07842891663312912,
-0.02467084489762783,
-4.3094896540196714e-8,
-0.11176032572984695,
0.058923449367284775,
-0.02292809821665287,
0.01163027435541153,
-0.001824844628572464,
-0.04090510681271553,
-0.002203253796324134,
0.0935605838894844,
-0.02545085735619068,
0.03667184337973595,
0.06070363521575928,
-0.0945574939250946,
-0.01747291348874569,
0.05388819798827171,
0.08635465055704117,
-0.025350799784064293,
0.08740970492362976,
0.02285180799663067,
-0.02589859813451767,
0.021996520459651947,
-0.009848692454397678,
0.013612381182610989,
0.030247638002038002,
-0.01377294771373272,
0.058631207793951035,
-0.07089724391698837,
-0.012628739699721336,
0.14026029407978058,
0.057112351059913635,
-0.018339360132813454,
-0.021963221952319145,
0.0030302959494292736,
0.003939672373235226,
-0.01101823803037405,
0.09700127691030502,
0.007835396565496922,
-0.018235666677355766,
0.01293307077139616,
0.008828189224004745,
-0.09582418203353882,
0.11186384409666061,
0.07040154188871384,
-0.0637994110584259,
0.06295369565486908,
0.07792695611715317,
0.06827601790428162,
0.02934298850595951,
-0.027670826762914658,
-0.021625196561217308,
0.032497767359018326,
0.04624060541391373,
0.037334144115448,
-0.0488649345934391,
0.07052090018987656,
0.01935514621436596,
0.02871476113796234,
-0.0016607068246230483,
-0.012760227546095848,
0.012523514218628407,
0.061409782618284225,
0.11667439341545105,
-0.06339854747056961,
-0.16636431217193604,
-0.05504869669675827
] |
IterComp: Iterative Composition-Aware Feedback Learning from Model Gallery for Text-to-Image Generation | https://openreview.net/forum?id=4w99NAikOE | [
"Xinchen Zhang",
"Ling Yang",
"Guohao Li",
"YaQi Cai",
"xie jiake",
"Yong Tang",
"Yujiu Yang",
"Mengdi Wang",
"Bin CUI"
] | Poster | Advanced diffusion models like Stable Diffusion 3, Omost, and FLUX have made notable strides in compositional text-to-image generation. However, these methods typically exhibit distinct strengths for compositional generation, with some excelling in handling attribute binding and others in spatial relationships. This disparity highlights the need for an approach that can leverage the complementary strengths of various models to comprehensively improve the composition capability. To this end, we introduce IterComp, a novel framework that aggregates composition-aware model preferences from multiple models and employs an iterative feedback learning approach to enhance compositional generation. Specifically, we curate a gallery of six powerful open-source diffusion models and evaluate their three key compositional metrics: attribute binding, spatial relationships, and non-spatial relationships. Based on these metrics, we develop a composition-aware model preference dataset comprising numerous image-rank pairs to train composition-aware reward models. Then, we propose an iterative feedback learning method to enhance compositionality in a closed-loop manner, enabling the progressive self-refinement of both the base diffusion model and reward models over multiple iterations. Detailed theoretical proof demonstrates the effectiveness of this method. Extensive experiments demonstrate our significant superiority over previous methods, particularly in multi-category object composition and complex semantic alignment. IterComp opens new research avenues in reward feedback learning for diffusion models and compositional generation. Code: https://github.com/YangLing0818/IterComp | Compositional text-to-image generation, Feedback learning for diffusion model | null | 2,273 | 2410.07171 | [
0.033288076519966125,
-0.10949865728616714,
0.05546880513429642,
0.07914041727781296,
0.03911953791975975,
0.05944768339395523,
-0.045219600200653076,
-0.023962654173374176,
0.08564558625221252,
-0.04314272105693817,
-0.039237625896930695,
-0.047255828976631165,
0.07179658114910126,
0.09696704894304276,
-0.006473578978329897,
0.03985457867383957,
0.09780068695545197,
0.032142382115125656,
-0.07387810200452805,
-0.053386151790618896,
-0.06591689586639404,
0.026726024225354195,
0.064844511449337,
-0.009724569506943226,
0.0019989481661468744,
0.042064063251018524,
-0.01108795590698719,
-0.04384799674153328,
0.04125385731458664,
-0.010370702482759953,
0.024008439853787422,
0.01331315841525793,
0.018672576174139977,
0.005792360752820969,
-0.08687915652990341,
0.0780211016535759,
-0.04581324756145477,
0.04392935335636139,
0.007821706123650074,
-0.024496447294950485,
-0.008424249477684498,
0.019695905968546867,
0.052222840487957,
0.023162104189395905,
0.05046403780579567,
-0.022942567244172096,
-0.03240695595741272,
-0.05958060920238495,
-0.007515950594097376,
0.05832686647772789,
-0.051085785031318665,
-0.03792916238307953,
-0.0678199976682663,
-0.00099883449729532,
0.022329991683363914,
0.05788566544651985,
0.06438293308019638,
0.01818659156560898,
-0.001869718311354518,
-0.08152129501104355,
-0.020700765773653984,
-0.11951815336942673,
-0.08607185631990433,
0.019895009696483612,
0.07184761017560959,
-0.033367257565259933,
0.020554738119244576,
0.06026841700077057,
-0.12890970706939697,
0.04516691714525223,
-0.004867285490036011,
0.045778948813676834,
-0.016661513596773148,
0.02453804947435856,
0.0743669867515564,
0.11622342467308044,
0.005503779277205467,
-0.0548221617937088,
-0.016598651185631752,
-0.06724228709936142,
0.03126230835914612,
0.08564169704914093,
0.033780716359615326,
-0.03437058627605438,
0.03317318111658096,
-0.02039860375225544,
0.05905473604798317,
0.06672339141368866,
0.028640029951930046,
-0.004531260579824448,
-0.0653751790523529,
0.015577155165374279,
0.006532987114042044,
-0.03914984315633774,
-0.018432334065437317,
0.08213093876838684,
0.0038753661792725325,
-0.08156607300043106,
-0.006326309870928526,
0.04251617193222046,
0.027780506759881973,
0.026212355121970177,
-0.0336608923971653,
-0.045038506388664246,
0.05050225928425789,
-0.05003536120057106,
0.03382405266165733,
0.0786941796541214,
0.008541331626474857,
-0.056805238127708435,
0.047139015048742294,
0.00997274648398161,
-0.018866822123527527,
0.024357646703720093,
-0.013062242418527603,
-0.051212042570114136,
0.08358296751976013,
-0.0720578283071518,
0.09457118064165115,
0.05945758521556854,
0.032183799892663956,
-0.04702496528625488,
-0.002680915640667081,
0.023571060970425606,
-0.03951870650053024,
-0.07542984187602997,
-0.010941778309643269,
-3.8186379388618376e-34,
0.012821285985410213,
-0.016114450991153717,
0.11639872193336487,
0.070194311439991,
0.07751573622226715,
-0.0054291957058012486,
-0.002741910982877016,
-0.02454538084566593,
-0.12671326100826263,
-0.05049356818199158,
0.01093925442546606,
0.03698832169175148,
-0.10124914348125458,
0.13860738277435303,
0.07302854210138321,
-0.039380501955747604,
-0.03256932273507118,
0.04354055970907211,
0.01723751612007618,
-0.03600000962615013,
0.01866660825908184,
-0.005018764641135931,
-0.022641345858573914,
-0.028021125122904778,
0.006332154385745525,
0.0008160141296684742,
0.06216242164373398,
-0.02717084437608719,
-0.12306979298591614,
-0.020199820399284363,
-0.08085040748119354,
0.0018037485424429178,
0.005271138623356819,
0.0013659747783094645,
-0.03991534933447838,
-0.03127625584602356,
-0.018377486616373062,
-0.017468847334384918,
0.027079131454229355,
-0.03537532687187195,
0.04119018837809563,
0.00843322928994894,
-0.08442744612693787,
0.049821753054857254,
-0.09195780754089355,
0.019998665899038315,
-0.008814984001219273,
-0.0001906885881908238,
-0.08653528243303299,
-0.03936881199479103,
0.07805348932743073,
0.020328128710389137,
-0.0038445976097136736,
-0.04557408392429352,
-0.04198217764496803,
0.004000289365649223,
-0.019064178690314293,
-0.028254136443138123,
0.07790178060531616,
-0.043177809566259384,
0.059414803981781006,
0.11768853664398193,
-0.0465119294822216,
0.03659705072641373,
0.00730043975636363,
0.04829765111207962,
0.016132108867168427,
0.023543549701571465,
0.024187186732888222,
0.03600442409515381,
-0.012131772935390472,
0.03094211034476757,
-0.1039164811372757,
-0.0955730751156807,
0.05954267084598541,
-0.02137047052383423,
-0.04701310396194458,
-0.1149386540055275,
0.038103137165308,
0.03203471377491951,
-0.07877090573310852,
-0.02731417678296566,
-0.04660215228796005,
-0.07538226991891861,
-0.05095938965678215,
-0.05292053893208504,
0.04654352366924286,
-0.07521741837263107,
-0.04316313564777374,
-0.02314625307917595,
-0.03190985321998596,
-0.022452589124441147,
-0.015471330843865871,
0.061034172773361206,
0.06444038450717926,
-2.1874207290263173e-34,
-0.00568903423845768,
-0.02058495208621025,
0.012412399984896183,
0.06417260318994522,
0.03143392875790596,
-0.020884202793240547,
-0.04309317469596863,
0.029820913448929787,
0.018395140767097473,
-0.06953801959753036,
-0.02171247825026512,
-0.022860651835799217,
-0.04614796116948128,
-0.0285468902438879,
0.037173863500356674,
-0.018949519842863083,
0.018334371969103813,
0.028989451006054878,
-0.026357905939221382,
0.04588161036372185,
0.03362547978758812,
0.06515232473611832,
-0.045621924102306366,
0.04677766561508179,
-0.09705409407615662,
0.07347045093774796,
0.020498167723417282,
0.05628592148423195,
0.02632957324385643,
0.04348602145910263,
-0.014107080176472664,
-0.0032816273160278797,
0.04378797113895416,
-0.05944349244236946,
-0.10834693163633347,
0.05126599222421646,
-0.026443228125572205,
-0.0507686585187912,
-0.032743871212005615,
0.13863879442214966,
0.005916874390095472,
-0.09738508611917496,
-0.05872721970081329,
0.06103283911943436,
-0.015088587068021297,
-0.042137011885643005,
-0.022754831239581108,
0.01835823245346546,
0.03454108163714409,
0.019403859972953796,
0.031908921897411346,
0.047187305986881256,
-0.11428645998239517,
0.04409899190068245,
-0.038755401968955994,
-0.06659961491823196,
-0.020609209313988686,
-0.038755059242248535,
0.12072973698377609,
-0.03247040882706642,
-0.043704234063625336,
-0.031443048268556595,
0.026762856170535088,
-0.06275784969329834,
-0.003574153408408165,
-0.029611019417643547,
-0.013218346983194351,
-0.04358990490436554,
-0.030110400170087814,
0.09443221241235733,
0.008344306610524654,
-0.0014749524416401982,
0.01725355163216591,
0.0011384518584236503,
0.07811290770769119,
-0.0481095090508461,
0.0323573462665081,
0.021708013489842415,
0.013112838380038738,
-0.0739109218120575,
-0.02698398567736149,
0.05805215239524841,
0.025009851902723312,
0.03277326375246048,
0.10086478292942047,
0.024913785979151726,
0.01317975576967001,
0.06226467713713646,
0.012209659442305565,
0.025333791971206665,
-0.019844362512230873,
0.003579036332666874,
0.050426289439201355,
0.03457185998558998,
-0.05661334842443466,
-4.03623445777157e-8,
-0.10777383297681808,
-0.02397271804511547,
0.01993708685040474,
0.0806511789560318,
-0.0702422708272934,
-0.012751402333378792,
0.08776866644620895,
0.0458621121942997,
0.03660799562931061,
-0.0018384746508672833,
0.05312853679060936,
0.017734773457050323,
-0.01734158582985401,
-0.05816938355565071,
-0.023495975881814957,
0.05003809183835983,
0.026356393471360207,
0.010628140531480312,
0.017254948616027832,
-0.005410211160778999,
-0.0001915805332828313,
0.010511702857911587,
0.0676107332110405,
-0.08495612442493439,
0.03693345934152603,
-0.04359108582139015,
-0.04829799011349678,
0.01975971646606922,
0.09505455940961838,
-0.01919841580092907,
0.061194345355033875,
-0.03811423107981682,
-0.051244959235191345,
-0.002762138843536377,
-0.008369291201233864,
0.04814419895410538,
0.026753535494208336,
-0.11055199056863785,
-0.013450690545141697,
-0.061268918216228485,
0.020903462544083595,
0.059670209884643555,
-0.005162591580301523,
0.0036994146648794413,
-0.030889561399817467,
0.04442468658089638,
0.045765649527311325,
-0.03328436240553856,
0.0016305377939715981,
-0.005861780606210232,
-0.03858566656708717,
-0.11301705986261368,
-0.03850233182311058,
0.04369772598147392,
0.07280830293893814,
-0.06452770531177521,
0.019279971718788147,
0.004684405867010355,
0.10003557801246643,
0.08166943490505219,
0.027874058112502098,
0.0016066397074609995,
-0.05311822518706322,
0.009539502672851086
] |
MambaQuant: Quantizing the Mamba Family with Variance Aligned Rotation Methods | https://openreview.net/forum?id=KI45uDnmzv | [
"Zukang Xu",
"Yuxuan Yue",
"Xing Hu",
"Dawei Yang",
"Zhihang Yuan",
"Zixu Jiang",
"Zhixuan Chen",
"JiangyongYu",
"XUCHEN",
"Sifan Zhou"
] | Poster | Mamba is an efficient sequence model that rivals Transformers and demonstrates significant potential as a foundational architecture for various tasks. Quantization is commonly used in neural networks to reduce model size and computational latency. However, applying quantization to Mamba remains underexplored, and existing quantization methods, which have been effective for CNN and Transformer models, appear inadequate for Mamba models (e.g., Quarot suffers a 21% accuracy drop on Vim-T$\dagger$ even under W8A8). We have pioneered the exploration of this issue and identified several key challenges. First, significant outliers arepresent in gate projections, output projections, and matrix multiplications. Second, Mamba’s unique parallel scan further amplifies these outliers, leading to uneven and heavy-tailed data distributions. Third, even with the application of the Hadamard transform, the variance across channels in weights and activations still remains inconsistent. To these ends, we propose MambaQuant, a post-training quantization (PTQ) framework consisting of: 1) Karhunen-Lo`eve Transformation (KLT) enhanced rotation, rendering the rotation matrix adaptable to diverse channel distributions. 2) Smooth-Fused rotation, which equalizes channel variances and can merge additional parameters into model weights. Experiments show that MambaQuant can quantize both weights and activations into 8-bit with less than 1% accuracy loss for Mamba-based vision and language tasks. To our knowledge, MambaQuant is the first comprehensive PTQ design for the Mamba family, paving the way for further advancements in its application. | Mamba, Quantization | null | 2,268 | 2501.13484 | [
-0.11967413127422333,
-0.061082109808921814,
-0.011796494945883751,
0.005115737672895193,
-0.08223831653594971,
0.05895981192588806,
-0.08805087953805923,
-0.012397346086800098,
0.04338500648736954,
-0.04182354733347893,
0.02049834467470646,
-0.003434040816500783,
-0.034867994487285614,
0.017069872468709946,
-0.05872116982936859,
-0.006080351769924164,
0.03538195416331291,
0.0928584486246109,
-0.1018485352396965,
-0.06739547103643417,
0.05652947351336479,
0.03725891560316086,
-0.013428173027932644,
-0.0051531414501369,
0.07678862661123276,
0.039845071732997894,
0.005951943341642618,
-0.014347703196108341,
0.008490249514579773,
-0.07309745997190475,
-0.006039989646524191,
0.03318880498409271,
0.0053177690133452415,
-0.026237376034259796,
-0.041511036455631256,
0.051718369126319885,
-0.0038654105737805367,
0.01580020599067211,
0.04226002097129822,
-0.09300904721021652,
0.06419316679239273,
-0.036910951137542725,
0.07878492772579193,
-0.02762882597744465,
0.04277408868074417,
-0.010857518762350082,
0.06677553057670593,
-0.0265501718968153,
0.059432078152894974,
0.00440513901412487,
0.005917556583881378,
0.018119415268301964,
-0.01410524919629097,
0.13450056314468384,
-0.04378576576709747,
-0.03837498649954796,
0.01401268970221281,
0.023776600137352943,
0.020560499280691147,
-0.031028157100081444,
-0.003240096615627408,
0.023969018831849098,
0.029993707314133644,
0.013839886523783207,
0.001826236373744905,
-0.0686807855963707,
0.05829748511314392,
0.009899245575070381,
-0.07247098535299301,
0.02667350508272648,
0.005303431302309036,
0.01147807203233242,
-0.03650236874818802,
0.09679876267910004,
0.06920783221721649,
0.06175742298364639,
0.12435183674097061,
0.003311657113954425,
0.09458668529987335,
-0.004863477777689695,
0.0754573866724968,
0.025357874110341072,
0.044972434639930725,
0.03335726633667946,
0.1009429544210434,
-0.00943491980433464,
-0.13072627782821655,
0.09082774817943573,
0.006135380361229181,
-0.01738228276371956,
-0.05527498945593834,
-0.005550093483179808,
0.008106919936835766,
-0.037108734250068665,
0.07355786859989166,
0.031592074781656265,
0.013629336841404438,
0.013951548375189304,
-0.03632026165723801,
0.08802313357591629,
0.07385321706533432,
-0.005647678393870592,
-0.0018378349486738443,
-0.002938798861578107,
0.055729545652866364,
-0.015836145728826523,
0.08192987740039825,
0.026609046384692192,
0.03746722266077995,
-0.10417386889457703,
0.00048706430243328214,
0.011972341686487198,
-0.024864275008440018,
-0.08945004642009735,
0.06042635813355446,
0.006630057003349066,
0.025731710717082024,
0.03614399954676628,
-0.04517596587538719,
-0.04035983979701996,
-0.05571608990430832,
-0.00596371665596962,
-0.012263420037925243,
0.09598792344331741,
0.06017053872346878,
-0.03350682929158211,
-0.16926859319210052,
4.7250523573957915e-33,
0.026495305821299553,
0.05849212408065796,
0.09190064668655396,
-0.04917100444436073,
0.06369929015636444,
-0.05104803293943405,
0.05278264731168747,
-0.0014464376727119088,
-0.008579634130001068,
0.03654688969254494,
-0.08778035640716553,
0.011913786642253399,
-0.0830967053771019,
-0.008982432074844837,
0.002770926570519805,
-0.03368251025676727,
-0.03634257987141609,
-0.014349118806421757,
-0.03231896460056305,
-0.07065828144550323,
0.0464838370680809,
-0.09677540510892868,
-0.031045744195580482,
0.07442064583301544,
-0.035987380892038345,
0.035660021007061005,
0.09510742872953415,
-0.023484617471694946,
-0.07851731777191162,
0.013376928865909576,
-0.07510785758495331,
-0.000406621111324057,
-0.05083996802568436,
-0.06352215260267258,
0.03673972934484482,
-0.07676638662815094,
-0.014344679191708565,
-0.027047110721468925,
0.02116432972252369,
-0.019721245393157005,
-0.03720777854323387,
0.006125226151198149,
-0.0062953713349998,
-0.03728953376412392,
-0.03232309967279434,
-0.021656673401594162,
0.04012955352663994,
-0.06389246135950089,
0.04547199606895447,
-0.014958715997636318,
0.06456698477268219,
-0.027464773505926132,
-0.0482359416782856,
0.008649329654872417,
0.01644444465637207,
0.012328381650149822,
0.1007298156619072,
0.044762611389160156,
0.03156503289937973,
0.06792005151510239,
-0.07431476563215256,
-0.04967712238430977,
-0.03854592517018318,
0.06663358211517334,
-0.040736064314842224,
-0.03793402761220932,
0.004475149791687727,
0.054424453526735306,
-0.013194357044994831,
0.0909891277551651,
-0.0445963479578495,
-0.00505922082811594,
-0.02843906171619892,
-0.05167366936802864,
0.09611287713050842,
0.024912789463996887,
0.043057456612586975,
0.01658695936203003,
-0.05698531121015549,
-0.023045659065246582,
0.007778366096317768,
0.04085192084312439,
0.012670718133449554,
0.014787421561777592,
-0.044009942561388016,
0.03484402224421501,
0.03353752940893173,
-0.03572770953178406,
-0.0245994720607996,
0.005427644122391939,
-0.03948846831917763,
0.006201001815497875,
-0.007542129140347242,
-0.03813592344522476,
-0.10394304245710373,
-3.354869326112455e-33,
-0.09785129129886627,
0.10335393995046616,
-0.04377327486872673,
0.042233362793922424,
-0.07073232531547546,
-0.11623334139585495,
0.012375685386359692,
0.06245626509189606,
-0.012347805313766003,
-0.11683192849159241,
-0.01735803484916687,
-0.036778293550014496,
0.02478458546102047,
-0.01921718753874302,
0.05292878299951553,
0.047088149935007095,
-0.0027701465878635645,
-0.00005366899858927354,
-0.029085488989949226,
0.018850114196538925,
0.029430953785777092,
0.06487350165843964,
-0.07400399446487427,
-0.053071748465299606,
-0.08535048365592957,
0.020143935456871986,
-0.029532579705119133,
0.08370266854763031,
0.02731386385858059,
-0.004099903162568808,
-0.03850695118308067,
-0.03974536433815956,
-0.05532516911625862,
-0.04823942109942436,
-0.03677109628915787,
0.022256415337324142,
0.014972533099353313,
0.004343262407928705,
-0.026180651038885117,
0.007816233672201633,
0.07025721669197083,
-0.026989249512553215,
-0.10308525711297989,
0.06487033516168594,
-0.056498803198337555,
0.038274481892585754,
-0.01031961478292942,
0.01853027008473873,
-0.030078865587711334,
0.07073622196912766,
0.019841810688376427,
0.05751532316207886,
-0.013035845942795277,
0.10187416523694992,
-0.00931430235505104,
-0.014221541583538055,
0.02846689522266388,
0.016759049147367477,
0.04950287938117981,
0.054898034781217575,
-0.08816493302583694,
-0.048961445689201355,
0.02615397237241268,
-0.1473175287246704,
-0.008015058003365993,
0.01295862253755331,
-0.016177915036678314,
0.0145585797727108,
0.004078658297657967,
0.07414136081933975,
-0.02699643187224865,
-0.06468790769577026,
0.013223711401224136,
-0.017658470198512077,
0.004753184970468283,
-0.08121379464864731,
-0.02191277965903282,
-0.048320092260837555,
0.013831856660544872,
0.021328043192625046,
-0.05373624339699745,
-0.009173576720058918,
0.0329151414334774,
-0.04168487340211868,
0.05975122004747391,
0.10556372255086899,
0.08361032605171204,
0.0036201346665620804,
0.06409317255020142,
0.04453476890921593,
-0.07170853018760681,
0.06766042858362198,
0.02093328721821308,
-0.020188482478260994,
-0.08038336038589478,
-4.891435168019598e-8,
-0.06614623963832855,
0.05362701788544655,
-0.018584420904517174,
-0.03757520765066147,
0.04992279037833214,
-0.0493655689060688,
0.010205589234828949,
0.13797368109226227,
-0.0119164464995265,
-0.02630755491554737,
0.08898914605379105,
-0.07314798980951309,
-0.05125776678323746,
-0.012644050642848015,
-0.005029746796935797,
0.09009496867656708,
0.028649453073740005,
0.03869492560625076,
0.0009736442007124424,
-0.04785653203725815,
-0.04033692181110382,
0.10229847580194473,
0.01466729398816824,
-0.013723364099860191,
0.04430954530835152,
-0.024483639746904373,
-0.0372966006398201,
0.055013108998537064,
0.09030420333147049,
0.0411209911108017,
-0.03591426461935043,
0.027360066771507263,
0.05004128813743591,
-0.021444054320454597,
-0.06058068200945854,
0.034437574446201324,
-0.00038932159077376127,
-0.046060021966695786,
-0.02119947411119938,
-0.003452255856245756,
0.013545216992497444,
0.034185655415058136,
-0.056344058364629745,
0.07118739187717438,
-0.05039852112531662,
0.04141385108232498,
0.04593110457062721,
-0.02490558661520481,
0.034055545926094055,
-0.04780496656894684,
0.006052260287106037,
-0.012574294582009315,
-0.004850989673286676,
0.018533747643232346,
-0.0008985152235254645,
0.009164368733763695,
0.08925455808639526,
-0.06965784728527069,
0.018966296687722206,
0.05700113624334335,
0.0021025214809924364,
-0.03043084777891636,
-0.04714047163724899,
-0.010314940474927425
] |
ST-GCond: Self-supervised and Transferable Graph Dataset Condensation | https://openreview.net/forum?id=wYWJFLQov9 | [
"Beining Yang",
"Qingyun Sun",
"Cheng Ji",
"Xingcheng Fu",
"Jianxin Li"
] | Poster | The increasing scale of graph datasets significantly enhances deep learning models but also presents substantial training challenges. Graph dataset condensation has emerged to condense large datasets into smaller yet informative ones that maintain similar test performance. However, these methods require downstream usage to match the original dataset and task, which is impractical in real-world scenarios. Our empirical studies show that existing methods fail in "cross-task" and "cross-dataset" scenarios, often performing worse than training from scratch. To address these challenges, we propose a novel method: Self-supervised and Transferable Graph dataset Condensation (ST-GCond). For cross-task transferability, we propose a task-disentangled meta optimization strategy to adaptively update the condensed graph according to the task relevance, encouraging information preservation for various tasks. For cross-dataset transferability, we propose a multi-teacher self-supervised optimization strategy to incorporate auxiliary self-supervised tasks to inject universal knowledge into the condensed graph. Additionally, we incorporate mutual information guided joint condensation mitigating the potential conflicts and ensure the condensing stability. Experiments on both node-level and graph-level datasets show that ST-GCond outperforms existing methods by 2.5% to 18.7% in all cross-task and cross-dataset scenarios, and also achieves state-of-the-art performance on 5 out of 6 datasets in the single dataset and task scenario. | Graph Neural Network; Graph Dataset Condensation | Current graph dataset condensation only designed for single task&dataset, showing poor performance in transferring scenarios. Hence, we redesign the supervised condensation framework and include self-supervised tasks, enhancing final transferability. | 2,264 | null | [
-0.06904395669698715,
-0.05104801803827286,
0.02597115747630596,
0.04146553948521614,
0.07117855548858643,
0.008011036552488804,
-0.02589651383459568,
-0.06496227532625198,
-0.05658367648720741,
-0.04080644249916077,
-0.005268708802759647,
-0.032203562557697296,
-0.055403608828783035,
0.09150601178407669,
-0.03270147368311882,
0.01606791466474533,
0.1398574411869049,
0.03168615326285362,
-0.1059088259935379,
-0.02697855420410633,
0.029256828129291534,
0.05215650424361229,
0.020584922283887863,
0.0797622948884964,
0.0843019038438797,
-0.002922394312918186,
-0.00864089373499155,
-0.08340082317590714,
0.056213151663541794,
-0.04084479808807373,
0.09580590575933456,
-0.03172822669148445,
-0.09785584360361099,
0.06594511866569519,
-0.038034770637750626,
0.07812368869781494,
-0.012953449040651321,
0.043413013219833374,
0.04496978595852852,
-0.03522113338112831,
0.04954877123236656,
0.007018472533673048,
0.019224220886826515,
0.06651841104030609,
0.023252803832292557,
0.09867311269044876,
-0.05098995938897133,
0.009579223580658436,
0.05068083852529526,
0.05742226541042328,
-0.006088846363127232,
-0.02177034318447113,
-0.09020254760980606,
0.07388141751289368,
0.023242758587002754,
0.05413777753710747,
0.09156523644924164,
-0.050800979137420654,
-0.0283480454236269,
0.039645563811063766,
0.03525329381227493,
-0.0597546212375164,
0.05319957435131073,
-0.028170118108391762,
0.0021500878501683474,
0.03947746753692627,
0.0013459601905196905,
0.08292177319526672,
-0.04599447920918465,
0.035381000488996506,
0.02257653698325157,
0.07278713583946228,
-0.04827136918902397,
0.03244888782501221,
0.09885978698730469,
0.08928444981575012,
0.013884352520108223,
0.010467598214745522,
0.0022781703155487776,
-0.04838205501437187,
0.02921530418097973,
0.06581950932741165,
0.042314063757658005,
0.0256488136947155,
0.07087714225053787,
-0.011136136017739773,
-0.005554774776101112,
0.013901833444833755,
0.0014848888386040926,
-0.009899086318910122,
-0.11208801716566086,
0.00012789227184839547,
-0.03969170153141022,
-0.006991267204284668,
0.013852938078343868,
-0.003813671413809061,
0.06305631250143051,
-0.022618144750595093,
-0.006301223300397396,
0.037144046276807785,
0.005633107386529446,
0.02955050580203533,
0.028950728476047516,
0.03794438764452934,
-0.03416787087917328,
0.015350925736129284,
-0.004772532731294632,
0.02383287250995636,
0.018705282360315323,
-0.09261982142925262,
0.045513737946748734,
0.07160564512014389,
0.009900097735226154,
-0.06582526862621307,
0.017718445509672165,
-0.02257225476205349,
-0.05008396506309509,
0.009509572759270668,
0.0036165493074804544,
0.04209304228425026,
-0.04714525118470192,
0.026585331186652184,
0.05443218722939491,
0.01628827676177025,
-0.04456133022904396,
-0.004211667459458113,
-0.05104180797934532,
1.0986889274720016e-33,
0.05059066414833069,
0.042411018162965775,
0.07280177623033524,
-0.014247897081077099,
-0.003028689417988062,
-0.02046475186944008,
-0.06453455984592438,
0.05529432371258736,
-0.03688132390379906,
0.030111055821180344,
-0.018620913848280907,
-0.014881358481943607,
-0.05374707654118538,
0.05232226476073265,
0.05342094972729683,
0.0333213284611702,
-0.019891729578375816,
-0.0011592194205150008,
0.03268645331263542,
-0.09979706257581711,
0.06487365067005157,
0.023885317146778107,
-0.02193889580667019,
0.011225342750549316,
0.07467366009950638,
0.01970633864402771,
0.04644474759697914,
-0.0007670857012271881,
0.03491929545998573,
-0.011946849524974823,
-0.14279891550540924,
0.06058737263083458,
0.027558842673897743,
0.0656794011592865,
0.011628876440227032,
-0.05680273100733757,
-0.04641173034906387,
-0.06017671525478363,
0.03866911679506302,
-0.006656746845692396,
0.05179243162274361,
0.05480387434363365,
0.024245765060186386,
-0.03262856975197792,
-0.04201049357652664,
-0.1708320677280426,
0.004626845009624958,
-0.03200683370232582,
-0.023440120741724968,
-0.05898096039891243,
-0.00861548725515604,
-0.015884073451161385,
-0.07817372679710388,
-0.033673301339149475,
-0.01044744998216629,
-0.034294482320547104,
0.05579688400030136,
0.08963475376367569,
0.04875757917761803,
0.03566931560635567,
0.04533141478896141,
0.00013421544281300157,
-0.01794613152742386,
0.11142229288816452,
0.0031858477741479874,
0.045139819383621216,
-0.059893228113651276,
0.05701088905334473,
0.06817968189716339,
-0.08302096277475357,
-0.08944899588823318,
0.02210172452032566,
-0.03230978548526764,
-0.04485645517706871,
0.052314069122076035,
-0.04021444916725159,
-0.019522104412317276,
-0.12130243331193924,
-0.031803205609321594,
-0.007177517283707857,
0.01956862397491932,
-0.0015579889295622706,
-0.05804500728845596,
-0.02981208637356758,
-0.07899408042430878,
-0.03707928955554962,
0.0558495819568634,
-0.05361644923686981,
0.009664604440331459,
-0.01790083758533001,
-0.05831850692629814,
0.018695548176765442,
-0.0050136796198785305,
-0.007958324626088142,
0.03248332068324089,
-7.752721361588924e-34,
0.013025481253862381,
0.07628356665372849,
-0.03269001841545105,
0.06681206077337265,
0.07171661406755447,
-0.0447736494243145,
0.05808673053979874,
-0.01144823431968689,
-0.0597880519926548,
0.010039275512099266,
-0.056673042476177216,
-0.09170203655958176,
-0.039430808275938034,
-0.026807503774762154,
-0.03263668343424797,
0.06216450780630112,
-0.055810317397117615,
0.025441128760576248,
-0.08157461136579514,
-0.00016881708870641887,
-0.018733277916908264,
0.09617768973112106,
-0.044246312230825424,
0.04536353424191475,
0.02448587864637375,
0.011705052107572556,
-0.03013661876320839,
0.013630565255880356,
0.04446347430348396,
0.07335663586854935,
-0.016355326399207115,
-0.0005912933847866952,
-0.039093151688575745,
-0.08403591066598892,
0.0030945155303925276,
0.052473973482847214,
0.0478048101067543,
-0.0825294554233551,
-0.059827595949172974,
0.026735378429293633,
0.017016874626278877,
-0.019043199717998505,
-0.08489381521940231,
0.02255058102309704,
0.003027831669896841,
0.03430495038628578,
-0.13739055395126343,
-0.013119803741574287,
-0.014376863837242126,
0.08043954521417618,
-0.05369789898395538,
0.048794426023960114,
-0.10311605036258698,
0.00895444955676794,
-0.022115709260106087,
-0.16437320411205292,
0.07672722637653351,
0.033254265785217285,
0.011610344052314758,
-0.0247175432741642,
-0.0238985363394022,
-0.047836750745773315,
0.01646967977285385,
-0.07088994979858398,
0.023849641904234886,
-0.08205980062484741,
-0.005918129812926054,
-0.03177032992243767,
0.056778937578201294,
0.05821961537003517,
0.01864580437541008,
-0.005844577681273222,
0.012517677620053291,
-0.05560925602912903,
0.045462992042303085,
-0.03954804688692093,
0.04865086078643799,
0.0005938304238952696,
-0.035406094044446945,
-0.0037695723585784435,
0.010401759296655655,
-0.016977230086922646,
0.008008996024727821,
0.10337816178798676,
0.05382690951228142,
0.06727809458971024,
0.020181814208626747,
0.11297357827425003,
-0.0034916598815470934,
-0.003488803282380104,
-0.040823109447956085,
-0.03010106459259987,
-0.08120524883270264,
0.024169836193323135,
-0.017165785655379295,
-3.491706124236771e-8,
-0.13622558116912842,
0.03161964192986488,
0.02893759123980999,
0.03439876437187195,
-0.02764020301401615,
-0.019156014546751976,
0.027325758710503578,
0.07109075039625168,
0.005827135406434536,
0.03885362669825554,
0.050081584602594376,
0.00812551099807024,
-0.01817452348768711,
-0.037887878715991974,
-0.06696146726608276,
-0.045988623052835464,
0.04160606116056442,
0.00011459193774498999,
0.0503903292119503,
-0.01388959027826786,
-0.04146425053477287,
-0.016785284504294395,
-0.013934866525232792,
-0.04422133415937424,
0.026823511347174644,
-0.10410532355308533,
0.05185616388916969,
0.060340192168951035,
0.03261714056134224,
0.004646773915737867,
-0.003366215154528618,
-0.028182290494441986,
0.039563246071338654,
-0.03132036328315735,
-0.00799782108515501,
0.08434432744979858,
0.06112516298890114,
0.06829959899187088,
-0.0253513902425766,
-0.02963130548596382,
0.000706906954292208,
0.18235613405704498,
-0.09087833017110825,
0.027823299169540405,
0.01711774803698063,
-0.0003793233190663159,
0.031927525997161865,
0.00614105723798275,
-0.007443289738148451,
0.014302844181656837,
-0.004442830570042133,
-0.05825583264231682,
-0.0012691018637269735,
0.04377302527427673,
0.06955122947692871,
0.01740444079041481,
-0.0355997309088707,
-0.013335250318050385,
0.011624660342931747,
0.02926175855100155,
-0.0027274847961962223,
-0.05961983650922775,
-0.10183044523000717,
-0.0423659086227417
] |
PPT: Patch Order Do Matters In Time Series Pretext Task | https://openreview.net/forum?id=7zwIEbSTDy | [
"Jaeho Kim",
"Kwangryeol Park",
"Sukmin Yun",
"Seulki Lee"
] | Poster | Recently, patch-based models have been widely discussed in time series analysis. However, existing pretext tasks for patch-based learning, such as masking, may not capture essential time and channel-wise patch interdependencies in time series data, presumed to result in subpar model performance. In this work, we introduce *Patch order-aware Pretext Task (PPT)*, a new self-supervised patch order learning pretext task for time series classification. PPT exploits the intrinsic sequential order information among patches across time and channel dimensions of time series data, where model training is aided by channel-wise patch permutations. The permutation disrupts patch order consistency across time and channel dimensions with controlled intensity to provide supervisory signals for learning time series order characteristics. To this end, we propose two patch order-aware learning methods: patch order consistency learning, which quantifies patch order correctness, and contrastive learning, which distinguishes weakly permuted patch sequences from strongly permuted ones. With patch order learning, we observe enhanced model performance, e.g., improving up to 7% accuracy for the supervised cardiogram task and outperforming mask-based learning by 5% in the self-supervised human activity recognition task. We also propose ACF-CoS, an evaluation metric that measures the *importance of orderness* for time series datasets, which enables pre-examination of the efficacy of PPT in model training. | Time Series Classification, Self-Supervised Learning, Pretext Task | Patch order aware self-supervised learning methodology for time series classification | 2,262 | null | [
-0.07301127165555954,
-0.08495543152093887,
0.11281576007604599,
0.004598020575940609,
-0.011831637471914291,
0.05115935578942299,
-0.03299293667078018,
-0.050601959228515625,
0.03801431134343147,
-0.023626521229743958,
-0.060809358954429626,
0.05249311402440071,
-0.1139107421040535,
0.10082016885280609,
-0.0595589280128479,
-0.05924353003501892,
0.0658840611577034,
0.01094615925103426,
0.004460313357412815,
-0.020275022834539413,
0.04483572393655777,
0.08131098747253418,
-0.010530431754887104,
0.06244860962033272,
0.03051019459962845,
0.007580467965453863,
0.015106048434972763,
0.03852409869432449,
0.05090588703751564,
0.0046957917511463165,
-0.04638366028666496,
0.003967208322137594,
0.03371839225292206,
0.022384922951459885,
-0.10395927727222443,
0.010262435302138329,
-0.05176560953259468,
-0.004884969908744097,
0.017653876915574074,
0.03764563426375389,
0.02578425034880638,
-0.002857885556295514,
0.025588693097233772,
-0.011162552051246166,
0.042006392031908035,
0.028924226760864258,
0.11360584944486618,
-0.05918875336647034,
-0.06416834145784378,
-0.020314114168286324,
-0.0016951020807027817,
0.0407966673374176,
-0.041331663727760315,
0.09691501408815384,
-0.03240944817662239,
-0.0011288649402558804,
0.04246874526143074,
-0.006448896136134863,
0.0020740560721606016,
0.008963174186646938,
0.012284788303077221,
-0.08590307831764221,
-0.03040754422545433,
-0.02073902077972889,
-0.024570036679506302,
-0.035060539841651917,
0.05123889446258545,
0.07645553350448608,
-0.02303641475737095,
-0.01052778959274292,
-0.03465322032570839,
0.1261846274137497,
-0.008316469378769398,
0.028982896357774734,
0.021922966465353966,
0.07192836701869965,
0.032965801656246185,
-0.0556451678276062,
0.02744346112012863,
-0.08493948727846146,
-0.0379846952855587,
0.022942936047911644,
0.05854899808764458,
-0.009006956592202187,
0.11868419498205185,
-0.032746803015470505,
0.02716502547264099,
0.07308521121740341,
-0.0229247584939003,
-0.04623214155435562,
-0.06833881139755249,
-0.016756344586610794,
0.026304790750145912,
-0.01338687352836132,
0.01340358704328537,
0.04298984631896019,
0.0069517530500888824,
-0.015772761777043343,
-0.009666875004768372,
0.09886150062084198,
0.03018355742096901,
0.02996169775724411,
-0.09053587913513184,
0.07610882073640823,
0.07640271633863449,
0.004667373839765787,
-0.0335499607026577,
0.006662961561232805,
0.09397439658641815,
-0.0786798894405365,
0.03735782951116562,
0.016277780756354332,
-0.010889573022723198,
-0.013324477709829807,
-0.02460394613444805,
-0.0005546269821934402,
-0.017911672592163086,
0.019197847694158554,
0.025020821020007133,
0.0902237743139267,
0.004568532109260559,
-0.008216118440032005,
0.03108549676835537,
0.0451953187584877,
0.03425148129463196,
-0.0302367452532053,
0.06285630166530609,
1.5535217627742193e-33,
0.08464496582746506,
0.05444543808698654,
0.07184600830078125,
-0.014796094037592411,
0.010518067516386509,
-0.01062659453600645,
-0.03436525911092758,
-0.053341832011938095,
0.06682375818490982,
0.059167057275772095,
-0.03669741004705429,
-0.0311634112149477,
-0.0659579485654831,
0.02395116165280342,
-0.0852019339799881,
-0.0063592130318284035,
-0.02707580104470253,
0.022493917495012283,
-0.015292254276573658,
-0.02095685712993145,
0.040156345814466476,
-0.0383305586874485,
-0.04283324256539345,
-0.06542691588401794,
0.029728587716817856,
0.0467250682413578,
0.05047473683953285,
-0.009033075533807278,
-0.006002551876008511,
0.008870778605341911,
-0.06153762713074684,
-0.005474001169204712,
0.06969492882490158,
0.00834753829985857,
0.0458814911544323,
-0.030292365700006485,
0.027063867077231407,
-0.03261502832174301,
0.0782725065946579,
-0.053254153579473495,
-0.001894337241537869,
-0.020583873614668846,
-0.045973580330610275,
-0.07649882882833481,
0.008070342242717743,
-0.12738746404647827,
0.0026145970914512873,
0.055305663496255875,
-0.009262602776288986,
0.013198909349739552,
0.019727367907762527,
-0.002846565330401063,
-0.032957449555397034,
-0.14168880879878998,
-0.03933284431695938,
-0.012569112703204155,
0.06127846986055374,
0.042042896151542664,
0.03400418162345886,
0.08796903491020203,
-0.005357385613024235,
0.026179635897278786,
-0.043925635516643524,
0.03660064563155174,
-0.025660065934062004,
0.005257735028862953,
0.05089268833398819,
0.04987897351384163,
0.03750048205256462,
-0.03241278976202011,
-0.05940805748105049,
0.03581231087446213,
-0.1059175580739975,
-0.05301033705472946,
0.07778104394674301,
0.040679268538951874,
0.08033805340528488,
-0.0027844589203596115,
-0.08446633815765381,
-0.0007897380855865777,
-0.03286491334438324,
-0.04055531322956085,
0.06422628462314606,
-0.10694721341133118,
-0.0963449701666832,
-0.035923201590776443,
0.09458722919225693,
0.020153608173131943,
0.04045933857560158,
-0.07265353947877884,
-0.021324969828128815,
0.0687297061085701,
0.08168304711580276,
0.05727452039718628,
0.04634863883256912,
-1.0873485293934359e-33,
-0.009076891466975212,
-0.04182001203298569,
-0.03299374878406525,
-0.009945410303771496,
0.051431525498628616,
-0.0877261683344841,
-0.019698144868016243,
0.06178824231028557,
-0.026048393920063972,
-0.00901342649012804,
-0.03513181954622269,
-0.03713230416178703,
0.013336585834622383,
-0.04551868513226509,
-0.004952638875693083,
0.022864535450935364,
-0.02684754692018032,
0.00489061139523983,
-0.034233398735523224,
0.022529030218720436,
-0.015770668163895607,
0.03633969649672508,
-0.06978114694356918,
0.05699287727475166,
-0.025672387331724167,
0.03353486582636833,
0.0667383000254631,
0.08057485520839691,
0.037437357008457184,
0.02207927592098713,
-0.00886225700378418,
-0.06632083654403687,
0.004766142927110195,
-0.012323619797825813,
-0.07021785527467728,
0.0876709595322609,
-0.0032804531510919333,
-0.005963498260825872,
-0.02134445123374462,
0.1311391294002533,
0.06229817867279053,
-0.017544684931635857,
-0.07600764185190201,
-0.027583250775933266,
-0.13361534476280212,
-0.03826863318681717,
-0.042775627225637436,
0.01734495908021927,
-0.068287692964077,
0.01046784594655037,
-0.03236307203769684,
0.10411661863327026,
-0.018497465178370476,
-0.03775707632303238,
-0.02403636835515499,
0.007161746267229319,
-0.005684512201696634,
0.00601129001006484,
0.05250445380806923,
-0.021130075678229332,
-0.022497257217764854,
0.01107796374708414,
0.07902126759290695,
-0.08027487993240356,
0.04526735842227936,
-0.014346702955663204,
-0.023874744772911072,
-0.09908527135848999,
0.04073597490787506,
0.008076204918324947,
0.020935844630002975,
-0.012487320229411125,
-0.05325956270098686,
-0.07117876410484314,
0.002338630147278309,
-0.01097172312438488,
0.002082373946905136,
-0.012161695398390293,
-0.0477115623652935,
-0.008257578127086163,
-0.06311804056167603,
-0.016257435083389282,
-0.01144458632916212,
0.11764460057020187,
-0.016912002116441727,
0.11262483149766922,
0.10560697317123413,
0.048137910664081573,
0.055012352764606476,
-0.04062892869114876,
-0.05606847628951073,
0.09902907907962799,
-0.018216770142316818,
0.0026346901431679726,
0.00335231376811862,
-3.757615019139848e-8,
-0.07299111783504486,
-0.027847129851579666,
-0.003432233352214098,
0.020751001313328743,
-0.010981043800711632,
-0.09714887291193008,
0.012755412608385086,
0.061062365770339966,
0.01445307582616806,
-0.024594178423285484,
0.13121552765369415,
0.03029647469520569,
-0.011117343790829182,
-0.1038111075758934,
-0.004203269258141518,
0.040216803550720215,
-0.02568592131137848,
-0.04753679782152176,
-0.06459755450487137,
-0.05205263942480087,
-0.02001906745135784,
0.05302368104457855,
0.09994085133075714,
-0.06977057456970215,
0.049894362688064575,
-0.06425274163484573,
0.0017651856178417802,
0.10997622460126877,
0.06686387956142426,
0.025624698027968407,
0.025775136426091194,
0.03524075448513031,
-0.008870474994182587,
-0.0496881902217865,
-0.05007229745388031,
0.0872659683227539,
0.070467010140419,
0.016453919932246208,
-0.035395585000514984,
-0.01084840390831232,
-0.03424837812781334,
0.01697422005236149,
-0.055753979831933975,
0.023615475744009018,
-0.05480699613690376,
-0.048312850296497345,
-0.03553784266114235,
-0.09500449150800705,
0.0010636276565492153,
-0.10151449590921402,
0.017839746549725533,
0.01673104055225849,
0.034259725362062454,
0.0474366769194603,
0.07565097510814667,
0.0932842344045639,
-0.006479931995272636,
-0.06250926107168198,
-0.0044923205859959126,
0.02488774061203003,
-0.04440145567059517,
-0.014787816442549229,
-0.02516603097319603,
-0.0033296348992735147
] |
Rethinking Graph Neural Networks From A Geometric Perspective Of Node Features | https://openreview.net/forum?id=lBMRmw59Lk | [
"Feng Ji",
"Yanan Zhao",
"Kai Zhao",
"Hanyang Meng",
"Jielong Yang",
"Wee Peng Tay"
] | Poster | Many works on graph neural networks (GNNs) focus on graph topologies and analyze graph-related operations to enhance performance on tasks such as node classification. In this paper, we propose to understand GNNs based on a feature-centric approach. Our main idea is to treat the features of nodes from each label class as a whole, from which we can identify the centroid. The convex hull of these centroids forms a simplex called the feature centroid simplex, where a simplex is a high-dimensional generalization of a triangle. We borrow ideas from coarse geometry to analyze the geometric properties of the feature centroid simplex by comparing them with basic geometric models, such as regular simplexes and degenerate simplexes. Such a simplex provides a simple platform to understand graph-based feature aggregation, including phenomena such as heterophily, oversmoothing, and feature re-shuffling. Based on the theory, we also identify simple and useful tricks for the node classification task. | Graph neural networks, node classification, feature centroid simplex, coarse geometry | We analyze the geometry of node features to understand graph-independent properties of node classification datasets. | 2,261 | null | [
-0.04449033737182617,
-0.03998415544629097,
0.028518855571746826,
-0.04992613568902016,
0.05108713358640671,
0.07314197719097137,
-0.04953779652714729,
-0.03213363513350487,
-0.04410502314567566,
0.007672022562474012,
-0.026669848710298538,
0.0004357380385044962,
0.023805050179362297,
0.04917977377772331,
-0.0682733878493309,
0.01133712101727724,
0.03550110012292862,
0.0357353575527668,
0.015530307777225971,
0.025959204882383347,
-0.001968185883015394,
-0.006683055777102709,
-0.04350017383694649,
0.057099759578704834,
0.046316567808389664,
0.0858430564403534,
-0.03866266459226608,
0.004036139231175184,
0.01371335331350565,
0.005648771766573191,
0.007607387378811836,
0.04651768133044243,
-0.00884817261248827,
0.0193879846483469,
-0.02848711982369423,
0.03246327489614487,
-0.039021942764520645,
0.027697347104549408,
0.04018978774547577,
-0.01938243955373764,
0.04644773155450821,
0.0288732647895813,
0.08049895614385605,
0.07782547175884247,
0.0849357396364212,
0.11838836967945099,
-0.059081774204969406,
0.022031717002391815,
0.0012056251289322972,
-0.035729240626096725,
-0.01715489849448204,
-0.07390975207090378,
-0.08313775062561035,
-0.0006484478362835944,
0.05612121522426605,
0.026542440056800842,
-0.01053778175264597,
-0.0153106190264225,
0.09320926666259766,
0.01708945818245411,
0.11155908554792404,
-0.07258645445108414,
0.08229882270097733,
0.022447921335697174,
-0.0016869892133399844,
0.01226934976875782,
0.019894663244485855,
-0.005523782689124346,
-0.09455863386392593,
0.016330355778336525,
0.08705547451972961,
0.05711282417178154,
-0.06882968544960022,
-0.013346109539270401,
0.058220140635967255,
0.011480036191642284,
0.015085596591234207,
0.016370171681046486,
-0.03356625884771347,
-0.0007034931913949549,
-0.08485165238380432,
0.057439375668764114,
0.09076245129108429,
0.015318219549953938,
0.018951060250401497,
0.031059125438332558,
-0.05971027538180351,
-0.025960488244891167,
-0.03511999547481537,
-0.013213773258030415,
-0.0823463648557663,
0.04042486473917961,
-0.030808167532086372,
-0.10188297182321548,
0.02727583795785904,
0.036682359874248505,
0.024541327729821205,
-0.023666473105549812,
-0.03799141198396683,
0.049382392317056656,
0.008100388571619987,
-0.028802374377846718,
0.02964046411216259,
-0.1256127953529358,
0.04017346724867821,
0.024803776293992996,
0.011485354043543339,
0.010476775467395782,
0.08736643940210342,
-0.07373448461294174,
-0.042158499360084534,
0.08716066926717758,
-0.04075218364596367,
-0.02240068092942238,
-0.06045466661453247,
-0.02441924624145031,
0.024317435920238495,
-0.014327092096209526,
0.11141827702522278,
0.022828243672847748,
-0.00974022876471281,
-0.0020095142535865307,
-0.034521836787462234,
0.018849270418286324,
0.019774753600358963,
0.08749932795763016,
-0.022783029824495316,
2.6292179289086408e-33,
0.01423400267958641,
-0.004359583370387554,
-0.0006690555601380765,
-0.025334162637591362,
0.030518466606736183,
0.011489922180771828,
-0.06270357966423035,
-0.03905118629336357,
0.054323289543390274,
0.03487124666571617,
-0.07011537998914719,
0.001817942364141345,
0.02802855335175991,
0.04607921093702316,
0.046011295169591904,
-0.0037408301141113043,
0.040600597858428955,
-0.01882116124033928,
-0.01105228066444397,
-0.1689039170742035,
-0.01262289471924305,
-0.0021218829788267612,
0.08333779871463776,
0.015397781506180763,
-0.018750879913568497,
0.00838817935436964,
-0.05111674591898918,
0.0017678453586995602,
-0.05328039824962616,
-0.027941780164837837,
-0.027698172256350517,
0.027673982083797455,
0.02732308767735958,
0.07347521185874939,
-0.051703937351703644,
0.0083308769389987,
0.02219036966562271,
-0.04730893671512604,
-0.005770900286734104,
-0.023547107353806496,
-0.07802481204271317,
0.035454876720905304,
-0.05213957279920578,
-0.09101260453462601,
-0.007199598476290703,
0.0724618136882782,
-0.03753693774342537,
-0.043113041669130325,
0.006026584655046463,
-0.07569441944360733,
-0.04392227903008461,
-0.0029958367813378572,
0.018388042226433754,
0.038327597081661224,
0.0034817217383533716,
0.06439273059368134,
0.02900351583957672,
0.04659869894385338,
-0.02991817332804203,
0.051814787089824677,
0.010332045145332813,
0.01855720579624176,
0.006688107270747423,
-0.01136555802077055,
-0.01070698443800211,
-0.01346229761838913,
0.030027218163013458,
0.07475315779447556,
0.10609854012727737,
0.005714411847293377,
-0.024111593142151833,
0.06717193871736526,
-0.02950569987297058,
-0.016623785719275475,
0.056042809039354324,
0.013374822214245796,
-0.00152644666377455,
-0.12714210152626038,
-0.05389709398150444,
0.042519837617874146,
-0.000009048244464793243,
-0.07643520087003708,
0.022168751806020737,
-0.03386230394244194,
-0.02372276782989502,
-0.00624645734205842,
0.0757216289639473,
-0.003423006273806095,
0.013430391438305378,
-0.005661809351295233,
-0.03589183837175369,
-0.02177847921848297,
0.04526747018098831,
0.03399353101849556,
-0.012325199320912361,
-2.2338277307110946e-33,
-0.09771036356687546,
0.0822390615940094,
-0.012394696474075317,
0.04114539176225662,
0.021753184497356415,
-0.02897815778851509,
-0.018297798931598663,
-0.05278851091861725,
-0.11229656636714935,
0.01200866885483265,
0.0017662362661212683,
-0.049931272864341736,
-0.0662086009979248,
-0.039965320378541946,
-0.00021563249174505472,
-0.005578089039772749,
-0.08549229800701141,
-0.04042985290288925,
-0.024689309298992157,
-0.009220288135111332,
0.04470448195934296,
0.03340829163789749,
-0.06489066034555435,
0.006514100357890129,
0.04277033358812332,
0.032864760607481,
-0.11655550450086594,
-0.0029686582274734974,
0.016258850693702698,
-0.004522743634879589,
-0.023523813113570213,
-0.042783547192811966,
-0.029535725712776184,
-0.0034260901156812906,
0.04961732402443886,
-0.0016155189368873835,
-0.05600103735923767,
-0.02701754681766033,
0.01605810970067978,
0.009615420363843441,
-0.016164105385541916,
-0.03885950520634651,
-0.08197956532239914,
0.036785803735256195,
0.007717246189713478,
0.027646685019135475,
-0.01907949708402157,
0.07447125762701035,
-0.1063082292675972,
0.010584639385342598,
-0.08969996124505997,
0.031888511031866074,
0.00805235281586647,
0.0496838241815567,
-0.002430389868095517,
-0.0015610731206834316,
0.04311753436923027,
0.06766077131032944,
0.04808804392814636,
-0.04898406192660332,
-0.036241501569747925,
-0.09946160763502121,
-0.056554507464170456,
0.049586664885282516,
0.012283065356314182,
0.041533421725034714,
-0.0021993278060108423,
-0.06312332302331924,
-0.054605159908533096,
0.1114373728632927,
0.03561191260814667,
0.0433410108089447,
-0.05504843592643738,
-0.0027535660192370415,
-0.07118932902812958,
0.001164515153504908,
0.03237440064549446,
-0.016783347353339195,
-0.03048618882894516,
-0.01817629113793373,
0.03602023422718048,
-0.016774822026491165,
0.010773013345897198,
0.10278946161270142,
0.11260062456130981,
0.03943423926830292,
0.049342043697834015,
0.05547843128442764,
0.08134424686431885,
0.016544600948691368,
-0.014181473292410374,
0.03473203256726265,
-0.0739051029086113,
0.03794063627719879,
-0.018085425719618797,
-4.1108243920007226e-8,
-0.22413182258605957,
-0.04643131047487259,
-0.006441314704716206,
-0.02614404819905758,
0.09152780473232269,
-0.014970466494560242,
0.11134602129459381,
0.15907785296440125,
-0.06757304817438126,
0.04009844735264778,
0.04750414937734604,
-0.042374588549137115,
-0.06565511226654053,
-0.02496633678674698,
0.014180734753608704,
0.046829115599393845,
0.02753695845603943,
0.004388187546283007,
0.06817230582237244,
-0.002350977621972561,
-0.04889998957514763,
-0.03954964876174927,
0.012272537685930729,
0.006653303746134043,
-0.03894574195146561,
-0.13057072460651398,
-0.022176086902618408,
0.1483696699142456,
0.12304418534040451,
-0.019800972193479538,
-0.04477278143167496,
0.007935534231364727,
0.024814337491989136,
0.07320430874824524,
-0.02108137309551239,
0.10152575373649597,
0.01671161875128746,
0.030996734276413918,
-0.03729483485221863,
-0.014393013902008533,
-0.0449347123503685,
0.08701936900615692,
0.008634977042675018,
-0.03980569913983345,
-0.031078115105628967,
0.02940460667014122,
0.07631345838308334,
-0.012459969148039818,
0.06240248680114746,
-0.029165878891944885,
-0.030096154659986496,
-0.057281818240880966,
-0.0154305100440979,
0.0206109881401062,
-0.057067278772592545,
0.029454093426465988,
-0.03582947328686714,
-0.07764069736003876,
0.046065554022789,
0.05321202799677849,
-0.032081592828035355,
0.05229486897587776,
-0.02928996831178665,
-0.07470106333494186
] |
KiVA: Kid-inspired Visual Analogies for Testing Large Multimodal Models | https://openreview.net/forum?id=vNATZfmY6R | [
"Eunice Yiu",
"Maan Qraitem",
"Anisa Noor Majhi",
"Charlie Wong",
"Yutong Bai",
"Shiry Ginosar",
"Alison Gopnik",
"Kate Saenko"
] | Poster | This paper investigates visual analogical reasoning in large multimodal models (LMMs) compared to human adults and children. A “visual analogy” is an abstract rule inferred from one image and applied to another. While benchmarks exist for testing visual reasoning in LMMs, they require advanced skills and omit basic visual analogies that even young children can make. Inspired by developmental psychology, we propose a new benchmark of 4,300 visual transformations of everyday objects to test LMMs on visual analogical reasoning and compare them to children (ages three to five) and to adults. We structure the evaluation into three stages: identifying what changed (e.g., color, number, etc.), how it changed (e.g., added one object), and applying the rule to new scenarios. Our findings show that while GPT-o1, GPT-4V, LLaVA-1.5, and MANTIS identify the “what” effectively, they struggle with quantifying the “how” and extrapolating this rule to new objects. In contrast, children and adults exhibit much stronger analogical reasoning at all three stages. Additionally, the strongest tested model, GPT-o1, performs better in tasks involving simple surface-level visual attributes like color and size, correlating with quicker human adult response times. Conversely, more complex tasks such as number, rotation, and reflection, which necessitate extensive cognitive processing and understanding of extrinsic spatial properties in the physical world, present more significant challenges. Altogether, these findings highlight the limitations of training models on data that primarily consists of 2D images and text. | large multimodal models, analogical reasoning, cognition, developmental psychology | We present a benchmark that closes a critical gap in current benchmarks for foundational models - visual analogical reasoning, which even young children can do but models perform poorly in. | 2,256 | 2407.17773 | [
0.06806370615959167,
-0.06463116407394409,
0.050549302250146866,
-0.009809741750359535,
0.11720089614391327,
-0.057067494839429855,
-0.055199019610881805,
0.025123516097664833,
0.06417948752641678,
0.03027952089905739,
-0.034275203943252563,
-0.09106042981147766,
-0.016916025429964066,
0.10638350248336792,
0.038653962314128876,
-0.00012384256115183234,
0.05301184579730034,
0.06552031636238098,
-0.0849531888961792,
-0.04572845250368118,
0.04908394441008568,
-0.05720609799027443,
0.04819473624229431,
-0.06558382511138916,
0.030013982206583023,
0.06865310668945312,
-0.008068843744695187,
-0.08528764545917511,
0.015199623070657253,
-0.0683475211262703,
0.03595419228076935,
0.010130726732313633,
-0.030358754098415375,
-0.04182717576622963,
-0.03487521409988403,
0.08053925633430481,
0.00388946570456028,
0.06837479025125504,
-0.02186846360564232,
-0.00779879093170166,
-0.05490925535559654,
-0.000041258157580159605,
-0.03083365596830845,
-0.05651526525616646,
0.061449967324733734,
-0.04145970940589905,
-0.00852837972342968,
-0.04742555692791939,
0.002656749449670315,
-0.014594470150768757,
-0.10798390954732895,
-0.07855518907308578,
-0.06753849238157272,
-0.002925173146650195,
0.00706250686198473,
0.06588893383741379,
0.058588165789842606,
-0.02946586348116398,
0.07073938101530075,
-0.04255044460296631,
-0.02115926332771778,
-0.042606547474861145,
-0.07226573675870895,
0.03552884981036186,
-0.005707466509193182,
0.02507859840989113,
-0.022660989314317703,
-0.027385877445340157,
0.020635128021240234,
0.012789378874003887,
-0.013948909938335419,
0.09998884797096252,
0.024257594719529152,
0.0436396487057209,
0.03034033626317978,
0.06790577620267868,
0.01066515687853098,
-0.0372619703412056,
0.06184810772538185,
-0.12431778013706207,
0.02231261134147644,
-0.007341460790485144,
-0.07827611267566681,
-0.053212955594062805,
0.12681053578853607,
0.07070201635360718,
-0.02186797931790352,
0.03423638269305229,
-0.04069678857922554,
-0.0013361811870709062,
-0.0033915042877197266,
-0.020497849211096764,
0.011917341500520706,
0.03114192932844162,
0.057246219366788864,
-0.00969784613698721,
0.005932803265750408,
-0.029600104317069054,
-0.018480923026800156,
0.027758687734603882,
0.04954306408762932,
0.049808792769908905,
0.018870145082473755,
0.04142920672893524,
0.04101995751261711,
-0.06608656048774719,
0.06959220767021179,
-0.03583855926990509,
0.05486069619655609,
-0.033821575343608856,
-0.022625653073191643,
0.03081057034432888,
0.028514942154288292,
0.04280754551291466,
-0.037928540259599686,
-0.05938977375626564,
0.04140164330601692,
0.0021376395598053932,
-0.007931600324809551,
-0.045104846358299255,
0.10836361348628998,
0.008901955559849739,
0.03144842013716698,
-0.006774564739316702,
0.06375277042388916,
0.00016553758177906275,
-0.06559797376394272,
7.097555452560795e-34,
0.03925995156168938,
0.01623438112437725,
0.11032211035490036,
0.12302908301353455,
0.07899502664804459,
-0.0007594986818730831,
-0.04503301903605461,
0.014670480042696,
0.021744834259152412,
0.029315456748008728,
-0.027797462418675423,
0.022937534376978874,
-0.10301651060581207,
0.116508848965168,
0.01274158526211977,
0.04943753033876419,
-0.1373095065355301,
0.003164420137181878,
-0.06462457031011581,
0.044933024793863297,
-0.06669783592224121,
-0.006934137549251318,
-0.05458836629986763,
0.048193372786045074,
-0.036078598350286484,
0.0610516332089901,
0.08554305136203766,
-0.052711229771375656,
-0.021403806284070015,
-0.01756836101412773,
-0.10866828262805939,
0.04128803312778473,
-0.0761566013097763,
0.0550989955663681,
-0.042794134467840195,
-0.12347134202718735,
0.07993382215499878,
-0.09795920550823212,
0.06180699169635773,
0.008031252771615982,
-0.015899159014225006,
-0.03697266802191734,
-0.07391314953565598,
0.04676627367734909,
0.06360960751771927,
0.03910575807094574,
-0.012544941157102585,
0.05728308483958244,
-0.09526751935482025,
0.0032405618112534285,
0.009229558520019054,
0.059213317930698395,
-0.042209722101688385,
-0.09610027074813843,
-0.09677287936210632,
0.015548287890851498,
0.0005366632831282914,
0.000329706264892593,
0.032786380499601364,
-0.046226199716329575,
0.049168288707733154,
0.06749750673770905,
0.028915531933307648,
0.06199882552027702,
-0.02130190283060074,
0.07251032441854477,
0.009205352514982224,
0.006265496835112572,
0.054749444127082825,
0.06421386450529099,
-0.03534513711929321,
0.05412669852375984,
-0.09075389802455902,
-0.07617117464542389,
0.02482479251921177,
-0.08351501077413559,
0.04205593094229698,
-0.017877956852316856,
-0.011549443937838078,
0.008541250601410866,
0.029629318043589592,
0.02165616676211357,
-0.025359265506267548,
-0.05176054313778877,
-0.04264816269278526,
-0.0754375234246254,
0.002288221614435315,
-0.15154728293418884,
-0.04377257451415062,
-0.09285280853509903,
0.014796258881688118,
-0.05588659271597862,
-0.03482403606176376,
-0.04566611349582672,
-0.003701221663504839,
-1.3353876638150275e-33,
-0.02636427991092205,
0.0004757018177770078,
0.010137634351849556,
0.07682584971189499,
-0.0345316007733345,
-0.09445753693580627,
0.04686393961310387,
0.015281397849321365,
0.006095776334404945,
-0.06966875493526459,
0.022392060607671738,
-0.016615254804491997,
-0.04972860589623451,
-0.013481014408171177,
0.014369888231158257,
-0.03816692903637886,
0.02213594876229763,
-0.01259738951921463,
-0.04144721478223801,
-0.03567510470747948,
0.044076643884181976,
0.1131042018532753,
-0.09887072443962097,
-0.0053696343675255775,
-0.06461414694786072,
0.07683856040239334,
-0.03986295312643051,
-0.027442632243037224,
-0.011515761725604534,
-0.016301529482007027,
0.029289545491337776,
-0.029700033366680145,
0.07008857280015945,
0.031369615346193314,
0.02351832389831543,
-0.0005450393073260784,
0.0697532519698143,
-0.04741976037621498,
-0.04938472434878349,
0.08349364995956421,
-0.02219104953110218,
-0.03602265194058418,
0.015727099031209946,
-0.013221624307334423,
-0.05318481847643852,
-0.022395284846425056,
-0.015541664324700832,
-0.02055874466896057,
-0.004417008254677057,
0.05263663828372955,
0.02985675446689129,
-0.005542263854295015,
-0.08943004906177521,
-0.06040751188993454,
-0.014645252376794815,
-0.028997404500842094,
-0.015287064015865326,
-0.06671688705682755,
0.06909135729074478,
0.09364130347967148,
-0.019172964617609978,
-0.04876201972365379,
-0.04555193707346916,
0.019674399867653847,
-0.08640391379594803,
0.031465981155633926,
-0.06072058156132698,
-0.02716217190027237,
0.032939765602350235,
0.023757541552186012,
-0.024868931621313095,
0.023072080686688423,
0.006321946159005165,
-0.011996250599622726,
0.0756097212433815,
0.011080666445195675,
-0.014202321879565716,
0.06218576803803444,
0.023525705561041832,
-0.03179335966706276,
-0.022471511736512184,
-0.12431452423334122,
0.04482926055788994,
0.03286334127187729,
-0.030787594616413116,
0.025195114314556122,
0.005625534802675247,
0.04918703809380531,
0.03155834600329399,
0.06506386399269104,
-0.056562699377536774,
0.004886820446699858,
-0.02739136666059494,
0.09430564194917679,
-0.03322631120681763,
-3.981136131869789e-8,
0.028193287551403046,
0.050114262849092484,
0.05777046084403992,
0.02954503335058689,
0.017577996477484703,
-0.021270403638482094,
-0.08125762641429901,
0.04342927783727646,
-0.02661019377410412,
0.0016818061703816056,
-0.01978944055736065,
0.014393007382750511,
-0.008134103380143642,
0.05548980087041855,
0.05822437256574631,
0.003313004272058606,
0.0031977430917322636,
-0.02235700562596321,
-0.004632991272956133,
0.025508126243948936,
0.0803489163517952,
0.061620794236660004,
-0.02158794365823269,
0.014792095869779587,
0.0015885810134932399,
-0.06113401800394058,
-0.0825401246547699,
0.09252604097127914,
0.011970742605626583,
0.04247559979557991,
0.0433945432305336,
0.04839619621634483,
-0.01282653957605362,
-0.011506214737892151,
0.005125090945512056,
-0.05417841672897339,
-0.028073750436306,
0.04746585339307785,
0.018502408638596535,
0.011658875271677971,
0.0031404143664985895,
0.0043840897269546986,
-0.009567629545927048,
0.06097846478223801,
0.04415806010365486,
-0.018285363912582397,
-0.04824669659137726,
-0.10506608337163925,
-0.002200136426836252,
0.07786412537097931,
-0.06818722188472748,
-0.009975539520382881,
-0.022243518382310867,
0.0649898573756218,
0.018294721841812134,
-0.05341354385018349,
0.008213397115468979,
0.05351465195417404,
-0.018301183357834816,
0.0616907924413681,
0.06316138058900833,
0.07447276264429092,
-0.05156383663415909,
0.03487842530012131
] |
Learning from weak labelers as constraints | https://openreview.net/forum?id=2BtFKEeMGo | [
"Vishwajeet Agrawal",
"Rattana Pukdee",
"Maria Florina Balcan",
"Pradeep Kumar Ravikumar"
] | Poster | We study programmatic weak supervision, where in contrast to labeled data, we have access to \emph{weak labelers}, each of which either abstains or provides noisy labels corresponding to any input. Most previous approaches typically employ latent generative models that model the joint distribution of the weak labels and the latent ``true'' label. The caveats are that this relies on assumptions that may not always hold in practice such as conditional independence assumptions over the joint distribution of the weak labelers and the latent true label, and more general implicit inductive biases in the latent generative models. In this work, we consider a more explicit form of side-information that can be leveraged to denoise the weak labeler, namely the bounds on the average error of the weak labelers. We then propose a novel but natural weak supervision objective that minimizes a regularization functional subject to satisfying these bounds. This turns out to be a difficult constrained optimization problem due to discontinuous accuracy bound constraints. We provide a continuous optimization formulation for this objective through an alternating minimization algorithm that iteratively computes soft pseudo labels on the unlabeled data satisfying the constraints while being close to the model, and then updates the model on these labels until all the constraints are satisfied. We follow this with a theoretical analysis of this approach and provide insights into its denoising effects in training discriminative models given multiple weak labelers. Finally, we demonstrate the superior performance and robustness of our method on a popular weak supervision benchmark. | unsupervised learning, weak supervision, learning theory | A new method for learning from weak labelers given information about their average errors. | 2,254 | null | [
-0.06786959618330002,
-0.035005733370780945,
-0.0009192521683871746,
0.05461771786212921,
0.07782945036888123,
0.05489015579223633,
0.03552230820059776,
-0.04314855858683586,
-0.08091399073600769,
-0.061487384140491486,
-0.018265875056385994,
-0.05271691456437111,
0.062430489808321,
0.06414546817541122,
-0.027023082599043846,
0.05023743584752083,
0.17105942964553833,
0.07800999283790588,
-0.09745804220438004,
-0.02678714506328106,
-0.06878611445426941,
0.08681086450815201,
-0.008718293160200119,
0.05206766352057457,
0.048173028975725174,
0.034641146659851074,
0.003024265170097351,
-0.013747787103056908,
0.0369461365044117,
0.013395117595791817,
-0.021363606676459312,
-0.03859862685203552,
-0.0642026886343956,
-0.0003956093278247863,
0.031118538230657578,
0.08311959356069565,
-0.053498681634664536,
0.06898920983076096,
0.04254526272416115,
0.029170941561460495,
0.035110145807266235,
0.09030725806951523,
-0.03472669795155525,
-0.013251439668238163,
-0.03595007210969925,
-0.00807218812406063,
0.010193298570811749,
-0.043670158833265305,
-0.06519202142953873,
0.04961216077208519,
0.009736121632158756,
-0.008801082149147987,
-0.0015130442334339023,
0.05169731751084328,
-0.028423257172107697,
-0.017675789073109627,
0.04409439489245415,
-0.05147832632064819,
0.025819584727287292,
0.08258488774299622,
-0.021668754518032074,
-0.08706128597259521,
-0.037693411111831665,
-0.08672401309013367,
-0.004546518437564373,
-0.10420525819063187,
0.0198002178221941,
0.005010268650949001,
-0.06878937035799026,
0.0881354957818985,
-0.008349441923201084,
0.06715095043182373,
-0.09321478754281998,
0.02030676230788231,
0.021174054592847824,
0.012800226919353008,
0.06213802844285965,
-0.026527855545282364,
0.038060881197452545,
-0.0664379820227623,
0.021116307005286217,
0.048237476497888565,
0.037473924458026886,
0.014658762142062187,
0.07934174686670303,
-0.02812051959335804,
0.0014726925874128938,
0.051160264760255814,
0.04936445131897926,
0.05993945524096489,
-0.03788193315267563,
-0.0014710515970364213,
0.007246313150972128,
0.05480144917964935,
0.03409950062632561,
-0.0023576945532113314,
0.032331641763448715,
-0.05077633634209633,
-0.019970448687672615,
0.11982744187116623,
-0.045814260840415955,
0.06744761019945145,
0.035639725625514984,
-0.07023972272872925,
0.03932224214076996,
-0.03578433766961098,
0.015406233258545399,
-0.02240641973912716,
0.006413760595023632,
-0.0888596847653389,
0.018220169469714165,
0.03442241623997688,
0.015638912096619606,
-0.01061573438346386,
-0.0018766946159303188,
-0.05269933119416237,
0.047517381608486176,
-0.0349600687623024,
0.0012527607614174485,
-0.00987254362553358,
-0.08531565219163895,
-0.03546479344367981,
0.03697643801569939,
0.03296961262822151,
0.0737544447183609,
-0.02770288661122322,
-0.06804294884204865,
8.648596830419283e-34,
0.043086059391498566,
0.027381066232919693,
0.004246970172971487,
-0.008749419823288918,
-0.01118814293295145,
0.031971532851457596,
-0.04964013397693634,
0.015608418732881546,
-0.024604996666312218,
0.0033734836615622044,
0.09270775318145752,
-0.019995318725705147,
-0.09912879019975662,
0.02695532888174057,
0.039220694452524185,
-0.05723981186747551,
0.019866297021508217,
0.07350421696901321,
-0.018839223310351372,
-0.07585334032773972,
-0.06613495945930481,
0.027053112164139748,
-0.05428120866417885,
-0.034936510026454926,
0.006198117043823004,
-0.027835257351398468,
0.04751996695995331,
-0.04822569712996483,
0.007248696871101856,
-0.02255236543715,
-0.1260872483253479,
-0.07538098096847534,
0.13939736783504486,
0.042699556797742844,
0.03231935575604439,
-0.009432200342416763,
-0.00778121268376708,
-0.027517568320035934,
-0.00011865091073559597,
0.013889292255043983,
0.01267766673117876,
0.08568769693374634,
0.0356561504304409,
-0.10135349631309509,
-0.02786458283662796,
-0.05526994168758392,
0.0014051354955881834,
-0.0028800282161682844,
0.015162918716669083,
-0.0871482565999031,
0.012839028611779213,
0.03407027944922447,
-0.008518502116203308,
-0.06400036066770554,
-0.05472172796726227,
0.009407829493284225,
-0.02958240546286106,
0.019517814740538597,
0.01818970963358879,
-0.006716582924127579,
-0.01956246793270111,
0.06638006120920181,
0.02288764901459217,
0.027621738612651825,
0.015838872641324997,
0.049938756972551346,
-0.014771447516977787,
-0.042238667607307434,
0.09711305797100067,
-0.06522440910339355,
-0.038356680423021317,
-0.0012382151326164603,
-0.006996551062911749,
-0.08270126581192017,
0.10058014839887619,
0.031008141115307808,
0.02587537094950676,
-0.05942551791667938,
0.047103118151426315,
0.01859131269156933,
-0.03166583180427551,
0.013311652466654778,
0.0040122028440237045,
0.005146614275872707,
-0.06797245144844055,
-0.03511783108115196,
0.03396493196487427,
-0.03488341346383095,
-0.01023910939693451,
0.0021583507768809795,
-0.09863580018281937,
-0.07818436622619629,
-0.093113973736763,
0.02101089246571064,
0.0074189892038702965,
-9.167811766822526e-34,
-0.01283235289156437,
0.03510013222694397,
0.005174708086997271,
-0.005128534510731697,
-0.00268132658675313,
-0.0712021216750145,
-0.021849513053894043,
0.050686806440353394,
0.004314248915761709,
-0.038560930639505386,
0.013839984312653542,
-0.04001335799694061,
-0.02086380496621132,
0.01705028861761093,
-0.023188192397356033,
0.03770012408494949,
-0.004170040134340525,
0.07472152262926102,
0.006439553573727608,
0.09442123770713806,
0.05529475212097168,
0.11542303115129471,
-0.0779733657836914,
0.11252719163894653,
0.024235950782895088,
0.03404969349503517,
0.014959363266825676,
0.1272878795862198,
-0.001839891541749239,
-0.008288450539112091,
-0.03812761977314949,
-0.05092015862464905,
-0.030214350670576096,
-0.07931535691022873,
-0.1279873251914978,
0.015560723841190338,
-0.028659066185355186,
0.07135698944330215,
-0.0698995515704155,
0.04030437394976616,
0.06331998109817505,
0.024288691580295563,
-0.1255752295255661,
0.040421392768621445,
-0.03277171403169632,
-0.013957537710666656,
-0.05122124403715134,
-0.05302693322300911,
0.0033746324479579926,
0.04169052839279175,
-0.046697333455085754,
-0.013681622222065926,
-0.05468016490340233,
0.08115346729755402,
-0.09143955260515213,
-0.15019871294498444,
0.024373380467295647,
-0.022073183208703995,
0.09265643358230591,
0.015123250894248486,
-0.07529071718454361,
0.039320241659879684,
-0.0020919900853186846,
-0.021379077807068825,
0.01813296601176262,
-0.01889447681605816,
-0.014558956027030945,
-0.055048517882823944,
0.04132002964615822,
0.025056948885321617,
0.0999208465218544,
0.06361827999353409,
0.0033805924467742443,
0.07328254729509354,
0.007588075939565897,
0.04610488936305046,
0.02349737100303173,
-0.055912576615810394,
-0.0025132668670266867,
-0.02390068955719471,
0.10369671881198883,
-0.011272060684859753,
0.034417491406202316,
0.026969708502292633,
0.0189273152500391,
-0.00006554143328685313,
0.03497409448027611,
0.019874459132552147,
0.027059314772486687,
-0.04299568012356758,
0.009142443537712097,
-0.05602717027068138,
0.043286312371492386,
0.05023030564188957,
-0.06524604558944702,
-4.197103820047232e-8,
-0.07985972613096237,
0.014832684770226479,
-0.01683521270751953,
-0.023707019165158272,
0.05179325491189957,
-0.05259731039404869,
0.0038676050025969744,
0.028409449383616447,
0.02338506653904915,
-0.04148183390498161,
0.0436851792037487,
-0.0533146858215332,
-0.03321696072816849,
-0.0818268433213234,
-0.02303500659763813,
0.07644228637218475,
0.014892534352838993,
0.018748199567198753,
0.03527555987238884,
0.048965759575366974,
-0.06018815562129021,
0.04003768414258957,
0.005586664658039808,
0.001954189967364073,
0.06863654404878616,
-0.07504623383283615,
0.05039755627512932,
0.05756800249218941,
0.07711915671825409,
0.026218395680189133,
0.013244460336863995,
0.017647715285420418,
0.06695712357759476,
0.008687123656272888,
-0.053125541657209396,
0.09850326925516129,
-0.019120408222079277,
-0.06671319901943207,
-0.08483140915632248,
-0.03917211294174194,
-0.014447633177042007,
0.09883233159780502,
-0.059232715517282486,
-0.06605156511068344,
0.04555428400635719,
0.024573972448706627,
0.04057617485523224,
-0.10042908042669296,
0.05790412798523903,
-0.04998837783932686,
0.02997850626707077,
-0.06654103845357895,
0.089008629322052,
-0.016216235235333443,
0.02243032678961754,
-0.002791299019008875,
-0.0006924336776137352,
-0.0021788799203932285,
-0.0428294837474823,
0.04251649230718613,
-0.008266087621450424,
0.034477248787879944,
-0.005576181225478649,
0.043823473155498505
] |
Bisimulation Metric for Model Predictive Control | https://openreview.net/forum?id=F07ic7huE3 | [
"Yutaka Shimizu",
"Masayoshi Tomizuka"
] | Poster | Model-based reinforcement learning (MBRL) has shown promise for improving sample efficiency and decision-making in complex environments. However, existing methods face challenges in training stability, robustness to noise, and computational efficiency. In this paper, we propose Bisimulation Metric for Model Predictive Control (BS-MPC), a novel approach that incorporates bisimulation metric loss in its objective function to directly optimize the encoder. This optimization enables the learned encoder to extract intrinsic information from the original state space while discarding irrelevant details. BS-MPC improves training stability, robustness against input noise, and computational efficiency by reducing training time. We evaluate BS-MPC on both continuous control and image-based tasks from the DeepMind Control Suite, demonstrating superior performance and robustness compared to state-of-the-art baseline methods. | Reinforcement Learning, Model-based reinforcement learning, optimal control, MPC | A new model-based reinforcement learning algorithm using bisimulation metric | 2,251 | 2410.04553 | [
-0.047755759209394455,
-0.09449029713869095,
0.00940035842359066,
0.060963623225688934,
0.004890758544206619,
0.05283179506659508,
0.0181833878159523,
-0.041749898344278336,
-0.01854427345097065,
0.04406877979636192,
-0.06941945850849152,
-0.05929020047187805,
0.030037663877010345,
0.040536221116781235,
-0.008475909009575844,
0.06789901852607727,
0.0988377183675766,
0.021450195461511612,
-0.0884200856089592,
-0.0621224045753479,
0.021674027666449547,
-0.0035780970938503742,
0.06951924413442612,
0.043061576783657074,
0.013356328941881657,
0.01608274318277836,
-0.021116815507411957,
-0.0396069772541523,
0.03626265376806259,
-0.03394988551735878,
0.047492798417806625,
-0.06566398590803146,
0.09799084067344666,
-0.10468582808971405,
-0.10125821828842163,
0.045186981558799744,
-0.06801021844148636,
-0.013408323749899864,
0.01701202243566513,
-0.05136728286743164,
-0.027173563838005066,
0.05204367637634277,
0.029595430940389633,
-0.005846434272825718,
0.055736083537340164,
0.07966548949480057,
0.04522814229130745,
-0.0337783619761467,
0.004673362709581852,
-0.023539021611213684,
-0.011639758013188839,
-0.04042043909430504,
0.054163314402103424,
-0.00493737543001771,
0.018306612968444824,
-0.0014140864368528128,
0.08757040649652481,
0.04665202647447586,
0.004070595372468233,
0.008165485225617886,
-0.006173097528517246,
-0.059048332273960114,
-0.04950130730867386,
-0.039409808814525604,
-0.06017722561955452,
0.022163553163409233,
0.017042411491274834,
0.029144620522856712,
0.047152694314718246,
-0.05057888105511665,
-0.015787070617079735,
0.050672922283411026,
0.0026386710815131664,
-0.029066160321235657,
0.03579314053058624,
-0.017096977680921555,
0.090421162545681,
-0.02636873722076416,
0.07576901465654373,
-0.04160081595182419,
0.03697037696838379,
-0.009258376434445381,
-0.055725958198308945,
-0.06458461284637451,
0.11090630292892456,
-0.01892223209142685,
0.024627964943647385,
0.06862398236989975,
0.08341341465711594,
-0.0177407655864954,
-0.02872476354241371,
-0.036225296556949615,
-0.002191626699641347,
-0.08648159354925156,
-0.02746376022696495,
-0.013012015260756016,
-0.06084946542978287,
-0.06492611020803452,
-0.06894119828939438,
0.022543560713529587,
-0.011659720912575722,
0.0282389298081398,
-0.053383391350507736,
0.016099732369184494,
0.07572953402996063,
-0.026605557650327682,
0.11857371777296066,
0.06014331802725792,
0.02781052701175213,
-0.034950509667396545,
0.034293774515390396,
-0.009010748006403446,
0.07911796867847443,
0.0899224653840065,
0.0159893911331892,
0.036011602729558945,
-0.027011718600988388,
-0.011499006301164627,
-0.022675639018416405,
0.05781938135623932,
-0.0594894103705883,
-0.05593812093138695,
-0.04789665341377258,
-0.0015154796419665217,
-0.013886001892387867,
-0.01646514981985092,
-0.02005748823285103,
1.761355776155323e-33,
0.023610884323716164,
-0.018357865512371063,
0.10897745937108994,
0.01855940744280815,
0.0021862543653696775,
0.03427618741989136,
-0.016207881271839142,
-0.023924943059682846,
-0.036082372069358826,
0.05035325512290001,
0.014270785264670849,
-0.06797146052122116,
-0.049789462238550186,
0.18700280785560608,
0.07977206259965897,
0.021290387958288193,
-0.09005412459373474,
0.0442858524620533,
0.009927093982696533,
-0.007128188852220774,
0.04929705336689949,
-0.023470740765333176,
-0.1087903082370758,
-0.08110535144805908,
-0.015412554144859314,
0.06436548382043839,
0.03250899538397789,
0.0929616168141365,
0.014937463216483593,
-0.019738975912332535,
-0.05475681275129318,
0.04215028136968613,
0.021601451560854912,
0.03252467140555382,
0.05646737664937973,
0.052523642778396606,
0.03244689851999283,
0.006682652048766613,
-0.06231369078159332,
0.02009613998234272,
0.04132169485092163,
0.054811667650938034,
0.02774597518146038,
-0.005994537845253944,
-0.04119528457522392,
-0.08738750219345093,
-0.011043391190469265,
-0.013145479373633862,
-0.08586856722831726,
0.035669200122356415,
0.06293235719203949,
0.03607628121972084,
-0.062368545681238174,
-0.07671062648296356,
-0.036832667887210846,
-0.05545257776975632,
0.04301391914486885,
0.025302670896053314,
-0.013973399065434933,
0.08438794314861298,
0.00494588864967227,
-0.007445118390023708,
0.01381420437246561,
0.006143625825643539,
0.03200540691614151,
0.044211119413375854,
-0.00022431502293329686,
0.08810838311910629,
0.0005523185827769339,
-0.020177066326141357,
-0.014979437924921513,
0.008176860399544239,
0.04030770808458328,
-0.0571434460580349,
0.05358467996120453,
-0.018466312438249588,
0.020308595150709152,
-0.09430186450481415,
-0.05006391927599907,
0.07657791674137115,
0.0073270974680781364,
0.06344632804393768,
-0.06044046953320503,
0.0020002550445497036,
-0.026123128831386566,
-0.05040040984749794,
0.04649869725108147,
-0.01939813233911991,
-0.05406464636325836,
0.01823258213698864,
-0.09722722321748734,
-0.00928429327905178,
-0.005270874593406916,
0.07340356707572937,
-0.07068059593439102,
-1.4747059685775388e-33,
0.06576889008283615,
0.05810469016432762,
0.011836695484817028,
0.020386043936014175,
-0.050180237740278244,
-0.044885799288749695,
0.023154083639383316,
0.010328443720936775,
-0.02137725055217743,
-0.05011105537414551,
0.011375746689736843,
-0.08214729279279709,
0.06144929304718971,
-0.00462843244895339,
0.025037050247192383,
-0.04245281592011452,
-0.07685791701078415,
-0.06255225837230682,
-0.025965673848986626,
0.029110662639141083,
0.06356287747621536,
0.052790287882089615,
-0.025572488084435463,
0.05140078067779541,
-0.049319956451654434,
0.03441890701651573,
-0.02896856516599655,
0.08750154823064804,
0.060593973845243454,
0.003306707600131631,
-0.05345045402646065,
-0.05442565307021141,
-0.03726356476545334,
-0.04937603697180748,
-0.06538495421409607,
0.04444565623998642,
0.05458880588412285,
0.009740361012518406,
-0.06429268419742584,
0.07399805635213852,
-0.003820420242846012,
0.014802795834839344,
-0.08160177618265152,
0.10018197447061539,
0.04313667118549347,
-0.03928293287754059,
-0.06060035154223442,
-0.00977497361600399,
0.04981614649295807,
0.032900042831897736,
-0.08134527504444122,
0.019128024578094482,
-0.09596068412065506,
-0.008448140695691109,
-0.006226971745491028,
-0.09213124215602875,
-0.07582928240299225,
0.01871676556766033,
0.07792983949184418,
0.009201939217746258,
-0.06325600296258926,
0.011543536558747292,
-0.022674616426229477,
0.016019312664866447,
-0.004614893812686205,
0.0708756297826767,
-0.04253431409597397,
0.03763383999466896,
0.09742838889360428,
-0.003440716303884983,
-0.019781410694122314,
0.014671201817691326,
0.0753089115023613,
0.014493297785520554,
0.007889371365308762,
-0.023083647713065147,
0.0026494418270885944,
-0.03995734080672264,
0.0013605679851025343,
0.026133447885513306,
0.01573026366531849,
0.021916350349783897,
0.03799721226096153,
0.007151701487600803,
0.02604367770254612,
0.07898757606744766,
-0.05362124368548393,
0.07935282588005066,
0.05270668864250183,
-0.067144975066185,
-0.10073583573102951,
-0.02761261537671089,
0.07887504249811172,
0.01615280844271183,
-0.027040330693125725,
-4.496963157407663e-8,
-0.09800158441066742,
0.05690630152821541,
0.07246147841215134,
-0.007488407660275698,
-0.06632449477910995,
0.003502253210172057,
-0.041876696050167084,
-0.07471512258052826,
0.04051338881254196,
0.032560739666223526,
0.10776122659444809,
0.013242566026747227,
-0.06941099464893341,
-0.04002134129405022,
-0.026607580482959747,
0.06961090862751007,
0.056622084230184555,
0.008057878352701664,
0.08461743593215942,
-0.001306031714193523,
0.006380475126206875,
-0.0558825358748436,
-0.025481805205345154,
-0.09551312029361725,
-0.0026978706009685993,
-0.082149937748909,
-0.04413282871246338,
0.08657079190015793,
0.021880943328142166,
0.010286821983754635,
0.062354933470487595,
-0.026536637917160988,
0.14803293347358704,
-0.010824199765920639,
0.07024627923965454,
-0.009080922231078148,
-0.016761651262640953,
-0.0008332894649356604,
0.051793988794088364,
0.01889948733150959,
-0.013469120487570763,
0.07782255113124847,
-0.07161557674407959,
-0.039225999265909195,
-0.039651211351156235,
0.05549110099673271,
0.03628592565655708,
-0.02420668490231037,
-0.003501809434965253,
-0.0081222765147686,
0.049731913954019547,
-0.0434403195977211,
0.020602881908416748,
0.04349198564887047,
0.08861839026212692,
0.004363100975751877,
0.006518231239169836,
-0.06662105768918991,
0.032730404287576675,
0.09364350885152817,
-0.07419411838054657,
-0.005925868172198534,
-0.09627863764762878,
-0.0031063908245414495
] |
Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved Sampling | https://openreview.net/forum?id=EgJhwYR2tB | [
"Wenda Xu",
"Rujun Han",
"Zifeng Wang",
"Long Le",
"Dhruv Madeka",
"Lei Li",
"William Yang Wang",
"Rishabh Agarwal",
"Chen-Yu Lee",
"Tomas Pfister"
] | Poster | Recent advances in knowledge distillation (KD) have enabled smaller student models to approach the performance of larger teacher models. However, popular methods such as supervised KD and on-policy KD, are adversely impacted by the knowledge gaps between teacher-student in practical scenarios. Supervised KD suffers from a distribution mismatch between training with a static dataset and inference over final student-generated outputs. Conversely, on-policy KD, which uses student-generated samples for training, can suffer from low-quality training examples with which teacher models are not familiar, resulting in inaccurate teacher feedback. To address these limitations, we introduce Speculative Knowledge Distillation (SKD), a novel approach that leverages cooperation between student and teacher models to generate high-quality training data on-the-fly while aligning with the student's inference-time distribution. In SKD, the student proposes tokens, and the teacher replaces poorly ranked ones based on its own distribution, transferring high-quality knowledge adaptively. We evaluate SKD on various text generation tasks, including translation, summarization, math, and instruction following, and show that SKD consistently outperforms existing KD methods across different domains, data sizes, and model initialization strategies. | LLM, Knowledge Distillation, On-policy | Knowledge distillation for language model | 2,249 | 2410.11325 | [
-0.09798749536275864,
-0.07948822528123856,
0.08387026935815811,
0.03087991289794445,
0.04614714905619621,
-0.039821844547986984,
-0.00180371617898345,
-0.01574183814227581,
0.035615358501672745,
0.036159541457891464,
0.01945541799068451,
0.037361033260822296,
0.049087103456258774,
-0.005892185028642416,
-0.0017107195453718305,
0.027205940335989,
0.08560845255851746,
0.07282231003046036,
-0.10070640593767166,
-0.032615210860967636,
0.03520294651389122,
-0.005141355097293854,
0.0009600772173143923,
0.042834486812353134,
0.0696583017706871,
0.010832082480192184,
0.03094172477722168,
-0.06999066472053528,
-0.01570904813706875,
-0.013402287848293781,
0.001008926541544497,
0.006675120908766985,
-0.09901349991559982,
-0.023434078320860863,
-0.02321266382932663,
0.034622352570295334,
0.0030991798266768456,
0.1365426927804947,
0.07976505160331726,
0.005607412196695805,
0.013937640935182571,
0.007881847210228443,
0.010737048462033272,
0.06371314823627472,
0.06217050924897194,
0.01664343848824501,
-0.023497462272644043,
-0.06855864822864532,
-0.011257766745984554,
0.04859945923089981,
-0.07336893677711487,
-0.05499083921313286,
-0.01761641353368759,
-0.0021637899335473776,
0.006742553319782019,
0.01402786374092102,
0.0180071834474802,
-0.004881762433797121,
0.04793267324566841,
-0.05329715088009834,
-0.08846709132194519,
-0.08647499978542328,
-0.01401499006897211,
-0.0410122312605381,
-0.030663730576634407,
-0.08994828164577484,
0.00006896600825712085,
0.13310398161411285,
0.02679116278886795,
0.023197196424007416,
-0.04093285650014877,
0.09501540660858154,
-0.01572110690176487,
0.050322603434324265,
0.024215858429670334,
0.025059634819626808,
0.04603499919176102,
0.034423865377902985,
0.03079206496477127,
-0.10342596471309662,
0.022608645260334015,
0.012877744622528553,
0.10073265433311462,
-0.06655897200107574,
0.09202077984809875,
-0.043797146528959274,
0.08603095263242722,
0.036271627992391586,
0.03305165469646454,
-0.01732897385954857,
-0.0720944032073021,
0.010871829465031624,
0.02397006005048752,
0.0012306675780564547,
0.01338456105440855,
0.07997812330722809,
0.016380762681365013,
-0.03737867996096611,
-0.023819338530302048,
0.12384690344333649,
-0.006356160156428814,
0.19657199084758759,
0.016441408544778824,
-0.10653272271156311,
-0.058110058307647705,
-0.08108390122652054,
0.04820983484387398,
0.05174091085791588,
0.004309921059757471,
-0.05182017758488655,
0.05717352032661438,
0.047761544585227966,
-0.017948204651474953,
0.005388682242482901,
0.0017424225807189941,
-0.024559931829571724,
0.06398467719554901,
-0.003931728657335043,
-0.05403955280780792,
0.052911702543497086,
0.036183636635541916,
0.018509481102228165,
0.025822194293141365,
0.03614462912082672,
-0.033333055675029755,
-0.0005471207550726831,
-0.008032620884478092,
1.3321466973201158e-33,
0.06618474423885345,
0.07582241296768188,
0.005604709964245558,
0.057546500116586685,
-0.04202611744403839,
-0.039330292493104935,
0.001367038581520319,
0.04549029842019081,
-0.03002597577869892,
-0.02365580014884472,
0.011779963038861752,
-0.04144492745399475,
-0.053943995386362076,
0.030028801411390305,
0.01949969306588173,
0.02766193263232708,
-0.08916174620389938,
0.012209007516503334,
-0.01683848351240158,
0.015120305120944977,
0.03695293515920639,
-0.034776072949171066,
-0.033003631979227066,
-0.06208011135458946,
-0.0034936240408569574,
0.058269694447517395,
0.05224458500742912,
0.016840552911162376,
0.0033900465350598097,
-0.005326464306563139,
-0.009172974154353142,
-0.038103632628917694,
0.00772009277716279,
0.03371909260749817,
0.004522526171058416,
-0.03961137309670448,
0.009578332304954529,
-0.07344519346952438,
-0.015475098975002766,
-0.0358370877802372,
-0.029374277219176292,
-0.0037212062161415815,
0.04256540909409523,
0.02848767675459385,
-0.07291693985462189,
-0.08148659020662308,
0.059646449983119965,
-0.03495417535305023,
0.006986843887716532,
-0.07055279612541199,
-0.026107463985681534,
0.012034266255795956,
-0.03688989579677582,
-0.06213976815342903,
-0.05882140249013901,
0.011615708470344543,
0.03202715888619423,
-0.011683303862810135,
0.014727928675711155,
0.025681549683213234,
-0.02494291588664055,
0.09603603184223175,
-0.030130600556731224,
0.08399974554777145,
0.014323500916361809,
0.013384365476667881,
-0.08674297481775284,
0.053849298506975174,
0.0970931425690651,
-0.03668966144323349,
-0.09520108997821808,
-0.027278030291199684,
-0.059369780123233795,
-0.06132909655570984,
0.04784100502729416,
-0.04762210324406624,
0.012564591132104397,
-0.070233054459095,
-0.005851748399436474,
-0.015318074263632298,
0.09592358022928238,
-0.15535634756088257,
-0.04461784288287163,
-0.08846331387758255,
-0.03632659092545509,
-0.005923858378082514,
0.025523383170366287,
-0.03531834855675697,
0.052371200174093246,
-0.05465440824627876,
-0.023009270429611206,
-0.017256498336791992,
-0.06180436536669731,
0.03021179512143135,
0.052821703255176544,
-4.430003826352374e-34,
-0.010573146864771843,
0.04195423796772957,
-0.04274317994713783,
0.11120365560054779,
0.03434710204601288,
0.02670464478433132,
-0.0020875767804682255,
0.04540230706334114,
-0.027510834857821465,
-0.11151448637247086,
-0.06668556481599808,
-0.0637439712882042,
-0.06404508650302887,
-0.006199244409799576,
-0.014473837800323963,
0.02404138445854187,
-0.0064077964052557945,
0.003961482085287571,
-0.0031418483704328537,
0.03258693963289261,
-0.015715466812253,
0.05486515909433365,
-0.0893767699599266,
0.04437887668609619,
0.02411932684481144,
0.04600188136100769,
-0.017612120136618614,
0.13443762063980103,
-0.025553163141012192,
0.04435526952147484,
0.013993281871080399,
-0.05726489797234535,
0.04855312407016754,
-0.013889588415622711,
-0.15018029510974884,
0.034054916352033615,
0.00999940000474453,
-0.03746921196579933,
-0.07071377336978912,
0.08831190317869186,
0.006277379579842091,
-0.059465598315000534,
-0.027261892333626747,
-0.04212849587202072,
-0.0529932901263237,
0.07941583544015884,
-0.04235518351197243,
-0.01585027389228344,
0.03198227658867836,
0.044306837022304535,
0.0025409141089767218,
0.04394271597266197,
-0.07482913881540298,
0.02451208606362343,
0.015227454714477062,
-0.07893648743629456,
-0.013830844312906265,
-0.01592136174440384,
-0.012857831083238125,
0.05332969129085541,
-0.004669450223445892,
0.0012446213513612747,
0.08170589059591293,
-0.06639382243156433,
0.037812139838933945,
-0.08053930848836899,
-0.06067999452352524,
0.08730442076921463,
-0.0063966247253119946,
0.05715084820985794,
-0.01837415061891079,
0.011546239256858826,
0.0030375823844224215,
-0.04976050928235054,
0.0536649227142334,
0.010595293715596199,
0.04768713191151619,
0.0041972738690674305,
-0.06076786294579506,
-0.007515591569244862,
0.031394705176353455,
-0.017936818301677704,
0.014867709949612617,
0.03605547919869423,
0.08228114247322083,
0.10341455042362213,
0.055244408547878265,
0.06751661747694016,
0.02370157279074192,
0.0036187181249260902,
-0.042690981179475784,
0.016176428645849228,
0.013795822858810425,
0.03457352891564369,
-0.007948020473122597,
-4.3752386602591287e-8,
-0.021599726751446724,
0.04510941356420517,
0.014799493364989758,
0.07941469550132751,
0.05594123527407646,
0.010671681724488735,
-0.053554512560367584,
0.040838368237018585,
-0.04030156880617142,
-0.039767179638147354,
0.04779348522424698,
-0.004470465239137411,
-0.024626294150948524,
-0.082903653383255,
0.07611190527677536,
0.09456245601177216,
-0.031660839915275574,
0.04787968844175339,
-0.019880974665284157,
-0.02133101411163807,
0.08517549186944962,
0.002398359589278698,
-0.01665700599551201,
-0.022688420489430428,
0.044703029096126556,
-0.03774068132042885,
0.05308496952056885,
0.1503690481185913,
0.033244844526052475,
-0.045181743800640106,
-0.04282117635011673,
-0.050262872129678726,
-0.017843054607510567,
-0.03837868571281433,
0.047461796551942825,
0.08909999579191208,
0.036308299750089645,
-0.015036258846521378,
0.02070474438369274,
-0.013071023859083652,
-0.018905028700828552,
0.02854684367775917,
-0.09843052923679352,
0.04922163486480713,
0.008222362957894802,
-0.028882719576358795,
-0.05141300708055496,
-0.0593520887196064,
0.013314696028828621,
0.033410072326660156,
0.05118551850318909,
-0.07777964323759079,
0.009110257029533386,
-0.013330280780792236,
0.018975209444761276,
0.05987727642059326,
-0.05764126032590866,
-0.015325640328228474,
-0.037449974566698074,
-0.006677292753010988,
0.05076657608151436,
-0.016003793105483055,
-0.06573022156953812,
0.07268791645765305
] |
NUDGE: Lightweight Non-Parametric Fine-Tuning of Embeddings for Retrieval | https://openreview.net/forum?id=MYw74B77KQ | [
"Sepanta Zeighami",
"Zac Wellmer",
"Aditya Parameswaran"
] | Poster | $k$-Nearest Neighbor search on dense vector embeddings ($k$-NN retrieval) from pre-trained embedding models is the predominant retrieval method for text and images, as well as Retrieval-Augmented Generation (RAG) pipelines. In practice, application developers often fine-tune the embeddings to improve their accuracy on the dataset and query workload in hand. Existing approaches either fine-tune the pre-trained model itself or, more efficiently, but at the cost of accuracy, train adaptor models to transform the output of the pre-trained model. We present NUDGE, a family of novel *non-parametric* embedding fine-tuning approaches that are significantly more accurate and efficient than both sets of existing approaches. NUDGE directly modifies the embeddings of data records to maximize the accuracy of $k$-NN retrieval. We present a thorough theoretical and experimental study of NUDGE's non-parametric approach. We show that even though the underlying problem is NP-Hard, constrained variations can be solved efficiently. These constraints additionally ensure that the changes to the embeddings are modest, avoiding large distortions to the semantics learned during pre-training. In experiments across five pre-trained models and nine standard text and image retrieval datasets, *NUDGE runs in minutes and often improves NDCG@10 by more than 10\% over existing fine-tuning methods. On average, NUDGE provides 3.3$\times$ and 4.3$\times$ higher increase in accuracy and runs 200$\times$ and 3$\times$ faster, respectively, over fine-tuning the pre-trained model and training adaptors.* | semantic similarity search, pre-trained embedding model, fine-tuning | We present NUDGE, a lightweight non-parametric method to fine-tuning embeddings from pre-trained models for nearest neighbor retrieval | 2,248 | 2409.02343 | [
-0.040347229689359665,
-0.08594118803739548,
0.02246851846575737,
0.04804893583059311,
0.043515726923942566,
0.05026023089885712,
-0.047590821981430054,
-0.0061698188073933125,
0.03094879910349846,
-0.07020778954029083,
0.0034984766971319914,
0.0005237169098109007,
0.06619296967983246,
0.06001618131995201,
-0.10749579966068268,
0.013998571783304214,
0.1423315703868866,
0.07797354459762573,
-0.07143272459506989,
-0.00322816614061594,
-0.06250900775194168,
0.022767210379242897,
0.10607682913541794,
-0.0939105749130249,
0.009214887395501137,
-0.07866509258747101,
0.002873318735510111,
-0.023662442341446877,
0.031756117939949036,
-0.028974276036024094,
0.012116321362555027,
0.033570386469364166,
-0.09279090166091919,
0.07036097347736359,
-0.028719037771224976,
0.04584880173206329,
-0.06493708491325378,
0.06597410887479782,
-0.03332003206014633,
-0.030151117593050003,
0.03269926831126213,
0.010944836772978306,
-0.005425563547760248,
0.09232381731271744,
0.05124461278319359,
0.002604976762086153,
-0.05072429031133652,
-0.03412077948451042,
0.018018918111920357,
-0.0011171146761626005,
-0.07417362183332443,
0.009356927126646042,
-0.06721081584692001,
0.021886374801397324,
0.002165074460208416,
-0.035037774592638016,
-0.006293957121670246,
-0.03753865510225296,
-0.007352127693593502,
-0.009788733907043934,
-0.0046327970921993256,
-0.13291141390800476,
0.02738013304769993,
-0.058652494102716446,
0.032634276896715164,
-0.03816961869597435,
0.042430248111486435,
0.011602190323174,
0.0274821724742651,
-0.0044985306449234486,
0.007171825040131807,
0.07325737923383713,
-0.06954638659954071,
0.014440114609897137,
-0.011222656816244125,
0.07741505652666092,
0.046899981796741486,
0.01162528246641159,
0.03223804011940956,
-0.06735212355852127,
-0.027787024155259132,
-0.022439206019043922,
0.038404274731874466,
0.022466909140348434,
0.10397183895111084,
0.03635476529598236,
0.04831426218152046,
-0.035658709704875946,
0.05940447747707367,
0.01985444501042366,
-0.05209359899163246,
-0.006669230293482542,
0.033739231526851654,
-0.04542722553014755,
0.029765978455543518,
0.026430439203977585,
0.07045987248420715,
0.010274067521095276,
-0.05919959396123886,
0.09506700187921524,
0.03632063418626785,
0.13146576285362244,
0.029842250049114227,
-0.017935240641236305,
0.016490884125232697,
0.02908124402165413,
0.0335102304816246,
0.031661879271268845,
0.006376476958394051,
-0.07737311720848083,
0.014775093644857407,
0.04905116185545921,
-0.02538946270942688,
-0.03152709826827049,
-0.008972187526524067,
-0.04343709722161293,
0.0029749495442956686,
-0.05058954656124115,
0.006655750796198845,
0.028003089129924774,
-0.05654536187648773,
0.06185999512672424,
-0.08550138026475906,
-0.008985578082501888,
-0.02605573832988739,
-0.001844103797338903,
-0.035187866538763046,
1.964221672865272e-33,
0.026654036715626717,
0.03152889385819435,
0.058579858392477036,
-0.009461619891226292,
0.01502836961299181,
-0.042861457914114,
-0.007372606545686722,
0.13241511583328247,
-0.10145371407270432,
0.017055967822670937,
-0.0539269745349884,
0.026388701051473618,
0.012579979375004768,
0.05174538865685463,
0.006272926460951567,
-0.03111538663506508,
-0.08053421974182129,
-0.008312750607728958,
-0.038348473608493805,
0.041941914707422256,
-0.02304803393781185,
-0.06915704160928726,
-0.01325124129652977,
0.012813973240554333,
-0.0506601482629776,
-0.031180618330836296,
0.05380268022418022,
-0.06668384373188019,
-0.04750199615955353,
-0.03555353358387947,
-0.1268964260816574,
-0.05220998078584671,
0.07758263498544693,
-0.020745782181620598,
0.015803445130586624,
-0.0726480782032013,
-0.004230217542499304,
0.021297482773661613,
-0.02088332548737526,
-0.004127825610339642,
0.004319141153246164,
0.018694864585995674,
0.005137541797012091,
-0.02052679844200611,
-0.06876490265130997,
-0.06673820316791534,
0.010749191045761108,
-0.005007348954677582,
-0.05547104775905609,
-0.021577123552560806,
0.07902036607265472,
0.041388802230358124,
-0.05333538353443146,
-0.028671786189079285,
0.04587169736623764,
-0.09061162173748016,
0.08470001816749573,
0.0317051000893116,
0.0320780910551548,
0.014665544964373112,
0.030789269134402275,
-0.003450270975008607,
0.01270800456404686,
0.09395979344844818,
-0.005018585827201605,
-0.04035445302724838,
0.055148929357528687,
0.06673870980739594,
0.08681900799274445,
0.018606606870889664,
-0.031065592542290688,
0.042388711124658585,
0.017737282440066338,
-0.07010402530431747,
0.06902103871107101,
0.021155212074518204,
-0.006232984364032745,
-0.0950118899345398,
0.04909491539001465,
0.006117059849202633,
0.03771113231778145,
-0.02819090709090233,
0.006709700915962458,
-0.048477087169885635,
-0.07019005715847015,
-0.07659044116735458,
0.02682916261255741,
-0.09708580374717712,
0.060741130262613297,
-0.006893955636769533,
-0.017341705039143562,
0.03339827433228493,
-0.13034600019454956,
-0.044680193066596985,
0.06895606964826584,
3.0220786561177593e-34,
0.06523656100034714,
-0.03642468526959419,
-0.029117165133357048,
0.0488518625497818,
0.008902457542717457,
0.003804013831540942,
0.01872887648642063,
0.007760381326079369,
-0.029099905863404274,
-0.0726362019777298,
-0.02290673740208149,
-0.07887039333581924,
0.07460039108991623,
-0.07408788055181503,
0.04205990582704544,
0.0885758325457573,
-0.12182652950286865,
0.045583419501781464,
-0.05095503851771355,
0.04870600625872612,
-0.03873048350214958,
0.019938530400395393,
-0.05250250920653343,
0.09431049227714539,
-0.014439844526350498,
0.09874652326107025,
-0.010698792524635792,
0.014094254933297634,
-0.07793345302343369,
-0.03794747591018677,
-0.037387311458587646,
-0.052327025681734085,
-0.00859859213232994,
-0.045258086174726486,
-0.0396936796605587,
0.05750777944922447,
0.08189597725868225,
0.038268934935331345,
-0.007478891406208277,
0.04537203907966614,
0.012112773954868317,
0.0293861273676157,
-0.03507927432656288,
0.052209362387657166,
-0.026666022837162018,
-0.030859386548399925,
-0.08916446566581726,
-0.03663034737110138,
0.07613274455070496,
0.031949181109666824,
0.02804313600063324,
0.03562387451529503,
-0.06033936142921448,
0.038091085851192474,
0.040961869060993195,
-0.08550211042165756,
-0.10592435300350189,
-0.007630630861967802,
0.09688850492238998,
0.020527200773358345,
0.011985493823885918,
-0.07784206420183182,
0.006630995776504278,
-0.056772518903017044,
-0.009939432144165039,
-0.05721850320696831,
-0.025917092338204384,
0.03047754056751728,
-0.07216528058052063,
0.035776302218437195,
-0.036253515630960464,
-0.037934545427560806,
0.05275546386837959,
-0.018628012388944626,
-0.00844559632241726,
0.01865866221487522,
0.09376498311758041,
0.008221163414418697,
-0.007318343501538038,
-0.08371730148792267,
-0.011486411094665527,
0.009538535960018635,
0.011851362884044647,
0.04890041425824165,
0.036489930003881454,
0.11296053975820541,
0.07096503674983978,
0.0011346046812832355,
0.03196894749999046,
0.016223996877670288,
-0.04053996130824089,
0.023117607459425926,
-0.03976713493466377,
0.03995298221707344,
0.044056206941604614,
-3.998675879302027e-8,
-0.05048698931932449,
0.06612975150346756,
-0.04322880879044533,
0.024084340780973434,
0.02300168201327324,
0.03216153383255005,
0.020603511482477188,
0.17120438814163208,
-0.07565812766551971,
0.00046410385402850807,
0.0577901192009449,
-0.014515235088765621,
-0.010044228285551071,
-0.010757580399513245,
0.024779217317700386,
0.029061635956168175,
0.02988690510392189,
-0.008947809226810932,
0.010899562388658524,
0.0008253610576502979,
-0.009694823995232582,
0.14387944340705872,
0.07087556272745132,
-0.033721789717674255,
-0.004702923819422722,
-0.022975580766797066,
-0.022204145789146423,
0.07818413525819778,
0.12934894859790802,
-0.061685819178819656,
0.007274869363754988,
0.003920596558600664,
0.023363342508673668,
-0.11736203730106354,
-0.02239118330180645,
0.07672176510095596,
-0.005244158674031496,
-0.02373604290187359,
-0.06083879619836807,
-0.011255219578742981,
0.06628814339637756,
0.06140647828578949,
-0.04638465493917465,
-0.014892395585775375,
-0.0016665689181536436,
-0.004430221859365702,
0.04220198467373848,
-0.035395972430706024,
0.016589520499110222,
0.028718827292323112,
-0.007237694691866636,
-0.09196732193231583,
-0.0156639963388443,
0.07520829886198044,
0.0031434521079063416,
0.0028995228931307793,
-0.026893246918916702,
-0.02372405305504799,
0.06543219834566116,
0.04611889645457268,
0.003545989515259862,
-0.003514717798680067,
-0.0050015985034406185,
0.045504361391067505
] |
Private Mechanism Design via Quantile Estimation | https://openreview.net/forum?id=JQQDePbfxh | [
"Yuanyuan Yang",
"Tao Xiao",
"Bhuvesh Kumar",
"Jamie Heather Morgenstern"
] | Poster | We investigate the problem of designing differentially private (DP), revenue-maximizing single item auction. Specifically, we consider broadly applicable settings in mechanism design where agents' valuation distributions are **independent**, **non-identical**, and can be either **bounded** or **unbounded**. Our goal is to design such auctions with **pure**, i.e., $(\epsilon,0)$ privacy in polynomial time.
In this paper, we propose two computationally efficient auction learning framework that achieves **pure** privacy under bounded and unbounded distribution settings. These frameworks reduces the problem of privately releasing a revenue-maximizing auction to the private estimation of pre-specified quantiles. Our solutions increase the running time by polylog factors compared to the non-private version. As an application, we show how to extend our results to the multi-round online auction setting with non-myopic bidders. To our best knowledge, this paper is the first to efficiently deliver a Myerson auction with **pure** privacy and near-optimal revenue, and the first to provide such auctions for **unbounded** distributions. | online auctions, differential privacy, mechanism design | we demonstrate how to efficiently learn a near-optimal mechanism with pure privacy from sample data for the single item setting. We further apply this mechanism to online settings involving non-myopic and strategic bidders. | 2,239 | null | [
-0.1111772283911705,
0.04706896096467972,
-0.05140838027000427,
-0.032397352159023285,
-0.031025974079966545,
-0.028915628790855408,
0.040468066930770874,
-0.018732814118266106,
0.041990604251623154,
0.052277058362960815,
-0.0002473488566465676,
-0.012837090529501438,
0.00803942047059536,
0.016658805310726166,
0.030571846291422844,
0.025299113243818283,
0.08074438571929932,
0.07595814764499664,
-0.06464081257581711,
0.0646124854683876,
0.07711951434612274,
-0.035464074462652206,
0.00001683407390373759,
-0.001124719507060945,
0.06774725764989853,
-0.07894793897867203,
0.02750101126730442,
-0.010902974754571915,
-0.0077713471837341785,
-0.028600826859474182,
-0.001880158670246601,
-0.017035027965903282,
-0.016970379278063774,
-0.004423591308295727,
0.051216479390859604,
0.01984541490674019,
-0.06072264164686203,
-0.007429988589137793,
-0.04902777075767517,
0.00585912587121129,
0.005219206679612398,
-0.02478730119764805,
-0.06831113249063492,
0.008716591633856297,
-0.029118601232767105,
0.04057946056127548,
0.04778878390789032,
0.10452046990394592,
0.011109388433396816,
-0.025360189378261566,
0.04526453837752342,
-0.021695733070373535,
0.010160158388316631,
0.020355023443698883,
-0.02687380276620388,
-0.039422959089279175,
-0.01892089657485485,
-0.07656212151050568,
-0.026925314217805862,
0.009976747445762157,
-0.07161653786897659,
-0.05258508026599884,
-0.021645164117217064,
0.04058070853352547,
-0.045164335519075394,
0.05161833390593529,
0.029578624293208122,
0.09209486842155457,
-0.07396122068166733,
0.0331408753991127,
0.043051138520240784,
0.01724940724670887,
-0.05382855609059334,
-0.0042204102501273155,
0.0876532718539238,
-0.03634420409798622,
0.034689318388700485,
-0.04563182219862938,
-0.048325929790735245,
-0.024891341105103493,
-0.018823668360710144,
-0.04233059659600258,
-0.04586324095726013,
-0.008657402358949184,
0.06647118180990219,
-0.04558715596795082,
-0.010336529463529587,
0.08598911017179489,
0.10851333290338516,
0.018777890130877495,
-0.06056991592049599,
0.09210194647312164,
-0.07579860836267471,
-0.024669213220477104,
0.06990117579698563,
-0.03075224719941616,
0.04389701411128044,
-0.028692152351140976,
0.016185926273465157,
0.11780696362257004,
-0.04510866478085518,
0.05413233861327171,
-0.0044594029895961285,
-0.022489303722977638,
0.04266311973333359,
0.02209555171430111,
-0.05390971526503563,
0.00379945058375597,
-0.006145077757537365,
0.004279098007827997,
0.006345764268189669,
-0.0034658624790608883,
0.0074707320891320705,
0.03743363544344902,
-0.007922133430838585,
0.0010420740582048893,
-0.030647363513708115,
0.06832433491945267,
-0.007477900478988886,
-0.06148882582783699,
0.01554737240076065,
-0.008364531211555004,
0.0408727303147316,
0.009678225964307785,
-0.026717839762568474,
0.02344835177063942,
-0.12357892841100693,
-2.6012953463018035e-33,
0.025466816499829292,
0.02116822451353073,
0.04933436959981918,
-0.10986832529306412,
-0.0012619707267731428,
0.08942806720733643,
0.06790301948785782,
0.05430338531732559,
-0.04790373891592026,
0.06752791255712509,
0.057366643100976944,
0.03736470267176628,
0.047172531485557556,
-0.0027884182054549456,
0.061532340943813324,
-0.005105858203023672,
-0.04096498340368271,
0.06900547444820404,
0.1177920252084732,
-0.06453030556440353,
0.07184427976608276,
-0.009958920069038868,
-0.01926828734576702,
-0.021511942148208618,
-0.03515305370092392,
-0.042012061923742294,
0.059928588569164276,
-0.046278614550828934,
0.05031534284353256,
0.043220967054367065,
-0.0681881383061409,
0.029692471027374268,
0.06577781587839127,
0.032816529273986816,
0.016842098906636238,
0.004642697982490063,
-0.08788229525089264,
-0.05795656517148018,
0.05100665241479874,
-0.0024848885368555784,
-0.051897358149290085,
-0.05297762528061867,
0.0641704872250557,
-0.0022652887273579836,
-0.14453288912773132,
0.011658238247036934,
-0.0337362065911293,
0.05190012976527214,
-0.015800055116415024,
-0.03800100460648537,
0.018230842426419258,
0.014851710759103298,
-0.06282016634941101,
-0.0662158727645874,
-0.015783047303557396,
-0.05129667744040489,
0.0035710798110812902,
0.017910214141011238,
-0.058561328798532486,
0.015522449277341366,
-0.01829393208026886,
0.029521889984607697,
0.04649638012051582,
-0.023314807564020157,
-0.06523925811052322,
0.02791116014122963,
-0.025555206462740898,
-0.01090118195861578,
0.05439901351928711,
-0.10155762732028961,
0.0021091599483042955,
-0.022649424150586128,
-0.008474146015942097,
-0.0637533888220787,
0.07174686342477798,
-0.007101402152329683,
0.12481195479631424,
0.10173115134239197,
0.03905816376209259,
-0.04129732772707939,
0.025666823610663414,
0.033556513488292694,
0.019709501415491104,
0.035211678594350815,
0.00012479399447329342,
-0.04957681894302368,
-0.029720863327383995,
0.03268291801214218,
-0.05664663016796112,
-0.00011932803317904472,
-0.08809851855039597,
-0.03799843043088913,
-0.108997642993927,
0.003791632829234004,
-0.029984595254063606,
3.542740231174231e-34,
-0.054010991007089615,
0.022131774574518204,
-0.051320262253284454,
0.04012920334935188,
0.0851655974984169,
-0.0566335991024971,
-0.07208143174648285,
0.002587111433967948,
0.028044071048498154,
-0.00419546477496624,
-0.08999238163232803,
0.059095319360494614,
0.020606614649295807,
-0.04024168848991394,
-0.016915153712034225,
0.04268898069858551,
0.007205050904303789,
0.03986269235610962,
-0.044813934713602066,
0.017382027581334114,
0.02930666320025921,
0.10858647525310516,
-0.01740802265703678,
0.07234978675842285,
0.041492659598588943,
-0.02448778972029686,
-0.10853918641805649,
0.09294665604829788,
0.01607896387577057,
0.058831579983234406,
-0.026249118149280548,
-0.09778256714344025,
-0.08461134880781174,
-0.1499318778514862,
-0.0017193249659612775,
-0.007270995993167162,
0.011979272589087486,
0.04358513280749321,
-0.015849048271775246,
0.04339488595724106,
-0.015618475154042244,
-0.08004466444253922,
-0.04444503039121628,
0.022390814498066902,
0.020975574851036072,
-0.004986774642020464,
-0.03334176540374756,
-0.022543206810951233,
0.06310273706912994,
0.07214413583278656,
-0.07436024397611618,
0.08774645626544952,
0.03288380056619644,
0.048761870712041855,
-0.10658151656389236,
-0.04705338925123215,
-0.025842400267720222,
0.007106916978955269,
0.06716221570968628,
0.04308343678712845,
0.01159716583788395,
0.014945105649530888,
-0.010830511339008808,
0.07659075409173965,
-0.03436120226979256,
-0.07243315875530243,
0.020684003829956055,
-0.0392906479537487,
-0.06102451682090759,
0.01125791110098362,
-0.026124313473701477,
-0.05777483806014061,
-0.0143863121047616,
-0.02363196760416031,
-0.02614293247461319,
0.11981084197759628,
0.007132555823773146,
0.01820969022810459,
0.02579585090279579,
0.05665213614702225,
0.020932961255311966,
0.0756290853023529,
0.07642165571451187,
-0.04741642624139786,
0.04327448457479477,
-0.11360248923301697,
0.1283506155014038,
-0.025600723922252655,
0.020796440541744232,
-0.02648754045367241,
0.023157412186264992,
0.01434998493641615,
0.037676114588975906,
0.024826811626553535,
-0.06302814185619354,
-4.0819173818817944e-8,
0.002197439782321453,
-0.012619156390428543,
0.07445571571588516,
0.001124369096942246,
0.07646571844816208,
0.005044996272772551,
-0.005708564538508654,
0.029885346069931984,
-0.11485671997070312,
-0.04297510161995888,
0.07835942506790161,
0.038622111082077026,
-0.025835495442152023,
-0.03705006092786789,
-0.09654894471168518,
0.04252303019165993,
-0.045924633741378784,
-0.051659077405929565,
-0.03268580883741379,
-0.043873824179172516,
-0.02621340937912464,
0.008835062384605408,
-0.008017865009605885,
-0.05416513979434967,
-0.0018780954414978623,
-0.033337242901325226,
0.019058264791965485,
-0.07073475420475006,
0.10580628365278244,
0.10209866613149643,
-0.03009490855038166,
-0.029783956706523895,
0.008758091367781162,
0.006215606816112995,
-0.05585714429616928,
0.12617337703704834,
-0.08446598052978516,
-0.015571611933410168,
-0.009386248886585236,
-0.02722688764333725,
0.011763964779675007,
-0.04462515935301781,
-0.044170644134283066,
0.01710548996925354,
0.07110164314508438,
-0.016353711485862732,
-0.03757479414343834,
-0.044627558439970016,
0.04806835204362869,
-0.026695547625422478,
0.015338179655373096,
-0.00870640017092228,
0.00026563555002212524,
-0.028347497805953026,
0.10458225011825562,
0.03466975688934326,
-0.003876536153256893,
-0.015484482981264591,
0.0900065153837204,
0.08616387844085693,
0.03367479145526886,
-0.03930072858929634,
-0.04337368905544281,
0.07513046264648438
] |
Reasoning-Enhanced Healthcare Predictions with Knowledge Graph Community Retrieval | https://openreview.net/forum?id=8fLgt7PQza | [
"Pengcheng Jiang",
"Cao Xiao",
"Minhao Jiang",
"Parminder Bhatia",
"Taha Kass-Hout",
"Jimeng Sun",
"Jiawei Han"
] | Poster | Large language models (LLMs) have demonstrated significant potential in clinical decision support. Yet LLMs still suffer from hallucinations and lack fine-grained contextual medical knowledge, limiting their high-stake healthcare applications such as clinical diagnosis. Traditional retrieval-augmented generation (RAG) methods attempt to address these limitations but frequently retrieve sparse or irrelevant information, undermining prediction accuracy. We introduce KARE, a novel framework that integrates knowledge graph (KG) community-level retrieval with LLM reasoning to enhance healthcare predictions. KARE constructs a comprehensive multi-source KG by integrating biomedical databases, clinical literature, and LLM-generated insights, and organizes it using hierarchical graph community detection and summarization for precise and contextually relevant information retrieval. Our key innovations include: (1) a dense medical knowledge structuring approach enabling accurate retrieval of relevant information; (2) a dynamic knowledge retrieval mechanism that enriches patient contexts with focused, multi-faceted medical insights; and (3) a reasoning-enhanced prediction framework that leverages these enriched contexts to produce both accurate and interpretable clinical predictions. Extensive experiments demonstrate that KARE outperforms leading models by up to 10.8-15.0\% on MIMIC-III and 12.6-12.7\% on MIMIC-IV for mortality and readmission predictions. In addition to its impressive prediction accuracy, our framework leverages the reasoning capabilities of LLMs, enhancing the trustworthiness of clinical predictions. | EHR Prediction, Large Language Models, Knowledge Graphs | null | 2,237 | 2410.04585 | [
-0.05031541734933853,
-0.027924351394176483,
-0.0186707004904747,
0.07048621773719788,
0.03803809732198715,
-0.008365009911358356,
-0.005232125520706177,
0.044437602162361145,
-0.010852975770831108,
-0.013741815462708473,
-0.05114438012242317,
-0.021272918209433556,
0.03994522988796234,
0.0934893861413002,
-0.03529560938477516,
0.03950788453221321,
0.056546226143836975,
-0.004191308282315731,
-0.031577423214912415,
-0.02564696967601776,
0.01764295995235443,
0.08880139887332916,
0.06069741025567055,
0.03919513523578644,
0.044251710176467896,
-0.000709537067450583,
-0.020945047959685326,
-0.016072692349553108,
0.030572038143873215,
0.040155332535505295,
0.08971426635980606,
0.07922911643981934,
0.011175533756613731,
0.049463901668787,
-0.041062820702791214,
0.03957243636250496,
-0.11981185525655746,
0.13108593225479126,
-0.004252160899341106,
0.00004218120375298895,
0.0391869954764843,
-0.0403745099902153,
0.05727550759911537,
0.11465122550725937,
0.05515452101826668,
0.025411317124962807,
-0.11760464310646057,
0.018698783591389656,
0.031539034098386765,
0.0594407357275486,
-0.12347935885190964,
-0.022776473313570023,
0.014283428899943829,
-0.00908363051712513,
0.023773275315761566,
0.018259968608617783,
-0.03912153095006943,
-0.007733040954917669,
-0.08915280550718307,
-0.09553191065788269,
0.018292956054210663,
-0.16133767366409302,
0.015086490660905838,
-0.005077720619738102,
-0.03326864168047905,
-0.008798398077487946,
0.061989594250917435,
0.055796150118112564,
-0.061690252274274826,
-0.010636914521455765,
0.016743719577789307,
-0.014284688048064709,
-0.017020754516124725,
0.1494491845369339,
0.007841452956199646,
0.08693227171897888,
0.06065649539232254,
-0.019426073879003525,
-0.01454803254455328,
-0.013851119205355644,
0.020804185420274734,
0.05216697230935097,
0.05430714040994644,
-0.020756002515554428,
0.018045499920845032,
-0.02037537470459938,
0.023785946890711784,
-0.05205678194761276,
-0.04541816562414169,
-0.03251178190112114,
-0.0130265261977911,
-0.058282554149627686,
0.05427444726228714,
-0.052814990282058716,
0.028022553771734238,
0.07683262228965759,
0.016529565677046776,
-0.07985769212245941,
-0.01838207058608532,
0.04977474361658096,
0.000003604510311561171,
0.02557889185845852,
0.027766462415456772,
-0.07812084257602692,
0.013948364183306694,
-0.06247131898999214,
-0.01650579459965229,
0.07296189665794373,
0.020859329029917717,
-0.030398210510611534,
0.019245849922299385,
0.1093098595738411,
0.059431083500385284,
0.006360601168125868,
0.05342855304479599,
-0.005393609870225191,
0.058514464646577835,
-0.0098236333578825,
0.050866249948740005,
0.0677875354886055,
-0.022248316556215286,
-0.01821601577103138,
0.02960004098713398,
-0.042424559593200684,
-0.016605010256171227,
0.030176004394888878,
-0.06105513125658035,
-1.4803330967691235e-33,
0.061704110354185104,
0.009608426131308079,
0.05854355916380882,
0.05435827746987343,
0.06161712482571602,
-0.042461007833480835,
-0.08259333670139313,
0.010127032175660133,
-0.027303947135806084,
-0.052593331784009933,
-0.09487269073724747,
0.05471004918217659,
0.03037911094725132,
0.017106838524341583,
-0.037247251719236374,
0.07506492733955383,
-0.10419237613677979,
0.02907794527709484,
-0.0015033885138109326,
-0.02355775609612465,
0.022563228383660316,
-0.0654221698641777,
-0.00425341771915555,
0.02823377028107643,
-0.049662671983242035,
0.0435258187353611,
0.011356549337506294,
0.003103731432929635,
0.009709330275654793,
-0.030276063829660416,
-0.13100361824035645,
0.03194805979728699,
0.045933328568935394,
-0.05491366237401962,
0.019257381558418274,
-0.007739013060927391,
-0.019249284639954567,
-0.08769391477108002,
0.01917405053973198,
0.008570929057896137,
-0.017840558663010597,
-0.0016534854657948017,
-0.009453455917537212,
-0.019317304715514183,
-0.003499754937365651,
-0.0445987768471241,
-0.09454762190580368,
-0.022096874192357063,
-0.03209935873746872,
-0.015879478305578232,
0.05617343261837959,
-0.03305404260754585,
-0.015233068726956844,
0.05913059413433075,
0.01997263915836811,
0.023240629583597183,
-0.002791560022160411,
0.008998542092740536,
0.07313375920057297,
0.038414955139160156,
0.0758192390203476,
0.0265590101480484,
0.040520600974559784,
0.018389619886875153,
-0.024389130994677544,
0.003750906093046069,
-0.019867390394210815,
0.05198461189866066,
0.07481522113084793,
-0.01684720627963543,
0.0018482167506590486,
0.06412198394536972,
0.035384006798267365,
-0.05931349843740463,
0.014486371539533138,
0.041718121618032455,
-0.016886547207832336,
-0.09822174906730652,
-0.04656976833939552,
0.01563885807991028,
0.03372757136821747,
-0.01290030125528574,
-0.038218673318624496,
0.025736533105373383,
0.0014624542091041803,
-0.05023196339607239,
0.02715565264225006,
-0.03875166177749634,
-0.0864364430308342,
-0.13984277844429016,
-0.0651763305068016,
0.022056791931390762,
0.013962182216346264,
0.021301910281181335,
0.025970159098505974,
2.0068767353545144e-34,
0.03439990431070328,
0.015509478747844696,
0.06080961227416992,
-0.02973227947950363,
0.0815451443195343,
-0.052558593451976776,
-0.04321597144007683,
-0.010916411876678467,
-0.044375769793987274,
-0.025453336536884308,
0.04554132744669914,
-0.0938345193862915,
0.06590957939624786,
-0.015069740824401379,
0.024615732952952385,
0.03182688355445862,
-0.04051298648118973,
-0.005918543785810471,
-0.07697346806526184,
0.10663464665412903,
-0.012200243771076202,
0.025211798027157784,
-0.13705813884735107,
0.060173336416482925,
0.024777423590421677,
0.04368820786476135,
0.007705464959144592,
-0.014023271389305592,
-0.01844382844865322,
-0.0202901903539896,
-0.08275621384382248,
-0.03816237673163414,
-0.022974301129579544,
-0.11789793521165848,
-0.0843401774764061,
0.01260280143469572,
0.02873414196074009,
-0.0959060862660408,
-0.07091351598501205,
-0.002843320369720459,
0.06359105557203293,
-0.004761949181556702,
-0.062095433473587036,
-0.01885315589606762,
-0.046164121478796005,
0.011857738718390465,
-0.0056389872916042805,
0.02657105028629303,
0.07154084742069244,
0.022397883236408234,
0.06419474631547928,
-0.03729938715696335,
0.0016522366786375642,
-0.000003628755166573683,
-0.008906388655304909,
-0.10891425609588623,
-0.06328777968883514,
0.019550440832972527,
-0.03287632390856743,
-0.03792209178209305,
-0.043410900980234146,
-0.037849310785532,
-0.0542965866625309,
0.009464295580983162,
-0.03250335901975632,
-0.058196455240249634,
0.03912217915058136,
-0.016378168016672134,
-0.0480833426117897,
0.03945581987500191,
-0.04167398065328598,
-0.00484883040189743,
-0.010074526071548462,
-0.028443630784749985,
0.1297069489955902,
0.024195637553930283,
0.02971048094332218,
-0.021207567304372787,
-0.038396358489990234,
-0.0643853172659874,
-0.0063916402868926525,
-0.0862831324338913,
0.05375524237751961,
0.07178798317909241,
0.03068009950220585,
-0.0466422513127327,
0.01521572470664978,
0.09003099054098129,
0.07406888902187347,
0.050247445702552795,
-0.05772800371050835,
-0.03569268062710762,
-0.06798999756574631,
0.009347948245704174,
-0.046539340168237686,
-3.828905192904131e-8,
-0.059558168053627014,
0.06290926784276962,
-0.04091230407357216,
0.053206659853458405,
0.024609679356217384,
-0.03544354811310768,
-0.05204807221889496,
0.09469731152057648,
0.0008635720587335527,
0.04409513249993324,
0.036037757992744446,
0.02797405794262886,
-0.07487490028142929,
-0.05392669886350632,
0.04176757112145424,
0.07932166755199432,
-0.029934842139482498,
0.06456711888313293,
0.024055665358901024,
-0.06356216967105865,
0.007607183884829283,
-0.030162950977683067,
0.027548642829060555,
0.0540471151471138,
0.1012921929359436,
-0.030952658504247665,
-0.11123750358819962,
0.11131266504526138,
0.02487562969326973,
-0.06831350177526474,
-0.03489965200424194,
-0.017545834183692932,
0.017930442467331886,
0.01716015674173832,
-0.011786310002207756,
0.005438188556581736,
0.027703193947672844,
-0.08379516005516052,
0.02501521073281765,
-0.030053481459617615,
0.03158360347151756,
0.059838730841875076,
-0.04114710912108421,
0.03535136580467224,
-0.047583479434251785,
0.0061229984275996685,
0.030104558914899826,
0.03945786505937576,
0.0158457700163126,
-0.06007599085569382,
-0.06160248816013336,
-0.08440152555704117,
0.04881339520215988,
0.01378070842474699,
0.05457251891493797,
0.07808241248130798,
-0.024691766127943993,
0.04699064418673515,
0.06698447465896606,
-0.07892018556594849,
0.055968958884477615,
0.04112590104341507,
-0.06595665961503983,
0.03518238663673401
] |
Brain Mapping with Dense Features: Grounding Cortical Semantic Selectivity in Natural Images With Vision Transformers | https://openreview.net/forum?id=yJ9QNbpMi2 | [
"Andrew Luo",
"Jacob Yeung",
"Rushikesh Zawar",
"Shaurya Rajat Dewan",
"Margaret Marie Henderson",
"Leila Wehbe",
"Michael J. Tarr"
] | Poster | We introduce BrainSAIL (Semantic Attribution and Image Localization), a method for linking neural selectivity with spatially distributed semantic visual concepts in natural scenes. BrainSAIL leverages recent advances in large-scale artificial neural networks, using them to provide insights into the functional topology of the brain. To overcome the challenge presented by the co-occurrence of multiple categories in natural images, BrainSAIL exploits semantically consistent, dense spatial features from pre-trained vision models, building upon their demonstrated ability to robustly predict neural activity. This method derives clean, spatially dense embeddings without requiring any additional training, and employs a novel denoising process that leverages the semantic consistency of images under random augmentations. By unifying the space of whole-image embeddings and dense visual features and then applying voxel-wise encoding models to these features, we enable the identification of specific subregions of each image which drive selectivity patterns in different areas of the higher visual cortex. This provides a powerful tool for dissecting the neural mechanisms that underlie semantic visual processing for natural images. We validate BrainSAIL on cortical regions with known category selectivity, demonstrating its ability to accurately localize and disentangle selectivity to diverse visual concepts. Next, we demonstrate BrainSAIL's ability to characterize high-level visual selectivity to scene properties and low-level visual features such as depth, luminance, and saturation, providing insights into the encoding of complex visual information. Finally, we use BrainSAIL to directly compare the feature selectivity of different brain encoding models across different regions of interest in visual cortex. Our innovative method paves the way for significant advances in mapping and decomposing high-level visual representations in the human brain. | fMRI, visual cortex, neuroscience, cognitive science, brain, vision transformer, semantic selectivity | We propose an efficient semantic distillation module and leverage ViTs to investigate selectivity in human visual cortex. | 2,236 | 2410.05266 | [
0.06847225874662399,
-0.12471926212310791,
0.12877441942691803,
0.03542473912239075,
0.08872438967227936,
-0.010096685029566288,
-0.0386316180229187,
-0.13093209266662598,
0.021904319524765015,
-0.01061074435710907,
-0.04297066852450371,
-0.1035793125629425,
-0.020963994786143303,
0.06392385065555573,
-0.012976774014532566,
0.03432287275791168,
0.09220852702856064,
0.0696568489074707,
-0.08514074981212616,
0.017097478732466698,
0.0012129201786592603,
0.01071664597839117,
-0.0029417641926556826,
0.01071365736424923,
0.07443658262491226,
0.015215561725199223,
0.0067702000960707664,
-0.013396600261330605,
-0.000026055917260237038,
-0.08877386897802353,
0.11565034091472626,
0.0019518767949193716,
-0.06627510488033295,
0.02863474190235138,
-0.015223048627376556,
0.060519590973854065,
0.004065312910825014,
-0.007940541952848434,
0.06890895217657089,
-0.01942158117890358,
0.04650474712252617,
0.02784567140042782,
-0.02203211933374405,
0.0694311335682869,
0.015478950925171375,
0.06615982949733734,
0.018597092479467392,
-0.06205148622393608,
-0.008972502313554287,
-0.039999209344387054,
-0.0005938945105299354,
0.03772585839033127,
-0.03780404105782509,
0.08129078894853592,
-0.026709413155913353,
0.03662974387407303,
0.03979627043008804,
-0.038290057331323624,
0.029523441568017006,
0.0264132060110569,
0.07413731515407562,
-0.020584488287568092,
0.051282308995723724,
0.028441909700632095,
0.0092465840280056,
0.005688106641173363,
-0.058687012642621994,
-0.017352532595396042,
-0.04193289205431938,
-0.03738201782107353,
0.04729586839675903,
0.04790320619940758,
-0.016934046521782875,
-0.03674665465950966,
0.0233061071485281,
0.04637450352311134,
-0.04063716530799866,
-0.09959757328033447,
0.06286215037107468,
-0.11570153385400772,
0.016283970326185226,
0.0979597270488739,
0.02116495370864868,
-0.0062011051923036575,
0.17279188334941864,
0.04550229012966156,
0.0025799416471272707,
0.008110662922263145,
-0.07829229533672333,
0.007319911383092403,
-0.03796650469303131,
-0.09112074971199036,
-0.011428908444941044,
-0.028607893735170364,
0.04361084848642349,
-0.0245622918009758,
-0.004322961438447237,
-0.13236045837402344,
-0.012782595120370388,
0.002525384770706296,
0.006566720549017191,
-0.002295688260346651,
0.04541019722819328,
-0.0015746124554425478,
0.019039342179894447,
-0.043239496648311615,
0.06463457643985748,
0.04222367703914642,
0.021940357983112335,
-0.018611155450344086,
0.01444962713867426,
0.022914348170161247,
0.006048223003745079,
0.06932564079761505,
-0.05230236053466797,
0.04112336412072182,
0.00015668640844523907,
0.061495304107666016,
0.06835378706455231,
-0.06900166720151901,
-0.017991304397583008,
0.03543483465909958,
-0.052068620920181274,
-0.0333416722714901,
0.03354382887482643,
-0.05337109789252281,
-0.08135892450809479,
1.0804687650342562e-33,
-0.005223037209361792,
-0.08636704087257385,
0.07981885224580765,
0.027415426447987556,
0.061013225466012955,
-0.03464791923761368,
-0.06616714596748352,
0.07273800671100616,
-0.007530910894274712,
0.05650045722723007,
-0.031801823526620865,
0.061266545206308365,
-0.00572583032771945,
0.14203029870986938,
-0.04834514483809471,
-0.014611574821174145,
-0.058051157742738724,
-0.005288238171488047,
0.01158199179917574,
-0.023576410487294197,
-0.010734016075730324,
-0.024586686864495277,
-0.07016447931528091,
0.015819648280739784,
-0.038217902183532715,
-0.027463706210255623,
0.03393051400780678,
-0.06198980286717415,
-0.008413158357143402,
0.01954091340303421,
-0.16055499017238617,
0.07278788089752197,
-0.00411746883764863,
0.010383793152868748,
0.014801335521042347,
-0.03645872324705124,
0.013153878971934319,
0.0026956479996442795,
-0.03307836875319481,
0.05497199669480324,
0.03972308337688446,
0.03402707353234291,
-0.06947857141494751,
-0.05056113377213478,
-0.04886544868350029,
0.05363308638334274,
0.0016944793751463294,
0.08079434186220169,
-0.08767493069171906,
0.021464860066771507,
0.07020808756351471,
-0.007391766179352999,
-0.027237264439463615,
-0.034652210772037506,
-0.011211719363927841,
-0.014932407066226006,
-0.0029047378338873386,
-0.006548726931214333,
0.06934720277786255,
-0.006904986221343279,
-0.008975109085440636,
-0.021645762026309967,
-0.007757753133773804,
0.0858188271522522,
0.049425337463617325,
-0.013964121229946613,
0.06205346807837486,
0.053539011627435684,
0.014985901303589344,
0.03301040828227997,
-0.022010253742337227,
0.10778811573982239,
-0.006121666636317968,
-0.07315394282341003,
-0.000024377415684284642,
-0.04855085909366608,
0.03669043630361557,
-0.032571595162153244,
0.031076623126864433,
0.07129736244678497,
-0.027546526864171028,
-0.046402424573898315,
-0.061488423496484756,
-0.060943178832530975,
-0.014631850644946098,
-0.020404426380991936,
0.008749167434871197,
-0.021929267793893814,
-0.041322365403175354,
-0.008436696603894234,
-0.01721707172691822,
-0.03425721079111099,
0.045947663486003876,
-0.04726569727063179,
-0.02337973564863205,
-1.721360866565541e-33,
-0.02714630216360092,
-0.03393623232841492,
0.0032908355351537466,
0.035130005329847336,
-0.07856228947639465,
-0.03452258184552193,
0.006814444903284311,
0.013619689270853996,
0.017216097563505173,
-0.00822971947491169,
-0.04330829530954361,
0.02410062775015831,
0.03442772105336189,
0.00902505312114954,
0.05801493301987648,
-0.06703423708677292,
-0.02661847323179245,
0.07509610801935196,
-0.0181274376809597,
0.023808524012565613,
-0.006726348306983709,
0.09224819391965866,
-0.06084882467985153,
0.038047078996896744,
-0.057456035166978836,
0.11064033210277557,
-0.08969533443450928,
0.037063777446746826,
-0.027603965252637863,
0.010726208798587322,
-0.002531731501221657,
0.026139700785279274,
-0.05008843168616295,
-0.08477719873189926,
-0.08489357680082321,
0.08064332604408264,
0.02945208176970482,
-0.02524407021701336,
-0.0972772166132927,
0.007295981515198946,
0.021834246814250946,
-0.07959461957216263,
-0.06310998648405075,
0.056546133011579514,
0.0006977951270528138,
-0.0757405161857605,
0.0019009856041520834,
-0.013963964767754078,
0.016407301649451256,
0.15179115533828735,
-0.026870548725128174,
0.006111919414252043,
-0.15065233409404755,
0.03127799555659294,
0.014131106436252594,
-0.018451552838087082,
-0.04034475237131119,
-0.01953233778476715,
0.08925890922546387,
-0.014706345275044441,
-0.036595381796360016,
-0.014534533023834229,
0.0008673648699186742,
0.008387676440179348,
-0.009513516910374165,
-0.015743259340524673,
-0.062285955995321274,
0.0016536794137209654,
0.07210083305835724,
0.0837908387184143,
-0.0583869144320488,
0.007069997023791075,
0.03324664756655693,
0.01972690224647522,
0.0073517486453056335,
-0.00494026206433773,
0.06856659799814224,
-0.021481139585375786,
0.060137175023555756,
0.0006240711663849652,
0.027886221185326576,
-0.10719798505306244,
-0.017506729811429977,
0.06018226593732834,
0.06810521334409714,
0.09654553979635239,
-0.05822722241282463,
-0.04341382905840874,
0.0403401255607605,
-0.02243729494512081,
-0.045433878898620605,
-0.000669489149004221,
-0.10812494903802872,
0.012516040354967117,
-0.07772840559482574,
-3.699034678561475e-8,
-0.07112795114517212,
0.014047829434275627,
-0.0047368500381708145,
-0.04422697052359581,
0.019253533333539963,
-0.042120449244976044,
0.026188530027866364,
0.10090066492557526,
-0.024912476539611816,
-0.07891149818897247,
0.09159892797470093,
-0.05282619595527649,
-0.015306791290640831,
-0.03519430011510849,
0.07913825660943985,
0.048230789601802826,
0.00737858098000288,
0.037596315145492554,
-0.0027696629986166954,
0.0219571590423584,
-0.011723804287612438,
0.021579289808869362,
0.06564595550298691,
-0.02994171530008316,
0.01648331992328167,
-0.036510366946458817,
-0.033518236130476,
0.06263727694749832,
0.025102559477090836,
0.010292144492268562,
0.10624329745769501,
-0.007614406291395426,
0.019377289339900017,
-0.03316458687186241,
0.011541470885276794,
0.05735541135072708,
0.049858175218105316,
0.005356981884688139,
-0.03479304164648056,
-0.062217533588409424,
-0.01875927485525608,
0.02982555516064167,
-0.011218426749110222,
-0.01122665498405695,
-0.03838065639138222,
0.03212431073188782,
0.04323457181453705,
-0.015291466377675533,
0.029134292155504227,
0.03193707391619682,
-0.1031765341758728,
0.00962889939546585,
0.03494763746857643,
0.14949055016040802,
0.014870533719658852,
-0.021474894136190414,
-0.03300832584500313,
-0.06873656809329987,
-0.019520511850714684,
0.02322767861187458,
-0.004427844192832708,
0.02326403744518757,
-0.028139246627688408,
-0.020429925993084908
] |
X-Drive: Cross-modality Consistent Multi-Sensor Data Synthesis for Driving Scenarios | https://openreview.net/forum?id=IEMmEd5Jgm | [
"Yichen Xie",
"Chenfeng Xu",
"Chensheng Peng",
"Shuqi Zhao",
"Nhat Ho",
"Alexander T. Pham",
"Mingyu Ding",
"Masayoshi Tomizuka",
"Wei Zhan"
] | Poster | Recent advancements have exploited diffusion models for the synthesis of either LiDAR point clouds or camera image data in driving scenarios. Despite their success in modeling single-modality data marginal distribution, there is an under- exploration in the mutual reliance between different modalities to describe com- plex driving scenes. To fill in this gap, we propose a novel framework, X-DRIVE, to model the joint distribution of point clouds and multi-view images via a dual- branch latent diffusion model architecture. Considering the distinct geometrical spaces of the two modalities, X-DRIVE conditions the synthesis of each modality on the corresponding local regions from the other modality, ensuring better alignment and realism. To further handle the spatial ambiguity during denoising, we design the cross-modality condition module based on epipolar lines to adaptively learn the cross-modality local correspondence. Besides, X-DRIVE allows for controllable generation through multi-level input conditions, including text, bounding box, image, and point clouds. Extensive results demonstrate the high-fidelity synthetic results of X-DRIVE for both point clouds and multi-view images, adhering to input conditions while ensuring reliable cross-modality consistency. Our code will be made publicly available at https://github.com/yichen928/X-Drive. | diffusion models; multi-modality data; autonomous driving | We propose a novel framework for the joint generation of point clouds and multi-view images with cross-modality consistency. | 2,234 | null | [
0.02597857266664505,
-0.11300241947174072,
0.014466783963143826,
-0.03088950552046299,
0.1096106693148613,
-0.016153499484062195,
0.026775268837809563,
-0.07878046482801437,
0.10416147857904434,
-0.030144309625029564,
-0.004344668705016375,
-0.0664994865655899,
0.03588613122701645,
0.02018859051167965,
0.0376729778945446,
0.030026357620954514,
0.028803473338484764,
0.05414285883307457,
-0.09026864916086197,
0.06467484682798386,
0.031035952270030975,
0.03444360941648483,
-0.03375682607293129,
0.0004970818408764899,
0.013786885887384415,
0.0933130756020546,
0.07124023884534836,
-0.030532199889421463,
0.0004516162152867764,
-0.0141405975446105,
0.044281914830207825,
0.03826121613383293,
-0.10523601621389389,
0.046449221670627594,
0.006402865517884493,
-0.018485847860574722,
-0.010095486417412758,
-0.030547402799129486,
-0.05810912698507309,
-0.06676042079925537,
0.029011320322752,
0.02031700871884823,
-0.038147419691085815,
-0.005768996197730303,
0.07308000326156616,
0.0011845538392663002,
-0.006366540212184191,
-0.04182743281126022,
0.059167247265577316,
0.025816308334469795,
0.03382868692278862,
-0.06975122541189194,
-0.042236849665641785,
0.018498115241527557,
0.026789195835590363,
0.06328875571489334,
-0.014973572455346584,
-0.06872846186161041,
0.0436931811273098,
-0.028499333187937737,
-0.008365650661289692,
-0.022780001163482666,
0.010148250497877598,
-0.009257043711841106,
0.03263532370328903,
-0.04477040469646454,
-0.014806769788265228,
-0.01344176847487688,
-0.005722026340663433,
-0.036435842514038086,
-0.08442168682813644,
0.054942917078733444,
0.0075884913094341755,
-0.11900954693555832,
-0.030372675508260727,
-0.010577804408967495,
0.07330550253391266,
-0.05612584948539734,
-0.035077665001153946,
-0.06647749245166779,
0.07650139927864075,
0.11199896782636642,
0.046061623841524124,
0.08293664455413818,
0.015322960913181305,
-0.04751676321029663,
-0.025758396834135056,
0.07880628854036331,
0.02445138990879059,
-0.0416613332927227,
-0.06616923958063126,
-0.026423731818795204,
-0.09297602623701096,
0.003202821360900998,
0.0006594689330086112,
-0.008456498384475708,
0.02581280656158924,
-0.08962884545326233,
0.08931582421064377,
0.055722836405038834,
-0.0196406040340662,
-0.010836885310709476,
0.010817811824381351,
0.0013876353623345494,
0.02392006851732731,
-0.06693288683891296,
0.05968083441257477,
0.02860897220671177,
-0.06663649529218674,
-0.05723295360803604,
0.04330690950155258,
-0.08204334229230881,
0.0207205880433321,
0.042915377765893936,
0.002919568680226803,
-0.08785513043403625,
0.05212981626391411,
0.015786757692694664,
0.004617339465767145,
-0.0506150908768177,
-0.08757367730140686,
-0.06886875629425049,
-0.031116817146539688,
0.06745094805955887,
0.045168567448854446,
-0.05751912295818329,
-0.007084956392645836,
1.594011091716442e-34,
-0.0033162692561745644,
0.026473218575119972,
0.07454132288694382,
0.0843140259385109,
0.12328457832336426,
0.010565754026174545,
-0.009108955040574074,
-0.01474014762789011,
-0.05466509982943535,
0.05080416798591614,
-0.03784425929188728,
-0.03622571751475334,
-0.039321016520261765,
0.03632178157567978,
0.04741261526942253,
-0.03603855520486832,
-0.03699574992060661,
-0.010660950094461441,
-0.01417719665914774,
0.08513785898685455,
-0.00553326727822423,
-0.026912013068795204,
-0.02967390976846218,
0.03336312994360924,
-0.02503720484673977,
0.041685573756694794,
0.05806097015738487,
-0.01313507929444313,
0.06384684890508652,
-0.013432727195322514,
-0.1185871884226799,
0.017945032566785812,
-0.027899034321308136,
0.02963346615433693,
-0.01683073118329048,
0.042094886302948,
-0.019609635695815086,
-0.037173569202423096,
-0.021185163408517838,
0.01128776278346777,
0.04319225251674652,
0.04882566258311272,
-0.02891407161951065,
-0.036617014557123184,
-0.1065247431397438,
-0.039818935096263885,
0.013281205669045448,
0.07598426192998886,
-0.06699251383543015,
0.0011679341550916433,
0.01902969554066658,
-0.036083612591028214,
-0.021809961646795273,
-0.07178248465061188,
-0.08122305572032928,
-0.04086478799581528,
-0.002206966746598482,
-0.022473158314824104,
0.06647611409425735,
-0.0005061478586867452,
-0.02862916700541973,
-0.059718675911426544,
-0.04192395880818367,
-0.018367435783147812,
0.023772140964865685,
0.02559073269367218,
0.022187357768416405,
0.05353774502873421,
0.05009891465306282,
0.03472859784960747,
-0.07378704100847244,
-0.002228141762316227,
-0.030975915491580963,
-0.02408858761191368,
0.10157011449337006,
-0.012950221076607704,
-0.0720083937048912,
0.03972920775413513,
0.09160011261701584,
0.10897602140903473,
-0.09752504527568817,
0.04525528848171234,
0.00819895975291729,
-0.03924916684627533,
-0.02821275219321251,
-0.010126003995537758,
-0.03768397495150566,
-0.03693687543272972,
-0.0754244476556778,
-0.0743543803691864,
-0.025451822206377983,
-0.04783153533935547,
0.0007909195846877992,
-0.010664976201951504,
-0.014907865785062313,
-3.188140880811668e-34,
0.05785742029547691,
-0.03747675195336342,
-0.02165202982723713,
0.038505010306835175,
-0.012280257418751717,
-0.024562016129493713,
0.006035235244780779,
-0.0298444926738739,
0.10817188769578934,
-0.015345273539423943,
-0.0378827229142189,
0.00352152856066823,
0.026453163474798203,
-0.023731602355837822,
0.013862048275768757,
0.05478517711162567,
0.010765600949525833,
-0.009885372593998909,
-0.077613465487957,
-0.02217099815607071,
0.04676848277449608,
0.010046718642115593,
0.014058244414627552,
0.03235065937042236,
-0.1144358366727829,
0.07075870782136917,
-0.031832292675971985,
0.05571165308356285,
0.028928928077220917,
-0.02305871993303299,
-0.00004156064096605405,
-0.031563740223646164,
0.05935893952846527,
-0.04564805328845978,
-0.058912791311740875,
0.06571156531572342,
0.006457035895437002,
0.012577695772051811,
-0.10699304193258286,
0.06724026799201965,
0.015330478549003601,
-0.07369956374168396,
-0.0007117607747204602,
0.000973450078163296,
0.03070421889424324,
0.020396973937749863,
0.03312051668763161,
-0.009407073259353638,
0.022893356159329414,
0.10395864397287369,
-0.022489162161946297,
0.06658043712377548,
-0.07012949883937836,
0.09163478761911392,
-0.01820032112300396,
-0.06386873871088028,
-0.10347535461187363,
-0.04534626007080078,
0.09538494050502777,
0.0026847838889807463,
-0.012199575081467628,
-0.07525166869163513,
-0.013796037063002586,
0.007381277158856392,
0.02114233188331127,
-0.0919441431760788,
-0.035552170127630234,
-0.038257427513599396,
0.01511288806796074,
0.09631364047527313,
0.036552052944898605,
-0.05817609280347824,
0.0008146645268425345,
-0.004497657995671034,
0.03620939701795578,
-0.08680815249681473,
0.05030015856027603,
-0.044307224452495575,
0.056181401014328,
-0.016133934259414673,
0.057732779532670975,
-0.0219360813498497,
0.06078019365668297,
0.1603565216064453,
0.030972272157669067,
0.1213449239730835,
0.022343432530760765,
-0.002154001733288169,
0.025671500712633133,
0.0037142925430089235,
0.01760195754468441,
0.010573397390544415,
0.0056558214128017426,
0.04130328819155693,
-0.05240172520279884,
-4.2291766533253394e-8,
-0.09090548753738403,
0.06568656861782074,
0.011277955956757069,
-0.02185136452317238,
-0.0710998922586441,
-0.09077125787734985,
0.10334022343158722,
0.027386728674173355,
0.02290356159210205,
-0.07421836256980896,
0.056415874511003494,
-0.09972492605447769,
-0.006466832477599382,
-0.05583890154957771,
-0.007593206129968166,
0.07052277028560638,
0.04668908566236496,
0.046084214001894,
0.030449997633695602,
0.03130574896931648,
-0.0509713850915432,
0.03056892566382885,
0.06882942467927933,
0.01307812612503767,
0.07945957034826279,
-0.006937271915376186,
-0.029200304299592972,
0.015131804160773754,
0.10546664893627167,
-0.06891096383333206,
-0.023471403867006302,
-0.010880032554268837,
0.055724725127220154,
0.07769551873207092,
-0.02981145866215229,
0.041872330009937286,
0.059202224016189575,
0.011971646919846535,
-0.008591809310019016,
0.0008686299552209675,
-0.032635513693094254,
0.05559276044368744,
-0.03400484099984169,
-0.02258058451116085,
0.03591650724411011,
0.11211762577295303,
0.11074720323085785,
-0.029750294983386993,
-0.027501223608851433,
0.0018708668649196625,
0.00805697962641716,
0.05596296116709709,
0.0028336667455732822,
0.06247708201408386,
0.027822647243738174,
0.041823916137218475,
0.03463220223784447,
-0.04368274658918381,
0.09021841734647751,
0.020065344870090485,
-0.03512469679117203,
-0.005766261834651232,
-0.08113773167133331,
-0.02180294506251812
] |
Selective Attention Improves Transformer | https://openreview.net/forum?id=v0FzmPCd1e | [
"Yaniv Leviathan",
"Matan Kalman",
"Yossi Matias"
] | Poster | Unneeded elements in the attention’s context degrade performance. We introduce Selective Attention, a simple parameter-free change to the standard attention mechanism which reduces attention to unneeded elements. Selective attention consistently improves language modeling and downstream task performance in a variety of model sizes and context lengths. For example, transformers trained with the language modeling objective on C4 with selective attention perform language modeling equivalently to standard transformers with ~2X more heads and parameters in their attention modules. Selective attention also allows decreasing the size of the attention’s context buffer, leading to meaningful reductions in the memory and compute requirements during inference. For example, transformers trained on C4 with context sizes of 512, 1,024, and 2,048 need 16X, 25X, and 47X less memory for their attention module, respectively, when equipped with selective attention, as those without selective attention, with the same validation perplexity. | selective attention, attention, transformer, llm, language model | null | 2,232 | 2410.02703 | [
0.038677725940942764,
0.02351812645792961,
0.014023317955434322,
0.026506666094064713,
0.03640282526612282,
-0.012205521576106548,
0.009088685736060143,
0.0690024197101593,
0.032205451279878616,
-0.08061867207288742,
-0.02672378160059452,
-0.026800349354743958,
0.031138326972723007,
0.017233487218618393,
0.03898623213171959,
0.03097587078809738,
0.09247545152902603,
0.07069370150566101,
-0.12124916911125183,
-0.0644778311252594,
0.039934780448675156,
-0.026061180979013443,
0.02712666615843773,
-0.05274023860692978,
-0.019413389265537262,
-0.023577915504574776,
-0.0664997473359108,
-0.08200016617774963,
0.014528841711580753,
-0.04458087682723999,
0.04222545772790909,
0.053970322012901306,
0.013863748870790005,
0.06284917891025543,
-0.04646020755171776,
0.0571550652384758,
-0.10544759780168533,
0.0077332681976258755,
0.02206592634320259,
-0.051672499626874924,
-0.004512061830610037,
0.022707197815179825,
-0.03062761016190052,
0.056935135275125504,
0.021916063502430916,
-0.008703983388841152,
0.03666751831769943,
-0.015635136514902115,
-0.004420912358909845,
-0.09606504440307617,
-0.03707787021994591,
-0.06310795247554779,
-0.037097591906785965,
0.055582474917173386,
-0.04056897759437561,
0.029108814895153046,
0.01007379125803709,
0.040115367621183395,
0.020640477538108826,
-0.012943888083100319,
-0.12300702184438705,
-0.08652763813734055,
-0.008142747916281223,
-0.053517114371061325,
-0.05695980042219162,
-0.017909718677401543,
0.07277964800596237,
-0.01656149886548519,
-0.08382143825292587,
-0.010712129063904285,
-0.027785325422883034,
0.07592036575078964,
0.03035556711256504,
0.016853686422109604,
0.07989832758903503,
-0.0031899046152830124,
0.011699420399963856,
-0.08099246770143509,
0.08462946116924286,
-0.04248474910855293,
-0.003929798025637865,
-0.0014083596179261804,
-0.019607895985245705,
-0.016355764120817184,
0.12484755367040634,
0.03417393937706947,
-0.013664625585079193,
-0.004236365668475628,
-0.009711985476315022,
0.02731902338564396,
-0.016985135152935982,
-0.12213534861803055,
0.03533145412802696,
0.008250094950199127,
-0.0036641110200434923,
0.02170214243233204,
0.01767979934811592,
-0.04615309089422226,
-0.08053971827030182,
0.07726730406284332,
0.060472261160612106,
-0.012429669499397278,
0.05907168611884117,
0.037056826055049896,
-0.025216732174158096,
-0.06104319915175438,
0.02599950321018696,
0.0698060691356659,
-0.029226958751678467,
-0.040758680552244186,
0.04306311160326004,
0.05323924869298935,
-0.02820175141096115,
-0.044605884701013565,
0.03822453320026398,
0.010825343430042267,
0.020114349201321602,
-0.03658408299088478,
-0.06122415140271187,
0.05395834520459175,
0.012361707165837288,
0.03805290535092354,
-0.020563600584864616,
0.0034036452416330576,
0.05545206740498543,
-0.018576495349407196,
-0.0631239041686058,
3.506466587724267e-33,
-0.021713661029934883,
0.03279002755880356,
0.013311217539012432,
0.0550006665289402,
0.0011486068833619356,
0.002481975592672825,
0.041222866624593735,
0.03307199105620384,
0.032425593584775925,
-0.001285172300413251,
-0.014192268252372742,
0.03702302277088165,
-0.047344036400318146,
0.056431904435157776,
0.005987725220620632,
-0.007329688407480717,
0.013040806166827679,
0.07543952763080597,
-0.04069460555911064,
-0.0097378334030509,
0.003947508055716753,
0.02083570882678032,
0.04010971263051033,
-0.06263743340969086,
-0.049521319568157196,
-0.01655389554798603,
0.032272662967443466,
-0.11929293721914291,
-0.04849548265337944,
0.005749743431806564,
-0.14299578964710236,
0.04936211183667183,
-0.03298313543200493,
0.03618238493800163,
0.05677773803472519,
-0.0777929350733757,
-0.002125194063410163,
-0.06164295971393585,
0.056925076991319656,
-0.008483881130814552,
0.002908328315243125,
0.1113976538181305,
0.02600996382534504,
0.02268029749393463,
-0.030805421993136406,
-0.08243158459663391,
0.014027297496795654,
0.02198946848511696,
-0.017084550112485886,
0.02254604920744896,
0.06478238850831985,
0.03862825781106949,
-0.01758839748799801,
-0.028012368828058243,
0.05819797143340111,
0.02312776818871498,
0.08572611212730408,
0.05862722545862198,
0.03187178447842598,
0.05099743977189064,
0.01177382841706276,
0.02219817414879799,
-0.04625307396054268,
0.09351852536201477,
0.09856169670820236,
0.008363150991499424,
-0.019034992903470993,
0.08118966966867447,
0.009375188499689102,
0.008334818296134472,
-0.04986079782247543,
-0.05533292144536972,
0.025151096284389496,
0.03554766625165939,
0.04311446100473404,
-0.025742461904883385,
0.07911890745162964,
-0.0352083221077919,
0.0166492722928524,
-0.019865598529577255,
0.03128160163760185,
0.01156364195048809,
-0.00008547813922632486,
-0.035937823355197906,
-0.014425283297896385,
-0.06581151485443115,
0.06501851975917816,
-0.049447666853666306,
-0.008424495346844196,
-0.06461641192436218,
0.005735023878514767,
0.008372509852051735,
-0.0522778183221817,
-0.04651735723018646,
-0.048950884491205215,
-2.2685938944916253e-33,
-0.00017697794828563929,
0.004767668433487415,
-0.0769912451505661,
0.06937699019908905,
-0.0704890638589859,
-0.023864086717367172,
0.0719655230641365,
0.010312650352716446,
-0.06914810836315155,
-0.0054406411945819855,
0.0357641726732254,
-0.030573764815926552,
0.06772682815790176,
-0.042721569538116455,
0.051539346575737,
-0.022077852860093117,
-0.03051942214369774,
-0.00795232504606247,
-0.026151524856686592,
0.03250488266348839,
-0.014350377023220062,
0.10056156665086746,
-0.04950163513422012,
-0.02300502359867096,
-0.11461001634597778,
-0.001098012551665306,
-0.11634359508752823,
0.04174165427684784,
0.06841320544481277,
-0.05781440809369087,
-0.06874091178178787,
0.051548730581998825,
0.04401946812868118,
0.03800526633858681,
-0.030828187242150307,
0.004508765414357185,
-0.0016354770632460713,
0.037867892533540726,
-0.031958986073732376,
0.074777752161026,
0.058358024805784225,
-0.019813798367977142,
-0.0044516632333397865,
0.07490866631269455,
-0.09464587271213531,
0.005263347644358873,
-0.02856718748807907,
-0.0870431512594223,
-0.030888643115758896,
0.017999961972236633,
-0.013727717101573944,
0.016565470024943352,
-0.06319554895162582,
-0.0676291435956955,
-0.03300655633211136,
-0.08085907250642776,
0.015578774735331535,
-0.05532659590244293,
0.10863572359085083,
-0.07096603512763977,
0.00769373495131731,
-0.00973389483988285,
0.06917361915111542,
-0.1228979304432869,
0.019598262384533882,
-0.03135531023144722,
-0.034153491258621216,
-0.06603141129016876,
0.06784915924072266,
0.014423402957618237,
0.03342292830348015,
0.019157633185386658,
0.010348515585064888,
-0.044325090944767,
-0.03312728554010391,
0.046042054891586304,
0.020601991564035416,
-0.05045926943421364,
0.00914119090884924,
-0.04354126751422882,
-0.08320602774620056,
0.04530996084213257,
-0.04212887957692146,
-0.016565658152103424,
0.0009211210999637842,
0.10565344244241714,
0.047789692878723145,
0.014369681477546692,
0.014119652099907398,
0.007498574908822775,
-0.05113418400287628,
0.07564613223075867,
0.04783032089471817,
0.05601850524544716,
-0.09427978098392487,
-4.431352351730311e-8,
-0.08975686132907867,
0.06667126715183258,
-0.025966864079236984,
0.04275784268975258,
0.051834240555763245,
-0.10883219540119171,
-0.0566781610250473,
0.14092066884040833,
-0.02464224584400654,
0.04159780591726303,
0.0764707401394844,
-0.021891899406909943,
0.003567986423149705,
-0.014440115541219711,
0.07689941674470901,
0.051157332956790924,
0.053503770381212234,
0.0171903595328331,
-0.020014550536870956,
-0.06921384483575821,
0.0009908300125971437,
0.05004512891173363,
-0.03387404605746269,
-0.027085378766059875,
0.08692613989114761,
-0.07879225164651871,
0.019342917948961258,
0.12277049571275711,
0.030464887619018555,
-0.041612930595874786,
0.03430245444178581,
0.0032645626924932003,
-0.04991926997900009,
-0.049703627824783325,
0.0012473994866013527,
0.10694430023431778,
0.05709661915898323,
-0.017937473952770233,
0.014111915603280067,
0.050664015114307404,
0.055689677596092224,
0.01046710554510355,
-0.06221262365579605,
0.0861852690577507,
-0.015264254063367844,
0.010904157534241676,
-0.03260847181081772,
-0.1563071310520172,
-0.024372797459363937,
-0.06432277709245682,
0.10219664871692657,
-0.015579276718199253,
-0.045959196984767914,
0.0433100126683712,
-0.00042602402390912175,
0.04199982061982155,
0.043660249561071396,
-0.05247816443443298,
-0.06931351125240326,
0.015272504650056362,
0.03127938508987427,
0.06802019476890564,
-0.05792738497257233,
-0.003272649832069874
] |
STAR: Stability-Inducing Weight Perturbation for Continual Learning | https://openreview.net/forum?id=6N5OM5Duuj | [
"Masih Eskandar",
"Tooba Imtiaz",
"Davin Hill",
"Zifeng Wang",
"Jennifer Dy"
] | Poster | Humans can naturally learn new and varying tasks in a sequential manner.
Continual learning is a class of learning algorithms that updates its learned model as it sees new data (on potentially new tasks) in a sequence.
A key challenge in continual learning is that as the model is updated to learn new tasks, it becomes susceptible to \textit{catastrophic forgetting}, where knowledge of previously learned tasks is lost. A popular approach to mitigate forgetting during continual learning is to maintain a small buffer of previously-seen samples, and to replay them during training. However, this approach is limited by the small buffer size and, while forgetting is reduced, it is still present. In this paper, we propose
a novel loss function STAR that exploits the worst-case parameter perturbation that reduces the KL-divergence of model predictions with that of its local parameter neighborhood to promote stability and alleviate forgetting. STAR can be combined with almost any existing rehearsal-based methods as a plug-and-play component. We empirically show that STAR consistently improves performance of existing methods by up to $\sim15\\%$ across varying baselines, and achieves superior or competitive accuracy to that of state-of-the-art methods aimed at improving rehearsal-based continual learning. Our implementation is available at https://github.com/Gnomy17/STAR_CL. | Continual Learning, Deep Learning, Weight Perturbation, Representation Learning | In Continual Learning, we regularize the divergence of the model output distribution with regards to past data via worst-case weight perturbation, alleviating forgetting in future updates. | 2,224 | 2503.01595 | [
-0.07278987765312195,
-0.04905490204691887,
0.02458341419696808,
0.10072645545005798,
-0.011883681640028954,
0.02625594288110733,
-0.02387865073978901,
-0.06040813773870468,
0.03543728217482567,
-0.017683327198028564,
-0.06198698654770851,
0.045125629752874374,
0.030086025595664978,
-0.006379700265824795,
-0.08555659651756287,
-0.01094883307814598,
-0.0009191475110128522,
0.07732730358839035,
-0.017771700397133827,
-0.03184809163212776,
-0.013004905544221401,
0.023239845409989357,
0.055388204753398895,
0.05755595117807388,
0.005982017610222101,
-0.02903313934803009,
-0.038645993918180466,
-0.05615982040762901,
0.06430497765541077,
-0.07452418655157089,
0.03036707267165184,
0.029972828924655914,
-0.10836461186408997,
-0.025052694603800774,
-0.10183105617761612,
0.03725077956914902,
-0.08849436789751053,
0.02922132983803749,
-0.0057755508460104465,
-0.03680261969566345,
0.009123647585511208,
0.0003054023254662752,
-0.038276780396699905,
0.020824307575821877,
0.03396932780742645,
0.01466620247811079,
0.0022769079077988863,
-0.0880565419793129,
0.049768563359975815,
0.04835081472992897,
-0.04345892742276192,
-0.024596557021141052,
-0.00911041721701622,
0.01161140762269497,
0.07821664214134216,
-0.0082201873883605,
0.06404048949480057,
0.04585446044802666,
-0.07784473896026611,
-0.007864007726311684,
0.03180839866399765,
-0.1037825271487236,
-0.028631487861275673,
-0.03582656756043434,
-0.07133372873067856,
-0.054218389093875885,
0.030205363407731056,
0.05151388421654701,
0.010442351922392845,
0.07522083073854446,
-0.03878038376569748,
0.09240096062421799,
-0.053279414772987366,
0.03197598084807396,
0.06554191559553146,
0.048178110271692276,
0.05873064696788788,
-0.09496340155601501,
0.038704995065927505,
0.013643286190927029,
0.021155107766389847,
0.0009920578449964523,
-0.02205456979572773,
-0.06858192384243011,
0.11316616088151932,
-0.01864473894238472,
0.05103643983602524,
0.0001240355195477605,
0.06145661324262619,
-0.039596617221832275,
-0.010965488851070404,
0.011121774092316628,
-0.0165812186896801,
-0.04568073898553848,
-0.0070198532193899155,
0.029973678290843964,
0.014267445541918278,
-0.055533986538648605,
-0.0075799464248120785,
0.10036761313676834,
-0.020766623318195343,
0.08463533967733383,
-0.027385476976633072,
0.011561715044081211,
0.023749003186821938,
-0.0700555294752121,
-0.00259411265142262,
0.060551028698682785,
0.0632256418466568,
-0.025417786091566086,
0.04810437560081482,
-0.003895207541063428,
-0.027809591963887215,
0.10550136119127274,
0.034792765974998474,
-0.011283142492175102,
0.014760201796889305,
-0.02922244369983673,
-0.06431157141923904,
0.11200488358736038,
0.03489510715007782,
0.0310390442609787,
0.009095562621951103,
0.037791118025779724,
-0.014775658026337624,
-0.002777093555778265,
-0.04613015428185463,
4.614436594108987e-34,
0.02177472971379757,
0.014762978069484234,
0.06726264953613281,
-0.02287592925131321,
0.01658596843481064,
-0.046886175870895386,
-0.002425124403089285,
0.057318173348903656,
-0.01650167815387249,
-0.08566464483737946,
0.0004269062483217567,
-0.033976536244153976,
-0.039049435406923294,
0.08778560906648636,
0.08596491068601608,
-0.03454894199967384,
-0.058257125318050385,
0.03420451655983925,
-0.005176458042114973,
-0.047713421285152435,
0.07205687463283539,
-0.037080712616443634,
-0.0027177894953638315,
-0.11443450301885605,
0.011638890951871872,
0.034759629517793655,
0.08211203664541245,
0.03985889256000519,
-0.047426845878362656,
0.022715620696544647,
-0.10886769741773605,
0.004347064532339573,
0.021533234044909477,
0.033686552196741104,
-0.01818324252963066,
-0.027293477207422256,
-0.016017908230423927,
-0.026280390098690987,
0.0201357901096344,
-0.06315219402313232,
0.008305473253130913,
0.05064062774181366,
-0.00900768581777811,
-0.010715336538851261,
-0.03740045800805092,
-0.17735636234283447,
0.03105301782488823,
0.02542322874069214,
-0.05624323710799217,
-0.05383521690964699,
-0.012705376371741295,
0.020301351323723793,
-0.06792546808719635,
-0.04604852572083473,
-0.06830570846796036,
0.016530459746718407,
0.0016896474407985806,
-0.02496262453496456,
-0.03468956798315048,
0.035077620297670364,
0.062445491552352905,
0.03265607729554176,
-0.04510995373129845,
0.11437518149614334,
0.021894479170441628,
0.0776950791478157,
0.02811710722744465,
0.058939095586538315,
-0.0288414154201746,
-0.022824501618742943,
0.008951319381594658,
-0.013529486954212189,
-0.0009593633003532887,
-0.07054028660058975,
0.06743835657835007,
-0.07266975194215775,
0.09223660826683044,
-0.10611691325902939,
-0.022369446232914925,
-0.009313601069152355,
0.07369556277990341,
-0.05295277386903763,
-0.024292556568980217,
-0.04400099068880081,
-0.034566331654787064,
-0.09450089186429977,
0.06906618922948837,
-0.0644126906991005,
-0.04529695585370064,
0.023578457534313202,
-0.07674282789230347,
-0.026327263563871384,
0.01229427196085453,
-0.02744768187403679,
-0.042371965944767,
-2.1896820403659636e-33,
-0.021536825224757195,
-0.02130894362926483,
-0.03829480707645416,
0.06593368947505951,
0.03362789377570152,
-0.019592076539993286,
-0.05082406476140022,
-0.0012194510782137513,
-0.1372867077589035,
-0.1385107785463333,
-0.05240211263298988,
-0.014890095219016075,
-0.026556631550192833,
0.0335700549185276,
0.04503043740987778,
-0.02180439420044422,
-0.061526212841272354,
0.04092331975698471,
-0.037408869713544846,
-0.0019148987485095859,
0.003576921531930566,
0.05715087056159973,
-0.03588571399450302,
0.042563073337078094,
-0.010670071467757225,
-0.012418132275342941,
-0.012967335060238838,
0.11572084575891495,
-0.04008777439594269,
-0.01258433610200882,
-0.025129608809947968,
-0.046511195600032806,
0.0018292935565114021,
-0.04864690080285072,
-0.004429914522916079,
0.03739123046398163,
0.06443453580141068,
0.04015921428799629,
-0.05185111612081528,
0.11753703653812408,
0.08965890854597092,
0.008493988774716854,
-0.020984981209039688,
0.013416815549135208,
0.051566045731306076,
-0.04957664757966995,
-0.052703607827425,
0.050961434841156006,
0.006382445804774761,
0.03657357767224312,
0.03480757027864456,
-0.011445969343185425,
-0.1421871930360794,
-0.018459031358361244,
-0.04984947666525841,
-0.008098135702311993,
-0.008499196730554104,
-0.010365145280957222,
0.04029981419444084,
0.0008035041391849518,
-0.06639358401298523,
-0.03861743584275246,
0.011143562383949757,
0.0018547100480645895,
0.024996081367135048,
-0.003086724551394582,
0.04529527947306633,
0.06046992912888527,
0.024869492277503014,
0.06949447095394135,
0.007009298074990511,
0.09009437263011932,
-0.02906516194343567,
-0.010728408582508564,
0.03015020117163658,
0.02358606830239296,
-0.05720088630914688,
-0.035657815635204315,
-0.019961485639214516,
-0.02673930488526821,
-0.019752318039536476,
-0.014931571669876575,
0.027309037744998932,
0.04992106184363365,
0.08758537471294403,
0.1050630584359169,
0.005811635404825211,
0.13154646754264832,
-0.008000249043107033,
-0.021111782640218735,
-0.03857135400176048,
0.031214363873004913,
-0.03100193478167057,
-0.02200508303940296,
-0.042332228273153305,
-4.646115314699273e-8,
-0.05426313355565071,
0.07080673426389694,
0.01311804261058569,
0.07738284021615982,
0.09181157499551773,
-0.06419361382722855,
-0.02457576058804989,
0.03356161341071129,
-0.012412597425282001,
-0.04749653860926628,
0.08204760402441025,
0.003015811089426279,
0.05085116997361183,
-0.0404479093849659,
-0.027259791269898415,
0.08300331234931946,
0.041404858231544495,
0.01910736970603466,
-0.043248724192380905,
0.012354911305010319,
0.01271874736994505,
0.00747306365519762,
0.02170572243630886,
-0.042495518922805786,
0.0024416653905063868,
-0.09444725513458252,
0.027956483885645866,
0.17599792778491974,
0.04833035543560982,
-0.02619113028049469,
0.005481458269059658,
-0.0022268714383244514,
0.011555097997188568,
-0.0046843234449625015,
0.02174891158938408,
0.05942012742161751,
0.12687449157238007,
-0.03008614107966423,
0.023941555991768837,
0.045936569571495056,
-0.0010114978067576885,
0.04904724285006523,
0.015989623963832855,
0.00523589551448822,
-0.008155198767781258,
0.024298084899783134,
-0.043622419238090515,
-0.07947961241006851,
0.01934451051056385,
-0.0543525293469429,
-0.0453006885945797,
0.03439175337553024,
-0.06908909976482391,
0.07396925240755081,
0.10832236707210541,
0.0035870117135345936,
-0.041136935353279114,
-0.04415284842252731,
-0.0037265813443809748,
0.0542689673602581,
0.06500018388032913,
-0.007291701622307301,
-0.03908536583185196,
0.0683172419667244
] |
OMNI-EPIC: Open-endedness via Models of human Notions of Interestingness with Environments Programmed in Code | https://openreview.net/forum?id=Y1XkzMJpPd | [
"Maxence Faldor",
"Jenny Zhang",
"Antoine Cully",
"Jeff Clune"
] | Poster | Open-ended and AI-generating algorithms aim to continuously generate and solve increasingly complex tasks indefinitely, offering a promising path toward more general intelligence. To accomplish this grand vision, learning must occur within a vast array of potential tasks. Existing approaches to automatically generating environments are constrained within manually predefined, often narrow distributions of environments, limiting their ability to create any learning environment. To address this limitation, we introduce a novel framework, OMNI-EPIC, that augments previous work in Open-endedness via Models of human Notions of Interestingness (OMNI) with Environments Programmed in Code (EPIC). OMNI-EPIC leverages foundation models to autonomously generate code specifying the next learnable (i.e., not too easy or difficult for the agent’s current skill set) and interesting (e.g., worthwhile and novel) tasks. OMNI-EPIC generates both environments (e.g., an obstacle course) and reward functions (e.g., progress through the obstacle course quickly without touching red objects), enabling it, in principle, to create any simulatable learning task. We showcase the explosive creativity of OMNI-EPIC, which continuously innovates to suggest new, interesting learning challenges. We also highlight how OMNI-EPIC can adapt to reinforcement learning agents’ learning progress, generating tasks that are of suitable difficulty. Overall, OMNI-EPIC has the potential to endlessly create learnable and interesting environments, further propelling the development of self-improving AI systems and AI-Generating Algorithms. | Open-endedness, Environment Generation, Reinforcement Learning | A framework that uses large language models to automatically generate diverse and interesting learning environments through code. | 2,223 | null | [
-0.05300794541835785,
-0.046577826142311096,
-0.013667881488800049,
0.06961695104837418,
0.08918190002441406,
-0.07685468345880508,
0.03649688884615898,
-0.04083780571818352,
-0.03339776024222374,
0.013704705983400345,
-0.06436843425035477,
-0.10066328197717667,
0.023965273052453995,
0.05030994862318039,
0.04057735204696655,
0.013433481566607952,
0.03659641742706299,
-0.007310555782169104,
-0.04609639570116997,
0.0060762930661439896,
0.037827350199222565,
0.007784971501678228,
0.0015880420105531812,
-0.012860853224992752,
-0.05310087278485298,
-0.03746822476387024,
0.03960663825273514,
-0.05373047664761543,
0.0660623237490654,
-0.04982464760541916,
0.09245920926332474,
-0.023331796750426292,
0.027763742953538895,
-0.017632871866226196,
-0.006625508889555931,
0.07937917858362198,
-0.10420932620763779,
-0.0004199405084364116,
-0.06674244999885559,
-0.09253451973199844,
-0.04658319056034088,
0.010390380397439003,
-0.025238754227757454,
0.026029326021671295,
0.08322905749082565,
-0.010061463341116905,
-0.03759482875466347,
-0.1163993701338768,
0.01701858825981617,
-0.03131776675581932,
-0.1160014271736145,
-0.09222348034381866,
-0.035517219454050064,
-0.04258863627910614,
0.025862306356430054,
0.05250968784093857,
0.01464616134762764,
-0.08904609829187393,
-0.013378267176449299,
-0.10454478114843369,
0.012437155470252037,
-0.07057487964630127,
-0.005654971580952406,
-0.03883221745491028,
-0.00364705384708941,
-0.10023950785398483,
-0.00016419838357251137,
0.07236305624246597,
0.011468945071101189,
-0.02988174557685852,
0.019825657829642296,
0.1010081022977829,
-0.004747479222714901,
0.04631757736206055,
0.08199108392000198,
0.04671711474657059,
-0.040103472769260406,
-0.001865229569375515,
0.0450284481048584,
0.013442420400679111,
-0.019492577761411667,
-0.041937075555324554,
-0.0193435400724411,
0.09333695471286774,
0.0809035375714302,
0.02704000473022461,
0.04794235900044441,
0.05297526344656944,
0.08166974782943726,
0.06690913438796997,
-0.04769989475607872,
-0.042355235666036606,
0.04469665512442589,
0.013262772001326084,
0.029026048257946968,
0.007894706912338734,
0.036229267716407776,
-0.10982085019350052,
-0.03518088534474373,
0.11706668138504028,
-0.03677738457918167,
0.06962154060602188,
-0.030531104654073715,
0.0002828961005434394,
0.024950122460722923,
0.033394087105989456,
0.05672633647918701,
-0.01556901540607214,
0.07174256443977356,
-0.03732488304376602,
0.026526987552642822,
-0.023384692147374153,
0.06463836878538132,
-0.05184275656938553,
-0.00316042872145772,
0.048424478620290756,
-0.010734349489212036,
0.022422296926379204,
0.024241285398602486,
0.06714737415313721,
0.08756771683692932,
0.03440895304083824,
-0.005168300122022629,
0.07477867603302002,
0.05965005233883858,
-0.0684533342719078,
0.006183706223964691,
-6.550750837224288e-34,
0.05268777534365654,
-0.027118414640426636,
0.017296375706791878,
0.06359301507472992,
0.11564090847969055,
-0.046894822269678116,
0.031976744532585144,
0.03596259281039238,
-0.03279365971684456,
-0.003972237464040518,
0.03432919457554817,
0.06514301896095276,
-0.04967447742819786,
0.11626113951206207,
0.050872813910245895,
-0.08470422774553299,
-0.034572262316942215,
0.028861153870821,
-0.023927388712763786,
-0.0259938295930624,
0.04989944398403168,
-0.08116015791893005,
-0.013536226004362106,
-0.053370777517557144,
0.043253093957901,
0.06888756901025772,
0.06942388415336609,
-0.04630490764975548,
0.017303938046097755,
-0.002617497695609927,
-0.02838936634361744,
0.007089682854712009,
-0.00912830512970686,
0.10698678344488144,
-0.0665023997426033,
0.043366868048906326,
-0.0748843401670456,
-0.001264719758182764,
0.03697993978857994,
0.02933209389448166,
0.027395784854888916,
0.027441781014204025,
0.025065837427973747,
0.0048980130814015865,
-0.04298293590545654,
-0.07511799782514572,
0.09177850186824799,
0.051934849470853806,
-0.09543922543525696,
-0.017100181430578232,
-0.025872765108942986,
-0.028128361329436302,
-0.038202133029699326,
-0.09107622504234314,
-0.08502748608589172,
0.036123987287282944,
-0.00244109146296978,
0.006796867586672306,
0.02938326634466648,
0.011698240414261818,
0.0027356520295143127,
0.041961196810007095,
-0.01910475082695484,
0.04691026732325554,
-0.00324600120075047,
0.03157433867454529,
0.02204900234937668,
0.024411451071500778,
0.11468987166881561,
0.042328137904405594,
-0.00377199100330472,
0.030085086822509766,
-0.029832076281309128,
-0.034228187054395676,
0.013463580049574375,
0.028078565374016762,
0.05213924124836922,
-0.09936612099409103,
0.04166938737034798,
-0.03401797264814377,
-0.017410574480891228,
-0.040161337703466415,
-0.03791720047593117,
-0.009384955279529095,
-0.009382206946611404,
-0.0345548577606678,
-0.01396601740270853,
-0.11411232501268387,
-0.07501452416181564,
0.0012235565809533,
-0.03045630268752575,
-0.04855978116393089,
0.016347412019968033,
0.0012948407093062997,
-0.05630560219287872,
-9.907139857127166e-35,
0.015743592754006386,
-0.04842981696128845,
-0.09549632668495178,
0.009816540405154228,
0.02573849819600582,
-0.02834482118487358,
-0.05600746348500252,
-0.022001968696713448,
-0.047569822520017624,
-0.016814356669783592,
-0.08007904142141342,
0.040122125297784805,
0.06849982589483261,
-0.013089063577353954,
0.03444511815905571,
-0.07884220033884048,
-0.046082206070423126,
0.052540265023708344,
-0.04028721898794174,
0.13614998757839203,
-0.02669626474380493,
0.09829162061214447,
-0.09480754286050797,
-0.04915664717555046,
0.044354233890771866,
0.08926209807395935,
-0.019837185740470886,
0.04571756720542908,
-0.007423673756420612,
0.05063565820455551,
-0.049771279096603394,
-0.02733408659696579,
-0.05405714362859726,
-0.050495609641075134,
-0.01364693883806467,
0.08957638591527939,
0.05416306108236313,
0.003792861010879278,
-0.05758703872561455,
0.11356016993522644,
0.032621946185827255,
-0.098973348736763,
-0.030134575441479683,
0.012201083824038506,
-0.04607898369431496,
0.04903094470500946,
-0.05746116489171982,
0.039434764534235,
-0.0007132583996281028,
-0.03266894444823265,
-0.013669339939951897,
0.02881900779902935,
-0.10139723122119904,
-0.027044393122196198,
-0.04145817831158638,
-0.03990790992975235,
-0.06214272603392601,
-0.02657865174114704,
0.02562178671360016,
0.04028545692563057,
-0.05116518959403038,
-0.02714068628847599,
0.024814004078507423,
0.06581947207450867,
-0.03511858731508255,
-0.04372774809598923,
-0.0759168192744255,
0.03208832070231438,
-0.10315301269292831,
-0.023816915228962898,
0.007005027029663324,
0.010982398875057697,
-0.02471109665930271,
-0.03811334818601608,
0.02458902820944786,
0.037484508007764816,
-0.03055589646100998,
-0.021522630006074905,
0.06864701956510544,
-0.04969541355967522,
0.01729445345699787,
-0.021846452727913857,
0.03121405653655529,
0.044157832860946655,
0.01012223307043314,
0.045724861323833466,
0.024372583255171776,
0.07898062467575073,
0.006550170481204987,
0.05007128417491913,
-0.04404183104634285,
0.09133312106132507,
-0.011104829609394073,
0.08323632180690765,
-0.049544159322977066,
-4.0061177486450106e-8,
-0.060224056243896484,
0.005375819746404886,
-0.005344009026885033,
0.011569111607968807,
0.03195497393608093,
-0.015801187604665756,
-0.029344599694013596,
0.015239882282912731,
-0.07559564709663391,
-0.011874419637024403,
0.029968714341521263,
0.040622178465127945,
0.0320783294737339,
0.008880085311830044,
-0.0072709652595222,
0.08329276740550995,
0.014181152917444706,
0.004510808736085892,
-0.008770699612796307,
0.007957917638123035,
0.03725530952215195,
0.03362571448087692,
-0.00777962151914835,
-0.07771971821784973,
-0.0431397520005703,
-0.0858113020658493,
-0.046348463743925095,
0.010406174696981907,
0.06295902281999588,
0.04685194417834282,
0.051786959171295166,
0.03764517605304718,
0.0511423796415329,
0.004350753966718912,
0.05190616101026535,
0.06591939926147461,
-0.022166773676872253,
-0.09894561767578125,
-0.03431742265820503,
-0.0381375253200531,
0.0377839095890522,
0.0789804607629776,
0.017673619091510773,
-0.016933847218751907,
-0.013664967380464077,
0.007846100255846977,
-0.028438685461878777,
-0.11926503479480743,
0.0324227400124073,
-0.026671847328543663,
-0.0347224660217762,
-0.044445011764764786,
-0.04669361561536789,
-0.01959978975355625,
0.11370006948709488,
0.03563676401972771,
-0.04484611749649048,
-0.03498232737183571,
0.02244074083864689,
0.11829765886068344,
0.05901027470827103,
0.04830724745988846,
-0.0033231882844120264,
0.013305929489433765
] |
CSA: Data-efficient Mapping of Unimodal Features to Multimodal Features | https://openreview.net/forum?id=6Mg7pjG7Sw | [
"Po-han Li",
"Sandeep P. Chinchali",
"ufuk topcu"
] | Poster | Multimodal encoders like CLIP excel in tasks such as zero-shot image classification and cross-modal retrieval. However, they require excessive training data.
We propose canonical similarity analysis (CSA), which uses two unimodal encoders to replicate multimodal encoders using limited data.
CSA maps unimodal features into a multimodal space, using a new similarity score to retain only the multimodal information.
CSA only involves the inference of unimodal encoders and a cubic-complexity matrix decomposition, eliminating the need for extensive GPU-based model training.
Experiments show that CSA outperforms CLIP while requiring $50,000\times$ fewer multimodal data pairs to bridge the modalities given pre-trained unimodal encoders on ImageNet classification and misinformative news caption detection.
CSA surpasses the state-of-the-art method to map unimodal features to multimodal features.
We also demonstrate the ability of CSA with modalities beyond image and text, paving the way for future modality pairs with limited paired multimodal data but abundant unpaired unimodal data, such as lidar and text. | multimodal, representation learning, relative representations | Canonical similarity analysis (CSA) achieves CLIP-level performance in multimodal tasks using much less data. It maps unimodal features into a multimodal space without extensive GPU training and supports all modalities. | 2,221 | 2410.07610 | [
-0.015866484493017197,
-0.15385596454143524,
0.018861092627048492,
0.0034584361128509045,
0.13209065794944763,
0.000370448746252805,
-0.11422406882047653,
-0.042053189128637314,
0.013568228110671043,
-0.06210760027170181,
0.014113355427980423,
-0.09115997701883316,
0.011181521229445934,
0.08638569712638855,
-0.029048603028059006,
-0.026524638757109642,
0.043536581099033356,
0.1187129020690918,
-0.11411827802658081,
0.004228362813591957,
0.0681808814406395,
0.013058243319392204,
0.0650433823466301,
0.005437042564153671,
0.05355151742696762,
0.0116810891777277,
-0.011226676404476166,
-0.06322867423295975,
0.028932999819517136,
-0.046198733150959015,
0.009694800712168217,
0.012939688749611378,
-0.05441988632082939,
0.02820909582078457,
-0.03002760000526905,
0.08071113377809525,
-0.019193727523088455,
0.02958364225924015,
0.008284252136945724,
-0.07099662721157074,
0.025473577901721,
0.04057306796312332,
-0.024438906461000443,
0.0918402224779129,
0.045528240501880646,
0.021141499280929565,
-0.05673045292496681,
-0.05206888169050217,
0.09171050041913986,
0.03219177573919296,
-0.045838478952646255,
0.05370583012700081,
0.012654275633394718,
0.06420528143644333,
-0.03748197853565216,
-0.1056215837597847,
0.0037070277612656355,
-0.05665145441889763,
0.06769313663244247,
-0.02013590931892395,
-0.011843455024063587,
-0.0804184302687645,
-0.02984374389052391,
-0.03923026844859123,
0.053657617419958115,
-0.046248067170381546,
0.01600668951869011,
-0.044602472335100174,
-0.0075884247198700905,
-0.030790172517299652,
-0.04247813671827316,
0.019096096977591515,
-0.03482183814048767,
0.04125398397445679,
-0.05671536922454834,
0.04888279736042023,
0.07613259553909302,
-0.03330443426966667,
-0.015449784696102142,
0.011130696162581444,
0.07793305069208145,
-0.010860957205295563,
0.09215020388364792,
-0.00935236643999815,
0.08258579671382904,
0.0056402599439024925,
-0.009043592028319836,
0.012323912233114243,
-0.018410470336675644,
0.03071640431880951,
-0.0944264605641365,
-0.024198882281780243,
0.07718712836503983,
-0.07598432153463364,
0.05186410993337631,
-0.011969026178121567,
0.002861613407731056,
-0.04572562500834465,
-0.033564768731594086,
0.09728986769914627,
0.009579287841916084,
0.028930263593792915,
-0.04603426903486252,
0.0010544662363827229,
0.06912451237440109,
0.05050714686512947,
0.0476418137550354,
-0.008151528425514698,
0.020079661160707474,
-0.042049575597047806,
0.04749204218387604,
0.05591271072626114,
-0.015575418248772621,
0.01641109213232994,
0.02484891191124916,
-0.0070123737677931786,
0.006216252222657204,
0.019038019701838493,
0.03496498242020607,
0.012403996661305428,
-0.05067949742078781,
0.004108561668545008,
-0.04535834491252899,
-0.032683659344911575,
-0.0007057218463160098,
-0.07998631149530411,
-0.08331543952226639,
4.6266241746396105e-33,
0.04392112046480179,
0.03084862232208252,
0.06796208024024963,
0.028556063771247864,
0.04937833547592163,
-0.056263409554958344,
-0.06575704365968704,
0.08468448370695114,
-0.015604391694068909,
-0.008708255365490913,
-0.024044297635555267,
0.03748837485909462,
-0.04861703887581825,
0.14167575538158417,
-0.033298559486866,
0.021291837096214294,
-0.05968121066689491,
-0.041497085243463516,
0.06378660351037979,
0.018790561705827713,
-0.030460495501756668,
-0.044714611023664474,
0.02298664301633835,
0.06077685207128525,
-0.019226599484682083,
-0.03659164905548096,
0.021798832342028618,
-0.1379098892211914,
-0.008182934485375881,
-0.011771448887884617,
-0.10177434980869293,
0.02196657471358776,
0.017611466348171234,
0.018311312422156334,
0.059477899223566055,
-0.06552048027515411,
0.010801058262586594,
-0.048918675631284714,
0.018306046724319458,
0.00046253608888946474,
-0.04977076128125191,
-0.02907663583755493,
-0.08562649041414261,
-0.06633076071739197,
-0.04908846691250801,
-0.0639636293053627,
0.08692473918199539,
-0.02244780771434307,
-0.040947746485471725,
-0.013087031431496143,
0.11853852868080139,
0.02241724170744419,
-0.06690345704555511,
-0.08159143477678299,
-0.06661150604486465,
-0.07551361620426178,
0.05441511422395706,
-0.031028974801301956,
0.06156565248966217,
0.00497118104249239,
-0.02383192628622055,
0.01916506700217724,
0.04179609566926956,
0.04234229400753975,
-0.006350107491016388,
0.020159004256129265,
-0.04088195413351059,
0.058848343789577484,
0.08997129648923874,
0.022788256406784058,
-0.07498546689748764,
0.019166426733136177,
-0.015053680166602135,
-0.029897527769207954,
0.021210430189967155,
0.04308591037988663,
0.03437274694442749,
-0.11519086360931396,
0.07195758074522018,
0.12539245188236237,
0.04179691895842552,
0.023508433252573013,
0.03562692552804947,
-0.0947609543800354,
-0.0660969689488411,
-0.06963993608951569,
0.01586274988949299,
-0.09450888633728027,
-0.01667587272822857,
-0.008634763769805431,
-0.027076266705989838,
0.058672647923231125,
-0.10273262113332748,
0.029905470088124275,
0.03413832187652588,
-3.611202767040548e-33,
0.03745418041944504,
0.020848622545599937,
-0.007110636215656996,
0.023945091292262077,
-0.012345921248197556,
-0.01709642820060253,
0.02504037134349346,
0.027915092185139656,
0.021554213017225266,
-0.05205933004617691,
0.01608358509838581,
-0.03365818411111832,
0.020196789875626564,
-0.035193171352148056,
0.047554682940244675,
0.02447429671883583,
-0.011772282421588898,
0.057879865169525146,
-0.1099117323756218,
0.03761440888047218,
0.0490729995071888,
0.07466673105955124,
-0.02666473016142845,
0.03409647196531296,
-0.046479731798172,
0.0848289430141449,
-0.05479252710938454,
-0.02819097600877285,
0.011550779454410076,
-0.03688312694430351,
-0.014228791929781437,
0.01772134192287922,
0.001727032009512186,
-0.03728555515408516,
-0.010368169285356998,
0.09460968524217606,
0.02414846234023571,
-0.01815851405262947,
0.015095962211489677,
0.11388267576694489,
0.08033145219087601,
0.011168447323143482,
-0.0326823927462101,
0.023267308250069618,
0.021305710077285767,
-0.07198306173086166,
-0.05403146520256996,
-0.04644303396344185,
0.07769445329904556,
0.08256904780864716,
0.009497166611254215,
0.018236154690384865,
-0.11156484484672546,
0.02803557738661766,
0.016465013846755028,
-0.14966343343257904,
-0.045658767223358154,
-0.015749339014291763,
0.03134122118353844,
0.04513876140117645,
-0.01818232610821724,
-0.055103085935115814,
-0.039394594728946686,
-0.024064693599939346,
0.032564084976911545,
-0.060860004276037216,
0.012965775094926357,
-0.0030930128414183855,
0.015392408706247807,
0.053777530789375305,
-0.020769504830241203,
0.0021334837656468153,
0.035352304577827454,
-0.04873087257146835,
0.014353829436004162,
0.017611730843782425,
0.04076157510280609,
-0.01868327334523201,
-0.006754628848284483,
-0.07064048945903778,
-0.00048473975039087236,
-0.06540357321500778,
0.011113188229501247,
0.06495273858308792,
0.08191804587841034,
0.15663661062717438,
0.04091143608093262,
-0.024494457989931107,
0.045285824686288834,
0.027703436091542244,
-0.00808517076075077,
0.022402677685022354,
0.05954695865511894,
0.09330102801322937,
0.024201249703764915,
-3.824290573106737e-8,
-0.03160180151462555,
0.006001731846481562,
-0.06790629029273987,
-0.0418032631278038,
-0.013784595765173435,
-0.04783551022410393,
-0.012535231187939644,
0.021467700600624084,
-0.02263563871383667,
-0.014148345217108727,
0.0616767443716526,
-0.0027952115051448345,
-0.04688912630081177,
-0.013244641944766045,
0.010419698432087898,
-0.025926729664206505,
0.022755008190870285,
0.028806593269109726,
0.05038287863135338,
0.019354140385985374,
0.012611228972673416,
-0.015765415504574776,
0.011929813772439957,
-0.03420347347855568,
-0.025709226727485657,
-0.07234275341033936,
-0.026001399382948875,
0.03746768832206726,
0.03783111646771431,
0.019208237528800964,
-0.009063241071999073,
-0.026731647551059723,
-0.01675356924533844,
-0.11633627861738205,
0.02751394920051098,
0.023235352709889412,
0.054120033979415894,
-0.05652044340968132,
-0.03314202278852463,
-0.026890279725193977,
0.0633898377418518,
0.04666982963681221,
-0.05033120512962341,
0.0495416484773159,
0.010068351402878761,
0.07791297137737274,
0.04326470196247101,
-0.0482025220990181,
0.01855028048157692,
-0.045971546322107315,
0.017859285697340965,
-0.012194005772471428,
0.03450217470526695,
0.10970355570316315,
0.02233102172613144,
-0.020980725064873695,
-0.02066061459481716,
0.026350267231464386,
0.10724213719367981,
0.0013819669838994741,
0.020533934235572815,
0.016077687963843346,
-0.09680002182722092,
-0.06586115062236786
] |
EC-Diffuser: Multi-Object Manipulation via Entity-Centric Behavior Generation | https://openreview.net/forum?id=o3pJU5QCtv | [
"Carl Qi",
"Dan Haramati",
"Tal Daniel",
"Aviv Tamar",
"Amy Zhang"
] | Poster | Object manipulation is a common component of everyday tasks, but learning to manipulate objects from high-dimensional observations presents significant challenges. These challenges are heightened in multi-object environments due to the combinatorial complexity of the state space as well as of the desired behaviors. While recent approaches have utilized large-scale offline data to train models from pixel observations, achieving performance gains through scaling, these methods struggle with compositional generalization in unseen object configurations with constrained network and dataset sizes. To address these issues, we propose a novel behavioral cloning (BC) approach that leverages object-centric representations and an entity-centric Transformer with diffusion-based optimization, enabling efficient learning from offline image data. Our method first decomposes observations into Deep Latent Particles (DLP), which are then processed by our entity-centric Transformer that computes attention at the particle level, simultaneously predicting object dynamics and the agent's actions. Combined with the ability of diffusion models to capture multi-modal behavior distributions, this results in substantial performance improvements in multi-object tasks and, more importantly, enables compositional generalization. We present BC agents capable of zero-shot generalization to perform tasks with novel compositions of objects and goals, including larger numbers of objects than seen during training. We provide video rollouts on our webpage: https://sites.google.com/view/ec-diffuser. | Diffusion, Object-Centric Representation, Robotic Manipulation | We propose a behavioral cloning method for multi-object manipulation that combines object-centric representations with diffusion models, enabling zero-shot generalization to novel object compositions. | 2,218 | null | [
-0.0456056147813797,
-0.11195237189531326,
0.0996234267950058,
0.061387114226818085,
0.058160483837127686,
-0.0301419235765934,
-0.054867472499608994,
-0.08611219376325607,
0.051384683698415756,
-0.01206628605723381,
-0.021179327741265297,
-0.1314881145954132,
-0.011799590662121773,
0.04249613732099533,
0.03560929372906685,
0.07199016958475113,
0.11952745914459229,
0.09384618699550629,
-0.1613094061613083,
-0.0333917997777462,
0.005622276104986668,
-0.0319829061627388,
0.017912395298480988,
-0.027713801711797714,
0.013237555511295795,
0.0519927516579628,
0.04002741724252701,
-0.152163565158844,
0.016025304794311523,
-0.09034615010023117,
0.0801132470369339,
0.015591534785926342,
-0.05944713205099106,
0.0345146507024765,
-0.06592757254838943,
0.08878009021282196,
-0.06299535185098648,
-0.018461979925632477,
-0.009622040204703808,
-0.037836626172065735,
0.05389717221260071,
0.024310197681188583,
-0.007384764961898327,
-0.009595696814358234,
0.039582040160894394,
0.01724827289581299,
0.012248395010828972,
-0.03327218070626259,
0.08205112814903259,
-0.0304142776876688,
-0.018569709733128548,
-0.019879944622516632,
-0.045074544847011566,
0.1112041100859642,
-0.04664042964577675,
0.027346035465598106,
0.08018844574689865,
-0.043579403311014175,
0.06608615815639496,
-0.029863979667425156,
-0.04783732071518898,
-0.0015005231834948063,
0.09144856035709381,
-0.04329845681786537,
-0.03605244308710098,
-0.06328517943620682,
0.04725699871778488,
-0.05269516631960869,
-0.033647824078798294,
-0.02376166172325611,
-0.027989763766527176,
-0.003409099765121937,
-0.022581515833735466,
0.008546019904315472,
0.09255465865135193,
0.010951711796224117,
0.016143159940838814,
-0.02697216160595417,
0.021356504410505295,
-0.08489292860031128,
0.04029293730854988,
0.022149603813886642,
0.06330136209726334,
0.004136628936976194,
0.09373017400503159,
0.0041993893682956696,
-0.02042599394917488,
0.10756008327007294,
-0.01765552908182144,
0.004746479447931051,
-0.06657271832227707,
-0.010535640642046928,
0.04541892185807228,
0.0011516583617776632,
0.01370914001017809,
-0.023045778274536133,
-0.04368206858634949,
-0.0637962818145752,
0.016883457079529762,
0.05704924091696739,
-0.014690871350467205,
0.04880096763372421,
0.06411364674568176,
0.01874692738056183,
-0.004929231479763985,
-0.08277589827775955,
0.057520315051078796,
0.03437101095914841,
-0.0008812008309178054,
-0.032532949000597,
0.09356849640607834,
0.08844710141420364,
0.006596826948225498,
-0.0001788754016160965,
0.06809654831886292,
-0.056597549468278885,
-0.02422976866364479,
0.011471053585410118,
-0.03368346020579338,
0.012155676260590553,
0.07155077904462814,
-0.0443941168487072,
-0.06467289477586746,
0.040974680334329605,
0.012422124855220318,
-0.01621345616877079,
-0.08228889852762222,
-5.386841993419812e-34,
0.08217871934175491,
-0.011780239641666412,
0.08182641118764877,
0.021955473348498344,
0.01631810888648033,
0.002347423229366541,
0.03097848780453205,
0.0613233707845211,
-0.03489818796515465,
-0.030560245737433434,
-0.03348846733570099,
0.010381758213043213,
-0.05562084913253784,
0.08752401173114777,
0.04235868528485298,
-0.030381817370653152,
-0.031035449355840683,
0.04903055727481842,
0.023027246817946434,
-0.00853955652564764,
-0.020065318793058395,
-0.016757020726799965,
-0.0825672596693039,
-0.04364480450749397,
0.028512757271528244,
-0.014844425953924656,
0.03554330766201019,
-0.09967096894979477,
0.0001107209682231769,
-0.025774050503969193,
-0.09836044162511826,
0.04178554192185402,
0.035418082028627396,
0.08153525739908218,
-0.004515429958701134,
-0.03844313323497772,
-0.04454919323325157,
-0.038505904376506805,
-0.007616592105478048,
-0.028064962476491928,
0.025326771661639214,
0.018947143107652664,
-0.04230697825551033,
-0.021771231666207314,
-0.11813338100910187,
-0.043653301894664764,
0.032221730798482895,
0.035818617790937424,
-0.04262696951627731,
-0.047679781913757324,
0.06768378615379333,
-0.00943891704082489,
0.013192365877330303,
-0.10042288154363632,
0.03333244472742081,
-0.0011892272159457207,
0.052135270088911057,
-0.035566918551921844,
0.04943911358714104,
-0.03816171735525131,
0.019154688343405724,
0.11402759701013565,
-0.05072248354554176,
0.08460880815982819,
0.02576165646314621,
-0.0011152413208037615,
-0.024022910743951797,
0.07104388624429703,
0.019862575456500053,
0.03710368648171425,
-0.03606042638421059,
0.11681490391492844,
-0.01829933002591133,
-0.07577122002840042,
0.10947557538747787,
-0.03596417233347893,
0.0358046293258667,
-0.07891089469194412,
-0.008040797896683216,
0.08718902617692947,
-0.049943484365940094,
-0.008889245800673962,
-0.023519271984696388,
-0.029341921210289,
0.0009497271967120469,
-0.019831929355859756,
0.03760940209031105,
-0.028896059840917587,
-0.05139588192105293,
-0.027170460671186447,
-0.005144590046256781,
-0.077088363468647,
-0.09604635834693909,
0.02605055831372738,
-0.06808139383792877,
6.576413348270677e-34,
-0.014209439978003502,
0.012907666154205799,
-0.039133623242378235,
0.07984904199838638,
0.015794208273291588,
-0.07808711379766464,
-0.018403185531497,
-0.013769560493528843,
-0.02085850201547146,
-0.049515195190906525,
-0.049139078706502914,
-0.020512780174613,
0.002454394707456231,
-0.07616665959358215,
0.03292076662182808,
-0.014191795140504837,
-0.03845135122537613,
0.015471423976123333,
-0.09407131373882294,
-0.0008676638826727867,
-0.022080078721046448,
0.07076457142829895,
-0.017553402110934258,
0.00010454458970343694,
-0.09546679258346558,
0.03250208497047424,
-0.03118184395134449,
0.035836588591337204,
0.06761377304792404,
-0.007544221822172403,
-0.04139964282512665,
0.006947108544409275,
0.03506904095411301,
-0.04182465746998787,
-0.09788631647825241,
0.05335812270641327,
0.004627893678843975,
0.023391831666231155,
-0.0032017945777624846,
-0.012937466613948345,
-0.007845439016819,
-0.07891502231359482,
-0.08421241492033005,
0.08286026120185852,
-0.011709651909768581,
0.00020350093836896122,
-0.022715618833899498,
0.01668849028646946,
0.03702991083264351,
0.10586781799793243,
-0.05413222312927246,
0.06420588493347168,
-0.09230341017246246,
0.0045687188394367695,
-0.029042575508356094,
-0.026836618781089783,
-0.023557549342513084,
-0.08048601448535919,
0.12077006697654724,
-0.07196386158466339,
-0.0646328404545784,
-0.03612067922949791,
0.05577384680509567,
-0.06562342494726181,
-0.020208002999424934,
0.00744564039632678,
-0.01582251489162445,
0.045419562608003616,
0.023688022047281265,
0.023214679211378098,
0.07002833485603333,
0.0019638310186564922,
0.05438503623008728,
-0.006243001203984022,
0.042858026921749115,
-0.04346519336104393,
0.042689383029937744,
-0.07816971838474274,
0.03694039583206177,
-0.015828067436814308,
0.016219768673181534,
-0.016114605590701103,
0.03504779934883118,
0.03255421295762062,
0.07825981825590134,
0.0442022979259491,
0.057764165103435516,
0.0253069456666708,
0.022837964817881584,
0.04390346631407738,
-0.03959794342517853,
0.0372118204832077,
0.035415008664131165,
0.06889854371547699,
-0.02124456875026226,
-4.036039769061972e-8,
-0.09816697984933853,
0.042257871478796005,
0.0466499999165535,
-0.005842225626111031,
0.030867600813508034,
-0.004670059774070978,
0.03635839745402336,
0.06160920485854149,
0.01726938784122467,
0.00035363598726689816,
0.05388004332780838,
-0.013080236501991749,
-0.014315910637378693,
-0.06469963490962982,
0.06775239109992981,
0.055290769785642624,
0.017960626631975174,
-0.050737638026475906,
0.028805751353502274,
0.02317115105688572,
-0.03437943756580353,
-0.005533342715352774,
0.03158201277256012,
-0.033704061061143875,
0.07229229062795639,
-0.03541318699717522,
-0.028132518753409386,
0.030418774113059044,
0.0747079998254776,
-0.009272255003452301,
0.009140326641499996,
0.02343866601586342,
0.010691607370972633,
0.04135648533701897,
-0.013339933007955551,
0.024230722337961197,
0.060676660388708115,
-0.004755402449518442,
-0.034940339624881744,
-0.06878901273012161,
0.05902819335460663,
0.03504300117492676,
-0.05236624553799629,
0.035561803728342056,
0.024078961461782455,
-0.019378511235117912,
-0.00798899307847023,
-0.07597121596336365,
0.04594546556472778,
0.003401384688913822,
0.07466667890548706,
-0.059834208339452744,
-0.014653648249804974,
0.11542539298534393,
0.04217371344566345,
0.031459733843803406,
0.056124139577150345,
-0.02414627932012081,
0.11137621104717255,
0.08872304111719131,
-0.022006327286362648,
0.020671285688877106,
-0.05803724005818367,
0.012442518956959248
] |
A Truncated Newton Method for Optimal Transport | https://openreview.net/forum?id=gWrWUaCbMa | [
"Mete Kemertas",
"Amir-massoud Farahmand",
"Allan Douglas Jepson"
] | Poster | Developing a contemporary optimal transport (OT) solver requires navigating trade-offs among several critical requirements: GPU parallelization, scalability to high-dimensional problems, theoretical convergence guarantees, empirical performance in terms of precision versus runtime, and numerical stability in practice. With these challenges in mind, we introduce a specialized truncated Newton algorithm for entropic-regularized OT. In addition to proving that locally quadratic convergence is possible without assuming a Lipschitz Hessian, we provide strategies to maximally exploit the high rate of local convergence in practice. Our GPU-parallel algorithm exhibits exceptionally favorable runtime performance, achieving high precision orders of magnitude faster than many existing alternatives. This is evidenced by wall-clock time experiments on 24 problem sets (12 datasets $\times$ 2 cost functions). The scalability of the algorithm is showcased on an extremely large OT problem with $n \approx 10^6$, solved approximately under weak entopric regularization. | Computational optimal transport, numerical optimization, numerical linear algebra | A high-performing, high-precision, GPU-parallel optimal transport solver based on truncated Newton methods. | 2,216 | null | [
-0.03715066984295845,
-0.06182291358709335,
0.02487141080200672,
0.013178492896258831,
0.009522657841444016,
-0.04971148446202278,
-0.045935750007629395,
0.01904156059026718,
-0.056218575686216354,
-0.004882060922682285,
-0.05189213529229164,
0.02820942923426628,
-0.02050408348441124,
0.043856602162122726,
-0.0771758034825325,
-0.01859465427696705,
0.05900898203253746,
0.044664833694696426,
-0.07723282277584076,
-0.020506368950009346,
-0.08416274189949036,
0.04505263268947601,
-0.013670201413333416,
0.05736342817544937,
0.003534903284162283,
-0.024296650663018227,
0.010464755818247795,
-0.05590047314763069,
-0.0019379578297957778,
0.042864855378866196,
0.025513652712106705,
-0.018445055931806564,
-0.04668416827917099,
-0.055424515157938004,
-0.09416748583316803,
0.10188611596822739,
0.011419306509196758,
-0.001139019732363522,
-0.004401837941259146,
0.009928176179528236,
0.029109906405210495,
-0.011570964008569717,
0.05562576651573181,
0.06161752715706825,
0.060195326805114746,
-0.07450737804174423,
0.07856695353984833,
-0.0410538986325264,
0.09790491312742233,
-0.05487491935491562,
0.03988621011376381,
-0.0032613056246191263,
-0.036730147898197174,
0.03049517050385475,
-0.01246099267154932,
-0.03994536027312279,
0.07676205039024353,
-0.024824338033795357,
-0.030015407130122185,
0.10562621802091599,
0.005733749829232693,
-0.12071472406387329,
-0.026944687590003014,
-0.010690123774111271,
-0.011534433811903,
-0.02491801790893078,
0.05719166621565819,
-0.019574541598558426,
0.004944894928485155,
0.11809086799621582,
0.0024008192121982574,
0.06949403882026672,
-0.10225582867860794,
0.036076080054044724,
0.028087925165891647,
-0.0005663267802447081,
0.06810858100652695,
0.0315718948841095,
0.011275934055447578,
-0.03576577827334404,
0.0354146845638752,
-0.05723346769809723,
-0.008292053826153278,
-0.016791563481092453,
0.016552120447158813,
-0.04650561511516571,
-0.034267980605363846,
0.06862203031778336,
0.03914106637239456,
-0.06602269411087036,
0.0409359484910965,
0.023946430534124374,
-0.029097972437739372,
-0.014531133696436882,
-0.02957203984260559,
-0.02471761405467987,
-0.009226749651134014,
-0.00732820387929678,
-0.007258635479956865,
0.10779741406440735,
-0.050574444234371185,
0.029735567048192024,
0.04518996179103851,
0.0443439781665802,
0.025579025968909264,
-0.018138572573661804,
0.025122202932834625,
0.07834062725305557,
-0.05789467319846153,
-0.062297604978084564,
0.0644427239894867,
0.038783520460128784,
0.07439347356557846,
0.08767545223236084,
0.009939388372004032,
-0.012232540175318718,
-0.022687789052724838,
-0.056180190294981,
0.04647213965654373,
0.11443530768156052,
-0.1281481683254242,
-0.025384996086359024,
0.08403794467449188,
0.03619128093123436,
0.024955863133072853,
0.043448761105537415,
-0.028996223583817482,
3.6679949398906036e-33,
-0.04412904381752014,
0.0754503458738327,
0.08936680853366852,
-0.0694524422287941,
-0.010084467008709908,
0.012948940508067608,
0.030705343931913376,
0.02246236801147461,
0.008322433568537235,
0.0004355314595159143,
-0.04572143405675888,
0.004515675362199545,
-0.06777771562337875,
0.10235486924648285,
0.00021038891281932592,
-0.07860274612903595,
-0.0045165917836129665,
0.030954740941524506,
0.014982260763645172,
-0.037248965352773666,
0.11996930837631226,
-0.08348274230957031,
-0.04158637672662735,
-0.07868555933237076,
-0.03138979151844978,
-0.032739199697971344,
-0.018465273082256317,
-0.03974921256303787,
0.11938861012458801,
0.03175375238060951,
-0.06338563561439514,
0.006217775866389275,
-0.044990215450525284,
0.08043507486581802,
0.05439142882823944,
0.021041112020611763,
0.05403511971235275,
-0.05820723995566368,
0.018215864896774292,
-0.0017581027932465076,
-0.06350906938314438,
0.04747726768255234,
-0.002998636569827795,
-0.03354472666978836,
-0.0863223597407341,
-0.04441976174712181,
0.016239840537309647,
-0.0067054154351353645,
0.037183962762355804,
-0.05796084553003311,
-0.021802684292197227,
-0.001613202621228993,
-0.08438224345445633,
-0.07705649733543396,
0.03323341906070709,
-0.1252674013376236,
0.04237223044037819,
0.002363728592172265,
0.04055725410580635,
0.11422362178564072,
0.07225457578897476,
0.029895678162574768,
0.016928521916270256,
0.00406308751553297,
-0.05047428980469704,
-0.012921072542667389,
-0.012644495815038681,
0.008675587363541126,
-0.02328668162226677,
-0.026094073429703712,
-0.017203429713845253,
0.009829285554587841,
0.06066542863845825,
-0.05271421745419502,
0.11156372725963593,
0.01839742250740528,
0.057536471635103226,
-0.10742029547691345,
-0.008652983233332634,
-0.0994410365819931,
-0.09326475858688354,
0.06434928625822067,
0.013879125006496906,
-0.030789274722337723,
-0.001954267267137766,
-0.08627334237098694,
-0.055909160524606705,
-0.011837000027298927,
-0.05598370358347893,
-0.11441850662231445,
-0.07929348945617676,
0.03059229999780655,
0.010124050080776215,
0.017704589292407036,
-0.03204237297177315,
-3.772503367176413e-33,
0.007018021307885647,
-0.029054945334792137,
0.026002077385783195,
0.11243095248937607,
0.002358099678531289,
-0.031222617253661156,
-0.030751921236515045,
-0.0591021291911602,
0.057553019374608994,
-0.04478957876563072,
0.009101435542106628,
0.0013171735918149352,
0.04779250919818878,
0.031113525852560997,
0.08719512820243835,
-0.1292014867067337,
-0.000784844218287617,
0.029880022630095482,
-0.05046944320201874,
-0.04942427948117256,
-0.03506355732679367,
-0.00330430013127625,
-0.023677218705415726,
-0.03719281405210495,
0.0130847729742527,
-0.027044305577874184,
-0.0456252321600914,
0.004574957303702831,
-0.029245946556329727,
0.04204671457409859,
-0.06314548850059509,
-0.025656603276729584,
-0.028742153197526932,
0.018448004499077797,
-0.01348808966577053,
0.13536646962165833,
0.01695951074361801,
0.11947936564683914,
-0.0010227493476122618,
0.03377500921487808,
0.03723660856485367,
0.0009254629258066416,
-0.028031237423419952,
0.04249570891261101,
0.026560978963971138,
0.01865317113697529,
-0.08766477555036545,
-0.043728917837142944,
0.048397764563560486,
0.08415737003087997,
0.01805201545357704,
-0.03473947197198868,
0.055270228534936905,
0.0442749559879303,
0.00706738093867898,
0.006859952118247747,
0.01761014759540558,
0.026205511763691902,
0.05590946599841118,
-0.0633213073015213,
-0.03891192004084587,
-0.039624664932489395,
-0.010687138885259628,
0.007830822840332985,
-0.01642964780330658,
-0.000804574228823185,
0.015099490992724895,
-0.0000036666776850324823,
-0.0036005456931889057,
0.04966080188751221,
-0.07018503546714783,
-0.01686420664191246,
0.02336002327501774,
0.07491932809352875,
-0.030969765037298203,
0.03997295722365379,
0.0914473682641983,
-0.06023804470896721,
-0.041019558906555176,
0.032139264047145844,
0.05517780780792236,
0.08116535097360611,
0.051008790731430054,
0.022284196689724922,
0.0306540597230196,
-0.013994565233588219,
0.0018136855214834213,
0.01978258602321148,
0.004977808799594641,
-0.06612222641706467,
0.010165153071284294,
-0.006786829326301813,
0.0748794823884964,
-0.048041630536317825,
-0.0825669914484024,
-4.459440461346276e-8,
-0.02927241288125515,
-0.03623224422335625,
0.010783828794956207,
-0.008274514228105545,
-0.014676760882139206,
-0.03832448646426201,
-0.010942723602056503,
-0.007828867062926292,
-0.017158042639493942,
-0.025374382734298706,
0.05294737219810486,
0.02299196831882,
-0.02461184561252594,
-0.03766947239637375,
-0.05236200988292694,
0.03894643113017082,
-0.01063345093280077,
-0.029842782765626907,
-0.013473954051733017,
-0.05275276303291321,
-0.09030898660421371,
0.05520356446504593,
0.004787680692970753,
-0.03253445774316788,
0.014704311266541481,
-0.039806485176086426,
0.010723985731601715,
0.08270157873630524,
0.1323256641626358,
-0.007517785765230656,
-0.027415402233600616,
-0.02971271239221096,
-0.023724494501948357,
0.03964132070541382,
-0.0063036722131073475,
0.03290301561355591,
0.011429696343839169,
0.13610917329788208,
-0.010834963992238045,
0.0026529242750257254,
0.016566570848226547,
-0.018060149624943733,
-0.003378802677616477,
-0.05528769642114639,
0.11510491371154785,
0.08615504205226898,
-0.06407487392425537,
0.06123034656047821,
-0.03963980823755264,
0.0642990991473198,
0.00006031330849509686,
0.05001498758792877,
0.021081464365124702,
0.09945167601108551,
0.06042711064219475,
-0.010224630124866962,
-0.10315841436386108,
-0.09770943969488144,
0.00453642662614584,
0.04106111079454422,
-0.07990746945142746,
0.016932092607021332,
-0.0396992452442646,
0.003389810211956501
] |
Automated Proof Generation for Rust Code via Self-Evolution | https://openreview.net/forum?id=2NqssmiXLu | [
"Tianyu Chen",
"Shuai Lu",
"Shan Lu",
"Yeyun Gong",
"Chenyuan Yang",
"Xuheng Li",
"Md Rakib Hossain Misu",
"Hao Yu",
"Nan Duan",
"Peng CHENG",
"Fan Yang",
"Shuvendu K Lahiri",
"Tao Xie",
"Lidong Zhou"
] | Poster | Ensuring correctness is crucial for code generation. Formal verification offers a
definitive assurance of correctness, but demands substantial human effort in proof
construction and hence raises a pressing need for automation. The primary obsta-
cle lies in the severe lack of data—there is much fewer proofs than code snippets
for Large Language Models (LLMs) to train upon. In this paper, we introduce
SAFE, a framework that overcomes the lack of human-written proofs to enable
automated proof generation of Rust code. SAFE establishes a self-evolving cycle
where data synthesis and fine-tuning collaborate to enhance the model capability,
leveraging the definitive power of a symbolic verifier in telling correct proofs from
incorrect ones. SAFE also re-purposes the large number of synthesized incorrect
proofs to train the self-debugging capability of the fine-tuned models, empowering
them to fix incorrect proofs based on the verifier’s feedback. SAFE demonstrates
superior efficiency and precision compared to GPT-4o. Through tens of thousands
of synthesized proofs and the self-debugging mechanism, we improve the capa-
bility of open-source models, initially unacquainted with formal verification, to
automatically write proofs for Rust code. This advancement leads to a signifi-
cant improvement in performance, achieving a 52.52% accuracy rate in a bench-
mark crafted by human experts, a significant leap over GPT-4o’s performance of
14.39%. | Large Language Models, Program Verification | null | 2,209 | 2410.15756 | [
-0.11535973846912384,
0.013476867228746414,
-0.010408436879515648,
0.04458533227443695,
0.08037545531988144,
-0.06270483881235123,
-0.0467752180993557,
0.06488200277090073,
-0.026038309559226036,
0.00696555944159627,
-0.04604332149028778,
-0.08842730522155762,
0.07230959832668304,
0.001285715727135539,
-0.04376239329576492,
-0.009681235067546368,
-0.03760294243693352,
0.04483051225543022,
-0.0039456491358578205,
-0.07317980378866196,
-0.04246188700199127,
0.019621772691607475,
0.02786722220480442,
0.05284113436937332,
0.027648167684674263,
0.019658803939819336,
0.017498062923550606,
-0.022525843232870102,
0.0029583873692899942,
-0.02337496168911457,
-0.02215026132762432,
0.020913485437631607,
0.013159041292965412,
0.0012520665768533945,
0.01721741072833538,
0.08331884443759918,
0.022608807310461998,
-0.015011582523584366,
-0.018971296027302742,
-0.022851359099149704,
-0.10570264607667923,
-0.014769382774829865,
-0.007122917100787163,
0.007104748860001564,
0.006006406154483557,
0.01613050512969494,
0.03607776761054993,
-0.004387323744595051,
-0.060347169637680054,
-0.031185820698738098,
-0.05783472955226898,
-0.03235986456274986,
-0.014482097700238228,
-0.06081718951463699,
0.04101663827896118,
-0.05281244218349457,
0.09354899078607559,
-0.009442409500479698,
-0.012622108682990074,
-0.016329120844602585,
0.026523850858211517,
0.020105930045247078,
-0.08673326671123505,
-0.008614500053226948,
0.04264162480831146,
0.085001640021801,
0.033487338572740555,
0.07375672459602356,
-0.005309887696057558,
0.01980464905500412,
-0.05217947065830231,
0.047273188829422,
-0.06888989359140396,
0.061436936259269714,
-0.026180488988757133,
0.13616986572742462,
0.010527253150939941,
0.017175232991576195,
-0.00006850917270639911,
-0.04485606029629707,
-0.05293206870555878,
0.05254872143268585,
-0.012063107453286648,
0.00567215122282505,
0.03906363621354103,
0.053880251944065094,
0.02975744940340519,
0.02517222985625267,
0.10096140205860138,
0.07967409491539001,
-0.01703951694071293,
-0.011774348095059395,
0.031664568930864334,
0.01686074584722519,
0.054300423711538315,
0.043951235711574554,
0.025483550503849983,
-0.02698242850601673,
0.007302234880626202,
0.07825517654418945,
-0.026573801413178444,
0.06683681160211563,
0.04403821378946304,
-0.09783296287059784,
0.06597006320953369,
0.012892212718725204,
0.07060187309980392,
0.0289449505507946,
-0.015056386590003967,
-0.04045749828219414,
0.08535834401845932,
0.010993406176567078,
0.023781079798936844,
0.004526367411017418,
-0.0014331225538626313,
0.020862165838479996,
-0.07414067536592484,
-0.02012019418179989,
0.016601981595158577,
0.03213408589363098,
0.01670115999877453,
0.03632805123925209,
0.008714932017028332,
0.0729636400938034,
0.030256472527980804,
-0.011553417891263962,
-0.04843101277947426,
1.8707397499321602e-33,
0.05674494802951813,
0.09046710282564163,
-0.007119236048310995,
0.09672528505325317,
0.022962259128689766,
0.011642823927104473,
0.003644936718046665,
-0.022364910691976547,
-0.07772033661603928,
0.02897830121219158,
0.05590759962797165,
0.05014117434620857,
-0.03925985470414162,
-0.025934455916285515,
0.05678880587220192,
0.009668828919529915,
-0.019638998433947563,
-0.008164583705365658,
-0.0931834727525711,
-0.005146010313183069,
0.07551959902048111,
-0.08659767359495163,
-0.034608788788318634,
-0.11223449558019638,
0.00383873307146132,
0.07557413727045059,
0.04215244576334953,
-0.04435458406805992,
-0.01623803935945034,
0.01214783824980259,
-0.07807106524705887,
-0.035261522978544235,
0.06762438267469406,
0.073041632771492,
0.04077896103262901,
-0.07397772371768951,
-0.02559535764157772,
-0.08264925330877304,
0.000577589962631464,
0.01787419430911541,
0.02797829918563366,
0.01982225477695465,
-0.008825148455798626,
0.026209725067019463,
0.09080315381288528,
-0.03694548457860947,
-0.05979917570948601,
0.019336087629199028,
-0.02278544008731842,
-0.08183598518371582,
0.0460357703268528,
0.10610794275999069,
-0.031993042677640915,
0.01110557746142149,
-0.04948388412594795,
-0.011056454852223396,
0.01663256622850895,
-0.002113820519298315,
0.0432870015501976,
0.1447944939136505,
0.009542745538055897,
0.14554999768733978,
-0.04852823168039322,
-0.005473252851516008,
-0.06231023371219635,
0.004786500707268715,
-0.07261081784963608,
0.048763152211904526,
0.06821794807910919,
0.07275024801492691,
-0.04482605308294296,
-0.03731657192111015,
-0.061041831970214844,
0.03516385331749916,
0.0026832965668290854,
-0.07598284631967545,
0.042928297072649,
-0.0856972187757492,
0.010312768630683422,
-0.07387803494930267,
0.04824874550104141,
0.07254461944103241,
0.022643543779850006,
0.03283357620239258,
-0.010368644259870052,
-0.10511508584022522,
-0.054189808666706085,
0.03556850552558899,
-0.07065463066101074,
-0.020624080672860146,
-0.0067728860303759575,
-0.09347441047430038,
0.06998732686042786,
-0.056604791432619095,
-0.02347051165997982,
-2.2721694176990405e-33,
-0.005131907295435667,
-0.05555842071771622,
-0.01904718205332756,
0.08237310498952866,
-0.13017626106739044,
-0.13453546166419983,
-0.05250043794512749,
-0.017530687153339386,
-0.013882370665669441,
-0.0574047714471817,
-0.07782160490751266,
-0.05895455181598663,
0.05771872028708458,
0.019433515146374702,
0.06553415954113007,
-0.07437068969011307,
-0.026438849046826363,
-0.0668548047542572,
0.008513975888490677,
0.05722234025597572,
0.033359725028276443,
0.07902960479259491,
-0.03253656625747681,
0.03629954904317856,
0.02928130514919758,
0.0019771293736994267,
-0.060827240347862244,
0.030861912295222282,
0.031128089874982834,
0.07890012115240097,
-0.009145921096205711,
0.04452670365571976,
-0.006161442957818508,
-0.08341032266616821,
-0.01586166024208069,
-0.027130326256155968,
0.09684879332780838,
0.0024282883387058973,
0.007072817999869585,
0.03419553115963936,
0.0679449662566185,
-0.04797305539250374,
-0.07419806718826294,
-0.01714547537267208,
-0.018584754317998886,
0.05788712948560715,
-0.08057838678359985,
0.03032253123819828,
0.027280939742922783,
-0.048828739672899246,
0.051268819719552994,
-0.04725196212530136,
0.0059805503115057945,
-0.04421732947230339,
-0.02248300239443779,
-0.10136464983224869,
-0.002916374709457159,
-0.031548354774713516,
-0.0275474414229393,
0.04199672117829323,
-0.038857344537973404,
-0.06735258549451828,
0.085483618080616,
-0.00510913971811533,
0.000505134230479598,
-0.05800529196858406,
-0.07372324168682098,
0.010027599520981312,
0.001889518927782774,
0.015715420246124268,
0.03361286595463753,
-0.059238556772470474,
-0.10095439851284027,
0.04041719064116478,
0.028606228530406952,
0.010093992576003075,
0.02918466180562973,
-0.03300946205854416,
0.009440827183425426,
0.01697133295238018,
0.015655923634767532,
-0.0033589715603739023,
0.062099389731884,
0.08502350747585297,
0.011991734616458416,
-0.023435641080141068,
0.03638309985399246,
0.04595359414815903,
-0.057362399995326996,
0.08409364521503448,
-0.09741808474063873,
0.0029815349262207747,
-0.04600755497813225,
0.12136857956647873,
-0.046434372663497925,
-4.1898701397258264e-8,
-0.01794874481856823,
-0.009877304546535015,
0.04129904508590698,
0.06601979583501816,
-0.00273183174431324,
0.028818458318710327,
-0.058256734162569046,
-0.05728796496987343,
0.00798588152974844,
-0.07009220123291016,
0.008743348531425,
0.021304883062839508,
-0.06623822450637817,
-0.05232812464237213,
-0.042211223393678665,
0.08205107599496841,
-0.029445849359035492,
-0.007432827726006508,
-0.058906763792037964,
-0.013993301428854465,
-0.012704388238489628,
0.056999336928129196,
-0.024213740602135658,
-0.0031040143221616745,
-0.0005826902925036848,
-0.06455894559621811,
0.0038869113195687532,
0.06112339347600937,
0.05738260596990585,
-0.014533604495227337,
0.042569030076265335,
-0.07369828224182129,
0.07025357335805893,
0.01486053504049778,
0.07500134408473969,
0.08377077430486679,
-0.010196846909821033,
0.07149452716112137,
0.035246592015028,
-0.06705789268016815,
-0.02161397784948349,
0.08113060891628265,
-0.05915782228112221,
-0.0025960486382246017,
-0.037962883710861206,
-0.03404794633388519,
0.0005300159100443125,
-0.07639355957508087,
-0.055584684014320374,
-0.0444331169128418,
0.0005458529340103269,
-0.050211310386657715,
-0.005677745211869478,
0.05897483229637146,
-0.03236794099211693,
0.07998953014612198,
-0.07006513327360153,
-0.026839876547455788,
0.02390376478433609,
0.03709446266293526,
-0.02790144272148609,
0.06229439377784729,
0.10456309467554092,
-0.005867721978574991
] |
Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems | https://openreview.net/forum?id=LkzuPorQ5L | [
"Guibin Zhang",
"Yanwei Yue",
"Zhixun Li",
"Sukwon Yun",
"Guancheng Wan",
"Kun Wang",
"Dawei Cheng",
"Jeffrey Xu Yu",
"Tianlong Chen"
] | Poster | Recent advancements in large language model (LLM)-powered agents have shown that collective intelligence can significantly outperform individual capabilities, largely attributed to the meticulously designed inter-agent communication topologies. Though impressive in performance, existing multi-agent pipelines inherently introduce substantial token overhead, as well as increased economic costs, which pose challenges for their large-scale deployments. In response to this challenge, we propose an economical, simple, and robust multi-agent communication framework, termed $\texttt{AgentPrune}$, which can seamlessly integrate into mainstream multi-agent systems and prunes redundant or even malicious communication messages. Technically, $\texttt{AgentPrune}$ is the first to identify and formally define the $\textit{Communication Redundancy}$ issue present in current LLM-based multi-agent pipelines, and efficiently performs one-shot pruning on the spatial-temporal message-passing graph, yielding a token-economic and high-performing communication topology.
Extensive experiments across six benchmarks demonstrate that $\texttt{AgentPrune}$ $\textbf{(I)}$ achieves comparable results as state-of-the-art topologies at merely $\\$5.6$ cost compared to their $\\$43.7$, $\textbf{(II)}$ integrates seamlessly into existing multi-agent frameworks with $28.1\\%\sim72.8\\%\downarrow$ token reduction, and $\textbf{(III)}$ successfully defend against two types of agent-based adversarial attacks with $3.5\\%\sim10.8\\%\uparrow$ performance boost. The source code is available at \url{https://github.com/yanweiyue/AgentPrune}. | Multi-agent collaboration, sparsification, LLM agents | null | 2,208 | 2410.02506 | [
-0.002080604201182723,
-0.01317213661968708,
-0.013236279599368572,
-0.01973281055688858,
0.033273790031671524,
-0.049876052886247635,
-0.05280115455389023,
-0.019597342237830162,
0.040752679109573364,
-0.02343646064400673,
-0.07721208035945892,
-0.037032995373010635,
0.08298633992671967,
0.029554301872849464,
0.06518150120973587,
0.01066256407648325,
0.09511739015579224,
0.08664576709270477,
-0.059101711958646774,
-0.06917046010494232,
-0.01352334301918745,
0.002693475689738989,
-0.02212485857307911,
-0.010982158593833447,
0.05509013682603836,
-0.019138794392347336,
-0.0616307370364666,
-0.07715733349323273,
-0.0330081507563591,
-0.005882405210286379,
0.06321953237056732,
-0.005658143665641546,
-0.004431723151355982,
0.0291899461299181,
-0.037569526582956314,
0.06400317698717117,
0.018940836191177368,
0.014316796325147152,
0.04948581010103226,
0.025924447923898697,
0.016384681686758995,
-0.02294160984456539,
-0.06451858580112457,
-0.0023264933843165636,
-0.05672712251543999,
-0.012349630706012249,
-0.0053327567875385284,
0.035519737750291824,
-0.07608349621295929,
-0.01924002170562744,
-0.06851286441087723,
-0.05888596922159195,
0.012321136891841888,
0.10913706570863724,
-0.027079956606030464,
-0.058879099786281586,
-0.07660332322120667,
0.029532482847571373,
-0.02066676877439022,
-0.028025247156620026,
0.007977533154189587,
-0.07266924530267715,
0.014433766715228558,
0.027465641498565674,
-0.02138584293425083,
-0.012536267749965191,
0.016528643667697906,
0.06117811053991318,
-0.03649681806564331,
0.05393579602241516,
0.058239877223968506,
-0.039056915789842606,
-0.030921926721930504,
-0.005111027043312788,
0.0036628697998821735,
0.0486023873090744,
0.043427418917417526,
0.025305280461907387,
0.03966458886861801,
-0.01529782172292471,
-0.011586125008761883,
0.012479481287300587,
-0.029803385958075523,
0.04334699735045433,
0.06581561267375946,
-0.019703639671206474,
-0.024052880704402924,
0.021091606467962265,
0.05190929397940636,
0.005267647560685873,
-0.11527837067842484,
-0.013392020016908646,
0.11252263188362122,
0.09069601446390152,
0.06924877315759659,
0.027173327282071114,
-0.06831023097038269,
-0.06138751283288002,
-0.028687238693237305,
0.10290802270174026,
-0.0627565085887909,
0.012181160040199757,
0.045010536909103394,
-0.07612627744674683,
-0.014811338856816292,
-0.04281380772590637,
0.003468368900939822,
0.023349735885858536,
-0.005506221204996109,
-0.03698679432272911,
-0.014384047128260136,
0.027936341241002083,
0.034247733652591705,
0.024516591802239418,
0.08310412615537643,
0.027030354365706444,
-0.014302346855401993,
0.045412104576826096,
0.061578284949064255,
0.013154716230928898,
0.010456595569849014,
-0.0756046399474144,
0.03684761002659798,
0.043807268142700195,
0.09294715523719788,
0.06143125146627426,
0.03566228970885277,
4.130096187371742e-34,
-0.0012731642927974463,
0.12605565786361694,
0.004262781701982021,
0.017580626532435417,
0.059681136161088943,
0.013895438984036446,
0.04984880983829498,
0.031940486282110214,
-0.06612439453601837,
-0.008294790051877499,
-0.028788693249225616,
0.028614025563001633,
0.007892552763223648,
0.06956037878990173,
-0.03634459897875786,
-0.042731478810310364,
0.03162787854671478,
0.04270288720726967,
0.033045556396245956,
-0.05269122123718262,
0.07911407202482224,
0.0571250282227993,
-0.013157766312360764,
-0.04353722557425499,
0.07366109639406204,
-0.0039746444672346115,
0.037592656910419464,
-0.1017570048570633,
0.07113951444625854,
0.029706334695219994,
-0.06524480134248734,
0.061193015426397324,
0.026577144861221313,
0.12875427305698395,
-0.03753204271197319,
-0.04686792939901352,
-0.08254751563072205,
-0.10598956793546677,
0.0009742737747728825,
0.044278427958488464,
-0.057961709797382355,
0.04905398562550545,
-0.00546731473878026,
-0.12524794042110443,
-0.04589112475514412,
-0.0766228660941124,
-0.060751207172870636,
-0.003921644762158394,
-0.006746804341673851,
-0.08368924260139465,
0.046222373843193054,
0.062276311218738556,
0.026550497859716415,
0.04945843294262886,
0.010722825303673744,
-0.05758236348628998,
-0.028889013454318047,
0.015416864305734634,
0.010098161175847054,
0.0883505567908287,
-0.045669589191675186,
0.015606881119310856,
-0.0029096759390085936,
0.07099132984876633,
0.1323941946029663,
-0.03717688098549843,
-0.026572996750473976,
0.11263676732778549,
0.021396707743406296,
-0.015582261607050896,
-0.013582267798483372,
0.06241755560040474,
-0.013818052597343922,
-0.02169426530599594,
0.029687920585274696,
-0.033120330423116684,
0.11824790388345718,
-0.09884489327669144,
0.07018143683671951,
0.026920832693576813,
-0.08230865001678467,
0.00886959582567215,
-0.03778395801782608,
0.026777686551213264,
0.01927538961172104,
-0.027495285496115685,
-0.04856657236814499,
-0.045530155301094055,
-0.04399915412068367,
0.07286267727613449,
-0.1049824133515358,
-0.05063234269618988,
-0.015545937232673168,
0.05145955830812454,
-0.011807223781943321,
-1.61450328728801e-33,
-0.05212986096739769,
-0.030439181253314018,
-0.0726495161652565,
0.11057028919458389,
-0.02638911083340645,
-0.08577638119459152,
-0.018216174095869064,
-0.04122374951839447,
-0.06001066416501999,
0.023120732977986336,
-0.14118586480617523,
-0.05285331979393959,
0.022138571366667747,
-0.0065542361699044704,
0.0507420189678669,
-0.05488266795873642,
0.06305621564388275,
-0.09856832027435303,
0.03989994153380394,
-0.022870536893606186,
0.05806625261902809,
0.05535091832280159,
-0.04888274148106575,
0.0050284238532185555,
-0.012656991370022297,
0.05275736749172211,
-0.020082399249076843,
0.001316057750955224,
0.027982180938124657,
0.01821785792708397,
0.003145834431052208,
-0.008929583244025707,
-0.017626045271754265,
-0.09315966069698334,
-0.06826820224523544,
0.12318561971187592,
-0.022990915924310684,
0.06790689378976822,
-0.019335856661200523,
0.05646934732794762,
0.055481359362602234,
-0.060399241745471954,
-0.07058749347925186,
-0.007200237829238176,
-0.013447805307805538,
0.03862706944346428,
-0.07406943291425705,
0.020463084802031517,
-0.0168197900056839,
0.030067792162299156,
-0.00002100265737681184,
-0.016352858394384384,
-0.04947317764163017,
0.0013214797945693135,
-0.049423400312662125,
-0.026976248249411583,
0.05660521239042282,
0.04713127389550209,
-0.061569780111312866,
-0.03503112867474556,
-0.02000509761273861,
-0.020919589325785637,
-0.02374803088605404,
0.010972646996378899,
0.009374156594276428,
0.021653173491358757,
0.0469679981470108,
0.007363358978182077,
0.004735112190246582,
-0.012370700016617775,
0.12294203042984009,
-0.02069776877760887,
-0.06792452186346054,
-0.009426116943359375,
0.07933370023965836,
0.10705471783876419,
-0.02380458451807499,
-0.038558702915906906,
-0.003166653448715806,
0.01003032922744751,
-0.013036992400884628,
0.07545105367898941,
-0.0054551418870687485,
-0.0000932499606278725,
0.06390590220689774,
0.0006412911461666226,
0.02070167474448681,
0.06453932821750641,
0.020354285836219788,
-0.0036556473933160305,
0.0037597562186419964,
0.027910057455301285,
-0.026977425441145897,
0.04614034667611122,
-0.0424419604241848,
-4.828598676454021e-8,
-0.06011287868022919,
0.0147005096077919,
-0.03009030781686306,
-0.00000567946244700579,
0.031061042100191116,
-0.07511531561613083,
-0.035362359136343,
0.0427321121096611,
0.06158901005983353,
-0.013202263973653316,
0.0966574028134346,
-0.09812689572572708,
-0.016354655846953392,
-0.04586155340075493,
0.04532543569803238,
0.0048377239145338535,
-0.06954001635313034,
-0.09776870161294937,
-0.05512703210115433,
-0.053494371473789215,
-0.02892184443771839,
0.03241898491978645,
-0.059048913419246674,
0.004220982547849417,
-0.00602785125374794,
-0.02960343286395073,
0.0320596918463707,
0.05502656102180481,
0.009979759342968464,
0.018275896087288857,
-0.07062789052724838,
-0.030594052746891975,
-0.015383969061076641,
0.05076827481389046,
-0.013435148634016514,
0.07415925711393356,
-0.027455873787403107,
-0.05078824236989021,
0.02418895997107029,
-0.05653148517012596,
0.0704621896147728,
0.025092002004384995,
-0.01830194890499115,
0.01927049458026886,
0.06944793462753296,
-0.0020849457941949368,
-0.10377681255340576,
-0.08926073461771011,
0.025638483464717865,
-0.05881834030151367,
-0.07282373309135437,
0.03080388903617859,
-0.018245786428451538,
0.12223218381404877,
0.03494398668408394,
-0.01509186066687107,
-0.008278042078018188,
-0.07363367080688477,
0.0837627723813057,
0.06448840349912643,
0.018186422064900398,
-0.030496742576360703,
-0.0533197820186615,
-0.06727342307567596
] |
Revealing and Reducing Gender Biases in Vision and Language Assistants (VLAs) | https://openreview.net/forum?id=oStNAMWELS | [
"Leander Girrbach",
"Stephan Alaniz",
"Yiran Huang",
"Trevor Darrell",
"Zeynep Akata"
] | Poster | Pre-trained large language models (LLMs) have been reliably integrated with visual input for multimodal tasks. The widespread adoption of instruction-tuned image-to-text vision-language assistants (VLAs) like LLaVA and InternVL necessitates evaluating gender biases. We study gender bias in 22 popular open-source VLAs with respect to personality traits, skills, and occupations. Our results show that VLAs replicate human biases likely present in the data, such as real-world occupational imbalances. Similarly, they tend to attribute more skills and positive personality traits to women than to men, and we see a consistent tendency to associate negative personality traits with men. To eliminate the gender bias in these models, we find that fine-tuning-based debiasing methods achieve the best trade-off between debiasing and retaining performance on downstream tasks. We argue for pre-deploying gender bias assessment in VLAs and motivate further development of debiasing strategies to ensure equitable societal outcomes. Code is available at https://github.com/ExplainableML/vla-gender-bias. | gender bias, vision-language-models | null | 2,205 | 2410.19314 | [
-0.013476192019879818,
-0.029973508790135384,
0.037251658737659454,
0.032486315816640854,
0.054440319538116455,
-0.04438120126724243,
-0.013601684011518955,
-0.03712236508727074,
-0.006402806844562292,
-0.02903207205235958,
-0.07075493037700653,
-0.05183033272624016,
-0.008282632566988468,
0.045163799077272415,
0.012831714935600758,
-0.02482481300830841,
0.08458079397678375,
0.04526105895638466,
-0.05623290687799454,
-0.047651395201683044,
0.04738781601190567,
-0.002580351196229458,
0.055234480649232864,
-0.10126811265945435,
0.007885819301009178,
-0.034134529531002045,
-0.00758848013356328,
-0.06153188645839691,
-0.013273452408611774,
-0.04053456336259842,
-0.013102663680911064,
0.02160722203552723,
0.06591177731752396,
0.025731444358825684,
-0.06733259558677673,
0.06897813826799393,
-0.03296184539794922,
0.008417243137955666,
-0.040962327271699905,
0.003871300956234336,
-0.08604971319437027,
-0.010020180605351925,
-0.004886898677796125,
0.006001037545502186,
0.046965017914772034,
-0.05596518889069557,
0.017091115936636925,
-0.06333129107952118,
-0.0653180256485939,
0.027697181329131126,
-0.04533567652106285,
-0.09880772978067398,
0.05666926130652428,
0.05205491557717323,
-0.033411577343940735,
0.019212130457162857,
0.057232148945331573,
0.06538829952478409,
-0.009694396518170834,
0.0489703044295311,
0.0024156789295375347,
-0.05786813795566559,
-0.0771247148513794,
0.024558939039707184,
-0.044883254915475845,
-0.01791706122457981,
0.06672158092260361,
0.021396268159151077,
-0.05763481557369232,
0.008943712338805199,
0.02045905403792858,
0.059745728969573975,
-0.018771113827824593,
0.04764300957322121,
0.048627760261297226,
0.008102894760668278,
0.04826686531305313,
-0.0983830913901329,
0.1242031380534172,
-0.09277887642383575,
-0.00022912600252311677,
-0.023950960487127304,
0.03989933431148529,
0.055864911526441574,
0.10680931806564331,
0.0063346633687615395,
-0.043426867574453354,
0.017674565315246582,
0.02668539248406887,
0.05266223102807999,
-0.11496418714523315,
-0.04581382870674133,
0.06805738806724548,
0.021560855209827423,
0.042865198105573654,
0.029853368178009987,
-0.030924076214432716,
-0.04684863239526749,
-0.03169101104140282,
0.06156345456838608,
-0.0024865313898772,
-0.025004183873534203,
-0.016472648829221725,
0.04454914480447769,
-0.02227119542658329,
-0.0348520390689373,
0.1037120595574379,
-0.007370073348283768,
-0.053461797535419464,
-0.044296037405729294,
0.005450582131743431,
0.03367949277162552,
-0.007545035798102617,
0.019121214747428894,
0.0182945653796196,
-0.06370576471090317,
0.04607579857110977,
0.01858626864850521,
0.010577967390418053,
0.009390615858137608,
0.021501213312149048,
-0.018526727333664894,
0.018286319449543953,
0.018214892596006393,
0.06786764413118362,
0.009253451600670815,
-0.04375449940562248,
-1.2270634257461055e-34,
0.038721714168787,
0.06146342679858208,
0.045298535376787186,
0.04537979140877724,
0.054717641323804855,
-0.00691212248057127,
0.03326074033975601,
0.059642333537340164,
-0.02882341668009758,
-0.03240165859460831,
-0.0359460674226284,
0.05900559574365616,
-0.09374731779098511,
0.06610801815986633,
-0.005283656995743513,
0.017588261514902115,
-0.01382288709282875,
0.09160233289003372,
-0.06983467191457748,
0.05353706330060959,
0.03275316208600998,
-0.018104571849107742,
-0.04913985729217529,
-0.05735039338469505,
0.0021527500357478857,
0.02551320753991604,
0.05365822836756706,
-0.02776087261736393,
-0.00853311363607645,
0.003320543095469475,
-0.08750662952661514,
-0.017723459750413895,
0.051461074501276016,
-0.015213877893984318,
0.0012740087695419788,
-0.0599433109164238,
-0.06510327756404877,
0.06072961539030075,
0.044856369495391846,
0.024822015315294266,
0.025437872856855392,
0.023068469017744064,
0.0641220211982727,
-0.04774976521730423,
-0.030180033296346664,
-0.03403094410896301,
0.057959094643592834,
0.028431426733732224,
-0.04098314791917801,
0.01848156750202179,
0.008089717477560043,
0.058628980070352554,
0.028455324470996857,
-0.012256182730197906,
-0.016968872398138046,
0.02458818443119526,
0.0718485489487648,
0.03215884789824486,
-0.015910571441054344,
-0.05210574343800545,
0.010560700669884682,
0.06140037626028061,
0.000543801870662719,
0.022554529830813408,
-0.004304883535951376,
-0.011563543230295181,
-0.023167602717876434,
0.014068424701690674,
0.009009766392409801,
0.019878702238202095,
-0.04674789682030678,
0.05422193557024002,
-0.0625540092587471,
-0.0008580957655794919,
-0.03945925086736679,
0.08181548863649368,
0.036873672157526016,
-0.02217576466500759,
0.054288968443870544,
-0.0035530433524399996,
0.061431512236595154,
0.0018876096000894904,
0.04330766573548317,
-0.1249176412820816,
-0.011745159514248371,
-0.05114562436938286,
0.029673602432012558,
-0.11937642097473145,
0.0827028751373291,
0.06965526193380356,
0.01684524491429329,
-0.052060700953006744,
-0.03203733265399933,
-0.008061124011874199,
0.0015783602138981223,
-9.880326647313917e-34,
-0.055353377014398575,
-0.002027779584750533,
-0.038502518087625504,
-0.0015445500612258911,
-0.029943477362394333,
-0.0027995898853987455,
0.13288027048110962,
-0.020131397992372513,
0.024169081822037697,
-0.02344108559191227,
-0.037775225937366486,
-0.056093305349349976,
-0.0018112454563379288,
0.02804800681769848,
0.04101109132170677,
-0.09256011992692947,
-0.016000008210539818,
-0.03150298073887825,
0.002241971902549267,
-0.05695748329162598,
0.051975496113300323,
0.12109444290399551,
-0.02366330847144127,
0.049542903900146484,
-0.04536354914307594,
0.02696612849831581,
0.020963208749890327,
-0.04827464371919632,
-0.005868162959814072,
0.025261061266064644,
-0.09124788641929626,
-0.031902119517326355,
-0.05015764385461807,
-0.004612477961927652,
0.04468613117933273,
-0.03955364599823952,
-0.047426071017980576,
-0.021169530227780342,
0.032615043222904205,
0.12010654807090759,
0.03304550424218178,
-0.0026918011717498302,
-0.09838229417800903,
-0.02880083955824375,
-0.057615626603364944,
-0.026313412934541702,
0.004021754954010248,
-0.03786253184080124,
-0.010918869636952877,
0.001256992807611823,
-0.06263083964586258,
-0.00695457449182868,
-0.10008301585912704,
0.009137511253356934,
-0.042844101786613464,
-0.17228221893310547,
0.04950984939932823,
-0.008084394037723541,
0.07991931587457657,
0.06822655349969864,
-0.06796497851610184,
-0.009966222569346428,
-0.012607994489371777,
-0.02319718711078167,
-0.08084859699010849,
-0.05169465020298958,
-0.04159407317638397,
-0.031037569046020508,
0.03476312756538391,
-0.011895522475242615,
0.1333724707365036,
-0.07626869529485703,
-0.044923536479473114,
0.008105187676846981,
-0.07059551775455475,
-0.01977986842393875,
0.022250305861234665,
0.007549635134637356,
-0.032039839774370193,
-0.06853582710027695,
0.043441809713840485,
-0.04532193765044212,
0.07012519240379333,
0.04045123979449272,
0.03654461354017258,
0.1567506045103073,
-0.019236143678426743,
0.06985623389482498,
-0.017448250204324722,
0.03027510456740856,
-0.09204994142055511,
-0.018715424463152885,
-0.045231640338897705,
0.05041604861617088,
-0.04226154088973999,
-4.009653764569521e-8,
-0.06297894567251205,
0.042997535318136215,
0.022800272330641747,
0.05277515947818756,
-0.020759955048561096,
-0.0667293518781662,
-0.07617799192667007,
-0.00642465241253376,
0.057741083204746246,
-0.003908362239599228,
0.03383241593837738,
0.02923230640590191,
0.008936876431107521,
-0.04825976490974426,
0.11287077516317368,
0.06475234031677246,
0.04384028911590576,
0.07046671956777573,
0.004974631126970053,
-0.04582248628139496,
0.15739697217941284,
0.004976353142410517,
-0.08982439339160919,
-0.05759616941213608,
0.01406804658472538,
-0.03731926903128624,
-0.07440048456192017,
0.10835482180118561,
0.01864703744649887,
-0.07138241082429886,
0.03315897285938263,
0.07955390959978104,
-0.038728587329387665,
-0.08581921458244324,
0.10956967622041702,
0.04654737189412117,
0.009894855320453644,
-0.01488801185041666,
0.024515744298696518,
0.021566225215792656,
0.04479120671749115,
0.06636357307434082,
0.00851815938949585,
0.026846421882510185,
-0.0024065538309514523,
0.029624149203300476,
-0.003048700513318181,
-0.06884205341339111,
-0.011986913159489632,
0.03917651250958443,
-0.028989896178245544,
-0.04505148157477379,
0.027808675542473793,
0.09097205102443695,
-0.03374439477920532,
-0.0566069595515728,
0.04574661701917648,
-0.03031773865222931,
-0.013868505135178566,
0.1318366378545761,
0.08019150793552399,
0.036056362092494965,
-0.06604650616645813,
-0.08992554992437363
] |
Zero-Shot Whole-Body Humanoid Control via Behavioral Foundation Models | https://openreview.net/forum?id=9sOR0nYLtz | [
"Andrea Tirinzoni",
"Ahmed Touati",
"Jesse Farebrother",
"Mateusz Guzek",
"Anssi Kanervisto",
"Yingchen Xu",
"Alessandro Lazaric",
"Matteo Pirotta"
] | Poster | Unsupervised reinforcement learning (RL) aims at pre-training models that can solve a wide range of downstream tasks in complex environments. Despite recent advancements, existing approaches suffer from several limitations: they may require running an RL process on each task to achieve a satisfactory performance, they may need access to datasets with good coverage or well-curated task-specific samples, or they may pre-train policies with unsupervised losses that are poorly correlated with the downstream tasks of interest. In this paper, we introduce FB-CPR, which regularizes unsupervised zero-shot RL based on the forward-backward (FB) method towards imitating trajectories from unlabeled behaviors. The resulting models learn useful policies imitating the behaviors in the dataset, while retaining zero-shot generalization capabilities. We demonstrate the effectiveness of FB-CPR in a challenging humanoid control problem. Training FB-CPR online with observation-only motion capture datasets, we obtain the first humanoid behavioral foundation model that can be prompted to solve a variety of whole-body tasks, including motion tracking, goal reaching, and reward optimization. The resulting model is capable of expressing human-like behaviors and it achieves competitive performance with task-specific methods while outperforming state-of-the-art unsupervised RL and model-based baselines. | reinforcement learning; foundation model; humanoid | We present a novel unsupervised reinforcement learning algorithm that leverages unlabeled behavior data to create a foundation model for controlling a humanoid agent in a zero-shot fashion. | 2,200 | null | [
-0.040122877806425095,
-0.012229017913341522,
0.059588074684143066,
0.03722980618476868,
0.06730917096138,
0.019352562725543976,
-0.01783057674765587,
-0.06293696165084839,
-0.0455101914703846,
0.03619597107172012,
0.006412992719560862,
-0.09440603852272034,
0.008043098263442516,
0.045221198350191116,
0.03167163208127022,
-0.016787853091955185,
0.11576158553361893,
0.032137610018253326,
-0.046221401542425156,
0.05784051492810249,
-0.0950550064444542,
-0.010371148586273193,
0.11337199062108994,
0.05677579715847969,
-0.044311705976724625,
-0.055988434702157974,
0.007261350285261869,
-0.13016299903392792,
0.000567399023566395,
-0.06117239221930504,
0.07706854492425919,
-0.03923102095723152,
0.009948180988430977,
-0.05620679259300232,
-0.08895643800497055,
0.08461353927850723,
0.003606140846386552,
0.025215068832039833,
-0.07273557782173157,
-0.015441750176250935,
0.004598964937031269,
0.024071425199508667,
0.0023004154209047556,
0.04433882609009743,
0.06826825439929962,
0.11684246361255646,
-0.013266826048493385,
-0.05978569760918617,
0.04925256595015526,
-0.024552293121814728,
-0.014638242311775684,
-0.023488730192184448,
0.05093395337462425,
0.09120192378759384,
0.028944194316864014,
0.014455524273216724,
-0.0014576246030628681,
-0.004530724138021469,
0.020077615976333618,
-0.049787506461143494,
-0.02627023495733738,
-0.13019852340221405,
-0.047473352402448654,
-0.06889908760786057,
-0.02932986617088318,
-0.01075691170990467,
-0.0245453380048275,
0.04532964527606964,
0.04098378121852875,
0.047373633831739426,
-0.010275699198246002,
-0.0013202273985370994,
-0.00792800821363926,
0.019194934517145157,
0.027683522552251816,
0.0054033538326621056,
-0.04222455248236656,
0.010061708278954029,
-0.01168854534626007,
-0.001037887646816671,
0.03876364603638649,
-0.05670328065752983,
-0.02298749051988125,
0.028640175238251686,
0.051174383610486984,
0.002816180931404233,
0.004028161987662315,
0.05552705004811287,
0.05449112132191658,
-0.00011961702693952248,
-0.053720563650131226,
0.06650445610284805,
0.0470968522131443,
-0.035839833319187164,
-0.03944810479879379,
-0.01573110744357109,
-0.049283672124147415,
-0.05601460114121437,
0.024801459163427353,
0.06233945116400719,
-0.0018913160311058164,
-0.004724928177893162,
0.034053631126880646,
0.08992742002010345,
0.002292533405125141,
-0.05410581827163696,
0.015093219466507435,
0.03808575123548508,
0.05595291405916214,
-0.014764124527573586,
0.04600469768047333,
-0.0382770299911499,
0.04267491027712822,
0.024871861562132835,
0.05098402872681618,
-0.02391424961388111,
-0.06908250600099564,
-0.02156819775700569,
0.01332506537437439,
0.005742897279560566,
-0.02998216450214386,
-0.024354439228773117,
0.01773003116250038,
0.01970195770263672,
0.010240030474960804,
-0.08361636102199554,
-0.02644650638103485,
2.174772731235706e-33,
0.06007532402873039,
0.05659698694944382,
0.09265659004449844,
-0.03699922561645508,
0.015026212669909,
0.010645359754562378,
0.018988333642482758,
0.012312930077314377,
-0.0004150095919612795,
0.03821493312716484,
0.05401165783405304,
-0.030908575281500816,
-0.0416841097176075,
0.1319248378276825,
0.028009751811623573,
-0.05975799635052681,
-0.04907926544547081,
0.028802670538425446,
0.006611226126551628,
0.04784548655152321,
0.10226248949766159,
-0.003726192284375429,
-0.05729864910244942,
-0.05154796317219734,
0.06987426429986954,
-0.004461000207811594,
-0.008859413675963879,
-0.023417547345161438,
-0.03794794902205467,
0.005754677578806877,
-0.06665486097335815,
0.01094615925103426,
0.0018202535575255752,
0.052413273602724075,
0.008890405297279358,
-0.008528994396328926,
-0.03641108423471451,
-0.03308609127998352,
0.0022957625333219767,
0.004701945465058088,
0.01804431900382042,
0.01264871470630169,
0.02539238892495632,
0.0015170280821621418,
-0.09007034450769424,
-0.13990280032157898,
0.03364042937755585,
0.07153663784265518,
-0.09546028822660446,
0.010303297080099583,
0.03153776377439499,
0.011765922419726849,
-0.041828904300928116,
-0.09902868419885635,
-0.09755595028400421,
0.005522725638002157,
-0.0034966012462973595,
-0.0034952752757817507,
-0.0502002015709877,
0.007448340300470591,
0.09282663464546204,
0.019896529614925385,
-0.04159970581531525,
0.02587367594242096,
-0.05473922938108444,
0.052398860454559326,
-0.05214676633477211,
0.06322913616895676,
0.10246311873197556,
-0.05342836305499077,
-0.016668878495693207,
0.03985702991485596,
0.008337249048054218,
-0.04952099174261093,
0.13845904171466827,
0.05765002965927124,
0.06253478676080704,
-0.10768773406744003,
-0.01242104358971119,
0.03924005851149559,
-0.024710090830922127,
0.015010007657110691,
-0.06201333552598953,
0.036181170493364334,
0.013019271194934845,
-0.05907796695828438,
0.07624109089374542,
-0.048270225524902344,
-0.04973023384809494,
-0.03614417463541031,
-0.05331074446439743,
-0.09663529694080353,
-0.11259686201810837,
0.08229345828294754,
0.005097867455333471,
-1.6249025552019488e-33,
0.10174194723367691,
0.00773497112095356,
-0.019902726635336876,
-0.0006177127943374217,
-0.03577661141753197,
0.0008016807842068374,
0.049966294318437576,
-0.00804251991212368,
0.09632375836372375,
-0.037238381803035736,
-0.05559713765978813,
-0.0343032106757164,
0.06807263940572739,
0.0071266889572143555,
0.01469912938773632,
0.005143019370734692,
0.009509231895208359,
0.008141543716192245,
-0.019664723426103592,
0.008615667000412941,
-0.025210004299879074,
0.08095526695251465,
-0.05452137067914009,
0.005407453514635563,
-0.04404177516698837,
0.05262290686368942,
0.02355344593524933,
0.09188351780176163,
0.014032823964953423,
0.041336789727211,
-0.03541421517729759,
-0.02989673800766468,
-0.005253627896308899,
-0.034924767911434174,
-0.1182328313589096,
0.07382958382368088,
-0.0281611829996109,
0.09087769687175751,
0.028618330135941505,
0.044096652418375015,
0.06495629251003265,
-0.05932089686393738,
-0.048519670963287354,
0.06032567843794823,
0.015509510412812233,
-0.05266953259706497,
-0.0916115865111351,
0.038919053971767426,
-0.027566365897655487,
0.016998445615172386,
-0.11834399402141571,
0.051959507167339325,
-0.07725585997104645,
-0.04476355388760567,
-0.07733481377363205,
-0.10332854837179184,
-0.0010393131524324417,
0.03216700628399849,
0.030458081513643265,
-0.021076440811157227,
-0.015521303750574589,
0.05949088931083679,
-0.005209681577980518,
0.05824613943696022,
-0.012517763301730156,
0.021224819123744965,
-0.003773157950490713,
0.03944208845496178,
0.052342284470796585,
-0.028202634304761887,
0.03493209555745125,
0.015185524709522724,
0.05633113160729408,
-0.027938615530729294,
0.029849642887711525,
0.02826261892914772,
-0.022495798766613007,
-0.004588688258081675,
0.017161324620246887,
-0.016048980876803398,
0.02839851751923561,
-0.052977196872234344,
0.04860001429915428,
0.05517439916729927,
0.07655439525842667,
0.12647268176078796,
-0.03038344718515873,
0.049873460084199905,
-0.010979815386235714,
-0.015681134536862373,
-0.04340527579188347,
0.01588428020477295,
0.045759715139865875,
0.08663100749254227,
-0.010810306295752525,
-4.061787706177711e-8,
-0.08649373054504395,
0.012593170627951622,
0.02825625240802765,
0.06690923869609833,
0.016065483912825584,
0.042566996067762375,
-0.009046708233654499,
-0.07098353654146194,
-0.018547307699918747,
-0.037967223674058914,
0.04093626141548157,
0.024546464905142784,
0.01233681384474039,
-0.07211809605360031,
-0.11845303326845169,
0.0578412227332592,
0.019275208935141563,
-0.01372298039495945,
-0.030798785388469696,
0.04578552022576332,
-0.007425714284181595,
-0.03229532763361931,
-0.014167112298309803,
-0.09007541090250015,
0.0435045100748539,
-0.07577506452798843,
-0.03367156907916069,
0.014138037338852882,
0.08482179790735245,
-0.06110648810863495,
0.006566766183823347,
-0.007813388481736183,
0.043156616389751434,
-0.008411325514316559,
0.020776554942131042,
0.06706356257200241,
0.02261848747730255,
-0.013176574371755123,
-0.04259604588150978,
-0.0741770938038826,
-0.03563784807920456,
0.10196666419506073,
-0.0180637389421463,
-0.02749692276120186,
-0.07881096005439758,
0.014365357346832752,
-0.028560034930706024,
-0.12130200862884521,
0.08487430214881897,
-0.04723582789301872,
0.020482484251260757,
-0.004388574976474047,
-0.013703126460313797,
0.12007499486207962,
0.09514021873474121,
0.04584728926420212,
-0.06416771560907364,
-0.02884661965072155,
0.03152291476726532,
0.07297281175851822,
0.02381434664130211,
-0.023154528811573982,
-0.07811246812343597,
0.008753083646297455
] |
Outlier Synthesis via Hamiltonian Monte Carlo for Out-of-Distribution Detection | https://openreview.net/forum?id=N6ba2xsmds | [
"Hengzhuang Li",
"Teng Zhang"
] | Poster | Out-of-distribution (OOD) detection is crucial for developing trustworthy and reliable machine learning systems. Recent advances in training with auxiliary OOD data demonstrate efficacy in enhancing detection capabilities. Nonetheless, these methods heavily rely on acquiring a large pool of high-quality natural outliers. Some prior methods try to alleviate this problem by synthesizing virtual outliers but suffer from either poor quality or high cost due to the monotonous sampling strategy and the heavy-parameterized generative models. In this paper, we overcome all these problems by proposing the Hamiltonian Monte Carlo Outlier Synthesis (HamOS) framework, which views the synthesis process as sampling from Markov chains. Based solely on the in-distribution data, the Markov chains can extensively traverse the feature space and generate diverse and representative outliers, hence exposing the model to miscellaneous potential OOD scenarios. The Hamiltonian Monte Carlo with sampling acceptance rate almost close to 1 also makes our framework enjoy great efficiency. By empirically competing with SOTA baselines on both standard and large-scale benchmarks, we verify the efficacy and efficiency of our proposed HamOS. | Trustworthy Machine Learning, Out-of-Distribution Detection, Outlier Detection | null | 2,198 | 2501.16718 | [
-0.07985281944274902,
-0.01427974458783865,
0.04073091223835945,
0.0738847628235817,
0.082514688372612,
-0.10121846944093704,
-0.03692862391471863,
0.006627388298511505,
-0.014708338305354118,
-0.036358002573251724,
-0.0519084706902504,
-0.02942667342722416,
0.06833948940038681,
-0.0770486518740654,
-0.006731455679982901,
-0.03156706690788269,
0.11775800585746765,
0.09275196492671967,
-0.009852216579020023,
-0.035284895449876785,
-0.04306920990347862,
0.06848356872797012,
-0.009644664824008942,
0.01540056336671114,
0.018843531608581543,
-0.013433362357318401,
0.038558878004550934,
-0.05358171835541725,
-0.07324816286563873,
-0.04711643233895302,
-0.025524111464619637,
0.0353405624628067,
-0.025947529822587967,
-0.1377428025007248,
0.046065058559179306,
0.022509030997753143,
-0.028643865138292313,
0.016072362661361694,
-0.007052095141261816,
-0.040218859910964966,
0.06039205938577652,
0.04539583623409271,
0.05446071922779083,
-0.009608692489564419,
-0.001796908676624298,
-0.027337227016687393,
0.04218902066349983,
-0.07109037786722183,
0.028227858245372772,
-0.03638002276420593,
-0.03162349760532379,
-0.008871008642017841,
-0.012066797353327274,
0.030512988567352295,
-0.05532640591263771,
-0.018507611006498337,
-0.045341432094573975,
-0.08770375698804855,
0.022146668285131454,
0.00728561682626605,
-0.0299686249345541,
-0.0940539762377739,
-0.015726326033473015,
-0.05357039347290993,
-0.00502783665433526,
-0.005752077791839838,
0.06456675380468369,
0.053774598985910416,
0.019830700010061264,
0.01912717893719673,
0.029242271557450294,
0.10850225389003754,
-0.0381004773080349,
0.013722257688641548,
0.019103942438960075,
0.039709072560071945,
-0.01910644955933094,
-0.05246072635054588,
0.028549350798130035,
-0.03074743039906025,
0.0177793949842453,
0.0033135628327727318,
0.032339904457330704,
-0.052610285580158234,
0.07177270948886871,
0.04302211478352547,
0.012645075097680092,
0.11618829518556595,
0.07829280942678452,
0.01564999297261238,
-0.06199897080659866,
0.03319340571761131,
0.00919346697628498,
-0.03332965821027756,
0.0208950936794281,
0.02548724040389061,
0.1317373812198639,
-0.01743718609213829,
0.10860993713140488,
0.10917908698320389,
-0.014427426271140575,
0.034960828721523285,
0.04609481617808342,
-0.027723345905542374,
0.09505623579025269,
-0.036230914294719696,
0.04856494069099426,
0.030347326770424843,
0.009590061381459236,
-0.07122547924518585,
0.003456573700532317,
-0.013739160262048244,
0.099125437438488,
-0.056321389973163605,
0.03978995606303215,
0.013316292315721512,
0.07593114674091339,
0.010090150870382786,
-0.11293477565050125,
0.019019465893507004,
-0.044229310005903244,
0.018208293244242668,
0.027625102549791336,
0.09676046669483185,
-0.027382992208003998,
-0.01885835826396942,
-0.07488491386175156,
3.4891614741702395e-33,
0.03486078232526779,
0.07127903401851654,
0.12138166278600693,
-0.02676030807197094,
0.1064600795507431,
-0.019293805584311485,
0.03496934473514557,
-0.044284190982580185,
-0.00788914691656828,
0.008816907182335854,
-0.02766144461929798,
-0.08697834610939026,
-0.0731498971581459,
0.026351863518357277,
0.06458724290132523,
0.014906158670783043,
-0.037008851766586304,
0.055185653269290924,
-0.037369079887866974,
-0.01999906450510025,
0.032125476747751236,
-0.11392329633235931,
-0.05756473168730736,
0.02233770862221718,
-0.0640118420124054,
0.09389226883649826,
0.04343673959374428,
-0.008614348247647285,
0.008538677357137203,
0.0002899552637245506,
0.02315199002623558,
0.058770954608917236,
-0.0430404357612133,
0.055180978029966354,
0.05292671546339989,
-0.0611991360783577,
-0.0796276107430458,
-0.029423903673887253,
-0.0367005318403244,
-0.01100146770477295,
-0.02998063713312149,
-0.0011315199080854654,
-0.02961273491382599,
-0.028634676709771156,
-0.0656498447060585,
-0.023119695484638214,
0.09613905102014542,
-0.05698410049080849,
0.03587966039776802,
-0.09063423424959183,
0.006630624178797007,
-0.007505872752517462,
0.03667733073234558,
0.02832331508398056,
-0.09347653388977051,
-0.06992333382368088,
0.07927452772855759,
0.0016431710682809353,
0.013502256944775581,
0.11940165609121323,
0.02414667047560215,
0.024498051032423973,
-0.018978893756866455,
0.04978901892900467,
-0.040752191096544266,
0.003476967103779316,
0.019378039985895157,
0.03471294417977333,
0.012150648981332779,
0.09208104759454727,
0.03193141520023346,
0.011219325475394726,
-0.09842317551374435,
-0.02832861617207527,
0.03545273095369339,
0.016953902319073677,
0.045163147151470184,
0.00709133455529809,
0.02167382463812828,
-0.02497069165110588,
-0.02751888893544674,
-0.0430111438035965,
-0.013295791111886501,
-0.0024374560452997684,
-0.004475110210478306,
-0.025980671867728233,
-0.03050205484032631,
0.015681974589824677,
-0.06926388293504715,
0.06367724388837814,
-0.09368572384119034,
-0.04917898401618004,
-0.06749606132507324,
-0.03684234991669655,
-0.021083446219563484,
-2.8811974489638833e-33,
-0.0325164869427681,
0.060212116688489914,
0.032890357077121735,
0.009211805649101734,
-0.030985653400421143,
-0.07062986493110657,
-0.07737869769334793,
0.02722414955496788,
0.07471928745508194,
-0.07275542616844177,
-0.15528930723667145,
-0.028427861630916595,
0.1038724035024643,
-0.033273618668317795,
0.0430169552564621,
-0.0361165851354599,
-0.019749991595745087,
0.00009777829109225422,
-0.05266333743929863,
-0.02203376777470112,
0.0028642723336815834,
-0.016382867470383644,
-0.0298598799854517,
-0.03444116562604904,
0.010447666980326176,
0.04874038323760033,
-0.0701262578368187,
0.06721284985542297,
-0.039828743785619736,
-0.00988752767443657,
-0.07800442725419998,
0.044258374720811844,
-0.01301646139472723,
0.005441382527351379,
-0.04622546210885048,
0.04831701144576073,
0.05803336575627327,
0.04450284689664841,
-0.011765338480472565,
0.07600273936986923,
-0.05376119166612625,
0.0036832906771451235,
-0.06527020782232285,
0.04948705434799194,
0.0177089124917984,
0.057703953236341476,
-0.0016365619376301765,
0.04184827208518982,
0.04553905129432678,
0.027818435803055763,
-0.007527276407927275,
0.10533460229635239,
-0.02417740970849991,
0.09332156926393509,
-0.06504411995410919,
-0.04031567648053169,
-0.06086095795035362,
0.027400163933634758,
0.06803234666585922,
0.023161934688687325,
-0.0675477609038353,
-0.059254277497529984,
0.02829592674970627,
-0.05608852580189705,
-0.02436707355082035,
-0.02524847909808159,
-0.04168511927127838,
-0.0047887456603348255,
0.009552699513733387,
0.047012172639369965,
-0.03608214110136032,
-0.04188874363899231,
0.006948638707399368,
-0.0027718269266188145,
-0.0028026122599840164,
-0.051388971507549286,
0.021113647148013115,
-0.08674906194210052,
0.0008278428576886654,
-0.024683229625225067,
0.016288142651319504,
-0.028354499489068985,
0.07027122378349304,
0.026933597400784492,
0.07098448276519775,
0.0525767020881176,
0.10885117948055267,
-0.034597743302583694,
0.006566551513969898,
-0.018551524728536606,
-0.02277488261461258,
0.032469768077135086,
-0.0010115456534549594,
0.053536638617515564,
-0.08487013727426529,
-4.0290725422664764e-8,
-0.05697491765022278,
0.025556214153766632,
-0.007602552883327007,
0.007371561136096716,
0.12342677265405655,
0.07177697867155075,
-0.028462303802371025,
0.08983638882637024,
0.004597736056894064,
-0.039805904030799866,
0.08598143607378006,
-0.0881824940443039,
0.01873747818171978,
-0.059316735714673996,
0.036920372396707535,
-0.0006412464426830411,
-0.037936750799417496,
0.10547207295894623,
-0.048917755484580994,
-0.00536021264269948,
-0.018209878355264664,
0.06273392587900162,
0.018859902396798134,
0.004564821720123291,
0.06856879591941833,
-0.05170901119709015,
0.0023160965647548437,
0.01323018316179514,
0.10943914949893951,
0.026632020249962807,
0.005874677561223507,
-0.05804159492254257,
0.0651242658495903,
0.016142183914780617,
0.002385079627856612,
0.09879156947135925,
-0.05636082962155342,
-0.03130175173282623,
0.0027673840522766113,
-0.08051956444978714,
-0.022516874596476555,
0.05026097595691681,
-0.09495427459478378,
0.03164706751704216,
-0.04544813930988312,
-0.01119066122919321,
-0.02795104868710041,
-0.05115433409810066,
0.03845972195267677,
-0.03652631491422653,
0.016356732696294785,
-0.07068982720375061,
0.09047980606555939,
0.027981150895357132,
0.005820188671350479,
-0.001862140721641481,
0.02353077195584774,
-0.02628127485513687,
0.010574989020824432,
0.008187933824956417,
0.03391829505562782,
-0.031984664499759674,
-0.008788841776549816,
0.0029120156541466713
] |
Enhancing Cognition and Explainability of Multimodal Foundation Models with Self-Synthesized Data | https://openreview.net/forum?id=lHbLpwbEyt | [
"Yucheng Shi",
"Quanzheng Li",
"Jin Sun",
"Xiang Li",
"Ninghao Liu"
] | Poster | Large Multimodal Models (LMMs), or Vision-Language Models (VLMs), have shown impressive capabilities in a wide range of visual tasks. However, they often struggle with fine-grained visual reasoning, failing to identify domain-specific objectives and provide justifiable explanations for their predictions. To address the above challenge, we propose a novel visual rejection sampling framework to improve the cognition and explainability of LMMs using self-synthesized data. Specifically, visual fine-tuning requires images, queries, and target answers. Our approach begins by synthesizing interpretable answers that include human-verifiable visual features. These features are based on expert-defined concepts, and carefully selected based on their alignment with the image content. After each round of fine-tuning, we apply a reward model-free filtering mechanism to select the highest-quality interpretable answers for the next round of tuning. This iterative process of synthetic data generation and fine-tuning progressively improves the model's ability to generate accurate and reasonable explanations. Experimental results demonstrate the effectiveness of our method in improving both the accuracy and explainability of specialized visual classification tasks. | Multimodal Foundation Models, Synthetic Data, Explainability, Visual Reasoning, Fine-grained Visual Categorization, Rejection Sampling | null | 2,192 | 2502.14044 | [
0.029716771095991135,
-0.07238834351301193,
0.024076512083411217,
0.027196034789085388,
0.1397610604763031,
-0.020006386563181877,
0.024345478042960167,
0.021761171519756317,
0.01132143009454012,
-0.015877308323979378,
-0.07497711479663849,
-0.16894984245300293,
0.03933893144130707,
0.05807584896683693,
-0.0295201875269413,
0.0012136389268562198,
0.07086490094661713,
0.05983070284128189,
-0.07024631649255753,
-0.02725224383175373,
0.06503893435001373,
-0.00257353694178164,
0.07928968220949173,
-0.0050758616998791695,
0.012324603274464607,
0.0223280880600214,
0.000898119353223592,
-0.043790802359580994,
-0.008883745409548283,
-0.05674847960472107,
0.026977550238370895,
-0.0159390140324831,
0.011452475562691689,
0.059711605310440063,
-0.04855263978242874,
0.06510987877845764,
-0.04269367456436157,
0.05213220417499542,
-0.053103648126125336,
-0.08431921899318695,
-0.06467732042074203,
-0.007140421774238348,
0.04613221436738968,
0.024160340428352356,
0.06991732120513916,
0.020835047587752342,
0.007734778802841902,
-0.0641385093331337,
0.03646925464272499,
-0.029131045565009117,
-0.09084978699684143,
-0.04475957155227661,
-0.05035751312971115,
-0.033496737480163574,
-0.03869566693902016,
0.05920831859111786,
0.03434169292449951,
0.002311549847945571,
0.03119725175201893,
-0.04151466116309166,
0.020973043516278267,
-0.04323191940784454,
-0.010919368825852871,
0.0331040658056736,
0.048277847468853,
-0.025544803589582443,
-0.00419317651540041,
0.00430271215736866,
-0.007295296993106604,
0.004137518350034952,
-0.060537274926900864,
0.03370881825685501,
0.028997160494327545,
0.02436012029647827,
0.03236129507422447,
0.10760332643985748,
-0.01702042669057846,
0.00016877707093954086,
0.052933353930711746,
-0.07223718613386154,
-0.027797458693385124,
-0.030071519315242767,
-0.013653941452503204,
0.03586908429861069,
0.10065074265003204,
0.04541003331542015,
0.005910915322601795,
0.05338689684867859,
0.004587194416671991,
-0.014648743905127048,
-0.08553265035152435,
-0.06754159927368164,
-0.0341932438313961,
-0.019372178241610527,
0.0644935891032219,
0.050535768270492554,
0.026762844994664192,
-0.14415580034255981,
-0.05679461359977722,
0.0375337190926075,
0.029804155230522156,
0.05957641452550888,
0.007815385237336159,
-0.03989627584815025,
0.052462391555309296,
-0.05066811665892601,
0.07790035754442215,
0.02291719615459442,
0.02629448100924492,
-0.07545500248670578,
0.017068304121494293,
0.039177898317575455,
0.009946567006409168,
-0.009867687709629536,
0.010687701404094696,
0.013811388052999973,
0.008597582578659058,
-0.002959331963211298,
0.016143888235092163,
-0.07086293399333954,
0.02596709318459034,
0.00438065268099308,
0.0014203917235136032,
-0.042284999042749405,
0.08406378328800201,
-0.041510675102472305,
-0.030208192765712738,
-4.995974050559574e-34,
0.04899667575955391,
0.018316475674510002,
0.11522546410560608,
0.07135436683893204,
0.09569980204105377,
-0.03331713750958443,
-0.026192698627710342,
0.04296748712658882,
-0.004261528141796589,
0.07286010682582855,
-0.01693638041615486,
0.013879280537366867,
-0.05838489532470703,
0.17276601493358612,
0.029593205079436302,
0.02123837359249592,
-0.0979190245270729,
0.0007586712599731982,
-0.05280434712767601,
0.02361472323536873,
0.0034633800387382507,
-0.011197634041309357,
-0.02792005054652691,
-0.04146946966648102,
0.006519701797515154,
-0.03554002568125725,
0.02308047190308571,
-0.04822208732366562,
-0.07908949255943298,
-0.011894085444509983,
-0.13682228326797485,
0.0217827707529068,
0.014811324886977673,
-0.03321722894906998,
0.01643642596900463,
-0.0904611274600029,
0.026683494448661804,
-0.014147913083434105,
0.028144199401140213,
-0.009409293532371521,
0.0216746274381876,
-0.0007297132979147136,
-0.009929999709129333,
0.06719453632831573,
-0.046005651354789734,
0.025996429845690727,
0.016374165192246437,
-0.02452133223414421,
-0.10446232557296753,
-0.022955462336540222,
0.041662972420454025,
0.022764721885323524,
-0.024682797491550446,
-0.08021929115056992,
-0.04245268180966377,
-0.011724526062607765,
0.01042979583144188,
0.02903209626674652,
0.01253833994269371,
-0.014621753245592117,
-0.0034306638408452272,
0.03877905011177063,
-0.04951168969273567,
0.1173478364944458,
-0.0034050634130835533,
0.006462811958044767,
0.017731714993715286,
0.04165671393275261,
0.036057475954294205,
0.05334734916687012,
-0.03557998687028885,
0.013549256138503551,
-0.0016734101809561253,
-0.024545060470700264,
0.0029087017755955458,
0.011444112285971642,
0.0656098797917366,
-0.0031784088350832462,
0.07283266633749008,
0.0060611325316131115,
0.06506931036710739,
-0.019854571670293808,
-0.07295501977205276,
-0.07110857218503952,
-0.04816219210624695,
-0.02385236881673336,
0.07941070944070816,
-0.16559270024299622,
-0.055287525057792664,
-0.07176413387060165,
-0.06042274460196495,
-0.0429404154419899,
0.03422866761684418,
-0.041624765843153,
-0.021599076688289642,
-1.6778833024138334e-34,
0.017712607979774475,
0.018539171665906906,
-0.009612408466637135,
0.024968909099698067,
-0.00888003595173359,
0.006146453320980072,
0.01397239975631237,
0.015744779258966446,
0.06512393802404404,
-0.036137450486421585,
-0.03321158140897751,
-0.01940799690783024,
0.006185309030115604,
0.013039831072092056,
-0.04767512530088425,
-0.040250230580568314,
-0.0042288959957659245,
-0.010089855641126633,
-0.047132331877946854,
0.0012840591371059418,
-0.030201802030205727,
0.1557166576385498,
-0.07717777788639069,
-0.0033623692579567432,
-0.06089318171143532,
0.1123306006193161,
-0.036830224096775055,
0.013951455242931843,
-0.022138245403766632,
0.005095649976283312,
-0.050120845437049866,
-0.027820458635687828,
0.004500876180827618,
-0.02216547727584839,
-0.0225335992872715,
0.03649220988154411,
0.06360414624214172,
-0.08081194013357162,
-0.05937548726797104,
0.11699759215116501,
0.03515760973095894,
-0.04024896025657654,
-0.09072054177522659,
-0.0053908028639853,
-0.02441376820206642,
-0.0963035374879837,
-0.004592099227011204,
-0.04689125344157219,
0.005413513630628586,
0.09900464862585068,
0.009805228561162949,
0.039197418838739395,
-0.10561656951904297,
0.009132112376391888,
0.0016635566717013717,
-0.10245822370052338,
-0.06187378987669945,
0.004382316488772631,
0.15773336589336395,
0.03613162785768509,
-0.05941779538989067,
0.002224525436758995,
-0.008244949392974377,
-0.02984005957841873,
-0.0006610528216697276,
-0.0078085921704769135,
-0.053644806146621704,
-0.00946780014783144,
0.07525824755430222,
-0.008191699162125587,
-0.018392041325569153,
-0.078140988945961,
0.038690585643053055,
0.018137816339731216,
0.07076182216405869,
-0.017722563818097115,
0.0630892738699913,
-0.0008939974941313267,
-0.013833414763212204,
-0.06849942356348038,
0.08151253312826157,
-0.05431237444281578,
0.06151942163705826,
0.117744579911232,
0.08416983485221863,
0.04482804611325264,
0.007008010521531105,
0.04098290205001831,
0.04904656857252121,
0.04016965255141258,
-0.0029069262091070414,
-0.03131015971302986,
-0.04226573184132576,
0.19311442971229553,
-0.0188932865858078,
-4.113649509918105e-8,
-0.07692944258451462,
0.07867725938558578,
-0.02863856963813305,
0.009436129592359066,
-0.007791766431182623,
-0.00826573558151722,
-0.061836205422878265,
0.01462451834231615,
-0.014200665056705475,
-0.073662668466568,
0.07803649455308914,
-0.004171621985733509,
-0.06171518936753273,
0.02128598466515541,
0.022359805181622505,
0.07057761400938034,
-0.018010299652814865,
0.07417483627796173,
-0.010355173610150814,
-0.007496465928852558,
-0.005367976147681475,
-0.006549779325723648,
-0.001281251898035407,
-0.03172270208597183,
0.021743077784776688,
-0.01620875485241413,
-0.041231025010347366,
0.05300310254096985,
0.009254478849470615,
0.05932680889964104,
0.03170866146683693,
0.004830712452530861,
0.008906163275241852,
-0.018144086003303528,
0.042234793305397034,
0.02146337926387787,
-0.02171327918767929,
-0.03347276151180267,
-0.0004938107449561357,
-0.02114199474453926,
-0.004041195847094059,
0.1389007568359375,
-0.031243247911334038,
0.05528198182582855,
0.014622891321778297,
-0.007117395754903555,
-0.030580660328269005,
-0.05947800725698471,
0.017610903829336166,
0.020314645022153854,
-0.02573281154036522,
-0.027787357568740845,
0.014507274143397808,
0.1133272647857666,
0.03913601487874985,
-0.06939742714166641,
-0.00009715818305267021,
0.04530922323465347,
0.04368411749601364,
0.052055247128009796,
0.035218723118305206,
0.028192035853862762,
0.02105131559073925,
-0.002255090046674013
] |
SMI-Editor: Edit-based SMILES Language Model with Fragment-level Supervision | https://openreview.net/forum?id=M29nUGozPa | [
"Kangjie Zheng",
"Siyue Liang",
"Junwei Yang",
"Bin Feng",
"Zequn Liu",
"Wei Ju",
"Zhiping Xiao",
"Ming Zhang"
] | Poster | SMILES, a crucial textual representation of molecular structures, has garnered significant attention as a foundation for pre-trained language models (LMs). However, most existing pre-trained SMILES LMs focus solely on the single-token level supervision during pre-training, failing to fully leverage the substructural information of molecules. This limitation makes the pre-training task overly simplistic, preventing the models from capturing richer molecular semantic information. Moreover, during pre-training, these SMILES LMs only process corrupted SMILES inputs, never encountering any valid SMILES, which leads to a train-inference mismatch. To address these challenges, we propose SMI-Editor, a novel edit-based pre-trained SMILES LM. SMI-Editor disrupts substructures within a molecule at random and feeds the resulting SMILES back into the model, which then attempts to restore the original SMILES through an editing process. This approach not only introduces fragment-level training signals, but also enables the use of valid SMILES as inputs, allowing the model to learn how to reconstruct complete molecules from these incomplete structures. As a result, the model demonstrates improved scalability and an enhanced ability to capture fragment-level molecular information. Experimental results show that SMI-Editor achieves state-of-the-art performance across multiple downstream molecular tasks, and even outperforming several 3D molecular representation models. | SMILES Language Model, SMILES Pre-training Model, Molecular Pre-training Model | null | 2,185 | null | [
-0.05379834026098251,
-0.05523836985230446,
0.08270895481109619,
0.08304598182439804,
0.08707746118307114,
-0.007142237387597561,
-0.0610056072473526,
-0.012257733382284641,
0.04106760025024414,
-0.022606244310736656,
-0.004885571543127298,
-0.04338439553976059,
-0.0196683406829834,
0.05409994721412659,
0.05829132720828056,
0.037416741251945496,
-0.06521529704332352,
0.06716582924127579,
-0.07077455520629883,
-0.045249905437231064,
0.07416832447052002,
0.06538215279579163,
0.019171904772520065,
0.028769485652446747,
0.0186742153018713,
0.047585729509592056,
-0.0023101798724383116,
0.02346402406692505,
0.04076381400227547,
-0.011653824709355831,
0.053326673805713654,
0.010249165818095207,
0.11032574623823166,
0.035665273666381836,
-0.021002143621444702,
0.05058356374502182,
-0.08567991107702255,
-0.005205897614359856,
0.04355766996741295,
-0.10038392245769501,
-0.018831653520464897,
-0.056273508816957474,
-0.061115823686122894,
0.04074471443891525,
0.07812797278165817,
-0.05138670280575752,
0.023787349462509155,
-0.003895342582836747,
-0.0191980991512537,
0.02878628857433796,
-0.07497694343328476,
-0.10297229886054993,
-0.04017171263694763,
0.018796270713210106,
-0.06763151288032532,
0.05296142026782036,
-0.03890956938266754,
-0.014966603368520737,
0.009026464074850082,
-0.031348615884780884,
-0.006279978901147842,
-0.03270567208528519,
-0.04027768597006798,
-0.01141724456101656,
-0.005699948873370886,
-0.04113001003861427,
0.07305340468883514,
0.04976864904165268,
-0.0059839230962097645,
0.022160034626722336,
0.005562877748161554,
0.02678852342069149,
-0.05542570352554321,
0.09113448113203049,
0.02796318754553795,
0.04781254380941391,
0.021779106929898262,
0.036511484533548355,
0.04039828106760979,
0.013779285363852978,
0.004200303927063942,
-0.006031348370015621,
0.08325634896755219,
0.02027004584670067,
0.0599038228392601,
0.0025761949364095926,
0.03999771550297737,
0.022171633318066597,
0.018351659178733826,
0.05422031506896019,
-0.04727645590901375,
-0.0638342797756195,
0.0472814105451107,
-0.07401103526353836,
-0.042218443006277084,
0.017132986336946487,
0.0012886825716122985,
-0.0043056076392531395,
0.03811844810843468,
0.09962889552116394,
0.030756056308746338,
0.05790751054883003,
-0.020191645249724388,
-0.06254222244024277,
0.008037208579480648,
-0.0307689867913723,
0.006668416317552328,
0.013955917209386826,
0.0358937606215477,
-0.04507022351026535,
0.02239609695971012,
0.059923071414232254,
0.020579397678375244,
-0.05971621721982956,
0.019380824640393257,
-0.010071342810988426,
-0.024091314524412155,
-0.06454044580459595,
0.031008919700980186,
0.06896112859249115,
-0.004453070927411318,
-0.032980918884277344,
-0.05092266947031021,
0.08678781241178513,
-0.03234868496656418,
-0.03493010252714157,
-0.14245343208312988,
-1.0731474785684228e-34,
0.13226348161697388,
0.07016512751579285,
0.06698009371757507,
0.0374457947909832,
0.04553065821528435,
-0.02253125235438347,
-0.05196910724043846,
0.005008595995604992,
-0.14089399576187134,
0.012109356001019478,
-0.008295954205095768,
-0.009845373220741749,
-0.025642575696110725,
0.037314753979444504,
-0.019960392266511917,
-0.03004397079348564,
-0.03038518875837326,
0.03035503253340721,
-0.0052911387756466866,
0.013473477214574814,
-0.09579300880432129,
-0.0357583649456501,
0.014454931020736694,
-0.03618847206234932,
-0.04157649725675583,
0.11244858056306839,
0.06607093662023544,
-0.007415022235363722,
0.022245733067393303,
-0.028533542528748512,
-0.09932125359773636,
-0.05856418237090111,
0.044047944247722626,
0.003930030390620232,
0.029950547963380814,
-0.03191789239645004,
-0.0026239841245114803,
-0.11046519130468369,
0.009506546892225742,
-0.11048366129398346,
0.009499120526015759,
0.05625656619668007,
0.06404039263725281,
-0.015727141872048378,
-0.09990153461694717,
0.017020825296640396,
-0.029663415625691414,
0.03140736743807793,
-0.05580660700798035,
0.0020518035162240267,
0.02711578458547592,
0.06876558065414429,
0.024954216554760933,
0.026530787348747253,
-0.06695898622274399,
0.035549767315387726,
0.0634826049208641,
0.023406963795423508,
0.0007062260992825031,
0.016566988080739975,
-0.0394950695335865,
0.05857965722680092,
-0.018535636365413666,
0.05669306963682175,
0.010742457583546638,
0.09722558408975601,
-0.08145657181739807,
0.05582518130540848,
0.05295862630009651,
0.002558122854679823,
-0.05990533530712128,
0.016682881861925125,
-0.05242684483528137,
-0.035149917006492615,
0.026335950940847397,
-0.026817725971341133,
0.0013569958973675966,
-0.05322049558162689,
0.035138972103595734,
0.031261082738637924,
0.004572566132992506,
-0.06776213645935059,
-0.0707373172044754,
-0.027689361944794655,
-0.0770479291677475,
-0.14490871131420135,
0.015222393907606602,
-0.11870904266834259,
-0.03830868378281593,
-0.02472684718668461,
-0.0084374425932765,
-0.1058773547410965,
0.04701690003275871,
0.01532532274723053,
-0.017111288383603096,
1.550830271011695e-34,
0.03473764657974243,
-0.001975821331143379,
-0.02181699313223362,
0.04014880210161209,
-0.08146927505731583,
-0.024504294618964195,
0.005671840626746416,
0.049035053700208664,
0.011947537772357464,
-0.01251040305942297,
0.05561225116252899,
-0.04739575460553169,
0.008162939921021461,
-0.06674692034721375,
-0.0017597312107682228,
0.024155544117093086,
0.026481620967388153,
0.002024024026468396,
0.039207350462675095,
0.0670272707939148,
-0.04271356761455536,
0.1055193617939949,
0.00629525538533926,
0.046957653015851974,
-0.02174452133476734,
0.08094270527362823,
0.008628809824585915,
0.0641956701874733,
0.012823555618524551,
-0.02138339914381504,
-0.04224054515361786,
-0.017092162743210793,
-0.037697598338127136,
-0.0458320751786232,
-0.04497309774160385,
-0.017405584454536438,
0.031839411705732346,
-0.03841608762741089,
0.03200725466012955,
0.008121440187096596,
0.09671132266521454,
-0.018320899456739426,
-0.05017184838652611,
0.13799448311328888,
0.04464399442076683,
0.025284327566623688,
-0.07992880046367645,
0.05906734615564346,
-0.0359712615609169,
-0.028151897713541985,
-0.059993620961904526,
-0.021905746310949326,
-0.08216023445129395,
-0.05524437502026558,
-0.014644007198512554,
-0.0473291277885437,
-0.07299850136041641,
-0.009338192641735077,
0.037020791321992874,
-0.07574925571680069,
-0.07184666395187378,
-0.07215357571840286,
-0.0001294293033424765,
-0.030114993453025818,
-0.01917591691017151,
-0.08738645911216736,
-0.0011069384636357427,
0.04363841190934181,
0.04185417667031288,
-0.06648435443639755,
0.05024930089712143,
0.04116952419281006,
0.037523116916418076,
0.0018201125785708427,
0.07380131632089615,
-0.07198593765497208,
-0.006612557452172041,
-0.0966038629412651,
-0.037729889154434204,
-0.0501055084168911,
-0.010829291306436062,
-0.013337320648133755,
-0.00034968051477335393,
0.048397261649370193,
0.08459247648715973,
0.07162898033857346,
0.024685323238372803,
0.07484467327594757,
-0.01789293810725212,
0.04585820436477661,
-0.05387549847364426,
0.13092702627182007,
0.05263521149754524,
0.12982723116874695,
-0.025892650708556175,
-3.347588162228021e-8,
-0.021822772920131683,
-0.009556171484291553,
-0.025852514430880547,
-0.024807609617710114,
0.02457938902080059,
-0.020166276022791862,
0.030989980325102806,
0.042452920228242874,
-0.020716341212391853,
-0.009586443193256855,
0.04883481562137604,
0.03561629354953766,
-0.09637777507305145,
-0.11267576366662979,
-0.020348677411675453,
0.06579148769378662,
-0.016543427482247353,
0.06797882169485092,
-0.025483140721917152,
0.003403659909963608,
0.015523879788815975,
0.07653651386499405,
0.005729658529162407,
0.02413046918809414,
0.07500597089529037,
-0.1172480434179306,
0.009693876840174198,
0.05375887453556061,
-0.030333809554576874,
0.0017053395276889205,
-0.05531657487154007,
0.019542710855603218,
0.07438429445028305,
0.09401074051856995,
0.05010655149817467,
0.04005551338195801,
0.037666045129299164,
-0.05350136384367943,
-0.0132782943546772,
-0.002147711580619216,
0.013995882123708725,
0.06558582186698914,
-0.13217608630657196,
0.016279248520731926,
-0.07929892092943192,
-0.0014944885624572635,
0.060118407011032104,
0.004178236238658428,
0.061296846717596054,
-0.05276961624622345,
0.006609607022255659,
0.026952555403113365,
-0.04420176520943642,
-0.020751405507326126,
-0.011368388310074806,
0.009232405573129654,
-0.0031876808498054743,
0.03182743862271309,
0.02483159489929676,
0.024006640538573265,
0.032767415046691895,
-0.018948113545775414,
0.01927551068365574,
-0.04104842618107796
] |
MMRole: A Comprehensive Framework for Developing and Evaluating Multimodal Role-Playing Agents | https://openreview.net/forum?id=FGSgsefE0Y | [
"Yanqi Dai",
"Huanran Hu",
"Lei Wang",
"Shengjie Jin",
"Xu Chen",
"Zhiwu Lu"
] | Poster | Recently, Role-Playing Agents (RPAs) have garnered increasing attention for their potential to deliver emotional value and facilitate sociological research.
However, existing studies are primarily confined to the textual modality, unable to simulate humans' multimodal perceptual capabilities.
To bridge this gap, we introduce the concept of Multimodal Role-Playing Agents (MRPAs), and propose a comprehensive framework, MMRole, for their development and evaluation, which comprises a personalized multimodal dataset and a robust evaluation approach.
Specifically, we construct a large-scale, high-quality dataset, MMRole-Data, consisting of 85 characters, 11K images, and 14K single or multi-turn dialogues.
Additionally, we present a robust evaluation approach, MMRole-Eval, encompassing eight metrics across three dimensions, where a reward model is designed to score MRPAs with the constructed ground-truth data for comparison.
Moreover, we develop the first specialized MRPA, MMRole-Agent.
Extensive evaluation results demonstrate the improved performance of MMRole-Agent and highlight the primary challenges in developing MRPAs, emphasizing the need for enhanced multimodal understanding and role-playing consistency.
The data, code, and models are all available at https://github.com/YanqiDai/MMRole. | Multimodal Role-Playing Agents, Large Multimodal Models | We propose MMRole, a comprehensive framework for developing and evaluating Multimodal Role-Playing Agents, comprising a personalized multimodal dataset and a robust evaluation approach. | 2,182 | 2408.04203 | [
-0.004662228282541037,
-0.08536709100008011,
0.03916757553815842,
0.0019538523629307747,
0.03341147303581238,
0.02884654700756073,
-0.02205735631287098,
0.0032121450640261173,
0.060190778225660324,
-0.0481012724339962,
-0.06472086906433105,
-0.08768036961555481,
0.01601358689367771,
0.11917486786842346,
0.04022248461842537,
-0.008624615147709846,
0.033680036664009094,
0.07830235362052917,
-0.06822379678487778,
0.0015318754594773054,
0.021449409425258636,
-0.006387884262949228,
0.02174270898103714,
-0.08767054229974747,
0.0010611884063109756,
-0.008429208770394325,
-0.017928175628185272,
-0.01962367817759514,
0.0016437176382169127,
0.002038520062342286,
0.07096938788890839,
0.033502157777547836,
0.06985490769147873,
0.04390547424554825,
-0.07664758712053299,
0.1120227500796318,
-0.07156199216842651,
0.018765032291412354,
-0.038438763469457626,
-0.05422014743089676,
-0.06709130853414536,
0.011180292814970016,
0.07855141162872314,
-0.00414580013602972,
0.02872335910797119,
-0.03333984687924385,
-0.07385459542274475,
-0.0750589594244957,
0.013140137307345867,
0.039118692278862,
-0.09596717357635498,
-0.04681183770298958,
0.03373175486922264,
0.09775657951831818,
-0.04615335538983345,
0.016941102221608162,
0.012156144715845585,
-0.0019898624159395695,
0.05222632735967636,
-0.01586521975696087,
-0.01897183246910572,
-0.06254388391971588,
-0.04249219596385956,
-0.0241322610527277,
0.005056808702647686,
-0.0500861331820488,
0.011307072825729847,
-0.06331484019756317,
-0.02244948036968708,
0.012656440027058125,
-0.06151732802391052,
-0.062346022576093674,
-0.0019654494244605303,
-0.01775902695953846,
-0.03137882426381111,
0.06344233453273773,
-0.017447778955101967,
-0.07421959191560745,
0.05709972232580185,
-0.045809678733348846,
0.027017628774046898,
-0.004622377455234528,
0.053281158208847046,
-0.00803395640105009,
0.014919664710760117,
0.008494395762681961,
0.006830329541116953,
0.038273170590400696,
-0.06949155032634735,
0.10371335595846176,
-0.11647947877645493,
0.039840880781412125,
0.05753837153315544,
-0.009987961500883102,
0.027591614052653313,
0.09014865756034851,
-0.0785508081316948,
-0.07150459289550781,
-0.035215117037296295,
0.08680526912212372,
0.030549680814146996,
0.054756760597229004,
-0.03369877114892006,
-0.0039383950643241405,
0.007744815666228533,
-0.01755393110215664,
0.06801502406597137,
-0.03538532182574272,
-0.023631231859326363,
-0.070847287774086,
-0.049965545535087585,
0.005936035420745611,
-0.019987007603049278,
-0.02564382366836071,
0.12697350978851318,
-0.023646455258131027,
0.05895140394568443,
0.040883515030145645,
0.04516013339161873,
0.05679547041654587,
0.07917744666337967,
-0.006907992996275425,
0.02242996357381344,
0.007192071992903948,
0.12195664644241333,
0.04899340867996216,
0.019741658121347427,
-1.218173244622783e-33,
0.07502373307943344,
-0.00011478931992314756,
0.11092562973499298,
0.12804749608039856,
0.03894893452525139,
-0.01948949135839939,
-0.022582754492759705,
-0.004992694593966007,
-0.09630490839481354,
-0.03667694702744484,
-0.023708516731858253,
0.034886617213487625,
-0.09199723601341248,
0.12084672600030899,
0.003381754970178008,
-0.0025366644840687513,
-0.07416065782308578,
0.012807343155145645,
0.03458297997713089,
0.035943713039159775,
0.00002515488267818,
0.005436049774289131,
0.07728160172700882,
0.03807597979903221,
0.006094350945204496,
-0.01024576835334301,
0.08628741651773453,
-0.10069405287504196,
0.0086887888610363,
-0.02864891104400158,
-0.15508650243282318,
0.0023086874280124903,
-0.020681284368038177,
-0.028283819556236267,
0.009617375209927559,
-0.030673015862703323,
-0.0950583890080452,
-0.043778195977211,
0.01636897213757038,
0.033635083585977554,
-0.0767519474029541,
-0.048921991139650345,
-0.00885679479688406,
0.024845095351338387,
-0.11867476999759674,
0.01781931146979332,
-0.01870405301451683,
0.024517126381397247,
-0.013220748864114285,
0.01878586784005165,
0.01966419443488121,
0.07861030846834183,
0.039348967373371124,
-0.056772273033857346,
-0.0074363150633871555,
-0.0037383073940873146,
0.04157145693898201,
0.0071237399242818356,
0.008839104324579239,
-0.10289087146520615,
0.08837571740150452,
0.01445040199905634,
0.028471410274505615,
0.0652143582701683,
0.03306586667895317,
0.050086237490177155,
0.02692127227783203,
-0.019737306982278824,
0.07829683274030685,
0.015822622925043106,
-0.042107511311769485,
0.04094908758997917,
0.057252369821071625,
-0.03279764950275421,
-0.050285499542951584,
0.01793304644525051,
-0.00046684793778695166,
-0.030745742842555046,
0.05327408388257027,
0.01351434551179409,
-0.01547548919916153,
-0.04375983029603958,
-0.029040148481726646,
-0.07334467023611069,
-0.014186257496476173,
-0.07552828639745712,
-0.004313731100410223,
-0.1528213620185852,
0.003182772547006607,
-0.017607582733035088,
-0.05099925026297569,
0.022126995027065277,
-0.07040395587682724,
0.045580290257930756,
-0.009741241112351418,
5.649004325124549e-34,
0.0015230877324938774,
0.001857729279436171,
-0.062263477593660355,
0.023749489337205887,
0.007669927552342415,
0.014452551491558552,
-0.013185041956603527,
0.018429014831781387,
0.05332256481051445,
0.05733536183834076,
0.014569534920156002,
-0.03723601996898651,
0.059845730662345886,
-0.0640489012002945,
0.04488729313015938,
-0.046696461737155914,
0.005275719799101353,
0.02899423986673355,
-0.0032404116354882717,
0.012242770753800869,
0.03786618262529373,
0.11840718239545822,
-0.043178945779800415,
0.07249394059181213,
-0.07434342801570892,
0.000664948602207005,
-0.016528552398085594,
-0.021605711430311203,
-0.029106665402650833,
0.022585131227970123,
0.018812458962202072,
-0.006067924667149782,
-0.02190902829170227,
-0.06434106826782227,
-0.01553117111325264,
0.018493136391043663,
0.03546511009335518,
-0.0764172226190567,
-0.02200956642627716,
0.05861510708928108,
0.0980856865644455,
-0.009714151732623577,
-0.12062795460224152,
0.031372226774692535,
-0.04279227927327156,
-0.03988458588719368,
-0.06327196210622787,
-0.022806696593761444,
0.03681586682796478,
0.04698581621050835,
-0.05659523978829384,
0.020258260890841484,
-0.04952296242117882,
-0.026088129729032516,
0.008285338059067726,
-0.12260311096906662,
-0.035336919128894806,
-0.053757842630147934,
0.026515847072005272,
0.008516136556863785,
-0.09330344945192337,
0.016616785898804665,
-0.05358686298131943,
-0.01866384968161583,
0.000930526468437165,
0.007026191800832748,
-0.018695857375860214,
-0.028366409242153168,
-0.04850003123283386,
0.01837148889899254,
0.030598940327763557,
-0.04218128323554993,
0.021364934742450714,
0.021238552406430244,
0.035304151475429535,
0.03851877152919769,
0.033130086958408356,
0.027322378009557724,
-0.02638832852244377,
-0.10317450016736984,
0.01669898070394993,
-0.04807532951235771,
0.05553163215517998,
0.061169859021902084,
0.026007121428847313,
0.12043602019548416,
0.048652149736881256,
0.09368164837360382,
-0.003670714097097516,
0.06808396428823471,
0.010344241745769978,
0.04758969321846962,
-0.005266875494271517,
0.08653382956981659,
0.01889503188431263,
-4.152560961756535e-8,
-0.04786036163568497,
-0.022555053234100342,
-0.01370206382125616,
-0.02803696133196354,
-0.06082058325409889,
-0.03901881352066994,
-0.002900328254327178,
-0.009076287969946861,
0.03832118585705757,
0.046826768666505814,
0.07022428512573242,
-0.04220360890030861,
0.02314460091292858,
-0.03352027386426926,
-0.02359749749302864,
0.05786861851811409,
-0.05036768317222595,
0.04712215065956116,
0.007500717416405678,
0.0010012179845944047,
0.11482000350952148,
0.02181008830666542,
0.011837433092296124,
-0.04926697164773941,
-0.050048828125,
0.04168424382805824,
-0.04281114041805267,
0.00411710049957037,
-0.0416366346180439,
0.018165208399295807,
0.07638846337795258,
0.012347680516541004,
0.026917025446891785,
-0.0606609471142292,
-0.020204909145832062,
0.02788029983639717,
0.020224647596478462,
-0.06216994673013687,
-0.017911644652485847,
0.014157846570014954,
-0.019362835213541985,
0.05699838697910309,
-0.009084927849471569,
-0.0073498934507369995,
-0.01074595283716917,
0.06648290902376175,
-0.06841332465410233,
-0.11069213598966599,
0.035845495760440826,
0.013364343903958797,
-0.04241359978914261,
-0.027276095002889633,
-0.030040781944990158,
0.0801745355129242,
0.0991223007440567,
-0.023194445297122,
0.031001603230834007,
0.11465487629175186,
0.12371180951595306,
0.036590348929166794,
0.015338785015046597,
0.02060214802622795,
-0.07740886509418488,
-0.04270951449871063
] |
REMEDY: Recipe Merging Dynamics in Large Vision-Language Models | https://openreview.net/forum?id=iX7eHHE5Tx | [
"Didi Zhu",
"Yibing Song",
"Tao Shen",
"Ziyu Zhao",
"Jinluan Yang",
"Min Zhang",
"Chao Wu"
] | Poster | Model merging has emerged as a powerful technique for combining task-specific vision models into a unified and multi-functional model. Previous methods represented by task arithmetic, have demonstrated effectiveness and scalability in this domain. When large vision-language models (LVLMs) arise with model size scaling up, this design becomes challenging to fuse different instruction-tuned LVLMs for generalization enhancement. The large scale and multi-modal nature of LVLMs present unique obstacles, including constructing reusable and modular components to accommodate the multi-component architecture of LVLMs and the requirement for dynamic fusion based on multi-modal input tokens. To address these challenges, we propose the \textbf{RE}cipe \textbf{ME}rging \textbf{DY}namics (REMEDY) method, a scalable and flexible paradigm for model merging in LVLMs. We first define reusable modules termed \textit{recipes} including the projector and shallow LLM layers, enhancing visual-language understanding. Then, we introduce a modality-aware allocator dynamically generates weights in a one-shot manner based on input relevance to existing recipes, enabling efficient cross-modal knowledge integration. REMEDY thus offers an adaptive solution for LVLMs to tackle both seen (i.e., multi-task learning) and unseen (i.e., zero-shot generalization) tasks. Experimental results demonstrate that our method consistently improves performance on both seen and unseen tasks, underscoring the effectiveness of REMEDY in diverse multi-modal scenarios. | Multi-Modal Large Language Models, Zero-shot Generalization, Model Merging | null | 2,164 | null | [
-0.008609026670455933,
-0.11639612168073654,
0.016851196065545082,
0.023300930857658386,
0.0794011577963829,
0.00029274504049681127,
0.011906079947948456,
-0.00015252384764607996,
0.06494297832250595,
-0.069811150431633,
-0.010304988361895084,
-0.08480857312679291,
0.06014809384942055,
0.062176406383514404,
0.012023967690765858,
0.027484627440571785,
0.03442482277750969,
0.10561662912368774,
-0.18533481657505035,
-0.02868148684501648,
0.08432270586490631,
-0.010291501879692078,
-0.013342669233679771,
0.011259264312684536,
0.039645981043577194,
0.10735335946083069,
-0.0296479519456625,
-0.08108096569776535,
0.062035225331783295,
-0.04610617086291313,
0.02385392226278782,
0.03880965709686279,
-0.05432246997952461,
0.0689733624458313,
-0.03344874829053879,
0.04808115214109421,
-0.07117115706205368,
0.020015036687254906,
-0.007317068055272102,
-0.08912472426891327,
-0.05584774911403656,
0.023707542568445206,
0.027853334322571754,
-0.018789153546094894,
0.004324460402131081,
0.015563893131911755,
-0.026588495820760727,
-0.07079705595970154,
0.012444994412362576,
-0.0005828675348311663,
-0.07417748123407364,
-0.04060149937868118,
-0.003995920997112989,
-0.005384379532188177,
-0.021654056385159492,
0.04403733089566231,
0.01565392315387726,
-0.035677067935466766,
-0.0008785253739915788,
-0.053290851414203644,
-0.05055826157331467,
-0.006513196509331465,
0.02495674230158329,
0.011499937623739243,
0.010879600420594215,
-0.058135442435741425,
0.02968350239098072,
-0.000772819563280791,
-0.0021137953735888004,
0.017594093456864357,
-0.02232895791530609,
0.024798130616545677,
0.056682102382183075,
-0.019295768812298775,
0.03768742457032204,
0.06899791955947876,
0.041052792221307755,
-0.03900088742375374,
0.08229922503232956,
-0.031491369009017944,
0.03169079124927521,
0.031406693160533905,
0.02646927908062935,
-0.05052688345313072,
0.12154140323400497,
0.02610183320939541,
-0.08390621095895767,
0.004450819455087185,
0.07640082389116287,
-0.028512634336948395,
-0.08314443379640579,
-0.07810548692941666,
0.035484571009874344,
0.013497484847903252,
0.09337300062179565,
-0.003598038339987397,
0.010964847169816494,
-0.1044730469584465,
0.022403214126825333,
0.047616779804229736,
-0.01858341321349144,
0.03782358020544052,
0.041642870754003525,
-0.039411164820194244,
0.034964364022016525,
-0.01799687370657921,
0.012915413826704025,
0.07769381999969482,
0.007358639966696501,
-0.15096794068813324,
0.048186544328927994,
0.01712784916162491,
0.06861334294080734,
0.040845394134521484,
0.026355259120464325,
-0.03377276659011841,
0.04235851764678955,
-0.03895414620637894,
0.05542454496026039,
0.008175313472747803,
-0.004269144032150507,
-0.01971219852566719,
-0.046681419014930725,
-0.008646315895020962,
0.041688550263643265,
-0.01072546187788248,
-0.03816024214029312,
6.513209404749863e-34,
0.03741254657506943,
0.06967054307460785,
0.023300638422369957,
0.08286658674478531,
0.0919322818517685,
-0.008577792905271053,
0.007103831507265568,
0.09823448210954666,
-0.09916166961193085,
0.018028967082500458,
-0.017372267320752144,
0.01849307119846344,
-0.07212791591882706,
0.13852743804454803,
0.019131692126393318,
-0.07260830700397491,
0.00545023288577795,
0.04757940396666527,
-0.07134116441011429,
0.006384686101227999,
-0.005854494869709015,
-0.020702989771962166,
0.012250514701008797,
0.016361448913812637,
0.037522148340940475,
0.027712276205420494,
0.1340140849351883,
-0.03451388701796532,
-0.0794217586517334,
0.007396992761641741,
-0.14013296365737915,
-0.026895003393292427,
-0.00843812432140112,
0.04976736009120941,
0.001691748620942235,
-0.016005486249923706,
-0.012896857224404812,
-0.05486922338604927,
0.028149208053946495,
-0.02096131443977356,
-0.0057652611285448074,
0.030381325632333755,
-0.10072670876979828,
-0.05206102877855301,
-0.08749552071094513,
0.002073735697194934,
0.006010040640830994,
0.1336587518453598,
-0.07012729346752167,
-0.10326970368623734,
0.05276207625865936,
0.03769967705011368,
0.010607481002807617,
-0.09537162631750107,
-0.018545107915997505,
-0.035717837512493134,
-0.019880499690771103,
0.030966270714998245,
0.08257201313972473,
0.015735609456896782,
-0.03999936580657959,
0.061467114835977554,
-0.07112136483192444,
0.12416433542966843,
0.05602981522679329,
0.03174914792180061,
0.019307374954223633,
0.0648002102971077,
0.08722417056560516,
-0.014706387184560299,
-0.044962767511606216,
0.017407191917300224,
-0.007375615648925304,
-0.03733356297016144,
0.044899940490722656,
-0.029350800439715385,
0.0284423790872097,
-0.14859312772750854,
0.029335997998714447,
0.09654313325881958,
-0.005659083835780621,
0.0535273402929306,
0.02229897677898407,
-0.0075110383331775665,
0.005659672897309065,
-0.057186368852853775,
0.024095972999930382,
-0.052666813135147095,
-0.07623255252838135,
-0.07453686743974686,
-0.023098174482584,
-0.08097121864557266,
0.051279935985803604,
-0.07260413467884064,
-0.011246788315474987,
-6.4323592706422905e-34,
-0.004428575746715069,
-0.010126623325049877,
-0.027440251782536507,
0.03332102671265602,
0.036395225673913956,
-0.08989772945642471,
0.032133717089891434,
-0.053663644939661026,
-0.020227747038006783,
-0.10170429199934006,
-0.01956138387322426,
-0.04475773498415947,
-0.007588768843561411,
0.0073593114502727985,
-0.0009023351012729108,
0.0368037149310112,
-0.03301757201552391,
0.013243001885712147,
0.015174801461398602,
0.018842823803424835,
-0.008095205761492252,
0.05625145882368088,
-0.04036898538470268,
-0.004731800407171249,
-0.03296913206577301,
0.058100610971450806,
-0.05967418849468231,
0.017270134761929512,
0.022110557183623314,
0.014194618910551071,
0.03446897119283676,
-0.054799631237983704,
0.04437526687979698,
-0.04380549117922783,
-0.02216739021241665,
0.05622377246618271,
-0.018924083560705185,
-0.06410124897956848,
0.025713922455906868,
0.059937916696071625,
0.022072479128837585,
-0.08683191984891891,
-0.04855381324887276,
0.015586676076054573,
0.03657101094722748,
-0.04157068207859993,
-0.055381208658218384,
-0.030619265511631966,
0.008379672653973103,
-0.037014201283454895,
-0.04507710039615631,
0.009166788309812546,
-0.10078977048397064,
0.02114959992468357,
-0.03753582015633583,
-0.07071113586425781,
0.00021271924197208136,
-0.02402338944375515,
0.07705782353878021,
-0.08024457097053528,
0.0013829728122800589,
-0.03252992033958435,
0.047155704349279404,
-0.013598823919892311,
0.013980797491967678,
0.0012158475583419204,
-0.003426923416554928,
-0.043408092111349106,
0.006561243906617165,
-0.025299372151494026,
0.046369463205337524,
-0.017076006159186363,
0.0233999565243721,
0.06986688822507858,
-0.0036436463706195354,
-0.06286222487688065,
0.07182952761650085,
-0.03695525974035263,
0.06126889958977699,
-0.010924614034593105,
-0.042870279401540756,
-0.08655306696891785,
0.07336422801017761,
0.040361933410167694,
-0.0010953574674203992,
0.05828787386417389,
0.04411502555012703,
0.1153891533613205,
0.03358927369117737,
0.028472639620304108,
-0.052091848105192184,
0.0014253678964450955,
0.054723165929317474,
0.08111433684825897,
0.025009214878082275,
-4.484024884732207e-8,
-0.021841205656528473,
0.07703275233507156,
-0.029609624296426773,
-0.01685451902449131,
-0.0011711132247000933,
-0.04811124876141548,
-0.006473747547715902,
0.029544632881879807,
0.02163228578865528,
-0.00790028739720583,
0.04805337265133858,
-0.04687443748116493,
-0.00525444932281971,
-0.038066741079092026,
0.02122773602604866,
0.06783200800418854,
-0.013904341496527195,
0.05386499688029289,
-0.015021029859781265,
-0.04711230471730232,
0.021588962525129318,
0.016139840707182884,
0.007938402704894543,
0.028692210093140602,
0.01261037029325962,
-0.05978601053357124,
-0.08019400388002396,
0.09448526054620743,
0.12235541641712189,
-0.022467589005827904,
0.01756642572581768,
0.05347015708684921,
-0.0027550416998565197,
-0.017643341794610023,
-0.024674901738762856,
0.01887453906238079,
0.015037893317639828,
0.012806463986635208,
0.037317682057619095,
0.011851197108626366,
0.006976803299039602,
0.04955495521426201,
-0.027433736249804497,
0.051868438720703125,
0.04598560184240341,
0.004508570767939091,
-0.04271174222230911,
-0.13163460791110992,
-0.016889844089746475,
-0.009475749917328358,
-0.02346300147473812,
0.03961421549320221,
-0.037901848554611206,
0.11527051776647568,
0.017372459173202515,
0.023833317682147026,
0.04180454835295677,
-0.0170600563287735,
0.054361842572689056,
0.013067688792943954,
-0.013930227607488632,
0.0012787795858457685,
-0.0832730084657669,
-0.06192600727081299
] |
TIPS: Text-Image Pretraining with Spatial awareness | https://openreview.net/forum?id=DaA0wAcTY7 | [
"Kevis-kokitsi Maninis",
"Kaifeng Chen",
"Soham Ghosh",
"Arjun Karpur",
"Koert Chen",
"Ye Xia",
"Bingyi Cao",
"Daniel Salz",
"Guangxing Han",
"Jan Dlabal",
"Dan Gnanapragasam",
"Mojtaba Seyedhosseini",
"Howard Zhou",
"Andre Araujo"
] | Poster | While image-text representation learning has become very popular in recent years, existing models tend to lack spatial awareness and have limited direct applicability for dense understanding tasks.
For this reason, self-supervised image-only pretraining is still the go-to method for many dense vision applications (e.g. depth estimation, semantic segmentation), despite the lack of explicit supervisory signals.
In this paper, we close this gap between image-text and self-supervised learning, by proposing a novel general-purpose image-text model, which can be effectively used off the shelf for dense and global vision tasks.
Our method, which we refer to as Text-Image Pretraining with Spatial awareness (TIPS), leverages two simple and effective insights.
First, on textual supervision: we reveal that replacing noisy web image captions by synthetically generated textual descriptions boosts dense understanding performance significantly, due to a much richer signal for learning spatially aware representations.
We propose an adapted training method that combines noisy and synthetic captions, resulting in improvements across both dense and global understanding tasks.
Second, on the learning technique: we propose to combine contrastive image-text learning with self-supervised masked image modeling, to encourage spatial coherence, unlocking substantial enhancements for downstream applications.
Building on these two ideas, we scale our model using the transformer architecture, trained on a curated set of public images.
Our experiments are conducted on $8$ tasks involving $16$ datasets in total, demonstrating strong off-the-shelf performance on both dense and global understanding, for several image-only and image-text tasks.
Code and models are released at https://github.com/google-deepmind/tips . | image representations, image-text, vision-language, dense understanding, computer vision | Vision+language learning with spatial awareness for off-the-shelf dense and global understanding tasks. | 2,163 | 2410.16512 | [
0.020735692232847214,
-0.13074292242527008,
0.027828196063637733,
0.05312323197722435,
0.102712482213974,
0.017541300505399704,
0.014283295720815659,
-0.02943314239382744,
0.027872474864125252,
-0.025516588240861893,
-0.038276080042123795,
-0.04420552775263786,
0.04067228361964226,
0.13695552945137024,
-0.017157122492790222,
-0.005782444961369038,
0.13188734650611877,
0.030065974220633507,
-0.08549533784389496,
-0.00660294946283102,
0.08981495350599289,
-0.026868604123592377,
0.10243429243564606,
0.03807578980922699,
0.07950715720653534,
-0.010138086974620819,
0.01600591279566288,
-0.05228276550769806,
-0.003465567249804735,
-0.006900933105498552,
-0.011118362657725811,
-0.0454770028591156,
0.019702674821019173,
0.030358167365193367,
-0.06095091998577118,
0.11244333535432816,
0.008866981603205204,
0.07301172614097595,
0.02443884313106537,
-0.04971479997038841,
0.03847837075591087,
-0.008721426129341125,
0.01793730817735195,
0.07557227462530136,
0.09468431025743484,
0.0057374294847249985,
-0.014538649469614029,
-0.03923415765166283,
0.026626549661159515,
-0.0019441330805420876,
-0.0569373182952404,
-0.05190746858716011,
-0.030851636081933975,
-0.034052494913339615,
-0.005490018054842949,
0.06923748552799225,
0.11118624359369278,
-0.03624897450208664,
0.046373769640922546,
-0.0009882345329970121,
0.03890606760978699,
-0.08555855602025986,
0.06401729583740234,
-0.009264074265956879,
0.017721477895975113,
-0.03336041420698166,
0.0007615606300532818,
0.03816507011651993,
-0.036518644541502,
-0.00811551883816719,
-0.01761491782963276,
0.05704410746693611,
-0.0007595141069032252,
0.002309792675077915,
0.013387705199420452,
0.011580544523894787,
-0.05474191904067993,
-0.08825618773698807,
0.025031065568327904,
-0.04059680551290512,
0.037837039679288864,
0.025522083044052124,
0.0918663963675499,
0.012387814000248909,
0.07227778434753418,
0.01602412760257721,
0.027361955493688583,
0.03880876302719116,
-0.03489112854003906,
-0.000380341662093997,
-0.012255026027560234,
-0.07836561650037766,
0.007742743939161301,
-0.01019708439707756,
-0.011896250769495964,
0.0003369622281752527,
-0.05035796016454697,
-0.10232019424438477,
-0.029699787497520447,
0.008644207380712032,
0.07472339272499084,
0.017712857574224472,
0.0026330091059207916,
-0.009798138402402401,
0.002717244904488325,
0.044538844376802444,
0.04780665785074234,
0.008620782755315304,
0.018732398748397827,
-0.043682388961315155,
0.07800028473138809,
-0.030246039852499962,
-0.07407792657613754,
0.01706404611468315,
-0.05408905819058418,
0.049810443073511124,
0.03418411314487457,
-0.014020142145454884,
0.03838108107447624,
0.037963416427373886,
-0.028696278110146523,
-0.001401454326696694,
-0.02071632444858551,
0.024597182869911194,
0.0013421220937743783,
-0.05849596485495567,
-0.02416273206472397,
-2.0666863930195813e-33,
0.06241784244775772,
-0.0028624737169593573,
0.09031394869089127,
0.019694318994879723,
0.028979256749153137,
-0.04692066088318825,
-0.05177940055727959,
0.04566526040434837,
-0.06734859943389893,
-0.03169950097799301,
0.05402199923992157,
-0.02648788131773472,
-0.021185826510190964,
0.09070959687232971,
0.020212989300489426,
-0.018388396129012108,
-0.06757065653800964,
0.015994805842638016,
0.013811303302645683,
0.02035466395318508,
-0.038086943328380585,
0.01141427829861641,
-0.016644909977912903,
0.006472220178693533,
0.06046723946928978,
-0.013898827135562897,
0.06746184080839157,
-0.06645046174526215,
-0.018845930695533752,
0.007980026304721832,
-0.14100946485996246,
0.02963641844689846,
0.014683559536933899,
0.041197121143341064,
0.01218915730714798,
-0.06321157515048981,
-0.057149820029735565,
-0.009228658862411976,
0.017578938975930214,
-0.038619961589574814,
-0.009806349873542786,
0.017622152343392372,
-0.020523639395833015,
-0.0647856742143631,
-0.06204710900783539,
-0.04770367592573166,
-0.030709844082593918,
0.04471103101968765,
-0.08317665010690689,
0.054057154804468155,
0.04336464777588844,
0.05489518120884895,
-0.07681622356176376,
-0.05018067732453346,
-0.02737884782254696,
-0.04559379070997238,
0.05187920480966568,
0.038931068032979965,
0.049141958355903625,
-0.03261897340416908,
0.05929192155599594,
0.016310138627886772,
-0.030133886262774467,
0.13834530115127563,
0.031830791383981705,
0.0016805502818897367,
0.04682392254471779,
0.03058953955769539,
0.0601603128015995,
-0.0033433614298701286,
-0.05918651446700096,
0.013742218725383282,
-0.007064802572131157,
-0.02554016001522541,
0.048950280994176865,
0.012857766821980476,
0.04435255751013756,
-0.09505843371152878,
0.012968461960554123,
0.033875055611133575,
0.01555321179330349,
-0.031004350632429123,
0.018736625090241432,
-0.08832836151123047,
-0.08897500485181808,
-0.027326587587594986,
0.05568844452500343,
-0.09051691740751266,
0.0421588160097599,
-0.006292706821113825,
-0.06538935005664825,
-0.07546006143093109,
-0.06133195012807846,
0.0029521319083869457,
-0.01227920688688755,
1.3766032508197416e-33,
0.08894672989845276,
0.0054238466545939445,
-0.031734220683574677,
-0.012481007725000381,
-0.034199558198451996,
-0.0293030496686697,
0.04197903349995613,
0.03486388549208641,
0.004821073729544878,
-0.07625170797109604,
-0.0682196319103241,
0.032512445002794266,
-0.015375600196421146,
0.033096738159656525,
0.029233695939183235,
-0.014963491819798946,
0.021624142304062843,
0.006386985071003437,
-0.09365532547235489,
0.046477507799863815,
-0.01129398588091135,
0.03916173055768013,
-0.018642013892531395,
0.055941373109817505,
-0.04056737199425697,
0.06471754610538483,
0.007338755298405886,
0.08607448637485504,
-0.06941700726747513,
0.029741423204541206,
0.003736025420948863,
0.02361144684255123,
0.06057585030794144,
-0.04642370715737343,
-0.11244290322065353,
0.060420453548431396,
0.05664064362645149,
-0.08500943332910538,
-0.042411528527736664,
0.0708511620759964,
0.09272400289773941,
-0.07310022413730621,
-0.012119069695472717,
-0.019507402554154396,
-0.14457440376281738,
-0.07540019601583481,
-0.05839431285858154,
-0.009536409750580788,
0.013987214304506779,
0.11960207670927048,
-0.0599004402756691,
0.04064595326781273,
-0.09915199875831604,
0.03510235622525215,
-0.05883653089404106,
-0.10690437257289886,
-0.06925973296165466,
0.030360974371433258,
0.0439850278198719,
-0.05345366150140762,
-0.037002068012952805,
-0.008735966868698597,
-0.01646627113223076,
-0.005221715662628412,
-0.034096892923116684,
-0.09443667531013489,
-0.09478304535150528,
-0.008866095915436745,
0.06638842076063156,
0.04365578293800354,
0.008684810250997543,
-0.016874020919203758,
-0.0005066656158305705,
0.0366683304309845,
0.014859207905828953,
0.005228639114648104,
0.05640583857893944,
0.029196837916970253,
-0.051227159798145294,
-0.06741268187761307,
0.06356462836265564,
-0.028200391680002213,
0.0027862410061061382,
0.08919649571180344,
0.13699857890605927,
0.08815371990203857,
-0.02125047706067562,
0.020648537203669548,
0.03557463362812996,
0.017762020230293274,
-0.08214086294174194,
-0.024615095928311348,
-0.10972997546195984,
0.00693379994481802,
-0.057673048228025436,
-3.505424928107459e-8,
-0.0986708477139473,
0.002788696438074112,
-0.0041390215046703815,
0.014358115382492542,
-0.009923113510012627,
-0.02915915474295616,
-0.02107245661318302,
0.07036393135786057,
-0.03391262888908386,
-0.058038800954818726,
0.05891451984643936,
-0.0375533252954483,
-0.024632595479488373,
-0.08661430329084396,
0.00008698702731635422,
0.049672551453113556,
0.018863748759031296,
0.019835425540804863,
0.02363283932209015,
0.03917090967297554,
-0.010279467329382896,
0.0404156856238842,
-0.009037329815328121,
-0.06772248446941376,
0.011648876592516899,
-0.009476440958678722,
-0.02981705032289028,
0.10286340117454529,
0.03765208646655083,
-0.04621234908699989,
0.11533673852682114,
-0.02009102702140808,
-0.015830134972929955,
-0.017485668882727623,
0.03253689780831337,
0.04664101451635361,
0.06967494636774063,
-0.044113077223300934,
0.01235862448811531,
-0.07356522977352142,
0.023018626496195793,
0.0589214563369751,
0.008826034143567085,
0.0293301772326231,
0.002807207405567169,
0.03817787766456604,
-0.007683411240577698,
-0.08378470689058304,
0.006046611815690994,
-0.009462178684771061,
-0.07892588526010513,
-0.020394107326865196,
-0.025179708376526833,
0.11659274250268936,
0.03883982449769974,
-0.006936331279575825,
0.014113489538431168,
0.011663367971777916,
0.001433931989595294,
0.1412128210067749,
0.0018211964052170515,
0.0503648966550827,
0.009771536104381084,
-0.011755464598536491
] |
From Promise to Practice: Realizing High-performance Decentralized Training | https://openreview.net/forum?id=lo3nlFHOft | [
"Zesen Wang",
"Jiaojiao Zhang",
"Xuyang Wu",
"Mikael Johansson"
] | Poster | Decentralized training of deep neural networks has attracted significant attention for its theoretically superior scalability compared to synchronous data-parallel methods like All-Reduce. However, realizing this potential in multi-node training is challenging due to the complex design space that involves communication topologies, computation patterns, and optimization algorithms. This paper identifies three key factors that can lead to speedups over All-Reduce training and constructs a runtime model to determine when and how decentralization can shorten the per-iteration runtimes. To support the decentralized training of transformer-based models, we introduce a decentralized Adam algorithm that overlaps communications with computations, prove its convergence, and propose an accumulation technique to mitigate the high variance caused by small local batch sizes. We deploy our solution in clusters with up to 64 GPUs, demonstrating its practical advantages in both runtime and generalization performance under a fixed iteration budget.
The experiment code is open-source at [https://github.com/WangZesen/Decentralized-Training-Exp](https://github.com/WangZesen/Decentralized-Training-Exp), and the extension code is open-source at [https://github.com/WangZesen/Decent-DP](https://github.com/WangZesen/Decent-DP). | distributed training, data parallelism, decentralized optimization | A system-level design for practical decentralized training in multi-GPU clusters | 2,160 | 2410.11998 | [
-0.06437715142965317,
-0.03546535223722458,
0.04952031746506691,
0.007934480905532837,
0.0073951128870248795,
-0.05404740199446678,
-0.060878101736307144,
0.015843922272324562,
-0.011013724841177464,
-0.07678915560245514,
-0.06597251445055008,
0.0324559286236763,
-0.0018046832410618663,
-0.009722975082695484,
-0.06461632251739502,
0.029066115617752075,
0.10949049144983292,
0.10300589352846146,
-0.0924728587269783,
-0.07566780596971512,
0.0006987969391047955,
-0.0014050219906494021,
0.029461130499839783,
-0.037913497537374496,
0.08045962452888489,
0.022861231118440628,
-0.0025113557931035757,
-0.10914801061153412,
0.006874812766909599,
0.01242891326546669,
0.08972780406475067,
-0.04716447740793228,
-0.014438755810260773,
0.04491119831800461,
-0.061267443001270294,
0.07341647893190384,
-0.04044003412127495,
-0.008077713660895824,
0.0029282462783157825,
0.027563445270061493,
0.09101725369691849,
-0.012375625781714916,
0.016068771481513977,
-0.040351491421461105,
0.040863871574401855,
0.013462903909385204,
0.04330849274992943,
-0.025845197960734367,
0.03985095024108887,
-0.059099819511175156,
-0.01937541551887989,
-0.07353444397449493,
-0.018644120544195175,
0.09972409904003143,
-0.020288271829485893,
-0.028960498049855232,
0.04799937456846237,
-0.004854798316955566,
-0.01096311118453741,
0.022951234132051468,
0.03545103222131729,
-0.049226656556129456,
-0.000015961464669089764,
-0.022788163274526596,
-0.016454434022307396,
-0.0020784747321158648,
0.12112241238355637,
0.05105850100517273,
-0.032366614788770676,
-0.056618474423885345,
0.1017007902264595,
0.11118964105844498,
-0.08816810697317123,
0.02356053702533245,
0.01171099953353405,
0.03938500955700874,
0.09251177310943604,
-0.017971303313970566,
0.03570899739861488,
-0.0573832243680954,
0.03697890788316727,
0.062151648104190826,
0.0333474762737751,
-0.014850721694529057,
0.05851477384567261,
-0.02584998868405819,
-0.015637995675206184,
0.018614575266838074,
0.01049670111387968,
-0.014030705206096172,
-0.08758915960788727,
0.04493800923228264,
-0.0585012286901474,
-0.045880820602178574,
0.00670892046764493,
0.008189844898879528,
-0.026444492861628532,
-0.06690923124551773,
-0.099341481924057,
0.11551599949598312,
-0.042982786893844604,
-0.002023711334913969,
0.039317358285188675,
-0.06004652753472328,
-0.004383380524814129,
0.017633536830544472,
0.07094134390354156,
0.056764036417007446,
0.04216821491718292,
-0.03548974543809891,
0.058873631060123444,
0.05413740128278732,
-0.024557696655392647,
0.027472641319036484,
-0.018152613192796707,
0.016577841714024544,
-0.02538667805492878,
0.020715218037366867,
-0.04575050622224808,
0.11641023308038712,
-0.031706057488918304,
0.02192702144384384,
-0.015134859830141068,
0.06672288477420807,
0.008417222648859024,
0.024289559572935104,
-0.11609102040529251,
2.850144745305997e-33,
0.0027877050451934338,
0.03218388557434082,
0.05439266562461853,
-0.03989134356379509,
0.10612929612398148,
-0.024197084829211235,
0.03513481467962265,
0.06880773603916168,
-0.07193402945995331,
0.02589372731745243,
-0.03355937823653221,
-0.0003699150984175503,
-0.028887365013360977,
0.05436674505472183,
-0.02063932456076145,
-0.07229170203208923,
0.03255626559257507,
-0.01564520038664341,
0.07063769549131393,
-0.0370645709335804,
0.09002908319234848,
-0.06396317481994629,
-0.017454342916607857,
0.031616512686014175,
-0.021589426323771477,
-0.043958380818367004,
0.048139289021492004,
-0.027101853862404823,
0.07641470432281494,
-0.024920377880334854,
-0.07575640082359314,
-0.019310299307107925,
0.0052176653407514095,
0.05302197486162186,
0.06785314530134201,
-0.05469053238630295,
-0.010147850029170513,
-0.011711972765624523,
-0.01785714365541935,
-0.021823158487677574,
-0.03878180682659149,
-0.0054857064969837666,
0.006071432493627071,
-0.09118000417947769,
-0.061469368636608124,
-0.10378734767436981,
0.0017351703718304634,
-0.04240754619240761,
0.022314274683594704,
-0.06752506643533707,
0.01744186133146286,
-0.017291149124503136,
-0.0635305866599083,
-0.029009809717535973,
0.0786212682723999,
-0.05607087165117264,
0.06773275136947632,
0.1086956113576889,
0.04950672388076782,
0.07921988517045975,
-0.024169359356164932,
-0.025343436747789383,
-0.058707740157842636,
0.04064030572772026,
-0.04549957811832428,
-0.03157828748226166,
-0.024974491447210312,
0.049565527588129044,
0.028772804886102676,
-0.027753913775086403,
-0.016024399548768997,
0.02313908562064171,
-0.007235921453684568,
0.01584382727742195,
0.1001608669757843,
-0.001876086345873773,
0.037186797708272934,
-0.04928210377693176,
-0.07838068157434464,
0.021463315933942795,
-0.04850459471344948,
0.0574500598013401,
-0.018971053883433342,
0.08177361637353897,
-0.01434938982129097,
-0.0009534482378512621,
-0.020072760060429573,
-0.009014214389026165,
0.04270448535680771,
0.025230834260582924,
-0.06722661107778549,
-0.04000163823366165,
-0.005830094683915377,
-0.043765828013420105,
-0.02073119394481182,
-2.1106438205975884e-33,
-0.03343049809336662,
0.028455089777708054,
-0.04999014362692833,
0.07063543796539307,
0.10382965207099915,
-0.060588330030441284,
-0.06623294949531555,
0.022805536165833473,
-0.04568500071763992,
-0.002742788288742304,
-0.03509359434247017,
-0.047241926193237305,
0.07011951506137848,
-0.05444899573922157,
0.056913409382104874,
-0.07207207381725311,
-0.011233003810048103,
0.01591896265745163,
0.021082688122987747,
0.00281867035664618,
-0.017010919749736786,
0.1113828495144844,
-0.012634317390620708,
-0.02351570501923561,
0.017910482361912727,
-0.000025565974283381365,
-0.10931088775396347,
0.07840070873498917,
0.0580412782728672,
-0.015138780698180199,
-0.04200125113129616,
-0.036212291568517685,
-0.0784204974770546,
-0.04077748581767082,
0.012557090260088444,
0.10844356566667557,
0.0151636628434062,
0.044181954115629196,
-0.0158297810703516,
0.04525582119822502,
0.0662568137049675,
-0.08142756670713425,
-0.03916771337389946,
0.04153774306178093,
-0.02754400484263897,
-0.021557403728365898,
-0.10582639276981354,
-0.025016872212290764,
-0.04841158539056778,
0.04389914125204086,
0.0005289220716804266,
0.0207489226013422,
-0.06682906299829483,
0.0328805185854435,
0.005759923718869686,
-0.054770004004240036,
0.054174553602933884,
0.06349383294582367,
0.08231575787067413,
0.015868602320551872,
-0.01678689941763878,
-0.10676819086074829,
0.024649329483509064,
-0.017518308013677597,
-0.029631340876221657,
0.004815036430954933,
-0.02454950474202633,
0.0649772435426712,
0.037003468722105026,
0.1078488901257515,
-0.016336655244231224,
-0.013947022147476673,
-0.045567676424980164,
0.061006810516119,
-0.08807201683521271,
0.0204373337328434,
0.058549582958221436,
-0.061508189886808395,
-0.03337845206260681,
-0.0048593818210065365,
0.017041528597474098,
0.04089676961302757,
-0.039666447788476944,
-0.05682900920510292,
0.12118759751319885,
0.04269459843635559,
0.08357497304677963,
0.047338277101516724,
0.024369604885578156,
0.030360110104084015,
-0.019074387848377228,
0.04926048219203949,
-0.030650343745946884,
0.033177487552165985,
-0.05608435720205307,
-4.5637392531716614e-8,
-0.055424757301807404,
0.049007002264261246,
0.00233696517534554,
0.025968389585614204,
0.09370655566453934,
-0.05848510190844536,
0.06540095061063766,
0.05931159481406212,
-0.011611320078372955,
-0.003152856370434165,
0.06475961953401566,
-0.07050725817680359,
-0.00400154571980238,
-0.09357874095439911,
0.040242813527584076,
0.021730655804276466,
-0.04027773067355156,
-0.036974236369132996,
0.007200843654572964,
-0.03352881222963333,
-0.01325283944606781,
0.03814320266246796,
-0.008409892208874226,
-0.021354157477617264,
0.008552337996661663,
-0.11143328249454498,
0.05494266003370285,
0.02401995100080967,
-0.010934104211628437,
-0.00934622623026371,
-0.05850513279438019,
0.02613687515258789,
-0.008775619789958,
-0.059185367077589035,
0.01160546112805605,
0.09857379645109177,
-0.020412607118487358,
0.07277504354715347,
0.012241866439580917,
-0.04239538684487343,
0.03298698738217354,
0.061231523752212524,
-0.025663577020168304,
-0.008442936465144157,
-0.019481580704450607,
-0.028731171041727066,
0.006600148510187864,
-0.007333429530262947,
0.06987334787845612,
0.011993898078799248,
0.03536335751414299,
0.012024154886603355,
-0.0007948661805130541,
0.0620831698179245,
0.007768444716930389,
-0.039508432149887085,
0.01082327589392662,
-0.1369057446718216,
0.038003452122211456,
0.13236220180988312,
-0.05771002918481827,
-0.044856444001197815,
-0.10949292778968811,
-0.0015460568247362971
] |
mPLUG-Owl3: Towards Long Image-Sequence Understanding in Multi-Modal Large Language Models | https://openreview.net/forum?id=pr37sbuhVa | [
"Jiabo Ye",
"Haiyang Xu",
"Haowei Liu",
"Anwen Hu",
"Ming Yan",
"Qi Qian",
"Ji Zhang",
"Fei Huang",
"Jingren Zhou"
] | Poster | Multi-modal Large Language Models have demonstrated remarkable capabilities in executing instructions for a variety of single-image tasks. Despite this progress, significant challenges remain in modeling long image sequences. In this work, we introduce the versatile multi-modal large language model, mPLUG-Owl3, which enhances the capability for long image-sequence understanding in scenarios that incorporate retrieved image-text knowledge, multimodal in-context examples, and lengthy videos. Specifically, we propose novel hyper attention blocks to efficiently integrate vision and language into a common language-guided semantic space, thereby facilitating the processing of extended multi-image scenarios. We conduct evaluations on 21 benchmarks that cover single/multi-image, and short/long video understanding. mPLUG-Owl3 achieves competitive performance with the state-of-the-art methods while reducing inference time and memory usage by 87.8\% and 48.5\% in average. Moreover, we propose a Distractor Resistance evaluation to assess the ability of models to maintain focus amidst distractions. mPLUG-Owl3 also demonstrates outstanding performance in distractor resistance on ultra-long visual sequence inputs. We hope that mPLUG-Owl3 can contribute to the development of more efficient and powerful multimodal large language models. | multimodal large language model, long sequence, efficient multimodal understanding | A effective and efficient multimodal large language model for various image-text interleaved scenario with hyper-attention | 2,156 | null | [
0.023401230573654175,
-0.10657263547182083,
0.009259339421987534,
0.014496715739369392,
0.177316814661026,
-0.004425200168043375,
-0.038819119334220886,
-0.046264540404081345,
0.019134510308504105,
-0.06481596827507019,
-0.010425692424178123,
-0.02105170674622059,
-0.0500241182744503,
0.08820561319589615,
0.050393976271152496,
-0.004204833414405584,
-0.019445344805717468,
0.025130636990070343,
-0.1168399378657341,
-0.0378631092607975,
0.10424326360225677,
-0.05835496261715889,
0.10105836391448975,
-0.021920479834079742,
-0.008007794618606567,
0.022909637540578842,
-0.07774404436349869,
-0.06518878042697906,
0.04243142902851105,
-0.029231637716293335,
0.03886578977108002,
0.03506380319595337,
0.002402432495728135,
0.06409773230552673,
-0.03391452133655548,
0.0804014727473259,
-0.06664438545703888,
0.010635046288371086,
-0.025787172839045525,
-0.08595093339681625,
0.0030697258189320564,
-0.00027293103630654514,
0.015714816749095917,
0.009149250574409962,
0.020537495613098145,
-0.01287610363215208,
0.019607318565249443,
-0.09803944081068039,
0.10712452977895737,
-0.009908022359013557,
-0.11627975106239319,
-0.04196517542004585,
-0.024077486246824265,
0.014063629321753979,
0.0194954015314579,
-0.03358400613069534,
-0.007692424580454826,
-0.018513040617108345,
-0.007466238457709551,
-0.039407163858413696,
-0.05570460483431816,
-0.08927090466022491,
0.0035191094502806664,
-0.010260707698762417,
0.007635768037289381,
-0.03334622085094452,
0.0646982491016388,
-0.05141357704997063,
-0.03852434456348419,
-0.027438918128609657,
-0.07531826198101044,
0.047669462859630585,
-0.020365281030535698,
0.019192766398191452,
0.004948367830365896,
0.015299095772206783,
0.02126600779592991,
-0.03386928141117096,
0.03270472213625908,
-0.12836812436580658,
0.07421812415122986,
-0.059642910957336426,
0.059798721224069595,
-0.007845785468816757,
0.12966977059841156,
-0.010346989147365093,
-0.009977624751627445,
0.049921926110982895,
-0.058599136769771576,
-0.007589270826429129,
-0.044811736792325974,
-0.10027684271335602,
0.05710965394973755,
0.024813005700707436,
0.06850704550743103,
0.03306868299841881,
-0.00557449646294117,
-0.13530199229717255,
-0.07579958438873291,
0.06633519381284714,
0.03694910928606987,
0.05878883972764015,
-0.028607305139303207,
-0.04410822317004204,
0.05694017559289932,
-0.028124431148171425,
0.03235999867320061,
0.05678325518965721,
-0.02510523982346058,
-0.023445190861821175,
0.06227969378232956,
0.07269516587257385,
0.052375540137290955,
0.00614177854731679,
0.006220973562449217,
0.009296514093875885,
-0.017760934308171272,
-0.022442733868956566,
0.05094015598297119,
0.022221913561224937,
0.021937843412160873,
-0.006400758866220713,
-0.03614061698317528,
-0.008718454279005527,
0.007842686958611012,
-0.030397625640034676,
-0.024832015857100487,
-1.1113891337421787e-33,
0.02762439474463463,
0.004153406713157892,
0.0500468984246254,
0.07034678757190704,
0.08551129698753357,
-0.0004627596354112029,
-0.013449756428599358,
0.0760325938463211,
-0.09459216892719269,
-0.007201491855084896,
0.005736672319471836,
-0.0305204838514328,
-0.04255443066358566,
0.1050242930650711,
0.030533915385603905,
0.008057661354541779,
-0.061667654663324356,
0.04326397925615311,
-0.0534941740334034,
0.06482957303524017,
-0.00543726934120059,
-0.0041595324873924255,
-0.042483020573854446,
0.04815300181508064,
0.01592468097805977,
0.03385613113641739,
0.02725338563323021,
-0.09736167639493942,
0.01503538154065609,
0.0016278850380331278,
-0.12221400439739227,
-0.009399835020303726,
-0.05054878816008568,
0.07607420533895493,
0.021001996472477913,
-0.07357113808393478,
-0.06668589264154434,
-0.06967144459486008,
0.01278830785304308,
-0.01809830032289028,
-0.01717362552881241,
-0.010699433274567127,
-0.008793143555521965,
0.00952137541025877,
-0.07564815878868103,
-0.06874869763851166,
-0.00988492276519537,
0.10988210886716843,
-0.0813777968287468,
0.0374794565141201,
0.08893077820539474,
0.023232104256749153,
-0.0424991138279438,
-0.07531341165304184,
-0.07071755081415176,
-0.00045726425014436245,
0.09733205288648605,
0.010292728431522846,
0.020190341398119926,
0.038474634289741516,
0.01304471306502819,
0.029932871460914612,
-0.013154749758541584,
0.13138090074062347,
0.026601482182741165,
0.029687047004699707,
-0.04388522356748581,
0.09785674512386322,
0.03311143070459366,
0.03376408666372299,
-0.02413775958120823,
0.04525957629084587,
0.04801440238952637,
-0.03317394480109215,
-0.007777692284435034,
0.017763271927833557,
0.043034426867961884,
-0.12162092328071594,
0.034628726541996,
0.09044626355171204,
0.03916391357779503,
0.04146551713347435,
0.035366736352443695,
-0.04558824002742767,
-0.039237990975379944,
-0.04588606208562851,
0.044309865683317184,
-0.12190524488687515,
0.01230704877525568,
-0.07654298841953278,
-0.027044076472520828,
0.05193657428026199,
-0.011432415805757046,
-0.0518481582403183,
-0.015813583508133888,
1.834574196861174e-33,
0.05921017751097679,
0.017543574795126915,
-0.002827346557751298,
-0.012706100940704346,
0.006652826443314552,
-0.033642206341028214,
0.049771495163440704,
0.0415341779589653,
-0.027255898341536522,
-0.06276091188192368,
0.012432623654603958,
-0.055174726992845535,
-0.005274887662380934,
-0.033869996666908264,
0.014461870305240154,
0.00004622885535354726,
0.02360311709344387,
-0.008269340731203556,
-0.05070224776864052,
0.057385366410017014,
0.03682353347539902,
0.05454466491937637,
-0.007926051504909992,
0.019256317988038063,
-0.12353155761957169,
0.03815613314509392,
-0.06369747966527939,
0.015453794039785862,
-0.03372865170240402,
-0.0424661785364151,
0.0031050669495016336,
-0.03008599579334259,
0.020898645743727684,
-0.04672243446111679,
-0.03239420801401138,
0.03629022464156151,
0.040145356208086014,
-0.021647939458489418,
-0.05502329766750336,
0.12957558035850525,
0.06633901596069336,
-0.043144769966602325,
-0.04077734425663948,
0.01850472390651703,
-0.044942717999219894,
-0.013921033591032028,
-0.06556595116853714,
-0.06217198818922043,
-0.04085511341691017,
0.07678788900375366,
-0.05426603928208351,
-0.019839085638523102,
-0.047998182475566864,
-0.014297614805400372,
-0.027541760355234146,
-0.08819060772657394,
-0.11621955782175064,
0.018081458285450935,
0.05156292766332626,
-0.04914180561900139,
0.002014177618548274,
-0.04054420068860054,
-0.03530983626842499,
0.010591319762170315,
0.05570673942565918,
-0.010054143145680428,
-0.0395776741206646,
-0.09745191782712936,
0.03929901868104935,
0.033354565501213074,
0.03248942270874977,
0.007843938656151295,
-0.007851241156458855,
0.044354863464832306,
0.01742699183523655,
0.06070457398891449,
0.003478182014077902,
-0.020117150619626045,
0.046594101935625076,
-0.0835852175951004,
-0.0022200094535946846,
-0.03255325183272362,
-0.008558188565075397,
0.02706913650035858,
0.032806362956762314,
0.1064797192811966,
0.06695225089788437,
0.10627645999193192,
0.02096564881503582,
0.006357491482049227,
-0.030514443293213844,
0.05848723277449608,
0.05881974473595619,
0.04345919191837311,
-0.006316318642348051,
-3.0175087317729776e-8,
-0.07082967460155487,
0.041500773280858994,
-0.01892971433699131,
-0.008009152486920357,
0.03864729776978493,
-0.05241885781288147,
0.0181400328874588,
0.0691433697938919,
0.0025551195722073317,
0.0024290084838867188,
0.03816210851073265,
-0.022512976080179214,
-0.01105241384357214,
-0.018246209248900414,
-0.06614948809146881,
0.06960424780845642,
0.07126787304878235,
0.039952777326107025,
0.02232375182211399,
0.04146692529320717,
0.03010430745780468,
0.03715495020151138,
0.004665647633373737,
-0.008332024328410625,
-0.02033700793981552,
-0.03281625360250473,
-0.005952267907559872,
0.09472391754388809,
0.09266108274459839,
-0.08060267567634583,
-0.007921249605715275,
0.008894827216863632,
-0.04281086474657059,
-0.08555404841899872,
-0.05703992769122124,
0.044689226895570755,
0.06558450311422348,
-0.017580492421984673,
0.03423116356134415,
-0.02789406105875969,
0.08656847476959229,
0.0027309043798595667,
0.013312631286680698,
0.0469009131193161,
0.041049979627132416,
0.04666515067219734,
-0.02339811436831951,
-0.0860200971364975,
0.02506081759929657,
0.0006002100417390466,
0.022327611222863197,
0.0021776079665869474,
0.04473584145307541,
0.0630282312631607,
0.011830703355371952,
0.00714712031185627,
0.023280741646885872,
0.07654669880867004,
0.07126045227050781,
0.07870617508888245,
0.046183302998542786,
0.02545364759862423,
-0.04010510817170143,
0.03458306938409805
] |
Long-Sequence Recommendation Models Need Decoupled Embeddings | https://openreview.net/forum?id=jkpGIxSsUD | [
"Ningya Feng",
"Junwei Pan",
"Jialong Wu",
"Baixu Chen",
"Ximei Wang",
"QianLi",
"Xian Hu",
"Jie Jiang",
"Mingsheng Long"
] | Poster | Lifelong user behavior sequences are crucial for capturing user interests and predicting user responses in modern recommendation systems. A two-stage paradigm is typically adopted to handle these long sequences: a subset of relevant behaviors is first searched from the original long sequences via an attention mechanism in the first stage and then aggregated with the target item to construct a discriminative representation for prediction in the second stage. In this work, we identify and characterize, for the first time, a neglected deficiency in existing long-sequence recommendation models: a single set of embeddings struggles with learning both attention and representation, leading to interference between these two processes. Initial attempts to address this issue with some common methods (e.g., linear projections---a technique borrowed from language processing) proved ineffective, shedding light on the unique challenges of recommendation models. To overcome this, we propose the Decoupled Attention and Representation Embeddings (DARE) model, where two distinct embedding tables are initialized and learned separately to fully decouple attention and representation. Extensive experiments and analysis demonstrate that DARE provides more accurate searches of correlated behaviors and outperforms baselines with AUC gains up to 9‰ on public datasets and notable improvements on Tencent's advertising platform. Furthermore, decoupling embedding spaces allows us to reduce the attention embedding dimension and accelerate the search procedure by 50\% without significant performance impact, enabling more efficient, high-performance online serving. Code in PyTorch for experiments, including model analysis, is available at https://github.com/thuml/DARE. | Recommender System, User Interest Modeling | Decoupled embeddings help solve the conflict between attention and representation in recommendation models. | 2,154 | 2410.02604 | [
-0.06497804820537567,
-0.14895547926425934,
0.014221522025763988,
0.04270681366324425,
0.07237495481967926,
0.10370699316263199,
0.004310143645852804,
0.0201773252338171,
0.02532309666275978,
-0.0932762399315834,
-0.03639352694153786,
0.014081332832574844,
0.028686223551630974,
0.04809378832578659,
-0.06289400905370712,
0.02995467372238636,
0.1014777347445488,
0.00037614331813529134,
-0.017275815829634666,
-0.025779813528060913,
0.04564529284834862,
-0.009449359029531479,
0.06400127708911896,
-0.03259744867682457,
0.01359604112803936,
0.01648002862930298,
-0.021524442359805107,
-0.04432821273803711,
0.0014691990800201893,
0.020476732403039932,
0.1201513409614563,
0.07305415719747543,
0.03439728543162346,
0.01761467009782791,
-0.09420546889305115,
0.047958917915821075,
-0.10466358810663223,
-0.009045761078596115,
-0.06858407706022263,
-0.0666729211807251,
0.010858194902539253,
0.014752749353647232,
0.014950808137655258,
0.11761651188135147,
0.03776393458247185,
-0.09339313209056854,
0.04385281354188919,
-0.07159696519374847,
0.016788424924016,
0.06070343032479286,
-0.048003047704696655,
0.016382941976189613,
0.02352837100625038,
0.08887362480163574,
-0.04419923946261406,
-0.06360740214586258,
-0.04082894325256348,
0.08122223615646362,
-0.0076072076335549355,
-0.028375569730997086,
-0.05826379358768463,
-0.09554687142372131,
0.002864098409190774,
-0.07364313304424286,
-0.01752793788909912,
-0.04750030115246773,
-0.06417515873908997,
0.04101419821381569,
-0.09340877830982208,
0.11985739320516586,
-0.051563676446676254,
0.06205720081925392,
-0.05096360668540001,
-0.005703548900783062,
0.058991361409425735,
0.05982618033885956,
-0.0013859120663255453,
-0.08387798815965652,
0.06408704817295074,
-0.048263125121593475,
-0.052565913647413254,
-0.026620758697390556,
0.051493726670742035,
-0.08564212173223495,
0.12500408291816711,
0.042116690427064896,
0.06430261582136154,
-0.03677241504192352,
-0.04203251004219055,
-0.0018850343767553568,
-0.012367983348667622,
-0.023109709843993187,
0.04509805515408516,
-0.04045694321393967,
0.02345879189670086,
0.06795798987150192,
0.03558652102947235,
-0.0140226474031806,
-0.03611316159367561,
0.04223193973302841,
-0.010279528796672821,
0.11670909821987152,
-0.010403735563158989,
-0.015926700085401535,
-0.0328708291053772,
-0.035876885056495667,
-0.04065470024943352,
0.09254124015569687,
-0.010124227032065392,
-0.08015822619199753,
-0.05693620815873146,
0.043266043066978455,
-0.0024392674677073956,
0.03338120877742767,
0.035676103085279465,
-0.08531905710697174,
0.02887505665421486,
-0.010117394849658012,
0.06971636414527893,
0.06893091648817062,
0.03438602760434151,
0.0681958869099617,
0.008645742200314999,
0.01246642880141735,
-0.0033044698648154736,
0.011384920217096806,
-0.009023282676935196,
1.2392904496223285e-33,
-0.01218778919428587,
-0.021145116537809372,
0.00526425102725625,
0.06204500421881676,
-0.033139992505311966,
-0.0033346866257488728,
-0.023911252617836,
0.09535463154315948,
-0.04485531523823738,
0.014340572990477085,
0.021109620109200478,
0.026500103995203972,
-0.014493525959551334,
0.15177609026432037,
-0.001890270272269845,
0.012075813487172127,
-0.03179261088371277,
0.015463593415915966,
-0.06713784486055374,
-0.07612904906272888,
-0.014479628764092922,
0.033468250185251236,
0.025829050689935684,
-0.012421448715031147,
0.03557530790567398,
-0.0368003286421299,
0.05550277605652809,
-0.0399872288107872,
-0.05014080926775932,
0.004673296585679054,
-0.12339496612548828,
0.029453599825501442,
-0.04493407532572746,
-0.010919707827270031,
0.002548349555581808,
-0.0705229714512825,
0.0583169087767601,
-0.029516562819480896,
0.06092764064669609,
-0.05581461265683174,
0.017933763563632965,
0.04494328424334526,
0.028075827285647392,
0.037448786199092865,
-0.07802903652191162,
-0.1175791472196579,
-0.004608256742358208,
0.035263385623693466,
-0.009934931993484497,
0.019528262317180634,
0.053253717720508575,
-0.0367329902946949,
-0.032223090529441833,
0.003519278485327959,
-0.05438956990838051,
-0.03359983488917351,
-0.007365372031927109,
0.05525427684187889,
-0.009808027185499668,
-0.028173664584755898,
0.04266946390271187,
0.02883356250822544,
0.02857685089111328,
0.027728518471121788,
-0.013911153189837933,
-0.0012910866644233465,
-0.0012238353956490755,
0.09564842283725739,
0.03888264298439026,
0.05624862760305405,
0.016216104850172997,
0.07135020196437836,
0.036021362990140915,
-0.004672711249440908,
0.01949462480843067,
-0.0025768305640667677,
0.016044126823544502,
-0.12137147039175034,
-0.011256051249802113,
0.014207748696208,
0.004295791499316692,
-0.02862301655113697,
-0.03229484707117081,
-0.03821192681789398,
-0.12364865839481354,
-0.04655929654836655,
0.041567761451005936,
-0.12248505651950836,
-0.015584520995616913,
-0.05837534740567207,
-0.028099482879042625,
-0.007540893740952015,
-0.06260601431131363,
0.020251445472240448,
0.04175072908401489,
-1.4380981357570557e-33,
0.021985668689012527,
-0.015196764841675758,
0.01646040566265583,
-0.03154384717345238,
0.0018536648713052273,
-0.0482921227812767,
-0.05470200255513191,
0.05094161257147789,
-0.03441065549850464,
-0.04515930265188217,
-0.02656634896993637,
-0.07252850383520126,
0.08901344984769821,
0.00710018677636981,
0.07404720783233643,
0.01819264516234398,
-0.0151057755574584,
0.09956101328134537,
-0.05010903626680374,
-0.006113121751695871,
0.015157349407672882,
0.055039871484041214,
-0.08507613092660904,
-0.002895108424127102,
-0.014484833925962448,
0.03440986946225166,
0.03532283008098602,
0.012144096195697784,
-0.023103756830096245,
-0.07497021555900574,
-0.08717504143714905,
0.006978112738579512,
-0.011847476474940777,
-0.06395404040813446,
-0.09312471747398376,
0.04635895416140556,
0.0017448262078687549,
-0.012738923542201519,
-0.09824350476264954,
0.04040251299738884,
0.11549922823905945,
-0.03234387934207916,
-0.05736256390810013,
0.050897013396024704,
-0.03188555687665939,
-0.05183608457446098,
-0.032694730907678604,
0.024217262864112854,
0.05202395096421242,
-0.004005589988082647,
-0.006472542881965637,
0.04240007326006889,
-0.057167839258909225,
0.027988234534859657,
-0.03409340977668762,
-0.03395175561308861,
-0.04939628764986992,
-0.01315502543002367,
0.06770135462284088,
-0.05358963832259178,
0.03294293209910393,
0.018789630383253098,
0.04804631695151329,
-0.06758245080709457,
0.030914753675460815,
-0.07421424239873886,
0.0077968379482626915,
-0.02002563327550888,
0.050439558923244476,
-0.025287920609116554,
0.045746810734272,
0.021804509684443474,
-0.027052395045757294,
0.03499695658683777,
0.0027607677038758993,
-0.0492713637650013,
0.030793409794569016,
-0.027043189853429794,
-0.03145827725529671,
-0.014627090655267239,
-0.0032299463637173176,
0.021455785259604454,
-0.03972085192799568,
-0.060204870998859406,
0.07641903311014175,
0.030668990686535835,
0.08724531531333923,
0.07433373481035233,
-0.032039545476436615,
-0.01636958308517933,
0.013533554971218109,
-0.002768681151792407,
-0.04690920561552048,
0.08897233009338379,
-0.031116807833313942,
-4.141972453908238e-8,
-0.07411783933639526,
0.047151755541563034,
-0.02230452559888363,
0.06979978084564209,
0.014448716305196285,
-0.019224947318434715,
-0.015906784683465958,
0.13507874310016632,
-0.03012763522565365,
-0.02469317987561226,
0.026831544935703278,
0.05279248207807541,
0.0010237216483801603,
-0.06431025266647339,
0.013816822320222855,
0.08073568344116211,
0.07136757671833038,
-0.040010157972574234,
-0.008400741033256054,
0.05118028447031975,
0.013494989834725857,
0.062131088227033615,
0.04172191768884659,
-0.0012803924037143588,
-0.04037994518876076,
0.0397045873105526,
-0.01365580316632986,
0.08101064711809158,
0.10360211879014969,
-0.023294169455766678,
0.01160157099366188,
0.043774887919425964,
0.009679457172751427,
-0.13055846095085144,
-0.06337630748748779,
0.03994191437959671,
0.006811877246946096,
-0.05300004407763481,
-0.0307113379240036,
-0.04195243865251541,
0.05563664063811302,
0.0548490509390831,
-0.0344080813229084,
0.03724103793501854,
0.026859911158680916,
0.024169811978936195,
0.00520606804639101,
-0.07628264278173447,
0.055073533207178116,
-0.016912361606955528,
0.03044368512928486,
-0.0948038399219513,
0.011746267788112164,
0.019730107858777046,
0.023127276450395584,
0.022567851468920708,
0.015892542898654938,
0.027052782475948334,
0.06765739619731903,
-0.002264173235744238,
0.035520680248737335,
-0.02503689005970955,
-0.03556325286626816,
0.09743307530879974
] |
MVTokenFlow: High-quality 4D Content Generation using Multiview Token Flow | https://openreview.net/forum?id=zu7cBTPsDb | [
"Hanzhuo Huang",
"Yuan Liu",
"Ge Zheng",
"Jiepeng Wang",
"Zhiyang Dou",
"Sibei Yang"
] | Poster | In this paper, we present MVTokenFlow for high-quality 4D content creation from monocular videos. Recent advancements in generative models such as video diffusion models and multiview diffusion models enable us to create videos or 3D models. However, extending these generative models for dynamic 4D content creation is still a challenging task that requires the generated content to be consistent spatially and temporally. To address this challenge, MVTokenFlow utilizes the multiview diffusion model to generate multiview images on different timesteps, which attains spatial consistency across different viewpoints and allows us to reconstruct a reasonable coarse 4D field. Then, MVTokenFlow further regenerates all the multiview images using the rendered 2D flows as guidance. The 2D flows effectively associate pixels from different timesteps and improve the temporal consistency by reusing tokens in the regeneration process. Finally, the regenerated images are spatiotemporally consistent and utilized to refine the coarse 4D field to get a high-quality 4D field. Experiments demonstrate the effectiveness of our design and show significantly improved quality than baseline methods. Project page: https://soolab.github.io/MVTokenFlow. | 4D Generation, Dynamic 3D Gaussian Splatting, Dynamic Reconstruction, Diffusion Models | null | 2,147 | 2502.11697 | [
0.014066211879253387,
-0.08496201038360596,
0.10364591330289841,
-0.0692109540104866,
0.007314915303140879,
-0.023996971547603607,
-0.1287803053855896,
-0.11482608318328857,
0.0748843178153038,
-0.04136860743165016,
-0.06088945269584656,
-0.0252111554145813,
-0.019996266812086105,
0.062122784554958344,
-0.05970902740955353,
0.008466063067317009,
0.0031885625794529915,
0.11400315165519714,
-0.12507930397987366,
-0.01288816798478365,
0.013162235729396343,
-0.08739794790744781,
0.00799559522420168,
-0.07244641333818436,
0.030577119439840317,
0.0575258731842041,
0.026250159367918968,
-0.08128068596124649,
0.08248180150985718,
-0.0377301424741745,
0.08168013393878937,
0.055531732738018036,
-0.04668574407696724,
0.032103318721055984,
-0.0551651231944561,
0.03563991189002991,
0.036414504051208496,
0.03303676098585129,
-0.02264353260397911,
0.0016718903789296746,
0.02997705712914467,
0.06079288199543953,
0.020175689831376076,
-0.013377250172197819,
0.01199983712285757,
0.045707136392593384,
0.04429005831480026,
-0.10094619542360306,
0.04320386052131653,
-0.027285220101475716,
-0.0125470831990242,
-0.020293323323130608,
-0.058641768991947174,
0.042731765657663345,
-0.019409164786338806,
0.052279360592365265,
0.025096513330936432,
-0.02874945104122162,
0.03331386297941208,
0.021562688052654266,
0.03854135051369667,
-0.016483088955283165,
0.06906447559595108,
0.009708637371659279,
-0.004159939009696245,
-0.06055855378508568,
-0.007922598160803318,
0.00017432462482247502,
0.008551962673664093,
-0.038845691829919815,
-0.05112041160464287,
0.05521632358431816,
0.00720448512583971,
-0.0917312428355217,
-0.049097247421741486,
0.015661170706152916,
0.06249942258000374,
-0.06151895225048065,
0.008478439413011074,
-0.06633836776018143,
0.10586870461702347,
0.04673147201538086,
0.0606364943087101,
-0.024378616362810135,
-0.0025236140936613083,
0.02292618528008461,
0.0069564166478812695,
0.043724190443754196,
-0.022737083956599236,
0.043595120310783386,
-0.08059891313314438,
0.0631297379732132,
-0.04518066346645355,
-0.012615952640771866,
-0.008962258696556091,
-0.036904964596033096,
0.019689584150910378,
-0.09758245199918747,
0.01837422139942646,
0.03515379875898361,
0.01686478592455387,
-0.03439962491393089,
-0.0015496535925194621,
0.01099222432821989,
0.0980129987001419,
-0.02744949981570244,
0.07546954602003098,
0.06731589138507843,
-0.03225253149867058,
0.07973244041204453,
0.08129173517227173,
0.06281358748674393,
-0.01634260080754757,
0.047114212065935135,
-0.018468398600816727,
0.039585232734680176,
-0.05247599259018898,
-0.03680839762091637,
-0.01952638477087021,
0.039166174829006195,
-0.034790702164173126,
-0.00020067817240487784,
-0.027058476582169533,
-0.015038875862956047,
0.004988221451640129,
-0.037594448775053024,
0.04197412356734276,
2.790539651208622e-33,
0.002559083513915539,
0.060211099684238434,
0.03321375325322151,
0.06682339310646057,
0.06422004848718643,
0.017785685136914253,
0.016438478603959084,
-0.03979557752609253,
-0.019100399687886238,
-0.11205174028873444,
0.0025708340108394623,
-0.09422047436237335,
-0.054410722106695175,
0.12488504499197006,
0.015516003593802452,
-0.09568046033382416,
-0.05144013464450836,
0.01309826597571373,
0.0334198959171772,
0.10332085937261581,
0.0001345537748420611,
-0.018333857879042625,
-0.04740219563245773,
0.027892494574189186,
-0.05279969796538353,
0.0068610189482569695,
0.08291085809469223,
-0.009094813838601112,
0.05408250167965889,
-0.006402616389095783,
-0.10545297712087631,
-0.04234747588634491,
0.011192811653017998,
-0.034860849380493164,
0.018882259726524353,
0.0001521941157989204,
-0.02834797464311123,
-0.01605849526822567,
-0.022417277097702026,
-0.03432140871882439,
-0.06217067316174507,
0.06607101112604141,
-0.048151589930057526,
-0.07718677073717117,
-0.08151943236589432,
0.03672138974070549,
0.0224815271794796,
0.039108216762542725,
-0.05610956996679306,
0.0074369232170283794,
0.09260232746601105,
0.04250693693757057,
0.059699561446905136,
-0.11803256720304489,
-0.05545266717672348,
-0.07069950550794601,
0.0005561415455304086,
-0.073196642100811,
0.06076526641845703,
0.04560505971312523,
0.012969167903065681,
0.054054055362939835,
-0.10217908769845963,
0.08886098861694336,
0.005127988290041685,
-0.036192335188388824,
0.07847824692726135,
0.019780155271291733,
0.029504695907235146,
0.03317178785800934,
-0.016940807923674583,
-0.006769205443561077,
0.03479207307100296,
-0.09379346668720245,
0.0457020103931427,
-0.012941951863467693,
0.058106232434511185,
-0.0004593881021719426,
0.04996354132890701,
0.049536947160959244,
-0.08905371278524399,
0.02365979738533497,
0.027702756226062775,
-0.05794087424874306,
0.02390557900071144,
-0.0449022613465786,
0.02328290231525898,
-0.03686077892780304,
-0.026711994782090187,
-0.06765059381723404,
0.0068952483125030994,
-0.06606115400791168,
0.03633979707956314,
-0.011380278505384922,
0.018910981714725494,
-2.1556375199142423e-33,
0.030306857079267502,
0.057638950645923615,
-0.03176858276128769,
0.04497169330716133,
-0.015632828697562218,
-0.002383254235610366,
-0.044122327119112015,
0.08467121422290802,
0.053062617778778076,
-0.10563545674085617,
-0.01736059971153736,
0.049359481781721115,
-0.08701152354478836,
-0.0024187227245420218,
-0.01276858989149332,
-0.05963558331131935,
0.02810719795525074,
-0.11964976787567139,
-0.08343628793954849,
-0.0033617764711380005,
0.08929003030061722,
0.029255276545882225,
-0.04139764606952667,
0.011808366514742374,
-0.030605658888816833,
0.09290461987257004,
-0.004902460612356663,
0.038338251411914825,
0.009966552257537842,
-0.0014419232029467821,
0.04357044771313667,
-0.05995743349194527,
0.049566179513931274,
-0.01055406965315342,
-0.08064287155866623,
0.004384322557598352,
0.036844607442617416,
0.0026427109260112047,
-0.04893193393945694,
0.04758048802614212,
-0.020510338246822357,
-0.031923286616802216,
-0.029004815965890884,
0.0770600438117981,
-0.044629961252212524,
0.03238995373249054,
-0.004132851026952267,
0.0454174242913723,
0.037241701036691666,
0.0696675181388855,
-0.0801640972495079,
0.0555313341319561,
-0.1153780147433281,
-0.008998136967420578,
0.019235700368881226,
-0.10550381988286972,
-0.07013457268476486,
-0.0678710862994194,
0.009499726817011833,
-0.04469076544046402,
-0.03802153468132019,
-0.11729054152965546,
-0.02584821917116642,
-0.022480793297290802,
-0.00015661802899558097,
0.0032867398113012314,
-0.04487307369709015,
-0.06235634163022041,
-0.03966090455651283,
0.08754885941743851,
-0.0026680564042180777,
-0.0024932497180998325,
-0.04706982523202896,
-0.004656424280256033,
0.09298595786094666,
-0.0504559762775898,
0.04456714913249016,
0.005244275089353323,
0.0854514017701149,
-0.03415116295218468,
-0.05256752297282219,
-0.00613444996997714,
0.050727177411317825,
0.08085618168115616,
0.07385757565498352,
0.0034216928761452436,
0.024407073855400085,
-0.005557468626648188,
0.05648510903120041,
0.05880960822105408,
-0.01584574207663536,
0.05609969049692154,
0.01266778539866209,
0.03229057788848877,
0.007710407488048077,
-3.7119221474313235e-8,
-0.017189063131809235,
0.038686349987983704,
0.029216883704066277,
0.001946319011040032,
-0.11320073902606964,
-0.052640631794929504,
0.02288590744137764,
0.08797822892665863,
0.10651299357414246,
-0.09135314077138901,
0.028380833566188812,
-0.007839557714760303,
-0.08954320102930069,
-0.036646246910095215,
0.02861570380628109,
0.03897421061992645,
-0.016259070485830307,
-0.024996554479002953,
-0.0037906800862401724,
-0.01630134880542755,
-0.08189272880554199,
0.04051794111728668,
0.007279636338353157,
-0.019887829199433327,
-0.02821677178144455,
0.07793731242418289,
0.016949966549873352,
-0.021070880815386772,
0.0814395397901535,
-0.06027550995349884,
0.04594460874795914,
0.03920861333608627,
0.02958747185766697,
-0.03369351848959923,
-0.07610733807086945,
-0.02370147407054901,
0.007482850458472967,
0.03799980878829956,
-0.021451614797115326,
-0.02751004323363304,
0.01461698766797781,
0.0801745131611824,
0.05384035035967827,
-0.001412348821759224,
0.0028619044460356236,
0.021965110674500465,
0.03734847158193588,
-0.0223772544413805,
0.013311225920915604,
0.023539766669273376,
-0.052225202322006226,
0.010030011646449566,
-0.03486280515789986,
0.06852944195270538,
0.06427189707756042,
0.009754431433975697,
0.00808686763048172,
-0.0017424005782231688,
0.07684780657291412,
0.0010589112062007189,
-0.003252534195780754,
0.03151840716600418,
-0.03595416247844696,
0.017146065831184387
] |
DenoiseVAE: Learning Molecule-Adaptive Noise Distributions for Denoising-based 3D Molecular Pre-training | https://openreview.net/forum?id=ym7pr83XQr | [
"Yurou Liu",
"Jiahao Chen",
"Rui Jiao",
"Jiangmeng Li",
"Wenbing Huang",
"Bing Su"
] | Poster | Denoising learning of 3D molecules learns molecular representations by imposing noises into the equilibrium conformation and predicting the added noises to recover the equilibrium conformation, which essentially captures the information of molecular force fields. Due to the specificity of Potential Energy Surfaces, the probabilities of physically reasonable noises for each atom in different molecules are different. However, existing methods apply the shared heuristic hand-crafted noise sampling strategy to all molecules, resulting in inaccurate force field learning. In this paper, we propose a novel 3D molecular pre-training method, namely DenoiseVAE, which employs a Noise Generator to acquire atom-specific noise distributions for different molecules. It utilizes the stochastic reparameterization technique to sample noisy conformations from the generated distributions, which are inputted into a Denoising Module for denoising. The Noise Generator and the Denoising Module are jointly learned in a manner conforming with the paradigm of Variational Auto Encoder. Consequently, the sampled noisy conformations can be more diverse, adaptive, and informative, and thus DenoiseVAE can learn representations that better reveal the molecular force fields. Extensive experiments show that DenoiseVAE outperforms the current state-of-the-art methods on various molecular property prediction tasks, demonstrating the effectiveness of it. | 3D Molecular pre-training via denoising, Molecular property prediction | We propose a novel denoising method for 3D molecular pre-training. | 2,138 | null | [
-0.03881777450442314,
-0.12886978685855865,
-0.006417304277420044,
0.01880159042775631,
-0.03154749423265457,
-0.03264869004487991,
-0.004515413660556078,
-0.11728689819574356,
0.0006192382425069809,
-0.02337120845913887,
-0.03857114538550377,
-0.03015349991619587,
-0.002844511764124036,
0.027037363499403,
-0.007212139666080475,
0.014560362324118614,
0.025405731052160263,
0.11195771396160126,
-0.08020805567502975,
-0.011312445625662804,
-0.0025436722207814455,
0.05788204446434975,
0.008335908874869347,
0.034383565187454224,
0.0791473239660263,
0.03135191649198532,
0.07516995072364807,
0.0350782573223114,
0.01057228073477745,
-0.010276129469275475,
0.10389328747987747,
-0.026349928230047226,
-0.036467112600803375,
-0.031174451112747192,
0.017322737723588943,
-0.03245359659194946,
-0.09985192120075226,
0.08164355158805847,
0.03517376258969307,
-0.00653173727914691,
-0.0067647481337189674,
0.10583982616662979,
-0.0427798256278038,
0.01080786157399416,
0.041414592415094376,
-0.057402193546295166,
0.014881611801683903,
-0.03725294768810272,
0.146066814661026,
0.0096275694668293,
0.07227692753076553,
-0.060911137610673904,
0.0029218746349215508,
-0.018953837454319,
0.059954579919576645,
0.04951731115579605,
0.022501632571220398,
-0.017941264435648918,
-0.000761949282605201,
-0.06243225932121277,
0.09634769707918167,
-0.0771014615893364,
0.000033940286812139675,
0.02025516703724861,
0.013191228732466698,
-0.0010146741988137364,
0.1209697425365448,
0.051008887588977814,
0.028561171144247055,
-0.007595638744533062,
-0.11094869673252106,
0.07104918360710144,
-0.03877723217010498,
0.0355922132730484,
0.033355697989463806,
0.04137859493494034,
-0.042369116097688675,
0.002504285890609026,
0.06532196700572968,
-0.044971030205488205,
0.05868574604392052,
0.028024831786751747,
0.09659935534000397,
0.041593458503484726,
0.052710212767124176,
0.007754459045827389,
0.03310811147093773,
-0.00725192204117775,
0.027679378166794777,
-0.03175138682126999,
-0.040355581790208817,
-0.006659177131950855,
-0.11151997745037079,
-0.04749448224902153,
-0.12293988466262817,
-0.005598724819719791,
0.08147022128105164,
0.03530570864677429,
0.05624202638864517,
0.07489614188671112,
0.05084605515003204,
-0.005341177340596914,
-0.04483184590935707,
0.04052085056900978,
0.026567570865154266,
-0.027725765481591225,
0.09213478863239288,
0.05817323550581932,
-0.014367712661623955,
0.051412150263786316,
-0.006772646214812994,
0.06332729011774063,
-0.0044110007584095,
0.023141875863075256,
0.04312864691019058,
0.03664238005876541,
0.06163400784134865,
0.011132122948765755,
-0.009829959832131863,
0.05355000123381615,
-0.05608340725302696,
-0.04266605153679848,
-0.05687696859240532,
0.005206098780035973,
0.08641330897808075,
-0.0058747632429003716,
-0.10177720338106155,
-9.770906490351593e-35,
0.058003392070531845,
0.02627420239150524,
0.09219859540462494,
0.01151325087994337,
0.08924675732851028,
-0.06934051215648651,
-0.017296522855758667,
-0.03644262254238129,
-0.060454197227954865,
0.03572722524404526,
-0.014189161360263824,
-0.0021993531845510006,
0.001734250457957387,
-0.04176295921206474,
-0.020175619050860405,
-0.025091474875807762,
-0.08176089078187943,
0.011471913196146488,
-0.010287179611623287,
0.006857073400169611,
-0.03285834193229675,
-0.00015161898045334965,
0.015988385304808617,
-0.0054228645749390125,
-0.04731064289808273,
0.026801466941833496,
0.009300644509494305,
0.07171517610549927,
0.02930746227502823,
-0.0054549649357795715,
-0.0773230493068695,
-0.01810133084654808,
-0.02378244139254093,
0.0017947641899809241,
0.004252785816788673,
-0.01723683625459671,
0.028827058151364326,
0.000792507897131145,
-0.03175780177116394,
-0.09604688733816147,
0.08538961410522461,
-0.0014119312399998307,
0.03647718206048012,
-0.06687141209840775,
-0.04396999627351761,
0.052991848438978195,
0.0388505794107914,
-0.037745699286460876,
-0.03421724587678909,
-0.026842130348086357,
0.013818916864693165,
-0.009382344782352448,
0.016151471063494682,
-0.03519279882311821,
-0.008881456218659878,
0.014622610062360764,
0.06510647386312485,
-0.021690716966986656,
0.05228400230407715,
-0.03324131667613983,
-0.07487993687391281,
-0.001518703531473875,
0.0012010153150185943,
0.04153066501021385,
0.07133518904447556,
0.09316597878932953,
-0.08428531140089035,
0.046531207859516144,
0.03471421077847481,
-0.026171285659074783,
0.015352212823927402,
0.0486229807138443,
-0.04641970247030258,
-0.013951589353382587,
0.033496007323265076,
0.07234156131744385,
0.007295205723494291,
-0.012623372487723827,
0.05027950927615166,
-0.008804108947515488,
-0.05575491487979889,
-0.04175069183111191,
-0.06158525124192238,
-0.04396452754735947,
-0.1374901980161667,
-0.07554660737514496,
0.0010471262503415346,
-0.09347233921289444,
-0.017774900421500206,
-0.0664619728922844,
-0.06476414203643799,
-0.10592815279960632,
0.029028134420514107,
-0.08107825368642807,
-0.04787002131342888,
-1.6385785136188045e-33,
-0.009235592558979988,
0.07684694230556488,
0.02407785691320896,
0.0394846610724926,
0.01674879901111126,
0.040163327008485794,
-0.06697539985179901,
0.00949977245181799,
0.04649658128619194,
-0.01330682821571827,
-0.02737024985253811,
-0.003952442202717066,
-0.0334828682243824,
-0.028332268819212914,
-0.018839431926608086,
0.029708869755268097,
-0.018896711990237236,
-0.04463210329413414,
0.04210280254483223,
-0.015804631635546684,
-0.007046543527394533,
0.05700306594371796,
0.05030513554811478,
0.04984124377369881,
-0.032929927110672,
0.04986751079559326,
0.017058957368135452,
0.04789537191390991,
0.05211804062128067,
0.03504030033946037,
-0.04775126650929451,
-0.016127949580550194,
-0.009740534238517284,
0.02574528008699417,
-0.10099250078201294,
0.058081936091184616,
-0.00008609317592345178,
0.026237554848194122,
0.04738841578364372,
0.02942058816552162,
-0.0076097226701676846,
-0.003315477631986141,
-0.07932057231664658,
0.014638622291386127,
0.05460616946220398,
-0.008792941458523273,
0.005315725691616535,
-0.0206107497215271,
0.013000324368476868,
-0.003231530310586095,
-0.01481456309556961,
0.052328016608953476,
-0.021697349846363068,
0.041233111172914505,
-0.05097774416208267,
-0.0028802070301026106,
-0.05135992914438248,
-0.0006394447409547865,
0.18075312674045563,
0.048462800681591034,
-0.09336919337511063,
-0.029873449355363846,
-0.02804076485335827,
-0.012824437581002712,
-0.14331327378749847,
-0.03934033215045929,
-0.06691054254770279,
0.0867500752210617,
0.05632612854242325,
-0.0010604648850858212,
-0.049515847116708755,
0.00961986929178238,
0.06856133788824081,
0.05086397007107735,
0.035016801208257675,
-0.094377301633358,
0.03263944014906883,
-0.1283200979232788,
-0.05094245448708534,
-0.033076211810112,
0.039209045469760895,
-0.01700921356678009,
-0.017996665090322495,
0.06680664420127869,
0.06393513828516006,
0.015042660757899284,
0.01387479342520237,
-0.02561563439667225,
0.0018233155133202672,
-0.056943509727716446,
0.03704759478569031,
0.044423677027225494,
0.038349319249391556,
0.03195460885763168,
-0.031548257917165756,
-3.528359115989588e-8,
0.07288829982280731,
-0.0004053626907989383,
0.07857244461774826,
-0.10089288651943207,
0.09148775786161423,
0.008419533260166645,
-0.007615010719746351,
0.013400474563241005,
0.006092280149459839,
-0.1107000857591629,
0.08836346119642258,
-0.05711402744054794,
-0.07159458845853806,
-0.03593309223651886,
0.013958947733044624,
0.0913451761007309,
-0.027103513479232788,
0.0018472466617822647,
-0.06099383905529976,
-0.026680294424295425,
-0.034687649458646774,
0.04010958597064018,
-0.057886626571416855,
-0.05763223022222519,
0.10607945173978806,
-0.05023336783051491,
-0.05487264692783356,
-0.030391385778784752,
-0.015671705827116966,
0.008175462484359741,
-0.03575176000595093,
0.04531639814376831,
0.0841517448425293,
0.050032202154397964,
-0.02831844426691532,
0.02530050091445446,
0.011588401161134243,
-0.03408672660589218,
-0.0975257158279419,
-0.00792991928756237,
-0.12055642157793045,
0.04050706699490547,
-0.06070162355899811,
-0.024548614397644997,
-0.043547116219997406,
0.051018133759498596,
-0.011540388688445091,
-0.0031094299629330635,
0.08463984727859497,
-0.011628705076873302,
0.03165071830153465,
0.006073205266147852,
-0.011382718570530415,
-0.053517214953899384,
-0.0003687014395836741,
0.030276557430624962,
-0.10524235665798187,
0.006591344252228737,
0.005351844243705273,
-0.02286948636174202,
-0.015068374574184418,
-0.015975458547472954,
-0.012023736722767353,
0.0029123618733137846
] |
Halton Scheduler for Masked Generative Image Transformer | https://openreview.net/forum?id=RDVrlWAb7K | [
"Victor Besnier",
"Mickael Chen",
"David Hurych",
"Eduardo Valle",
"Matthieu Cord"
] | Poster | Masked Generative Image Transformers (MaskGIT) have emerged as a scalable
and efficient image generation framework, able to deliver high-quality visuals with
low inference costs. However, MaskGIT’s token unmasking scheduler, an essential
component of the framework, has not received the attention it deserves. We analyze
the sampling objective in MaskGIT, based on the mutual information between
tokens, and elucidate its shortcomings. We then propose a new sampling strategy
based on our Halton scheduler instead of the original Confidence scheduler. More
precisely, our method selects the token’s position according to a quasi-random,
low-discrepancy Halton sequence. Intuitively, that method spreads the tokens
spatially, progressively covering the image uniformly at each step. Our analysis
shows that it allows reducing non-recoverable sampling errors, leading to simpler
hyper-parameters tuning and better quality images. Our scheduler does not require
retraining or noise injection and may serve as a simple drop-in replacement for
the original sampling strategy. Evaluation of both class-to-image synthesis on
ImageNet and text-to-image generation on the COCO dataset demonstrates that the
Halton scheduler outperforms the Confidence scheduler quantitatively by reducing
the FID and qualitatively by generating more diverse and more detailed images.
Our code is at https://github.com/valeoai/Halton-MaskGIT. | Image Synthesis, MaskGIT, Halton Sequence | We propose a new scheduler for MaskGIT leveraging the Halton sequence, ensuring uniform token coverage and enhancing quality and diversity without noise injection. | 2,130 | 2503.17076 | [
-0.10853862017393112,
-0.02223874256014824,
0.050918515771627426,
0.06471151858568192,
0.05074520409107208,
-0.0848267525434494,
-0.0371062308549881,
-0.06759540736675262,
0.037745799869298935,
-0.037799570709466934,
-0.024331606924533844,
-0.05404776334762573,
0.043678656220436096,
-0.05099880322813988,
-0.00998102780431509,
-0.04381364583969116,
-0.009089629165828228,
0.10464116185903549,
-0.03207745775580406,
-0.09317715466022491,
0.08553916215896606,
0.02009550668299198,
0.07914087176322937,
-0.03444189205765724,
0.02970872074365616,
-0.019306369125843048,
-0.017124244943261147,
-0.09254967421293259,
0.02280595153570175,
-0.054189108312129974,
0.01831386238336563,
-0.011735985986888409,
0.017618803307414055,
0.0017016461351886392,
0.03207390755414963,
0.08498702198266983,
-0.026444805786013603,
0.03788938373327255,
0.041058510541915894,
-0.02236303500831127,
0.003851694753393531,
0.042785923928022385,
0.01457955501973629,
-0.02948668599128723,
0.0021275999024510384,
-0.006483137141913176,
0.03965515270829201,
-0.018959028646349907,
-0.009282228536903858,
0.008992304094135761,
-0.020645176991820335,
-0.030001316219568253,
-0.06762518733739853,
0.056923992931842804,
-0.011529655195772648,
0.06545034050941467,
0.05076967552304268,
-0.04730250686407089,
0.04473760724067688,
0.02436681278049946,
-0.11180733889341354,
-0.016408979892730713,
0.021842284128069878,
-0.005974208936095238,
-0.01907884143292904,
-0.017443601042032242,
0.05942990630865097,
-0.003242786042392254,
0.09544792771339417,
-0.07338147610425949,
-0.07202551513910294,
0.045158516615629196,
-0.044495753943920135,
0.10263560712337494,
-0.0404287651181221,
0.1034904271364212,
0.012435734272003174,
-0.05252116173505783,
0.04209134727716446,
-0.09627646952867508,
0.07079903036355972,
-0.05588580667972565,
0.11054527014493942,
0.056720562279224396,
0.04526224359869957,
0.0020566973835229874,
-0.0035581677220761776,
0.13288067281246185,
-0.007205150555819273,
0.035432204604148865,
-0.11051109433174133,
-0.003785697976127267,
-0.0054004015401005745,
0.025891387835144997,
-0.02415246143937111,
0.00012203461665194482,
0.02015395648777485,
-0.0170299019664526,
-0.031600166112184525,
0.0949467271566391,
0.00339724263176322,
-0.016460364684462547,
0.0356883779168129,
-0.05126270279288292,
0.02174658700823784,
-0.05245441943407059,
0.06238280236721039,
-0.009220163337886333,
-0.010807894170284271,
-0.0008013544138520956,
0.04171905666589737,
0.07333813607692719,
-0.024652311578392982,
-0.029284708201885223,
0.03914351016283035,
0.035923223942518234,
-0.05309327691793442,
0.01057881023734808,
-0.03214652091264725,
0.02579602412879467,
0.04801703616976738,
0.004791950806975365,
-0.05235815420746803,
-0.009430850856006145,
-0.06656694412231445,
-0.07500310242176056,
0.04569537192583084,
2.4306925450891033e-33,
-0.05284746363759041,
-0.002987787825986743,
0.05052347853779793,
-0.0038103791885077953,
0.08362110704183578,
-0.0008418564102612436,
0.024388812482357025,
0.028554851189255714,
-0.04142182692885399,
-0.0063145095482468605,
0.018213830888271332,
-0.11966903507709503,
-0.05202728509902954,
0.14430633187294006,
0.042567163705825806,
-0.03741583228111267,
0.008064714260399342,
0.04661993682384491,
-0.03188659995794296,
-0.037322644144296646,
0.009720591828227043,
-0.02804696187376976,
-0.04019111767411232,
-0.03700532764196396,
0.0008369914721697569,
0.010920430533587933,
-0.007525253109633923,
-0.024163099005818367,
0.0667228028178215,
-0.0139229167252779,
-0.11118461191654205,
0.041396044194698334,
0.076978400349617,
0.04241018742322922,
0.04159899801015854,
-0.04818863049149513,
0.05675581097602844,
-0.036002710461616516,
-0.0601610392332077,
0.06948971748352051,
-0.0011119338450953364,
0.061589356511831284,
0.0066811093129217625,
-0.017703156918287277,
-0.026375919580459595,
-0.06350056082010269,
0.012641802430152893,
0.008653069846332073,
0.016957083716988564,
0.05061289668083191,
0.08622703701257706,
0.023514846339821815,
-0.030150936916470528,
-0.05381976440548897,
0.027284080162644386,
-0.051922090351581573,
0.08114775270223618,
-0.03542705252766609,
0.03230854496359825,
0.07116703689098358,
0.017702139914035797,
0.05090077593922615,
-0.06229522451758385,
0.05756954848766327,
-0.053337305784225464,
0.02947843074798584,
0.01745360903441906,
-0.05178971216082573,
-0.05523952841758728,
0.010439612902700901,
-0.003273633075878024,
0.02138037607073784,
-0.08274092525243759,
-0.04827912524342537,
-0.010241538286209106,
0.058574844151735306,
0.035763420164585114,
0.055627137422561646,
-0.07225947082042694,
0.05286400020122528,
0.009939935058355331,
0.029828142374753952,
-0.06160074472427368,
-0.02488320879638195,
-0.053474899381399155,
0.007773288059979677,
0.016341112554073334,
0.031138915568590164,
-0.020108681172132492,
-0.0024352699983865023,
0.00929178111255169,
-0.0010253931395709515,
-0.02879774011671543,
0.023995863273739815,
0.002022245666012168,
-4.417707971771781e-34,
0.04942625015974045,
0.08011531084775925,
0.025987423956394196,
0.08971092104911804,
-0.04139656946063042,
-0.043791066855192184,
0.010235286317765713,
0.008732331916689873,
0.05675555020570755,
-0.08517871052026749,
-0.10247419029474258,
0.019014529883861542,
-0.03716045990586281,
-0.05926165357232094,
-0.007156751584261656,
-0.05987939611077309,
0.02512536570429802,
0.006958735641092062,
-0.12460976094007492,
0.047023069113492966,
0.03863120451569557,
0.11609391868114471,
-0.04483562335371971,
-0.038588136434555054,
-0.06713403761386871,
0.1075064018368721,
-0.05485718697309494,
0.09472481161355972,
0.07766573131084442,
-0.015508166514337063,
-0.03734623268246651,
-0.05114239081740379,
0.06314291059970856,
-0.02131565473973751,
-0.006856569554656744,
0.013813481666147709,
0.014944769442081451,
0.09047450125217438,
-0.03939139470458031,
0.06425445526838303,
0.04321353882551193,
-0.002164627192541957,
-0.09373597055673599,
0.06225964426994324,
-0.03783750534057617,
-0.011311190202832222,
-0.017306657508015633,
-0.023087847977876663,
0.0481034480035305,
0.06138516217470169,
-0.05544541776180267,
0.03900882601737976,
-0.08632108569145203,
0.03680767863988876,
0.02610846236348152,
-0.09265438467264175,
-0.11310075223445892,
-0.018583375960588455,
0.09124287962913513,
-0.023914415389299393,
-0.010827901773154736,
-0.04237741231918335,
-0.016102878376841545,
-0.12769033014774323,
0.03503304347395897,
-0.01802719011902809,
-0.05208383500576019,
-0.03189248964190483,
0.021548127755522728,
0.06255541741847992,
0.008500437252223492,
-0.0469893179833889,
0.05710480362176895,
0.04870151728391647,
-0.017946356907486916,
-0.0031458279117941856,
0.037095341831445694,
-0.025456659495830536,
-0.005400659516453743,
-0.029785308986902237,
0.018865568563342094,
0.017346380278468132,
-0.04058396816253662,
0.08136506378650665,
0.05959916487336159,
0.09012755751609802,
-0.0078285438939929,
-0.005542574916034937,
0.061027221381664276,
0.0319211520254612,
-0.04494994506239891,
0.026830442249774933,
0.06355292350053787,
0.1015937551856041,
0.04253099113702774,
-4.63320510846188e-8,
0.017632491886615753,
0.013409455306828022,
-0.03711192309856415,
-0.00370756140910089,
0.003203572938218713,
-0.06099613755941391,
0.037873320281505585,
-0.028619321063160896,
0.01557126734405756,
-0.07753299921751022,
0.1496264487504959,
-0.04272262379527092,
-0.01733952946960926,
-0.0743047297000885,
0.001958363689482212,
0.012946924194693565,
0.03975507989525795,
0.07707998901605606,
0.013177028857171535,
-0.0573170930147171,
-0.10315141081809998,
0.02384158968925476,
-0.03368528187274933,
-0.09527979791164398,
-0.019705919548869133,
-0.027521995827555656,
0.0380406528711319,
0.04935033991932869,
0.059409935027360916,
-0.02522682212293148,
-0.00931848119944334,
-0.004630210809409618,
0.037750184535980225,
-0.009674077853560448,
-0.013784444890916348,
0.04020189866423607,
-0.06013229489326477,
-0.007597730495035648,
0.03178748860955238,
-0.09809809923171997,
0.052904676645994186,
0.04542061313986778,
-0.031097691506147385,
-0.005366332828998566,
-0.010614877566695213,
-0.06539519876241684,
0.08490452170372009,
-0.05881279706954956,
-0.0251900777220726,
0.021292131394147873,
-0.04558584839105606,
-0.060840804129838943,
-0.0018108295043930411,
0.04979187995195389,
0.05664440989494324,
0.004074336960911751,
0.04197514429688454,
-0.022340336814522743,
0.021965859457850456,
0.14353133738040924,
0.03424231335520744,
-0.02596123330295086,
-0.061404988169670105,
-0.034727126359939575
] |
GS-CPR: Efficient Camera Pose Refinement via 3D Gaussian Splatting | https://openreview.net/forum?id=mP7uV59iJM | [
"Changkun Liu",
"Shuai Chen",
"Yash Sanjay Bhalgat",
"Siyan HU",
"Ming Cheng",
"Zirui Wang",
"Victor Adrian Prisacariu",
"Tristan Braud"
] | Poster | We leverage 3D Gaussian Splatting (3DGS) as a scene representation and propose a novel test-time camera pose refinement (CPR) framework, GS-CPR. This framework enhances the localization accuracy of state-of-the-art absolute pose regression and scene coordinate regression methods. The 3DGS model renders high-quality synthetic images and depth maps to facilitate the establishment of 2D-3D correspondences. GS-CPR obviates the need for training feature extractors or descriptors by operating directly on RGB images, utilizing the 3D foundation model, MASt3R, for precise 2D matching. To improve the robustness of our model in challenging outdoor environments, we incorporate an exposure-adaptive module within the 3DGS framework. Consequently, GS-CPR enables efficient one-shot pose refinement given a single RGB query and a coarse initial pose estimation. Our proposed approach surpasses leading NeRF-based optimization methods in both accuracy and runtime across indoor and outdoor visual localization benchmarks, achieving new state-of-the-art accuracy on two indoor datasets. | Visual Localization, Camera Pose Estimation, 3D Gaussian Splatting | we present GS-CPR, a novel test-time camera pose refinement framework leveraging 3DGS for scene representation. | 2,120 | null | [
-0.06192164123058319,
-0.08533408492803574,
0.052005402743816376,
0.008005489595234394,
0.1301039308309555,
-0.03361060097813606,
-0.014375445432960987,
-0.0809665098786354,
-0.054056815803050995,
-0.03201267123222351,
0.006730301305651665,
-0.0730263739824295,
0.022340379655361176,
0.07221721857786179,
0.02645215578377247,
-0.02140587568283081,
0.1245579719543457,
0.08840522170066833,
-0.017529187723994255,
0.020945481956005096,
-0.03533980995416641,
0.015955781564116478,
-0.013498583808541298,
0.00057509500766173,
0.07443244755268097,
-0.075297512114048,
0.050690654665231705,
-0.09201610833406448,
-0.0003038439026568085,
0.007427725940942764,
0.021881451830267906,
-0.024134162813425064,
-0.06695745885372162,
0.0033833652269095182,
-0.04756179824471474,
0.047784820199012756,
0.034843605011701584,
-0.021687058731913567,
-0.015607333742082119,
0.003099680645391345,
-0.005063426680862904,
0.05886979401111603,
0.005537798628211021,
0.02239375188946724,
0.07926622033119202,
0.13898541033267975,
0.055218733847141266,
-0.016966411843895912,
0.0468149296939373,
-0.03571028262376785,
-0.008216222748160362,
0.007815039716660976,
0.03000580333173275,
-0.002017856575548649,
-0.023827608674764633,
0.020077673718333244,
0.013235698454082012,
-0.06689384579658508,
0.11278923600912094,
0.0073563712649047375,
-0.009868070483207703,
-0.07180383056402206,
0.010607847943902016,
0.005514162126928568,
-0.01438963320106268,
-0.05812019854784012,
0.015077657997608185,
-0.0015370595501735806,
-0.000021716363335144706,
0.047947611659765244,
-0.06424209475517273,
0.043688639998435974,
0.014580683782696724,
-0.03849750757217407,
-0.013398239389061928,
0.030788354575634003,
-0.03188437968492508,
-0.006486330647021532,
0.016531895846128464,
-0.029577454552054405,
0.049296215176582336,
-0.060248661786317825,
0.031046785414218903,
0.0264089684933424,
0.05613357201218605,
-0.018714170902967453,
-0.021660175174474716,
0.10926387459039688,
0.006647815462201834,
-0.005449504125863314,
-0.04472583159804344,
-0.03374286741018295,
-0.07448432594537735,
-0.017230629920959473,
-0.011202529072761536,
-0.02737366408109665,
0.004997153300791979,
-0.03904666751623154,
0.023981211706995964,
0.023907043039798737,
-0.012112317606806755,
0.047828882932662964,
0.03149261698126793,
-0.009192748926579952,
0.05020919069647789,
-0.016823813319206238,
-0.021595682948827744,
0.041156455874443054,
0.05106441676616669,
0.04313989728689194,
0.018840918317437172,
-0.0013016104931011796,
0.04144100099802017,
-0.01834709197282791,
-0.007436577696353197,
0.03729170188307762,
-0.023905275389552116,
0.010236511006951332,
-0.006236999295651913,
-0.02963310480117798,
0.03566904366016388,
0.017154620960354805,
-0.009695906192064285,
-0.039577167481184006,
0.019487472251057625,
-0.011199552565813065,
-0.0026430231519043446,
1.0744404993954533e-33,
0.029980044811964035,
0.08779814094305038,
0.0607377365231514,
0.003226117929443717,
0.037536557763814926,
-0.016461675986647606,
-0.039055515080690384,
0.07313153892755508,
-0.030694177374243736,
0.0704648569226265,
0.014392085373401642,
-0.08237216621637344,
-0.01546207070350647,
0.0873374417424202,
0.011120829731225967,
0.0396956130862236,
-0.05693184956908226,
-0.04076547548174858,
-0.011736509390175343,
0.14156053960323334,
-0.015419863164424896,
-0.04687141254544258,
-0.03590511530637741,
-0.014855444431304932,
-0.042025577276945114,
-0.0009851480135694146,
0.023134935647249222,
0.04400426521897316,
-0.031057968735694885,
-0.027299819514155388,
0.014131695032119751,
-0.008690821938216686,
-0.061941977590322495,
0.038943275809288025,
0.02770322561264038,
-0.08804323524236679,
0.029558738693594933,
-0.07815736532211304,
-0.036568865180015564,
-0.003801562823355198,
-0.02645132876932621,
0.08980855345726013,
-0.08057879656553268,
-0.018951360136270523,
-0.03426795452833176,
-0.029510730877518654,
0.05512641370296478,
0.027334772050380707,
-0.04646891728043556,
0.0906158909201622,
0.11361019313335419,
0.020386584103107452,
-0.10995160788297653,
-0.0404403880238533,
-0.10722462087869644,
-0.01508801244199276,
0.024371013045310974,
-0.021234825253486633,
0.03751182183623314,
-0.004621362779289484,
-0.03548888862133026,
0.025713883340358734,
-0.017579341307282448,
0.024484436959028244,
-0.047648243606090546,
-0.024963220581412315,
0.025143569335341454,
0.019827215000987053,
0.05786040052771568,
0.09725384414196014,
0.009032762609422207,
-0.01742769591510296,
0.05251699313521385,
0.03312671557068825,
0.04523429274559021,
0.024659695103764534,
-0.020428407937288284,
-0.01060634758323431,
0.016094088554382324,
0.08739843964576721,
-0.06744387000799179,
0.07781624048948288,
-0.021188892424106598,
0.04993860423564911,
-0.11259320378303528,
-0.13029533624649048,
-0.01712006703019142,
-0.03282459080219269,
-0.02996324747800827,
-0.031754132360219955,
-0.008579554036259651,
-0.04828775301575661,
-0.04092532768845558,
-0.03398796170949936,
-0.03943144902586937,
-2.9467700480399155e-35,
0.09727838635444641,
-0.023152589797973633,
0.015558086335659027,
0.005079460330307484,
-0.027689959853887558,
-0.07917600125074387,
-0.004446313250809908,
0.033726926892995834,
0.12162598222494125,
-0.04914521053433418,
-0.007090750616043806,
0.031195344403386116,
0.06417439132928848,
0.0865502804517746,
0.08977662771940231,
0.04463661089539528,
-0.06971687823534012,
0.013371982611715794,
-0.12836462259292603,
-0.057848889380693436,
0.07502511888742447,
0.04161127284169197,
0.06803618371486664,
0.032554492354393005,
-0.12405000627040863,
0.12523183226585388,
-0.020247288048267365,
0.04137950763106346,
0.015714135020971298,
-0.01253802701830864,
-0.008295745588839054,
0.010839044116437435,
0.01441027782857418,
-0.0631030797958374,
-0.008154832758009434,
0.05502234399318695,
0.008078768849372864,
0.034405216574668884,
-0.004089858382940292,
0.0316447950899601,
0.04234685003757477,
0.01738741248846054,
-0.05229611694812775,
0.06813878566026688,
0.033257439732551575,
0.034834492951631546,
0.057264067232608795,
-0.011285467073321342,
0.014613469131290913,
0.06797277927398682,
-0.0915018692612648,
0.07039696723222733,
-0.11412028223276138,
0.006971427705138922,
0.05482730641961098,
-0.09405093640089035,
-0.1519048660993576,
-0.09058676660060883,
0.08450224995613098,
0.03192330524325371,
0.0013289913767948747,
0.00428966386243701,
-0.10234329104423523,
0.08944661915302277,
-0.03396116942167282,
0.005814914591610432,
0.007223732769489288,
-0.006949148140847683,
0.010593035258352757,
0.08574888110160828,
0.0026857105549424887,
-0.023670418187975883,
-0.005793125834316015,
0.022855142131447792,
-0.055179350078105927,
-0.02051527611911297,
-0.0044303652830421925,
-0.006040313746780157,
0.0066613247618079185,
-0.018432674929499626,
0.084286168217659,
-0.040697358548641205,
-0.02200550213456154,
0.11926814913749695,
0.10968995094299316,
0.06355595588684082,
-0.016908420249819756,
0.01731647178530693,
0.031832363456487656,
-0.07640708982944489,
-0.08447769284248352,
0.036232322454452515,
-0.020383117720484734,
0.04123228043317795,
-0.060052596032619476,
-3.754699307023657e-8,
-0.05474128946661949,
0.08172114938497543,
0.007684258744120598,
-0.021892372518777847,
-0.11532841622829437,
-0.04818849638104439,
-0.0018286611884832382,
0.06220873072743416,
-0.01432445365935564,
-0.04083549231290817,
0.06877890974283218,
-0.052469223737716675,
0.01969931647181511,
-0.04869290813803673,
0.026890167966485023,
0.061168257147073746,
-0.0010920775821432471,
0.0005598391871899366,
-0.039688315242528915,
0.055304497480392456,
-0.039248764514923096,
0.03655145689845085,
0.04105972498655319,
0.008027167990803719,
0.002724778838455677,
-0.007361819036304951,
-0.01474193949252367,
0.0734863355755806,
0.1398572325706482,
0.008165444247424603,
-0.008317289873957634,
0.02046794816851616,
0.052731551229953766,
0.032696276903152466,
-0.03520531952381134,
-0.05999289080500603,
-0.04087924584746361,
-0.004485125653445721,
0.06723885983228683,
-0.019992349669337273,
-0.07244888693094254,
0.012467989698052406,
-0.06127961724996567,
0.02286137454211712,
-0.020796602591872215,
0.06609459966421127,
0.014015581458806992,
-0.07280102372169495,
-0.022592628374695778,
-0.0459243468940258,
-0.03212495520710945,
0.00950726680457592,
-0.02336663194000721,
0.08738251030445099,
0.08842682838439941,
0.01948556862771511,
-0.029703395441174507,
-0.04994149133563042,
0.0479966476559639,
-0.0060182069428265095,
-0.07156447321176529,
-0.038399193435907364,
-0.07903055846691132,
0.011157123371958733
] |
Black Sheep in the Herd: Playing with Spuriously Correlated Attributes for Vision-Language Recognition | https://openreview.net/forum?id=g1fkhbhHjL | [
"Xinyu Tian",
"Shu Zou",
"Zhaoyuan Yang",
"Mengqi He",
"Jing Zhang"
] | Poster | Few-shot adaptation for Vision-Language Models (VLMs) presents a dilemma: balancing in-distribution accuracy with out-of-distribution generalization. Recent research has utilized low-level concepts such as visual attributes to enhance generalization. However, this study reveals that VLMs overly rely on a small subset of attributes on decision-making, which co-occur with the category but are not inherently part of it, termed spuriously correlated attributes. This biased nature of VLMs results in poor generalization. To address this, 1) we first propose Spurious Attribute Probing (SAP), identifying and filtering out these problematic attributes to significantly enhance the generalization of existing attribute-based methods; 2) We introduce Spurious Attribute Shielding (SAS), a plug-and-play module that mitigates the influence of these attributes on prediction, seamlessly integrating into various Parameter-Efficient Fine-Tuning (PEFT) methods. In experiments, SAP and SAS significantly enhance accuracy on distribution shifts across 11 datasets and 3 generalization tasks without compromising downstream performance, establishing a new state-of-the-art benchmark. | Few-shot Adaptation, Prompt Learning, Vision-Language Models | null | 2,119 | 2502.15809 | [
0.03277931734919548,
-0.0896749496459961,
0.06205885112285614,
0.04548768326640129,
0.12740713357925415,
0.03400730714201927,
0.04006950929760933,
-0.060354799032211304,
-0.0017676394199952483,
-0.045841898769140244,
-0.018497968092560768,
-0.11714480072259903,
0.005530036985874176,
0.07063453644514084,
0.006639749277383089,
-0.07918764650821686,
0.10542263090610504,
0.018679898232221603,
-0.07359994947910309,
-0.011329964734613895,
-0.02566894330084324,
-0.03341863676905632,
0.043516989797353745,
-0.005957961082458496,
0.047713752835989,
-0.028840776532888412,
-0.04452258348464966,
-0.06444130092859268,
-0.043107591569423676,
-0.04958469048142433,
-0.037603460252285004,
0.0411401093006134,
-0.0354926735162735,
0.03535410761833191,
-0.057964883744716644,
0.021356910467147827,
-0.03745061159133911,
0.0722106322646141,
-0.005834076553583145,
0.008936714380979538,
-0.017137734219431877,
-0.025158323347568512,
0.039888039231300354,
0.01613008603453636,
0.006077589001506567,
0.02876720204949379,
0.04592352733016014,
-0.050977375358343124,
-0.03129744157195091,
0.02209327183663845,
-0.11918369680643082,
0.007563656195998192,
-0.013500102795660496,
0.038086019456386566,
-0.008600163273513317,
-0.010501652956008911,
0.06811559945344925,
0.050763849169015884,
-0.024315595626831055,
-0.0033177263103425503,
-0.056212566792964935,
-0.11631941795349121,
-0.0014015648048371077,
0.0621911883354187,
-0.04092080146074295,
0.02230634167790413,
0.020596561953425407,
0.0017245315248146653,
-0.07568623870611191,
-0.011045441962778568,
0.006171884015202522,
0.10839051008224487,
0.034257084131240845,
0.050603967159986496,
0.06128327175974846,
0.09863832592964172,
0.019612165167927742,
-0.037878528237342834,
0.08833644539117813,
-0.12186071276664734,
0.017536921426653862,
0.07371673732995987,
-0.062035441398620605,
0.03439759835600853,
0.17622756958007812,
0.04463515803217888,
-0.01719171181321144,
0.009128418751060963,
-0.06309393793344498,
-0.03863373398780823,
-0.04387917369604111,
-0.09377660602331161,
0.06497756391763687,
-0.008897880092263222,
0.036790501326322556,
0.06349802017211914,
-0.002345856511965394,
-0.12193796783685684,
-0.021250035613775253,
0.04936071112751961,
-0.00884440541267395,
-0.006788669619709253,
0.043110672384500504,
-0.051030129194259644,
-0.015355521813035011,
-0.06181707978248596,
0.033833544701337814,
0.04662622883915901,
-0.0026853035669773817,
-0.07411528378725052,
0.01785328797996044,
0.07323890924453735,
0.04029161110520363,
0.025626445189118385,
0.004950800910592079,
0.04356463998556137,
-0.002744407393038273,
-0.03741086274385452,
0.0024778235238045454,
0.038917940109968185,
0.0252018291503191,
0.006287753582000732,
0.030804893001914024,
-0.02503654547035694,
0.03210907056927681,
0.06303200870752335,
0.015730520710349083,
1.4862375681591054e-33,
0.014575236476957798,
-0.03009451925754547,
0.051526494324207306,
0.014283168129622936,
0.03701268509030342,
-0.0737372636795044,
-0.012747116386890411,
-0.01670139469206333,
-0.04034124314785004,
0.04870578646659851,
0.03293122723698616,
0.015027509070932865,
-0.0585530549287796,
0.10501089692115784,
0.06305950880050659,
0.004620798863470554,
-0.04083305224776268,
0.07738681137561798,
0.0012466518674045801,
0.018732624128460884,
0.041697464883327484,
-0.05373239517211914,
0.010423821397125721,
-0.04889661446213722,
-0.007385964971035719,
0.009448857977986336,
0.039032042026519775,
-0.05541035905480385,
-0.07502308487892151,
0.03624795749783516,
-0.09240099787712097,
0.01929260417819023,
0.07103636115789413,
-0.012187914922833443,
0.0328655019402504,
-0.13892331719398499,
-0.04314210265874863,
-0.07914875447750092,
0.03801680728793144,
0.020648863166570663,
-0.014564397744834423,
0.03719336539506912,
-0.03159220889210701,
0.01997629925608635,
-0.03432447090744972,
-0.0008163040038198233,
0.004295960068702698,
0.04661499336361885,
-0.09640863537788391,
-0.010097726248204708,
0.06229052320122719,
0.009865967556834221,
0.032553158700466156,
-0.02364732138812542,
-0.002767737954854965,
0.042340245097875595,
0.0263957716524601,
-0.027393199503421783,
-0.051921457052230835,
-0.07777661085128784,
0.024730302393436432,
0.04262005165219307,
-0.06060333177447319,
0.0015751553000882268,
0.028747430071234703,
-0.013164710253477097,
0.03341829776763916,
0.04280371218919754,
0.0005990398349240422,
0.017409980297088623,
0.014153102412819862,
0.012282169423997402,
-0.10543374717235565,
-0.054553717374801636,
0.06484924256801605,
-0.009433028288185596,
0.10034158080816269,
-0.018974443897604942,
0.07009345293045044,
0.018499068915843964,
-0.025489985942840576,
0.04026474058628082,
-0.001987351570278406,
0.010081163607537746,
-0.049065493047237396,
-0.053345341235399246,
0.08308715373277664,
-0.04222773015499115,
-0.017641063779592514,
-0.09316359460353851,
0.012377455830574036,
0.02059444785118103,
-0.07402648031711578,
-0.08542905002832413,
0.012071488425135612,
-2.296902737195303e-33,
-0.06757752597332001,
0.013926032930612564,
0.015186827629804611,
0.015762362629175186,
-0.026858607307076454,
-0.07676136493682861,
0.06724033504724503,
0.007722171489149332,
0.05279536545276642,
-0.043866951018571854,
-0.04940434917807579,
-0.0599074512720108,
-0.06499084830284119,
0.005596791859716177,
0.050876278430223465,
-0.009904686361551285,
-0.04632163792848587,
0.05811502784490585,
0.05051293969154358,
-0.0011539342813193798,
0.04147011786699295,
0.0747651532292366,
-0.021216772496700287,
0.027189968153834343,
-0.018642107024788857,
0.04433564469218254,
-0.06764140725135803,
0.04642322659492493,
-0.05018492415547371,
-0.07031216472387314,
-0.011891324073076248,
0.04741964489221573,
-0.0390733927488327,
-0.030901312828063965,
-0.054380692541599274,
0.04654473438858986,
-0.01957920379936695,
-0.04309798404574394,
-0.031266115605831146,
0.13552792370319366,
0.041617121547460556,
-0.027342643588781357,
-0.09080905467271805,
-0.03754626214504242,
-0.02483946457505226,
-0.03906959295272827,
-0.029688259586691856,
-0.01203962229192257,
0.015569299459457397,
0.029992470517754555,
-0.07534307986497879,
0.05380172282457352,
-0.05972839891910553,
0.04716174304485321,
-0.031533289700746536,
-0.07619766891002655,
0.03459453210234642,
-0.019449763000011444,
0.05250414088368416,
0.015076819807291031,
-0.057996392250061035,
0.04677658528089523,
0.011659553274512291,
-0.05178907513618469,
-0.004185773432254791,
-0.011224070563912392,
-0.016162067651748657,
-0.0361640527844429,
0.0597354918718338,
-0.07256587594747543,
0.033606257289648056,
-0.06647932529449463,
-0.030084028840065002,
-0.01915372908115387,
-0.06960439682006836,
0.012223923578858376,
0.012818428687751293,
-0.004234348889440298,
0.023764081299304962,
0.007607726380228996,
0.042870376259088516,
-0.06422612816095352,
0.028188815340399742,
0.11729062348604202,
0.051774762570858,
0.11527083069086075,
0.0384821891784668,
0.09085685014724731,
0.10534358024597168,
-0.05737589672207832,
-0.08498643338680267,
0.0002970276400446892,
0.005358607042580843,
0.10455676913261414,
-0.04559534788131714,
-4.8346425529643966e-8,
-0.08193250745534897,
0.07008076459169388,
-0.03832800313830376,
0.048454735428094864,
0.0016895001754164696,
0.005415165796875954,
-0.17534087598323822,
0.08743076026439667,
0.013416371308267117,
-0.008574347011744976,
0.04667351394891739,
0.02063736692070961,
0.0008457567309960723,
0.05703306570649147,
0.005316285416483879,
0.05752483382821083,
0.0019650484900921583,
0.05780750885605812,
-0.039073776453733444,
0.033995818346738815,
-0.003127716016024351,
0.008549343794584274,
-0.006280315574258566,
-0.02653638646006584,
-0.016551494598388672,
-0.1026298999786377,
-0.047401994466781616,
0.05132066830992699,
0.10315568000078201,
-0.060399241745471954,
0.00456198351457715,
0.04332950338721275,
-0.016373418271541595,
-0.038199931383132935,
0.03906626999378204,
0.09435443580150604,
-0.0525607131421566,
0.00624040374532342,
0.04227575287222862,
0.06657326221466064,
0.008411349728703499,
0.0739755928516388,
-0.006822898518294096,
0.02614966779947281,
-0.016424603760242462,
-0.009623362682759762,
-0.04085763543844223,
-0.07254986464977264,
0.01913486234843731,
-0.037285301834344864,
0.0063619473949074745,
-0.03143429383635521,
-0.03179014101624489,
0.11146354675292969,
0.014347651973366737,
-0.059188470244407654,
0.026263490319252014,
-0.04598874971270561,
0.04865977168083191,
0.031426310539245605,
0.01255708560347557,
-0.012384776957333088,
-0.051499009132385254,
0.012166297994554043
] |
Democratic Training Against Universal Adversarial Perturbations | https://openreview.net/forum?id=4M0BRyGMnJ | [
"Bing Sun",
"Jun Sun",
"Wei Zhao"
] | Poster | Despite their advances and success, real-world deep neural networks are known to be vulnerable to adversarial attacks. Universal adversarial perturbation, an input-agnostic attack, poses a serious threat for them to be deployed in security-sensitive systems. In this case, a single universal adversarial perturbation deceives the model on a range of clean inputs without requiring input-specific optimization, which makes it particularly threatening. In this work, we observe that universal adversarial perturbations usually lead to abnormal entropy spectrum in hidden layers, which suggests that the prediction is dominated by a small number of ``feature'' in such cases (rather than democratically by many features). Inspired by this, we propose an efficient yet effective defense method for mitigating UAPs called \emph{Democratic Training} by performing entropy-based model enhancement to suppress the effect of the universal adversarial perturbations in a given model. \emph{Democratic Training} is evaluated with 7 neural networks trained on 5 benchmark datasets and 5 types of state-of-the-art universal adversarial attack methods. The results show that it effectively reduces the attack success rate, improves model robustness and preserves the model accuracy on clean samples. | Neural network adversarial attack; Universal adversarial perturbation; Adversarial attack defense | A novel method to mitigate the effect of UAPs via democratic training. | 2,118 | 2502.05542 | [
-0.08299766480922699,
0.004372905474156141,
0.01740488037467003,
0.013526893220841885,
0.030321462079882622,
0.05268051475286484,
-0.0023432401940226555,
-0.14423075318336487,
0.005426843650639057,
-0.023395102471113205,
0.026293713599443436,
-0.04712453484535217,
0.075727678835392,
-0.008769516833126545,
-0.10358934104442596,
0.020499270409345627,
0.07360422611236572,
0.10898657143115997,
-0.10975498706102371,
0.009475182741880417,
0.030521059408783913,
0.02814597077667713,
-0.020316345617175102,
0.06050102785229683,
0.007756607607007027,
-0.06781736016273499,
-0.003891689470037818,
-0.010539568029344082,
-0.038608066737651825,
-0.06594245135784149,
0.03221610561013222,
-0.08114777505397797,
0.0014605737524107099,
0.014491412788629532,
-0.06812065094709396,
0.020084837451577187,
-0.04897688329219818,
-0.008264562115073204,
-0.021940214559435844,
-0.08056730031967163,
-0.007093658205121756,
-0.037269048392772675,
-0.028582287952303886,
0.02614671178162098,
0.023457637056708336,
-0.006844279356300831,
0.0373864509165287,
-0.03665605187416077,
0.03342689946293831,
-0.04381518065929413,
0.013465975411236286,
0.0624878816306591,
0.026799941435456276,
0.038430355489254,
0.014306225813925266,
-0.08585771173238754,
-0.024771979078650475,
0.043497923761606216,
-0.04110510274767876,
0.00031587923876941204,
-0.018709149211645126,
-0.016670603305101395,
-0.03816645219922066,
-0.0035976346116513014,
0.0046672578901052475,
0.007258042227476835,
0.01820877194404602,
0.025326265022158623,
0.0023341593332588673,
0.0385325588285923,
0.02961062826216221,
0.08705856651067734,
-0.031933579593896866,
0.006810274440795183,
0.08562611043453217,
0.11370013654232025,
-0.038235124200582504,
-0.013294295407831669,
0.063556969165802,
-0.03392787650227547,
0.1135164275765419,
-0.004398922435939312,
0.014686541631817818,
-0.013489040546119213,
0.14507080614566803,
0.019832290709018707,
-0.039940401911735535,
-0.01395599078387022,
0.10458647459745407,
0.0036510496865957975,
-0.09986702352762222,
-0.03589334338903427,
0.049207642674446106,
0.03389760106801987,
0.08016844838857651,
-0.011832588352262974,
-0.0007862706552259624,
-0.08016331493854523,
-0.08701501786708832,
0.10359791666269302,
-0.053376976400613785,
-0.017127612605690956,
-0.031681548804044724,
-0.014573334716260433,
0.09788632392883301,
0.031090890988707542,
0.07343336939811707,
-0.027594776824116707,
0.05349689722061157,
-0.0746036171913147,
0.07420232146978378,
0.030165230855345726,
-0.010088561102747917,
-0.001930421800352633,
0.04971776530146599,
0.0006250953301787376,
-0.041645750403404236,
0.025138527154922485,
-0.08844339847564697,
0.0793699100613594,
-0.10048303753137589,
0.021065035834908485,
0.0016830879030749202,
0.028119675815105438,
0.03135862573981285,
0.005005244631320238,
-0.09647898375988007,
2.751169039321722e-33,
-0.013424509204924107,
0.013979703187942505,
-0.023595180362462997,
-0.020903069525957108,
0.008077863603830338,
0.012751070782542229,
-0.002111536916345358,
0.050597257912158966,
-0.02519838698208332,
0.06172599270939827,
-0.09874594956636429,
0.0054809399880468845,
-0.01600014604628086,
0.09076933562755585,
0.07663971185684204,
0.0764877051115036,
-0.004017733037471771,
-0.025505300611257553,
0.05807022005319595,
-0.020948363468050957,
0.042187731713056564,
0.013678059913218021,
-0.017733268439769745,
-0.09441109746694565,
0.024224411696195602,
-0.029199764132499695,
0.0083995396271348,
-0.007630397565662861,
0.01091128308326006,
0.06682948023080826,
-0.10552962869405746,
0.039573606103658676,
0.005519399885088205,
0.013318704441189766,
0.0315658263862133,
-0.018163597211241722,
-0.04519377648830414,
-0.0212189182639122,
0.01198669895529747,
0.018556829541921616,
-0.0004674278025049716,
0.008753285743296146,
-0.04549336060881615,
-0.011111559346318245,
0.004068367648869753,
-0.003875701455399394,
0.018116116523742676,
0.02492094784975052,
-0.12831257283687592,
-0.03887127712368965,
0.00858217477798462,
0.05030589550733566,
-0.05055560916662216,
-0.07152774184942245,
-0.06087569147348404,
-0.10815692692995071,
0.0523858442902565,
0.01018037274479866,
-0.03713018819689751,
0.027899058535695076,
-0.02629384770989418,
-0.028340209275484085,
-0.01679542474448681,
0.01370553020387888,
-0.02797565795481205,
-0.051017411053180695,
-0.0044133663177490234,
0.023488717153668404,
0.050745412707328796,
-0.050911031663417816,
-0.01168645266443491,
0.03432423621416092,
0.0048126233741641045,
-0.014112583361566067,
-0.028251051902770996,
-0.008258256129920483,
0.05379413440823555,
-0.020656727254390717,
0.028772836551070213,
0.029089750722050667,
-0.009491599164903164,
0.002110252855345607,
0.011421455070376396,
0.007934434339404106,
-0.06604669988155365,
0.024030517786741257,
0.008613158017396927,
-0.031649086624383926,
-0.027994411066174507,
0.027231106534600258,
-0.05850515142083168,
-0.03421342372894287,
0.023007279261946678,
0.01930440403521061,
-0.01772017776966095,
-3.817645288983866e-33,
-0.07491917908191681,
0.04128650575876236,
-0.05263689160346985,
0.031015384942293167,
-0.058666668832302094,
-0.030410852283239365,
0.013281396590173244,
0.08366240561008453,
-0.04813884198665619,
-0.020555438473820686,
-0.010603330098092556,
0.008621273562312126,
0.10137458145618439,
-0.04503149539232254,
0.051623471081256866,
-0.01050922553986311,
-0.06652005016803741,
0.07933907210826874,
0.026766879484057426,
0.023636139929294586,
0.023621585220098495,
0.10748399049043655,
-0.1077059879899025,
0.11414094269275665,
0.05324823781847954,
-0.06890711188316345,
-0.04980241134762764,
0.06550519913434982,
0.058937542140483856,
0.00557694211602211,
-0.051609449088573456,
-0.03426109254360199,
-0.028981508687138557,
-0.020927084609866142,
-0.03709811344742775,
0.049210891127586365,
0.08683614432811737,
0.02457684464752674,
-0.008969184942543507,
0.07752871513366699,
0.04507969692349434,
-0.01774914748966694,
-0.09112802892923355,
0.002910117618739605,
-0.028850549831986427,
-0.026415783911943436,
-0.015997344627976418,
0.027929818257689476,
0.03997558355331421,
0.002553327474743128,
0.048790253698825836,
-0.023456556722521782,
-0.011485574766993523,
0.039964526891708374,
-0.06867630034685135,
-0.03458026051521301,
0.03317447006702423,
-0.016410646960139275,
0.1170477643609047,
0.10177953541278839,
-0.021922016516327858,
-0.04275522753596306,
-0.019366903230547905,
0.014720486477017403,
-0.03284188359975815,
0.03599392995238304,
-0.00024570373352617025,
0.1246965080499649,
0.028688576072454453,
-0.004837315995246172,
0.062352098524570465,
-0.007071941625326872,
-0.024048693478107452,
0.06422491371631622,
-0.03686123713850975,
-0.041352588683366776,
-0.0006252243765629828,
-0.0643574669957161,
0.02077891305088997,
0.023708391934633255,
0.0405409038066864,
-0.029035991057753563,
0.02824123203754425,
-0.013798551633954048,
0.08474243432283401,
0.04400821402668953,
0.11154500395059586,
-0.05991598963737488,
-0.021261194720864296,
0.026746951043605804,
-0.04307274892926216,
0.0778965950012207,
-0.036486584693193436,
0.012050865218043327,
-0.00754469633102417,
-4.230341588140618e-8,
-0.02472413331270218,
0.06336759775876999,
0.04932691901922226,
-0.002145507838577032,
0.031949032098054886,
0.021408749744296074,
-0.045753542333841324,
0.027260681614279747,
-0.002777643036097288,
-0.056930530816316605,
0.07639683783054352,
0.03411121666431427,
-0.023369386792182922,
-0.0736672431230545,
-0.02723202295601368,
0.0588359460234642,
-0.02698693796992302,
0.025280848145484924,
-0.0155370207503438,
-0.0675278902053833,
0.01681290939450264,
-0.010139592923223972,
-0.051501091569662094,
-0.07319465279579163,
0.021715626120567322,
-0.09289414435625076,
-0.06757233291864395,
0.032501112669706345,
0.04588937386870384,
0.0983986034989357,
-0.06352345645427704,
-0.06358994543552399,
-0.01222232822328806,
-0.012074077501893044,
-0.040588200092315674,
0.17624612152576447,
-0.06656496226787567,
-0.09275341778993607,
0.027490898966789246,
-0.04306243360042572,
0.004425283055752516,
-0.006340584252029657,
0.05456674098968506,
-0.05155470594763756,
-0.02590375579893589,
-0.03091450408101082,
0.04014159366488457,
-0.12415913492441177,
0.047711700201034546,
0.02431461215019226,
0.07407884299755096,
0.01773262768983841,
-0.00814279355108738,
0.1368778795003891,
0.048678841441869736,
-0.030737482011318207,
0.002163897268474102,
-0.12115106731653214,
0.028737328946590424,
0.11177214980125427,
0.03948384150862694,
-0.0005753277218900621,
-0.06064555421471596,
-0.04501030594110489
] |
Framer: Interactive Frame Interpolation | https://openreview.net/forum?id=Lp40Z40N07 | [
"Wen Wang",
"Qiuyu Wang",
"Kecheng Zheng",
"Hao OUYANG",
"Zhekai Chen",
"Biao Gong",
"Hao Chen",
"Yujun Shen",
"Chunhua Shen"
] | Poster | We propose Framer for interactive frame interpolation, which targets producing smoothly transitioning frames between two images as per user creativity. Concretely, besides taking the start and end frames as inputs, our approach supports customizing the transition process by tailoring the trajectory of some selected keypoints. Such a design enjoys two clear benefits. First, incorporating human interaction mitigates the issue arising from numerous possibilities of transforming one image to another, and in turn enables finer control of local motions. Second, as the most basic form of interaction, keypoints help establish the correspondence across frames, enhancing the model to handle challenging cases (e.g., objects on the start and end frames are of different shapes and styles). It is noteworthy that our system also offers an "autopilot" mode, where we introduce a module to estimate the keypoints and refine the trajectory automatically, to simplify the usage in practice. Extensive experimental results demonstrate the appealing performance of Framer on various applications, such as image morphing, time-lapse video generation, cartoon interpolation, etc. The code, model, and interface are publicly accessible at https://github.com/aim-uofa/Framer. | Video Frame Interpolation; Interactive; Diffusion Model; Correspondence Modeling | We propose Framer, an interactive frame interpolation method that allows users to produce smoothly transitioning frames between two images by customizing the trajectory of selected keypoints, enhancing control and handling challenging cases. | 2,116 | 2410.18978 | [
-0.07829704135656357,
-0.050104688853025436,
0.012393875047564507,
-0.062301963567733765,
0.06122219190001488,
-0.02070043608546257,
-0.01307754311710596,
0.003013255074620247,
0.029544061049818993,
-0.07784201204776764,
-0.035354092717170715,
-0.03769676387310028,
-0.00798789132386446,
0.029460377991199493,
0.003297798801213503,
-0.02800326608121395,
-0.004683550912886858,
0.05711501091718674,
-0.02227383852005005,
-0.022464964538812637,
-0.02286459505558014,
-0.09548021852970123,
-0.02658931538462639,
-0.07924910634756088,
0.01646244339644909,
0.03722020611166954,
-0.01672004535794258,
-0.02729533612728119,
0.008007858879864216,
-0.04191891849040985,
0.020972730591893196,
-0.066904217004776,
0.0001775860000634566,
-0.04022952541708946,
-0.08094736188650131,
-0.009566434659063816,
0.04490086808800697,
0.0257798433303833,
-0.09577842056751251,
-0.046691033989191055,
-0.005965027492493391,
0.012302463874220848,
0.08705689758062363,
0.03810758516192436,
0.061015043407678604,
0.04940415918827057,
-0.011093765497207642,
-0.035054389387369156,
-0.013702904805541039,
0.0722346156835556,
-0.08678945153951645,
-0.03912061080336571,
0.0040814136154949665,
-0.011801246553659439,
0.012096880935132504,
0.07386316359043121,
0.09152843058109283,
-0.04736459255218506,
0.08726256340742111,
-0.04192836955189705,
-0.04372464865446091,
0.046959150582551956,
0.029667116701602936,
0.031805045902729034,
-0.043417125940322876,
-0.08040751516819,
0.09240074455738068,
-0.04532669857144356,
0.056662194430828094,
-0.023540660738945007,
-0.10126975178718567,
-0.02859625034034252,
0.07697049528360367,
-0.04081495851278305,
-0.09404460340738297,
-0.012817857787013054,
-0.017991503700613976,
-0.008419969119131565,
-0.027615848928689957,
-0.0874875858426094,
0.07116472721099854,
-0.0463249608874321,
-0.009356414899230003,
-0.006125290412455797,
-0.030561773106455803,
-0.005433974787592888,
-0.014600735157728195,
0.051243171095848083,
0.027900973334908485,
0.07255980372428894,
-0.0512426532804966,
0.021143892779946327,
0.00393540458753705,
-0.04012859985232353,
0.0197820533066988,
-0.020885752514004707,
0.012451527640223503,
-0.09698863327503204,
0.04937838762998581,
0.07744041830301285,
-0.01504703238606453,
-0.05902739614248276,
0.027243876829743385,
-0.004257121589034796,
0.08610396832227707,
0.05746270716190338,
-0.007816364988684654,
0.02329108491539955,
-0.009446623735129833,
0.08018457144498825,
-0.02815386839210987,
-0.0638207420706749,
-0.02797415480017662,
0.02112480066716671,
0.0951080247759819,
0.031908705830574036,
0.01959878019988537,
-0.017834635451436043,
0.08241847157478333,
-0.016356218606233597,
0.07264681160449982,
-0.008800595067441463,
-0.04310189560055733,
0.02211049571633339,
0.045560840517282486,
-0.051431600004434586,
-0.011200076900422573,
3.1033304327677786e-33,
0.02811376564204693,
-0.012401963584125042,
0.05776534602046013,
0.04992213100194931,
0.047106072306632996,
0.0005431075696833432,
-0.03877938166260719,
-0.016389256343245506,
-0.02657872810959816,
-0.044462256133556366,
-0.01941261813044548,
-0.048440054059028625,
-0.005456368438899517,
0.06730661541223526,
0.05418696999549866,
-0.05180270969867706,
-0.02743331529200077,
0.058766480535268784,
0.055523261427879333,
0.11655547469854355,
-0.008793502114713192,
-0.03621647506952286,
-0.01932905800640583,
-0.031349778175354004,
-0.05816776677966118,
0.08635801821947098,
0.03701474517583847,
0.029016438871622086,
0.04223218187689781,
-0.0328473225235939,
-0.05709093436598778,
-0.049353450536727905,
-0.03990332409739494,
-0.02133852243423462,
-0.00007563017425127327,
-0.048030976206064224,
0.019725514575839043,
-0.04830152168869972,
-0.016799671575427055,
0.02874280884861946,
0.01713503524661064,
-0.02108163759112358,
-0.03044435754418373,
0.022331511601805687,
-0.04860668256878853,
-0.004593867342919111,
0.05940403416752815,
0.19567684829235077,
-0.13442939519882202,
0.06924490630626678,
0.047436852008104324,
0.023326754570007324,
-0.014794833958148956,
-0.06352642923593521,
-0.029262997210025787,
-0.024930933490395546,
0.03826507553458214,
-0.022447684779763222,
-0.00918070413172245,
0.006880664266645908,
-0.042803239077329636,
0.05838368833065033,
-0.057111844420433044,
-0.027813371270895004,
-0.020127568393945694,
-0.05616161599755287,
0.04042500630021095,
0.014168520458042622,
-0.017555713653564453,
0.06017834320664406,
-0.017429398372769356,
-0.02158261276781559,
0.038568731397390366,
-0.01618107594549656,
0.03132358938455582,
0.044741079211235046,
-0.03793209418654442,
0.004530075937509537,
-0.007844521664083004,
0.018126048147678375,
-0.05060773715376854,
0.00752493878826499,
0.013345031999051571,
-0.02405003271996975,
0.055586688220500946,
-0.06670695543289185,
0.01720118150115013,
0.013720620423555374,
-0.04642264172434807,
-0.07775818556547165,
-0.03915977105498314,
0.013302180916070938,
0.014316512271761894,
0.12220686674118042,
0.013502778485417366,
-1.4887126266160594e-33,
0.09636226296424866,
-0.0006601718487218022,
0.04738194867968559,
0.11212646961212158,
0.025822874158620834,
-0.007041957229375839,
0.06812497973442078,
-0.010919640772044659,
0.07242655009031296,
-0.07354516535997391,
-0.04108385369181633,
0.0005889025051146746,
0.040548015385866165,
-0.04819595441222191,
-0.0016085404204204679,
-0.026075270026922226,
0.01620638370513916,
-0.04040824994444847,
-0.047237955033779144,
0.005251876078546047,
0.055556394159793854,
0.0014643378090113401,
-0.09785238653421402,
-0.014980290085077286,
-0.11711454391479492,
0.088616281747818,
0.01738908886909485,
-0.0171638373285532,
-0.030484523624181747,
0.027168266475200653,
-0.013558931648731232,
-0.057596925646066666,
0.05957575514912605,
-0.08048521727323532,
0.0006801604176871479,
0.05080527812242508,
0.010131727904081345,
-0.024383725598454475,
-0.042940542101860046,
0.019930120557546616,
0.025195728987455368,
0.0009460539440624416,
0.02099372074007988,
0.039595670998096466,
-0.001743728993460536,
0.06160742789506912,
-0.04144342616200447,
0.08965140581130981,
-0.06611151248216629,
0.032986633479595184,
-0.10579648613929749,
0.028522511944174767,
-0.060761742293834686,
-0.10246798396110535,
0.0028721923008561134,
-0.05884158983826637,
-0.06935735791921616,
-0.021154792979359627,
0.06193576380610466,
0.016696549952030182,
0.017801599577069283,
-0.015216815285384655,
-0.074051134288311,
0.009404199197888374,
0.0329282321035862,
-0.015484290197491646,
-0.07136201858520508,
-0.10882332921028137,
-0.04995483160018921,
0.05897632986307144,
0.11278324574232101,
-0.03141019493341446,
0.003915757406502962,
0.019708756357431412,
0.06949126720428467,
-0.07065179198980331,
0.10159685462713242,
-0.020452141761779785,
0.07348141074180603,
-0.04344554245471954,
0.007445904426276684,
-0.04156392812728882,
0.07905887067317963,
0.04209160432219505,
-0.06786829978227615,
0.15030986070632935,
-0.07159432023763657,
0.04743245616555214,
0.09598974883556366,
0.04791773483157158,
-0.021235676482319832,
0.08390962332487106,
0.010424891486763954,
0.07567007094621658,
-0.006621909327805042,
-4.553234589366184e-8,
-0.051946595311164856,
0.05778343975543976,
0.02895013429224491,
0.0032569975592195988,
-0.11579243838787079,
-0.02787020429968834,
0.08288674801588058,
0.02647830732166767,
0.07889460772275925,
-0.07761210948228836,
0.08545676618814468,
0.009267960675060749,
0.055707357823848724,
0.03535062074661255,
-0.018027110025286674,
0.08976984769105911,
0.05746453255414963,
0.033178724348545074,
0.003309656400233507,
0.022406307980418205,
-0.03615496680140495,
0.026441041380167007,
-0.05549365282058716,
-0.06921941041946411,
-0.07561405748128891,
-0.008158056996762753,
0.007468611933290958,
0.06350284814834595,
0.050826478749513626,
-0.05492330342531204,
-0.02777491696178913,
0.017743119969964027,
0.046696923673152924,
0.02359267883002758,
0.024084851145744324,
-0.020532092079520226,
-0.03843947499990463,
0.004636493511497974,
0.012635034509003162,
-0.03252815827727318,
0.03099553845822811,
0.05298509821295738,
-0.04439244791865349,
-0.01706547476351261,
0.023937532678246498,
0.0344625860452652,
0.03566160425543785,
-0.08972334861755371,
-0.06657007336616516,
-0.0033589669037610292,
0.010149689391255379,
-0.055961091071367264,
-0.011653618887066841,
-0.013520626351237297,
0.09635872393846512,
0.033502575010061264,
0.029246937483549118,
0.028627080842852592,
0.10753650218248367,
0.10029672086238861,
-0.0739220529794693,
0.016284698620438576,
-0.053183380514383316,
0.014295900240540504
] |
LucidPPN: Unambiguous Prototypical Parts Network for User-centric Interpretable Computer Vision | https://openreview.net/forum?id=BM9qfolt6p | [
"Mateusz Pach",
"Koryna Lewandowska",
"Jacek Tabor",
"Bartosz Michał Zieliński",
"Dawid Damian Rymarczyk"
] | Poster | Prototypical parts networks combine the power of deep learning with the explainability of case-based reasoning to make accurate, interpretable decisions. They follow the this looks like that reasoning, representing each prototypical part with patches from training images. However, a single image patch comprises multiple visual features, such as color, shape, and texture, making it difficult for users to identify which feature is important to the model.
To reduce this ambiguity, we introduce the Lucid Prototypical Parts Network (LucidPPN), a novel prototypical parts network that separates color prototypes from other visual features. Our method employs two reasoning branches: one for non-color visual features, processing grayscale images, and another focusing solely on color information. This separation allows us to clarify whether the model's decisions are based on color, shape, or texture. Additionally, LucidPPN identifies prototypical parts corresponding to semantic parts of classified objects, making comparisons between data classes more intuitive, e.g., when two bird species might differ primarily in belly color.
Our experiments demonstrate that the two branches are complementary and together achieve results comparable to baseline methods. More importantly, LucidPPN generates less ambiguous prototypical parts, enhancing user understanding. | xai, interpretability, prototypical parts | Enhancing the interpretability of prototypical parts networks by aligning prototypical parts with object parts and separating color and non-color visual features through a novel LucidPPN architecture. | 2,112 | 2405.14331 | [
0.022956063970923424,
-0.0023960124235600233,
0.09051274508237839,
0.005621007643640041,
0.13772331178188324,
-0.0344427190721035,
0.01684202067553997,
-0.0526309609413147,
0.019142145290970802,
-0.08723024278879166,
-0.08831944316625595,
-0.11044252663850784,
0.03076748736202717,
0.06985417753458023,
-0.013802744448184967,
-0.011877444572746754,
0.0448133610188961,
0.029955152422189713,
-0.12861202657222748,
0.02525748498737812,
-0.03150518983602524,
-0.012180706486105919,
0.03969353809952736,
-0.0012104043271392584,
-0.0702492892742157,
-0.009367491118609905,
0.020222323015332222,
-0.016720090061426163,
-0.006622764281928539,
-0.0482967235147953,
0.07212959975004196,
-0.0027909446507692337,
-0.04222169518470764,
0.07778050005435944,
-0.04886535182595253,
0.051494937390089035,
-0.026382695883512497,
0.04762112721800804,
-0.042726557701826096,
-0.034907955676317215,
0.005493186414241791,
0.006320194341242313,
-0.06279432028532028,
0.07050967216491699,
0.1302039474248886,
0.0718291774392128,
-0.008988941088318825,
-0.005781949032098055,
0.019407054409384727,
0.03189524635672569,
-0.08236321061849594,
-0.057250380516052246,
-0.06786289811134338,
0.05942806601524353,
0.011676176451146603,
0.11802642047405243,
0.05847219377756119,
-0.028601400554180145,
0.021321430802345276,
-0.03247688710689545,
0.042923539876937866,
-0.0293801911175251,
-0.0018807280575856566,
-0.012200948782265186,
-0.0012484390754252672,
-0.010011298581957817,
0.0395842045545578,
-0.036855340003967285,
0.006796577945351601,
-0.07633303105831146,
0.06309301406145096,
0.10799148678779602,
-0.014623649418354034,
0.04071458429098129,
-0.01856829784810543,
0.10977952927350998,
0.0811692401766777,
-0.00016717682592570782,
0.007916828617453575,
-0.09676408767700195,
-0.0029457630589604378,
0.025229373946785927,
-0.006848445162177086,
-0.006861910689622164,
0.12221364676952362,
0.029457347467541695,
-0.036196451634168625,
-0.0007913521258160472,
-0.11412693560123444,
-0.0035796714946627617,
-0.06370500475168228,
-0.01599508710205555,
0.05714455619454384,
-0.05845537781715393,
0.0662524625658989,
0.05674039199948311,
0.02317190170288086,
-0.11947394907474518,
0.006357699632644653,
0.06857287138700485,
-0.046645574271678925,
-0.046806640923023224,
0.0918058529496193,
-0.0599130243062973,
0.034953247755765915,
-0.029017465189099312,
0.03290373459458351,
-0.040424834936857224,
0.08472029119729996,
-0.05706888437271118,
0.050991930067539215,
0.025909243151545525,
-0.0650341734290123,
0.005891888868063688,
-0.038618046790361404,
-0.011823786422610283,
0.00635992456227541,
-0.011883698403835297,
0.020915407687425613,
0.02009744569659233,
0.010969097726047039,
-0.006980585400015116,
-0.009104236960411072,
0.007503416854888201,
0.028764469549059868,
-0.06732184439897537,
-0.05929366499185562,
3.072665825916656e-33,
0.043264321982860565,
-0.07317075878381729,
0.023240460082888603,
0.013250197283923626,
0.13441859185695648,
-0.03577015921473503,
-0.09060265123844147,
-0.008502650074660778,
-0.03128309175372124,
0.023784616962075233,
-0.036877065896987915,
-0.004992566537111998,
-0.1053280383348465,
0.13450664281845093,
0.07912766933441162,
0.006292047910392284,
-0.06175535172224045,
-0.013378150761127472,
-0.039238546043634415,
-0.002971762092784047,
-0.08893004059791565,
-0.009897451847791672,
-0.02789839170873165,
0.0237409770488739,
-0.017530344426631927,
-0.03395157679915428,
0.04896509647369385,
-0.003161821747198701,
0.02685633674263954,
-0.03565429151058197,
-0.027629870921373367,
0.025975046679377556,
0.05764518305659294,
0.07567524164915085,
0.0059185754507780075,
-0.08655961602926254,
0.004267573356628418,
-0.024407552555203438,
0.08329356461763382,
0.047105323523283005,
0.02307196520268917,
0.0023529720492661,
-0.0016542378580197692,
0.039218030869960785,
-0.06100017949938774,
-0.027122700586915016,
-0.0000040750933294475544,
0.0005489678005687892,
-0.09339532256126404,
0.06037173047661781,
0.06001769378781319,
0.0018602210329845548,
0.060978375375270844,
-0.13050781190395355,
-0.093439482152462,
-0.01767071709036827,
0.03016211837530136,
-0.007170408498495817,
0.021490519866347313,
-0.054104894399642944,
-0.030002065002918243,
0.04649582877755165,
0.01719491556286812,
-0.00006389980262611061,
-0.024163519963622093,
0.03258943557739258,
-0.0009648322011344135,
0.061156030744314194,
0.0013699757400900126,
-0.0064513105899095535,
-0.04770534485578537,
0.034099675714969635,
0.04224548861384392,
-0.0845993310213089,
0.00964030809700489,
0.013042409904301167,
0.03614053502678871,
-0.012157639488577843,
-0.006958672311156988,
0.03759143874049187,
-0.04595823958516121,
0.0933036282658577,
-0.03163621202111244,
-0.01986929215490818,
-0.1222933754324913,
0.016589339822530746,
0.06150933727622032,
-0.0258195623755455,
0.005516937933862209,
-0.06820306181907654,
-0.029714412987232208,
0.025800392031669617,
-0.02909281849861145,
-0.04554186761379242,
0.04186670854687691,
-2.487666735892544e-33,
-0.06304062902927399,
0.01658693328499794,
-0.056790318340063095,
0.03449578955769539,
-0.0030831010080873966,
-0.03438533842563629,
-0.0009345115977339447,
0.018331507220864296,
-0.012560051865875721,
-0.00382434600032866,
0.01475022453814745,
0.02281791716814041,
-0.03663048520684242,
-0.03348763287067413,
-0.03273051232099533,
0.019379062578082085,
-0.06772556155920029,
0.08369140326976776,
0.0023815929889678955,
0.024314941838383675,
0.0007705434109084308,
0.07471570372581482,
-0.1236523911356926,
-0.00846803467720747,
-0.13877178728580475,
0.1151285246014595,
-0.01940549910068512,
0.024510212242603302,
-0.02655058167874813,
-0.04655889794230461,
-0.03780445083975792,
-0.06333710998296738,
-0.006619592197239399,
-0.04384215921163559,
0.004603104200214148,
0.03843104466795921,
0.028961120173335075,
-0.06357348710298538,
-0.030034903436899185,
0.08228325098752975,
0.018930530175566673,
-0.10049887001514435,
-0.06078094244003296,
0.014710654504597187,
-0.0683865025639534,
-0.05305317044258118,
0.061327945441007614,
-0.03160993754863739,
-0.0449087917804718,
0.07347685843706131,
0.026998115703463554,
0.037702955305576324,
-0.028161220252513885,
-0.006311090663075447,
0.022861821576952934,
-0.021676763892173767,
-0.010231571272015572,
0.0621473491191864,
0.11970509588718414,
0.05060650408267975,
-0.04130364954471588,
0.0020501166582107544,
-0.05318477004766464,
0.01844973675906658,
-0.005920174531638622,
-0.019080014899373055,
-0.06890101730823517,
-0.05052899196743965,
0.0540664941072464,
-0.012071224860846996,
0.01888558641076088,
-0.0053945970721542835,
0.009000984951853752,
0.001363261602818966,
0.05571143329143524,
-0.026404093950986862,
0.006586863659322262,
-0.013077174313366413,
0.010619140230119228,
0.022370923310518265,
0.026059690862894058,
-0.1060895249247551,
0.025893699377775192,
0.05414099618792534,
0.060425907373428345,
0.06884405016899109,
-0.037749018520116806,
-0.009048844687640667,
0.035899896174669266,
-0.040029119700193405,
-0.04431794956326485,
0.07256923615932465,
-0.04963303357362747,
0.11112861335277557,
-0.029103651642799377,
-4.269601561190939e-8,
-0.012441400438547134,
0.06267248839139938,
0.051127202808856964,
-0.03981481492519379,
0.06079595535993576,
-0.06245391070842743,
-0.023251555860042572,
0.04438021406531334,
-0.023043392226099968,
0.06405945122241974,
0.03224626183509827,
-0.024021806195378304,
-0.058289580047130585,
-0.05740819498896599,
-0.015613062307238579,
0.06330865621566772,
-0.019130662083625793,
0.013310965150594711,
0.00003908177313860506,
0.05002414807677269,
-0.032705824822187424,
0.05146076902747154,
-0.00401557981967926,
-0.02322939783334732,
0.025166144594550133,
-0.07912901788949966,
-0.03276357427239418,
0.15233327448368073,
-0.011060462333261967,
0.04936690255999565,
0.003433992387726903,
0.061183393001556396,
-0.0039953626692295074,
-0.009145921096205711,
0.070567287504673,
0.041618023067712784,
-0.005337814800441265,
0.044539641588926315,
-0.03341653198003769,
-0.00985668133944273,
-0.010072646662592888,
0.042800966650247574,
-0.028452295809984207,
0.06008969247341156,
-0.007778904400765896,
-0.030387796461582184,
0.0491274893283844,
-0.054614532738924026,
0.023348933085799217,
0.11100596934556961,
-0.0482693687081337,
-0.016365692019462585,
-0.02168244868516922,
0.08585033565759659,
-0.01817735843360424,
-0.07389586418867111,
0.004697130993008614,
-0.03349296376109123,
0.04545097053050995,
0.09529276937246323,
0.012879408895969391,
0.046669214963912964,
-0.012015982531011105,
-0.021076146513223648
] |
Certifying Language Model Robustness with Fuzzed Randomized Smoothing: An Efficient Defense Against Backdoor Attacks | https://openreview.net/forum?id=USI3ZbuFaV | [
"Bowei He",
"Lihao Yin",
"Huiling Zhen",
"Jianping Zhang",
"Lanqing HONG",
"Mingxuan Yuan",
"Chen Ma"
] | Poster | The widespread deployment of pre-trained language models (PLMs) has exposed them to textual backdoor attacks, particularly those planted during the pre-training stage. These attacks pose significant risks to high-reliability applications, as they can stealthily affect multiple downstream tasks. While certifying robustness against such threats is crucial, existing defenses struggle with the high-dimensional, interdependent nature of textual data and the lack of access to original poisoned pre-training data. To address these challenges, we introduce **F**uzzed **R**andomized **S**moothing (**FRS**), a novel approach for efficiently certifying language model robustness against backdoor attacks. FRS integrates software robustness certification techniques with biphased model parameter smoothing, employing Monte Carlo tree search for proactive fuzzing to identify vulnerable textual segments within the Damerau-Levenshtein space. This allows for targeted and efficient text randomization, while eliminating the need for access to poisoned training data during model smoothing. Our theoretical analysis demonstrates that FRS achieves a broader certified robustness radius compared to existing methods. Extensive experiments across various datasets, model configurations, and attack strategies validate FRS's superiority in terms of defense efficiency, accuracy, and robustness. | Language Model, Textual Backdoor Attack, Certified Robustness, Fuzzed Randomized Smoothing | This paper introduces Fuzzed Randomized Smoothing, a novel defense method that combines randomized smoothing with fuzzing techniques to provide certified robustness against textual backdoor attacks in pre-trained language models. | 2,097 | 2502.06892 | [
-0.08885873109102249,
-0.05837774649262428,
-0.02082240954041481,
0.050968896597623825,
0.06668230891227722,
-0.0019391292007640004,
0.027461936697363853,
-0.012323858216404915,
-0.012972450815141201,
-0.05514544993638992,
-0.029995854943990707,
-0.018698271363973618,
0.11429440230131149,
-0.026628540828824043,
-0.04042000696063042,
-0.030799657106399536,
0.04059027135372162,
0.08659202605485916,
-0.06378073245286942,
-0.018933584913611412,
-0.010530228726565838,
0.07930023223161697,
-0.030953319743275642,
-0.0019100351491943002,
-0.011625210754573345,
-0.040835749357938766,
-0.03916017711162567,
-0.027253858745098114,
-0.026553520932793617,
-0.016975037753582,
0.027704577893018723,
-0.013407552614808083,
0.006064168177545071,
-0.0018174294382333755,
0.04723780229687691,
-0.0014709510141983628,
-0.023962071165442467,
-0.042191050946712494,
0.00108275783713907,
0.008117398247122765,
-0.06498293578624725,
0.014631877653300762,
0.0011609200155362487,
0.016763491556048393,
0.00998004898428917,
-0.07916910201311111,
0.005161306820809841,
-0.018664831295609474,
-0.058747876435518265,
-0.0030950475484132767,
-0.09378241002559662,
-0.038305412977933884,
-0.043183159083127975,
0.03898170590400696,
-0.0031070394907146692,
-0.04626348614692688,
0.008451762609183788,
0.016508664935827255,
-0.0181247740983963,
0.03018568456172943,
-0.07799123972654343,
-0.08606025576591492,
-0.0376407690346241,
0.025944290682673454,
-0.024288136512041092,
-0.003379219677299261,
0.029315005987882614,
0.0470237098634243,
0.02995915152132511,
0.06430133432149887,
-0.04770638421177864,
0.10343754291534424,
-0.07184489071369171,
0.06580458581447601,
-0.007643588352948427,
0.13623183965682983,
-0.026794616132974625,
-0.05575430765748024,
0.061917755752801895,
-0.08450693637132645,
-0.005162573419511318,
-0.015953876078128815,
0.07973020523786545,
-0.013294967822730541,
0.05768347531557083,
0.061900075525045395,
0.030573848634958267,
0.04184486344456673,
0.10459893196821213,
0.0872054323554039,
0.007717888802289963,
-0.03893841803073883,
0.07339426130056381,
0.08174001425504684,
-0.016723353415727615,
0.05506749078631401,
0.016297470778226852,
-0.04526741802692413,
0.006573319435119629,
0.10873381048440933,
-0.03722138702869415,
0.002684702631086111,
0.05152614414691925,
-0.05403557047247887,
0.037743330001831055,
-0.0028028422966599464,
0.07825452089309692,
-0.007714676205068827,
-0.005253923125565052,
-0.08552569150924683,
-0.0088912732899189,
0.05023634806275368,
0.0675082579255104,
-0.032715845853090286,
0.013186577707529068,
-0.00945382658392191,
0.03164193034172058,
-0.05339404195547104,
-0.02510133571922779,
0.042774517089128494,
0.022569281980395317,
0.046344026923179626,
0.043574508279561996,
-0.00640882970765233,
-0.0010308524360880256,
0.012689338065683842,
-0.04646184295415878,
1.4454858422454853e-33,
0.07124713808298111,
0.08785445988178253,
-0.008336580358445644,
0.014934484846889973,
0.041374847292900085,
-0.03757680952548981,
-0.002645375207066536,
0.07461782544851303,
-0.027331825345754623,
0.01754973642528057,
-0.0010583767434582114,
0.0015735211782157421,
-0.019080471247434616,
-0.026136130094528198,
0.04708319157361984,
0.03649879992008209,
-0.024460606276988983,
0.05112844333052635,
-0.08448527008295059,
0.0034425160847604275,
0.08080893009901047,
-0.09729911386966705,
0.02866976521909237,
-0.06582409143447876,
0.020546261221170425,
0.06656942516565323,
0.027682775631546974,
0.018801521509885788,
-0.0002678249147720635,
0.06825640052556992,
-0.08014025539159775,
-0.058007605373859406,
0.015340007841587067,
0.003539724973961711,
0.038760315626859665,
-0.02791295386850834,
-0.050920430570840836,
-0.059144921600818634,
-0.05592536926269531,
-0.01885727420449257,
-0.00526663102209568,
0.0021978800650686026,
0.013818031176924706,
0.011382265947759151,
-0.00876287929713726,
-0.023676127195358276,
-0.060325752943754196,
-0.0018886539619415998,
-0.025536293163895607,
-0.09631635248661041,
0.04419095069169998,
0.024428853765130043,
0.023478202521800995,
-0.0015022874576970935,
-0.11185620725154877,
0.016169840469956398,
0.011205016635358334,
0.007968742400407791,
0.03156162425875664,
0.138984814286232,
-0.006201544310897589,
0.03599221259355545,
-0.005572354421019554,
-0.013487864285707474,
0.001592672080732882,
-0.025555351749062538,
-0.03852633759379387,
0.007808494381606579,
0.06072206050157547,
-0.024922911077737808,
-0.029812326654791832,
-0.05969651788473129,
-0.009302365593612194,
0.0168511513620615,
-0.05828399583697319,
-0.042950280010700226,
0.09616638720035553,
-0.043216876685619354,
0.047481510788202286,
-0.0530804768204689,
0.0236905999481678,
-0.04775495454668999,
0.029075542464852333,
-0.051779262721538544,
-0.0793522298336029,
-0.07989352941513062,
-0.0278374832123518,
-0.05733862519264221,
-0.060690462589263916,
0.02539900317788124,
-0.01740157976746559,
-0.06724357604980469,
-0.000826415722258389,
-0.017061671242117882,
-0.0211056899279356,
-1.9607692092227083e-33,
-0.09668377786874771,
-0.034221675246953964,
-0.07482995837926865,
0.07150813192129135,
-0.08003918826580048,
-0.06733463704586029,
-0.02239103615283966,
0.03523031994700432,
0.047150515019893646,
-0.09226490557193756,
-0.061564866453409195,
-0.052860599011182785,
0.09766194224357605,
-0.05152208358049393,
0.0782667025923729,
-0.015869906172156334,
-0.027602003887295723,
0.020572934299707413,
-0.011894161812961102,
0.11167332530021667,
0.030753755941987038,
0.024576080963015556,
-0.12792149186134338,
0.03977798670530319,
0.03420373797416687,
0.023384112864732742,
-0.05966007336974144,
0.03860556706786156,
0.07500436156988144,
0.07018943876028061,
-0.05561955273151398,
0.06499835848808289,
-0.07139284908771515,
0.005543358623981476,
-0.07603886723518372,
-0.0099465511739254,
0.07011141628026962,
-0.020007947459816933,
0.06610388308763504,
0.057243138551712036,
0.1126105859875679,
0.03280966356396675,
-0.09781335294246674,
0.035919200628995895,
0.00947113148868084,
0.07701296359300613,
-0.04577386751770973,
-0.010575467720627785,
0.06074625626206398,
-0.02813912369310856,
-0.023477919399738312,
-0.004779013805091381,
0.0052156709134578705,
0.05995658040046692,
-0.04777708649635315,
-0.06990800052881241,
-0.05695774033665657,
-0.07702977955341339,
0.00011431827442720532,
0.04244695231318474,
-0.0428779274225235,
-0.030852416530251503,
-0.012467478401958942,
-0.029217295348644257,
0.025131789967417717,
-0.08353657275438309,
-0.02289939858019352,
0.027427511289715767,
0.026444142684340477,
-0.08977260440587997,
0.043414898216724396,
-0.03760335594415665,
-0.04772166907787323,
0.03277638182044029,
0.010446029715240002,
-0.02796235680580139,
-0.039125293493270874,
-0.12241724878549576,
-0.05570783093571663,
0.054292380809783936,
0.054948028177022934,
0.03248940780758858,
-0.001423688605427742,
0.056608255952596664,
0.10475552082061768,
0.04924866184592247,
0.07330300658941269,
0.026661310344934464,
0.0375838577747345,
-0.00684269517660141,
-0.08223572373390198,
0.02738644927740097,
-0.017631912603974342,
0.08871158957481384,
-0.06173204630613327,
-3.8370476573845735e-8,
-0.0005733626312576234,
0.005848236847668886,
-0.014109257608652115,
0.06405658274888992,
0.07482168078422546,
0.04548411816358566,
-0.12944991886615753,
0.01674327626824379,
0.003378107212483883,
-0.05291946232318878,
0.0704147070646286,
-0.011971129104495049,
-0.039118558168411255,
-0.10339200496673584,
-0.03770314157009125,
0.08814097195863724,
-0.10568756610155106,
0.06624044477939606,
-0.06330050528049469,
-0.0007815525750629604,
0.0065344288013875484,
0.08796592056751251,
-0.006311124190688133,
0.04512616991996765,
0.044797904789447784,
-0.055241137742996216,
0.01882934384047985,
0.07723500579595566,
0.06520600616931915,
0.03495596721768379,
-0.035130832344293594,
-0.019684327766299248,
0.003216167213395238,
0.000295269419439137,
-0.05991382896900177,
0.1820135861635208,
0.030647899955511093,
-0.01142861321568489,
-0.0005956272361800075,
0.004556264728307724,
0.0460997074842453,
0.014392618089914322,
-0.0360567532479763,
-0.017714960500597954,
-0.042601052671670914,
-0.01620267890393734,
0.008433934301137924,
-0.07660028338432312,
0.05767300724983215,
-0.019468538463115692,
0.058796804398298264,
-0.03841613233089447,
-0.00170599773991853,
0.0615692213177681,
-0.006040505599230528,
0.05179222300648689,
-0.03400446102023125,
-0.0572691336274147,
-0.018617777153849602,
0.027660483494400978,
0.014557232148945332,
0.0017799168126657605,
0.08601295948028564,
-0.025334402918815613
] |
Kolmogorov-Arnold Transformer | https://openreview.net/forum?id=BCeock53nt | [
"Xingyi Yang",
"Xinchao Wang"
] | Poster | Transformers stand as the cornerstone of mordern deep learning. Traditionally, these models rely on multi-layer
perceptron (MLP) layers to mix the information between channels. In this paper, we introduce the Kolmogorov–Arnold
Transformer (KAT), a novel architecture that replaces MLP layers with Kolmogorov-Arnold Network (KAN) layers to
enhance the expressiveness and performance of the model. Integrating KANs into transformers, however, is no easy
feat, especially when scaled up. Specifically, we identify three key challenges: (C1) Base function. The standard B-spline
function used in KANs is not optimized for parallel computing on modern hardware, resulting in slower inference speeds.
(C2) Parameter and Computation Inefficiency. KAN requires a unique function for each input-output pair, making the
computation extremely large. (C3) Weight initialization. The initialization of weights in KANs is particularly challenging
due to their learnable activation functions, which are critical for achieving convergence in deep neural networks. To
overcome the aforementioned challenges, we propose three key solutions: (S1) Rational basis. We replace B-spline functions
with rational functions to improve compatibility with modern GPUs. By implementing this in CUDA, we achieve faster
computations. (S2) Group KAN. We share the activation weights through a group of neurons, to reduce the computational
load without sacrificing performance. (S3) Variance-preserving initialization. We carefully initialize the activation weights
to make sure that the activation variance is maintained across layers. With these designs, KAT scales effectively and readily
outperforms traditional MLP-based transformers. We demonstrate the advantages of KAT across various tasks, including
image recognition, object detection, and semantic segmentation. It consistently enhances performance over the standard
transformer architectures of different model sizes. | Kolmogorov-Arnold Network; Transformer | In this paper, we introduce the Kolmogorov–Arnold Transformer (KAT), a novel architecture that replaces MLP layers with Kolmogorov-Arnold Network (KAN) layers to enhance the expressiveness and performance. | 2,086 | 2409.10594 | [
-0.12621243298053741,
-0.06864941865205765,
-0.0023733843117952347,
-0.013542828150093555,
-0.08468680083751678,
0.05389139801263809,
-0.012692734599113464,
-0.06981217861175537,
0.037480808794498444,
-0.044626932591199875,
-0.024653242900967598,
-0.05253000184893608,
-0.010506154038012028,
0.014048328623175621,
-0.014687391929328442,
0.055605996400117874,
-0.04539668187499046,
0.09967944771051407,
-0.13852278888225555,
0.013293663039803505,
0.013601697981357574,
0.024839671328663826,
-0.06763407588005066,
0.014240961521863937,
0.09299905598163605,
-0.036920640617609024,
-0.0012865607859566808,
-0.008983904495835304,
-0.004381377715617418,
-0.01022704690694809,
0.056456416845321655,
-0.10139472037553787,
0.010384727269411087,
0.018809519708156586,
-0.03836173936724663,
0.04215915501117706,
0.00976536888629198,
-0.0419103167951107,
-0.015298365615308285,
-0.016330759972333908,
0.007176203187555075,
0.011099551804363728,
-0.01752215437591076,
0.05424780771136284,
0.07651734352111816,
0.052816811949014664,
0.05625329166650772,
-0.0389786958694458,
0.009281421080231667,
-0.02839433029294014,
-0.0062760766595602036,
0.006214923691004515,
-0.016428787261247635,
0.09434942156076431,
-0.02372531220316887,
0.04959835857152939,
-0.03231848031282425,
0.017587540671229362,
-0.025213386863470078,
-0.04072331637144089,
-0.07623712718486786,
-0.07652947306632996,
0.08048432320356369,
-0.0006792659987695515,
-0.0831686332821846,
-0.05184856802225113,
0.03280794247984886,
0.013918200507760048,
0.05850805342197418,
0.015839334577322006,
-0.02339744195342064,
0.02516704611480236,
-0.03936063498258591,
0.0275257620960474,
0.08085410296916962,
-0.024032261222600937,
0.1090744361281395,
-0.021751681342720985,
0.016617905348539352,
-0.07745397090911865,
0.05818025767803192,
0.004980127327144146,
-0.03804169222712517,
-0.004926199093461037,
0.10250144451856613,
0.013871182687580585,
-0.04050484299659729,
0.014546436257660389,
-0.043645311146974564,
0.03905005380511284,
-0.04644928500056267,
0.015812594443559647,
0.038865070790052414,
-0.10864551365375519,
-0.04768753796815872,
-0.0039380257949233055,
-0.050325796008110046,
-0.05029614642262459,
-0.04772993177175522,
0.12389726936817169,
-0.058336492627859116,
-0.016293367370963097,
0.09815404564142227,
-0.02879153937101364,
-0.013313814997673035,
-0.061821725219488144,
0.01513333898037672,
0.030594389885663986,
-0.02815754897892475,
-0.09252291917800903,
0.06192466989159584,
-0.00387683161534369,
0.06434722989797592,
0.008365738205611706,
0.0958448052406311,
0.026795420795679092,
0.02352977730333805,
-0.018390176817774773,
0.020736565813422203,
0.07777638733386993,
-0.05248770862817764,
-0.018785525113344193,
-0.08573052287101746,
0.12708142399787903,
-0.08805453777313232,
0.014568648301064968,
-0.08296516537666321,
3.19451720298091e-33,
-0.031501900404691696,
0.06583705544471741,
0.10585206001996994,
-0.008165079168975353,
-0.02095760405063629,
-0.07069102674722672,
-0.013750017620623112,
-0.03754419833421707,
-0.05667843669652939,
-0.019003819674253464,
-0.10567566752433777,
0.023625899106264114,
-0.04186338186264038,
0.07511623203754425,
0.018058910965919495,
-0.0004725295875687152,
-0.04397905245423317,
0.002142983255907893,
0.013158740475773811,
-0.07018980383872986,
0.061918407678604126,
-0.07469813525676727,
0.0003979166503995657,
-0.030652394518256187,
-0.008687560446560383,
-0.02269105426967144,
0.08509277552366257,
-0.044501107186079025,
-0.03342730551958084,
0.009520999155938625,
-0.0781211107969284,
-0.0015815807273611426,
-0.014623465947806835,
-0.03448662906885147,
-0.009940130636096,
-0.048988040536642075,
0.023589301854372025,
-0.05299673601984978,
-0.017823824658989906,
0.012793848291039467,
0.009342959150671959,
0.03566689044237137,
0.003451298689469695,
0.021353868767619133,
-0.029960613697767258,
0.022823793813586235,
0.06042063981294632,
0.04779725894331932,
0.03443541005253792,
-0.021061042323708534,
0.05478472635149956,
-0.023390162736177444,
-0.028592219576239586,
0.036725759506225586,
0.08816764503717422,
0.03977024927735329,
0.03882671520113945,
0.04262283444404602,
0.050084616988897324,
0.08230594545602798,
-0.03929918259382248,
0.03877568244934082,
-0.030526239424943924,
0.07718603312969208,
-0.0001102555324905552,
-0.012807625345885754,
0.03742941468954086,
0.09027241170406342,
0.004732498899102211,
-0.08897781372070312,
-0.04441870003938675,
0.03074084222316742,
0.07478154450654984,
-0.045846737921237946,
0.021737707778811455,
-0.01383736077696085,
-0.07107647508382797,
-0.11108557134866714,
-0.06344667077064514,
0.08855223655700684,
-0.04228754714131355,
0.054159753024578094,
0.002918656449764967,
0.029153062030673027,
0.037523385137319565,
-0.008486669510602951,
0.08008871972560883,
-0.025506513193249702,
-0.010109837166965008,
0.032686952501535416,
-0.058888982981443405,
-0.008829199708998203,
0.01562131941318512,
-0.03544463962316513,
-0.05210614204406738,
-2.6573074684474932e-33,
-0.032611116766929626,
0.010165177285671234,
-0.0745023861527443,
0.06760341674089432,
-0.038803327828645706,
-0.06053810939192772,
-0.016472460702061653,
-0.05783942714333534,
0.031895291060209274,
-0.004785933066159487,
0.013011137023568153,
-0.05663258209824562,
-0.005926411598920822,
-0.030246585607528687,
0.13191109895706177,
-0.008678337559103966,
-0.061324592679739,
-0.025299502536654472,
0.025290483608841896,
-0.012669449672102928,
0.03024958446621895,
0.025207463651895523,
-0.09029387682676315,
0.04085228964686394,
-0.05609021335840225,
0.01618306152522564,
-0.07084294408559799,
0.12926170229911804,
0.06061865761876106,
-0.00871992763131857,
-0.054897747933864594,
-0.06110699847340584,
-0.016128910705447197,
-0.02177661843597889,
-0.007225808687508106,
0.00985165499150753,
0.011999410577118397,
0.04573763161897659,
-0.011090326122939587,
-0.021254373714327812,
0.03105044923722744,
-0.04733619466423988,
-0.04717603698372841,
-0.013237064704298973,
-0.022786814719438553,
-0.016201863065361977,
-0.006222050171345472,
-0.035643063485622406,
-0.004353654570877552,
0.03757680207490921,
0.00260232319124043,
0.04769781231880188,
-0.07538844645023346,
0.024776415899395943,
-0.008410723879933357,
-0.01836824230849743,
-0.02967733144760132,
0.05593016743659973,
0.06230858713388443,
-0.011415381915867329,
-0.062285780906677246,
-0.08980592340230942,
0.04705656319856644,
-0.0608682781457901,
-0.004951808601617813,
-0.001317778485827148,
0.0545877143740654,
0.022251449525356293,
0.029297009110450745,
0.004895532503724098,
-0.04528083652257919,
-0.029046066105365753,
0.03671683371067047,
0.013535370118916035,
-0.1147901639342308,
-0.024101190268993378,
-0.01483622845262289,
-0.06015152111649513,
-0.007842515595257282,
0.019242573529481888,
-0.016990065574645996,
-0.041404079645872116,
0.016828179359436035,
0.09851746261119843,
0.12049975246191025,
0.1318797618150711,
0.07238239049911499,
0.01150126475840807,
0.053036730736494064,
-0.01294758077710867,
0.022529494017362595,
0.1296408772468567,
-0.00368660781532526,
0.06712733954191208,
-0.05278594046831131,
-4.683009890982248e-8,
-0.03462618216872215,
0.03788311779499054,
-0.011839781887829304,
-0.0359472930431366,
0.07677090913057327,
-0.027650136500597,
0.0715804472565651,
0.02281387336552143,
-0.04056788980960846,
-0.06287536770105362,
0.050671301782131195,
0.00621004169806838,
0.030097752809524536,
-0.03056800737977028,
0.02418874017894268,
0.08420317620038986,
0.033850833773612976,
-0.03324718028306961,
0.0021030886564403772,
-0.07765810191631317,
-0.09935692697763443,
0.09437499940395355,
-0.05715765804052353,
-0.031162245199084282,
-0.01793510466814041,
-0.07231345772743225,
-0.02510875090956688,
0.09060980379581451,
0.09031114727258682,
0.01002900768071413,
-0.04467643052339554,
-0.018927941098809242,
-0.006319024134427309,
0.035910751670598984,
0.04352963715791702,
0.0484979972243309,
0.06426981836557388,
0.04996176064014435,
-0.024063406512141228,
0.025084687396883965,
0.05600622668862343,
0.04122928902506828,
0.00932940375059843,
0.007393692620098591,
0.01759028062224388,
0.01053351815789938,
-0.01858731545507908,
-0.06512132287025452,
0.011405153200030327,
0.0736841931939125,
0.06897160410881042,
0.06501273810863495,
-0.006382754072546959,
0.06742986291646957,
0.026360919699072838,
0.023456193506717682,
-0.012466349638998508,
-0.1008380576968193,
-0.013186769559979439,
0.14105285704135895,
-0.06121867150068283,
0.013828851282596588,
-0.02459004335105419,
0.015658007934689522
] |
BrainUICL: An Unsupervised Individual Continual Learning Framework for EEG Applications | https://openreview.net/forum?id=6jjAYmppGQ | [
"Yangxuan Zhou",
"Sha Zhao",
"Jiquan Wang",
"Haiteng Jiang",
"Shijian Li",
"Tao Li",
"Gang Pan"
] | Poster | Electroencephalography (EEG) is a non-invasive brain-computer interface technology used for recording brain electrical activity. It plays an important role in human life and has been widely uesd in real life, including sleep staging, emotion recognition, and motor imagery. However, existing EEG-related models cannot be well applied in practice, especially in clinical settings, where new patients with individual discrepancies appear every day. Such EEG-based model trained on fixed datasets cannot generalize well to the continual flow of numerous unseen subjects in real-world scenarios. This limitation can be addressed through continual learning (CL), wherein the CL model can continuously learn and advance over time. Inspired by CL, we introduce a novel Unsupervised Individual Continual Learning paradigm for handling this issue in practice. We propose the BrainUICL framework, which enables the EEG-based model to continuously adapt to the incoming new subjects. Simultaneously, BrainUICL helps the model absorb new knowledge during each adaptation, thereby advancing its generalization ability for all unseen subjects. The effectiveness of the proposed BrainUICL has been evaluated on three different mainstream EEG tasks. The BrainUICL can effectively balance both the plasticity and stability during CL, achieving better plasticity on new individuals and better stability across all the unseen individuals, which holds significance in a practical setting. | Continual Learning; EEG Applications | null | 2,075 | null | [
0.006236464250832796,
-0.013175010681152344,
0.07701136916875839,
0.04811934381723404,
0.04271011799573898,
-0.024659816175699234,
0.019156109541654587,
-0.07821956276893616,
0.045210111886262894,
-0.07504494488239288,
0.007365727331489325,
-0.029363222420215607,
-0.03036605380475521,
-0.020273420959711075,
-0.03196096420288086,
-0.021711310371756554,
-0.033292341977357864,
0.021169373765587807,
-0.023151349276304245,
-0.01879253052175045,
0.019013691693544388,
0.02124548703432083,
0.013806345872581005,
0.06234263628721237,
-0.03570574149489403,
0.05882500484585762,
-0.006640451028943062,
-0.06088576093316078,
0.10260950773954391,
-0.024634258821606636,
0.061984650790691376,
0.05855153128504753,
-0.03863631561398506,
0.036149267107248306,
-0.12385360151529312,
-0.023667728528380394,
-0.0165424682199955,
0.06037405505776405,
-0.04796585068106651,
-0.024096908047795296,
0.03684264421463013,
-0.007118314038962126,
-0.009293044917285442,
-0.018065722659230232,
0.06036033108830452,
0.007107118144631386,
0.0034120504278689623,
-0.14072474837303162,
0.06925269961357117,
0.015964217483997345,
-0.02280343323945999,
-0.033153485506772995,
0.03682054951786995,
0.0935133844614029,
0.06975259631872177,
0.03073529712855816,
-0.041088663041591644,
0.019029052928090096,
-0.05135859176516533,
0.03539552167057991,
-0.03582281246781349,
0.004397616721689701,
0.0064169843681156635,
-0.023141048848628998,
-0.050674740225076675,
0.07677434384822845,
-0.0013533029705286026,
0.013940488919615746,
0.03339347243309021,
0.0266279224306345,
-0.05138138681650162,
0.005037638824433088,
-0.027218207716941833,
0.004666556604206562,
0.03409350663423538,
0.022285768762230873,
-0.03501256927847862,
-0.0908365547657013,
0.10438311100006104,
0.005379657726734877,
0.014092838391661644,
0.02636018954217434,
0.0059815663844347,
-0.08087919652462006,
0.1286177635192871,
0.01751408539712429,
0.10264969617128372,
0.01621166802942753,
0.013365385122597218,
0.02498210407793522,
0.004105303902179003,
-0.028018636628985405,
0.04414055868983269,
-0.026556529104709625,
0.033397164195775986,
0.037078652530908585,
-0.010725054889917374,
-0.023564046248793602,
-0.035715632140636444,
0.04188500717282295,
-0.0048716929741203785,
0.048574868589639664,
-0.046744730323553085,
0.009822453372180462,
-0.07418273389339447,
-0.06833130866289139,
-0.0097556933760643,
0.045356761664152145,
0.06318321079015732,
-0.01543646864593029,
0.018847856670618057,
0.0006901994347572327,
-0.02627948857843876,
0.010727492161095142,
0.03432859107851982,
0.004176371730864048,
-0.030885864049196243,
0.10352973639965057,
0.04311632737517357,
0.11425229161977768,
0.0430813692510128,
0.02034619264304638,
-0.004369014874100685,
-0.0007639594841748476,
0.08323591947555542,
-0.00874294526875019,
-0.06512519717216492,
4.9865847894323806e-33,
-0.012719455175101757,
-0.03145957365632057,
0.02412029355764389,
0.03460753336548805,
-0.03290018439292908,
0.00212112651206553,
-0.032905176281929016,
0.09752757102251053,
0.004879445768892765,
0.011786519549787045,
-0.030525080859661102,
0.06626089662313461,
-0.03404855728149414,
0.08657819777727127,
0.058336034417152405,
-0.04936158284544945,
-0.10522586852312088,
0.029199251905083656,
0.04557592421770096,
-0.08887262642383575,
0.09277433902025223,
-0.014198875986039639,
-0.0004012082063127309,
-0.07896003127098083,
-0.01763126440346241,
0.012888087891042233,
0.02237403392791748,
0.0552804172039032,
0.006104104686528444,
0.03331220522522926,
-0.07865357398986816,
0.02933664247393608,
-0.0054865386337041855,
-0.0014688465744256973,
-0.011667486280202866,
-0.04059228301048279,
0.015632541850209236,
-0.017400050535798073,
0.077266626060009,
-0.021601514890789986,
0.0009958504233509302,
0.007645485922694206,
-0.04512004554271698,
-0.013408324681222439,
-0.06665972620248795,
-0.04852315038442612,
0.1016874760389328,
0.04905465245246887,
-0.04417070746421814,
-0.021354438737034798,
-0.05108068138360977,
-0.014049786143004894,
-0.09102672338485718,
-0.06476493179798126,
-0.10380791127681732,
0.03381144627928734,
-0.0426543690264225,
0.027578994631767273,
0.049385279417037964,
-0.014711344614624977,
0.026622043922543526,
0.07026582956314087,
-0.023641357198357582,
0.08354473114013672,
0.05872591957449913,
0.09091342985630035,
-0.023468228057026863,
0.05977848172187805,
-0.0077680000104010105,
-0.07651478052139282,
0.03539770841598511,
-0.013246450573205948,
0.008581535890698433,
-0.09408567100763321,
0.03971342369914055,
-0.07393301278352737,
0.02289230190217495,
-0.014602657407522202,
-0.023614702746272087,
0.023531971499323845,
0.060463909059762955,
-0.049577172845602036,
-0.08819502592086792,
-0.02159525640308857,
0.04589937627315521,
0.01884143240749836,
0.05073034390807152,
-0.03712800145149231,
-0.12991458177566528,
0.0051019624806940556,
-0.006270830985158682,
-0.022752637043595314,
0.05050244927406311,
-0.034503281116485596,
-0.028031187132000923,
-4.620107160490255e-33,
0.059072062373161316,
-0.045528437942266464,
-0.011725107207894325,
0.0012046302435919642,
0.06303389370441437,
-0.035241398960351944,
0.006952481810003519,
0.006520227529108524,
-0.0682859867811203,
-0.08167047798633575,
0.024800026789307594,
-0.05892990157008171,
0.04244484007358551,
0.05031494423747063,
-0.06601112335920334,
0.005029757507145405,
-0.11154042929410934,
-0.01346286665648222,
-0.03373190015554428,
0.09720401465892792,
-0.05297994241118431,
0.0026371695566922426,
-0.040345966815948486,
-0.05375058203935623,
-0.025371046736836433,
0.08012381941080093,
0.011341639794409275,
0.11398310214281082,
0.017242684960365295,
-0.018842071294784546,
-0.1068408265709877,
-0.040334302932024,
-0.0679573118686676,
-0.14707189798355103,
-0.026832742616534233,
0.07305334508419037,
0.05530231073498726,
-0.0002845328126568347,
-0.029029490426182747,
0.039624638855457306,
0.052701905369758606,
-0.02155257575213909,
-0.007282075006514788,
-0.021890789270401,
0.06572292000055313,
-0.020761337131261826,
-0.08557869493961334,
0.044109541922807693,
-0.03824405372142792,
0.08625642955303192,
0.012874403037130833,
0.04298119246959686,
-0.12314410507678986,
-0.06173764169216156,
-0.09499851614236832,
-0.031066520139575005,
-0.04134391248226166,
-0.005983585026115179,
-0.01326735969632864,
-0.06369855999946594,
-0.04953066259622574,
-0.06674563884735107,
0.010154206305742264,
0.04434235766530037,
0.014676488004624844,
0.03041713312268257,
0.015605696476995945,
0.06020490452647209,
-0.028733745217323303,
-0.054066337645053864,
-0.0019550553988665342,
0.03794092312455177,
-0.0015226479154080153,
-0.11471808701753616,
0.012127459980547428,
0.04922853782773018,
-0.02798611670732498,
-0.09266526997089386,
-0.02272331528365612,
-0.022832931950688362,
-0.039086077362298965,
-0.034085679799318314,
-0.06194101646542549,
0.06419206410646439,
0.005862879101186991,
0.02995348908007145,
0.015939051285386086,
0.08728284388780594,
0.0577147901058197,
0.02507752552628517,
-0.07760336250066757,
0.03477854281663895,
0.022313978523015976,
0.044608935713768005,
0.003153821686282754,
-4.429004363259992e-8,
-0.10572772473096848,
-0.03255505859851837,
0.044771432876586914,
0.03188090771436691,
0.13362640142440796,
-0.11182579398155212,
-0.005904742516577244,
0.01554721686989069,
-0.053369246423244476,
-0.032322585582733154,
0.12352750450372696,
-0.003650830127298832,
0.029749998822808266,
0.005977740976959467,
0.013489597477018833,
0.06332423537969589,
0.09585213661193848,
-0.013884341344237328,
-0.002228817204013467,
-0.09530080854892731,
0.033327724784612656,
-0.061561647802591324,
0.05897512659430504,
-0.018184950575232506,
0.02757415920495987,
-0.025750165805220604,
-0.005789448041468859,
0.03328939899802208,
-0.06378334760665894,
-0.0657167062163353,
0.011395799927413464,
-0.021616945043206215,
0.05078530311584473,
-0.002944422885775566,
0.018594441935420036,
0.0031964851077646017,
0.04907870292663574,
-0.03277333080768585,
-0.03992033377289772,
-0.053450897336006165,
-0.05485619604587555,
0.012299096211791039,
-0.01964745856821537,
0.03343226760625839,
0.007937302812933922,
-0.01951874978840351,
0.02557883784174919,
-0.04142424464225769,
-0.02099275030195713,
0.03173314779996872,
0.018738102167844772,
-0.02706655114889145,
0.05621270090341568,
0.09768392145633698,
0.158127561211586,
0.02627306990325451,
0.03128870204091072,
0.042230941355228424,
-0.04608798027038574,
0.06009679660201073,
0.028592068701982498,
0.021560322493314743,
-0.07129909843206406,
0.042803820222616196
] |
Oryx MLLM: On-Demand Spatial-Temporal Understanding at Arbitrary Resolution | https://openreview.net/forum?id=ODiY6pbHZQ | [
"Zuyan Liu",
"Yuhao Dong",
"Ziwei Liu",
"Winston Hu",
"Jiwen Lu",
"Yongming Rao"
] | Poster | Visual data comes in various forms, ranging from small icons of just a few pixels to long videos spanning hours. Existing multi-modal LLMs usually standardize these diverse visual inputs to fixed-resolution images or patches for visual encoders and yield similar numbers of tokens for LLMs. This approach is non-optimal for multimodal understanding and inefficient for processing inputs with long and short visual contents. To solve the problem, we propose Oryx, a unified multimodal architecture for the spatial-temporal understanding of images, videos, and multi-view 3D scenes. Oryx offers an on-demand solution to seamlessly and efficiently process visual inputs with arbitrary spatial sizes and temporal lengths through two core innovations: 1) a pre-trained OryxViT model that can encode images at any resolution into LLM-friendly visual representations; 2) a dynamic compressor module that supports 1x to 16x compression on visual tokens by request. These designs enable Oryx to accommodate extremely long visual contexts, such as videos, with lower resolution and high compression while maintaining high recognition precision for tasks like document understanding with native resolution and no compression. Beyond the architectural improvements, enhanced data curation and specialized training on long-context retrieval and spatial-aware data help Oryx achieve strong capabilities in image, video, and 3D multimodal understanding simultaneously. | Multi-modal Large Language Model, Multi-Modal Understanding, Arbitary Resolution | null | 2,074 | 2409.12961 | [
0.01280768122524023,
-0.06307286024093628,
0.033811505883932114,
-0.045700520277023315,
0.10921765118837357,
-0.05020911619067192,
-0.09616159647703171,
-0.048796121031045914,
0.02899295464158058,
-0.01996731571853161,
-0.047237612307071686,
-0.02234172262251377,
-0.0036135984119027853,
0.06825581192970276,
0.01709466613829136,
0.03901910409331322,
0.020584437996149063,
0.11582706868648529,
-0.11992853134870529,
0.018225133419036865,
0.06044250726699829,
-0.05622962862253189,
0.030906112864613533,
-0.0025376109406352043,
0.06428524106740952,
0.05083684250712395,
0.02394258789718151,
-0.027541721239686012,
0.08131849765777588,
-0.04291335865855217,
0.02610182948410511,
0.01949770376086235,
0.002546686679124832,
0.08282501995563507,
-0.09257784485816956,
0.045456334948539734,
0.0072447145357728004,
-0.013047694228589535,
-0.1200786828994751,
-0.09835387766361237,
-0.028971387073397636,
0.0289764367043972,
0.006019986234605312,
-0.01675424538552761,
0.029649781063199043,
-0.004100636579096317,
0.031123578548431396,
-0.08404964953660965,
0.004021106753498316,
0.03301268443465233,
-0.08647218346595764,
-0.07484009861946106,
-0.02068803273141384,
0.004181721713393927,
-0.057995498180389404,
0.06806281208992004,
0.0006308552110567689,
-0.0539616234600544,
-0.028142957016825676,
0.010351564735174179,
-0.01289311982691288,
-0.018636178225278854,
0.02324536070227623,
0.04876982420682907,
-0.03665821999311447,
-0.03196493163704872,
0.043657105416059494,
-0.03418188542127609,
0.009928598999977112,
-0.05134683474898338,
-0.029276352375745773,
-0.053029149770736694,
0.007880334742367268,
0.006445427890866995,
-0.0318872444331646,
0.024038581177592278,
0.0365886352956295,
-0.023745039477944374,
0.013811271637678146,
-0.07691266387701035,
0.11894150823354721,
0.024782495573163033,
-0.006482776720076799,
0.04601067304611206,
0.04460982605814934,
-0.03480997681617737,
0.003638543887063861,
0.095176100730896,
-0.011008606292307377,
0.0027392359916120768,
-0.02172721177339554,
-0.04802941903471947,
0.048495229333639145,
-0.04021509364247322,
0.060039401054382324,
-0.042432162910699844,
0.10110913962125778,
-0.07117876410484314,
0.00362695031799376,
0.06704986840486526,
0.0028922755736857653,
-0.0030039744451642036,
-0.009207621216773987,
0.002455282723531127,
0.050913095474243164,
-0.01845468208193779,
0.09973150491714478,
0.05175095796585083,
-0.004747648257762194,
-0.002188669051975012,
0.046868178993463516,
-0.03643692657351494,
0.00685448432341218,
-0.016090286895632744,
0.023645548149943352,
0.0009560599573887885,
-0.013977188616991043,
0.02695329301059246,
0.11050636321306229,
0.011034952476620674,
-0.002869036514312029,
-0.016973068937659264,
-0.12326386570930481,
0.006157375406473875,
0.07938839495182037,
-0.0053433640860021114,
-0.03603142872452736,
2.126991824610657e-33,
0.03551463782787323,
0.057071536779403687,
0.01329721324145794,
0.05127179995179176,
0.03719763830304146,
0.05598282068967819,
0.013012714684009552,
0.010570675134658813,
0.0220012329518795,
0.00585662666708231,
0.002291082637384534,
-0.03620343282818794,
-0.034560322761535645,
0.16449478268623352,
-0.0033455321099609137,
-0.06083546578884125,
-0.004000268876552582,
0.032741956412792206,
-0.010190188884735107,
0.045051876455545425,
0.004972320050001144,
-0.033032435923814774,
-0.022463537752628326,
0.03492533415555954,
-0.017781421542167664,
0.01412529218941927,
0.05734886974096298,
-0.045765843242406845,
0.015692513436079025,
-0.0461820624768734,
-0.05486712232232094,
-0.005257716868072748,
-0.003722947556525469,
-0.03846201300621033,
0.03635173663496971,
-0.04451045021414757,
-0.034346941858530045,
-0.00887115579098463,
-0.04079112783074379,
-0.03564882650971413,
-0.03895464166998863,
0.033398423343896866,
-0.06740961968898773,
-0.09761466085910797,
-0.10738345235586166,
-0.006241115741431713,
0.025770211592316628,
0.14159950613975525,
-0.11296513676643372,
-0.005096058361232281,
0.09722919017076492,
0.07578559219837189,
-0.06740805506706238,
-0.0775643065571785,
-0.03209971636533737,
-0.051383089274168015,
0.045731425285339355,
0.017130356281995773,
0.0711941346526146,
-0.013822093605995178,
-0.02999248169362545,
-0.004051573574542999,
0.01702810823917389,
0.0842643454670906,
0.019249293953180313,
-0.011552615091204643,
0.045836325734853745,
0.09207955747842789,
0.025306155905127525,
0.009608950465917587,
-0.037967003881931305,
0.04128884896636009,
0.07178543508052826,
-0.07938915491104126,
-0.03570646792650223,
-0.012180816382169724,
0.00432058610022068,
-0.07672236859798431,
0.0075295125134289265,
0.07032086700201035,
-0.03303554654121399,
-0.009314946830272675,
0.03738943114876747,
-0.00928696058690548,
0.048658668994903564,
-0.048800498247146606,
0.017611294984817505,
-0.06907716393470764,
0.0041718860156834126,
-0.10332566499710083,
-0.04524783417582512,
-0.04640065133571625,
0.009432227350771427,
-0.0409010611474514,
-0.005107278004288673,
-1.1026553943865614e-33,
0.06640221923589706,
-0.00885382667183876,
-0.035882368683815,
0.047431234270334244,
0.045201625674963,
-0.02692704275250435,
-0.015563399530947208,
0.06178921088576317,
-0.03144650161266327,
-0.06004635989665985,
-0.05407986789941788,
-0.028156017884612083,
-0.018738949671387672,
-0.07085376977920532,
0.0055623226799070835,
0.023833753541111946,
0.025782635435461998,
-0.04910138249397278,
-0.050143130123615265,
-0.0026379874907433987,
0.005397847853600979,
0.05980602279305458,
-0.02487807348370552,
-0.020765548571944237,
-0.06149386987090111,
0.08950996398925781,
-0.013956639915704727,
0.007437046151608229,
-0.06363993138074875,
0.04035606235265732,
0.036640990525484085,
-0.036795005202293396,
0.07364430278539658,
-0.05112338811159134,
-0.00968928076326847,
0.03060116246342659,
0.08077698945999146,
-0.029834305867552757,
-0.044536106288433075,
0.05830583721399307,
0.03203899413347244,
-0.07345551997423172,
-0.016557538881897926,
0.015649784356355667,
0.019033143296837807,
-0.0113249272108078,
-0.05604972690343857,
0.0064763412810862064,
0.03745004162192345,
0.06696508079767227,
-0.0567907877266407,
0.04248066991567612,
-0.07489056885242462,
-0.060271237045526505,
0.060143981128931046,
-0.10331542044878006,
-0.10867025703191757,
-0.027852069586515427,
0.10800958424806595,
-0.050564851611852646,
-0.03235302120447159,
-0.10972476005554199,
-0.07915118336677551,
0.016439080238342285,
-0.024677664041519165,
0.0181121863424778,
-0.03785581886768341,
-0.10722365975379944,
-0.014053883962333202,
0.05357775837182999,
0.003560124197974801,
-0.03735434636473656,
-0.04102803394198418,
-0.004600778222084045,
0.02654842846095562,
0.027327515184879303,
0.02251405268907547,
-0.021363580599427223,
0.0499795563519001,
-0.029589885845780373,
-0.029246173799037933,
-0.06048399209976196,
0.03640259802341461,
0.14224675297737122,
0.07458669692277908,
0.0870136022567749,
0.009725655429065228,
0.06892392039299011,
0.04416317865252495,
0.03009078837931156,
-0.00694731529802084,
0.07431859523057938,
0.019481472671031952,
0.06667503714561462,
0.044528208673000336,
-3.9128025264290045e-8,
-0.049490440636873245,
0.019431939348578453,
-0.0488341748714447,
-0.0033735332544893026,
-0.016147678717970848,
-0.0391424186527729,
0.039425935596227646,
0.08656149357557297,
0.06410182267427444,
-0.01474503893405199,
0.12450049817562103,
-0.08703815191984177,
-0.04562227055430412,
-0.020854279398918152,
0.049408622086048126,
0.0431097038090229,
0.06100470572710037,
-0.003245249856263399,
-0.013318256475031376,
0.010263362899422646,
0.009480875916779041,
0.01647830381989479,
-0.016203084960579872,
-0.024987170472741127,
-0.012025764212012291,
-0.039408307522535324,
-0.036721039563417435,
0.14808882772922516,
0.0583009272813797,
-0.0484476201236248,
-0.01567348837852478,
0.029320042580366135,
0.02370249293744564,
-0.051011618226766586,
-0.011447173543274403,
-0.054533302783966064,
0.05752452090382576,
0.019928839057683945,
0.01860833540558815,
-0.09120561182498932,
-0.014568597078323364,
0.012341972440481186,
-0.0409354493021965,
0.0021703944075852633,
0.01953764632344246,
0.009124333038926125,
0.04215544834733009,
-0.10187121480703354,
-0.06544747203588486,
0.03017829917371273,
-0.06017626076936722,
0.008670647628605366,
0.06011711806058884,
0.08613844960927963,
0.11596491932868958,
0.017368735745549202,
0.046437423676252365,
0.07257381826639175,
0.05410657450556755,
0.05119042098522186,
-0.028001433238387108,
-0.0027879304252564907,
-0.062152035534381866,
-0.051239676773548126
] |
Is In-Context Learning Sufficient for Instruction Following in LLMs? | https://openreview.net/forum?id=STEEDDv3zI | [
"Hao Zhao",
"Maksym Andriushchenko",
"Francesco Croce",
"Nicolas Flammarion"
] | Poster | In-context learning (ICL) allows LLMs to learn from examples without changing their weights: this is a particularly promising capability for long-context LLMs that can potentially learn from many examples. Recently, Lin et al. (2024) proposed URIAL, a method using only three in-context examples to align base LLMs, achieving non-trivial instruction following performance. In this work, we show that, while effective, ICL alignment with URIAL still underperforms compared to instruction fine-tuning on established benchmarks such as MT-Bench and AlpacaEval 2.0 (LC), especially with more capable base LLMs. We then uncover the most relevant elements for successful in-context alignment, finding the crucial role of the decoding parameters. Based on these insights, we show that the approach of URIAL can indeed be improved by adding more, potentially carefully selected, high-quality demonstrations in context, getting closer to the performance of instruct models. Finally, we provide the first, to our knowledge, systematic comparison of ICL and instruction fine-tuning (IFT) for instruction following in the low data regime. Overall, our work advances the understanding of ICL as an alignment technique and its relationship to IFT. | Large Language Models, In-context Learning, Alignment | null | 2,072 | 2405.19874 | [
-0.014032121747732162,
-0.06289175897836685,
0.013617630116641521,
-0.05692094936966896,
-0.0060784355737268925,
-0.08561732620000839,
-0.029981043189764023,
0.017720118165016174,
0.035159632563591,
-0.020196329802274704,
-0.04136631637811661,
-0.018293509259819984,
0.031197145581245422,
-0.03330918401479721,
0.04097786173224449,
-0.034258484840393066,
0.06931418925523758,
0.10362351685762405,
-0.09962929040193558,
-0.04813868924975395,
0.07325399667024612,
0.04015348479151726,
0.050633106380701065,
-0.011759390123188496,
0.014470268040895462,
0.057744547724723816,
-0.0331679992377758,
0.0263616181910038,
0.10616041719913483,
-0.05885039642453194,
0.02586224488914013,
0.098879374563694,
-0.01476222276687622,
0.04977652430534363,
0.004759443458169699,
0.1066865548491478,
-0.0335395373404026,
-0.026515919715166092,
-0.01944330707192421,
-0.016112100332975388,
-0.03256813809275627,
0.022297093644738197,
0.05196690931916237,
0.02822001278400421,
0.06428728252649307,
-0.007093577645719051,
0.007630699779838324,
-0.02660306729376316,
-0.049381744116544724,
-0.001096361898817122,
-0.09747297316789627,
0.016677727922797203,
-0.04761762544512749,
-0.02992182970046997,
-0.05174245685338974,
0.06756790727376938,
0.056003838777542114,
0.013315737247467041,
0.028271036222577095,
-0.003865367965772748,
-0.039829693734645844,
-0.08310997486114502,
-0.04892491176724434,
0.05108911916613579,
-0.03582776337862015,
-0.023639123886823654,
0.03514990210533142,
0.03342925012111664,
0.01696394942700863,
-0.02722037024796009,
-0.10038019716739655,
-0.0323164165019989,
-0.03128618374466896,
0.14439760148525238,
-0.028236765414476395,
0.06494513899087906,
0.10426905751228333,
-0.03545614331960678,
0.03383456915616989,
-0.07793208956718445,
0.030595475807785988,
-0.004369278904050589,
0.03410704806447029,
-0.014655078761279583,
0.06826981902122498,
-0.0487595871090889,
0.06679178774356842,
0.014400499872863293,
0.046097349375486374,
0.04386042430996895,
0.032565753906965256,
-0.08749750256538391,
-0.02950982376933098,
0.016293346881866455,
0.07315181195735931,
0.038551829755306244,
0.058725789189338684,
-0.06158611923456192,
-0.07000569999217987,
0.08996473252773285,
0.019791726022958755,
0.07962127774953842,
-0.013178831897675991,
-0.0730057954788208,
-0.056661006063222885,
-0.03716807812452316,
0.024776993319392204,
0.07346970587968826,
-0.026609521359205246,
-0.10921406000852585,
0.02817581035196781,
0.04649053141474724,
0.014685744419693947,
0.038580622524023056,
0.022920183837413788,
-0.08113347738981247,
0.04612123966217041,
-0.13348722457885742,
0.03765735402703285,
0.01873978227376938,
-0.0560501292347908,
-0.013126343488693237,
-0.054714541882276535,
0.027736030519008636,
-0.0834854319691658,
-0.09806519746780396,
-0.05092533305287361,
1.3189330387909194e-33,
0.060073696076869965,
0.07538054883480072,
0.0243437010794878,
0.019427714869379997,
0.03622497245669365,
-0.01590893045067787,
0.06352999061346054,
0.06762292981147766,
-0.0342136025428772,
0.0646105483174324,
0.033244259655475616,
0.007783129811286926,
-0.008133632130920887,
0.026373406872153282,
0.022470908239483833,
-0.023295434191823006,
-0.05223018676042557,
0.04301191121339798,
-0.08479228615760803,
0.01652715727686882,
0.01647760346531868,
-0.0032648369669914246,
-0.003402300411835313,
-0.08799371123313904,
0.05588291212916374,
0.05010871961712837,
0.05872082710266113,
-0.04783942177891731,
-0.05173880606889725,
0.012607389129698277,
-0.07681463658809662,
-0.0691128745675087,
-0.07933410257101059,
0.043850596994161606,
0.08748029917478561,
-0.06293556094169617,
0.01390279270708561,
-0.035620108246803284,
-0.022648068144917488,
-0.045212287455797195,
0.00012273230822756886,
0.04287215322256088,
0.04419581964612007,
0.004504199605435133,
-0.10164520144462585,
-0.09857238829135895,
-0.005748949013650417,
-0.0189803596585989,
0.026029616594314575,
-0.03956263139843941,
0.06090162694454193,
0.04464391991496086,
-0.06713098287582397,
-0.07604078948497772,
0.007860179990530014,
0.005885271355509758,
0.012282048352062702,
0.02962956391274929,
0.029661504551768303,
0.1381392478942871,
-0.029065843671560287,
0.039317719638347626,
-0.04518004134297371,
0.1210022121667862,
0.036283351480960846,
0.0034179305657744408,
-0.053600020706653595,
0.045997828245162964,
0.031316619366407394,
-0.06422140449285507,
-0.037518586963415146,
-0.012399856932461262,
0.03460152819752693,
-0.006530069746077061,
0.05943595618009567,
-0.007313182111829519,
0.0031327451579272747,
-0.08821635693311691,
-0.02388070337474346,
0.006381611339747906,
0.09585077315568924,
0.10855165868997574,
0.018434373661875725,
-0.008388334885239601,
-0.045716140419244766,
-0.0889676958322525,
0.027635913342237473,
-0.08541591465473175,
-0.041763607412576675,
-0.008762054145336151,
-0.017975283786654472,
-0.02795233763754368,
-0.046347059309482574,
0.03815709054470062,
-0.022239230573177338,
-1.619853806965167e-33,
0.03555493801832199,
-0.023910831660032272,
0.034441810101270676,
0.06632678955793381,
-0.04213258996605873,
-0.03955848887562752,
0.06002497673034668,
0.008803529664874077,
-0.03681259602308273,
-0.038865335285663605,
0.002570126671344042,
0.00547584006562829,
-0.01655714213848114,
-0.01444877590984106,
0.014825855381786823,
-0.04388437047600746,
-0.034906212240457535,
-0.009336505085229874,
0.0582086518406868,
0.042661912739276886,
0.038837745785713196,
0.06806547939777374,
-0.10025269538164139,
-0.010917594656348228,
-0.0244592297822237,
0.02840406633913517,
-0.0552731491625309,
0.031552791595458984,
0.033621180802583694,
0.032942432910203934,
0.032036904245615005,
0.03849954158067703,
0.00809771940112114,
0.039014339447021484,
-0.05965833365917206,
0.06561292707920074,
0.05120210722088814,
0.06121546030044556,
-0.03717007115483284,
0.026042867451906204,
0.11182007193565369,
0.023008553311228752,
-0.023168185725808144,
-0.008428279310464859,
0.04091468080878258,
-0.018317686393857002,
-0.04967004805803299,
-0.06212077662348747,
0.001540978206321597,
0.02269265614449978,
-0.02484963834285736,
0.008153950795531273,
-0.029951494187116623,
-0.02416880987584591,
-0.012882526963949203,
-0.06115161255002022,
-0.04127775505185127,
-0.07987894862890244,
0.018175968900322914,
-0.06284525990486145,
0.010336189530789852,
-0.010547949001193047,
0.04905128851532936,
-0.09562619030475616,
0.06685583293437958,
0.0015460680006071925,
-0.07370662689208984,
-0.008686347864568233,
-0.06789356470108032,
-0.018302693963050842,
0.00975106842815876,
-0.05208476632833481,
-0.05675318092107773,
-0.05213528871536255,
0.011425548233091831,
-0.045002225786447525,
0.0423971451818943,
-0.08830747753381729,
-0.02848402038216591,
-0.020595623180270195,
0.006339948158711195,
-0.0009166361414827406,
-0.05558202415704727,
0.08981475234031677,
-0.03258250653743744,
0.09167567640542984,
0.07682392001152039,
0.09110807627439499,
-0.02733134850859642,
0.004985589534044266,
-0.03734500706195831,
0.06384771317243576,
0.03708455339074135,
0.040877122431993484,
-0.06688369810581207,
-4.5841133555768465e-8,
-0.09695611149072647,
0.01749642752110958,
0.01539347693324089,
0.06962981820106506,
0.014289312064647675,
0.049893781542778015,
-0.03555269166827202,
0.07078167051076889,
0.03943930193781853,
0.003469988005235791,
0.043107811361551285,
-0.03228120505809784,
-0.04489254206418991,
-0.029613088816404343,
0.029013773426413536,
0.11802983283996582,
-0.005937475245445967,
-0.016349641606211662,
-0.0010152419563382864,
0.04237360134720802,
0.03285006806254387,
0.07503226399421692,
0.021748388186097145,
-0.015313387848436832,
0.03130017966032028,
-0.09423621743917465,
-0.0028269411996006966,
0.132438063621521,
0.05649259313941002,
-0.014663119800388813,
-0.046391405165195465,
-0.009337612427771091,
0.03339792415499687,
-0.04527388885617256,
-0.018243882805109024,
0.04232440143823624,
0.044916070997714996,
-0.013752630911767483,
0.05338165536522865,
-0.07263324409723282,
-0.01663321629166603,
-0.0018809091998264194,
-0.030778151005506516,
0.06470810621976852,
0.040290310978889465,
-0.014319988898932934,
-0.02116522379219532,
-0.08242446184158325,
-0.03773999214172363,
-0.08093626797199249,
0.0032683818135410547,
-0.011238731443881989,
-0.056377630680799484,
-0.010319780558347702,
0.05169517919421196,
0.06540674716234207,
-0.03213522583246231,
-0.06718223541975021,
0.028298798948526382,
0.11057952791452408,
-0.006148250307887793,
0.041961781680583954,
-0.09169924259185791,
0.014537579379975796
] |
OVTR: End-to-End Open-Vocabulary Multiple Object Tracking with Transformer | https://openreview.net/forum?id=GDS5eN65QY | [
"Jinyang Li",
"En Yu",
"Sijia Chen",
"Wenbing Tao"
] | Poster | Open-vocabulary multiple object tracking aims to generalize trackers to unseen categories during training, enabling their application across a variety of real-world scenarios. However, the existing open-vocabulary tracker is constrained by its framework structure, isolated frame-level perception, and insufficient modal interactions, which hinder its performance in open-vocabulary classification and tracking. In this paper, we propose OVTR (End-to-End Open-Vocabulary Multiple Object Tracking with TRansformer), the first end-to-end open-vocabulary tracker that models motion, appearance, and category simultaneously. To achieve stable classification and continuous tracking, we design the CIP (Category Information Propagation) strategy, which establishes multiple high-level category information priors for subsequent frames. Additionally, we introduce a dual-branch structure for generalization capability and deep multimodal interaction, and incorporate protective strategies in the decoder to enhance performance. Experimental results show that our method surpasses previous trackers on the open-vocabulary MOT benchmark while also achieving faster inference speeds and significantly reducing preprocessing requirements. Moreover, the experiment transferring the model to another dataset demonstrates its strong adaptability. | Open-Vocabulary, Multiple-Object Tracking, Multimodal, Transformer, End-to-End | null | 2,059 | 2503.10616 | [
0.02305244654417038,
-0.1499658226966858,
-0.018466845154762268,
0.04017849639058113,
0.1775391548871994,
-0.04801272228360176,
-0.06518201529979706,
-0.04088440537452698,
-0.010601961985230446,
-0.08610499650239944,
0.003990167751908302,
-0.08922690898180008,
-0.048440393060445786,
0.0431455597281456,
0.023558082059025764,
0.008065111935138702,
0.09570688009262085,
0.16225215792655945,
-0.15602998435497284,
-0.009682142175734043,
0.05223727598786354,
-0.01633271761238575,
0.11304279416799545,
-0.014306089840829372,
0.032464757561683655,
0.023789189755916595,
-0.06046988442540169,
-0.08748941868543625,
0.013095693662762642,
-0.08515182137489319,
0.0430634580552578,
-0.032104793936014175,
-0.05531023442745209,
0.06897100806236267,
-0.0711149200797081,
-0.01125215645879507,
-0.0191422738134861,
0.003865477629005909,
-0.06688123941421509,
-0.00891965813934803,
0.0111968033015728,
0.06880583614110947,
0.010662679560482502,
0.04149946570396423,
-0.015367946587502956,
0.03864876180887222,
0.019306691363453865,
-0.04410874843597412,
0.03683040291070938,
-0.0008561488939449191,
0.010116999968886375,
-0.0576312318444252,
0.02474297396838665,
0.06582044064998627,
-0.04553467035293579,
0.04124967381358147,
0.008813909254968166,
-0.10857553780078888,
0.031235817819833755,
-0.004117798060178757,
-0.024536922574043274,
-0.00880705751478672,
0.021021762862801552,
0.007545426953583956,
-0.05601903423666954,
-0.08865366876125336,
0.014098619110882282,
-0.018468599766492844,
-0.016561409458518028,
0.03442741185426712,
-0.026543477550148964,
0.0473327673971653,
0.08024196326732635,
0.018184302374720573,
0.03839404135942459,
-0.0228157676756382,
0.01861862652003765,
-0.05010417848825455,
0.04010309651494026,
-0.06042569875717163,
0.06900034844875336,
-0.04531214013695717,
0.021372999995946884,
-0.05972466245293617,
0.11684586107730865,
0.031057745218276978,
-0.07235591113567352,
0.05564827844500542,
-0.05418490618467331,
0.0402994230389595,
-0.07322714477777481,
-0.06893954426050186,
0.0785597637295723,
-0.04687860235571861,
0.008981977589428425,
-0.09347153455018997,
-0.039951011538505554,
-0.03204609453678131,
0.03280521556735039,
0.04860829561948776,
-0.001284659723751247,
-0.0464133620262146,
-0.015305583365261555,
0.04679218307137489,
-0.009213700890541077,
-0.049450118094682693,
0.04780711606144905,
0.09033626317977905,
-0.002378383418545127,
0.021990273147821426,
0.08494404703378677,
-0.022253068163990974,
0.06096253916621208,
-0.04550504684448242,
0.01312676165252924,
0.09986530989408493,
-0.056704889982938766,
0.04269478842616081,
0.0357421338558197,
0.023564843460917473,
0.0001689924974925816,
0.0012356922961771488,
0.014298963360488415,
-0.00968259759247303,
-0.00043901678873226047,
0.030724771320819855,
-0.011619525961577892,
-1.7043891158625672e-33,
0.049645714461803436,
-0.034605178982019424,
0.024702055379748344,
-0.03075990453362465,
0.03233402222394943,
0.03756505623459816,
-0.06178414076566696,
0.026850935071706772,
-0.0686998963356018,
0.020281845703721046,
0.03173062205314636,
0.014276922680437565,
-0.10690213739871979,
0.08459898829460144,
0.03491842374205589,
-0.0488160103559494,
-0.0670020580291748,
0.06229112669825554,
-0.03769916296005249,
0.07355386018753052,
0.05850542336702347,
-0.043158531188964844,
-0.042778436094522476,
-0.017955288290977478,
0.04047827422618866,
0.028869744390249252,
0.07194440811872482,
-0.05770227685570717,
0.010868076235055923,
0.01961483806371689,
-0.07064849138259888,
0.02802545763552189,
0.028146782889962196,
0.044316232204437256,
0.05743694305419922,
-0.06878730654716492,
-0.041903212666511536,
-0.049781281501054764,
0.03136565536260605,
-0.034972235560417175,
-0.03776681795716286,
-0.019312720745801926,
-0.06546970456838608,
-0.039543285965919495,
-0.0764380395412445,
-0.03012954629957676,
0.0593065470457077,
0.10888130962848663,
-0.09336625784635544,
0.02606217935681343,
0.03813162073493004,
-0.007413601037114859,
-0.08566350489854813,
-0.040491633117198944,
-0.03163826838135719,
-0.040150851011276245,
0.0004060211649630219,
-0.029192686080932617,
-0.022002214565873146,
-0.02009214274585247,
-0.04875974357128143,
0.08287308365106583,
-0.021943314000964165,
-0.01773928292095661,
0.024211430922150612,
-0.016110608354210854,
0.0014837547205388546,
0.09561695158481598,
0.024784348905086517,
0.06935577839612961,
-0.03986980393528938,
0.04986812174320221,
-0.03275329992175102,
-0.06585776060819626,
0.08254765719175339,
0.06320226937532425,
0.03477753698825836,
-0.020811256021261215,
0.0034338587429374456,
0.06085633859038353,
-0.014353340491652489,
-0.0478120893239975,
0.03259662166237831,
0.009636731818318367,
-0.01238000113517046,
0.03491725027561188,
0.007644776254892349,
-0.009711835533380508,
-0.022671621292829514,
0.028468577191233635,
-0.013629279099404812,
-0.01879679225385189,
-0.06457671523094177,
0.011234931647777557,
-0.005628828890621662,
1.7235333270626544e-33,
0.02886035107076168,
0.060550589114427567,
0.05095783248543739,
-0.008364580571651459,
0.008199659176170826,
-0.048025742173194885,
-0.031525805592536926,
-0.01551862619817257,
0.06205790117383003,
-0.07701392471790314,
-0.04215608164668083,
-0.036871183663606644,
0.02307901345193386,
0.009586106985807419,
0.08579324185848236,
-0.02938302792608738,
-0.048041097819805145,
0.0763183981180191,
-0.0344572551548481,
-0.013367366045713425,
0.016396867111325264,
0.020834650844335556,
-0.09430642426013947,
-0.021256867796182632,
-0.022796038538217545,
0.047026846557855606,
-0.042157646268606186,
-0.0042732167057693005,
0.02337827906012535,
-0.04160105809569359,
0.043049588799476624,
0.014216496609151363,
0.046607159078121185,
-0.05322551354765892,
-0.06194635480642319,
0.054411351680755615,
0.04703538864850998,
-0.02322658710181713,
-0.024448202922940254,
0.04630718007683754,
0.02640693634748459,
-0.038442760705947876,
-0.05609945207834244,
0.01861012540757656,
-0.047997619956731796,
-0.03648502007126808,
-0.08218224346637726,
0.05390165373682976,
0.007171793840825558,
0.024396216496825218,
-0.081461600959301,
0.03426799923181534,
-0.05490722507238388,
-0.02294331230223179,
0.008523453027009964,
-0.021447768434882164,
-0.03627549484372139,
-0.04617643356323242,
0.023758625611662865,
-0.04897301271557808,
0.01564420759677887,
0.036152251064777374,
-0.001222029561176896,
-0.028294116258621216,
0.020604267716407776,
-0.028628841042518616,
0.029983144253492355,
-0.011185132898390293,
0.010602808557450771,
0.018084172159433365,
-0.07377529889345169,
0.017659833654761314,
-0.014154365286231041,
-0.016364574432373047,
0.0075266980566084385,
0.01905161142349243,
-0.0004412480629980564,
-0.05126342177391052,
0.07944095879793167,
0.026262184605002403,
-0.017036575824022293,
-0.1141047328710556,
0.07808490842580795,
-0.010465350933372974,
0.1093241423368454,
0.05005631595849991,
0.057547103613615036,
0.07330092787742615,
0.09203345328569412,
0.06974299997091293,
-0.03672073408961296,
0.033565398305654526,
0.026814918965101242,
0.049570243805646896,
-0.057471562176942825,
-3.654780655892864e-8,
-0.14383761584758759,
0.020418765023350716,
-0.050481002777814865,
-0.03878369554877281,
-0.04110553860664368,
0.028666313737630844,
0.015813205391168594,
0.06848876178264618,
0.0010370200034230947,
-0.011474788188934326,
0.033165670931339264,
0.08907835185527802,
0.02632208541035652,
0.004412508569657803,
0.012939543463289738,
0.029253745451569557,
0.012660530395805836,
-0.03556813299655914,
0.046135857701301575,
0.06999930739402771,
-0.02890748158097267,
-0.018998950719833374,
0.029927747324109077,
0.010510009713470936,
-0.0671560987830162,
-0.03717850521206856,
0.02172579988837242,
0.1389949917793274,
0.10698724538087845,
-0.025423701852560043,
-0.030015187337994576,
0.041238993406295776,
0.010929599404335022,
-0.06643054634332657,
-0.01093311421573162,
0.05638115480542183,
0.028234204277396202,
0.008465064689517021,
0.022999588400125504,
0.02272249571979046,
0.09395954012870789,
0.002412663074210286,
0.006607615388929844,
0.07437001913785934,
-0.03556369990110397,
0.03634866699576378,
-0.011260846629738808,
-0.12162966281175613,
-0.005539058707654476,
-0.03508398309350014,
-0.0011105951853096485,
0.011056982912123203,
-0.005140634719282389,
0.16528649628162384,
0.03804630786180496,
-0.0073164282366633415,
0.018489016219973564,
-0.02309281937777996,
0.01599145494401455,
0.07826333492994308,
-0.03565486893057823,
-0.06920340657234192,
-0.017009105533361435,
0.09105151891708374
] |
Learning Evolving Tools for Large Language Models | https://openreview.net/forum?id=wtrDLMFU9v | [
"Guoxin Chen",
"Zhong Zhang",
"Xin Cong",
"Fangda Guo",
"Yesai Wu",
"Yankai Lin",
"Wenzheng Feng",
"Yasheng Wang"
] | Poster | Tool learning enables large language models (LLMs) to interact with external tools and APIs, greatly expanding the application scope of LLMs. However, due to the dynamic nature of external environments, these tools and APIs may become outdated over time, preventing LLMs from correctly invoking tools. Existing research primarily focuses on static environments and overlooks this issue, limiting the adaptability of LLMs in real-world applications. In this paper, we propose ToolEVO, a novel framework designed to enhance the adaptive and reflective capabilities of LLMs against tool variability. By leveraging Monte Carlo Tree Search, ToolEVO facilitates active exploration and interaction of LLMs within dynamic environments, allowing for autonomous self-reflection and self-updating of tool usage based on environmental feedback. Additionally, we introduce ToolQA-D, a benchmark specifically designed to evaluate the impact of tool variability. Extensive experiments demonstrate the effectiveness and stability of our approach, highlighting the importance of adaptability to tool variability for effective tool learning. | Tool Learning, Monte Calro Tree Search, Large Language Models | null | 2,051 | 2410.06617 | [
-0.06163601577281952,
-0.07624717056751251,
0.006490633357316256,
0.08073332905769348,
0.06434079259634018,
-0.0890359878540039,
-0.06222352012991905,
-0.018272191286087036,
0.01066984049975872,
-0.03143657371401787,
-0.09782560169696808,
-0.07291742414236069,
0.052621111273765564,
-0.016419151797890663,
0.06112094968557358,
-0.016356589272618294,
0.12076026201248169,
0.00037313459324650466,
-0.011825635097920895,
-0.06132889911532402,
0.019996099174022675,
0.06755036115646362,
0.056974008679389954,
0.010699812322854996,
0.006368022877722979,
-0.006063405890017748,
-0.00536659499630332,
0.0036885386798530817,
0.02240576222538948,
0.007618337403982878,
-0.03678975999355316,
0.06487314403057098,
-0.05463922768831253,
-0.010294362902641296,
0.011821632273495197,
0.038724351674318314,
-0.07467219233512878,
0.01428284589201212,
0.003789674723520875,
-0.05680917948484421,
-0.10168182849884033,
0.011694422923028469,
0.0020290298853069544,
-0.09644519537687302,
0.051323190331459045,
-0.07694579660892487,
-0.02831105701625347,
-0.06282510608434677,
-0.05012393370270729,
0.05979293957352638,
-0.060682833194732666,
-0.14019478857517242,
-0.05325858294963837,
-0.030759217217564583,
0.02052694745361805,
0.04057232663035393,
0.05850134789943695,
0.021013127639889717,
-0.025898577645421028,
-0.00979518610984087,
0.008870985358953476,
-0.06631402671337128,
-0.05290469154715538,
0.07244724035263062,
-0.018206708133220673,
0.012295638211071491,
0.02168930508196354,
0.08143141865730286,
0.010300121270120144,
-0.03543195500969887,
-0.0898522138595581,
0.0816730260848999,
0.003330638399347663,
0.01778094470500946,
0.02731584757566452,
0.029383404180407524,
-0.0591132752597332,
-0.014274333603680134,
0.07564350217580795,
-0.11338084936141968,
0.0031071340199559927,
0.037985995411872864,
-0.018899859860539436,
0.06585361063480377,
0.08174804598093033,
0.05389660596847534,
0.11218000948429108,
0.036173563450574875,
0.10226346552371979,
0.06808710843324661,
0.01880837418138981,
-0.07137995213270187,
0.04383135959506035,
0.0072875856421887875,
-0.026527039706707,
-0.002520298585295677,
0.05698634311556816,
-0.09303701668977737,
-0.017404163256287575,
0.0654226765036583,
0.015660110861063004,
0.048892825841903687,
0.020140696316957474,
-0.08330047875642776,
-0.03555523231625557,
0.04532553255558014,
0.0037340978160500526,
-0.011854905635118484,
-0.012799658812582493,
-0.03877212852239609,
-0.0036692628636956215,
-0.0006650753784924746,
0.013499882072210312,
-0.03015974909067154,
-0.0041247643530368805,
-0.05020010471343994,
0.08705639094114304,
-0.09460345655679703,
-0.014659837819635868,
0.0667751207947731,
0.024222955107688904,
0.056891050189733505,
-0.021105336025357246,
-0.00999003741890192,
0.034544285386800766,
0.05695914849638939,
-0.04321004077792168,
-1.6146545633090564e-34,
0.14343708753585815,
-0.04418451711535454,
-0.005276472307741642,
0.06727775931358337,
0.07041997462511063,
-0.019456367939710617,
0.0852232277393341,
0.07571911066770554,
-0.06410752236843109,
-0.03496047481894493,
0.010458569973707199,
0.0686505138874054,
-0.05408418923616409,
-0.008871774189174175,
0.11408179253339767,
0.024106139317154884,
-0.07297704368829727,
0.10503046959638596,
-0.06696638464927673,
-0.01969616860151291,
-0.0027822668198496103,
-0.07190028578042984,
-0.0037883769255131483,
-0.04129086807370186,
0.09610383957624435,
0.09193184226751328,
0.07740876078605652,
0.02247323840856552,
-0.029317496344447136,
0.003960584755986929,
-0.038160428404808044,
-0.06770855188369751,
0.041880372911691666,
0.062025994062423706,
-0.007714258506894112,
-0.010272437706589699,
-0.06343864649534225,
-0.024393117055296898,
0.008171755820512772,
-0.044563841074705124,
-0.020628418773412704,
-0.014429431408643723,
0.04037172347307205,
-0.06407987326383591,
-0.030893009155988693,
-0.06166968122124672,
0.0018490643706172705,
0.012050852179527283,
0.017220281064510345,
-0.05945907160639763,
0.04503045976161957,
0.04204203188419342,
0.04843045398592949,
0.0349816232919693,
-0.005965696182101965,
0.006822490133345127,
-0.007390236482024193,
0.06678959727287292,
0.008006289601325989,
0.02344880811870098,
-0.01222284883260727,
0.0539773665368557,
0.001960485242307186,
0.06531289219856262,
0.10520811378955841,
0.06692973524332047,
0.050988439470529556,
0.10847986489534378,
0.027158085256814957,
-0.015855468809604645,
-0.005533881019800901,
-0.01788492500782013,
-0.042560070753097534,
-0.024433773010969162,
0.055508993566036224,
-0.06942009180784225,
0.05156155675649643,
-0.0963883027434349,
0.04156240075826645,
-0.0164315365254879,
0.009682590141892433,
0.006047375034540892,
0.028971750289201736,
-0.04724686220288277,
-0.0141641516238451,
-0.08061497658491135,
0.059522729367017746,
-0.11043433845043182,
-0.007777482736855745,
0.04185857996344566,
-0.02725999429821968,
-0.06526307761669159,
-0.05466056242585182,
0.05066950246691704,
-0.008556322194635868,
-1.649048678535777e-33,
-0.03339358791708946,
-0.0723721832036972,
0.00970771536231041,
0.09822622686624527,
-0.05335143581032753,
-0.06356589496135712,
-0.07181967794895172,
-0.003085864707827568,
0.0006543354247696698,
-0.08304038643836975,
-0.049579039216041565,
0.026782169938087463,
0.01357088703662157,
-0.01657971739768982,
0.09185221791267395,
0.005435646511614323,
-0.07123833149671555,
-0.013059895485639572,
0.014180026017129421,
0.06831580400466919,
-0.017853116616606712,
0.054618142545223236,
-0.015942992642521858,
-0.011784123256802559,
-0.0577094629406929,
-0.03931507468223572,
-0.013440604321658611,
0.001487136585637927,
-0.02300846390426159,
0.0013315451797097921,
-0.017591020092368126,
0.014817333780229092,
-0.057606905698776245,
-0.020473642274737358,
-0.04186839982867241,
0.03496552258729935,
0.059193406254053116,
-0.012599612586200237,
0.02635791338980198,
0.07867757976055145,
0.04088746756315231,
-0.05670502409338951,
-0.05309449881315231,
-0.028241019695997238,
-0.005164007656276226,
0.049270711839199066,
-0.11237987875938416,
0.03440437838435173,
0.029544951394200325,
-0.022354625165462494,
0.04267130047082901,
-0.008124136365950108,
-0.030967338010668755,
-0.06019855663180351,
-0.052154380828142166,
-0.03068375773727894,
-0.05561353638768196,
-0.035139817744493484,
0.00030519545543938875,
0.045630957931280136,
-0.05710720270872116,
-0.05221579596400261,
0.01353534683585167,
-0.03533908724784851,
-0.030325336381793022,
-0.028880445286631584,
-0.06188385561108589,
-0.003584772115573287,
-0.06930864602327347,
-0.0069883293472230434,
0.08356599509716034,
-0.035704776644706726,
-0.050585031509399414,
-0.06206529214978218,
0.031430285423994064,
-0.031206686049699783,
0.022540729492902756,
-0.10675972700119019,
-0.041037146002054214,
-0.056587811559438705,
0.04836149141192436,
0.08732607960700989,
0.06537037342786789,
0.03418426960706711,
-0.030291080474853516,
0.0659305676817894,
-0.032742615789175034,
0.11783096194267273,
0.033432260155677795,
0.011011811904609203,
-0.029994644224643707,
0.009124765172600746,
-0.01191645860671997,
0.05823475122451782,
-0.06786470860242844,
-3.5728312752780766e-8,
-0.012496610172092915,
0.09092932194471359,
0.014806156046688557,
0.09957592189311981,
0.009994715452194214,
0.03906829655170441,
-0.03878152370452881,
0.11050135642290115,
0.01356352400034666,
0.02127659134566784,
0.06453093886375427,
-0.05093136429786682,
-0.001617683214135468,
0.021379467099905014,
-0.026591463014483452,
0.06235197186470032,
-0.07254309952259064,
0.026690948754549026,
-0.047554291784763336,
-0.06052500009536743,
0.05247952789068222,
0.05182403698563576,
0.03471958264708519,
-0.004174119792878628,
0.030705291777849197,
-0.10013547539710999,
-0.07798564434051514,
0.058051954954862595,
0.07038414478302002,
-0.016068778932094574,
0.02951807714998722,
0.0035457781050354242,
-0.024210816249251366,
0.02122584357857704,
0.03797820210456848,
0.033289577811956406,
-0.040472615510225296,
-0.02823217213153839,
0.01918286643922329,
0.012914716266095638,
0.03207199275493622,
0.03902241215109825,
-0.0621594563126564,
0.02481713518500328,
-0.03878939524292946,
-0.015383194200694561,
-0.02948955073952675,
-0.07931910455226898,
0.02121714875102043,
-0.07115888595581055,
-0.03602098673582077,
-0.035964496433734894,
-0.024092594161629677,
-0.013005495071411133,
0.08506446331739426,
0.04513223469257355,
0.005081025883555412,
-0.025390254333615303,
0.07544109970331192,
0.06561112403869629,
-0.04036908224225044,
-0.006348421331495047,
-0.006338541395962238,
0.056209150701761246
] |
Towards Effective Evaluations and Comparisons for LLM Unlearning Methods | https://openreview.net/forum?id=wUtCieKuQU | [
"Qizhou Wang",
"Bo Han",
"Puning Yang",
"Jianing Zhu",
"Tongliang Liu",
"Masashi Sugiyama"
] | Poster | The imperative to eliminate undesirable data memorization underscores the significance of machine unlearning for large language models (LLMs). Recent research has introduced a series of promising unlearning methods, notably boosting the practical significance of the field. Nevertheless, adopting a proper evaluation framework to reflect the true unlearning efficacy is also essential yet has not received adequate attention. This paper seeks to improve the evaluation of LLM unlearning by addressing two key challenges---a) the robustness of evaluation metrics and b) the trade-offs between competing goals. The first challenge stems from findings that current metrics are susceptible to various red teaming scenarios. It indicates that they may not reflect the true extent of knowledge retained by LLMs but rather tend to mirror superficial model behaviors, thus prone to attacks. We address this issue by devising and assessing a series of candidate metrics, selecting the most robust ones under various types of attacks. The second challenge arises from the conflicting goals of eliminating unwanted knowledge while retaining those of others. This trade-off between unlearning and retention often fails to conform the Pareto frontier, rendering it subtle to compare the efficacy between methods that excel only in either unlearning or retention. We handle this issue by proposing a calibration method that can restore the original performance on non-targeted data after unlearning, thereby allowing us to focus exclusively on assessing the strength of unlearning. Our evaluation framework notably enhances the effectiveness when assessing and comparing various LLM unlearning methods, further allowing us to benchmark existing works, identify their proper hyper-parameters, and explore new tricks to enhance their practical efficacy. | llm unlearning | null | 2,049 | 2406.09179 | [
0.025278223678469658,
-0.07361584901809692,
-0.06100180745124817,
0.0703861191868782,
0.035959482192993164,
0.06949340552091599,
0.008535954169929028,
-0.004004350397735834,
0.007529571186751127,
-0.05799173563718796,
-0.04231547191739082,
-0.03364618495106697,
0.07691515237092972,
-0.019221318885684013,
0.03455186262726784,
0.01988382823765278,
0.07632214576005936,
0.06637072563171387,
-0.05307133123278618,
-0.058160677552223206,
-0.009046629071235657,
-0.00038926786510273814,
-0.002388234017416835,
0.028133679181337357,
0.02742817811667919,
-0.057730454951524734,
-0.03158194199204445,
0.0035260368604213,
-0.04992549493908882,
-0.03515874221920967,
-0.011176664382219315,
0.03468392789363861,
-0.0016004418721422553,
0.02577236481010914,
-0.02650274522602558,
0.02813529223203659,
-0.05699187144637108,
0.0325140655040741,
0.014028022065758705,
-0.05880969017744064,
-0.08936094492673874,
-0.018776027485728264,
0.008955119177699089,
-0.021603994071483612,
0.05443977937102318,
-0.02357606031000614,
-0.04897679015994072,
0.049272194504737854,
-0.11160233616828918,
-0.02833457477390766,
-0.029091723263263702,
-0.01815984770655632,
-0.039559170603752136,
0.05318187177181244,
-0.02016325853765011,
0.02360278181731701,
0.019096845760941505,
0.01572445034980774,
-0.043753787875175476,
-0.04818634316325188,
-0.06968309730291367,
-0.11506617069244385,
-0.08902566879987717,
0.005063771735876799,
-0.06938820332288742,
-0.035429537296295166,
0.05866105854511261,
0.0575866624712944,
-0.03159896284341812,
0.09013521671295166,
0.004195164889097214,
0.014180942438542843,
-0.017789054661989212,
0.009298066608607769,
0.0567752905189991,
0.10850432515144348,
0.021983115002512932,
0.01395572442561388,
0.06916555017232895,
-0.08476214110851288,
0.01812608726322651,
0.04419118911027908,
0.01921718195080757,
0.03801232948899269,
0.11576246470212936,
0.01185758225619793,
-0.00027041573775932193,
-0.016924941912293434,
0.05149398744106293,
0.027054166421294212,
-0.0008089844486676157,
0.03766000643372536,
0.09196595102548599,
0.006663879379630089,
0.0429946705698967,
0.0057168821804225445,
-0.004565032199025154,
-0.03353545814752579,
-0.047309648245573044,
0.07917285710573196,
0.009439635090529919,
0.100065216422081,
0.031760212033987045,
-0.09089235216379166,
-0.022338546812534332,
-0.019706889986991882,
0.00046713417395949364,
-0.016911640763282776,
0.04854721948504448,
-0.08470093458890915,
0.0374123677611351,
0.07135523855686188,
0.0246121846139431,
0.011440671980381012,
0.024213368073105812,
0.0061812507919967175,
0.025161854922771454,
-0.01870046555995941,
-0.05357656627893448,
0.1077069491147995,
-0.01907208375632763,
-0.010486917570233345,
0.10310167819261551,
0.02397957071661949,
0.021630506962537766,
0.01119964849203825,
-0.05240565165877342,
4.5480353977894325e-33,
0.04010117053985596,
0.024207210168242455,
-0.05145532637834549,
0.012176544405519962,
0.052324093878269196,
-0.046228062361478806,
0.007694246247410774,
0.021030526608228683,
-0.05842972174286842,
-0.01512078009545803,
-0.03779095038771629,
0.06369739770889282,
0.03124195523560047,
0.05719054862856865,
0.08892269432544708,
0.03795343264937401,
-0.012187021784484386,
0.08356984704732895,
-0.09655405580997467,
-0.03282250091433525,
0.06066298857331276,
-0.023587996140122414,
0.03706071525812149,
-0.0999063029885292,
0.01631856895983219,
-0.022115398198366165,
0.01657487452030182,
0.003311944892629981,
-0.05102715268731117,
0.05488106235861778,
-0.06431183964014053,
-0.035532496869564056,
0.04939579218626022,
0.10698102414608002,
0.03796902298927307,
-0.021341634914278984,
-0.08094419538974762,
-0.053245045244693756,
0.04076191782951355,
0.0065192896872758865,
-0.04156206548213959,
0.027963912114501,
0.00024213486176449805,
-0.04313776269555092,
-0.010182738304138184,
-0.01094930712133646,
-0.017847789451479912,
-0.027362175285816193,
0.007391090039163828,
-0.07571340352296829,
0.05993756279349327,
0.0014481000835075974,
0.0009793221252039075,
0.03589106723666191,
-0.04664863646030426,
0.0316447913646698,
-0.008445787243545055,
0.05112113431096077,
-0.010522008873522282,
0.07483987510204315,
-0.01470126211643219,
0.07271618396043777,
-0.035196639597415924,
0.05747712403535843,
0.056995753198862076,
-0.03293992578983307,
-0.011912615969777107,
0.06368190050125122,
0.015036900527775288,
-0.07814543694257736,
-0.03639146685600281,
-0.06628584116697311,
-0.040381185710430145,
-0.04107511416077614,
0.013886000961065292,
-0.07756521552801132,
0.10973663628101349,
0.008465511724352837,
0.07142835855484009,
-0.02915959060192108,
0.07744232565164566,
-0.010344681330025196,
0.0006375456578098238,
-0.093778595328331,
-0.08008061349391937,
-0.04780331999063492,
0.05822483077645302,
-0.055445872247219086,
-0.03304398059844971,
0.0882808044552803,
-0.025526657700538635,
-0.030480794608592987,
-0.07773394137620926,
-0.0038770728278905153,
0.015096253715455532,
-5.495382458164438e-33,
-0.051449019461870193,
-0.0060571967624127865,
-0.01536162756383419,
0.11779455840587616,
-0.024602822959423065,
-0.06296424567699432,
0.020779311656951904,
0.02355591580271721,
-0.08980421721935272,
-0.07283356040716171,
-0.013584191910922527,
-0.03902830183506012,
-0.005868818610906601,
-0.011784280650317669,
0.03462495654821396,
-0.02374761551618576,
-0.02473783679306507,
0.013039556331932545,
0.009959262795746326,
0.07743381708860397,
0.03341151773929596,
0.04157467931509018,
-0.04791439324617386,
0.04463624209165573,
-0.07077289372682571,
-0.04127929359674454,
-0.06765784323215485,
-0.041741471737623215,
0.0006375248194672167,
-0.027868160977959633,
0.0248189028352499,
0.003899641800671816,
-0.051201604306697845,
-0.02767757698893547,
-0.03828301653265953,
0.03547872602939606,
0.05970083922147751,
-0.021860212087631226,
-0.011483981274068356,
0.1262792944908142,
0.047071125358343124,
0.017650503665208817,
-0.13309866189956665,
-0.06331442296504974,
0.014757715165615082,
-0.05623830854892731,
-0.08085603266954422,
-0.011250046081840992,
0.02662261389195919,
-0.06860526651144028,
0.0038905732799321413,
0.003914805594831705,
0.004634378477931023,
0.042083658277988434,
-0.07740217447280884,
-0.06886682659387589,
0.06845459342002869,
-0.09840980917215347,
0.016168510541319847,
0.04953746497631073,
-0.04636751487851143,
0.01634947396814823,
-0.015338241122663021,
-0.03062455542385578,
0.05263068899512291,
-0.06445048749446869,
-0.0027346096467226744,
0.10597263276576996,
-0.05669521540403366,
0.017792958766222,
-0.0035996658261865377,
0.0036685995291918516,
-0.03830553591251373,
-0.011990010738372803,
-0.052958980202674866,
-0.021363072097301483,
-0.06265030056238174,
-0.07567848265171051,
-0.040434058755636215,
-0.049466632306575775,
0.038063667714595795,
0.012815483845770359,
0.012576360255479813,
0.08850531280040741,
-0.007632181514054537,
0.06313300132751465,
0.10405857861042023,
0.0941910669207573,
-0.013047389686107635,
-0.03130511939525604,
-0.011036011390388012,
-0.07096832990646362,
0.011758479289710522,
0.06757032126188278,
-0.037904031574726105,
-5.3860230053714986e-8,
-0.0055658710189163685,
-0.01000414788722992,
0.0312785767018795,
0.040691353380680084,
0.026065843179821968,
-0.07462681829929352,
-0.15694692730903625,
0.09968159347772598,
-0.013728071935474873,
-0.0031522889621555805,
0.09288658201694489,
-0.025551527738571167,
-0.11631971597671509,
-0.024731406942009926,
-0.00604755524545908,
0.07579100131988525,
0.0042776563204824924,
0.07521211355924606,
-0.0555192194879055,
-0.032360341399908066,
0.04357410594820976,
0.02568357065320015,
-0.03225726634263992,
-0.012961338274180889,
0.06690452992916107,
-0.04580205678939819,
-0.04921753332018852,
0.08468616008758545,
0.05505324527621269,
0.012945027090609074,
0.00807269662618637,
-0.04758451133966446,
-0.015517167747020721,
-0.022524720057845116,
0.04728804901242256,
0.12655965983867645,
-0.014847314916551113,
-0.07036720216274261,
0.030853701755404472,
0.020906345918774605,
0.006225269287824631,
0.11292117089033127,
-0.022927967831492424,
0.09328466653823853,
0.0037361322902143,
-0.03494039177894592,
-0.07327877730131149,
-0.11115582287311554,
-0.012397954240441322,
-0.08326626569032669,
0.033476777374744415,
-0.02144562266767025,
0.0030309376306831837,
0.0832272320985794,
0.03012687899172306,
0.06526797264814377,
-0.019923483952879906,
-0.00855667982250452,
0.025601191446185112,
0.04338478669524193,
0.07379218190908432,
-0.005452196579426527,
-0.049252692610025406,
0.01706734485924244
] |
Distilling Reinforcement Learning Algorithms for In-Context Model-Based Planning | https://openreview.net/forum?id=BfUugGfBE5 | [
"Jaehyeon Son",
"Soochan Lee",
"Gunhee Kim"
] | Poster | Recent studies have shown that Transformers can perform in-context reinforcement learning (RL) by imitating existing RL algorithms, enabling sample-efficient adaptation to unseen tasks without parameter updates. However, these models also inherit the suboptimal behaviors of the RL algorithms they imitate. This issue primarily arises due to the gradual update rule employed by those algorithms. Model-based planning offers a promising solution to this limitation by allowing the models to simulate potential outcomes before taking action, providing an additional mechanism to deviate from the suboptimal behavior. Rather than learning a separate dynamics model, we propose Distillation for In-Context Planning (DICP), an in-context model-based RL framework where Transformers simultaneously learn environment dynamics and improve policy in-context. We evaluate DICP across a range of discrete and continuous environments, including Darkroom variants and Meta-World. Our results show that DICP achieves state-of-the-art performance while requiring significantly fewer environment interactions than baselines, which include both model-free counterparts and existing meta-RL methods. | reinforcement learning, in-context learning | This work is the first to propose model-based planning for in-context RL that imitates a source RL algorithm, leveraging Transformers to simultaneously learn environment dynamics and improve policy in-context. | 2,042 | 2502.19009 | [
-0.055232707411050797,
-0.035373155027627945,
0.08339492231607437,
0.05564573034644127,
0.06007835641503334,
-0.09788443148136139,
-0.04414118826389313,
0.008552677929401398,
0.022009819746017456,
0.0290400180965662,
-0.09522189199924469,
-0.05167875438928604,
0.018192345276474953,
0.05683373287320137,
0.08358577638864517,
0.019338348880410194,
0.08632054179906845,
0.054832108318805695,
-0.1138010323047638,
-0.004242857918143272,
-0.006625333800911903,
0.024039767682552338,
0.0527530200779438,
-0.012809615582227707,
0.018881823867559433,
0.020523037761449814,
0.03952276334166527,
0.021334247663617134,
0.04746130481362343,
-0.009751087054610252,
0.0557737722992897,
0.003173927078023553,
-0.05766158178448677,
-0.08839765191078186,
-0.041796810925006866,
0.12763844430446625,
-0.07991795986890793,
-0.00258835731074214,
-0.02153080329298973,
-0.021072756499052048,
0.003658606903627515,
0.049980904906988144,
-0.014128652401268482,
0.053255993872880936,
0.03598590940237045,
0.0015054605901241302,
0.007630798500031233,
0.012400830164551735,
-0.0027387177105993032,
-0.06014753133058548,
-0.03163259103894234,
-0.004560857079923153,
0.01697658561170101,
0.0341610349714756,
0.0031920531764626503,
0.04279039800167084,
0.0512663871049881,
0.013131409883499146,
0.0033919967245310545,
-0.07336386293172836,
-0.0320318229496479,
-0.058806274086236954,
-0.052763354033231735,
-0.02728821337223053,
-0.13925550878047943,
0.027320824563503265,
0.0010973858879879117,
0.09354753792285919,
0.03753725811839104,
-0.0794159322977066,
-0.040904849767684937,
0.02583189308643341,
0.02210514061152935,
0.015882866457104683,
0.05337468534708023,
0.0036521817091852427,
-0.03238316997885704,
0.0012114866403862834,
0.03868822380900383,
-0.005817390512675047,
0.11119423806667328,
0.011775313876569271,
-0.05135384574532509,
0.009274769574403763,
0.08498824387788773,
-0.029828092083334923,
0.007917499169707298,
0.04701227694749832,
0.06767577677965164,
0.0057943155989050865,
0.024167906492948532,
-0.016677575185894966,
-0.02240123599767685,
-0.031663913279771805,
-0.03252346068620682,
0.03242814168334007,
-0.0308183953166008,
-0.06494662910699844,
-0.039174795150756836,
0.11319553107023239,
-0.011210941709578037,
0.01589593105018139,
-0.016956262290477753,
0.008492257446050644,
-0.035011935979127884,
-0.048410478979349136,
0.006905258633196354,
0.0362391397356987,
-0.028850959613919258,
-0.05254526436328888,
0.06998404860496521,
-0.006020525004714727,
0.10211190581321716,
0.049932848662137985,
-0.011510099284350872,
0.010143253020942211,
0.04448765516281128,
-0.034563098102808,
-0.03480949625372887,
0.10826421529054642,
0.003076908877119422,
-0.06660059094429016,
0.023188557475805283,
0.020507214590907097,
-0.06732787191867828,
-0.05171779915690422,
-0.012517470866441727,
8.75818780140446e-34,
0.038678769022226334,
-0.03307570144534111,
0.028818853199481964,
0.05655645206570625,
0.07713861018419266,
0.0011511609191074967,
0.016499105840921402,
0.00961310788989067,
-0.02230365015566349,
-0.0145376306027174,
0.08227775990962982,
0.038308653980493546,
-0.07201137393712997,
0.014081573113799095,
0.06202264130115509,
-0.007357982452958822,
-0.056107088923454285,
0.08017287403345108,
-0.023826010525226593,
-0.021552128717303276,
0.02694675326347351,
-0.04261019825935364,
-0.05176205188035965,
-0.07694313675165176,
0.02358250878751278,
0.08773592114448547,
0.04737411439418793,
0.050465650856494904,
0.08367650210857391,
-0.0039450195617973804,
0.00281537720002234,
0.055233824998140335,
-0.07058284431695938,
0.10727895796298981,
-0.019403627142310143,
0.037662383168935776,
-0.05821344256401062,
-0.04947183281183243,
-0.03723246604204178,
-0.01878407970070839,
-0.017107874155044556,
0.00013095598842483014,
-0.0039048087783157825,
0.05227251723408699,
-0.0967273935675621,
-0.07864903658628464,
0.05293154716491699,
-0.02520027756690979,
-0.10693011432886124,
0.014196990057826042,
0.05236489325761795,
0.07210949063301086,
-0.02589591033756733,
-0.15284165740013123,
-0.0301728006452322,
-0.005492111202329397,
0.042231932282447815,
0.0038327090442180634,
0.04559651389718056,
0.08690807223320007,
0.04508088156580925,
-0.0021977119613438845,
-0.0606582909822464,
0.14221611618995667,
0.04097280278801918,
0.04880036786198616,
-0.02012370526790619,
0.07632022351026535,
0.07119117677211761,
-0.045240603387355804,
-0.042727287858724594,
-0.0023898109793663025,
-0.037385277450084686,
-0.0022062945645302534,
0.04388012737035751,
-0.061040643602609634,
0.018353475257754326,
-0.03344351425766945,
-0.0652787834405899,
0.02843802236020565,
0.03800579532980919,
0.05898318067193031,
-0.06926018744707108,
0.014886364340782166,
0.04590462148189545,
-0.02581755258142948,
-0.03165086731314659,
-0.0501866415143013,
-0.03937561810016632,
-0.07168442755937576,
-0.038975559175014496,
0.006651720032095909,
-0.03181208670139313,
0.031242696568369865,
0.05680637061595917,
-1.6140768032438525e-33,
0.026511888951063156,
-0.08895061910152435,
-0.0535825677216053,
0.035681720823049545,
0.027088157832622528,
-0.020895687863230705,
-0.01635286957025528,
-0.09146244078874588,
0.041792307049036026,
-0.06310625374317169,
-0.07355552911758423,
0.03072582371532917,
0.028163766488432884,
0.03454384580254555,
0.013186844997107983,
-0.03698277100920677,
0.013115781359374523,
-0.052850861102342606,
-0.03529966250061989,
0.010422212071716785,
0.024977177381515503,
0.035476673394441605,
-0.09333238750696182,
0.033452391624450684,
-0.08872748166322708,
0.02380971424281597,
-0.018945282325148582,
0.07252561300992966,
0.06818848103284836,
0.03791790455579758,
-0.0846574679017067,
0.011746851727366447,
0.01731809228658676,
-0.03075420670211315,
-0.06446128338575363,
0.07100541144609451,
0.02421562932431698,
-0.018131166696548462,
-0.12094871699810028,
0.07741335034370422,
0.008279266767203808,
-0.08651959151029587,
-0.026823658496141434,
0.028304079547524452,
-0.03629250079393387,
0.030244993045926094,
-0.05273103713989258,
-0.045302603393793106,
-0.0219173151999712,
-0.005086504388600588,
-0.02466655522584915,
0.07055409997701645,
-0.12507200241088867,
-0.06885642558336258,
-0.00016316748224198818,
-0.057153210043907166,
-0.014352444559335709,
0.009274237789213657,
0.08205460011959076,
0.040038034319877625,
-0.040351562201976776,
0.008719312027096748,
-0.012166326865553856,
0.007840239442884922,
-0.05396091192960739,
0.042336273938417435,
-0.012135101482272148,
-0.03211139887571335,
0.03597641363739967,
-0.0025952099822461605,
-0.03942466899752617,
-0.006368446163833141,
0.009827143512666225,
-0.07508687674999237,
0.06974863260984421,
-0.04009547829627991,
-0.004724220372736454,
-0.05853696167469025,
-0.0142340874299407,
-0.026989521458745003,
0.0010596350766718388,
-0.021245423704385757,
-0.010924961417913437,
0.054584961384534836,
-0.019973326474428177,
0.036000560969114304,
-0.033235952258110046,
0.11179684847593307,
0.002077666576951742,
0.03151671588420868,
-0.017759492620825768,
0.020368022844195366,
-0.011239021085202694,
0.05696462094783783,
-0.06255955249071121,
-4.413664811409035e-8,
-0.05610993131995201,
0.08555418252944946,
0.06057307496666908,
0.001680887769907713,
0.04175211489200592,
0.004646475426852703,
0.041287612169981,
0.03598639369010925,
0.017744557932019234,
0.07919907569885254,
0.027676885947585106,
-0.020182104781270027,
0.061517659574747086,
-0.032581876963377,
-0.030320407822728157,
0.11749272048473358,
0.10334354639053345,
-0.04465977102518082,
0.013319217599928379,
0.0174730122089386,
0.06427668780088425,
0.0430196188390255,
0.015807202085852623,
-0.05360524728894234,
0.0950954258441925,
-0.09641383588314056,
-0.10322708636522293,
0.04655344784259796,
0.09948796778917313,
0.028957121074199677,
0.007229306735098362,
-0.005712446756660938,
0.029576042667031288,
0.015876196324825287,
-0.10238765925168991,
0.02353939786553383,
0.05741652473807335,
-0.06311067938804626,
-0.0009216422913596034,
-0.08686622232198715,
0.003709229873493314,
0.046176932752132416,
-0.051867641508579254,
0.005016892217099667,
-0.09549108147621155,
0.07267751544713974,
0.018290452659130096,
-0.08704278618097305,
0.007226551417261362,
-0.006359729915857315,
-0.017232170328497887,
-0.04132583737373352,
-0.02058354951441288,
0.018035635352134705,
0.09950828552246094,
0.02329353056848049,
-0.005698502063751221,
-0.06902828067541122,
0.04466359689831734,
0.06825622171163559,
-0.054402731359004974,
-0.010279691778123379,
-0.05864090099930763,
0.019824080169200897
] |
Chain-of-Focus Prompting: Leveraging Sequential Visual Cues to Prompt Large Autoregressive Vision Models | https://openreview.net/forum?id=noidywkBba | [
"Jiyang Zheng",
"Jialiang Shen",
"Yu Yao",
"Min Wang",
"Yang Yang",
"Dadong Wang",
"Tongliang Liu"
] | Poster | In-context learning (ICL) has revolutionized natural language processing by enabling models to adapt to diverse tasks with only a few illustrative examples. However, the exploration of ICL within the field of computer vision remains limited. Inspired by Chain-of-Thought (CoT) prompting in the language domain, we propose Chain-of-Focus (CoF) Prompting, which enhances vision models by enabling step-by-step visual comprehension. CoF Prompting addresses the challenges of absent logical structure in visual data by generating intermediate reasoning steps through visual saliency. Moreover, it provides a solution for creating tailored prompts from visual inputs by selecting contextually informative prompts based on query similarity and target richness. The significance of CoF prompting is demonstrated by the recent introduction of Large Autoregressive Vision Models (LAVMs), which predict downstream targets via in-context learning with pure visual inputs. By integrating intermediate reasoning steps into visual prompts and effectively selecting the informative ones, the LAVMs are capable of generating significantly better inferences. Extensive experiments on downstream visual understanding tasks validate the effectiveness of our proposed method for visual in-context learning. | Visual In-context Learning | null | 2,040 | null | [
0.05024885758757591,
-0.08616006374359131,
0.04542505368590355,
0.01686035469174385,
0.14602915942668915,
-0.0121464217081666,
0.030047405511140823,
0.032484956085681915,
0.04510999470949173,
-0.039645928889513016,
-0.06345169991254807,
-0.10775286704301834,
0.03200475126504898,
0.03875116631388664,
0.03651498630642891,
0.013945367187261581,
0.07815437763929367,
-0.02046779915690422,
-0.08506885915994644,
-0.022388221696019173,
0.05053291842341423,
-0.06219407916069031,
0.07022212445735931,
0.0028277388773858547,
-0.03795228153467178,
0.01984010636806488,
0.02437080256640911,
-0.038100264966487885,
0.034269534051418304,
0.015982672572135925,
0.017456455156207085,
0.00784410908818245,
0.021689364686608315,
0.08236309885978699,
-0.016721894964575768,
0.0700598880648613,
-0.03663580119609833,
0.044989701360464096,
-0.01502686832100153,
-0.0386631116271019,
-0.06081633269786835,
-0.025754205882549286,
0.014518210664391518,
0.022293327376246452,
0.045745544135570526,
-0.059677671641111374,
-0.038312263786792755,
-0.0037818243727087975,
0.030571915209293365,
-0.002948484616354108,
-0.12903033196926117,
-0.051334578543901443,
-0.01807296834886074,
-0.01102392002940178,
-0.04282870888710022,
0.09682773053646088,
-0.0000989846812444739,
-0.015453112311661243,
0.05908960849046707,
-0.022211603820323944,
-0.07360666990280151,
-0.0795544981956482,
0.0031443273182958364,
0.006735902279615402,
-0.06110704317688942,
-0.0033330563455820084,
0.04062316566705704,
-0.018768228590488434,
-0.03341415524482727,
0.029684701934456825,
-0.01418633759021759,
0.06840593367815018,
-0.016449041664600372,
0.006361254490911961,
0.029653605073690414,
0.0967964306473732,
-0.0488271601498127,
-0.11157923191785812,
0.013859616592526436,
-0.06486165523529053,
0.02776874601840973,
0.0791606679558754,
0.03424486890435219,
0.0025644905399531126,
0.15521293878555298,
0.08261004835367203,
-0.00537100387737155,
-0.0047609382309019566,
-0.01746077835559845,
-0.032122958451509476,
-0.06638482958078384,
-0.12413262575864792,
0.017595283687114716,
0.029442431405186653,
0.08827345818281174,
0.08643431961536407,
-0.011944634839892387,
-0.0917883887887001,
-0.028232403099536896,
0.0036794201005250216,
0.08023006469011307,
0.006097816862165928,
-0.010161525569856167,
-0.07781074941158295,
-0.006298974622040987,
-0.026761194691061974,
0.047433122992515564,
0.0669594332575798,
0.04769118130207062,
-0.08478965610265732,
0.05638524144887924,
0.015043370425701141,
0.048004526644945145,
-0.005263125523924828,
-0.0394330695271492,
-0.0390489287674427,
0.016704929992556572,
-0.05688402056694031,
0.0749608650803566,
-0.04235147312283516,
0.06978469341993332,
0.027204480022192,
-0.04574543237686157,
-0.03573072329163551,
0.039806000888347626,
-0.12427063286304474,
-0.00338730332441628,
6.428010400721729e-34,
0.04297832399606705,
0.01596849039196968,
0.11767075955867767,
0.03052869252860546,
0.06363051384687424,
-0.030466964468359947,
-0.0774616226553917,
0.08401408046483994,
-0.030342496931552887,
0.02784501574933529,
0.014694398269057274,
-0.0019946880638599396,
-0.06663784384727478,
0.08853112906217575,
-0.041327692568302155,
-0.06601575762033463,
-0.061888642609119415,
0.049317408353090286,
-0.08851249516010284,
0.01949378103017807,
-0.0720050036907196,
-0.005169768817722797,
0.011568203568458557,
0.008770929649472237,
0.03353327885270119,
-0.004392966162413359,
0.02261245623230934,
-0.013982649892568588,
-0.012704899534583092,
0.0022398794535547495,
-0.09460023790597916,
-0.010088124312460423,
-0.034847863018512726,
0.05379287898540497,
-0.03187108784914017,
-0.04127207770943642,
0.016948338598012924,
-0.07834658026695251,
0.0415986143052578,
0.008162816986441612,
0.007795429322868586,
0.04258526489138603,
0.0463336706161499,
0.04535960778594017,
-0.08157451450824738,
-0.05799582973122597,
-0.03896990790963173,
0.0908392146229744,
-0.09410331398248672,
0.00048707370297051966,
0.07087539881467819,
0.01073600072413683,
-0.029604827985167503,
-0.06753558665513992,
-0.02191046066582203,
-0.021937938407063484,
0.024036603048443794,
0.02888507768511772,
0.05154893174767494,
-0.016490988433361053,
0.025934314355254173,
0.0052391658537089825,
0.0019341728184372187,
0.11182410269975662,
0.000832182529848069,
0.05347549542784691,
0.0016643459675833583,
0.07606925815343857,
0.053940847516059875,
0.039496373385190964,
-0.009038024581968784,
0.06495659053325653,
-0.06791748851537704,
-0.05326543748378754,
0.07335178554058075,
0.0291033573448658,
-0.0020271255634725094,
-0.045545365661382675,
0.05812937021255493,
0.07303083688020706,
0.014878624118864536,
-0.06282185018062592,
0.0010886503150686622,
-0.0011995017994195223,
-0.01212976686656475,
-0.036359574645757675,
0.0675971508026123,
-0.11427194625139236,
-0.021539464592933655,
-0.060107793658971786,
-0.009441499598324299,
-0.0007991718011908233,
-0.03297210484743118,
-0.01825730688869953,
-0.012706119567155838,
-4.426672034551762e-34,
0.0472591370344162,
0.010912097990512848,
-0.0006167178507894278,
-0.006193696055561304,
-0.04257889837026596,
-0.0095819141715765,
0.07979881018400192,
-0.02828330732882023,
-0.019381234422326088,
-0.1014578714966774,
0.014830972999334335,
-0.0007861992926336825,
-0.0056841932237148285,
0.016920151188969612,
0.004447256680577993,
-0.010228188708424568,
0.0320441797375679,
0.05616588518023491,
-0.03518937528133392,
0.06168961152434349,
0.023150954395532608,
0.02087360993027687,
-0.07939785718917847,
-0.025040388107299805,
-0.10515474528074265,
0.06524590402841568,
-0.03945081681013107,
0.03146236389875412,
-0.027821769937872887,
-0.01616622321307659,
-0.016322100535035133,
0.015145058743655682,
-0.008826845325529575,
-0.03938582167029381,
-0.0976172685623169,
0.07725314050912857,
0.06919185072183609,
-0.0861940011382103,
-0.0966423824429512,
0.09010546654462814,
0.08245047181844711,
-0.08354610204696655,
-0.017285147681832314,
-0.005963269155472517,
-0.09101538360118866,
0.005886419210582972,
0.03070971556007862,
-0.02755350060760975,
-0.003497971687465906,
0.08325202763080597,
-0.05374449864029884,
0.04369070753455162,
-0.08246985822916031,
0.021362947300076485,
-0.07230202108621597,
-0.03491857647895813,
-0.07158102095127106,
-0.00525883212685585,
0.10046329349279404,
-0.05975883826613426,
-0.03884194418787956,
0.04196229204535484,
0.044046998023986816,
-0.053350068628787994,
0.028605125844478607,
-0.061541903764009476,
-0.06851495057344437,
-0.04045279696583748,
0.03623510152101517,
-0.040831200778484344,
0.022169439122080803,
-0.016766052693128586,
0.0012270478764548898,
-0.01642463542521,
0.031739797443151474,
0.0230184905230999,
0.059473782777786255,
-0.05191668123006821,
-0.019920771941542625,
-0.026324506849050522,
0.06003449857234955,
-0.08667390048503876,
0.024148063734173775,
0.10731648653745651,
0.0019667744636535645,
0.06661941111087799,
0.03815782070159912,
0.0560016855597496,
0.05174602195620537,
0.029542144387960434,
-0.04149463400244713,
-0.011641327291727066,
-0.011888198554515839,
0.09775533527135849,
-0.062256697565317154,
-3.565861916854374e-8,
-0.08316246420145035,
0.11355289071798325,
-0.0028210601303726435,
0.01724730059504509,
0.035037554800510406,
-0.07366818189620972,
-0.05277056246995926,
0.08969922363758087,
-0.0055810133926570415,
-0.02228778786957264,
-0.02162947505712509,
-0.02237606793642044,
-0.003980405628681183,
-0.05507486313581467,
0.0062732272781431675,
0.060867790132761,
0.034087713807821274,
0.01174584124237299,
0.02692728117108345,
-0.024966171011328697,
0.02341819927096367,
0.01627669483423233,
-0.012714243493974209,
0.0495285801589489,
0.0028581859078258276,
-0.0331493504345417,
-0.033133577555418015,
0.1241975948214531,
0.06591816246509552,
-0.014051166363060474,
0.033152688294649124,
0.03309598192572594,
-0.009915824979543686,
-0.04934482276439667,
-0.00019726523896679282,
0.024979423731565475,
-0.004224782809615135,
0.014715757220983505,
0.025162195786833763,
-0.0617687813937664,
0.006285998038947582,
0.04135208576917648,
-0.014470656402409077,
0.048607319593429565,
-0.03271283209323883,
0.022463034838438034,
0.02403506077826023,
-0.11150961369276047,
0.025377050042152405,
-0.03668252006173134,
-0.030023710802197456,
-0.04585624486207962,
0.02282026782631874,
0.08970164507627487,
0.027507618069648743,
-0.06016485020518303,
0.023195398971438408,
-0.0009211272117681801,
0.021284371614456177,
0.04212344065308571,
0.04869920387864113,
0.08033878356218338,
-0.033875010907649994,
0.02862200327217579
] |
BOFormer: Learning to Solve Multi-Objective Bayesian Optimization via Non-Markovian RL | https://openreview.net/forum?id=UnCKU8pZVe | [
"Yu Heng Hung",
"Kai-Jie Lin",
"Yu-Heng Lin",
"Chien-Yi Wang",
"Cheng Sun",
"Ping-Chun Hsieh"
] | Poster | Bayesian optimization (BO) offers an efficient pipeline for optimizing black-box functions with the help of a Gaussian process prior and an acquisition function (AF). Recently, in the context of single-objective BO, learning-based AFs witnessed promising empirical results given its favorable non-myopic nature. Despite this, the direct extension of these approaches to multi-objective Bayesian optimization (MOBO) suffer from the hypervolume identifiability issue, which results from the non-Markovian nature of MOBO problems. To tackle this, inspired by the non-Markovian RL literature and the success of Transformers in language modeling, we present a generalized deep Q-learning framework and propose BOFormer, which substantiates this framework for MOBO via sequence modeling. Through extensive evaluation, we demonstrate that BOFormer constantly achieves better performance than the benchmark rule-based and learning-based algorithms in various synthetic MOBO and real-world multi-objective hyperparameter optimization problems. | Multi-Objective Bayesian Optimization, Transformers, Hyperparameter Optimization, Reinforcement Learning, Acquisition Function | null | 2,034 | null | [
-0.06684374809265137,
-0.056662049144506454,
-0.014570488594472408,
-0.02432076260447502,
-0.058871377259492874,
0.0003535226860549301,
-0.02938397414982319,
-0.02905530296266079,
0.04966988414525986,
-0.025408770889043808,
-0.030414974316954613,
-0.04460413381457329,
0.013163024559617043,
0.03178876265883446,
-0.04357733577489853,
0.07938841730356216,
0.04841390624642372,
0.06379294395446777,
-0.060221116989851,
-0.08347685635089874,
0.06927356868982315,
0.10223455727100372,
0.0641314759850502,
0.004719574004411697,
0.05577496066689491,
0.029690468683838844,
0.041957903653383255,
-0.00981216225773096,
-0.018391475081443787,
0.00930603314191103,
0.04559806361794472,
0.012992708012461662,
0.06641606241464615,
-0.0685446634888649,
0.04497816041111946,
0.08642534911632538,
-0.10192565619945526,
-0.01417576801031828,
0.02250761352479458,
-0.025799714028835297,
-0.007355478126555681,
0.015440150164067745,
-0.00735047459602356,
0.033272046595811844,
0.07497408241033554,
-0.026675548404455185,
0.02682294137775898,
-0.0642816573381424,
0.06587968021631241,
-0.011021416634321213,
-0.03754328191280365,
-0.015828914940357208,
-0.07946987450122833,
0.14271125197410583,
0.016001801937818527,
0.10276015847921371,
0.0018943478353321552,
-0.03618175536394119,
0.025359846651554108,
-0.03072899766266346,
-0.12311381846666336,
-0.11463817209005356,
0.02583051100373268,
0.008419777266681194,
-0.005345245823264122,
0.016519544646143913,
-0.018345266580581665,
0.05044938251376152,
-0.07597548514604568,
0.09034699201583862,
-0.05567093938589096,
0.073442243039608,
0.029232008382678032,
-0.00023784206132404506,
0.00044909544521942735,
0.043149739503860474,
0.06707894802093506,
-0.061077289283275604,
0.01987197995185852,
-0.04743688926100731,
0.04771644249558449,
-0.02006206102669239,
0.0175690446048975,
0.0019301867578178644,
0.14946743845939636,
0.027793290093541145,
-0.019215920940041542,
-0.014883385971188545,
0.029996121302247047,
-0.003544391831383109,
-0.05299440771341324,
-0.04242463782429695,
0.013577922247350216,
-0.06895565986633301,
-0.01999768242239952,
-0.03211262449622154,
0.013997812755405903,
-0.04964090883731842,
0.029173314571380615,
0.07767364382743835,
-0.01612689346075058,
0.00822366215288639,
0.09555172175168991,
-0.06007689610123634,
-0.04161350801587105,
-0.06427538394927979,
-0.017766118049621582,
0.04420674964785576,
0.0515613928437233,
-0.05233408883213997,
0.023348722606897354,
0.024553606286644936,
0.09362655133008957,
0.06211477145552635,
-0.04025570675730705,
0.06652431190013885,
0.04626477137207985,
-0.02300918474793434,
-0.05916266888380051,
0.05669691413640976,
-0.06704244017601013,
-0.0637085884809494,
-0.001691400189884007,
0.06821338832378387,
-0.015219350345432758,
0.028625918552279472,
-0.0336444228887558,
1.8138911900874635e-33,
0.00987290684133768,
-0.02412799559533596,
0.05350339040160179,
-0.02371829003095627,
0.007928009144961834,
0.018396183848381042,
0.045854948461055756,
-0.040422648191452026,
-0.03689359128475189,
-0.011537006124854088,
-0.02151140198111534,
0.01974712684750557,
-0.08900411427021027,
-0.0036955196410417557,
0.041290588676929474,
-0.01438696775585413,
-0.09551720321178436,
0.05186939612030983,
0.0028887572698295116,
-0.026795130223035812,
0.04576525837182999,
-0.04066290333867073,
-0.06848929077386856,
-0.02558831125497818,
-0.014595436863601208,
0.04055684804916382,
0.13704735040664673,
-0.06148712709546089,
-0.04312170669436455,
0.016186874359846115,
-0.1430652141571045,
-0.011289580725133419,
-0.04899677634239197,
0.007627403363585472,
0.034687574952840805,
-0.022836964577436447,
-0.05014267936348915,
-0.011384217999875546,
-0.051227957010269165,
-0.026374293491244316,
-0.060771774500608444,
-0.009637209586799145,
0.04208210110664368,
-0.026535823941230774,
-0.11225897818803787,
-0.08152417093515396,
0.010199423879384995,
-0.05518464744091034,
0.041231974959373474,
-0.07628977298736572,
0.04990560933947563,
-0.06730211526155472,
-0.014384928159415722,
0.009604553692042828,
0.07082663476467133,
-0.010456591844558716,
0.00012607622193172574,
0.05401235446333885,
0.026658836752176285,
0.08796561509370804,
0.03245310112833977,
0.07348324358463287,
-0.006001129746437073,
0.06516741216182709,
-0.0030533240642398596,
0.0513821616768837,
-0.003497620113193989,
0.03794423118233681,
0.06712058931589127,
-0.03139800578355789,
0.07205928862094879,
-0.014476362615823746,
0.0561719685792923,
-0.027564208954572678,
0.1046830490231514,
0.0060040648095309734,
0.06383247673511505,
-0.04798101633787155,
-0.0013803148176521063,
0.004467702005058527,
-0.001436376478523016,
0.06897200644016266,
0.016135243698954582,
0.04953524097800255,
0.03171129897236824,
-0.026553906500339508,
0.0016786776250228286,
-0.048981744796037674,
-0.07424288988113403,
0.019982073456048965,
-0.07142820209264755,
-0.00020018735085614026,
-0.0892876610159874,
0.025160223245620728,
-0.036283381283283234,
-1.194084242967565e-33,
-0.014484116807579994,
0.0034449370577931404,
-0.0013545871479436755,
0.040567800402641296,
-0.008316746912896633,
-0.08767823874950409,
-0.018027288839221,
-0.04521218687295914,
0.043037887662649155,
-0.08964632451534271,
-0.03641004115343094,
0.0029186627361923456,
-0.0010631944751366973,
0.005686950869858265,
0.062055569142103195,
0.01914261467754841,
-0.05823405832052231,
-0.013517023995518684,
0.01980886049568653,
0.0040297843515872955,
0.02342926897108555,
0.05204290896654129,
-0.05775173008441925,
0.01186111755669117,
-0.06126148998737335,
0.009118610061705112,
-0.04977212846279144,
0.09422805905342102,
0.021742399781942368,
0.035880375653505325,
-0.06096968799829483,
-0.019143911078572273,
-0.08606404066085815,
-0.01085348054766655,
-0.10774267464876175,
0.11914318054914474,
0.02307077683508396,
0.04329311475157738,
-0.03199096396565437,
0.06645133346319199,
0.004635122139006853,
-0.0049660964868962765,
-0.06456460803747177,
-0.06372734904289246,
-0.031542133539915085,
0.012928280048072338,
-0.001621658680960536,
-0.03297833353281021,
0.07027797400951385,
0.008794697932898998,
-0.06277114152908325,
0.07678961008787155,
-0.06463146954774857,
0.008293320424854755,
-0.11560409516096115,
-0.045532628893852234,
-0.04204398766160011,
-0.04206440970301628,
0.08163701742887497,
0.017790982499718666,
-0.002929139183834195,
-0.004952665418386459,
0.03763779252767563,
-0.07107201963663101,
0.010838400572538376,
0.045976895838975906,
0.06922256201505661,
0.04003814980387688,
0.014239943586289883,
-0.00268532638438046,
0.0167038906365633,
-0.03605285659432411,
0.03477194160223007,
0.08061454445123672,
0.02193663828074932,
-0.012864516116678715,
-0.061362121254205704,
-0.017347192391753197,
0.006722302176058292,
-0.020518982782959938,
0.08070728927850723,
0.08417132496833801,
-0.037227924913167953,
0.10983406007289886,
0.016402380540966988,
0.04945645108819008,
0.09437106549739838,
0.029279524460434914,
0.07258972525596619,
-0.05889567360281944,
0.013810884207487106,
0.013334473595023155,
0.04479552432894707,
0.04222502559423447,
-0.10389762371778488,
-4.186664526173445e-8,
-0.04716021195054054,
-0.005220646969974041,
0.07179279625415802,
-0.014437713660299778,
0.12005090713500977,
-0.020026061683893204,
-0.05015945062041283,
0.028015533462166786,
0.06435336172580719,
-0.08954090625047684,
0.04799239709973335,
-0.01846596784889698,
0.020273856818675995,
-0.04715985432267189,
-0.09044582396745682,
0.06290115416049957,
0.01926547847688198,
-0.02075827307999134,
0.005611536558717489,
-0.06902805715799332,
0.016821688041090965,
0.08341587334871292,
-0.002286250703036785,
0.01757686212658882,
0.032917317003011703,
-0.07040710002183914,
-0.01885451376438141,
-0.02026209607720375,
0.10539110004901886,
0.008754316717386246,
-0.026589414104819298,
0.04408400505781174,
0.02442440576851368,
0.014480890706181526,
-0.05432109162211418,
0.060895271599292755,
0.0015654824674129486,
-0.004653481300920248,
-0.03572995588183403,
-0.04504147171974182,
-0.03563616797327995,
0.07076601684093475,
-0.050077229738235474,
-0.007488916162401438,
0.0019036972662433982,
0.046077027916908264,
-0.07328164577484131,
-0.07446927577257156,
0.057173073291778564,
0.024823907762765884,
0.07496015727519989,
-0.006455320864915848,
0.017234403640031815,
0.048929207026958466,
0.05141373723745346,
-0.06017100811004639,
0.06505919247865677,
-0.061891864985227585,
0.07181044667959213,
0.06115947291254997,
0.00317595642991364,
0.01131298579275608,
-0.055654656141996384,
0.023528626188635826
] |
Deep Signature: Characterization of Large-Scale Molecular Dynamics | https://openreview.net/forum?id=xayT1nn8Mg | [
"Tiexin Qin",
"Mengxu ZHU",
"Chunyang Li",
"Terry Lyons",
"Hong Yan",
"Haoliang Li"
] | Poster | Understanding protein dynamics are essential for deciphering protein functional mechanisms and developing molecular therapies. However, the complex high-dimensional dynamics and interatomic interactions of biological processes pose significant challenge for existing computational techniques. In this paper, we approach this problem for the first time by introducing Deep Signature, a novel computationally tractable framework that characterizes complex dynamics and interatomic interactions based on their evolving trajectories. Specifically, our approach incorporates soft spectral clustering that locally aggregates cooperative dynamics to reduce the size of the system, as well as signature transform that collects iterated integrals to provide a global characterization of the non-smooth interactive dynamics. Theoretical analysis demonstrates that Deep Signature exhibits several desirable properties, including invariance to translation, near invariance to rotation, equivariance to permutation of atomic coordinates, and invariance under time reparameterization. Furthermore, experimental results on three benchmarks of biological processes verify that our approach can achieve superior performance compared to baseline methods. | Molecular dynamics; representation learning; graph neural network; path signature | null | 2,032 | 2410.02847 | [
-0.05408176779747009,
-0.1384781301021576,
0.014323118142783642,
-0.04311343654990196,
-0.016252966597676277,
-0.02661452628672123,
-0.027267664670944214,
-0.012250476516783237,
0.05666474252939224,
-0.1373593509197235,
-0.02891734428703785,
-0.07845905423164368,
-0.03241654857993126,
0.10121749341487885,
-0.04967321828007698,
-0.006480365991592407,
0.009912010282278061,
0.07453721016645432,
-0.0995238721370697,
-0.06439733505249023,
-0.03281031548976898,
-0.0018680235370993614,
0.00713958078995347,
0.02049950137734413,
0.058131903409957886,
0.05472791939973831,
0.01833527721464634,
-0.011093667708337307,
0.013179928995668888,
-0.05686825141310692,
0.04320615157485008,
0.0162644200026989,
-0.041997235268354416,
-0.027042746543884277,
0.026740452274680138,
0.08194175362586975,
0.010254809632897377,
0.004669381771236658,
0.031142672523856163,
-0.05762114375829697,
0.12334570288658142,
0.04828513786196709,
0.036272209137678146,
0.05135481804609299,
0.01717970333993435,
0.027969174087047577,
0.0012206463143229485,
-0.0280712079256773,
-0.0018810391193255782,
0.012924831360578537,
-0.024428589269518852,
0.000038269969081738964,
-0.009338809177279472,
0.06560446321964264,
0.020187683403491974,
0.01123163104057312,
0.022981587797403336,
-0.042294468730688095,
0.02761492319405079,
-0.11851949989795685,
0.06571196764707565,
-0.06499948352575302,
0.04970763996243477,
0.0212747510522604,
0.012005629017949104,
0.007576684467494488,
0.04116363450884819,
0.05870999023318291,
0.001097221509553492,
0.035450346767902374,
-0.027127232402563095,
0.06527436524629593,
-0.06225898489356041,
0.021386681124567986,
0.04987334832549095,
0.023010944947600365,
-0.041730090975761414,
0.03923787549138069,
0.008409849368035793,
-0.056364141404628754,
0.05748912692070007,
-0.013150662183761597,
0.011081038974225521,
0.0000311532239720691,
0.10490130633115768,
0.0019051010021939874,
-0.06631433963775635,
0.013933070935308933,
0.021141229197382927,
0.024120323359966278,
-0.03858567401766777,
0.000954694754909724,
-0.005066609010100365,
-0.06225775182247162,
-0.04098103940486908,
0.003935397136956453,
0.06169577315449715,
-0.07157531380653381,
0.04022812843322754,
0.07683346420526505,
-0.09303638339042664,
0.00619130115956068,
0.013033134862780571,
0.028497181832790375,
0.09957265853881836,
-0.0237283892929554,
0.01557725016027689,
0.016920626163482666,
0.036065198481082916,
0.019764089956879616,
0.04307279735803604,
0.0382721908390522,
0.06807968765497208,
0.12669840455055237,
0.06577777117490768,
0.01019066572189331,
0.05540076270699501,
-0.0022554085589945316,
-0.007350033614784479,
-0.023933419957756996,
-0.033483538776636124,
-0.06894790381193161,
-0.09140970557928085,
0.04448369890451431,
-0.02215605042874813,
0.048641569912433624,
-0.1177247166633606,
3.619692866824858e-34,
0.03911297768354416,
0.033212810754776,
0.06662161648273468,
0.060735322535037994,
0.032162781804800034,
-0.09712359309196472,
-0.05077206343412399,
-0.046988166868686676,
-0.09369497001171112,
0.08024361729621887,
-0.01886129565536976,
0.05673864856362343,
0.049163129180669785,
0.07706556469202042,
-0.030322004109621048,
0.0029859100468456745,
-0.009129778482019901,
0.04918889328837395,
-0.02234465442597866,
-0.05673619359731674,
-0.012131399475038052,
0.039706312119960785,
-0.0415513850748539,
-0.055279653519392014,
-0.07365702092647552,
0.07165216654539108,
-0.019177289679646492,
0.02505023591220379,
0.03262827545404434,
-0.013031075708568096,
-0.05170537531375885,
-0.0081279082223773,
-0.05864842236042023,
0.02215166576206684,
0.10669707506895065,
-0.033752214163541794,
0.09132343530654907,
-0.06507889181375504,
-0.003999833017587662,
-0.06314726918935776,
0.0028293796349316835,
-0.012995514087378979,
-0.11826666444540024,
-0.07179442048072815,
-0.028847353532910347,
0.03217507526278496,
0.040305886417627335,
0.011896051466464996,
-0.03360546752810478,
-0.05306657403707504,
-0.029690539464354515,
-0.07031862437725067,
0.06833542883396149,
-0.027276596054434776,
-0.06057652458548546,
-0.008053203113377094,
0.012222290970385075,
-0.016754524782299995,
-0.004714188165962696,
0.048297859728336334,
-0.014185148291289806,
-0.0007469321717508137,
-0.024881048128008842,
-0.020448295399546623,
0.032443445175886154,
0.02586514689028263,
-0.0825410857796669,
0.05761019140481949,
0.13224104046821594,
0.07683701813220978,
0.013440461829304695,
0.09723932296037674,
-0.006365851499140263,
-0.1071038767695427,
0.06782903522253036,
0.01902003586292267,
0.021476056426763535,
-0.040925532579422,
-0.05330086499452591,
0.10292313247919083,
-0.020324451848864555,
-0.015165169723331928,
-0.025762265548110008,
0.04194667562842369,
-0.01976928487420082,
-0.056562960147857666,
0.0019794711843132973,
-0.06391247361898422,
-0.09057117998600006,
-0.025656072422862053,
-0.03379658982157707,
-0.014919407665729523,
0.023117315024137497,
0.002942961174994707,
-0.017328979447484016,
-2.4778493376822625e-33,
0.03183402493596077,
-0.013551210053265095,
0.03732237219810486,
0.05498794838786125,
0.015337295830249786,
-0.031709250062704086,
-0.0672907680273056,
0.003068929770961404,
0.03253517672419548,
-0.04987557604908943,
0.07720384746789932,
-0.029433762654662132,
-0.01494993083178997,
-0.02424849197268486,
0.04471660777926445,
-0.04214055463671684,
-0.04214225336909294,
0.00946536473929882,
0.1004713624715805,
-0.036920223385095596,
-0.03961661830544472,
-0.025107795372605324,
0.010312463156878948,
0.05877663940191269,
-0.006115565076470375,
0.08096186816692352,
-0.005087761208415031,
-0.007162625901401043,
-0.008967741392552853,
0.012807345017790794,
-0.07901041209697723,
-0.0010878945467993617,
-0.08036786317825317,
0.0031361773144453764,
-0.09958598017692566,
0.08612444996833801,
-0.030752915889024734,
0.00026334921130910516,
0.05082051455974579,
-0.0636480525135994,
-0.005850483197718859,
-0.03030199371278286,
-0.08663444221019745,
0.04454433172941208,
0.05572056025266647,
0.0197931956499815,
-0.0075777387246489525,
0.08564849942922592,
-0.0018905941396951675,
0.06412578374147415,
0.0817204937338829,
-0.0001776947610778734,
-0.018712712451815605,
0.02283537946641445,
-0.046665579080581665,
0.030404014512896538,
0.03879902884364128,
-0.06814807653427124,
0.0818968117237091,
-0.02825227938592434,
-0.07736924290657043,
-0.048167478293180466,
-0.0642494484782219,
0.012425396591424942,
-0.05379472300410271,
0.016369357705116272,
0.05786336213350296,
-0.0427858792245388,
0.011674846522510052,
0.03411054238677025,
0.014817859046161175,
-0.0028513965662568808,
-0.014425477012991905,
0.05985785648226738,
0.08440586924552917,
-0.05892343074083328,
0.01812531240284443,
-0.1300431191921234,
-0.03888261318206787,
-0.07994509488344193,
-0.025472158566117287,
-0.0027055873069912195,
-0.01957782730460167,
-0.03341323137283325,
0.05667649209499359,
0.03834246098995209,
-0.04590008780360222,
0.04361019283533096,
-0.023614609614014626,
0.04740523174405098,
-0.003122097346931696,
0.056149158626794815,
0.003920791670680046,
0.00955031905323267,
-0.05274432897567749,
-4.380924778502049e-8,
-0.011416669003665447,
0.05235113203525543,
0.046409834176301956,
-0.07987435162067413,
0.026413753628730774,
0.06540658324956894,
0.012864181771874428,
0.04216548800468445,
0.028206776827573776,
0.020457733422517776,
0.09189770370721817,
-0.09877456724643707,
-0.014036632142961025,
-0.05036735534667969,
-0.035395096987485886,
0.059619445353746414,
-0.037432897835969925,
-0.02755153924226761,
-0.08822726458311081,
-0.03270157799124718,
-0.0346740186214447,
0.06532080471515656,
-0.03945111110806465,
-0.041762370616197586,
0.04187089949846268,
-0.08805359154939651,
-0.02641848288476467,
0.0353083573281765,
0.040730852633714676,
-0.03582141548395157,
-0.029885033145546913,
0.06719990074634552,
0.051265887916088104,
0.018419507890939713,
-0.08881230652332306,
0.038809485733509064,
-0.022604482248425484,
-0.016820739954710007,
-0.03900282829999924,
-0.009464231319725513,
-0.04920722916722298,
0.14048945903778076,
-0.0994272232055664,
-0.035531021654605865,
-0.03455941379070282,
0.060688361525535583,
-0.011478726752102375,
0.01252852100878954,
0.07452114671468735,
0.05116838589310646,
0.05025697126984596,
-0.016729062423110008,
0.011175403371453285,
0.021240321919322014,
-0.018173011019825935,
0.04702165722846985,
-0.054031018167734146,
-0.07274135202169418,
0.11563277989625931,
0.014424142427742481,
-0.013896123506128788,
-0.05443210154771805,
-0.07218621671199799,
-0.0706895962357521
] |
RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation | https://openreview.net/forum?id=yAzN4tz7oI | [
"Songming Liu",
"Lingxuan Wu",
"Bangguo Li",
"Hengkai Tan",
"Huayu Chen",
"Zhengyi Wang",
"Ke Xu",
"Hang Su",
"Jun Zhu"
] | Poster | Bimanual manipulation is essential in robotics, yet developing foundation models is extremely challenging due to the inherent complexity of coordinating two robot arms (leading to multi-modal action distributions) and the scarcity of training data. In this paper, we present the Robotics Diffusion Transformer (RDT), a pioneering diffusion foundation model for bimanual manipulation. RDT builds on diffusion models to effectively represent multi-modality, with innovative designs of a scalable Transformer to deal with the heterogeneity of multi-modal inputs and to capture the nonlinearity and high frequency of robotic data. To address data scarcity, we further introduce a Physically Interpretable Unified Action Space, which can unify the action representations of various robots while preserving the physical meanings of original actions, facilitating learning transferrable physical knowledge. With these designs, we managed to pre-train RDT on the largest collection of multi-robot datasets to date and scaled it up to $1.2$B parameters, which is the largest diffusion-based foundation model for robotic manipulation. We finally fine-tuned RDT on a self-created multi-task bimanual dataset with over $6$K+ episodes to refine its manipulation capabilities. Experiments on real robots demonstrate that RDT significantly outperforms existing methods. It exhibits zero-shot generalization to unseen objects and scenes, understands and follows language instructions, learns new skills with just 1$\sim$5 demonstrations, and effectively handles complex, dexterous tasks. We refer to https://rdt-robotics.github.io/rdt-robotics/ for the code and videos. | robot learning, diffusion models, foundation models, bimanual manipulation | null | 2,031 | null | [
0.002718918025493622,
-0.1586717665195465,
0.02922035939991474,
-0.020965436473488808,
0.008466132916510105,
-0.04094501584768295,
-0.0477195642888546,
-0.034604109823703766,
0.029746918007731438,
0.015571928583085537,
0.038686029613018036,
-0.0005676838918589056,
-0.0014603475574404001,
0.07435085624456406,
0.04299163818359375,
0.03106137365102768,
0.06623061746358871,
0.05868136137723923,
-0.1470683515071869,
-0.006338253617286682,
0.018890736624598503,
0.01578589715063572,
0.03166532143950462,
-0.012849066406488419,
-0.016576755791902542,
0.05540278181433678,
0.03355535864830017,
-0.14636506140232086,
-0.04141255468130112,
-0.09867870807647705,
0.03719460964202881,
-0.017367014661431313,
-0.07754313945770264,
-0.028273576870560646,
-0.06012819707393646,
0.07550149410963058,
-0.03175203502178192,
-0.004484830889850855,
-0.013122235424816608,
-0.07300777733325958,
0.005082655232399702,
0.026926327496767044,
0.05335095152258873,
0.019016843289136887,
0.05589992552995682,
0.0704227015376091,
-0.008714483119547367,
-0.0839543342590332,
0.0870196595788002,
0.06650040298700333,
-0.024380717426538467,
-0.02704487182199955,
0.053412508219480515,
0.06507164984941483,
-0.029647663235664368,
-0.02494148723781109,
0.03653091937303543,
0.009827236644923687,
0.0845661610364914,
-0.08054910600185394,
-0.0017061269609257579,
-0.020047731697559357,
0.04354565590620041,
-0.05386274680495262,
-0.05791134387254715,
-0.0574546679854393,
0.04021450877189636,
-0.00987029355019331,
-0.01294130552560091,
-0.0035736027639359236,
0.004848993383347988,
-0.004308388568460941,
-0.08917859941720963,
-0.009811336174607277,
0.03380238637328148,
-0.031761012971401215,
-0.0022522632498294115,
-0.03423603996634483,
-0.02504952996969223,
0.004421908874064684,
0.006435369607061148,
-0.017919428646564484,
0.04915003851056099,
0.05601757392287254,
0.06128867715597153,
-0.003567627863958478,
0.012185142375528812,
0.06655805557966232,
0.0029645187314599752,
0.022824568673968315,
-0.05415334925055504,
0.03912457078695297,
0.010393860749900341,
-0.06622537225484848,
-0.0408061221241951,
-0.023154404014348984,
-0.0387125089764595,
-0.025643501430749893,
0.028452277183532715,
0.06633897125720978,
-0.008145436644554138,
0.03266765549778938,
-0.030819131061434746,
0.05107995867729187,
0.0016168879810720682,
-0.08789858967065811,
0.038281846791505814,
-0.022228917106986046,
0.04347769916057587,
-0.0018602465279400349,
0.042491938918828964,
-0.03913216292858124,
0.04719478636980057,
0.007669358979910612,
0.04402628540992737,
-0.020354771986603737,
0.01988532766699791,
0.017833467572927475,
-0.00454509723931551,
0.0679711177945137,
0.03959006071090698,
-0.006689384113997221,
-0.04477265104651451,
0.0620296448469162,
0.061838652938604355,
-0.011536039412021637,
-0.04143445938825607,
5.37994721395701e-35,
0.043990153819322586,
0.024121100082993507,
0.0683836042881012,
0.05581526830792427,
0.04105983301997185,
-0.025373606011271477,
0.047013767063617706,
0.007462742272764444,
0.02124708518385887,
-0.09086869657039642,
-0.011244367808103561,
0.09932871162891388,
-0.051251135766506195,
0.10864086449146271,
-0.014051287434995174,
-0.028565851971507072,
0.0297758299857378,
0.003434192854911089,
0.013486572541296482,
0.01503387838602066,
0.04365220293402672,
0.016895338892936707,
-0.07784438878297806,
0.008701146580278873,
-0.0041078077629208565,
-0.034812081605196,
0.029376626014709473,
0.023348307237029076,
0.01253857184201479,
-0.011180747300386429,
-0.17352943122386932,
0.027499007061123848,
-0.06230705976486206,
0.023502487689256668,
0.05243772640824318,
-0.034510355442762375,
-0.06778492778539658,
-0.05797542259097099,
0.0010490297572687268,
-0.051528725773096085,
0.021367361769080162,
-0.004328893963247538,
-0.0704631581902504,
-0.07079778611660004,
-0.11149603128433228,
-0.03942761942744255,
0.032007668167352676,
0.04280192032456398,
-0.02910006232559681,
-0.06968016177415848,
-0.0035781837068498135,
0.03340664878487587,
-0.024288825690746307,
-0.06974497437477112,
0.035477302968502045,
-0.05571013316512108,
0.05517246946692467,
-0.0058292183093726635,
0.06353713572025299,
0.05934123694896698,
0.012528759427368641,
0.08546589314937592,
0.012404709123075008,
0.13433024287223816,
0.03603322431445122,
-0.019465787336230278,
-0.008518502116203308,
0.10530412942171097,
0.1323545128107071,
-0.017245393246412277,
-0.09962218999862671,
0.012638358399271965,
0.02176431380212307,
-0.0846216157078743,
0.0281805619597435,
-0.045445188879966736,
0.020851606503129005,
-0.09754641354084015,
0.017208416014909744,
0.037608224898576736,
-0.060674864798784256,
-0.023371580988168716,
0.002707043197005987,
-0.028080737218260765,
-0.009853926487267017,
-0.012884081341326237,
-0.012290406972169876,
-0.09353876113891602,
-0.0905834436416626,
-0.03583424165844917,
-0.07953698188066483,
-0.07066264003515244,
-0.1223207414150238,
0.03227157145738602,
0.02457774430513382,
-1.2733520514763352e-33,
0.019137633964419365,
0.02672618255019188,
0.03225129842758179,
0.08180209994316101,
-0.0014822219964116812,
-0.03332636132836342,
-0.032817091792821884,
0.014970174059271812,
0.05556754395365715,
-0.013259200379252434,
-0.03530769422650337,
-0.06443871557712555,
-0.023776991292834282,
-0.02187269553542137,
0.06468962877988815,
-0.010252811014652252,
0.01735428161919117,
0.023454707115888596,
-0.032712556421756744,
-0.0014000150840729475,
-0.04561275616288185,
0.11329329013824463,
-0.0823211744427681,
0.048463065177202225,
-0.09885552525520325,
0.0489070862531662,
-0.0675458237528801,
-0.009391178376972675,
0.028835514560341835,
0.054210115224123,
-0.06675680726766586,
-0.03763227164745331,
0.0473615862429142,
0.03984379768371582,
-0.05751889944076538,
0.06836961209774017,
0.012011343613266945,
0.0452287383377552,
0.035871535539627075,
-0.005194587167352438,
0.016563642770051956,
-0.0891360267996788,
-0.06556377559900284,
0.04241698235273361,
0.023305023089051247,
-0.07923835515975952,
-0.08244638890028,
0.021830210462212563,
-0.02891770750284195,
0.048421092331409454,
-0.005422627553343773,
0.09849425405263901,
-0.08194834738969803,
-0.06680446863174438,
-0.009637037292122841,
-0.09531795233488083,
0.008569381199777126,
-0.034658920019865036,
0.07618238776922226,
-0.040585704147815704,
-0.047482557594776154,
0.032382939010858536,
0.018769290298223495,
0.040582217276096344,
-0.042415980249643326,
-0.0074020130559802055,
-0.026725322008132935,
-0.02340429462492466,
0.012250836938619614,
0.061836615204811096,
-0.002799038775265217,
0.018626874312758446,
0.10160525888204575,
-0.019806262105703354,
0.07616449147462845,
-0.04582921788096428,
0.048283081501722336,
-0.06366287916898727,
-0.005566858220845461,
0.0057329353876411915,
0.0005012205801904202,
0.011663644574582577,
0.06438003480434418,
0.05457286536693573,
-0.0049581220373511314,
0.11303962767124176,
-0.01161987241357565,
0.03411119803786278,
0.01391698233783245,
0.04324674606323242,
0.005235469434410334,
-0.00856755767017603,
0.011529285460710526,
0.11260251700878143,
-0.05544569715857506,
-4.233421790900138e-8,
-0.058632444590330124,
0.07947277277708054,
0.035724613815546036,
0.0413472056388855,
-0.028031470254063606,
0.007623033132404089,
0.034016747027635574,
0.05085406079888344,
-0.03217298537492752,
0.055338043719530106,
0.07857061177492142,
-0.023079009726643562,
-0.018884098157286644,
-0.07386123389005661,
-0.021587753668427467,
0.07936146855354309,
0.02069367654621601,
-0.019553573802113533,
0.03537464886903763,
0.044805631041526794,
0.05267607048153877,
-0.029975904151797295,
0.03155013546347618,
-0.046367347240448,
0.056854888796806335,
-0.020917434245347977,
-0.13938821852207184,
0.04479920491576195,
0.037598010152578354,
0.024037448689341545,
0.03274634853005409,
-0.06692053377628326,
0.039724864065647125,
-0.0002191343519371003,
-0.004984243307262659,
-0.02252490073442459,
0.018711905926465988,
-0.023107357323169708,
-0.04604437202215195,
-0.038729894906282425,
0.01019029039889574,
0.0922182947397232,
-0.06777223944664001,
0.03736159950494766,
-0.04408128932118416,
0.0387086495757103,
-0.044027917087078094,
-0.08827967196702957,
0.00568963261321187,
0.014673145487904549,
0.0011790608987212181,
-0.012494340538978577,
0.006893930025398731,
0.09134524315595627,
0.06519342958927155,
0.10530871152877808,
0.058472249656915665,
-0.0643821656703949,
0.041143693029880524,
0.028354700654745102,
0.022251568734645844,
-0.01113726943731308,
-0.071940116584301,
0.030632352456450462
] |
VILA-U: a Unified Foundation Model Integrating Visual Understanding and Generation | https://openreview.net/forum?id=02haSpO453 | [
"Yecheng Wu",
"Zhuoyang Zhang",
"Junyu Chen",
"Haotian Tang",
"Dacheng Li",
"Yunhao Fang",
"Ligeng Zhu",
"Enze Xie",
"Hongxu Yin",
"Li Yi",
"Song Han",
"Yao Lu"
] | Poster | VILA-U is a Unified foundation model that integrates Video, Image, Language understanding and generation. Traditional visual language models (VLMs) use separate modules for understanding and generating visual content, which can lead to misalignment and increased complexity. In contrast, VILA-U employs a single autoregressive next-token prediction framework for both tasks, eliminating the need for additional components like diffusion models. This approach not only simplifies the model but also achieves near state-of-the-art performance in visual language understanding and generation. The success of VILA-U is attributed to two main factors: the unified vision tower that aligns discrete visual tokens with textual inputs during pretraining, which enhances visual perception, and autoregressive image generation can achieve similar quality as diffusion models with high-quality dataset. This allows VILA-U to perform comparably to more complex models using a fully token-based autoregressive framework. | Unified Visual Language Model, Autoregressive Model | VILA-U is a Unified foundation model that integrates Video, Image, Language understanding and generation. It employs a single autoregressive next-token prediction framework for both tasks. | 2,027 | null | [
0.03699488565325737,
-0.10723876953125,
-0.005505210720002651,
-0.024822335690259933,
0.05800383538007736,
0.018442170694470406,
-0.07461950182914734,
-0.022452589124441147,
0.07685541361570358,
-0.06966240704059601,
0.009578241035342216,
-0.10297486186027527,
0.0035053391475230455,
0.0576552115380764,
-0.01805521361529827,
-0.04570765048265457,
0.014352284371852875,
0.05091509968042374,
-0.08335421979427338,
-0.08582441508769989,
0.03995606675744057,
-0.008200092241168022,
-0.04589144140481949,
-0.025766979902982712,
0.05387667939066887,
0.09947150200605392,
-0.0029103984124958515,
-0.060207054018974304,
0.10081899166107178,
-0.06954190880060196,
-0.01360530685633421,
-0.011194399558007717,
0.07725542038679123,
0.03666168823838234,
-0.0790458545088768,
0.0348360538482666,
-0.03886512666940689,
0.003964030649513006,
-0.0462995246052742,
-0.0712943747639656,
-0.0726342424750328,
-0.0049075959250330925,
0.039363931864500046,
0.04552527517080307,
0.1462707221508026,
0.018023282289505005,
0.020647555589675903,
-0.006054116412997246,
0.006198618095368147,
-0.011432355269789696,
-0.1000836193561554,
-0.05352050065994263,
-0.0026471104938536882,
-0.011207925155758858,
-0.07576414942741394,
0.0920986607670784,
0.05087814852595329,
-0.000614935765042901,
0.06680003553628922,
-0.06672999262809753,
-0.03908693790435791,
-0.02277463860809803,
0.020797381177544594,
0.02212545834481716,
-0.007604943588376045,
-0.02921346202492714,
0.04821182042360306,
0.03971119225025177,
0.014306488446891308,
-0.01041021291166544,
-0.055075474083423615,
0.037849582731723785,
0.0020129031036049128,
0.017351092770695686,
0.0018486261833459139,
0.08583562076091766,
0.03389037400484085,
-0.04118678718805313,
0.05956004187464714,
-0.05978662893176079,
0.06461093574762344,
0.043273430317640305,
0.08065370470285416,
-0.030956709757447243,
0.0977870523929596,
0.03438596427440643,
0.009305275045335293,
0.014574022963643074,
-0.012058664113283157,
-0.022656463086605072,
-0.1031872034072876,
-0.07785477489233017,
0.03918919339776039,
-0.008724157698452473,
0.035190731287002563,
0.017699044197797775,
0.003086603945121169,
-0.08014608919620514,
-0.009394676424562931,
0.05760834738612175,
0.03479147329926491,
0.02072484791278839,
0.04143788293004036,
0.00720618013292551,
0.06413517892360687,
-0.012132855132222176,
0.13113261759281158,
0.0733625739812851,
0.04966948553919792,
-0.03182217851281166,
0.024459360167384148,
0.040906015783548355,
-0.01830758899450302,
0.011069797910749912,
0.00022605755657423288,
-0.07016152143478394,
-0.011166735552251339,
-0.058158185333013535,
0.09277652204036713,
0.04032246023416519,
-0.009559988975524902,
-0.021757619455456734,
-0.060985930263996124,
0.01640062779188156,
0.02160360850393772,
-0.08095484226942062,
0.017561711370944977,
1.7035840859032463e-33,
-0.02491471730172634,
0.008198525756597519,
0.02513444423675537,
0.03133761137723923,
0.06861067563295364,
-0.019737141206860542,
-0.056464433670043945,
-0.05919425189495087,
-0.006563597358763218,
-0.039453212171792984,
-0.005714929662644863,
0.010798518545925617,
-0.08189217746257782,
0.2208264321088791,
0.044930290430784225,
-0.028971796855330467,
-0.03501319885253906,
0.0028266161680221558,
-0.011881282553076744,
0.00011955407535424456,
0.0006238119094632566,
0.012726475484669209,
0.016245638951659203,
-0.061825573444366455,
0.019245781004428864,
0.0848030298948288,
0.07268748432397842,
-0.05348917841911316,
-0.04296224191784859,
0.023815451189875603,
-0.1316688507795334,
0.005846004001796246,
-0.03250190615653992,
0.008930908516049385,
0.013206446543335915,
-0.010726909153163433,
0.03779127448797226,
-0.060049936175346375,
0.0005435916827991605,
-0.0215382669121027,
-0.0346413291990757,
0.08672100305557251,
-0.04476321488618851,
-0.06729932129383087,
-0.07703433930873871,
0.031625859439373016,
0.03735378012061119,
0.07366963475942612,
-0.10503439605236053,
-0.028921963647007942,
0.010725464671850204,
0.001416057231836021,
-0.06393004953861237,
-0.05393442511558533,
-0.00640584621578455,
-0.030322005972266197,
0.031084831804037094,
0.020137403160333633,
0.03216183930635452,
-0.023213330656290054,
-0.020756935700774193,
0.031839217990636826,
0.027707437053322792,
0.1100088432431221,
0.00116728397551924,
-0.004946139175444841,
0.02266993746161461,
0.06182314082980156,
0.0291224904358387,
-0.02532075345516205,
-0.08719789981842041,
0.02650357410311699,
0.004346549976617098,
-0.08335991203784943,
-0.007173852063715458,
-0.026355931535363197,
0.043224502354860306,
-0.02259395457804203,
-0.0023330405820161104,
0.12308064848184586,
-0.0700971707701683,
-0.04787697643041611,
0.012312682345509529,
-0.10764847695827484,
0.03490006923675537,
-0.02015138417482376,
0.09056229889392853,
-0.07837031036615372,
-0.04815908521413803,
-0.11359408497810364,
-0.014972887001931667,
0.000012282571333344094,
0.07496959716081619,
0.06036914139986038,
0.04548484832048416,
-2.258385460738703e-33,
0.023131268098950386,
0.09257952123880386,
0.05795879289507866,
0.03807378560304642,
-0.044756531715393066,
-0.06278442591428757,
0.052297577261924744,
0.013416082598268986,
-0.021986065432429314,
-0.08862901479005814,
-0.011461933143436909,
0.006074282806366682,
-0.03945978730916977,
-0.0083994772285223,
-0.004845588468015194,
-0.02755652740597725,
0.0661339983344078,
-0.03287307918071747,
-0.0023787873797118664,
0.030449384823441505,
0.0030623022466897964,
0.019535241648554802,
-0.1195668876171112,
-0.02250724285840988,
-0.032432787120342255,
0.058651071041822433,
-0.020615991204977036,
0.06070525571703911,
-0.07847951352596283,
-0.04491526633501053,
0.006824378855526447,
0.009566057473421097,
0.043507643043994904,
-0.019529011100530624,
-0.09054691344499588,
0.03170763701200485,
0.038178734481334686,
-0.06275711208581924,
-0.031065288931131363,
0.1046430841088295,
0.073628731071949,
-0.019603248685598373,
-0.07433327287435532,
-0.03470950946211815,
-0.07365601509809494,
0.03708520531654358,
-0.026210399344563484,
-0.008356224745512009,
-0.00508374348282814,
-0.03977359086275101,
0.02298043854534626,
0.07559289783239365,
-0.10815750807523727,
-0.005018238909542561,
-0.045945413410663605,
-0.07774914801120758,
0.00666175177320838,
0.03883695602416992,
0.034277282655239105,
-0.02202019840478897,
-0.01948658376932144,
-0.00396647909656167,
0.012688097544014454,
-0.05786474049091339,
-0.013053765520453453,
-0.034161195158958435,
-0.03584326058626175,
-0.012397866696119308,
-0.012969676405191422,
0.015508627519011497,
0.040982987731695175,
-0.0018237062031403184,
0.002631603041663766,
0.06570509076118469,
0.014492107555270195,
-0.06789667904376984,
0.010413105599582195,
-0.03481258824467659,
0.021049439907073975,
-0.06689099967479706,
-0.045750971883535385,
-0.08242665976285934,
0.042147208005189896,
0.08991248905658722,
0.0328977145254612,
0.0813656821846962,
0.03185706585645676,
-0.005735163111239672,
0.07118722051382065,
0.04751870036125183,
-0.040129199624061584,
0.037973981350660324,
0.00408431189134717,
0.08953530341386795,
-0.048213984817266464,
-4.2449809001254835e-8,
-0.09757161885499954,
0.004571774508804083,
-0.014258998446166515,
-0.023368634283542633,
-0.023596487939357758,
-0.03733859956264496,
-0.02348579838871956,
0.061363063752651215,
-0.004944636486470699,
-0.029897566884756088,
0.010237139649689198,
-0.03914593532681465,
-0.08607850223779678,
-0.03463411703705788,
0.02076428011059761,
0.0899769738316536,
0.03941800445318222,
0.051437973976135254,
0.029486950486898422,
-0.018421899527311325,
0.00904130470007658,
0.02915181964635849,
-0.0035272410605102777,
-0.013757283799350262,
-0.018877390772104263,
-0.02926204539835453,
-0.02912629209458828,
0.05840336158871651,
0.051390379667282104,
-0.08479730784893036,
0.029812928289175034,
0.09812112152576447,
0.014641975983977318,
-0.05437145754694939,
-0.00438443198800087,
0.050537459552288055,
0.036118291318416595,
-0.039391789585351944,
-0.01164672989398241,
-0.012699494138360023,
0.040457338094711304,
0.05468664690852165,
-0.02666575275361538,
0.015120119787752628,
0.00233480054885149,
0.032193880528211594,
0.0746980756521225,
-0.0826343297958374,
-0.0055264318361878395,
-0.05988387018442154,
-0.0010653040371835232,
0.004261775407940149,
-0.005282538942992687,
0.07903973013162613,
0.09863583743572235,
0.015052716247737408,
0.046970028430223465,
-0.028471730649471283,
0.0936238244175911,
0.06121615320444107,
0.034851640462875366,
-0.016418281942605972,
-0.04786008596420288,
-0.030438734218478203
] |
Feast Your Eyes: Mixture-of-Resolution Adaptation for Multimodal Large Language Models | https://openreview.net/forum?id=1EnpStvBU8 | [
"Gen Luo",
"Yiyi Zhou",
"Yuxin Zhang",
"Xiawu Zheng",
"Xiaoshuai Sun",
"Rongrong Ji"
] | Poster | In existing multimodal large language models (MLLMs), image resolution plays a significant role for granular visual recognition. However, directly increasing image resolution leads to expensive computational cost for MLLMs. In this paper, we reveal that a combination of low- and high-resolution visual features can efficiently mitigate this shortcoming. Based on this principle, we propose a novel and efficient method for MLLMs, termed Mixture-of-Resolution Adaptation (MRA). In particular, MRA adopts two visual pathways for images of different resolutions, where high-resolution visual information is embedded into the low-resolution pathway via the novel mixture-of-resolution adapters (MR-Adapters). This design also greatly reduces the input sequence length of MLLMs. To validate MRA, we apply it to a recent MLLM called LLaVA, and term the new model LLaVA-HR. We conduct extensive experiments on 17 vision-language (VL) tasks, which show that LLaVA-HR outperforms existing MLLMs on 15 VL tasks, e.g., +5.2\% on TextVQA. More importantly, both training and inference of LLaVA-HR remain efficient with MRA, e.g., 20 training hours and faster inference speed than LLaVA-NeXT. Source codes are released at: https://github.com/luogen1996/LLaVA-HR. | high-resolution adaptation, multimodal large language models | null | 2,025 | 2403.03003 | [
0.015425525605678558,
-0.08414404839277267,
0.063376285135746,
-0.04816924408078194,
0.08498265594244003,
-0.03390609100461006,
-0.046539418399333954,
-0.006464557256549597,
0.048263922333717346,
-0.05467985197901726,
-0.049964260309934616,
-0.044788118451833725,
0.0381486639380455,
0.03447234258055687,
-0.00861483532935381,
-0.00031264591962099075,
0.033262357115745544,
0.10906021296977997,
-0.11602570116519928,
-0.04962649568915367,
0.055092256516218185,
-0.05074159801006317,
0.04629943147301674,
-0.04653629660606384,
0.038600385189056396,
0.05737590417265892,
-0.02087368443608284,
-0.056630223989486694,
0.0034709821920841932,
-0.043492428958415985,
-0.008486481383442879,
0.09877567738294601,
0.02759401872754097,
0.029414601624011993,
-0.07140226662158966,
0.04455431178212166,
-0.04705459997057915,
0.07383803278207779,
-0.03814351186156273,
-0.0799747109413147,
0.0034976606257259846,
-0.010065360926091671,
0.026238810271024704,
-0.03591779246926308,
0.030111467465758324,
0.05165194347500801,
0.021352684125304222,
-0.02518276683986187,
0.02049758844077587,
-0.046942345798015594,
-0.07678060233592987,
-0.05037474259734154,
-0.021973511204123497,
0.06581123918294907,
-0.09352580457925797,
0.004890777636319399,
0.015527894720435143,
0.0020371705759316683,
0.03897463530302048,
0.05772673711180687,
-0.024669000878930092,
-0.02514113113284111,
0.018463047221302986,
0.004405986983329058,
0.018554789945483208,
-0.030448634177446365,
0.034783024340867996,
-0.07081259042024612,
-0.034001994878053665,
-0.004987853113561869,
-0.05676284432411194,
0.04132432863116264,
-0.009194313548505306,
0.010843227617442608,
-0.020614592358469963,
0.01586257293820381,
0.030876748263835907,
-0.05323272570967674,
0.12779974937438965,
-0.03617188334465027,
0.008158029988408089,
0.06022031232714653,
-0.023303749039769173,
-0.015005302615463734,
0.16786202788352966,
-0.005964288953691721,
-0.0062246499583125114,
0.023801345378160477,
0.02462032437324524,
-0.048801034688949585,
-0.10990697890520096,
-0.08401094377040863,
0.00207910081371665,
-0.008025643415749073,
-0.044080689549446106,
0.008860413916409016,
0.05921438708901405,
-0.07856584340333939,
0.023462139070034027,
0.0757879689335823,
0.030586732551455498,
-0.03393504396080971,
0.045067984610795975,
-0.011913839727640152,
0.042453162372112274,
-0.009301699697971344,
0.10177453607320786,
0.0438946932554245,
0.01962130144238472,
-0.052073460072278976,
0.08505584299564362,
0.052149251103401184,
0.07252843677997589,
0.03324095159769058,
0.014528324827551842,
-0.050698742270469666,
0.018493829295039177,
-0.0007424128125421703,
0.01300265546888113,
0.0008409571601077914,
-0.0209935512393713,
-0.009547637775540352,
-0.009303692728281021,
-0.07726723700761795,
0.06753557920455933,
0.005593383219093084,
-0.014668893069028854,
1.329246807857935e-33,
0.028964051976799965,
0.04642251133918762,
0.037492044270038605,
0.08341419696807861,
0.03783544898033142,
0.003239770419895649,
-0.01354713924229145,
-0.0063576954416930676,
0.06143864616751671,
-0.007345432415604591,
-0.006709489040076733,
-0.014814314432442188,
-0.10627290606498718,
0.16899171471595764,
-0.01976054534316063,
0.011290322989225388,
-0.01124480925500393,
0.06250452995300293,
0.02569831907749176,
-0.013020898215472698,
0.03252927586436272,
-0.06868760287761688,
0.051164526492357254,
-0.011820263229310513,
-0.016465360298752785,
0.02343970164656639,
0.06196358799934387,
-0.11655593663454056,
-0.023122195154428482,
-0.02002696320414543,
-0.10029081255197525,
0.01055955421179533,
0.04786901921033859,
0.04059427231550217,
0.010550680570304394,
-0.06310673803091049,
-0.0024803904816508293,
-0.022344935685396194,
-0.0068780118599534035,
0.003175517776980996,
-0.024256249889731407,
0.016929341480135918,
0.005846851039677858,
-0.06773925572633743,
-0.05256963148713112,
0.010341450572013855,
0.01905764825642109,
0.08738084882497787,
-0.0744975134730339,
-0.031223146244883537,
0.03769165650010109,
-0.05190145969390869,
-0.03148750215768814,
-0.07705572247505188,
-0.021075695753097534,
-0.029422344639897346,
0.008124816231429577,
0.013808449730277061,
0.05621550232172012,
-0.011663164012134075,
-0.001513439929112792,
-0.024839457124471664,
-0.03385402262210846,
0.1367669403553009,
0.11848696321249008,
0.01354095246642828,
0.000039937607652973384,
0.06524666398763657,
-0.04037494212388992,
0.00352339050732553,
-0.00723362248390913,
0.04600135609507561,
0.057993534952402115,
-0.002394805196672678,
-0.026040438562631607,
-0.016738442704081535,
0.0638689324259758,
-0.052774567157030106,
0.04583295062184334,
0.11068196594715118,
0.013004723004996777,
0.03271491825580597,
0.01936856284737587,
-0.08283565193414688,
-0.00559916440397501,
0.025935158133506775,
0.02200888656079769,
-0.09694024920463562,
-0.025403421372175217,
-0.1422511488199234,
-0.03343002125620842,
0.0029286693315953016,
0.016502847895026207,
-0.057770371437072754,
0.03539164736866951,
-1.2579039434901187e-33,
0.04252518713474274,
-0.007536131888628006,
0.06101056560873985,
0.010245779529213905,
-0.026366978883743286,
-0.01581715978682041,
0.09360482543706894,
0.002456316025927663,
-0.01000627875328064,
-0.0880330428481102,
-0.05823485553264618,
-0.04328148439526558,
0.004354905337095261,
-0.060512930154800415,
0.027106888592243195,
-0.02917073853313923,
0.021289188414812088,
0.018675781786441803,
-0.01943615823984146,
0.04506289213895798,
0.008517717942595482,
0.07931079715490341,
-0.05651085823774338,
-0.007559103425592184,
-0.08834667503833771,
-0.0007389771053567529,
-0.058577679097652435,
0.06688423454761505,
-0.055317435413599014,
-0.00200779945589602,
0.03313878923654556,
-0.004166330676525831,
-0.036279890686273575,
-0.060886647552251816,
-0.03442779928445816,
0.06530603766441345,
0.02802330069243908,
-0.05662141740322113,
-0.059086255729198456,
0.12489189952611923,
0.06566761434078217,
-0.06602168828248978,
-0.023001743480563164,
0.005903271492570639,
0.007631086744368076,
-0.08750824630260468,
-0.0829535648226738,
-0.09185755997896194,
0.057054366916418076,
0.0334971658885479,
-0.018991759046912193,
-0.032732684165239334,
-0.042237788438797,
-0.016270486637949944,
-0.012684663757681847,
-0.0467895083129406,
-0.04714161902666092,
-0.027315041050314903,
0.06341969221830368,
-0.03093677945435047,
-0.040996696799993515,
-0.05803549662232399,
-0.09442856162786484,
-0.026474617421627045,
-0.013842769898474216,
0.061238761991262436,
-0.007835421711206436,
-0.08079390972852707,
0.045407239347696304,
0.01318309549242258,
0.052940431982278824,
-0.13299882411956787,
0.011052235960960388,
-0.02947728894650936,
0.003436377504840493,
-0.018046988174319267,
-0.003015882568433881,
0.055296141654253006,
0.025034412741661072,
-0.050877466797828674,
0.004425927065312862,
-0.07590675354003906,
-0.0028783029410988092,
0.04050680622458458,
0.08311984688043594,
0.09445624053478241,
-0.0030373944900929928,
0.029216265305876732,
0.08391004800796509,
0.03462514653801918,
-0.05425116792321205,
0.060822993516922,
0.02486267127096653,
0.02855340763926506,
0.01904924027621746,
-4.311074874863152e-8,
-0.05635477602481842,
0.027047473937273026,
-0.09159326553344727,
-0.042931169271469116,
0.024398958310484886,
-0.042624332010746,
-0.10520625114440918,
0.10134820640087128,
0.06271805614233017,
0.006243468727916479,
0.07994907349348068,
-0.029285935685038567,
0.00022373825777322054,
0.002745606005191803,
0.06267020851373672,
0.006899961270391941,
0.047372013330459595,
0.055456846952438354,
-0.01400834321975708,
-0.031416524201631546,
0.023821571841835976,
0.0052537331357598305,
0.014249581843614578,
-0.017334062606096268,
0.028715621680021286,
-0.09679566323757172,
0.005363428965210915,
0.08401314914226532,
0.014160209335386753,
-0.05646596476435661,
-0.01645948924124241,
0.08114165812730789,
-0.0016187572618946433,
-0.051253583282232285,
0.02347654663026333,
0.023604510352015495,
0.0502072349190712,
0.051624249666929245,
-0.009814702905714512,
-0.028508245944976807,
0.023913946002721786,
0.09176041185855865,
-0.0490369088947773,
0.005296089220792055,
0.04121912270784378,
0.029959924519062042,
-0.05187882110476494,
-0.12136635184288025,
-0.019586831331253052,
0.0014595974935218692,
0.009610033594071865,
0.05811597406864166,
0.0689338892698288,
0.08152022212743759,
0.050816603004932404,
-0.0393141508102417,
0.08034868538379669,
0.007325510494410992,
0.07257489860057831,
0.04279547184705734,
0.013931344263255596,
-0.002440200187265873,
-0.0581497848033905,
-0.07231564074754715
] |
PAL: Sample-Efficient Personalized Reward Modeling for Pluralistic Alignment | https://openreview.net/forum?id=1kFDrYCuSu | [
"Daiwei Chen",
"Yi Chen",
"Aniket Rege",
"Zhi Wang",
"Ramya Korlakai Vinayak"
] | Poster | Foundation models trained on internet-scale data benefit from extensive alignment to human preferences before deployment. However, existing methods typically assume a homogeneous preference shared by all individuals, overlooking the diversity inherent in human values. In this work, we propose a general reward modeling framework for pluralistic alignment (PAL), which incorporates diverse preferences from the ground up. PAL has a modular design that leverages commonalities across users while catering to individual personalization, enabling efficient few-shot localization of preferences for new users. Extensive empirical evaluation demonstrates that PAL matches or outperforms state-of-the-art methods on both text-to-text and text-to-image tasks: on Reddit TL;DR Summary, PAL is 1.7% more accurate for seen users and 36% more accurate for unseen users compared to the previous best method, with 100× less parameters. On Pick-a-Pic v2, PAL is 2.5% more accurate than the best method with 156× fewer learned parameters. Finally, we provide theoretical analysis for generalization of rewards learned via PAL framework showcasing the reduction in number of samples needed per user. | alignment, preference learning, foundation model, reward model, ideal point model, plurality | A novel alignment framework to learn from heterogeneous human preferences | 2,020 | null | [
-0.06496421992778778,
-0.06618060171604156,
-0.014993125572800636,
-0.019554492086172104,
0.05880814418196678,
0.050984594970941544,
0.006629629991948605,
0.029569212347269058,
0.08431480079889297,
0.042005911469459534,
0.013387585058808327,
-0.11224520951509476,
0.10117935389280319,
0.06643747538328171,
0.056575268507003784,
-0.01134194154292345,
0.08450373262166977,
-0.006354887969791889,
-0.06525421887636185,
-0.04801102355122566,
-0.10005857795476913,
0.006856995634734631,
0.08339160680770874,
0.005036030430346727,
-0.03678229823708534,
-0.018148060888051987,
0.0005842441460117698,
-0.013271467760205269,
-0.024062227457761765,
-0.04668561741709709,
-0.017015188932418823,
-0.022544171661138535,
0.0567411445081234,
0.01848960667848587,
-0.07061802595853806,
0.11202722787857056,
-0.10636080801486969,
-0.008192546665668488,
-0.023230794817209244,
-0.03273584693670273,
-0.06976199150085449,
-0.0435154065489769,
0.024511272087693214,
0.03470231592655182,
0.08437900245189667,
0.03190268948674202,
0.0010517628397792578,
-0.013484887778759003,
0.053309597074985504,
0.012337081134319305,
-0.028812943026423454,
0.005197656806558371,
0.002145550213754177,
0.03406628593802452,
-0.04954985901713371,
0.032320890575647354,
0.03659214824438095,
-0.023581210523843765,
0.03956156224012375,
-0.031426385045051575,
-0.03469070419669151,
-0.10983896255493164,
-0.024760406464338303,
0.052074186503887177,
0.031956903636455536,
-0.07747295498847961,
0.05701245367527008,
-0.022070908918976784,
-0.04616168141365051,
-0.03802066296339035,
-0.020223623141646385,
0.09073340892791748,
0.03509361296892166,
0.03554248809814453,
0.012165305204689503,
0.0538800023496151,
0.006454510614275932,
-0.066256083548069,
0.03846590965986252,
-0.026284432038664818,
-0.03315697982907295,
0.04452180862426758,
0.0771978348493576,
-0.011097792536020279,
0.08176974207162857,
-0.02746172621846199,
0.003579967189580202,
0.013089497573673725,
0.003893618006259203,
-0.012555887922644615,
-0.07545590400695801,
-0.011736349202692509,
0.077639140188694,
-0.03766639158129692,
-0.021569516509771347,
0.07154911756515503,
-0.03442112356424332,
-0.05011722445487976,
-0.09906762838363647,
0.10162854194641113,
0.02890118956565857,
0.09375344961881638,
-0.0042636822909116745,
-0.07815239578485489,
0.0527581162750721,
0.04816615581512451,
-0.02965138480067253,
0.010568473488092422,
0.08676477521657944,
-0.10869178175926208,
-0.0027691186405718327,
-0.009871731512248516,
-0.07413937151432037,
0.05235324427485466,
-0.008282601833343506,
-0.07465078681707382,
0.029176753014326096,
-0.038941822946071625,
0.00770168611779809,
0.007357001304626465,
0.03222968801856041,
-0.020272210240364075,
-0.021954230964183807,
-0.03650585561990738,
-0.033046141266822815,
0.016207659617066383,
-0.0020860829390585423,
4.0219300441094556e-33,
0.04489942267537117,
0.09059286117553711,
0.12123259156942368,
0.0619429387152195,
0.039138756692409515,
-0.015345035120844841,
-0.061404038220644,
0.04956585913896561,
-0.12644387781620026,
-0.058881476521492004,
-0.0034967639949172735,
0.03771628811955452,
-0.02995195798575878,
0.15969495475292206,
0.01527159009128809,
-0.038012225180864334,
-0.06745842844247818,
0.05382377281785011,
0.006001363042742014,
0.020388493314385414,
0.03399654105305672,
0.025088200345635414,
0.023151790723204613,
-0.012863636016845703,
-0.002889086026698351,
-0.025580234825611115,
0.06887607276439667,
-0.004533391445875168,
-0.05724462866783142,
-0.04310331121087074,
-0.0653219223022461,
-0.0180645938962698,
0.08881130814552307,
-0.0005925321020185947,
0.01944720558822155,
-0.04168561473488808,
0.04485410451889038,
-0.050077881664037704,
0.027239669114351273,
-0.010334311053156853,
-0.02199467457830906,
0.04786605387926102,
0.007197096943855286,
-0.022152388468384743,
-0.07984187453985214,
0.0022308635525405407,
0.025422660633921623,
0.01038964744657278,
-0.02947564423084259,
0.0007157582440413535,
0.03345850855112076,
0.014768507331609726,
-0.017447931692004204,
0.010225404985249043,
-0.06562768667936325,
0.006154363043606281,
-0.005147692747414112,
-0.022930702194571495,
0.02428709715604782,
-0.09236496686935425,
0.025371888652443886,
-0.008966062217950821,
0.011946450918912888,
0.03281385824084282,
0.010563591495156288,
0.04499702528119087,
0.05893184617161751,
-0.0225878544151783,
0.04105212539434433,
0.038436710834503174,
0.030359704047441483,
0.055925365537405014,
-0.04985712096095085,
-0.04717789962887764,
0.02226974256336689,
0.04256558418273926,
0.07210292667150497,
-0.09478948265314102,
0.047641534358263016,
-0.02052818052470684,
-0.014755542390048504,
-0.016305221244692802,
-0.008548167534172535,
-0.10830137878656387,
-0.03143111988902092,
-0.05224061384797096,
0.07542966306209564,
-0.15290755033493042,
-0.018026206642389297,
0.03391502797603607,
0.014609827660024166,
0.040938850492239,
-0.043976735323667526,
0.021552855148911476,
0.03347355127334595,
-3.474966645200091e-33,
-0.042822569608688354,
-0.018459511920809746,
0.03924408555030823,
0.015299370512366295,
0.007421426475048065,
0.018411222845315933,
0.06547945737838745,
0.025504454970359802,
-0.025829758495092392,
-0.05760906636714935,
-0.047490060329437256,
0.024344440549612045,
0.008865857496857643,
-0.07324668020009995,
-0.004891857970505953,
0.07246777415275574,
0.022998468950390816,
0.06124201416969299,
-0.036255985498428345,
0.003250607056543231,
0.05718430131673813,
0.07997187972068787,
-0.15738020837306976,
0.02348523586988449,
0.018325287848711014,
0.04757465422153473,
0.016360094770789146,
-0.02298213168978691,
0.008740930818021297,
-0.019643057137727737,
-0.06436027586460114,
-0.05948692932724953,
-0.1080377995967865,
-0.03081732802093029,
-0.021740809082984924,
0.06210336089134216,
-0.035849299281835556,
0.0008964089211076498,
-0.011683540418744087,
0.1697237193584442,
-0.012721942737698555,
-0.010166846215724945,
-0.10874781757593155,
-0.01651451177895069,
-0.05528848618268967,
-0.03693624585866928,
-0.11376993358135223,
-0.04721912369132042,
-0.00025625326088629663,
0.011532573029398918,
-0.054244514554739,
0.031177835538983345,
-0.11175605654716492,
-0.0019391450332477689,
-0.06086822599172592,
-0.08259078115224838,
-0.049148693680763245,
-0.022850174456834793,
0.0627603754401207,
-0.03990654647350311,
-0.07345035672187805,
0.024003632366657257,
-0.007701345719397068,
0.011892725713551044,
-0.021983513608574867,
-0.0451427586376667,
0.02618541568517685,
0.017039164900779724,
-0.09110192954540253,
0.01938614994287491,
0.05675879493355751,
-0.07121534645557404,
0.03453771397471428,
-0.03951901197433472,
0.016374919563531876,
-0.01799609325826168,
0.03454102948307991,
0.015034460462629795,
-0.0066634491086006165,
-0.06814810633659363,
0.06151750683784485,
0.037600647658109665,
0.039526332169771194,
-0.005015792790800333,
0.06849350780248642,
0.0720180869102478,
0.10281776636838913,
0.11726666986942291,
0.019198084250092506,
0.04065131023526192,
-0.019974488765001297,
0.044705212116241455,
0.03554067015647888,
0.05195498466491699,
-0.03928293660283089,
-4.859698421455505e-8,
-0.07390210777521133,
0.036528080701828,
-0.008559801615774632,
0.07855836302042007,
-0.09796413034200668,
0.003411715617403388,
-0.0126039432361722,
0.02054925076663494,
-0.016144361346960068,
-0.021981148049235344,
0.00932467170059681,
0.09206677228212357,
-0.02159452810883522,
-0.06160786747932434,
-0.038462646305561066,
0.06665235757827759,
-0.047114137560129166,
0.07671235501766205,
-0.0044495160691440105,
0.024052660912275314,
0.025005534291267395,
0.06613892316818237,
0.009118768386542797,
-0.057160742580890656,
-0.012800611555576324,
-0.03930627554655075,
-0.05300799384713173,
0.07548590004444122,
0.02025163732469082,
-0.02770283631980419,
0.01859753578901291,
-0.024765532463788986,
0.006742848549038172,
-0.011358743533492088,
0.08650209754705429,
0.02456025592982769,
-0.05361516401171684,
-0.018120011314749718,
-0.004424739629030228,
-0.0030563068576157093,
0.041403189301490784,
0.031132487580180168,
0.002494564512744546,
0.02092728018760681,
0.025872373953461647,
0.06033756211400032,
-0.053094442933797836,
-0.027736680582165718,
0.003561608958989382,
-0.031961530447006226,
-0.003923974931240082,
-0.06814956665039062,
-0.03381984308362007,
0.05893958732485771,
0.07673590630292892,
-0.04092804715037346,
0.042753420770168304,
-0.006796468049287796,
0.11722789704799652,
0.05956171080470085,
0.05790625885128975,
-0.0026210627984255552,
-0.09923043847084045,
0.0035262247547507286
] |
Dynamic Diffusion Transformer | https://openreview.net/forum?id=taHwqSrbrb | [
"Wangbo Zhao",
"Yizeng Han",
"Jiasheng Tang",
"Kai Wang",
"Yibing Song",
"Gao Huang",
"Fan Wang",
"Yang You"
] | Poster | Diffusion Transformer (DiT), an emerging diffusion model for image generation,
has demonstrated superior performance but suffers from substantial computational
costs. Our investigations reveal that these costs stem from the static inference
paradigm, which inevitably introduces redundant computation in certain diffusion
timesteps and spatial regions. To address this inefficiency, we propose Dynamic
Diffusion Transformer (DyDiT), an architecture that dynamically adjusts its compu-
tation along both timestep and spatial dimensions during generation. Specifically,
we introduce a Timestep-wise Dynamic Width (TDW) approach that adapts model
width conditioned on the generation timesteps. In addition, we design a Spatial-
wise Dynamic Token (SDT) strategy to avoid redundant computation at unnecessary
spatial locations. Extensive experiments on various datasets and different-sized
models verify the superiority of DyDiT. Notably, with <3% additional fine-tuning it-
erations, our method reduces the FLOPs of DiT-XL by 51%, accelerates generation
by 1.73×, and achieves a competitive FID score of 2.07 on ImageNet. | Diffusion Transformer, Dynamic Neural Network, Efficiency | null | 2,019 | 2410.03456 | [
-0.009049749933183193,
-0.006013385020196438,
0.054638586938381195,
0.04020846262574196,
0.02008717507123947,
-0.11082296073436737,
-0.09009270370006561,
0.007024161983281374,
0.07001087069511414,
-0.03267650306224823,
-0.029548441991209984,
-0.027666399255394936,
-0.008057908155024052,
0.05447503551840782,
-0.000922260107472539,
0.006091486196964979,
-0.04892519488930702,
0.06330425292253494,
-0.0912676528096199,
-0.08752205222845078,
-0.02573464997112751,
-0.052905939519405365,
0.02403736487030983,
-0.0545613057911396,
0.04674043133854866,
0.03328101336956024,
0.09142932295799255,
-0.11268770694732666,
0.05201480910181999,
-0.015421216376125813,
0.06036631762981415,
-0.029255524277687073,
-0.06524535268545151,
-0.008120765909552574,
-0.03199191018939018,
0.011080116033554077,
0.01684112288057804,
0.011770177632570267,
0.08992456644773483,
-0.030123382806777954,
0.02285294607281685,
-0.006073520053178072,
0.06959152221679688,
0.02379237487912178,
0.006068405695259571,
0.02297452837228775,
0.027076635509729385,
0.018233872950077057,
-0.007054396439343691,
-0.014811807312071323,
0.03196882829070091,
0.011416468769311905,
-0.045235708355903625,
0.06619492173194885,
-0.05318448320031166,
0.06211395934224129,
0.09337428957223892,
0.08566325902938843,
0.018228331580758095,
-0.040887270122766495,
-0.0836896002292633,
-0.04006589204072952,
0.0715898647904396,
0.003793425625190139,
-0.07181379199028015,
-0.010323493741452694,
0.03146912902593613,
0.02972477301955223,
-0.00789661519229412,
-0.05210144445300102,
-0.05182213336229324,
0.05580070987343788,
-0.06340339034795761,
0.0242479145526886,
-0.018180089071393013,
0.03019712306559086,
0.1066039577126503,
-0.0274141076952219,
0.025387689471244812,
-0.024455389007925987,
0.10721280425786972,
0.0015624177176505327,
0.058785248547792435,
-0.0449393056333065,
0.04389456659555435,
-0.01936975121498108,
-0.039563097059726715,
0.05033067241311073,
0.004695989657193422,
-0.01052221842110157,
-0.023025140166282654,
-0.001897546462714672,
-0.06765268743038177,
-0.01292700320482254,
0.004024798050522804,
-0.02364019863307476,
-0.03619720786809921,
-0.03175678476691246,
-0.035653986036777496,
0.06354157626628876,
-0.039461709558963776,
-0.018498290330171585,
0.00019480939954519272,
0.02218935452401638,
0.05138716474175453,
-0.09532984346151352,
0.05831427499651909,
0.05478160083293915,
0.035226430743932724,
-0.032209258526563644,
0.0984419733285904,
0.0650988519191742,
0.04248810559511185,
0.044122252613306046,
0.03278101980686188,
-0.05593545734882355,
-0.06365582346916199,
-0.005859836470335722,
0.03346161171793938,
-0.015563691966235638,
0.049525126814842224,
-0.013857703655958176,
-0.10545872151851654,
0.054563969373703,
-0.030132614076137543,
-0.07431608438491821,
0.006981099024415016,
1.474382340289078e-33,
-0.11061903834342957,
-0.006034835707396269,
0.03395936265587807,
0.06451498717069626,
0.027031324803829193,
0.0124738160520792,
0.047950368374586105,
-0.03761003166437149,
-0.04240133985877037,
-0.07012265175580978,
-0.042459845542907715,
-0.06442192941904068,
-0.07181567698717117,
0.12701331079006195,
0.09159421175718307,
-0.038639772683382034,
-0.007317240349948406,
0.06986021250486374,
0.021346956491470337,
-0.011337928473949432,
0.052238862961530685,
-0.011143331415951252,
-0.028905030339956284,
-0.023439759388566017,
-0.02336849458515644,
-0.01623700000345707,
0.033645033836364746,
-0.05343170464038849,
0.05021876469254494,
-0.010089889168739319,
-0.16857187449932098,
-0.04774266481399536,
-0.029469534754753113,
0.012174539268016815,
0.08337315917015076,
-0.012869938276708126,
-0.062226977199316025,
-0.042318496853113174,
-0.012984111905097961,
0.003869408741593361,
-0.017761005088686943,
0.050699979066848755,
-0.016466764733195305,
-0.0020088497549295425,
-0.09781681001186371,
0.02329101413488388,
0.010241713374853134,
0.011686138808727264,
-0.05055800452828407,
0.02369154989719391,
0.07774319499731064,
0.0034932997077703476,
-0.025023970752954483,
-0.03786758333444595,
0.06679996848106384,
-0.05761980637907982,
0.0507630854845047,
-0.09069499373435974,
0.0817561149597168,
0.10829035937786102,
0.04925236478447914,
0.05235728248953819,
-0.06718306243419647,
0.04633404687047005,
0.01753869093954563,
-0.07169204950332642,
0.052899278700351715,
0.024819865822792053,
0.007097249384969473,
0.06033029407262802,
0.037313707172870636,
0.015232997946441174,
0.02629452385008335,
-0.10882207006216049,
0.03416472673416138,
-0.039015043526887894,
0.11656296998262405,
0.042949989438056946,
-0.07198274880647659,
-0.00429847277700901,
-0.05848458781838417,
0.005097469314932823,
0.01620699279010296,
-0.053522344678640366,
0.07070768624544144,
0.0018001977587118745,
-0.03139984607696533,
-0.04754137620329857,
-0.09571666270494461,
-0.09878762811422348,
0.025008050724864006,
-0.024811113253235817,
0.02664973959326744,
0.026287395507097244,
-0.03822968527674675,
-1.5698144818186008e-33,
-0.023223478347063065,
0.08299213647842407,
0.04422830790281296,
0.06085651367902756,
-0.07258837670087814,
-0.039595913141965866,
-0.02039000391960144,
0.000021564301277976483,
0.09603232890367508,
-0.0971691906452179,
-0.08685274422168732,
-0.018895020708441734,
-0.054869189858436584,
-0.013341608457267284,
0.027377106249332428,
-0.024104606360197067,
0.02186054177582264,
-0.022760536521673203,
-0.07441843301057816,
0.0020815988536924124,
0.025122113525867462,
0.07049120217561722,
-0.030254244804382324,
-0.01085254829376936,
-0.08144594728946686,
0.06628942489624023,
-0.06666475534439087,
0.0543975755572319,
-0.011627578176558018,
-0.018586985766887665,
-0.06371577084064484,
0.014806592836976051,
0.06124395504593849,
-0.048131201416254044,
-0.012651119381189346,
0.0357709676027298,
0.0209600068628788,
-0.0015389621257781982,
0.026756688952445984,
0.07219932973384857,
-0.03347862511873245,
-0.07282284647226334,
-0.04319765791296959,
0.08985424786806107,
-0.06063484400510788,
0.03350896015763283,
-0.009403284639120102,
-0.0030021413695067167,
0.02738332748413086,
0.04397027567028999,
0.028095997869968414,
0.07268894463777542,
-0.11103662103414536,
0.054098185151815414,
0.026604415848851204,
-0.03557397425174713,
-0.0989171639084816,
0.018441196531057358,
0.05823950469493866,
-0.017875798046588898,
-0.0028634604532271624,
-0.08744161576032639,
0.06844792515039444,
-0.06864351779222488,
-0.058626845479011536,
-0.010041485540568829,
-0.055328600108623505,
-0.039746228605508804,
-0.017326805740594864,
0.16972394287586212,
-0.02236437425017357,
-0.006548813544213772,
-0.03497811779379845,
-0.02704467624425888,
-0.034654777497053146,
-0.02686617523431778,
0.08003827929496765,
-0.05538538098335266,
0.05471431836485863,
0.01741984486579895,
-0.08734369277954102,
0.03487124666571617,
-0.016611387953162193,
-0.05455269664525986,
-0.022903868928551674,
-0.004388913512229919,
0.04071991518139839,
-0.010890995152294636,
0.019355861470103264,
0.10178060084581375,
-0.017412075772881508,
0.029533062130212784,
0.07964630424976349,
0.052807435393333435,
-0.026998370885849,
-5.019272464323876e-8,
-0.014605528675019741,
0.036441635340452194,
0.021559441462159157,
0.01395319402217865,
-0.025157321244478226,
-0.061170872300863266,
0.1341605931520462,
0.08097698539495468,
0.007948582991957664,
-0.033638399094343185,
0.06489282846450806,
-0.08612864464521408,
-0.030639799311757088,
-0.04050685092806816,
-0.01503416895866394,
0.051843058317899704,
0.012213269248604774,
-0.030996736139059067,
-0.02142917923629284,
-0.0017265500500798225,
-0.11204925179481506,
0.053453732281923294,
0.00808727741241455,
-0.022299708798527718,
0.016021618619561195,
-0.03364564850926399,
-0.032717298716306686,
0.009505223482847214,
0.1166895180940628,
-0.018475251272320747,
0.024202845990657806,
0.027077946811914444,
-0.021797029301524162,
0.019441859796643257,
-0.0197611041367054,
0.012409161776304245,
0.0030630279798060656,
0.02046339586377144,
-0.013448168523609638,
0.001644573756493628,
-0.0251210555434227,
0.060005903244018555,
0.016050206497311592,
0.050590235739946365,
-0.010544110089540482,
-0.003371571656316519,
0.07979505509138107,
0.02541539818048477,
-0.03341231122612953,
0.027131082490086555,
0.03287119045853615,
-0.027185702696442604,
-0.05890413373708725,
0.004451484885066748,
0.06056502088904381,
0.012743543833494186,
0.02109747938811779,
0.012732974253594875,
0.05610015243291855,
0.08972994238138199,
-0.039488960057497025,
0.03442863002419472,
-0.04551710560917854,
-0.005931224208325148
] |
GS-LiDAR: Generating Realistic LiDAR Point Clouds with Panoramic Gaussian Splatting | https://openreview.net/forum?id=RMaRBE9s2H | [
"Junzhe Jiang",
"Chun Gu",
"Yurui Chen",
"Li Zhang"
] | Poster | LiDAR novel view synthesis (NVS) has emerged as a novel task within LiDAR simulation, offering valuable simulated point cloud data from novel viewpoints to aid in autonomous driving systems. However, existing LiDAR NVS methods typically rely on neural radiance fields (NeRF) as their 3D representation, which incurs significant computational costs in both training and rendering. Moreover, NeRF and its variants are designed for symmetrical scenes, making them ill-suited for driving scenarios. To address these challenges, we propose GS-LiDAR, a novel framework for generating realistic LiDAR point clouds with panoramic Gaussian splatting. Our approach employs 2D Gaussian primitives with periodic vibration properties, allowing for precise geometric reconstruction of both static and dynamic elements in driving scenarios. We further introduce a novel panoramic rendering technique with explicit ray-splat intersection, guided by panoramic LiDAR supervision. By incorporating intensity and ray-drop spherical harmonic (SH) coefficients into the Gaussian primitives, we enhance the realism of the rendered point clouds. Extensive experiments on KITTI-360 and nuScenes demonstrate the superiority of our method in terms of quantitative metrics, visual quality, as well as training and rendering efficiency. | Gaussian Splatting, LiDAR Simulation | null | 2,018 | null | [
0.0065909442491829395,
-0.05461083725094795,
0.03425604850053787,
0.009115699678659439,
0.08825244754552841,
-0.012553774751722813,
0.03857515752315521,
-0.10287371277809143,
0.01966901868581772,
-0.04389527812600136,
-0.07105771452188492,
-0.027611516416072845,
-0.015281408093869686,
0.008893352001905441,
-0.010866457596421242,
-0.0014391826698556542,
0.03431734815239906,
0.04446421191096306,
-0.013988468796014786,
0.05803494155406952,
0.015064134262502193,
0.01704981178045273,
-0.06211638078093529,
0.024486549198627472,
0.04589509591460228,
0.024898285046219826,
0.0674537643790245,
-0.07453104108572006,
-0.00936858355998993,
0.04076823219656944,
0.08667145669460297,
-0.015662016347050667,
-0.10797690600156784,
-0.0036534564569592476,
-0.027883782982826233,
-0.06339807063341141,
0.014394821599125862,
0.01735946163535118,
-0.017644833773374557,
-0.0687607154250145,
0.022260401397943497,
-0.025479523465037346,
-0.021898968145251274,
0.07840485870838165,
0.07955892384052277,
0.0790572389960289,
0.006167038809508085,
-0.11355213820934296,
0.06827997416257858,
0.0017660522134974599,
-0.01014532521367073,
-0.025986909866333008,
0.012046406045556068,
-0.03291131928563118,
0.026650862768292427,
0.08924052864313126,
-0.08657173812389374,
-0.10813038796186447,
0.08170434087514877,
-0.023564830422401428,
0.024472396820783615,
-0.015211883001029491,
-0.031708087772130966,
-0.0033402624540030956,
0.029688145965337753,
-0.07848242670297623,
0.05185761675238609,
-0.062053363770246506,
0.010621966794133186,
0.036087535321712494,
-0.05680133402347565,
0.07426615804433823,
0.0179241094738245,
-0.0883384421467781,
-0.041534673422575,
0.07448509335517883,
-0.009629200212657452,
-0.014930461533367634,
-0.06318043917417526,
-0.01266286801546812,
0.07033972442150116,
0.08287100493907928,
0.026859043166041374,
0.07767030596733093,
-0.059481386095285416,
-0.006979994475841522,
-0.01054640393704176,
0.19203056395053864,
0.034512463957071304,
0.04455140605568886,
-0.055754173547029495,
0.04841270297765732,
-0.040631797164678574,
-0.09822095930576324,
-0.06003660336136818,
-0.03667023405432701,
0.00938212126493454,
-0.08763176202774048,
0.045600757002830505,
0.04495083913207054,
0.03554218262434006,
-0.04799278825521469,
0.02655796892940998,
0.0008466750150546432,
-0.005633071530610323,
0.03640231117606163,
0.028491489589214325,
0.0013274982338771224,
-0.03333405405282974,
0.04157724976539612,
0.04077466204762459,
0.012081718072295189,
-0.01777925342321396,
-0.022990144789218903,
0.006174714304506779,
0.02014417201280594,
-0.001457281643524766,
0.009742794558405876,
0.007880819030106068,
-0.020022323355078697,
-0.0035819564945995808,
-0.03386528417468071,
0.013200034387409687,
0.09045477956533432,
0.005247172433882952,
-0.0329919159412384,
-0.03653567284345627,
1.7223982403301416e-33,
0.027242518961429596,
0.10012783110141754,
0.026607124134898186,
0.029794640839099884,
0.08598750829696655,
-0.009644672274589539,
-0.010857424698770046,
0.051029615104198456,
-0.06207140162587166,
0.05362200736999512,
-0.08926931768655777,
-0.04106388986110687,
-0.008916960097849369,
-0.006665649823844433,
0.04030189663171768,
0.059233952313661575,
0.017684824764728546,
-0.031706396490335464,
-0.09987804293632507,
0.03063010983169079,
-0.09466022998094559,
0.023821718990802765,
0.03945451229810715,
0.008532039821147919,
-0.08089862018823624,
0.09217794239521027,
0.026788482442498207,
-0.001340105663985014,
-0.008516660891473293,
-0.023918572813272476,
0.0197654590010643,
0.051869191229343414,
-0.014542382210493088,
0.020493067800998688,
-0.0023707780055701733,
-0.01954825222492218,
-0.0028606050182133913,
-0.023839257657527924,
-0.021360043436288834,
-0.010009228251874447,
-0.01891944371163845,
0.07255373150110245,
0.00985556561499834,
0.06299491226673126,
-0.04358668997883797,
0.02189946547150612,
0.061934035271406174,
0.09349599480628967,
-0.02095036953687668,
0.030392026528716087,
0.00768349040299654,
-0.0021328015718609095,
-0.06943512707948685,
-0.047895513474941254,
-0.026556797325611115,
-0.009025027975440025,
0.01980668120086193,
-0.008211715146899223,
0.07321380823850632,
-0.033350907266139984,
-0.018504571169614792,
-0.09581255167722702,
-0.00707766180858016,
-0.011155596002936363,
-0.07872888445854187,
0.0021740945521742105,
-0.006700186990201473,
0.054048508405685425,
0.050838254392147064,
0.03385555371642113,
-0.025595715269446373,
0.010319186374545097,
0.01940838433802128,
0.022424785420298576,
0.06341813504695892,
-0.03751901909708977,
0.01728787273168564,
-0.006363822612911463,
0.00857737846672535,
0.11971940100193024,
-0.046616557985544205,
0.04836595430970192,
-0.009745919145643711,
-0.07287617772817612,
-0.057533200830221176,
-0.03492815047502518,
-0.05746090039610863,
0.02907632477581501,
-0.05667603388428688,
-0.07985007017850876,
-0.05688630789518356,
-0.08332592248916626,
-0.021964896470308304,
-0.05176100507378578,
-0.06864342093467712,
-2.185279814022134e-33,
0.04929615184664726,
-0.03901122510433197,
-0.13962361216545105,
0.05905331298708916,
0.018590059131383896,
0.02484179101884365,
-0.018002351745963097,
-0.030413661152124405,
0.02969241701066494,
-0.07816700637340546,
-0.009922136552631855,
0.01686176285147667,
0.047123707830905914,
0.035503488034009933,
-0.004470230080187321,
-0.027723165228962898,
-0.011160026304423809,
-0.05868059769272804,
-0.10207269340753555,
-0.03620085120201111,
0.027132563292980194,
0.061589308083057404,
-0.02399798482656479,
-0.024364374577999115,
-0.06333579123020172,
0.062471769750118256,
0.013074573129415512,
0.09849168360233307,
-0.0065132807940244675,
0.06062815338373184,
-0.00014469468442257494,
0.038661327213048935,
0.03820136934518814,
-0.06491442769765854,
-0.006804316770285368,
0.04789792373776436,
0.11131471395492554,
0.026784295216202736,
-0.10000794380903244,
0.00014203164028003812,
0.0045727090910077095,
-0.032511480152606964,
0.030874477699398994,
-0.005911465268582106,
0.011465530842542648,
-0.011502907611429691,
0.06591116636991501,
-0.018928619101643562,
-0.04346328601241112,
0.06496063619852066,
-0.039606474339962006,
-0.02000425010919571,
-0.12606379389762878,
0.06127169728279114,
0.020961938425898552,
-0.08223135024309158,
-0.11748813092708588,
-0.005679115653038025,
0.059556279331445694,
-0.016176892444491386,
-0.016121508553624153,
-0.01214294321835041,
-0.02633688785135746,
0.0160180926322937,
-0.011240148916840553,
-0.008788247592747211,
-0.06556334346532822,
0.016516435891389847,
-0.017184220254421234,
0.08952777087688446,
-0.08569730073213577,
-0.05223527550697327,
0.019068323075771332,
0.02138192020356655,
-0.04420314356684685,
-0.022378453984856606,
0.07270123809576035,
-0.03180946037173271,
0.011463436298072338,
-0.004512494429945946,
0.04041144996881485,
-0.016000567004084587,
0.04302592575550079,
0.1100834310054779,
0.059321705251932144,
0.05472154915332794,
0.02206503599882126,
-0.034534141421318054,
-0.028383085504174232,
0.03244941681623459,
0.0077827745117247105,
0.06685040891170502,
-0.038181666284799576,
0.08593884855508804,
-0.03151949122548103,
-3.884904131723488e-8,
-0.05016167461872101,
0.0777774229645729,
0.006387124769389629,
0.0026602493599057198,
-0.034368228167295456,
-0.02281658537685871,
0.05363629758358002,
0.00842907465994358,
-0.053175415843725204,
-0.1199849396944046,
-0.004024899564683437,
-0.14174462854862213,
0.030428897589445114,
-0.024688176810741425,
0.02089155837893486,
0.11679010093212128,
0.007871163077652454,
0.02933073416352272,
-0.009131176397204399,
0.06691090017557144,
-0.061712391674518585,
0.07320566475391388,
0.03546001762151718,
-0.01296598557382822,
-0.048332273960113525,
0.08434019237756729,
-0.0315975546836853,
0.046448905020952225,
0.1259423792362213,
-0.03216620162129402,
-0.026491964235901833,
-0.012629156932234764,
0.07396291196346283,
0.051715172827243805,
-0.03479352220892906,
0.0044678994454443455,
-0.05951697379350662,
0.008141196332871914,
0.05505705997347832,
-0.03015868179500103,
-0.03724285960197449,
0.07442501932382584,
-0.06002693623304367,
-0.014434666372835636,
-0.021545974537730217,
0.10280437022447586,
0.016312554478645325,
-0.07842832803726196,
-0.043083008378744125,
0.047794606536626816,
-0.006997235119342804,
0.06202879920601845,
-0.051273245364427567,
0.06672816723585129,
0.06944189220666885,
-0.002184596611186862,
0.05000326409935951,
-0.071713887155056,
0.023591553792357445,
-0.02475166693329811,
-0.027113258838653564,
-0.012895156629383564,
-0.10709349811077118,
0.009695027954876423
] |
LongVILA: Scaling Long-Context Visual Language Models for Long Videos | https://openreview.net/forum?id=wCXAlfvCy6 | [
"Yukang Chen",
"Fuzhao Xue",
"Dacheng Li",
"Qinghao Hu",
"Ligeng Zhu",
"Xiuyu Li",
"Yunhao Fang",
"Haotian Tang",
"Shang Yang",
"Zhijian Liu",
"Yihui He",
"Hongxu Yin",
"Pavlo Molchanov",
"Jan Kautz",
"Linxi Fan",
"Yuke Zhu",
"Yao Lu",
"Song Han"
] | Poster | Long-context capability is critical for multi-modal foundation models, especially for long video understanding. We introduce LongVILA, a full-stack solution for long-context visual-language models by co-designing the algorithm and system. For model training, we upgrade existing VLMs to support long video understanding by incorporating two additional stages, i.e., long context extension and long video supervised fine-tuning. However, training on long video is computationally and memory intensive. We introduce the long-context Multi-Modal Sequence Parallelism (MM-SP) system that efficiently parallelizes long video training and inference, enabling 2M context length training on 256 GPUs without any gradient checkpointing. LongVILA efficiently extends the number of video frames of VILA from 8 to 2048, achieving 99.8% accuracy in 6,000-frame (more than 1 million tokens) video needle-in-a-haystack. LongVILA-7B demonstrates strong accuracy on 9 popular video benchmarks, e.g., 65.1% VideoMME with subtitle. Besides, MM-SP is 2.1x - 5.7x faster than ring style sequence parallelism and 1.1x - 1.4x faster than Megatron with a hybrid context and tensor parallelism. Moreover, it seamlessly integrates with Hugging Face Transformers. | Large language models, Long context, Multi-modality, Video understanding | null | 2,006 | 2408.10188 | [
0.025725871324539185,
-0.12893402576446533,
-0.013875309377908707,
-0.045369744300842285,
0.12973684072494507,
-0.0013288318878039718,
-0.05086629092693329,
-0.0163233894854784,
0.008108961395919323,
-0.10353641211986542,
-0.004016043618321419,
-0.06539333611726761,
-0.0392439067363739,
0.046277858316898346,
-0.0058982460759580135,
-0.011841293424367905,
0.02942637912929058,
0.10435893386602402,
-0.1160367801785469,
-0.05268516018986702,
0.05105559527873993,
-0.04019275680184364,
0.028333915397524834,
0.024121448397636414,
0.023439746350049973,
0.06237826496362686,
-0.019576532766222954,
-0.0447060726583004,
0.08826260268688202,
-0.029462002217769623,
0.07473509013652802,
0.0446995384991169,
0.005794811062514782,
0.008768725208938122,
-0.05907421186566353,
0.033443041145801544,
-0.04838642105460167,
-0.03499571606516838,
-0.03800816461443901,
-0.06079608574509621,
-0.0017629382200539112,
0.050940729677677155,
0.06312650442123413,
0.013588991947472095,
0.01687648333609104,
-0.013281011953949928,
0.06639279425144196,
-0.04452735558152199,
0.08856440335512161,
0.02322821132838726,
-0.09243470430374146,
-0.026992298662662506,
-0.013162747956812382,
0.03716402128338814,
-0.07214726507663727,
0.0016195665812119842,
0.026256388053297997,
0.0013005784712731838,
0.03011152148246765,
0.0435514934360981,
-0.024160459637641907,
-0.08862147480249405,
-0.07190536707639694,
0.0332779660820961,
-0.011927179992198944,
-0.0386284776031971,
0.0771932303905487,
-0.030550038442015648,
-0.007899750955402851,
-0.014742450788617134,
-0.03292516618967056,
0.0287909097969532,
-0.026756301522254944,
0.0867794007062912,
-0.004740728996694088,
0.03188842535018921,
0.060643214732408524,
-0.025956355035305023,
0.05247153714299202,
-0.12877966463565826,
0.08023146539926529,
-0.0022754294332116842,
0.1247585192322731,
-0.018573107197880745,
0.10959142446517944,
-0.01877891831099987,
0.00008148673077812418,
0.034521277993917465,
-0.01371717918664217,
0.009413189254701138,
-0.005025835707783699,
-0.04770878329873085,
0.07252936065196991,
-0.008217636495828629,
0.028106119483709335,
0.003347212914377451,
0.013344821520149708,
-0.05289508402347565,
-0.0675826370716095,
0.09705229103565216,
-0.026844745501875877,
-0.005719154141843319,
-0.041419897228479385,
0.03253417834639549,
0.039019230753183365,
0.014663120731711388,
0.04409098997712135,
0.11368915438652039,
-0.008542037568986416,
-0.03307364135980606,
0.06215081736445427,
0.10777165740728378,
0.027038145810365677,
0.03037276864051819,
0.06485135853290558,
-0.0021090256050229073,
-0.03006383590400219,
-0.060368455946445465,
-0.01623125746846199,
0.049038197845220566,
-0.037207815796136856,
-0.025331543758511543,
-0.024950487539172173,
-0.029327645897865295,
-0.008563986979424953,
-0.044321343302726746,
-0.010669665411114693,
4.766319647784843e-34,
0.03679654747247696,
0.039283715188503265,
0.08769439905881882,
0.03320219740271568,
0.04508461430668831,
0.0434904471039772,
0.004012719262391329,
0.035087428987026215,
-0.07259875535964966,
0.030553722754120827,
0.017317168414592743,
-0.015835145488381386,
-0.056612443178892136,
0.12196429818868637,
0.0061713471077382565,
-0.03976978361606598,
-0.004042213782668114,
0.0016201970865949988,
-0.01843058317899704,
0.07156077027320862,
0.05968199297785759,
-0.03279442340135574,
0.01277459878474474,
0.01605214551091194,
0.026367299258708954,
0.006004459224641323,
0.08148246258497238,
-0.04729006066918373,
0.016891933977603912,
0.024677110835909843,
-0.12891453504562378,
-0.015495305880904198,
-0.022582799196243286,
0.030659204348921776,
0.06548787653446198,
-0.06902160495519638,
0.004704657010734081,
-0.06471936404705048,
-0.030710699036717415,
-0.030338246375322342,
-0.06087658926844597,
0.011913741938769817,
0.00882588978856802,
-0.06308536976575851,
-0.1459149569272995,
-0.060782577842473984,
0.013490580953657627,
0.07805513590574265,
-0.07124477624893188,
0.018607247620821,
0.0742531269788742,
0.037205327302217484,
-0.054310403764247894,
-0.04099214822053909,
-0.026228247210383415,
0.006870339624583721,
0.05878434330224991,
0.05835595726966858,
0.06507661193609238,
0.0359683483839035,
-0.003953170031309128,
-0.002494948450475931,
-0.018703684210777283,
0.16400684416294098,
0.0012185215018689632,
-0.022551145404577255,
-0.06822309643030167,
0.0422825962305069,
0.043085452169179916,
-0.011891089379787445,
-0.06030312553048134,
0.021122783422470093,
0.07038614898920059,
0.02050040476024151,
-0.004010090138763189,
-0.02473740093410015,
0.021032599732279778,
-0.04662267118692398,
-0.011963305063545704,
0.10994867235422134,
0.0010231304913759232,
-0.006548874545842409,
0.04558315873146057,
-0.06673643738031387,
0.0006608202238567173,
-0.07247773557901382,
-0.014667542651295662,
-0.09134861081838608,
-0.02258889190852642,
-0.10765065997838974,
-0.06779934465885162,
-0.06863187998533249,
0.0436411090195179,
-0.023477395996451378,
-0.029454533010721207,
7.83341629370799e-34,
0.06923572719097137,
0.0060849799774587154,
-0.024744762107729912,
0.0485244058072567,
0.02828029729425907,
-0.09238073974847794,
0.03001418709754944,
0.06677474081516266,
-0.06417886167764664,
-0.07599639147520065,
0.0368107371032238,
-0.018713481724262238,
-0.000822089146822691,
0.019829602912068367,
0.050164129585027695,
-0.02672935463488102,
0.04508981481194496,
-0.0022075576707720757,
-0.02649628557264805,
-0.0384177565574646,
0.03353823721408844,
0.026723874732851982,
-0.07567797601222992,
-0.029829518869519234,
-0.045647092163562775,
0.004308633506298065,
-0.07476849853992462,
0.034762829542160034,
-0.017677972093224525,
0.009439554065465927,
0.0026722464244812727,
-0.007375050336122513,
0.011270810849964619,
-0.04011745750904083,
-0.03224848955869675,
0.060158319771289825,
0.058257292956113815,
-0.05486711114645004,
-0.017459925264120102,
0.0959252268075943,
0.09398627281188965,
-0.022457364946603775,
-0.012978076003491879,
-0.018681133165955544,
0.029283400624990463,
0.02358420379459858,
-0.07768481969833374,
-0.027722297236323357,
-0.00829879567027092,
-0.01420119684189558,
-0.06662821024656296,
0.06238673999905586,
-0.03789987787604332,
0.012974345125257969,
0.052172113209962845,
-0.15984995663166046,
-0.07630516588687897,
0.029681513085961342,
0.0643354058265686,
-0.0715106874704361,
-0.0015225320821627975,
-0.008613076992332935,
-0.023198258131742477,
-0.03723534196615219,
0.09051625430583954,
0.001515039592050016,
0.0007150759920477867,
-0.044293638318777084,
0.016341060400009155,
0.002851495984941721,
0.008740803226828575,
-0.042043644934892654,
-0.07505394518375397,
0.06804616749286652,
-0.024196602404117584,
0.0018157243030145764,
0.017579609528183937,
-0.06629770249128342,
0.049855589866638184,
-0.04997880384325981,
-0.004280315712094307,
-0.11219324916601181,
0.025746216997504234,
0.023357193917036057,
0.08353278785943985,
0.09736956655979156,
0.07042394578456879,
0.0692022293806076,
0.06867029517889023,
0.04786108806729317,
-0.016754817217588425,
0.04763834550976753,
0.0508812814950943,
0.010149638168513775,
0.00015796131629031152,
-3.3874535176892095e-8,
-0.0491100512444973,
0.06701839715242386,
-0.04087555781006813,
-0.007598812226206064,
0.02684529311954975,
0.01202524732798338,
-0.00623337272554636,
0.063173808157444,
0.026739045977592468,
-0.020812606438994408,
0.03255793824791908,
-0.019780844449996948,
-0.025176797062158585,
-0.01514360960572958,
-0.0398559495806694,
0.046849120408296585,
0.014197849668562412,
-0.02803715504705906,
-0.0018845145823433995,
0.075896255671978,
0.03356942906975746,
0.03413967788219452,
0.043731268495321274,
0.0052886344492435455,
-0.06780747324228287,
-0.07002566009759903,
0.005841243080794811,
0.12482801824808121,
0.07737977057695389,
-0.08393974602222443,
-0.03766121715307236,
0.04634660482406616,
-0.042321112006902695,
-0.1061134859919548,
-0.04566319286823273,
0.03243500739336014,
0.04615245386958122,
0.061634644865989685,
0.04033632203936577,
0.0035390867851674557,
0.022073136642575264,
-0.005402969662100077,
0.009532155469059944,
0.006256347056478262,
0.020274311304092407,
0.09784303605556488,
-0.024385662749409676,
-0.08958624303340912,
-0.013038912788033485,
-0.030952665954828262,
-0.011730101890861988,
0.009713990613818169,
0.00045429368037730455,
0.09716001152992249,
0.0226439256221056,
0.01477998960763216,
0.00837295688688755,
0.0055084917694330215,
0.07598818838596344,
0.010750461369752884,
-0.04992331564426422,
-0.03536835312843323,
-0.04214107617735863,
-0.015181795693933964
] |
McEval: Massively Multilingual Code Evaluation | https://openreview.net/forum?id=UunCPtPOlZ | [
"Linzheng Chai",
"Shukai Liu",
"Jian Yang",
"Yuwei Yin",
"JinKe",
"Jiaheng Liu",
"Tao Sun",
"Ge Zhang",
"Changyu Ren",
"Hongcheng Guo",
"Noah Wang",
"Boyang Wang",
"Xianjie Wu",
"Bing Wang",
"Tongliang Li",
"Liqun Yang",
"Sufeng Duan",
"Zhaoxiang Zhang",
"Zhoujun Li"
] | Poster | Code large language models (LLMs) have shown remarkable advances in code understanding, completion, and generation tasks. Programming benchmarks, comprised of a selection of code challenges and corresponding test cases, serve as a standard to evaluate the capability of different LLMs in such tasks. However, most existing benchmarks primarily focus on Python and are still restricted to a limited number of languages, where other languages are translated from the Python samples degrading the data diversity. To further facilitate the research of code LLMs, we propose a massively multilingual code benchmark covering 40 programming languages (McEval) with 16K test samples, which substantially pushes the limits of code LLMs in multilingual scenarios. The benchmark contains challenging code completion, understanding, and generation evaluation tasks with finely curated massively multilingual instruction corpora McEval-Instruct. In addition, we introduce an effective multilingual coder mCoder trained on McEval-Instruct to support multilingual programming language generation. Extensive experimental results on McEval show that there is still a difficult journey between open-source models and closed-source LLMs in numerous languages. The instruction corpora and evaluation benchmark are available at https://github.com/MCEVAL/McEval. | Benchmark, Code Intelligence, Multilingual, Large Language Model, Multilingual Multitask Code Evaluation | To facilitate the development of code LLMs, we introduce a complete framework that includes the multilingual code instruction corpora, multilingual coder (mCoder), and multilingual code evaluation benchmark. | 2,004 | 2406.07436 | [
-0.0859983041882515,
-0.06953758746385574,
0.019210707396268845,
0.031717490404844284,
0.013455266132950783,
-0.08386791497468948,
-0.06750485301017761,
-0.019575495272874832,
-0.012390246614813805,
-0.04796808585524559,
-0.025137435644865036,
-0.09037472307682037,
0.04976833984255791,
0.003610297804698348,
0.03986165300011635,
-0.019987767562270164,
-0.022437015548348427,
0.08279169350862503,
-0.025330206379294395,
-0.11819032579660416,
0.04456038028001785,
0.03989933431148529,
0.015821073204278946,
0.013880143873393536,
0.09259716421365738,
-0.02932724729180336,
-0.011808842420578003,
-0.0010096578625962138,
0.021759456023573875,
-0.025008367374539375,
0.04200104624032974,
0.06360545754432678,
-0.01744953542947769,
0.08945715427398682,
0.05274718627333641,
0.06792381405830383,
-0.09248185902833939,
-0.03714181110262871,
-0.05195848271250725,
-0.004608004353940487,
-0.0820692703127861,
0.05047468841075897,
0.09349139779806137,
-0.03668787702918053,
0.041215598583221436,
-0.023791920393705368,
-0.039386995136737823,
-0.05208009108901024,
-0.025861945003271103,
-0.021505005657672882,
-0.09751297533512115,
-0.030230168253183365,
-0.004407442174851894,
-0.05401983857154846,
0.006961762439459562,
-0.06375956535339355,
0.08638665825128555,
-0.0386352464556694,
0.008237401954829693,
-0.09119928628206253,
-0.09845711290836334,
-0.0661470964550972,
0.009673705324530602,
0.034983694553375244,
-0.03455757722258568,
-0.04135572537779808,
0.0408715195953846,
0.07696777582168579,
-0.009394971653819084,
0.012029437348246574,
-0.07649832963943481,
-0.0015557213919237256,
-0.02852645516395569,
0.1242852434515953,
-0.005614192225039005,
0.03291868418455124,
0.014364906586706638,
-0.017775118350982666,
0.05040830001235008,
-0.0743466317653656,
0.013409934937953949,
0.01816345937550068,
-0.012334469705820084,
0.04728284850716591,
0.06531679630279541,
-0.002731871325522661,
0.1027931421995163,
0.08341962099075317,
0.02132837474346161,
0.11103783547878265,
0.012701642699539661,
-0.03140485659241676,
0.022608237341046333,
0.05181041359901428,
0.021286707371473312,
0.04915044084191322,
0.04191438853740692,
-0.05658148229122162,
-0.04482164978981018,
0.055035486817359924,
0.004952321294695139,
0.038270529359579086,
0.036241717636585236,
-0.05458394065499306,
-0.0318804569542408,
0.002746148267760873,
0.11889862269163132,
-0.030422257259488106,
0.03947906941175461,
-0.07150113582611084,
0.05004822090268135,
0.055399149656295776,
0.017603948712348938,
-0.03216272592544556,
0.05018897354602814,
0.04960912838578224,
0.05704103782773018,
0.01118666585534811,
0.08438601344823837,
0.02618270181119442,
-0.03583507612347603,
0.04646268114447594,
0.012858827598392963,
-0.04197448492050171,
-0.013347236439585686,
0.04817258566617966,
-0.0006227829726412892,
1.678076593174372e-33,
0.04476982727646828,
0.006786825601011515,
-0.02041817083954811,
0.05235770717263222,
0.06007853522896767,
-0.07306943088769913,
0.0842873826622963,
0.07076311111450195,
-0.1390167772769928,
-0.03525479510426521,
0.005147515330463648,
0.034705065190792084,
-0.053795453161001205,
0.07779247313737869,
0.006209805142134428,
-0.01105180662125349,
0.018480908125638962,
0.053634077310562134,
-0.0477600023150444,
0.04993823170661926,
0.04550667852163315,
0.012628694996237755,
0.049214139580726624,
-0.0460418239235878,
0.022952362895011902,
0.030658278614282608,
0.07296178489923477,
-0.06498400121927261,
-0.009746952913701534,
0.02341349422931671,
-0.10042057186365128,
-0.07467721402645111,
0.016253754496574402,
0.0759555995464325,
0.05904000997543335,
-0.0324011854827404,
-0.049535878002643585,
-0.04676162078976631,
-0.015774082392454147,
0.06147736683487892,
-0.020946500822901726,
-0.0038857797626405954,
-0.02838846668601036,
0.027582664042711258,
-0.045068856328725815,
-0.08061414211988449,
-0.02876441553235054,
0.0023823475930839777,
-0.014686630107462406,
-0.0246158204972744,
0.037906281650066376,
0.030416877940297127,
-0.025610094889998436,
0.018282469362020493,
-0.006808878853917122,
0.045861948281526566,
0.02445504069328308,
-0.03157297149300575,
0.047073401510715485,
0.048661258071660995,
-0.04955831542611122,
0.0579136423766613,
0.030127832666039467,
0.03266970440745354,
0.09136118739843369,
0.020612508058547974,
-0.024139195680618286,
0.06944648176431656,
0.07188136875629425,
0.012838456779718399,
-0.02682931162416935,
-0.0084017189219594,
-0.006688375025987625,
-0.016334816813468933,
0.026053963229060173,
0.008855661377310753,
0.024577956646680832,
-0.11570705473423004,
0.05464422330260277,
-0.0460037961602211,
0.026398791000247,
-0.02073521539568901,
0.02048400230705738,
-0.04440758004784584,
-0.022453974932432175,
-0.05023900046944618,
0.03248990699648857,
-0.03506968915462494,
0.029050270095467567,
0.0007097062771208584,
-0.045613814145326614,
-0.11871064454317093,
0.00581154553219676,
-0.07696995139122009,
-0.004216247703880072,
-1.1872258762789901e-33,
0.01523875817656517,
-0.05843212828040123,
-0.03598205745220184,
0.08309782296419144,
-0.07663891464471817,
-0.12110049277544022,
-0.0004812921688426286,
0.025121673941612244,
-0.039072148501873016,
-0.02287968061864376,
-0.03430101275444031,
-0.06750631332397461,
0.06330293416976929,
0.026586201041936874,
0.07312532514333725,
-0.033185623586177826,
-0.013695603236556053,
0.046589139848947525,
-0.003406072035431862,
0.1161523312330246,
-0.06634626537561417,
0.11244269460439682,
-0.006195360329002142,
-0.021782133728265762,
-0.05859630927443504,
0.03213835880160332,
-0.0609770342707634,
0.013554987497627735,
-0.011723987758159637,
-0.046772368252277374,
0.030986761674284935,
0.062092360109090805,
-0.07472853362560272,
0.015990151092410088,
-0.021304691210389137,
0.040492236614227295,
0.06995952129364014,
0.01838100142776966,
-0.01602526381611824,
0.07258923351764679,
0.10743992030620575,
0.007461461704224348,
-0.0022285685408860445,
0.037591151893138885,
0.0020636997651308775,
0.05121847242116928,
-0.08615831285715103,
0.027321340516209602,
-0.0006929663359187543,
-0.03672067075967789,
-0.07267771661281586,
-0.019459299743175507,
-0.012686382979154587,
0.017158985137939453,
0.013824176974594593,
-0.15439164638519287,
-0.04065423831343651,
-0.06308048218488693,
-0.04124707356095314,
-0.040965210646390915,
-0.07714688777923584,
-0.035226788371801376,
0.07967128604650497,
-0.0343446210026741,
0.03215016797184944,
-0.019659824669361115,
-0.09624148905277252,
-0.027954258024692535,
-0.06795331835746765,
-0.05548613518476486,
0.033418651670217514,
-0.09141533076763153,
-0.06105232238769531,
0.0003816636744886637,
-0.00711046950891614,
0.03714900091290474,
-0.013940677978098392,
-0.003126970725134015,
-0.02097642607986927,
-0.029310252517461777,
0.06183586269617081,
-0.02593325451016426,
-0.01333111897110939,
-0.029865864664316177,
-0.008527395315468311,
0.057823240756988525,
0.027270875871181488,
0.10304635018110275,
0.031801532953977585,
0.0384308286011219,
-0.03384626656770706,
0.08871123194694519,
0.01842801459133625,
-0.005694300401955843,
0.03825973719358444,
-3.374408308332022e-8,
-0.028977854177355766,
0.02968592941761017,
-0.026958536356687546,
0.01879182644188404,
-0.03538567200303078,
0.0055441223084926605,
-0.029751379042863846,
0.06489044427871704,
0.010075615718960762,
0.005926655605435371,
0.04202483966946602,
0.004776454996317625,
-0.023234285414218903,
-0.05825808271765709,
-0.0429275743663311,
0.09246986359357834,
-0.03547544777393341,
0.0331096388399601,
0.04917469993233681,
0.039820775389671326,
0.025041943415999413,
0.07217918336391449,
0.02293791063129902,
-0.013505443930625916,
-0.06614522635936737,
-0.029204552993178368,
0.033861927688121796,
0.07054433226585388,
0.0516895093023777,
-0.1000080332159996,
0.06805983930826187,
-0.0219886414706707,
-0.04269438236951828,
-0.04484705626964569,
0.04185659810900688,
0.04048663005232811,
-0.06385871022939682,
0.006721089594066143,
0.06328406929969788,
0.010461202822625637,
0.025897199288010597,
0.018620973452925682,
-0.03756590187549591,
0.03499661013484001,
-0.00976166594773531,
-0.07757294178009033,
-0.08461843430995941,
-0.0700870230793953,
-0.012590521946549416,
-0.0962260365486145,
-0.029548777267336845,
-0.08691577613353729,
-0.06191035360097885,
0.023394813761115074,
0.061179086565971375,
0.08593009412288666,
-0.11576782166957855,
-0.01606644131243229,
0.011475182138383389,
0.103282131254673,
0.017138801515102386,
-0.002840303350239992,
0.01896318793296814,
0.020191706717014313
] |
Re-Evaluating the Impact of Unseen-Class Unlabeled Data on Semi-Supervised Learning Model | https://openreview.net/forum?id=WPsnH6875d | [
"Rundong He",
"Yicong Dong",
"Lan-Zhe Guo",
"Yilong Yin",
"Tailin Wu"
] | Poster | Semi-supervised learning (SSL) effectively leverages unlabeled data and has been proven successful across various fields. Current safe SSL methods believe that unseen classes in unlabeled data harm the performance of SSL models. However, previous methods for assessing the impact of unseen classes on SSL model performance are flawed. They fix the size of the unlabeled dataset and adjust the proportion of unseen classes within the unlabeled data to assess the impact. This process contravenes the principle of controlling variables. Adjusting the proportion of unseen classes in unlabeled data alters the proportion of seen classes, meaning the decreased classification performance of seen classes may not be due to an increase in unseen class samples in the unlabeled data, but rather a decrease in seen class samples. Thus, the prior flawed assessment standard that "unseen classes in unlabeled data can damage SSL model performance" may not always hold true. This paper strictly adheres to the principle of controlling variables, maintaining the proportion of seen classes in unlabeled data while only changing the unseen classes across five critical dimensions, to investigate their impact on SSL models from global robustness and local robustness. Experiments demonstrate that unseen classes in unlabeled data do not necessarily impair the performance of SSL models; in fact, under certain conditions, unseen classes may even enhance them. | Safe Semi-Supervised Learning, Unseen-Class Unlabeled Data | Re-Evaluating the Impact of Unseen-Class Unlabeled Data on Semi-Supervised Learning Model | 2,001 | 2503.00884 | [
-0.06174786761403084,
-0.018024388700723648,
-0.05318306013941765,
0.12365448474884033,
0.0566633939743042,
0.06220804899930954,
-0.026275059208273888,
-0.07043109089136124,
0.04115757718682289,
-0.03737460449337959,
0.03148237615823746,
0.01623544842004776,
0.1171150878071785,
-0.050489045679569244,
-0.007713133469223976,
-0.07943180203437805,
0.07335474342107773,
0.029475878924131393,
-0.15237422287464142,
-0.02532682567834854,
0.07162827253341675,
0.0014426250709220767,
0.0031366373877972364,
0.04949198663234711,
-0.0037081995978951454,
-0.05252009630203247,
0.03145560249686241,
0.033499814569950104,
-0.05693111568689346,
-0.07672080397605896,
-0.050462204962968826,
0.06602957844734192,
-0.04381578043103218,
0.04705679789185524,
-0.028665360063314438,
-0.07289344817399979,
0.01355712953954935,
0.09514840692281723,
0.02149701677262783,
-0.038429535925388336,
-0.005796842277050018,
-0.021462956443428993,
-0.05729980766773224,
-0.006252681370824575,
-0.040450531989336014,
-0.023982804268598557,
-0.01997307315468788,
-0.04772208631038666,
-0.08665341883897781,
-0.09040050208568573,
-0.060508422553539276,
-0.060836534947156906,
-0.01928526535630226,
0.07275212556123734,
-0.11670288443565369,
-0.05280381441116333,
0.03559921681880951,
-0.05825149640440941,
-0.059479791671037674,
0.056395597755908966,
-0.09119334071874619,
-0.03267662227153778,
-0.03088143840432167,
0.01659291423857212,
-0.006854940205812454,
0.0436604842543602,
-0.03343961387872696,
-0.015081021934747696,
-0.013560239225625992,
0.12652407586574554,
0.04884013533592224,
0.087104931473732,
-0.049618691205978394,
0.023333322256803513,
0.05399604141712189,
-0.06971891969442368,
-0.022793272510170937,
0.014164578169584274,
0.08925454318523407,
-0.0767756998538971,
-0.006172023713588715,
0.028054192662239075,
-0.04458516463637352,
0.027878008782863617,
0.05200790613889694,
0.008490703999996185,
0.03160681575536728,
0.02166426181793213,
0.008450343273580074,
0.039356768131256104,
0.03823353722691536,
0.045126575976610184,
0.027549801394343376,
-0.0035462442319840193,
0.011235005222260952,
-0.03604615479707718,
0.01876935362815857,
-0.028652526438236237,
0.0373031385242939,
-0.002021436346694827,
-0.02088603936135769,
0.06733565777540207,
0.030705299228429794,
-0.03679772466421127,
0.01962919905781746,
0.02581142634153366,
0.07383875548839569,
0.01733609288930893,
0.034811485558748245,
-0.06207044795155525,
0.04657424986362457,
0.00011124707816634327,
-0.09287162870168686,
-0.0002516209497116506,
-0.053333211690187454,
0.021343309432268143,
0.03942624479532242,
-0.021267302334308624,
-0.041245393455028534,
0.0037404040340334177,
-0.019950291141867638,
0.019289249554276466,
0.059002120047807693,
0.06386914849281311,
-0.060170698910951614,
0.04935050383210182,
-0.12477369606494904,
6.619194730484885e-33,
0.01606236957013607,
-0.025528766214847565,
-0.09031309187412262,
-0.07391805946826935,
0.03127273917198181,
-0.016910633072257042,
-0.02237899787724018,
0.0330762080848217,
-0.012067290022969246,
0.08148553222417831,
0.029594719409942627,
0.038970451802015305,
-0.0008123813895508647,
0.03459249436855316,
0.09116582572460175,
0.016549699008464813,
-0.013179264031350613,
0.10441502928733826,
-0.05832255631685257,
-0.026159385219216347,
0.04941928759217262,
-0.026801183819770813,
0.07120718061923981,
-0.06325710564851761,
-0.027782021090388298,
0.0023932787589728832,
-0.04007711634039879,
0.036164380609989166,
-0.022534674033522606,
0.0184829980134964,
0.05214489623904228,
-0.021160386502742767,
0.06716766953468323,
0.02980721741914749,
0.06865470856428146,
0.06783311069011688,
0.03992636129260063,
0.03796074539422989,
0.054387178272008896,
0.016819903627038002,
0.007839140482246876,
0.04656793922185898,
0.02536037378013134,
0.015806663781404495,
-0.0264540147036314,
-0.01244346983730793,
0.055911723524332047,
-0.015896739438176155,
-0.046255409717559814,
-0.017350059002637863,
-0.0032342232298105955,
0.031020760536193848,
-0.00280612800270319,
-0.08012417703866959,
-0.11333465576171875,
0.054929137229919434,
0.0633048340678215,
0.016653606668114662,
-0.044455431401729584,
0.026394814252853394,
-0.11330004036426544,
-0.004548834171146154,
0.009725159965455532,
0.008508006110787392,
0.042877111583948135,
0.031208934262394905,
0.0333501435816288,
-0.04919621720910072,
-0.008135714568197727,
-0.09315373748540878,
-0.009035835973918438,
0.023292794823646545,
-0.029073206707835197,
-0.017694098874926567,
0.05866193771362305,
-0.029406074434518814,
-0.00465236185118556,
0.05504963919520378,
0.05626694858074188,
0.028590872883796692,
0.07220368087291718,
-0.050361841917037964,
0.011509743519127369,
-0.02846471779048443,
-0.07079175859689713,
-0.04357517510652542,
0.04803008586168289,
0.008063464425504208,
-0.060680702328681946,
-0.016457319259643555,
0.024148281663656235,
-0.04410222917795181,
0.0057545006275177,
-0.034149400889873505,
0.03248445317149162,
-6.968219574395161e-33,
-0.10540766268968582,
-0.031115930527448654,
-0.006166035309433937,
0.048887766897678375,
-0.10219477862119675,
-0.04273037239909172,
-0.01345054991543293,
0.06587296724319458,
-0.011184425093233585,
-0.03806028142571449,
0.07271172851324081,
-0.014600323513150215,
-0.00006323255365714431,
-0.004102121107280254,
0.018662188202142715,
-0.013874657452106476,
-0.053072426468133926,
0.022306300699710846,
-0.022939365357160568,
0.08758175373077393,
0.06169790402054787,
0.045457202941179276,
-0.019119558855891228,
0.1146443784236908,
-0.020888283848762512,
0.027284227311611176,
-0.0805153176188469,
0.046807046979665756,
0.03046686388552189,
-0.009597556665539742,
-0.007894872687757015,
-0.008958781138062477,
-0.01746516115963459,
-0.044729817658662796,
-0.09118589013814926,
-0.002427720231935382,
0.0362745076417923,
-0.017166340723633766,
-0.009455417282879353,
0.08960608392953873,
0.016890551894903183,
0.051455579698085785,
-0.1777123510837555,
-0.02630731649696827,
0.03743357956409454,
-0.07690765708684921,
0.01852130889892578,
0.07404019683599472,
0.017742251977324486,
-0.01428976096212864,
-0.028462572023272514,
0.05061034858226776,
-0.025119945406913757,
0.03227367252111435,
0.007454629987478256,
-0.07270582765340805,
0.02958797849714756,
-0.07982021570205688,
-0.034224867820739746,
0.05239056423306465,
-0.03655748441815376,
0.048656538128852844,
-0.03838996961712837,
0.020248888060450554,
0.014821171760559082,
-0.08653904497623444,
-0.08578270673751831,
0.04894266277551651,
0.03711145743727684,
-0.03860863670706749,
-0.005834689363837242,
-0.010240839794278145,
-0.0350165031850338,
-0.11737244576215744,
-0.026492947712540627,
-0.03652624413371086,
-0.003562880214303732,
0.07071390002965927,
-0.005490812938660383,
0.09237870573997498,
0.03188887611031532,
-0.048711538314819336,
-0.018349623307585716,
0.04719024896621704,
0.049840789288282394,
0.002351182047277689,
0.07196959853172302,
-0.03271612524986267,
-0.01626906730234623,
-0.0031739131081849337,
-0.07756663113832474,
0.026734231039881706,
-0.047689665108919144,
0.10756257176399231,
0.0019223723793402314,
-4.4276607269466695e-8,
-0.004336213227361441,
0.05387163907289505,
0.06067468225955963,
0.017696408554911613,
0.11284970492124557,
-0.03162938356399536,
-0.06309999525547028,
0.07973561435937881,
-0.06122063100337982,
0.005164641886949539,
-0.00807694811373949,
-0.022314654663205147,
-0.018326597288250923,
-0.009073324501514435,
-0.027603494003415108,
0.0269315205514431,
0.014758162200450897,
0.026871198788285255,
0.012370697222650051,
0.027907587587833405,
0.06406845897436142,
-0.05154314264655113,
0.04839656129479408,
-0.01458805613219738,
0.06650296598672867,
-0.08782566338777542,
-0.016499213874340057,
0.017131196334958076,
0.014843503013253212,
0.0035582503769546747,
-0.034579046070575714,
-0.022297339513897896,
-0.002681079087778926,
0.04922286421060562,
0.15337780117988586,
0.08832893520593643,
-0.04678629711270332,
-0.015513131394982338,
-0.04193929210305214,
0.13527846336364746,
-0.0059816534630954266,
0.038248106837272644,
-0.03231463581323624,
0.026548797264695168,
0.06954552978277206,
-0.03853209316730499,
0.007685533724725246,
-0.07892190665006638,
0.040558572858572006,
-0.010682845488190651,
-0.054532844573259354,
-0.028547167778015137,
-0.024611985310912132,
0.027259506285190582,
-0.022362206131219864,
0.05464301258325577,
0.034827202558517456,
0.0014362371293827891,
-0.01498881820589304,
0.09860872477293015,
0.0910153016448021,
-0.058998432010412216,
-0.09685450792312622,
0.024895619601011276
] |
Cross-Attention Head Position Patterns Can Align with Human Visual Concepts in Text-to-Image Generative Models | https://openreview.net/forum?id=1vggIT5vvj | [
"Jungwon Park",
"Jungmin Ko",
"Dongnam Byun",
"Jangwon Suh",
"Wonjong Rhee"
] | Poster | Recent text-to-image diffusion models leverage cross-attention layers, which have been effectively utilized to enhance a range of visual generative tasks. However, our understanding of cross-attention layers remains somewhat limited. In this study, we introduce a mechanistic interpretability approach for diffusion models by constructing Head Relevance Vectors (HRVs) that align with human-specified visual concepts. An HRV for a given visual concept has a length equal to the total number of cross-attention heads, with each element representing the importance of the corresponding head for the given visual concept. To validate HRVs as interpretable features, we develop an ordered weakening analysis that demonstrates their effectiveness. Furthermore, we propose concept strengthening and concept adjusting methods and apply them to enhance three visual generative tasks. Our results show that HRVs can reduce misinterpretations of polysemous words in image generation, successfully modify five challenging attributes in image editing, and mitigate catastrophic neglect in multi-concept generation. Overall, our work provides an advancement in understanding cross-attention layers and introduces new approaches for fine-controlling these layers at the head level. | text-to-image diffusion model, diffusion model, text-to-image generative model, cross-attention | null | 1,999 | 2412.02237 | [
0.03332173824310303,
-0.0268253181129694,
0.05973934382200241,
0.05963839590549469,
0.10632537305355072,
0.002410931047052145,
-0.047089505940675735,
-0.0004981855163350701,
0.05794471874833107,
-0.05736439675092697,
-0.04274451732635498,
-0.025245126336812973,
0.031111236661672592,
0.008822022005915642,
-0.044642459601163864,
0.004054309334605932,
0.030469229444861412,
0.0676315501332283,
-0.09515217691659927,
-0.006852343678474426,
0.03808344528079033,
0.006692162714898586,
0.08509552478790283,
-0.03704223036766052,
-0.010953119024634361,
0.046540211886167526,
-0.07819658517837524,
-0.03870338201522827,
0.024035239592194557,
-0.0693444237112999,
0.01269474346190691,
-0.03923628851771355,
-0.014501522295176983,
0.128847137093544,
-0.1190171018242836,
0.10255447030067444,
-0.07121852785348892,
0.09851746261119843,
-0.013682716526091099,
-0.06550320982933044,
0.03489900752902031,
0.02059127949178219,
0.005303491838276386,
0.010804685764014721,
0.0773564949631691,
-0.008708089590072632,
-0.03975319117307663,
-0.09792207181453705,
0.0007440603221766651,
-0.01900813914835453,
-0.093061164021492,
-0.0370914451777935,
-0.055627141147851944,
0.060329701751470566,
-0.019142085686326027,
0.06944365799427032,
0.07033902406692505,
0.011477433145046234,
0.02601712942123413,
-0.03202800825238228,
-0.0738828256726265,
-0.00038122173282317817,
0.062233757227659225,
-0.043564315885305405,
0.053400978446006775,
-0.060677871108055115,
0.035947974771261215,
0.0080205537378788,
-0.0680512860417366,
0.060793180018663406,
0.02487064152956009,
0.029066935181617737,
-0.04431556165218353,
-0.024396926164627075,
0.04684777930378914,
0.1257675141096115,
-0.010873123072087765,
-0.043076325207948685,
0.005006833001971245,
-0.0796983614563942,
-0.002961315680295229,
0.029282649978995323,
0.07872362434864044,
-0.012421111576259136,
0.12269814312458038,
0.03595516085624695,
-0.02149251475930214,
-0.014633430168032646,
-0.05312094837427139,
0.018065201118588448,
-0.06377453356981277,
-0.07663938403129578,
0.05321831628680229,
-0.02390247769653797,
0.12385337799787521,
0.021686367690563202,
-0.03075908124446869,
-0.039471399039030075,
-0.020971832796931267,
0.008056257851421833,
0.041770029813051224,
0.01561765931546688,
-0.02136058546602726,
-0.05501493066549301,
0.07814577966928482,
-0.043109577149152756,
0.04202873632311821,
0.035255078226327896,
-0.010591414757072926,
-0.07227113842964172,
0.0342530757188797,
-0.002725435420870781,
-0.020931335166096687,
-0.05377328395843506,
0.008925754576921463,
-0.07655128836631775,
0.06468556821346283,
0.0020785450469702482,
0.03893590718507767,
-0.02731931582093239,
0.03542168810963631,
0.04843780770897865,
-0.04549279063940048,
-0.030231863260269165,
0.07399304211139679,
-0.06795795261859894,
0.009668863378465176,
-3.454486089777661e-34,
-0.00890804547816515,
-0.026308031752705574,
0.052472423762083054,
0.06581902503967285,
0.062438517808914185,
0.009913958609104156,
-0.036981940269470215,
-0.03137553110718727,
-0.08915568888187408,
-0.0030124899931252003,
-0.014806565828621387,
0.004620156716555357,
-0.08723472058773041,
0.09280967712402344,
0.016650494188070297,
-0.07898429036140442,
-0.015101185999810696,
-0.0019825082272291183,
-0.03973058983683586,
-0.001305034034885466,
-0.06877097487449646,
-0.006127023138105869,
-0.04165675491094589,
-0.04218389466404915,
-0.01616002433001995,
0.011992520652711391,
0.07159090787172318,
-0.09633432328701019,
-0.05860920995473862,
-0.0024930196814239025,
-0.16064725816249847,
-0.0021762382239103317,
0.0481095016002655,
-0.035846058279275894,
-0.008842040784657001,
-0.03764099255204201,
-0.05702107027173042,
-0.01706528849899769,
0.07945463061332703,
0.029631884768605232,
-0.023264411836862564,
0.05661695450544357,
-0.06174097955226898,
-0.004284239374101162,
-0.05949736386537552,
0.07077004760503769,
0.02366442233324051,
0.09853624552488327,
-0.05498034507036209,
-0.031177673488855362,
0.01257410179823637,
0.029560446739196777,
0.0136958546936512,
-0.04522031173110008,
-0.019590554758906364,
-0.024874644353985786,
0.049779925495386124,
0.01292488444596529,
0.006770163308829069,
-0.011120963841676712,
0.1200706958770752,
0.05537376552820206,
-0.04458491876721382,
0.08366189897060394,
0.04476211592555046,
-0.02915048785507679,
0.029485639184713364,
0.1007261872291565,
0.005081238225102425,
0.061927974224090576,
-0.019016260281205177,
0.031029460951685905,
-0.007331703323870897,
-0.047397322952747345,
0.018774772062897682,
0.0003868054482154548,
0.02827785164117813,
-0.06285359710454941,
0.10467798262834549,
0.00016750814393162727,
-0.006295568775385618,
-0.026892025023698807,
0.027428243309259415,
-0.043846890330314636,
-0.04729900509119034,
-0.012930730357766151,
0.026818037033081055,
-0.080818310379982,
-0.07419473677873611,
-0.017135130241513252,
0.043113771826028824,
-0.029550982639193535,
-0.004457784816622734,
-0.03948782756924629,
-0.04891176149249077,
5.495716188357476e-34,
-0.008141313679516315,
0.05132704973220825,
0.011081020347774029,
0.005435756873339415,
-0.013086955063045025,
0.016032585874199867,
0.0015544728375971317,
0.027798134833574295,
-0.036097049713134766,
-0.03152814134955406,
0.004734097979962826,
-0.06344611197710037,
-0.026528429239988327,
-0.04616006091237068,
-0.005545985884964466,
0.013285046443343163,
0.059810470789670944,
0.0365871824324131,
-0.07050042599439621,
0.02728840336203575,
0.006603062152862549,
0.09337121993303299,
-0.05154208093881607,
0.04060071334242821,
-0.05901181697845459,
0.05215837061405182,
0.01257361564785242,
0.03257884830236435,
0.013085132464766502,
-0.10279489308595657,
-0.037077825516462326,
0.022315217182040215,
0.023960430175065994,
0.007478856481611729,
-0.10479898750782013,
0.03808305785059929,
-0.002619874896481633,
-0.07803263515233994,
-0.05241936817765236,
0.1647864580154419,
0.05934707820415497,
-0.0594787634909153,
-0.013201143592596054,
-0.04008312150835991,
-0.03500671312212944,
-0.04476998746395111,
-0.09661243110895157,
-0.01171726081520319,
-0.03704505413770676,
0.11364448815584183,
-0.03689068183302879,
0.03445965796709061,
-0.11240451782941818,
0.02687169425189495,
-0.1158619076013565,
-0.04805273190140724,
-0.07715177536010742,
-0.01859980821609497,
0.11853784322738647,
-0.06771795451641083,
0.0031504433136433363,
-0.006811704020947218,
0.020247608423233032,
-0.033895108848810196,
-0.02605275623500347,
0.008258935995399952,
-0.03052925318479538,
-0.06008991226553917,
0.00544637069106102,
-0.04895402863621712,
0.04375210776925087,
-0.028761450201272964,
0.019523775205016136,
-0.022798912599682808,
0.04418332502245903,
-0.013970846310257912,
0.03357446938753128,
0.004597249440848827,
0.028116079047322273,
-0.1103362888097763,
0.02123812399804592,
-0.0015722678508609533,
-0.012191638350486755,
0.03493356332182884,
0.03579156473278999,
0.09926736354827881,
0.021592961624264717,
0.04054297134280205,
0.012730938382446766,
-0.03553410619497299,
-0.020212972536683083,
-0.03422553464770317,
-0.0159983541816473,
0.10088157653808594,
-0.06300889700651169,
-3.630088585282465e-8,
-0.07338424026966095,
-0.01247399765998125,
-0.012201190926134586,
0.0467342734336853,
0.01680768094956875,
-0.04675793647766113,
-0.00879314448684454,
0.10112657397985458,
-0.034927256405353546,
0.008914166130125523,
0.029979148879647255,
-0.007299200166016817,
0.0006168640684336424,
-0.02793724648654461,
0.015929367393255234,
0.05029049515724182,
-0.020942864939570427,
0.06589483469724655,
0.012506776489317417,
-0.07771851867437363,
-0.05471659451723099,
0.0740693211555481,
0.004376057535409927,
0.051592838019132614,
0.01636883057653904,
-0.0059957318007946014,
-0.06411953270435333,
0.059729501605033875,
0.06402207911014557,
-0.08472707867622375,
0.092941515147686,
0.07103472948074341,
-0.020616644993424416,
0.00025888491654768586,
0.008716415613889694,
0.07865208387374878,
0.02018275111913681,
0.022671161219477654,
-0.036843955516815186,
0.04161500930786133,
-0.002058679237961769,
0.07395656406879425,
0.018061593174934387,
0.08220041543245316,
0.05581487715244293,
0.02042095921933651,
0.05106378719210625,
-0.04508509486913681,
-0.013479686342179775,
-0.07284622639417648,
0.0410541296005249,
0.0064870924688875675,
0.027207165956497192,
0.05172758176922798,
0.038503844290971756,
-0.035314496606588364,
0.062120094895362854,
0.035783663392066956,
0.009660575538873672,
0.059405576437711716,
0.04899153485894203,
0.007115856744349003,
0.024335620924830437,
0.03711197152733803
] |
FormalAlign: Automated Alignment Evaluation for Autoformalization | https://openreview.net/forum?id=B5RrIFMqbe | [
"Jianqiao Lu",
"Yingjia Wan",
"Yinya Huang",
"Jing Xiong",
"Zhengying Liu",
"Zhijiang Guo"
] | Poster | Autoformalization aims to convert informal mathematical proofs into machine-verifiable formats, bridging the gap between natural and formal languages. However, ensuring semantic alignment between the informal and formalized statements remains challenging. Existing approaches heavily rely on manual verification, hindering scalability. To address this, we introduce FormalAlign, a framework for automatically evaluating the alignment between natural and formal languages in autoformalization. FormalAlign trains on both the autoformalization sequence generation task and the representational alignment between input and output, employing a dual loss that combines a pair of mutually enhancing autoformalization and alignment tasks. Evaluated across four benchmarks augmented by our proposed misalignment strategies, FormalAlign demonstrates superior performance. In our experiments, FormalAlign outperforms GPT-4, achieving an Alignment-Selection Score 11.58\% higher on \forml-Basic (99.21\% vs. 88.91\%) and 3.19\% higher on MiniF2F-Valid (66.39\% vs. 64.34\%). This effective alignment evaluation significantly reduces the need for manual verification. | Large Language models, Autoformalization, Lean 4, Formal Math, AI for Math | FormalAlign is a framework that automatically evaluates the alignment between informal and formal mathematical proofs, significantly improving performance and reducing reliance on manual verification. | 1,997 | 2410.10135 | [
-0.09262554347515106,
0.015338348224759102,
-0.011660777032375336,
-0.022068535909056664,
-0.021306650713086128,
-0.02931426465511322,
-0.05448809266090393,
0.03601916879415512,
0.005562779027968645,
-0.0005646006902679801,
-0.03683941811323166,
-0.13055996596813202,
-0.009195261634886265,
0.06213128939270973,
-0.009249897673726082,
-0.07011301070451736,
-0.06494966149330139,
0.07597365230321884,
-0.02184220589697361,
-0.039816927164793015,
-0.014843400567770004,
0.025855261832475662,
-0.015443878248333931,
0.06126216799020767,
0.04273305833339691,
0.007105756551027298,
0.017314322292804718,
-0.02876647748053074,
0.015592915005981922,
-0.066155806183815,
-0.06442456692457199,
-0.009090407751500607,
0.03910858929157257,
0.015581030398607254,
0.017492124810814857,
0.058463022112846375,
-0.05151565745472908,
-0.026307765394449234,
0.0040270425379276276,
-0.10903476923704147,
-0.014019266702234745,
-0.07808949798345566,
0.028864391148090363,
0.00011239585001021624,
-0.010166898369789124,
0.04993743449449539,
-0.03148030489683151,
-0.0008221245370805264,
-0.006205282174050808,
0.019644999876618385,
-0.09936897456645966,
0.03290167823433876,
-0.007058513350784779,
0.017449714243412018,
-0.04012643173336983,
0.09792567044496536,
0.08451665937900543,
0.02531730942428112,
-0.01555326022207737,
-0.02457069233059883,
0.005240059457719326,
-0.022362934425473213,
-0.07011342793703079,
0.029461462050676346,
0.0025084479711949825,
0.02572333812713623,
0.03672075644135475,
0.03058922104537487,
-0.021219147369265556,
0.06330953538417816,
-0.10196869820356369,
0.059724658727645874,
-0.09080778807401657,
0.07589710503816605,
-0.012007325887680054,
0.0787806510925293,
-0.04591807350516319,
-0.01810298301279545,
0.0005797157646156847,
0.016368869692087173,
-0.006840561516582966,
0.06303637474775314,
0.035655438899993896,
0.05406677722930908,
0.06554538756608963,
-0.06544395536184311,
-0.004877464380115271,
-0.015038340352475643,
0.08353590220212936,
0.018195968121290207,
-0.050644468516111374,
-0.07242166250944138,
0.05243424326181412,
0.06745858490467072,
0.06054123118519783,
0.04516147822141647,
0.05892156437039375,
-0.024028325453400612,
0.03122309036552906,
0.044176001101732254,
0.02534778043627739,
0.10253240168094635,
-0.023782512173056602,
-0.12199082970619202,
0.039271555840969086,
-0.024790681898593903,
0.03653593361377716,
-0.029739126563072205,
0.053389377892017365,
-0.11247939616441727,
0.07572026550769806,
-0.018195614218711853,
0.02159285731613636,
0.021601606160402298,
-0.057744912803173065,
0.08563374727964401,
-0.0605311393737793,
-0.0225348100066185,
0.026225248351693153,
-0.014657050371170044,
-0.08253755420446396,
0.008383662439882755,
0.0016069484408944845,
0.010648541152477264,
0.08659278601408005,
-0.05880632624030113,
0.003488735994324088,
3.153323840166266e-33,
0.030620357021689415,
0.07768340408802032,
0.00980261992663145,
0.10032430291175842,
0.005892846267670393,
-0.004202230833470821,
-0.06805451959371567,
-0.009752820245921612,
-0.017524853348731995,
0.02449430711567402,
-0.0037346186581999063,
0.010786807164549828,
-0.04209413379430771,
0.04421094059944153,
-0.01420053094625473,
0.027088256552815437,
-0.013861091807484627,
0.0610615536570549,
-0.08753661811351776,
0.0574486069381237,
0.06353772431612015,
0.04223466292023659,
-0.020266955718398094,
-0.07375447452068329,
0.04199669510126114,
0.07361721992492676,
0.02788640931248665,
-0.024051619693636894,
-0.08455293625593185,
-0.0036286707036197186,
-0.06503952294588089,
-0.13612625002861023,
0.03190118819475174,
0.03401883319020271,
0.03462601453065872,
0.01346142590045929,
0.034179847687482834,
-0.11818503588438034,
0.0038195792585611343,
0.018688738346099854,
0.029768237844109535,
0.08320428431034088,
-0.009955200366675854,
0.026271872222423553,
0.031002739444375038,
-0.022168360650539398,
0.04307780787348747,
0.05186104774475098,
0.05038905888795853,
-0.06628860533237457,
0.02668480947613716,
0.06947546452283859,
-0.03530022129416466,
-0.017822010442614555,
-0.048226695507764816,
0.03829061985015869,
0.008994114585220814,
0.007048817817121744,
0.02815309166908264,
0.04741581901907921,
-0.013248385861515999,
0.06945992261171341,
-0.023087313398718834,
0.01612897217273712,
-0.03699744865298271,
-0.021049022674560547,
-0.0027933332603424788,
0.04052392393350601,
0.08515475690364838,
0.050384633243083954,
-0.054234180599451065,
-0.06405291706323624,
-0.07452629506587982,
0.016038646921515465,
0.08880282938480377,
-0.018670784309506416,
0.08004306256771088,
-0.12950122356414795,
0.023679807782173157,
-0.014838780276477337,
0.011784070171415806,
0.12593448162078857,
-0.005889510735869408,
-0.010106140747666359,
-0.022286154329776764,
-0.11715996265411377,
0.058229174464941025,
-0.02019786834716797,
-0.028879351913928986,
-0.011987796053290367,
-0.03834350407123566,
-0.012161589227616787,
0.01436707004904747,
0.05396803095936775,
-0.0032713976688683033,
-3.985233681870676e-33,
-0.00019932388386223465,
-0.00785659346729517,
0.020504305139183998,
0.13142380118370056,
-0.07253105193376541,
-0.05216532200574875,
0.043506134301424026,
-0.0014579802518710494,
0.0015144780045375228,
-0.0019929560367017984,
0.0159137099981308,
-0.02094079926609993,
0.011730577796697617,
-0.031953223049640656,
-0.038772594183683395,
-0.04720621928572655,
-0.06915616244077682,
0.06598195433616638,
0.010812954977154732,
0.045101411640644073,
0.061843547970056534,
0.08182696998119354,
-0.09092725068330765,
0.039433952420949936,
0.08368802070617676,
0.04183186963200569,
-0.01934441551566124,
0.037185151129961014,
-0.03029584512114525,
0.01780126430094242,
0.026209497824311256,
0.03358796611428261,
-0.025716494768857956,
-0.029730280861258507,
-0.03987743332982063,
-0.010803161188960075,
0.07719103246927261,
0.015633130446076393,
-0.0036172408144921064,
0.040431298315525055,
0.029397394508123398,
-0.0067022391594946384,
-0.04348481819033623,
0.019958648830652237,
-0.012788509950041771,
0.0014656116254627705,
-0.07508715242147446,
-0.03319345787167549,
-0.057307448238134384,
0.022190634161233902,
0.0054925354197621346,
0.008725566789507866,
-0.07403921335935593,
0.0029071406461298466,
-0.02211194485425949,
-0.04033275693655014,
-0.060469549149274826,
0.027380190789699554,
-0.1096733808517456,
0.05121516063809395,
-0.03147679194808006,
-0.02236029878258705,
0.11655234545469284,
-0.029766853898763657,
0.05666642636060715,
-0.04843148961663246,
-0.0010220876429229975,
0.026611458510160446,
0.018550945445895195,
0.016064489260315895,
0.04091989994049072,
-0.02116081863641739,
-0.050009988248348236,
0.09132207930088043,
0.03873152285814285,
-0.051188308745622635,
0.055624060332775116,
-0.007868684828281403,
0.04309677332639694,
-0.0069158561527729034,
0.027541037648916245,
0.034873053431510925,
0.029291845858097076,
0.055565301328897476,
-0.08321260660886765,
0.011785926297307014,
0.057548727840185165,
0.058912090957164764,
-0.02844352461397648,
0.11141416430473328,
-0.04969974607229233,
0.05028814822435379,
0.02851537987589836,
0.0435500368475914,
-0.13498663902282715,
-4.6427487632172415e-8,
-0.011774290353059769,
0.043553490191698074,
0.03926066681742668,
0.04203863441944122,
-0.10165969282388687,
-0.061929889023303986,
-0.08032630383968353,
-0.02218463458120823,
0.02437140978872776,
-0.10850456357002258,
-0.005127190612256527,
-0.04399843141436577,
-0.06009676679968834,
-0.07561417669057846,
-0.04429136589169502,
0.03729533404111862,
-0.024718351662158966,
-0.002360251732170582,
-0.10129737854003906,
0.041000865399837494,
-0.04997461289167404,
0.0932057648897171,
-0.07767092436552048,
-0.005228254944086075,
-0.026169346645474434,
-0.04769374430179596,
0.02664998173713684,
0.13839855790138245,
0.029830675572156906,
-0.056909505277872086,
-0.03184279426932335,
0.031892526894807816,
0.061956532299518585,
-0.013676491566002369,
-0.057944390922784805,
0.059403516352176666,
0.052000828087329865,
0.03131452202796936,
0.05991716310381889,
-0.0034087609965354204,
-0.02586153708398342,
0.08398634940385818,
-0.11139801889657974,
0.002681651385501027,
-0.012412945739924908,
-0.07781878858804703,
-0.04913458973169327,
-0.002888655522838235,
-0.041694026440382004,
-0.09042882174253464,
0.041544582694768906,
-0.050342947244644165,
-0.012046086601912975,
-0.028091927990317345,
0.019716564565896988,
0.03819149360060692,
-0.07606188207864761,
-0.024265654385089874,
0.04511597380042076,
0.013824227266013622,
-0.02444133535027504,
0.034822285175323486,
-0.009485115297138691,
-0.060896482318639755
] |
SEAL: Safety-enhanced Aligned LLM Fine-tuning via Bilevel Data Selection | https://openreview.net/forum?id=VHguhvcoM5 | [
"Han Shen",
"Pin-Yu Chen",
"Payel Das",
"Tianyi Chen"
] | Poster | Fine-tuning on task-specific data to boost downstream performance is a crucial step for leveraging Large Language Models (LLMs). However, though fine-tuning enhances the model performance for specialized applications, previous studies have demonstrated that fine-tuning the models on several adversarial samples or even benign data can greatly comprise the model's pre-equipped alignment and safety capabilities. In this work, we propose SEAL, a novel framework to enhance safety in LLM fine-tuning. SEAL learns a data ranker based on the bilevel optimization to up rank the safe and high-quality fine-tuning data and down rank the unsafe or low-quality ones. Models trained with SEAL demonstrate superior quality over multiple baselines, with 8.5\% and 9.7\% win rate increase compared to random selection respectively on Llama-3-8b-Instruct and Merlinite-7b models. Our code is available on github https://github.com/hanshen95/SEAL. | LLM fine-tuning, LLM safety, bilevel optimization | null | 1,995 | 2410.07471 | [
-0.05437107011675835,
-0.09983377158641815,
0.013160841539502144,
-0.0091977808624506,
0.02321169339120388,
-0.0003513367846608162,
-0.011385306715965271,
-0.02290170080959797,
-0.05313373729586601,
-0.05043361335992813,
-0.04503113031387329,
-0.036165304481983185,
0.07135694473981857,
-0.02417871728539467,
-0.05098668485879898,
0.00785414781421423,
0.09736577421426773,
0.09209953993558884,
-0.08172916620969772,
-0.054884884506464005,
0.010135134682059288,
0.031670935451984406,
0.08851910382509232,
-0.007520191837102175,
-0.02553841844201088,
-0.04957963898777962,
-0.009531018324196339,
0.031003493815660477,
0.01702459156513214,
-0.1259375363588333,
-0.06193748489022255,
-0.015126336365938187,
0.021761175245046616,
0.009950275532901287,
-0.14994943141937256,
0.014181816019117832,
-0.05918172746896744,
0.016884787008166313,
0.02073822356760502,
-0.056390997022390366,
-0.017092358320951462,
-0.018837817013263702,
-0.028953930363059044,
-0.016799284145236015,
0.024621766060590744,
-0.045716363936662674,
0.038847532123327255,
-0.052429720759391785,
-0.01164659857749939,
-0.014160077087581158,
-0.06638671457767487,
-0.030164668336510658,
0.0002887161390390247,
0.09030772000551224,
-0.05580213665962219,
0.019511474296450615,
0.0563945472240448,
-0.03724456951022148,
-0.043882809579372406,
-0.015527967363595963,
0.014863437041640282,
-0.02773263491690159,
-0.024376440793275833,
-0.013304498046636581,
0.018474452197551727,
-0.05974392592906952,
0.03309668228030205,
0.048406634479761124,
-0.05373190715909004,
0.018704421818256378,
0.0014680136227980256,
0.025244522839784622,
-0.02150050364434719,
0.034122347831726074,
0.019054146483540535,
0.0923723429441452,
0.021972952410578728,
-0.027126388624310493,
0.09788420796394348,
-0.00964268483221531,
-0.00029105565045028925,
-0.039052803069353104,
0.009755214676260948,
0.06250261515378952,
0.12356850504875183,
-0.006965507287532091,
0.011592982336878777,
-0.03950614109635353,
0.08133777230978012,
0.009105625562369823,
-0.0023818607442080975,
-0.0010239939438179135,
0.062055204063653946,
0.04761727154254913,
-0.0023411803413182497,
0.019597018137574196,
0.05158250778913498,
-0.03534324839711189,
-0.04555094242095947,
0.09529101848602295,
-0.010892217978835106,
0.03618844226002693,
-0.023981429636478424,
-0.039744678884744644,
0.08089809119701385,
-0.05552050471305847,
0.03344271332025528,
-0.009287404827773571,
0.04547911509871483,
-0.07955776154994965,
0.038911137729883194,
0.04938128590583801,
0.003974551800638437,
-0.0035259451251477003,
-0.04581611230969429,
0.03408970311284065,
-0.14747293293476105,
-0.013694358989596367,
-0.020278537645936012,
0.04087779298424721,
-0.061320722103118896,
0.021179605275392532,
0.012783165089786053,
0.0030845485161989927,
0.035028308629989624,
-0.02392078936100006,
-0.02945391647517681,
2.0536870785384176e-33,
0.07121738791465759,
0.020593887194991112,
-0.01957409642636776,
-0.011947701685130596,
0.04189276695251465,
-0.009858699515461922,
0.022882401943206787,
0.040964554995298386,
-0.08858024328947067,
0.021871302276849747,
-0.08035928010940552,
-0.05041445046663284,
-0.07058157026767731,
0.06716443598270416,
0.03909913823008537,
0.0270441472530365,
-0.018919644877314568,
0.03603645786643028,
-0.09247394651174545,
-0.05022360384464264,
0.11782907694578171,
-0.04545801505446434,
0.001120027620345354,
-0.10621130466461182,
0.04319624975323677,
0.026897897943854332,
0.0527460090816021,
-0.020933182910084724,
-0.06512663513422012,
0.05687422305345535,
-0.11045639216899872,
-0.02710672654211521,
0.0490562804043293,
0.02020701952278614,
0.06486958265304565,
0.03399791195988655,
-0.09157275408506393,
-0.02187512256205082,
0.0111087029799819,
-0.04282744601368904,
0.000048141540901269764,
0.02080029435455799,
0.02629736252129078,
-0.04903324693441391,
-0.06992006301879883,
-0.007484296336770058,
-0.05337102338671684,
0.0029991022311151028,
-0.013101542368531227,
-0.057229749858379364,
0.03895063325762749,
0.028507789596915245,
-0.03185472637414932,
0.05624114349484444,
-0.10864720493555069,
-0.0008204439072869718,
0.07634496688842773,
0.07757534086704254,
-0.040787920355796814,
0.07259272783994675,
-0.08578575402498245,
0.03648567199707031,
-0.03689403831958771,
0.05778726562857628,
0.06171165034174919,
0.017904093489050865,
0.031168896704912186,
0.01697327010333538,
0.039142362773418427,
-0.014209314249455929,
-0.027949564158916473,
-0.0038182265125215054,
-0.023686908185482025,
0.0073887682519853115,
-0.04668299853801727,
-0.036278072744607925,
0.10854817926883698,
-0.04194152355194092,
0.05810314044356346,
-0.03745646774768829,
0.04801907762885094,
0.02008071169257164,
0.01968306116759777,
-0.0813000276684761,
-0.12425930798053741,
-0.07771703600883484,
0.028312528505921364,
-0.02495565265417099,
0.0028820959851145744,
0.047465238720178604,
-0.02898598462343216,
0.032173920422792435,
0.054082583636045456,
-0.009552483446896076,
-0.04992785304784775,
-1.4329300932107647e-33,
-0.027331775054335594,
-0.0252663716673851,
0.0009704124531708658,
0.1403646469116211,
-0.05348929017782211,
-0.04257421940565109,
0.00021847145399078727,
0.07017569243907928,
-0.07858043164014816,
-0.048032451421022415,
-0.022691890597343445,
0.008291902020573616,
0.05825150012969971,
-0.03801598399877548,
0.061148177832365036,
-0.0004185604921076447,
-0.05349597707390785,
0.035977527499198914,
0.010020757094025612,
-0.002651829971000552,
0.03407333791255951,
0.09033985435962677,
-0.11690124124288559,
0.0845855325460434,
-0.04064171388745308,
-0.0018292962340638041,
-0.05215131491422653,
0.025992512702941895,
0.03324180096387863,
0.012430539354681969,
0.030853191390633583,
0.06882943212985992,
0.001745225046761334,
0.011716021224856377,
-0.051640577614307404,
0.012477836571633816,
0.09435436874628067,
-0.007829470559954643,
0.01444754283875227,
0.10448659211397171,
0.061895955353975296,
0.018417976796627045,
-0.08795013278722763,
-0.00839062687009573,
-0.017110291868448257,
-0.002804090501740575,
-0.038915470242500305,
-0.05061563476920128,
0.025352289900183678,
-0.0249563567340374,
0.02424442581832409,
0.004227968864142895,
-0.012119232676923275,
0.02318100444972515,
0.0037960477638989687,
-0.02164999209344387,
-0.048868320882320404,
-0.04489099979400635,
0.04076170548796654,
0.04220670834183693,
-0.016416216269135475,
-0.023197071626782417,
0.02684454806149006,
-0.1044069454073906,
-0.011168927885591984,
-0.009879274293780327,
-0.010521642863750458,
-0.020609894767403603,
-0.019838182255625725,
-0.04744221270084381,
0.027637368068099022,
-0.018138691782951355,
0.0023073959164321423,
0.04282722249627113,
-0.061126381158828735,
-0.052599918097257614,
-0.0334005169570446,
-0.059215642511844635,
-0.010337394662201405,
-0.004137413110584021,
0.04733040928840637,
0.0021586548537015915,
0.03374510258436203,
0.08891098946332932,
0.050158899277448654,
0.07644159346818924,
0.10068059712648392,
0.03685764595866203,
0.007659100461751223,
-0.02167631685733795,
-0.034027349203825,
0.04607076197862625,
-0.026929065585136414,
-0.03014271706342697,
-0.04158246889710426,
-4.562155098142284e-8,
-0.04879556596279144,
0.0034009921364486217,
-0.029361257329583168,
0.1271679699420929,
-0.016421038657426834,
-0.033067092299461365,
-0.03363792225718498,
0.07519230246543884,
0.048038680106401443,
-0.028194308280944824,
0.0679149180650711,
-0.00874136108905077,
-0.12430489808320999,
-0.051617398858070374,
0.05366549268364906,
0.047555454075336456,
-0.026718851178884506,
0.11956771463155746,
-0.044540517032146454,
-0.03906907141208649,
0.014904140494763851,
0.0852433443069458,
0.057697828859090805,
-0.056003350764513016,
0.05995063856244087,
-0.07265008985996246,
-0.04053472355008125,
0.12965461611747742,
0.089595727622509,
0.0727391466498375,
-0.012646539136767387,
-0.045154158025979996,
0.05915447324514389,
-0.0038507592398673296,
0.04247279465198517,
0.13637840747833252,
-0.0444328635931015,
-0.03304296359419823,
0.011313310824334621,
0.06914950162172318,
0.06781559437513351,
0.05215512961149216,
-0.05035225301980972,
0.033404361456632614,
-0.031551558524370193,
0.013326022773981094,
0.0016910373233258724,
-0.09236004203557968,
0.04127699136734009,
-0.01618606224656105,
0.08071507513523102,
-0.025350768119096756,
0.06531669944524765,
0.03874192759394646,
0.03509257361292839,
0.08643947541713715,
-0.009883029386401176,
-0.011812101118266582,
0.031313467770814896,
0.030462130904197693,
0.06530691683292389,
-0.07400508970022202,
-0.014657238498330116,
0.005753868259489536
] |
SeRA: Self-Reviewing and Alignment of LLMs using Implicit Reward Margins | https://openreview.net/forum?id=uIGnuyDSB9 | [
"Jongwoo Ko",
"Saket Dingliwal",
"Bhavana Ganesh",
"Sailik Sengupta",
"Sravan Babu Bodapati",
"Aram Galstyan"
] | Poster | Direct alignment algorithms (DAAs), such as direct preference optimization (DPO), have become popular alternatives to Reinforcement Learning from Human Feedback (RLHF) due to their simplicity, efficiency, and stability. However, the preferences used by DAAs are usually collected before alignment training begins and remain unchanged (off-policy). This design leads to two problems where the policy model (1) picks up on spurious correlations in the dataset (as opposed to only learning alignment to human preferences), and (2) overfits to feedback on off-policy trajectories that have less likelihood of being generated by the updated policy model. To address these issues, we introduce Self-Reviewing and Alignment (SeRA), a cost-efficient and effective method that can be readily combined with existing DAAs. SeRA comprises of two components: (1) sample selection using implicit reward margin to alleviate over-optimization on such undesired features, and (2) preference bootstrapping using implicit rewards to augment preference data with updated policy models in a cost-efficient manner. Extensive experiments, including on instruction-following tasks, demonstrate the effectiveness and generality of SeRA in training LLMs with diverse offline preference datasets and and DAAs. | Preference Alignment, Large Language Models, Implicit Reward Margin, Sample Selection, Preference Bootstrapping | null | 1,984 | null | [
-0.08107993006706238,
-0.04899398982524872,
-0.027427466586232185,
0.006631794851273298,
0.012971845455467701,
0.06018611788749695,
-0.007747902534902096,
0.00408230721950531,
0.04875804856419563,
0.058128777891397476,
-0.05120152607560158,
-0.06078379228711128,
-0.016444137319922447,
-0.004350490868091583,
0.01749737374484539,
-0.015654943883419037,
0.12727029621601105,
0.04467758163809776,
-0.047620873898267746,
0.004491292871534824,
-0.006013610400259495,
0.05578407645225525,
0.0769163966178894,
0.05313641205430031,
-0.04702373221516609,
0.019406184554100037,
0.02606453374028206,
0.039209041744470596,
0.006779050454497337,
-0.07035991549491882,
0.05724210664629936,
-0.027189375832676888,
0.05295452103018761,
-0.04281677305698395,
-0.08158896863460541,
0.039943594485521317,
-0.1317511945962906,
-0.0054739671759307384,
-0.07090751081705093,
-0.0526723749935627,
-0.043646737933158875,
-0.006131381262093782,
-0.008339213207364082,
0.025565557181835175,
0.06428616493940353,
0.024760333821177483,
-0.02987397462129593,
-0.020309781655669212,
0.01846645027399063,
0.04525492712855339,
-0.028917118906974792,
-0.004900634288787842,
0.019754430279135704,
-0.012399161234498024,
0.05503569915890694,
0.09655048698186874,
0.0751611739397049,
0.050622932612895966,
0.011464404873549938,
-0.04257709160447121,
0.011845833621919155,
-0.12552794814109802,
-0.04858727008104324,
0.024150867015123367,
-0.044894538819789886,
-0.03156168758869171,
0.02892226353287697,
0.014160852879285812,
-0.05513683706521988,
0.018498588353395462,
-0.08178305625915527,
0.07194151729345322,
0.05447802692651749,
0.020329870283603668,
0.022667517885565758,
0.08975089341402054,
0.028030233457684517,
-0.0016312700463458896,
0.07750952988862991,
-0.03909190371632576,
-0.03834257647395134,
0.023305697366595268,
0.052907176315784454,
0.09177745133638382,
0.03376489505171776,
-0.07119426131248474,
0.03202793002128601,
-0.0005232849507592618,
0.13316723704338074,
-0.030886705964803696,
0.008649790659546852,
-0.02144896797835827,
-0.01734117977321148,
-0.03434213995933533,
-0.049041226506233215,
0.0973343551158905,
-0.007025304716080427,
-0.024421613663434982,
-0.065728560090065,
0.06100856140255928,
0.059389811009168625,
0.0985240787267685,
-0.08562982082366943,
-0.015633435919880867,
-0.010152249597012997,
-0.06440092623233795,
-0.002574603771790862,
0.039980556815862656,
0.09484682977199554,
-0.09934160113334656,
0.027630692347884178,
-0.011446875520050526,
-0.02770180068910122,
0.044405821710824966,
-0.044228143990039825,
-0.0316348634660244,
0.03722824901342392,
-0.0884494036436081,
0.04072222113609314,
0.00959562323987484,
-0.010282519273459911,
-0.10478080809116364,
-0.0014904746785759926,
-0.024722754955291748,
0.016225598752498627,
-0.023266732692718506,
-0.02466961368918419,
1.903006518849255e-33,
0.02882290817797184,
0.0530773289501667,
0.050057120621204376,
-0.04007475823163986,
-0.022454893216490746,
-0.030964167788624763,
0.029291238635778427,
0.030899565666913986,
-0.07154817879199982,
0.032199420034885406,
-0.03396793454885483,
0.01337715145200491,
-0.04794710502028465,
0.1150282695889473,
0.00495087169110775,
-0.041377365589141846,
-0.05122103542089462,
0.08115708827972412,
-0.07159176468849182,
0.006523753982037306,
0.045681655406951904,
0.05542378127574921,
0.025089196860790253,
-0.04745600372552872,
0.06623297929763794,
-0.03701237216591835,
0.016686134040355682,
-0.01770925708115101,
-0.06676755845546722,
-0.0006701007951050997,
-0.06722081452608109,
-0.0433824360370636,
0.02856207825243473,
-0.006682353559881449,
-0.016284188255667686,
0.0039489432238042355,
-0.053227491676807404,
-0.02142215333878994,
0.020854132249951363,
-0.0010762815363705158,
0.0646190196275711,
0.06313174962997437,
0.04999745637178421,
0.04138380289077759,
-0.09167585521936417,
-0.06697939336299896,
0.024011066183447838,
-0.01781543530523777,
-0.0033121576998382807,
-0.03380841016769409,
0.0172381903976202,
0.06107025220990181,
0.0030804607085883617,
-0.05263475701212883,
-0.1083228811621666,
0.022786764428019524,
0.003745035966858268,
0.0707208439707756,
-0.01625128835439682,
0.011495061218738556,
-0.02019835077226162,
0.02513505332171917,
-0.009425867348909378,
0.017287319526076317,
0.009519346058368683,
0.03430820629000664,
0.015381862409412861,
0.04808298870921135,
0.05784304812550545,
-0.030741356313228607,
-0.005353026557713747,
0.04956904798746109,
-0.04137590155005455,
-0.0741630345582962,
0.06021063029766083,
-0.008532825857400894,
0.08559950441122055,
-0.0748453140258789,
0.04266776889562607,
-0.05375663936138153,
0.019283641129732132,
0.021428674459457397,
-0.038871996104717255,
-0.0551617406308651,
-0.002390957670286298,
-0.12394048273563385,
0.08727533370256424,
-0.07797324657440186,
-0.02729962021112442,
0.0249251127243042,
-0.03240960091352463,
0.0486435703933239,
0.00018084960174746811,
0.06932342797517776,
0.057914379984140396,
-2.0312575442987745e-33,
0.011580091901123524,
-0.02483944594860077,
0.06698240339756012,
0.03205230459570885,
-0.042179595679044724,
0.048835836350917816,
0.004915309604257345,
-0.03670033439993858,
-0.0069977412931621075,
-0.08224999904632568,
-0.08274085819721222,
-0.047391749918460846,
0.012726503424346447,
-0.02198495715856552,
-0.030413763597607613,
0.052787236869335175,
-0.08836472034454346,
0.036489639431238174,
0.030787276104092598,
-0.018248550593852997,
0.03543146327137947,
0.10486683249473572,
-0.10258160531520844,
0.049309249967336655,
-0.013448890298604965,
0.004975526127964258,
0.06604576110839844,
0.04393146187067032,
-0.0333675742149353,
0.028783759102225304,
-0.04003705456852913,
-0.004114433657377958,
-0.09295769780874252,
-0.0038416963070631027,
-0.048825472593307495,
0.08659126609563828,
-0.006876550614833832,
-0.02218204364180565,
-0.02825482003390789,
0.13027089834213257,
-0.026421230286359787,
0.02414211630821228,
-0.04280533641576767,
0.018498875200748444,
0.02008374035358429,
-0.02649344690144062,
-0.10363167524337769,
-0.013404224067926407,
-0.03847417235374451,
0.051381878554821014,
-0.06874384731054306,
0.03820463642477989,
-0.08772094547748566,
0.05907684192061424,
-0.07077070325613022,
-0.07838016748428345,
-0.038251616060733795,
-0.0052664028480648994,
0.06523288041353226,
0.015113793313503265,
-0.018319815397262573,
0.06231303885579109,
0.013636544346809387,
-0.06039523333311081,
0.03347371518611908,
0.023729361593723297,
0.10042620450258255,
-0.012319200672209263,
0.022336773574352264,
-0.008079475723206997,
0.017848296090960503,
-0.015652531757950783,
-0.008555183187127113,
-0.03018292412161827,
0.07487413287162781,
-0.0052140504121780396,
0.04011359065771103,
-0.04136814922094345,
0.005496429745107889,
-0.058068037033081055,
0.022596701979637146,
0.07471373677253723,
0.03611325845122337,
0.03849850967526436,
-0.03442809358239174,
0.08137770742177963,
0.07181872427463531,
0.131961390376091,
-0.030369536951184273,
0.026593685150146484,
-0.04825670272111893,
-0.014924809336662292,
0.07329410314559937,
0.005229689180850983,
-0.07278788834810257,
-5.055848717461231e-8,
-0.07711523026227951,
-0.01945994235575199,
0.056669868528842926,
0.1515471190214157,
-0.034817833453416824,
-0.03544255718588829,
0.03548717498779297,
0.06120337173342705,
-0.01012333296239376,
-0.07063265889883041,
0.05497392639517784,
0.08149030804634094,
-0.017648465931415558,
-0.09878726303577423,
-0.004055664408951998,
0.02492280676960945,
0.004817533772438765,
0.06535577028989792,
-0.0561375766992569,
0.03460328280925751,
0.02559846080839634,
0.037021003663539886,
-0.027756238356232643,
-0.042950063943862915,
0.07167761027812958,
-0.06495469808578491,
-0.007444167509675026,
0.09128575772047043,
0.0637013241648674,
0.0031482407357543707,
0.019761325791478157,
-0.06321071088314056,
0.02630077488720417,
-0.011746128089725971,
0.014596306718885899,
0.004124937113374472,
0.00304871192201972,
-0.016749752685427666,
0.054619766771793365,
-0.033079952001571655,
0.013621028512716293,
0.06217997521162033,
-0.06055091693997383,
-0.007122992072254419,
-0.03687373176217079,
-0.0019178733928129077,
-0.05107570439577103,
-0.09655351936817169,
0.015031453222036362,
-0.10090593993663788,
0.04914891719818115,
-0.09475717693567276,
0.00455255014821887,
0.027315780520439148,
0.024881023913621902,
-0.03452577441930771,
-0.009237361140549183,
-0.010967070236802101,
0.08761242032051086,
0.06610683351755142,
0.05694097280502319,
-0.05661015957593918,
-0.07664570212364197,
-0.012321875430643559
] |
Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling | https://openreview.net/forum?id=qZmn2hkuzw | [
"Yuejiang Liu",
"Jubayer Ibn Hamid",
"Annie Xie",
"Yoonho Lee",
"Max Du",
"Chelsea Finn"
] | Poster | Predicting and executing a sequence of actions without intermediate replanning, known as action chunking, is increasingly used in robot learning from human demonstrations. Yet, its effects on the learned policy remain inconsistent: some studies find it crucial for achieving strong results, while others observe decreased performance. In this paper, we first dissect how action chunking impacts the divergence between a learner and a demonstrator. We find that action chunking allows the learner to better capture the temporal dependencies in demonstrations but at the cost of reduced reactivity to unexpected states. To address this tradeoff, we propose Bidirectional Decoding (BID), a test-time inference algorithm that bridges action chunking with closed-loop adaptation. At each timestep, BID samples multiple candidate predictions and searches for the optimal one based on two criteria: (i) backward coherence, which favors samples that align with previous decisions; (ii) forward contrast, which seeks samples of high likelihood for future plans. By coupling decisions within and across action chunks, BID promotes both long-term consistency and short-term reactivity. Experimental results show that our method boosts the performance of two state-of-the-art generative policies across seven simulation benchmarks and two real-world tasks. Code and videos are available at https://bid-robot.github.io. | Robot Learning, Action Chunking, Action Decoding, Test-Time Sampling | We present a thorough analysis of action chunking and a decoding algorithm to improve it | 1,980 | null | [
-0.020360397174954414,
-0.09237797558307648,
0.07871909439563751,
0.014987904578447342,
0.08434783667325974,
-0.025432424619793892,
-0.032120686024427414,
-0.07802658528089523,
0.042712192982435226,
0.003573654219508171,
-0.02047961764037609,
0.003809912595897913,
-0.05278311297297478,
0.016851169988512993,
0.022093217819929123,
-0.033878393471241,
0.051383815705776215,
0.05018439516425133,
-0.03441251069307327,
0.005716462153941393,
0.05586273968219757,
-0.00297656562179327,
0.05324974283576012,
0.030513811856508255,
0.030698752030730247,
0.02007598616182804,
0.03152427077293396,
-0.08103766292333603,
0.017845474183559418,
-0.01116596907377243,
0.03225911408662796,
-0.04319734871387482,
-0.014571265317499638,
-0.08464325219392776,
-0.05767276883125305,
0.10554762929677963,
-0.07935363054275513,
-0.014003404416143894,
-0.05683930218219757,
0.0010310625657439232,
-0.03253453969955444,
-0.019618622958660126,
0.030960384756326675,
-0.008715982548892498,
0.0435723252594471,
0.047000691294670105,
0.052906736731529236,
-0.07939159125089645,
0.03286582976579666,
0.02660900168120861,
-0.01325907465070486,
0.03456060215830803,
0.05582654848694801,
-0.05015983805060387,
0.0077810585498809814,
0.029259370639920235,
0.060201145708560944,
0.06242818012833595,
-0.004769610241055489,
-0.0041314843110740185,
-0.009523370303213596,
-0.09944511950016022,
-0.03874209523200989,
-0.05842269957065582,
-0.03316403180360794,
-0.01058926247060299,
0.01399242039769888,
0.024168262258172035,
-0.0003324663557577878,
0.07014335691928864,
-0.019330553710460663,
0.025960266590118408,
0.006650221534073353,
0.012784288264811039,
-0.010660742409527302,
0.05264245346188545,
0.005935653578490019,
-0.0317976213991642,
0.025518016889691353,
-0.07375650852918625,
-0.016584882512688637,
-0.07289529591798782,
-0.025310497730970383,
0.007105755619704723,
0.08030440658330917,
-0.004054737742990255,
0.03379616141319275,
0.0950641855597496,
0.010353093035519123,
0.002269152784720063,
-0.019095933064818382,
-0.06647446751594543,
-0.005742761306464672,
-0.02399655245244503,
-0.030194547027349472,
0.0251860823482275,
-0.01541568711400032,
-0.061988797038793564,
-0.012340460903942585,
0.09232044965028763,
-0.050850771367549896,
0.0932290330529213,
-0.0060980967245996,
0.010088787414133549,
-0.06766124814748764,
-0.07023536413908005,
0.03500290587544441,
-0.01732758618891239,
0.049894336611032486,
0.00410464545711875,
0.01703071780502796,
-0.045783400535583496,
0.0829264298081398,
0.04781755805015564,
0.04783986136317253,
-0.012680095620453358,
-0.04772328585386276,
-0.02688043937087059,
-0.018360689282417297,
0.025208983570337296,
0.02300521917641163,
-0.0277795922011137,
-0.007017073687165976,
0.038812488317489624,
0.06880062073469162,
-0.029451612383127213,
-0.012313401326537132,
9.282831133985602e-34,
0.06764873117208481,
0.03836815059185028,
0.10441427677869797,
0.04606184363365173,
0.005638502538204193,
0.022773930802941322,
0.10564184933900833,
-0.01164042018353939,
0.00003862757876049727,
0.015044238418340683,
0.05379476398229599,
-0.06792135536670685,
-0.025234133005142212,
0.13672776520252228,
-0.01373172365128994,
0.028417283669114113,
-0.04178820177912712,
0.03156843036413193,
-0.07181382924318314,
0.029305001720786095,
0.0572177954018116,
-0.01738656498491764,
-0.07122679054737091,
-0.048958227038383484,
-0.01746251806616783,
-0.11678503453731537,
-0.0022005869541317225,
-0.013470286503434181,
-0.05300595611333847,
-0.011027062311768532,
-0.13710324466228485,
0.0067207389511168,
-0.06930328905582428,
0.06700389832258224,
0.06507622450590134,
-0.06988047063350677,
-0.01215688418596983,
-0.06886091828346252,
-0.055341944098472595,
-0.02388201281428337,
-0.020754439756274223,
-0.033117733895778656,
-0.004423867911100388,
-0.009373147040605545,
-0.07194318622350693,
-0.1083083376288414,
-0.002058079233393073,
-0.03686181455850601,
-0.03756088763475418,
-0.0040148720145225525,
0.04427080973982811,
0.06151556968688965,
0.00003906365600414574,
-0.1013358011841774,
0.014331674203276634,
-0.020402010530233383,
0.03899853676557541,
0.049880336970090866,
0.015517804771661758,
0.05107248201966286,
-0.0014079191023483872,
0.02455349825322628,
-0.0031678357627242804,
0.11427170783281326,
-0.028483251109719276,
0.1490173041820526,
-0.06010042503476143,
0.0873272493481636,
0.07876721769571304,
-0.032417625188827515,
-0.06412870436906815,
0.00893661193549633,
-0.011047585867345333,
-0.06442590802907944,
0.0759635716676712,
-0.01582968607544899,
0.04760991036891937,
-0.07695992290973663,
0.04674648493528366,
0.019882211461663246,
-0.006521906238049269,
-0.06290752440690994,
-0.0469362810254097,
0.01134916115552187,
-0.023358944803476334,
-0.02606274001300335,
0.027126824483275414,
-0.12194141000509262,
-0.018972013145685196,
-0.034237589687108994,
-0.12515704333782196,
-0.013163190335035324,
-0.05312342569231987,
0.03537842631340027,
0.08062616735696793,
-2.9270951769047164e-33,
0.052039340138435364,
-0.014163427986204624,
0.02262602187693119,
0.018288053572177887,
-0.021502364426851273,
-0.034114353358745575,
-0.040140263736248016,
0.004284861497581005,
0.04947996884584427,
-0.06489049643278122,
-0.06508517265319824,
-0.060274504125118256,
0.0403079017996788,
-0.02290959842503071,
-0.01402979250997305,
-0.014584806747734547,
0.03147045522928238,
0.01046384871006012,
0.09064091742038727,
0.008664223365485668,
0.005777042359113693,
0.10499045997858047,
-0.03786486014723778,
-0.011931913904845715,
-0.06408291310071945,
-0.006990728434175253,
0.02110031433403492,
0.09238124638795853,
0.022223204374313354,
0.004457481205463409,
-0.05468129366636276,
-0.03747770935297012,
-0.020222371444106102,
0.05564117804169655,
-0.06786275655031204,
0.08417578786611557,
0.03587054833769798,
0.013195096515119076,
-0.02638402208685875,
0.07605354487895966,
0.05464183911681175,
-0.07874894142150879,
-0.014059147797524929,
0.048383548855781555,
0.030163289979100227,
-0.07137960940599442,
-0.06936658918857574,
-0.010449015535414219,
-0.016667384654283524,
0.04349543899297714,
-0.02253558114171028,
0.08169718086719513,
-0.0999498963356018,
-0.02429027110338211,
-0.036966025829315186,
-0.05554067716002464,
-0.02295885793864727,
0.030556386336684227,
0.06460220366716385,
0.006734106689691544,
-0.06090031936764717,
0.0724431574344635,
0.022309906780719757,
0.0006992704584263265,
0.016331013292074203,
0.007726948708295822,
-0.007652918808162212,
-0.016265716403722763,
0.1321011781692505,
-0.013375557027757168,
0.030840832740068436,
-0.003778620855882764,
0.013347428292036057,
0.036792024970054626,
0.1197194829583168,
-0.027814464643597603,
-0.0079942112788558,
0.012088454328477383,
-0.03255889192223549,
-0.01167407725006342,
-0.004080158192664385,
0.00926899816840887,
0.02339084818959236,
0.05709732323884964,
0.00026697697467170656,
0.14535276591777802,
-0.09289932996034622,
0.09526576846837997,
-0.02055255137383938,
-0.03052675537765026,
-0.023514147847890854,
0.013528041541576385,
0.04930993169546127,
0.07308733463287354,
-0.030881404876708984,
-4.302560441260539e-8,
-0.04023415222764015,
0.08897102624177933,
0.06576990336179733,
0.00329611636698246,
0.03791315108537674,
-0.02450387366116047,
-0.013392604887485504,
-0.011004236526787281,
-0.020692389458417892,
-0.05778368189930916,
0.12821167707443237,
0.00045280728954821825,
0.01138367597013712,
-0.04671420529484749,
-0.06394925713539124,
0.0346834622323513,
0.006456010043621063,
-0.05290002375841141,
-0.041096534579992294,
0.016700830310583115,
0.034143611788749695,
-0.05608024820685387,
0.046150606125593185,
-0.02912631258368492,
-0.028771119192242622,
-0.05363449081778526,
0.0013071771245449781,
0.09535381197929382,
0.021042214706540108,
-0.027442779392004013,
0.04466163367033005,
0.010319062508642673,
0.0017483303090557456,
-0.0070742773823440075,
0.01330570224672556,
0.043716952204704285,
0.020808616653084755,
-0.04328914359211922,
0.04853956401348114,
0.008178440853953362,
-0.030109936371445656,
0.03944023326039314,
-0.02884039469063282,
0.010299069806933403,
-0.018899841234087944,
-0.028220416978001595,
-0.08641241490840912,
-0.1449964940547943,
0.0377337746322155,
-0.11681424826383591,
0.011139987036585808,
-0.035263996571302414,
-0.024654971435666084,
0.09801294654607773,
0.13466128706932068,
0.07382812350988388,
0.015487873926758766,
-0.06460347026586533,
-0.0261974073946476,
0.09485487639904022,
0.021300306543707848,
0.02010640688240528,
-0.09818123281002045,
-0.009710319340229034
] |
6DGS: Enhanced Direction-Aware Gaussian Splatting for Volumetric Rendering | https://openreview.net/forum?id=sUvBTEYXGt | [
"Zhongpai Gao",
"Benjamin Planche",
"Meng Zheng",
"Anwesa Choudhuri",
"Terrence Chen",
"Ziyan Wu"
] | Poster | Novel view synthesis has advanced significantly with the development of neural radiance fields (NeRF) and 3D Gaussian splatting (3DGS). However, achieving high quality without compromising real-time rendering remains challenging, particularly for physically-based rendering using ray/path tracing with view-dependent effects. Recently, N-dimensional Gaussians (N-DG) introduced a 6D spatial-angular representation to better incorporate view-dependent effects, but the Gaussian representation and control scheme are sub-optimal. In this paper, we revisit 6D Gaussians and introduce 6D Gaussian Splatting (6DGS), which enhances color and opacity representations and leverages the additional directional information in the 6D space for optimized Gaussian control. Our approach is fully compatible with the 3DGS framework and significantly improves real-time radiance field rendering by better modeling view-dependent effects and fine details. Experiments demonstrate that 6DGS significantly outperforms 3DGS and N-DG, achieving up to a 15.73 dB improvement in PSNR with a reduction of 66.5\% Gaussian points compared to 3DGS. The project page is: https://gaozhongpai.github.io/6dgs/. | 3D Gaussian splatting, 6D Gaussian splatting, volumetric rendering | The paper introduces 6D Gaussian Splatting (6DGS) for real-time radiance field rendering. It achieves up to a 15.73 dB PSNR boost while using 66.5% fewer Gaussian points over 3DGS. | 1,975 | 2410.04974 | [
-0.019071562215685844,
-0.13765756785869598,
0.03911622613668442,
-0.028953440487384796,
-0.009230975061655045,
0.010905330069363117,
-0.04631396755576134,
-0.07039713859558105,
-0.0009623344521969557,
-0.017072351649403572,
-0.0647834911942482,
0.03808489814400673,
-0.023598995059728622,
-0.0017140855779871345,
-0.030531249940395355,
0.007165775168687105,
0.0622132271528244,
0.10989071428775787,
-0.07502096146345139,
0.03955448418855667,
0.012291091494262218,
-0.030306117609143257,
-0.0587257519364357,
-0.031196149066090584,
0.08589968085289001,
-0.018097255378961563,
0.13221697509288788,
-0.04691097140312195,
0.026070887222886086,
0.017402177676558495,
0.029845604673027992,
0.01229774858802557,
-0.05944869667291641,
-0.007067559752613306,
-0.11876242607831955,
-0.014569205231964588,
0.04545297846198082,
-0.00602220231667161,
-0.03956879302859306,
-0.023236453533172607,
0.026939669623970985,
0.06461067497730255,
0.013573210686445236,
0.006104269530624151,
0.024366246536374092,
0.03037690557539463,
0.030055340379476547,
-0.08717934042215347,
0.01925409957766533,
-0.005019816104322672,
0.04131320118904114,
-0.06297200918197632,
0.003494822420179844,
-0.01822223700582981,
0.018911750987172127,
0.07882994413375854,
-0.053538136184215546,
-0.04713983088731766,
0.010789284482598305,
-0.009259594604372978,
-0.00149531289935112,
0.02289869636297226,
0.06684377789497375,
-0.001889890874736011,
0.014659382402896881,
-0.05388132110238075,
0.03967708349227905,
-0.05006207898259163,
-0.04919353500008583,
-0.012651447206735611,
-0.09127341210842133,
0.05510294437408447,
-0.017992625012993813,
-0.1062854528427124,
-0.07433914393186569,
0.045799072831869125,
-0.03589167073369026,
-0.07866000384092331,
-0.008269313722848892,
-0.0357406884431839,
0.062328778207302094,
0.041597433388233185,
-0.02798830159008503,
0.07548937946557999,
-0.0010910265846177936,
0.0009627751423977315,
0.008905965834856033,
0.17415305972099304,
0.006460872478783131,
0.014527087099850178,
0.011850504204630852,
0.03957164287567139,
-0.05539074167609215,
-0.11710819602012634,
-0.10648661851882935,
-0.030424900352954865,
-0.010910396464169025,
-0.07583228498697281,
0.023493247106671333,
-0.04146739840507507,
0.07391158491373062,
-0.042879942804574966,
0.02002573572099209,
0.05464998632669449,
0.08992768824100494,
-0.007315973285585642,
0.0595637708902359,
0.10933446884155273,
-0.042830437421798706,
0.03913591429591179,
-0.003834260394796729,
0.0510527528822422,
-0.021137595176696777,
0.014762152917683125,
0.011074553243815899,
0.07624306529760361,
-0.07112289220094681,
-0.016699550673365593,
0.027116257697343826,
0.008218508213758469,
-0.00490508321672678,
0.012952707707881927,
-0.07383006066083908,
0.06701051443815231,
0.06766253709793091,
-0.059140611439943314,
-0.035820409655570984,
2.8783679973272556e-33,
-0.031083082780241966,
0.06318977475166321,
0.0009100160677917302,
-0.007809891831129789,
0.04265512898564339,
0.04525146633386612,
0.011944863013923168,
0.008215348236262798,
-0.01641019433736801,
0.037032049149274826,
0.01260377001017332,
0.0010008621029555798,
0.002128082560375333,
0.054517026990652084,
0.04879763722419739,
0.026221299543976784,
-0.015898967161774635,
0.04434919357299805,
-0.04977340251207352,
0.03136514127254486,
-0.016723565757274628,
-0.010645647533237934,
-0.00014488704618997872,
-0.009524594061076641,
-0.11099527776241302,
0.0978875383734703,
-0.04033612087368965,
0.026216816157102585,
-0.056660786271095276,
-0.03888508677482605,
-0.03687627613544464,
0.015904363244771957,
-0.016396569088101387,
-0.00671864440664649,
-0.030031327158212662,
-0.04410420358181,
-0.03795094043016434,
-0.06631095707416534,
-0.0062164925038814545,
0.0196126326918602,
-0.07376352697610855,
0.10474187135696411,
0.021033691242337227,
0.019914988428354263,
-0.08642555773258209,
0.12135026603937149,
0.07227779924869537,
0.04398228973150253,
-0.018355127424001694,
0.04361080750823021,
0.026780355721712112,
-0.031188344582915306,
0.004461219999939203,
-0.027027331292629242,
0.07321778684854507,
-0.08048698306083679,
0.008872441947460175,
0.053820665925741196,
0.08481821417808533,
0.015049844048917294,
-0.015886636450886726,
0.008861283771693707,
-0.003222062950953841,
0.05817971006035805,
-0.08261726051568985,
0.004549712873995304,
0.013779218308627605,
0.047055382281541824,
0.010844378732144833,
0.01494202483445406,
-0.060005635023117065,
0.016411706805229187,
0.021530091762542725,
0.030952779576182365,
-0.0007389705860987306,
-0.02828424982726574,
0.060990165919065475,
-0.08106440305709839,
0.004227516707032919,
0.02827710285782814,
-0.06350051611661911,
0.039813749492168427,
-0.03329531475901604,
-0.04584487900137901,
-0.02241751179099083,
-0.05038583651185036,
-0.07211726903915405,
0.05694591626524925,
-0.07282058149576187,
-0.11225266009569168,
0.019262367859482765,
-0.10577601194381714,
-0.017951833084225655,
-0.06277871876955032,
-0.05074968561530113,
-2.1134530684250614e-33,
0.02463408373296261,
0.0013153974432498217,
-0.06809494644403458,
0.02070990949869156,
-0.00043273690971545875,
-0.005472665652632713,
0.09064266085624695,
-0.0010017849272117019,
0.06193574145436287,
-0.10901767760515213,
0.009370938874781132,
0.08969666808843613,
0.028880467638373375,
0.05760178342461586,
0.008788797073066235,
-0.04293903708457947,
-0.002937859855592251,
-0.026857763528823853,
0.03595166653394699,
-0.07831541448831558,
-0.032380491495132446,
0.00040731311310082674,
-0.0340825617313385,
0.022614972665905952,
-0.03147686645388603,
0.13439567387104034,
0.01640540547668934,
-0.006215604022145271,
-0.0356941744685173,
0.006942125968635082,
-0.007872575893998146,
0.06654710322618484,
0.04010765999555588,
-0.03752771392464638,
-0.007958324626088142,
0.017713114619255066,
0.10968340933322906,
-0.0010622821282595396,
-0.08551950752735138,
-0.037236880511045456,
0.011929196305572987,
0.03392232954502106,
0.07385530322790146,
0.06168322265148163,
-0.034671131521463394,
0.08781416714191437,
0.017219889909029007,
-0.029954006895422935,
-0.018959341570734978,
0.08515762537717819,
-0.04034135490655899,
0.024351315572857857,
-0.041151564568281174,
0.015984289348125458,
0.01261370163410902,
-0.1620301455259323,
-0.0729188621044159,
-0.013017090037465096,
0.11115768551826477,
-0.009544016793370247,
0.02229766547679901,
0.055503543466329575,
-0.05107155069708824,
-0.07584676146507263,
-0.09106579422950745,
0.07097406685352325,
0.03249504417181015,
0.04548293352127075,
0.08771885931491852,
0.0414593406021595,
-0.04012982174754143,
-0.007048873230814934,
0.014256501570343971,
-0.019061779603362083,
0.008151879534125328,
-0.008952097035944462,
0.0760439932346344,
-0.0010918788611888885,
0.02009144239127636,
0.0796930342912674,
0.03669392690062523,
-0.0030562984757125378,
-0.011491183191537857,
0.11348401755094528,
0.1098211333155632,
0.01745881512761116,
-0.06311050057411194,
-0.022595005109906197,
0.05154022201895714,
-0.007445518393069506,
-0.00972315575927496,
0.0660034716129303,
0.011631449684500694,
0.03780917450785637,
0.01761871576309204,
-4.10601010969458e-8,
-0.006519294809550047,
-0.02814354933798313,
0.04648704081773758,
-0.04157189652323723,
0.00007386007928289473,
-0.07560587674379349,
0.014286927878856659,
0.029740512371063232,
-0.02723696641623974,
-0.07025005668401718,
0.05254029855132103,
-0.12184129655361176,
0.012949402444064617,
-0.05706740915775299,
0.08219349384307861,
0.06385748088359833,
-0.04849652200937271,
-0.04283145070075989,
-0.01664935052394867,
0.030623411759734154,
-0.06771339476108551,
0.015505091287195683,
-0.028732461854815483,
-0.019050000235438347,
0.01654735393822193,
0.08123042434453964,
-0.029122672975063324,
0.04013042151927948,
0.13601692020893097,
-0.021392127498984337,
0.03353524208068848,
-0.0035793776623904705,
0.02438190206885338,
0.03949655964970589,
0.015402893535792828,
0.011224146001040936,
-0.02020907960832119,
-0.00021137385920155793,
0.03451838716864586,
-0.0031024101190268993,
-0.09689383208751678,
0.017269417643547058,
-0.013084194622933865,
-0.01453934796154499,
-0.05603417381644249,
0.010477501899003983,
0.01391998678445816,
-0.10502779483795166,
0.011529103852808475,
0.03421344980597496,
-0.03984200954437256,
0.022467894479632378,
-0.029686138033866882,
0.04909961298108101,
0.09154996275901794,
-0.034608371555805206,
0.056773003190755844,
-0.028170112520456314,
-0.0008976667304523289,
0.004200858995318413,
-0.03160293400287628,
-0.0005952058709226549,
-0.04792746156454086,
0.04284295067191124
] |
Credit-based self organizing maps: training deep topographic networks with minimal performance degradation | https://openreview.net/forum?id=wMgr7wBuUo | [
"Amirozhan Dehghani",
"Xinyu Qian",
"Asa Farahani",
"Pouya Bashivan"
] | Poster | In the primate neocortex, neurons with similar function are often found to be spatially close. Kohonen's self-organizing map (SOM) has been one of the most influential approaches for simulating brain-like topographical organization in artificial neural network models. However, integrating these maps into deep neural networks with multitude of layers has been challenging, with self-organized deep neural networks suffering from substantially diminished capacity to perform visual recognition. We identified a key factor leading to the performance degradation in self-organized topographical neural network models: the discord between predominantly bottom-up learning updates in the self-organizing maps, and those derived from top-down, credit-based learning approaches. To address this, we propose an alternative self organization algorithm, tailored to align with the top-down learning processes in deep neural networks. This model not only emulates critical aspects of cortical topography but also significantly narrows the performance gap between non-topographical and topographical models. This advancement underscores the substantial importance of top-down assigned credits in shaping topographical organization. Our findings are a step in reconciling topographical modeling with the functional efficacy of neural network models, paving the way for more brain-like neural architectures. | Computer vision, Neuroscience, Convolutional Networks, topographical organization, self-organizing maps, functional organization | We developed a new topographical neural network model that replicates the functional organization of the visual ventral stream while retaining high object recognition performance | 1,964 | null | [
-0.011943320743739605,
-0.09806670993566513,
0.06608813256025314,
0.009273173287510872,
0.007878056727349758,
0.034234609454870224,
-0.06094585731625557,
-0.01822265237569809,
0.039273716509342194,
0.001879398012533784,
-0.03295242413878441,
-0.033063359558582306,
-0.005909232888370752,
0.06771711260080338,
-0.06158865988254547,
-0.062464114278554916,
-0.03233970329165459,
0.08892294019460678,
-0.0419512614607811,
-0.025531703606247902,
0.04999833554029465,
0.010276852175593376,
0.06257456541061401,
0.054943785071372986,
-0.021729886531829834,
0.03903030976653099,
-0.006408029235899448,
0.025738243013620377,
-0.0012466234620660543,
-0.09005742520093918,
0.12122739851474762,
-0.005390087608247995,
0.05489424243569374,
0.033041492104530334,
-0.04428093880414963,
0.05728156864643097,
-0.029373258352279663,
0.029597459360957146,
-0.013166985474526882,
0.002752650761976838,
-0.017110675573349,
0.04321691766381264,
-0.015258904546499252,
0.01799415796995163,
0.09826697409152985,
0.07994244992733002,
0.03155035525560379,
-0.05244532600045204,
-0.012693352065980434,
-0.002637444995343685,
-0.01468774676322937,
0.010677173733711243,
-0.07272251695394516,
0.071485236287117,
-0.00570178497582674,
0.15876953303813934,
0.037102896720170975,
0.01773696020245552,
-0.05646407976746559,
0.01580102927982807,
0.06249522790312767,
-0.023788295686244965,
0.0487278588116169,
0.03697292506694794,
0.11839299649000168,
0.08071860671043396,
0.015446157194674015,
0.08004781603813171,
-0.047734975814819336,
0.0006982722552493215,
0.1291237771511078,
0.020200345665216446,
0.000324633700074628,
-0.008396357297897339,
0.0724090039730072,
0.04501635953783989,
-0.053353507071733475,
0.02355613373219967,
0.012775007635354996,
-0.037999071180820465,
0.0035146332811564207,
0.1473231315612793,
0.01710757240653038,
0.023228757083415985,
0.025083238258957863,
0.005470623727887869,
-0.038248106837272644,
-0.032068148255348206,
-0.06092192977666855,
-0.0403585247695446,
-0.011433983221650124,
-0.014246953651309013,
-0.04176231101155281,
-0.08824769407510757,
0.0294637493789196,
-0.019056331366300583,
-0.00680128438398242,
-0.08313378691673279,
-0.00854673981666565,
-0.009813475422561169,
-0.01798032782971859,
-0.05320431664586067,
0.015128492377698421,
0.0030979083385318518,
0.09728501737117767,
-0.03145753592252731,
0.09971331804990768,
0.021535279229283333,
0.07345891743898392,
-0.04679149016737938,
0.01868576928973198,
-0.01356495637446642,
-0.0032466126140207052,
0.04313972592353821,
-0.007619098760187626,
0.023158498108386993,
-0.006562723312526941,
0.0424032136797905,
0.07061067968606949,
0.03391137719154358,
-0.02669477090239525,
-0.013404753059148788,
-0.0033131963573396206,
-0.04449835792183876,
0.0017160032875835896,
0.04923282563686371,
-0.10780733078718185,
1.687927419505255e-33,
0.01501193642616272,
0.01064868364483118,
0.04549664631485939,
-0.015744376927614212,
0.04041927680373192,
-0.03624112159013748,
-0.03257578983902931,
0.033612191677093506,
0.0023934675846248865,
0.06071673706173897,
-0.10316968709230423,
0.08088840544223785,
-0.03019120916724205,
0.1340484619140625,
0.053104933351278305,
-0.00022642243129666895,
-0.04071471840143204,
-0.0311414897441864,
-0.05571986734867096,
-0.10322804003953934,
0.02635575458407402,
0.030837586149573326,
0.018419798463582993,
-0.0790325254201889,
0.006710823625326157,
0.03543534129858017,
-0.007293289992958307,
-0.009241029620170593,
-0.08962582051753998,
-0.0035159848630428314,
-0.03752945736050606,
0.055600568652153015,
-0.04497644677758217,
-0.038114190101623535,
0.03640275076031685,
0.017447175458073616,
0.005837992299348116,
0.000971382949501276,
0.06726057827472687,
-0.009237982332706451,
0.033948637545108795,
0.0273674838244915,
-0.04579002037644386,
-0.028570502996444702,
-0.06018014997243881,
-0.06739552319049835,
0.036152925342321396,
-0.0016822089673951268,
-0.01884356141090393,
-0.04970485344529152,
0.02828057110309601,
-0.01184443011879921,
0.010504759848117828,
-0.043741852045059204,
-0.055777814239263535,
0.02344525046646595,
-0.010114897973835468,
0.043620865792036057,
0.05399861931800842,
0.026498446241021156,
0.03585192933678627,
-0.07316137105226517,
-0.029444899410009384,
0.10803068429231644,
0.04565303027629852,
-0.05268176272511482,
0.029288483783602715,
0.026317700743675232,
0.03490300849080086,
-0.03181280940771103,
-0.02422657422721386,
0.0774272158741951,
-0.08724100887775421,
-0.058481622487306595,
-0.000900621060281992,
-0.04908428713679314,
-0.015306217595934868,
-0.0396655909717083,
-0.0791739970445633,
-0.014371708035469055,
-0.006162418983876705,
0.030274735763669014,
-0.06153256446123123,
-0.0312887467443943,
0.03730739280581474,
0.02896016836166382,
0.12762011587619781,
-0.021205760538578033,
-0.019253317266702652,
0.06764515489339828,
-0.04470431059598923,
-0.04202539473772049,
0.08580446243286133,
-0.01718886010348797,
-0.03078761324286461,
-1.9230174734620813e-33,
-0.05945169925689697,
0.05140775814652443,
-0.03651261702179909,
0.012860293500125408,
0.013736126013100147,
-0.011274253018200397,
0.04872246831655502,
0.020722173154354095,
-0.05677985027432442,
0.000484804913867265,
-0.025021212175488472,
0.027467699721455574,
0.06677386164665222,
-0.01813177578151226,
0.06681474298238754,
-0.023342249915003777,
-0.06869897246360779,
0.05311109125614166,
0.037386249750852585,
-0.10218285769224167,
-0.018898004665970802,
0.05041774362325668,
-0.03831498697400093,
-0.019566869363188744,
-0.024447746574878693,
0.034997131675481796,
-0.11489633470773697,
0.06687181442975998,
0.02301935851573944,
0.08622211962938309,
-0.081194207072258,
-0.001159902079962194,
-0.11411415785551071,
-0.01719563826918602,
-0.005121887195855379,
0.09682242572307587,
-0.03852052986621857,
-0.03888575732707977,
-0.06679333746433258,
0.010304807685315609,
-0.005958082154393196,
-0.005866229999810457,
-0.04238273575901985,
-0.04924723878502846,
0.06759235262870789,
-0.0975855365395546,
-0.03985074535012245,
0.012492779642343521,
-0.05843557417392731,
0.14522159099578857,
-0.06638920307159424,
-0.02502780593931675,
-0.09453460574150085,
-0.01941036805510521,
-0.04956766217947006,
0.026738891378045082,
0.026219278573989868,
0.000038638783735223114,
0.009926524944603443,
-0.024108819663524628,
-0.08951491862535477,
-0.07551742345094681,
-0.028396090492606163,
-0.01940133608877659,
-0.0482940636575222,
-0.026524338871240616,
0.018567431718111038,
0.037821605801582336,
0.0479513555765152,
0.027304336428642273,
0.019983159378170967,
-0.0147284846752882,
0.0521976538002491,
-0.0003774119832087308,
-0.04895899072289467,
-0.013089523650705814,
-0.01843518763780594,
0.04187087342143059,
-0.05083319544792175,
-0.08022084832191467,
-0.00828215666115284,
-0.05217979475855827,
-0.058416206389665604,
0.09111712127923965,
0.10335279256105423,
0.08907485753297806,
0.029960891231894493,
0.03495455160737038,
0.0007432700367644429,
0.032181836664676666,
0.06048762798309326,
-0.0805465430021286,
-0.06429333239793777,
0.048583123832941055,
-0.09038404375314713,
-3.630457712233692e-8,
-0.13052615523338318,
0.009514404460787773,
0.06115654110908508,
0.04463852196931839,
0.09277216345071793,
-0.03681371361017227,
0.08853071928024292,
0.07577206939458847,
-0.02734883315861225,
-0.041219551116228104,
0.07367675006389618,
-0.03414670005440712,
-0.06493919342756271,
-0.028927970677614212,
0.004893223289400339,
0.05545320361852646,
0.0208112969994545,
0.08482549339532852,
-0.027547013014554977,
0.0013756508706137538,
-0.025533104315400124,
-0.03392414748668671,
0.011210998520255089,
-0.020245404914021492,
0.021300897002220154,
-0.05332821235060692,
-0.011758923530578613,
0.05911938473582268,
-0.0439336821436882,
-0.050699878484010696,
0.06336978077888489,
0.045570503920316696,
0.036059148609638214,
0.02924683503806591,
-0.01056126318871975,
0.04640315845608711,
0.0009718031506054103,
0.11367280781269073,
-0.027282340452075005,
-0.0459868460893631,
-0.06348866969347,
0.07223675400018692,
0.04394179582595825,
0.013397438451647758,
0.01057558786123991,
0.013770041055977345,
0.04725969210267067,
0.03517642617225647,
0.059999193996191025,
-0.06878520548343658,
-0.041595980525016785,
0.011774465441703796,
-0.026553692296147346,
0.06192939728498459,
-0.006603751797229052,
-0.017748761922121048,
0.02488136477768421,
-0.02815825492143631,
-0.012296488508582115,
0.016526026651263237,
-0.004502404946833849,
-0.023558534681797028,
-0.09906630963087082,
-0.09004440903663635
] |
An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels | https://openreview.net/forum?id=tjNf0L8QjR | [
"Duy Kien Nguyen",
"Mido Assran",
"Unnat Jain",
"Martin R. Oswald",
"Cees G. M. Snoek",
"Xinlei Chen"
] | Poster | This work does not introduce a new method. Instead, we present an interesting finding that questions the necessity of the inductive bias of locality in modern computer vision architectures. Concretely, we find that vanilla Transformers can operate by directly treating each individual pixel as a token and achieve highly performant results. This is substantially different from the popular design in Vision Transformer, which maintains the inductive bias from ConvNets towards local neighborhoods (e.g., by treating each 16x16 patch as a token). We showcase the effectiveness of pixels-as-tokens across three well-studied computer vision tasks: supervised learning for classification and regression, self-supervised learning via masked autoencoding, and image generation with diffusion models. Although it's computationally less practical to directly operate on individual pixels, we believe the community must be made aware of this surprising piece of knowledge when devising the next generation of neural network architectures for computer vision. | locality, convolutional networks, transformers | we show locality is not a fundamental inductive bias for vision architectures | 1,958 | 2406.09415 | [
-0.049972862005233765,
-0.03371317684650421,
0.0695287361741066,
-0.03828001767396927,
0.08298707753419876,
-0.038443561643362045,
0.012277609668672085,
-0.05686967074871063,
-0.0073419907130301,
-0.07557663321495056,
-0.04070433974266052,
-0.07953903824090958,
-0.011703967116773129,
0.07624627649784088,
-0.07852055132389069,
0.0327916294336319,
0.04793054983019829,
0.07947909086942673,
-0.10760650038719177,
-0.06512473523616791,
-0.03677477687597275,
-0.06695570796728134,
-0.013213660567998886,
0.02263575792312622,
0.09331048280000687,
0.07607696205377579,
0.03733592480421066,
-0.08510638028383255,
0.027498304843902588,
-0.046182360500097275,
0.00835163053125143,
-0.020963311195373535,
-0.02317134290933609,
0.03965293616056442,
-0.021005606278777122,
0.027207424864172935,
-0.0017248092917725444,
0.059348829090595245,
0.03332693129777908,
-0.07275551557540894,
0.035922907292842865,
0.02498711459338665,
0.03749783709645271,
-0.009227435104548931,
0.11323750764131546,
0.07784657180309296,
0.05398501083254814,
-0.04703056812286377,
-0.008107513189315796,
-0.04880446195602417,
0.000390782137401402,
-0.008155390620231628,
-0.054347943514585495,
0.07928788661956787,
-0.06405364722013474,
0.016920946538448334,
0.06929782032966614,
0.0008778056944720447,
0.05317681282758713,
0.05508130416274071,
0.030170835554599762,
-0.07104220241308212,
0.025389021262526512,
0.06905330717563629,
-0.0837947279214859,
-0.05047475919127464,
0.0404401458799839,
-0.032411497086286545,
-0.033079713582992554,
-0.012874370440840721,
0.025901639834046364,
0.08703579753637314,
-0.024918628856539726,
0.01585010066628456,
0.07006455957889557,
0.04091214761137962,
0.10854543745517731,
-0.05155837535858154,
0.09933847188949585,
-0.07362381368875504,
0.010538504458963871,
-0.0010382921900600195,
0.05851277336478233,
0.013596282340586185,
0.10659919679164886,
-0.003042263211682439,
-0.030784234404563904,
0.04289630800485611,
0.010135866701602936,
0.035778991878032684,
-0.07902386784553528,
-0.05071337893605232,
-0.05128813162446022,
-0.058514442294836044,
0.05552441626787186,
-0.06354845315217972,
0.03585544973611832,
-0.09713903814554214,
-0.03317752480506897,
0.08580280095338821,
-0.050918012857437134,
-0.00751907704398036,
0.015933841466903687,
-0.02127368561923504,
0.025045203045010567,
-0.03256038576364517,
0.03302735835313797,
0.027901194989681244,
0.09016431123018265,
-0.013521384447813034,
0.03976142778992653,
0.014155580662190914,
-0.040679339319467545,
-0.044377245008945465,
0.04752441123127937,
-0.04687662422657013,
0.011487521231174469,
0.053627341985702515,
0.02961333468556404,
0.015308822505176067,
-0.033233534544706345,
-0.011716723442077637,
-0.0802900418639183,
-0.004358876030892134,
0.0485122948884964,
0.01122062373906374,
-0.08433635532855988,
6.413458721972384e-35,
-0.02181963250041008,
0.01276689488440752,
0.031349048018455505,
0.005762709304690361,
0.07853931933641434,
-0.020170094445347786,
0.02654031105339527,
0.060182612389326096,
0.020152654498815536,
-0.023932386189699173,
-0.011621139012277126,
0.0054313382133841515,
-0.05001223459839821,
0.13396050035953522,
0.013426138088107109,
-0.02217540144920349,
-0.0028629349544644356,
-0.02468634769320488,
-0.031989093869924545,
-0.027577225118875504,
0.009263943880796432,
0.021095035597682,
-0.005915809888392687,
-0.04161830246448517,
-0.06405799835920334,
-0.07755114883184433,
0.03680143505334854,
0.01717047207057476,
0.04190724715590477,
0.0020568235777318478,
-0.07130611687898636,
0.05012688413262367,
0.002782197669148445,
0.036017823964357376,
0.0299032311886549,
0.005859090015292168,
0.0051063965074718,
-0.01452132873237133,
0.0673479288816452,
0.016162646934390068,
-0.007065154146403074,
0.07523863762617111,
-0.025532761588692665,
-0.07783427834510803,
-0.039214856922626495,
0.012153544463217258,
0.011161082424223423,
0.020379995927214622,
-0.06199467554688454,
-0.05066355690360069,
0.04820817708969116,
0.009738415479660034,
-0.08645089715719223,
-0.09791642427444458,
0.03415687754750252,
-0.011521954089403152,
0.02798892930150032,
-0.021278010681271553,
0.0799030140042305,
-0.022618664428591728,
0.02740880660712719,
0.02989942952990532,
-0.06197650358080864,
0.06481882184743881,
-0.007665010169148445,
-0.052913323044776917,
0.09486553817987442,
0.060707852244377136,
-0.06603746116161346,
0.021888263523578644,
-0.050817977637052536,
0.08639346063137054,
-0.08354415744543076,
-0.09569279849529266,
0.03988682106137276,
0.02122822217643261,
0.03693453222513199,
0.010386784560978413,
-0.007444359362125397,
0.05578240379691124,
-0.0066041527315974236,
-0.01921609230339527,
-0.021062133833765984,
-0.04140946641564369,
0.0026641744188964367,
0.011139404028654099,
0.02062174864113331,
-0.07265812158584595,
-0.016011152416467667,
-0.025636468082666397,
-0.007811684161424637,
-0.06908053159713745,
0.044347576797008514,
0.002837234875187278,
-0.02920667827129364,
-1.518385042766691e-33,
-0.010581056587398052,
0.05783271789550781,
-0.019927658140659332,
0.0585562139749527,
-0.06432632356882095,
-0.040683481842279434,
0.03248213604092598,
0.026073090732097626,
0.020648986101150513,
-0.03629278764128685,
-0.024335822090506554,
0.0325436107814312,
-0.03397451341152191,
0.006976820528507233,
0.00464634457603097,
-0.040558140724897385,
-0.00941421091556549,
0.004098708741366863,
-0.09966608881950378,
-0.022034578025341034,
-0.025046437978744507,
0.12437509745359421,
-0.0635138675570488,
0.05598452314734459,
-0.12283284962177277,
0.052698150277137756,
-0.05481617525219917,
0.04450416937470436,
-0.013926433399319649,
-0.015922848135232925,
-0.009340655989944935,
-0.00647641159594059,
0.013199638575315475,
-0.041816383600234985,
-0.05788346379995346,
0.04153572767972946,
0.0006608095136471093,
-0.029652679339051247,
-0.011759285815060139,
0.030393091961741447,
0.028751086443662643,
-0.07227253913879395,
-0.060530927032232285,
0.0764908492565155,
-0.04317834973335266,
-0.0880621150135994,
0.009883257560431957,
0.012517216615378857,
-0.03587871044874191,
0.10494141280651093,
-0.03568436577916145,
0.049025408923625946,
-0.14083246886730194,
0.016325632110238075,
-0.03226014971733093,
-0.02491685189306736,
-0.06243184208869934,
0.028080014511942863,
0.15091419219970703,
0.004616086371243,
-0.03112148307263851,
-0.07967167347669601,
-0.01585860177874565,
-0.04375482723116875,
-0.04595579579472542,
-0.03184351325035095,
-0.1291630119085312,
0.018980760127305984,
0.02017061784863472,
0.10737257450819016,
0.003692892612889409,
-0.05008631572127342,
-0.05538145452737808,
-0.04933830723166466,
-0.06612411141395569,
-0.055056553333997726,
0.04433388262987137,
-0.04668581485748291,
0.013611282221972942,
0.005541389342397451,
-0.015714162960648537,
-0.05884036421775818,
0.07176914811134338,
0.0325012281537056,
0.09029076248407364,
0.07353803515434265,
0.0003591578279156238,
-0.021397212520241737,
0.06633986532688141,
0.06510362029075623,
-0.04915127903223038,
0.055283334106206894,
-0.010179250501096249,
0.006040807813405991,
-0.03371547907590866,
-4.1036923192905306e-8,
-0.04527537152171135,
0.07095853239297867,
0.0007201189873740077,
-0.014930485747754574,
-0.017134614288806915,
-0.09877528995275497,
0.037528857588768005,
0.10994932055473328,
-0.03347552940249443,
0.017721500247716904,
0.1159474328160286,
-0.0345442034304142,
-0.0393555611371994,
-0.04665384069085121,
0.032142382115125656,
0.006992190144956112,
0.023147717118263245,
0.024262046441435814,
0.008552523329854012,
-0.0032666646875441074,
-0.07074462622404099,
0.05017330124974251,
0.0074554793536663055,
-0.04454205185174942,
0.01258146297186613,
-0.07885274291038513,
-0.003703893395140767,
0.0556478314101696,
0.015654491260647774,
0.009277110919356346,
0.0033789959270507097,
0.05869263410568237,
0.026103923097252846,
-0.048615772277116776,
0.029080264270305634,
0.06651414185762405,
0.04822870343923569,
0.030482139438390732,
0.0044347308576107025,
-0.057119276374578476,
-0.02023444138467312,
0.002442818833515048,
-0.03055192157626152,
0.025223560631275177,
-0.07871592789888382,
0.04407326132059097,
0.0829753503203392,
-0.07580163329839706,
-0.06774626672267914,
0.009610789828002453,
0.05413638427853584,
-0.012956099584698677,
0.02463510073721409,
0.12623761594295502,
0.09596483409404755,
-0.027304742485284805,
-0.0007691577193327248,
-0.0701095461845398,
0.041775986552238464,
0.12872157990932465,
-0.03944685682654381,
0.0007484466186724603,
-0.0015448133926838636,
-0.01864342764019966
] |
Model-Free Offline Reinforcement Learning with Enhanced Robustness | https://openreview.net/forum?id=QyVLJ7EnAC | [
"Chi Zhang",
"Zain Ulabedeen Farhat",
"George K. Atia",
"Yue Wang"
] | Poster | Offline reinforcement learning (RL) has gained considerable attention for its ability to learn policies from pre-collected data without real-time interaction, which makes it particularly useful for high-risk applications. However, due to its reliance on offline datasets, existing works inevitably introduce assumptions to ensure effective learning, which, however, often lead to a trade-off between robustness to model mismatch and scalability to large environments. In this paper, we enhance both aspects with a novel double-pessimism principle, which conservatively estimates performance and accounts for both limited data and potential model mismatches, two major reasons for the previous trade-off. We then propose a universal, model-free algorithm to learn a policy that is robust to potential environment mismatches, which enhances robustness in a scalable manner. Furthermore, we provide a sample complexity analysis of our algorithm when the mismatch is modeled by the $l_\alpha$-norm, which also theoretically demonstrates the efficiency of our method. Extensive experiments further demonstrate that our approach significantly improves robustness in a more scalable manner than existing methods. | offline RL, robust, scalability, model-free | null | 1,945 | null | [
-0.07425998151302338,
-0.07341606169939041,
0.016741786152124405,
0.06730297207832336,
0.07144485414028168,
0.022307518869638443,
-0.072262704372406,
0.005280863028019667,
-0.056372638791799545,
0.03144286945462227,
-0.016786985099315643,
-0.002971900859847665,
0.02873963490128517,
-0.012395367957651615,
-0.02219531685113907,
0.04009092226624489,
0.054340653121471405,
0.0360003225505352,
-0.11433826386928558,
0.009782448410987854,
-0.022151513025164604,
-0.007034219801425934,
-0.011973117478191853,
0.02093954011797905,
0.06719335913658142,
-0.003485599299892783,
-0.011264515109360218,
0.02149704471230507,
-0.01637471280992031,
-0.002630992792546749,
0.035567257553339005,
-0.06360003352165222,
0.00334364315494895,
-0.10791002959012985,
-0.06275324523448944,
0.05618086829781532,
-0.06784219294786453,
-0.022947343066334724,
-0.04506874457001686,
0.011708008125424385,
-0.04294116422533989,
0.05128084495663643,
0.004046041052788496,
-0.004694828763604164,
0.045021556317806244,
-0.02336898259818554,
0.0006068159127607942,
-0.07830668985843658,
0.014469435438513756,
0.01100938767194748,
-0.018009107559919357,
-0.02466399222612381,
0.01767132803797722,
0.046978406608104706,
-0.020494850352406502,
0.03844160959124565,
-0.008838283829391003,
0.04472244903445244,
-0.023487843573093414,
-0.026538273319602013,
-0.033085450530052185,
-0.09958946704864502,
-0.0028503835201263428,
-0.015139271505177021,
-0.07458280026912689,
-0.02271869219839573,
-0.03769446909427643,
0.07356450706720352,
0.029471617192029953,
0.0590885765850544,
-0.0795043334364891,
0.060469020158052444,
-0.05619315803050995,
-0.012556037865579128,
0.004823463503271341,
-0.030885176733136177,
0.04745390638709068,
-0.0005754027515649796,
0.018147913739085197,
0.028400445356965065,
0.019827082753181458,
-0.040511492639780045,
-0.09719539433717728,
0.008087719790637493,
0.10346736013889313,
-0.031247101724147797,
-0.008317780680954456,
0.03502897173166275,
0.2066194862127304,
-0.018517976626753807,
0.021996991708874702,
0.07767084985971451,
0.047005265951156616,
0.0033652177080512047,
0.009918615221977234,
0.05038321390748024,
0.0013838360318914056,
-0.0781848356127739,
-0.06980740278959274,
0.13258245587348938,
-0.0001087239506887272,
0.11371541768312454,
-0.01108049787580967,
-0.0046081505715847015,
0.013222225941717625,
-0.05845257639884949,
0.030211815610527992,
0.012838145717978477,
0.04264434054493904,
-0.07394415885210037,
0.08437439799308777,
0.019312435761094093,
0.05488595739006996,
0.03388146311044693,
-0.03671251982450485,
-0.014951801858842373,
0.024767836555838585,
-0.07703018933534622,
-0.033464569598436356,
0.04993053153157234,
0.016130484640598297,
-0.02222112938761711,
0.08441342413425446,
0.03623547777533531,
0.001944714575074613,
-0.03015720471739769,
-0.032308243215084076,
2.4485985842122403e-34,
0.08938011527061462,
0.023167869076132774,
0.047488387674093246,
-0.0025315277744084597,
0.07480762153863907,
-0.02706068754196167,
0.015511364676058292,
-0.02078777179121971,
-0.04578106477856636,
-0.005852863658219576,
0.038087908178567886,
0.01886163093149662,
-0.031808219850063324,
0.0341988280415535,
0.0659644603729248,
0.04320872575044632,
-0.020695574581623077,
0.04080493748188019,
0.06557703763246536,
-0.03552217409014702,
0.038876477628946304,
-0.1237025037407875,
-0.003445314709097147,
-0.061396095901727676,
0.05641857907176018,
0.02032792754471302,
0.04514232650399208,
-0.006635969504714012,
0.03454787656664848,
-0.012921682558953762,
0.011276580393314362,
0.03802623599767685,
-0.028079986572265625,
0.06405657529830933,
0.030436556786298752,
0.009081797674298286,
-0.1280936598777771,
-0.04389791935682297,
0.00836118869483471,
0.03090560808777809,
-0.017136411741375923,
0.01076095923781395,
0.07653164118528366,
0.02410479635000229,
-0.006365688517689705,
-0.12970663607120514,
0.07358523458242416,
-0.03003508411347866,
-0.09829974919557571,
-0.06763575971126556,
0.00539590697735548,
0.005987888667732477,
-0.012850643135607243,
-0.03545495122671127,
-0.08503258228302002,
0.009753412567079067,
-0.005095290951430798,
0.08502919971942902,
-0.02659708447754383,
0.08636318147182465,
-0.019575636833906174,
0.025753287598490715,
0.07402392476797104,
0.03291482478380203,
-0.01581508107483387,
0.05354361608624458,
-0.04714425653219223,
0.06685788184404373,
-0.002042141044512391,
-0.0414680615067482,
0.0318269282579422,
0.03394504636526108,
-0.0031209259759634733,
-0.08050619065761566,
0.09106569737195969,
-0.01733625866472721,
0.06647398322820663,
-0.07113251090049744,
0.040046751499176025,
-0.017557378858327866,
0.015832973644137383,
-0.029199562966823578,
0.04593190550804138,
-0.015600166283547878,
-0.022230802103877068,
-0.10182922333478928,
-0.008125773631036282,
-0.023744165897369385,
-0.0514415018260479,
-0.016502434387803078,
-0.11061158776283264,
-0.008773166686296463,
-0.03286299481987953,
-0.009823376312851906,
0.02444840781390667,
-8.159524140565016e-34,
-0.03656012937426567,
-0.11562347412109375,
0.037485428154468536,
0.048999711871147156,
-0.038487277925014496,
-0.014708936214447021,
-0.009307347238063812,
-0.015352223999798298,
0.015147005207836628,
-0.010439003817737103,
-0.0734180435538292,
0.08139052987098694,
0.03912748768925667,
-0.056979257613420486,
0.07347820699214935,
-0.06421873718500137,
-0.015307671390473843,
0.0030553345568478107,
0.032152898609638214,
-0.005900056101381779,
0.07918719202280045,
-0.0345270037651062,
-0.08420262485742569,
0.12055756896734238,
-0.0023527336306869984,
0.006407878361642361,
-0.049940358847379684,
0.03020944632589817,
-0.010250533930957317,
0.031208788976073265,
-0.08871444314718246,
-0.026755308732390404,
-0.01519753597676754,
-0.038101498037576675,
-0.05225866660475731,
0.10357847064733505,
0.02447241172194481,
0.04726618528366089,
-0.07021304965019226,
0.08442040532827377,
-0.0221564918756485,
-0.0036761045921593904,
-0.1369047611951828,
-0.045265085995197296,
0.008671889081597328,
0.009949230588972569,
-0.031961433589458466,
0.014474393799901009,
0.019819555804133415,
-0.001011078362353146,
-0.03887605294585228,
0.03371376171708107,
-0.032659195363521576,
0.05239289999008179,
-0.01326753105968237,
-0.03323580324649811,
-0.07331808656454086,
0.012834989465773106,
-0.00949904602020979,
0.02326425537467003,
-0.021544599905610085,
-0.014111274853348732,
-0.13471786677837372,
0.09883417934179306,
-0.034747298806905746,
0.018358558416366577,
-0.04276648536324501,
0.024186071008443832,
0.03943381831049919,
0.024894412606954575,
0.028910387307405472,
0.013877485878765583,
0.0022033401764929295,
-0.05390341207385063,
0.014546085149049759,
0.023723697289824486,
0.05896143242716789,
-0.010498708114027977,
0.002030914416536689,
0.0651310607790947,
0.06388084590435028,
0.0249236598610878,
0.11133576184511185,
0.025588057935237885,
0.08284155279397964,
0.06331049650907516,
0.010473093017935753,
0.1135832741856575,
-0.004367133602499962,
-0.03364906832575798,
-0.02564062736928463,
0.027301596477627754,
0.0036427867598831654,
0.042262546718120575,
-0.03680254518985748,
-4.807126430250719e-8,
-0.0288148932158947,
0.029939517378807068,
0.11138688772916794,
-0.03321443498134613,
0.08092550188302994,
-0.013947260566055775,
-0.018281036987900734,
-0.017883285880088806,
-0.027566898614168167,
-0.041944462805986404,
-0.03671349957585335,
0.010297046042978764,
0.002575403079390526,
-0.09380270540714264,
-0.020984910428524017,
0.08556848019361496,
-0.0032946576829999685,
-0.033339910209178925,
-0.02430431731045246,
0.07034319639205933,
0.04609451815485954,
0.014009455218911171,
-0.03420097380876541,
-0.07457999885082245,
0.06869915872812271,
-0.045399367809295654,
-0.022091813385486603,
0.10798502713441849,
0.035213273018598557,
0.012419560924172401,
-0.008960063569247723,
-0.07736251503229141,
0.06428896635770798,
-0.01799602620303631,
0.050192221999168396,
0.06155341863632202,
0.07810302823781967,
-0.08101674914360046,
-0.019257890060544014,
-0.08622414618730545,
-0.028433889150619507,
0.058846719563007355,
0.009557942859828472,
-0.033983469009399414,
0.027768190950155258,
0.011652638204395771,
-0.0420847125351429,
-0.11114011704921722,
0.07096932083368301,
-0.06394065171480179,
0.05126864090561867,
0.013166319578886032,
-0.015423478558659554,
0.05368916317820549,
0.08917298167943954,
-0.03465142101049423,
-0.007312511559575796,
-0.044732093811035156,
0.06077433377504349,
0.08041369169950485,
0.005133859813213348,
-0.03568139672279358,
-0.08035451918840408,
0.03275708481669426
] |
Draw-and-Understand: Leveraging Visual Prompts to Enable MLLMs to Comprehend What You Want | https://openreview.net/forum?id=bfa58H1nQ8 | [
"Weifeng Lin",
"Xinyu Wei",
"Ruichuan An",
"Peng Gao",
"Bocheng Zou",
"Yulin Luo",
"Siyuan Huang",
"Shanghang Zhang",
"Hongsheng Li"
] | Poster | In this paper, we present the Draw-and-Understand framework, exploring how to integrate visual prompting understanding capabilities into Multimodal Large Language Models (MLLMs). Visual prompts allow users to interact through multi-modal instructions, enhancing the models' interactivity and fine-grained image comprehension. In this framework, we propose a general architecture adaptable to different pre-trained MLLMs, enabling it to recognize various types of visual prompts (such as points, bounding boxes, and free-form shapes) alongside language understanding. Additionally, we introduce MDVP-Instruct-Data, a multi-domain dataset featuring 1.2 million image-visual prompt-text triplets, including natural images, document images, scene text images, mobile/web screenshots, and remote sensing images. Building on this dataset, we introduce MDVP-Bench, a challenging benchmark designed to evaluate a model's ability to understand visual prompting instructions. The experimental results demonstrate that our framework can be easily and effectively applied to various MLLMs, such as SPHINX-X and LLaVA. After training with MDVP-Instruct-Data and image-level instruction datasets, our models exhibit impressive multimodal interaction capabilities and pixel-level understanding, while maintaining their image-level visual perception performance. | Multimodal Large Language Model, Visual Prompting | null | 1,940 | null | [
0.005015389062464237,
-0.1068943589925766,
0.021276401355862617,
-0.05601609870791435,
0.11838064342737198,
-0.016048256307840347,
0.0023402837105095387,
0.008550891652703285,
0.0054961564019322395,
-0.031220516189932823,
-0.014298100024461746,
-0.07175033539533615,
0.05490261688828468,
0.06227441877126694,
0.07093454152345657,
-0.018865665420889854,
0.0368233397603035,
0.01405246090143919,
-0.10119281709194183,
0.006532544270157814,
0.029732221737504005,
-0.01325023453682661,
0.08212843537330627,
-0.023331722244620323,
0.03785713016986847,
0.0348370298743248,
-0.007910309359431267,
-0.062221769243478775,
0.04896727204322815,
-0.0237288735806942,
0.0151197649538517,
-0.0112279262393713,
0.02993720769882202,
0.10247621685266495,
-0.0406528003513813,
0.07954738289117813,
-0.013620187528431416,
0.029546141624450684,
-0.05381704866886139,
-0.09757784754037857,
-0.09211785346269608,
-0.04655981808900833,
0.03911741077899933,
-0.009275919757783413,
0.06971286237239838,
-0.028211573138833046,
-0.030172210186719894,
-0.06852003186941147,
0.06949658691883087,
0.02321406826376915,
-0.0846107229590416,
-0.08177672326564789,
0.014140019193291664,
0.02641202323138714,
-0.05427102744579315,
0.026152128353714943,
0.06523547321557999,
-0.024265363812446594,
0.020947786048054695,
-0.00858292542397976,
-0.05969369411468506,
-0.03573200851678848,
-0.010222983546555042,
0.04041111841797829,
-0.002878944855183363,
-0.024581143632531166,
0.011367052793502808,
-0.033802397549152374,
0.03633356839418411,
-0.062351785600185394,
-0.0653015747666359,
0.022580578923225403,
0.024723699316382408,
-0.005164098925888538,
0.0026071243919432163,
0.041850388050079346,
-0.012354047037661076,
-0.04072466492652893,
0.004759237170219421,
-0.11017495393753052,
0.005242220591753721,
0.03509135544300079,
0.025547588244080544,
0.03538420796394348,
0.09403540939092636,
-0.022914083674550056,
-0.023812923580408096,
0.07359784841537476,
0.0014285342767834663,
0.028585977852344513,
-0.05636883154511452,
-0.058514710515737534,
0.006861652713268995,
0.035823721438646317,
0.02014201506972313,
0.07280579954385757,
0.03318727761507034,
-0.15054912865161896,
-0.08333174139261246,
0.06508001685142517,
0.005925196688622236,
0.014659498818218708,
0.01103734690696001,
-0.056074097752571106,
0.017201118171215057,
-0.02556622214615345,
0.04595654085278511,
0.006787918508052826,
0.014115258120000362,
-0.07103072106838226,
0.05712324380874634,
0.0029845258686691523,
-0.004430720582604408,
-0.01256579253822565,
0.03407139331102371,
-0.0322476364672184,
0.01483284030109644,
-0.05118381604552269,
0.06993397325277328,
-0.03875507786870003,
-0.0034956473391503096,
0.03329360857605934,
-0.03935547173023224,
-0.01760590635240078,
0.07254400104284286,
-0.08103740960359573,
-0.04999201372265816,
-1.8149032172551246e-33,
0.0926155224442482,
0.028601914644241333,
0.06434197723865509,
0.07860029488801956,
0.0777958482503891,
-0.002411490073427558,
-0.034905027598142624,
0.026789965108036995,
-0.05859121307730675,
-0.0009716779459267855,
0.0016923228977248073,
-0.026172032579779625,
-0.03210724517703056,
0.13300539553165436,
0.0013387298677116632,
0.00012958569277543575,
-0.04204073175787926,
0.04310092702507973,
-0.01999342441558838,
0.03968976438045502,
-0.00986627209931612,
-0.03102077916264534,
0.021644901484251022,
0.05860969424247742,
0.07523642480373383,
0.04853697493672371,
0.06448384374380112,
-0.029325146228075027,
-0.02327299490571022,
0.017337340861558914,
-0.11835028976202011,
-0.03732208535075188,
-0.016514478251338005,
0.012117798440158367,
-0.009251742623746395,
-0.07280005514621735,
-0.021623484790325165,
-0.058802470564842224,
0.028526270762085915,
0.006047682836651802,
-0.007862620986998081,
-0.05005370080471039,
-0.007900637574493885,
-0.010926431976258755,
-0.10863588750362396,
0.0041120704263448715,
0.024061305448412895,
0.07540695369243622,
-0.10108769685029984,
0.010884649120271206,
0.023390650749206543,
0.014006508514285088,
-0.042335961014032364,
-0.04612313210964203,
-0.01643340475857258,
-0.05837168172001839,
0.02521071583032608,
0.02630615420639515,
0.04326784610748291,
-0.02883511409163475,
0.02703751251101494,
0.08052258938550949,
-0.011521593667566776,
0.09579634666442871,
-0.004188536666333675,
-0.031096167862415314,
0.009538344107568264,
0.06824231147766113,
0.049384355545043945,
0.014649858698248863,
-0.09049350768327713,
0.012883102521300316,
0.03629089519381523,
-0.06459914147853851,
0.002101079560816288,
-0.010081080719828606,
0.016464097425341606,
-0.07042942196130753,
0.11605435609817505,
0.07608301937580109,
0.010134278796613216,
-0.012726426124572754,
0.02660471946001053,
-0.06166267767548561,
-0.010410387068986893,
-0.03075961396098137,
0.06980203092098236,
-0.15003861486911774,
-0.03186710178852081,
-0.07282072305679321,
-0.03157182037830353,
0.0038314175326377153,
-0.002663946710526943,
-0.021566025912761688,
-0.011218064464628696,
1.5166847085553258e-33,
0.0763278529047966,
0.019863462075591087,
-0.058081477880477905,
0.03362570330500603,
0.011721107177436352,
-0.05172989144921303,
0.04067837819457054,
0.04656076058745384,
-0.016852818429470062,
-0.08275096863508224,
-0.010958625003695488,
0.02311922423541546,
-0.035694222897291183,
-0.017724137753248215,
0.024253014475107193,
0.012494263239204884,
0.014320334419608116,
-0.019815871492028236,
-0.005116220563650131,
0.013308202847838402,
-0.024854639545083046,
0.06499705463647842,
-0.08059938997030258,
-0.0041018640622496605,
-0.11048077791929245,
0.029243487864732742,
-0.07338175922632217,
0.021510034799575806,
-0.049070265144109726,
0.01344676036387682,
-0.01053922064602375,
0.008712604641914368,
0.02635204792022705,
-0.011243315413594246,
-0.034962352365255356,
0.028815699741244316,
0.08422569185495377,
-0.08471102267503738,
-0.034514766186475754,
0.13085149228572845,
0.13274680078029633,
-0.061434149742126465,
-0.05274738371372223,
-0.0067744688130915165,
-0.07111942023038864,
-0.010098513215780258,
-0.06005815044045448,
0.010169215500354767,
0.009908263571560383,
0.08797576278448105,
-0.01837138459086418,
-0.006552315782755613,
-0.04499790817499161,
-0.04181600362062454,
-0.03464528173208237,
-0.09847867488861084,
-0.06831064075231552,
-0.0052179391495883465,
0.08353649824857712,
-0.0442592017352581,
-0.04790731891989708,
-0.03879911080002785,
-0.03311145678162575,
-0.013488300144672394,
0.016215136274695396,
-0.026498734951019287,
-0.05590807646512985,
-0.047777313739061356,
0.012673198245465755,
0.0025847055949270725,
0.04483762010931969,
-0.07669483125209808,
-0.009235248900949955,
0.013467398472130299,
0.06596031039953232,
0.027420159429311752,
0.013069267384707928,
-0.038686174899339676,
0.00169481395278126,
-0.07814855128526688,
0.0799829438328743,
-0.02470257133245468,
0.06952034682035446,
0.09245298057794571,
0.09254767000675201,
0.12434032559394836,
-0.028370797634124756,
0.0787801742553711,
0.022929104045033455,
0.03259637579321861,
-0.05396267771720886,
0.08409107476472855,
0.03911673650145531,
0.13092458248138428,
0.00453161122277379,
-3.504828782752156e-8,
-0.07813982665538788,
0.05920163542032242,
-0.015558414161205292,
-0.0278333630412817,
0.017433730885386467,
-0.02209329791367054,
-0.032160114496946335,
0.08227000385522842,
0.05882107466459274,
-0.017198894172906876,
0.0748489648103714,
-0.03203563764691353,
-0.04135274514555931,
-0.04200055077672005,
-0.024244002997875214,
0.08796840906143188,
-0.024644222110509872,
0.045017991214990616,
-0.005631464999169111,
-0.0356934554874897,
0.08396420627832413,
0.011894901283085346,
-0.008385597728192806,
-0.010906318202614784,
0.009051776491105556,
-0.019484203308820724,
-0.05389058589935303,
0.08955767750740051,
0.018384162336587906,
-0.04183575138449669,
0.003790169255807996,
-0.006837007589638233,
0.02069655805826187,
-0.023738941177725792,
0.028468746691942215,
0.03849120810627937,
-0.00748345535248518,
-0.06314542144536972,
0.027733370661735535,
-0.04960815981030464,
0.06363295763731003,
0.052726440131664276,
-0.0352790467441082,
-0.004729119595140219,
0.02049027383327484,
0.010072200559079647,
0.03009170852601528,
-0.09777870774269104,
-0.012095334008336067,
-0.03496764600276947,
-0.0845578983426094,
-0.001609105966053903,
0.04614821821451187,
0.13812577724456787,
0.055035900324583054,
-0.015149781480431557,
0.016512274742126465,
0.047697823494672775,
0.07453513890504837,
0.12609019875526428,
-0.010187205858528614,
0.06369824707508087,
-0.04019339010119438,
-0.03460434079170227
] |
Subsets and Splits