View
48
Download
9
Category
Preview:
Citation preview
Page3
UnderstandingRepresentationintheCognitiveSciencesDoesRepresentationNeedReality?
Editedby
AlexanderRieglerFreeUniversityofBrussels
Brussels,Belgium
MarkusPeschlUniversityofVienna
Vienna,Austria
and
AstridvonSteinUniversity/ETHZurichZurich,Switzerland
Page4
BasedontheproceedingsoftheconferenceNewTrendsinCognitiveScience(NTCS'97),heldMay1416,1997,inVienna,Austria
ISBN:0306462869
1999KluwerAcademic/PlenumPublishers233SpringStreet,NewYork,N.Y.10013
10987654321
AC.I.P.recordforthisbookisavailablefromtheLibraryofCongress.
Allrightsreserved
Nopartofthisbookmaybereproduced,storedinaretrievalsystem,ortransmittedinanyformorbyanymeans,electronic,mechanical,photocopying,microfilming,recording,orotherwise,withoutwrittenpermissionfromthePublisher
PrintedintheUnitedStatesofAmerica
Page5
Preface
Currentlyaparadigmshiftisoccurringinwhichthetraditionalviewofthebrainasrepresentingthe"thingsoftheworld"ischallengedinseveralrespects.Thepresentvolumeisplacedattheedgeofthistransition.Basedonthe1997conference"NewTrendsinCognitiveScience"inVienna,Austria,ittriestocollectandintegrateevidencefromvariousdisciplinessuchasphilosophyofscience,neuroscience,computationalapproaches,psychology,semiotics,evolutionarybiology,socialpsychologyetc.,tofosteranewunderstandingofrepresentation.
Thesubjectiveexperienceofanoutsideworldseemstosuggestamappingprocesswhereenvironmentalentitiesareprojectedintoourmindviasomekindoftransmission.Whileaprofoundcritiqueofthisideaisnearlyasoldasphilosophy,ithasgainedconsiderablesupportwiththeadvancementofempiricalscienceintothestudyofmentalprocesses.Evidencesuchasthediscoveryofsinglecellsthatrespondtoparticularenvironmentalfeatures,orspecificareasofthebrainthatlightupduringspecificmentalprocessesinimagingstudies,havesupportedthenotionofamappingprocess,andprovidedadeepfoundationformaterialismand"localism".
Buttheideaofaclearandstablereferencebetweenarepresentationalstate(e.g.,inaneuron,aHebbianensemble,anactivationstate,etc.)andtheenvironmentalstatehasbecomequestionable.Already,weknowthatlearnedexperiencesandexpectationscanhaveanimpactontheneuralactivitythatisasstrongasthestimulusitself.Sincetheseinternallystoredexperiencesareconstantlychanging,thenotionofreferentialrepresentationsischallenged.Thegoalofthisbookistodiscussthephenomenonofrepresentationonvariouslevelsofinvestigation,aswellasitsimplications.
Inordertogivemuchroomtoconceptualandepistemologicalquestions(andlesstotechnicaldetails)thebookstartswithourpositionpaper"DoesRepresentationNeedReality?"Itopensthegroundinreviewingevidencethatcreateproblemsfortheconventionalunderstandingofrepresentations.Thepaperalsosummarizestherationalefortheselectionofcontributionstothisvolume,whichwillroughlyproceedfromrelatively"realist"conceptionsofrepresentationtomore"constructivist"interpretations.Thefinalchapterofdiscussions,tapedduringandattheendoftheconference,providesthereaderwiththepossibilitytoreflectuponthedifferentapproachesandthuscontributestobetterandmoreintegrativeunderstandingoftheirthoughtsandideas.
Thisbookhasatrulyinterdisciplinarycharacter.Itispresentedinaformthatisreadilyaccessibletoprofessionalsandstudentsalikeacrossthecognitivesciencessuchasneuroscience,computerscience,philosophy,psychology,andsociology.Wehopethatitwillpavethewayforabetterunderstandingofrepresentationandinspireitsreadersintheirfieldofstudy.
Acknowledgements
WehavegreatlybenefitfromtheAustrianSocietyofCognitiveSciencewhichattractedboththeidealismofparticipantsandthemoneyoffinanciers.ItpleasesusverymuchtothanktheAustrianFederalMinistryofScience,TransportandtheArts,theformerChristianDopplerLaboratoryforExpertSystems(ViennaUniversityofTechnology),theOsterreichischeForschungsgemeinschaft,andtheCityofVienna.AlexRiegleracknowledgesthefinancialsupportoftheOesterreichischeNationalbank,projectnumber5722/4.ThetitlegraphicsisreproducedwithkindpermissionofAmandaHeitler(neePask).Last,butnotleast,wewouldliketothankthefollowingpeoplewhohavebeeninvolvedintheprocessofpreparingthisbook:LianeGabora,BrigitteRmmer,JohannesSarnthein,andVroniqueWilquet.
THEEDITORS,MARCH1999
Page7
Tableofcontents
PositionPaper
MarkusFPeschl&AlexanderRieglerDoesRepresentationNeedReality? 9
OverviewofContributions 19
DifferentFacetsofRepresentation
GeorgDorffnerTheConnectionistRoutetoEmbodimentandDynamicism 23
MatthiasScheutzTheOntologicalStatusofRepresentations 33
AnthonyChemeroEmpiricalandMetaphysicalAntiRepresentationalism 39
AlfredoPereira,Jr.RepresentationinCognitiveNeuroscience 49
DanielD.HuttoCognitionwithoutRepresentation? 57
ComputationalApproaches
GeorgSchwarzOnComputingSystemsandTheirEnvironment 75
WilliamS.RobinsonRepresentationandCognitiveExplanation 83
RobertM.French
WhenCoffeeCupsAreLikeOldElephants,orWhyRepresentationModulesDon'tMakeSense
93
L.AndrewCowardTheRecommendationArchitecture:RelatingCognitiontoPhysiology 101
CognitionasaDynamicalSystem
PimHaselagerNeurodynamicsandtheRevivalofAssociationisminCognitiveScience 115
StevenL.BresslerTheDynamicManifestationofCognitiveStructuresintheCerebralCortex 121
KenMogiResponseSelectivity,NeuronDoctrine,andMach'sPrincipleinPerception 127
MariusUsher&ErnstNieburMentalRepresentations:AComputationalNeuroscienceScheme 135
RelevanceofActionforRepresentation
ErichHarthSketchpadsinandBeyondtheBrain 143
MarkWexlerInductiveLearningwithExternalRepresentations 147
AstridVonStein
DoestheBrainRepresenttheWorld?EvidenceAgainsttheMappingAssumption
161
RalfMller
PerceptionThroughAnticipation.ABehaviourBasedApproachtoVisualPerception
169
Page8
SymbolGroundingandLanguage
TomZiemkeRethinkingGrounding 177
N.Chandler,VBalendran,L.Evett,&K.SivayoganathanReality:APrerequisitetoMeaningfulRepresentation 191
ChristianBalkenius&SimonWinterExplorationsinSyntheticPragmatics 199
CommunicationandSocialCoupling
PeterGrdenforsDoesSemanticsNeedReality? 209
J.RichardEiserEmpiricismandSocialReality:CanCognitiveScienceBeSocialized? 219
TomRoutenHabitusandAnimats 229
S.Weiss,H.M.Mller,&P.Rappelsberger
ProcessingConceptsandScenarios:ElectrophysiologicalFindingsonLanguageRepresentation
237
HannaRiskuConstructivistConsequences:TranslationandReality 247
QualitativeAspectsofRepresentationandConsciousness
WolfSingerTheObserverintheBrain 253
MichaelPauenRealityandRepresentation.Qualia,Computers,andthe"ExplanatoryGap" 257
Constructivism
AnnikaWallinCanaConstructivistDistinguishbetweenExperienceandRepresentation? 269
SverreSjolanderHowAnimalsHandleRealityTheAdaptiveAspectofRepresentation 277
ErnstVonGlasersfeldPiaget'sLegacy:CognitionasAdaptiveActivity 283
Appendix
DiscussionsofRepresentationandConstructivism 289
SubjectIndex 301
Page9
DoesRepresentationNeedReality?RethinkingEpistemologicalIssuesintheLightofRecentDevelopmentsandConceptsinCognitiveScience
MarkusF.PeschlDept.forPhilosophyofScience,Univ.ofVienna,Sensengasse8/10,A1090Wien,AustriaEmail:franzmarkus.peschl@univie.ac.at
AlexanderRieglerCLEA,FreeUniversityBrussels,ruedelaStrategie33,B1160Brussels,BelgiumEmail:ariegler@vub.ac.be
AbstractThispaperdiscussesthenotionofrepresentationandoutlinestheideasandquestionswhichledtotheorganizationofthisvolume.Weargueforadistinctionbetweentheclassicalviewofreferentialrepresentation,andthealternativeconceptofsystemrelativerepresentation.Thelatterreferstosituatedcognitiveprocesseswhosedynamicsaremerelymodulatedbytheirenvironmentratherthanbeinginstructedanddeterminedbyit.
Introduction
Howis"realityoutthere"representedinourheads?Doescognitionworkwithinformationfromthe"outside"world?Isourthinkingexclusivelydeterminedbytheworld,isitexposedtoperturbationsfromanapparentlyendlessenvironment,ordoescognitionactivelygenerateandconstructthe"world"?
Toapproachthesequestions,letusstartwithPlato'swellknownallegoryofthecave(TheRepublic.BookVII).Inthisallegory,prisonersareboundinacaveinsuchawaythattheycannotturntheirheadsormovearound.Theycanonlyseeawallinfrontofthem.Thelightofadistantfirebehindthemcastsshadowsonthecavewallofthemselvesandotherpeoplewanderingaround.Theprisonershavebeenrestrictedtothisperspectivesincebirth.Therefore,theironlyperceptionofthemselvesandtheirworldisthroughthemovingshadowsonthewall.Hence,theprisonersperceiveandtaketheseshadowstobetheactualobjectsintheworldratherthanrecognizingthemasmereshadowsofthe"real"environment.
Obviously,whenlookingattheproblemofrepresentationwefaceasimilarsituation:wearetheprisonershavingonlyamediatedaccesstothe"real"world.Howcanwedistinguishthe"true"realityfromthefake?WittgensteinapproachedthisquestioninhisTractatusPhilosophicus:"Inordertotellwhetherapictureistrueorfalsewemustcompareitwithreality."(2.223)
Thismakestheproblemofreferencetoa"realouterworld"clear.Fromsuchaperspectivewehavetodistinguishbetweentheworldasthedomainofourexperience(actuality,wirklichkeitW)andtheworldasthedomainofthingsinthemselves(realittR):
RealittfromtheLatin"res"(thing)connotestheontologicallygivenenvironmenteveryrealistmakesreferenceto.By"ontology"werefertothephilosophicaltraditionofclaimingtheexistenceandrecognizabilityofanindependentworldoutside,theexistenceofthingsinthemselves(DingeansichinthesenseofKant)likethe"real"peoplewhoseshadowsareperceivedbytheprisoners.
Wirklichkeit(actuality)ontheotherhandstemmingfromtheGermanverb"wirken",meaningtohaveaneffectoncanbeusedtodesignatethe"constructed"worldinourminds,madeupbyourexperiencesand(genetic)predispositions.
TheepistemologicallymosttrivialversionoftherelationshipbetweenWandRsuggests:W=R.Here,oneassumesthatanorganismperceivesitsenvironmentdirectlyandfreefromdistortion,andtheworldofexperienceisdirectlyrelatedtoanobjectiveworld(radicalrealism).
Page10
Thisisonlyacommonsenserelationshipandseemstobeepistemologicallynotveryplausible.Inthefollowingweintroducemoresophisticatedconceptsthathavebeendeveloped1.
ReferentialRepresentation
Fortheclassicalrepresentationaltheory,wirklichkeitisafunctionofrealitt,W=f(R).Inthisview,ourexperiencesareadistortedandnonproportionalimageofrealitt.Thistheoryisunsatisfactoryinsofarasitdoesnottellusalotaboutthenatureofthedistortionsandtheirrelationshiptotheobserver.Furthermore,itisfromaneuroscientificaswellassystemtheoreticperspectivenotveryplausiblethattherepresentationWdependsonlyontherealittR.Suchaview,i.e.,W=f(R),mightimplysomekindofdistortion.2Therepresentationalrelationship,however,remainsbasicallyreferential(whichseemstobecontradictorytoourexperiences.3
ContextDependentRepresentation
Withinadialecticmaterialisticworldviewtheconstructionofwirklichkeitisestablishedbyaninteractionbetweenanobserverandtheobserved(environment/phenomenon):W=f(R,O,C).0denotesthepropertiesoftheobserverandCtheculturalhistoricalbackground.Suchaperspectivetakesintoaccountthepropertiesoftheobserverinthesensethatheractivityofconstructionplaysanactiveroleintheformationoftherepresentation.However,itmaybedifficulttosufficientlydefinethepropertiesofCand0.
SelfReferentialRepresentation
Theconceptofcognitiveselfreferencedescribesperceptionandrepresentationasperceptionofrelations.Stimuliaremereperipheralenergeticconditions(i.e.,perturbationsP)forasemanticallyclosedandselforganizingcognitivesystem.Thestructureofthecognitivesystemdetermineswhichstructuralconfigurationsofitssurroundingsareperturbationstothesystem,andwhicharenot.Theideaisthatthecognitivesystemisinadynamicalequilibrium.Thismeansthattheperpetuallyactingcomponentsofthesystem(e.g.,neurons)respondsolelytotheactivityofothercomponents.Thereisnootherwayofinfluencingtheirstatewithoutdestroyingthem,muchasadetectorofradiowavesdisplaysactivityonlyinthepresenceofwavesofacertainlength.Fromtheperspectiveofanobserver,anorganism(withitsselfreferentialcognitiveequipment)isembeddedwithinitsenvironment.Someoftheprocesseswithintheenvironment(andexternaltotheorganism)actasperturbationstotheorganism:Throughthesensorysurfacetheyaretransducedintoneuronalactivitywhichinturnmayhaveanimpactonthedynamicalequilibriumofthecognitivesystem.Generallyspeaking,cognitiveselfreferencemeansthatwirklichkeitisafunctionofthreevariables:W=f(W,E,P).Edenotestheindividualbackgroundexperienceofaspecificorganismfromtheperspectiveofsystemstheoryorcomputationalneuroscience,Ereferstothestructureofthestatespace.Inotherwords,Edeterminesthespaceofpossiblesuccessorstatesofeveryparticularstateofthesystem.4NotethatRisnotpartoftheequation,astheneuronsreferonlytoeachother!Inaconstructivistcontext,weinterprettheabsenceofRasaconsequenceoftheoperationalclosureofthecognitiveapparatus.Thatis,thestatesofneuralactivityalwaysoriginatefromandleadtootherstatesofneuronalactivityinarecurrent,selfreferringmanner(Maturana&Varela1979Winograd&Flores1986).Fromthispointofview,todealwithanobjectmeansforanorganismtodealwithitsowninternalstates.5Laterinthispaperwewillelaboratetheideaofcognitiveselfreferenceand
1AsimilarcategorizationcanbefoundinStadlerandKruse(1990).2Thedistortioniscausedbythefunction (.)whichcanbethoughtofasadescriptionofthedistortiontakingplace,forinstance,inthesensorysystemorintheprocessesoccurringinthetransductionorprimaryprocessingoftheenvironmentalsignal.3Think,forinstanceofthephenomenonofcolorconstancymoregenerallyspeaking,thinkofallthesituationsinwhichoneparticularenvironmentalevent/phenomenonisexperiencedindifferentways(accordingtoourpresentinternalstate).4Ofcourse,thisisadescriptiveperspective.Dynamicallyspeaking,thenervoussystemdoesnotrefertopaststates,e.g.,itsstate10secondsago(cf.thenontemporalcharacterizationofthenervoussystem,Maturana&Varela1979).5Thisofcoursedoesnotrefertoasolipsisticworldview,sincethisequationdescribesamappingontowirklichkeitWratherthanrealitt!
Page11
arguethatitsupersedestheprevailingreferentialconceptofrepresentationinthecognitivesciences.
Atthispoint,aninterestingquestionappears:WhatisthenatureoftheperturbationsPinthecognitiveselfreferenceperspective?Thisquestiongiveswaytomanyinterpretations.Arathercautiousposition(e.g.,Roth1994)suggeststhatperturbationsarecertainpatternsofenergytransducedbythesensororgansthatgiverisetoamosaicofelementaryeventsthatthebraintriestomakesenseoutof.Thiswouldmeanthatobjectsinthetraditionalsense(chairs,mountains,etc.)donotexistasobjectiveentitiesbutenergy(suchaselectromagneticwaves)does(cf.Chandler'srejectionofthisconcept,thisvolume).Asomewhatmoreradicalinterpretation(suchasGlasersfeld's)acknowledgesthearbitrarinessofacceptingtheabsoluteexistenceofenergywhiledismissing"ordinary"objects.Let'sinvestigatethisalittledeeper.Aboveweemphasizedthatthisviewistheviewofanobserverwhodistinguishesbetweenanorganismanditsenvironment,bothbeingexternaltoherself.However,wearriveatacrucialpointwhenspeakingofone'sowncognitiveselfreference:Anobserverwhoobservesherselfcannevertranscendthisperspective.Thatis,wecanplacebothourselvesandourenvironmentonly"exernal"toourselveswhendescribingtheirrelationship.Thereforewecannothelpbutassumethatthereisindeedanexternalrealittwhichistheoriginatorofperturbationsthatinfluenceournervoussystemthroughmediationofoursensors.However,aswealwayshavetodothesomewhatbizarrestepofthinkingofourselvesasathirdperson,suchaperspectivecanneverrevealthe"true"natureofperturbations,evenworse,wecannotevenclaimnordenytheexistenceofrealitt.
Implications
Whataretheimplicationsforouroriginalquestionofwhetherrepresentationneedsreality?Obviously,thesolutiontotheproblemofappropriaterepresentationdependsontheperspectivefromwhichwelookattheagent.Onewellknownexamplefora,cognitivelyspeaking,"misleading"approacharemicroworldmodelsusedinartificialintelligence(Dreyfus&Dreyfus1988):i.e.,theattempttodeterminebasiccognitivemechanismsbyreducingthecomplexityoftherealworldtothesimplicityofatoyworldandstartingfromthissimplifiedand"cognitivelypreprocessedworld"tobuilduparepresentationalsystemasanetworkofpropositions.Itisobviousthatallthetaskof"understanding"wasactuallydonebytheprogrammersthemselves.Theyputalargeamountofpreprocessedinformationintothesystem,thuscreatingauniverseofinterrelatedfacts(i.e.,purelysyntacticalstructures),buttheydidnotcreateameaningfulworld.Asonecanseeinhighlycomplexexpertsystems,suchastrategyisnotlimitedtosmalltoyworlds,butcangoveryfaronthelevelofthecomplexityofknowledgeithastobeclear,however,thatthisimmenseandquantitativeincreaseincomplexity(i.e.,morerules,facts,relations,etc.)doesnotimplyanewlevel(e.g.,"semantics").Hence,itdoesnotfromanepistemologicalperspectivebringaboutanewquality.
Notonlyhassymboliccognitivesciencecometoadeadend(inthecontextoftheefforttoexplaincognitiveprocesses).Theeuphoricdaysofthenewpossibilitiesofconnectionistnetworks,includingthesimulationofcognitiveactivitiesandlearning,havepassed,too.ItseemsthatcognitivesciencehasreturnedbacktoinKuhn'ssense(1970)"normalscience"orto"puzzlesolving".Asanimplicationofthisdevelopment,connectionistcognitivesciencehasalmostlostitselfin(technical)details,suchaslearningfactors,minimaladjustments,andoptimizationsinlearningalgorithmsoractivationfunctions,andsoon.Intacklingonlythesetechnical"microproblems",itseemstohavefailedtoaddressthereallyinteresting,pressing,qualitative,and"big"questionsaboutcognition.
Nevertheless,inthecontextofrecentdevelopmentsincognitivescience(e.g.,computationalneuroscience,artificiallife,autonomousagents,situatedaction,robotics,etc.)aconsiderableamountofepistemologicalpotentialcanbefoundbasicissueshavereceivednewinputsandnewinterpretationsoverthelastyears.Thegoalofthisvolumeistoinvestigatesomeofthesetrendsandtomakethemmoreexplicitinordertoachievesomeclarityastowherecognitivesciencemightdevelopinthefuture.Amongtheseissuesare:
Thenecessityofrethinkingtheconceptofrepresentationinthelightofdynamical,topdown,andrecurrentprocessesinnaturalandartificialcognitivesystems,inparticularinneuralsystems
Theembodimentofknowledgeinitssubstratum,e.g.,inaneuralstructure
Page12
Theimportanceofsimulationasmethodologicaltoolfortheorydevelopmentincognitivescienceandforamoreprofoundconceptualunderstandingofcognition.
Amongothers,theseissueshaveimpactson
theevolutionofsemanticsandsymbolgrounding
thedesignofautonomoussystemsactinginthe"real"world,eitherasrobotsorsoftwareagents
theempiricalresearchin(cognitive)neuroscience(asfarastheirassumptions,experiments,andinterpretationofdatagoes)
thephilosophicalinterpretationofmodelsbeingproposedbycognitivescience
Inthefollowingsectionsadiscussionwillbegivenonthesepointsmentionedabove.Thisisdoneinordertounderstandthe"conceptualatmosphere"andthemotivationforthisvolume.
ReviewingtheNotionofRepresentation
Theunderstandingoftheconceptofrepresentationhasreceivednewstimulifromthedevelopmentsinconnectionism/computationalneuroscienceaswellasin(empirical)neuroscience.6Thefindingsandconceptsstemmingfromthesefieldsseemtoseriouslyquestionthetraditionalunderstandingofrepresentation.Thereareatleastthreepointswhichseemtobeofgreatimportance:
(a)givinguplinguistictransparencyinfavorofdistributedrepresentation
(b)givinguptheclassicalreferentialviewofrepresentationinfavorofagenerativeparadigmandtheconceptoffunctionalfitness
(c)theconceptsofembodimentandconstructioninknowledgerepresentation.
LinguisticTransparency
Theclassicalunderstandingofrepresentationislargelybasedontheideathatpropositions(e.g.,Fodor1981)representthe(internalandexternal)environmentinamoreorlesslinguisticallytransparentway.Withtheadventof"trulydistributed"connectionistmodels7,linguistictransparencyhadtobegivenupinfavoroftheconceptofdistributedrepresentationandsubsymbolicrepresentationinmicrofeatures(cf.Singer'scontribution,aswellasGelder1992,Hintonetal.1986,Rumelhartetal.1986,Smolensky1988,andmanyothers).
Onasuperficiallevel,givinguplinguistictransparencyseemstobeoneofthebigdisadvantagesofdistributedrepresentation(seealsoFodor&Pylyshin1988andmanyothers),asitimpliesthatwehavetogiveupthetraceabilityofthesystem'sdynamicsaswell.Takingacloserlookreveals,however,that(a)itisbynomeansclearwhyourbraindoesitsworkbymakinguseofthesamesemanticcategoriesasourlanguagedoes.(b)Whatisreferredtoasapropositionistheresultofextremelycomplexprocessesoccurringintheneuraldynamicsandleadingtotheexternalizationof"propositionalcategories"(e.g.,informofsymbols,language,etc.).Hence,itisnotatallclearwhytheprocessesresponsibleforgeneratingtheselinguisticcategoriesnecessarilyhavetobebasedonandhavetorelyonexactlythesecategories.Fromtheperspectiveofphilosophyofscienceitseemsratherquestionable,ifitisjusttouseinanexplanationofsocalledhighercognitiveabilities(suchaslanguage)thesamemechanisms/structuresbothintheexplanatorymechanism/explanans(e.g.,symbolprocessingmechanisms)andintheresultingbehavior/explanandum(e.g.,linguisticstructures).(c)Furthermore,itisknownfromempiricalneurosciencethattheactivityofmostneuronscannotbeexplicitlyrelatedtosemanticallytransparentphenomenaorevents.
Alltheseconsiderationsleadtotheconclusionthatthecriterionoflinguisticallytransparentrepresentationisperhapstheresultofourcommonsenseexperience,"autointrospection",andourcommonsenseassumptionsaboutrepresentation.Itseemsthatitarisesfromourneedtosomehowboxeverythingandeveryprocessintolinguisticcategories(and,thus,alsotheseprocesseswhichareleadingtothesecategories)inordertomakeitavailableforcognitivemanipulation.Oneoftheobjectivesofthisbookistoinvestigate,ifthereisanecessityoflinguisticcategoriesforexplainingcognitivephenomenaandtosearchforalternativeviews.
6Thetermsconnectionismandcomputationalneurosciencewillbeusedsynonymouslythroughoutthispaperforthepurposeoftheargumentstofollowthedifferencebetweenthesetwotermscanbeneglected.7"Trulydistributedrepresentation"ismainlyduetoweightconfigurationswhicharetheresultoflearningalgorithms,suchasthebackpropagationlearningrule.
Page13
ReferentialRepresentationinRecurrentArchitectures?
Thesecondandepistemologicallymoreimportantimplicationofconnectionistsystemsconcernstheveryconceptofrepresentation:duetothehighlyrecurrentarchitectureinthebrainweareconfrontedwithaphenomenonwhichquestionstheclassicalviewofreferentialrepresentation,i.e.,W=f(R):here,anystatewithinthecognitivesystemreferstoan(internalorexternal)environmentalstateinamoreorlessstablemanner(independentlywhetherithappensinalinguisticallytransparentmannerornot).Therecurrentarchitectureimplies,however,thattheneuralsystemfindsitselfinacertaininternalstateatanypointintime.8Thisinternalstatehasanindirectordirectinfluenceonthefollowinginternal/representationalstates,astheresultingtopdownprocessesarefeedingbackontheincomingstimuli.
Thisimpliesthattheenvironmentalstatedoesnotnecessarilydeterminetherepresentational(=internal)stateanymore,becausethepresentinternalstatehastobeseenassomekindofdispositionforthepossiblesuccessorstate.Moreprecisely,thepresentinternalstatedeterminesthespaceofpossiblerepresentationalsuccessorstatesandthecurrentenvironmentalstateonlychoosesoneoutofthem.Hence,adifferentpresentinternal/representationalstatedeterminesadifferentspaceofpossiblesuccessorstateswhichimpliesthatthesameenvironmentalstatemightleadintodifferentrepresentationalstates(seealsoPeschl1997).Inotherwords,asingleenvironmentalstate/eventcanberepresentedindifferentrepresentationalstates(dependingontheprecedentinternalstate).Thisimpliesthattheconceptofastablereferentialrelationshipofrepresentationcannotbefoundanymore.
Inotherwords,theclassicalideaofanenvironmentalstatedeterminingtheinternal/representationalstatehastobeforsakenduetothefeedbackinfluenceofpreviousinternal/representationalstatesintherecurrentneuralsystem.Theideaofamoreorlessstablerelationshipbetweenarepresentationalstateandan(internalorexternal)environmentalstatehastobeabandoned(forfurtherdetailsseePeschl1997).Rather,theinfluenceoftheenvironmentalinputhastobereducedtothemodulationoftheinternalrepresentationaldynamics(perturbationsPasdiscussedabove).Unfortunately,theimportanceofthisfarreachingepistemologicalissuehasnotbeenacknowledgedbyalargegroupwithinthecognitivesciencecommunity.
Representation,Construction,andGenerationofBehavior
So,iftheaimofrepresentationisnolongertomaptheenvironmentasaccuratelyaspossiblewehavetocharacterizeitasthegenerationofbehaviorintermsoffunctionalfitness(cf.contributionofvonGlasersfeldwhospeaksoftheviabilityofrepresentations).I.e.,behaviorwhich(i)facilitatestheorganism'ssurvival(inthebroadestsense)and(ii)functionallyfitsintotheparticularenvironmentalcontext.9Itisthereforenolongernecessarytosearchforneurons(orgroupsofneurons)whoseactivationscorrelatewithexternaleventsinastablereferentialmanner.
Astherepresentationalstructureistheresultofasystemrelativeconstructionprocess,itisnowonderthatweareexperiencingdifficultiesidentifying(traditional,referential)representationsinnaturalandartificialneuralsystems.Itseemsthatitissimplythewrongthingtosearchfor.Understandingrepresentationfromtheperspectiveofconstructivismandtheconceptoffunctionalfitness(e.g.,Glasersfeld1984,1995)givesusaclueastowhatwehavetolookforintherepresentationalsubstratumnamely,mechanismswhichallowthegenerationofadequatebehavior.Traditionallythesemechanismshavebeenthoughtofintermsofmanipulationsonreferentialrepresentations,however,thereisneitherneuroscientificnorepistemologicalevidenceinfavorofsuchaview.
Inthiscontextitisimportanttonotethatasimilarproblemarisesinmostapproachesinartificiallife.AsarguedinRiegler(1997),atypicaldeficiencyofmanyartificiallifemodelsisthePacMansyndrome:Simulatedorganismsinteractwithanthropomorphicallydefinedentities,suchas''food"and"enemy".Suchmodelsperformamereoptimizingtaskwhichyieldsamaximumgainofenergytogetherwithaminimumlossofhealth.Noattentionispaidto
8Thisalsoappliestofeedforwardarchitectureshowever,inthesearchitecturestheinternalstatedoesnothaveaninfluenceontheincomingactivationsbecausetheinternalstateis"shiftedout"ofthenetworkineachtimestep.9Recentdevelopmentsinagentbasedarchitecturesspeakofsituatednessofreactiveagents(cf.Clancey1997).
Page14
questionslike:Howhaveorganismsarrivedattheideathatsomethingisasourceoffood?Howdothey"know"thatanothercreatureisadangerousoponent?Predatorsdonotrunaroundwithalabelsaying"I'myourenemy".Evenifthiswouldbethecasehowwouldhavecognitivebeingslearnedtounderstandthemeaningofthoselabels?(Thinkofthedifficultiestounderstandsignsinacountrywhoselanguageandlettersyoudon'tknowatall.)However,thingsseemtobedifferentwithinmathematicalmodels.Ifwelookatmathematicalformulaewe(usually)knowwhatthemeaningofthelabels(variables)isalthoughwearenottheauthoroftheequations.E.g.,arrivingataresultofm=8.3weknowwhatmis,howmeaninggotattached(namelyaprioritothecalculation).Inlogicalcalculus,semanticsdefinesmeaningandtruthintermsofanunderlyingmodel,ontology,orlogicalinterpretation.Cognitivesciences,however,transcendthispurelysymbolicframework.Whileweasdesignerofartificiallifemodelswouldliketheupperleftpixelonthecomputerscreentobeafoodpillforthepixelinthelowerrightcornerrepresentingthecognitivecreature,thisisnotnecessarilytruefromtheperspectiveofthecreatureanditscognitiveapparatus.Thequestionregardingthephylogeneticandontogeneticemergenceofasystemrelativerepresentation,i.e.,''meanings"fortheorganisms,isnottouched.
Furthermore,itisimportanttonotethatthedynamicsofarecurrentcognitivearchitectureneednotnecessarilybeimplementedinatypicalneuralnetworkmanner.Riegler(1994,1997)describestheimplementationofarulebasedsystem:thealgorithmoperatesexclusivelyonasetof"internalstatecells"(havingnoexplicitlinguisticreference)ratherthanonanthropomorphicallypredefinedsensorandmotorstates.Asthispreservestheideaofoperationalclosure(MaturanaandVarela1979)itis,too,animplementationofcognitiveselfreferenceandhencetranscendspurelyareferentialrepresentation.Thus,weconcludethatthereisnoepistemologicaldifferencebetweenrulebasedsystemsandconnectionistapproachesaslongasweavoidforcingareferentialrepresentationscheme.
Embodiment,Construction,andDynamicsofKnowledge
Inthecontextofthesequestionsonecanobserveashiftininterestsinthefieldofknowledgerepresentationinthelastdecade:thefocushaschangedfromtryingtocaptureordepictenvironmentalstructuresstaticallyontotherepresentationalstructuretowardsstressingmorethequestionofthegenesis,development,anddynamicsofknowledge.Connectionistapproaches(i.e.,theirfocusonlearningstrategies),geneticalgorithms,andthecombinationofboth(e.g.,Elmanetal.1996,Cangelosietal.1994,andmanyothers)hadacrucialimpactonthedevelopmentofrepresentationmechanismsmodelinglearningandthedynamicalaspectofknowledge.Furthermore,resultsfromempiricalneuroscience(sometimeshavingbeentriggeredbycomputationalapproachesandconcepts)havebroughtaboutabetterunderstandingofthelearningmechanismswhichareresponsibleforthedynamicsofknowledgeinourbrains10.
Oneofthemostimportantepistemologicalimplicationsofthisdevelopmentandtheabovediscussionsistheinsightthatknowledgeistheresultofanactiveconstructionprocessratherthanofamoreorlesspassivemapping(cf.Sjlander,thisvolume).Inthisview,theorganismactivelyextractsandconstructstheseenvironmentalregularitieswhicharerelevantforitsparticularsurvival.Theenvironmentnolongerinstructsordeterminesthestructureoftherepresentationsystem,butonlyplaystheroleofconstrainingtheconstructionprocesses.Inotherwords,theknowledgecanbefreelyconstructedaslongasitdoesnot"violate"theenvironmentalconstraints.Theresultisasystemrelativerepresentationofknowledge(aboutenvironmentalregularities),whereW=f(W,E,P),aspresentedintheintroduction:representationdoesnotneedrealityasaninstructiveinstance!Hittingaflywithaflapyieldsadifferenteffectthanbeatinganelephantwiththesameflap:Representationdependsonthestructureofthecognitivesystemratherthanonoutsideentities.
Inthiscontextitseemsthatweneedtotaketheconceptofembodimentofknowledgemoreseriously.Knowledge(representation)canatleastinneurallybasedcognitivesystemsnolongerbeunderstoodassomethingabstractandcompletely
10Unfortunately,wearestandingonlyattheverybeginningtofullyunderstandtheseprocesses.However,thebasicprinciples(e.g.,longtermpotentiation/depression(LTP,LTD),Hebb'sconcepts,etc.)seemtobequitepromisingandhavebroughtaboutanewunderstandingofknowledge.
Page15
detachedfromthe(neural)substratum.Rather,wehavetomaketheeffortandtrytounderstandneuralstructures,architectures,anddynamicsintermsofcontributingtotheproductionoffunctionallyfittingbehavioronanonreferentialbasisofrepresentation.Inotherwords,thesocalledenvironmentalregularitiesarenotstoredexplicitlyinthestructureofthesynapticweights.Aparticularneuralarchitecturehastobeunderstoodastheresultofalongphyloandontogeneticadaptation/constructionprocesswhichaimsatrelatingtheorganismrelevantenvironmentalregularitieswiththeorganism'srequirementsforproductionofbehaviorensuringitssurvival.
Attheheartoftheseconstructionprocesses,neurallearning,adaptation,andplasticity,aswellasphylogeneticprocessescanbefound.Alltheseprocessesarethesubstratumforanylearningdynamicscontinuouslyoccurringinacognitivesystem.Consequently,knowledge(representation)inacognitivesystemcannotbeconsideredsomethingstatic,buthastobeseenasahighlydynamicalprocesscontinuouslyadaptingtothechanging(internalandexternal)environmentalconstraints.Oneoftheintentsofthisvolumeistostudyexactlythisrelationshipbetweentheneurophysiologicalprocesses(oflearningandadaptation)anditsepistemologicalimplications.
SimulationasMethodologicalTool
Thelastyearshaveshownasharpincreaseintheimportanceofthemethodofsimulationinthecontextoftheorydevelopmentincognitivescience.Theextensiveuseofsimulationbringsaboutawholenewmethodologicalapproachanddynamicsindisciplineswhichformerlywereworkingalmostexclusivelyempirically,suchasneuroscience,psychology,biology,andphysics.Theinterestinginsightswhichareachievedbysimulationarenotsomuchresultsaboutdetails,butconcernconceptualknowledgewhichcanbeusedasinputandstimulationforbothempiricalandepistemologicalinvestigations.
Oneofthemainpurposesofpsychology,(cognitive)neuroscience,linguistics,andmanyother"cognitivedisciplines"hasalwaysbeenabetterunderstandingofsocalledcognitiveprocesses.Mostoftheresultingapproachestocognitionwerebasedonempiricalinvestigationsand/ormoreorlessspeculativeandcommonsenseinterpretationsofcognitivephenomena.Progressinempiricalsciencesisbasedonacontinuousprocessofconstruction,negotiation,andadaptationtothe"empiricaldata".Thetargetofthisprocessistoreachastateof(epistemological)equilibriuminwhichthetheoryfitsintotheenvironmentaldynamics,meaningthatthetheoryatleastpredictstheenvironmentaldynamicscorrectlywithinsomemarginoferror.Oftenthecomplexityofcognitiveprocessesandtheirsubstratumdoesnotmatchthecomparablypoorempiricalapproachesandunderstandingofcognitivephenomena(cf.Dorffner'scontributionwhichstressestheimportanceofconnectionismasahelpfulmodelingframeworktounderstandcognition).Therefore,muchroomisopenedupforratherspeculativeconceptsinthisfield.
Fortunately,thesimulationmethodintroducesanewdimensiontocognitivescienceand,morespecifically,tocomputationalneuroscience/connectionism.Simulationmodelsareespeciallyinterestinginthecontextofcognitiveneuroscience,asitsempiricalresultsandtheoriesaresometimessorichindetail(e.g.,dataonthereleaseofneurotransmitter,theoriesonamolecularlevel,etc.)thatitisalmostimpossibletorelatethemtocognitivephenomena.Inotherwords,thereisanexplanatorygapandastrongtensionbetweentheepistemologicallyinspiredquestionsoncognition(e.g.,aboutknowledgerepresentation)andtheempiricalandhighlydetailedresultsfromneuroscience.Inthiscontexttheconnectionistapproachinthebroadestsenseplaysacrucialroleasmediator:itstandsbetweenthetwopolesoftheratherspeculativeepistemologicaltheoriesandtheempiricallygroundedneuroscientificdetailsandinmanycasesmakesthemcompatible.Thiscompatibilityisachievedbythetrickoffocusingontheconceptuallevelofneuralprocesses.Bydoingso,themostimportantcharacteristicsandstructuresofneuralsystems,suchasparallelprocessing,networkarchitectureandmassiveconnectivity,anddistributedrepresentation,arecapturedinamoreorlesssimplifiedcomputationalmodelwhosedynamicscanberelatedtoandisdirectlyrelevantforepistemologicaland"cognitive"issues.
So,whydowestresstheimportanceofsimulationmodelsofcognitioninthisbook?Itisnotsomuchthetechnicaldetailsofsimulationwhichweareinterestedin,butratherintheconceptualimplicationswhichthesemodelshaveontheproblemofknowledgerepresentation.Hence,oneoftheobjec
Page16
tivesofthisvolumeistoshow,howfocussingonexactlythisconceptuallevelcanbringaboutbothanempiricallyandepistemologicallysoundunderstandingoftheancientproblemofrepresentationincognitivesystems.Furthermore,includingsimulationtechniquesasanecessarytoolfortheoryconstructioncanguideempiricalresearchnotonlyontheleveloftechnicaldetails,butandthisseemstobeevenmoreimportantonaconceptuallevel(e.g.,concerningtheassumptions/premisesofaresearchstrategy,theepistemologicalframeworkandfoundations,etc.).
Conclusion
Wehavearguedinfavorofcompletelyrethinkingbasicissuesincognitivescienceinthecontextofrecentdevelopmentsinthisfield.Themainissueseemstobethequestionofknowledgerepresentation,whichchangesdramaticallywiththeadventoftheconceptsofconnectionismandartificiallife.Wehavesuggestedthattheconceptofareferentialunderstandingofrepresentationshouldbereplacedbyasystemrelativeformofrepresentationwhichisnotnecessarilysemanticallytransparent.
Wearguedfromanepistemologicalandneuroscientificperspectivethatthetaskofgeneratingbehaviorismoreimportantthantheaccuratemappingofenvironmentalstructurestorepresentationalstructures.Itisbynomeansclearwhatthe"pointofreference"couldbeforan"accuratemapping".Isitourownperceptionandconceptualizationoftheworld,orthatofarat,ortheworlditself,...?Oneistemptedtoassumethattheoutsideworld(inthesenserealitt)actsassomekindofconstraintforourconstruction/representationprocesses.Butassumingtheexistenceofthatrealitt(orpartsofitsuchaselectromagneticwavesseeabove)wouldputusintothecampofreferentialrealistsaswe(thoughnottrivially)mapourexperienceontotheideaofanoutsideworld.Assumingthenonexistenceoftheworld,onthecontrary,wouldstigmatizeusassolipsists.Asasolutionwehavethereforetoacceptthatclaimsabouttheexistenceofanobjectiverealittmightnotbenecessaryforthepurposeofscientificexplanations.
Wehaveseenthatsuchaviewofknowledgerepresentationiscloselyrelatedtoconstructivistconcepts.Inthisapproachtoepistemology,theonlycriterionforsuccessfulknowledge(representation)isitsfunctionalfitness.Furthermore,ithasbecomeclearthatknowledgeisnotastaticstructure,butiscontinuouslychanging.Thisdynamicscanbedescribedasaprocessofconstructionandadaptationandfindsitssubstrateintheneuraldynamics/plasticity.
Themaingoalofthisvolumeistodiscussthesefundamentalshiftsincognitivescienceandtosketchtheimplicationsonanepistemologicalandmethodologicallevelforcognitivescienceanditsrelateddisciplines.
References
Cangelosi,A.,Parisi,D.&Nolfi,S.(1994)Celldivisionandmigrationinagenotypeforneuralnetworks.Network:computationinneuralsystems5(4):497516.
Clancey,W.J.(1997)SituatedCognition.OnHumanKnowledgeandComputerRepresentations.NewYork:CambridgeUniversityPress.
Dreyfus,H.L.&Dreyfus,S.E.(1988)MakingaMindversusModellingtheBrain:ArtificialIntelligenceBackataBranchPoint.ArtificialIntelligence117:30933.
Elman,J.etal.(1996)Rethinkinginnateness.Aconnectionistperspectiveondevelopment.Cambridge,MA:MITPress.
Fodor,J.A.(1981)Representations:philosophicalessaysonthefoundationsofcognitivescience.Cambridge,MA:MITPress.
Fodor,J.A.&Pylyshin,Z.W.(1988)Connectionismandcognitivearchitecture:acriticalanalysis.Cognition20.Reprintedin:Beakley,B.etal.(eds.)(1992)Thephilosophyofmind.Cambridge:MITPress.
Gelder,T.von(1992)Defining"distributedrepresentation".ConnectionScience4(3/4):175191.
Glasersfeld,E.von(1984)Anintroductiontoradicalconstructivism.In:Watzlawick,P.(ed.)Theinventedreality.NewYork:Norton,pp.1740.
Glasersfeld,E.von(1995)Radicalconstructivism:awayofknowingandlearning.London:FalmerPress.
Hinton,G.E.,McClelland,J.L.&Rumelhart,D.E.(1986)Distributedrepresentations.In:Rumelhart,D.E.&McClelland,J.L.(eds.)ParallelDistributedProcessing:explorationsinthemicrostructureofcognition.Foundations,VolumeI.Cambridge,MA:MITPress,pp.77109.
Kuhn,T.S.(1970)TheStructureofScientificRevolutions(2nded.)Chicago:Univ.ofChicagoPress.
Page17
Maturana,H.R.&Varela,F.J.(1979)AutopoiesisandCognition:TheRealizationoftheLiving.Boston:Reidel.
Peschl,M.F.(1997)TheRepresentationalRelationBetweenEnvironmentalStructuresandNeuralSystems:AutonomyandEnvironmentalDependencyinNeuralKnowledgeRepresentation.NonlinearDynamics,Psychology,andLifeSciences1(2):99121.
Riegler,A.(1994)ConstructivistArtificialLife:Theconstructivistanticipatoryprincipleandfunctionalcoupling.In:Hopf,J.(ed.)Proceedingsofthe18thGermanConferenceonArtificialIntelligence(KI94)WorkshoponGeneticAlgorithmswithintheFrameworkofEvolutionaryComputation.MaxPlanckInstituteReportNo.MPII94241,pp.7383.
Riegler,A.(1997)EinkybernetischkonstruktivistischesModellderKognition.In:Mller,A.,Mller,K.H.&Stadler,F.(eds.)KonstruktivismusundKognitionswissenschaft.KulturelleWurzelnundErgebnisse.Wien,NewYork:Springer,pp.7588.
Roth,G.(1994)DasGehirnundseineWirklichkeit.Frankfurt:Suhrkamp.
Rumelhart,D.E.,P.Smolensky,McClelland,J.L.&Hinton,G.E.(1986)SchemataandsequentialthoughtprocessesinPDPmodels.In:McClelland,J.L.&Rumelhart,D.E.(eds.)ParallelDistributedProcessing:explorationsinthemicrostructureofcognition.Psychologicalandbiologicalmodels,VolumeII.Cambridge,MA:MITPress,pp.757.
Smolensky,P.(1988)Onthepropertreatmentofconnectionism.BehavioralandBrainSciences11:174.
Stadler,M.&Kruse,P.(1990)berWirklichkeitskriterien.In:Riegas,V.&Vetter,C.(eds.)ZurBiologiederKognition.Frankfurta.M.:Suhrkamp,pp.133158.
Winograd,T.&Flores,F.(1986)UnderstandingComputersandCognition:ANewFoundationforDesign.Norwood,NJ:Ablex.
Page19
OverviewofContributions
Thefirstpapersaddresstheepistemologicalstatusofrepresentationandpossiblecategorizations.Theserathergeneralconsiderationsaboutrepresentationarefollowedbycontributionsthatfocusoncomputationalapproaches,especiallyonhowmeaningcanbesimulatedinartificialdevices,andhowmeaningcanbegrounded.Solutionsareofferedonvariouslevels,amongthemsymbolgroundingandsystemtheoreticalconsiderationstorepresentation.Variousdisciplinesindependentlyarriveatemphasizingtheimportanceofactionsforrepresentationsandthenecessitytoclosethesensorymotorloopexamplesfromtheperspectiveofcomputationalapproaches("embodiment"intoanenvironment),theneuroscientificperspective("topdown"processes),thepsychologicalperspectiveandothersarepresented.Severalpapersaddresstheroleofsocialinteractionsinparticularlanguageasameansofstabilizingsystems,andintheemergenceofmeaning.Thedeepepistemologicalimplicationsconnectedwiththeaboveconsiderationsarediscussedinthefinalchapterofthevolumededicatedtoconstructivistapproaches.Followingtheideaofatrueinterdisciplinaryapproach,thecontributionstothevarioussectionsareorganizedexclusivelywithrespecttothematiccontentratherthantoscientificdiscipline.Theremainderofthisoverviewpresentsthepapersandtheirmutualrelationshipinmoredetail.
Inthefirstchapter,dedicatedtotheoreticalconsiderations,GeorgDorffnerprovidesanintroductiontotheproblemofrepresentationbydefiningthreetypesofrepresentation.Heshowsthatconnectionismsolvesseveralproblemsofrepresentationandactuallyhelpstoabandononeofthesetypesofrepresentation.Similarily,AlfredoPereirainvestigatestwotypesoftherepresentationsusedintheneurosciences,perceptualandexecutiveprocesses.AnthonyChemerotriestoclassifypossiblecritiquesregardingrepresentations.Hepresentstwotypesofantirepresentationalism:eithertherepresentingstructurecanbequestioned("empiricalantirepresentationalism"),ortherepresentedstructurecanbequestioned("metaphysicalantirepresentationalism").Evidenceforthefirstcomese.g.fromconnectionismwhichhasdemonstratedthatbrainlikestructurescanfunctionwithoutexplicitrepresentations.Argumentsforthesecondareofaphilosophicnaturedenyingthepossibilityofspeakingoffixedentitiesoftheworld,and,instead,proposingthatthe"things"aredefinedonlywithrespecttotheiruserandthuscannotbe''mapped"ontothebrainofthatuser.MatthiasScheutzinvestigatestheontologicalstatusofrepresentations,andquestionswhethertheyareentitiesontheirownhearguesthatwhethersomethingcountsasrepresentationofsomethingelseisdependentonthelevelofdescription.Bytalkingaboutamentalstructurerepresentingsomethingintheworld,oneimpliesthatthesetwothingsexistontwodifferentontologicallevels.Thisisnottrue,however,sincethemindandthethingsoftheworldarebothentitieswithinourcognitiveexperience.Thus,theproblemofrepresentationarisesonlyasartifactofourdescribingthephenomenonofcognition.DanielHuttoexploresthedifferencebetweennonconceptualrepresentationsandconceptualrepresentationsandasksthequestionwhetheritmakessensetodefinerepresentationonthelowestlevelofabstractionoronthehighestlevel.
Isitpossibletosimulatemeaningorcognition?Theissueofcomputationtorepresentationispresentedinthechapter"ComputationalApproaches".ThepaperbyGeorgSchwarzdrawsconclusionsforthefunctioningofthebrainfromacomputationalperspective.Fromaphilosophicalpointofview,WilliamRobinsontriestoaddressacontroversyincomputationalapproachestocognitionbyintroducingadistinctionbetweencognitionandcognitiveabilities.RobertFrenchmakestheargumentthatcreatingrepresentationscannotbeseparatedfrommanipulatingthem.Moreover,heemphasizesthecontextdependencyoflinguisticterms(oneofthereasonsthatcomputationallinguisticsisnontrivialmatter)and,asaconsequence,thedistractivecharacterofcontextladenrepresentations.AndrewCowardproposesthatnaturalpressureshaveresultedinbiologicalbrainshaving
Page20
simplefunctionalarchitectures.Thetypeofarchitectureconstrainsthetypeofrepresentationswhicharepossible.
WhileconnectionismovercamesomeoftheproblemsderivedfromconsideringthebrainasavonNeumanComputer,seriousnewquestionscameintofocus.Withsimulatingrepresentationsinneuronalnetworksitbecameevenmoreobviousthattheproblemtobesolvedwasthatthemeaningwasputintothesystembytheengineer.Whetherthestateofthenetworkwouldstandforsomethingisdefinedbyanexternalobserverbutisnotinherenttothesystem.Thisproblemofthefoundationofmeaningisreferredtoasthegroundingproblem,whichistreatedinthechapteron"SymbolGroundingandLanguage".TomZiemkeprovidesareviewaboutthegroundingproblemandit'stwoproposedsolutions.Thefirstsolutionofgroundingisgroundingasymboltoaninput,whichwasfirstproposedbyStevanHarnad.ThepapersbyNathanChandleretal.andChristianBalkenius&SimonWinterpresentsuchmodelsongroundedlanguagesystems.AsdescribedinthepaperbyZiemke,this,however,isonlyapartialsolutiontherealbreakthroughisonlyobtainedwhenarobotisreallyembeddedintoanenvironmentandinteractingwithit.ExamplesofthiskindofgroundingaregivenbyMarkWexlerandRalfMller(seebelow).
NotonlyinclassicalAIbutalsoinconnectionism,meaningandrepresentationsaremostlyofastaticnature.Thus,althoughdistributed,representationsarestillverymuchtreatedlike"boxes".Cognition,however,isdynamicanditneedstime(seealsodiscussion,page290).Systemstheoryprovidesaframeworkthatencompassestheseissues(chapter"CognitionasaDynamicalSystem").Fromapsychologicalpointofview,PimHaselagerprovidesareviewabouttherelevanceofdynamicalsystemstheorytothetopicofrepresentationandits(philosophical)classificationinthecontextofbehaviorism.MariusUsher&ErnstNieburpresentneuroscientificevidenceandamodeldemonstratingthatneuralrepresentationsareactiveprocesseswhichcanmediatecontextualcomputationandbindrelationalproperties.KenMogidiscussestheconceptoftimeandcausalityintherelationofperceptionandneuralfiring.Hisprincipleofinteractionsimultaneityprovidesanexplanationfortheoriginofsubjectivetime.SteveBresslerpresentsacomplexsystemstheoryapproachtorepresentationandcognition.Inspiredbyempiricalfindingsoflocalandlargescalecorticalinteractions,hederivesadynamicalviewofcorticalrepresentationbasedoninterarealpatternconstraintsandlargescalerelaxation.
Inarelatedrealm,awayfromrepresentationsunderstoodasprojectionsofenvironmentalentitiesontomentalstructures,severaldisciplinesindependentlydiscoveredtheimportanceofactionsforrepresentations.Theyproposetoclosethesensorymotorloop.(Forthistopicofparallelismsinscientific/philosophicaldevelopments:seethecommentbyWolfSingerinthediscussions,page293).Inmodeling,thegroundingproblemdescribedaboveintroducedtherelevanceofinteractionforrepresentationsembodiedsystemsthatinteractwiththeirenvironmentinasensorymotorlooparecreated.Philosophically,constructivismandthetheoryofautopoieticsystemsdemonstratedthatwhatweperceiveasrepresentationsofapredefinedexternalworldisbetterdescribedasconstructionsthatservetosuccessfullygeneratebehavior.Inneuroscience,theoldconceptthatenvironmentalobjectsaremappedontoneuronsinafeedforwardmannerischallengedbyanatomicalandphysiologicalevidence.Ithasbeenshownthatinformationdoesnotexclusivelyflowfromthesensorstothehigherbrainareas,wheretherepresentationsarethenthoughttobeusedfor"thinking"andthefinalactivationofmotorprograms("bottomup").Rather,massiveactivityistransportedtheotherwayaround(''topdown").Thusmotoractivity,i.e.,plannedactions,canbeviewedassupervisingtheincomingsignalsandthusstructuringcorticalrepresentationsthemselves.Psychophysicalevidence,suchasthefigurevaseambiguity,showthatperceptionisambiguous,dependentonprimingandexpectancy.Thus,severalapproachescometotheconclusionthatperceptionisnotapassive,feedforwardprocessofmappingbutanactiveconstruction,wheretheneedforactioncreatesanhypothesisabouttheenvironmentwhichisthencomparedtotheincomingsensorysignals.Thechapter"RelevanceofActionforRepresentation"isdedicatedtotheseproblems.
Theideaof"theinversionofsensoryprocessingbyfeedbackpathways"wasintroducedbyErichHarth.Inhispaper,hedescribesneurophysiologicalevidenceforascenariowhereprocessingisnotfromtheperipherytothehigherbrainareasbutrathertheotherwayaround.Hepresentsamodelwhereacorticalhypothesis("internalsketchpad")mightbecontrolledbytheincomingsignalsina
Page21
hillclimbingalgorithm.Fromtheperspectiveofneurobiology,AstridvonSteininterpretsthecorticalarchitecturewithitsfeedforwardandfeedbackconnectionsasamediumofinterwovensensorymotorloopsinteractingondifferentlevelsofahierarchy.Representationsdonot"sit"oneitherendoftheconnections,butarestablestatesinthesesensorymotorloops.ThisisconsistentwithPiaget'sfindingthatanobjectisdefinedthroughtheperceivedchangeinsensationthatanactionperformedontheobjectinduces.Accordingtothisapproach,whatweperceiveasentitiesisneitherintheworldnorinthebrainbutintheinteractionbetweenboth.Basedontheideaof"perceptionthroughanticipation",RalfMllerperformsasimulationwithasimilarapproach:hepresentsasystemwheretherepresentationisnotabuildingblockwithinthesystem,withperceptionbeingaprojectionfromoutside,butwhereperceptionisaprocessofanticipatingthesensoryconsequencesofactions.TomWexlergivesempiricalevidencefortheimportanceofactionsinrepresentationsusingbothpsychophysicalexamplesandsimulation(representationasstrategyforactingratherthanasamappingoftheenvironment).Hedemonstratesexampleswherepreviouslyunresolvedproblems(e.g.the"nparityproblem'')canbesolvedifthelearningsystemisdirectlyinteractingwiththeenvironmentinasensorymotorrelation,i.e.,iftheenvironmentisincludedintotherepresentation.
Whenafirststepindefiningrepresentationsonanewgroundistoconsiderthesensorymotorcouplingwithanorganisms'senvironment,asecondstepistoconsiderthecouplingnotonlywithastaticenvironmentbutwithotherrepresentingsystems.Theroleofsocialinteractionsinparticularlanguageforstabilitieswithinasocioculturalnetwork,andtheemergenceofrepresentationsand"meaning"isasubjectwhichisfurthertreatedbyseveralauthorsinthechapter"CommunicationandSocialCoupling".Fromasocialpsychologicalpointofview,RichardEiserasksthequestionhowindividualsacquireknowledgeorrepresentationofsocialrealitythatadaptwithnewexperience,andalsohowgroupscancoordinatetheirbehaviorinamanneradaptivetotheirenvironment.Hepresentsatheoryabouttheself,wheretheselfisseenastheproductofthecognitivesystematwork,i.e.,asexplanatorymodelforalreadyperformedortobeperformedbehavior.TomRoutendemonstratesthattheadvantageoftheadaptivebehaviorapproachoverclassicalAIisonlyapartialone,andthatlanguagehastobeincluded.HeintroducestheideasofBourdieu,andMcDowell,whotakelanguagenotasastaticsetofwordsandgrammaticalrulesbutasa"wayoflife",aHabitus.Habitus(Bourdieu)is"asystemofdispositionswhichservestosettheindividual'sunderstandingandactionsinagreementwithitssocialandphysicalenvironment".SinceHabitusisinstalledbysocialization,itwillbethesamewithinagroup,i.e.,theexperiencingwillbesimilarandacommonworldviewwilldevelop.Thus,languageasaHabitathastobeincludedintomodelsofcognitiveagents.Aprototypicalexampleofrepresentation,languageandsemanticsisgivenbyPeterGrdenfors.Incontrasttorealisttheoriesofsemanticshesuggeststhatitisnotreasonabletospeakabouttherelationbetweensignsandthingsintheouterworld.Instead,hearguesthatitmakessensetospeakabouttherelationbetweensigns(i.e.,"linguisticexpressions")andconcepts,i.e.,meaningintheheadoftheindividuals.Thatlinguisticexpressionsmeanthesameforeverybodyisnotbecausetheymapafixedworldbutbecausetheyemergeduringinteractionfromtheconceptsintheheadsofthelanguageusers.ExamplesofneurophysiologicalcorrelatesoflanguageprocessingaregivenbySabineWeissetal.TheproblemoftransferringmeaningbetweentwolanguagestheproblemoftranslationisconsideredbyHannaRisku.
Tillnowwereferredtorepresentationasmentalenvironmentalstabilitiesthathaveevolvedtosuccessfullyguidethebehavioroforganismsobjectswouldinthissensebeconsideredasadaptiveconstructsinthisorganismenvironmentrelation.This,however,doesnotaddressthequestionofthesubjectiveexperienceofrepresentations,i.e.,theawarenessthattherearethingsintheoutsideworld,ortheawarenessofonesselfasactivepartintheinteractionwiththeworld.Thechapter"QualitativeAspectsofRepresentationandConsciousness"dealswiththisproblemofconsciousness.WolfSingeraddressestheissueofconsciousexperiencefromaneuroscientificperspective,endingwiththeconclusionthatpropertiessuchasselfawarenessandconsciousnesscannotbeexplainedintermsofneuronalcorrelatesalone,butonlybyadditionallyconsideringthe(social)interactionsamongbrains.Hispositionis,thus,bothdualisticandmaterialistic,sayingthatconsciousnessishousedbyaneuronalsubstratebutcanneverbeexplainedbyneuroscienceitselfsince
Page22
itconsidersaphenomenonwhichcannotbefoundinsinglebrains(Seealsodiscussions,page291).MichaelPauenstressestheimportanceofstructuralsimilaritiesinrepresentations:differencesintheworldhavetoscaletodifferencesintheneuronalactivities.Qualiaisproposedtoserveastheaspectorganizingthesestructures.Theproblemofqualiadissolvesifoneabandonstheartificialdistinctionbetweensubjectiveexperienceandobjectiveoutsideworld.
Finally,thechapter"Constructivism"concludeswiththedeepepistemologicalimplicationstheaboveconsiderationspresent.Ifweagreethatwhatisrepresentedinthebrainisnotjustaprojectionfromentitiesintheoutsideworldbutratherarepertoireofhypothesisaboutfutureactions,groupedandstructuredinanefficientwaytosuccessfullyactwhat,then,isthatrealityoutthere?
ErnstvonGlasersfeldhasaradicalapproachtothatquestion:hearguesthatwhatweperceiveasoutsideworldisnothingbutconstructionsofourcognitiveapparatus.Inhisarticleheproposestonotusethewordrepresentationatall,sinceitiscommonlyunderstoodasa"copy"or"reproduction"suggestingthereferencetoanoriginal.Whilethisimpliesthatwecancompareamentalexperiencetosomethingoutsideourmentalexperiencetheoutsideworldhearguesthatthisisnotpossiblesincementalcontentcannotrefertoanythingbuttofurthermentalcontent.Hesuggeststousetheword"presentation"instead,toimplythatitisaproposalthebrainofferstotheoutsideworldinadaptingtothatworldaworld,however,thatinprincipalcannotbeknown.EmpiricalevidencefortheconstructivistnatureofrepresentationispresentedbySverreSjlander.Heinvestigatesthesubjectfromanevolutionaryperspectiveanddescribeshowobjectcentered(invariant)representationsdevelopfromsimpleinputoutputprocessinginloweranimals.Heshowsthattheevolutionarypurposeofsensoryprocessingisnotamappingoftheworldbutproducingadaptivebehavior.AnnikaWallinshowsthatthefactthatwehavetwotypesofrepresentationsimmediateperceptionsandrepresentationssuchasduringimagerycreatesaseriousproblemforconstructivism:accordingtoconstructivisttheories,theconceptsconstructedtosubserveadaptivebehaviorareadjustedaccordingtotheconstraintspresentedtothesystem.
Asthisoverviewshows,ourgoalwastobringtogetherinterdisciplinarilyworkingscientistsinordertocontributetothisvolume.Webelievethatourselectionofpaperswillhelpfosterabetterunderstandingandfruitfulapplicationoftheubiquitousnatureofrepresentation.
Page23
TheConnectionistRoutetoEmbodimentandDynamicism
GeorgDorffnerAustrianResearchInstituteforArtificialIntelligence,andDept.ofMedicalCyberneticsandArtificialIntelligence,UniversityofVienna,AustriaEmail:georg@ai.univie.ac.at
AbstractInthispaperIdemonstratethatconnectionismis,orcanbe,largelyinlinewithmostrecenttrendsincognitivescience.Thecoreofmyargumentisadistinctionofseveraltypesorusesofrepresentationincognitivescience.Idemonstratehowconnectionismhelpsinabandoningoneofthesetypesthenotionthatrepresentationisamirrorofanobjectivelyexistentworldwhilemaintainingtheimportantothertwotypescausalcorrelatesofphysicalstates,andinternalmentalstatesstandinginforpastperceptions.Withthisdistinctionthequestion"Doesrepresentationneedreality?"canbeanswered.Ifurtherdepictaconnectionistroutetoembodiedandsituatedcognitivemodels,astheyareputforwardbyrecentcognitivetheories.Afterashortdiscussionofconnectionism'srolefordynamicisttheoriesofcognition,Iconcludethatmuchofcurrentconnectionistresearchishighlyrelevanttomoderncognitivescience,evenifthemodelsareapparentlyremotefromtrulyembodiedorsituatedones.
NewTrendsinCognitiveScience
Inrecentyears,anumberofnoveltrendshaveemergedincognitivescience,whichalthoughrunningbydifferentnamesandunderdifferentheadingsappeartohaveacommondenominator.Theyputthebodybackintothefocusofcognitivescience,emphasizetheinteractionbetweenacognitiveagentanditsenvironment,viewexperienceandindividualdifferencesasatleastasimportantasdetachedlogicalprinciples,andapplyamoreholisticstancetowardresearchonhumancognition.Thesetrendsarecharacterizedbykeywordssuchas
embodiment
situatedness
emergence
grounding
selforganization
nonrepresentationalism
dynamicism
andmanymore.Theyhavedecisivelyinfluencedthewaycognitivescientiststhinkaboutintelligenceandcognitivebehavior,andsomeclaimthattheywillmainlyshapecognitiveresearchyettocome.
Examplesareworkonautonomousrobots(Brooks1991Pfeifer&Verschure1992Nehmzow&Smithers1991andothers),constructivisttheoriesofcognition(e.g.Varelaetal.1991),dynamicmodelsofcognition(Port&vanGelder1995Skarda&Freeman1987andothers),workongrounding(Harnad1990),andothers.
Onequestionmightariseinthisdiscussion:Whateverhappenedtoconnectionism(Rumelhart&McClelland1986,McClelland&Rumelhart),whichinthelate1980swasseenassomethingclosetoarevolutionincognitivescience(Smolensky1988)oratleastamajorthreat(Fodor&Pylyshyn1988).Whatrolecanordoesconnectionismplayinthesechangingtimesincognitivescience?Hasitbecomeanestablishedmodelingmethodology,whichhasrunintosimilarshortfallsaspreviouscognitivistapproaches?Isitjustanewversionofclassicaltheoryviewingcognitionascomputationonstylizedrepresentationsoftheworld?Bickhard&Terveen(1995)seemtoargueforthelatter,similarlytoVarela(1990)whodistinguishesfourphasesof20thcenturycognitivescience,connectionismbeingthethird,replaced(orextended)bythelatestemergentistphase.
InthispaperIarguethatconnectionismisfarfrombeingreplacedbyanyofthelatesttrends.Tothecontrary,itcanbeamajormodelingframeworkthatencompassesmostorevenalloftheideasexpressedbytheabovekeywords.Oratleast,itcanprovidethepavementforaroutetowardrealizingthoseideas.Take,forinstance,thefollowingimag
Page24
inarydialogusedbyElmanetal.(1996)astheintroductiontotheirbook.Onepersoniswonderingaboutanotherperson'sinterestinconnectionism:
"...Ithoughtyoubelievedinconstructivism,interactionism,epigenesisandallthatmurkystuff."
"Oh,I'mstillabeliever!Buttheconnectionistframeworkallowsmetocomeupwithamuchmoreprecisenotionofwhatallthatstuffreallymeans..."
Thisisexactlythespiritofthispaper.Iwanttodemonstratethatconnectionismcancontributemuchtothelatesttrendsincognitivescience,bydiscussingtwoimportanttopics:Representation,it'suse,misuseandunderstandingincognitivescienceandhowconnectionismcancontributetoclarifyingthings.Andthedynamicistviewofcognitionandhowconnectionismrelatestoit.
ThreeMeaningsofRepresentation
Likemostepistemologicaltermsincognitivescience,'representation'hasbeenusedandabusedinmanydifferentwaysinliterature.Whilethenotionplaysacentralandunambiguousroleinclassicalcognitivistapproaches,ithasbeenmodified,analyzedandcontestedinmorerecentstreamsofcognitivescience.AuthorslikeBrooks(1991)andVarelaetal.(1991)havegoneasfarasspeakingaboutcognitionwithoutrepresentation.Thisleadstoanapparentcontradiction.If,asitiswidelybelieved,representationisseenasmediatingbetweentheimmediatelyperceptibleenvironmentandasstandinginforit(Markman&Dietrich1998),thuspermittingcomplexcognitivebehavior,howcancognitionworkwithoutit?Ontheotherhand,representationthewayitisusedinartificialintelligencebearsawiderangeofconceptuallimitations,whichwasoneofthemajormotivationsforcognitivescientiststomovetowardsuchnotionsasembodiedcognition,groundedrepresentations(yes,representationsagain),andthelike.
Now,isconnectionismnonrepresentational,orcanitbe?Ifnot,whatisdifferentaboutitandhowdoesitrelatetoembodiedmodelsofcognition?
Theanswerliesinaclarificationofthenotion'representation'.Idistinguishthreequitedifferenttypes(orbetterstill,threeuses)of'representation',anditwillbecomeclearthatnovelviewsoncognitionarestrivingtoreplaceonlyoneofthesetypes,whileitisperfectlyconsistenttomaintaintheother
Figure1:Representationsoftype1.
two.Inordernottorunintothetrapofusingoverloadedtermsandraisemoreconfusion,Isimplycallthemtype1,2and3:
RepresentationofType1
By'type1'Irefertotheclassicalnotionofrepresentationasaconceptualmirroroftheworld(seefigure1).Thisisthemainbasisofsymbolicartificialintelligence,andofmanymodelsfromcognitivepsychologyalike.Theunderlyingviewisthatthereisanobjectivelyexistentoutsideworld(reality)whichmustbemappedontoafaithful(albeitoftensimplified)imageinthecognitiveagentinorderforthelattertoactintelligently.As,forinstance,Charniak&McDermott(1985)putit,such"representationsarestylizedversionsoftheworld".Cognitionisthenseenasakindofcomputationontheserepresentations,whichcanbelargelydefinedindependentlyfromtheactualimplementation(beitasymbolicprogramorthebrain).
Theimportantthingtonoteisthatinordertouseandspeakofsuchkindsofrepresentations,onemustastheobserveridentifytwothings:therepresentationanditsmedium(e.g.symbols,oractivationpatternsinaconnectionistnet)ononehand,andtheobjectsorconceptsreferredto(thedenotata)ontheother.Forinstance,tosaythatasymbol'CHAIR'representsthecategoryofchairs,onemustnotonlyspecifythesymbol,butmustalsoassumethatacategorychairexistsintheworld,independentlyfromwhethertheobserverortheagenttobemodeledinteractswiththeworld.Theimplicitassumptionisthusthatofastructuredworldthatexistsevenwithoutintelligentagentscognizingaboutit.
Page25
Figure2:Representationsoftype2.
RepresentationofType2
WhatIcallrepresentationoftype2referstointernalmentalstatesthatenableacognitiveagenttoactevenintheabsenceofimmediateperception.Whilethisisimplicitlycontainedwithintheclassicalviewoftype1,thisconceptisnotnecessarilytiedtoarepresentationoftype1.Type2representationscanexistwithoutpresupposinganobjectiveoutsidereality,sincetheycanbetheresultofacomplexindividualinteractionoftheagentwithitsenvironment,withoutthenecessityofpostulatingthatitmustdirectlycorrespondtoanythingintheworld(seefig.2).
Thisisthecrucialpoint.Ifweobserveahumanactinabsenceofimmediateperceptionofthebasicdrivingforcebehindtheiractionwecanandmustassumethatthepersonactsoninternalmentalstatesstandinginfortheirconceptsofwhatdrivestheiraction.Forinstance,ifsomebodyasksusforthenearestcoffeeshopbecausethatpersonistired,weingeneralcansafelyassumethatthepersonisactingonaninternalrepresentationoftheirconceptofcoffeeandtheeffectitcanhave.Thisiswhatdistinguisheshumansfromlowerlifeformsthatmainlyorsolelyactonperceptionsinareactiveway.
Totalkaboutandmodelsuchakindofrepresentationonedoesnothavetopresupposeanypregivenstructureintheworldandthattherepresentationreflectsthatstructure.Representationsinthisviewdonotrefertoanythinginreality,theyarerealityforthecognitiveagent.Theyaretheresultoftheagent'sinteractionswiththeworld,itsperceptionsanditsexperiencewiththem.Thisisthemajordifferencetotype1.Whenusingtheconceptofrepresentationsoftype2,onedoesnot(ornotnecessarily)assumetheexistenceofanobjectivestructuredreality,butinsteadseesrepresentationalwayswithrespecttotheindividualagent.Whateveritisthatpersonreferstoas'coffee',beitthesamethatweastheobservercall'coffee',oronlysimilar,orsomethingcompletelydifferent,itisencompassedinaspecificinternalmentalstatethatpermitsthatpersontoactonandspeakaboutcoffee.Thisstatestandsinforwhateverthepersonhadpreviouslyperceived(orconceived),andthusitisakindofrepresentation.Contrarytotype1,however,itisrepresentationwithnoidentifiabledenotata,unlessonehasaverydetailedinsightintotheperson'spastandpresentperceptionsandexperiences.
RepresentationofType3
Thisistheloosestsenseofrepresentation,butthisuseisnotuncommon.Type3referstoimmediatecausalrelationshipsbetweenastimulusanditseffect.Forinstance,thefiringofaneuronintheretinarepresentstheamountoflightthathasfallenontoit.Itisthiskindofrepresentationmerelyreactivelifeformsactupon.Similarlytotype2,onedoesnothavetoassumeaprestructuredworld.Evenmoreso,theonlyassumptiononeneedstospeakabouttype3representationsistheexistenceofphysicalsignalsintheworldcapableofactingasstimulitotheagent.
Sincetype2representationsweredefinedasbeingtheresultofperceptionandactionbytheagent,type3representationsareaprerequisitefortype2statestoemerge.
TheImportanceofType2Representations
Itisworthnotingthatthedistinctionofseveraltypesofrepresentationissomewhatconsistentwitharecentsuggestionof"representationasamediatingstate"byMarkman&Dietrich(1998).Firstofall,theseauthorsalsoseetheimportanceofexamininghow"theconcept[ofrepresentation]isusedincognitivemodelsinpractice"(par.4).Theythenproposerepresentationasamediatingstatebetweenperceptualstimuliandanagent'sreactionstothem.Thisviewappliestobothmyrepresentationsoftype2and3.Markman&DietrichfurtherarguethatrepresentationsinartificialintelligencewhatIcalltype1lackthelinktoentitiesoutsidethesystemandthusarenotmediating.
Page26
Figure3:Whathappenswhendesigningrepresentationsoftype1.
Although,atfirst,Markman&Dietrichdonotdistinguishbetweentype2and3,theythenmakeanimportantobservationaboutenduringmediatingstates.Forthem,thisisanimportantfeaturedistinguishingsimplereactivefromtrulycognitivesystems.Myargumentiscastalongsimilarlines.Bysayingthatcognitiveagentsdonotneedrepresentations,oneeasilythrowsoutthebabywiththebathwater.Inordertoexplaincognition,oneneedstoexplainenduringmediatingstates(inMarkman&Dietrich'sterminology),orrepresentationsoftype2(inmyterminology).Sincemanyauthorshavefailedtorecognizethis,anapparentgapbetweenresearchonembodiedandsituatedsystemsononehand,andresearchon"higherlevel"cognition,suchaslanguageorreasoningcapabilities,hasemerged.Howconnectionismcanfillthisgap,isexplainedbelow.
ConnectionismandRepresentations
Thedistinctionofthreetypesofrepresentationhelpsclarifyingthemaindifferencebetweenconnectionistandmoreclassicalcognitivistmodels.Firstitisworthlookingatthepicturebehindrepresentationsoftype1alittlemoreclosely(fig.3).Whenassumingthattherepresentationsreflectstructureintheoutsideworld,onemakesthemistakeofincludingoneself(astheobserver)intotheloop.Inorderfor'arepresentationstandsfor(orevenrefersto)astateintheworld'tomakesense,onemustassumethestate(thedenotata)toexistinanexactlyspecifiedway.However,onsecondlook,itbecomesclearthattheonlyplacewherethatstatecanreliablybeseenasexistingisinthemindoftheobserverwhoidentifiedthereferringrelationshipoftherepresentationtobeginwith.
Amoreconsistentpicture(inlinewiththeconstructivisttheoryofcognition(Maturana&Varela1987vonGlasersfeld1988seealsovonGlasersfeld,thisvolume)isthis.Ifwepushtheobserveroutofthisloop,weneedonlyassumetheworldtoconsistofphysicalstatesandsignals,whichacognitiveagentinteractswith.Whenadesignerofacognitivemodel(oranAIprogram)insertsrepresentationsoftype1(betheysymbolicorconnectionistorotherwise)theonlythingtheyrefertoisthedesigner'sownconceptualizationsoftheworld(therealityasperceivedbythedesigner)seefig.3.However,thatindividuallyperceivedrealityneednotbeusefulforthecognitiveagenttobemodeled.Especiallyifthatagenthasnothadthesametypeofexperiences(orisnotevenabletohavethem,sinceitssensorsandmotorcapacityarequitedifferent,asinarobot),insertingrepresentationsoftype1(orassumingsuchobjective,insertablerepresentationstoplayaroleintheagent)cruciallyseverstherepresentationsfromwhattheyaresupposedtoserveinnamelyinhelpingtheagenttoactintelligently.
Connectionismisamodelingframeworkthatifusedproperlycanhelpinpushingthedesignerorobserveroutoftheloop.Onemainideabehindconnectionistmodelsistheirinherentfocusonlearningandselforganization.Withthisfocus,thedesignofmodelsforagentswhichthemselvesactontheirownperceptionsoftheworld(therepresentationsoftype3)andbuildtheirowninternalstatestoactuponappropriately(theirrepresentationsoftype2)ispossible.
Asimpleexampleofrepresentationsoftype2inconnectionsttermsisthefamousmodelbyElman(1990).HeusesarecurrentneuralnetworktolearnthetaskofpredictingsubsequentwordsinasequenceofsimpleEnglishsentences.Thewordsarerepresentedattheinputandtargetoutputaslocalactivationpatterns.Thesearerepresentations
Page27
oftype1,sinceonemustassumethatthosewordsexistassuch,andtheyaredirectlymappedtosingleunitsintheinputandoutputlayers.Wordsarepresentedoneaftertheotherandthenetworkistrainedtooutputaprobabilityofwhatthemostlikelynextwordinthesequenceis.
TheinterestingobservationcomesfromElman'sinspectionofthehiddenlayeractivationpatternsaftertraining.Foreachword,hetooktheaveragepatternthatisactivatedwhenthewordisinput,andperformedaclusteranalysisonthesepatterns.Hethenobservedthatthepatternsclusteredinawayroughlycorrespondingtoaclassificationintoverbsandnouns.Looselyspeaking,thenetworkhasdiscoveredconceptsofgrammaticalcategories,andhasarepresentationofthemintermsofhiddenlayersubspaces.Severalaspectsareimportanttonote:
Elmanasthemodeldesignerdidnothavetoassumetheexistenceofgrammaticalcategories(onlyofwords).
Onecannotsafelysaythatthoserepresentationsrefertoanyobjectivelyexistentconcept.Whatlinguistscall'nouns'or'verbs'isdecisivelydifferentfromwhatthenetworkhasdiscovered.Forinstance,nosemanticaspectsofgrammaticalcategoryisincluded,andwouldnotevenbeusefulfortheparticulartaskofthenetwork.
Nevertheless,therepresentationsstandinforsomethingusefulintreatingsentencesandgrammarandthusplayanimportantroleforthenetwork's"actions".
Forthesereasons,therepresentationsarenotoftype1,butoftype2.Theonlyexceptiontotheabovediscussionisthatthenetworkdoesnotexhibitanysignificantactionintheabsenceof"perception"(itsinputs).Thismakestherepresentationshardtodistinguishfromtype3inthiscase.Butthesteptowardtruetype2representationsisobvious.1
Thepicturebecomessomewhatclearernow.Apparentlyitisthefirstnotionrepresentationoftype1thatmanyrecentapproachestocognitivescienceseektoabandon.Whilerepresentationsoftype2and3seemtobestillconsistentwithanemergentist,constructivistviewofcognition,itisthefallacyofseeingrepresentationasmirroringorsomewhatreferringtoanobjectiveoutsideworldwhichmustbeovercome.Thus,whenauthorsspeakabout'nonrepresentationalism'theyshouldinsteadrefertonontype1representations.
TheConnectionistRoutetoEmbodiment
Thediscussionaroundrepresentationinthecontextofconnectionistmodelsmightstillappearabitfarfetched.Takeanytypicalconnectionistmodelsuchasthosebasedonmultilayerperceptronsandbackpropagation(e.g.Elman1990Seidenberg&McClelland1989,Plunkett&Marchman1991).Thosemodelsarerelativelyremotefromanyideaofasituatedembodiedcognitiveagent.Firstofall,thereisnoaction,letaloneinteractionbetweenthesystemandtheenvironment,otherthantheadaptationofthesystem'sweightstopatternsthathavebeencarefullydesignedbytheprogrammer.Thosepatternsindeedmustbeviewedasrepresentationsoftype1,sincetheycanonlybedesigned(e.g.byhandcraftingdistributedpatternsbasedonmicrofeatures)throughreferencetoworldconceptsasconceivedbytheprogrammer.Onlytheinteriorofthemodel,althoughlackinganyinterestingdynamicsinmanycases,appearstooffersomethingnewastheElmanmodelhasexemplified.
Isthiswhereconnectionismends,asVarela(1990)apparentlysuggests?Theanswerprovidedbythispaperisaclear'no'.Asarguedearlier(Dorffner1997),connectionismhasalltherequisitesofgoingallthewaytowardabasisfortrulyembodiedcognitivemodels.Fig.4highlightsthemajorstepsonsucharoutetowardembodiment.Startingfromclassicalcognitivistapproacheswhicharefirmlybasedonrepresentationsoftype1,"traditional"connectionistmodelstakeadecisivesteptowardpermittingatype1notiononlyattheirinputsandoutputs.Suchmodels,throughmechanismsoflearningandselforganization,demonstratehow"internal"representationscanbebuiltfromthetaskoflearninginputoutputbehaviorwhichmustbeseenasonlybeingoftype2,not1.
Buttheinputsandoutputstoaconnectionistmodeldonothavetoberepresentationsoftype1either.Bymovingthoseendsallthewaytosensory
lBycomparingmysuggestionsagainwithMarkman&Dietrich(1998),onenotesthefollowing.Whiletheyseeconnectionistweightmatricesastheenduringmediatingstates,Idistinctivelysuggestactivationpatternsor,inalaterstep,dynamicattractorsinactivationstatespace,asrepresentationsoftype2.
Page28
Figure4:Theconnectionistroutetoembodiment.
input(e.g.realacousticinputinsteadofphonologicalfeatures)andoutputs(e.g.articulatormovementsinsteadoffeaturesagain),intheoryamodelisconceivablethatfreesitselfofrepresentationsoftype1altogether.Thisisthecasesincesensoryinputandmotoroutputifstillcalledrepresentationsaremerelyoftype3causalcorrelatesofphysicalsignalsintheenvironment.Thisissymbolizedinfig.4byanextended"towerbridge"metaphorbyCummins&Schwarz(1987)(seealsoDorffner1991).Thelastdecisivestepalmostcomesautomatically.Inordertobuildandtrainsuchamodel,itmustitselfinteractwithitsenvironment,becausepredigestedtasks(suchasmappingawordtoitspasttenseform)becomemoreandmoredifficult,ifnotimpossible,whengoingtosuchinputsandoutputs.Inafirststep,eventhetypeoftaskwillshift,suchasawayfromlanguagetowardmorebasicautonomousbehaviorsymbolizedinfig.4throughanimageofthesimpleneuralnetworkbasedagentsbyPfeifer&Verschure(1992).Butthebasicpremisesforbeingacognitivemodel(andnotjustamodelofreactivebehavior)arestillthereandthecognitiveresearcherisnowfacedwiththeformidabletaskofdesigningtheinteriorofthemodel.Ihavepreviouslycalledthisdesign"metalevelrepresentations"aconceptIwillcomebackto.
Insummary,eventhoughconnectionismwillhavetoshiftsomeofitsfocus,andperhapsalsoitsmodelbuildingblocks,itprovidesthebasicingredientsofbecomingsomethingthatcouldbecalled"appliedconstructivism".Theconstructivistpartisthatofamodelbuildingitsownrepresentations(oftype2),basedonitsownexperience,valueschemesandperceptions,andbeingoptimalandrelevantforthemodelitself,independentofanobserverordesigner.
ImplicationsforConnectionistCognitiveResearch
Twoimportantquestionscomeupwhenviewingconnectionistresearchinthislightofrecenttrendsincognitivescience.
1.Dowehavetostoplookingatempiricaldataandforminghypothesesaboutaspectsofcognition?Oneofthemostimportantrevolutionsincognitivism(ascomparedtoitspredecessor,behaviorism)wasthenewlygainedpossibilitytospeakabouttheinnardsofthemind.Thushypothesesaboutcognitivefunctionscouldbeformedwhichcouldbetestedagainstempiricaldata.Whenresortingtorepresentationsoftype2,whichemergefromtheagent'slearningbehavior,thispossibilityseemstobelostagain.
This,however,isnotthecase.Afterall,cognitivescientistsasmentionedabovestillhavetodesignthemodelstructure(the"metalevel"representations).Itwouldbeabigmistaketoexpectafullygroundedandembodiedconnectionistmodeltolooklikeahugebackpropagationnetwork.Currentmodelsonlypresentuswithtinybitsandpiecesofalargermodel,thestructure,connectivityanddynamicsofwhichwillhavetobebuiltbasedonhypothesestestedagainstempiricaldata.
Furthermore,computationallearningtheory(Anthony&Biggs1992)showsusthatinorderforaproblemtobelearnable,thelearnerneedsbias(i.e.apregiven,orinnate,constraint).Elmanetal.(1996)haveusedconnectionismtospawnanewdis
Page29
cussionaboutinnateness.Theyargueforreplacingolder,e.g.Chmoskyan,notionsofinnatenesswithstructuralandotherconstraintsinconnectionistnetworks.Andconstraintsneedtobethere,anditisuptohypothesesandempiricaldatatoidentifywhich.
Agoodexampleishumans'conceptualizationcapability(compareLakoff1987).Ifweobservethatinhumancognitioncategorizationofenvironmentalstimuliplaysanimportantrole,theninsertingabasiccategorizationfacility(as,forinstance,inmodelsbyDorffneretal.1996Kruschke1993Murreetal.1989),intothemodelmeansrepresentingthiscapability.Thisrepresentationisonametalevel,sincenoconreteconceptisrepresentedbutonlythemetaconceptofcategory(and,perhaps,propertiesofhumancategories,suchasprototypicality,basicleveleffectsetc.cf.Dorffner1996).Withinsucharchitecturalconstraints,representationsoftype2canevolvethroughinteractiveexperiencebytheagent.
Onemustnote,however,thatthepermissionofthedesignofarchitecturalconstraintsmeansthattheprogrammerissomehowbackintheloop.Theargumentalsoputforwardpreviously(Dorffner1997)isthatthisnowhappensonadecisivelydifferentlevel.Theonlyalternativewouldbetheinclusionofamodelofevolutionarydevelopmentofarchitectureintotheapproach(asVarelaetal.1991argue).Borrowingterminologyfromlinguistics,onecansaythattheapproachfavoredhereisatypeofsynchroniccognitivescience,whichdoesnotworryabouthowandwhyacertaincognitivearchitecturehasdevelopedasopposedtoamorediachroniccognitivesciencefavoredbyVarelaetal.(1991).Byadoptingasynchronicview,onegetstheopportunityoftestinghypothesesagainstempiricaldataandtodevoteone'seffortstocomplexphenomenaobservedincognitiontoday.Thefeewehavetopayforthisisaframework,whichisnotentirelyfreeofthemodelersownperceptionoftheworld.Somerepresentationoftype1(onametalevel,though)remains.
2.Dowehavetosticktosimplereactivebehavior?Ashasbeenarguedabove,whilegoingalongtheconnectionistroutetoembodiment,thefocusseemstoshiftfrom"higherlevel"cognitionlikeaspectsoflanguagetomore"lowerlevel"behaviorofautonomousrobots.Cognitivescience,however,appearstobemoreabouttheformerthanaboutthelatter.
Thequestioncanagainbeanswerednegatively.Thereisnoreasonwhyoneshouldnotstudyphenomenalikelanguageinthecontextofanembodiedagent(whoseprimarytask,ofcourse,istoavoidobstaclesandthelike).Wejusthavetorealizethatitisnotnecessarilyhumancognitioninallitsextentwearestudying.Agoodexample,althoughnotwithinconnectionism,istheworkonlanguageinautonomousagentsbySteels(1996)andcolleagues.Whyshouldconnectionistsbepreventedfromdoingsimilarthings?
Onecantakeyetanothertwistinansweringthequestion.Ascanbeseenfromlookingattheconnectionistroutetoembodiment,connectionistresearchishighlyrelevanttomanynewtrendsincognitivescience,evenifthemodelsarenotfullyembodied,groundedorsituated.Aslongascognitivescientistskeepthebasicsettingofamodelasbeingembodiedinmind,theyaresafeinbuildingandstudyingmodelswhilelargelyleavingoutimportantfacetsofanembodiedagent(e.g.thatithasaphysicalbodyandisinteractingwithitsenvironment).Theystillarriveatimportantresultsforanewcognitivescience.TheworkbyElman(1990),Seidenberg&McClelland(1989)orPlunkett&Marchman(1991)onlanguagelearningaregoodexamples.Noneofthemapparentlysharemanyfeatureswithworkonembodiedrobotsorsimulatedautonomousagents,butneverthelesscontributemuchtoacognitivesciencethatseesrepresentationsoftype1asobsolete.Inthissense,thesetwobodiesofresearcharemuchclosertoeachotherthantheyaretoclassicalcognitivistapproaches.
WhataboutDynamicism?
Ihavestartedbylistingseveralthemesbelievedtobethemajortrendsinrecentcognitivescience.Amongthemtherewasthenotionofdynamicism.Thisreferstoworkpromotingtheconsiderationofacognitivesystemasaninherentlydynamicone,whichcanappropriatelybedescribedusingthevocabularyandformalismsofnonlineardynamicalsystemstheory(e.g.Port&vanGelder1995vanGelder1995Bickhard&Terveen1995Elmanetal.1996).
VanGelder(1995),forinstance,hasarguedthatdynamicalsystemsneitherinvolverepresentationsnorcomputations.Hepredictedthatafewyearsfromnow,thedistinctionwillnolongerbebetween
Page30
symbolicandconnectionistmodels,butonlybetweencomputationalanddynamicones.Connectionistmodelscanbeofeithersideofthisdivision.
Giventheargumentsabove,onecanmaketwoobservations.First,lookingatcognitionfromadynamicviewpointdoesnotmakerepresentationsoftype2obsolete.VanGelder'sfamousexampleofadynamicalsystemtheWattgovernorwould,incognitivescienceterms,correspondtoasimplereactivesystemwithnorealcognitiveabilities.Inotherwords,foradynamicmodelofhighercognitivefunctionsonewouldexpecttosee(dynamic)states(orstatetransitions)thatcanstandinforimmediateperceptionsandthusfulfilltheroleofarepresentationasmentionedabove(compareMarkman&Dietrich1998,par.13).SimilartoBrooks(1991)orVarelaetal.(1991),vanGelderapparentlyreferstotype1whenstatingtherearenorepresentations.
Secondly(assomeauthorsinPort&vanGelder1996exemplify),connectionsimhastheperfectprerequisitesforformingthebasisfordynamicalmodels,especiallywhencontainingrecurrentcomponents,asinElman's(1990)case(seealsoElmanetal.1996).Therefore,connectionismshouldcontinuetoplayalargeroleincognitivescienceevenunderthedynamicistparadigm.
WhereisReality?
Finallywearriveattheimportantquestionofthisvolume:Doesrepresentationneedreality?Giventhedistinctionoftype1through3,thisquestionisrelativelyeasytoanswer.
Representationsoftype3needtoassumearealityasacollectionofphysicalsignals.
Representationsoftype2needarchitecturalconstraints,whichareshapedbysomeabstractconceptofrealityasperceivedbythemodeler(e.g.thatcognitivesystemscategorizetheworld).Theactualrepresentations,however,canbeconceivedofwithoutthenecessityofassuminganobjectiverealityonthelevelofconcreteentities(suchasconcretecategories).
Representationsoftype1arehighlydependentonanobjectivereality,buttheycanbelargelyabandoned.
Thequestionariseswhetheracognitivescientistisstillpermittedtotalkaboutrepresentations,analyzethemormakehypothesesaboutthem.Ifrealityonaconcretelevelcannolongerbeassumed,thebasisforreasoningaboutrepresentationsseemstohavevanished.Orputdifferently,anythingwouldcountasavalidrepresentation,aslongasitguidesbehaviorinsomeway.
Thisadditionalquestioncanalsobeansweredpositively.Asobserversofbothrealcognitivebeingsandcomputersimulatedmodels,noharmisdoneifwetalkaboutrepresentations(oftype2),asiftheywereoftype1(i.e.asiftheycorrespondtoobjectivelyexistingitemsinreality).Thisistrueaslongasourdoingsodoesnotinfluencethemodel'sbehavior.Therefore,contrarytobehaviorism,cognitivescientistsarestillentitledtoreasonaboutmentalstates,aslongastheyrealizethatthosestatesdonotexactlycorrespondtoanyofthescientist'sownconceptsoftheworld.Forinstance,wearesafetoreasonaboutElman'smodelasiftheinternalrepresentationscorrespondtoalinguist'snotionofgrammaticalcategories,andcaninferimportantpropertiesoflanguagelearning.Wearesafeaslongasweonlydesignthegeneralprerequisitesforlearningsuchrepresentations(throughthenetworkarchitecture)insteadofhandcraftingtype1representationsofnounsandverbspriortothetasktobemodeled.
Conclusions
InthispaperIhavearguedthatconnectionistresearch,asamodelingframework,isconsistentwithmostrecenttrendsincognitivescience,fromresearchonembodiedandsituatedagentstodynamicism.Ihavecenteredmyargumentaroundaclarificationoftheconceptofrepresentationbyintroducingthreetypes.Recentapproachestocognitivemodelingabandononlythefirstofthesetypesthenotionofrepresentationasbeingamirrorofanobjectivelyexistentworld.Theothertwotypesareconsistentwithanembodied,dynamicistorconstructivistviewoncognition.Ihaveparticularlystressedtheimportanceoftype2representationasaninternalmentalstatestandinginforanagent'spastperceptionsorexperienceswhichmuchrecentresearchappearstoerronouslyabandonaswell.
StartingfromthisdistinctionIhavearguedthatconnectionismcannotonlyincorporatetheabandonmentoftype1representations,butthatitcanalsoprovideanaturalroutetoembodiedandsituatedmodels.Fromthisconnectionistroutetoembodimentitbecameclearthatcurrentconnec
Page31
tionistwork,althoughapparentlydetachedfrommostideasofembodiedandsituatedcognition,isneverthelesshighlyrelevantforthesemodernstreamsincognitivescience.Therefore,aradicaldeparturefromcurrentissuesinconnectionismisnotnecessary,aslongasthebasicsettingofacognitiveagentinteractingwiththeoutsideworldiskeptinmind.
Lastbutnotleastthequestion''Doesrepresentationneedreality"couldbeanswered.Ifweabandonrepresentationsoftype1,theonlyrealitywemustassumeistheexistenceofphysicalsignalsintheenvironment.
References
Anthony,M.&Biggs,N.(1992)ComputationalLearningTheory.Cambridge:CambridgeUniversityPress.
Bickhard,M.H.&Terveen,L.(1995)FoundationalIssuesinArtificialIntelligenceandCognitiveScience.ElsevierSciencePublishers.
Brooks,R.A.(1991)IntelligencewithoutRepresentation.ArtificialIntelligence,SpecialVolume:FoundationsofArtificialIntelligence47(13):139160.
Charniak,E.&McDermott,D.(1985)IntroductiontoArtificialIntelligence.Reading,MA:AddisonWesley.
Cummins,R.&Schwarz,G.(1987)RadicalConnectionism.Proc.ofSpindelConf.1987:ConnectionismandthePhilosophyofMind.
Dorffner,G.(1991)Konnektionismus.Stuttgart:Teubner.
Dorffner,G.(1996)CategorizationinearlylanguageacquisitionAccountsfromaconnectionistmodel.OsterreichischesForschungsinstitutfrArtificialIntelligence,Wien,TR9616.
Dorffner,G.(1997)RadicalConnectionismANeuralBottomUpApproachtoAI.In:Dorffner,G.(ed.)NeuralNetworksandaNewAl.London:InternationalThomsonComputerPress.
Dorffner,G.,Hentze,M.&Thurner,G.(1996)AConnectionistModelofCategorizationandGroundedWordLearning.In:Koster,C.&Wijnen,F.(eds.)ProceedingsoftheGroningenAssemblyonLanguageAcquisition(GALA'95).
Elman,J.L.(1990)FindingStructureinTime.CognitiveScience2(14):179212.
Elman,J.L.,Bates,E.A.,Johnson,M.H.,KarmiloffSmithA.,Parisi,D.&Plunkett,K.(1996)RethinkingInnateness.Cambridge,MA:MITPress.
Fodor,J.A.,Pylyshyn,Z.W.(1988)ConnectionismandCognitiveArchitecture:ACriticalAnalysis,Cognition28:371.
Gelder,T.van(1995)Modeling,connectionistandotherwise.In:Niklasson,L.&Boden,M.(eds.)CurrentTrendsinConnectionism.Hillsdale,NJ:LawrenceErlbaum,pp.217235.
Glasersfeld,E.von(1988)TheConstructionofKnowledge.Seaside:IntersystemsPublications.
Harnad,S.(1990)TheSymbolGroundingProblem.PhysicaD42:335346.
Kruschke,J.K.(1993)HumanCategoryLearning:ImplicationsforBackpropagationModels.ConnectionScience5(1),336.
Lakoff,G.(1987)Women,FireandDangerousThingsWhatCategoriesRevealabouttheMind.Chicago:UniversityofChicagoPress.
Markman,A.B.&Dietrich,E.(1998)Indefenseofrepresentationasmediation.PSYCOLOQUY98.9.48.
Maturana,H.R.&Varela,F.J.(1987)TheTreeofKnowledge.Boston:Shambhala.
McClelland,J.L.,Rumelhart,D.E.(1986)ParallelDistributedProcessing,ExplorationsintheMicrostructureofCognition,VolII:PsychologicalandBiologicalModels.Cambridge,MA:MITPress.
Murre,J.M.J.,Phaf,R.H.&Wolters,G.(1992)CALM:CategorizingandLearningModule.NeuralNetworks5(1):5582.
Nehmzow,U.&Smithers,T.(1991)MapbuildingusingSelfOrganisingNetworksin"ReallyUsefulRobots".In:Meyer,J.A.&Wilson,S.W.(eds.)FromAnimalstoAnimats.Cambridge,MA:MITPress.
Pfeifer,R.,Verschure,P.(1992)BeyondRationalism:SelfOrganizing,SensoryBasedSystems.ConnectionScience4(3&4):313326.
Plunkett,K.&Marchman,V.(1991)Ushapedlearningandfrequencyeffectsinamultilayeredperceptron:Implicationsforchildlanguageacquisition.Cognition38:43102.
Port,R.F.&Gelder,T.J.van(eds.)(1995)MindasMotion.Cambridge,MA:MITPress.
Rumelhart,D.E.&McClelland,J.L.(1986)ParallelDistributedProcessing,ExplorationsintheMicrostructureofCognition,Vol1:Foundations.Cambridge,MA:MITPress.
Seidenberg,M.S.&McClelland,J.L.(1989)Adis
Page32
tributed,developmentalmodelofwordrecognitionandnaming.PsychologicalReview96(4):523568.
Skarda,A.&Freeman,W.J.(1987)Howbrainsmakechaosinordertomakesenseoftheworld.BehavioralandBrainSciences2(10)161196.
SmolenskyP.(1988)OntheProperTreatmentofConnectionism.BehavioralandBrainSciences11(88):174.
Steels,L.(1994)EmergentFunctionalityinRoboticAgentsthroughOnLineEvolution.In:Brooks,R.A.&Maes,P.(eds.)ArtificialLifeIVCambridge,MA:MITPress,pp.816.
Steels,L.(1996)TheSpontaneousSelforganizationofanAdaptiveLanguage.In:Muggleton,S.(ed.)MachineIntelligence15.Oxford:OxfordUniv.Press.
Varela,F.J.(1990)KognitionswissenschaftKognitionstechnik.Frankfurt/Main:Suhrkamp.
Varela,F.J.,Thompson,E.&Rosch,E.(1991)TheEmbodiedMind.Cambridge,MA:MITPress.
Page33
TheOntologicalStatusofRepresentations
MatthiasScheutzComputerScience&EngineeringDepartment,UniversityofNotreDameNotreDame,IN46556,USAEmail:mscheutz@cse.nd.edu
AbstractThegoalofthispaperistoarguethattheontologicalstatusofrepresentationscanonlybeevaluatedwithinatheory.Inotherwords,whatcountsasrepresentation,orwhetheracertainrepresentationisbetterthananotherone,dependssolelyonthe(levelof)descriptionofthephenomenonunderscrutiny.Itisshownhow"representation",beingasemanticnotion,canbedefinedintermsofthenotion"meaning".Forcognitivescience,inparticular,itfollowsthatrepresentations,functioningasmeredescriptivedevicestofacilitateone'sgoalofexplainingandmodelingbrain/thoughtprocesses,cannotinandbythemselvesgiverisetoontologicalorepistemologicalclaims.
Introduction
Representation,aswithallwidelyusedterms,isaratherambiguousnotionthatcarriesalotofsemanticoverweightwithit:inwhatsensedoesaword,forexample,representitsmeaning,alettergradestudents'abilities,theretinalimageobjects,ornuclearmissilespeace?Itisobviousthatonehastorestrictallpossibleconnotationsofrepresentationifonewantstodiscusstheapplicabilityandusefulnessofthisnotionforcognitivescience.
Mostcognitivescientistsuse"representation"todescribetherelationofcertainmind/brain1statestoeventsintheworld,i.e.,the"inner"tothe''outer"perspective,theassumptionbeingthatminds/brainshavetosomehow"represent"whatisgivenintheworldtomakesenseoutofit,orasGardnerputsit:
"[...]thecognitivescientistrestshisdisciplineontheassumptionthat,forscientificpurposes,humancognitiveactivitymustbedescribedintermsofsymbols,schemas,images,ideas,andotherformsofmentalrepresentation."2
Theterm"represent"alreadysuggeststhisusage,since"re""present",beingcomposedof"re"and"present",literallymeans"presentagain".3Itindicatesthatsomethingwhichisnotpresent,butwhichhadbeenpresentedatsomepoint,istobepresentedagain(seealsoGlasersfeld1995).Whatthisreadingmeansforcognitivescienceisthatthebrainexploitsmechanismstokeeptrackofperceivedobjectsandstorethis"impression"inawaythatitcanberetrievedandusedifnecessary.This"storedimpression"isthenlabeled"representationofthestimulu
Recommended