Computational Cognitive Neuroscience and Its Appli

ComputationalCognitiveNeuroscienceandItsApplicationsLaurentIttiUniversityofSouthernCaliforniaIntroductionandmotivationAnumberoftaskswhichcanbeeffortlesslyachievedbyhumansandotheranimalshaveuntilrecentlyremainedseeminglyintractableforcomputers.Strikingexamplesofsuchtasksexistinthevisualandauditorydomains.Theseincluderecognizingthefaceofafriendinaphotograph,understandingwhetheranew,never-seenobjectisacaroranairplane,quicklyfindingobjectsorpersonsofinterestlikeone'schildinacrowd,understandingfluentspeechwhilealsodecipheringtheemotionalstateofthespeaker,orreadingcursivehandwriting.Infact,severalofthesetaskshavebecomethehallmarkofhumanintelligence,whileother,seeminglymorecomplexandmorecognitivelyinvolvedtasks,suchasplayingcheckersorchess,solvingdifferentialequations,orprovingtheoremshavebeenmasteredbymachinestoareasonabledegree(Samuel,1959;Newell&Simon,1972).Aneverydaydemonstrationofthisstateofaffairsistheuseofsimpleimage,character,orsoundrecognitioninCAPTCHAtests(CompletelyAutomatedPublicTuringTeststoTellComputersandHumansApart)usedbymanywebsitestoensurethatahuman,ratherthanasoftwarerobot,isaccessingthesite(CAPTCHAtests,forexample,areusedbywebsitesprovidingfreeemailaccountstoregisteredusers,andareasimpleyetimperfectwaytopreventspammersfromopeningthousandsofemailaccountsusinganautomatedscript).Tosomeextent,thesemachine-intractabletasksarethecauseforourfallingshortontheearlypromisesmadeinthe1950sbythefoundersofArtificialIntelligence,ComputerVision,andRobotics(Minsky,1961).Althoughtremendousprogresshasbeenmadeinjustahalfcentury,andoneisbeginningtoseecarsthatcandriveontheirownorrobotsthatvacuumthefloorwithouthumansupervision,suchmachineshavenotyetreachedmainstreamadoptionandremainhighlylimitedintheirabilitytointeractwiththerealworld.Althoughintheearlyyearsonecouldblamethepoorperformanceofmachinesonlimitationsincomputingresources,rapidadvancesinmicroelectronicshavenowrenderedsuchexcuseslessbelievable.Thecoreoftheproblemisnotonlyhowmuchcomputingcyclesonemayhavetoperformatask,buthowthosecyclesareused,inwhatkindofalgorithmandofcomputingparadigm.Forbiologicalsystems,interactingwiththevisualworld,inparticularthroughvision,audition,andothersenses,iskeytosurvival.Essentialtaskslikelocatingandidentifyingpotentialprey,predators,ormatesmustbeperformedquicklyandreliablyifananimalistostayalive.Takinginspirationfromnature,recentworkincomputationalneurosciencehashencestartedtodeviseanewbreedofalgorithms,whichcanbemoreflexible,robust,andadaptivewhenconfrontedwiththecomplexitiesoftherealworld.Iherefocusondescribingrecentprogresswithafewsimpleexamplesofsuchalgorithms,concernedwithdirectingattentiontowardsinterestinglocationsinavisualscene,soastoconcentratethedeploymentofcomputingresourcesprimarilyontotheselocations.ModelingvisualattentionPositivelyidentifyinganyandallinterestingtargetsinone'svisualfieldhasprohibitivecomputationalcomplexity,makingitadauntingtaskevenforthemostsophisticatedbiologicalbrains(Tsotsos,1991).Onesolution,adoptedbyprimatesandmanyotheranimals,istobreakdownthevisualanalysisoftheentirefieldofviewintosmallerregions,eachofwhichiseasiertoanalyzeandcanbeprocessedinturn.Thisserializationofvisualsceneanalysisisoperationalizedthroughmechanismsofvisualattention:Acommon(athoughsomewhatinaccurate)metaphorforattentionisthatofavirtual''spotlight,''shiftingtowardsandhighlightingdifferentsub-regionsofthevisualworld,sothatoneregionatatimecanbesubjectedtomoredetailedvisualanalysis(Treisman&Gelade,1980;Crick,1984;Weichselgartner&Sperling,1987).Thecentralprobleminattentionresearchthenbecomeshowtobestdirectthisspotlighttowardsthemostinterestingandbehaviorallyrelevantvisuallocations.Simplestrategies,likeconstantlyscanningthevisualfieldfromlefttorightandfromtoptobottom,likemanycomputeralgorithmsdo,maybetooslowforsituationswheresurvivaldependsonreactingquickly.Recentprogressincomputationalneurosciencehasproposedanumberofnewbiologically-inspiredalgorithmswhichimplementmoreefficientstrategies.Thesealgorithmsusuallydistinguishbetweenaso-called''bottom-up''driveofattentiontowardsconspicuousor''salient''locationsinthevisualfield,andavolitionalandtask-dependentso-called''top-down''driveofattentiontowardbehaviorallyrelevantsceneelements(Desimone&Duncan,1995;Itti&Koch,2001).Simpleexamplesofbottom-upsalientandtop-downrelevantitemsareshowninFigure1.Figure1:(left)Examplewhereoneitem(aredandroughlyhorizontalbar)inanarrayofitemsishighlysalientandimmediatelyandeffortlesslygrabsvisualattentionattentioninabottom-up,image-drivenmanner.(right)Examplewhereasimilaritem(aredandroughlyverticalbar)isnotsalientbutmaybebehaviorallyrelevantifyourtaskistofinditasquicklyaspossible;top-down,volition-drivenmechanisms,mustbedeployedtoinitiateasearchfortheitem.(AlsoseeTreisman&Gelade,1980).KochandUllman(1985)introducedtheideaofasaliencymaptoaccomplishpreattentiveselectionintheprimatebrain.Thisisanexplicittwo-dimensionalmapthatencodesthesaliencyofobjectsinthevisualenvironment.Competitionamongneuronsinthismapgivesrisetoasinglewinninglocationthatcorrespondstothemostsa