Beyond Modularity: A Developmental Perspective on Cognitive Science

BEHAVIORAL AND BRAIN SCIENCES (1994) 17, 693-745Printed in the United States of America

Precis of Beyond modularity: Adevelopmental perspective oncognitive science

Annette Karmiloff-SmithMedical Research Council Cognitive Development Unit and UniversityCollege London, London WC1H0BT, United KingdomElectronic mail: [email protected]

Abstract: Beyond modularity attempts a synthesis of Fodor's anticonstructivist nativism and Piaget's antinativist constructivism.Contra Fodor, I argue that: (1) the study of cognitive development is essential to cognitive science, (2) the module/central processingdichotomy is too rigid, and (3) the mind does not begin with prespecified modules; rather, development involves a gradual process of"modularization." Contra Piaget, I argue that: (1) development rarely involves stagelike domain-general change and (2) domain-specific predispositions give development a small but significant kickstart by focusing the infant's attention on proprietary inputs.Development does not stop at efficient learning. A fundamental aspect of human development ("representational redescription") isthe hypothesized process by which information that is in a cognitive system becomes progressively explicit knowledge to that system.Development thus involves two complementary processes of progressive modularization and progressive "explicitation." Empiricalfindings on the child as linguist, physicist, mathematician, psychologist, and notator are discussed in support of the theoreticalframework. Each chapter concentrates first on the initial state of the infant mind/brain and on subsequent domain-specific learning ininfancy and early childhood. It then goes on to explore data on older children's problem solving and theory building, with particularfocus on evolving cognitive flexibility. Emphasis is placed throughout on the status of representations underlying different capacitiesand on the multiple levels at which knowledge is stored and accessible. Finally, consideration is given to the need for more formaldevelopmental models, and a comparison is made between representational redescription and connectionist simulations ofdevelopment. In conclusion, I consider what is special about human cognition by speculating on the status of representationsunderlying the structure of behavior in other species.

Keywords: cognitive development; connectionism; domain-specific/domain-general; modularity; nativism/constructivism; otherspecies; phase versus stage; representational redescription

It is less illogical than it first appears to speakof instincts for inventiveness.(Marler 1991, p. 63)

1. Taking the developmentalperspective seriously

Beyond modularity: A developmental perspective on cog-nitive science (Kanniloff-Smith 1992a) not only aims toreach developmental psychologists, but also strives topersuade cognitive scientists to treat cognitive develop-ment as a serious theoretical science contributing to thediscussion ofhoiv the human mind/brain develops and isorganized internally, and not merely as a cute empiricaldatabase addressing the question of the age at whichexternal behavior can be observed. Nowadays much of theliterature focuses on what cognitive science can offer thestudy of development. In Beyond modularity, I concen-trate, on what a developmental perspective can offercognitive science and attempt to pinpoint what is specifi-cally human about human cognition.

As Piaget's conception of the sensorimotor infant isbeing severely undermined by new paradigms for study-

ing infancy, the battle between nativism and constructiv-ism once again rears its rather unconstructive head. InBeyond modularity, I do not choose between these twoepistemological stances, one arguing for predominantlybuilt-in, domain-specific knowledge, and the other for aminimum innate underpinning to subsequent domain-general learning. Rather, I suggest that nativism (whenredefined within a truly epigenetic perspective of geneticexpression rather than genetic unfolding), on the onehand, and Piaget's constructivism, on the other, arecomplementary in fundamental ways, and that the ulti-mate theory of human cognition will encompass aspects ofboth. Beyond modularity is intended to excite the readerabout the possibilities of a developmental perspectiveembracing both domain-specific predispositions and con-structivism and to demonstrate that one can attributevarious innate processes/structures to the human neonatewithout denying the crucial roles of the physical andsociocultural environments and without jeopardizing thedeep-seated conviction that we are special - creative,cognitively flexible, capable of conscious reflection, novelinvention, and occasional inordinate stupidity!

Developmental psychologists of the Piagetian school

© 1994 Cambridge University Press 0140-525X194 $5.00+ .00 693

Karmiloff-Smith: Beyond modularity

are loath to attribute domain-specific predispositions tothe human infant, yet they would not hesitate to do sowith respect to the ant, the spider, the bee, or thechimpanzee. Why would Nature have endowed everyspecies except the human with some domain-specificpredispositions? Yet, if it turns out that all species havesuch predispositions, that most can maintain a goal in theface of changing environmental conditions, and that mosthave the capacity for learning on the basis of interactionwith conspecifics and the physical environment, what isspecial about human cognition? Is it simply that thecontent of knowledge differs between species? Is it lan-guage that makes humans special? Or, compared to otherspecies, are there qualitatively different processes atwork across many domains of the human mind? Doeshuman cognitive change affect all domains of knowledgemore or less simultaneously, or does development occurin a domain-specific fashion? These are some of thequestions addressed in Beyond modularity.

I argue that domain-specific predispositions give devel-opment a small but significant kickstart by focusing theyoung infant's attention on proprietary inputs. The earlyperiod is followed by intricate interaction with environ-mental input which in turn critically affects brain develop-ment as subsequent learning takes place. But develop-ment does not stop at efficient learning. A fundamentalaspect of human development is the hypothesized pro-cess by which information that is in a cognitive systembecomes progressively explicit knowledge to that system.I call this the "representational redescription" hypothesis(henceforth RR). Support for the theoretical discussions ofChapter I is explored in Chapters 2 through 6, calling onempirical findings on the child as a linguist, a physicist, amathematician, a psychologist, and a notator. Each chap-ter concentrates first on the initial state of the infantmind/brain and on subsequent domain-specific learningin infancy and early childhood, and then goes on toexplore empirical data on older children's problem solv-ing and theory building, with particular focus on evolvingcognitive flexibility and metacognition. Throughout, Iplace particular emphasis on the status of representationsunderlying different capacities and on the multiple levelsat which knowledge is stored and accessible.

In Chapters 7 and 8, I reconsider the reconciliationbetween nativism and Piagets constructivism, and I dis-cuss the need for more formal developmental models.Here, I compare aspects of the RR framework withconnectionist simulations of development. The book endswith a final look at the RR framework and conjecturesabout the status of representations underlying the struc-ture of behavior in nonhumans, who never become re-describers of the implicit knowledge embedded in theirbehavior, no matter how complex the behavior.

If our focus is on cognitive flexibility and consciousaccess to knowledge, why not explore the data from adultpsychology? Surely adults are far more flexible cog-nitively than children, so what justifies a developmentalperspective? Not, rest assured, the fact that child data are"cute"! One need only glance at the developmental litera-ture to notice that many researchers are absorbed withthe ages at which children reach cognitive milestones.Decades of developmental research were wasted, in myview, because the focus was entirely on lowering the age

at which children could perform a task successfully, with-out concern for how they processed the information. Ionce began an article (Karmiloff-Smith 1981, p. 151) asfollows: "The enticing yet awful fact about child develop-ment is that children develop! Awful, because it hasprovoked a plethora of studies, totally unmotivated theo-retically, accepted for publication in certain types ofjournal because the results are 'significant'- significantstatistically, since it is indeed easy to obtain differentialeffects between, say, 5 and 7 year olds, but questionableas to their significance scientifically." Some researchers,however, use the study of development as a theoreticaltool for exploring the human mind/brain from a cognitivescience perspective. We are not really interested in chil-dren per se but in human cognition in general, which webelieve can be more fully understood via its develop-ment.

A developmental perspective is essential to the analysisof human cognition because understanding the predis-positions of the human mind/brain, the constraints onsubsequent learning, and how representations changeprogressively over time can provide subtle clues to repre-sentational format in the adult mind. The work of Spelke(1991), which I discuss in Chapter 3, has been particularlyinfluential in pointing to the importance of a developmen-tal perspective on cognitive science. For example, theprocesses for segmenting visual arrays into objects areoverlaid, in preschool children and adults, by other pro-cesses for recognizing object categories. But by focusingon how very young infants segment visual arrays intoobjects before they are able to categorize certain objectkinds, Spelke is able to generate new hypotheses abouthow the visual system may actually function beyondinfancy and in adults.

Another area in which the developmental perspectivecan change our view of the adult mind concerns the statusof different types of representations. Distinctions such asdeclarative/procedural, conscious/unconscious, explicit/implicit, and controlled/automatic, which are often usedto explain cognitive processing in adujts, turn out toinvolve far more than a dichotomy when explored within adevelopmental context. But in assuming a developmentalperspective we must take the notion "developmental"seriously. Paradoxically, studies on neonates and infantsare often not developmental at all. Like studies on adults,they frequently focus not on change but on real-timeprocessing within steady-state systems, It is of courseessential to determine the initial stat£ of the humanmind/brain, but the "developmental" notion goes beyondthe specification of initial predispositions. It does notsimply mean a focus on learning in children of differentages rather than the adult. When one makes theoreticaluse of development in cognitive science the specific age atwhich children can successfully perform a task is, to someextent, irrelevant.

A developmental perspective focuses on behavioral andrepresentational change over time. I bften use a laterphase in a developmental sequence to understand thestatus of representations underlying earlier behavior -particularly in the interesting cases where child and adultbehaviors are practically identical. This, notion of repre-sentational change over time is the focus throughoutBeyond modularity. It is for all these reasons that I hold

694 BEHAVIORAL AND BRAIN SCIENCES (1994) 17:4


that a developmental perspective is essential to cognitivescience's efforts to understand the human mind morefully.

2. Is the initial architecture of the infantmind/brain modular?

Fodor's 1983 book, The modularity of mind, made asignificant impact on developmental theorizing by sug-gesting how the nativist thesis and the domain-specificityof cognition are relevant to constraints on the architectureof the human mind/brain. In Beyond modularity, I crit-ically discuss Fodor's thesis at some length but, since ithas been the subject of a BBS treatment (Fodor 1985) it isunnecessary to reiterate all the details in the presentPre'cis. A brief summary suffices to recall that according toFodor the mind/brain is made up of genetically specified,independently functioning, special-purpose "modules"(or input systems). Each functionally distinct module hasits own dedicated processes and proprietary inputs. In-formation from the external environment first passesthrough a system of sensory transducers, which transformthe data into formats that each special-purpose modulecan process. Each module, in turn, outputs data in acommon format suitable for central, domain-general pro-cessing. The modules are deemed to be hard-wired (notassembled from more primitive processes), of fixed neuralarchitecture, domain specific, fast, autonomous, manda-tory, automatic, stimulus driven, giving rise to shallowoutputs; they are informationally encapsulated and insen-sitive to central cognitive goals. For Fodor, it is the co-occurrence of all the properties that defines a module.Modules, then, are the parts of the human mind that areinflexible and unintelligent. They are the stupidity in themachine - but they are just what a young organism mightneed to get initial cognition off the ground speedily andefficiently.

Fodor posits a built-in dichotomy between what iscomputed blindly by the modules and what the organism"believes." It is in "central processing" that computationsrelevant to the human belief system are processed, byderiving top-down hypotheses about what the world islike from the interlace between the outputs of modulesand what is already stored in long-term memory. Fodorconsiders central processing, in contrast to modules, to beinfluenced by what the system already knows, and there-fore to be relatively unencapsulated, slow, nonmandatory,controlled, often conscious, and influenced by globalcognitive goals. Central processing receives outputs fromeach module which are automatically translated into acommon representational format, a language of thought(Fodor 1975). Central processing, then, is general-purpose. It is devoted to the fixation of belief, the build-ing up of encyclopedic knowledge, and the planning ofintelligent action, in contrast to the special-purpose,domain-specific computations of modules.

Although I endorse the importance of some aspects ofFodor's thesis for understanding the human mind/brain, Ido not maintain the notion that modules are prespecifiedin detail, and I question the strictness of the dichotomythat Fodor draws between modules and central process-ing. I also challenge his contention that the outputs of

modules are automatically encoded into a single commonlanguage of thought. I focus on the argument that a crucialaspect of development involves the RR process of goingbeyond modularity.

3. Prespecified modules versus a processof gradual modularization

Fodor's detailed account of the encapsulation of modulesfocuses predominantly on their role in on-line processing.There is little discussion of ontogenesis. I draw a distinc-tion between the notion of prespecified modules versusthat of a process of "modularization" (which, I speculate,occurs repeatedly as the product of development). Here Idiffer from Fodor's strict nativist conception. I hypothe-size that if the human mind/brain ends up with anymodular structure, then this is the result of a process ofmodularization as development proceeds. My positiontakes account of the plasticity of early brain development(Johnson 1990; 1993; Neville 1991), suggesting that afairly limited number of innately specified, domain-specific predispositions would be sufficient to constrainthe classes of inputs that the infant mind computes. Thesepredispositions can operate at many different levels anddo not have to be limited to representational content (seeKarmiloff-Smith 1992b for more recent discussion). It canthus be hypothesized that, with time, brain circuits areprogressively selected for different domain-specific com-putations. In certain cases, relatively encapsulated mod-ules would be formed as a product of development. Inother cases, there would be more room for influence fromother computations.

Only future research using on-line brain activationstudies with neonates and young infants can distinguishbetween the two hypotheses. If Fodor's thesis of pre-specified modules is correct, such studies should showthat, from the very outset (or the moment at which theinfant shows sensitivity to particular forms of input),specific brain circuits are activated in response to domain-specific inputs. By contrast, if the modularization thesis iscorrect, activation levels should initially be relativelydistributed across the brain, and only with time (and thiscould be a short or relatively long time during infancy,depending on the domain) would specific circuits beactivated in response to domain-specific inputs. The mod-ularization thesis allows us to speculate that, althoughthere are maturationally constrained attention biases anddomain-specific predispositions that channel the infant'searly development, this endowment involves far morethan mere triggering. Rather, it interacts richly with, andis in return affected by, the environmental input.

Research with other species also demonstrates thebrain's plasticity. In studies of the rat, for example,Greenough et al. (1987) have shown that the brain's lossesand gains of synapses are a function of different types ofexperience. Thus, when placed merely for exercise in atreadmill, the rat shows an increase in blood capillaries inthe cerebellum, but a decrease in synapses (due to prun-ing of existing neural pathways, because of the lack ofstimulation other than physical exercise). However, whenthe rat is placed in a rich environment that challenges it tolearn, substantial increases in dendritic growth and syn-aptic connectivity are generated.

BEHAVIORAL AND BRAIN SCIENCES (1994) 17:4 695


Despite my reservations regarding Fodor's modularitythesis, I, together with a number of cognitive develop-mentalists, believe that Fodor's thesis has pointed towhere a domain-general view of development such asPiaget's is likely to be wrong. In Beyond modularity,however, I argue for a more dynamic view of developmentthat Fodor's modularity of mind and I challenge Fodor'sdismissal of the relevance of a developmental perspectiveon cognitive science. Moreover, I question Fodor's oftencited claim that "the limits of modularity are also likely tobe the limits of what we are going to be able to understandabout the mind" (1983, p. 126). I argue that cognitivescientists can go beyond modularity to study the morecreative aspects of human cognition. But my contention isthat such an endeavor is greatly enhanced by a develop-mental perspective on the problem.

4. Development from a domain-generalperspective

Fodor's nativist thesis is in sharp contrast with domain-general theories of learning, such as Piaget's constructiv-ist epistemology, once so popular in the developmentalliterature. According to Piagetian theory neither process-ing nor storage is domain specific. Of course, implicitly atleast, Piagetians acknowledge that there are differentsensory transducers for vision, audition, touch, and soforth. They do not accept, however, that the transducerstransform data into innately specified, domain-specificformats for modular processing. Rather, for Piagetians, alldata are processed by the same mechanisms and develop-ment involves domain-general changes in representa-tional structures.

By opposing the domain-general view to the domain-specific explanation of development, I suggest that Piagetand behaviorism have much in common. Neither thePiagetian nor the behaviorist grants the infant any innatestructures or domain-specific knowledge. Each grantsonly some domain-general, biologically specified pro-cesses: for the Piagetians, a set of sensory reflexes andthree functional processes (assimilation, accommodation,and equilibration); for the behaviorists, inherited physi-ological sensory systems and a complex set of laws ofassociation. These domain-general learning processes areheld to apply across all areas of linguistic and nonlinguisticcognition. Piaget and the behaviorists thus concur on anumber of conceptions about the initial state of the infantmind/brain. The behaviorists saw the infant as a tabularasa with no built-in knowledge (Skinner 1953). Piaget'sview of the young infant as assailed by "undifferentiatedand chaotic" inputs (Piaget 1955) is substantially the same.

Needless to say, there are fundamental differencesbetween these two schools. Piagetians view children asactive information constructors; behaviorists view themas passive information storers. Piagetians conceive ofdevelopment as involving fundamental stagelike changesin logical structure, whereas behaviorists invoke a pro-gressive accumulation of knowledge. However, in thepresent state of developmental theorizing, Piagetians andbehaviorists have much in common in their view of theneonate's "knowledge-empty" mind and their claims thatdomain-general learning explains subsequent develop-ment across all aspects of language and cognition.

5. Development from a domain-specificperspective

The domain-specific thesis projects a very different pic-ture of the young infant. Rather than being assailed byincomprehensible, chaotic data from many competingsources, the neonate is seen as having domain-specificpredispositions allowing it to process specific types ofinputs. Contrary to the Piagetian or the behavioristtheses, the domain-specific thesis gives the infant a verygood start. This does not, of course, mean that nothingchanges during infancy and beyond; the infant has muchto learn, but subsequent learning is guided by innatelyspecified, domain-specific principles, aijd these princi-ples determine how subsequent learning takes place(Gelman 1990a; Spelke 1991). :

Irrespective of whether they agree with Fodor's strictmodularity thesis, many psychologists now considerdevelopment to be domain-specific. Indeed, much de-pends on what one understands by "domain," and it isimportant not to confuse domain with "module." Fromthe point of view of the child's mind, a domain is the set ofrepresentations sustaining a specific area of knowledge:language, number, physics, and so forth, A module is aninformation-processing unit that encapsulates that knowl-edge and the computations on it. thus, considering devel-opment to be domain specific does not necessarily implyconsidering it modular. In other words; the storing andprocessing of information may be domain specific withoutbeing encapsulated, hardwired, mandatory, and so on.Throughout Beyond modularity, I argue for the domainspecificity of development rather than modularity in thestrict Fodorian sense. I retain the term "domain" to coverlanguage, physics, mathematics, and so forth. I alsodistinguish "microdomains" such as gravity within thedomain of physics and pronoun acquisition within thedomain of language. These microdomains can be thoughtof as subcomponents within particular domains.

The need for this finer distinction of what constitutes adomain stems from the fact that I put forward a phasemodel of development, rather than a stage model. In astage model, such as Piaget's, overarching changes occurmore or less contemporaneously acro$s different do-mains. One alternative view is that broad changes occurwithin a domain - for example, that a particular type ofchange occurs first with respect to language and later withrespect to physics. The model discussed in Beyond mod-ularity differs from both of these. It invokes recurrentphase changes at different times across different microdo-mains and repeatedly within each domain.

The domain specificity of cognitive systems is alsosuggested by developmental neuropsychology, that is theexistence of children in whom one or more domains arespared or impaired. For example, high functioning autis-tic individuals show a serious deficit in communicationand reasoning about mental states (theory of mind), [seeGopnik: "How We Know our Minds" BBS 16(1) 1993;Tomaselloetal.: "Cultural Learning" BBS 16(3) 1993.] therest of their cognition being relatively unimpaired (Frith1989). Individuals with Williams syndrome, by contrast,display a very uneven cognitive profile in which language,face recognition, and theory of mind seem relativelyspared, whereas number, spatial cognition, and problem


solving are severely retarded (Bellugi et al. 1988;Karmiloff-Smith et: al., submitted). Whether autism andWilliams syndrome involve domain-specific representa-tional deficits or computational deficits, or both, remainsan open question. There are also numerous cases of idiotsavants in whom only one domain (such as drawing orcalender calculation) functions at a high level, whilecapacities are extremely restricted over the rest of thecognitive system (Hermelin & O'Connor 1986). Domain-general theorists have difficulty explaining such within-domain and across-domain dissociations.

Adult brain damage also points to domain specificity. Itis remarkably difficult to find examples in the neuropsy-chological literature of an across-the-board, domain-general disorder (Marshall 1984), although a case could bemade for an overall deficit in planning in patients withprefrontal damage (Shallice 1988). In many instances,however, disorders of higher cognitive functions, as aconsequence of brain damage, are often domain-specific -that is, they affect only face recognition, number, lan-guage, or some other facility, leaving the other systemsrelatively intact.

So if adults manifest domain-specific damage, and if itcan be shown that infants come into the world with somedomain-specific predispositions, does that not mean thatthe nativists have won the debate over the developmen-talists still ensconced on the theoretical shores of LakeGeneva (Piaget's former bastion of antinativism and anti-modularity)? Not necessarily, for two reasons. First, mostnativist accounts call on detailed genetic unfolding, sim-ply triggered by environmental stimuli. An epigeneticview is very different (see the excellent discussion inOyama 1985). Second, it is important to bear in mind thatthe greater the number of the fixed domain-specificproperties of the infant mind/brain, the less creative andflexible the subsequent system would be (Chomsky 1988).Although the fixed constraints provide an initial adaptiveadvantage, there is a tradeoff between efficiency andautomaticity, on the one hand, and relative inflexibility,on the other. This leads me to a crucial point: The morecomplex the picture we ultimately build of the innatelyspecified predispositions of the infant mind, the moreimportant it becomes for us to explain the flexibility ofsubsequent cognitive development. It is toward such anend - exploring the flexibility and creativity of the humanmind beyond the initial state - that my work in languageacquisition and cognitive development has been concen-trated, in an attempt to determine both the domain-specific and the domain-general contributions to develop-ment. It is implausible that development will turn out tobe entirely domain specific or entirely domain general.And, although I will need to invoke some initial con-straints, development clearly involves a more dynamicprocess of interaction between mind/brain and environ-ment than the strict nativist stance presupposes.

6. Reconciling nativism and Piaget'sconstructivism

What theory of development could encompass the dy-namics of a rich process of interaction between mind/brain and environment? At first blush, a theory with acentral focus on epigenesis and constructivism, like Pi-


aget's, would seem the most appropriate. The notion ofconstructivism in Piaget's theory is the equivalent at thecognitive level of the notion of epigenesis at the level ofgene expression. For Piaget, both gene expression andcognitive development are emergent products of a self-organizing system that is directly affected by its interac-tion with the environment. Fodor (1983, p. 33) uses theterm "constructivism" very differently from Piaget. ForFodor, it is a form of empiricism, whereas Piaget arguedthat his constructivist genetic epistemology was an alter-native to both nativism and empiricism. This generalaspect of Piaget's theory, if more formalized, may wellturn out to be appropriate for future explorations of thenotion of progressive modularization discussed above.Much of the rest of Piaget's theory, however, has comeunder a great deal of criticism.

A growing number of cognitive developmentalists havebecome disenchanted with Piaget's account of the infantas a purely sensorimotor organism. For Piaget, the new-born has no domain-specific knowledge, merely sensoryreflexes and the three domain-general processes of assim-ilation, accommodation, and equilibration. By contrast,the infancy research I discuss in the first part of Chapters2 through 6 of Beyond modularity suggests that there isconsiderably more to the initial state of the mind/brainthan Piaget's theory posits. But the exclusive focus ofnativists like Fodor and Chomsky on biologically specifiedmodules suggests that they think there is nothing ofinterest to say about development beyond modularity.Moreover, Fodor's concentration on input systems -hehas far less to say about either output systems or centralprocessing - does not help us to explore the ways in whichchildren turn out to be active participants in the construc-tion of their own knowledge.

Although for Chomsky (1988) and Spelke (1991) anativist/modularity stance precludes constructivism, Iargue that nativism and Piaget's epigenetic constructiv-ism are not necessarily incompatible, with certain pro-visos. First, to Piaget's view one must add some innatelyspecified predispositions that would give the epigeneticprocess a head start in each domain. This does not implymerely adding a little more domain-general structurethan Piaget supposed. Rather, it means adding domain-specific biases to the initial endowment. But the secondproviso for the marriage of constructivism and nativism isthat the initial endowment involves far less detailed speci-fications than some nativists presuppose, and a moreprogressive process of modularization (as opposed toprespecified modules) where the structure of the inputplays an essential role in the structure of the resultingmodule. Fodor does not, for instance, discuss the cases inwhich the operation of one of his prespecified modulescannot be triggered by its proprietary input (e.g., audi-tory input in the case of the congenitally deaf). We knowthat in such cases the brain selectively adapts and recon-figures itself to receive other (e.g., visuomanual) nonaudi-tory inputs (Changeux 1985; Neville 1991; Poizner et al.1987). Many cases of early brain damage indicate thatthere is far more plasticity in the brain than Fodor's strictmodularity would imply. The brain is not prestructuredwith ready-made representations which are simply trig-gered by environmental stimuli; it is channeled to pro-gressively develop representations via interaction withboth the external environment and its own internal envi-



ronment. Furthermore, it is important not to equateinnateness with presence at birth or with the notion of astatic genetic blueprint for maturation. Whatever innatecomponent we invoke, it becomes part of our biologicalpotential only through interaction with the environment;it is latent until it receives input (Johnson 1988; 1993;Marler 1991; Oyama 1985; Thelen 1989) and the inputrequired is either relatively specific or simply in the formof environmental stimuli per se (Greenough et al. 1987;Johnson & Bolhuis 1991). The interaction with the inputcrucially in turn affects the development of the brain.

Nativists argue that development follows similar pathsbecause all normal children start life with the sameinnately specified structures. The role of the environmentis reduced to that of a mere trigger. [See Lightfoot: "TheChild's Trigger Experience: Degree-0 Learnability" BBS12(2) 1989; Crain: "Language Acquisition in the Absenceof Experience" BBS 14(4) 1991. ] But the fact that develop-ment proceeds in similar ways across normal childrendoes not necessarily mean that development must beinnately specified in detail, because it is also true that allchildren evolve in a species-typical environment (John-son & Morton 1991) and we are discovering that environ-ments are more structured than was originally thought(see Elman 1990; 1993). Thus, it is the interaction be-tween similar innate constraints and similar environmen-tal constraints that gives rise to common developmentalpaths.

The proposed reconciliation of nativism and construc-tivism will allow us to adhere to Piaget's epigenetic-constructivist view of the developmental process but todrop his insistence on domain generality in favor of a mor,edomain-specific approach. Furthermore, the Piagetianfocus on output systems (i.e., on the infant's and thechild's action on the environment) is an important addi-tion to the nativist's accent on input systems. But Piaget'sstrong antinativism and his arguments for across-the-board major structural stages no longer constitute a viabledevelopmental framework.

The need to invoke domain specificity is apparentthroughout Beyond modularity. For example, domain-general sensorimotor development alone cannot explainthe acquisition of language. Syntax does not derive simplyfrom exploratory problem solving with toys, as somePiagetians claim, [cf. interesting discussion by Green-field: "Language, Tools and Brain" BBS 14(4) 1991.]Lining up objects does not form the basis for word order.Trying to fit one toy inside another has nothing to do withembedded clauses. General sensorimotor activity alonecannot account for specifically linguistic constraints. If itcould, then it would be difficult to see why chimpanzees,who manifest rich sensorimotor and representational abil-ities, do not acquire anything remotely resembling thecomplex structure of human language despite very exten-sive training (Premack 1986).

Despite these criticisms of Piaget's view of early infancyand my rejection of his stage view of development, I hopethat Beyond modularity will persuade readers that impor-tant aspects of Piaget's epistemology should be salvaged,and that there is far more to cognitive development thanthe unfolding of a genetically specified program simplytriggered by environmental stimuli. If we are to under-stand the human mind, our focus must stretch wellbeyond any innate specifications and embrace the interac-

tion of both domain-specific constraints and domain-general processes.

7. The empirical data

Because of the space limitations of a Pre'cis, I refer thereader to Chapters 2 through 6 in Beyond modularity fordiscussions of the empirical data and the, literature refer-enced therein. New infancy research and the representa-tional status of infant knowledge form the detailed focusof the first part of each chapter, showing the linguistic,physical, mathematical, psychological, and notationaldomain-specific constraints on early development. Fu-ture research may lead to reinterpretations of the presentinfancy data, but I remain convinced that we will have toinvoke some domain-specific predispositions which ini-tially constrain the infant mind/brain. For each cognitivedomain, I go on to consider data suggesting that develop-ment involves much more than the domain-specific con-straints. My research strategy has always been ratherdifferent from that of developmentalists who study agiven capacity, from failure to partial success through tocomplete mastery. By contrast, I focus on an age group ineach domain where the particular capacity under study isalready proficient. I then attempt to trace subsequentrepresentational change. The most important and subtledata in Chapters 2-6 are, in my view, those pointing to alevel of representation in which knowledge is explicitlydefined (i.e., represented differently from the informa-tion embedded in special-purpose domain-specific proce-dures of the earlier phase) but not yet available to con-scious access and verbal report. Spontaneous repairs tolinguistic output, unsuccessful problem solving subse-quent to success, redundant behaviors, and so forth (dataoften ignored in developmental and adult research) are allused as vital clues to this phase of development.

At several points throughout Beyond modularity, Iallude to abnormal development. Nature, alas, oftenpresents the scientist with experiments of its own, inwhich different capacities are either spared or impaired.Such cases warrant study in their own right, but they alsohelp us gain a deeper understanding of normal develop-ment and domain specificity/modularity. Again, for spacereasons I merely allude to them here (for more recentdetailed discussion, see Karmiloff-Smith 1992c).

Development involves, then, two complementary pro-cesses of progressive modularization and progressive ex-plicitation. In the remainder of this Precis, I will concen-trate on the second of these two processes, that is, on myhypothesis that development involves' representationalredescription, a process that increases the flexibility andmanipulability of the knowledge stored in the mind, byturning information that is in the mind into progressivelymore explicit knowledge to the mind.

8. Beyond domain-specific constraints: How newknowledge gets into the mind

How does information get stored in the child's mind? Iargue that there are several different ways. One is viainnate specification as the result of evolutionary pro-cesses. Predispositions can be either specific or non-specific (Johnson & Bolhuis 1991). In both cases, environ-


mental input is of course necessary. Should an innatecomponent be specified in detail (if it ever is), then it islikely that the environment acts simply as a trigger for theorganism to select one parameter or circuit over others(Changeux 1985; Chomsky 1981; Piatelli-Palmerini 1989).By contrast, when a predisposition is specified merely as abias or as a skeletal outline, then the environment acts asmuch more than a trigger; it influences the subsequentstructure of the brain via a rich epigenetic interactionbetween the mind/brain and the physical/socioculturalenvironment (for discussions, see Johnson & Karmiloff-Smith 1992). The skeletal outline involves attention biasestoward particular inputs and a certain number of predis-positions constraining the computation of those inputs.

There are several other ways in which new informationgets stored in the child's mind. One occurs when the childfails to reach a goal and must take information from thephysical environment into account. New knowledge isalso acquired when the child has to take into accountand to represent information provided by the socio-cultural environment, often in the form of a direct lin-guistic statement. These are both external sources ofchange from environmental input, but there are alsointernal sources of change. One is illustrated by theabove-mentioned process of modularization when inputand output processing become progressively less influ-enced by other processes in the brain. This causes knowl-edge to become more encapsulated and less accessible toother systems. Another essential facet of cognitive changegoes in the opposite direction, however, with knowledgebecoming progressively more accessible.

My claim is that a specifically human way to gainknowledge is for the mind to exploit internally the infor-mation that it has already stored, by redescribing itsrepresentations or, more precisely, by iteratively re-representing in different representational formats whatits internal representations represent. This is what Ihypothesize is particular to human cognition (see detailsin section 9 below).

Finally, there is a form of knowledge change that is farmore obviously restricted to the human species: explicittheory change, which involves conscious constructionand exploration of analogies, thought experiments andreal experiments, typical of older children and adults(Carey 1985; Klahr 1992; Kuhn et al. 1988). I argue,however, that this more obvious characteristic of humancognition is possible only on the basis of the more subtleprior representational redescription, which turns implicitinformation embedded in special-purpose proceduresinto explicit knowledge but is not yet available to con-scious verbal report.

To give a more tangible feel for the theoretical discus-sion on which I am about to embark, let's consider thepathway to learning to play the piano. There is a firstperiod during which a sequence of separate notes islaboriously practiced. The beginning pianist pays con-scious attention to particular notes. There is a secondperiod during which chunks of several notes are playedtogether as blocks, until finally the whole piece can beplayed more or less automatically. In other words, thesequence gradually becomes proceduralized (see Ander-son 1980). It is something like this that I call "reachingbehavioral mastery." But the automaticity is constrainedby the fact that the learner can neither start in the middle


of the piece nor play variations on a theme (Hermelin &O'Connor 1989). The performance is generated, I hy-pothesize, by procedural representations which are sim-ply run off in their entirety. There is little flexibility.At best, in a third period, the learner is able to play thewhole piece softer, louder, slower, or faster. The pianist's"knowledge" is embedded in the procedural representa-tions sustaining the execution. But most learners do notstop there. During a fourth period, the learner caninterrupt the piece and start at, say, the third bar withouthaving to go back to the beginning and repeat the entireprocedure from the outset.

I hypothesize that this-fourth period cannot take placeon the basis of the automatized procedural representa-tions. Rather, it involves a process of representationalredescription such that the knowledge of the differentnotes and chords (rather than simply their run-off se-quence) becomes available as manipulable data. It is onlyafter a period of behavioral mastery that the pianist cangenerate variations on a theme, change sequential orderof bars, introduce insertions from other pieces, and soforth. This differentiates, for instance, jazz improvisationfrom strict adherence to sheet music. The end result isrepresentational flexibility and control, which allows forcreativity. Also important is the fact that the earlierproceduralized capacity is not lost: for certain goals,pianists can call on the automatic skill; for others, they callon the more explicit representations that allow for flex-ibility and creativity. (Of course, the playing of somepianists remains simply at the procedural level.)

This movement from implicit information embedded inan efficient problem-solving procedure, to rendering theknowledge progressively more explicit, is a theme thatrecurs throughout Beyond modularity. And this, togetherwith the process of modularization discussed earlier, isprecisely what I think development is about. Children arenot satisfied with success in learning to talk or to solveproblems; they want to understand how they do thesethings. In seeking such understanding, they become littletheorists and to do so they have to change the nature oftheir internal representations.

Development and learning, then, seem to take twocomplementary directions. On the one hand, they in-volve the gradual process of proceduralization and attimes modularization (that is, rendering behavior moreautomatic and less accessible). On the other hand, theyinvolve a process of "explicitation" and increasing acces-sibility (that is, explicitly representing information that isimplicit in the procedural representations). Both arerelevant to cognitive change, but the main focus of Be-yond modularity is the process by which the representa-tional "explicitation" which, I posit, occurs in a variety oflinguistic and cognitive domains throughout develop-ment.

9. The process of representational redescription

For a number of years I have been trying to understand howinternal representations change in the course of develop-ment, even when overt behavior may look identical. In thisattempt, I have developed the hypothesis of a reiterativeprocess of representational redescription(RR). First, I willmake some general points about the hypothesis; then I willprovide a summary.



The notion of RR attempts to account for the way inwhich children's representations become progressivelymore manipulable and flexible. Ultimately, this leads, ineach domain at different times, to the emergence ofconscious access to knowledge and children's theorybuilding. RR involves a cyclical process by which informa-tion already present in the organism's independentlyfunctioning, special-purpose representations is madeprogressively available, via redescriptive processes, toother parts of the cognitive system, first within a domainand then sometimes across domains.

The RR process is posited to occur spontaneously aspart of an internal drive toward the creation of intrado-main and interdomain relationships. Although I stress theendogenous nature of representational redescription,clearly the process may at times also be triggered byexternal influences.

The actual process of RR is domain general, but it iscrucially affected by the form and level of explicitness of therepresentations supporting particular domain-specificknowledge at a given time. When I state that RR is domain-general, I do not mean to imply that it involves a simul-taneous change across domains. Rather, I mean that,within each domain, the RR process operates in a similarway.

Let us look now at the RR hypothesis in some detail.Development, I argue, involves three recurrent phases.During the first phase the child focuses predominantly oninformation from the external environment. This initiallearning is data driven. Phase 1 culminates in consistentlysuccessful performance on whatever microdomain hasreached that level. This is what I term "behavioral mas-tery. " Behavioral mastery does not necessarily imply thatthe underlying representations are equivalent to theadult's, even though the behavioral output may be thesame. The same performance (say, correctly producing aparticular linguistic form, or managing to balance blockson a narrow support) can be generated at various ages byvery diiferent representations. Later (phase 3) behaviormay appear identical to phase 1 behavior. We thus need todraw a distinction between "behavioral change" (whichsometimes gives rise to a U-shaped developmental curve)and "representational change" because behavioral mas-tery is not tantamount to the end point of the develop-mental progression in a given microdomain.

Phase 1 is followed by an internally driven phase duringwhich the child no longer focuses on the external data.Rather, system-internal dynamics take over such thatinternal representations become the focus of change. Inphase 2, the current state of the child's representations ofknowledge in a microdomain predominate over informa-tion from the incoming data. The temporary disregard forfeatures of the external environment during phase 2 canlead to new errors and inflexibilities. This can, but doesnot necessarily, give rise to a decrease in successfulbehavior - a U-shaped developmental curve. This isdeterioration at the behavioral level, not at the represen-tational level.

Finally, during phase 3, internal representations andexternal data are reconciled, and a balance is achievedbetween the quests for internal and external control. Inthe case of language, for example, a new mapping is madebetween input and output representations in order torestore correct usage.

But what about the format of the internal representa-tions that sustain these reiterated phases? The RR frame-work argues for at least four levels at which knowledge isrepresented and re-represented. I have termed themImplicit (I), Explicit-l(El), Explicit-2 (E2), and Explicit-3(E3). The RR framework postulates different representa-tional formats at different levels. At level I, representa-tions are in the form of procedures or action patterns forresponding to stimuli in the external environment. Anumber of constraints operate on the representationaladjunctions that are formed at this level:1

Information is encoded in procedural form.The procedure-like encodings are sequentially spe-cified.New representations are independently stored.Level-I representations are bracketed, and hence nointradomain or interdomain representational links canyet be formed.

Information embedded in level-I representations istherefore not available to other operators in the cognitivesystem. Thus, if two procedures or action patterns containidentical information, this potential interi*epresentationalcommonality is not yet represented in the child's mind. Aprocedure as a whole is available as data to other opera-tors; however, its component parts are not. It takesdevelopmental time and representational redescription(see discussion of level El below) for corrtponent parts tobecome available for the marking of potential intradomainand interdomain relationships, a process which ulti-mately leads to interrepresentational flexibility and cre-ative problem-solving capacities (see discussion of levelsE2/E3). At this first level, however, the potential repre-sentational links and the information embedded in proce-dures or action patterns remain implicit. This gives rise tothe ability to compute specific inputs in preferential waysand to respond rapidly and effectively to the environ-ment. But the behavior generated from level-I represen-tations is relatively inflexible.

Level-El representations are the result of redescrip-tion, into a new format, of the procedurally encodedrepresentations at level-I. The redescriptions are abstrac-tions and, unlike level-I representations, they are notbracketed (that is, the component parts are now open topotential intradomain and interdomain representationallinks). The El representations are reduced descriptionsthat lose many of the details of the procedurally encodedinformation. As a nice example of what I have in mindhere, consider the details of the grated image delivered tothe perceptual system of a person who sees a zebra(Mandler 1992). A redescription of this into "stripedanimal" (either linguistic or image-like) has lost much ofthe perceptual precision. To Mandler's discussion, Iwould add that the redescription allows the cognitive (asopposed to the perceptual) system to understand theanalogy between an actual zebra and the road sign for a so-called "zebra crossing" (a European crosswalk with broad,regular, black and yellow stripes), although the zebra andthe road sign deliver very different inputs to the percep-tual system. A species without representational re-descriptions would not make the analogy between thezebra and the zebra crossing sign. The redescribed repre-sentation is, on the one hand, simpler and less special-purpose but, on the other, more flexible cognitively(because it is transportable to other goals and useable to



make other inferences). Unlike perceptual representa-tions, conceptual redescriptions are productive; theymake possible the invention of new terms (e.g., "zebrin,"the antibody which stains certain classes of cells in stripedpatterns).

Note that the original level-I representations remainintact in the child's mind and can continue to be called forparticular cognitive goals which require speed and auto-maticity. The redescribed representations are used forother goals where explicit knowledge is required.

As representations are redescribed into the E l format,we witness the beginnings of a flexible cognitive systemupon which the child's nascent theories can subsequentlybe built. Level-El representations go beyond the con-straints imposed at level I, where procedure-like repre-sentations are simply used in response to external stimuli.Once knowledge previously embedded in procedures isexplicitly defined, the potential relationships betweenprocedural components can then be marked and repre-sented internally. Moreover, once redescription hastaken place and explicit representations become manipu-lable, children can introduce violations to their data-driven, veridical descriptions of the world - violationswhich allow, for instance, for pretend play, false belief,and the use of counterfactuals.

It is important to stress that although E l representa-tions are available as data to the system, they are notavailable to conscious access and verbal report. Through-out the book I examine examples of the formation ofexplicit representations which are not yet accessible toconscious reflection and verbal report, but which areclearly beyond the procedural level. In general, develop-mentalists have not distinguished between implicitlystored knowledge and E l representations in whichknowledge is explicitly represented but is not yet con-sciously accessible. Rather, they have drawn a dichotomybetween an undefined notion of something implicit inbehavior (as if information were not represented in anyform) and consciously accessible knowledge that can bestated in verbal form. According to the RR framework,the human representational system is far more complexthan a mere dichotomy. It is particularly via a develop-mental perspective that one can pinpoint this multiplicityof levels of representational formats.

In the RR framework, conscious access and verbalreport are possible only at levels beyond El . At level E2,it is hypothesized, representations are available to con-scious access but not to verbal report (which is possibleonly at level E3). Although for some theorists conscious-ness is reduced to verbal reportability, in the RR frame-work E2 representations are accessible to consciousnessbut they are in a representational code similar to that ofthe E l representations of which they are redescriptions.Thus, for example, E l spatial representations are recodedinto consciously accessible E2 spatial representations.(We often draw diagrams of problems we cannot easilyverbalize.)

At level E3, knowledge is recoded into a cross-systemcode. This common format is hypothesized to be closeenough to natural language for easy translation into sta-table, communicable form. It is possible that some knowl-edge learned directly in linguistic form is immediatelystored at level E3. Children learn a lot from verbalinteraction with others, but knowledge may be stored in

linguistic code and not yet linked to similar knowledgestored in other representational formats. Linguisticknowledge (e.g., a mathematical principle governing sub-traction) often fails to constrain nonlinguistic knowledge(e.g., an algorithm used for actually doing subtraction)until both have been redescribed into a similar format, sothat interrepresentational constraints can operate (Hen-nessy 1986).

The empirical examples throughout Beyond mod-ularity, illustrate levels I, E3, and particularly the subtle-ties of level El . In the book, I do not distinguish betweenlevels E2 and E3, both of which, I believe, involveconscious access, because thus far research has not beendirectly focused on level E2 (conscious access withoutverbal report). Most, if not all, metacognitive studiesfocus on verbal report (i.e., level E3). Thus, E2 remains tobe tested empirically. Nevertheless, I do not wish toforeclose the possibility of spatial, kinesthetic, and othernonlinguistically encoded representations that are avail-able to conscious access, and it may well be that E2 andE3 redescriptional formats are both made directly on thebasis of the El format, rather than E3 being a redescrip-tion of E2. This is discussed fully in Chapter 1.

The end result of these various redescriptions is theexistence in the mind of multiple representations ofsimilar knowledge at different levels of detail and explicit-ness. This notion of multiple encoding is important; thedevelopment of the mind does not seem to be a drive foreconomy. Indeed, the human mind may turn out to be avery redundant store of knowledge and processes.

Let me stress again the concept of reiterative develop-mental phases. There is no such thing as a "phase E2child." The child's representations are in different repre-sentational formats with respect to particular microdo-mains.

Although the process of representational redescriptioncan occur on line, I suggest that it also takes place withoutongoing analysis of incoming data or production of output.Thus, change can occur outside normal input/output rela-tions, that is, simply as the product of system-internaldynamics, when there are no external pressures. Repre-sentational change within phases involves adding repre-sentations; here, negative feedback (failure, incomple-tion, inadequacy, mismatch between input and output,etc.) plays an important role, leading progressively tobehavioral mastery. But in the transition between phases,it is hypothesized that positive feedback is essential to theonset of representational redescription. In other words,according to this success-based view of cognitive change,it is representations that have reached a stable state (thechild having reached behavioral mastery) that are re-described. Representational redescription is a process of"appropriating" stable states to extract the informationthey contain, which can then be used more flexibly forother purposes. Many of the studies discussed in Beyondmodularity, and new data from Siegler and Crowley(1991), show that change often follows success, not onlyfailure. In other words, children explore domain-specificenvironments beyond their successful interaction withthem.

This is not to deny the importance of instability, failure,conflict, and competition as generators of other types ofchange (Bates & MacWhinney 1987; Piaget 1967a; Thelen1989). It is worth reiterating this point. Competition can



occur on line between different processes and can causebehavioral change, but the hypothesis I develop through-out Beyond modularity is that competition leading torepresentational change takes place after each of thepotential competitors has been consolidated (i.e., is sta-ble in its own right). In Chapter 3, for example, it is shownhow counterexamples are not taken into account (do nothave the status of a counterexample) until the child'stheory about a particular microdomain has been consoli-dated. Similar examples are to be found in the history ofscience and in children's strategies of scientific experi-mentation (Klahr & Dunbar 1988; Kuhn et al. 1988; Kuhn& Phelps 1982; Schauble 1990), as well as across thevarious domains of knowledge discussed throughout Be-yond modularity.

10. Are there domain-general processes at work?

Invoking domain-specific constraints on developmentdoes not deny the existence of some domain-generalmechanisms. The infancy tasks explored in each chaptermake it very clear that infants can call on complex inferen-tial processes across different domains. Moreover, younginfants go well beyond sensorimotor encodings and makeuse of domain-general processes such as representationalredescription to recode sensorimotor input into accessi-ble formats (see also Mandler 1992). Domain-generalprocesses sustaining inference and representational re-description operate throughout development, but invok-ing general processes that are the same across differentdomains is not equivalent to invoking domain-generalstages of change. It is the latter that Beyond modularityrejects.

Yet there might turn out to be some across-the-boarddomain-general changes also, perhaps linked to majormaturation of particular regions of the brain (e.g. prefron-tal cortex). One such change suggested by an abundanceof empirical data seems to occur around 18 months of age.This holds for several domains, particularly with respectto holding two representations simultaneously in mindand representing hypothetical events in general (Meltzoff1990; Perner 1991), rather than theory-of-mind computa-tions in particular (Leslie 1992). Eighteen months is alsothe age Piaget singled out for a change in representationalstructure which allowed for the onset of pretend play,language, and mental imagery. The precise way in whichPiaget accounted for such a change in terms of the closureof a purely sensorimotor period is likely to be wrong, butthe conviction that something fundamental occurs around18 months may turn out to be well-founded.

The other age at which an across-the-board, domain-general change may occur is somewhere around three anda half to four years. This age does not correspond to a stagechange in Piagetian theory, but it seems to be whenfundamental changes occur in various domains. More-over, this is also roughly the age at which the human childdiffers radically from the chimpanzee. As Premack (1991,p. 164) put it, "a good rule of thumb has proved to be: if thechild of three and a half years cannot do it, neither can thechimpanzee."

If it turns out that across-the-board, domain-generalchanges do occur, we may be able to use them as adiagnostic for fundamental neural changes in the brain,

and vice versa. This of course remains ail open question,but the flourishing new field of developmental cognitiveneuroscience may soon provide some relevant answers.Even if some across-the-board changes were to hold,however, it is important to recall that their effects wouldbe manifest somewhat differently across domains, sincethey would interact with domain-specific constraints.Development will not turn out to be either domain spe-cific or domain general. It is clearly the intricate interac-tion of both - more domain-general than is presupposedby most nativist/modularity views of development, butmore domain-specific than Piagetian theory envisages.

So, does Piagetian theory retain any role in develop-mental theorizing? To me, the answer is affirmative.Theories of cognitive development (and recent connec-tionist modeling of cognitive development [McClelland& Jenkins 1990; Parisi 1990; Plunkett & Sinha 1992],which I discuss in Chapter 8) continue to draw inspirationfrom Piaget's epistemology - his quest to understandemergent properties and his general stance with regard toepigenesis and the importance of the child's action on theenvironment. It is the details of his psychological descrip-tion of across-the-board stagelike changes in logico-mathematical structure that are no longer viable. I be-lieve that it is possible to retain the essence of Piagetiantheory while doing away with stage and structure. Theproblem with Piaget's theory (and indeed, with the RRframework too), however, is that it is underspecified incomparison with, say, theories expressed as computermodels. I now turn briefly to this issue.

11. Modeling development

One of the aims of Beyond modularity is to persuadecognitive scientists of the value of a developmental per-spective for understanding the workings of the humanmind. Yet, at the heart of much of the work in cognitivescience is the use of computer models to test psychologi-cal theories. It is therefore essential to devote some spaceto a discussion of how the RR framewofk might be rele-vant to attempts to express developmental theories in theform of computer simulations.

What type of framework is RR? Throughout Beyondmodularity, I describe RR in verbal terms. It is, as Klahr(1992) has put it, at the "soft-core" end of the modeling ofcognitive development, the "hard-cor^' end being theimplementation of theories as computer programs.Klahr's contrast captures an important distinction be-tween a focus on general principles of development and afocus on the specification of precise mechanisms. Klahrargues that the very process of simulating development inthe form of computer programs leads to insights about themechanisms underlying developmental change, whereasverbal descriptions generally underspecify the mecha-nisms. I agree, but soft-core and hard-core approachesshould not be considered mutually exclusive.

In my view, soft-core approaches often lead to a broaderintuitive understanding of general principles of change,whereas both the information-processing use of the flowchart and the symbolic approach to computer simulationrun the risk of reifying into one or more boxes or single-named operators what is in fact the product of a highlyinteractive system. Nonetheless, at the hard-core end of



modeling there have been a number of interesting at-tempts to express developmental theories in variousinformation-processing terms - for example, in the formof scripts (Nelson 1986; Schank& Abelson 1977), develop-mental contingency models (Morton 1986), and self-modifying production systems (Klahr et al. 1987). InChapter 8, however, I take as my main example somerecent connectionist simulations, since they seem to beclosest to the spirit of epigenesis and constructivism (forfuller discussions, see Bates & Elman 1993; Clark &Karmiloff-Smith 1993; Elman et al., in press; Karmiloff-Smith 1992b; 1992c; Karmiloff-Smith & Clark 1993; Mc-Clelland & Jenkins 1990; Parisi 1990; Plunkett & Sinha1992). Connectionist simulations also address the prob-lems I raise in Beyond modularity with respect to stagetheories, in that they show that by incremental learningone can obtain stagelike shifts in overt behavior withoutthe need for qualitatively different structures and mecha-nisms (McClelland & Jenkins 1990).

Although the connectionist framework has come undersevere criticism (Pinker & Mehler 1988), a growing num-ber of cognitive developmentalists see within this frame-work a considerable theoretical potential for explicatingthe more general tenets of Piaget's epistemology (e.g.,Bates & Elman 1993; Bechtel & Abrahamsen 1991; El-man etal., in press; Karmiloff-Smith 1992b; 1992c; Clark& Karmiloff-Smith 1993; Karmiloff-Smith & Clark 1993;McClelland & Jenkins 1990; Plunkett & Sinha 1992).Moreover, a number of features of the RR framework,developed quite independently in the 1970s and early1980s, map interestingly onto features of recent connec-tionist simulations.

Chapter 8 of Beyond modularity describes the mainfeatures of connectionist models, but since a BBS treat-ment has dealt extensively with such models (Smolensky1988), I will not repeat the description in this Precis.Instead, I will go on to explore directly the extent towhich connectionist simulations can and cannot capturewhat I deem to be crucial to a model of developmentalchange. To the extent that they can, connectionism wouldoffer the RR framework a powerful set of hard-core toolsby applying the mathematical theory of complex dynami-cal systems to cognitive development (van Geehrt 1991).And to the extent that connectionist models fail to modeldevelopment adequately, the RR framework suggestssome crucial modifications.

Many of the details of phase 1 learning, which leads tobehavioral mastery and level-I representations, turn outto be captured particularly well in a connectionist model.However, the very aspect of development on which Be-yond modularity focuses - the process of representationalredescription - is precisely what seems to be missingfrom connectionist simulations of development.

12. Connectionism: The starting state, the role ofthe input, and the process of representationalredescription

Let us now look at some of the specific issues discussedthroughout Beyond modularity and how they can beinformed by, as well as inform, the connectionist frame-work.

12.1. The starting state

Most connectionist researchers adopt a non-nativist viewas their research strategy. This makes it possible to ex-plore the extent to which developmental phenomena canbe simulated from a tabula rasa starting state - that is,from random weights and random activation levels, withno domain-specific knowledge. This has led some tointerpret the results of connectionist modeling as strongevidence for the antinativist position, but there is nothingabout the connectionist framework that precludes theintroduction of initial biased weights and connections(i.e., the equivalents of innately specified predispositionsas a result of evolution) rather than random weights andconnections. Also, specific architectures, learning algo-rithms, learning rates, and so on which are part of thestarting state, clearly affect how an input set is learned.

Various ways of simulating developmental change havebeen proposed. One is to start a network with a smallnumber of hidden units and, as "development" proceeds,to recruit more and more units or an extra hidden layer tocompress the data even further (Shultz 1991). This israther like the neo-Piagetians' notion that processingcapacity increases with age (Case 1985; Halford 1982).Other researchers (Bechtel & Abrahamsen 1991) havesuggested the equivalent of "maturational" change, suchthat the network would start by using one learning algo-rithm (e.g., contrastive Hebbian learning)and, with mat-uration, come to use a different learning algorithm (e.g.,backpropagation). Incremental learning has also beenused, such that the network first sees only part of theinput at a time, rather than the whole input set in one go(Elman 1990; Plunkett & Marchman 1991). These are alldomain-general solutions to developmental change, butwe are beginning to witness an increasing tendency onthe part of connectionists to explore the ways in whichdomain-specific constraints might also shape learning.This is in my view, likely to be a future focus for connec-tionist models of development.

It might seem at present that connectionist modelsdeny, either implicitly or explicitly, the need for domain-specific learning. In favor of domain generality, connec-tionists stress that their models use the same learningalgorithms for different categories of input presented todifferent networks. But, in effect, architectures are fine-tuned to specific types of input. For example, a recurrentarchitecture is used for sequential input (see Elman 1990)whereas an associative network is used for concept learn-ing (see Plunkett & Sinha 1992). To my knowledge, littlework has been done on networks which progressivelydevelop their own architecture as a function of the inputthey happen to process. Moreover, no single network hasbeen presented with an array of inputs from differentdomains (e.g., language, spatial tasks, tasks involvingphysical principles). Networks designed to simulate lan-guage acquisition (e.g., Elman 1990; 1993) see only lin-guistic strings. A similar network could be used forphysics input, but the very same network could not beused without totally upsetting the language learning thathas already taken place unless it also continues to betrained on the original set. In other words, the fact thateach network is dedicated to a specific type of input, in aspecific learning task, with a specific architecture andlearning algorithm, turns out to be equivalent to domain



specificity in the human. Infants seem to process propri-etary, domain-specific inputs separately, and so do net-works. We will probably end up requiring multiple net-works with different architectures and different learningalgorithms.

A final point with respect to the starting state: networksare not "modules" in the sense of the distinction I drewbetween modules and a process of modularization. Infact, networks mimic the process of modularization be-cause, with few or no built-in representational biases, it isonly as learning proceeds that they become increasinglylike special-purpose modules.

12.2. The role of the Input

Although connectionist models have potential for devel-opmental theorizing, they have several shortcomings.One concerns the input presented to networks. First,decisions about input representation are entirely externalto the network and often are not motivated theoretically.Second, with some exceptions, connectionists have, untilnow, not really modeled development; they have mod-eled tasks. This becomes particularly apparent if we lookat the example of the balance scale that is so popular in allkinds of computer modeling, including connectionist(Langley et al. 1987; McClelland & Jenkins 1990; Newell1990; Shultz 1991; Siegler & Robinson 1982). The modelshave focused on children's performance on the balance-scale task, not on how children learn about general physi-cal phenomena in real life (see also Shultz 1991 fordiscussion). Many children come to a balance scale exper-iment with no experience of balance scales, but this doesnot mean that they bring no relevant knowledge to thetask. They may focus on weight in tasks using the tradi-tional balance scale because weights are what the experi-menter more obviously manipulates. But in other blockbalancing tasks not presented in the form of a balancescale, many young children ignore weight and focus solelyon length. Children come to such tasks having alreadylearned something about how rulers fall from tables, howseesaws work, and so forth. But a seesaw is not a balancescale. It does not have a neat line of equidistant pegs onwhich children of absolutely equal weight can be placedone on top of another! Development is not simply task-specific learning. It involves deriving knowledge frommany sources and using that knowledge in a goal orientedway. Thus, in my view, far richer input vectors and thesimulation of goal oriented behaviors are needed if we areto model the ways in which real children learn in realenvironments.

12.3. Behavioral mastery

Chapter 3, on the child as a physicist, and Chapter 6, onthe child as a notator, give particularly clear examples ofhow a lengthy period of behavioral mastery precedesrepresentational change. Indeed, throughout Beyondmodularity, I argue that behavioral mastery is a prerequi-site for representational change. An analysis of learning ina connectionist network, however, already reveals in thehidden units the existence of some representation ofsubsequent change before it is observable in the output.This suggests a way in which connectionist modelingmight change the RR framework in that full behavioral

mastery may not be a prerequisite to change; that is,representational change may start to occur prior to overtbehavioral mastery.

12.4. Implicit to explicit representational change

It has often been difficult to convey, particularly to devel-opmental psychologists, precisely what I meant by "level-I implicit representations." [See also Shanks & St. John:"Characteristics of Dissociable Human Learning Sys-tems" BBS 17(3) 1994.] Researchers have often used theterm "implicit" to explain away efficient behavior thatappears "too early" for the tenets of a particular theory,but no definition of implicit has been offered. The connec-tionist framework may help to give a more precise defini-tion. Indeed, some recent connectionist simulations oflanguage learning (Elman 1990; 1993),' for instance, areparticularly illustrative of the status df implicit level-Irepresentations. Elman's model is discussed fully in Be-yond modularity. It demonstrates how grammatical func-tion (noun/verb, transitive/intransitive verb, singular/plural, etc.) can be progressively inferred from statisticalregularities of the input set and can be represented in thehidden units as learning proceeds. The full details of thelearning process need not concern us,here, rather, weshould focus on the status of the representations that thenetwork progressively builds. First, Elman shows that, aswith most connectionist networks using nonlinear func-tions, a lengthy initial period is essential to learning. Atfirst, the network's predictions are random. However,with time the network learns to predidt, not necessarilythe actual next word, but the correct category of word(noun vs. verb; if noun, animate vs. inanimate, edible vs.nonedible, etc.), as well as the correct subcategorizationframe for the next verb (transitive or intransitive), and thecorrect number marking on both noun and verb (singularor plural). This cannot be done by mere associationbetween adjacent surface elements. For example,whereas in the case of the simple strings, a network couldlearn always to predict that strings without an "s" (pluralverb) follow strings with an "s" (plural noun), it cannot doso for embedded relative clause strings. Here, a pluralverb may follow a singular noun (e.g., "the boys that chasethe girl see the dog"). In such cases, the network mustmake structure-dependent predictions. Thus, the net-work moves progressively from processing mere surfaceregularities to representing something more abstract, butwithout this being built in as a prespecified linguisticconstraint.

This seemingly impressive grammatical knowledge isonly implicit in the system's internal representations.Note, however, that this does not mean that the gram-matical knowledge is not represented. As in the case ofearly learning in the child, I would argue that it isrepresented in level-I format. But it is we, as externaltheorists, who use level-E formats to label the trajectoriesthrough weight space as nouns, verbs, subjects, objects,intransitives, transitives, plurals, singulars, and so on.The network itself never goes beyond the formation of theequivalent of stable (but unlabeled) level-I representa-tions. In other words, it does not spontaneously go be-yond its efficient behavioral mastery. It does not rede-scribe the representations that are stored in its activationtrajectories. Unlike the child, the network does not spon-


taneously "appropriate" the knowledge it representsabout different linguistic categories. It cannot directly usethe higher-level, more abstract knowledge for any otherpurpose than the one it was designed for, nor can itengage directly in internetwork knowledge transfer be-cause its representations are input/task specific. Thenotion of, say, nounhood, always remains implicit in thenetwork's system dynamics. The child's initial learning islike this, too. But, as several examples throughout Beyondmodularity show, children go on to redescribe sponta-neously their linguistic (and other) knowledge. This per-vasive process of representational redescription gives riseto the manipulability and flexibility of the human repre-sentational system.

Now, it is not difficult to build a network, inspired byRR, that would redescribe stable states in weight spacesuch that the implicit information represented in trajecto-ries could be used as knowledge by the same or othernetworks. However, this would suggest a change in thearchitecture of the network, involving perhaps the cre-ation of special nodes not implicated in other aspects ofthe on-line processing. Furthermore, the RR frameworksuggests that what is abstracted during the redescriptiveprocess involves a loss of detail and a gain in accessibility.Thus, one would not want the entire trajectories of thenetwork to be redescribed - only the product of the mostimportant ones. (This would be equivalent to, say, label-ing the phase-state portraits of the principal-componentanalysis.) The RR framework postulates that redescribedknowledge capturing abstract notions such as "verb" and"noun" must be in a format different from that of theoriginal level-I representations. In other words, re-descriptions would have to be in a representational formatusable across networks which had previously processeddifferent representations at the input level: hence theneed for representational redescription into different(level E) formats. Simple copies of level-I representationswould not be usable/transportable from one network toanother because they would be too dependent on thespecific features of their inputs.

In Chapter 2, I discuss a particularly relevant exampleof what progressive RR might look like in the human case.When three- to six-year-olds are asked to repeat the lastword the experimenter had said before a story was inter-rupted, some of the youngest subjects (three years old)could not do the task at all, despite lengthy modeling andhelp from the experimenter. Yet their fluent language andtheir lack of segmentation errors suggest that they dorepresent formal word boundaries for the majority ofwords they use and understand, but that they are not yetready to go beyond that behavioral mastery. There wereother children (four to five years old) who could not do thetask im mediately but who, with one-off modeling for a fewopen-class words, were able immediately to extend thenotion of "word" to all open-class and closed-class catego-ries. Their level-I representations were ready for level-Elredescription triggered by the experimenter from out-side, but slightly older children (five to six years) who hadnever had a grammar lesson had spontaneously under-gone the redescriptive process on their own. Theseshowed immediate success, even on the practice story.Finally, six- to seven-year-olds' representations showedsigns of having undergone further redescription into theE3 format; these children were able to access their knowl-


edge consciously and to provide verbal explanations ofwhat counts as a word and why. This process of multipleredescription of knowledge that becomes increasinglyaccessible to different parts of the system is an essentialcomponent of human development and one that connec-tionist modelers need to take into account.

It seems plausible that connectionist models can lendprecision to an account of what I have called phase 1learning - the phase that results in behavioral mastery(i.e., the period of rich interaction with the environmentduring which level-I representations are built and consol-idated). However, there is much more to developmentthan this. I have intimated at various points that connec-tionist simulations stop short of accounting for certainessential components of human development. Indeed,whereas behavioral mastery is the endpoint of learning inconnectionist models, in the RR framework it is thestarting point for new flexibility - that is, for generatingredescriptions of implicitly defined level-I representa-tions. Until now, connectionist models of developmenthave had little to say about how to move from implicitrepresentations to explicit ones, an essential processcalled for by RR. How could a network appropriate itsown stable states? Clark (1989), Dennett (1993), andMcClelland (1991) have argued that all that would have tobe added to a connectionist network is another networkthat uses the equivalent of public language, implying thatthe only difference between implicit and explicit knowl-edge is that the latter is linguistically encoded. I have,however, provided numerous examples of children'sknowledge that is explicitly represented, but with chil-dren unable to articulate it linguistically. The RR frame-work offers a far more complex view of multiple levels ofrepresentational redescription, of which language is butone manifestation, finally, the fact that most connectionistmodels blend structure and content makes it difficult forthe network to exploit knowledge components. Yet, inseveral chapters I show that children extract knowledgecomponents from the procedural representations inwhich they are embedded, re-represent them, and usethem in increasingly manipulable ways.

How representational redescription might be modeledin a connectionist network remains an open question. Canit be done simply by adding layers to the architecture of asingle network, or by creating, say, a hierarchy of inter-connected networks? Should a node, external to the on-line processing, be fed gradually with information fromthe developing internal representations when hiddenunits reach a certain threshold of stability? How caninternet relations be introduced while keeping in mindthe constraints suggested by RR regarding common trans-portable representational format? Or will we have to optfor hybrid models containing both parallel distributedprocessing and more classical sequential manipulation ofdiscrete symbols (see discussions in Clark & Karmiloff-Smith 1993; Karmiloff-Smith 1987; 1992b; 1992c;Karmiloff-Smith & Clark 1993; Schneider 1987)? As con-nectionist networks become more complex, I think theissue of whether something is truly "hybrid" will loserelevance. Future developmental modeling must, in myview, simulate both the benefits of rapid processing viaimplicit representations and the benefits gained by fur-ther representational redescription - a process that Isuggest makes possible human creativity (for a BBS treat-



ment of creativity, see Boden: "Precis of: The CreativeMind" BBS 17(3) 1994).

13. Concluding remarks

I started Beyond modularity distinguishing between therepresentations that sustain complex behavior and thethings that a given species can do with that complexity.My argument throughout has been that, far more per-vasively even than that of its near cousin the chimpanzee,the human mind exploits its representational complexityby re-representing its implicit knowledge into manylevels of explicit form. The knowledge thereby becomesapplicable beyond the special-purpose goals for which it isnormally used and representational links across differentdomains can be forged.

This is rarely if ever true of other species. The plover(discussed in Chapter 5), for example, displays a complexset of behaviors to keep competitors at bay - behaviorsthat, in human terms, would be called deceit. But thesebehaviors (keeping competitors away from their hatchingeggs) are not available for other, even closely related,purposes (keeping competitors away from food). Whatabout the chimpanzee, with whom we share close to 100%of our genetic makeup? Do chimpanzees, like children,play with knowledge, just as they play with physicalobjects and conspecifics? According to discussions I havehad with Premack, there are no obvious indicators ofrepresentational redescription in the behavior of thechimpanzee. There are numerous examples of how thechimpanzee goes beyond a specified task; for example,when the task is to assemble the pieces of a puzzle of achimp face, a chimpanzee might, after succeeding, addextra pieces as decoration to form a hat or a necklace(Premack 1975). But Premack could find no example thatrevealed that the chimpanzee spontaneously analyzes thecomponents of its successful behavior in the way a childdoes. It is, of course, not immediately obvious how wewould recognize representational redescription in thechimpanzee if it did exist. The higher levels of redescrip-tion (into, say, linguistic format) are obviously ruled out.We know, however, that in many instances children de-velop explicit representations (El) which lie between theimplicit representations and the verbally reportable data.In the child, level El representational redescription isfrequently manifest after overt behavioral mastery. Thechimpanzee, by contrast, seems to be content to repeat itssuccesses continuously; it does not go beyond behavioralmastery. Yet, throughout Beyond modularity, examplesare explored of how human children spontaneously seekto understand their own cognition, and of how this leadsto the sort of representational manipulability that eventu-ally allows them to become folk linguists, physicists,mathematicians, psychologists, and notators.

My conjecture is that either the process of representa-tional redescription is not available to other species or, if itis (perhaps to the chimpanzee), the higher-level codesinto which representations are translated during re-description are very impoverished. It is conceivable that"language-trained" chimpanzees will show signs of repre-sentational redescription, but this would be due, not tothe existence of a languagelike code per se, but to thepossibility of redescription into any other more explicit

code (for fuller discussion, see Karmiloff-Smith 1983).RR is basically a hypothesis about the specifically hu-

man capacity to enrich itself from within, by exploitingknowledge already stored rather than by simply exploit-ing the human and physical environment. Intradomainand interdomain representational relations are the hall-mark of a flexible and creative cognitive system. Thepervasiveness of representational redescription is whatmakes human cognition specifically human. This is, ofcourse, a challenge to ethologists and One I look forwardto pursuing in the future. What indices should we beseeking in other species? What machinery would we haveto add to the plover, the ant, the spider, the bee, or thechimpanzee to make the process of representational re-description possible?

In the final pages of Beyond modularity, I present acaricature drawing of the difference between humans andother species.. In the top half of the caricature is drawn ahuman and an animal in reciprocal interaction with theexternal environment. In the bottom half, the drawingshows just the human figure with an arrow going aroundthe head from one side to the other. This (rather silly)caricature is intended to illustrate that level-I representa-tions exist as cognitive tools, allowing an organism (humanor nonhuman) to act on the environment and to beaffected by it in return. The second part of the figure is notmeant to suggest that, in the human, knowledge goes inone ear and out the other! Rather, it is a reminder that, inthe human, internal representations become objects ofcognitive manipulation such that the mind extends wellbeyond its environment and is capable of creativity. Letme go so far as to say that the RR process is, in Marler's(1991) terms, one of the human instincts for inventive-

ness.In Beyond modularity, and even in this short Precis, I

hope to have convinced the reader that the flourishingnew domain of cognitive science needs to go beyond thetraditional nativist-empiricist dichotomy that permeatesmuch of the field, in favor of an epistemology that em-braces both innate predispositions and constructivism.Cognitive science also has much to gain by going beyondmodularity and taking developmental change seriously.

A Precis necessarily makes conceptual leaps, misses outon the richness of the empirical data, as well as onnumerous references to relevant literature. It also, alas,leaves no room for the humor. Nevertheless, it has given arelatively complete idea of the theoretical issues raised inthe book. I began this Precis with a quotation from Marlerthat I find particularly conducive to my thinking, and thatI used as a colophon for one of the book chapters, but Ibegan the actual book with a quotation from Fodor, andended with the following one: "Deep down, I'm inclinedto doubt that there is such a thing as cognitive develop-ment in the sense that developmental cognitive psycholo-gists have in mind" (Fodor 1985, p. 35).

If this Precis has encouraged you tb read the book infull, I hope that by the time you reach the end, deep downyou will disagree with Fodor's statement and, with me,you will conclude that development goes far beyond thetriggered unfolding of a genetic program, that wheremodularity occurs it is the result of a gradual process ofmodularization, and that representational redescriptionallows the human mind to go beyond modularity.


Commentary IKarmiloff-Smith: Beyond modularity

A C K N O W L E D G M E N TMany people have read and commented on Beyond modularityand they are thanked in the preface to the book. Here I wouldlike to express my gratitude to my colleague, Geoff Hall, forhaving read and provided useful comments on this Pre'cis.

Open Peer Commentary

Commentary submitted by the qualified professional readership of thisjournal will be considered for publication in a later issue as ContinuingCommentary on this article. Integrative overviews and syntheses areespecially encouraged. All page references are to Karmiloff-Smith'sBeyond modularity unless otherwise indicated.

The real problem with constructivism

Paul Bloom and Karen WynnDepartment of Psychology, University of Arizona, Tucson, AZ [email protected] and [email protected]

Karmiloff-Smith is right: Fodor has it in for cognitive develop-ment. Her book focuses on Fodor's claim that input systems arespecial encapsulated modules whose functions, representa-tions, and procedures are innately specified. Although somedevelopmental change ensues, it is constrained by domain-specific properties (e.g., parameter setting in the case of syntax).If this is true, and if Fodor is justified in his belief that we areunlikely to learn anything interesting about nonmodular parts ofthe mind, there is little hope for a general theory of cognitivedevelopment. We can study the acquisition and maturation ofdistinct domains, but there will not be anything interesting tosay about cognitive development per se.

In contrast, Karmiloff-Smith argues that humans have thepotential for "representational redescription" (RR), in whichinformation originally implicit within a module can becomeincreasingly more explicit and accessible to other systems in thecourse of development. If so, then there does exist a develop-mental process that spans different domains, and one centralaspect of Piaget's constructivist epistemology can be retained.This provocative claim will prompt considerable debate, but wewant to focus on a more fundamental attack on constructivism,also by Fodor, that has nothing to do with modules.

The core of the attack is that we have no theory of howgenuinely new concepts or conceptual systems can emergethrough a developmental process. Fodor (1981) argues as fol-lows: concepts are either primitive themselves or are "built out"of primitive concepts. Current theories of categorization positthat primitives (or "properties," "features," or "microfeatures")are innate, whereas complex concepts are composed either ofprimitives that are singly necessary and jointly sufficient forcategory membership (definitions), or of primitives of whichonly a proper subset is required for category membership(prototypes) (e.g., Armstrong et al. 1983; McClelland &Rumelhart 1986; Rosch & Mervis 1975). A common assumptionin cognitive science is that there exists a relatively small set ofprimitives (e.g., those linked to perceptual systems) that com-bine to form concepts corresponding to most of the words withinnatural language.

Fodor points out, however, that we have had no success in thisprogram, despite over 2,000 years of trying. Consider, forinstance, the claim that the concept associated with the word

"dog" is a prototype, encoded through a set of features (4-legs,barks, has-a-tail, etc.); something is a dog if and only if it hassome subset of these features. Unfortunately for this theory,adults know that something can be a dog even if it has none of thefeatures associated with typical dogs (it is silent, has three legs,no tail, etc.), and we know something might not be a dog even ifit has all the features (it might be a robot that looks like a dog).Young children possess a similar understanding, and such con-siderations undermine decompositional theories of natural kindand artifact terms (see Keil 1989). This has led Fodor to theradical conclusion that the concept associated with "dog" is justdog, where dog is primitive, and therefore innate.

The same argument applies to conceptual domains in general,and it is here that the traditional constructivist program runsinto serious problems. Consider the development of logicalcompetence. Adults can reason using higher-order logic, soeither (1) the representations underlying this capacity areunlearned, or (2) they are constructed out of simpler unlearnedrepresentations. Piaget chose the second option, but was fa-mously vague as to exactly how this construction took place,leading many to reject the constructivist proposal in this domain(e.g., Macnamara 1976). Or consider syntactic development; itis sometimes suggested that children do not initially possesslinguistic categories, but rather have some "simpler" conceptualcategories like "object name" which are somehow transformedinto syntactic parts of speech. Again, however, there are fewplausible theories of how such an abstraction mechanism couldwork, which suggests that syntactic categories are present ini-tially (e.g., Bloom 1994a; Pinker 1984).

Much of this is plausible. On evolutionary grounds, onewould expect specialized innate mechanisms to underlie com-munication, categorizing, and reasoning. After all, as Karmiloff-Smith points out, we have little hesitation in attributing similarstructures to spiders and chimps. And this might lead us to aconception of development where the only source of change isKarmiloff-Smith's redescription process, in which knowledgealready in the child's mind becomes accessible to higher-orderprocesses. In this view, mental mechanisms capable of generat-ing new concepts or structures simply do not exist.

This, however, is surely too minimalist a conception of cogni-tive development. People readily learn names that correspondto concepts that could not have been preset by natural selection:"modem," "virus," "escrow," and so forth. As with "dog,' thesedo not appear to decompose into simpler concepts, but theycannot be primitive, and so their mere existence suggests theremust be some way out of Fodor's paradox. There is also evidencefrom both the history of science and the development of individ-ual children that conceptual structures do undergo radicalchange, complete with intermediate stages. Carey (1985; 1988)argues this for the domain of biological thought; Wynn (1992a;1992b) makes a similar claim for the domain of numericalcognition.

In Karmiloff-Smith's own discussion of number, she arguesthat the RR model can explain children's difficulties with earlycounting while preserving Gelman & Gallistels (1978) "count-ing principles" theory of innate numerical knowledge. Heraccount is interesting and might be plausible if it were true thatthe foundations of numerical knowledge correspond to thoseposited by Gelman & Gallistel. [See also Davis & P(5russe:"Numerical Competence in Animals." BBS 11(4) 1988.] There isevidence, however, that our initial representations of numbermay have a qualitatively different structure from those producedby the counting system (Wynn 1990; 1992a; 1992b). One pro-posal is the "accumulator mechanism" (initially proposed byMeek & Church 1983, as the basis of numerical competence inrats). On this model, representations of number are viewed asaccumulations within a container. For each entity counted, oneincrement is added to the container. All increments are of thesame fixed size; the final fullness of the container represents the


Commentary/Karmi\oS-Smith: Beyond modularity

number of items counted. Thus, numerical relationships arereproduced by the magnitudes that represent the numbers. Forexample, the number 4 is twice as large as the number 2; therepresentation for 4 is twice as large (the accumulator is twice asfull) as the representation for 2.

In the linguistic counting system, however, the representa-tions for the numbers gain their meaning by virtue of theirordinal position within a list; the word "four" is not twice as largeas the word "two," but occurs twice as far along in the orderedlist of number words. Thus, the task for children in learning tocount is not to render more explicit their initial numericalrepresentations, but to somehow relate them to those embodiedby the counting system of their language. Essentially, they mustmake an analogy: larger numbers are represented by symbolsoccurring later in a list, smaller numbers by symbols occurringearlier in the list. This kind of restructuring cannot be accountedfor by Karmilloff-Smith's theory; it requires a more powerfulconstructivist mechanism, of the sort that Fodor argues cannotexist.

There do exist proposals for how new representational andconceptual structures could develop. One is that conceptualchange in children can be viewed as akin to theory change inadult scientists (e.g., Carey 1985; 1988). Here, concepts do notdecompose into features, but are understood in terms of theirroles within larger conceptual structures, sometimes termed"naive theories." A different theory is that of "analogical map-ping" (e.g., Gentner 1983), in which a person exploits theabstract structure of one domain to learn about another, newdomain. This process might be what allows children to map thecounting system onto their initial numerical representations.Similarly, older children might come to appreciate the "genera-tive" nature of number, and thus numerical infinity, throughlearning the recursive structure of the counting system andsomehow using this to radically restructure their (initially non-generative) understanding of number (Bloom 1994b).

Karmiloff-Smith is herself sympathetic to parallel distributedprocessing (PDP) models as a way to resuscitate some of Piaget'sepistemology. However, although they might conceivably un-derlie the RR process, connectionist models show little prospectof explaining constructivist change. As she notes, they areexplicitly built to include the representations unique to thedomain they are modeling. Networks simulating reading, forinstance, might have preexisting nodes corresponding to En-glish letters, whereas those simulating acquisition of words fornatural kinds might have features like barks and 4-legs. This isplausible for domains in which representations might be in-nately present, and connectionist models do offer a usefulimplementation of prototype theories of conceptual representa-tion (e.g., McClelland & Rumelhart 1986), but they tell usnothing about how new representations can emerge develop-mentally (see also, Bloom, in press).

Karmiloff-Smith concludes her book by arguing for the central-ity of cognitive development in the study of the mind. She makesan excellent case for this, but the focus of her book is on theredescription of knowledge already present in the child's mind.This commentary presents a different problem for the study ofdevelopment - that of explaining how new concepts or concep-tual systems can emerge through a developmental process.

Representational redescription: A questionof sequence

Margaret A. BodenSchool of Cognitive and Computing Sciences, University of Sussex,Brighton BN1 9QH, England, [email protected]

Abstract: Karmiloff-Smith describes behavioral changes, attributed torepresentational redescription (RR) as occurring in a fixed order. For

example, the shape and size of parts of a drawing can be altered beforethe shape of the whole, and before any parts can be deleted, inserted, orreoriented. A theory of RR should explain this sequencing (and mightilluminate differential difficulties within adult creativity).

Karmiloff-Smith presents a powerful argument for the occur-rence of representational redescription (RR) in many differentdomains. Her discussion of possible underlying mechanisms isunavoidably no more than suggestive. I want to mention just oneintriguing question that will need to be taken into account indiscussing such mechanisms.

Her evidence shows that there is a regular sequence ofbehavioral changes. For example, the shape and size of the partsof "funny houses" or "funny men" can be altered before theshape of the whole can be changed, and before any parts can bedeleted. Furthermore, insertion of new elements comes beforechanges in the position or orientation of parts or wholes, whichprecede cross-category mixing. Why these changes, in thisorder?

Intuitively, one might say that it is easier to vary a parameterin a procedure than to vary the structure of the procedure itself,so that it is not surprising that size/shape changes of parts occurbefore deletion or insertion. This implies that whatever the RR-generating processes are, they act more easily on the parame-ters of the newly articulated parts of the skill than on theoccurrence or deletion of those parts themselves. What sort ofmechanism could make this distinction? And what sort of devel-oping mechanism would pick out the parameters before theparts? ;

Similarly, what sort of RR mechanism would be able todistinguish parts sufficiently for them to be dropped or (later)inserted, but not (yet) changed in position or orientation? Once adoor can be omitted, or an extra door inserted, why can it not beplaced in the middle of the wall? Perhaps this is because a door is(still) represented as having its lower edge coincident with thelower edge of the'whole house, whereas later it comes to berepresented as a shape of a certain sort which need not beconstrained in this way. But why? Perhaps the door is normallyadded only after the house frame has been drawn, and (once thedoor has been represented as a potentially separable part ofthe procedure as a whole) the "cue" for the addition of the door isthe existence of the line depicting the bottom of the house.What sort of mechanism would be able to (or would have to)represent a door in these ways? i

Skeletonization (Mozer & Smolensky 1989) and recurrentnetworks (Elman 1991), for instance, suggest (schematically)how a procedure initially executable only a$ a whole might cometo be represented in an articulated part-whole fashion, but theyoffer no clue as to the reason for the ordering changes observedby Karmiloff-Smith.

Karmiloff-Smith's work is of interest not only for developmen-tal psychology, but for the study of adult learning and creativityas well (Boden 1990, Ch. 4). Much as RR enables children todraw imaginatively, creating pictures of things they (literally)could not have drawn before, so the exploration and transforma-tion of conceptual spaces in adults requires those spaces to beinternally mapped, in various ways. Possibly, the mechanismsinvolved in RR in children are also involved in creative adultthinking. If so, we might find similar orders (sequences) ofdifficulty when adults try to think creatively. Or perhaps weshould expect this only when the domain is a novel one.(Karmiloff-Smith reports studies showing that adults learning toread, or to play the piano, show sequential changes similar tothose observed in children.) Once the skill of creatively alteringa lower-level skill is itself fluent, much of the psychologicalinterest will lie in the details of the repres0ntations and heuris-tics specific to that domain, rather than in general principles ofRR.


Commentary/Karmiloff-Smith: Beyond modularity

A Fodorian guide to Switzerland: Jungand Piaget combined?P6ter Bodor° and Csaba PI6hb

•Institute of Sociology and ^Department of General Psychology, LorindEOtvds University. Budapest, Hungary, [email protected]

This book is certainly a serious attempt at a new synthesis indevelopmental studies, a field that has undergone drasticchanges during recent decades. In its time, the now traditionalconception of Piaget (.1923; 1926) contrasted both with acceptededucational wisdom and with behaviorist views on developmentin positing a sophisticated mental organization in children aswell as qualitative changes in development. More recent devel-opmental theory has moved further down the time scale: Now itis infants rather than preschoolers who are smarter than weexpected.

Karmiloff-Smith's book is a synthesis in two senses. Sheaccepts how smart infants are at the outset and she is willing toattribute to them (some) innate organization, following themodular approach of Fodor (1983), but she at the same timeaccepts the idea of development. Development, however,though constructive, remains domain-specific and is charac-terized by a strong "meta" component. A combination of Piagetand Fodor, as well as a synthesis of early cleverness and meta-cognitive development is postulated, the latter being responsi-ble for human flexibility and creativity (pp. 25-26).

The book is certainly a good move toward giving cognitivescience a more sensible developmental framework (this was bythe way, the aim of Piaget [e.g., 1967b; 1970] himself; he merelylacked the term "cognitive science" for his endeavors), as well asmaking the study of development more theoretical. The realissue concerns how this relates to Fodor (1983) and his anti-developmental theory of modularity. Do we indeed have a newtheory here, with one of the features of modules, innateness,acknowledged, but the module's encapsulation attributed todevelopment (p. 5) rather than to the starting point? Let usexamine some critical points in Karmiloff-Smith's presentation.

A certain lack of criticism: Junglan allusions. The pathKarmiloff-Smith offers us around the field of cognition anddevelopmental psychology lies just beyond two notable Swissmodules (regarded by some as mountains). The path is avowedlyin the shadow of Jean Piaget; it is, horribile dictu, in the shadowof C. C. Jung as well.

There is an interesting parallel between contemporary nativ-ist developmental theory and the development of psychoanaly-tic thought: for Sigmund Freud, the message concerned how"sexual" and complex children already are at, say, 8 months ofage (a remarkably early age in his time), but the later "objectrelations" theory as well as the Budapest School of psycho-analysis started to raise the issue of how complex the mentalorganization of perineonates was. (For a direct comparisonbetween the modern Freudian and cognitivist views of develop-ment see Cergely 1992.)

There are even closer parallels between the neonativist infantresearch program and the depth psychology of C. G. Jung. Boththeories stipulate a stock of inborn human knowledge. For Jung,this is the "collective unconscious," for the neonativist re-searchers it is covered by the not less attractive term "architec-ture" (p. 2). The vehicle which carries this information in bothapproaches seems to be the same, namely, the brain. SinceJung's position in this respect is probably less widely known, wewill cite one of its versions: "archetypes are systems of prepared-ness. . . . They are inherited with the brain structure of whichthey represent the psychic aspects" (Jung 1928, p. 112). Fur-thermore, both Jung and the neonativists divide their ultimateexplanatory concepts into parts. What is for Jung articulated intoarchetypes (Mother, Father, Travelling), for the neonativistsconsists of domains (Physics, Mathematics, Language, Psychol-ogy) or, in procedural terms, input systems.

There might certainly be substantial differences between a

neonativist and a Jungian approach. At least two theoretical-methodological criteria could be used to differentiate them. Thefirst would be conceptual clarity. The second, a more criticalhandling of data by cognitive science. Unfortunately, Karmiloff-Smith is not far from Jung in these respects. With regard toconceptual clarity, some of the key notions are not entirely clearin the book. Take the concept of implicit knowledge. From thestart (pp. 15-16) on, this sometimes seems to be interpreted in aliteral, prepositional way; at other times as any kind of disposi-tion. The term "behavioral mastery," used for / level, is maskingthe issue of whether it is a "knowing how" type of knowledge or apropositionally organized "knowing that." Similar problemscharacterize the treatment of data. Jung was certainly unwillingto consider data from mythology that would place doubt on hisproposed archetypes. KarmilofF-Smith, in a somewhat similarmanner, is often unwilling to consider data that would questionthe neonativist claims. This is most characteristic of the fewanimal data presented. For example, data on early speech sounddiscrimination in infants are not contrasted with studies onrodent sound discrimination. When she considers higher animalachievements, as in studies on numbers (pp. 112-14), she isunwilling to consider how deeply these studies call into questionthe domain-specificity of some supposed innate "knowledgesystems" in humans.

As a consequence, not only its content but also the type ofargument used by Karmiloff-Smith supports the impression thatthe psychology advocated by Beyond modularity is in partsimilar to Jung's psychology.

The Issue of phases and Piaget. Piaget is sometimes misrepre-sented and sometimes misguided at important intersections.His views on initial endowment are simplified to make him looklike an antiquated and complete empiricist (p. 7), whereas in factfrom his early work on Piaget (1936) acknowledged some sophis-ticated innate endowments that constrained development (e.g.,Euclidian geometry); and he certainly considered the wholeprocess of development (universal course and stages) as drivenby an internal program.

Most prominently missing is Piaget's (1976) theory of cogni-zance. This is a crucial omission because it is related to a criticalinternal problem of the book: the transition between implicitand explicit knowledge. The original contribution of Karmiloff-Smith is the representational redescription (RR) theory. Themain features of explicit representations are: they "are reduceddescriptions that lose many of the details of the procedurallyencoded information" (p. 21) and they "can be manipulated"(p. 21); sometimes they "are not yet accessible to consciousreflection" (p. 22, El) and sometimes accessible to conscious-ness (nonverbally, E2, or verbally, E3).

The RR theory has difficulties in providing motivation for thechanges between phases and in giving a precise structuralaccount of them. Karmiloff-Smith, for the most part, over-emphasizes the internally driven nature of the transition asopposed to conflict-driven changes (e.g., p. 25, pp. 163-64).Somehow, the picture Karmiloff-Smith gives of phase transitionslacks function (except that of increasing human creativity); itseems to be a kind of Cartesian self-explanatory species-specificrealization of reflectivity in human consciousness (pp. 192-93).

A conceptually and empirically more sophisticated account(in other words, a more plausible one) of the same developmen-tal problems can be found in some late Piagetian writings(mostly in Piaget 1976); unfortunately, these are not appreciatedby Karmiloff-Smith. According to Piaget, the grasping of con-sciousness functionally "is always triggered by the fact thatautomatic regulations . . . are no longer sufficient" (Piaget1976, p. 333); structurally, "cognizance (or the act of becomingconscious) of an action scheme transforms it into a concept"(p. 332), specifically "cognizance proceeds from the periphery tothe center" (p. 334). By taking Piaget's theory of the "grasping ofconsciousness " seriously, RR would be more at ease in renderingthe issue of transition between phases more functional. It would


Commentary/K.armi\off-Smith: Beyond modularity

also be helpful to clarify the structure of E2 and E3. Wheneverthe book provides examples, it comes out that those phasesinvolve a sign system. In the framework of domain-specificity,one would like to have a clear image of whether in one of thesehigher phases one domain (e.g., language) penetrates all theothers. Or is there a general semiotic function a la Piaget (1946)hidden in the domain-specific theory?

Areas missing. Central areas are totally ignored in the "empir-ical survey" part of the book. One of them is the "child as anovelist," that is, storyteller. Recently, people like Bruner(Bruner & Haste 1987; Bruner & Luciarello 1989) and Nelson(1989) have suggested that storytelling monologues are tools formaking the world of the child coherent and creating the socialself. It would be interesting to see how RR deals with domainsthat seem to be driven from top down (creating an I level from Elevel representations), and how it would deal with social con-structivism, instead of downplaying it, as suggested on pages121-122.

A missing related issue is the child as a social theoretician.Think of early achievements of social classification in the forma-tion of relational concepts (father of, mother of, etc.). This ismore than an aspect of the child's theory of mind since it impliespower and attachment relations. Do these later concepts havetheir own domain for RR or are they merely an aspect of thechild as a psychologist?

Finally, logic and many later achievements are also missing.Neither this word nor that of "inference" figures anywhere in thebook. This is rather curious because logic certainly played acentral role for Piaget (1967b). If one wishes to show that he wasnot right about the central place of logic in development oneshould show this, or at least discuss it.

Owing to the repeated attempts to demonstrate that the childis born equipped with certain predispositions that form thestarting points for development, Karmiloff-Smith tends to ig-nore those domains that definitely manifest themselves ratherlate. In a truly constructivist approach, however, this should notbe a real obstacle. There might be domains that take shape onlyafter substantial development in others. Nothing precludeseither that some domains have late critical periods for manifesta-tion or that some may emerge from the interaction of others(provided one takes seriously the idea that modularization is aprocess where encapsulation is not present initially).

What's getting redescribed?

Robert L. CampbellDepartment of Psychology, Clemson University, Clemson, SC [email protected]

Shunning the empiricist game of "getting developmental differ-ences," Karmiloff-Smith (1981) has been unwavering in herfocus on the processes at work in development. Indeed, devel-opmentalists will find something to think about on nearly everypage of Beyond modularity.

The meter is running, however, so I must concentrate on thecore epistemological question raised by this long-awaited book.Karmiloff-Smith posits a process of representational redescrip-tion (RR). It begins with implicit or I-level representations,which are embedded in procedures that must be run off as awhole and cannot be compared with one another by the knowingsystem. The first phase of redescription converts level-I repre-sentations to the least advanced sort of explicit representationsat level El. These have been recharacterized in a "compressed"format that allows them to be compared with other representa-tions and their parts to be manipulated, but they are notaccessible to consciousness. At level E2, El representationshave been recast further to enable them to become accessible toconsciousness, but not verbally stateable. At level E3, they are

converted into a "cross-system code" (p. 23) which is close, if notidentical, to natural language.

Karmiloff-Smith uses RR to explain some major changes inthe course of development. Whether it can explain them de-pends on what is getting redescribed. Karmiloff-Smith's en-dorsement of the claims made by Spelke (1990) and Gelman(1990b), that object perception and knowledge of number re-quire the innate presence of domain-specific "principles," andher continual references to "procedurally encoded" knowledgeand "linguistic encoding" indicate an adherence to the conven-tional view that human knowledge is encoded.

An encoding is supposed to be an object, or structure ofobjects, in the mind that represents objects, or structures ofobjects, in the environment by virtue of cprrespondence withthem. Encoding pose no problems as long as they are under-stood as stand-ins for something already known. In the realm of"external representations" or notations, Morse code is an exam-ple: ". . .", which we already know, stands ,in for "s," which wealready know. We can define X as an encoding only in terms ofsome other representation Y that it stands in for: X representswhat Y represents. We might in turn define Y as an encoding interms of some other representation Z that it stands in for, butthis progression has to terminate - in a representation that is notitself an encoding.

An encoding-based conception of representation asserts theexistence of foundational encodings, that is encodings that donot stand in for other representations. A foundational encodingwould have to stand in for things in the environment that arealready known, yet it is supposed to be the means by whichthose things are known. If Z is a foundational encoding, it cannotbe given a coherent definition: Z represents whatever Z repre-sents will not qualify (Bickhard 1980a; 1993; Bickhard & Richie1983).

The incoherence of "encodingism" is not a new or exoticproblem; it has gotten many partial recognitions in the litera-ture, from Hamad's (1990) concerns about Symbol grounding toMaturana and Varela's (1980) concession that encoded represen-tations can be defined only from an observer's standpoint. Howdo the problems of encodingism visit themselves on RR?

If an I-level representation is an encoding, then because thereis no more basic kind of representation, it must be a foundationalencoding, and that is impossible. In any case, encodings do nothave epistemically implicit properties that could be made ex-plicit at a higher level; the representational content of anencoding is what it corresponds with, no more or no less, andthat is entirely explicit. In consequence, a process of recoding Iinto El or El into E2, and so forth could n6t introduce any newknowledge (Bickhard & Campbell 1989).

If a level-I representation is not an encoding, what kind ofrepresentation is it? Connectionist networks (whose develop-mental potentials Karmiloff-Smith explores in some depth) areno help here. They do not solve the "mystery of originalmeaning" (Kosslyn & Hatfield 1984). As a rule, either entirenetworks or specified portions of them are taken to represent bycorrespondence (Bickhard 1993).

There is an alternative conception of interactive representa-tion. Let us begin with a system that is capable of acting in anenvironment - one that can take in inputs, make transitions todifferent internal states, and produce outputs. If such a systemcan end up in different internal states depending on the sort ofenvironment it is in, then it can be said to represent thosedifferent environments. No knowledge is presupposed as towhat the environment is: if the system ends up in state A, it is ina type A environment, and if it ends up in state B, it is in a type Benvironment. Period. Insofar as the system can modify itself,however, it can learn what to do in different kinds of environ-ments. For example, if in a type A environment, carry outprocedure P72 to reach internal goal G27; if in a type Benvironment, do P234 to reach G33, and so on. Such expecta-tions can be falsified if, for instance, the system reaches state A



and tries P72 but does not reach G27 (Bickhard 1980a; 1980b;1993).

Treating level-I (and many level-E) representations as inter-active would honor Karmiloff-Smith's procedural intuitions, butit would require profound changes elsewhere, including a totaldeparture from the legacy of Chomsky (1980) and other lin-guists. If representation is basically interactive then languagecannot be a logically arbitrary system of rules for recoding anddecoding encoded mental contents. For starters, there willoften be no encoded contents to recode. Besides, if it were alogically arbitrary system, not only could language not belearned, it could not have merged through evolution. Interactiv-ism proposes reconceiving language as a social action system inwhich utterances function as operators to change shared under-standings (Bickhard 1980a; Brickhard & Campbell 1992; Camp-bell & Bickhard 1992a).

Interactivism has more to say about the developmental phe-nomena Karmiloff-Smith has uncovered: about the derivativeencoding (purely internal encoding of interactive representa-tions) that may generate some level-El representations (Bick-hard & Richie 1983); the creation of new learning heuristics andnew goals for learning that is also involved at level El (Campbell& Bickhard 1992b); and the different set of processes andconstraints that seems to be responsible for ascension to theconsciously accessible levels E2 and E3 (Campbell & Bickhard1986). In fact, the across-the-board, maturational shift at age 4 towhich Karmiloff-Smith refers was predicted by interactivism(which construes it as a shift to second-level knowing) quite awhile back (Bickhard 1978; 1992; Campbell 1992).

Karmiloff-Smith's (1979a; 1986; 1992) work takes an importantstep forward in our understanding of developmental processesand constraints. Her account of the transitions from I to E l to E2to E3 ought to supplant the customary sloppy discourse about"implicit" and "explicit" knowledge. The RR model is moreprecise than Piaget's (1977) presentation of reflective abstractionand it cuts loose from his lingering entanglements with equili-bration. However, Karmiloff-Smith still needs to replace Pi-aget's cognitive structures (and other people's concepts, princi-ples, or features) with an adequate conception of representa-tion.

Representational redescription and cognitivearchitectures

Antonella CarassaQ and Maurizio Tirassa"•Dipartimento di Psicologia Generate, University di Padova, 35122 Padua,Italy; •'Istituto di Psicologia delta Facoltd Medica, University di Milano,20122 Milan, Italy; '-"Centro di Scienza Cognitiva, University di Torino,10123 Turin, Italy, •[email protected] and &[email protected]

Abstract: Karmiloff-Smith's representational redescription poses somearchitectural problems. We discuss the implicit/explicit dichotomy andthe relation between natural language and the language of thought,arguing that the model addresses how knowledge is used rather thanhow it is represented in the system.

Karmiloff-Smith's representational redescription (RR) modelaccounts for changes across learning and development in termsof knowledge translation into increasingly higher-level formats.A basic feature of the model is the distinction between implicit,procedural, and explicit declarative knowledge (respectively,formats I and E1/E3, in terms of RR). Learning of proceduresmight well be the first to take place when an individual starts toexplore a new domain, but this implies nothing about how therelevant knowledge is actually represented. Procedural repre-sentations should be distinguished from (possibly declarative)representations of procedures (Rumelhart & Norman 1985). Theformer are actually encapsulated, similar to what RR posits; on

the other hand, it is hard to see how they might be decomposedinto basic steps for redescription if not through an "external"observation of one's own actions (which would open a sort ofregression, because it requires procedurally representing one'sown procedures, and so on). On the contrary, representations ofprocedures, if expressed declaratively, can be more easily disas-sembled into basic steps. Thus, decomposition requires notranslation into different codes. Another advantage of using asingle declarative code is that it eliminates the necessity forevery learning to start from a procedural phase.

A further problem with declarative/explicit formats is theirrelation to natural language. Whatever their nature, mentalrepresentations are not coded in a literally linguistic format.Language of thought theories (e.g., Fodor 1975) simply assertthat thought shares a number of properties with natural lan-guages, such as constituency, compositionality, and so on (Fodor& Pylyshyn 1988). The distinction between natural languagesand the language of thought (LOT) is mandatory for manyobvious reasons. For example, it is difficult to imagine theItalians, French, and so forth, each having their own LOT; andthere is no reason to postulate that thought requires anythinglike "inner speech" beyond the simple use of representations.But then, what does it mean to say that formats E2 and E3 differin that the latter is "close enough to natural language for easytranslation into statable, communicable form" (p. 23)? If it justmeans that E3 is mentalese, what differentiates the threeexplicit codes, conscious access aside? Is it the case that thereare two (or three) different kinds of mentalese? This is not atheoretical impossibility, but needs careful consideration. Asingle LOT poses a great many problems; what will it be likewith a number of LOTs? Another critical point concerns theformation of interdomain links which, according to Karmiloff-Smith, depends on the use of a pseudolinguistic (E3) format. Wesee no necessity to postulate this, both because such links onlyrequire a LOT with constituency, and because there is no reasonto exclude the existence of cross-domain links in higher animals.

Regarding the latter point, we would like to argue against theidea that the chimpanzee must stop at level I (i.e., that "it doesnot go beyond behavioral mastery," p. 192). This seems toostrong a statement, at least because at level I "procedural-likerepresentations are simply used in response to external stimuli,"(p. 22) and such procedures are "independently stored," (p. 20);that is, no generalization can be made from them, nor is anycross-domain link possible. There is a great deal of discussion ofthese topics; our position is that it is premature to exclude thepossibility of higher-level representations in animals (for discus-sion see Prato Previde et al. 1992).

All of our considerations bring us to the underlying topic - thehypothesized architecture of the system. The RR model positsthree or four different kinds of codes, and a correspondingnumber of translating devices. Thus, there must be a great dealof machinery (which must be built in, since the system cannot beexpected to learn new codes); furthermore, a huge controlstructure is needed that is in charge of deciding what to re-describe, and when. The general architecture of the system isnot described in detail, so it is difficult to assess its plausibility.An obvious problem concerns efficiency; although we agree onnot being obsessed with parsimony, the system seems quitefarraginous.

Our suggestion is that the content of representations mattersmore than their format in determining their implicitness orexplicitness, which might be viewed as just the presence or theabsence of "brackets" around the content itself. The interpreterof representations must be very complicated anyway, so whybelieve that the representations it operates on are complicatedas well?

Analogous considerations hold for conscious accessibility:why should it depend on the use of a particular code, rather thanon properties of the consciousness itself? When we first learnprocedures in a new domain, we are deeply (and painfully)


CommentaryVKarmiloff-Smith: Beyond modularity

conscious of the difficulties of the task, of our need for betterknowledge, and so on. Thus, the unavailability to consciousaccess only concerns the content of the representation, not therepresentation itself. In the end, the content of any representa-tion, once reduced to its basic components, is no longer accessi-ble to consciousness. So, one might think that the only differ-ence between inaccessible and accessible representations (I/Eland E2/E3, respectively) consists of the possibility of con-sciously accessing their contents rather than the formats them-selves.

To conclude, the RR model seems to deal with how knowledgecontents are used, rather than how they are actually representedin the system: in Newell's terms (1982), it resides at the knowl-edge level, rather than on the symbol level. When one goesdown to the latter, some problems start to emerge. This does notmean that the RR model would be discarded. The psychologicalevidence it brings is impressive, and the psychological plau-sibility of the model is not impaired by our considerations; butmuch work will be needed to answer questions about thearchitecture of an RR system.

Redescribing redescription

Terry DartnallSchool of Computing and Information Technology, Griffith University,Brisbane, Queensland 4111, Australia, [email protected]

Karmiloff-Smith argues that development involves the comple-mentary processes of modularization and explication. Her repre-sentational redescription (RR) hypothesis focuses on the latter:we are endogenously driven to redescribe our implicit pro-cedural knowledge as explicit declarative knowledge, and tocontinue to redescribe it in increasingly abstract terms. We beginwith knowledge that is in us, that underlies our skills and abilities,and that drives our behaviour, and we redescribe it as knowledgethat is available to us - as structures that we can access,manipulate, and map onto structures from other domains.

This gives us a rich picture of native endowment, knowledge,creativity, and mental life that explicitly or implicitly includes,(1) an epistemological third route: knowledge can be innate orlearned or it can be a re-representation of what we already know,(2) the possibility that procedural knowledge can become declar-ative knowledge, (3) an account of what it is to have a mental life:a system has a mental life if it can present its implicit knowledgeto itself as explicit structures (thoughts), (4) an account ofcreativity: a system is creative if it can articulate its domain-specific skills as accessible structures that it can manipulate andmap onto structures from other domains (see also, Karmiloff-Smith 1993 and BBS multiple book review of Boden's TheCreative Mind, BBS 17(3) 1994), and (5) a challenge: to modelthe mechanism that makes all this possible.

There are deep philosophical implications. Karmiloff-Smith'semphasis on theory construction reinstates the rationalist claimthat knowledge can develop from within; it counters, in a wayreminiscent of both Kant and Popper, the dreary empiricismaccording to which science and experience start with observa-tion and slowly and cautiously proceed towards theory. But therationalist story is subtly reworked. Out goes innate content. Incomes an innate propensity to construct theories and to re-represent to ourselves what we already know. Kant set out toreconcile rationalism and empiricism. Karmiloff-Smith sets outto reconcile nativism and constructivism. Both say that knowl-edge requires (rationalist) theory and (empiricist) sensory input."Concepts without percepts are empty, percepts without con-cepts are blind," as Kant (Kant 1785, B75) put it. Pritna facie,they disagree about innate content; Kant says we bring a prioricategories to experience, but he does not explicitly say they areinnate. Why then should they not be the results of an innatepropensity to (re)organise our experience?

This reworking of rationalism by no means weakens its claims.In enriching itself from within, the mind can still gain newknowledge about the world. The Pre-Socratic philosopher Anax-imander argued that the world is floating in the void. He came tothis conclusion by examining Thales's theory that the earth isfloating on water. This will not do, said Anaximander, because itleads to a regress. Using this and other theoretical principles hedeveloped his theory. (He then opened his eyes, saw a flat earth,and concluded that the world is shaped like a barrel. Popper(1964) says he should have kept his eyes shut longer!) The pointis that in formulating his theory he didn't use any new facts: here-represented what he already knew.

The RR hypothesis also points toward' new problems. Itsuggests that when knowledge is re-represe!nted it remains thesame knowledge. This leads to a problem that is new to philoso-phy: knowledge-identity. How can procedural knowledge anddeclarative knowledge be the same knowledge? What does iteven mean to say this? (Remember, too, that Karmiloff-Smithsays that RR gives us new knowledge. But it cannot be both newand the same!)

We can add to the RR hypothesis to give us a still broaderpicture. In Dartnall (1992; 1993; 1994) I sugjgest that RR throwslight on intentionality (the symbol grounding problem; Hamad1990). We can combine RR with an "information theoretic"account of intentionality (Dretske 1980; 19$1; Heil 1983; Sayre1986). This exploits the mathematical theory of information,which says that one state carries information about another justto the extent that it is lawfully dependent bn that other state.Applying this to intentionality we can say, "Any physical systemwhose internal states are lawfully dependent . . . on the valueof an external magnitude . . . qualifies as an intentional system"(Dretske 1980, p. 286). :

Consider a connectionist system such as NETtalk (Sejnowski& Rosenberg 1986). When NETtalk's peripheries are bom-barded with phonemes, it settles into internal states that are"lawfully dependent . . . on the value of an external magni-tude." It settles into states that are about phonemes, and thataccount for its ability to pronounce words. Of course, it re-quires, for example, cluster-analysis to identify these states; thestates are not available to NETtalk.

Now suppose the system is a redescriber. It redescribes itsimplicit knowledge as explicit structures - as thoughts. Theseare about the world because they are redescriptions of knowl-edge that came about by being causally situated in the world;their intentionality lies in their pedigree. !

The RR hypothesis, then, gives us a rich picture of nativeendowment, knowledge, creativity, and mental life, and with alittle help from its friends throws light on epistemologicalproblems both ancient and modern.

The task is to model the underlying mechanism. The problemis that we do not know how, and are not really sure how to framethe question properly (see Clark 1993; Clark & Karmiloff-Smith,with commentaries, 1993). I will suggest that we have miscon-ceived the problem and that we need to redescribe representa-tional redescription.

Clark (1993) argues that the concepts we use to think aboutthought and cognition are inevitably folk-th.eoretic, that cogni-tive science is conceptually dependent op concepts such as"thought," "proposition," "concept," "knowledge," and "belief."I agree; but I also believe that these concepts need to beanalysed. When we analyse them we find that they are theoryladen. I believe that this theoretical component is itself theresult ofRR: RR has idealised the phenomena we are trying tounderstand, so that we have an overly abstract conception ofcognition. We have constructed theories about our own behav-iour and now try to model a hypothetical mechanism that mapscognitive states onto theories about the behaviour generated bythose states. This is a category mistake.

I will illustrate this by looking at our concept of language. Thishas been influenced (very recently, as these things go) by the


Com?nentart//Karmiloff-Smith: Beyond modularity

development of writing, grammar, and linguistic theory. Be-cause we have a written culture, we think of a natural languageas a set of sentences built out of a set of words, which are built outof a set of letters. A lot of linguistic theory takes this further andcharacterises natural languages as formal systems in which setsof sentences are generated out of a finite set of elements by afinite number of rules (see especially, Chomsky 1957).

These are theories that we have brought to linguistic behav-iour. They give us an excessively abstract conception of languageand direct our attention away from what we need to explain:linguistic behaviour. Rather than looking at writing and formalsystems, we should look at speech or Sign. (For an excellentintroduction to Sign, see Sacks 1989). It is not really paradoxicalto say that language is the idealisation, the theoretical redescrip-tion, of linguistic behaviour. In modelling the emergence oflanguage we need to model the way we have progressivelymodelled our own behaviour.

I provide this as an accessible, though very recent, example ofthe way we have idealised our behaviour. At a deeper level wehave idealised the concept of thought itself. Philosophical anal-ysis can unravel this redescription to give us a leaner, behav-iourally oriented set of concepts. I believe I can show, forinstance, that propositional attitude talk is a convenient but, inprinciple, dispensible way of talking about ability. In light ofthese leaner concepts we can see RR as a process that hasevolved by constructing implicit theories about our behaviour.

The evolutionary story that falls out of this account is thatindividuals born with an ability to modify their behaviour have abetter chance of survival, so that the ability to redescribebecomes an inherited characteristic. Modelling the evolvedmechanism in the individual is likely to be too difficult, but itmight be feasible to discover the kind of mechanism that evolveswhen we select from individuals born with an ability to modifytheir behaviour. This relates research in developmental psy-chology to attempts to model the origins of language andcognition in the species (e.g., Edelman 1992; Edelman et al.1992; Reeke et al. 1990a; 1990b). We should take not just adevelopmental but an evolutionary approach to the emergenceof mind.

Such an approach throws light on the vexing question oflanguage. Karmiloff-Smith argues that RR precedes languagelearning in the individual. Others (e.g., Dennett in press) arguefor the importance of language in understanding RR. An evolu-tionary approach resolves this conflict. We can conjecture thatthe development and exploitation of what we might loosely call"interactive behaviour" in the species played a vital role inevolving innate redescriptive abilities in the individual. Thus, atthe level of individual development Karmiloff-Smith is right:there is an innate redescriptive ability that precedes languagelearning and to a large extent makes it possible. Interactive,"protolinguistic," behaviour got its finger into the pie early onthough. Individuals born with a capacity to modify their behav-iour, especially their interactive behaviour, had a better chanceof survival, and RR became an inherited characteristic. Itstarted with very basic: abilities. But you gotta start somewhere.

The risks of rationalising cognitivedevelopmentBeatrice de GelderDepartment of Psychology, Tilburg University, 5000 LE Tilburg, TheNetherlands, and Laboratoire de Psychology Sxp6rimentale, UniversityLibre de Bruxelles, Belgium, [email protected]

Abstract: The notion that cognitive development might be a redescrip-tive and demodularising process raises two issues: (1) the apparentsymmetry between initial state and adult state modularity, and (2) thecontinuity and temporal logic assumed to link implicit and explicitrepresentations.

Beyond modularity (henceforth, Modularity) presents a post-modular picture of mental representations, arguing that modu-lar initial state representations implicit in behavioral masteryand in domain-specific processing undergo successive phases ofrecasting. This process ultimately brings about the uniquelymensch-like achievement of explicit representation in symbolicthought.

1. Development as demodularisatlon. At first blush, the claimthat the notion of a modular mind is not useful for understandingcognitive development is difficult to integrate with a host ofneuropsychological observations over the last two decades thatseem to fit the modularity thesis. Modules whose candidacyappears best supported by the facts from adult neuropsychologyare also those for which there is currently the best evidence inearly infancy. As a matter of fact, the chapter titles of Modularitywhich detail the demodularisation process (e.g., the develop-ment of linguistic, physical, mathematical, psychological knowl-edge), read like a list of the major cognitive domains for whichadult disorders have been documented. If initial-state mod-ularity is minimal and quickly superseded, how is one to explainthese striking continuities in functional architecture betweenthe infant and the adult brain? The continuity becomes evenmore puzzling when Modularity also claims that the modularityof adult cognitive architecture is a developmental product andresults from representational redescription (RR).

2. Does explicit representation result upward mobility of Im-plicit representations? The RR approach incorporates a tempo-ral logic where explicit representation depends on upwardmobility of implicit representation. Alternatives to such a conti-nuity view could be that implicit and explicit representationsexist independently, operate in interaction or in parallel, and ifso, that they belong to representational systems that maydissociate, with one possibly present without the other. There isample empirical evidence of this alternative and much morecomplex relationship between implicit and explicit processesand representations in adult studies - normal as well as neuro-psychological - perhaps most clearly in the study of memory(Weiskrantz 1987). In the area of developmental phenomena, acritical observation would be one showing that knowledge asinferred from conscious report and knowledge as inferred fromimplicit performance may dissociate.

The initial state competence for speech processing is in placein early infancy. Moreover, as reported in Modularity, explicitand conscious knowledge about language seems to arise veryearly on. Four-year-olds show explicit phonological knowledgein word games, rhyming, and so forth, and a couple of yearslater, they easily segment words into phonemes. Where doesthis explicit knowledge come from? A resoundingly popularanswer in the early seventies was enshrined in expressions like"the child as a young grammarian," admitting simply that overtime infants came to have access to their implicit phonemicrepresentations. Since those days, the list of problems with thisview of explicit as access to implicit has grown longer and longer.

To begin with, it can no longer be assumed that speechperception consists of a simple mapping from the signal tolinearly ordered phoneme-like segments. Second, it turns outthat achieving explicit representation generally requires tuition(de Gelder et al. 1993; Read et al. 1987). Third, modularity maybe far more solidly entrenched in the functional architecture ofthe modular initial state than even Fodor suggests, a possibilitysuggested by cross-modal interactions in the language modulethat cut across the sensory systems. Representational systemsare thus likely to be abstract from the beginning and to remain soall along. Fourth, the assumption of implicit phonemes waschallenged by connectionist models, with the very same impli-cations for the link between implicit and explicit representa-tions, and with extra uncertainty about the nature of the implicitrepresentations that may (or may not) underlie a given explicitmanifestation. For example, with backpropagation models ofword recognition, there is no control over the kinds of implicit



representations the network will develop. If phonemic repre-sentations are developed in the hidden units, this still will notprovide a basis for explicit phoneme representations, unlessanother representational system takes the network as its objectand projects an interpretation for it in a metalanguage, equatingpatterns of activation with phonemes (Norris 1992).

What, then, are the constraints on such metaprocesses? Arethey still domain-specific or will any kind of instructional envi-ronment promote metarepresentation? Evidence about specificphonological deficits in developmental dyslexias suggests theformer. There may then be a trade-off which is problematic forany RR kind of view: the less domain-specific the input to thedevelopment of metarepresentation, the looser the connectionbecomes between implicit and explicit representation. Thisopens the gates for dissociations between implicit and explicitrepresentations. It is generally observed that young, poorreaders lack explicit segmental representations of spokenwords. Is this a delay in an RR-like process? Do adult poorreaders catch up and have explicit representation in adulthood?As a matter of fact, there is evidence that they do (de Gelder &Vroomen 1991); but do these explicit representations result froma process of bringing implicit representations out? Not only doreading skills of this group remain very poor, but there areindications that implicit representations of speech are anoma-lous. If so, we would have a case of explicit phonemic represen-tations without implicit ones. A model of adult language process-ing where implicit processes are first and fastest and explicitones come last and take longest, would hardly be acceptable.Likewise, one seriously hesitates to consider that in develop-ment implicit representations simply precede explicit ones. Itwould seem that neither chronometry not chronology offers aunique cue to implicit versus explicit processes andrepresentations.

One might view Modularity as a developmental theory of thecentral processor, noting that it was Piaget who imposed suchphilosophical a prioris on development psychology, leaving itwith the mission of providing an empirical solution to theepistemological question of the possibility of objective knowl-edge (subject-object interactions grounded in sensorimotor be-haviour with representational systems growing out of it throughgradually incorporating the external object of knowledge intothe cognitive structures of subject). Philosophical credos makeheavy saddles for empirical theories. Of course, we are attachedto the thought of the human mind as an integrated whole, butthat is quite a mouthful for a scientific program. Many currentapproaches to adult cognitive architecture portray this as afederation of autonomous and interactive systems, with distri-bution of processes over separate subprocessors, multiple sys-tems underlying what appears superficially as the same behav-iour and modular processes across vastly different sensorialsystems. For the time being, understanding development maybe better served with the same working hypothesis.

Representation: Ontogenesisand phylogenesis

Merlin DonaldDepartment of Psychology, Queen's University, Kingston, Ontario K7L 3N6,Canada. [email protected]

This is one of the more enjoyable books I have read in some time;it is clear, provocative, and positive. Many of the ideas pre-sented need further testing, but this is precisely the point ofconstructing an integrative hypothesis - to provoke empiricaltests of the hypothesis. RR theory addresses two structuralissues central to human cognitive ontogenesis: (1) how memoryrepresentations become explicit, and (2) how the modular struc-ture of mind can be affected by ontogenetic factors. I have

argued elsewhere (Donald 1991; 1993a; 1993b) that both ofthese structural issues are also central to human cognitivephylogenesis. I wish to explore very briefly some of the parallelsand differences between the ontogenetic and phylogeneticapproaches.

1. The accessibility/retrievability Issue. The crucial memoryevent in human cognitive evolution, as in development, wasgaining explicit access to memory. Our closest relatives, theapes, have very poor explicit memory retrieval, whereas hu-mans voluntarily call up items from their own memory banks,reflect on them, alter them, and store the products of their ownreflection. The only apes showing a more advanced capacity forexplicit retrieval are those raised in human culture, and evenunder these circumstances, where symbols and languages areprovided, there are severe limitations to what they can achieve.In this aspect of their cognition, apes seem more like othermammals than like humans; they seem to lack the kinds ofaccessible representations needed for explicit memory re-trieval. In contrast, all neurologically normal humans are capa-ble of performing explicit memory tasks. It follows that homi-nids must have passed through a series of cognitive adaptationsthat eventually gave modern humans the powerful retrievaldevices that support explicit memory. This places the questionof explicit access to memory at the center of human cognitiveevolution.

Converging evidence from several disciplines suggests thatearly hominids evolved at least two different ways to constructautocuable, that is, self-triggerable, memory representations;these consisted of a nonverbal, or mimetic route, and a verbal,or linguistic route (Donald 1991; 1993a; 1993b). These retrievalcapacities were probably not acquired as adaptations in theirown right; they were byproducts of a more general change inrepresentational capacity, driven primarily by sociocultural se-lection pressures. Accordingly, there are a number of possibleways - both verbal and nonverbal - to gain explicit access torepresentations in adult humans. In this respect, I agree fullywith Karmiloff-Smith, at least in principle.

2. Modularity. The modularity issue is just as central to evolu-tion as it is to development. Did human cognitive evolutionalways occur within specified domains, or was it domain-general? I have argued that so-called nonmodular or domain-general adaptations, those that Fodor places in the "centralprocessor" are in fact quasimodular in humans. Retrievablerepresentations are superordinate models that serve to describeknowledge encoded in traditional mammalian systems, such asthe perceptual systems. They are thus scaffolded on top of amore fundamental, episodic level of representation, which itselfremains implicit. ;

Explicit access to episodic memories is a uniquely humanskill; in animals, episodic memories are implicit (Donald 1993b).In effect, to retrieve a given memory, the latt0r must be recodedin accessible form using either of the two autocuable pathsacquired by humans in evolution. Karmiloff-Smith's RR theoryseems quite compatible with this way of thinking, at least on firstview; her proposed process of redescription even has a nonver-bal dimension that seems to correspond to what I call the"mimetic mode," as shown in her E2 phase, where the child hasconscious access to representations but cannot yet speak aboutthem. What is impressive about this convergence of views is thather database, terminology, and theoretical starting points wereso different from my own.

Another feature of the RR model that appeals to me is that ithas room to accommodate the impact of literacy training onontogenesis. I have proposed that one of the seminal develop-ments in recent human cognitive evolution has been the exter-nalization of memory. Since the emergence of external symbolicstorage in the Upper Paleolithic, the human representationaluniverse has not only been expanded greatly, but it has alsoundergone a basic structural change. The thousands of hoursdevoted to acquiring symbolic literacy skills must have a tre-


Commentany/Karmiloff-Smith: Beyond modularity

mendous effect on the brain's growth process, and especially onthe tertiary regions of association cortex which have such a highdegree of plasticity. Thus, culturally imposed changes in thepattern of early experience can affect the modular structure ofcognition, and especially of explicit memory, in the modernadult. In this there is a convergence of phylogenetic factors withthe developmental factors described in detail by Karmiloff-Smith.

One obligatory caveat: although there can be many points ofconceptual agreement between ontogenetic and phylogeneticapproaches- to the structure of human cognition, it would bedangerous to generalize from onto- to phylogenesis, or viceversa, and especially dangerous to generalize with respect to theorder of emergence of specific kinds of representations. Modernhuman infants may have all the cognitive adaptations acquired inhuman evolution, but they need not develop in the order of theirevolutionary succession, except perhaps, in certain cases, intheir first point of emergence in development. The best exam-ple of a discrepancy in order of emergence is the very rapidappearance of speech, early in ontogenesis, despite the lateappearance of speech in phylogenesis. Because of this, languageand other modes of representation are in continuous interactionthroughout development, as Karmiloff-Smith emphasizes. Theresulting hybrid systems of representation are special to humansand would not necessarily resemble what preceded them inphylogenesis.

Developmental psychologyfor the twenty-first century

David EstesDepartment of Psychology, University of Wyoming, Laramie, Wyoming82071. [email protected]

Abstract: Karmiloff-Smith's landmark work decisively addresses centralproblems in developmental psychology and cognitive science. Her mostvaluable contributions include specifying how nativist and constructivistviews of development are complementary, how development can beboth domain-specific and domain-general, and how we are special. Hersolution to the latter problem correctly focuses on the development ofmetacognition. Two questions about Karmiloff-Smith's theory areraised: Can it explain its own development? Can it enhance the acquisi-tion of expertise?

One of the first articles I read in graduate school was Karmiloff-Smith's (1979b) inspiringly clever research on children's sponta-neous refinement of representations that were already entirelyadequate for the task at hand. I remember being struck then bythe idea that here was evidence for a natural predispositiontoward more explicit understanding and, ultimately, eleganceand simplicity - motivations often ascribed to scientists andmathematics but here already evident in the untutored nota-tional behavior of children. Now Karmiloff-Smith has given us aricher and more explicit account of her ideas on cognitivedevelopment and representational change. Let me begin bygoing out on what looks to be a very sturdy limb to predict thatthis book will be recognized as a landmark, one that providesfurther evidence for the maturity of developmental psychologyand its potential for illuminating not only the course of cognitivedevelopment but the nature of the human mind as well.

An especially admirable aspect of Karmiloff-Smith's work isher willingness to tack head-on the question of "how we arespecial.' This crucial issue is too often either ignored altogetherin cognitive science, or worse, it is argued that in fact we are notspecial. Karmiloff-Smith reviews convincing evidence on theoverwhelming cognitive limitations of other species comparedto humans, but otherwise does not waste our time belaboringthe obvious fact that we are special. (Notify us when either

chimps or chips begin to establish peer reviewed journals orresearch institutes.) Karmiloff-Smith's response to the questionof how we are special is to provide a strong case for the centralrole in human life and human development of "metacognition'and the process of "explicitation." Her account of "representa-tional redescription" (the RR model) is in my view the mostdetailed and plausible existing model for the development ofwhat is arguably our most important and quintessentially humancharacteristic - the ability to make what we already know insome form increasingly explicit and conscious, and conse-quently more meaningful, flexible, and useful.

Karmiloff-Smith deals decisively with several other centralissues in cognitive development and cognitive science. Sheprovides an appealing alternative for those of us who are uncom-fortable attributing to the infant even a small set of innate"theories" but who were beginning to wonder how to avoid it,given accumulating research demonstrating remarkable earlyunderstanding in several domains. Her distinction between"innate modules" and a "process of modularization" is importantbecause the latter retains the essential core of Piagetian con-structivism, something that many developmentalists, myselfincluded, are unwilling to discard along with strict Piagetianstages and domain-general, all-encompassing cognitive struc-tures. At the same time, positing a process of modularizationrather than a set of innate modules allows Karmiloff-Smith toaccount both for recent findings on infant perception and cogni-tion, as well as for evidence that much knowledge is domain-specific. No doubt many in the field of cognitive developmenthave had the intuition that nativist and constructivist theories ofdevelopment are complementary rather than contradictory, orthat development is somehow both domain-specific anddomain-general. Karmiloff-Smith has done something aboutthese widely shared intuitions by presenting a detailed accountof how they could be true.

One important unanswered question about Karmiloff-Smith'sRR model concerns its intended scope. Her primary goal isobviously to explain major aspects of cognitive development,but the model might explain other phenomena as well. Perhapsthe most interesting additional application of the RR modelwould be to explain its own development. It would seem todescribe rather well the work of individual scientists such asKarmiloff-Smith herself as they move from a kind of "behavioralmastery" of specific research findings to the abstraction of trendsand general principles, and then progressively formalize, gener-alize, refine, connect, and make more explicit, ideas that exist insome nascent or implicit form. In fact, we might wish that somesuch criterion - some reasonable potential to explain its owndevelopment in the minds of individual scientists - could beapplied to any theory of human cognition that wishes to be takenseriously. This would narrow the field considerably. Most gen-eral frameworks purporting to explain human thought andaction (behaviorism and radically reductionistic forms of neuro-science leap immediately to mind) could do little more thanappeal to some vague faith in future progress if they wererequired to explain the scientific activity of their adherents. Incontrast, scientific theory development might be seen as just aspecial case of Karmiloff-Smith's general RR model.

Another question one might ask about the RR model is really aversion of what Piaget called "the American question": can weuse the insights provided by this theory to enhance develop-ment and learning? Karmiloff-Smith is clear regarding thereason that development takes time. Behavioral mastery, inmost cases an inherently slow process, is required before repre-sentational redescription can occur, so even if we wanted to tryto speed up development (and who wants to shorten childhood?)it might be difficult or impossible. What about nondevelopmen-tal phenomena, however, such as the acquisition of expertise?Karmiloff-Smith notes that the process of representational re-description occurs in adulthood "for some kinds of new learning"(p. 18), but this application of the theory is not emphasized. One


Commentan//Karmiloff-Smith: Beyond modularity

wonders what the effect would be on both teacher and learnerif they knew in advance, thanks to a detailed model like this one,that the acquisition of expertise (in jazz piano or computerprogramming, for example) requires an initial level of behavior-al mastery followed by increasingly flexible, abstract, and ex-plicit levels of representational redescription (the nature ofwhich might be specified in detail for a given area). AlthoughKarmiloff-Smith describes representational redescription asprimarily an endogenous and spontaneous phenomenon, mak-ing the process of explicitation itself explicit, and therebyproviding insight into and external support for the cognitiverepresentational changes that occur when we master some newskill or area of learning, might have profound implications foreducational practice. Karmiloff-Smith is admirably restrained inher speculations about the implications of her work, but thatdoes not mean the rest of us have to be. In fact, we might invokethe RR model itself to provide those of us who need it withtheoretical justification for trying to make more explicit ourvague intuitions, for actively trying to push some level-I or -Elrepresentations toward level E3.

No one else, to my knowledge, including those who are nowfinally attempting to grapple directly with the nature of con-sciousness and its role in human cognition, has presented asdetailed, plausible, empirically-grounded, and potentially fruit-ful a model of how implicit understanding becomes explicit andconscious. This alone is sufficient reason for cognitive science totake developmental psychology seriously, something whichKarmiloff-Smith recommends and which her work should in-sure.

Arguments against linguistic"modularization"

Susan H. Foster-CohenDepartment of English, Northern Arizona University, Flagstaff, AZ86011-6032. [email protected]

Karmiloff-Smith's position that there is emergent modularity forlanguage is untenable, both empirically and logically. For pro-gressive encapsulation to be truly explanatory of language de-velopment, Karmiloff-Smith would have to demonstrate thatdomain specificity can be distinguished empirically from encap-sulation, that domain specificity exists independently of encap-sulation in the early stages of language acquisition, and thatevidence to the contrary is invalid. She has done none of these.

Take the evidence of the acquisition of the binding principlesfrom studies such as Chien and Wexler (1990), Grimshaw andRosen (1990), and McKee (1992). Their results may appear toprovide a confused picture of children's understanding of thebinding restrictions on pronouns and anaphors, leading those ofKarmiloff-Smith s persuasion to posit a lack of modularity in thisarea of the grammar. However, a closer look reveals goodevidence for modularity.

As Grodzinsky and Reinhart (1993) have argued, most of theproblems children have with tests of the binding principles areactually problems with pragmatic coreference in which childrenwrongly assume a local antecedent, rather than syntactic prob-lems based on not knowing the binding principles. I argue(Foster-Cohen 1994) that these problems are aided and abettedby pragmatic problems of relevance (Sperber & Wilson 1986;1987). Thus, a full understanding of the acquisition of pronouncoreference requires an appeal to two separate modules. Thatthis is the case is supported by the following four facts aboutchildren's behavior.

First, young children are sensitive to the relevance of thecontext in which a test sentence is presented; they perform lesswell when the linguistic and nonlinguistic contexts prior to thepresentation of the test sentence are at odds with the test

sentence. The processing load of trying to compute the prag-matic relevance of the context sentences (forced on them by thepresumption of optimal relevance predicted by relevance the-ory [RT]) supersedes syntactic computation. Second, testingparadigms in which there is major redundancy of informationpresented in the pragmatic context for the syntactic test, reducethe success rates because the processing costs of computing therelevance of repeated material interferes with children's abilityto process the syntax of the next utterance. Third, testingparadigms that do not respect (relevance-driven) given-newinformation structure cause an overload in the processing ofrelevance. So, in a paradigm such as Grimshaw and Rosen's, inwhich an irrelevant individual is referred to in sentence-finalfocus position in the sentence immediately prior to a testsentence, children perform more poorly. All three of these factssuggest that children have trouble abandoning pragmatic pro-cessing and moving on to the syntactic processing of the nextutterance. Finally, children show a sensitivity to conceptualinformation (real-world gender) in their interpretation of pro-nouns, but not in their interpretation of anaphors (Chien &Wexler 1990). This only makes sense in a framework in whichpronoun coreference is carried out in a pragmatic module,whereas anaphoric reference is purely syntactic.

A modular account presumes that children process eachutterance syntactically first and then move to the pragmaticinterpretation; and the data support this interpretation. Becauseof the encapsulation and ordering of the separate modules, theyfind it hard to attend to the syntax of the next utterance untilthey have completed the pragmatic interpretation of the pre-vious one. In addition, if that next utterance is the test sentence,the consequence is potential failure on the syntactic test.

There are those who suggest that pragmatics is not encapsu-lated at all - that it is, rather, part of a generalized centralprocessing mechanism. This remains to be seen from subse-quent empirical research. However, the linguistic specificity ofReinhart's Rule I (which says that locally rqlated noun phrasescan corefer, provided this yields a pragmatic interpretationdistinct from the parallel anaphor-noun phrase construction)suggests that we are more likely to find that there is indeed apragmatic module covering some, though maybe not all, aspectsof the interpretation of communicated messages.

Finally, Karmiloff-Smith's progressive "modularization" ac-count has some specific logical flaws. First, she is willing tointerpret evidence that children are innately predisposed todistinguish linguistic from nonlinguistic facts as evidence ofdomain specificity, but not modularity. However, without sys-tematic arguments (which she does not provide) for how todistinguish domain specificity from encapsulation, it is equallylegitimate to use the same evidence as support for modularity.For example, she interprets evidence from the acquisition ofASL (American Sign Language) in which children show a com-plete obliviousness to any connection between gestures that canbe used to indicate "you" and "me,' and the ASL "words " for thesame, as evidence for domain specificity without providing anyargumentation against that same evidence being interpreted asmodularity.

A second flaw is that Karmiloff-Smith provides no discussionof how a child could learn from experience that there aresubmodules within the linguistic system, for example, that theconstraints of binding, bounding, theta roles, and so forth, aredifferent from the constraints of morphology or phonology. Thiswould seem to be at least as severe a logical problem as those forwhich she advocates an innateness argument (e.g., the need todistinguish between the sensorimotor understanding of nestedcups and the similarly named, but actually radically different,notion of nested clauses).

Finally, it is not clear why Karmiloff-Smith needs the notion ofprogressive encapsulation at all. If it is needed simply to allowspace in her theory for "constructivism" of a fieo-Piagetian type,it is misguided. Surely there is already plenty of scope within an



"innatist" view for the role of input, and for children to deduce,at various stages of development, many of the rules of theirlanguage from the interaction of what they already know withnewly encountered input. Progressive modularization wouldonly be of value to a theory of development if it provided aconvincing reanalysis of data currently covered by universalgrammar modules. If the only data appealed to are those thatwould be regarded as peripheral and "learned" in any case, thenthe entire concept of modularization misses the mark.

Redescription of intentionality

Norman H. FreemanDepartment of Psychology, University of Bristol, Bristol BS8 1TN, [email protected]

Abstract: Concepts of level of exploitation and of representationalredescription are useful in explaining cognitive transformations in sev-eral domains. An excellent framework theory results, but its nestedspecific theories are difficult to assess in the absence of an explicitaccount of representation. Treatment of the domain of intentionalmental states can be straightforwardly strengthened, but the domain ofdepiction merits reworking.

In Beyond modularity (hereafter Modularity), we get an excellentadvance in analyzing what internal resources the cognitive systemaffords in the furtherance of intellectual transformations.

The child as psychologist. The analysis of the child's construc-tion of intentional psychology simply does not go far enough inusing tools provided in earlier sections. Several "establishment"accounts are characterised by a single-level dichotomy of "eitherthe child has it or does not." Perner (1991) hypothesised anepistemological revolution around age four years, whereby aconcept of representation comes into existence and thereafterthe child works out the consequences. Gopnik and Wellman(1992) presented an even closer mapping of changes in behav-ioural mastery onto changes in the child's beliefs about inten-tionality. Since Modularity has distinguished between levels of"explicitation," why has it held back from deploying the sameanalysis of The Child as a Psychologist? It is easy to maketestable predictions concerning levels of "explicitation."

First, it should be that children have implicit representationsthat enable them to make correct computations in the domainlong before there is verbalisable access to the success. There isvalue in Povinelli and cleBlois's (1992) demonstration that 3-year-olds followed an experimenter's pointing gesture in order to finda surprise, yet none could verbalise how they managed tosucceed. All that was required was to report that the experimen-ter had pointed, and the children had months of behaviouralmastery of deictic proto-imperatives. The suggestion is thatrepresentational redescription is extremely slow in this inten-tional domain. Let us accordingly follow an analogue ofKarmiloff-Smith's analysis of linguistic self-repair and askwhether there is anything that looks like self-repair in children'sproduction of intentional actions.

Freeman et al. (1991) used a false belief test with dolls playinghide and seek. One doll hid at A, and then, unseen by theseeker, it moved to B. As usual, a large number of preschoolerserroneously said that the seeker would search at B, or that theseeker thought the hider was at B. Yet, with 186 children tested,66 such children, when asked to take hold of the seeker doll,spontaneously made it go to A - which is what the doll would doin accord with its false belief (with three counter-instances ofverbalising the false belief yet taking the seeker straight to B).Clements and Perner (1994) reported that young 3-year-oldswould anticipate a protagonist's erroneous search by turningtheir eyes to A.

Second, Lewis (1994) discussed studies in which repeating

most of a story, before reaching the final episode in which theagent develops an intention to seek the desired object, made itpossible for 3-year-olds to verbalise subsequently the agent'sfalse belief. The technique worked best when the childrenthemselves narrated the story back, thus verbally recycling thedata explicitly through their processing system. Does explicit-ness prime access to "explicitated" representations? It would benice to believe that one gains access to knowledge by talkingthings through.

Third, Yoon and Yoon (1993) reported that children could takeaccount of a false belief in answering the question of where Bettywill look for her doll before they could correctly answer "wheredoes Betty think her doll is?" That is, many children couldpredict mistaken action before being able to articulate thereason for the prediction. Lewis (1994) discussed a similareffect. That advantage appeared very weakly in Chandler andHala (1994) and not in early studies (Hogrefe et al. 1986; Perneret al. 1987). No one has reported the reverse effect. Conceiv-ably, the look question allows some children to gain access totheir representation of the agent's intentionality at the "primor-dial level" (as Searle 1983, has it) of computations on perceptionand action before they can disembed the think component(Freeman 1994).

In sum, Karmiloff-Smith's insistence on levels of "explicita-tion" points the way. There is no discernible grip for representa-tional redescription, however, unless one counts the Povinelli-deBlois demonstration that the child's deictic knowledge has yetto be redescribed in higher-level intentional terms - perhaps inthe way that "pronouns are not an extension of gesture" (Mod-ularity, p. 39). Redescription generates new orderliness.

The child as notator. My critique of the analysis of the child'sgrasps of graphic systems is that Modularity has forced thedomain into a format to match other chapters. That is not to saythat the account we are given is false. Rather, the account fails todo justice to the richness of transformations in cognisance of thedomain. I will focus on the child's noncartographic pictureproduction.

I suspect that the weakness of the section attests to recapitula-tion of its ontogeny. The precursor is a convincing analysis oflanguage. On pages 44-45, Modularity makes the point that"Although infants are sensitive to word order, one should notconfound this with sensitivity to serial order. . . . Childrenorder abstract linguistic categories." The fact about depiction isthat there is no abstract equivalent of a lexeme, no equivalent ofa function word, nor of modal-operator brackets.

So what one finds in the Modularity analysis of drawings isfalling back on serial order. That might seem entirely reason-able, yet Modularity contains the dignified retreat that "thesequential constraint on the first level of redescription is, partic-ularly in domains like drawing, considerably weaker than Ioriginally predicted" (Modularity, p. 162). What went wrong?

Depiction should be a suitable case for Modularity treatmentsince it is downright implausible that that there is a dedicatedinput processor to feed a Fodorian module (pace Ellis 1989). Isuggest that a wrong turn occurred with an earlier overstate-ment: "Theory-of-mind computations are special in that it isdifficult to find another area of human cognition in which thedistinction between propositional contents and prepositionalattitudes is a crucial component" (Modularity, p. 127). Philo-sophical aesthetics is drenched with that, however, and cognatedistinctions as well. There has been a long struggle over ontol-ogy concerning whether observers discover or construct picto-rial meaning, and the latter is "an appropriate response to thecentral fact about art: that it is an intentional manifestation ofmind" (Wollheim 1993, p. 134). Children seem to grapple withthis fact (Freeman 1995). Let me cite the comment of a 7-year-old; when we asked whether a pretty picture of a snake wouldlead someone to revalue their negative attitude to real snakes,the reply was, "Not exactly, because that's only a picture, that'ssomeone's mind being put on paper." The mind on paper



formulation catches something of the intentionality of depiction.What Modularity should have done is to grasp the inten-

tionality issue, upgrade the analysis of the child's construction ofpsychology as I suggested earlier, and sweep into the net theemergence of the child's intentional theory. The clue to advanceis given in: "To have theoretical status, knowledge must beencoded in a format usable outside normal input/output rela-tions. It is these redescriptions that can be used for buildingexplicit theories" (Modularity, p. 78). Therefore, let us notconfine the study of pictorial reasoning to the normal output ofpicture production. Even preschoolers may have interestinghypotheses about how a picture serves to unite an artist, anobserver, and the world in an intentional net (Freeman 1995). Itis that terrain on which representational redescription can andshould be tested.

Do you have to be right to redescribe?

Susan Goldin-Meadow and Martha Wagner AlibaliDepartment of Psychology, University of Chicago, Chicago, IL [email protected] and [email protected]

Abstract: Karmiloff-Smith's developmental perspective forces us torecognize that there are many levels at which knowledge can berepresented. We first offer empirical support for a distinction made ontheoretical grounds between two such levels. We then argue that"redescription' onto a new level need not await success (as Karmiloff-Smith proposes), and that this modification of the theory has importantimplications for the role redescription plays in development.

One of the many insights in Karmiloff-Smith's book is thatdevelopment does not stop at behavioral mastery. For example,several years after children succeed at balancing a wide varietyof oddly shaped blocks on a narrow support, they begin to makeerrors on the same task. They now ignore the proprioceptivecues used so effectively years earlier and are able to balance onlythose blocks with weight evenly distributed around their geo-metric center. Is this developmental progress?

Karmiloff-Smith says "yes." She argues that such changes inchildren's performance reflect the fact that they have begun toredescribe that knowledge. In the process of redescription,particular aspects of the "data" are highlighted (in this case, thegeometric center of the block) and incorporated into a theory inaction. Phenomena of this sort (and the book contains manycompelling examples in a variety of domains) point to a level ofrepresentation in which certain aspects of the child's knowledgeare explicitly defined. By "explicit" Karmiloff-Smith means thatthe information is no longer embedded in the special purposeprocedures of the earlier period. This is the first step in makingknowledge accessible beyond the particular task in which it wasdeveloped (level El), the first step in the process called "re-description."

1. Gesture offers empirical support for a distinction made ontheoretical grounds. Studies of adults tend to distinguish twolevels of knowing (implicit vs. explicit; unconscious vs. con-scious; automatic vs. controlled). Karmiloff-Smith's develop-mental perspective, however, forces us to recognize that thereare many levels at which knowledge can be represented. Shedistinguishes two levels beyond the first level of redescription(level El) - one in which knowledge is available to consciousaccess but not to verbal report (level E2), and one in whichknowledge is available to both conscious access and verbalreport (level E3). The book collapses the theoretical distinctionbetween these two levels on the grounds that there is very littleempirical research to support this claim. However, empiricalevidence does exist for such a distinction. The evidence comesfrom gesture.

When children are able to conserve number on a Piagetiantask, they often justify their correct responses with a verbal

rationale, and that verbal rationale is frequently accompaniedby gestures that convey the same information. For example, achild might describe one-to-one correspondence between tworows of checkers in speech and simultaneously point out thesame correspondence in gesture. Such a child would presum-ably be at level E3. However, even before children are able toexpress some concepts in speech, they may do so in gesture. Forexample, a child might persist in saying that one row of checkersis longer than the other, while at the same time pointing out acorrespondence between the two rows in gesture (Church &Goldin-Meadow 1986). Such a child is not yet able to expressverbally one-to-one correspondence, but the knowledge is stillthere, in some sense. The question, according to Karmiloff-Smith's formulation, is whether the knowledge is consciouslyavailable, as it ought to be to qualify for level E2. [See alsoShanks & St. John: "Characteristics of [dissociable HumanLearning Systems" BBS 17(3) 1994.] I

"Consciously available" is, of course, a slippery term. Typ-ically, speakers are not aware of the fact that they are gesturing asthey speak (cf. McNeill 1992; indeed, gestures change in formand function if they become the focus of communicative atten-tion; see Singleton et al. 1994). Consequently, if awareness is arequirement for consciousness, gesturing does not qualify.However, speakers do appear to have access to the informationthey express in gesture and, in this sense (a sense that webelieve captures the essence of the distinction Karmiloff-Smithis making), information conveyed in gesture is "conscious." Forexample, we have presented to children a solution to a mathproblem and then asked them to judge whether the solution wasan acceptable one for that problem. The solution presented wasgenerated either by a procedure that the child had (on a pretest)expressed in gesture but not in speech, or by a procedure thatthe children had not expressed in either modality. Childrenwere consistently more likely to accept solutions of the first typethan of the second type (Garber et al. 1994; Goldin-Meadow etal. 1993), suggesting that they did have access to the informationthey conveyed in gesture, and that they were able to apply thatinformation in a different context. Observations such as theseprovide evidence that some knowledge may be consciouslyavailable, though not able to be verbalized.

2. When does redescription begin, and what Is It good for? Ourstudies of gesture lead us to question a basic assumption ofKarmiloff-Smith's theory - that the child must master a taskbefore redescription of that task can begin. According to thetheory, redescription starts with correct information (repre-sented at the implicit or I level). The proces? takes this informa-tion and repackages it, highlighting certain aspects and omittingothers, but it does not add new information. This repackagingleads to cognitive flexibility but does not, according toKarmiloff-Smith, affect behavioral mastery. Thus, the theory ofredescription has nothing to say about the long period of devel-opment prior to behavioral mastery. It might, however, ifredescription were not tied to behavioral mastery.

We suggest that behavioral mastery is not necessary forredescription to occur. Well before children have mastered atask, they are able to articulate (in speech or gesture) beliefsabout the task - which implies that redescription has alreadygone on (e.g., Alibali & Goldin-Meadow 1993). It is, of course,possible that the child has achieved behavioral mastery by thistime, and that we do not know how to tap into this knowledge.This, however, is a weak way to save the theory. Even Karmiloff-Smith herself suggests (in the Pr6cis) that behavioral masterymay not be a prerequisite for redescription.

This slight modification of Karmiloff-Smith's theory has ob-vious implications for the onset of redescription and, less obvi-ously, implications for the role that redescription plays in devel-opment. When children are wrong, they are often systematic-ally wrong; that is, their incorrect answers are consistent andmake sense within their own framework. (This insight is one ofPiaget's most important contributions to developmental psy-


Commentart//Karmi]off-Smith: Beyond modularity

chology.) Stable states such as these could well set the stage forredescription. We suggest, however, that redescription of incor-rect knowledge could, in the end, be a destabilizing force. Forexample, a redescription resulting in newly framed explicitknowledge might feed back and encourage the child to alterimplicitly encoded knowledge, particularly if that knowledge isincorrect or incomplete. If implicit knowledge is altered, whatthen happens to the redescriptions that were originally formu-lated from it? Once we allow the possibility of redescription ofincorrect knowledge, we are forced to think about change - notonly at the higher levels where the same information is re-packaged, but also at the lower levels where new informationmay be added.

Although complicating Karmiloff-Smith's theory, our datasuggest that redescription of incorrect knowledge does indeedtake place. It seems to us likely that redescription beginswhenever a stable state is achieved, be it a correct or incorrectstable state. As a result, you probably do not have to be right toredescribe - but if you are wrong, beware of the consequences.

Dissociation, self-attribution,and redescription

George GrahamDepartment of Philosophy, University of Alabama at Birmingham,Birmingham, AL 35294. [email protected]

What's in a child? Karmiloff-Smith has written a book which mayhold the answer. The book is first-rate, bold in conception,articulate in telling, and synoptic in evidential sweep. I want tocomment on one and only one theme within it, although it is abig one, that of representational redescription. And I want tofocus on only one topic within that theme, although it is a topiccapable of bubbling over into other themes.

Karniiloff-Smith claims that information in a child, includinginformation about the child's own mind, can become pro-gressively explicit to that child, and once explicit it may bedeployed by the child in understanding self and world.Karmiloff-Smith calls this claim the representational redescrip-tion hypothesis. The information may be relatively confined tospecific areas of knowledge: number, language, human psychol-ogy, and so forth. Such confinement she dubs "domain specific-ity." Information may cross domains at different phases ofdevelopment. There may also be domain-specific damage in theform of degraded or destroyed information or computationalcapacities within a domain.

Why does Karmiloff-Smith call information in a child whichhas become explicit to the child "redescribed representation"?At first I was confused by this locution, for when a child hasinformation only inexplicitly or implicitly, the information is notdescribed by the child. So the coining to be of explicit informa-tion seems synonymous with coming to describe (not re-describe) one's otherwise implicit information. Her hypothesisshould really be called the representational description hypo-thesis.

So I thought - but then I discovered the wisdom in thephrase. For Karmiloff-Smith, exploiting internal information,making it explicit, is not an all-or-none affair. Informationbecomes progressively more explicit, in phases, fits, starts andstops, and in chunks, the normatively ideal end result beingrepresentational flexibility and control. No single descriptionmay or even should satisfy the child. One and the same segmentof information may be described and used in different ways, atdifferent times, for different purposes. Dynamically under-stood, the best term for the affair is "redescription."

My question concerns a form of redescription to which youngchildren seem particularly prone or vulnerable, that is, theprocess of dissociating themselves from their own minds bylinking or attributing their own mental states to other minds or

selves both real and imaginary. Children "play tricks" withimaginary playmates and even with themselves.

Sadly, dissociation may be anything but playful. It can betragic. Consider, by means of illustration, the following influen-tial explanatory hypothesis for the childhood origins of multiplepersonality disorder or MPD (Rhue & Lynn 1991).

Childhood abuse is a developmental antecedent common incases of MPD. Some abused children may use dissociation as apsychic safety valve to minimize negative affect (guilt, anger,anxiety) and to avoid conflict. They may invent or constructpersonae that enable them to disavow engagement in abusiveparental relations. These personae, or believed-in other selves,create a credible sense of separation or personal distance fromsubjective involvement in such relations.

For Karmiloff-Smith, powers of fantasy and imagination (in-cluding, arguably, the tendency to dissociate) are central cata-lysts in the dynamic process of representational redescription.As cases of nascent MPD and other dissociation phenomenareveal, making information in the child explicit to the child is notnecessarily making this information explicit for the child as thechild's. The move from implicit to explicit - the process ofrepresentational redescription - allows children to attend totheir own states of mind but this may happen without theiracknowledging or realizing that such states truly are their own.Redescription is one thing; self-attribution of explicit informa-tion about the self is another.

Karmiloff-Smith notes that "children spontaneously seek tounderstand their own cognition" (p. 192). She notes that thisallows them to become folk psychologists. Alas, however, it alsoallows them to become folk psychopathologists, as it were, sothat in order to keep pain at bay they may distance themselvesfrom information about themselves. Dissociation prevents infor-mation in the child from being (what may be called) self-referentially available. As such, it represents a departure fromthe normative ideal of representational flexibility and control.More exactly, dissociation may be locally adaptive by enabling achild to adapt to a current crisis but globally dysfunctional byundermining the child's ability eventually to face adult demands

and responsibilities.Throughout her chapter (Ch. 5) on the child as psychologist,

Karmiloff-Smith, to her credit, talks about children coming toterms with their own attitudes and with the contents of thoseattitudes. What is missing in Karmiloff-Smith's discussion isrecognition of a special form of metacognitive achievement,namely knowing my mind as mine, and the role this form mayplay in cognitive development. This is a topic on which I haveruminated (Graham & Stephens 1994; Stephens & Graham1994), and which, if I am right, deserves the attention ofcognitive scientists in general and developmentalists inparticular.

It remains an open question how the theory of representa-tional redescription may describe or explain the achievement ofself-attribution, and what role the phenomenon may play in theemergence of representational flexibility and control. Havingacknowledged the existence of childhood dissociation, whatlessons should a Karmiloff-Smith-inspired theorist draw? Onelesson has just been mentioned: making internal informationexplicit to the child is not necessarily making the child's own selfexplicit for the child. Self-attribution is a separable achieve-ment. Another lesson concerns the fragility of self-explicitnessand self-attribution. In line with Karmiloff-Smith's views of theprogressive and fitful emergence of explicitness, self-attributiondoes not appear as a sudden takeover of the child's own internalinformation; it more closely resembles a long and slowly matur-ing process, vulnerable to breakdown, and compounded by aneed for constant reconstitution.


Commentary/Karmi\off-Smith: Beyond modularity

Beyond connectionist versus classical Al: Acontrol theoretic perspective on developmentand cognitive science

Rick GrushDepartments of Philosophy and Cognitive Science, University of California,San Diego, La Jolla, CA 92093-0302. [email protected]

I agree in spirit with almost all the content of Beyond mod-ularity; it is in some of the details that I think Karmiloff-Smith'saccount could be improved. Specifically, the representationalredescription (RR) model fails to distinguish clearly two distinctaspects of learning, the forward problem and the inverse prob-lem. The forward problem is: given some system in a certainstate and some actions on that system, what will happen (predic-tion)? The inverse problem is: given some system in a certainstate, and a goal state, what actions must be performed to reachthat goal (control)? I will refer to solutions to the forwardproblem as emulators (they are also called, in the controlliterature, forward models or forward mappings) because theyemulate the input/output function of the target system. Solu-tions to the inverse problem I will call controllers (also known asinverse models or inverse mappings). Finally, I will call thesystem which is controlled or emulated the target system orsimply the target (this might be a Rubik's Cube, the mus-culoskeletal system, a VCR, or whatever).

Consider learning to accurately control one's arm. Minimally,what one would like is a good controller which, when given thegoal (e.g., locating an object one wishes to grasp) and the arm'sinitial state, computes a motor sequence - efferent signals whichspecify the timing and amplitudes of the contractions of themany muscles involved - such that the execution of that motorsequence puts the arm in the goal state. However, in addition tolearning a reliable controller, there are good reasons why fast,accurate control also requires learning an emulator of the rele-vant musculoskeletal dynamics.1 Such an emulator (in effect aninternal model of the arm) might be a central circuit thatreceives an efferent copy of the motor sequence and computesthe effect the motor sequence will have if executed properly.

Once one has learned a good emulator, many benefits accrue.First, the emulator can provide the controller with predictedfeedback much faster than real feedback is available from theperiphery. Second, the controller can operate the emulator "off-line" for planning purposes.2 Finally, the controller can "prac-tice" with the emulator, rather than being constrained to prac-tice on the real target.3 In short, having both a controller and anemulator allows the control loop to be internalized, allowing forthe possibilities of imagination, and so forth. The importantpoint is that this internalization depends on having an emulator,and not just a good controller. Even a very sophisticated para-metric controller with the capacity to generalize accurately topreviously unencountered examples cannot by itself internalizethe control loop.

Of course, as learning proceeds, the representational status ofemulators and controllers can change. We might assume thatduring initial experimentation with some novel domain, experi-ences get stored as a sort of "bracketed sequence" of an initialstate, some subsequent actions, and a final state:

[ S 2 3 » ^ 1 1 > ^ 2 3 8 . •••.

[S4 5 . A41 , A8

1 ^ * 8 3 » " ^ 1 8 ' ^ 1 2 4 > ••• '

etc.

: S,]

(1)

And in fact, as experience proceeds, we can imagine theaccumulation of a large set of these bracketed sequences, per-haps even enough so that for most encountered initial states andfor most goals, an appropriate sequence can be recalled from thelookup table. This, I take it, is something like behavioralmastery supported by level I-representations.

The extraction of implicit information from this table mightinvolve the determination ofafunction which takes initial statesand goals as inputs and computes an appropriate action se-quence (i.e., one that leads from the former to the latter) asoutput. With such a function one could replace a large list ofsequences such as (1) with a single parametric function (thecontroller) of the form:

AJ (2)

Of course, one can extract a different sort of information fromtables like (1) specifically, an emulator. Such a function, which,modulo the initial state specification, is the; inverse of/, wouldtake the form:

{g: (S, , A, , A2 A,,)->($,)} (3)

The conflation I see in Beyond modularity is that the RRmodel seems to collapse the two orthogonal processes of (a)internalization via learning an emulator, and (b) parametriza-tion, that is, the extraction, from an unstructured lookup table ofexperience, of a parametric function (emulator or controller)which uses separable, individually adjustable parameters. As Ihave argued, however, these processes are distinct.

This very schematic view of developmental learning has thevirtue, like the RR model, of being applicable to many levels ofbrain activity, from motor control to programming a VCR (a taskat which my parents refuse to advance beyond a lookup table-based controller). Another commonality with the RR model isthat it takes no stand on architectural implementation. Aspectsof the emulation-control model outlined here have been imple-mented successfully by connectionist architectures. For exam-ple, Rumelhart et al. (1986) discuss connectionist models ofemulators, and mention their use for modeling an opponent in agame of tic-tac-toe. Such a connectionist model seems clearlyapplicable to the child as developing a theory of mind. Kawato(1990) has built fascinating connectionist arm controllers whichlearn their task by way of first learning the forward model of thearm. Furthermore, Kawato reports that the hidden units in theemulator come to represent explicitly such quantities as inertia,viscous friction force, Coulomb friction force, and so on. That is,as learning of the emulator proceeds, the model extracts repre-sentations of independent parameters which are important forcomputing trajectories (as opposed to a list of past motor se-quences, which, at best, implicitly contain information aboutfrictional forces, etc.)

The control-theory-inspired view of learning and cognitionvery briefly outlined here might provide a way, contrary to thedoubts expressed by Karmiloff-Smith, for connectionist modelsto shed light on the transition from Ievel-I representations tolevel El , .and in fact to do so in a way which makes finerdistinctions than the RR model itself. This is not to say that Ithink the RR model is wrong. Indeed, if the control-theory-inspired model I have outlined is right, then the RR model fallsout as a sort of special case.

ACKNOWLEDGMENTSI would like to thank Robert Hecht-Nielsen for valuable discussions 011these topics. I would also like to thank the Cecil and Ida GreenFoundation and the ARCS Foundation for financial support.

NOTES1. For considerations to this effect, see Kawato et al. (1987) and

Kawato (1990).2. An example of a model which uses emulation-supported imagery

to plan complex limb movements, see Mel (1988).3. For an example, see Nguyen and Widrow (1989).



Representational redescription, memory,and connectionism

P. J. HampsonDepartment of Applied Psychology, University College Cork, Cork, Ireland.p|[email protected]

I strongly support Karmiloff-Smith's claim that a developmentalperspective is needed in cognitive science. For too long, cogni-tive science has examined the nature of fully developed, adultcognitive abilities and skills, and has shown too little interest inthe emergence of, and changes in, such abilities. The represen-tational redescription (RR) approach, therefore, is a timelyoffering, and one which cognitive scientists would do well toheed. The emphasis on progressive modularisation is also help-ful. It offers a route which avoids the pitfalls associated withmore extreme positions, such as a too rigid nativism.

It would have been interesting had Beyond modularity(henceforth, Modularity) explored the relation between the RRmodel and the development of, and ongoing change in, thehuman memory system. Specifically, how would Karmiloff-Smith envisage, if at all, the relation between the RR model andthe putative subsystems of memory: procedural, semantic, andepisodic? The question is pertinent in that early views claimingthese subsystems to be functionally dependent.(e.g., Tulving1983) have been superseded by accounts emphasising theirfunctional and developmental dependence. For example, Tulv-ing (1985) has claimed that procedural memory may be develop-mentally and ontogenetically prior to semantic memory, out ofwhich, in turn, episodic memory derives. It is interesting torecast this account in terms of the RR model so that behaviouralmastery and possibly level-I description are equated with pro-cedural memory, and progressively higher levels of RR withsemantic and episodic memory.

An example might make this clearer. Note the differencebetween a young child learning to form and use correctly aprocedural memory such as the name "Clancy" in the presenceof a new dog. Presumably this requires only simple behaviouralmastery. The (semantically based) ability to remember theproposition that "the name of the dog is Clancy," then derivesfrom this simple skill via RR. Finally, the full episodic recollec-tion of the circumstances, mental stance, and feelings associatedwith learning this new fact, and possibly those of the otherparticipants in the situation, occurs following further RR. Inter-preted thus, a full episodic memory appears to require quiteextensive RR, including, perhaps, the formation of a representa-tion which refers back to the original source modality of theinformation received ("I recollect hearing/reading that the dog'sname is Clancy,") and to the environmental, psychological, andsocial context, including the mental state of the recipient andprovider of the information. The developmental emergence ofthese memory skills could be worth further investigation in lightof RR. Also, in a full recollective experience the remembererswould seem, at times, to need to adopt complex prepositionalattitudes towards their own and others' previous mental states:"I remember that I wrongly believed that the dog was calledRover, but I was told, truthfully, by Y, that its name was Clancy,"and so on. Comment on these issues would be interesting.

I strongly agree with the views on connectionism expressed inModularity. Karmiloff-Smith is probably right to claim that mostcontemporary models do not go beyond level I in their level ofrepresentation. She is also right that connectionism could beprofitably used to examine the development, emergence, andchange of cognitive abilities. In our laboratory we have devel-oped simple models which track the growth of various inferen-tial skills (Barnes & Hampson 1993a; 1993b). However, it is truethat any complex representations they form are still at level I.Also, in our models we had to supply relevant abstract structuralinformation about the incoming stimuli that in a more developedmodel would be extracted through a process of RR. We have,however, demonstrated developmental shifts in the behaviour

of connectionist systems and have described such effects as"learning to learn."

Genes, development, and the "innate"structure of the mind

Timothy D. JohnstonDepartment of Psychology and Center for Critical Inquiry in the LiberalArts, University of North Carolina at Greensboro, Greensboro, NC27412-5001. [email protected]

In Beyond modularity, Karmiloff-Smith combines elements oftwo kinds of explanation for cognitive development: one (repre-sented by Piagetian constructivism) that attributes the emer-gence of structure and order to the influence of an orderedenvironment, and another (represented by Fodor's modularity)that attributes this emergence to the unfolding of preexistingorder in the genes. Her argument is that although children'sexperiences with their physical, biological, and social environ-ments play an important role in the development of thesecapabilities, they also bring preexisting capabilities to the task -native predispositions that can not be explained by appeal toPiagetian kinds of processes.

A major theme of the book is a plea for "taking developmentseriously" (the title of its first chapter); and the need for asynthesis of nativism and constructivism is a prominent andrecurrent aspect of that theme. In this commentary, I want tofocus on the conceptual underpinning of the book's develop-mental theme, in particular on the paradox of using nativism as adevelopmental construct in the way Karmiloff-Smith does.

The nativist strategy is a common one in psychology. Thoughit takes a number of forms, its basic structure is to partitionaspects of the phenotype (whether behavioral or cognitive) intotwo classes, attributing one of these to extrinsic sources ofinformation (typically learning or other environmental influ-ences) and the other to intrinsic sources (in modern discussion,typically the genes). The mechanisms by which learned ele-ments of the phenotype develop are generally spelled out inconsiderable detail, but nativist accounts are usually silent onthe question of how the other (innate) elements come intobeing. The strategy has been criticized by several generations ofdevelopmental theorists (including, among many others, Car-michael 1925; Dent 1990; Gottlieb 1976; Johnston 1987; Kuo1921; 1924; Lehrman 1953; Oyama 1985) on the grounds thatinvoking the concept of innateness amounts to abdicating re-sponsibility for explaining development. In various ways, all ofthese theorists have argued that when we describe a behavior ora cognitive ability as innate we are simply declining to give adevelopmental explanation for it.

Although Karmiloff-Smith claims to be making more carefuland subtle use of the concept of innateness than other nativisttheorists (such as Fodor), she really has as little to say about thedevelopment of innate competencies as anyone else. Thus,having said (on p. 5), "When I use the term 'innately specified' inthis book, I do not mean to imply anything like a geneticblueprint for prespecified modules, present at birth," she goeson to say, "Nature [presumably she means the genes] specifiesinitial biases or predispositions that channel attention to rele-vant environmental inputs, which in turn affect subsequentbrain development." This is only a difference in degree: Fodor iswilling to attribute more cognitive structure to innateness thanis Karmiloff-Smith, but neither has anything to say about howthat innate structure develops. If we are going to requirecognitive scientists to take development seriously (and I fullyagree with Karmiloff-Smith that we should), we can not do it byproviding a nondevelopmental category of phenomena labeled"innateness," into which we place any capability that our theoryof development cannot explain.



It might be argued that to call something innate is not to avoida developmental explanation. Surely innate capabilities areexplained by appeals to developmental mechanisms involvingthe genetic regulation of development. Indeed, Karmiloff-Smith alludes to such explanations in several places in her bookand the implicit promise is that, if we want a developmentalexplanation of, for example, the 4-month-old infant's under-standing of physical principles such as gravity and solidity(p. 72), it can be provided by some theory involving the geneticcontrol of development. There is, of course, no question thatgene activity plays an important role in development, includingthe development of the cognitive capabilities that Karmiloff-Smith discusses in her book. If we really want to take develop-ment seriously, however, we shall have to give an account ofhowgene activity contributes to cognitive development; designatingcertain features of cognition as innate does not accomplish that.(Incidentally, interposing brain circuitry between the genes andcognition does not alter the nature of the problem. All cognitivedevelopment must entail some underlying change in the infant'sphysiology, and that change is in just as much need of adevelopmental explanation as the cognitive change it makespossible.)

Just how we are going to provide an account of the role of geneactivity in cognitive development is far from clear, but anapproach that attributes cognitive (or underlying neural) struc-ture to information in the genes is certainly not going to work.Explaining the development of even relatively simple mor-phological features by appeal to genetic information (or geneticcontrol, or genetic programs) turns out to be immensely prob-lematic (e.g. Nijhout 1990). Using that strategy to explaincognitive development simply obscures problems that are manyorders of magnitude more complicated. In part, it is that verycomplexity that encourages talk of innate cognitive structures.The task of providing a sufficiently elaborated model of brain-cognition relationships at the cellular and molecular level (thelevel where gene activity actually occurs) appears so challengingthat it might be decades before we know enough to speak inother than metaphorical terms about genetic contributions tothe development of the mind. As Nijhout (1990, p. 441) pointsout, however, the use of the kinds of genetic metaphors impliedby talk of innate or "genetically programmed" structures leads"to highly distorted pictures of developmental processes" (seealso, Johnston 1987; Oyama 1985).

Even if we are currently unable to specify in useful detail howgene activity contributes to the development of cognitive as-pects of the phenotype, I would argue that we should search fora way of framing the problem that may eventually permit itssolution. History shows us that classifying aspects of the phe-notype as innate or genetically determined retards rather thanadvances the search for solutions to developmental problems.For example, the tendency to equate innate with both "presentat birth' and "genetically specified" discourages investigation ofthe prenatal development of the behavior or capacity in ques-tion. I found no reference in Karmiloff-Smith's book to prenataldevelopment, although the capabilities she describes as innatemust have their origins in this period (e.g., Fifer & Moon 1988).

This willingness to accept a nondevelopmental construct likeinnateness as part of a developmental framework for under-standing infant cognition mars what is otherwise a fine book.Karmiloff-Smith is to be congratulated for organizing a wealth ofdata on infant cognition into a clearly focused theoretical presen-tation. My own reservations about some aspects of that presenta-tion notwithstanding, I plan to recommend the book highly tomy students and colleagues and anticipate that it will generatelively and productive discussion.

The power of explicit knowing

Deanna KuhnDepartment of Psychology, Teachers College, Columbia University, NewYork, NY 10027. [email protected]

Abstract: The contribution of Karmiloff-Smith's book lies not in itsbearing on the debate over generality and specificity, but rather in itsmessage that development does not stop with the first rudimentarymanifestations of cognitive competency. Explicit knowing is at least assignificant as knowing in more implicit forms, i

Despite its title promising to take us beyond modularity,Karmiloff-Smith's book is likely to be received on many frontslargely as more support on the side of domain-specificity, in adebate that has preoccupied developmental psychology forsome time now. As a result, the book's real contribution may beneglected. Karmiloff-Smith's stance in the specificity-generalitydebate is a sensible one that has been adopted by many develop-mentalists: a domain-general process governs developmentalchange that takes place largely within domains. Yet, for anytheory that posits an interaction of domain-general and domain-specific processes to be useful, we are going to need some moreprincipled way to define a domain than a "specific area ofknowledge,' the customary vague definition that Karmiloff-Smith does not try to improve upon. We also need to ask to whatdegree, and how, developmental processes taking place withindomains affect one another, a question that remains to beaddressed by comprehensive empirical data. Karmiloff-Smithrecognizes that this question is about process and must beexamined by methods capable of addressing process questions,but this book presents no new data to address it.

Instead, the important contribution that Karmiloff-Smithmakes with this book is to highlight a critical and neglectedaspect of what develops (understanding the how is still to come).There are multiple forms of knowing, she stresses, and the lion'sshare of our research attention has gone to implicit, proceduralforms of knowing, rather than to explicit, reflective knowing.This focus has led to a research thrust that Karmiloff-Smithlaments, one that has produced earlier and 0arlier exhibitions ofcompetence in more and more rudimentary forms. Many ofthese studies are testaments to experiment0r ingenuity as muchas subjects' competence. They extend the data base that must beaccommodated in a developmental account, but, as Karmiloff-Smith aptly notes, they do not offer any account of how thedisplayed competence arose or how it will be elaborated. Inshort, they are not developmental studies. Karmiloff-Smithmakes her view clear - that our attention should be focused ondevelopmental process.

Karmiloff-Smith appropriately conceives of a continuum be-tween implicit and explicit knowing, and a major aspect ofdevelopment is the gradual "explicitation" of implicit knowl-edge. Still, I think she does not give enough emphasis to what Iwould claim is a major qualitative shift that occurs in the course ofthis process. Although Karmiloff-Smith places much emphasison children's "theory building" as a means of cognitive advance,she does not stress the critical development from thinking withone's theories to thinking about these theories. Although thisshift cannot be tied to a single point in development, it is no lesssignificant than the two very early shifts in representationalcapacity that she identifies (at approximately 18 months and 4years of age. It also documents well the enhanced flexibility thatKarmiloff-Smith links to the developmental process she calls"explicitation." It is this flexibility that enables theories tobecome objects (rather than only vehicles) of cognition.

The increased cognitive power this advance entails is hard tooverstate. Only with this metacognitive advance are we in aposition to know what we think, to exercise control over what webelieve and why. It is an advance especially interesting todevelopmentalists because of its long developmental course.The recognition of assertions as the product of (and therefore


Commentary/K.armi\off-Smith: Beyond modularity

relative to) a knower's mind is critical in this development. As aresult, these become subject to scrutiny within a framework ofalternatives and evidence.

Such recognition does not come as a single, sudden insight.Metacognition, like cognition, is not a zero-one, present-absentphenomenon that emerges in full bloom at a particular point indevelopment. Relevant understanding is observable in rudi-mentary forms within the first few years of life. The 2-year-oldwho calls her parents into her bedroom with the claim that it is aghost in her closet that is the cause of a soft "whooshing" noisethat is keeping her awake understands as well as her parents thatopening the closet door will provide the evidence capable ofdisconfirming this causal hypothesis. Similarly the 4-year-oldwho recognizes that the candy believed to be in the cupboard isin truth elsewhere has made at least a rudimentary differentia-tion between what a mind theorizes to be true and informationfrom the external world that bears on this theory. False beliefs,by definition, are subject to disconfirmation by evidence.

Yet into and throughout their adult lives, people remainchallenged to treat their beliefs as objects of cognition to anextent enabling them to be aware of the bases of their beliefs, tobe able to justify them to themselves and others, and to distin-guish firmly the implications of evidence from what they believeto be true. Many adolescents and adults, as well as children,exhibit serious deficiencies in these respects (Kuhn 1991). Theytreat their beliefs as self-evident statements of the way the worldis, rather than as claims to be examined within a framework ofalternatives that compete with them and evidence that bears onthem. The social impact is one of individual minds unable toengage in discourse in ways involving more than a static confron-tation of opposing claims. A further implication of this weakmetacognitive development is a lack of control of the interactionof theories and evidence in the individuals own thinking.Although thinking with theories is essential for coming to knowthe world, it does not allow any control of the revision thesetheories undergo as new evidence is encountered. These all toocommon developmental failures should be a prime concern ofdevelopmental psychologists, as well as educators, and shouldmotivate us toward further understanding of the developmentalprocess Karmiloff-Sinith calls "explicitation. '

Beyond methodological solipsism?

Michael LosonskyDepartment of Philosophy, Colorado State University, Ft. Collins, CO80523. [email protected]

1. Introduction. Karmiloff-Smith maintains that innate anddomain-specific predispositions are developed by "a dynamicprocess of interaction between mind and environment" (p. 9).During development "the environment acts as inuch more thana trigger . . . it actually influences the subsequent structure ofthe brain via rich epigenetic interaction between the mind andthe physical/sociocultural environment" (p. 15). This is one ofthe ways in which Karmiloff-Smith moves beyond Fodorianmodularity. Does she also move beyond another thesis associ-ated with Fodor, namely methodological solipsism (Fodor 1980;Putnam 1975)? She does not address this issue explicitly, but heraccount suggests that the answer is "yes.' After a brief discussionof solipsism I will highlight some antisolipsistic features of heraccount.

2. Methodological solipsism. This is the view that "no psycho-logical state, properly so called, presupposes the existence ofany individual other than the subject to whom the state isascribed" (Putnam 1975, p. 220). This suggests the autonomythesis: "Any differences between organisms which do not mani-fest themselves as differences in their current, internal, physicalstates ought to be ignored by a psychological theory" (Stich1983, p. 164). This also suggests that for the science of mind

"there is no necessary or deep individuative relation" betweenbeing in a psychological state of a certain sort and "the nature ofthe individual's physical or social environments" (Burge 1986,p. 4).

Fodor takes this to imply that psychological states are individ-uated only by their formal or computational properties withoutrespect to their semantic properties (Fodor 1987). He assumesthat the computational states that play a role in any adequateexplanation of individual behavior do not include environmentalstructures. He assumes, in other words, that the individualembedded in the environment does not make up an integratedcomputational system (Davies 1991). Karmiloff-Smith's workseems to undermine this assumption.

3. Procedural representations. The first phase of developmentinvolves implicit "procedural representations" that sustain thestructure of behavior (p. 17 ff). Karmiloff-Smith's discussion ofprocedural representations is sparse. One illustration she offersby way of explanation is Elman's (1991) recurrent network forgrammatical structure. She suggests that the weight spaces ofhidden units encode information, but the information is onlyimplicit because it is not available to that network (p. 185). Onlythe second phase would encode this information in explicitrepresentations that can be manipulated by the system.

This illustrates how representations are implicit in the firstphase, but it does not explain what is procedural about pro-cedural representations. How do they encode information in the"form of procedures" (pp. 20; 161)? Karmiloff-Smith's discussionof particular cases and examples suggests that these procedureswill sometimes involve environmental structures.

4. Examples. She writes that her initial solution to the Rubik'sCube puzzle was "sort of proprioceptive" and it was "informationembedded in an efficient problem-solving procedure" (p. 17).These claims need elucidation. Where and how is the informa-tion embedded? What is the problem-solving procedure? Astraightforward answer is that her interaction with the Cube isthe problem-solving procedure and the information is embed-ded in this interaction. Without the discrete states of the Rubik'sCube playing a role in the procedure, she would not have aproprioceptive solution or a representation of a solution.

Another example is the microstructure of youngsters' behav-ioral mastery of gravity and torque. When 4 and 5-year-olds areasked to balance different blocks with different points of bal-ance, they "pick up each block, move it along the support untilthey feel the direction of imbalance, and correct the positioningof the block by using proprioceptive feedback about the direc-tion of fall until the block balances" (p. 84). The procedure that issustained by the procedural representation involves a sequenceof both states of mind and external physical states. Developmentfrom behavioral mastery to representational redescription alsoinvolves "externalized markers which act as a sort of cognitiveprop to sustain the internal processes" (p. 138).

Assigning a role to externalized computations is consistentwith procedural representations being "internal to the mind"(p. 139; p. 140, n.l). The internal representations that sustainbehavioral mastery are embedded in a procedure that can becharacterized as a feedback loop consisting of a sequence ofinternal mental states as well as external states of the environ-ment. The procedural representation depends on there beingthe appropriate sequence of external and internal states.

5. Computational Interaction. Karmiloff-Smith touches onthese issues when discussing Reddy (1991), for whom an infant'sknowledge of teasing and humor is situated or embedded in theinteractional framework that includes the child's internal repre-sentations as well as the interaction with a laughing adult.Karmiloff-Smith grants that this "may well be true of earlymoments of knowledge acquisition," but she maintains that"ultimately, knowledge is represented in individual minds"(p. 123). The adult's external laughter may in the early stages"change and complete the child's representation," but it is"gradually incorporated into the child's internal representation



to subsequently mark the humor explicitly" (p. 123, emphasisadded). So at first, the complete procedural representation mayinvolve internal as well as environmental structures.

This takes us beyond solipsism in the early stages of develop-ment. Later explicit representations also will not be solipsisticinsofar as they are individuated by their epigenesis and theirfunction to make explicit information implicit in the earlierinteractional framework. However, they are solipsistic in thatthey are completely in the head and supervene only on currentinternal properties of the organism. This suggests that theautonomy and individuality of psychological states, as well as themodularity of mind, is a product of development.

Computational accounts of interaction with the environmentare defended by Rutkowska (1984; 1988; 1993b), who citesKarmiloff-Smith's earlier work with approval. For Rutkowska,interaction is best understood computationally, and "the notionof rule-governed structure manipulations must be taken toinclude environmental as well as intra-subject structures"(Rutkowska 1984, p. 1). The environmental structures are state-spaces that include the agent, the environment, and operatorsthat link the state-spaces (Rutkowska 1988, pp. 76-79). Kaelb-ling's work on autonomous embedded systems (1993) offers anexact model for understanding such interaction. The algorithmsfor learning behavior she offers are loops that involve theexternal world's state transitions as well as internal informationalstates of the system (Kaelbling 1993, p. 24).

Representational change, generality versusspecificity, and nature versus nurture:Perennial issues in cognitive research

Stellan OhlssonLearning Research and Development Center, University of Pittsburgh,Pittsburgh, PA 15206. [email protected]

Three perennial issues in cognitive research appear as thethemes of Karmiloff-Smith's book on cognitive development:representational change, generality versus specificity, and na-ture versus nurture.

1. Representational change. Brunei- (1966) suggested thatknowledge moves from enactive to iconic and into symbolicrepresentations. Before him, Piaget (1985) postulated a processof reflective abstraction and Bartlett (1932) wrote about "turningaround" on one's schemata. A process of restructuring is oftenclaimed to be central in cognitive development, particularly inthe development of scientific knowledge. How does Karmiloff-Smith's concept of representational redescription (RR) go be-yond these formulations?

A hypothesis about representational change ought to answertwo questions: first, how is a representation changed? Unfor-tunately, Karmiloff-Smith deliberately sets the question ofmechanism aside (p. 190). Second, why is a representationchanged? Karmiloff-Smith says that RR translates implicit (pro-cedural?) knowledge into explicit (declarative?) knowledge andthat this increases flexibility. However, it is unclear what shemeans by "flexibility" and why flexibility is improved by animplicit to explicit conversion. To illustrate the problem, flex-ible cannot mean "generally applicable," because an explicit(declarative) representation of a cognitive skill is not moregenerally applicable than an implicit (procedural) representa-tion of that same skill.

The most interesting thing Karmiloff-Smith has to say on thistopic is that representational change continues after behavioralmastery has been attained. This idea has not received muchattention recently, although classical learning research fromEbbinghaus onwards established that overlearning, that is,

continued practice after behavioral mastery, has strong effectson forgetting. Perhaps the need for postma^tery representa-tional change is the reason why educational programs that aimfor behavioral mastery tend to have poor results in terms ofconceptual understanding? However, the explanatory potentialof the RR hypothesis remains dormant as long as the functionand mechanism of RR remain unspecified. (The mechanism forthe opposite transformation from declarative to executable rep-resentations has been formally specified by several researchers;see, for example, Anderson 1982; Hayes-Roth et al. 1981;Ohlsson & Rees 1991.)

2. Generality versus specificity. Karmiloff-Smith joins Fodor's(1983) and Gardner's (1983) attempts to describe the mindin terms of semi-independent, special purpose modules. Themain questions to be answered by any theory that claims that themind consists of parts are Which parts? and "Why those parts?The chapter headings indicate that Karmiloff*Smith's answer tothe first question includes language, physics, mathematics,psychology, and notations. It is not obvious which dimension isused to distinguish these modules from one another. In Fodor'sand Gardner's proposals, modules are individuated on the basisof the types of information they process (linguistic, spatial,visual, etc.). "Physics" and "psychology," however, are not typesof information - they are aspects of reality. If these domains areprocessed by specialized modules, how about meteorology?Surely children have interesting ideas about the weather earlyon. And what about clothing, food, sports, toys - do they havemodules? Any attempt to divide reality into distinct modules isarbitrary. Furthermore, to put "language" an<;l "physics" side byside must be a category mistake, because language can be usedto talk about physics. To illustrate these difficulties, consider awritten story about the fright of a sky diver in free fall; is itprocessed by the language, physics, psychplogy, or notationmodule?

The most interesting thing Karmiloff-Smith has to say on thistopic is that modularity is not innate, but emerges as a result ofdevelopment and learning. This conjecture i$ supported by thefact that skills that are practiced over and over acquire a certaindegree of autonomy and insularity; witness the old Einstellungphenomenon as well as more recent work an automaticity incognitive skills.

3. Nature versus nurture. The third theme of Karmiloff-Smith'sbook is that purely empiricist and purely natjvist positions haveboth ceased to be intellectually viable. A, theory of humancognition must describe how innate cognitive structures interactwith experience. Knowledge is neither impressed on the mindby the environment nor imposed on experience by the mind; itemerges in the meeting between mind and world. This is theposition that Piaget argued for when he deplored the "experi-ence without structure" of the behaviorists as! well as the "struc-tures without experience" postulated by the Gestalt psycholo-gists. Although Karmiloff-Smith's thesis that human cognitioncan only be understood in terms of the interaction betweennature and nurture is unassailable, her book does little toadvance our understanding of that interaction.

4. Empirical grounding? The bulk of the book consists ofchapter-length reviews that are supposed to provide empiricalsupport for Karmiloff-Smith's hypotheses. It is not clear, how-ever, what counts as evidence for or against these hypotheses.What kind of observations prove that RR occurs? What kind ofevidence would prove that RR does not occur? What kind of datafavors the emergent modularity view over either the innatemodularity or general capacity views? In spite of the prepon-derance of empirical material over conceptual analysis,Karmiloff-Smith does not address such questions. As far as I cansee, most of the studies she reviews are as compatible with thehypotheses she proposes as with those she rejects.

5. Summary. Karmiloff-Smith proposes that knowledge un-dergoes representational changes during development (in par-



ticular, an implicit to explicit conversion), that these changescontinue after the attainment of behavioral mastery, that cogni-tive development produces semi-independent modules, andthat mind emerges in the interaction between innate structuresand experience. These are interesting and useful ideas whichdeserve the careful conceptual analysis and thorough empiricalgrounding that Karmiloff-Smith does not provide.

Where redescriptions come from

David R. OlsonCentre for Applied Cognitive Science, Ontario Institute for Studies inEducation, Toronto, Ontario, Canada M5S 1V6. [email protected]

Karmiloff-Smith puts a great deal of weight on what she dubs"representational redescription" (RR) as an engine of cognitivedevelopment. By means of RR a child can move from perceptualprimitives to image schemas and again from image schemas tolinguistic representations. RR is something that cognizers do totheir representations in the course of development.

What 1 see as missing from the account is the importance ofthe adults' provision of the categories or concepts in terms ofwhich lower-order representations are redescribed. I will com-ment on the development of the metalinguistic concepts wordand sound. Karmiloff-Smith points out that even 3-year-oldshave some implicit knowledge of word boundaries (and presum-ably of phonology) which they exploit in self-corrections. How-ever "they seem to know little if anything explicit about whatcounts as a word" (p. 51). Although they have some success injudging nouns to be words, when asked if the is a word, theyreply in the negative. They conflate, we may say, a word with athing. Representational redescription turns such functional en-tities as function words into a metalinguistic class of words.Karmiloff-Smith argues that this is a normal developmentalprocess, one of the things the mind does. Although she does notdiscuss phonemic awareness, that is, the ability to subdividelexical entities into phonemic constituents (if you removed the/(/ sound from /fish/ you would get /ish/), the same procedureswould presumably apply; with development and overlearning,language shifts from being a means of communication to beingan object of analysis, reflection, and redescription.

Although it cannot be denied that the cognizer is the one whodoes the redescription, my suggestion is that this is not aninternally generated or internally motivated action. It is rather amatter of using a model as a framework for reanalysis. In the caseof word, that model, I have argued (Olson 1994), is a script; onlyin word segmented scripts do we find clear, explicit representa-tions of words. Word is a problematic concept in Chinesebecause written characters do not correspond to words, butrather to morphemes. Furthermore, readers of syllabic scriptswhich do not mark word boundaries fail to acquire a concept ofword as a lexical entity.

The case for phonemic awareness is even clearer. Onlyreaders of alphabetic scripts are able to segment the flow ofspeech into the phonemic constituents represented (roughly) byletters of the alphabet (Morais et al. 1986; Read et al. 1986). Whyshould this be? Why does such redescription not take placenaturally and spontaneously? Again, my suggestion is that thescript provides a model for thinking about speech; the scriptprovides the model in terms of which the output of one system -the stream of speech - is redescribed or re-represented. Ratherthan a spontaneous or developmental achievement, it is thedirect product of learning to do something, in this case, to read.

The process may be illustrated by reference to the studies ofchildren's invented spellings reported by Read (1971). Charac-teristics of such invented spellings were the following:

nglish

day

lady

feel

bait

bet

bat

igloo

fell

Invented Spelling

DA

LADE

FEL

BAT

EGLIOW

FAL

Read rightly interpreted these findings as indications of chil-dren's implicit phonological knowledge. It is the redescriptionthat is of concern here, however. The children seem to becoming to hear their speech in terms of the phonemic categoriesoffered by the names of the letters of the alphabet.

Consider how this would work. All. the children in Read'sstudy knew the alphabet, that is, the names of the letters andhow to draw them. Their task, as they saw it (if I may speak forthese preschool children) was to interrogate their pronunciationof an utterance in terms of the letter names they knew. Thus,knowing that the letter a was called /ae/ they listened to theirpronunciation of such words as "day" and "lady" and detectingthe sound corresponding to the letter name, namely, /ae/, usedit in producing DA and LADE. The same is true for the majorityof consonants and for all the so-called tense vowels, for in thesecases the name of the letter corresponds to the sound it repre-sents. So too for "bait, bet, and bat," for which the sound of theletter name /ae/ led them to represent those sounds with theletter a, to produce BAT in all three cases. A slightly morecomplex story has to be told for the so-called lax vowels in wordssuch as "fell." Here, the children appear to "hear" the lax vowelas discrepant from the sound of the letter name /eel and closerto that of the letter a, resulting in FAL. That is, they hear theirspeech in terms of a similarity relation between the names of theletters and the sounds in their speech.

These are clear cases of representational redescription inKarmiloff-Smith's terms. Yet, far from being a spontaneousredescription, it is a matter of learning to hear and analyze anexisting linguistic product in terms of a model provided by someother system, in this case, writing. The English alphabetic scriptprovided the model - the set of constituent forms and sounds -in terms of which the children analyzed their speech. The unitsof speech they detect are not the underlying phonemes of theirlanguage (which remain implicit unless one becomes a phonolo-gist) but rather the sounds corresponding to the names of theletters of the alphabet. The writing system, then, provides amodel in the form of a set of categories in terms of which speechsounds are represented, thereby bringing those aspects ofspeech into consciousness.

Consequently, although representational redescription doesinvolve considerable intellectual resourcefulness and does pro-vide higher-order levels of representation, the extent to which itis "one of the human instincts for inventiveness" (p. 193) may beoverstated. It is perhaps more of an indication of a child's abilityto represent some aspect of a cognitive practice in terms of thecategories offered by such cultural artifacts as the alphabet.

Beyond modularity: Neural evidence forconstructivist principles in developmentSteven R. Quartz and Terrence J. SejnowskiHoward Hughes Medical Institute, Computational Neurobiology Laboratory,The Salk Institute for Biological Studies, San Diego, CA [email protected]

Karmiloff-Smith has outlined an exciting new developmentalperspective on the nativist debate. She cites evidence for brain



plasticity in infancy against initial modularity, but holds onto thedomain-specific predispositions of nativism. We think her con-cessions to nativism depend on still viewing modularity asessentially a psychological theory. During the last several years,however, advances in our knowledge of how brains develop hasprovided an additional source of constraints on the issue ofmodularity that has largely been overlooked by the psychologi-cal community. Although this new evidence points in the con-structivist directions Karmiloff-Smith advocates, the psycho-logical theory consistent with these data suggests a strongerrejection of nativism and the modularity thesis.

Since modularity is ultimately a thesis about how informationis represented in the brain, it strikes us as natural to look atevidence from neurobiology as a source of support for thisposition, although we suspect that the functionalist belief thatneurobiological evidence undermines this issue may still holdsway. From a developmental perspective, the modularity thesisbecomes a question of whether some cortical regions are pre-specified for domain-specific knowledge that determines a pri-ori the computation a region may perform. This appears to bethe case for many subcortical structures such as the retina andsuperior colliculus; however, recent neurobiological evidencehas established that this view is implausible for cerebral cortex.In addition these experiments suggest that it is the pattern ofafferent activity that determines regional identity and function.To cite briefly some striking examples, clonal analysis (e.g.,Walsh & Cepko 1988; 1992) has determined that clonally relatedcells migrate divergently across large regions of cortex, indicat-ing that patterns of cell migration do not specify the laterfunctional organization of cortex and suggesting that neuro-genesis produces a cortex that is initially functionally equipoten-tial. This has been supported by fetal transplant studies in whichsmall portions of cortex have been transplanted into differentregions with the result that the transplanted cortex developedthe features specific to the new region (e.g., Schlagger &O'Leary 1991). Cross-modal rewiring experiments, in whichprojections of one sensory modality have been rerouted intocortices normally receiving a different modality, also supportthis view (e.g., Frost & Metin 1985; Sur et al. 1988; 1990). Inthese rewiring experiments, the recipient cortex develops func-tional properties appropriate for the sensory modality of theafferent projection and not the properties it would have had itreceived its usual modality. The striking capacity of maturebrains to reorganize in response to changes in afferent stimula-tion (reviewed in Merzenich & Sameshima 1993) suggests thatmany of the same principles underlying plasticity in develop-ment and the dependence of cortical organization on environ-mentally derived patterns of stimulation may be lifespan phe-nomena.

These results are not limited to a passive imprinting of theworld on the cortex. Instead, contrary to the nativist view ofdevelopment as something that just happens to a system(Chomsky 1980; Piatelli-Palmarini 1989), neurobiology andcomputational modeling have come to recognize, as did Piaget,the importance of the active role of biological systems in devel-opment. For example, the computer view of vision as a systemthat simply extracts and processes information contained in thevisual signal, has been rejected in favor of a view in whichsensory representations are organized to serve action systems(Churchland et al. 1994). A developmental example of this is thecomputational study of Montague et al. (1993) that showed howvisually guided action (such as foveation) and its consequencescan guide map development in other sensorimotor systems.

In our opinion, these studies turn the modularity thesis on itshead. Not only is the brain not initially modular, but thedetermination of the information-processing properties of cor-tex depends crucially on the nature of the environmentallyderived activity cortical regions receive. The fact that variousregions of cortex receive different patterns of afferent termina-tion and activity appears sufficient to explain the development of

specialized cortical functions. If so, then does this capture whatKarmiloff-Smith calls "prespecified dispositions"? If it does,then this seems barely to be a concession to the nativist, as itneed not suppose that the recipient cortex anticipates anythingin terms of domain-specific knowledge. Ratli0r, it is the differingpattern of afferent activity, reflective of different sensory modal-ities, that confers area-specific properties onto the cortex - notpredispositions that are somehow embedded in the recipientcortical structure. Furthermore, the fact that many of theseprocesses operate before birth, as in the case of spontaneousvisual activity (Maffei & Galli-Resta 1990; Meister et al. 1991),suggests that cortical specification could begin by the verymechanisms that will be used postnatally through interactionwith an environment. This two-staged approach, divided be-tween pre- and postnatal periods, has been shown in a computa-tional model to account for the development of two distinctclasses of cells in visual cortex, monocular and binocular cells,and their disparity selectivity (Berns et al. 1993). These featuresarose because the model underwent two distinct periods ofdevelopment, where the first period set a number of importantparameters that then constrained postnatal development. Thenewborn could, therefore, already display skills that did notdepend on any prespecification in the sense of domain-dependent knowledge built into recipient cortical structures.This staging of development may also serve as a foundation forsubsequent development, which would be under the influenceof environmentally derived activity.

Viewing the issue of modularity from a neurobiological per-spective, the nativist position and the related modularity thesisare highly implausible. This may appear t6 conflict with theprimary motivation for them: the results of'Vigorous" learningtheories demonstrating that complex skills cannot be learned(e.g., Wexler & Culicover 1980). It needs to be pointed out,however, that these results all depend on the assumption thatthe learning mechanism's resources are fixed. The underlyingappeal of constructivism lies precisely in its rejection of thisassumption and hence in having fundamentally different learn-ing properties (Quartz 1993). By extending the neural perspec-tive we have been urging, constructivism identifies these learn-ing processes in terms of structural growth mediated byenvironmentally derived activity. We are currently exploring anumber of such constructivist algorithms. We suspect that theinteraction between structural growth and its regulation byenvironmentally derived activity will be sufficient to explain theacquisition of mature skills without the need to posit domain-specific predispositions in recipient cortical structures.

Situating representational redescriptionin infants' pragmatic knowledge

Julie C. RutkowskaSchool of Cognitive and Computing Sciences, University of Sussex,Brighton BN1 9OH, England, [email protected]

Abstract: Attributing infant abilities to conceptual mechanisms, inpreference to (Fodorean) input perceptual systems or to a (Piagetian)deficit view of action, makes it harder to understand how representa-tional redescription might work. Mapping between conceptual abstrac-tion and representational "explicitation" is unclear. It is better to treatthe infant as a situated agent, whose pragmatic knowledge is groundedin a computational model of action.

Clarifying what representational format(s) could support younginfants' abilities is important to Karmiloff-Smith's proposals,insofar as understanding what representational redescriptionworks on is essential to understanding how it works. Karmiloff-Smith draws on research that concludes, contra Piaget, thatyoung infants' organization of stationary, moving, and interact-


Commentery/Karmiloff-Smith: Beyond modularity

ing objects goes well beyond literal sensory impressions. How-ever, thinking about what mechanisms might underlie thesenew data is heavily influenced by traditional views of "superior"functioning. Image-schematic representation is proposed as aform of conceptual structure generated from primitives extrac-ted by analysis of spatial structure (Mandler 1988). A centralconceptual system is thought responsible for organizing anunsegmented perceptual array into discrete objects on the basisof unchanging principles such as cohesion and boundedness(Spelke 1990). Conceptually grounded "belief in object perma-nence" is assumed (Baillargeon 1986). Such concept talk seemsto me to introduce more problems than it solves. Does the bestexplanation demand deciding between properties of isolatedperceptual (input) systems and conceptual knowledge that li-cences inferences about physical objects?

Given these options, Karmiloff-Smith chooses the latter.How, then, do infant mechanisms fit the levels of representationthat comprise the representational redescription model? Sincelevel-I representations are limited to mediating the context-bound input-output relations that underlie behavioral mastery,at least the El level might be expected. The increasingly flexiblefunctioning of E representations appears comparable to theflexibility and interrelatedness of knowledge that concepts havebeen assumed to enable through their predictability, that is,explicit representation of invariances as properties of things(Kirsh 1991), and systematicity, that is, syntactic combinatorialcapacity (Evans 1982). Karmiloff-Smith doubts such equivalenceas far as infants are concerned, however. Spelke's principles, shesuggests, are likely to be at level I, embedded in response toenvironmental stimuli. Baillargeon's demonstrations may re-quire image-schematic representation in Mandler's sense, butare not yet thought to be at the El level where flexibility isattributed to extracting components of representations for useoutside their original input-output context.

Such mapping difficulties suggest to me that young infants notonly operate without theories, as Karmiloff-Smith argues, butwithout concepts as philosophers generally characterize them.Mandlers work appears closest to demonstrating propertiesakin to "strictly" conceptual functioning; but the second half ofthe first year yields the clearest findings, for example, discrimi-nation between stationary objects of similar form along abstractline such as "self-motion" and "caused motion." Such tasks maybe tapping into mechanisms that are an outcome of some form ofrepresentational redescription, not its foundation. Key in-stances of younger infants' purportedly conceptual functioningmay be amenable to explanation in terms of a perceptualprocessing option, where perceptual processes are viewed asone component of an action system.

For example, Marr's (1982) ideas about the object knowledgethat is implicit in low-level vision's construction of a 2£-D sketchconverge with infants' successes and errors in parsing scenes andin prehension (Rutkowska 1991; 1993a; 1993b). This representa-tional format is assumed to make explicit the depth and orienta-tion of patches of visible surfaces and their discontinuities, anddiscontinuities in surface depth are crucial to the way infantssegment scenes into "unitary objects" (Spelke 1990). Marr'sproposals for algorithms converge with infants' grouping to-gether things that undergo common motion, even in the face ofunaligned major axes and differences in colour, texture, andshape. All levels of description produced by such (input system)processing may be involved directly in supporting purposiveactivity. For example, in this case infants reach for surfaceswhose boundaries are specified by separation in depth or bymotion (von Hoftsen & Spelke 1985).

If infants' object understanding was based in a Fodoreancentral system, change by enrichment of core principles wouldseem straightforward, but redescription to a qualitatively newlevel of representation would be difficult to envisage. It may beeasier to find a place in infancy for representational redescrip-tion by viewing the infant as a situated agent, concurring with

Karmiloff-Smith s emphasis on maintaining the Piagetian focuson output systems and the subject's action (Rutkowska 1993b).Then central processes are no longer directed at the fixation ofprepositional beliefs. Instead, they can be thought of as actionprograms: virtual, modifiable structures that underlie the func-tional coordination of perceptual and behavioral components ofaction in a supporting environment. Redescription can thenoperate at this computational level, altering its selective use ofperceptual and behavioral processes, hence the infant's contri-bution to control of action.

Applying this pragmatic perspective to infants' appreciation ofobject size-weight covariation generates an interesting example.When infants grasp and lift a series of objects whose size andweight are proportional, three distinct levels of reaction arefound to an inappropriately light "trick" object (Mounoud &Hauert 1982). Initially, they act as they do with "normal"objects, persisting with a local, one-off adjustment to the cur-rent circumstances. Subsequently, lifting is disrupted, for exam-ple, by rapid upward arm movement, and affective responsessuggest the detection of an anomaly. Finally, initial disruption israpidly compensated for. This pattern closely parallels the threephases of representational redescription that Karmiloff-Smithfinds in older children. Overall, we appear to see the infantmaking explicit, or becoming attuned to, constraints on success-ful action. This involves information in recurring patterns ofperceptual and behavioral activity becoming explicit at theaction program level, supporting anticipatory rather than reac-tive control. There appears to be abstraction of novel perceptualand motor variables from a range of local problem solutions. Theprocess is conservative, that is, initial mechanisms are supple-mented but not replaced by the development of anticipatorymechanisms.

Computationally rich models of action are needed to deepenour understanding of how viable patterns of activity can be fixedas permanent adaptive changes. Karmiloff-Smith relates thisissue to connectionism, asking how a connectionist networkcould appropriate its own stable states. My conclusion is thatsuch work needs to focus on situated (causally embedded)systems if it is to prove of import to epigenetic models ofrepresentational change. A promising area for computationalinsights may be research on autonomous agents designed forreal environment adaptive behavior (e.g., Maes 1990).

Redescribing development

Ellin Kofsky ScholnickDepartment of Psychology, University of Maryland, College Park, MD20742. [email protected]

Abstract: Beyond modularity describes how infants whose behaviors fitthe assumptions of cognitive science transform themselves into Piage-tians, with the aid of a redeseribing device formatted in computerlanguage but producing directed, qualitative change. The success of thisventure depends on paying closer attention to the nature of modules,the initial and final steps in development, and the redescriptive process.

Developmental psychologists who have abandoned the Piage-tian paradigm cannot find a new theory that respects his claimsabout the usefulness of developmental analysis. Piaget assumedthere was directional change towards an ideal end state whichconstituted an explicit, well-organized, and complete theory ofthe world. Validation of the end state depended on finding apopulation whose theoretical representations were incompleteand less cohesive, and then assessing their deficits in perfor-mance. Study of the factors that produced representationalchange provided information about the causal mechanisms un-derlying cognitive processing. Hence the legacy of the Piagetiantradition is the belief that developmental observation is a prime



method of testing theories of representation. Nativist ap-proaches underestimate change by characterizing the infant as afully formed adult masquerading in diapers. Connectionisttheories eliminate directionality because the end point is arbi-trary.

The demise of Piagetian theory posed yet another problem.Without commonality in processes or destinations, there is nodevelopmental analysis. The overthrow of Piagetian theory isanalogous to the dissolution of the Soviet Union. The currentcognitive map consists of fractionated territories, each inhabitedby processors who use specialized and untranslatable languageswhich prevent them from assimilating potentially useful contri-butions from other cultures. If each domain is acquired inunique ways and each representation is inaccessible to others,there is no development to describe, only a list of changes.

Beyond modularity is a response to these concerns.Karmiloff-Smith proposes a theory that maintains the definingcharacteristics of a developmental approach while linking it totwo seemingly incompatible paradigms: cognitive science andconnectionism. She tries to persuade the cognitive scientist thatdirected change exists and is revealing. She also suggests thatearly developments can be modelled by a network. Laterqualitative transformations towards an end state arise through areorganization of hidden layers during the course of learning.She tells developmentalists that attributing innate special-purpose biases and procedures to the infant does not precludedirected developmental change towards a more general andaccessible cognitive system.

Her bold strategy is to redescribe the end state and thechange process in terms compatible with cognitive science. Shethen translates the architecture of the cognitive processor intoan associative neural network. The redescription of develop-ment ought to appeal to many developmentalists because itcaptures elements of the Piagetian credo. Development movesfrom specificity to generality, from the implicit to the explicit,from stereotyped procedures to creativity, and from what Piaget(1978) termed "success to understanding." The path of develop-ment is constrained. Although the timing of change varies acrosstasks and environments, development always involves rewritinginitial representations into more flexible formats. The impetusfor rewriting is the need to reflect on accomplishments. Thus,the organism is not a collection of modules because there is animplicit, unified self, seeking to understand its own operations.

The reader of Beyond modularity will inevitably ask whetherthe rewriting is successful. Karmiloff-Smith may not convincecognitive scientists of the existence of development. The trans-lation of each code into the next level requires that the twolanguages be compatible but that the new code be more ab-stract, dissociable, and accessible. The cognitive scientist willdoubt that complex codes can originate from simpler ones.

Developmental psychologists acknowledge qualitative changebut note the anomalies of a fully developed redescriber existingin an organism working with primitive codes and a general-purpose redescriptive device coexisting with all-purpose mod-ules. They will also be dissatisfied with the vague analysis of thenature of modules, representations, and redescriptions.Karmiloff-Smith postulates innate constraints or attentionalbiases for domains as diverse as perceptual organization, mathe-matics, and drawing. She lacks a principled way to separatemodules immediately available to the infant from domainswhere attentional biases quickly produce encapsulated pro-cedural knowledge and from domains that are never mod-ularized. Yet a theory of the architecture of cognition must beprecisely situated.

Even in innately structured domains, the structure is oftenunderspecified. The infant intuitive physicist quickly under-stands that two objects cannot occupy the same place and thatdropped objects fall until they hit a surface. These notions guidethe baby's search for missing objects, but what is the representa-tion? Is it a search routine or an image schema? How is under-

standing formatted? Without answers to these questions we donot know whether and how the representations of the infantphysicist lead to adult knowledge of the world.

The first phase of growth translates implicit procedures intoexplicit codes. Without a rich model of the implicit and earlyexplicit codes we will remain ignorant of the redescriptiveprocess and difficulties of translation. This problem is seriousbecause in some areas, such as understanding the referents ofnouns, the initial biases are discarded for new principles.Moreover, the requirements of translation may differ so exten-sively across domains that it would be hard to assign translationto a common processor. Similarly, the conversion from implicitto explicit knowledge differs radically from tjie redescription ofrepresentations into verbal and conscious forms. In addition,not all knowledge reaches this final state. Karjniloff-Smith needsto specify the limits of cognitive transfer and conscious access tocognitive structure, as well as constraints on the format ofmature representations. Precision about the re-representationaldevice is important. The "redescriber" must be equipped withskills or knowledge that will allow rewriting from a simpler andqualitatively different format to the more abstract and moreorganized representation. Where does this knowledge comefrom? The role of culture and social experience as a source ofknowledge is underemphasized. She also offers a skeletal expla-nation for change. Mastery of a field may be necessary forchange, but it may not prompt further resnalysis unless re-description satisfies some goal.

Karmiloff-Smith's considerable contributipn is the use of thefindings and language of cognitive science to redescribe andcorrect Piaget's theory. Without introducing a new level ofspecificity, however, it is not easy to understand how cognitivechange can be represented in a language common to the pur-poses of cognitive scientists and developmentalists.

The challenge of representationalredescription

Thomas R. ShultzDepartment of Psychology, McGill University, Montrial, Ou6bec, CanadaH3A 1B1. [email protected]

Abstract: Representational redescription (RR) poses a significant chal-lenge to cognitive science; but Karmiloff-Smith underestimates theextent to which some current computational models already engage inRR. Moreover, a large part of the existing challenge is to produceconvincing psychological evidence that deserves tp be modeled. Finally,task constraints are essential for success in both psychological theorizingand modeling.

One of the most interesting aspects of Karmiloff-Smith's Beyondmodularity is her notion of representational redescription (RR).RR characterizes how representations change with develop-ment, becoming progressively more explicit and accessible toother parts of the cognitive system. She rejects a traditionalstage approach to development in favor of a recurrent system inwhich three phases characterize the mastery of each conceptualdomain.

While acknowledging that connectionist models have beenparticularly successful in capturing the implicit representationsand behavioral mastery of her phase 1, Karmiloff-Smith cor-rectly argues that these models generally have not made muchheadway in simulating the more explicit representations ofphases 2 and 3. However, one connectionist approach to model-ing cognitive development already implements a recurrentmultiphase RR process that might capture some of the datahighlighted by Karmiloff-Smith. Cascade-correlation is a gener-ative connectionist algorithm (Fahlman & Lebiere 1990) that hasbeen successfully applied to a variety of developmental domains



including the balance scale (Shultz et al. 1994b; Shultz &Schmidt 1991), seriation (Mareschal & Shultz 1993), predictionof effect size, integration of velocity, time, and distance cues(Shultz et al., in press), and acquisition of personal pronouns(Shultz etal. 1994a). Like other generative algorithms, cascade-correlation constructs a network topology as it learns a domain.For cascade-correlation, this construction entails the recruitingof new hidden units whose activations correlate with the behav-ioral error that the network is experiencing. Such newly re-cruited hidden units receive input from the network's inputunits and from any previously installed hidden units, thuseffectively redescribing developmentally earlier computations.Because high-level hidden units receive both raw descriptionsof inputs and interpreted descriptions from previous hiddenunits, they permit ever more sophisticated interpretations ofproblems in the domain being learned. Such cascaded hiddenunits afford the construction of increasingly powerful knowledgerepresentations that were not available to developmentallyearlier instantiations of the network.

Moreover, cascade-correlation goes through two recurrentphases that would appear to capture many of the phenomenacited by Karmiloff-Smith. The network begins in what is calledthe output phase, reducing behavioral error based on environ-mental feedback by adjusting output-side weights (those con-nection weights leading into output units). When this errorreduction stagnates, the network enters the so-called inputphase, where the focus shifts to building new hidden units. Theinput phase adjusts input-side weights to candidate hidden unitsso as to maximize the correlation between network error andcandidate unit activation. When these correlations level off, thecandidate unit with the highest absolute correlation with net-work error is installed into the cascade, just beyond the lasthidden unit. Then the algorithm reverts to the output phase, inwhich the network must adjust to this new representation of theproblem domain by again training the output-side weights.

Karmiloff-Smith's phases 1 and 3 appear to be analogous to theoutput phase of cascade-correlation. In both cases, there isconcentration on reducing behavioral error. In all but the firstoutput phase of cascade-correlation, there is also focus onreconciling new RR with performance demands, analogous toKarmiloff-Smith s phase 3. Her phase 2 appears to correspond tothe input phase of cascade-correlation. In both cases, one sees afocus on RR of current computation.

The transition between Karmiloff-Smith's phases 1 and 2appears analogous to the increase in error typically observed incascade-correlation nets just after the transition from inputphase back to output phase, following the installation of a hiddenunit that represents the network's output in a novel way. In bothcases, new representations temporarily interfere with previousperformance. The transition to Karmiloff-Smith's phase 3 islikewise similar to cascade-correlation's adjustment of output-side weights after hidden unit installation. Error decreases andthe network's representation of the problem being learned isbetter tuned than ever.

A major difference between the two approaches is that theredescription in Karmiloff-Smith's phases 2 and 3 is driven by anas yet unspecified analysis of earlier procedures. In contrast,both the input and output phases in cascade-correlation aredriven by the necessity to reduce error. Karmiloff-Smith maybeunderestimating the extent to which RR in children is motivatedby the need to reduce error on newly posed tasks that differ fromthe original behavioral task. Another difference between theapproaches is that standard cascade-correlation may not createthe explicit awareness that is often credited to children, forexample in Karmiloff-Smith's phase 3. It is unclear how orwhether connectionist networks can model such awareness.Nonetheless, available simulations display some important as-pects of RR and suggest that phase transitions in developmentmay occur continuously.

A second point is that much of the psychological evidence for

RR offered by Karmiloff-Smith could be attributed to eitherphase 2 RRs or to phase 1 behavioral adjustments. For example,the tendency to balance off-center-weighted blocks at theirgeometric center (pp. 84-87) would be a natural characteristic ofnetworks trained to balance large numbers of center-weightedblocks. Likewise, decreased reaction time on conservation tasks(p. 110) could just as easily be simulated by networks doing mereweight adjustment as by networks doing RR. Most cognitivemodelers believe that it is unproductive to successfully modelnonexistent phenomena. A significant part of the challenge ofRR must be borne by those purporting to document it psycho-logically. Verbalizations may be quite easy to document as RR,but nonverbal explicit representations seem to be more chal-lenging.

A final point concerns Karmiloff-Smith's criticism that con-nectionist efforts suffer from modeling individual tasks, ratherthan modeling development. This is unfair, because virtually alldevelopmental research, including that cited by Karmiloff-Smith, is restricted to individual tasks. At the present state ofthe art, task constraints are necessary for success, whether doingmodeling or empirical psychology. Surely the next reasonablestep is to deal with multiple tasks rather than venturing intotask-free development.

Modal knowledge and transmodularity

Leslie SmithDepartment of Educational Research, Lancaster University, Lancaster LA14YL, Great Britain. [email protected]

Abstract: Necessary knowledge is modal knowledge. The modal featuresof the phase model of representational redescription (RR) in Karmiloff-Smith's Beyond modularity are not squarely addressed. First, one mainepistemological problem is to explain the temporal construction ofatemporal knowledge. The RR model is silent here. Second, the RRmodel is primarily domain-specific. Yet the construction of modalknowledge is a universal, though not a general, process. Third, truth-value is distinct from modality, yet the RR model pays more respect tothe former than to the latter, even in its account of the construction ofnovel knowledge.

Necessary knowledge is one main form of modal knowledge.How does the transmodular model in Karmiloff-Smith's (1992a)Beyond modularity, (henceforth, Modularity) deal with thedevelopment of such knowledge?

Consider three features of the representational redescription(RR) model outlined in Modularity. First, this is a phase modeland so carries no implications about the simultaneity of age-related changes (pp. 6, 173). Second, the model operatesthrough reiterative cycles within a domain, where a domain istaken to be a set of representations bound to a specific area ofknowledge, such as number and language (p. 6). The RR processis stated to be domain-specific (p. 11). Third, the model charac-terises the capacity of the human mind to enrich itself (pp. 28,190). My argument will be that each of these three features hasmodal implications which are not squarely addressed inModularity.

1. Phase model. There may be stage models which make ageclaims (Demetriou et al. 1992) but Piaget's (1960) model is notone of them; his five stage criteria (constant order, overarchingstructure, integration, consolidation, and equilibration) do notinclude age. This point was not lost on Brainerd (1978). Age is anindicator, not a criterion, of developmental level (Smith 1993,sect. 18). Unlike a criterion, which must be exceptionless, anindicator has relative utility, for example, in sample selection oreducation. Modularity, denies that age claims have scientificinterest (p. 28), but exactly why such a position is incompatiblewith Piaget's model, which is stated to be "likely wrong" (p. 167),



requires clarification. The reason age is not a criterion is itsexplanatory inadequacy: how can temporal construction gener-ate atemporal knowledge (Smith 1993, sect. 1)? The age of onsetof modal understanding is simply irrelevant to the resolution ofthis paradox.

2. Domaln-speclficity. The transmodular theory in Modularityimplies that modal knowledge is domain-specific, its successfulconstruction occurring "first within a domain and then some-times across domains" (p. 18). It could be argued that modalknowledge is not constrained in this way for Kantian reasons (cf.Kant 1787/1933, sect. B4). First, modal concepts are universal;they are domain-invariant. Yet a transmodular theory relativisesany necessity to a specified domain. Note that number isfrequently regarded as a distinct domain: there may be evidencethat nonlinguistic "animals are capable of various forms ofnumerical competence" (Davis & Perusse 1988) or that humaninfants are innately equipped with some numerical knowledge(Wynn 1992b). Numerical knowledge could not count as modalknowledge unless it is also rationally legitimate and so applica-ble across domains. For example, number conservation requireschildren to realise that a proposition in the language domain(recall the protocol where Bon says there are "six" bottles and"six" glasses) is the same as the proposition "6 = 6" in thenumber domain (Piaget 1952, p. 43). According to Piaget (1977p. 50/1995), rational norms constitute a distinct domain, which isuniversal. A universal domain is not thereby general, however,because "it is well known in logic that 'universal' and 'general' donot mean the same thing" (Piaget 1977, p. 199/1995).

Second, the modal concept of necessity is a paradigm exampleof necessity. Since a proposition is necessary just in case itsnegation is impossible, it could have exceptions neither in theactual world nor in "possible worlds" (Sainsbury 1991). There is amajor difference between truth-value (a true proposition is notfalse) and modality (a necessary proposition could not be false).Young children do understand modal words (Byrnes & Duff1988), and they do acquire modal knowledge: in the task wherechildren were required to read a map upside down, a childcalled Pie remarks that "it always has to be the opposite" (Piaget& Karmiloff-Smith 1992a, p. 117). Note that Pie's remark is notdomain-restricted. How does a transmodular theory account forthe construction of such knowledge? If modal knowledge isdomain-specific, how can it be rationally legitimate knowledgesince a necessary proposition is true across "possible worlds"which are not domain-restricted? Yet if modal knowledge is aconstruction across a rational domain, this domain is universaland not domain-specific in the way outlined in Modularity.

3. Mental enrichment The RR model concerns a process whichgenerates genuinely novel knowledge, but "all knowledge . . .presupposes explicitly or implicitly principles of conservation"(Piaget 1952, p. 3, as amended in Smith 1993, p. 57). The point isthat modality is central to conservation (Smith 1993, sect. 16).Concepts are embodied in knowledge and conservation is re-quired by the Leibnizean salva veritate principle: "those con-cepts are identical if either can be substituted for the otherwithout change in truth-value" (adapted from Ishiguro 1972,pp. 17-21). This principle requires that pairs of descriptionsinvoke the same concept just in case both extensional equiva-lence and intensional coincidence hold. There can be neither anactual nor a possible case where the descriptions diverge, if theyinvoke the same concept. Where this principle is satisfied, oneand the same concept is conserved through its successive, andalways contextually different, epistemic uses. Where this prin-ciple is not satisfied, different concepts are in use, evident asnonconservation. A related issue is methodological and is rele-vant to the distinct norms relating to truth-value and modality.The RR model demarcates level E2 (conscious access withoutverbal report) and level E3 (conscious access with verbal report).This leaves open the question of the justification of the knowl-edge at level E2. The issue is not whether a justification isconscious or verbally accessed but rather how a reasoning with a

modal character can be rational in the absence of justification(Smith 1993, sect. 13). Finally, connectionist rnodels, which areinductive, are constitutionally incapable of providing a deduc-tively necessary product. In Modularity, connectionist model-ing is left an open question (p. 189) - could it be otherwise?

Is there an implicit level of representation?

Annie Vinter and Pierre PerruchetLEAD, CNRS 1838, Faculty des Sciences, Universit6 de Bourgogne,21000 Dijon, France, [email protected] [email protected]

Abstract: We suggest limiting the relevance of the representationalredescription model to the successive levels of explicit representations.Behavioral mastery either results from prior explicit representations, asillustrated by drawing behavior in children, or reflects some kind ofdirect sensitivity to the product of the rules structuring the world,without any embedded genuine knowledge of these rules, as shown byimplicit learning in adults.

Karmiloff-Smith offers the scientific community an excitingbook that addresses one of the main challenges in psychology:accounting for knowledge acquisition through the integration ofnativism and constructivism. Throughout the book, Karmiloff-Smith repeatedly illustrates a three-phase' iterative cycle ofdevelopmental change that takes behavioral mastery in onemicrodomain as "a prerequisite for passing from procedurallyencoded representations to the first level <̂f representationalredescription' (p. 157). Our comment is related to this notion of"procedurally encoded" or implicit representations; it is orga-nized around two main issues.

The first pertains to the extent to which some of the examplesof behavioral mastery discussed by the author involve onlyimplicit representations, elaborated at level I. For the sake ofillustration, let us consider the graphic domain. In Chapter 6,Karmiloff-Smith reports an experiment illustrating representa-tional changes in drawing behavior between 4 and 10 years ofage, and considers that the drawing routines efficiently run byyoung children aged 4-6 years correspond to level I of themodel. We would suggest that these routines are not assembledby means of local and implicit adaptations,, but rather are theproducts of a first representational redescription (RR) process,involving at least E l level (with explicit knowledge available toother parts of the cognitive system). We have studied in detailthis first phase of development (3-6 years) in the mastery ofdrawing (Vinter, in press). Around the age of 5-6, efficientdrawing routines are established, as indicated by three criteria:the presence of a quite rigid, ordered sequence of movements,the stability of the routine for each item, and sufficient precisionin the drawings to allow easy identification. However, theseroutines are structured according to common graphic produc-tion rules (such as starting at the top), which also characterizehandwriting but differ radically from the regularities observedin the very first drawing movements. Behavioral mastery thusseems to emerge as a consequence of a first RR process, duringwhich the graphic routines in elaboration must be permeable tothe influence of a "common internal rule generator." It is be-yond the scope of this commentary to discuss whether such arule generator is implemented somewhere in the cognitivesystem (symbolic solution) or emerges in the course of thedynamics operating inside the planning and executive systemsthat control drawing behavior (dynamical solution).

Early drawing behavior, although inefficient in terms ofcorrespondence to the model, nonetheless produces rudimen-tary drawings located in the graphic space beginning around 3years of age, thus showing a certain degree of mastery. Thisbehavior appears to be organized as a function of basic bio-



mechanical constraints affecting movement production. Theregularities observed at this level are indeed immediate ordirect consequences of global biomechanical constraints, suchas the tendency to prefer outward to inward movements. Whendrawing verticals for instance, such a constraint induces apreferential starting position at the bottom of the figure, as if animplicit syntactical graphic rule were in force. This last inter-pretation raises a legitimate question, however: to what extentdoes early drawing behavior embed implicit knowledge of suchbiomechanical constraints? This introduces our second issue:are the initial behavioral adaptations based on genuine repre-sentations that may further evolve into explicit knowledge?

Recent models and data from experimental studies of implicitlearning in adults provide some unexpected insights into thisquestion. In these studies, subjects are faced with situationsgoverned by complex, arbitrary rules, without being promptedfor an explicit analysis of the rules. They are asked to perform atask chosen in such a way that their performance testifies to theirknowledge about the experimental situations. In these condi-tions, performance improves, although subjects remain unableto articulate the rules governing the displayed material. InKarmiloff-Smith's terms, they reach some degree of behavioralmastery without concomitant explicit representations. Theearlier and still advocated position accounts for improved perfor-mance by positing that subjects implicitly abstract rules embod-ied in the experimental situations, a claim that echoes Karmiloff-Smith's claim that behavioral mastery embeds some kind ofimplicit knowledge.

A growing set of experimental data, however, shows that earlyadaptive changes are not due to the acquisition of an implicitknowledge base representative of the actual structure of thesituation. The initial improved performance is due to some kindof direct sensitivity to the product of the rules and not to theimplicit encoding of the rules themselves. For example, sub-jects may improve their familiarity with some particularly fre-quent fragments of the displayed material. Of course, thefrequency of the different fragments of the material is a directconsequence of the generating rules, but sensitivity to thisfrequency effect may be regarded as having no intrinsic relation-ship with the internal representation of the rules (Shanks & St.John 1994). A genuine internal representation of the rules maybe reached only when explicit, conscious processing is engaged(Perruchet, in press).

To conclude, we suggest that the relevance of the RR model islimited to the successive levels of explicit representations.When Karmiloff-Smith claims that explicit representationsemerge from the knowledge implicit in early behavioral mas-tery, she may be wrong for at least two reasons: behavioralmastery is grounded on prior explicit representations moreoften than she assumes, and, when no explicit representationsare available, behavioral mastery may reflect some kind of directsensitivity to the product of the rules structuring the world - aphenomenon which in no way embeds the genuine knowledgeof this structure.

ACKNOWLEDGMENTThis work was supported by the CNRS, the University of Bourgogne,and a grant from INSERM.

From the decline of developmentto the ascent of consciousness

Philip David ZelazoDepartment of Psychology, University of Toronto, Toronto, Ontario, CanadaM5S 1A1. [email protected]

Somewhere down the line that leads from Chomsky to Spelke toa recent cover story in Life magazine (July 1993), the notion of

development seems to have gotten lost. Admittedly the searchfor complex behavior and domain-specific predispositions ininfancy and early childhood has met with considerable success;surely some types of information processing reflect the functionof prespecified modules that operate on mental representations.However, too many psychologists have been playing a theoreti-cal endgame in which even infants are viewed as biologicallyprepared to reason in an adultlike manner, and the dramaticdifferences between the newborn "mewling and puking in thenurse's arms" and the 12-month-old, between the toddler andthe school-aged child, are downplayed or even, at times, ig-nored. Karmiloff-Smith should be applauded for promptingpsychologists to take development seriously. Her model ofrepresentational redescription (RR) goes beyond both simplemodularity and "adultomorphism" and begins to address thedevelopment of representational flexibility and consciously con-trolled behavior.

Karmiloff-Smith's triphasic RR model captures the insightthat an important type of knowledge acquisition and corre-sponding behavioral change involves increasing awareness thatemerges from the redescription of implicitly represented infor-mation. Although I am extremely sympathetic to this idea, and Ibelieve that the book as a whole is largely a success, it seems tome that several facets of the model need to be refined, and thatthe model needs to be located within an expanded conception ofthe growth of representational flexibility and control.

As Karmiloff-Smith has shown, aspects of the RR model canaccount for a variety of phenomena occurring at various ages.Additional examples may be adduced. P. D. Zelazo and Spin-azzola (1992) explored the phenomenon of relapse during re-learning in 24-month-olds using a visible-hiding A-not-B task inwhich children executed a three-step sequence to retrieve ahidden object. When the object was retrieved at one locationseveral times prior to being conspicuously hidden at a newlocation, children tended to search at the old location. Thisphenomenon has been explored in adults in a slightly differentsituation (Schwarz 1927; see Heckhausen & Beckmann 1990, fora discussion) and is probably fairly ubiquitous. People may havedifficulty analyzing initially implicit, procedural representationsinto components. A second example illustrates how neatly theprocess of RR based on behavioral mastery can account fornumerous age-related changes. Repeated elicitation of neuro-motor patterns such as stepping and sitting in 6-week-old infantsproduces potentiation that is specific to the patterns elicited(N. A. Zelazo et al. 1993), and this elicitation promotes theearlier intentional integration of these patterns towards the endof the first year of life (P. R. Zelazo etal. 1972). Mastery may leadto RR, which may in turn allow for integration.

Despite its success, however, the model fails to captureseveral characteristics of the development of cognitive control.First, although mastery may in some cases be a prerequisite forRR and concomitant flexibility, in other cases, it seems topromote rigidity. P. D. Zelazo et al. (1993) examined children'sability to switch between pairs of explicit rules. In one task,children were required to sort colored shapes first by (say) color(i.e., they were told, "If it's red, put it here, but if it's blue, put itthere"), and then by (say) shape. Children received either 1 or10 trials prior to the switch to the second set of rules. Three-year-olds perseverated on the preswitch rules even after a singletrial, but 4-year-olds were more likely to switch between rulepairs after one trial than after 10. These results disconfirm aprediction from the RR model. On the other hand, they areconsistent with the suggestion that flexibility and control re-quire resistance to interference from prepotent response ten-dencies (Diamond 1990). This idea echoes the oft-repeatedsuggestion that creativity in a particular domain declines withoverlearning (see Stemberg & Lubart 1992).

Research on rule use in preschoolers can also be used toillustrate two additional limitations of the RR model. First,3-year-olds can represent rules explicitly and use them to govern


Response/Karmiloff-Smith: Beyond modularity

their behavior (P. D. Zelazo & Reznick 1991). Nonetheless, asmentioned above, they sometimes perseverate on these explicitrules (cf. the phenomenon of ruminative thinking). The RRmodel is perhaps too rigid in its insistence that inflexibilitiesdepend on implicit, procedural representations.

Second, perhaps the most significant omission in the RRmodel is its failure to account for age-related, domain-generalchanges in reflection and the control of behavior. Douglas Fryeand I have found changes between 3 and 5 years of age on anumber of tasks, such as a ramp task, a lever task, and theory ofmind tasks, that have the same logical structure as the bi-dimensional card sort described above. Furthermore, there arepositive correlations among these tasks when age is controlled,despite the fact that they are drawn from different domains (e.g.,Frye et al. 1992; 1994; P. D. Zelazo et al. 1993). It would seemthat it is not until age five that children can construct a higher-order rule governing which of two incompatible pairs of rules toemploy. Similarly, the development of walking occurs in con-junction with a variety of other changes (P. R. Zelazo et al. 1989),and may be another example of integration that results from ageneral refinement of control structures. Perhaps processescaptured by the RR model interact with age-related increases inthe capacity for intentional control per se.

Changes in reflection probably occur in a variety of ways, andas a function of a variety of factors. The RR model would seem todescribe one important type of change. The general ability toconstruct higher-order rules that operate on extant rules is aform of reflection that yields flexibility and control, but need notinvolve RR. However, even if its range of application is some-what more restricted than Karmiloff-Smith maintains (e.g.,p. 16), a number of questions remain. First, what drives RR?Karmiloff-Smith's unexplicated suggestion that RR is "one of thehuman instincts for inventiveness" (p. 193) hardly goes beyondsaying that it just happens. One alternative that receives shortshrift is the possibility that redescription occurs in the course ofsolving problems that require information integration. Second,the process of RR itself needs to be clarified. For example, it isunclear whether the implicit representations themselves arebeing translated directly into new representations, or whetherRR occurs when one notices regularities in ones behavior. Also,the evidence for level-El representations (explicit but uncon-scious) is far from compelling; late-occurring errors (e.g., confu-sion of the numeral "un" and the indefinite article, pp. 56-7)could result simply from competition among alternative implicitprocedures, as these procedures are gradually generalized to awider variety of contexts. More important, the characterizationof these representations as unconscious could result from theconflation of consciousness with children's ability to articulatetheir knowledge.

These questions are constructive; part of the appeal of the RRmodel consists in its speculative nature. In fact, one of the mostimportant functions that Karmiloff-Smith's book serves is tobring issues to our attention. Unlike the sensorimotor infant ofold, today's infant is said to engage in remarkably sophisticatedforms of mental activity. Unfortunately, many implications ofthese claims are often ignored. By addressing the role of con-sciousness in representational change, Karmiloff-Smith has re-described at least one fundamental question that heretoforeremained embedded in implicit assumptions.

Author's Response

Transforming a partially structured braininto a creative mind

Annette Karmiloff-Smith ;Medical Research Council Cognitive Development Unit and UniversityCollege London, London WC1H OBT, United [email protected]

Abstract: This response discusses progressive modularity, the breadth ofthe notion of domain; whether Beyond modularity goes beyond Fodor,misrepresents Piaget, and gives sufficient prominence to the externalenvironment and to the early phases of learning; whether the represen-tational redescription (RR) framework has been fully extended to inten-tional psychology and notational competence; the problems of theimplicit/explicit dichotomy, of specifying mechanisms; the role ofsocio-cultural influences on representational change, (Jonnectionist models,and the use of the term "innate."

R1. Introduction. When I was invited tp prepare a Pre'cisfor peer commentary on Beyond modularity (henceforthModularity), I agreed in the hope that it would attractcritical review and suggestions from researchers in disci-plines outside my own. Indeed, one of the aims of Mod-ularity was to encourage the nondevelopmental cognitivescience community to go beyond the on-line processing ofsteady state systems and to take developmental changeseriously — developmental change at the cognitive/repre-sentational level and not merely at the behavioral level. Iendeavored to bring to the attention of the field specificissues that only a developmental perspective highlights,even if my particular speculations on developmentalchange may turn out to be only partially correct. Therepresentational redescription hypothesis (RR) was in-tended as a framework - rather than as a precise theory- for exploring possible generalities in developmentalchange across a variety of domains. My hope was that thepeer commentary would give rise to concrete suggestionsfor better specifying the concepts with which I have beengrappling, since modelling is not my field of expertise. Inthis respect, I have been somewhat disappointed in that,with a few exceptions, many commentators reiteratedwhat I had already stated in Modularity, by simply point-ing to the underspecification of my framework rather thansuggesting potential solutions from th^ir domain of exper-tise. On the other hand, I was delighted to read ofimportant new evidence either supporting or challengingRR (Bloom & Wynn, Freeman, Goldin-Meadow & Ali-bali, Rutkowska, Zelazo), of analogies and differencesbetween my framework and other theories (Campbell,Donald), as well as suggestions about additional represen-tational distinctions (Graham, Scholnick), together withpointers to more explicit models (Dartnall, Losonsky,Shultz) which are relevant to RR. Thfc peer commentaryon Modularity illustrates the huge diversity of existingviews on human development (compare, for example,Estes, and Goldin-Meadow & Alibali and Grush withCampbell and Ohlsson on the usefulness of the RRframework, or Johnston and Quartz & Sejnowski with deGelder and Foster-Cohen on modularity and nativism, totake just a couple of examples).

Owing to length limitations, it is obviously impossi-



ble to respond to every question that has arisen. I havetherefore drawn out a number of recurring themes acrossthe different commentaries. Before moving on, however,one general point: quite a number of commentators (e.g.,Bodor & Pl6h, Campbell, de Gelder, Foster-Cohen,Olson, Smith) seem to argue implicitly for. what I wouldterm an "either/or" view of development (e.g., it must beeither domain specific or domain general, RR must begenerated either by external pressures or by internalpressures, etc.) rather than contemplating the more plau-sible view that development is likely to be an intricateinteraction of several different coexisting solutions, aposition endorsed by other commentators (e.g.,Dartnall, Estes, Kuhn, Zelazo).

R2. Prespecified modules versus progressive modulariz-ation. In Modularity I placed a great deal of emphasis onthe hypothesis that where modules exist, they are theproduct of development, not its starting point. Estes andHampson find the notion of progressive modularizationuseful in avoiding the pitfalls of rigid nativism, but severalother commentators do not. First, de Gelder has seri-ously misread my claims. I do not argue that the notion ofa modular mind is not useful. I attempted to integratesome of Fodor's ideas because I believe he is right that theadult mind operates in part via a number of modularstructures, together with a number of domain-generalprocesses. Also, contrary to de Gelder's statement, myclaim is not that RR is the process by which all modulesare created, nor that human development undergoes aprocess of demodularization. On the contrary. RR oper-ates on the outputs of modules, obviously not on theinternal workings of the modules themselves, as well as onthe representations formed in nonmodularized parts ofthe brain. The process is conservative, however. Themodules themselves remain intact once formed and con-tinue to operate as modules after RR has taken place. RRis a process of redescription, not replacement. De Gelderseems to have conflated what I deem to be two distinct,parallel developmental processes: one of progressivemodularization, the other of progressive explicitness ofknowledge representations.

Quartz & Sejnowski are right that until recently muchof the psychological community has taken insufficientaccount of what is now known from neuroscience. Acareful reading of their commentary, however, shows thatat times they move too quickly from denying that pre-specification exists at the level of cortical structures todenying it of the brain as a whole. Although they arecertainly right about the equipotentiahty of some parts ofthe cerebral cortex, I find it difficult to accept that theinfant brain is totally unstructured, and I see no reason torestrict our discussion of predispositions to those thatmight be cortically specified. Certain behaviors dependcrucially on the nature of the environmentally derivedactivity that cortical regions receive, as they argue,whereas others depend simply on nonspecific input (Gott-lieb 1981). Moreover, if subcortical structures specifyinformation in the neonate brain (namely the discussion offace processing in Ch. 5 of Modularity,) then this gives theinfant brain a head start on what to pay attention to in itsenvironment and thereby represents a domain-specificpredisposition to development. It is on the basis of suchminimal predispositions that I argue for a process of

modularization. I certainly agree with much of the rest ofQuartz & Sejnowski's commentary and with their rejec-tion of initial modularity, but that is not the same as theprogressive modularization of certain parts of the brain.

Although some commentators were sensitive to thedistinction between modularity and modularization (e.g.,Estes, Ohlsson), the conflation does obtain in Foster-Cohen's commentary. She argues that the notion of mod-ularization is superfluous, although she discusses theissue only with respect to modularity. Biologically speak-ing, if the mind has any modular structure at all (a pointchallenged by Quartz & Sejnowski), then, knowing whatwe do about the brain's plasticity, progressive modulariza-tion, rather than prespecified modularity, is more plausi-ble. By definition, precoded encapsulated modules couldnot reconfigure themselves. By contrast, modularizationis a function of multiple levels of environment/organisminteractions. Where I find Foster-Cohen's commentaryuseful is in her insistence on the need to identify data thatwould differentiate domain specificity from modularity.Ohlsson raises a similar question about data distinguish-ing innate modularity from emergent modularity. I agreethat this often is not easy, but why opt for the lessbiologically plausible of the two options prior to findingclear-cut empirical distinctions? I would define as "do-main specific" a predisposition that could, in the absenceof appropriate input, reconfigure itself (as we know partsof the cortex can) to process other input, but it is not"modular" because of the very fact that it can reconfigureitself; it is neither of fixed neural architecture nor encap-sulated. A predisposition can become encapsulated as aproduct of progressive modularization. In Modularity Isuggest that future research on neonates and young in-fants using brain activation studies (ERP, PET, MRI, etc.)might offer such data. Foster-Cohen's discussion seems toconfound the on-line processes which may indeed end upas highly specialized and modular, on the one hand, and aprogressive developmental process, on the other. A plau-sible hypothesis discussed throughout Modularity is thatdevelopment starts out with processes that preferentiallycompute certain types of input but may be quite capable ofprocessing other inputs if the original one is not available(as is true of the congenitally deaf) and, with time andexperience, these processes become more and more spe-cialized, ultimately closing themselves off to other influ-ences. The predispositions can be at one or more differentlevels: the architecture, the learning algorithms, thespeed of change of parameters, the number of neurons orcircuits, and so forth. Progressively, with repeated pro-cessing of a specific class of input, these different levelsmay congeal to form a module dedicated to processingthose inputs. This ultimately gives rise to the rapid in-put/output typical of such processes in adults.

Scholnick rightly points out that Modularity lacks aprincipled way to separate domains where attention bi-ases rapidly produce encapsulated procedural knowledgefrom those which are never modularized. My only at-tempt to address this question was via the work ondiscourse processes discussed in Chapter 2. The datasuggest that when two processes compete for simul-taneous computational resources, one becomes mod-ularized while the other is open to RR and subsequentconscious reflection, but the mechanisms by which suchprocesses interact clearly require more detailed study.



R3. How broad is the notion of a domain? A number ofcommentators expressed concern about constraints onthe concept of a domain (e.g., Bodor & Ple"h, Foster-Cohen, Kuhn, Ohlsson, Scholnick). Bodor & Pleh askhow narrow or broad the theory-of-mind domain is, andwhether, for example, attachment relations and conceptssuch as "father-of/mother-of" are to be included. Here iswhere abnormal development can, in my view, play aparticularly important role. As discussed in Chapter 5,individuals with autism display a deficit in the attributionof beliefs to other people. They also show deficits inattachment relations. However, to my knowledge, theydo not have problems with concepts such as "parent of,"suggesting that relations like the latter are not part of thetheory-of-mind domain but involve general conceptualdevelopment. Another example can be gleaned from thecase of Williams syndrome (WS). A serious deficit in thearea of grammatical gender-marking in French-speakingWS subjects can coexist with relatively intact syntacticprocessing, suggesting that these two aspects of languageare processed differently. Depending on one's theoreticalleanings, this can be explained either as a deficit in an"agreement module" or as drawing on domain-generalmechanisms that are generally weak in these subjects. Insupport of the latter view, of the total population with thisdeficit, the only WS subject who showed good responseson the gender-agreement tasks was also the only subjectwith normal scores on an inspection time task - a domain-general capacity. Even though a single case is obviouslynot statistically meaningful, it is epistemologicallysuggestive.

Like Scholnick, Ohlsson rightly raises questions aboutthe parts of the cognitive system which might be said toconstitute domains. He answers, however, with the fol-lowing rather tendentious example: "consider a writtenstory about the fright of a sky diver in free fall; is itprocessed by the language, physics, psychology, or no-tation module [my italics]?" Why is this an either/orquestion? Surely it is clear that the products of manyprocesses must interact to understand such a story. Mod-ularity in no way implies that it should boil down to asingle one. Of course, Ohlsson is right in suggesting thatto place language and physics side by side could be acategory mistake, but this is only the case when oneconsiders language at its end point, as a communicativevehicle. Placing side by side the acquisition of knowledgeabout the linguistic system and the acquisition of knowl-edge about the physical world is not, in my view, acategory error. And as Ohlsson suggests, children mayhave interesting ideas about meteorology, but this clearlydoes not entail a "meteorology domain." Such ideas areexplicitly formulated and call on domain-general pro-cesses. As Kuhn reminds us, domain specificity is aquestion about process; explicit ideas about meteorologyobviously call on domain-general central processes.

R4. Beyond Fodor? It was not my main intention inModularity to defend or reject Fodor's nativist modularitytheory. Rather, it was to see what, in this influentialtheory, could be usefully salvaged while keeping a trulydevelopmental perspective in mind. On this generalissue, Bodor & Pleh, Hampson, and Zelazo welcome thefact that, contrary to Fodor, I stress the importance of adevelopmental perspective. Carassa & Tirassa, by con-

trast, challenge the multiple levels of the RR frameworkand opt for a Fodorian single Language of Thought (Fodor1976). If all the products of modules were automaticallytranslated into a common LOT, however, then I fail to seewhy it takes so long to translate knowledge expressed ingesture into knowledge expressed in the vqrbal mode (seeGoldin-Meadow & Alibali; Clements & Perner 1994).

Campbell argues that if we take a cons|ructivist view,we must simply get rid of any Chom^kian/Fodorianviews, but he fails to reconcile this with a critical discus-sion of some of the known constraints on syntactic pro-cessing referred to in Modularity. Quartz & Sejnowskitake a similar view in rejecting any form of modularity, butas mentioned above, they do not really discuss my argu-ments for progressive modularization in my attempt to gobeyond Fodor. De Gelder, by contrast, argues that weshould go even further than Fodor and opt for an almosttotally modular view of the entire brain. I disagree in thatthe more modular the human brain, the more it resem-bles the brain of species which engage only in relativelyrigid, special-purpose behaviors. How, under de Gelder'svery strict prespecified modular view, does the humanbrain achieve flexibility and creativity?

Rather than focus on the modularity debate, as I did inModularity, Bloom & Wynn focus on another aspect ofFodor's theory: his arguments regarding the incapacity ofconstructivism to explain the human mind. If I haveunderstood their commentary correctly, with Foster-Cohen they share Fodor's view that various aspects oflanguage are innately specified and are in no interestingway constructed. By contrast, Bloom & Wynn argue thatFodor's reasoning cannot be extended to number orconceptual development, which both involve the con-struction of novel structures. They argue that RR is notpowerful enough to account for theory change and citeCarey's (1985) and Gentner's (1983) alternative solutions.In Modularity I do not deny the important role of analogyand theory reconstruction in understanding how newrepresentational and conceptual structures could de-velop. On the contrary. Nowhere in Modularity do Isuggest that RR is the only source of change. But I doargue that processes such as analogical mapping cannotoperate on implicitly defined representations; they mustoperate on explicitly defined representations.

The process of RR, then, provides the reconstructiveand analogical mechanisms invoked by Carey and Gent-ner with the right representational formats on which tocarry out their mappings. In other words, RR acts as abridge for differently formatted pieces 6f information tobe mapped on to one another. This, I believe, helps torespond to Bloom & Wynn's example of number. First, Iam not entirely clear that magnitudes exclude sequentialorder, but in any case, RR can result either in newinformation being available from the redescription ofcomponents of procedural representations or in newmappings becoming possible across differently repre-sented information. If early number is based on magni-tudes which do not map directly onto the counting system,then both must ultimately be in a similar representationalformat for a mapping between them to take place. In otherwords, RR is just part of another complex process but, Iwould argue, a necessary part of it. Bodor & Pteh also pointout that knowledge may emerge from interactions betweendifferent domains. I, of course, agree with them, but as just


Resporwe/Karmiloff-Smith: Beyond modularity

pointed out, I. argue that this is only possible if theinteractions take place on representations which havebeen redescribecl in formats that render such interactionpossible.

I fail to see why redescription of knowledge alreadypresent in the child's mind necessarily precludes thatknowledge from being considered "new" (Bloom & Wynn)although, as Dartnall points out, this opens a thornyphilosophical question. If new relationships are formedacross previously unlinked representations (because theywere in different formats), then the resulting representa-tions are truly new and reflect relationships that may not bediscernible in the real world. Rather, they are discoverableonly in hypothetical, possible-world reasoning.

Losonsky, like Bloom & Wynn, raises another aspect ofFodorian theory. He maintains that I not only go beyondFodor's modularity thesis, but also beyond his meth-odological solipsism. Losonsky embraces the RR notionthat autonomy and individuality of psychological statesare a product of development and not its starting point.For Losonsky, however, level-El representations are alsononsolipsistic in so far as they are individuated by theirepigenesis - by being redescriptions of level-I represen-tations which were created via interaction with the envi-ronment - a point also made by Dartnall. However, likeRutkowska, Losonsky suggests that I give a more promi-nent role to external states in the environment as essentialcomponent parts of the initial procedural representa-tions. What he envisages for level-I representations is afeedback loop consisting of appropriate sequences ofstates of mind and external physical states. This, accord-ing to Dartnall and Losonsky, circumvents the symbolgrounding problem (Harnad 1990) and may also helpaddress Campbell's worries about "encodingism" a laFodor. According to Dartnall, by combining RR with aninformation-theoretic account of intentionality, E l isworld-inducing because it is a redescription of knowledgethat came about originally by being causally situated inthe world. The important role of external states in gener-ating behavior at the I level can be seen in the data fromstudies of external notations (discussed in Ch. 6), in whichchildren often cannot carry out an expressed intentionbecause they are driven by the external state of thedepiction (perhaps Boden had something along theselines in mind in part of her commentary). The young childannounces she will make a funny drawing, but, data-driven by the sequential states of the trace of her produc-tion, her drawing ends up just like her normal one.

Together with many commentators (e.g., Campbell,Dartnall, Donald, Freeman, Losonsky, Quartz & Sej-nowski, Rutkowska), Modularity stresses the need to shiftfrom Fodor's exclusive concern with input systems to afocus on output systems. I have given considerablethought to the importance of the difference betweeninput and output systems in infants, but I found Free-man's commentary particularly illuminating with respectto certain facets of the role of output systems in olderchildren. He cites research by Lewis (1994) showing thatwhen children narrated to themselves a story they hadbeen listening to (i.e., when they recycled the dataexplicitly through the processing system via their ownverbal output) then their performance was significantlybetter than that of a control group of children who merelylistened. The experimental children were able to base

their theory-of-mind reasoning on explicit representa-tions.

The focus on the redescription of input data into anoutput system may have pedagogical implications, as Estessuggests, although in my view explicit teaching often doesnot result in the same progress as spontaneous redescrip-tion (see Hennessy 1986), which is why I disagree with deGelder's interpretation of data from adult dyslexics. Theremay be other pedagogical implications I have not consid-ered, however, such as those stressed by Goldin-Meadow& Alibali and Ohlsson, concerning the importance ofunderstanding the process of learning that representa-tional change continues after behavioral mastery.

R5. Piaget misrepresented? I am accused by Bodor &¥\6h of misrepresenting Piaget. Having spent over adecade working closely with Piaget and his colleagues, Ifeel authorized to disagree strongly. Never did Piagetargue that Euclidian geometry was innately specified inthe infant brain. On the contrary, he constantly endeav-ored to invoke the most minimal domain-general pro-cesses possible. Nor in Piaget's theory is "the wholeprocess of development . . . driven by an internal pro-gram. " On the contrary, Piaget's genetic epistemologywas an attempt to obviate both maturational accounts ofdevelopment and teleological ones. I therefore disagreewith Scholnick's teleological view of Piagetian theory.

Bodor & Ple"h accuse me of treating Piaget as anantiquated empiricist. I would encourage them to rereadthat section of Chapter 1. They will see that I havecompared Piaget and behaviorists solely in the light ofcurrent nativist claims about development, but I go on tostress fundamental differences between Piaget's theoryand that of the behaviorists. I still maintain that there areseveral similarities between the two theories at the levelof domain-general change and the claim that the infantbrings only domain-general predispositions to the learn-ing situation. What these domain-general predispositionsare and how learning takes place is, of course, verydifferent in the two theories, as Modularity makes clear.

Bodor & He'll are also unhappy that I do not discussexplicitly Piaget (1976) on the grasp of consciousness.They feel that I would have found a more precise accountof metacognition in Piagetian theory, whereas Campbellcongratulates me on being more precise than Piaget onsuch issues. It is obviously up to the reader to decide, butmy aim in Modularity was neither to demolish Piaget byan exhaustive discussion of every aspect of his theory norto give a disciple's view. Rather, I tried to retain thoseaspects of the Piagetian enterprise which I had alwaysfound inspirational, and to dispense with those whicheither data or theorizing have shown to be questionable.

Smith still holds to a strictly Piagetian view. I fail totallyto grasp why he maintains that universality necessarilyexcludes domain specificity. One need only consider thecase of phonological abilities, which are clearly bothuniversal and domain specific, to see that such an infer-ence cannot hold. And I find it very unfair to link mycriticism (on p. 28 of Modularity) of certain types ofdevelopmentalists who focus only on age with a totallydifferent point made some 140 pages later regarding whatwas wrong with Piagetian theory. Nowhere in Modularitydo I claim that Piagetian stages are strictly age-related.Nor do I imply anywhere that modal knowledge is domain



specific, nor that RR is domain specific. What I do claim isthat RR is a domain-general process that operates domainspecifically, that is, at different times according to therepresentational status of the knowledge on which itoperates. Likewise, had I discussed modal knowledge, Iwould have equated it with the upper levels of RR,arguing that necessity can arise only in the case of explic-itly defined representations which become progressivelymore domain general.

I am not convinced, however that modal knowledge isthe quintessential feature of human cognition, as Smithmaintains. When I was still in Geneva, a physicist,R. Carreras, carried out a project (to my knowledge, alas,unpublished) on the limits of cognition in subjects ofnormal intelligence. He asked children and adults a se-ries of questions about the infinitesimally small. Thoughyoung children found the questions difficult, older chil-dren and adult subjects approached some (but not all) ofthe questions with arguments of necessity. For example,when asked to imagine a building filled with billions ofgrains of salt to which one grain was added, they allanswered that this would necessarily increase the cardinalsum of the grains. In such contexts, number appears to beprocessed modally. By contrast, when asked whether theaddition of one salt grain would change the total weight orthe total volume, many older children and adults hesi-tated and were clearly unsure. They also expressed uncer-tainty when asked whether one drop of ink in LakeGeneva would change the color of the water. In many ofthe contexts involving the infinitesimally small, they didnot reason according to modal logic. This seems to chal-lenge Smith's (and Piaget's) views in this respect. Ingeneral, much of the reasoning of children and adults lacksa modal quality; they have problems extending theirthinking to unobservable cause-effect relations (Karmiloff-Smith 1984), as well as difficulties in relating theory andevidence as Kuhn points out. Smith's vision of the humanmind seems to point to the ideal subject, not the realpsychological one.

Some commentators (Campbell, Smith) do not acceptthe recurrent phase model of change and argue stronglyin favor of a Piagetian view of across-the-board change.Graham, by contrast, explores the area of self-attributionand concludes that a phaselike view of change is indeedappropriate. Once again, we must avoid taking an ei-ther/or approach to change. Development may involveboth recurrent phaselike change and some stagelikechange, as Zelazo suggests in invoking a general refine-ment of control structures around age 5. For Zelazo, theRR model fails to capture several aspects of the develop-ment of age-related domain-general changes in cognitivecontrol, but he agrees that development may involve bothdomain-specific and across-the-board changes, and RR isaimed at accounting for aspects of domain-specificchange. As I argued in Modularity, it is possible that suchacross-the-board changes, where they occur, are discern-ible from major maturational changes in the brain.

R6. Beyond the implicit/explicit dichotomy. Modularity'sgoal was to discuss the multiple levels at which knowledgeis represented, but for Carassa & Tirassa the content ofrepresentations is more important than their format.They ask why conscious access should depend on the useof a particular format rather than on properties of con-

sciousness itself. I could answer their question if they (oranyone) had enumerated what the properties of con-sciousness are! I wish we knew. Whereas some re-searchers question the very existence of implicit knowl-edge (see the BBS treatment of Shanks & St. John [1994]and Vinter & Perruchet), de Gelder accepts the distinc-tion, but suggests that there is no relation between im-plicit and explicit knowledge and that they develop in anunconnected, parallel fashion. If this were the case, thenwe would again be faced with the symbol groundingproblem as far as explicit representations are concerned(Harnad 1990), a problem to which de Gelder proposes noalternative solution.

According to Dartnall and Losonsky, RR circumventsthe grounding problem, but for Campbell this remains aserious lacuna. Mandler has discussed these issues indetail with respect to infancy and her thinking and experi-ments have led her to change from a parallel view (Man-dler 1988) to one consistent with the redescriptive frame-work (Mandler 1992) developed in Modularity. The merefact that something ends up dissociated in adults does notautomatically tell us how it started off in development, asde Gelder maintains.

When criticizing various aspects of the RR framework,several commentators fall back on a simple dichotomybetween implicit and explicit knowledge (Bodor & Pl^h,Carassa & Tirassa, de Gelder, Ohlssori, Vinter & Per-ruchet). For example, Bodor & Pleh ask whether I-levelrepresentations are equivalent to "knowing how" or"knowing that." If I were to accept that simple dichotomy,then obviously the I level is about "knowing how." Butwhat Modularity tries to show is that there are multiplelevels of "knowing that," which a simple dichotomy doesnot capture. Carassa & Tirassa remind us that it is essen-tial to distinguish between representations of proceduresand procedural representations, but in my view Mod-ularity takes this distinction into account and discussesthe different uses of the term "procedure" (pp. 161-62)and the definition of I level in terms of bracketed repre-sentations (p. 20), as further discussed by Carassa &Tirassa and by Grush.

Several commentators lamented the lack of data inModularity supporting the hypothesized level-E2 repre-sentations as well as the difficulty of clearly distinguishinglevels (Ohlsson, Scholnick, Shultz, Zelazo). In my view,Goldin-Meadow & Alibali have provided exciting newevidence in support of the RR multiple level distinctions.They demonstrate that children may display knowledgein their gestures that is not yet in their verbal responses.If this were all they had shown, it would be difficult todecide whether the gestural knowledge should be classi-fied as level I, El , or E2. However, their subsequentresearch on children's explicit judgments neatly ad-dressed this question. Children used the knowledge dis-played in their gestures to make judgments. WhatGoldin-Meadow & Alibali have not yet suggested, butwhich might well exist, is an even earlier stage of gestur-ing without the concomitant ability to use the knowledgein judgment. Smith argues against the multiple level RRmodel, maintaining that there can be no explicit repre-sentations without justification and necessity, but Goldin-Meadow & Alibali's research is an elegant demonstrationof how explicit representations can exist prior to verbalones.


For Freeman, Graham, and Kuhn, Modularity's dis-cussion of the child's construction of intentional psychol-ogy is not detailed enough. Graham introduces an impor-tant distinction between representational content and selfattribution, suggesting that these can dissociate in certainforms of pathology in which representations are not ex-plicitly marked with self. It would be interesting to usehis framework to rethink the question raised in Mod-ularity about whether autism involves a representation orinformation-processing deficit. Freeman argues that Ihave not used the multiple redescriptive frameworkenough when discussing (in Ch. 5) the child's constructionof intentional psychology, an issue also raised byDartnall. The RR model predicts that children shouldhave implicit knowledge of theory-of-mind computationsbefore such knowledge is explicitly represented and ver-bally expressible. Freeman cites new work by Povinelliand deBIois (1992), Freeman et al. (1991) and Clementsand Perner (1994) that shows this to be the case. Youngchildren often look towards the correct location, whereasin their statements they indicate the wrong location,suggesting that knowledge expressible in one modalityhas not yet been redescribed into another.

To address Shultz's above mentioned question about therepresentational status of these different types of re-sponse, it might be profitable for theory-of-mind studies toincorporate a judgment trial such as the one used byGoldin-Meadow & Alibali in their research. The new datafrom the number and theory-of-mind domains supportthe view discussed in Modularity that a simple im-plicit/explicit dichotomy will not suffice to account fordevelopmental change.

Whereas Bodor & Pl6h return to a Piagetian positionand suggest that RR can be boiled down to a single generalsemiotic function that penetrates all others, Hampsonand Donald find the concept of multiple levels in differentformats useful. Hampson asks whether level-I represen-tations can be thought of as procedural memory (I defineprocedural representations on p. 20 of Modularity asbracketed) and levels E1-E3 as semantic and episodicmemory (see also Donald). To do so, however, would shiftthe focus from (1) the process of representational re-description which can be used to change knowledge inone type of memory store into knowledge in another typeto (2) the product of memory stores. This shift would losethe important fact, stressed by Donald, and the RRframework, that level-I representations are data-driven,whereas higher levels can be self-triggered.

R7. The problem of mechanism. As Dartnall stresses, theremaining challenge is to model the mechanism thatmakes RR possible. Boden also notes that the discussionof mechanisms in Modidarity is unavoidably vague. I ofcourse agree, and explicitly mentioned this in Chapter 8.But neither offers any suggestions regarding mechanismsto replace the lack of specification. I did take some time inModularity to discuss the difference between precisemechanisms governing on-line behavior, and generalprinciples governing developmental change. AlthoughOhlsson dismisses much of the content of Modularity,particularly the lack of a specified mechanism, in the workof his own cited in the commentary (Ohlsson & Rees 1991)he readily admits that he has not provided a developmen-tal mechanism at all: "We have tried to formulate a theory


of how knowledge of principles, once acquired, can beused in learning procedures; we have not tried to explainthe acquisition of principles" (p. 175).

Let me reiterate how essential it is to distinguishbetween the use of knowledge in on-line phenomena in asteady state system and the acquisition of knowledge viamacro-developmental change. RR was an attempt to ad-dress the latter. It is also worth noting that whenever astrictly nativist solution is proposed, as Johnston pointsout, the mechanism is never specified. As Braine (1994)stresses, the projection principle, the theta-criterion,subjacency, the case filter, the empty category principle,to name but a few, have never been accompanied by adiscussion of how such principles would be embodied inmental operations or structures in the child's mind. Con-trary to what Foster-Cohen implies, we simply do notknow what innate mental resources are actually beingposited. I believe that RR goes a little further than this.

The developmental period leading up to behavioralmastery is one avenue of research that might, according toGoldin-Meadow & Alibali and Vinter & Perruchet, helpin advancing our thinking about the issue of mechanism.They are right that these are essential complements to thedevelopmental picture, but one researcher cannot doeverything, and Modularity's strategy was to look beyondbehavioral mastery - the point, as Ohlsson recognizes, atwhich most other studies stop.

Vinter & Perruchet suggest that behavioral masterymay emerge as a consequence of a first RR process or adirect sensitivity to the product of rules, rather than to theimplicit rules themselves. Goldin-Meadow & Alibali sug-gest that redescription may occur when stable states havenot given rise to successful performance. In this way,redescriptions may provide an internal destabilizing force(along the lines of Shultz's error reduction). It is possiblethat different solutions apply to different domains. Also,there is a dominant tendency in developmental theoriz-ing to seek failure-driven explanations of change.

As I pointed out in Modularity, this is certainly onesource of change, but not necessarily the only one. None-theless, Goldin-Meadow & Alibali's suggested modifica-tion to the constraints of behavioral mastery on RR is aninteresting one. The question then becomes Why isbehavioral mastery a prerequisite in some domains andnot others? A first guess is that mathematical reasoningcannot be informed by external feedback in the same waythat problem solving and language can. The children inGoldin-Meadow & Alibali's study do not know their rea-soning is incorrect, so it becomes stabilized. In non-mathematical problem solving, for example, children are

, usually aware during phase 1 of their failure to reach agoal. In such cases, redescription seems to occur only asbehavioral mastery has been attained. The timing of RRmay therefore be influenced by the contexts in which itoccurs.

Three commentators (Grush, Rutkowska, Zelazo) raisesome interesting issues regarding the distinction betweencontrol structures and prediction. Grush, while agreeingin spirit with almost all aspects of the RR framework,suggests that it fails to differentiate between the forwardproblem (prediction) and the inverse problem (control)and that it collapses two processes - internalization andparametrization. I agree that it is useful to separate theseprocesses. I found particularly helpful his argument that



predicted feedback is faster than real feedback and that itallows for off-line planning and practice. This is an issue Ihave been grappling with, concerning El-level represen-tations. Rutkowska also explores the implications of RR tosupport anticipation and reactive control in infancy, show-ing that as infants get older they shift from reactive controlto the general refinement of control structures, a shift thatalso occurs, according to Zelazo, in older infants andchildren.

I found Shultz's discussion of the new cascade-correlation connectionist model particularly illuminating,especially the notion that cascaded hidden units afford theconstruction of increasingly powerful knowledge repre-sentations that were not available to developmentallyearlier instantiations of the network. However, the mech-anisms of error reduction that move a system from I to E lalways occur simply as the product of processing theinput. Moreover, in the human case, there exist data fromabnormal phenotypes suggesting that although behav-ioral mastery is often necessary for representationalchange, it is not sufficient. This is suggested by a study(Cromer 1991) of a retarded but extremely fluently speak-ing adolescent with hydrocephalus and associated my-elomeningocele who showed no signs whatsoever ofmetalinguistic awareness. Also relevant is a case study of10-year-old with Down syndrome who, across repeatedsessions of successful block balancing (discussed in Ch. 3),gave no indication of moving beyond level-I success.Thus, one can reach behavioral mastery without automat-ically moving on to RR, suggesting that, unlike in Shultz'ssimulation, there exists a mechanism particular to RRwhich is not part of normal input/output processing. Thismay be a predisposition that the child brings to learning,or it may arise in ways similar to those connectionistnetworks which modify their architecture as a function oflearning.

R8. The role of written language in generating representa-tional change. For some commentators (Oartnall, deGelder, Olson, Scholnick), I have underestimated the factthat we have a written culture. But Modularity in no waydisputed the fact that literacy training during ontogenyaffects development and causes reconfiguring of parts ofthe brain, as Donald recognizes. In fact, Modularitymade a strong case for the importance of external nota-tions in general and devoted a whole chapter (5) to thattopic.

Olson appears to agree with the RR framework as ageneral process of representational change but arguesthat it is not internally generated. It has certainly beendemonstrated that familiarity with the alphabet enhancesphonemic awareness, but this cannot be achieved withoutthe involvement of some endogenous factors (as evi-denced by cases of dyslexia). There is a difference be-tween having explicit knowledge of phonemes and beingable to map phonemes onto some other (orthographic)representation, and then to manipulate the other repre-sentation. For example, adults and children, once theycan spell, will say there are more sounds in "pitch" than in"rich." Like de Gelder, Olson draws attention to the workof the Brussels team on normally intelligent adult illiter-ates, also discussed in Modularity. Illiterates performpoorly on phoneme awareness tasks, suggesting thatphonemic awareness is a function of literacy. I have no

quarrel with this and have never argued that RR is solelygenerated endogenously. However, both Olson and deGelder slip from discussing the general issue of phonolog-ical awareness to the more restricted issue of phonemicawareness.

My understanding is that it is far from clearcut what thephonemic level of representation of spoken languagereally is. It is true that infants discriminate between "/ba/and /ga/," but it is not clear that in the auditory modalitythey, or we, ever hear /ba/ as two separate phonemesmade up of Ihl and /a/; these may be considered primi-tives by the auditory system and indeed only be availablefor redescription once alphabetic writing has been ac-quired. This, however, does open up the question of thehistorical beginnings of the alphabetic system. By con-trast, the syllabic and word levels are represented inspeech; pauses can be inserted between: them and, as deGelder herself notes, 4-year-olds play word games, userhyme, and show other such signs of haying redescribedthese aspects of their spoken language. By contrast, it isnot until the age of 6 that children seem to solve phonemicawareness tasks, suggesting that phonemic awareness, butnot phonological awareness, may indeed be a function oflearning the alphabet, as Olson also argues. To find outwhether explicit tuition is always required for phonemicawareness, as de Gelder suggests, it would be necessaryto explore the question with children who spontaneouslyteach themselves to read very early.

For Olson, even awareness of the word level requiresliteracy. Here I disagree. Preliminary results with adultilliterates suggested that they perform well on the wordtask discussed in Chapter 2. Moreover, our results showthat preliterate 4-year-olds have awareness of both openand closed class words. The written form is not essentialto the word level of metalinguistic awareness becausethere are other phonological and morphological cues tothe word level in spoken language. Many of the testspurporting to get at children's metalinguistic knowledgeare too off line and require nonlinguistic processes, apoint discussed at some length in Chapter 2.

R9. Drawing and other forms of external notation. Someintriguing suggestions about the order of developmentalchange are made by Boden, but alas her interpretation ofmy reported data is incorrect and thus her suggestionscannot be taken further. The data show no developmentalsequence between change of whole/parts/deletion, norbetween deletion/insertion/change of position/orienta-tion. The former three occur simultaneously in the young-er age group's development, and the latter occur simul-taneously in the older age group. The previously publishedhistograms and the ages of the illustrations in Modularitymake this very clear. No significant within-age groupdifferences between the categories of change were everreported. The difference lies between the two age groups.The order in which they appear on the histogram withineach age group is totally arbitrary, as is obvious from thetiny percentage differences. I originally explained theyounger children's solutions in terms of a sequentialconstraint, but in Modularity (p. 162) I point out thatsubsequent data showed the sequential constraint to beless rigid than I initially thought for drawing (though seeVinter & Perruchet). A sequential constraint does seem tohold for early phonemic awareness, (as discussed in the


original article, and reiterated in Boden's commentary)just as it does for the development of other cognitive skillssuch as seriation. But, as Boden suggests, it is certainlyworth pursuing the similarities and differences betweendevelopmental change and adult creativity.

Freeman is absolutely right to pull me up on the factthat propositional content/propositional attitude distinc-tion is not only relevant to theory-of-mind computationsbut also to the notational domain (see Dartnall for adifferent view). In general, I agree with Freeman that Ihave not applied the RR framework to the notationaldomain as consistently as I have to others. In fact, ifautism researchers are right in suggesting that proposi-tional attitudes are not available to individuals with au-tism (Leslie 1987), then the notational domain mightturn out to be a very interesting test case.

R10. Other influences on representational change.Several commentators feel that I downplayed importantsociocultural influences and overestimated the internallydriven nature of representational change (e.g., Bodor &Pl<5h, Campbell, de Gelder, Olson, Scholnick). First, letme reiterate, as I did in both the book and the Precis,that I do not deny that some representational change isindeed generated by externally driven sociocultural fac-tors (see p. 122 of Modularity, and above comments onthe role of written language). What I argue is that socio-cultural factors are not the sole factor influencing repre-sentational change.

Whereas Freeman agrees that we should analyze theinternal resources that cognitive systems afford in thefurtherance of intellectual transformations, other com-mentators argue that RR is not endogenously generatedbut is either driven by error or conflict reduction (Bodor& Pl6h, Ohlsson, Shultz), by information integrationduring problem solving (Zelazo), by noticing regularitiesin one's behavioral output (Rutkowska, Vinter & Per-ruchet, Zelazo), or by external factors such as the writtensystem (Olson, de Gelder). I do not disagree with any ofthese suggestions. But, as I stressed in the introduction,development should not be seen in "either/or" terms.There are many different processes by which we learn,some of which are spontaneous and endogenously gener-ated. Sometimes error reduction is indeed a generator ofchange and, as Shultz suggests, the error may be moresubtle than at the output level. Indeed, children maynotice regularities in their behavioral output (see thediscussion of the map task in Ch. 6), but they also noticeregularities in the input, and in their own representa-tions of the input/output relations. RR is predominantlyabout the latter.

R11. Connectionist principles of development. Piagetiantheory and connectionist theories are incompatible,Scholnick argues, in that the latter eliminate direction-ality because the end point is arbitrary. By contrast, Isee in the connectionist principles some very closeanalogies with Piagets intuitions of assimilation, ac-commodation, and equilibration (see McClelland, sub-mitted).

It is a pity that no Fodorian disciple or Chomskianpsycholinguist (e.g., Foster-Cohen) took on the task ofdiscussing solutions to language learning other than theparty line. I would be curious to know how they would


deal with Elman's (1990; 1993) work on the structure ofembedded sentences with long-distance dependencies(discussed in Ch. 8). Of particular interest is that whenthe network was fed with the full input set from thebeginning, it did not learn. However, either by restrict-ing the input incrementally or, more realistically, by in-creasing the short-term memory of the network incre-mentally, it did learn long-distance dependencieswithout any prespecified module to direct that learning.Furthermore, whereas connectionists argue for parallelprocessing, Foster-Cohen views the brain as involvingsequential processes. She talks of children having trou-ble abandoning processing in the pragmatic module,prior to moving on to processing in the syntactic moduleof the next utterance. Since we know that the brain is amassively parallel system, it is rather surprising thatchildren do not process syntactic and pragmatic informa-tion in parallel.

In Modularity, my position on connectionism was pos-itive with respect to how it suggests ways of modellingthe processes leading to behavioral mastery and I-levelrepresentations, as well as ways of rethinking the nature/nurture debate. I was more cautious about whether theconnectionist framework could account for the process ofexplicitation. Since Modularity was written, I have beenable to study recent connectionist models in far moredetail and would now withdraw my initial criticism abouttheir sole focus on tasks (see Shultz). Many connectionisttheorists do indeed focus on understanding developmen-tal principles (Bates & Elman 1993; McClelland forth-coming; Plunkett & Sinha 1992; Shultz 1991). Somecommentators endorsed my positive evaluation of theconnectionist implications for development (Grush,Hampson, Quartz & Sejnowski, Shultz), whereas othersfocused on my caution and were entirely negative (e.g.,Campbell, Smith). According to Bloom & Wynn, for ex-ample, connectionist models tell us nothing about hownew representations can emerge developmentally. Insaying this, they seem to neglect the complex interactionbetween the architecture, learning algorithms, timingfactors of the network, and the structure of the inputswith which the network interacts, together with the factthat networks can change their architecture as a functionof learning (e.g., Shultz). De Gelder makes a point simi-lar to that made in Modularity, that is, that another net-work would have read off the representations created bythe initial one. I would go further, however, and add thatin order for the new network to be able to use the re-described representations for broader purposes and totransport the knowledge to the products of other net-works, the format must be redescribed into a commonone. The implicit representations of any particular net-work continue to reflect the specific properties of theinput set that they have processed and therefore are nottransportable. However, both Grush and Shultz countermy caution by suggesting that connectionism does offersolutions to modelling the passage to explicit representa-tions at least at level El. In particular, Shultz details amodel of cascade correlation which seems to capture thefirst level of redescription in ways that are very close tothe intuitions underlying RR. By contrast, Rutkowskasees a problem in the connectionist framework becauseit is not situated causally in embedded systems, whereasfor Dartnall the grounding problem is not relevant to



connectionism because its representations are the resultof processing external input.

Many connectionists argue for a domain-general viewof learning, although in Modularity I argue that net-works are in effect often domain specific. So-called struc-tural connectionists (e.g., Shastri & Ajjanagadde BBStreatment 1993; see also, discussion in Quartz 1993) doactually argue explicitly, on both computational and rep-resentational grounds, against the domain-general viewof learning. Partially structured connectionism combinesstructured and unstructured network design techniques,which make it possible to capture both the flexibilityafforded by unstructured networks and the tractability inlearning and improved generalization that result frommore structured network design. This is a nice exampleof circumventing an "either/or" view of learning.

R12. Should developmental science banish the term "in-nate"? Whereas for some commentators Modularity hasfurthered the nature/nurture debate (Estes, Dartnall),Ohlsson strongly disagrees. In an important commen-tary, Johnston suggests that the term "innate" be totallybanned from developmental discussion (see also, Oyama1985). He suggests that, despite myself, I tend to fallback on the nature/nurture dichotomy. Let me reiteratemy position. In the state of the art of developmentaltheorizing, where so much is explained (away) by invok-ing innateness (e.g., Bloom & Wynn, Foster-Cohen, onsyntactic structure), I felt it essential to encouragereaders of that persuasion to rethink nativism and todiscuss how minimal the innate specifications might be.

From the point of view of the genes, Johnston is ofcourse right: although genes play some role in every-thing, that role is exceedingly indirect. Obviously nocognitive outcome is genetically coded; there is certainlyno "gene (or set of genes) for syntax" or for "past tense"!Genes simply supply cells and organisms with chemicalmaterials; they are passive sources of material on which acell can draw (Nijhout 1990). Genes do not code forlevels of organization higher than the primary structureof proteins. The brain is not totally unstructured at birth,however; there is considerable plasticity, but not totalplasticity. Recent developments in the study of patternsof cell division and determination suggest that distinctregions of the neuraxis may use different mechanisms ofregional specification which differ in their degree of plas-ticity (Levitt & O'Leary 1993). Although cortical circuitsare not immutably specified for their functional roles, itis also true that not all features of the cortex can bealtered by altering the pattern of input activity (Sur1994); this questions the total initial functional equipo-tentiality of the total cortex. I find Quartz & Sejnowski'srejection of any predispositions difficult to reconcile withthis and with the examples of dissociations in abnormaldevelopment discussed in Modularity.

We are of course left with the problem of how to un-pack influences on development. One solution proposedby Johnson & Morton (1991, pp. 130-31) is to distin-guish between specific and nonspecific aspects of thespecies-typical environment and thereby to delineatedifferent types of process that result from development.They introduce the term "primal" to distinguish thesource of influence arising from the interaction of thegenome with the internal environment from nonspecific

external environmental input - necessary stimulationthat is unrelated to the structures whpse maturation ittriggers. They distinguish this from "genetic" influencesfrom the genome and "innate" influences from the inter-nal environment, as well as from specific aspects of thespecies-typical environment which give rise to normallearning. Similarly, the interactionist approach does notclaim that all behavior is learned, but rather that allbehavior results from a complex of epigenetic interac-tions within the developing animal-environment system(not the gene-environment), interactions that may in-volve both obvious and nonobvious contributions by theenvironment (Johnston 1987, pp. 172-73).

Johnston and Quartz & Sejnowski feel that a trulyepigenetic approach must embody prenatal develop-ment. It is true that the fetus has sorne surprising dis-criminatory capacities, but Dartnall, Dpnald, and Camp-bell suggest that even considering prenatal developmentdoes not suffice without the addition of an evolutionaryperspective on the human mind. Donald warns rightlyagainst any simple mapping of phylogeny onto on-togeny, but argues that one thing that nonhuman pri-mates have in common with the first phase of humanmicrodevelopments is that they cannot self-triggerstored information. This stimulus-driven memory re-trieval corresponds to phase 1 in the reiterated 3-phaseprocess invoked in Modularity. Donald's other twostages of evolutionary development (the recoding of rec-ognition memory into mimetic and verbal auto-cuablepaths) are, according to him, specifically human. By con-trast, Carassa & Tirassa argue that it is premature toexclude the possibility of explicit representations inother species. I agree that it has been shown that"language-trained" chimpanzees have explicit represen-tations, but I would argue that RR helps us to under-stand why such representations always turn out to bevery limited in comparison to the human case. Indeed,chimpanzees are massively trained with external stimuli;their explicit representations are not the result of inter-nal redescription of already stored knowledge. Nonethe-less, I fail to see why the existence of a capacity in otherspecies necessarily calls into question its domain speci-ficity in humans (Bodor & PU;h). Domain specificity doesnot necessarily imply species specificity, but simply thatthe behavior has biological generality and evolutionaryroots.

R13. Concluding comment. I should like to thank allthose who took the trouble to write commentaries. Theyserve as a guide to important unresolved issues. RR wasmy best shot at giving an integrated account of dataacross a wide variety of domains (which in isolation maybe better accounted for by a series of different explana-tions), and at the question of what is special about thehuman brain/mind compared to that of other species. Ifnothing else, this treatment serves to highlight the hugediversity of opinion there is in develppmental cognitivescience. Together with Johnston, I also hope it will gen-erate lively and productive discussion about the benefitsthat accrue from a truly developmental approach to hu-man cognition.


References

Letters a and r appearing before authors' initials refer to target article andresponse respectively.

Alibali, M. W. & Coldin-Meadow, S. (1993) Gesture-speech mismatch andmechanisms of learning: What the hands reveal about a child's state ofmind. Cognitive Psychology 25:468-523. [SG-M]

Anderson, J. R. (1980) Cognitive psychology and its implications.Freeman. |aAK-S]

(1982) Acquisition of cognitive skills. Psychological Review 89:369-406. [SO]

Armstrong, S., Cleitman, L. & Gleitman, H. (1983) What some conceptsmight not be. Cognition 13:263-308. (PBI)

Baillargcon, R. (1986) Representing the existence and the location of hiddenobjects: Object permanence in 6- and 8-month-old infants. Cognition23:21-41. [JCR]

Barnes, D. P. & Hampson, P. J. (1993a) Learning to learn: The contribution ofbehaviour analysis to connectionist models of inferential skills in humans.In: Neural computing: Research and applications, ed. G. Orchard. AdamHilger. [PJH]

(1993b) Stimulus equivalence and connectionism: Implications for behaviouranalysis and cognitive science. Psychological Record 43:617-38. [PJH]

Bartlett, F. C. (1932) Remembering. Cambridge University Press. [SO]Bates, E. 6r Etman, J. L. (1993) Connectionism and the study of change. In:

Rrain development and cognition: A reader, ed. M. H. Johnson.Blnckwell. [arAK-S]

Bates, E. & MacWhinney, B. (1987) Competition, variation and languagelearning. In: Mechanisms of language acquisition, ed. B. MacWhinney.Erlbamn. luAK-S]

Bechtel, W. & Abrahamscn, A. (1991) Connectionism and the mind: Anintroduction to parallel processing in networks. Blackwell. [aAK-S]

Bellugi, U., Marks, S., Bihrle, A. M. & Sabo, H. (1988) Dissociation betweenlanguage and cognitive functions in Williams syndrome. In: Languagedevelopment in exceptional circumstances, ed. D. Bishop & K. Mogford.Churchill Livingstone. [aAK-S]

Berns, C. S., Dayau. P. & Sejnowski, T. J. (1993) A correlational model forthe development of disparity selectivity in visual cortex that depends onprenatal and postnatal phases. Proceedings of the National Academy ofSciences 90:8277-81. [SRQ]

Bickhurd, M. H. (1978) The nature of developmental stages. HumanDevelopment 21:217-33. [RLC]

(1980a) Cognition, convention, and communication. Praegcr. [RLC](1980b) A model of developmental and psychological processes. Ccnetic

Psychology Monographs 102:61-116. [RLC](1992) Commentary on "The Age 4 Transition." Human Development

35:182-92. [RLC](1993) Representational content in humans and machines. Journal of

Experimental and Theoretical Artificial Intelligence 5:285-333. [RLC]Bickhnrd, M. H. & Campbell, R. L. (1989) Interactivism and genetic

cpistemology. Archives dc Psychologic 57:99-121. [RLC](1992) Some foundational questions concerning language studies: With a

focus on categorial grammars and model theoretic possible worldssemantics. Journal of Pragmatics 17:401-33. [RLC]

Bickhnrd, M. H. & Ritchie, D. M. (1983) On the nature of representation: Acase study of James Gibson's theory of perception. Praeger. [RLC]

Bloom, P. (1994a) Recent controversies in the study of language acquisition.In: Handbook of psycholinguistics, ed. M. A. Gernsbacher. AcademicPress. [PBI]

(1994b) Cenerativity within language and other cognitive domains.Cognition 51:177-89. [PBI]

(in press) Theories of word learning: Rationalist alternatives toassociationism. In: Handbook of language acquisition, ed. T. K. Bhatia 6cW. C. Ritchie. Academic Press. [PBI]

Bodcn, M. A. (1990) The creative mind: Myths and mechanisms. London:Weidenfeld & Nicolson. [MAB]

Brnine, M. D. S. (1994) Is nativism sufficient? Journal of Child Language21:9-31. [rAK-S]

Braincrd, C. (1978) The stage question in cognitive-developmental theory.Behavioral and Brain Sciences 2:173-82. [LS]

Brunei', J. (1966) Toward a theory of instruction. Harvard UniversityPress. (SO)

Bruner, J. & Haste, H., eds. (1987) Making sense: The child's construction ofthe world. Methuen. iPBo]

Bruner, J. & Lucariello, J. (1989) Monologue as narrative recreation of the

world. In: Narratives from the crib. ed. K. Nelson. Harvard UniversityPress. [PBo]

Re/erences/Karmiloff-Smith: Beyond modularity

Burge, T. (1986) Individualism and psychology. Philosophical Review 95:3-45. | ML]

Byrnes, J. & Duff, M. (1988) Young children's comprehension of modalexpressions. Cognitive Development 4:369-87. [LS]

Campbell, R. L. (1992) A shift in the development of natural-kind categories.Human Development 35:156-64. [RLC]

Campbell, R. L. & Bickhard, M. H. (1986) Knowing levels and developmentalstages. Karger. [RLC]

(1992a) Clearing the ground: Foundational questions once again. Journal ofPragmatics 17:557-602. [RLC]

(1992b) Types of constraints on development: An interactivist approach.Developmental Review 12:311-38. [RLC]

Carey, S. (1985) Conceptual change in childhood. MIT Press. [arAK-S, PBI](1988) Conceptual differences between children and adults. Mind and

Language 3:167-81. [PBI]Carmichael, L. (1925) Heredity and environment: Are they antithetical?

Journal of Abnormal and Social Psychology 20:245-60. [TDJ]Case, R. (1985) Intellectual development: Birth to adulthood. Academic

Press. [aAK-S]Chandler, M. & Hala, S. (1994) The role of personal involvement in the

assessment of early false-belief skills. In: Children's early understandingof mind, ed. C. Lewis & P. Mitchell. Erlbaum. [NHF]

Changeux, J. P. (1985) Neuronal man: The biology of mind. Pantheon.[aAK-S]

Chomsky, N. (1957) Syntactic structures. The Hague: Mouton. |TD](1980) Rules and representations. Columbia University Press. . [RLC, SRQ](1981) Lectures on government and binding. Foris. [aAK-S](1988) Language and problems of knowledge. MIT Press. [aAK-S]

Church, R. B. & Goldin-Meadow, S. (1986) The mismatch between gestureand speech as an index of transitional knowledge. Cognition 23:43-71. [SG-M]

Churchland, P. S., Ramachandran, V. S. & Sejnowski, T. J. (1994) A critiqueof pure vision. In: Large-scale neuronal theories of the brain, cd.C. Koch & J. Davis. MIT Press. [SRQ]

Clark, A. C. (1989) Microcognition: Philosophy, cognitive science, and paralleldistributed processing. MIT Press. [aAK-S]

(1993) Associative engines: Connectionism, concepts and representationalchange. MIT Press. [TD]

Clark, A. C. & Karmiloff-Smith, A. (1993) The cognizer's innards: Apsychological and philosophical perspective on the development ofthought. Mind and Language 8(3):487-568. [aAK-S, TD]

Clements, W. & Punier, J. (1994) Implicit understanding of false belief.Presentation at the Tenth Annual Conference of the Cognitive Section ofthe British Psychological Society, Cambridge, September. [arAK-S]

Cohen, A., Chien, Y. C. & Wexler, K. (1990) Children's knowledge of localityconditions in binding as evidence for the modularity of syntax andpragmatics. Language Acquisition 1:225-95. [SHF-C]

Cramer, R. F. (1991) Language and thought in normal handicapped children.Basil Blackwell. [rAK-S]

Dartnall, T. H. (1992) After Lamington: Artificial intelligence, creativity andrepresentational redescription. Artifwial Intelligence and Simulation ofBehavior Quarterly 81:35-39. [TD]

(1993) AI, creativity, representational redescription, intentionality, mentallife: An emerging picture. In: American Association of ArtifwialIntelligence technical report on artificial intelligence and creativity.American Association of Artificial Intelligence Press. [TD]

(1994) Creativity, thought and representational redescription. In: Artifwialintelligence and creativity: An interdisciplinary approach, ed. T. H.Dartnall. Kluwer Academic Publishers. [TD]

Davies, M. (1991) Individualism and perceptual content. Mind 100:461-84. [ML]

Davis, H. & Perusse, R. (1988) Numerical competence in animals:Definitional issues, current evidence, and a new research agenda.Behavioral and Brain Sciences 11:561-615. [LS]

de Gelder, B. & Vroomen, J. (1991) Phonological deficits: Beneath the surfaceof reading-acquisition problems. Psychological Research 53:88-97. [BdC]

de Gelder, B., Vroomen, J. & Bertelson, P. (1993) Effects of alphabeticreading competence on language representation in bilingual Chinesesubjects. Psychological Research 55:315-21. [BdG]

Demetriou, A., Shayer, M., Efklides, A. (1992) The nco-Piagetian theories ofcognitive development. Routledge. [LS]

Dennett, D. C. (1993) Learning and labeling. Peer commentary on Clark &Karmiloff-Smith "The cognizer's innards." Mind and Language 8(3):540-

48. [aAK-S, TD]Dent, C. H. (1990) An ecological approach to language development: An

alternative functionalism. Developmental Psychobiology 23:679-703. [TDJ]



Diamond, A. (1990) Developmental time course in human infants and infantmonkeys, and the neural bases of inhibitory control in reaching. In:The development and neural bases of higher cognitive functions, ed.A. Diamond. Annals of the New York Academy of Sciences 608:637-76. [PDZ]

Donald, M. (1991) Origins of the modern mind. Harvard UniversityPress. [MD]

(1993a) What we were, what we are becoming: Cognitive evolution. SocialResearch 60:143-70. [MD]

(1993b) Precis of Origins of the modern mind: Three stages in the evolutionof culture and cognition. Behavioral and Brain Sciences 16:737-91. [MD]

Dretske, F. (1980) The intentionality of cognitive states. Midwest Studies inPhilosophy 5:280-95. " [TD]

(1981) Knowledge and the flow of information. MIT Press. [TD]Edelman, G. (1992) Bright air, brilliant fire: On the matter of the mind. Basic

Books. [TD]Edelman, G. M., Reeke, G. N., Gall, W. E., Tononi, G., Williams, D. &

Spoms, O. (1992) Synthetic neural modelling applied to a real-worldartifact. Proceedings of the National Academy of Science, USA 89:7267-71. [TD]

Ellis, H. D. (1989) The distinction between object recognition and picturerecognition. Behavioral and Brain Sciences 12:81—82. [NHF]

Elman, J. L. (1990) Finding structure in time. Cognitive Science 14:179-211. [arAK-S]

(1991) Distributed representations, simple recurrent networks, andgrammatical structure. Machine Learning 7:195-225. [MAB, ML]

(1993) Learning and development in neural networks: The importance ofstarting small. Cognition 48 1:71-99. [aAK-S]

Elman, J. L., Bates, E., Johnson, M. H., Karmiloff-Smith, A., Parisi, D. &Plunkett, K. (in press) Rethinking innateness: Connectionism in adevelopmental framework. MIT Press. [aAK-S]

Evans, G. (1982) The varieties of reference. Oxford University Press. [JCR]Fahlman, S. E. & Lebiere, C. (1990) The cascade-correlation learning

architecture. In: Advances in neural information processing systems 2,ed. D. Touretzky. Morgan Kaufmann. [TRS]

Fifer, W. P. & Moon, C. (1988) Auditory experience in the fetus. In: Behaviorof the fetus, ed. W. P. Smotherman & S. R. Robinson. TelfordPress. [TDJ]

Fodor, J. A. (1975) The language of thought. Thomas Crowell. [arAK-S, AC](1980) Methodological solipsism considered as a research strategy in

cognitive psychology. Behavioral and Brain Sciences 3:63-73. [ML](1981) The present status of the innateness controversy. In: Representations,

ed. J. A. Fodor. MIT Press. [PB1](1983) The modularity of mind. MIT Press. [aAK-S, PBo, SO](1985) Precis of The modularity of mind. Behavioral and Brain Sciences

8(1): 1-46. [aAK-S](1987) Psychosemantics. MIT Press. [ML]

Fodor, J. A. & Pylyshyn, Z. W. (1988) Connectionism and cognitivearchitecture: A critical analysis. Cognition 28:3-71. [AC]

Foster-Cohen, S. (1994) Exploring the boundary between syntax andpragmatics: Relevance and the binding of pronouns. Journal of ChildLanguage 21(l):237-55. [SHF-C]

Freeman, N. H. (1994) Associations and dissociations in theories of mind. In:Children's early understanding of mind, C. Lewis & P. Mitchell.Erlbaum. [NHF]

(1995) The emergence of a framework theory of pictorial reasoning. In:

Drawing and looking, ed. C. Lange-Keuttner & G. Thomas. Harvester,in press. [NHF]

Freeman, N. H., Lewis, C. & Doherty, M. (1991) Preschoolers' grasp of adesire for knowledge in false-belief prediction: Practical intelligence andverbal report. British Journal of Developmental Psychology 9:139-57. [rAK-S, NHF]

Frith, U. (1989) Autism: Explaining the enigma. Blackwell. [aAK-S]Frost, D. O. & Metin, C. (1985) Induction of functional retinal projections to

the somatosensory system. Nature 317:162-64. [SO, SRQ]Frye, D., Zelazo, P. D., Brooks, P. J. & Samuels, M. C. (submitted-a)

Inference and action in early causal reasoning. [PDZ]Frye, D., Zelazo, P. D. & Palfai, T. (submitted-b) The cognitive basis of

children's theory of mind. [PDZ]Garber, P., Alibali, M. W. & Goldin-Meadow, S. (in preparation) Does

gesture reflect implicit knowledge? [SG-M]Gardner, H. (1983) Frames of mind. Basic Books. [SO]Gelman, R. (1990a) Structural constraints on cognitive development. Cognitive

Science 14:39. [aAK-S](1990b) First principles organize attention to and learning about relevant

data: Number and the animate-inanimate distinction as examples.

Cognitive Science 14:79-106. [RLC]

Gelman, R. & Gallistel, C. R. (1978) The child's understanding of number.Harvard University Press. [PBI]

Centner, D. (1983) Structure-mapping: A theoretical framework for analogy.Cognitive Science 7:155-70. [rAK-S, PBI]

Gergely, G. (1992) Developmental reconstructions: Infancy from the point ofview of psychoanalysis and developmental psychology. Psychoanalysis andContemporary Thought 15:3-55. [PBo]

Coldin-Meadow, S., Alibali, M. W. & Church, R. B. (1993) Transitions inconcept acquisition: Using the hand to read the mind. PsychologicalReview 100(2):279-97. [SG-M] '

Gopnik, A. & Wellman, H. (1992) Why the child's theory of mind really is atheory. Mind and Language 7:145-71. [NHF]

Gottlieb, G. (1976) The roles of experience in the development of behaviorand the nervous system. In: Studies on the development of behavior andthe nervous system, vol. 3: Neural and behavioral specificity, ed.G. Gottlieb. Academic Press. [TDJ]

(1981) Roles of early experience in species-specific perceptual development.In: Development of perception, vol. 1, ed. R. N. Aslin, J. R. Alberts &M. P. Peterson. Academic Press. [rAK-S] '

Craham, G. & Stephens,.C. L. (1994) Mind and mine,. In: Philosophicalpsychopathology, ed. G. Graham & G. L. Stephens. MIT Press. [GC]

Greenough, W. T., Black, J. E. & Wallace, C. S. (1987) Experience and braindevelopment. Child Development 58:539-59. [aAK-S]

Grimshaw, J. & Rosen, S. (1990) Knowledge and obedience: Thedevelopmental status of the binding theory. Linguistic Inquiry 21:187—222. [SHF-C]

Crodzinsky, J. & Reinhart, T. (1993) The innateness of binding andcoreference. Linguistic Inquiry 24:69-101. [SHF-C]

Halford, G. S. (1982) The development of thought. Erlbaum. [aAK-S]Hamad, S. (1990) The symbol grounding problem. Physica D 42:335-

46. [rAK-SRLC, TD]Hayes-Roth, F., Klahr, P. & Mostow, D. (1981) Advice taking and knowledge

refinement: An iterative view of skill acquisition. In: Cognitive skills andtheir acquisition, ed. J. Anderson. Erlbaum. [SO]

Heckhausen, H. & Beckmann, J. (1990) Intentional action and action slips.Psychological Review 97:36-48. [PDZ]

Heil, J. (1983) Perception and cognition. University of California Press,Berkeley. [TD]

Hennessy, S. (1986) The role of conceptual knowledge in the acquisition ofarithmetic algorithms. Ph.D. dissertation, University CollegeLondon. [aAK-S]

Hermelin, B. & O'Connor, N. (1986) Idiot savant calendrical calculators: Rulesand regularities. Psychological Medicine 16:885-93. [aAK-S]

(1989) Intelligence and musical improvisation. Psychological Medicine19:447-57. [aAK-S]

Hogrefe, G. J., Wimmer, H. & Perner, J. (1986) Ignorance versus false belief:A developmental lag in attribution of epistemic states. Child Development57:567-82. [NHF]

Ishiguro, H. (1972) Leibniz's philosophy of logic and language. London:Duckworth. [LS]

Ito, M. (1984) The cerebellum and neural control. Raven Press. [RG]Johnson, M. H. (1988) Memories of mother. New Scientist 18:60-62.

[aAK-S](1990) Cortical maturation and the development of visual attention in early

infancy. Journal of Cognitive Neuroscience 2:81-95. [aAK-S](1993) Constraints on cortical plasticity. In: Brain development and

cognition: A reader, ed. M. H. Johnson. Blackwell. [aAK-S]Johnson, M. H. & Bolhuis, J. J. (1991) Imprinting, predispositions and filial

preference in the chick. In: Neural and behavioral plasticity, ed. R. J.Andrew. Oxford University Press. [aAK-S]

Johnson, M. H. & Karmiloff-Smith, A. (1992) Can neural selectionism beapplied to cognitive development and its disorders? New Ideas inPsychology 10:35-46. [aAK-S]

Johnson, M. H. & Morton, J. (1991) Biology and cognitive development: Thecase of face recognition. Blackwell. [arAK-S]

Johnston, T. D. (1987) The persistence of dichotomies in the study ofbehavioral development. Developmental Review 7:149-82. [rAK-S, TDJ]

Jung, C. G. (1928) Contributions to analytical psychology. Kegan Paul. [PBo]Kaelbling, L. P. (1993) Learning in embedded systems. MIT Press. [ML]Kant, I. (1787/1933) Critique of pure reason. 2d ed. Macmillan. [TD, LS]Karmiloff-Smith, A. (1979a) A functional approach to child language: A study

of determiners and reference. Cambridge University Press. [RLC](1979b) Micro- and macro-developmental changes in language acquisition

and other representational systems. Cognitive Science 3:81-118. [DE](1981) Getting developmental differences or studying child development?

Cognition 10:151-58. [aAK-S, RLC](1983) A new abstract code or the new possibility pf multiple codes?

Behavioral and Brain Sciences 6(l):149-50. [aAK-S]



(1984) Children's problem solving. In: Advances in developmentalpsychology, vol. 3, cd. M. E. Lamb, A. L. Brown & B. Rogoff.Erlbaum. [rAK-S]

(1986) From mctaprocesses to conscious access: Evidence from children'smetalinguistic and repair data. Cognition 23:95-147. [RLC]

(1987) A developmental perspective on human consciousness. InvitedAddress, British Psychological Society Annual Conference,Sussex. [aAK-S]

(1992a) Beyond modularity: A developmental perspective on cognitivescience. MIT Press. [aAK-S, LS]

(1992b) Nature, nurture and PDP: Preposterous developmental postulates?Connection Science 4(3/4):253-69. [aAK-S]

(1992c) Abnormal behavioural phenotypes and the challenges they pose toconnectionist models of development. Technical Reports in ParallelDistributed Processing and Cognitive Neuroscience, TR.PDP.CNS.92.7,Carnegie Mellon University. [aAK-S]

(1993) Is creativity domain-specific or domain general? Clues from normaland abnormal development. In: Special issue on artificial intelligence andcreativity, eel. T. H. Dartnall. AISB Quarterly (Autumn):26-31. [TD]

Karmiloff-Smith, A., Klima, E., Bellugi, U., Grant, J. & Baron-Cohen,S. (in press) Is there a social interaction module? Language, faceprocessing and theory of mind in subjects with Williams syndrome.journal of Cognitive Neuroscience. [aAK-S]

Karmilofl'-Smith, A. & Clark, A. (1993) What's special about the humanmind/brain? A reply to Abrahamsen, Bechtel, Dennett, Plunkett, Scutt &O'Hara. Af iW and Language 8(3):569-82. [aAK-S]

Kawato, M. (1990) Computational schemes and neural network models forformation and control of multijoint arm trajectory. In: Neural networksfor control, cd. W. T. Miller III, R. Sutton & P. Werbos. MITPress. [RC]

KaWato, M., Furukawa, K. & Suzuki, R. (1987) A hierarchical neural networkmodel for the control and learning of voluntary movement. BiologicalCybernetics 57:169-85. [RG]

Keil, F. C. (1989) Concepts, kinds, and cognitive development. MITPress. [PB1]

Kirsh, D. (1991) Today the earwig, tomorrow man? Artificial Intelligence47:161-84. [JCR]

Klahr, D. (1992) Information-processing approaches to cognitive development.In: Developmental psychology: An advanced textbook, 3d ed., ed. M. H.Bornstein & M. E. Lamb. Erlbaum. [aAK-S]

Klahr, D. & Dunbar, K. (1988) Dual search space during scientific reasoning.Cognitive Science 12:148. [aAK-S]

Klahr, D., Langlcy, P. & Neches, R., eds. (1987) Production system models oflearning and development. MIT Press. [aAK-S]

Kosslyn, S. M. & Hatfield, G. (1984) Representation without symbol systems.Social Research 51:1019-45. [RLC]

Kuhii, D. (1991) The skills of argument. Cambridge University Press. [DK]Kuhn, D., Amsel, E. & O'Loughlin, M. (1988) The development of scientific

thinking skills. Academic Press. [aAK-S]Kuhn, D. & Phelps, E. (1982) The development of problem-solving strategies.

In: Advances in child development and behavior, vol. 17, ed. H. Reese.Academic Press. [aAK-S]

Kuo, Z. Y. (1921) Giving up instincts in psychology. Journal of Philosophy18:645-64. [TDJ]

(1924) A psychology without heredity. Psychological Review 31:427-48. [TDJ]

Langley, P., Simon, H. A., Bradshaw, G. L. & Zytkow, J. M. (1987) Scientificdiscovery: Computational explorations of the creative processes. MITPress. [aAK-S]

Lehrman, D. S. (1953) A critique of Konrad Lorenz's theory of instinctivebehavior. Quarterly Review of Biology 28:337-63. [TDJ]

Leslie, A. M. (1987) Pretense and representation: The origins of theory ofmind. Psychological Review 94:412-26.

(1992) Pretense, autism and the theory of mind module. Current Directionsin Psychological Science 1:18-21. [aAK-S]

Levitt, P. & O'Leary, D., eds. (1993) Mechanisms of cortical specification.Perspectives on Developmental Neurobiology 1(2):63—113. [rAK-S]

Lewis, C. (1994) Events, episodes, and narratives in the child's understandingof mind, ed. C. Lewis & P. Mitchell. Erlbaum. [rAK-S, NHF]

Macnamara, J. (1976) Stomachs assimilate and accommodate, don't they?Canadian Psychological Review 17:167-73. [PBI]

Maes, P., ed. (1990) Designing autonomous agents. MIT Press. [JCR]Maffei, L. & Galli-Resta, L. (1990) Correlation in the discharges of

neighboring rat retinal ganglion cells during prenatal life. Proceedings ofthe National Academy of Sciences 87:2861-64. [SRQ]

Mandlcr, J. M. (1988) How to build a baby: On the development of an

accessible representational system. Cognitive Development 3:113-36.|rAK-S, JCR]

(1992) How to build a baby II: Conceptual primitives. Psychological Review99(4):587-604. [aAK-S]

Mareschal, D. & Shultz, T. R. (1993) A connectionist model of thedevelopment of seriation. In: Proceedings of the Fifteenth AnnualConference of the Cognitive Science Society. Erlbaum. [TRS]

Marler, P. (1991) The instinct to learn. In: Epigenesis of the mind: Essays inbiology and knowledge, ed. S. Carey & R. Gelman. Erlbaum. [aAK-S]

Marr, D. (1982) Vision. Freeman. [JCR]Marshall, J. C. (1984) Multiple perspectives on modularity. Cognition 17:209-

42. [aAK-S]Maturana, H. R. & Varela, F. (1980) Atttopoiesis and cognition: The

realization of the living. Reidel. [RLC]McClelland, J. L. (1991) Paper given at meeting of the Society for Research in

Child Development, Seattle. [aAK-S](submitted) The interaction of nature and nurture in development: A

parallel distributed processing perspective. [rAK-S]McClelland, J. L. & Jenkins, E. (1990) Nature, nurture and connectionism:

Implications for connectionist models for cognitive development. In:Architectures for intelligence, ed. K. van Lehn. Erlbaum. [aAK-S]

McClelland, J. L. & Rumclhart, D. E. (1986) A distributed model of humanlearning and memory. In: Parallel distributed processing, ed. J.McClelland & D. Rumelhart. MIT Press. [PBI]

McKee, C. (1992) A comparison of pronouns and anaphors in Italian andEnglish acquisition. Language Acquisition 2:21-54. [SHF-C]

McNeill, D. (1992) Hand and mind. University of Chicago Press. [SG-M]Meek, W. H. & Church, R. M. (1983) A mode control model of counting and

timing processes, journal of Experimental Psychology: Animal BehaviorProcesses 9:320-34. [PBI]

Meister, M., Wong, R., Baylor, D. & Shatz, C. J. (1991) Synchronous burstsof action potentials in ganglion cells of the developing mammalian retina.Science 252:939-43. [SRQ]

Mel, B. (1988) MURPHY: A robot that learns by doing. In: Neuralinformation processing systems, ed. D. Anderson. American Institute ofPhysics, New York. (RG]

Meltzoff, A. N. (1990) Towards a developmental cognitive science: Theimplications of cross-modal matching and imitation for the development ofmemory in infancy. Annals of the New York Academy of Sciences 608:1-37. [aAK-S]

Merzenich, M. M. & Sameshima, K. (1993) Cortical plasticity and memory.Current Opinion in Neurobiology 3:187-96. [SRQ]

Montague, P. R., Dayan, P., Nowlan, S. J., Pouget, A. & Sejnowski, T. J.(1993) Using a periodic reinforcement for directed self-organization duringdevelopment. In: Advances in neural information processing systems 5,ed. C. L. Giles, S. J. Hanson & J. D. Cowan. MorganKaufmann. [SRQ]

Morais, J., Bertelson, P., Cary, L. & Alegria, J. (1986) Literacy training andspeech segmentation. Cognition 24:45-64. [DRO]

Morton, J. (1986) Developmental contingency modelling. In: Theory buildingin developmental psychology, ed. P. van Geehrt. Elsevier. [aAK-S]

Mounoud, P. & Hauert, C. A. (1982) Development of sensorimotororganization in young children: Grasping and lifting objects. In: Actionand thought: From sensorimotor schemes to thought operations, ed. G.E. Forman. Academic Press. [JCR]

Mozer, M. & Smolensky, P. (1989) Using relevance to reduce network sizeautomatically. Connection Science 1:3-17. [MAB]

Nelson, K. (1986) Event knowledge, structure and function in development.Erlbaum. [aAK-S]

(1989) Monologue as the linguistic construction of self in time. In:Narratives from the crib, ed. K. Nelson. Harvard UniversityPress. [PBo]

Neville, H. J. (1991) Neurobiology of cognitive and language processing:Effects of early experience. In: Brain maturation and cognitivedevelopment: Comparative and cross-cultural perspectives, ed. K. R.Gibson & A. C. Peterson. Aldine de Gruyter. [aAK-S]

Newell, A. (1982) The knowledge level. Artificial Intelligence 18:87-127. [AC]

(1990) Unified theories of cognition. Harvard University Press.[aAK-S]

Nguyen, D. & Widrow, B. (1989) The truck backer-upper: An example of self-learning in neural networks. In: Proceedings of the International JointConference on Neural Networks II. Institute of Electrical and ElectronicsEngineers Press. [RC]

Nijhout, H. F. (1990) Metaphors and the role of genes in development.BioEssays 12:441-46. [rAK-S, TDJ]

Norris, D. (1992) Connectionism: A new breed of bottom-up model? In:Connectionist approaches to language processing, ed. N. Sharkey &R. Reiley. Erlbaum. [BdG]

Ohlsson, S. & Rees, E. (1991) The function of conceptual understanding in

BEHAVIORAL AND BRAIN SCIENCES (1994) 17.4 743


the learning of arithmetic procedures. Cognition and Instruction 8:103-79. [rAK-S, SO]

Olson, D. R. (1994) The world on paper: The conceptual and cognitiveimplications of writing and reading. Cambridge University Press.[DRO]

Oyama, S. (1985) The ontogeny of information: Developmental systems andevolution. Cambridge University Press. [arAK-S, TDJ]

Parisi, D. (1990) Connectionism and Piaget's sensory-motor intelligence. Paperpresented at conference on Evolution and Cognition: The Heritage ofJean Piaget's Epistemology, Bergamo, Italy. [aAK-S]

Perner, J. (1991) Understanding the representational mind. MITPress. [aAK-S, NHF]

Perner, J., Leekam, S. & Wimmer, H. (1987) 3-year-olds' difficulty inunderstanding false belief: Cognitive limitation, lack of knowledge orpragmatic misunderstanding. British Journal of Developmental Psychology5:125-37. [NHF]

Perruchet, P. (in press) Learning from complex rule-governed environments:On the proper functions of nonconscious and conscious processes. In:Attention and performance XV: Conscious and nonconscious informationprocessing, ed. C. Umilta & M. Moscovitch. MIT Press. [AV]

Piaget, j . (1923) Le langage et la pens&e chez I'enfant. Neuchatel: Delachaux

et Niestle. [PBo](1926) La representation du monde chez I'enfant. Paris: Alcan. [PBo](1936) La naissance de intelligence chez I'enfant. Neuchatel: Delachaux et

Niestle. [PBo](1946) La formation du symbole chez I'enfant. Neuchatel: Delachaux et

Niestle. [PBo](1952) The child's conception of number. Routledge & Kegan Paul. [LS](1955) The child's construction of reality. Routledge & Kegan Paul. [aAK-S](1960) The general problems of the psychobiological development of the

child. In: Discuss folks' of child development, vol. 4, ed. J. Tanner & B.Inhelder. London: Tavistock. [LS]

(1967a) Biologie et connaissance. Callimard. [aAK-S](1967b) Logique et connaissance sdentifique. Encyclopddie de la Pleiade.

Paris: Callimard. [PBo](1970) Psychologie et dpistemologie. Paris: Denoel. [PBo](1976) The grasp of consciousness. Harvard University Press. [PBo](1977) Recherches sur I'abstraction rifUchissante [Studies in reflecting

abstraction], 2 vol. Presses Universitaires de France. [RLC](1977/1995) Etudes sociologiques. Geneva: Droz/Sociological studies.

London: Routledge (in press). [LS](1978) Success and understanding. Harvard University Press. [EKS](1985) The equilibration of cognitive structures. University of Chicago

Press. [SO]Piaget, J. & KarmilofT-Smith, A. (1992) A special case of inferential symmetry:

Reading a road map upside down. In: Morphisms and categories, ed.J. Piaget, G. Henriques & E. Ascher. Erlbaum. [LS]

Piatelli-Palmerini, M. (1989) Evolution, selection, and cognition: From"learning" to parameter setting in biology and the study of language.Cognition 31:1-44. [aAK-S, SRQ]

Pinker, S. (1984) Language learnability and language development. HarvardUniversity Press. [PBI]

Pinker, S. & Mehler, J., eds. (1988) Connectionism and symbol systems:Special edition. Cognition 28. [aAK-S]

Plunkett, K. & Marchman, V. (1991) U-shaped learning and frequency effectsin a multilayered perception: Implications for child language acquisition.Cognition 38:43-102. [aAK-S]

Plunkett, K. & Sinha, C. (1992) Connectionism and developmental theory.British Journal of Developmental Psychology 10:209-54. [arAK-S]

Poizner, H.t Klima, E. S. & Bellugi, U. (1987) What the hands reveal about

the brain. MIT Press. [aAK-S]Popper, K. (1963) Back to the presocratics. In: Conjectures and refutations.

Routledge & Kegan Paul. [TD]Povinelli, D. J. & deBlois, S. (1992) Young children's (Homo sapiens)

understanding of knowledge formation in themselves and others. Journalof Comparative Psychology 106:228-38. [rAK-S, NHF]

Prato Previde, E., Colombetti, M., Poli, M. & Cenami Spada, E. (1992) The

mind of organisms: Some issues about animal cognition. InternationalJournal of Comparative Psychology 6:79-100, 116-19. [AC]

Premack, D. (1975) Putting a face together. Science 188:228-36. [aAK-S]

(1986) Cavagai! Or the future history of the animal language controversy.MIT Press. [aAK-S]

(1991) "Does the chimpanzee have a theory of mind?" revisited. In:Machiavellian intelligence, ed. R. Byrne & A. Whiten. Oxford SciencePublications. [aAK-S]

Putnam, H. (1975) The meaning of "meaning." Mind, language and reality:Philosophical papers, vol. 2. Cambridge University Press. [ML]

Quartz, S. R. (1993) Nativism, neural networks, and the plausibility ofconstructivism. Cognition 48:123-44. [rAK-S, SRQ]

Read, C. (1971) Pre-school children's knowledge of English phonology.Harvard Educational Review 41(l):l-34. [DRO]

Read, C , Zhang, Y., Nie, H. & Ding, B. (1986) The apility to manipulatespeech sounds depends on knowing alphabetic reading. Cognition 24:31-44. [DRO] '

Reddy, V. (1991) Playing with others' expectations: Teasing and mucking aboutin the first year. In: Natural theories of mind, ed. A. Whiten.Blackwell. [ML]

Reeke, C. N., Jr., Sporns, O. & Edelman, G. M. (199ba) Synthetic neuralmodelling: The Darwin series of recognition automata. Proceedings of theInstitute of Electrical and Electronics Engineers 78(9): 1498-1530. [TD]

Reeke, G. N. Jr., Finkel, L. H., Sporns, O. & Edelmdn, G. M. (1990b)Synthetic neural modelling: A multilevel approach to the analysis of braincomplexity. In: Signal and sense: Local and global order in perceptualmaps, ed. C. M. Edelman, W. E. Gall & W. M. Cowan. Wiley. [TD]

Rhue, J. W. & Lynn, S. J. (1991) Fantasy proneness, fyypnotizability, andmultiple personality. In: Human suggestibility, ed, J. F. Schumaker.Routledge. [GG] i

Rosch, E. & Mervis, C. B. (1975) Family resemblances: Studies in theinternal structure of concepts. Cognitive Psychology 7:573-605. [PBI]

Rumelhart, D. E. & Norman, D. A. (1985) Representation of knowledge. In:Issues in cognitive modeling, ed. A. M. Aitkenhc^d & J. M. Slack.Erlbaum. [AC]

Rumelhart, D., Smolensky, P., McClelland, J. L. & Hinton, G. (1986)Schemata and sequential thought processes in PDP models. In: Paralleldistributed processing: Explorations in the microstructurc of cognition,vol. 2: Psychological and biological models, ed. J. L. McClelland 6:D. Rumelhart. MIT Press. [RG] ;

Rutkowska, J. C. (1984) Explaining infant perception: Insights from artificialintelligence. Cognitive Studies Research Paper 005. University ofSussex. [ML]

(1988) Perception, action and representation in infancy: A computationalapproach. Doctoral Dissertation, University of Sussex. [ML]

(1991) Looking for 'constraints' in infants' perceptual-cognitive development.Mind and Language 6:215-38. [JCR]

(1993a) Ontogenetic constraints on scaling up sensory-motor systems.Proceedings of the Second European Conference on Artificial Life 2:970-79. [JCR]

(1993b) The computational infant: Looking for developmental cognitivescience. Harvester Wheatsheaf. [ML, JCR]

Sacks, O. (1989) Seeing voices. Picador. [TD]Sainsbury, M. (1991) Logical forms. Blackwell. [LS]Sayre, K. M. (1986) Intentionality and information processing: An alternative

model for cognitive science. Behavioral and Brain Sciences 9:121-66. [TD]

Schank, R. C. & Abelson, R. P. (1977) Plans, goals aiid understanding: Aninquiry into human knowledge structures. Erlbaum. [aAK-S]

Schauble, L. (1990) Belief revision in children: The role of prior knowledgeand strategies for generating evidence. Journal of Experimental ChildPsychology 1:31-57. [aAK-S]

Schlaggar, B. L. & O'Leary, D. D. M. (1991) Potential of visual cortex todevelop an array of functional units unique to somatosensory cortex.Science 252:1556-60. [SRQ]

Schneider, W. (1987) Connectionism: Is it a paradigm shift for psychology?Behaviour Research Methods, Instruments and tiomputcrs 19(2):73-83. [aAK-S]

Schwarz, G. (1927) Uber Ruckfalligkeit bei Umgewohnung: 1. Tiel:

Ruckfalltendenz und Verwechselungsgefahr [On relapse during relcarning:Part 1: The tendency to relapse and confound elements]. PsychologischeForschung 9:86-158. [PDZ]

Searle, J. R. (1983) Intentionality. Cambridge University Press. [NHF]Sejnowski, T. & Rosenberg, C. (1986) NETtalk: A parallel network that learns

to read aloud. Johns Hopkins Electrical Engineering and ComputerScience Technical Report JHU/EEC-86/01. [TD]

Shallice, T. (1988) From neuropsychology to mental structure. CambridgeUniversity Press. [aAK-S]

Shanks, D. R. & St. John, M. F. (1994) Characteristics of dissociable humanlearning systems. Behavioral and Brain Sciences 17(3):367-447. [rAK-S,AV]

Shastri, L. & Ajjanagadde, V. (1993) From simple associations to systematicreasoning: A connectionist representation of rules, variables and dynamicbindings using temporal synchrony. Behavioral and Brain Sciences16:417-94. [rAK-S]

Shultz, T. R. (1991) Simulating stages of human cognitive development with

connectionist models. In: Machine learning: Proceedings of the eighth



international workshop, ed. L. Birnbaum & G. Collins. MorganKnufmann. [arAK-S]

Shultz, T. R., Buckingham, D. & Oshima-Takane, Y. (1994a) A connectionistmodel of the teaming of personal pronouns in English. In: Computationallearning theory and natural learning systems, ed. S. J. Hanson, M.Kearns, T. Petsche & R. L. Rivest. MIT Press. [TRS]

Shultz, T. R., Marcschal, D. & Schmidt, W. C. (1994b) Modeling cognitivedevelopment on balance scale phenomena. Machine Learning. [TRS]

Shultz, T. R. & Schmidt, W. C. (1991) A cascade-correlation model of balancescale phenomena. In: Proceedings of the Thirteenth Annual Conference ofthe Cognitive Science Society. Erlbaum. [TRS]

Shultz, T. R., Schmidt, W. C , Buckingham, D. & Mareschal, D. (in press)Modeling cognitive development with a generative connectionistalgorithm. In: Developing cognitive competence: New approaches toprocess modeling, ed. T. Simon & G. Halford. Erlbaum. [TRS]

Siegler, R. S. & Crowley, K. (1991) The microgenetic method: A direct meansfor studying cognitive development. American Psychologist 46(6):606-20. [aAK-S]

Siegler, R. S. & Robinson, M. (1982) The development of numericalunderstanding. In: Advances in child development and behavior, vol. 16,ed. H. W. Reese & L. P. Lipsett. Academic Press. [aAK-S]

Singleton, J. U, Goldin-Meadow, S. & McNeill, D. (in press) The cataclysmicbreak between gesticulation and sign: Evidence against an evolutionarycontinuum of manual communication. In: Language, gesture, and space,ed. K. Emmorey & J. Reilly. Erlbaum. [SG-M]

Skinner, B. F, (1953) Science and human behaviour. Macmillan. [aAK-S]Smith, L. (1993) Necessary knotvlcdge: Piagetian perspectives on

constructivism. Erlbaum. [LS]Smolcnsky, P. (1988) On the proper treatment of connectionism. Behavioral

and Brain Sciences 11:1-74. [aAK-S]Spelke, E. S. (1990) Principles of object perception. Cognitive Science 14:29-

56. [RLC, JCR](1991) Physical knowledge in infancy: Reflections on Piaget's theory. In:

Epigcncsis of the mind: Essays in biology and knowledge, ed. S. Carey &R. Gelman. Erlbaum. |aAK-S]

Sperber, D. 6c Wilson, D. (1986) Relevance: Communication and cognition.Harvard University Press. [SHF-C]

(1987) Precis of Relevance: Communication and cognition. Behavioral andBrain Sciences 10:697-754. [SHF-C]

Stephens, G. L. & Crnham, G. (1994) Self-consciousness, mental agency, and

the clinical psychopathology of thought-insertion. Philosophy, Psychiatry,and Psychology I. [C.G|

Sternberg, R. J. & Lubart, T. I. (1992) Buy low and sell high: An investmentapproach to creativity. Current Directions in Psychological Science1:1-5. [PDZ]

Stich, S. (1983) From folk psychology to cognitive science. CambridgeUniversity Press. [ML]

Sur, M. (1994) Cortical specification: Microcircuits, perceptual identity and anoverall perspective. Perspectives on Developmental Neurobiologyl(2):109-13. [rAK-S]

Sur, M., Garraghty, P. E. & Roe, A. \V. (1988) Experimentally induced visualprojections into auditory thalamus and cortex. Science 242:1437—41. [SRQI

Sur, M., Pallas, S. L. & Roe, A. W. (1990) Cross-modal plasticity in cortical

development: Differentiation and specification of sensory neocortex.Trends in Neuroscience 13(6):227-33. [SRQ]

Thelen, E. (1989) Self-organization in developmental processes: Can systemsapproaches work? In: Systems and development. Minnesota symposium inchild psychology, vol. 22, ed. M. Gunnar & E. Thelen. Erlbaum.[aAK-S]

Tulving, E. (1983) Elements of episodic memory. Oxford UniversityPress. [PJH]

(1985) Memory and consciousness. Canadian Psychology 26:1-12. [PJH]Van Geehrt, P. (1991) A dynamic systems model of cognitive and language

growth. Psychological Review 98:3-53. [aAK-S]Vinter, A. (in press) Hierarchy between graphic production rules in the

drawing of elementary figures. In: Advances in handwriting and drawing:A multidisciplinary approach, ed. C. Faure, P. Keuss, G. Lorette &A. Vinter. Paris: Europia. [AV]

von Hofsten, C. & Spelke, E. S. (1985) Object perception and object-directedreaching in infancy, journal of Experimental Child Psychology 114:192-212. [JCR]

Walsh, C. & Cepko, C. L. (1988) Clonally related cortical cells show severalmigration patterns. Science 241:1342-45. [SRQ]

(1992) Widespread dispersion of neuronal clones across functional regions ofthe cerebral cortex. Science 255:434-40. [SRQ]

Weiskrantz, L. (1987) Neuroanatomy of memory and amnesia: A case formultiple memory systems. Human Neurobiology 6:93-105. [BdG]

Wexler, K. & Culicover, P. (1980) Formal principles of language acquisition.MIT Press. [SRQ]

Wollheim, R. (1993) The mind and its depths. Harvard UniversityPress. [NHF]

Wynn, K. (1990) Children's understanding of counting. Cognition 36:155-93. [PB1]

(1992a) Children's acquisition of the number words and the countingsystem. Cognitive Psychology 24:220-51. [PB1]

(1992b) Evidence against empiricist accounts of the nature of numericalknowledge. AfiW and Language 7:315-32. [PB1, LS]

Yoon, M. G. & Yoon, M. (1993) Correct prediction of intentional actiondevelops earlier than attribution of false belief in three-year-olds.Unpublished manuscript, Dalhousie University. [NHF]

Zelazo, N. A., Zelazo, P. R., Cohen, K. & Zelazo, P. D. (1993) Specificity ofpractice effects on elementary neuromotor patterns. DevelopmentalPsychology 29:686-91. [PDZ]

Zelazo, P. D., Frye, D. & Reznick, J. S. (1993) Age-related changes in theexecution of explicit rules: The roles of logical complexity and executivefunction. Submitted. [PDZ]

Zelazo, P. D. & Reznick, J. S. (1991) Age-related asynchrony of knowledgeand action. Child Development 62:719-35. [PDZ]

Zelazo, P. D. & Spinazzola, J. F. (1992) Representational flexibility andresponse control in a search task. Infant Behavior and Development(special International Conference on Infant Studies) 15:785. [PDZ]

Zelazo, P. R., Weiss, M. J. & Leonard, E. (1989) The development of unaidedwalking: The acquisition of higher-order control. In: Challenges todevelopmental paradigms, ed. P. R. Zelazo & R. Barr. Erlbaum. [PDZ]

Zelazo, P. R., Zelazo, N. & Kolb, S. (1972) "Walking" in the newborn. Science176:314-15. [PDZ]


Think how you think ...

What is Intelligence?What isligence?

Edited byJean Khalfa

Edited by JEAN KHALFA

How we define intelligence, whatit consists in, how it evolved, andhow we can enhance it in thefuture, are all questions addressedby the eight expert contributors

to this remarkable book.

The volume originates in theseventh annual series of DarwinCollege Lectures, delivered inCambridge in 1992.

What is Intelligence? will be ofinterest to linguists, behaviouralbiologists, and psychologists, aswell as cognitive and brain scien-tists, and to anyone who wishes toexamine the nature and power ofhuman intelligence.

Contributors include S IMH A AROM on musical intelligence), G BO KG E ]BUTTERWORTH on perception k infants DAN DENNETT on language jandintelligence, RICHARD GREGORY on the intelligence of the eye, NiCnotASMACKINTOSH on animal intelligence, RO^ER FEN ROSE on mathematicalintelligence, ROGER SCHANK and LAWRENCE BXRNBAXTM on artificial!intelligence* and DAN SPERBBR on understanding verbal understanding'̂

NKW ISSAYS BY

Richard Gregory

Roger Schank

George liuuerworih

Roger Penrose

Dan Sperber

Nicholas Mackintosh

Si mil a Arom

Daniel Dennett

£19.95 HB 0 521 43307 X 208 pp.

CAMBRIDGEUNIVERSITY PRESS

The Edinburgh Building, Cambridge CB2 2RU

Documents

Beyond Modularity: A Developmental Perspective on Cognitive Science