A Biologically-BasedComputational Model of Working Memorypsych.colorado.edu/~oreilly/papers/OReillyBraverCohen99.pdfA Biologically-BasedComputational Model of Working Memory Randall

A Biologically-BasedComputationalModelof WorkingMemory

RandallC. O’Reilly�, ToddS.Braver

�, & JonathanD. Cohen

�

�Departmentof Psychology

Universityof ColoradoatBoulderCampusBox 345

Boulder, CO80309-0345

�Departmentof PsychologyandCenterfor theNeuralBasisof Cognition

CarnegieMellon UniversityBaker Hall

Pittsburgh,PA [email protected],[email protected],[email protected]

December22,1997

Draft Chapterfor Miyake,A. & Shah,P. (Eds)Modelsof WorkingMemory:Mechanismsof ActiveMaintenanceandExecutiveControl. New York: CambridgeUniversityPress.

O’Reilly, Braver, & Cohen 1

FiveCentralFeaturesof theModel

Wedefineworkingmemoryascontrolledprocessinginvolving activemaintenanceand/orrapidlearning,wherecontrolledprocessingis anemergentpropertyof thedynamicinteractionsof multiplebrainsystems,but theprefrontalcortex (PFC)andhippocampus(HCMP)areespeciallyinfluentialdueto theirspecializedprocessingabilitiesandtheirprivilegedlocationswithin theprocessinghierarchy(boththePFCandHCMParewell connectedwith a wide rangeof brainareas,allowing themto influencebehavior at a globallevel).Thespecificfeaturesof ourmodelinclude:

1. A PFCspecializedfor active maintenanceof internalcontextual informationthat is dynamicallyup-datedandself-regulated,allowing it to bias (control) ongoingprocessingaccordingto maintainedinformation(e.g.,goals,instructions,partialproducts,etc).

2. A HCMP specializedfor rapidlearningof arbitraryinformation,whichcanberecalledin theserviceof controlledprocessing,while theposteriorperceptualandmotorcortex (PMC)exhibits slow, long-termlearningthatcanefficiently representaccumulatedknowledgeandskills.

3. Controlthatemergesfrom interactingsystems(PFC,HCMPandPMC).

4. Dimensionsthatdefinecontinuaof specializationin differentbrainsystems:e.g.,robustactive main-tenance,fastvsslow learning.

5. Integrationof biologicalandcomputationalprinciples.

2 Biologically-BasedComputationalModel

Introduction

Working memoryis an intuitively appealingtheoreticalconstruct— perhapsdeceptively so. It hasproven difficult for the field to converge on a fully satisfying,mechanisticallyexplicit accountof whatexactly working memoryis andhow it fits into a larger modelof cognition(hencethemotivation for thisvolume). Existing theoreticalmodelsof working memorycan be tracedto ideasbasedon a traditionalcomputer-like mentalarchitecture,whereprocessingis centralizedand long-termmemoryis essentiallypassive. In this context, it makes senseto have RAM or cache-like working memorybuffers dedicatedto temporarilystoringitemsthat areneededduring processingby the central executive(Baddeley, 1986).Alternative processingarchitectureshave beenproposed,bothwithin thecomputationalandpsychologicalliteratures(e.g.,Anderson,1983;Newell, 1990),in which working memoryis definedfunctionally — asthe activatedcomponentof long term memoryrepresentations— ratherthanstructurallyasa dedicatedcomponentof the system. However, thesetypically include a structuraldistinction betweenprocessingandmemory. Noneof thesearchitecturesseemsto correspondcloselyto the architectureof the brain, inwhich processingandmemoryfunctionsaretypically distributedwithin andperformedby thesameneuralsubstrate(Rumelhart& McClelland,1986).

Webelieve thatconsideringhow workingmemoryfunctionmightbeimplementedin thebrainprovidesa uniqueperspective that is informative with regardto both the psychologicalandbiologicalmechanismsinvolved. This is what we attemptto do in this chapter, by providing a biologically-basedcomputationalmodelof working memory. Our goal is not only to provide anaccountthat is neurobiologicallyplausible,but alsoonethatis mechanisticallyexplicit, andthatcanbeimplementedin computersimulationsof specificcognitive tasks. We sharethis goal with othersin this volumewho have alsocommittedtheir theoriestomechanisticallyexplicit models,at both the symbolic (Lovett, Reder, & Lebiere,this volume;Young&Lewis, this volume;Kieras,Meyer, Mueller, & Seymour, this volume)andneural(Schneider, this volume)levels.

It is possibleto identify a coresetof informationprocessingrequirementsfor many working memorytasks:1) Taskinstructionsand/orstimuli mustbeencodedin sucha form that they caneitherbeactivelymaintainedover time,and/orlearnedrapidly andstoredoffline for subsequentrecall;2) Theactive mainte-nancemustbebothdynamicandrobust,sothatinformationcanbeselectively maintained,flexibly updated,protectedfrom interference,andheldfor arbitrary(althoughrelatively short)durations;3) Themaintainedinformationmustbeableto rapidly andcontinuallyinfluence(bias)subsequentprocessingor actionselec-tion. 4) Therapid learningmustavoid theproblemof interferencein orderto keepevenrelatively similartypesof informationdistinct. In additionto thesespecificationsfor anactivememorysystemanda rapidlearning system,we think that the working memoryconstructis generallyassociatedwith tasksthat re-quirecontrolledprocessing, whichgovernstheupdatingandmaintenanceof activememoryandthestorageandretrieval of rapidly learnedinformationin a strategic or task-relevant manner. This is consistentwiththe original associationof working memorywith central-executive like functions. Taken together, thesefunctionalaspectsof workingmemoryprovideabasicsetof constraintsfor ourbiologicallybasedmodel.

Our approachinvolvestwo interrelatedthreads.Thefirst is a focuson thefunctionaldimensionsalongwhichdifferentbrainsystemsappearto havespecialized,andtheprocessingtradeoffs thatresultasaconse-quenceof thesespecializations.Thesecondis asetof computationalmodelsin whichwehaveimplementedthesefunctionalspecializationsasexplicit mechanisms.Throughsimulations,wehave endeavoredto showhow theinteractionsof thesespecializedbrainsystemscanaccountfor specificpatternsof behavioral per-formanceonawiderangeof cognitive tasks.Wehavepostulatedthatprefrontalcortex (PFC),hippocampus(HCMP) andposteriorandmotor cortex (PMC) representthreeextremesof specializationalongdifferentfunctionaldimensionsimportantfor workingmemory:sensoryandmotorprocessingbasedoninferenceandgeneralization(PMC); dynamicandrobustactive memory(PFC);andrapid learningof arbitraryinforma-


tion�

(HCMP).Sinceeachof thesespecializationsinvolve tradeoffs, it is only throughinteractionsbetweenthesesystemsthatthebraincanfulfill theinformationprocessingrequirementsof workingmemorytasks.

As anexampleof how thesecomponentswork together, considera simplereal-world taskthat involvescontributionsfrom thesedifferentbrainsystems.Imagineyou arelooking for someinformation(thenameof a college friend’s child) containedin anemailmessageyou receiveda yearagoandhave storedin oneof your many messagefolders.You canrememberseveral thingsaboutthatemail,like who sentit (a goodfriend who knows thecollege friend), whatelsewashappeningat aroundthat time (you hadjust returnedfrom aconferencein Paris),but youdon’t rememberthesubjectline or whereyoufiled it. This informationabouttheemailis retrievedfrom theHCMPsystem,whichwasableto bind togethertheindividual featuresof the memoryandstoreit asa uniqueevent or episode.Oncerecalled,thesefeaturesmustbe usedtoguide the processof searchingthroughthe folders and email messages.We think that this happensbymaintainingrepresentationsof thesefeaturesin an active statein the PFC,which is able to keepthemactive for thedurationof thesearch,andprotectthemfrom beingdislodgedfrom active memoryby all theotherinformationyou read. Meanwhilethe basicabilities of readinginformationandissuingappropriatecommandswithin theemailsystemaresubservedby well-learnedrepresentationswithin thePMC, guidedby representationshelpedactive in PFC.

Onceinitiated,thesearchrequirestheupdatingof itemsin active memory(collegefriend’s name,goodfriend’sname,Parisconference)andits interactionwith informationencounteredin thesearch.Forexample,whenyou list all of the folders,you selecta small subsetasmostprobable.This requiresan interactionbetweentheitemsin activememory(PFC),long-termknowledgeaboutthemeaningsof thefolders(PMC),andspecificinformationaboutwhatwasfiled into them(HCMP). Theresultis theactivationof a new setof itemsin active memory, containingthenamesof thenew setof foldersto search.You mayfirst decideto look in a folder that will containan email telling you exactly whenyour conferencewas,which willhelp narrow the search.As you do this, you may keepthat datein active memory, andnot maintaintheconferenceinformation.Thus,theitemsin activememoryareupdated(activatedanddeactivated)asneededby the taskat hand. Finally, you iteratethroughthe foldersandemail messages,matchingtheir dateandsenderinformationwith thosemaintainedin active memory, until thecorrectemailis found.

All of thishappenedasa resultof strategic, controlledprocessing,involving theactivationandupdatingof goals(theoverallsearch)andsub-goals(e.g.,findingthespecificdate).Themaintenanceandupdatingofgoals,likethatof theotheractivememoryitems,is dependentonspecializedmechanismsin thePFCsystem.Thus,thePFCsystemplaysa dominantrole in bothactive memoryandcontrolledprocessing,which aretwo centralcomponentsof theworkingmemoryconstruct.However, othersystemscanplayequallycentralroles.For example,if youwereinterruptedin yoursearchby aphonecall, thenyou mightnot retainall thepertinentinformationin active memory(“Now, wherewasI?”). TheHCMP systemcanfill in themissinginformationby frequently(andrapidly)storingsnapshotsof thecurrentlyactiverepresentationsacrossmuchof thecortex, which canthenberecalledafteraninterruptionin orderto pick up whereyou left off. Thus,workingmemoryfunctionalitycanbeaccomplishedby multiplebrainsystems,thoughthespecializedactivememorysystemof thePFCremainsacentralone.

Wehavestudiedasimpleworkingmemorytaskbasedon thecontinuousperformancetest(CPT),whichinvolvessearchingfor target lettersin a continuousstreamof stimuli (typically letters). For example,inthe AX versionof the CPT (AX-CPT), the target is the letter ’X’, but only if it immediatelyfollows theletter ‘A’. Thus, the ’X’ aloneis inherentlyambiguous,in that it is a non-target if precededby anythingotherthanan’A’. Like theemailsearchtask,this requiresthedynamicupdatingandmaintenanceof activememoryrepresentations(e.g.,thecurrentstimulusmustbemaintainedin orderto performcorrectlyon thesubsequentone),whichmakesthisaworkingmemorytask.Activemaintenanceis evenmoreimportantforamoredemandingversionof thistaskcalledtheN-back, in whichany lettercanbeatargetif it is identicaltotheoneoccurringN trialspreviously (whereN is pre-specifiedandis typically 1, 2, or 3). Thus,moreitems


need� to bemaintainedsimultaneously, andacrossinterveningstimuli. The N-backalsorequiresupdatingtheworkingmemoryrepresentationsaftereachtrial, in orderto keeptrackof theorderof thelastN letters.

Thereareseveralotherrelevantdemandsof this task.For example,uponreceiving thetaskinstructions,subjectsmustrapidly learntheotherwisearbitraryassociationbetweenthe letter ’A’ and’X’. We assumethat this is carriedout by the HCMP. Of course,subjectsmustalsobe ableto encodeeachstimulus,andexecutetheappropriateresponse,whichweassumeis carriedoutby PMC.Together, therapidassociationofthecuewith thecorrectresponseto thetarget (HCMP), theactive maintenanceof informationprovidedbythespecificcuepresentedin eachtrial (PFC),andtheuseof thatinformationto guidetheresponse(PMC),constitutea simpleform of working memoryfunction. In Figure1, we presenta computationalmodelofperformancein this task,thatillustratesour theoryregardingthefunctionalrolesof PFC,PMCandHCMP.We have implementedcomponentsof this model,anddemonstratedthat it canaccountfor detailedaspectsof normalbehavior in theAX-CPT, aswell asthatof patientswith schizophreniawhoarethoughtto sufferfrom PFCdysfunction(Braver, Cohen,& Servan-Schreiber, 1995;Cohen,Braver, & O’Reilly, 1996;Braver,Cohen,& McClelland,1997a).

In themodel,thePMC layerof thenetwork performsstimulusidentification,andresponsegeneration.Thus,in panela), whenthe‘A’ stimulusis presented,anunequivocalnon-targetresponse(heremappedontotheright hand,but counterbalancedin empiricalstudies)is generated.However, thePFCis alsoactivatedby this ‘A’ stimulus,sinceit servesasa cuefor a possiblesubsequenttarget ‘X’ stimulus.During thedelayperiodshown in panelb), thePFCmaintainsits representationin anactive state.This PFCrepresentationencodesthe informationthattheprior stimuluswasa cue,andthusthat if an‘X’ comesnext, a target (lefthand)responseshouldbe made. Whenthe ‘X’ stimulusis thenpresented(panelc), the PFC-maintainedactive memorybiasesprocessingin thePMC in favor of theinterpretationof the‘X’ asa target, leadingtoa target(left hand)response.We think thattheHCMPwouldalsoplayanimportantrole in performingthistask,especiallyin early trials, by virtue of its ability to rapidly learnassociationsbetweentheappropriatestimuli (e.g.,‘A’ in thePMC and‘left-to-X’ in thePFC)basedon instructions,andprovide a link betweentheseuntil directcorticalconnectionshave beenstrengthened.However, we have not yet implementedthisimportantcomponentof workingmemoryin this model.

In the following sections,we first elaborateour theoryof working memoryin termsof a morecom-prehensive view of how informationis processedwithin neuralsystems.While we believe it is importantthatour theorybasedon mechanisticmodelsof cognitive functionwhosebehavior canbecomparedwithempiricaldata,a detailedconsiderationof individual modelsor empiricalstudiesis beyond the scopeofthis chapter. Furthermore,many of the featuresof our theoryhave not yet beenimplemented,andremaina challengefor futurework. Thus,our objective in this chapteris to provide a high level overview of ourtheory, andhow it addressesthetheoreticalquestionsposedfor thisvolume.

A Biologically-BasedComputationalModelof Cognition

Ourmodelof workingmemoryis aunifiedonein thatthesameunderlyingcomputationalprinciplesareusedthroughout.Weandothershaverelieduponthesecomputationalprinciplesin previouswork to addressissuesregardingcognitive functionandbehavioral performancein many othertaskdomains,includingonesinvolving responsecompetition,classicalconditioning,andcovertspatialattention.Moreover, thefunctionalspecializationsthatwepostulatefor differentbrainsystemsemergeasdifferentparametricvariationswithinthis unified framework, giving rise to a continuumof specializationalong thesedimensions.The basiccomputationalmechanismsarerelatively simple,includingstandardparallel-distributed-processing (PDP)ideas(Rumelhart& McClelland,1986;McClelland,1993;Seidenberg, 1993),themostrelevantbeing:

The Brain usesparallel, distributed processinginvolving many relatively simple elements(neuronsor


X Q P L Left Right

Stimulus Response

PMC

PFCHCMP

A�

Leftto X

Rightto X

a) Cue b) Delay�

c) Probe

X Q P L Left Right

Stimulus�

Response

PMC

PFCHCMP

A�

Leftto X

Rightto X

X Q P L Left Right

Stimulus�

Response

PMC

PFCHCMP

A�

Leftto X

Rightto X

Figure 1: Neural network model of the AX-CPT task, showing roles of PMC (posterior& motor cortex), PFC(pre-frontalcortex) andHCMP (hippocampusandrelatedstructures:not actually implementedin our modelsyet).Activation is shown by black units, and the weightsbetweensuchunits arehighlightedto emphasizethe flow ofinformation throughthe network. Lateral inhibition exists within eachof the layers. a) The cue stimulus‘A’ ispresented,resultingin the activation of a PFC representationfor “Respondwith left (target) handif an X comesnext.” b) During thedelay, thePFCrepresentationis actively maintained,providing top-down supportfor thetargetinterpretationof the ‘X’ stimulus.c) Whenthe ‘X’ comes,it resultsin a target(left) response(whereasa non-target(right) responsewould have occurredwithout the top-down PFCactivationduring thedelayperiod). In early trials,theHCMPprovidesappropriateactivationto varioustaskcomponents,asa resultof its ability for rapidlearning.

neuralassemblies),eachof which is capableof performinglocal processingandmemory, andwhicharegroupedinto systems.

Systemsarecomposedof groupsof relatedelements,thatsubserve a similar setof processingfunctions.Systemsmaybedefinedanatomically(by patternsof connectivity) and/orfunctionally(by specializa-tion of representationor function).

Specialization arisesfrom parametricvariationsin propertiespossessedby all elementsin thebrain(e.g.,patternsof connectivity, time constants,regulation by neuromodulatorysystems,etc.). Parametervariationsoccur along continuousdimensions,and thus subsystemspecializationcan be a gradedphenomenon.

Knowledge is encodedin the synapticconnectionstrengths(weights) betweenneurons,which typicallychangeslowly comparedwith thetimecourseof processing.Thismeansthatneuronshave relativelystable(dedicated) representationsover time.

Cognition resultsfrom activationpropagation throughinterconnectednetworks of neurons. Activity isrequiredto directly influenceongoingprocessing.

Learning occursby modifyingweightsasa functionof activity (which canconvey error andreward feed-backinformationfrom theenvironment).

Memory is achievedeitherby therelatively short-termpersistenceof activationpatterns(activememory)or longer-lastingweightmodifications(weight-basedmemory).

Representationsare distributed over many neuronsand brain systems,and at many different levels ofabstractionandcontextualization.

Inhibition betweenrepresentationsexistsat all levels,bothwithin andpossiblybetweensystems,andin-creasesasa (non-linear)function of the numberof active representations.This resultsin attention


System Function InternalRelation ExternalRelation Act Capacity LearnRatePMC inference,processing distributed,overlapping embedded many slowPFC maintenance,control isolated,combinatorial global few slowHCMP rapidlearning separated,conjunctive context sensitive one fast

Table1: Critical parameterizationsof the threesystems.Function specifiesthe functionoptimizedby this system.Inter nal Relation indicateshow representationswithin eachsystemrelateto eachother. External Relation indicateshow representationsrelateto othersystems.Act Capacity indicateshow many representationscanbeactive at anygiventime. Learn Rate indicatescharacteristicrateof learning.Seetext for fuller description.

phenomena,andhasimportantcomputationalbenefitsby enforcingrelatively sparselevelsof activa-tion.

Recurrence (bidirectionalconnectivity) existsamongtheelementswithin a systemandbetweensystems,allowing for interactivebottom-upandtop-down processing,constraint satisfactionsettling,andthecommunicationof errorsignalsfor learning.

A centralfeatureof this framework, asoutlinedabove, is that differentbrain systemsarespecializedfor differentfunctions. In orderto characterizethesespecializations(andunderstandwhy they may havearisen),wefocusonbasictradeoffs thatexist within thiscomputationalframework (e.g.,activity- vsweight-basedstorage,or rapid learningvsextractionof regularities). Thesetradeoffs leadto specialization,sincea homogeneoussystemwould requirecompromisesto bemade,whereasspecializedsystemsworking to-gethercanprovide thebenefitsof eachendof thetradeoff without requiringcompromise.Thisanalysishasled to thefollowing setof coincidentbiologicalandfunctionalspecializations,which arealsosummarizedin Figure2 andTable1:

Posterior perceptual and motor cortex (PMC): Optimizesknowledge-dependentinferencecapabilities,which dependon denseinterconnectivity, highly distributed representations,and slow integrativelearning(i.e., integratingover individual learningepisodes)in orderto obtaingoodestimatesof theimportantstructural/statistical propertiesof theworld, uponwhich inferencesarebased(McClelland,McNaughton,& O’Reilly, 1995). Similarity-basedoverlapamongdistributedrepresentationsis im-portantfor enablinggeneralizationfrom prior experienceto new situations.Thesesystemsperformsensory/motorandmoreabstract,multi-modalprocessingin a hierarchicalbut highly interconnectedfashion,resultingin theability to perceive andact in theworld in accordancewith its salientandre-liable properties.We take this to bethecanonicaltypeof neuralcomputationin thecortex, andviewtheothersystemsin referenceto it.

Pre-frontal cortex (PFC): Optimizesactivememoryvia restrictedrecurrentexcitatoryconnectivity andanactive gatingmechanism(discussedbelow). This resultsin theability to bothflexibly updateinter-nal representations,maintaintheseover time andin the faceof interferenceand,by propagationofactivationfrom theserepresentations,biasPMCprocessingin a task-appropriatemanner. PFCis spe-cializedbecausethereis afundamentaltradeoff betweentheability to actively sustainrepresentations(in theabsenceof enduringinput or thepresenceof distractinginformation)anddenseinterconnec-tivity underlyingdistributed(overlapping)representationssuchasin thePMC (Cohenet al., 1996).As a result,the individual self-maintainingrepresentationsin PFCmustbe relatively isolatedfromeachother(asopposedto distributed). They canthusbeactivatedcombinatoriallywith lessmutualinterferenceor contradiction,allowing for flexible andrapid updating.Becausethey sit high in thecorticalrepresentationalhierarchy, they arelessembeddedandmoregloballyaccessibleandinfluen-tial. Becausethey areactively maintainedandstronglyinfluencecognition,only a relatively smallnumberof representationscantypically beconcurrentlyactive in thePFCatagiventime in orderfor


PMC PFC

HCMP

Figure2: Diagramof key propertiesof thethreeprincipalbrainsystems.Active representationsareshown in grey;highly overlappingcircles are distributed representations;non-overlappingare isolated;in betweenare separated.Weightsbetweenactive units are shown in solid lines; thoseof non-active units in dashed. Threeactive featurevaluesalongthreeseparate“dimensions”(e.g.,modalities)arerepresented.PMC representationsaredistributedandembeddedin specialized(e.g.,modality specific)processingsystems.PFC representationsare isolatedfrom eachother, andcombinatorial,with separateactive unitsrepresentingeachfeaturevalue.Unlike othersystems,PFCunitsarecapableof robustself-maintenance,asindicatedby therecurrentconnections.HCMP representationsaresparse,separated(but still distributed),andconjunctive,sothatonly a singlerepresentationis activeata time,correspondingto theconjunctionof all active features.

coherentcognitionto result.Thus,inhibitory attentionalmechanismsplay animportantrole in PFC,andfor understandingtheorigin of capacityconstraints.

Hippocampusand relatedstructures(HCMP): Optimizes rapid learning of arbitrary information inweight-basedmemories.Thispermitsthebindingof elementsof anovel association,includingrepre-sentationsin PFCandPMC,providing a mechanismfor temporarystorageof arbitrarycurrentstatesfor laterretrieval. Thereis a tradeoff betweensuchrapidlearningof arbitraryinformationwithout in-terferingwith prior learning(retroactive interference),andtheability to developaccurateestimatesofunderlyingstatisticalstructure(McClellandetal.,1995).To avoid interference,learningin theHCMPusespatternseparation (i.e., individual episodesof learningareseparatedfrom eachother),asop-posedto the integrationcharacteristicof PMC. This separationprocessrequiressparse, conjunctiverepresentations,whereall theelementscontribute interactively (not separably)to specifyinga givenrepresentation(O’Reilly & McClelland,1994). This conjunctivity is theoppositeof thecombinato-rial PFC,wheretheelementscontributeseparably. Conjunctivity leadsto context specificandepisodicmemories,which bind togethertheelementsof a context or episode.This alsoimplies that thereisa singleHCMPrepresentation(consistingof many active neurons)correspondingto anentirepatternof activity in thecortex. Sinceonly oneHCMP representationcanbeactive at any time, reactivationis necessaryto extractinformationfrom multiplesuchrepresentations.

While thereareundoubtedlymany otherimportantspecializedbrainsystems,we think that thesethreeprovide central,andcritical contributionsto working memoryfunction. However, brainstemneuromodula-tory systems,suchasdopamineandnorepinephrine,play an importantsupportingrole in our theory, asaresultof theircapacityto modulatecorticalprocessingaccordingto reward,punishment,andaffectivestates.In particular, aswe discussfurtherbelow, we have hypothesizedthatdopamineactivity playsa critical rolein working memoryfunction,by regulatingactive maintenancein PFC(Cohen& Servan-Schreiber, 1992;Cohenetal., 1996).


It shouldalsobeemphasizedthattheabovearerelatively broadcharacterizationsof largebrainsystems,which (especiallyin the caseof the neocorticalsystems)may have subsystemswith different levels ofconformanceto thesegeneralizations.Further, theremay be other importantdifferencesbetweenthesesystemsthatarenot reflectedin ouraccount.Nevertheless,thesegeneralizationsareconsistentwith a largecorpusof empiricaldataandideasfrom othertheorists(e.g.,Fuster, 1989;Shallice,1982;Goldman-Rakic,1987;Squire,1992).Finally, wenotethattherearestill importantportionsof this accountthathave notyetbeenimplementedin computationalmodels,andthesufficiency of theseideasto performcomplex cognitivetasks,especiallythoseinvolving extendedsequentialbehavior, remainsuntestedasof now. Nevertheless,encouragingprogresshasbeenmadeimplementingandtestingmodelssomeof themorebasicfunctionswehave described,suchastheactive maintenancefunctionof PFCandthebindingfunctionof hippocampus(seeCohenet al.,1996;McClellandet al.,1995;Cohen& O’Reilly, 1996for reviews).

In whatfollows, we will elaboratethewaysin which thesebrainsystemsinteractto producecontrolledprocessingandworking memory, andmake morecleartheir relationshipto otherconstructssuchascon-sciousnessandactive memory. Wewill thenfocusonasetof importantissuessurroundingtheoperationofthePFCactive memorysystem,followedby anapplicationof theseideasto understandingsomestandardworkingmemorytasks.This thenprovidesa sufficient setof principlesto addressthetheoreticalquestionsposedin thisvolume.

ControlledProcessingandBrain SystemInteractions

We considercontrolledprocessingto be an importantaspectof our theoryof working memory. Thishasclassicallybeendescribedin contrastwith automaticprocessing(Shiffrin & Schneider, 1977;Posner&Snyder, 1975),andhasbeenthoughtto involve a limited capacityattentionalsystem.However, morerecenttheorieshave suggestedthata continuummayexist betweenbetweencontrolledandautomaticprocessing,(Kahneman& Treisman,1984;Cohen,Dunbar, & McClelland,1990),andweconcurwith thisview. Thus,working memoryalsovariesalongthis samecontinuum.In particular, we have conceptualizedcontrolledprocessingastheability to flexibly adaptbehavior to thedemandsof particulartasks,favoring theprocess-ing of task-relevant informationover othersourcesof competinginformation,andmediatingtask-relevantbehavior overhabitualor otherwiseprepotentresponses.In ourmodels,thisis operationalizedastheuseandupdatingof actively maintainedrepresentationsin PFCto biassubsequentprocessingandactionselectionwithin PMCin atask-appropriatemanner. For example,in theAX-CPT modeldescribedabove, thecontextrepresentationactively maintainedin PFCis ableto exert controlover processingby biasingthe responsemadeto anambiguousprobestimulus.

While it is temptingto equatecontrolledprocessingwith theoreticalconstructssuchasa centralexec-utive (Gathercole,1994;Shiffrin & Schneider, 1977),therearecritical differencesin theassumptionsandcharacterof thesemechanismsthathave importantconsequencesfor our modelof working memory. Per-hapsthe most importantdifferencebetweenour notion of controlledprocessingandtheoriesthat posit acentralexecutive is that we view controlledprocessingasemerging from the interactionsof several brainsystems,ratherthan the operationof a single,unitary CPU-like construct. We believe that our interac-tive, decentralizedview is moreconsistentwith thegradedaspectof controlledprocessing,aswell asthecharacterof neuralarchitectures.However, aspectsof our theoryarecompatiblewith othermodels. Forexample,Shallice(1982)hasproposedatheoryof frontal function,andtheoperationof acentralexecutive,in termsof asupervisoryattentionalsystem(SAS).Hedescribesthisusingaproductionsystemarchitecture,in which theSASis responsiblefor maintaininggoalstatesin working memory, in orderto coordinatethefiring of productionsinvolved in complex behaviors. This is similar to therole of goalstacksandworkingmemoryin ACT (Anderson,1983;Lovettetal., thisvolume).Similarly, our theoryof workingmemoryandcontrolledprocessingdependscritically on actively maintainedrepresentations(in PFC).This centralrolefor active maintenancein achieving controlledprocessingcontrastswith a view whereactive maintenance


Fam

iliar

(Hip

poca

mpu

s I

ndep

ende

nt)

N

ovel

(Hip

poca

mpu

s

Dep

ende

nt)

Bin

ding

Bias

Automatic

Controlled

Familiar, sustainedor weak processing

Novel, sustainedor weak processing

Weak/Sustained(PFC Dependent)

Novel, transientprocessing

Routine, transientprocessing

Strong/Transient(PFC Independent)

Figure3: Waysin which theHCMP andPFCcontributeto theautomaticvscontrolled-processingdistinction(afterCohen& O’Reilly, 1996).Bias is providedby thePFC,andcanbeusedto performsustainedprocessing,canfacilitatethe processingof weakly-learned(i.e., relatively infrequent)tasks,and can serve to coordinateprocessingacrossdifferentsystems.Binding is provided by the HCMP, andcan be usedto rapidly learnandstorethe informationnecessaryto performnovel tasksor processing.Controlledprocessingcaninvolveeitheror bothof thesecontributions,while automaticprocessingcanbeperformedindependentof them.

andexecutive controlarestrictly segregated(Baddeley & Logie,thisvolume).

We considercontrolledprocessingto arisefrom theinterplaybetweenPFCbiasingandHCMP binding(Cohen& O’Reilly, 1996). Figure3 illustratesthe centralideasof this account,which is basedon thefunctionalcharacterizationsof thePFCandHCMP asdescribedabove. Accordingto this view, thedegreeto which controlledprocessingis engagedby a taskis determinedby theextent to which eitheror bothofthefollowing conditionsexist:

� Sustained,weakly-learned(i.e., relatively infrequent),or coordinatedprocessingis required.

� Novel informationmustberapidlystoredandaccessed.

SincethePFCcanbiasprocessingin therestof thesystem,sustainedactivity of representationsin PFCcanproducea focusof activity amongrepresentationsin PMC neededto performa given task. This canbe usedto supportrepresentationsin PMC over temporallyextendedperiods(e.g., in delayedresponsetasks),and/orweakly-learnedrepresentationsthat might otherwisebe dominatedby strongerones(e.g.,in the Strooptask,wherehighly practicedword-readingdominatesrelatively infrequentcolor naming—Cohenetal.,1990).This functionof PFCcorrespondscloselyto Engle,Kane,andTuholski’s (thisvolume)notion of controlledattentionandto Cowan’s (this volume)notion of focusof attention. In contrast,theHCMP contributestheability to learnnew informationrapidly andwithout interference,binding togethertask-relevant information(e.g.,taskinstructions,particularcombinationsof stimuli, intermediatestatesofproblemsolutions,etc) in sucha form thatit beretrievedat appropriatejuncturesduringtaskperformance.Thismayberelevantfor EricssonandDelaney’s (thisvolume)notionof longtermworkingmemory, YoungandLewis’ (this volume)productionlearningmechanism,andMoscovitch andWinocur’s (1992)notionof“working-with-memory”.Weproposethatthecombinationof thesetwo functions(PFCbiasingandHCMPbinding) canaccountfor the distinctionbetweencontrolledand automaticprocessing.On this account,


automatic� processingis whatoccursvia activation propagationthroughintrinsic PMC connectivity, whilecontrolledprocessingreflectstheadditionalconstraintson theflow of activity broughtto bearby thePFCand/orHCMP.

ActivationPropagationandMultiple Constraint Satisfaction

While someaspectsof behavior canbeunderstoodin termsof relatively localprocesseswithin thebrain,weassumethat,undermostcircumstances,behavior is determinedby arich anddynamicsetof interactionsinvolving thewidespreadpropagationof activationto multiple,distributedbrainsystems.While thedetailedoutcomeof suchprocessingin a particularcasemaybedifficult, if not ultimately impossible,to describe,its generalcharactercan be understoodin termsof multiple constraint satisfaction: the activation statethat resultsfrom this propagationof informationover weightedneuronalconnectionsis likely to be onethatsatisfiesvariousconstraints,includingthoseimposedby threecritical components:1) externalstimuli;2) sustainedactivity in PFC;and3) recalledinformationfrom the HCMP. Thus, representationsin PFCandHCMP act as“control signals,” insofar as theseinfluencethe flow of activity andtherebyshapetheconstraintsatisfactionprocessthatis takingplacein therestof thebrain.Furthermore,theiractivationstatesarethemselvesinfluencedby similarconstraintsatisfactionmechanismsbasedonactivationsfrom thePMC(thougha presumedgatingmechanismin the PFCcanmake it moreor lesssusceptibleto this “bottom-up” influence— seediscussionbelow). All of theseconstraintsaremediatedby thesynapticconnectionsbetweenneurons,whichareadaptedthroughexperiencein suchawayasto resultin betteractivationstatesinsimilarsituationsin thefuture.Thus,muchof therealwork beingdonein ourmodel(andouravoidanceof ahomunculusor otherwiseunspecifiedcentralexecutivemechanismswhenwediscusscontrolledprocessing),lies in theseactivation dynamicsandtheir tuning asa function of experience.Computationalmodelsareessentialin demonstratingtheefficacy of thesemechanisms,whichmayotherwiseappearto havemysteriousproperties.

AccessibilityandConsciousness

In our model,oneof the dimensionsalongwhich brain systemsdiffer is in the extent to which theirrepresentationsare globally accessibleto a wide rangeof otherbrain systems,asopposedto embeddedwithin morespecificprocessingsystemsand lessglobally accessible.We view this differenceasarisingprincipally from a system’s relative positionwithin anoverall hierarchyof abstractnessof representations.Thishierarchyis definedby how farremovedasystemis from directsensoryinputor motoroutput.Systemssupportinghigh level, moreabstractrepresentationsaremorecentrallylocatedwith respectto the overallnetwork connectivity, resultingin greateraccessibility. Accordingly, becauseboththePFCandHCMP areat the top of the hierarchy(Squire,Shimamura,& Amaral,1989;Fuster, 1989)they aremoreinfluentialand accessiblethan subsystemswithin PMC. Like the other dimensionsalong which thesesystemsarespecialized,we view this asa gradedcontinuum,andnot asan all-or-nothing distinction. Furthermore,we assumethat the PMC hasrich “lateral” connectivity betweensubsystemsat the samegenerallevel ofabstraction(at leastbeyondthefirst few levelsof sensoryor motorprocessing).Nevertheless,thePFCandHCMPassumeapositionof greateraccessibility, andthereforegreaterinfluence,relative to othersystems.

Accessibilityhasmany implications,which relateto issuesof consciousawareness,andpsychologicaldistinctionslike explicit vs implicit or declarativevs procedural. We view thecontentsof consciousexpe-rienceasreflectingtheresultsof globalconstraintsatisfactionprocessingthroughoutthebrain,with thosesystemsor representationsthataremostinfluentialor constrainingon thisprocesshaving greaterconscioussalience(c.f., Kinsbourne,1997). In general,this meansthathighly accessibleandinfluentialsystemslikePFCandHCMP will tendto dominateconsciousexperienceover the moreembeddedsubsystemsof thePMC. Consequently, thesesystemsaremostclearlyassociatedwith notionsof explicit or declarative pro-cessing,while thePMC andsubcorticalsystemsareassociatedwith implicit or proceduralprocessing.We


endorse this distinction,but addthe importantcaveatthatPFCandHCMP areparticipantsin anextendedinteractive system,andthat,onceagain,suchdistinctionsshouldbeconsideredalonga continuum.Thus,our theory is not compatiblewith strongassumptionsaboutinformationallyencapsulatedmodules(e.g.,Fodor, 1983;Moscovitch & Winocur, 1992).

ActiveMemoryvsWorkingMemory

In our model,we usethe termactive memoryto refer to informationthat is representedasa patternofactivity (neuronalspiking)acrossa setof units(neuralassembly)which persistsover some(possiblybrief)time interval. We view working memoryasrelying on active memory, by virtue of theneedto rapidly andfrequentlyaccessstoredinformationovershortintervalsandusethisinformationto biasprocessingin anon-goingway in otherpartsof thesystem.However, theHCMP, becauseit is capableof rapidly formingnovelassociationsandretrieving thesein task-relevant contexts, is alsousefulfor working memory. Conversely,we do not assumethat actively maintainedrepresentationsare invoked exclusively within the context ofworking memory. Sustainedactivity canoccurandplay a role in automaticprocessingaswell. For exam-ple,it is notdifficult to imaginethatrelatively automatictaskssuchastypingwouldrequirepersistentactiverepresentations,andsustainedactivity hasindeedbeenobserved in areasoutsideof PFC(Miller & Desi-mone,1994).We assumethatactively maintainedrepresentationsparticipatein working memoryfunctiononly underconditionsof controlledprocessing— that is, whensustainedactivity is theresultof represen-tationscurrentlybeingactively maintainedin thePFC,or retrievedby theHCMP. Thiscorrespondsdirectlyto thedistinctions,proposedby Cowan(this volume)andEngleet al. (this volume),betweencontrolledorfocusedattentionvsothersourcesof activationandattentionaleffects.

Regulationof ActiveMemory

It haslongbeenknown from electrophysiological recordingsin monkeysthatPFCneuronsremainactiveover delaysbetweena stimulusanda contingentresponse(e.g.,Fuster& Alexander, 1971). Furthermore,while suchsustainedactivity hasbeenobservedin areasoutsideof PFC,it appearsthatPFCactivity is robustto interferencefrom processinginterveningdistractorstimuli, while activity within thePMC is not (Miller,Erickson,& Desimone,1996;Cohen,Perlstein,Braver, Nystrom,Jonides,Smith,& Noll, 1997).Althoughtheprecisemechanismsresponsiblefor activemaintenancein PFCarenotyetknown, onelikely mechanismis strongrecurrentexcitation. If groupsof PFCneuronsarestronglyinterconnectedwith eachother, thenstrongmutualexcitation will leadto both sustainedactivity, andsomeability to resistinterference.Thisideahasbeendevelopedin anumberof computationalmodelsof PFCfunction(e.g.,Dehaene& Changeux,1989;Zipser, Kehoe,Littlewort, & Fuster, 1993).However, webelieve thatthissimplemodelis inadequateto accountfor bothrobustactive maintenance,andthekind of rapidandflexible updatingthat is necessaryfor complex cognitive tasks.

Theunderlyingproblemreflectsabasictradeoff — to theextentthatunitsaremadeimperviousto inter-ference(i.e., by makingtherecurrentexcitatoryconnectionsstronger),this alsopreventsthemfrom beingupdated(i.e., new representationsactivatedandexisting onesdeactivated). Conversely, weaker excitatoryconnectivity will makeunitsmoresensitive to inputsandcapableof rapidupdating,but will notenablethemto be sustainedin the faceof interference.To circumvent this tradeoff, we think that the PFChastakenadvantageof midbrainneuromodulatorysystems,which canprovide a gating mechanismfor controllingmaintenance.Whenthegateis opened,thePFCrepresentationsaresensitive to their inputs,andcapableofrapidupdating.Whenthegateis closed,thePFCrepresentationsareprotectedfrom interference.Suchagatingmechanismcanaugmentthecomputationalpowerof recurrentnetworks(Hochreiter& Schmidhuber,1997),andwe have hypothesizedthatdopamine(DA) implementsthis gatingfunctionin PFC,basedon asubstantialamountof biologicaldata(Cohenet al., 1996).


synthesizer

tele

visi

on

a) b)

?

?

?terminal

key−board

spea−ker

mon−itor

+

+

+

key−board

spea−ker

mon−itor

Figure4: Illustrationof difficultieswith active maintenancevia recurrentexcitationwith distributedrepresentations.a) No valueof the excitatory weightswill enablean appropriatesubsetof two featuresto be maintained,withoutalsoactivating the third. b) If representationsaremadeindependent,thenmaintenanceis no problem,but semanticrelatednessof thefeaturesis lost. Onecouldalsojust maintainthehigher-level items(i.e., synthesizer, terminalandtelevision).

Thus,we proposethat themidbrainDA nuclei (theventraltegmentalarea,VTA), undercontrolof de-scendingcortical projections,enablethe PFC to actively regulatethe updatingof its representationsbycontrollingthe releaseof DA in a strategic manner. Specifically, we proposethat theafferentconnectionsinto the PFCfrom otherbrain systemsareusuallyrelatively weakcomparedto strongerlocal excitation,but thatDA enhancesthestrengthof theseafferents1 at timeswhenupdatingis necessary. This would pre-dict that theVTA shouldexhibit phasicfiring at thosetimeswhenthePFCneedsto beupdated.Schultz,Apicella,andLjungberg (1993)have foundthatindeed,theVTA exhibitstransient,stimulus-lockedactivityin responseto stimuli which predictedsubsequentmeaningfulevents(e.g.,reward or othercuesthat thenpredictreward). Further, we arguethatthis role of DA asa gatingmechanismis synergistic with its widelydiscussedrole in reward-basedlearning(e.g.,Montague,Dayan,& Sejnowski, 1996).As will bediscussedfurtherin thefinal discussionsection,this learninghelpsusto avoid theneedto postulateahomunculus-likemechanismfor controlledprocessing.

General Natureof ActiveMemoryRepresentations

Ourtheoryplacesseveralimportantdemandsonthenatureof representationswithin thePFC,in additionto therapidlyupdatableyetrobustactivemaintenancediscussedabove. In general,weview thePFC’sroleincontrolledprocessingasimposingasustained,taskrelevant,top-down biasonprocessingin thePMC.Thus,in complex cognitiveactivities,thePFCshouldbeconstantlyactivatinganddeactivatingrepresentationsthatcanbiasa largenumberof combinationsof PMC representations,while sustaininga coherentandfocusedthreadof processing.ThismeansthatthePFCneedsavastrepertoireof representationsthatcanbeactivatedondemand,andtheserepresentationsneedto beconnectedwith thePMCin appropriateways.Further, theremustbesomewayof linking togethersequencesof representationsin acoherentway.

Our initial approachtowardsunderstandingPFC representationshasbeendominatedby an interest-ing coincidencebetweentheabove functionalcharacterizationof thePFC,anda consequenceof anactivemaintenancemechanismbasedon recurrentexcitation. Distributedrepresentations,which arethoughttobecharacteristicof thePMC, areproblematicfor this kind of active maintenancemechanism,asthey relycritically onafferentinput in orderto selecttheappropriatesubsetof distributedcomponentsto beactivated.

1In addition,it appearsthat inhibitory connectionsarealsoenhanced,which would provide a meansof deactivating existingrepresentations.


In

theabsenceof this afferentselection(e.g.,duringa delayperiod),recurrentexcitationamongthecom-ponentswill spreadinappropriately, andresultin thelossof theoriginal activity pattern.This is illustratedin Figure4, whereadistributedrepresentationis usedto encodethreedifferentitems,whicheachsharetwoout of threetotal features.If thesedistributionshave thestrongrecurrentexcitatoryconnectionsnecessaryfor activemaintenance,thenit will bedifficult to keepauniquesubsetof two featuresactivewithoutalsoac-tivatingthethird: activationwill spreadto thethird unit via theconnectionsnecessaryto maintainit in othercircumstances.The alternative (shown in panelb) is to useisolatedrepresentations,which maintainonlythemselves. However, what is missingfrom theseisolatedrepresentationsis therich interconnectivity thatencodesknowledgeaboutthe relationshipsbetweenthe features,which could be usedfor performingtheknowledge-dependentinferencethatwe think is characteristicof thePMC. We obtaintheoreticalleveragefrom this basictradeoff, whichcanbeavoidedby having two specializedsystems(PMCandPFC).

TheideathatPFCrepresentationsarerelatively isolatedfrom eachotherhasimportantfunctionalcon-sequencesbeyond theactive maintenanceof information. For example,in orderto achieve flexibility andgenerativity, PFCrepresentationsmustbeusefulin novel contexts andcombinations.Thus,individualPFCrepresentationsshouldnot interferemuchwith eachothersothatthey canbemoreeasilyandmeaningfullycombined— thisis justwhatonewouldexpectfrom relatively isolatedrepresentations.Wethink thatlearn-ing in thePFCis slow andintegrative like therestof thecortex, so that this graduallearningtakingplaceover many yearsof humandevelopmentproducesa rich anddiversepaletteof relatively independentPFCrepresentationalcomponents,which eventuallyenablethe kind of flexible problemsolving skills that areuniquelycharacteristicof adulthumancognition.

This view of PFCrepresentationscanbeusefullycomparedwith thatof humanlanguage.In termsofbasicrepresentationalelements,languagecontainswords,which have a relatively fixedmeaning,andcanbecombinedin a hugenumberof differentwaysto expressdifferent(andsometimesnovel) ideas.Wordsrepresentthingsat many different levels of abstractionandconcreteness,andcomplicatedor particularlydetailedideascanbe expressedby combinationsof words. We think that similar propertieshold for PFCrepresentations,andindeedthata substantialsubsetof PFCrepresentationsdo correspondwith word-likeconcepts.However, we emphasizethatword meaningshave highly distributedrepresentationsacrossmul-tiple brainsystems(Damasio,Grabowski, & Damasio,1996),andalsothatthePFCundoubtedlyhasmanynon-verbal representations.Nevertheless,it may be that the PFCcomponentof a word’s representationapproachesmostcloselythenotionof adiscrete,symbol-like entity.

Takingthis languageideaonestepfurther, it mayprovide usefulinsightsinto thekindsof updatingandsequentiallinking that PFCrepresentationsneedto undergo during processing.For example,languageisorganizedat many different levels of temporalstructure,from shortphrases,throughlongersentencestoparagraphs,passages,etc. Theselevels aremutually constraining,with phrasesaddingtogetherto buildhigher-level meaning,andthis accumulatedmeaningbiasingtheinterpretationof lower-level phrases.Thissameinteractive hierarchicalstructureis presentduringproblemsolvingandothercomplex cognitive activ-ities,andis critical for understandingthedynamicsof PFCprocessing.We think thatall of thesedifferentlevels of representationcanbe active simultaneously, andmutually constrainingeachother. Further, it ispossiblethat the posterior-anteriordimensionof the PFCmay beorganizedroughlyaccordingto level ofabstraction(andcorrespondingly, temporalduration).For example,thereis evidencethatthemostanteriorareaof PFC,thefrontal pole,is only activatedin morecomplicatedproblemsolvingtasks(Baker, Rogers,Owen,Frith, Dolan,Frackowiak, & Robbins,1996),andthatposteriorPFCreceivesmostof theprojectionsfrom PMC,andthenprojectsto moreanteriorregions(Barbas& Pandya,1987,1989).Finally, thisnotionofincreasinglyabstractlevelsof planor internaltaskcontext is consistentwith theprogressionfrom posteriorto anteriorseenwithin themotorandpremotorareasof the frontal cortex (Rizzolatti,Luppino,& Matelli,1996).


Specialization�

of ActiveMemoryRepresentations

An importantissuebothwithin theworkingmemoryliteratureandwith regardto theoriesof PFCfunc-tion is that of specializationalongfunctionaland/orrepresentationaldimensions.For example,Baddeley(1986)proposedthattherearetwo separateworkingmemorybuffers: a phonological loop anda visuospa-tial scratchpad, whichmight itself besubdividableinto objectandspatialcomponents.It hasbeenproposedthat this functionalspecializationreflectsanunderlyingspecializationin thebrainsystemssubservingthedifferentbuffer systems(Gathercole,1994). For example,theremaybespecificbrainsystemssubservingverbalrehearsal(e.g.,Broca’s areaand/orangulargyrus). Thereis alsoa well recognizedsegregationofprocessingof objectandspatialinformationinto ventral(temporal)anddorsal(parietal)streamswithin thePMC(Ungerleider& Mishkin, 1982),thatmaycorrespondto thetwo subdivisionsof Baddeley’s visuospa-tial sketchpad.At a somewhatmoregenerallevel, ShahandMiyake (1996)have foundevidenceconsistentwith theideaof separablespatialandverbalcapacities.

Specializationmayalsoplay a role in PFCorganizationandfunction. Argumentsin the literaturehavecenteredaroundtwo dimensionsalongwhich PFCmaybeorganized:functionalandcontent-based.How-ever, we arguethat it is difficult to draw a cleandistinctionbetweenfunctionandcontent.Indeed,a basicprinciple of neuralnetworks is that processingandknowledge(content)are intimately intertwined. Forexample,thefunctionaldistinctionbetweenmemory(in thedorsolateralPFC)andinhibition (in theorbitalareas;Fuster, 1989;Diamond,1990),could alsobe explainedby a content-baseddistinction in termsoftherepresentationof affective, appetitive andsocialinformationin theorbital areas,which might bemorefrequentlyassociatedwith theneedfor behavioral inhibition. Similarly, thefunctionaldissociationbetweenmanipulation(in dorsalateralareas)andmaintenancein (ventrolateralareas;Petrides,1996)may be con-foundedwith the needto representsequentialorder informationin mosttasksthat involve manipulation.To furthercomplicatethe issue,onetypeof functionalspecializationcanoftengive rise to otherapparentfunctionalspecializations.For example,we have arguedthatthememoryandinhibitory functionsascribedto PFCmaybothreflecttheoperationof asinglemechanism(i.e., inhibition canresultfrom maintainedtop-down activation of representationswhich then inhibit othercompetingpossibilitiesvia lateral inhibition;Cohen& Servan-Schreiber, 1992;Cohenetal., 1996).

Given theseproblems,we find it moreuseful to think in termsof the computationalmotivationsde-scribedabove in orderto understandhow thePFCis specialized(i.e., in termsof thetradeoff betweenactivemaintenanceanddistributedrepresentations).Thus,we supporttheideathat thePFCis specializedfor thefunction of active maintenance.Consequently, representationswithin the PFCmustbe organizedby thecontentof therepresentationsthataremaintained.A numberof neurophysiologicalstudieshave suggesteda content-basedorganizationthatreflectsananteriorextensionof theorganizationfoundin thePMC,withdorsalregionsrepresentingspatialinformation(Funahashi,Bruce,& Goldman-Rakic,1993)andmoreven-tral regionsrepresentingobjector patterninformation(Wilson,Scalaidhe,& Goldman-Rakic,1993).Othercontentdimensionshavealsobeensuggested,suchassequentialorderinformation(Barone& Joseph,1989),and“dry” cognitive vsaffective,appetitive and/orsocialinformation(Cohen& Smith,1997).However, thedatadoesnot consistentlysupportany of theseideas. For example,Rao,Rainer, andMiller (1997)haverecordedmorecomplex patternsof organizationin neurophysiologicalstudies,with significantdegreesofoverlapandmultimodality of representations.Recentfindingswithin the humanneuroimagingliteraturearealsoconfusing,asearlyreportsthat indicateddistinctionsin theareasactivatedby verbalvsobjectandverbalvs spatialinformation(e.g.,Smith,Jonides,& Koeppe,1996)have not beenreliably replicated(asreportedin anumberof recentconferenceproceedingsandin unpublisheddatafrom our lab).

In light of thisdata,we suggestthatthePFCmaybeorganizedaccordingto moreabstract,multimodal,andlessintuitive dimensionsthathave beenconsideredto date(i.e., thatdo not correspondsimply to sen-sory modalitiesor dimensions).This seemslikely, given the relatively high-level positionof the PFCintheprocessinghierarchy(seediscussionabove), which would give it highly processedmulti-modalinputs.


Further�

, this typeof input mayinteractwith thelearningmechanismsandotherconstraintson thedevelop-mentof representationswithin PFC.For example,wehaveshown thattaskdemandsandtrainingparameters(i.e., blockedvs interleaved exposure)canplay an importantrole in determiningwhethera simulatedPFCdevelopsuni- or multimodalrepresentationsof objectandspatialinformation(Braver & Cohen,1995).

ExampleWorkingMemoryTasks

Earlier, we provided an exampleof how the mechanismswe have proposedareengagedin a simpleworkingmemorytask(theAX-CPT). Here,weconsiderhow they maycomeinto play in two tasksthatarecommonlyusedto measureworking memorycapacity, andcontrastthemwith onesthatarethoughtto notinvolve working memory. The verbalworking memoryspantask(Daneman& Carpenter, 1980)involvesreadingaloudasetof sentencesandrememberingthefinal wordsfrom eachsentencefor laterrecall.Thus,thesefinal wordsmustbemaintainedin the faceof subsequentprocessing,which makesthis taskheavilydependentontherobustPFCactivemaintenancemechanisms.A spatialversionof thistask(Shah& Miyake,1996)involvesidentifying letterspresentedat differentnon-uprightorientationsasbeingeithernormalormirror-reversed,which appearsto requiresomeamountof mentalrotationto theuprightorientation,whilerememberingthe orientationsof the lettersfor later recall. This mentalrotation requiresthe driving ofPMC-basedvisual transformations(learnedover extensive experienceseeingvisual transformationssuchasrotation,translation,etc.) in a task-relevant manner, presumablyvia actively maintainedPFCtop-downbiasing.Further, theorientationinformationmustbemaintainedin thefaceof subsequentprocessingof thesamekindsof information,whichagainrequiresrobustPFCactive maintenance.

Theseworking memoryspantaskshave beencontrastedwith othersthatarenot consideredto involveworkingmemory. For example,theverbalworkingmemoryspantaskhasbeencomparedwith asimpledigitspantask,whichpresumablyonly requiresactive maintenance,but notcontrolledprocessing.Behaviorally,theverbalworking memoryspantaskis bettercorrelatedwith otherputative verbalworkingmemorytasksthatalsoinvolvecontrolledprocessing,comparedto thissimpledigit spantask(Daneman& Merikle,1996).Theotherhalf of thisargumenthasbeenmadein thecaseof aspatialequivalentof thedigit spantask(whichinvolved rememberingthe orientationsof a setof arrows), that wassignificantlycorrelatedwith a simplevisualprocessingtask,while thespatialworking memoryspanmeasurewasnot (Shah& Miyake, 1996).SeeEngleet al. (this volume)for a moredetaileddiscussionof this issueandotherrelevant experimentalresults.

Anotherexample,involving the useof the HCMP system,is the comprehensionof extendedwrittenpassages.Becauseof limited capacityin the PFC active memorysystem,it is likely that someof therepresentationsactivatedby the comprehensionof prior paragraphsareencodedonly within the HCMP,andmustbe recalledasnecessaryduring laterprocessing(e.g.,whenencounteringa referencelike, “thiswould be impossible,given Ms. Smith’s condition,” which refersto previously introducedinformationthatmaynot have remainedactive in thePFC).The ideais that this later referencecanbeusedto triggerrecallof thepreviousinformationfrom theHCMP, perhapswith theadditionof somestrategic activationofotherrelevant informationwhich haspersistedin the PFC(e.g.,the fact that Ms. Smith livesin Kansas).A successfulrecall of this informationwill result in the activation of appropriaterepresentationswithinthe PFCandPMC, which combinedwith the currenttext resultsin comprehension(e.g.,Ms. Smith washit by a tornado,andcan’t comeinto work for animportantmeeting).In contrastwith theoriesthatdraw astrongdistinctionbetweenactivememoryandHCMPweight-basedmemories(e.g.,Moscovitch & Winocur,1992),we think thata typicalcognitive taskanalysismaynotdistinguishbetweenthesetypesof memoryinmany situations,makingthegenericworking memorylabelmoreappropriatefor both. Finally, YoungandLewis (this volume)presentwhatappearsto bea roughlysimilar role for rapid learningin their theoryofworkingmemory, andEricssonandDelaney (thisvolume)describerelatively long-lastingworkingmemoryrepresentationswhich would seemto involve theHCMP (aswell astheeffectsof extensive experienceon


underlyingcorticalrepresentations).

Answersto TheoreticalQuestions

This sectionsummarizesour answersto a set of eight basicquestionsaboutour theory of workingmemory. Thequestionsaresummarizedby thesectionheaders,andTable2 providesaconcisesummaryofouranswersto thesequestions.

BasicMechanismsandRepresentationsin WorkingMemory

Active maintenance(for which thePFCis specialized),rapidlearning(for which theHCMP is special-ized),andcontrolledprocessing(biasingandbindingbasedon these)arethebasicmechanismsof workingmemoryin our account. Controlledprocessingemergesfrom the interactionsbetweenall threeprimarybrainsystems(PFC,HCMP, PMC), but is moststronglyinfluencedby thePFCandHCMP. For purposesof comparison,we describeour basicmechanismsin termsof standardmemoryterminologyof encoding,maintenance,andretrieval:

Encoding: Due to slow learning,the cortical systems(PFCandPMC) have relatively stablerepresenta-tionalcapability. Thus,encodingin thesesystemsreliesontheselectionandactivation(via constraintsatisfactionprocessingoperatingoverexperience-tunedweights)of thosepre-existingrepresentationsthat are most relevant in a particularcontext. In the HCMP, encodinginvolves the rapid bindingtogetherof a novel conjunctionof the representationsactive in the restof the brain. An importantinfluenceon thisprocess,andacritical componentof controlledprocessing,is thestrategic activation(undertheinfluenceof thePFC)of representationsthatinfluenceHCMPencodingin task-appropriateways(e.g.,activatingdistinctive featuresduringelaborative encodingin amemorytask).

Maintenance: Only thePFCis thoughtto becapableof sustainingactivity over longerdelaysandin thefaceof otherpotentiallyinterferingstimuli or processing.However, underconditionsof shorterde-laysandtheabsenceof interference,PMCcanexhibit sustainedactivememories(Miller etal.,1996).We include in our definition of the PFCthe frontal languageareaswhich have beenshown to beactive in neuroimagingstudiesinvolving active memoryasdiscussedabove. For example,consider-ableevidencesupportsthe ideathatmaintenancein this systemis implementedby anphonologicalloop (Baddeley & Logie, this volume;Baddeley, 1986),which mayinvolve morehighly specializedmechanismsthanthosehypothesizedto exist in otherareasof PFC.We do not think that theHCMPmaintainsinformationin an active form, but ratherthroughrapid weight changesmadeduring en-coding. Theseweight-basedHCMP memoriescan persistover much longer intervals than activememories(c.f., Ericsson& Delaney, thisvolume).

Retrieval: For active memories,retrieval is not anissue,but for HCMP weight-basedmemories,retrievaltypically requiresmultiple cuesto trigger a particularhippocampalmemory(dueto its conjunctivenature).As with encoding,thestrategic activationof suchcuesconstitutesanimportantpartof con-trolledprocessing.

As for the natureof the representationsin our modelof working memory, we have characterizedthedistinctive propertiesof representationsin eachof thethreemainbrainsystems(seeTable1 andFigure2).However, sincewedonotadhereto abuffer-basedor any otherdistinctsubstrateview of workingmemory,this questionis difficult to address.Essentially, thespaceof possibledifferentrepresentationsfor workingmemoryis aslarge asthe spaceof all representationsin the neocortex andhippocampus,sinceany suchrepresentationcouldbeactivatedin a controlledmanner, thussatisfyingourdefinitionof workingmemory.


Ho

wever, we think thatbrainsystemsspecializedfor languagemayprovide anexceptionallypowerful andgeneral-purposerepresentationalsystemfor encodingarbitrary information,and are likely to be usedtoencodeeven superficiallynon-verbal information. Similarly, it may also be that abstractspatialand/ornumericalrepresentationsareusefulfor encodingrelationalandperhapstemporalinformation.

TheControl andRegulationof WorkingMemory

In ourtheory, controlresultsfromthebiasingfunctionof thePFC,andthebindingfunctionof theHCMP.Thesesystems,in turn, areregulatedby eachother, the PMC, andascendingbrainstemneuromodulatorysystems.Thus,control andregulationareinteractive anddistributedphenomena,that involve all partsofthesystem.While theseinteractionsarenecessarilycomplex, it is possibleto identify characteristiccontri-butionsmadeby eachcomponentof the system.The PFCplaysa dominantrole in controlledprocessingby virtue of its characteristicfeatures:its ability to maintainactivation over time; the flexible andrapidupdatingof representationsdueto theircombinatorialandactive nature;andits positionhigh in thecorticalprocessinghierarchy. Notethatunlike modelswhich separatecontrol(e.g.,acentralexecutive) from activestorage(e.g.,buffers),active maintenanceplaysacentralrole in controlin ourmodel.

Representationswithin the PFCarethemselvessubjectto the influenceof processingwithin the PMCandHCMP, by wayof specializedcontrolmechanismsthatregulateaccessto thePFC.As describedabove,we suggestthat the midbrain dopamine(DA) systemprovides an active gating of PFC representations,controllingwhenthey canbeupdated,andprotectingthemfrom interferenceotherwise.We think that thePFC,togetherwith thePMC andpossiblytheHCMP, controlsthefiring of theDA gatingsignal,throughdescendingprojections.Furthermore,we assumethat theseprojectionsaresubjectto learning,so that thePFCandPMCcanlearnhow to controlthegatingsignalthroughexperience.

TheUnitary vs Non-UnitaryNatureof WorkingMemory

We take the view that working memoryis not a unitary construct— insteadwe suggestthat it is thecombinationof activememory, rapidlearning,andemergentcontrolledprocessingoperatingoverdistributedbrain systems. Insteadof the moving of information from long-termmemoryinto and out of workingmemorybuffers,we think that informationis distributedin a relatively stableconfigurationthroughoutthecortex, andthatworkingmemoryamountsto thecontrolledactivationof theserepresentations.As wenotedat the outset,this view sharessomesimilaritieswith the view of working memoryofferedby productionsystemaccounts(e.g.,ACT — Anderson,1983;Lovettetal., thisvolume).However, it doesnot includethestructuraldistinctionbetweendeclarative andproceduralknowledgeassumedby suchaccounts.

This non-unitaryview is consistentwith findingslike thoseof ShahandMiyake (1996),who foundnosignificantcorrelationbetweenan individual’s verbalandspatialworking memorycapacities.We wouldfurther predict that working memorycapacitywill vary alonga variety of dimensions,dependingon thequality of therelevantPMC andPFCrepresentationsdevelopedover experience(c.f., Ericsson& Delaney,this volume). However, working memoryis alsoaffectedby moredomain-general,controlledprocessingmechanisms(suchasthosesupportedby brainstemneuromodulatorysystems),sothatsomecharacteristicsof workingmemoryfunctionmightexhibit moreunitary-like features(c.f., Engleetal., thisvolume).Thus,theactualperformanceof agivensubject,underagivensetof taskconditions,will dependonacombinationof bothdomainspecificandmoregeneralfactors.

TheNatureof WorkingMemoryLimitations

Thepresenceof capacitylimitationsseemsto beoneof thefew pointsaboutwhichthereis consensusintheworkingmemoryliterature.Despitethisagreement,thereis relatively little discussionof whysuchlimi-tationsexist. Are thesetheunfortunateby-productof fundamentallimitationsin theunderlyingmechanisms


(e.g., insufficient metabolicresourcesto sustainadditionalmentalactivity), or do they reflectsomemoreinterestingcomputationalconstraint?We adhereto the latter view. We believe that capacitylimitationsin working memoryreflecta tradeoff betweentwo competingfactors: the accessibility, andwide-spreadinfluenceof PFCrepresentations— necessaryto implementits biasingfunctionasa mechanismof control— andtheneedto constraintheextentof activationthroughoutthePMC, to avoid “runaway” activity, andpromotefocusedandcoherentprocessing.We assumethat this tradeoff is managedby inhibitory mech-anisms,that constrainspreadingactivation, andprevent the runaway activity that would otherwiseresultfrom the positive feedbackloopswithin the cortex. This is particularlyimportantin the PFC,becauseofits widespreadandinfluentialprojectionsto therestof thebrain. We have begunto explorethis possibilityin explicit computationalmodelingwork (Usher& Cohen,1997). This accountemphasizesthe potentialbenefitsof what otherwisemight appearto be arbitrarylimitations (c.f., Lovett et al., this volume). NotethatYoungandLewis (this volume)andSchneider(this volume)presentfunctionalmotivationssimilar toourown.

As we statedabove, we think thereareboth domainspecificandmoregeneralcontributions to work-ing memoryfunction. Similarly, thereare likely to be both experiencedependentandgenetically-basedcontributions. Further, it is likely that thereare interactionsbetweenthesefactors. For example,exten-sive experiencewill producea rich andpowerful setof domain-specificrepresentationsthat supporttheability to encodemoredomain-specificinformationboth in active memoryandvia rapid learningin theHCMP. However, this is unlikely to affectmoregeneralfactors(e.g.,neuromodulatoryfunction,or theover-all level of inhibition within the PFC).This is consistentwith the generallack of cross-domaintransferfrom experience-basedworkingmemorycapacityenhancementsasdiscussedin EricssonandDelaney (thisvolume),andwith therelatively domaingenerallimitationsobservedby Engleetal. (this volume).

The role of experience-basedlearning(which is an importantcomponentof our overall model) in en-hancingdomainspecificworking memorycapacitycanbe illustratedby consideringthe following phasesof experience:

Novel phase: HCMP is requiredto storeand recall novel task-relevant information, so that capacityisdominatedby theconstraintof having only oneHCMPrepresentationactiveatatime,with significantcontrolledprocessingrequiredto orchestratetheuseof this informationwith ongoingtaskprocessing.This is like the first time onetries to drive a car, wherecompleteattentionis required,everythinghappensin slow serialorder, andmany mistakesaremade.

Weakphase: PFCis requiredto biastheweakPMC representationsunderlyingtaskperformance,sothatcapacityis dominatedby therelatively moreconstrainedPFC.Thus,it is difficult to performmultipletasksduringthisphase,or maintainotheritemsin active memory. This is like theperiodafterseveraltimesof driving, whereonestill hasto devote full attentionto the task(i.e., usePFCto coordinatebehavior), but at thebasicoperationsarereasonablyfamiliarandsomecanbeperformedin parallel.

Expert phase: Weightswithin thePMC have beentunedto thepoint thatautomaticprocessingis capableof accomplishingthe task. Sincethe PMC representationsarerelatively moreembedded,they canhappilycoexist with activity in otherareasof PMC, resultingin high capacity. This is thecasewithexpert drivers, who cancarry on conversationsmoreeffectively thannovices while driving. Notethat slow improvementswithin this phaseoccur with continuedpractice,resultingin experience-baseddifferencesin strengthandsophisticationof underlyingrepresentations,which contribute toindividual differencesin capacityand performance.This is true in the PFCaswell, wherefeweractive representationsneedbemaintainedif amoreconcise(e.g.,“chunked”) representationhasbeenlearnedoverexperience.


TheRoleof WorkingMemoryin Complex CognitiveActivities

Complex cognitive activities involve controlledprocessingandthus,by our definition,involve workingmemory. According to our accountof the roles of the PFC (biasing)and HCMP (binding), controlledprocessingoccursunderconditionsof temporallyextendedand/ornovel tasks,andin caseswhich requirecoordinatedprocessingamongmultiplesystems.Typically, complex tasksinvolve thetemporally-extendedcoordinationof multiplestepsof processing,oftenin novel combinationsandsituations,andthestorageofintermediateproductsof computation,subgoals,etc. Active memorytogetherwith thecontrolledencodingandretrieval of HCMPmemoriescanbeusedto retaintheintermediateresultsof theseprocessingstepsforsubsequentuse.

Wehave yet to applyourmodelto specificcomplex tasks,aswehave yet to producesatisfactoryimple-mentationsof theentiresetof neuralsystemsthatwould berequired.Our overarchinggoal in developingsuchmodelsis theability to accountfor complex taskperformancewithout resortingto a homunculusofoneform or another. While many accountsof executive controlremainpurelyverbalandareobviouslysus-ceptibleto thehomunculusproblem,evenmechanisticallyexplicit accountsof complex taskperformanceinproductionsystemarchitectures(e.g.,Young& Lewis, thisvolume;Lovettetal., thisvolume)haveahiddenhomunculusin the form of the researcherwho builds in all theappropriateproductionsin orderto enablethesystemto solve thetask.As wediscussin greaterdetailbelow, ourcurrentmodelingeffortsarefocusedon developinglearningmechanismsthat would give rise to a rich anddiversepaletteof PFCrepresenta-tions(andcorrespondingPMCsubsystems),whichshouldbecapableof performingcomplex taskswithoutresortingto ahomunculusof any form.

TheRelationshipof WorkingMemoryto Long-TermMemoryandKnowledge

We view working memoryasbeingtheactive portionof long-termmemory, wherelong-termmemoryrefersto theentirenetwork of knowledgedistributedthroughoutthecortex, HCMP, andotherbrainsystems.Asnotedearlier, thisis similarto productionsystemtheoriesof workingmemory(suchasACT— Anderson,1983;Lovett et al., this volume). However, we alsospecifythat thetermworking memoryonly appliestothoserepresentationsthatareactivatedasaresultof controlledprocessing.Thus,it is possibleto haveactiverepresentationsthat exist outsideof working memory(c.f. Cowan, this volume;Engleet al., this volume,for similar views). Becauseof this intimaterelationshipbetweenworking memoryandlong-termmemory,we expectworking memoryto beheavily influencedby learningin the long-termmemorysystem(seethediscussionin thecapacitysectionabove andEricsson& Delaney, thisvolume).

Wedonotthink thatall componentsof long-termmemoryareequallylikely to berepresentedin workingmemory. As discussedpreviously, languageprovidesa particularlyusefulmeansof encodingarbitraryin-formation,andis thusheavily involvedin workingmemory. In contrast,moreembedded,low-level sensoryandmotorprocessingis lesslikely to comeundertheinfluenceof controlledprocessing,andis not typicallyconsideredto be involved in working memory. Thus, the generallevel of accessibilityassociatedwith agivenbrainsystemis correlatedwith theextentto which it is likely to beinvolved in working memory. Asdiscussedearlier, thismeansthatmoredeclarativeor explicit longtermknowledgeis likely to beinvolvedinworkingmemory, whereasimplicit or procedural knowledgeis moreassociatedwith automaticprocessing.

TheRelationshipof WorkingMemoryto AttentionandConsciousness

Working memory, attention,andconsciousnessareclearly relatedin importantways. We view theun-derlyingconstraintthatgivesriseto attentionasresultingfrom theinfluenceof competitionbetweenrepre-sentations,implementedby inhibitory interneuronsthroughoutthecortex (andpossiblyalsoby subcorticalmechanismsin the thalamusandbasalganglia). This inhibition providesa mechanismthat causessomethingsto beignoredwhile othersareattendedto, andis a critical aspectof attentionthat is not strictly part


of� working memory. However, assumingthis constraint,controlledprocessingplaysan importantrole indeterminingwhat is active in a given context (andvia competitionandinhibition, alsowhat is ignored).Thus,workingmemoryandattentionarerelatedin thatthey arebothdefinedin partby themechanismsthatdeterminewhatis activatedin aparticularcontext.

Consciousnessis also relatedto both working memoryandattention. As statedpreviously, we viewconsciousexperienceasreflectingtheoutcomeof globalconstraintsatisfactionprocessing,wheresalienceis afunctionof thedegreeof influenceover thisprocessattributableto agivenrepresentation.Thus,systemswhicharegloballyaccessiblelikethePFCandHCMParealsohighly influential,andthuslikely to dominateconsciousexperience.This meansthat the controlledprocessing-basedactivation (attention)mediatedbythesesystemsis mostrelevantfor consciousness,andthatthecontentsof consciousexperiencearelikely toreflectthatof workingmemoryaswehave definedit (seealsoprevioussectionandKinsbourne,1997).

TheBiological Implementationof WorkingMemory

Ourmodelis basedlargely onbiologicaldata,andits neuralimplementationhasbeendescribedbothintermsof basicpropertiessuchasactivation,inhibition, andlearning,andin termsof theinteractionsof thespecializedbrainsystemsdescribedabove (PFC,HCMP, PMC).By virtue of thesebiologicalfoundations,thereis awealthof datawhich is consistentwith ourmodel,from anatomyandphysiologyto neuroimagingandneuropsychological work. Wewill just review someof themostrelevantdatahere.

With respectto theinvolvementof thePFCin workingmemorytasks,our labhasfocusedonneuroimag-ing andschizophrenicpatientperformanceon theAX-CPT taskdescribedin the introduction.By makingthetargetA-X sequencevery frequent(80%),andthedelaybetweenstimuli longer(5 secs),we predictedthatschizophrenicpatientssuffering animpairmentof PFCfunctionwould make a relatively largenumberof falsealarmsto B-X sequences(where‘B’ is any non-‘A’ stimulus)due to a failure of PFC-mediatedworking memoryfor the prior stimulus. This was confirmed,with unmedicatedschizophrenicpatientsshowing thepredictedincreasein falsealarms,while medicatedschizophrenicsandcontrolsubjectsdid not(Servan-Schreiber, Cohen,& Steingard,in press).In addition,neuroimagingof healthysubjectsperformingtheAX-CPT showedthatPFCincreasedactivity with increasesin delayinterval (Barch,Braver, Nystrom,Forman,Noll, & Cohen,1997).NeuroimagingduringN-backperformancerevealedthatPFCactivity alsoincreaseswith workingmemoryload(Braver, Cohen,Nystrom,Jonides,Smith,& Noll, 1997b),andis sus-tainedacrosstheentiredelayperiod(Cohenetal.,1997).Thesedatatogetherwith otherconsistentfindingsfrom monkey neurophysiology(e.g.,Fuster, 1989;Miller et al., 1996), frontally-damagedpatients,(e.g.,Damasio,1985)all supportthe ideathat thePFCis critically importantfor working memory. Also, Engleetal. (this volume)discusstheimportanceof thePFCin workingmemory.

With respectto the role of the HCMP in working memory, it haslong beenknown that the HCMPis critical for learningnew information (Scoville & Milner, 1957; seeSquire,1992; McClelland et al.,1995,for recentreviews). Recentneuroimagingdatasuggeststhatthecontrolledencodingandretrieval ofinformationin theHCMP dependson interactionsbetweenthePFCandtheHCMP (e.g.,Tulving, Kapur,Craik,Moscovitch, & Houle,1994).Further, patientswith frontal lesionsshow impairedability to performstrategic encodingandretrieval on standardmemorytests(Gershberg & Shimamura,1995). All of this isconsistentwith our view that PFCandHCMP interactionsareimportantfor the controlledprocessingofmemorystorageandretrieval, whichcanbeusedasanon-active form of workingmemory.

RecentDevelopmentsandCurrentChallenges

Our theoryof working memoryrepresentsan attemptto understandthis constructin termsof a setofbiologically-based, computationalmechanisms.This hasresultedin a novel set of functionalprinciples


(1) BasicMechanismsandRepresentationsin WorkingMemory:Active memoryandrapid learningvia controlledprocessing,asimplementedby thepre-frontalcor-tex (PFC),hippocampusandrelatedstructures(HCMP),andtheposteriorperceptual& motorcortex(PMC). Representations,distributedthroughoutsystem,areencodedby controlledactivation; main-tainedby robust PFCmechanismsandweight-basedHCMP learning;andretrieved in the caseofHCMP by controlledactivationof cues.Verbalandperhapsspatialand/ornumericalrepresentationsareespeciallyusefulwaysof encoding.

(2) TheControlandRegulationof WorkingMemory:Workingmemoryis notseparatedfrom control,sincecontrolledprocessingandactivememoryarein-timatelyrelated.Controlis alsonotcentralized,emerging insteadfrom interactionsbetweendifferentbrainsystems.PFCplaysimportantrole dueto: robust maintenancecapabilities;flexible andrapidupdatingof representations;positionat thetopof thecorticalprocessinghierarchy(with HCMP).

(3) TheUnitaryvsNon-UnitaryNatureof WorkingMemory:Working memoryis not unitary: consistingof active memory, rapidlearningandcontrolledprocess-ing, anddistributedover several brain systems.Commonuseof controlledprocessingmechanismsmaycontributeaunitary-like componentto performance.

(4) TheNatureof WorkingMemoryLimitations:Two mechanisms:inhibition, andinterference.PFChasgreaterinhibition to promotecoherentpro-cessing,thus lower capacity. Capacityhasdomainspecificand generalcomponents(see3), andcorrespondingexperienceandgeneticbases.Capacityis highly dependenton amountandtype ofcontrolledprocessingnecessary, andefficiency of underlyingrepresentationslearnedoverexperience.

(5) TheRoleof WorkingMemoryin Complex Cognitive Activities:Workingmemoryis critical,assuchactivitiesaredefinedby theinvolvementof controlledprocessing,andrequireactive memory/rapidlearningto maintainintermediateresults.Distributedbrainsystemsareinvolvedasrelevantin particulartasks,with morecommoninvolvementof PFCandHCMP.

(6) TheRelationshipof WorkingMemoryto Long-TermMemoryandKnowledge:Working memoryis largely just the active portion of long-termmemory, which is itself distributedover many brainareas.More globally accessiblesystemsandthosethat provide particularlyusefulrepresentations(e.g.,language)aremorelikely to be involved in working memory, leadingto a biastowardsdeclarativeor explicit representationsinsteadof implicit or procedural ones.

(7) TheRelationshipof WorkingMemoryto AttentionandConsciousness:Working memoryis thesubsetof representationsattendedto by virtue of controlledprocessing.At-tentionalsorefersto a constrainingmechanism(inhibition), andcanbeinfluencedby automaticpro-cessing.Consciousnessreflectstheglobalconstraintsatisfactionprocess,which is disproportionatelyinfluencedby controlled-processing systems.Thus,thecontentsof consciousexperiencearelikely toreflectthatof workingmemory.

(8) TheBiological Implementationof WorkingMemory:Our modelis basedon the biology, including neural-level propertieslike activation, inhibition, andlearning, and a computationalaccountof specializedbrain systemfunction, including the PFC,HCMP, andPMC. A large amountof empiricaldatafrom patients,neuroimaging,neurophysiology,andanimalstudiesis consistentwith ourmodel.

Table2: Summaryof ouranswersto theeightdesignatedquestions.


that�

explain many of thesamephenomenaastraditionalworking memoryconstructs,but in a mannerthatcontrastswith existingtheoreticalideasin importantways.Ourexistingcomputationalwork hasinstantiatedand validateda numberof aspectsof our theory, including: the gradednatureof controlledprocessing(Cohenet al., 1990); the ability of PFCrepresentationsto biassubsequentprocessing(Cohen& Servan-Schreiber, 1992);therole of PFCin active maintenance(Braver et al., 1995);andtherole of theHCMP inrapidlearning(O’Reilly, Norman,& McClelland,1998;O’Reilly & McClelland,1994).However, wehavenot yet implementeda computationalmodelthatcapturesall of our ideasregardingworking memoryandcontrolledprocessing.Moreover, therearea numberof importantissuesraisedby our overall modelthathave notbeenproperlyaddressedin ourprior work, andwhichform thecurrentfocusof our research.

Theseunresolved issuescanbe describedat two generallevels of analysis— one level involves thedevelopmentof bettermodelsof eachof the individual brainsystemsthatplay a role in our overall model(PMC, PFC,HCMP), andthe other level involvescharacterizingthe natureof interactionsbetweenthesesystems.Obviously, thelattereffort dependscritically on thesuccessof theformer, which is wherewehavebeenprimarily focused.Underlyingtheentireendeavor areissuesof thecomputationalsufficiency necessaryto learnandperformtemporallyextendedcontrolled-processing tasksusingneuralnetwork models.

Modelsof thePMC,PFC,andHCMP

Becauseit representsthecanonicalform of corticalprocessing,our modelof thePMC lies at thefoun-dationof theothermodels.We have recentlymadeimportantadvancesin characterizingthenatureof pro-cessingandlearningin cortex, andnow have a computationalframework (calledLeabra; O’Reilly, 1996b,1996a)which containsall of the basicmechanismsandpropertiesrequiredby our model. In particular,the Leabraframework combinesrecurrence,inhibition, and integratederror-driven andHebbianlearningmechanismsin asimple,principled,robust,andbiologicallyplausiblemanner. While thesepropertieshavebeenimplementedseparatelyin differentmodelsbefore,Leabraintegratesthemall in a unified,coherentframework. Our modelof theHCMP systemis relatively well-developedconceptually, andpartsof it havebeenmodeledat a very detailedlevel (O’Reilly & McClelland,1994). Recently, we have createda com-pleteHCMP modelusingtheLeabraframework (O’Reilly et al., 1998),andhave modeledtherecollectivecontribution to many of thebasicrecognitionmemoryphenomena(list length,list strength,etc).

It is thePFCwhichhasreceivedmostof our recenttheoreticalattention,building onpreviouswork thatestablishesa basicframework for understandingthe computationalrole of PFCin controlledprocessing.Therearetwo primarythreads:1) Therole of a dopamine(DA) mediatedactive gatingmechanismasde-scribedpreviously;and2) Thenatureof PFCrepresentationsnecessaryto accomplishcontrolledprocessingin complex tasks.We have recentlyimplementeda DA-like gatingmechanismin a computationalmodelof PFC,andshown thatit cansuccessfullyaccountfor all of thephenomenaaccountedfor by our previousmodels,while makingnew predictionsbasedon the phasicnatureof the DA gate(Braver et al., 1997a).Thismodelis beingextendedto morecomplex tasksthatwill bettertestthegatingmechanismby requiringboth rapid updatingandsustainedmaintenancein the faceof interference.Our currentwork on the PFCrepresentationsis investigatingthetradeoff betweendistributedrepresentationsandactive maintenanceasafunctionof differenttaskdemands.

Reward-basedLearning, Goals,andthePFC

Oneof the most importantunresolved challengesto modelsof working memory(andcognitionmoregenerally)is specifyingthemechanisticbasisof executivecontrol(controlledprocessing)in awaythatdoesnot resortto a homunculus.While we have generallycharacterizedour view of how controlledprocessingemergesfrom constraintsatisfactionandthespecializedpropertiesof thePFCandHCMP, actuallyshowingthat this works in real tasksremainsa challenge. We think that the solution to this problemrequiresapowerful learningmechanismwhichis capableof developingsomethinglike the“productions”thatunderlie


the�

performanceof complex cognitive tasks(thusavoiding the hiddenhomunculusof the researcherwhobuilds in theappropriateproductionsfor eachtask).Thefollowing is onesetof ideasregardingthenatureofthis learningmechanism,which emergesfrom a synthesisof our basicideasabouta DA-basedmechanismfor active gating,andthenatureof therepresentationsin thePFC.

Theseideascanbemotivatedby thinking abouttheessentialdifferencebetweenhumancognition,andthat of even our closestprimaterelatives. It is obvious that language,abstraction,problemsolving, andtool useareimportantbehavioral differences.However, we suggestthat thesemayall befacilitatedby theability to internalize,abstract,and chain togetherrepresentationsof reward (andpunishment). In short,the realdifferencebetweenhumansandotherprimatesmaybe thatwe canestablishelaboratesystemsofinternalizedreward that motivateus to learnandengagein thesemoreabstractbehaviors, whereasotherprimates,whocanlearnimpressively complex andabstracttasks,mustneverthelessbeconstantlymotivatedby externalforces(e.g.,food,juice)to doso.Thus,insteadof beinga“pure” cognitivesystemdivorcedfromall emotionalor motivationalconcerns,thePFCmay insteadbecentrallyinvolved in thedirty businessofmotivation,emotion,pleasure,andpain(Davidson& Sutton,1995;Bechara,Tranel,Damasio,& Damasio,1996).

This observation provestantalizingin thecontext of our ideasaboutthe role of dopamine(DA) in thePFC.In particular, if thecritical specializationof thePFCis that it hastaken controlover theDA systemin orderto regulateits own active maintenancefunction, thenit is alsoin a positionto take over andin-ternalizethe deploymentof DA-mediatedreinforcement.It is well known that DA playsa critical role inreinforcement-basedlearning(Schultzet al., 1993;Montagueet al., 1996). If theactivationof PFCrepre-sentationscorrespondessentiallyto goalswhicharemaintainedin anactive andrelatively protectedstateintheabsenceof DA firing, thentheactof satisfyingagoalshouldsimultaneouslyresultin reinforcementandgating(i.e., thedeactivation of that goal representationandtheopportunityto activatea subsequentone).The firing of DA underPFCcontrol would provide both,andthe influenceof this DA signalon learningshouldresultin moreeffective ellicitationandefficientexecutionof thatgoalin thefuture.

Further, we have arguedabove that the PFC hasthe capacityfor the simultaneousrepresentationofmany levels of temporalextentandabstraction,which would beneededto accountfor thegoal structuresunderlyingcomplex humancognition. Sincereward is underthedescendingcontrolof thePFCitself, theneedfor externalrewardis reduced,allowing for thedevelopmentof elaboratemeans(interveninggoals)toaccomplishremoteandabstractends.In contrast,otheranimalsdependto amuchgreaterextentonconstantexternalinput to drive theDA rewardsystem,andthuscannotbuild theseelaborateinternalgoalstructures.

Therearemany differentways in which the internalizedcontrol of DA could be implementedin thePFC,but unfortunatelylittle is known aboutthe relevant biological details. Thus,we areusingcomputa-tionalmodelsto determinetherelativeadvantagesanddisadvantagesof differentimplementations.Anotherimportantimplementationalissuehasto do with learningon thebasisof actively-maintained,isolatedrep-resentationslike thosein thePFC,which have a morediscrete,binarycharacterandthusdo not appeartobeamenableto thetypesof gradient-basedlearningmechanismsthatwork sowell in thedistributed,gradedrepresentationscharacteristicof the PMC. In summary, the specializationsof the PFCwill likely requirespecializedlearningmechanisms,whicharethefocusof ourcurrentresearch.

Conclusion

In summary, our overall modelof thebrainsystemsunderlyingworking memory, including thePMC,PFC,andHCMP, is still underconstruction,but we have a broadandcompellingblueprintfor futureex-ploration. This modelprovidesmany exampleswherecomputationalprinciples(e.g.,basictradeoffs) areusedto understandbiologicalproperties,in waysthat,while consistentwith existing ideasin many cases,canachieve a new level of synthesisandclarity. We hopethat this approachwill continueto prove useful,


despite�

theinevitablerevisionof many of thespecificideasproposedherein.

Acknowledgments

Wewould like to thankAndrew Conway, PeterDayan,Yuko Munakata,KenNorman,Mike Wolfe, andall of theconferenceparticipantsfor usefulcomments.Weweresupportedby NIH GrantMH47566-06,andR.O.wasalsosupportedby a McDonnellPew PostdoctoralFellowshipat MIT in theDepartmentof BrainandCognitive Science.


References

Anderson,J.R. (1983).Thearchitecture of cognition. Cambridge,MA: HarvardUniversityPress.

Baddeley, A. D. (1986).Workingmemory. New York: OxofordUniversityPress.

Baker, S.C.,Rogers,R. D., Owen,A. M., Frith, C. D., Dolan,R. J.,Frackowiak, R. S.J.,& Robbins,T. W.(1996). Neuralsystemsengagedby planning: A PET studyof the Tower of Londontask. Neuropsy-chologia, 34, 515.

Barbas,H., & Pandya,D. N. (1987). Architectureandfrontal corticalconnectionsof thepremotorcortex(area6) in therhesusmonkey. Journalof ComparativeNeurology, 256, 211–228.

Barbas,H., & Pandya,D. N. (1989). Architectureof intrinsic connectionsof the prefrontalcortex in therhesusmonkey. Journalof ComparativeNeurology, 286, 353–375.

Barch,D. M., Braver, T. S.,Nystrom,L. E.,Forman,S.D., Noll, D. C.,& Cohen,J.D. (1997).Dissociatingworkingmemoryfrom taskdifficulty in humanprefrontalcortex. Neuropsychologia, 35, 1373.

Barone,P., & Joseph,J. P. (1989). Prefrontalcortex andspatialsequencingin macaquemonkey. Experi-mentalBrain Research, 78, 447–464.

Bechara,A., Tranel, D., Damasio,H., & Damasio,A. R. (1996). Failure to respondautonomicallytoanticipatedfutureoutcomesfollowing damageto prefrontalcortex. Cerebral Cortex, 6, 215–225.

Braver, T. S.,& Cohen,J. D. (1995). A modelof thedevelopmentof objectandspatialworking memoryrepresentationsin prefrontalcortex. CognitiveNeuroscienceAbstracts, Vol. 2 (p. 95).

Braver, T. S.,Cohen,J.D., & McClelland,J.L. (1997a).An integratedcomputationalmodelof dopaminefunctionin reinforcementlearningandworkingmemory. Societyfor NeuroscienceAbstracts,23(p.775).SanDiego,CA: Societyfor Neuroscience.

Braver, T. S.,Cohen,J.D., Nystrom,L. E., Jonides,J.,Smith,E. E., & Noll, D. C. (1997b).A parametricstudyof frontalcortex involvementin humanworkingmemory. NeuroImage, 5, 49–62.

Braver, T. S., Cohen,J. D., & Servan-Schreiber, D. (1995). A computationalmodelof prefrontalcortexfunction.In D. S.Touretzky, G.Tesauro,& T. K. Leen(Eds.),Advancesin neural informationprocessingsystems(pp.141–148).Cambridge,MA: MIT Press.

Cohen,J.D., Braver, T. S.,& O’Reilly, R. C. (1996).A computationalapproachto prefrontalcortex, cogni-tivecontrol,andschizophrenia:Recentdevelopmentsandcurrentchallenges.PhilosophicalTransactionsof theRoyalSocietyof LondonSeriesB (Biological Sciences), 351, 1515–1527.

Cohen,J. D., Dunbar, K., & McClelland,J. L. (1990). On thecontrolof automaticprocesses:A paralleldistributedprocessingmodelof thestroopeffect. Psychological Review, 97(3), 332–361.

Cohen,J.D., & O’Reilly, R. C. (1996). A preliminarytheoryof theinteractionsbetweenprefrontalcortexandhippocampusthatcontribute to planningandprospective memory. In M. Brandimonte,G. O. Ein-stein,& M. A. McDaniel (Eds.),Prospectivememory:Theoryandapplications. Mahwah,New Jersey:LawrenceEarlbaumAssociates.

Cohen,J.D., Perlstein,W. M., Braver, T. S.,Nystrom,L. E.,Jonides,J.,Smith,E. E.,& Noll, D. C. (1997).Temporaldynamicsof brainactivity duringaworkingmemorytask.Nature, 386, 604–608.

Cohen,J.D., & Servan-Schreiber, D. (1992). Context, cortex, anddopamine:A connectionistapproachtobehavior andbiology in schizophrenia.Psychological Review, 99, 45–77.

Cohen,J.D., & Smith,E. E. (1997). Responseto OwenAM, Tuningin to thetemporaldynamicsof brainactivationusingfunctionalmagneticresonanceimaging(fMRI). Trendsin CognitiveSciences, 1, 124.


Damasio,�

A. R. (1985). The frontal lobes. In K. M. Heilman,& E. Valenstein(Eds.),Clinical neuropsy-chology (pp.339–375).New York: OxfordUniversityPress.

Damasio,H., Grabowski, T. J.,& Damasio,A. R. (1996). A neuralbasisfor lexical retrieval. Nature, 380,499.

Daneman,M., & Carpenter, P. A. (1980). Individual differencesin working memoryandreading.Journalof VerbalLearningandVerbalBehavior, 19, 450–466.

Daneman,M., & Merikle, P. (1996). Working memoryand languagecomprehension:A meta-analysis.PsychonomicBulletin& Review, 3, 422–433.

Davidson,R. J., & Sutton,S. K. (1995). Affective neuroscience:theemergenceof a discipline. CurrentOpinionin Neurobiology, 5, 217–224.

Dehaene,S.,& Changeux,J. P. (1989). A simplemodelof prefrontalcortex functionin delayed-responsetasks.Journalof CognitiveNeuroscience, 1, 244–261.

Diamond,A. (1990). The developmentandneuralbasesof memoryfunctionsasindexed by the a-not-btask:Evidencefor dependenceondorsolateralprefrontalcortex. In A. Diamond(Ed.),Thedevelopmentandneural basesof highercognitivefunctions(pp.267–317).New York: New York Academyof SciencePress.

Fodor, J.(1983).Themodularityof mind. Cambridge,MA: MIT/BradfordPress.

Funahashi,S.,Bruce,C. J.,& Goldman-Rakic,P. S.(1993).Dorsolateralprefrontallesionsandoculomotordelayed-responseperformance:Evidencefor mnemonic’scotomas’.Journalof Neuroscience, 13, 1479–1497.

Fuster, J.M. (1989).Theprefrontal cortex. anatomy, physiology andneuropsychology of thefrontal lobe.n.New York: RavenPress.

Fuster, J. M., & Alexander, G. E. (1971). Neuronactivity relatedto short-termmemory. Science, 173,652–654.

Gathercole,S.E. (1994). Neuropsychologyandworkingmemory:A review. Neuropsychology, 8(4), 494–505.

Gershberg, F. B., & Shimamura,A. P. (1995). Impaireduseof organizationalstrategiesin free recall fol-lowing frontal lobedamage.Neuropsychologia, 33, 1305.

Goldman-Rakic,P. S.(1987).Circuitry of primateprefrontalcortex andregulationof behavior by represen-tationalmemory. Handbookof Physiology —TheNervousSystem, 5, 373–417.

Hochreiter, S.,& Schmidhuber, J. (1997).Longshorttermmemory. Neural Computation, 9, 1735–1780.

Kahneman,D., & Treisman,A. (1984). Changingviews of attentionand automaticity. SanDiego, CA:AcademicPress,Inc.

Kinsbourne,M. (1997).Whatqualifiesarepresentationfor arole in consciousness?In J.D. Cohen,& J.W.Schooler(Eds.),Scientificapproachesto consciousness(pp.335–355).Mahway, NJ:LawrenceErlbaumAssociates.

McClelland,J. L. (1993). The GRAIN model: A framework for modelingthe dynamicsof informationprocessing.In D. E. Meyer, & S. Kornblum(Eds.),Attentionand performanceXIV: Synergies in ex-perimentalpsychology, artifical intelligence, and cognitive neuroscience(pp. 655–688).Hillsdale, NJ:LawrenceErlbaumAssociates.

McClelland,J. L., McNaughton,B. L., & O’Reilly, R. C. (1995). Why therearecomplementarylearningsystemsin the hippocampusand neocortex: Insightsfrom the successesand failuresof connectionstmodelsof learningandmemory. Psychological Review, 102, 419–457.


Miller�

, E. K., & Desimone,R. (1994).Parallelneuronalmechanismsfor short-termmemory. Science, 263,520–522.

Miller, E. K., Erickson,C. A., & Desimone,R. (1996). Neuralmechanismsof visualworking memoryinprefontalcortex of themacaque.Journalof Neuroscience, 16, 5154.

Montague,P. R., Dayan,P., & Sejnowski, T. J. (1996). A framework for mesencephalicdopaminesystemsbasedonpredictive hebbianlearning.Journalof Neuroscience, 16, 1936–1947.

Moscovitch, M., & Winocur, G. (1992).Theneuropsychologyof memoryandaging.In T. A. Salthouse,&F. I. M. Craik (Eds.),Thehandbookof agingandcognition. Hillsdale,NJ:Erlbaum.

Newell, A. (1990).Unifiedtheoriesof cognition. Cambridge:HarvardUniversityPress.

O’Reilly, R. C. (1996a).Biologically plausibleerror-drivenlearningusinglocalactivationdifferences:Thegeneralizedrecirculationalgorithm.Neural Computation, 8(5), 895–938.

O’Reilly, R. C. (1996b).Theleabra modelof neural interactionsandlearningin theneocortex. PhDthesis,Carnegie Mellon University, Pittsburgh,PA, USA.

O’Reilly, R.C.,& McClelland,J.L. (1994).Hippocampalconjunctiveencoding,storage,andrecall:Avoid-ing a tradeoff. Hipocampus, 4(6), 661–682.

O’Reilly, R. C.,Norman,K. A., & McClelland,J.L. (1998).A hippocampalmodelof recognitionmemory.In M. I. Jordan(Ed.), Advancesin neural informationprocessingsystems10. Cambridge,MA: MITPress.

Petrides,M. E. (1996).Specializedsystemsfor theprocessingof mnemonicinformationwithin theprimatefrontalcortex. PhilosophicalTransactionsof theRoyalSocietyof London,SeriesB., 351, 1455–1462.

Posner, M. I., & Snyder, C. R. R. (1975).Attentionandcognitive control. In R. L. Solso(Ed.),Informationprocessingandcognition (pp.55–85).Hillsdale,NJ:LawrenceErlbaumAssociates.

Rao,S. C., Rainer, G., & Miller, E. K. (1997). Integrationof what andwherein the primateprefrontalcortex. Science, 276, 821.

Rizzolatti,G., Luppino,G., & Matelli, M. (1996). Theclassicsupplementarymotorareais formedby twoindependentareas.Avancesin Neurology, 70, 45–56.

Rumelhart,D. E., & McClelland,J. L. (1986). On learningthe pasttensesof English verbs. In J. L.McClelland,D. E. Rumelhart,& PDPResearchGroup(Eds.),Parallel distributedprocessing. volume2:Psychological andbiological models(pp.216–271).Cambridge,MA: MIT Press.

Schultz,W., Apicella, P., & Ljungberg, T. (1993). Responsesof monkey dopamineneuronsto rewardand conditionedstimuli during successive stepsof learninga delayedresponsetask. The Journal ofNeuroscience, 13, 900–913.

Scoville, W. B., & Milner, B. (1957). Lossof recentmemoryafterbilateralhippocampallesions.Journalof Neurology, Neurosurgery, andPsychiatry, 20, 11–21.

Seidenberg, M. (1993).Connectionistmodelsandcognitive theory. Psychological Science, 4(4), 228–235.

Servan-Schreiber, D., Cohen,J. D., & Steingard,S. (in press).Schizophrenicperformancein a variantoftheCPT-AX: A testof theoreticalpredictionsconcerningtheprocessingof context. Archivesof GeneralPsychiatry.

Shah,P., & Miyake, A. (1996). the separabilityof working memoryresourcesfor spatialthinking andlanguageprocessing:anindividualdifferencesapproach.Journalof ExperimentalPsychology: General,125, 4.


Shallice,T. (1982). Specificimpairmentsof planning. PhilosophicalTransactionsof theRoyalSocietyofLondon, 298, 199–209.

Shiffrin, R. M., & Schneider, W. (1977). Controlledand automatichumaninformationprocessing:II.Perceptuallearning,automaticattending,andageneraltheory. Psychological Review, 84, 127–190.

Smith,E. E., Jonides,J.,& Koeppe,R. A. (1996). Dissociatingverbalandspatialworking memoryusingPET. Cerebral Cortex, 6, 11–20.

Squire,L. R. (1992). Memoryandthe hippocampus:A synthesisfrom findingswith rats,monkeys, andhumans.Psychological Review, 99, 195–231.

Squire,L. R., Shimamura,A. P., & Amaral,D. G. (1989). Memoryandthehippocampus.In J. H. Byrne,& W. O. Berry(Eds.),Neural modelsof plasticity: Experimentalandtheoreticalapproaches. SanDiego,CA: AcademicPress,Inc.

Tulving,E.,Kapur, S.,Craik,F. I. M., Moscovitch, M., & Houle,S.(1994).Hemisphericencoding/retrievalasymmetryin episodicmemory: Positronemissiontomographyfindings. Proceedingsof the NationalAcademyof Sciences,USA, 91, 2016–2020.

Ungerleider, L. G., & Mishkin, M. (1982). Two corticalvisualsystems.In D. J. Ingle, M. A. Goodale,&R. J.W. Mansfield(Eds.),Theanalysisof visualbehavior. Cambridge,MA: MIT Press.

Usher, M., & Cohen,J. D. (1997). Interference-basedcapacitylimitationsin active memory. AbstractsofthePsychonomicsSociety, Vol. 2 (p. 11).

Wilson, F. A. W., Scalaidhe,S. P. O., & Goldman-Rakic,P. S. (1993). Dissociationof objectandspatialprocessingdomainsin primateprefrontalcortex. Science, 260, 1955–1957.

Zipser, D., Kehoe,B., Littlewort, G., & Fuster, J. (1993). A spiking network modelof short-termactivememory. Journalof Neuroscience, 13, 3406–3420.

Documents

A Biologically-BasedComputational Model of Working Memorypsych.colorado.edu/~oreilly/papers/OReillyBraverCohen99.pdfA Biologically-BasedComputational Model of Working Memory Randall