Human-level concept learning through probablistic program in(2)

2020-03-27 02:26

RESEARCH|RESEARCHARTICLES

thesestructuressupport;fractalstructures,suchasriversandtrees,wherecomplexityarisesfromhighlyiteratedbutsimplegenerativeprocesses;andevenabstractknowledge,suchasnaturalnumber,naturallanguagesemantics,andintuitivephysicaltheories(17,46–48).

Capturinghowpeoplelearnalltheseconceptsatthelevelwereachedwithhandwrittencharac-tersisalong-termgoal.Inthenearterm,applyingourapproachtoothertypesofsymbolicconceptsmaybeparticularlypromising.Humanculturesproducemanysuchsymbolsystems,includinggestures,dancemoves,andthewordsofspokenandsignedlanguages.Aswithcharacters,theseconceptscanbelearnedtosomeextentfromoneorafewexamples,evenbeforethesymbolicmeaningisclear:Considerseeinga“thumbsup,”“fistbump,”or“highfive”orhearingthenames“BoutrosBoutros-Ghali,”“KofiAnnan,”or“BanKi-moon”forthefirsttime.Fromthislimitedexperience,peoplecantypicallyrecognizenewexamplesandevenproducearecognizablesem-blanceoftheconceptthemselves.TheBPLprin-ciplesofcompositionality,causality,andlearningtolearnmayhelptoexplainhow.

ToillustratehowBPLappliesinthedomainofspeech,programsforspokenwordscouldbeconstructedbycomposingphonemes(subparts)systematicallytoformsyllables(parts),whichcomposefurthertoformmorphemesandentirewords.Givenanabstractsyllable-phonemeparseofaword,realisticspeechtokenscanbegener-atedfromacausalmodelthatcapturesaspectsofspeechmotorarticulation.ThesepartandsubpartAlphabet of charactersi)New machine-generated characters in each alphabetii)Human or Machine?iii)121212i)Human or Machine?12ii)1212121212iii)componentsaresharedacrossthewordsinalanguage,enablingchildrentoacquirethemthroughalong-termlearning-to-learnprocess.Wehavealreadyfoundthataprototypemodelexploitingcompositionalityandlearningtolearn,butnotcausality,isabletocapturesomeaspectsofthehumanabilitytolearnandgeneralizenewspokenwords(e.g.,EnglishspeakerslearningwordsinJapanese)(49).Furtherprogressmaycomefromadoptingarichercausalmodelofspeechgeneration,inthespiritofclassic“analysis-by-synthesis”proposalsforspeechperceptionandlanguagecomprehension(20,50).

Althoughourworkfocusedonadultlearners,itraisesnaturaldevelopmentalquestions.Ifchil-drenlearningtowriteacquireaninductivebiassimilartowhatBPLconstructs,themodelcouldhelpexplainwhychildrenfindsomecharactersdifficultandwhichteachingproceduresaremosteffective(51).Comparingchildren’sparsingandgeneralizationbehavioratdifferentstagesoflearningandBPLmodelsgivenvaryingback-groundexperiencecouldbetterevaluatethemodel’slearning-to-learnmechanismsandsuggestimprove-ments.Bytestingourclassificationtasksoninfantswhocategorizevisuallybeforetheybegindrawingorscribbling(52),wecanaskwhetherchildrenlearntoperceivecharactersmorecausallyandcomposi-tionallybasedontheirownproto-writingexperi-ence.CausalrepresentationsareprewiredinourcurrentBPLmodels,buttheycouldconceivablybeconstructedthroughlearningtolearnatanevendeeperlevelofmodelhierarchy(53).

Last,wehopethatourworkmayshedlightontheneuralrepresentationsofconceptsandthedevelopmentofmoreneurallygroundedlearningmodels.Supplementingfeedforwardvisualpro-cessing(54),previousbehavioralstudiesandourresultssuggestthatpeoplelearnnewhandwrittencharactersinpartbyinferringabstractmotorprograms(55),arepresentationgroundedinpro-ductionyetactiveinpurelyperceptualtasks,independentofspecificmotorarticulatorsandpotentiallydrivenbyactivityinpremotorcortex(56–58).Couldwedecoderepresentationsstruc-turallysimilartothoseinBPLfrombrainimagingofpremotorcortex(orotheraction-orientedregions)inhumansperceivingandclassifyingnewchar-actersforthefirsttime?Recentlarge-scalebrainmodels(59)anddeeprecurrentneuralnetworks(60–62)havealsofocusedoncharacterrecogni-tionandproductiontasks—buttypicallylearningfromlargetrainingsampleswithmanyexamplesofeachconcept.Weseetheone-shotlearningcapacitiesstudiedhereasachallengefortheseneuralmodels:oneweexpecttheymightrisetobyincorporatingtheprinciplesofcomposition-ality,causality,andlearningtolearnthatBPLinstantiates.

REFERENCESANDNOTES

Fig.7.Generatingnewconcepts.(A)Humansandmachinesweregivenanovelalphabet(i)andaskedtoproducenewcharactersforthatalphabet.Newmachine-generatedcharactersareshownin(ii).Humanandmachineproductionscanbecomparedin(iii).Thefour-charactergridsineachpairthatweregeneratedbythemachineare(byrow)1,1,2;1,2.(B)Humansandmachinesproducednewcharacterswithoutareferencealphabet.Thegridsthatweregeneratedbyamachineare2;1;1;2.SCIENCEsciencemag.org

1.B.Landau,L.B.Smith,S.S.Jones,Cogn.Dev.3,299–321(1988).2.E.M.Markman,CategorizationandNaminginChildren(MIT

Press,Cambridge,MA,1989).

3.F.Xu,J.B.Tenenbaum,Psychol.Rev.114,245–272(2007).4.S.Geman,E.Bienenstock,R.Doursat,NeuralComput.4,1–58

(1992).

5.Y.LeCun,L.Bottou,Y.Bengio,P.Haffner,Proc.IEEE86,

2278–2324(1998).

11DECEMBER2015?VOL350ISSUE6266

1337

RESEARCH|RESEARCHARTICLES

6.G.E.Hintonetal.,IEEESignalProcess.Mag.29,82–97(2012).7.A.Krizhevsky,I.Sutskever,G.E.Hinton,Adv.NeuralInf.

Process.Syst.25,1097–1105(2012).

8.Y.LeCun,Y.Bengio,G.Hinton,Nature521,436–444(2015).9.V.Mnihetal.,Nature518,529–533(2015).

10.J.Feldman,J.Math.Psychol.41,145–170(1997).11.I.Biederman,Psychol.Rev.94,115–147(1987).12.T.B.Ward,Cognit.Psychol.27,1–40(1994).

13.A.Jern,C.Kemp,Cognit.Psychol.66,85–125(2013).14.L.G.Valiant,Commun.ACM27,1134–1142(1984).

15.D.McAllester,inProceedingsofthe11thAnnualConferenceon

ComputationalLearningTheory(COLT),Madison,WI,24to26July1998(AssociationforComputingMachinery,NewYork,1998),pp.230–234.

16.V.N.Vapnik,IEEETrans.NeuralNetw.10,988–999(1999).17.N.D.Goodman,J.B.Tenenbaum,T.Gerstenberg,Concepts:

NewDirections,E.Margolis,S.Laurence,Eds.(MITPress,Cambridge,MA,2015).

18.Z.Ghahramani,Nature521,452–459(2015).

19.P.H.Winston,ThePsychologyofComputerVision,P.H.Winston,

Ed.(McGraw-Hill,NewYork,1975).

20.T.G.Bever,D.Poeppel,Biolinguistics4,174(2010).21.L.B.Smith,S.S.Jones,B.Landau,L.Gershkoff-Stowe,

L.Samuelson,Psychol.Sci.13,13–19(2002).

22.R.L.Goldstone,inPerceptualOrganizationinVision:Behavioral

andNeuralPerspectives,R.Kimchi,M.Behrmann,C.Olson,Eds.(LawrenceErlbaum,City,NJ,2003),pp.233–278.23.H.F.Harlow,Psychol.Rev.56,51–65(1949).

24.D.A.Braun,C.Mehring,D.M.Wolpert,Behav.BrainRes.206,

157–165(2010).

25.C.Kemp,A.Perfors,J.B.Tenenbaum,Dev.Sci.10,307–321

(2007).

26.R.Salakhutdinov,J.Tenenbaum,A.Torralba,inJMLRWorkshop

andConferenceProceedings,vol.27,UnsupervisedandTransferLearningWorkshop,I.Guyon,G.Dror,V.Lemaire,G.Taylor,D.Silver,Eds.(Microtome,Brookline,MA,2012),pp.195–206.27.J.Baxter,J.Artif.Intell.Res.12,149(2000).

28.Y.LeCunetal.,NeuralComput.1,541–551(1989).

29.R.Salakhutdinov,J.B.Tenenbaum,A.Torralba,IEEETrans.

PatternAnal.Mach.Intell.35,1958–1971(2013).

30.V.K.Mansinghka,T.D.Kulkarni,Y.N.Perov,J.B.Tenenbaum,

Adv.NeuralInf.Process.Syst.26,1520–1528(2013).

31.K.Liu,Y.S.Huang,C.Y.Suen,IEEETrans.PatternAnal.Mach.

Intell.21,1095–1100(1999).

32.M.-P.Dubuisson,A.K.Jain,inProceedingsofthe12thIAPR

InternationalConferenceonPatternRecognition,Vol.1,ConferenceA:ComputerVisionandImageProcessing,Jerusalem,Israel,9to13October1994(IEEE,NewYork,1994),pp.566–568.

33.G.Koch,R.S.Zemel,R.Salakhutdinov,paperpresentedatICML

DeepLearningWorkshop,Lille,France,10and11July2015.34.M.Revow,C.K.I.Williams,G.E.Hinton,IEEETrans.Pattern

Anal.Mach.Intell.18,592–606(1996).

35.G.E.Hinton,V.Nair,Adv.NeuralInf.Process.Syst.18,515–522

(2006).

36.S.-C.Zhu,D.Mumford,FoundationsTrendsComput.Graphics

Vision2,259–362(2006).

37.P.Liang,M.I.Jordan,D.Klein,inProceedingsofthe27th

InternationalConferenceonMachineLearning,Haifa,Israel,21to25June2010(InternationalMachineLearningSociety,Princeton,NJ,2010),pp.639–646.

38.P.F.Felzenszwalb,R.B.Girshick,D.McAllester,D.Ramanan,

IEEETrans.PatternAnal.Mach.Intell.32,1627–1645(2010).39.I.Hwang,A.Stuhlmüller,N.D.Goodman,http://arxiv.org/abs/

1110.5667(2011).

40.E.Dechter,J.Malmaud,R.P.Adams,J.B.Tenenbaum,in

Proceedingsofthe23rdInternationalJointConferenceonArtificialIntelligence,F.Rossi,Ed.,Beijing,China,3to9August2013(AAAIPress/InternationalJointConferencesonArtificialIntelligence,MenloPark,CA,2013),pp.1302–1309.

41.J.Rule,E.Dechter,J.B.Tenenbaum,inProceedingsofthe37th

AnnualConferenceoftheCognitiveScienceSociety,D.C.Noelleetal.,Eds.,Pasadena,CA,22to25July2015(CognitiveScienceSociety,Austin,TX,2015),pp.2051–2056.42.L.W.Barsalou,Mem.Cognit.11,211–227(1983).

43.J.J.Williams,T.Lombrozo,Cogn.Sci.34,776–806(2010).44.A.B.Markman,V.S.Makin,J.Exp.Psychol.Gen.127,331–354

(1998).

45.D.N.Osherson,E.E.Smith,Cognition9,35–58(1981).

46.S.T.Piantadosi,J.B.Tenenbaum,N.D.Goodman,Cognition

123,199–217(2012).

47.G.A.Miller,P.N.Johnson-Laird,LanguageandPerception

(Belknap,Cambridge,MA,1976).

48.T.D.Ullman,A.Stuhlmuller,N.Goodman,J.B.Tenenbaum,

inProceedingsofthe36thAnnualConferenceoftheCognitiveScienceSociety,QuebecCity,Canada,23to

26July2014(CognitiveScienceSociety,Austin,TX,2014),pp.1640–1645.

49.B.M.Lake,C.-y.Lee,J.R.Glass,J.B.Tenenbaum,in

Proceedingsofthe36thAnnualConferenceoftheCognitiveScienceSociety,QuebecCity,Canada,23to26July2014(CognitiveScienceSociety,Austin,TX,2014),pp.803–808.50.U.Neisser,CognitivePsychology(Appleton-Century-Crofts,

NewYork,1966).

51.R.Treiman,B.Kessler,HowChildrenLearntoWriteWords

(OxfordUniv.Press,NewYork,2014).

52.A.L.Ferry,S.J.Hespos,S.R.Waxman,ChildDev.81,472–479

(2010).

53.N.D.Goodman,T.D.Ullman,J.B.Tenenbaum,Psychol.Rev.

118,110–119(2011).

54.S.Dehaene,ReadingintheBrain(Penguin,NewYork,2009).55.M.K.Babcock,J.J.Freyd,Am.J.Psychol.101,111–130

(1988).

56.M.Longcamp,J.L.Anton,M.Roth,J.L.Velay,Neuroimage19,

1492–1500(2003).

57.K.H.James,I.Gauthier,Neuropsychologia44,2937–2949

(2006).

58.K.H.James,I.Gauthier,J.Exp.Psychol.Gen.138,416–431

(2009).

59.C.Eliasmithetal.,Science338,1202–1205(2012).60.A.Graves,http://arxiv.org/abs/1308.0850(2014).

61.K.Gregor,I.Danihelka,A.Graves,D.J.Rezende,D.Wierstra,

inProceedingsoftheInternationalConferenceonMachineLearning(ICML),Lille,France,6to11July2015(InternationalMachineLearningSociety,Princeton,NJ,2015),pp.1462–1471.

62.J.Chungetal.,Adv.NeuralInf.Process.Syst.28,(2015).

ACKNOWLEDGMENTS

ThisworkwassupportedbyaNSFGraduateResearchFellowshiptoB.M.L.;theCenterforBrains,Minds,andMachinesfundedbyNSFScienceandTechnologyCenterawardCCF-1231216;ArmyResearchOfficeandOfficeofNavalResearchcontractsW911NF-08-1-0242,W911NF-13-1-2012,andN000141310333;theNaturalSciencesandEngineeringResearchCouncilofCanada;theCanadianInstituteforAdvancedResearch;andtheMoore-SloanDataScienceEnvironmentatNYU.WethankJ.McClelland,T.PoggioandL.SchulzformanyvaluablecontributionsandN.Kanwisherforhelpinguselucidatethe

threekeyprinciples.WearegratefultoJ.GrossandtheOmniglot.comencyclopediaofwritingsystemsforhelpingtomakethisdatasetpossible.OnlinearchivesareavailableforvisualTuringtests(http://github.com/brendenlake/visual-turing-tests),Omniglotdataset(http://github.com/brendenlake/omniglot),andBPLsourcecode(http://github.com/brendenlake/BPL).

SUPPLEMENTARYMATERIALS

www.sciencemag.org/content/350/6266/1332/suppl/DC1MaterialsandMethodsSupplementaryTextFigs.S1toS11

References(63–80)

1June2015;accepted15October201510.1126/science.aab3050PHYSICALCHEMISTRYSpectroscopiccharacterizationofisomerizationtransitionstatesJoshuaH.Baraban,1*P.BryanChangala,1?GeorgCh.Mellau,2JohnF.Stanton,3AnthonyJ.Merer,4,5RobertW.Field1?Transitionstatetheoryiscentraltoourunderstandingofchemicalreactiondynamics.Wedemonstrateamethodforextractingtransitionstateenergiesandpropertiesfromacharacteristicpatternfoundinfrequency-domainspectraofisomerizingsystems.Thispattern—adipinthespacingsofcertainbarrier-proximalvibrationallevels—canbeunderstoodusingtheconceptofeffectivefrequency,weff.Themethodisappliedtothecis-transconformationalchangeintheS1stateofC2H2andthebond-breakingHCN-HNCisomerization.Inbothcases,thebarrierheightsderivedfromspectroscopicdataagreeextremelywellwithpreviousabinitiocalculations.Wealsoshowthatitispossibletodistinguishbetweenvibrationalmodesthatareactivelyinvolvedintheisomerizationprocessandthosethatarepassivebystanders.hecentralconceptofthetransitionstateinchemicalkineticsisfamiliartoallstudentsofchemistry.SinceitsinceptionbyArrhe-nius(1)andlaterdevelopmentintoafulltheorybyEyring,Wigner,Polanyi,andEvans(2–5),theideathatthethermalratedependsprimarilyonthehighestpointalongthelowest-DepartmentofChemistry,MassachusettsInstituteofTechnology,Cambridge,MA02139,USA.2Physikalisch-ChemischesInstitut,Justus-Liebig-Universit?tGiessen,D-35392Giessen,Germany.3DepartmentofChemistryandBiochemistry,UniversityofTexas,Austin,TX78712,USA.4DepartmentofChemistry,UniversityofBritishColumbia,Vancouver,BCV6T1Z1,Canada.5InstituteofAtomicandMolecularSciences,AcademiaSinica,Taipei10617,Taiwan.

*Presentaddress:DepartmentofChemistryandBiochemistry,UniversityofColorado,Boulder,CO80309,USA.?Presentaddress:JILA,NationalInstituteofStandardsandTechnology,and

DepartmentofPhysics,UniversityofColorado,Boulder,CO80309,USA.?Correspondingauthor.E-mail:rwfield@mit.edu

1Tenergypathfromreactantstoproductshasre-mainedessentiallyunchanged.Mostofchemicaldynamicsisnowfirmlybasedonthisideaofthetransitionstate,notwithstandingtheemergenceofunconventionalreactionssuchasroaming(6,7),whereaphotodissociatedatomwandersbeforeabstractingfromtheparentfragment.Despitetheclearimportanceofthetransitionstatetothefieldofchemistry,directexperimentalstudiesofthetransitionstateanditspropertiesarescarce(8).Here,wereporttheobservationofavibrationalpattern,adipinthetrendofquantumlevelspac-ings,whichoccursattheenergyofthesaddlepoint.Thisphenomenonisexpectedtoprovideagenerallyapplicableandaccuratemethodforcharacterizingtransitionstates.Onlyasubsetofvibrationalstatesexhibitadip;thesestatescontainexcitationalongthereactioncoordinateandarebarrier-proximal,meaningthattheyaremore

sciencemag.orgSCIENCE

1338

11DECEMBER2015?VOL350ISSUE6266

Human-level concept learning through probabilistic programinduction

Brenden M. Lake et al.Science 350, 1332 (2015);

DOI: 10.1126/science.aab3050

This copy is for your personal, non-commercial use only.

Permission to republish or repurpose articles or portions of articles can be obtained byfollowing the guidelines here.

The following resources related to this article are available online at

www.sciencemag.org (this information is current as of December 10, 2015 ):

Updated information and services, including high-resolution figures, can be found in the onlineversion of this article at:

http://www.sciencemag.org/content/350/6266/1332.full.html

Supporting Online Material can be found at:

http://www.sciencemag.org/content/suppl/2015/12/09/350.6266.1332.DC1.html

A list of selected additional articles on the Science Web sites related to this article can befound at:

http://www.sciencemag.org/content/350/6266/1332.full.html#related This article cites 50 articles, 2 of which can be accessed free:

http://www.sciencemag.org/content/350/6266/1332.full.html#ref-list-1 This article appears in the following subject collections:Psychology

http://www.sciencemag.org/cgi/collection/psychology

Science (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by theAmerican Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. Copyright2015 by the American Association for the Advancement of Science; all rights reserved. The title Science is aregistered trademark of AAAS.

Downloaded from www.sciencemag.org on December 11, 2015If you wish to distribute this article to others, you can order high-quality copies for yourcolleagues, clients, or customers by clicking here.


Human-level concept learning through probablistic program in(2).doc 将本文的Word文档下载到电脑 下载失败或者文档不完整,请联系客服人员解决!

下一篇:污水处理中级工试题

相关阅读
本类排行
× 注册会员免费下载(下载后可以自由复制和排版)

马上注册会员

注:下载文档有可能“只有目录或者内容不全”等情况,请下载之前注意辨别,如果您已付费且无法下载或内容有问题,请联系我们协助你处理。
微信: QQ: