dissociable processes for learning the surface structure and abstract structure of sensorimotor...

Dissociable Processes for Learning theSurface Structure and Abstract Structure ofSensorimotor Sequences

Peter F Dominey and Taiumlssia LelekovCNRS Institut des Sciences Cognitives Lyon France

Jocelyne Ventre-DomineyINSERM Vision et Motricit Lyon France

Marc JeannerodCNRS Institut des Sciences Cognitives Lyon France

Abstract

A sensorimotor sequence may contain information struc-ture at several different levels In this study we investigated thehypothesis that two dissociable processes are required for thelearning of surface structure and abstract structure respec-tively of sensorimotor sequences Surface structure is the sim-ple serial order of the sequence elements whereas abstractstructure is dened by relationships between repeating se-quence elements Thus sequences ABCBAC and DEFEDF havedifferent surface structures but share a common abstract struc-ture 123213 and are therefore isomorphic Our simulations ofsequence learning performance in serial reaction time (SRT)tasks demonstrated that (1) an existing model of the primatefronto-striatal system is capable of learning surface structurebut fails to learn abstract structure which requires an addi-tional capability (2) surface and abstract structure can belearned independently by these independent processes and(3) only abstract structure transfers to isomorphic sequences

We tested these predictions in human subjects For a sequencewith predictable surface and abstract structure subjects ineither explicit or implicit conditions learn the surface struc-ture but only explicit subjects learn and transfer the abstractstructure For sequences with only abstract structure learningand transfer of this structure occurs only in the explicit groupThese results are parallel to those from the simulations andsupport our dissociable process hypothesis Based on the syn-thesis of the current simulation and empirical results with ourprevious neuropsychological ndings we propose a neuro-physiological basis for these dissociable processes Surfacestructure can be learned by processes that operate underimplicit conditions and rely on the fronto-striatal systemwhereas learning abstract structure requires a more explicitactivation of dissociable processes that rely on a distributednetwork that includes the left anterior cortex

INTRODUCTION

Essentially all aspects of cognitive function are embed-ded in a sequential context as seen in the perceptionand production of language game playing problem solv-ing and motor control In such domains humans regu-larly perceive and produce novel legal sequences relyingon a capacity to manipulate rules that permit generaliza-tion to new instances An open question in cognitivescience is whether there are distinct forms of knowledgerelated to rules versus instances with the correspondingquestion in cognitive neuroscience as to whether thesedistinct forms of knowledge must be treated by distinctneural systems (eg Gomez amp Schvaneveldt 1994Knowlton amp Squire 1993 1996 Shanks amp St John 1994)The current research explores this question in the do-

copy 1998 Massachusetts Institute of Technology Journal of Cognitive Neuroscience 106 pp 734ndash751

main of sensorimotor sequence learning with resultsthat argue for the necessity of distinct processes toaccommodate instance versus rule-based knowledge

We will explore this processing distinction in thedomain of serial reaction time (SRT) sequence learningIn the SRT paradigm subjects perform a series of manualresponses to the successive elements in a series of visualstimuli Nissen and Bullemer (1987) demonstrated thatthe reaction times (RTs) for these responses are sig-nicantly reduced if the stimuli appear in a repeatingsequence as opposed to in a random order This reduc-tion in RTs between responses for sequential versusrandom series is a measure of the sequence learning

Within the SRT domain we approach the instanceversus rule dissociation in terms of ldquosurfacerdquo and ldquoab-stractrdquo structure in sensorimotor sequences Surface

structure is dened by the serial order of the sequenceelements at the level of verbatim and aggregate struc-ture as dened by Stadler (1992) In contrast abstractstructure is dened by the relations between elementsthat repeat within a sequence In this context the twosequences ABCBAC and DEFEDF share the same abstractstructure 123213 and have completely unrelated surfacestructures (ie they are isomorphic) If we consider twosuch isomorphic sequences in terms of what is pre-dicted by the abstract structure we observe that in bothsequences the last three elements are predictable by therst three which themselves are unpredictable In anSRT protocol with a repeating sequence like ABCBACsurface structure learning should be manifest by a uni-form RT reduction for both the predictable and unpre-dictable elements Learning of the abstract structureshould be manifest by a nonuniform prole with addi-tional RT reduction for the predictable versus unpre-dictable elements that should transfer to newisomorphic sequences

A Theoretical Basis for Dissociable Processes

From a theoretical perspective there is an importantrelation between the structure of information to learn(eg surface or abstract structure) and the class of ma-chines or architectures that are capable of learning orprocessing this structure (eg Chomsky 1959 Turing1936) This would predict for example that there is asystem capable of learning surface structure that cannotlearn abstract structure Here we describe a neural net-work model based on the primate fronto-striatal system(Dominey 1995 Dominey Arbib amp Joseph 1995) andpresent simulation results demonstrating that this systemis capable of learning surface structure but is not capableof learning and transferring abstract structure whichrequires separate processing capabilities (Dominey1997b Dominey Ventre-Dominey Broussolle amp Jean-nerod 1995a 1995b)

Based on these results and our initial hypothesis wepredict that there are two independent and behaviorallydissociable mechanisms for learning surface and abstractstructure in humans As noted by Shanks and St John(1994) the learning of rules or abstract structure ischaracterized by a conscious effort to discover and ex-ploit the appropriate rules an effort that can be invokedby specic instructions to make such a conscious effort(Gick amp Holyoak 1983) This position is supported bystudies of analogical transfer in problem solving thatinvolve the extraction of an abstract structure commonto several problems with different surface structures(Gick amp Holyoak 1983 Holyoak Junn amp Billman 1984Holyoak Novick amp Melz 1994) Such studies demon-strate that this process requires the explicit intention tond the abstract structure In contrast the learning ofinstances or surface structure is oriented toward memo-rization of the instances themselves (Shanks amp St John

1994) without a conscious processing effort to searchfor common rule-based structure (eg Cohen Ivry ampKeele 1990 Curran amp Keele 1993) These studies suggestthe existence of dissociable mechanisms for processingsurface and abstract structure

Several studies of articial grammar learning (AGL)have also provided convincing evidence for the exist-ence of dissociable mechanisms for surface versus ab-stract structure representations (Gomez 1997 Gomez ampSchvaneveldt 1994 Knowlton amp Squire 1996) Knowl-ton and Squire demonstrated that both rule adherence(abstract structure) and chunk strength that is similarityof letter bigram and trigram distribution (surface struc-ture) inuence grammaticality judgments LikewiseGomez and Schvaneveldtrsquos results indicate that trainingon legal letter pairs is sufcient for classication withthe same letter set but that training with longer stringsis required to allow transfer of abstract structure to achanged letter set This suggests that pairs provide asource of surface structure whereas abstract structure isonly available in strings These results thus indicate thatin AGL there are dissociable forms of representation forsurface and abstract structure respectively

Although AGL tasks are often considered to test im-plicit learning it is important to note that during the testphase the subjects are explicitly instructed to apply a setof rules to classify the new objects and it is likely thatsome rule abstraction takes place during this explicittesting phase (Perruchet amp Pacteau 1991 Reddington ampChater 1996) Likewise Mathews et al (1989) have dem-onstrated that this grammatical knowledge can becomeat least partially explicit and that for grammars thatexploit relational properties like the ones used in ourcurrent studies learning can only occur in truly explicitconditions This relation between explicit processing andAGL has recently been further claried by Gomez(1997) who demonstrated that subjectsrsquo ability to trans-fer abstract structure learning to changed letter sets wasinvariably accompanied by explicit knowledge as re-vealed in direct tests Conversely subjects who learnedrst-order surface structure dependencies but failed todisplay transfer of the abstract structure in the changedletter set condition did not differ from naive controls onthe direct tests Finally Gomez demonstrated that for thesame testing materials presented either in a whole-stringAGL task or a letter-by-letter sequence learning tasktransfer and the associated acquisition of explicit knowl-edge occurred only in the AGL task This indicates thatespecially in sequencing tasks the acquisition of abstractstructure and its transfer to isomorphic sequences in-volves explicit processing

We thus predict for our sequence learning task thatin completely uninformed implicit conditions processesthat are capable of learning surface structure but notabstract structure will be accessible whereas in explicitconditions both will be accessible In response to thesepredictions we will report the results of three experi-

Dominey et al 735

ments in human subjects demonstrating that surfacestructure can be learned in implicit conditions but thatlearning and transferring abstract structure occur only inexplicit conditions Based on a synthesis of these obser-vations with our previous neuropsychological ndings(Dominey amp Georgieff 1997 Dominey amp Jeannerod1997 Dominey et al 1995a 1995b Dominey Ventre-Dominey Broussolle amp Jeannerod 1997) we propose aneurophysiological basis for the dissociable processingof surface and abstract structure observed in the simula-tion and human experiments

A DUAL PROCESS MODEL OF SURFACE ANDABSTRACT STRUCTURE LEARNING

An important class of sequence learning models demon-strates the capability to predict future events by encod-ing the context or the history of previous events viarecurrent connections or self-connections (eg Cleere-mans amp McClelland 1991 Dominey 1995 1998a 1998bElman 1990) In these recurrent networks the contextis sensitive to events several positions in the past allow-ing the networks to resolve ambiguities as in determin-ing the successor to B in the sequence ABCBAC Moregenerally these recurrent systems are ideally suited forlearning surface structure However this recurrent con-text mechanism appears to be insufcient to representthe common relation between two isomorphic se-quences ABCBAC and DEFEDF The ability to representthis abstract structure requires the capacity to recognizethe structure of repeating elements and to let this infor-mation be the source of the context This distinction willbe claried in the following description of the dissoci-able models for surface and abstract structure processingthat together make up the dual process model

The surface model (Figure 1 amp ldquoAppendixrdquo) consistsof the Input State and Output units and falls into thegeneral category of recurrent networks described above(Dominey 1995 1998a Dominey Arbib et al 1995) Themodel architecture is based closely on the neuro-anatomy of the primate fronto-striatal system and themodel has been used to explain detailed electrophysi-ological activity in the primate prefrontal cortex andbasal ganglia during sequence processing tasks(Dominey Arbib et al 1995 Dominey amp Boussaoud1997) Input Output and State each consist of 5 times 5arrays of leaky integrator units whose response latency(ie reaction time) is a function of the strength of theirinput signals State is inuenced by Input and Output(Equations 1 amp 2 see ldquoAppendixrdquo for all equations) andhas recurrent excitatory and inhibitory connectionsState thus plays the crucial role of encoding sequencecontext corresponding to the prefrontal cortex in thefronto-striatal system whereas Output corresponds tothe striatum (Dominey Arbib et al 1995) Sequences arepresented to the model by activating individual units(1 to 25) in the Input array Input units are directly

connected to their corresponding Output units Theseconnections are responsible for the baseline reactiontimes (RTs) for responses in Output to stimuli presentedto Input This baseline reaction time is modulated byconnections from State to Output that are modiable bylearning (Equations 3 amp 4) These connections corre-spond to cortico-striatal synapses that are modiable byreward-related dopamine release in the striatum(Dominey Arbib et al 1995)

Figure 1 Schematic representation of an anatomically structuredmodel for learning surface and abstract structure Surface modelPresentation of sequence stimuli to Input activates both State (corre-sponding to prefrontal cortex in Dominey Arbib et al 1995) andOutput (corresponding to Caudate nucleus of striatum in DomineyArbib et al 1995) State is a recurrent network whose activity overtime encodes the sequence context that is the history of previoussensory (Input) and motor (Output) events At the time of each Out-put response there is a specic pattern of activity or context inState Connections from State to Output (dotted line) are modiedduring sequence learning thus binding each sequence context inState with the corresponding response in Output for each sequenceelement yielding reduced reaction times (RTs) Abstract model exten-sion A short-term memory (STM) encodes the 7 plusmn 2 previous re-sponses Recog detects repetitions between current response andprevious responses in and provides this input to State Modiableconnections (dotted line) from State gate the contents of appropri-ate STM elements to Output when repetitive structure is predict-able reducing RTs for predictable elements in isomorphicsequences Left The surface model can learn the serial order ofsequence elements 613163 but cannot transfer this knowledge toisomorphic sequence 781871 Right The abstract model learns therelations between repeating elements in 613163 as u u u n - 2n - 4 n - 3 (where ldquourdquo signies unpredictable and ldquon - 2rdquo indi-cates a repetition of the element two places behind etc) Thisabstract structure transfers completely to the isomorphic sequence781871

736 Journal of Cognitive Neuroscience Volume 10 Number 6

Learning Surface Structure

In the SRT task simulations each element in a sequenceis presented in Input which thus generates a responsein Output with the baseline reaction time When a re-sponse is generated active units in State encode thecurrent sequence context and connections from theseState units to units in Output coding the response arestrengthened thus binding the sequence context to thecurrent element in the sequence The result of this learn-ing is that the next time this same pattern of sequencecontext activity in State occurs (ie the next time thesequence of elements that precede the current elementis presented) the strengthened State-Output connectionswill cause State to increasingly activate the appropriateOutput units (effectively predicting the current se-quence element) resulting in a reduced RT for thisresponse By this mechanism the learning of surfacestructure will be demonstrated as RT reductions for allelements in the learned repeating sequence Indeed wehave recently demonstrated that this model is robust inits ability to simulate SRT learning in normal and ldquodis-tractionrdquo conditions for surface structure based on itssensitivity to temporal as well as serial sequence organi-zation (Dominey 1998a 1998b) The model fails how-ever to learn abstract structure (Dominey 1997bDominey et al 1995a 1995b)

Learning Abstract Structure

To permit the representation of abstract structure aswersquove dened it the model must be capable of compar-ing the current response with previous responses torecognize repetitive structure (ie u u u n - 2 n - 4n - 3 for ABCBAC where ldquourdquo signies unpredictable andldquon - 2rdquo indicates a repetition of the element two placesbehind etc) These functions would rely on the morenonsensorimotor associative areas of the anterior cortexand would permit the generalization of grammarlikerules to new but ldquolegalrdquo sensorimotor sequences(Dominey 1997b) To make this possible we introduceda short-term memory (STM) mechanism (Equation 5)that is continuously updated to store the previous 7 plusmn 2responses (see Lisman amp Idiart 1995) and a Recognitionmechanism (Equation 6) that compares the current re-sponse to the stored STM responses to detect any re-peated elements (Dominey et al 1995a 1995b) Thispermits the recoding of sequences in terms of theirabstract structure that is now provided as input to State(Equation 1 ) Thus in terms of the recoded abstractstructure representation provided to State the two se-quences ABCBAC and DEFEDF are equivalent u u u n -2 n - 4 n - 3 For sequences that follow this ldquorulerdquo thepattern of activation (context) produced in State bysubsequence u u u n - 2 will reliably be followed bythat context associated with n - 4 To exploit this pre-dictability the system should then take the contents of

the STM for the n - fourth element and direct it to theoutput yielding an RT reduction This is achieved in thefollowing manner For each STM element (ie the struc-tures that store the n - 1 n - 2 frac14 responses) there isa unit that modulates (Equation 8) the contents of thisstructure to Output If one of these units is active thecontents of the corresponding STM structure is directedto Output (Equation 4 )

Now during learning each time a match is detectedbetween the current response and an STM element(Equation 6) the connections are strengthened betweenState units encoding the current context and the modu-lation unit for the matched STM element (Equation 7)The result is that the next time this same pattern ofactivation in State occurs (ie before a match n - 4corresponding to the learned rule) the contents of theappropriate STM will be directed to Output in anticipa-tion of the predicted match thus yielding a reduced RTIn the same sense that the surface model learns toanticipate specic elements that dene a given se-quence the abstract model learns to anticipate a repeti-tive abstract structure that denes a class of isomorphicsequences

We have now dened two formal models for thetreatment of surface and abstract structure respectivelythat together make up the dual process model By usingdifferent randomized starting conditions for the modelsrsquoconnections we can generate multiple model subjectsthat permit the same statistical analysis to be performedin parallel on human and model populations In thefollowing section we will demonstrate a double dissocia-tion in the modelsrsquo processing capabilities That is wewill show that the surface model learns surface but notabstract structure and that the opposite is true for theabstract model We will make a special effort to arguethat indeed there is no feasible combination of parame-ter settings etc that can yield a single model that iscapable of both types of processing These observationsconcerning the dual process model then provide thebasis for a series of predictions that are subsequentlytested in human subjects

SIMULATION 1 RESULTS LEARNINGSURFACE AND ABSTRACT STRUCTURE

This SRT test using sequences with both surface andabstract structure was performed on groups of ve sur-face and ve abstract models with results supporting thehypothesis that dissociable systems are required to treatsurface and abstract structure The mean RTs for predict-able and unpredictable responses in blocks 1 through 10are presented in Figure 2 for the two model groupsRecall that ldquopredictablerdquo and ldquounpredictablerdquo refer to theabstract structure whereas all responses are ldquopredict-ablerdquo by the surface structure The abstract model(group) displays reduced RTs only for the predictable

Dominey et al 737

versus unpredictable responses in blocks 1 through 6evidence for pure abstract structure learning The surfacemodel displays reductions for both evidence for surfacestructure learning In the test of transfer to randommaterial in block 7 both models display negative transferfor the predictable elements whereas for the unpre-dictable elements only the surface model shows negativetransfer In the test of transfer to a new isomorphicsequence with the same abstract structure in blocks 9and 10 the abstract model displays complete transfer tothe new sequence whereas the surface model displaysnegative transfer in block 9 with some improvement inblock 10

Statistical Conrmation of Observations

The observations about learning and transfer wereconrmed by a 2 (Model abstract surface) times 2 (Predict-ability) times 6 (Block 1ndash6) analysis of variance (ANOVA)The interaction between Model and Predictability wasreliable F(1 336) = 289 p lt 00001 The effect of ModelF(1 336) = 183 p lt 00001) was also reliable as werethe effects for Predictability F(1 336) = 430 p lt 00001and for Block F(5 336) = 118 p lt 00001

The observations about negative transfer to a randomseries for both models was conrmed by a 2 (Modelabstract surface) times 2 (Predictability) times 2 (Block 6 and 7)ANOVA The three main effects were reliable as was the

interaction between Model Block and Predictability F(1112) = 55 p lt 00001 reecting the striking absence oflearning and negative transfer for the abstract modelwith unpredictable elements

The observation about transfer of acquired knowledgeto the new sequence was conrmed by a 2 (Predict-ability) times 2 (Model surface abstract) times 2 (Block 9 and10) ANOVA The Model effect was reliable F(1 112) =458 p lt 005 as were those for Predictability F(1112) = 7512 p lt 00001 and Block F(1 112) = 693p lt 001 The interaction between Model and Predict-ability was reliable F(1 112) = 1163 p lt 00001 Inseparate ANOVAs for the two models we conrmed asignicant Predictability effect for the abstract model(F(1 56) = 2956 p lt 00001) but not for the surfacemodel (F(1 56) = 16 p = 02) This indicates that onlyfor the abstract model did the abstract knowledge of therule transfer to the new isomorphic sequence

Discussion

In theory different types of information structure mustbe treated by different processes and the inverse agiven processing architecture must be capable of treat-ing some but not other information structures Thesesimulation results demonstrate that for surface and ab-stract structure as we have dened them there are twocorresponding sequence learning models each of whichis capable of learning only one of these types of sequen-tial structure Specically the abstract model was sensi-tive only to the sequence elements that were predictableby the abstract structure and demonstrated strictly nolearning for the sequence elements that were unpre-dictable by the abstract structure despite the fact thatthese elements recurred in a regular fashion in thesurface structure For this reason it was not sensitive tothe change in surface structure in blocks 9 and 10 Incontrast the surface model was insensitive to thepredictableunpredictable distinction in the abstractstructure displaying reduced RTs for both types of ele-ments indicative of learning the surface structure Thisknowledge however was seen to be inadequate forsupporting transfer to the isomorphic sequence ofblocks 9 and 10 which was effectively treated as a newsequence

SIMULATION 2 RESULTS LEARNINGABSTRACT STRUCTURE ALONE

Here we consider performance of the two models whenonly an abstract structure is present in repeating se-quences Figure 3 displays the mean RT values for pre-dictable and unpredictable elements for the Surface andAbstract models During training in blocks 2 through 4the predictable structure had a large effect on RTs pri-marily for the abstract model although there appears tobe some effect of predictability in the surface models as

Figure 2 Simulation 1 Mean RTs for predictable and unpredictableresponses in the 10 blocks of trials for Surface and Abstract modelgroups Blocks 1 through 6 and 8 have the same surface and ab-stract structure The isomorphic sequence in blocks 9 and 10 retainsthis abstract structure but has a different surface structure Block 7is random The critical blocks for learning and transfer assessmentare marked in the rounded boxes The surface model learns only thesurface structure and does not transfer to the isomorphic sequenceThe abstract model learns only the abstract structure and transfersthis knowledge to the isomorphic sequence Rtime expressed insimulation time units (stu) where one network update cycle corre-sponds to 0005 stu


well There was a negative transfer to the nal randomblock but only for the predictable elements in the ab-stract model


The observations about learning were conrmed by a 2(Model surface abstract) times 2 (Predictability predictableunpredictable) times 3 (Block 2ndash4) ANOVA The Model timesPredictability interaction was reliable F(1 78) = 537p lt 00001 reecting the performance benet of thepredictable abstract structure only for the abstractmodel The main effect of Model F(1 78) = 59 p lt 005was reliable as was that for Predictability F(1 78) = 969p lt 00001

The observations about negative transfer to the ran-dom material in block 5 were conrmed by a 2 (Modelsurface abstract) times 2 (Predictability predictable unpre-dictable) times 2 (Block 4 and 5) ANOVA The Model timesPredictability times Block interaction was reliable F(1 52) =68 p lt 005 A posthoc test (Scheffe) revealed that thenegative transfer was signicant only for the abstractmodel with predictable elements

The observation that the surface model may havedisplayed a prediction effect was conrmed by a 2 (pre-dictable unpredictable) times 3 (Block 2ndash4) ANOVA Indeedthe surface model displayed a reliable effect for Predict-ability F(1 39) = 535 p lt 005 indicating that thisarchitecture is capable of acquiring some of the predict-able structure of the isomorphic sequences

Discussion

These results conrm that even in the absence of surfacestructure the abstract model learns and transfers knowl-edge of this abstract structure and that the surface modelfails to do so at the same level However the observationthat the surface model does acquire some knowledge ofthe abstract structure requires explanation

This observation is counter to our position that thesurface model should be incapable of exploiting theabstract structure The explanation for this observationlies in the potential capacity of the surface model toexploit the weak but present surface structure of thesesequences in terms of single-item recency Consider thefollowing sequence fragment ABC BCD CDE inwhich the rst two elements of each triplet are predict-able and the third is unpredictable by the abstract struc-ture ldquon - 2 n - 2 urdquo Note the single-item recency of lag- 1 for predictable elements (eg B and C in BCD)versus the single-item recency of lag - 19 for unpre-dictable elements (eg D in BCD) All sequence ele-ments that are predictable by the abstract structure arealso favored in terms of single-item recency Because thesurface model cannot represent the abstract structure asdened above its small but signicant predictabilityeffect appears to be purely a function of single-itemrecency differences between predictable and nonpre-dictable elements (surface structure) and does not cor-respond to the abstract structure and thus cannotprovide the basis for transfer of learning to isomorphicsequences Indeed in Simulation 1 the single-item re-cency difference for predictable versus unpredictableelements is smaller (lag - 2 and lag - 8 versus lag - 1and lag - 19 in Simulation 2) and there is no predict-ability effect for the surface model in Simulation 1 (F(1196) = 318 p gt 01) nor is there any transfer to theisomorphic sequence in blocks 9 and 10

With respect to our argument that two models arenecessary to accommodate surface and abstract struc-ture one might argue that in Simulation 2 a single modelcould produce both types of behavior based on single-item recency with a simple gain reduction responsiblefor the smaller predictability effect in the implicitsurface condition This argument fails however when weapply it to Simulation 1 There the difference betweenthe Abstract and Surface models is a difference of kindnot quantity demonstrating behaviors that cannot beattributed to a common system

TESTING SIMULATION PREDICTIONS INHUMANS

By extending these observations which support dissoci-able systems to human SRT performance we can nowtest the hypothesis that surface and abstract structureare processed by dissociable systems in humans Animportant issue in testing this hypothesis is to develop

Figure 3 Simulation 2 Mean RTs for predictable and unpredictableresponses in the ve blocks of trials for surface and abstract modelgroups The rst (Rand1) and last (Rand2) of the ve blocks of 120trials are random Block 2ndash4 (S1 S2 and S3) are made up of threedifferent repeating isomorphic sequences one per block The criticalblocks for learning and transfer assessment are marked in therounded boxes The abstract model learns and transfers the abstractstructure and the surface model does not Rtime expressed in simula-tion time units (stu) where one network update cycle correspondsto 0005 stu

Dominey et al 739

a method to isolate such processes in humans As men-tioned above several recent studies demonstrate thatabstract rule processing involves explicit intentional ef-fort in mapping the appropriate rule onto the targetproblem (Gick amp Holyoak 1983 Gomez 1997 Mathewset al 1989 Shanks amp St John 1994) whereas surfacestructure processing is likely to be a more automaticimplicit function (eg Cohen et al 1990 Curran ampKeele 1993) We consider that in implicit conditionsprimarily processes capable of learning surface but notabstract structure will be accessible whereas in explicitconditions both will be accessible

We thus employ two experimental conditions to dis-sociate the effects of surface structure versus abstractstructure learning In the Explicit condition the subjectswere shown a diagram visually depicting the abstractstructure in question and were told before and onceduring the examination that such a rulelike structuremight be found in the subsequent testing and thatsearching for and nding such a structure could aid theirperformance In the Implicit condition the subjects weresimply told that they should respond as quickly andaccurately as possible In the following three experi-ments we compare explicit and implicit human perfor-mance in SRT tasks that manipulate surface and abstractstructure with the goal of demonstrating that learningsurface and abstract structure rely on functionally disso-ciable mechanisms Experiment 1 thus repeats the pro-tocol that was used in Simulation 1

Experiment 1 Results Learning Surface andAbstract Structure

The analysis focused on the RT data for predictable andunpredictable events The mean RTs for predictable andunpredictable responses in blocks 1 through 10 arepresented in Figure 4 for the Explicit and Implicitgroups Both groups display progressively reducing RTsin blocks 1 through 6 with an additional reduction forthe predictable elements seen only in the Explicit groupThis suggests that although both groups prot from thesurface structure only the Explicit group benets addi-tionally from the abstract structure This predictabilityadvantage in the Explicit group does not appear tochange over blocks 1 through 6 Both groups displaynegative transfer to random material in block 7 andpositive transfer to block 8 which is constructed asblocks 1 through 6 The test of transfer involves newmaterial with different surface structure but the sameabstract structure in blocks 9 and 10 The explicit grouptransfers knowledge of the abstract structure as revealedby signicantly reduced RTs for predictable elements inblock 9 and 10 whereas the implicit group shows noevidence of learning or transfer of the abstract informa-tion

In posttest interviews all of the Explicit group sub-jects reported awareness of a repeating structure in the

sequence blocks and were able to sketch a gure thatreected ABCBAC structure The sketches were consid-ered to reect reasonable awareness if they included atleast two of the three repetitions n - 2 n - 4 and n -3 None of these subjects reported a specic awarenessof the 12-element sequence ABCBACDEFEDF and ourinterview did not attempt to quantify partial knowledgeNone of the Implicit group subjects reported any aware-ness of the underlying abstract structure or a specicawareness of the 12-element sequence ABCBACDEFEDF


The observations about learning in blocks 1 through 6were conrmed by a 2 (Group Explicit Implicit) times 2(Predictability) times 6 (Block 1ndash6) times ANOVA The effect ofGroup F(1 696) = 1558 p lt 00001 was reliable aswere the effects for Predictability F(1 696) = 107 p lt0005 and for Block F(5 696) = 360 p lt 00001 Theinteraction between Group and Predictability was reli-able F(1 696) = 142 p lt 00005 corresponding to theobservation of a Predictability effect in the Explicit butnot Implicit group This was conrmed by the absenceof a Predictability effect for the Implicit group (F(1348) = 011 p = 07) The observation that the predict-ability effect did not vary with block was conrmed bythe nonsignicant interaction between Predictability andBlock (1 through 6) for the Explicit group F(5 348) =018 p gt 09

The observations about negative transfer to a randomseries in block 7 for both groups were conrmed by a

Figure 4 Experiment 1 Mean RTs for predictable and unpre-dictable responses in the 10 blocks of trials for Explicit and Implicitsubjects Blocks 1 through 6 and 8 have the same surface and ab-stract structure Blocks 9 and 10 retain this abstract structure buthave a different surface structure Block 7 is random The criticalblocks for learning and transfer assessment are marked in therounded boxes Explicit subjects learn surface and abstract structureand display transfer to blocks 9 and 10 Implicit subjects learn onlysurface structure with no transfer


2 (Group explicit implicit) times 2 (Predictability) times 2(Block 6 and 7) ANOVA The interaction between Groupand Block was reliable F(2 232) = 224 p lt 00001Planned comparisons revealed signicant differences inboth groups between RTs in blocks 6 and 7 for predict-able and unpredictable events The effect for Block wasalso reliable F(1 232) = 2579 p lt 00001 as was thatfor Group F(1 232) = 67 p lt 005

The observation about transfer of acquired knowledgeto the new sequence was conrmed by a 2 (Predict-ability) times 2 (Group explicit implicit) times 2 (Block 9 and10) ANOVA The Group effect was reliable F(1 232) =387 p lt 00001 as were those for Predictability F(1232) = 1537 p lt 00005 and Block F(1 232) = 185p lt 00001 The interaction between Group and Predict-ability was reliable F(1 232) = 52 p lt 005 SeparateANOVAs for the two groups conrmed a signicant Pre-dictability effect for the Explicit group (F(1 116) = 176p lt 00001) but not for the Implicit group (F(1 116) =17 p = 023)

Discussion

The results from the Implicit and Explicit groups provideclear evidence for a dissociation between processing ofthe surface and abstract structures of a sequence thathas both structures Explicit subjects acquired and usedboth types of knowledge and transferred the abstractknowledge to a new isomorphic sequence Note thattransfer does not imply improvement on the transferblock but simply the maintenance of the previouslyestablished performance The Implicit group acquiredonly the surface structure and did not transfer this infor-mation to the isomorphic sequence It is important tonote that with respect to the abstract structure thegroup difference is a difference of kind not of degreeThere is no evidence of acquisition of the abstract struc-ture in the Implicit group In contrast both groups dem-onstrate a signicant learning of the surface structureThe fact that surface structure can be acquired inde-pendently of abstract structure argues strongly for theexistence and use of distinct processes for treating sur-face and abstract structure

Because our Explicit subjects were briefed on the ruletheir performance improvement for predictable ele-ments could be attributed to their learning to apply therule rather than their learning or discovery of the ruleitself The point of interest however is that knowledgeof the rule yields a performance advantage for predict-able but not unpredictable elements both in the initialtraining sequence and in a new isomorphic sequenceindicating a transfer of the rule-based information

Comparing Simulation and Human Performance

It is of interest to note the difference between thehuman and simulation data in these conditions Whereas

the surface model predicts the behavior of the Implicitgroup rather well the abstract model differed from theExplicit group in two major ways First it displayed nolearning for the unpredictable elements whereas theExplicit group showed signicant learning To under-stand this difference we note that the simulations allowa complete isolation of the surface and abstract systemsbut this is not possible in humans That is for explicitlearning in humans we cannot prevent the surface (im-plicit) system from operating in parallel with the abstract(explicit) system This is demonstrated by the signicantlearning effect in the Explicit group for elements thathave surface structure but are unpredictable by the ab-stract structure This along with the Group times Block (6and 7) interaction allows us to consider that there is anadditive effect between these two systems in humans(Mathews et al 1989)

The second major difference is the lack of negativetransfer for predictable elements in block 9 for theAbstract model as compared to the Explicit subjects Theabstract model does not in fact make any distinctionbetween blocks 8 and 9 because it operates entirely interms of abstract structure which is identical for thesetwo isomorphic sequences Explicit subjectsrsquo perfor-mances on predictable elements of the abstract struc-ture on block 9 however are inuenced by the changein surface structure even though the abstract structureremains the same demonstrating the additive effect be-tween surface and abstract structure processing mecha-nisms in humans

As we previously stated simulation of the abstractmodel alone is psychologically invalid in the sense thatalthough the processes related to abstract structure canbe engaged (or not) as a function of the pretrial instruc-tions and intention the processes that treat surfacestructure are engaged by default Thus we should con-sider the combined behavior of both of the dual proc-esses to simulate what occurs in explicit conditions Wesimulate the performance of such a dual process modelas a nonlinear combination of the performance of thesurface and abstract models This performance and thatof explicit human subjects were compared by apiecewise linear regression on the 20 data pointsdened by the RTs for predictable and unpredictableelements for the dual-process model and the Explicitgroup yielding a signicant correlation r2 = 079 p lt000001 versus the correlation r2 = 033 p = 00093 forthe Abstract versus Explicit comparison Figure 5 dis-plays the resulting behavior for the dual process modelWe now see humanlike performance regarding learningfor the unpredictable events and negative transfer inblock 9 is now seen in the hybrid model A linear regres-sion analysis for the Surface model and Implicit grouprevealed a correlation of r2 = 0853 p lt 00001

Dominey et al 741

Experiment 2 Results Learning AbstractStructure Alone

We now test the learning and transfer of abstract struc-ture (ABCBAC) with continuously changing surfacestructure The mean RTs for predictable and unpre-dictable responses in blocks 1 through 9 are presentedin Figure 6 for the Explicit and Implicit groups TheExplicit group displays greatly reduced RTs for the pre-dictable versus unpredictable responses in blocks 1through 6 whereas such a reduction is not seen for theImplicit group In the test of transfer to a new but similarabstract structure (ABACBC) in block 7 and transfer torandom material in block 9 the Explicit group displayeda large RT increase for the predictable elements but notfor the unpredictable ones This effect appears absent inthe Implicit group This difference in behavior for trans-fer to the new abstract structure (block 7) indicates thatthe Explicit group learns the abstract structure itself butthe Implicit group does not

In the posttest interviews all of the Explicit subjectsreported awareness of a repeating structure in the se-quence blocks and were able to sketch a gure thatreected some knowledge of the ABCBAC structure Thesketches were considered to reect reasonable aware-ness if they included at least two of the three repetitionsn - 2 n - 4 and n - 3 As in Experiment 1 none of theImplicit subjects reported any awareness of the underly-ing structure


The observations about learning and transfer of the ab-stract structure were conrmed by a 2 (Group explicitimplicit) times 2 (Predictability predictable unpredictable)times 6 (Block 1ndash6) ANOVA The interaction between Group

and Predictability was reliable F(1 336) = 705 p lt00001 corresponding to the observation that the pre-diction effect is seen primarily in the Explicit group Theeffect of Group F(1 336) = 48 p lt 005 was alsoreliable as were the effects for Predictability F(1 336) =1068 p lt 00001 and for Block F(5 336) = 85 p lt00001

The observations about negative transfer to a differentabstract structure and to a random series for the Explicitsubjects were conrmed by a 2 (Group explicit Im-plicit) times 2 (Predictability predictable unpredictable) times 3(Block 7ndash9) times ANOVA The Group effect was reliable F(1168) = 428 p lt 00001 as were those for PredictabilityF(1 168) = 5311 p lt 00001 and Block F(2 168) = 126p lt 00001 and all three of the two-way interactionsPlanned comparisons revealed signicant differences be-tween RTs in blocks 7 and 8 (new versus learned ab-stract structure) and blocks 8 and 9 (learned abstractstructure versus random) only for the Explicit subjectsand only for the predictable elements

The observation about a possible Predictability effectin the Implicit group was conrmed by a 2 (Predict-ability) times 6 (Block 1ndash6) ANOVA The effect for Predict-ability was just reliable F(1 168) = 397 p = 0048However the lack of negative transfer in block 7 asrevealed by planned comparison indicates that the pre-dictability effect for the Implicit group is in fact due tosingle-item recency as described in Simulation 2 ratherthan learning that is specic to the abstract structure

Figure 5 Experiment 1 Comparison of explicit human perfor-mance and dual process model performance Same format as Fig-ure 4 see text

Figure 6 Experiment 2 Mean RTs for predictable and unpre-dictable responses in the nine blocks of trials for Explicit and Im-plicit subjects There is no repetitive surface structure throughoutthe nine blocks Block 7 uses a different abstract structure than thatin blocks 1 through 6 and 8 and block 9 is random The criticalblocks for learning and transfer assessment are marked in therounded boxes Only Explicit subjects learn the abstract structureand display negative transfer to the novel abstract structure inblock 7


Discussion

This experiment demonstrated that abstract knowledgecan be acquired by the Explicit subjects even in theabsence of repeating surface structure and that thisknowledge is specic to a given abstract structure anddoes not transfer to related but different abstract struc-ture Indeed for the predictable elements the Explicitgroup showed negative transfer to a new abstract struc-ture and to random material ruling out the possibilitythat their predictability effect is due to single-item re-cency In contrast in the Implicit group there was nochange in the small predictability effect when the ab-stract structure was changed indicating the use of re-cency cues in the surface structure This allows us toconclude that the relatively small predictability effect inthe Implicit group is not due to weak learning of theabstract structure but rather to single item recency ef-fects Note again however that the single-item recencyexplanation does not hold for the Explicit group becausethe negative transfer in block 7 is unexplained


Figure 7 displays the mean RT values for predictable andunpredictable elements for the Explicit and Implicitgroups in the task described in Simulation 2 Duringtraining in blocks 2 through 4 the predictable structurehad a large effect on RTs primarily for the Explicit groupas learning accumulated over the three sequence blocksalthough there appears to be some effect of predict-ability in the Implicit group as well There was a largenegative transfer to the nal random block but only forthe predictable elements in the Explicit group

In the posttest interviews all of the Explicit subjectsreported awareness of a repeating structure in the threesequence blocks and were able to sketch a gure thatreected the ldquon - 2 n - 2 urdquo structure The sketcheswere considered to reect reasonable awareness if theyincluded a directed graph representing the pattern A-B-A-B-C In contrast none of the Implicit subjects reportedany awareness at all of the underlying analogical schemaSeveral reported that if there was any pattern it was toolong and complicated to be learned


The observations about training were conrmed by a 2(Group explicit implicit) times 2 (Predictability predictableunpredictable) times 3 (Block 2ndash4) ANOVA The Group timesPredictability interaction was reliable F(1 78) = 373p lt 00001 indicating that the performance benet ofthe predictable abstract structure is dependent as pre-dicted on Group The main effect of Group F(1 78) =51 p lt 00001 was reliable as were those for Predict-ability F(1 78) = 746 p lt 00001 and for Block F(2 78)

= 72 p lt 0005 and for the Predictability times Blockinteraction F(2 78) = 448 MSE = p lt 0005 The obser-vations about a progressive improvement in the Explicitgroup were conrmed by a 2 (Predictability) times 3 (Block2ndash4) ANOVA with a signicant interaction (F(2 39) =569 p lt 00001)

The observations about negative transfer to the ran-dom material in block 5 were conrmed by a 2 (Groupexplicit implicit) times 2 (Predictability predictable unpre-dictable) times 2 (Block 4 and 5) ANOVA The Group timesPredictability times Block interaction was reliable F(1 52) =114 p lt 0005 reecting the observed negative transferfrom block 4 to 5 only for the Explicit group withpredictable elements A post hoc test (Scheffe) revealedthat the negative transfer was signicant only for theExplicit group with predictable elements

The observation that the Implicit group may havedisplayed a Predictability effect was not conrmed by a2 (predictable unpredictable) times 3 (block 2ndash4) ANOVAHowever the implicit group displayed a nearly reliableeffect for Predictability F(1 39) = 39 MSE = 8880 p =0055 suggesting that like the Surface model they ex-ploited single-item recency effects in the isomorphicsequences Neither the Block effect nor the interactionwere reliable

Discussion

To compare the results of the surface model with thosefrom the Implicit subjects a linear regression was ap-plied to the 10 RT means (predictable and unpredictablein the ve blocks) demonstrating a signicant correla-

Figure 7 Experiment 3 Mean RTs for predictable and unpre-dictable responses in the ve blocks of trials for Explicit and Im-plicit groups The rst and last of the ve blocks of 120 trials arerandom (Rand) Block 2ndash4 (S1 S2 and S3) are made up of three dif-ferent repeating isomorphic sequences one per block The criticalblocks for learning and transfer assessment are marked in therounded boxes Only Explicit subjects learn the abstract structure

Dominey et al 743

tion with r 2 = 076 p = 0001 The same analysis for theabstract model and Explicit subjects also yielded sig-nicant correlation with r2 = 093 p = 0000005

These data reconrm the observation that explicitlearning conditions are necessary to encode and transferabstract structure Likewise simulation data indicate thatonly the abstract model architecture is adequate forlearning the abstract structure that is common to thethree sequence blocks and also imply that small predict-ability effects in the Implicit group might be due tosingle-item recency This suggests that in explicit learninga processing capabilitymdashcorresponding to that of theabstract modelmdashis enabled whereas it is not enabled inthe implicit learning conditions

GENERAL DISCUSSION

A given sequence of stimuli can encode different typesof information or structure Depending on the type ofencoded structure different mechanisms may be re-quired to extract that structure (eg Chomsky 1959Turing 1936) We dene surface structure in terms of thestraightforward serial order of sequence elements Incontrast abstract structure is dened in terms of orderedrelations between repeating sequence elements Thusalthough the two sequences ABCBAC and DEFEDF havedifferent surface structures their abstract structures areidentical The purpose of the current research has beento test the hypothesis that as dened surface and ab-stract sequential structure are processed by distinct anddissociable systems Separate results from simulation andrelated human experiments support this hypothesis andcombined with recent neuropsychological results pro-vide a framework for an initial specication of the un-derlying neurophysiology

Simulation Evidence for Dissociable Processes

It has been clearly demonstrated that although our ldquosur-face modelrdquo of sequence learning based on the func-tional neuroanatomy of the primate fronto-striatal system(Dominey Arbib et al 1995) is quite capable of display-ing humanlike performance for learning surface struc-ture (Dominey 1995 1998a 1998b) as well as explainingprimate cortical electrophysiological results in cognitivesequencing tasks (Dominey Arbib et al 1995 Domineyamp Boussaoud 1997) it fails to learn abstract structure(Dominey 1997b Dominey et al 1995a 1995b) Thisprovides a formal argument for the independence ofprocesses that learn surface and abstract structure re-spectively Part of what is missing in the surface modelis a specialized short-term or working memory that hasbeen proposed to be necessary to construct the map-ping from source to target problems in analogical rea-soning (Holyoak et al 1994 Holyoak amp Thagard 1989Thagard Holyoak Nelson amp Gochfeld 1990) When thesurface model is modied to include a short-term mem-

ory of the previous responses that can be comparedwith current responses so as to encode the repetitivestructure the resulting abstract model is capable of learn-ing and exploiting the abstract structure of sequencesindependently of the surface structure (Dominey 1997a1997b 1998c Dominey et al 1995a 1995b) This pro-vides the second half of the dissociation The surfacemodel learns surface but not abstract structure and theabstract model learns abstract but not surface structurewith the surface model simulating performance of theimplicit group and the combined surface and abstractmodels simulating performance of the explicit group

One might argue however that the more powerfulAbstract model could simulate both groupsrsquo perfor-mance If the extent of the short-term memory is in-creased to accommodate the 12 previous events theAbstract model could learn the surface structure of se-quence ABCBACDEFEDF from Experiment 1 based onthe abstract structure ldquon - 12 n - 12 frac14 n - 12rdquo In thissense one might suggest that the same model couldexplain behavior attributed to distinct processes for sur-face and abstract structure with explicit and implicithuman performance modeled by a simple gain modica-tion in the single model There are however at least twoaws in this approach

First as we recall from Experiment 1 the Implicit andExplicit groupsrsquo performances are different in kind notin degree The highly visible predictability effect in theExplicit group does not exist in the Implicit group Thusa simple ldquogainrdquo change in the abstract model can accountfor this difference Second for the implicit group thedata could be explained by the abstract model only if (1)the size of the STM is raised to 12 elements (versus aminimal STM size of only 4 required for the explicitgroup) and (2) the implicit group is modeled by a muchmore complex and memory-intensive system than theexplicit group (ie STM extent of 12 versus 7 plusmn 2) Thusthe single-model approach requires one to defend theidea that a system that is quite ldquooverqualiedrdquo is at workin all manipulations of surface structure In taking thisposition one is forced to reject a large body of work onrecurrent network learning (eg Cleeremans amp McClel-land 1991 Dominey 1995 1998a 1998b Elman 1990)and obliged to say that even though recurrent networkscan simulate SRT and related results a more complexmodel must be proposed to avoid a dual-process expla-nation

We clearly admit however that the dual-processmodel as proposed is by no means complete That isthere are related forms of rule-based abstract structuresthat cannot be processed by the abstract model Forexample consider the sequence ldquoA-B-C left of A left ofB left of Crdquo The rst three elements are unpredictableand the next three are predictable based their spatialrelations with the rst three Although humans are likelyto pick up on this rule and transfer it to isomorphicsequences the abstract model in its current state would


not because it doesnrsquot have the representational hard-ware to exploit these spatial relations

Experimental Evidence for Dissociable Processesfrom Healthy Subjects

The results of manipulation of surface and abstract struc-ture in three experiments with human subjects provideevidence that two distinct mechanisms are required totreat surface and abstract structure respectively in hu-mans Experiment 1 provided the crucial test of whetherknowledge of surface structure could be acquired inde-pendently of knowledge of abstract structure while bothwere present In agreement with the proposed hypothe-sis implicit subjects although capable of signicantlearning of surface structure did not learn the abstractstructure nor display transfer to the new isomorphicsequences Experiments 2 and 3 demonstrated that inthe absence of surface structure only explicit subjectsare capable of extracting the abstract structure and trans-ferring it to isomorphic sequences and Experiment 2demonstrated that this learning is highly specic to thelearned abstract structure

It is important to note that we do not claim thatexplicit learning is equal to abstract structure learningnor that implicit learning is equal to surface structurelearning Our point is to demonstrate the existence ofdistinct processes for distinct types of information proc-essing To do so we choose to observe performance inthe functional modes (implicit versus explicit) in whichthese processes may be expressed or liberated Ourchoice is supported by the observation of Gomez (1997)that in an implicit sequence learning task surface struc-ture was learned but no learning or transfer of abstractstructure occurred This suggests that holistic exposureto entire strings permits additional processing that is notpossible in an element-by-element presentation as in ourSRT tasks Likewise in an AGL task using the same mate-rial but presented in letter strings Gomez observed thatsurface structure (rst-order dependencies) learning canoccur without explicit awareness whereas learning ab-stract structure (supporting transfer to changed lettersets) is invariably linked to explicit knowledge The de-bate over the possibility of transfer with implicit learn-ing however is not yet resolved (Gomez 1997Knowlton amp Squire 1993) and is likely to require the useof control subjects in the testing conditions and a carefulcontrol over nongrammatical cues that could bias trans-fer scores In this framework although AGL has beenoften considered a test of implicit learning we recall thatthe testing phase includes an instruction to exploit rule-based regularities that probably invokes explicit abstrac-tion processes (Perruchet amp Pacteau 1991 Reddingtonamp Chater 1996) It is just this kind of explicit instructionto directs onersquos attention that can lead subjects in prob-lem-solving studies to abstract a shared rule based onprevious problems to solve the current problem (Gick

amp Holyoak 1983) Such behavioral process shifting isconsistent with recent observations that attentional stateand awareness can inuence the selection of neuro-physiological cognitive processes (Grafton Hazeltine ampIvry 1995 Treue amp Maunsell 1996)

Toward a Neurophysiological Basis forDissociable Systems

We can now begin to specify a neurophysiological basisfor these dissociable systems for surface and abstractstructure processing in terms of recent results fromneuropsychological experiments Patients with fronto-striatal dysfunction due to Parkinsonrsquos disease are sig-nicantly impaired in a SRT task that requires learningsurface structure under implicit conditions yet theyhave near normal performance when the task becomesexplicit (Pascual-Leone et al 1993) More interestinglythese patients display an intact capability to learn ab-stract sequential structure in explicit conditions(Dominey et al 1997 Dominey amp Jeannerod 1997) Thisindicates that although the fronto-striatal system pro-vides part of the neurophysiological basis for implicitprocesses that can acquire surface structure it is lessinvolved in explicit processes that treat abstract struc-ture This is supported by the demonstration that ourmodel of the primate cortico-striatal system that explainsprefrontal electrophysiology during primate sequencelearning tasks (Dominey Arbib et al 1995 Dominey ampBoussaoud 1997) is also capable of learning surfacestructure yet fails to learn abstract structure (Dominey1997b Dominey et al 1995a 1995b)

In comparison to the Parkinsonrsquos disease patientsschizophrenic patients yield an opposite prole and an-other piece of the puzzle That is although schizophrenicpatients display signicant learning of surface structurethey fail to learn abstract structure (Dominey amp Geor-gieff 1997) Because schizophrenia is characterized inpart by a hypoactivity of the left anterior cortex (SuzukiKurachi Kawasaki Kiba amp Yamaguchi 1992) we canconsider that the impaired abstract structure learning inthese participants might be related to this left hypofron-tality Such an interpretation ts well with the initialmotivation behind the development of the abstractmodel which was to account for the ability to generalizebetween ldquogrammaticallyrdquo related but novel sensorimotorsequences (Dominey 1997b) This work was based inpart on related ideas from Greeneld (1991) suggestingthat linguistic syntax and abstract aspects of motor con-trol are treated in Brocarsquos area and adjacent corticalmotor areas in the left anterior cortex It is thus ofinterest to note that the left hypofrontality may contrib-ute not only the failed processing of abstract structurethat we observed in schizophrenic patients (Dominey ampGeorgieff 1997) but also to their impairment in gram-matical language processing (eg Portnoff 1982) Wethus propose that surface structure can be learned by

Dominey et al 745

processes that can operate under implicit conditions andrely on the fronto-striatal system whereas learning ab-stract structure requires a more explicit activation ofdissociable processes that rely on a network that in-cludes the left anterior cortex

Conclusion

From the theoretical perspective that fundamentally dif-ferent information structures must be treated by distinctcomputational machines we predicted the existence ofcomputationally behaviorally and neurophysiologicallydissociable systems for treating surface and abstractstructure in sensorimotor sequences Starting with a re-current network that is based on the functionalneuroanatomy of the fronto-striatal system we demon-strated that surface structure can be learned inde-pendently of abstract structure and that only afterundergoing modications that provide distinct repre-sentational capabilities can the updated model learnabstract structure We propose that the surface (fronto-striatal) model corresponds to human processes that areaccessible in implicit conditions and that the abstractmodel corresponds to human processes that must bedeliberately put into play in explicit conditions Thisconcurs with an increasing body of evidence that trans-fer performance in articial grammar and sequencelearning is linked to explicit knowledge and processing(Gomez 1997 Mathews et al 1989 Perruchet amp Pacteau1991 Reddington amp Chater 1996) Three human experi-ments designed around these assumptions demonstratedthat the processing of surface and abstract structure canbe behaviorally dissociated in human subjects corre-sponding to the dissociation between the surface andabstract models These results support our initial hy-pothesis that surface and abstract structure are treatedby distinct mechanisms Combined with our neuropsy-chological results the current results are consistent withthe position that implicit processes for surface structurelearning depend on the fronto-striatal system whereasprocesses for abstract structure learning that are re-vealed in explicit conditions rely in a dissociated fashionon a network that includes the left anterior cortex Thisis in agreement with the general position that instancesversus rules or categories are probably treated by disso-ciable information processing systems in humans (egKnowlton amp Squire 1993 1996 Shanks amp St John 1994)

METHODS

Simulation 1 Learning Surface and AbstractStructure

The rst simulation is designed to test the modelsrsquo abilityto learn surface and abstract structure when both arepresent in the same sequence to determine if it is pos-sible to learn surface structure independently of abstract

structure The SRT test we employ consists of 10 blocksof 108 trials each where each trial corresponds to thepresentation of a single-sequence element Blocks 1through 6 and 8 use an abstract structure of the form uu u n - 2 n - 4 n - 3 (where ldquourdquo signies unpredictableand ldquon - 2rdquo indicates a repetition of the element twoplaces behind etc) This abstract structure recurs twicein the 12-element sequence ABCBACDEFEDF that re-peats nine times in each block Thus each repetition ofthe sequence contains two repetitions of the abstractstructure and one repetition of the surface structure Inthis sequence predictable elements have a mean single-item recency of lag - 2 while the unpredictable ele-ments are lag - 8 Block 7 is a random series of elementsBlocks 9 and 10 each use nine repetitions of a new12-element sequence isomorphic to that used in blocks1 through 6 and 8 (ie with the same abstract structurebut with a different surface structure) by using a differ-ent mapping of A through F to elements in the inputarray

In evaluating the performance of the models the fol-lowing constraints should be kept in mind For the ab-stract structure in question (ABC BAC) the rst threeelements (ABC) are considered to be unpredictablewhereas the second three are completely predictable(BAC) Pure abstract structure learning should be mani-fest as a reduction in RTs for the predictable but not theunpredictable elements with a high degree of transferto the isomorphic sequence in blocks 9 and 10 Puresurface structure learning should be manifest as an RTreduction for both predictable and unpredictable ele-ments because that distinction is valid only with respectto the abstract structure In addition there should not besignicant transfer of surface structure learning to theisomorphic sequence in blocks 9 and 10 This experi-ment was performed on ve surface and ve abstractmodels

Simulation 2 Learning Abstract Structure Alone

The rst simulation experiment demonstrated the disso-ciation of surface and abstract structure processingwhen both are present in the same sequence In thissimulation we address two resulting questions First inthe absence of surface structure is the abstract modelstill capable of learning the abstract structure Secondis there truly no information that is available to thesurface model in a sequence that has abstract but notsurface structure

These questions are addressed through the use ofsequences that have a low degree of surface structureand a high degree of abstract structure The experimentconsists of ve blocks of 120 trials each Blocks 1 and 5are randomly organized and blocks 2 through 4 aresequence blocks Sequence blocks are based on 24-element sequences of the form ABC BCD CDE DEF EFGFGH GHA HAB repeated ve times to make a block of


120 trials To appreciate the surface structure complexityof this 24-element sequence note that each of the eight(A through H) elements appears three times with twodifferent successors yielding a complex or ambiguoussequence We thus consider that this sequence has asurface structure that should be relatively difcult tolearn In contrast a clear form of abstract structure be-comes evident if we note that the rst two elements ofeach triplet (eg B and C in BCD) are predictable repe-titions of the elements two places behind them (n - 2)whereas the third is unpredictable (u) This abstractstructure ldquon - 2 n - 2 urdquo repeats throughout the se-quence The predictable elements have a mean single-item recency of lag - 1 whereas the unpredictableelements are at lag - 19

To study the transfer of abstract structure knowledgewe construct three isomorphic sequences by using the24-element pattern described above with three differentmappings of A through H onto eight locations on the 5times 5 Input array Thus the three resulting 24-elementsequences differ completely in their surface structure(ie the serial ordering of their spatial targets) Howeverthey are isomorphic in that they all share the abstractstructure ldquon - 2 n - 2 urdquo This sequence learning experi-ment was performed on ve surface and ve abstractmodels

Human Experiments

Apparatus

In each of the three experiments subjects are seated infront of a touch-sensitive computer screen (Micro-TouchTM) on which we can display the sequence ele-ments (25 cm2) and record response time (from targetonset until subjectrsquos contact with the screen) The eightsequence elements are spatially distributed in apseudorandom fashion over a 25- times 25-cm surface of thescreen as illustrated in Figure 1 The tasks are piloted bya PC using Cortex software (NIH Robert Desimone) Thetasks are based on the SRT protocol (Nissen amp Bullemer1987) and involve pointing to successively illuminatedsequence elements on the touch-sensitive screen asquickly and accurately as possible In a given trial oneof the eight sequence elements is illuminated After anelement is touched it is extinguished the reaction timeis recorded and the next element is displayed

Experiment 1 Learning Surface and AbstractStructure

Simulation 1 demonstrated that the surface modellearned the surface structure of the sequence but madeno distinction between elements that were predictableversus unpredictable by the abstract structure and dis-played no transfer of performance to the new isomor-phic sequence In contrast the abstract model learnedonly the elements that were predictable by the abstract

structure and transferred this knowledge quite effec-tively to the isomorphic sequence Experiment 1 em-ploys the same task in Explicit and Implicit groups ofhuman subjects to determine if the same processingdissociation demonstrated in the models can be repli-cated in humans in order to further demonstrate thedissociation between processes for treating surface ver-sus abstract structure

Experiment 1 Method This experiment uses the sameprocedure as that of Simulation 1 with two groups of 10subjects each in the Implicit (surface) and Explicit (ab-stract) conditions as dened above Blocks 1 through 6and 8 use an abstract rule of the form u u u n - 2 n -4 n - 3 that recurs in the 12-element sequence ABCBAC-DEFEDF (where elements ABCDEF are mapped to ele-ments 285613 in Figure 1) that repeats nine times ineach block Thus each repetition of the sequence con-tains two repetitions of the abstract structure and onerepetition of the surface structure Predictable elementshave single-item recency of lag - 2 and unpredictableelements lag - 8 Block 7 is a random series of elementsBlocks 9 and 10 each use nine repetitions of a new12-element sequence (ABCBACDEFEDF with A throughF mapped to elements 781365) isomorphic to that usedin blocks 1 through 6 and 8 (ie with the same abstractstructure but with a different surface structure) Thedelay between a response and the next stimulus (re-sponse to stimulus interval or RSI) was 200 msec withina six-element subsequence and 500 msec between sub-sequences

Subjects in the Explicit group (N = 10) were shown aschematic representation of the rule and asked to dem-onstrate knowledge of the rule by pointing to BAC givenABC They were told to actively try to use such a rule tohelp them go as fast as possible Subjects in the Implicitgroup (N = 10) were simply told to go as fast as possibleand were given no hint that there might be an underly-ing structure in the stimuli

Experiment 2 Learning Abstract Structure Alone

Experiment 1 provides evidence that abstract structurecan be learned and exploited in an SRT setting whenboth surface and abstract structure are present There aretwo potential criticisms to this interpretation First thelearning of the abstract structure may still require acoherent repeating surface structure and in the absenceof this surface structure the learning of abstract struc-ture may fail Second performance that we are attribut-ing to abstract structure learning may be somethingmuch simpler related to a single-item recency effectThat is the reduced RTs for predictable elements maysimply be due to the fact that all predictable elementshave a mean single-item recency of lag - 2 whereas theunpredictable elements have a recency of lag - 8 Thusthe reduced RT for predictable elements may simply be

Dominey et al 747

due to a recency effect rather than the learning of aspecic abstract structure If so this effect should beinsensitive to a change in abstract structure in transfertests with new sequences provided that the new se-quence maintains a similar degree of exploitable recencyinformation Experiment 2 specically addresses thesequestions in the following manner During training acontinuously changing surface structure and a xed ab-stract structure are used Testing then occurs with acontinuously changing surface structure and a new xedabstract structure that maintains roughly the same de-gree of single-item recency If the abstract structure itselfis being learned we will see negative transfer to the newabstract structure with no such negative transfer if it isprimarily recency information that is being exploited Itis worth mentioning that although some related studiesfocus on the learning of rules versus smaller ldquochunksrdquo ofinformation (eg Knowlton amp Squire 1994) we rule outany learning effect due to element chunking in thecurrent experiment because there are no regularly re-peating chunks (ie no repeating surface structure)

Experiment 2 Method The methods used in Experi-ment 2 are similar to those in Experiment 1 with thefollowing exceptions The test is divided into nine blocksof 90 trials each Blocks 1 through 6 and 8 use an abstractstructure of the form ABCBAC (u u u n - 2 n - 4 n -3) that repeats 15 times Although the abstract structureis always the same the surface structure changes con-tinuously within each block so that the same sequenceis never repeated Specically the mapping between ABCand elements 1 through 8 changes for each sequence andis never repeated Block 7 uses an abstract structure ofthe form ABACBC (u u n - 2 u n - 4 n - 2) also withcontinuously changing surface structure and serves as atransfer test Block 9 uses a random series Predictableelements in the rst abstract structure have a meansingle-item recency of lag - 2 versus lag - 16 for thetransfer abstract structure

Subjects in the Explicit group (N = 5) were shown aschematic representation of the rule and told to activelytry to use such a rule to help them go as fast as possibleas in Experiment 1 Subjects in the Implicit group (N =5) were simply told to go as fast as possible and weregiven no hint that there might be an underlying struc-ture in the stimuli


Simulation 2 demonstrated that for sequences with a lowdegree of surface structure and a high degree of abstractstructure only the abstract model could learn the ab-stract structure and neither model learned the surfacestructure Experiment 3 uses the same protocol as Simu-lation 2 and allows further testing of the prediction thatabstract structure can be learned independently of sur-face structure by the Explicit group and that some re-

cency information may be extracted from the surfacestructure by the Implicit group

Experiment 3 Method As in Simulation 2 the task isdivided into ve blocks of 120 trials Blocks 1 and 5 arerandomly organized and blocks 2 through 4 are se-quence blocks Each of the three isomorphic sequenceblocks is made by taking the 24-element sequenceABCBCDCDE and mapping A through H to differentelements and repeating it ve times yielding a total of120 trials per block The mappings of A through H forthe three isomorphic sequences are respectively28561374 54123867 and 71486253 The test starts witha block of 120 trials in random order followed by thethree isomorphic sequence blocks of 120 trials each anda nal block of 120 trials in random order The RSI was500 msec

Subjects in the Explicit group (N = 5) were shown aschematic representation of the rule and told to activelytry to use such a rule to help them go as fast as possibleSubjects in the Implicit group (N = 5) were simply toldto go as fast as possible and were given no hint that theremight be an underlying structure in the stimuli

Appendix Specication of Surface and AbstractModels

Surface Model

Recurrent State Representation The model is imple-mented in Neural Simulation Language (NSL 21 Weitzen-feld 1991) Equation 1a and 1b describe how therepresentation of sequence context in the 5 times 5 State isinuenced by external inputs from Input responses fromOut and a recurrent inputs from StateD In Equation 1athe leaky integrator s( ) corresponding to the membranepotential or internal activation of State is described InEquation 1b the output activity level of State is generatedas a sigmoid function f( ) of s(t) The term t is the time

t is the simulation time step and is the time constantAs increases with respect to t the charge and dis-charge times for the leaky integrator increase The t is5 msec For Equations 1 through 4 the time constantsare 10 msec except for Equation 2 which has ve timeconstants that are 100 600 1100 1600 and 2100 msec(see below)

si (t + t) = ( 1 - t) si (t) +

t (

j = 1

n

wijIS Input j (t) +

j = 1

n

wijSS StateD j (t) +

j = 1

n

wijOS Out j (t) ) (1a)

State(t) = f(s(t)) (1b)

The connections wIS wSS and wOS dene the projec-tions from units in Input StateD and Out to State Theseconnections are one-to-all are mixed excitatory and in-


hibitory and do not change with learning This mix ofexcitatory and inhibitory connections ensures that theState network does not become saturated by excitatoryinputs and also provides a source of diversity in codingthe conjunctions and disjunctions of input output andprevious state information Recurrent input to State origi-nates from the layer StateD StateD (Equation 2a and 2b)receives input from State and its 25 leaky integratorneurons have a distribution of time constants from 100to 2100 msec (20 to 420 simulation time steps) whereasState units have time constants of 10 msec (2 simulationtime steps) This distribution of time constants in StateD

yields a range of temporal sensitivity similar to thatprovided by using a distribution of temporal delays(KQhn amp van Hemmen 1992)

sdi (t + t) = (1 - t) sdi (t) +

t (Statei (t)) (2a)

StateD = f(sd(t)) (2b)

Associative Memory During learning for each correctresponse generated in Out the pattern of activity in Stateat the time of the response becomes linked via reinforce-ment learning in a simple associative memory to theresponding element in Out The required associativememory is implemented in a set of modiable connec-tions (wSO) between State and Out Equation 3 describeshow these connections are modied during learning Therst response in Out above a certain threshold is selectedby a ldquowinner take allrdquo (WTA) function and is evaluatedThus Equation 3 is executed only once for each re-sponse When a response is evaluated the connectionsbetween units encoding the current state in State and theunit encoding the current response in Out are strength-ened as a function of their rate of activation and learningrate R R is positive for correct responses and negativefor incorrect responses Weights are normalized to pre-serve the total synaptic output weight of each State unit

wijSO (t + 1) = wij

SO (t) + R Statei Out j (3)

The network output is thus directly inuenced by theInput and also by State via learning in the wSO synapsesas in Equation 4a and 4b where f( ) includes a WTAfunction

oi (t + t) = (1 - t) oi (t) +

t (Inputi (t) +

j = 1

n

wijSOStatej (t))

(4a)

Out = f(o(t)) (4b)

Abstract Structure Learning Model To represent andlearn abstract structure a system must (1) compare the

current sequence element with previous elements torecognize repetition and (2) maintain a representation ofthis recoded context to predict future repetitions Toprovide a record of the previous responses with whichthe current response can be compared the model ofFigure 1 is augmented with a continuously updatedshort-term memory (STM) of the ve previous responsesEach time a response is generated in Out the STM isupdated as described in Equation 5a and 5b so that STMalways contains the ve previous responses in Out Eachof the ve STM elements is thus a 5 times 5 array

for i = 5 to 2 STM(i) = STM(i - 1) (5a)

STM(1) = Out (5b)

To detect if the current response is a repetition of oneof the ve previous responses it is compared with eachof the ve previous responses encoded in STM (prior toupdate of Equation 5) The result is stored in a six-element vector called Recognition (Recog in Figure 1)Each Recognition element i for 1 i 5 is either 0 ifOut is different from STM(i) or 1 if they are the same asdescribed in Equation 6 If no match is detected in STMfor a given response in Out Recog0 is set to 1 indicatingthat a unique (u) response has occurred

Recogi = STM(i) Out (6)

After Recognition compares a response in Outputwith the contents of the STM the results of this com-parison (eg u u u n - 2 n - 4 n - 3 for ABCBAC) isprovided as input to State from Recognition as de-scribed in the updated version of Equation 1a Thus theabstract structure is encoded and represented in thesequence context

si (t + t) = (1 - t) si (t) +

t (

j = 1

n

wifIS Input j (t) +

j = 1

n

wijSS StateDj (t) +

j = 1

n

wijOS Outj (t) +

j = 1

n

wijRS Recog j (t)) (1a )

The result is that now the abstract structure of se-quences is represented in State and thus can be ex-ploited For example after training with sequence(s)with the abstract structure u u u n - 2 n - 4 n - 3when the model is exposed to a subsequence with thestructure u u u n - 2 it should predict that the nextelement will be the same as the element stored in STM4

(ie n - 4) To exploit this predictive abstract structurethat is represented in the State pattern we want toselectively take the contents of the STM that are pre-dicted to match the upcoming element and direct thisSTM content to Out To achieve this a new learning rule

Dominey et al 749

is developed Each time a repetition is recognized con-nections are strengthened between the active context(State) units and units in a four-element vector Modula-tion (described below) that modulate the contents ofthe matched STM element to the output The result isthat this State pattern becomes increasingly associatedwith and thus predicts the occurrence of the matchedelement before it occurs This learning rules is describedin Equation 7

wijSM (t + 1) = wij

SM (t) + Statei Recog j (7)

The goal of this learning is to allow State to modulatethe contents of specic predicted STM elements intoOut To permit this a ve-element vector Modulation isintroduced such that for i = 1 to 5 if Modulationi isnonzero the contents of STM(i) are modulated or di-rected to Out This leads to an updated version of Equa-tion 4a

oi (t + t) = (1 - t) oi (t) +

t (Input i (t) +

j = 1

n

wijSO Statej (t) +

k = 1

m

Modulationk STM(k)i ) (4a )

Based on the learning in Equation 7 State now directsthis modulation of the STM contents into Out via Statersquosinuence on Modulation as described in Equation 8

Modulationi = j = 1

n

wijSM Statej (t) (8)

After training on sequence ABCBAC when the modelis exposed to a new isomorphic sequence DEFEDF theabstract structure u u u n - 2 n - 4 n - 3 will berecognized After exposure to subsequence DEFE theactive State units will drive the Modulation unit corre-sponding to the n - 4 STM element STM(4) directing itscontents D to the output leading to a reduced RT forthe response to D By the same type of context codingwith which the surface model predicts elements of alearned sequence the abstract model can predict ele-ment repetitions in a learned class of isomorphic se-quences

Acknowledgments

PFD was supported by the Fyssen Foundation (Paris) and theGIS Sciences de la Cognition (Paris) TL is supported by theMinistRre de lrsquoEducation Nationale de la Recherche et de laTechnologie (Paris) We gratefully acknowledge Keith Holyoakand Richard Ivry for insightful discussions concerning analogi-cal transfer and SRT learning and Mike Stadler Tim Curran andJim Neely for their constructive comments on a previous ver-sion of this manuscript

Reprint requests should be sent to Peter F Dominey Institut

des Sciences Cognitives CNRS UPR 9075 67 Blvd Pinel 69675BRON Cedex France or via e-mail domineyisccnrsfr

REFERENCES

Brooks L R amp Vokey J R (1991) Abstract analogies and ab-stracted grammars Comments on Reber (1989) andMathews et al (1989) Journal of Experimental Psychol-ogy General 120 316ndash323

Catrambone R amp Holyoak K J (1989) Overcoming contex-tual limitations on problem-solving transfer Journal of Ex-perimental Psychology Learning Memory andCognition 15 1147ndash1156

Chomsky N (1959) On certain formal properties of gram-mars Information and Control 2 137ndash167

Cleeremans A amp McClelland J L (1991) Learning the struc-ture of event sequences Journal of Experimental Psychol-ogy General 120 235ndash253

Cohen A Ivry R I amp Keele S W (1990) Attention andstructure in sequence learning Journal of ExperimentalPsychology Learning Memory and Cognition 16 17ndash30

Curran T amp Keele S W (1993) Attentional and nonatten-tional forms of sequence learning Journal of Experimen-tal Psychology Learning Memory and Cognition 19189ndash202

Dominey P F (1995) Complex sensory-motor sequence learn-ing based on recurrent state-representation and reinforce-ment learning Biological Cybernetics 73 265ndash274

Dominey P F (1997a) Analogical transfer reduces problemcomplexity Behavioral and Brain Sciences 20 71ndash77

Dominey P F (1997b) An anatomically structured sensory-motor sequence learning system displays some general lin-guistic capacities Brain and Language 59 50ndash75

Dominey P F (1998a) Inuences of temporal organizationon sequence learning and transfer Comments on Stadler(1995) and Curran and Keele (1993) Journal of Experi-mental Psychology Learning Memory and Cognition24 234ndash248

Dominey P F (1998b) A shared system for learning serialand temporal structure of sensori-motor sequences Evi-dence from simulation and human experiments CognitiveBrain Research 6 163ndash172

Dominey P F (1998c) From double-step and colliding sac-cades to pointing in abstract space Towards a basis foranalogical transfer Behavioral and Brain Sciences 20745

Dominey P F Arbib M A amp Joseph J P (1995) A model ofcortico-striatal plasticity for learning oculomotor associa-tions and sequences Journal of Cognitive Neuroscience7 311ndash336

Dominey P F amp Boussaoud D (1997) Encoding behavioralcontext in recurrent networks of the frontostriatal systemA simulation study Cognitive Brain Research 6 53ndash65

Dominey P F amp Georgieff N (1997) Schizophrenics learnsurface but not abstract structure in a serial reaction timetask NeuroReport 8 2877ndash2882

Dominey P F amp Jeannerod M (1997) Contribution of fron-tostriatal function to sequence learning in Parkinsonrsquos dis-ease Evidence for dissociable systems NeuroReport 8iiindashix

Dominey P F Ventre-Dominey J Broussolle E amp JeannerodM (1995a) Analogical transfer in sequence learning Hu-man and neural-network models of fronto-striatal functionAnnals of the New York Academy of Sciences 769 369ndash373

Dominey P F Ventre-Dominey J Broussolle E amp JeannerodM (1995b) Representation and computation for analogical


transfer in sequence learning (ATSL) Human and neuralmodels of cortico-striatal function In J M Bower (Ed)Computational neuroscience Trends in research(pp 335ndash341) San Diego CA Academic Press

Dominey P F Ventre-Dominey J Broussolle E amp JeannerodM (1997) Analogical transfer is effective in a serial reac-tion time task in Parkinsonrsquos disease Evidence for a dissoci-able sequence learning mechanism Neuropsychologia 351ndash9

Elman J L (1990) Finding structure in time Cognitive Sci-ence 14 179ndash211

Gick M L amp Holyoak K J (1983) Schema induction andanalogical transfer Cognitive Psychology 15 1ndash38

Gomez R L (1997) Transfer and complexity in articialgrammar learning Cognitive Psychology 33 154ndash207

Gomez R L amp Schvaneveldt R W (1994) What is learnedfrom articial grammars Transfer tests of simple associa-tion Journal of Experimental Psychology Learning Mem-ory and Cognition 20 396ndash410

Grafton S T Hazeltine E amp Ivry R (1995) Functional map-ping of sequence learning in normal humans Journal ofCognitive Neuroscience 7 497ndash510

Greeneld P M (1991) Language tools and brain The ontog-eny and phylogeny of hierarchically organized sequentialbehavior Behavioral and Brain Science 14 531ndash595

Holyoak K J Junn E N amp Billman D O (1984) Develop-ment of analogical problem-solving skill Child Develop-ment 55 2042ndash2055

Holyoak K J Novick L R amp Melz E R (1994) Compo-nent processes in analogical transfer Mapping patterncompletion and adaptation In K Holyoak amp J Barnden(Eds) Advances in connectionist and neural computa-tion theory Vol 2 Analogical connections Norwood NJAblex

Holyoak K J amp Thagard P (1989) Analogical mapping byconstraint satisfaction Cognitive Science 13 295ndash355

Knowlton B J amp Squire L R (1993) The learning of catego-ries Parallel brain systems for item memory and categoryknowledge Science 262 1747ndash1749

Knowlton B J amp Squire L R (1994) The information ac-quired during articial grammar learning Journal of Eperi-mental Psychology Learning Memory and Cognition20 79ndash91

Knowlton B J amp Squire L R (1996) Articial grammarlearning depends on implicit acquisition of both abstractand exemplar-specic information Journal of Experimen-tal Psychology Learning Memory and Cognition 22169ndash181

KQhn R amp van Hemmen J L (1992) Temporal associationIn E Domany J L van Hemmen amp K Schulten (Eds) Phys-ics of neural networks (pp 213ndash280) Berlin Springer-Verlag

Lisman J E amp Idiart M A P (1995) Storage of 7 plusmn 2 short-term memories in oscillatory subcycles Science 2671512ndash1515

Mathews R C Buss R R Stanley W B Blanchard-Fields FCho J R amp Druhan B (1989) Role of implicit and ex-plicit processing in learning from examples A synergisticeffect Journal of Experimental Psychology LearningMemory and Cognition 15 1083ndash1100

Nissen M J amp Bullemer P (1987) Attentional requirementsof learning Evidence from performance measures Cogni-tive Psychology 19 1ndash32

Novick L R (1988) Analogical transfer problem similarityand expertise Journal of Experimental Psychology Learn-ing Memory and Cognition 14 510ndash520

Pascual-Leone A Grafman J Clark K Stewart M Mas-saquoi S Lou J-S amp Hallett M (1993) Procedural learn-ing in Parkinsonrsquos disease and cerebellar degenerationAnnals of Neurology 34 594ndash602

Perruchet P amp Pacteau C (1991) Implicit acquisition of ab-stract knowledge about articial grammar Some methodo-logical and conceptual issues Journal of ExperimentalPsychology General 120 112ndash116

Portnoff L A (1982) Schizophrenia and semantic aphasia Aclinical comparison International Journal of Neurosci-ence 16 189ndash197

Reddington M amp Chater N (1996) Transfer in articial gram-mar learning A reevaluation Journal of Experimental Psy-chology Learning Memory and Cognition 125 123ndash138

Shanks D R amp St John M F (1994) Characteristics of disso-ciable memory systems Behavioral and Brain Sciences17 367ndash447

Stadler M A (1992) Statistical structure and implicit seriallearning Journal of Experimental Psychology LearningMemory and Cognition 18 318ndash327

Suzuki M Kurachi M Kawasaki Y Kiba K amp YamaguchiN (1992) Left hypofrontality correlates with blunted af-fect in schizophrenia Japanese Journal of Psychiatryand Neurology 46 653ndash657

Thagard P Holyoak K J Nelson G amp Gochfeld D (1990)Analog retrieval by constraint satisfaction Articial Intelli-gence 46 259ndash310

Treue S amp Maunsell J H R (1996) Attentional modulationof visual motion processing in cortical areas MT and MSTNature 382 539ndash541

Turing A M (1936) On computable numbers with an appli-cation to the entscherdungsproblem Proceedings of theLondon Mathematical Society 42 238ndash265 Correctionibid 43 544ndash546

Weitzenfeld A (1991) NSL neural simulation language Ver-sion 21 University of Southern California Brain Simula-tion Laboratory Technical Report 91-05

Dominey et al 751

structure is dened by the serial order of the sequenceelements at the level of verbatim and aggregate struc-ture as dened by Stadler (1992) In contrast abstractstructure is dened by the relations between elementsthat repeat within a sequence In this context the twosequences ABCBAC and DEFEDF share the same abstractstructure 123213 and have completely unrelated surfacestructures (ie they are isomorphic) If we consider twosuch isomorphic sequences in terms of what is pre-dicted by the abstract structure we observe that in bothsequences the last three elements are predictable by therst three which themselves are unpredictable In anSRT protocol with a repeating sequence like ABCBACsurface structure learning should be manifest by a uni-form RT reduction for both the predictable and unpre-dictable elements Learning of the abstract structureshould be manifest by a nonuniform prole with addi-tional RT reduction for the predictable versus unpre-dictable elements that should transfer to newisomorphic sequences

A Theoretical Basis for Dissociable Processes

From a theoretical perspective there is an importantrelation between the structure of information to learn(eg surface or abstract structure) and the class of ma-chines or architectures that are capable of learning orprocessing this structure (eg Chomsky 1959 Turing1936) This would predict for example that there is asystem capable of learning surface structure that cannotlearn abstract structure Here we describe a neural net-work model based on the primate fronto-striatal system(Dominey 1995 Dominey Arbib amp Joseph 1995) andpresent simulation results demonstrating that this systemis capable of learning surface structure but is not capableof learning and transferring abstract structure whichrequires separate processing capabilities (Dominey1997b Dominey Ventre-Dominey Broussolle amp Jean-nerod 1995a 1995b)

Based on these results and our initial hypothesis wepredict that there are two independent and behaviorallydissociable mechanisms for learning surface and abstractstructure in humans As noted by Shanks and St John(1994) the learning of rules or abstract structure ischaracterized by a conscious effort to discover and ex-ploit the appropriate rules an effort that can be invokedby specic instructions to make such a conscious effort(Gick amp Holyoak 1983) This position is supported bystudies of analogical transfer in problem solving thatinvolve the extraction of an abstract structure commonto several problems with different surface structures(Gick amp Holyoak 1983 Holyoak Junn amp Billman 1984Holyoak Novick amp Melz 1994) Such studies demon-strate that this process requires the explicit intention tond the abstract structure In contrast the learning ofinstances or surface structure is oriented toward memo-rization of the instances themselves (Shanks amp St John

1994) without a conscious processing effort to searchfor common rule-based structure (eg Cohen Ivry ampKeele 1990 Curran amp Keele 1993) These studies suggestthe existence of dissociable mechanisms for processingsurface and abstract structure

Several studies of articial grammar learning (AGL)have also provided convincing evidence for the exist-ence of dissociable mechanisms for surface versus ab-stract structure representations (Gomez 1997 Gomez ampSchvaneveldt 1994 Knowlton amp Squire 1996) Knowl-ton and Squire demonstrated that both rule adherence(abstract structure) and chunk strength that is similarityof letter bigram and trigram distribution (surface struc-ture) inuence grammaticality judgments LikewiseGomez and Schvaneveldtrsquos results indicate that trainingon legal letter pairs is sufcient for classication withthe same letter set but that training with longer stringsis required to allow transfer of abstract structure to achanged letter set This suggests that pairs provide asource of surface structure whereas abstract structure isonly available in strings These results thus indicate thatin AGL there are dissociable forms of representation forsurface and abstract structure respectively

Although AGL tasks are often considered to test im-plicit learning it is important to note that during the testphase the subjects are explicitly instructed to apply a setof rules to classify the new objects and it is likely thatsome rule abstraction takes place during this explicittesting phase (Perruchet amp Pacteau 1991 Reddington ampChater 1996) Likewise Mathews et al (1989) have dem-onstrated that this grammatical knowledge can becomeat least partially explicit and that for grammars thatexploit relational properties like the ones used in ourcurrent studies learning can only occur in truly explicitconditions This relation between explicit processing andAGL has recently been further claried by Gomez(1997) who demonstrated that subjectsrsquo ability to trans-fer abstract structure learning to changed letter sets wasinvariably accompanied by explicit knowledge as re-vealed in direct tests Conversely subjects who learnedrst-order surface structure dependencies but failed todisplay transfer of the abstract structure in the changedletter set condition did not differ from naive controls onthe direct tests Finally Gomez demonstrated that for thesame testing materials presented either in a whole-stringAGL task or a letter-by-letter sequence learning tasktransfer and the associated acquisition of explicit knowl-edge occurred only in the AGL task This indicates thatespecially in sequencing tasks the acquisition of abstractstructure and its transfer to isomorphic sequences in-volves explicit processing

We thus predict for our sequence learning task thatin completely uninformed implicit conditions processesthat are capable of learning surface structure but notabstract structure will be accessible whereas in explicitconditions both will be accessible In response to thesepredictions we will report the results of three experi-

Dominey et al 735

















Dominey et al 737







Discussion











Discussion







Dominey et al 739














Discussion








Dominey et al 741












Discussion










Discussion



Dominey et al 743



GENERAL DISCUSSION

















Dominey et al 745


Conclusion


METHODS











Human Experiments

Apparatus









Dominey et al 747










Surface Model



si (t + t) = ( 1 - t) si (t) +

t (

j = 1

n

wijIS Input j (t) +

j = 1

n


j = 1

n







sdi (t + t) = (1 - t) sdi (t) +

t (Statei (t)) (2a)



wijSO (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Inputi (t) +

j = 1

n

wijSOStatej (t))

(4a)

Out = f(o(t)) (4b)




STM(1) = Out (5b)




si (t + t) = (1 - t) si (t) +

t (

j = 1

n

wifIS Input j (t) +

j = 1

n

wijSS StateDj (t) +

j = 1

n

wijOS Outj (t) +

j = 1

n




Dominey et al 749


wijSM (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Input i (t) +

j = 1

n

wijSO Statej (t) +

k = 1

m



Modulationi = j = 1

n



Acknowledgments




REFERENCES


















































Dominey et al 751

















Dominey et al 737







Discussion











Discussion







Dominey et al 739














Discussion








Dominey et al 741












Discussion










Discussion



Dominey et al 743



GENERAL DISCUSSION

















Dominey et al 745


Conclusion


METHODS











Human Experiments

Apparatus









Dominey et al 747










Surface Model



si (t + t) = ( 1 - t) si (t) +

t (

j = 1

n

wijIS Input j (t) +

j = 1

n


j = 1

n







sdi (t + t) = (1 - t) sdi (t) +

t (Statei (t)) (2a)



wijSO (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Inputi (t) +

j = 1

n

wijSOStatej (t))

(4a)

Out = f(o(t)) (4b)




STM(1) = Out (5b)




si (t + t) = (1 - t) si (t) +

t (

j = 1

n

wifIS Input j (t) +

j = 1

n

wijSS StateDj (t) +

j = 1

n

wijOS Outj (t) +

j = 1

n




Dominey et al 749


wijSM (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Input i (t) +

j = 1

n

wijSO Statej (t) +

k = 1

m



Modulationi = j = 1

n



Acknowledgments




REFERENCES


















































Dominey et al 751







Discussion











Discussion







Dominey et al 739














Discussion








Dominey et al 741












Discussion










Discussion



Dominey et al 743



GENERAL DISCUSSION

















Dominey et al 745


Conclusion


METHODS











Human Experiments

Apparatus









Dominey et al 747










Surface Model



si (t + t) = ( 1 - t) si (t) +

t (

j = 1

n

wijIS Input j (t) +

j = 1

n


j = 1

n







sdi (t + t) = (1 - t) sdi (t) +

t (Statei (t)) (2a)



wijSO (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Inputi (t) +

j = 1

n

wijSOStatej (t))

(4a)

Out = f(o(t)) (4b)




STM(1) = Out (5b)




si (t + t) = (1 - t) si (t) +

t (

j = 1

n

wifIS Input j (t) +

j = 1

n

wijSS StateDj (t) +

j = 1

n

wijOS Outj (t) +

j = 1

n




Dominey et al 749


wijSM (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Input i (t) +

j = 1

n

wijSO Statej (t) +

k = 1

m



Modulationi = j = 1

n



Acknowledgments




REFERENCES


















































Dominey et al 751






Discussion







Dominey et al 739














Discussion








Dominey et al 741












Discussion










Discussion



Dominey et al 743



GENERAL DISCUSSION

















Dominey et al 745


Conclusion


METHODS











Human Experiments

Apparatus









Dominey et al 747










Surface Model



si (t + t) = ( 1 - t) si (t) +

t (

j = 1

n

wijIS Input j (t) +

j = 1

n


j = 1

n







sdi (t + t) = (1 - t) sdi (t) +

t (Statei (t)) (2a)



wijSO (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Inputi (t) +

j = 1

n

wijSOStatej (t))

(4a)

Out = f(o(t)) (4b)




STM(1) = Out (5b)




si (t + t) = (1 - t) si (t) +

t (

j = 1

n

wifIS Input j (t) +

j = 1

n

wijSS StateDj (t) +

j = 1

n

wijOS Outj (t) +

j = 1

n




Dominey et al 749


wijSM (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Input i (t) +

j = 1

n

wijSO Statej (t) +

k = 1

m



Modulationi = j = 1

n



Acknowledgments




REFERENCES


















































Dominey et al 751














Discussion








Dominey et al 741












Discussion










Discussion



Dominey et al 743



GENERAL DISCUSSION

















Dominey et al 745


Conclusion


METHODS











Human Experiments

Apparatus









Dominey et al 747










Surface Model



si (t + t) = ( 1 - t) si (t) +

t (

j = 1

n

wijIS Input j (t) +

j = 1

n


j = 1

n







sdi (t + t) = (1 - t) sdi (t) +

t (Statei (t)) (2a)



wijSO (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Inputi (t) +

j = 1

n

wijSOStatej (t))

(4a)

Out = f(o(t)) (4b)




STM(1) = Out (5b)




si (t + t) = (1 - t) si (t) +

t (

j = 1

n

wifIS Input j (t) +

j = 1

n

wijSS StateDj (t) +

j = 1

n

wijOS Outj (t) +

j = 1

n




Dominey et al 749


wijSM (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Input i (t) +

j = 1

n

wijSO Statej (t) +

k = 1

m



Modulationi = j = 1

n



Acknowledgments




REFERENCES


















































Dominey et al 751












Discussion










Discussion



Dominey et al 743



GENERAL DISCUSSION

















Dominey et al 745


Conclusion


METHODS











Human Experiments

Apparatus









Dominey et al 747










Surface Model



si (t + t) = ( 1 - t) si (t) +

t (

j = 1

n

wijIS Input j (t) +

j = 1

n


j = 1

n







sdi (t + t) = (1 - t) sdi (t) +

t (Statei (t)) (2a)



wijSO (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Inputi (t) +

j = 1

n

wijSOStatej (t))

(4a)

Out = f(o(t)) (4b)




STM(1) = Out (5b)




si (t + t) = (1 - t) si (t) +

t (

j = 1

n

wifIS Input j (t) +

j = 1

n

wijSS StateDj (t) +

j = 1

n

wijOS Outj (t) +

j = 1

n




Dominey et al 749


wijSM (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Input i (t) +

j = 1

n

wijSO Statej (t) +

k = 1

m



Modulationi = j = 1

n



Acknowledgments




REFERENCES


















































Dominey et al 751



GENERAL DISCUSSION

















Dominey et al 745


Conclusion


METHODS











Human Experiments

Apparatus









Dominey et al 747










Surface Model



si (t + t) = ( 1 - t) si (t) +

t (

j = 1

n

wijIS Input j (t) +

j = 1

n


j = 1

n







sdi (t + t) = (1 - t) sdi (t) +

t (Statei (t)) (2a)



wijSO (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Inputi (t) +

j = 1

n

wijSOStatej (t))

(4a)

Out = f(o(t)) (4b)




STM(1) = Out (5b)




si (t + t) = (1 - t) si (t) +

t (

j = 1

n

wifIS Input j (t) +

j = 1

n

wijSS StateDj (t) +

j = 1

n

wijOS Outj (t) +

j = 1

n




Dominey et al 749


wijSM (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Input i (t) +

j = 1

n

wijSO Statej (t) +

k = 1

m



Modulationi = j = 1

n



Acknowledgments




REFERENCES


















































Dominey et al 751









Dominey et al 745


Conclusion


METHODS











Human Experiments

Apparatus









Dominey et al 747










Surface Model



si (t + t) = ( 1 - t) si (t) +

t (

j = 1

n

wijIS Input j (t) +

j = 1

n


j = 1

n







sdi (t + t) = (1 - t) sdi (t) +

t (Statei (t)) (2a)



wijSO (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Inputi (t) +

j = 1

n

wijSOStatej (t))

(4a)

Out = f(o(t)) (4b)




STM(1) = Out (5b)




si (t + t) = (1 - t) si (t) +

t (

j = 1

n

wifIS Input j (t) +

j = 1

n

wijSS StateDj (t) +

j = 1

n

wijOS Outj (t) +

j = 1

n




Dominey et al 749


wijSM (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Input i (t) +

j = 1

n

wijSO Statej (t) +

k = 1

m



Modulationi = j = 1

n



Acknowledgments




REFERENCES


















































Dominey et al 751


Conclusion


METHODS











Human Experiments

Apparatus









Dominey et al 747










Surface Model



si (t + t) = ( 1 - t) si (t) +

t (

j = 1

n

wijIS Input j (t) +

j = 1

n


j = 1

n







sdi (t + t) = (1 - t) sdi (t) +

t (Statei (t)) (2a)



wijSO (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Inputi (t) +

j = 1

n

wijSOStatej (t))

(4a)

Out = f(o(t)) (4b)




STM(1) = Out (5b)




si (t + t) = (1 - t) si (t) +

t (

j = 1

n

wifIS Input j (t) +

j = 1

n

wijSS StateDj (t) +

j = 1

n

wijOS Outj (t) +

j = 1

n




Dominey et al 749


wijSM (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Input i (t) +

j = 1

n

wijSO Statej (t) +

k = 1

m



Modulationi = j = 1

n



Acknowledgments




REFERENCES


















































Dominey et al 751



Human Experiments

Apparatus









Dominey et al 747










Surface Model



si (t + t) = ( 1 - t) si (t) +

t (

j = 1

n

wijIS Input j (t) +

j = 1

n


j = 1

n







sdi (t + t) = (1 - t) sdi (t) +

t (Statei (t)) (2a)



wijSO (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Inputi (t) +

j = 1

n

wijSOStatej (t))

(4a)

Out = f(o(t)) (4b)




STM(1) = Out (5b)




si (t + t) = (1 - t) si (t) +

t (

j = 1

n

wifIS Input j (t) +

j = 1

n

wijSS StateDj (t) +

j = 1

n

wijOS Outj (t) +

j = 1

n




Dominey et al 749


wijSM (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Input i (t) +

j = 1

n

wijSO Statej (t) +

k = 1

m



Modulationi = j = 1

n



Acknowledgments




REFERENCES


















































Dominey et al 751










Surface Model



si (t + t) = ( 1 - t) si (t) +

t (

j = 1

n

wijIS Input j (t) +

j = 1

n


j = 1

n







sdi (t + t) = (1 - t) sdi (t) +

t (Statei (t)) (2a)



wijSO (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Inputi (t) +

j = 1

n

wijSOStatej (t))

(4a)

Out = f(o(t)) (4b)




STM(1) = Out (5b)




si (t + t) = (1 - t) si (t) +

t (

j = 1

n

wifIS Input j (t) +

j = 1

n

wijSS StateDj (t) +

j = 1

n

wijOS Outj (t) +

j = 1

n




Dominey et al 749


wijSM (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Input i (t) +

j = 1

n

wijSO Statej (t) +

k = 1

m



Modulationi = j = 1

n



Acknowledgments




REFERENCES


















































Dominey et al 751



sdi (t + t) = (1 - t) sdi (t) +

t (Statei (t)) (2a)



wijSO (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Inputi (t) +

j = 1

n

wijSOStatej (t))

(4a)

Out = f(o(t)) (4b)




STM(1) = Out (5b)




si (t + t) = (1 - t) si (t) +

t (

j = 1

n

wifIS Input j (t) +

j = 1

n

wijSS StateDj (t) +

j = 1

n

wijOS Outj (t) +

j = 1

n




Dominey et al 749


wijSM (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Input i (t) +

j = 1

n

wijSO Statej (t) +

k = 1

m



Modulationi = j = 1

n



Acknowledgments




REFERENCES


















































Dominey et al 751


wijSM (t + 1) = wij



oi (t + t) = (1 - t) oi (t) +

t (Input i (t) +

j = 1

n

wijSO Statej (t) +

k = 1

m



Modulationi = j = 1

n



Acknowledgments




REFERENCES


















































Dominey et al 751































Dominey et al 751

dissociable processes for learning the surface structure and abstract structure of sensorimotor...

Documents