Annette Karmiloff-Smith
"It is less illogical than it first appears to speak of instincts for inventiveness." (Marler 1991, p. 63)
1. TAKING THE DEVELOPMENTAL PERSPECTIVE SERIOUSLY
Beyond Modularity: A Developmental Perspective on Cognitive Science (Karmiloff-Smith, 1992a) aims not only to reach developmental psychologists, but also to persuade students and scientists in other areas of cognitive science - philosophy, anthropology, linguistics, ethology, adult cognitive psychology, neuroscience, computer science - to treat cognitive development as a serious theoretical science contributing to the discussion of how the human mind/brain develops and is organized internally, and not as merely a cute empirical database about what age external behavior can be observed. Nowadays much of the literature focuses on what cognitive science can offer the study of development. In Beyond Modularity, I concentrate on what a developmental perspective can offer cognitive science and attempt to pinpoint what is specifically human about human cognition.
As Piaget s conception of the sensorimotor infant is being severely undermined by new paradigms for studying infancy, the battle between nativism and constructivism once again rears its rather unconstructive head. In Beyond Modularity, I do not choose between these two epistemological stances, one arguing for predominantly built-in, domain-specific knowledge, and the other for a minimum innate underpinning to subsequent domain-general learning. Rather, I submit that nativism (when redefined within a truly epigenetic perspective of genetic expression rather than genetic unfolding), on the one hand, and Piaget s constructivism, on the other, are complementary in fundamental ways, and that the ultimate theory of human cognition will encompass aspects of both. Beyond Modularity is intended to excite the reader about the possibilities that a developmental perspective embracing both domain-specific predispositions and constructivism might yield, and to demonstrate that one can attribute various innate processes/structures to the human neonate without negating the crucial roles of the physical and sociocultural environments and without jeopardizing the deep-seated conviction that we are special - creative, cognitively flexible, capable of conscious reflection, novel invention, and occasional inordinate stupidity!
Developmental psychologists who come from the Piagetian school are loath to attribute domain-specific predispositions to the human infant. Yet they would not hesitate to do so with respect to the ant, the spider, the bee, or the chimpanzee. Why would Nature have endowed every species except the human with some domain-specific predispositions? Yet, if it turns out that all species have such predispositions, that most can maintain a goal in the face of changing environmental conditions, and that most have the capacity for learning on the basis of interaction with conspecifics and the physical environment, what is special about human cognition? Is it simply that the content of knowledge differs between species? Is it language that makes humans special? Or, compared to other species, are there qualitatively different processes at work across many domains of the human mind? Does human cognitive change affect all domains of knowledge more or less simultaneously, or does development occur in a domain-specific fashion? These are some of the questions addressed in Beyond Modularity.
I argue that domain-specific predispositions give development a small but significant kickstart by focusing the young infant s attention on proprietary inputs. The early period is followed by intricate interaction with environmental input which crucially affects brain development in return as subsequent learning takes place. But development does not stop at efficient learning. A fundamental aspect of human development is the hypothesized process by which information that is in a cognitive system becomes progressively explicit knowledge to that system. I call this the Representational Redescription hypothesis (henceforth RR). Support for the theoretical discussions of chapter 1 is explored in chapters 2 through 6 calling on empirical findings on the child as a linguist, a physicist, a mathematician, a psychologist, and a notator. Each chapter concentrates first on the initial state of the infant mind/brain and on subsequent domain-specific learning in infancy and early childhood. Each chapter then goes on to explore empirical data on older children s problem solving and theory building, with particular focus on evolving cognitive flexibility and metacognition. Throughout, I place particular emphasis on the status of representations underlying different capacities and on the multiple levels at which knowledge is stored and accessible.
In chapters 7 and 8 I reconsider the reconciliation between nativism and Piaget s constructivism, and discuss the need for more formal developmental models. Here I compare aspects of the RR framework with connectionist simulations of development. I end the book by taking a final look at the RR framework and speculating on the status of representations underlying the structure of behavior in nonhumans, which - however complex their behaviors - never become redescribers of the implicit knowledge embedded in their behavior.
If our focus is on cognitive flexibility and conscious access to knowledge, why not explore the data from adult psychology? Surely adults are far more cognitively flexible than children, so what justifies a developmental perspective? Not, rest assured, the fact that child data are cute ! One only has to glance at the developmental literature to notice that many researchers are absorbed with the ages at which children reach cognitive milestones. Decades of developmental research were wasted, in my view, because the focus was entirely on lowering the age at which children could perform a task successfully, without concern for how they processed the information. I once began an article (Karmiloff-Smith 1981, p. 151) as follows: The enticing yet awful fact about child development is that children develop! Awful, because it has provoked a plethora of studies, totally unmotivated theoretically, accepted for publication in certain types of journal because the results are significant - significant STATISTICALLY, since it is indeed easy to obtain differential effects between, say, 5 and 7 year olds, but questionable as to their significance SCIENTIFICALLY. However, some researchers use the study of development as a theoretical tool for exploring the human mind/brain from a cognitive science perspective. We are not really interested in children per se but in human cognition in general which we believe can be more fully understood via its development.
A developmental perspective is essential to the analysis of human cognition because understanding the predispositions of the human mind/brain, the constraints on subsequent learning, and HOW REPRESENTATIONS CHANGE PROGRESSIVELY OVER TIME can provide subtle clues to representational format in the adult mind. The work of Spelke (1991), which I discuss in chapter 3, has been particularly influential in pointing to the importance of a developmental perspective on cognitive science. For example, the processes for segmenting visual arrays into objects are overlaid, in preschool children and adults, by other processes for recognizing object categories. But by focusing on how very young infants segment visual arrays into objects before they are able to categorize certain object kinds, Spelke is able to generate new hypotheses about how the visual system may actually function beyond infancy and in adults.
Another area in which the developmental perspective can change our view of the adult mind is with respect to the status of different types of representations. Distinctions such as declarative/procedural, conscious/unconscious, explicit/implicit, and controlled/automatic, which are often used to explain cognitive processing in adults, turn out to involve far more than a dichotomy when explored within a developmental context. But in assuming a developmental perspective we must take the notion RdevelopmentalS seriously. Paradoxically, studies on neonates and infants are often not developmental at all. Like studies on adults, they frequently focus not on change but on real-time processing within steady-state systems. It is of course essential to determine the initial state of the human mind/brain, but the notion developmental goes beyond the specification of initial predispositions. It does not simply mean a focus on learning in children of different ages rather than the adult. When making theoretical use of development within a cognitive science perspective, the specific age at which children can successfully perform a task is, to some extent, irrelevant.
The fundamental implication of a developmental perspective involves behavioral and representational change over time. I often use a later phase in a developmental sequence to understand the status of representations underlying earlier behavior - particularly in the interesting cases where child and adult behaviors are practically identical. This notion of REPRESENTATIONAL CHANGE OVER TIME is my focus throughout Beyond Modularity. It is for all these reasons that I maintain that a developmental perspective is essential to cognitive science s efforts to understand more fully the human mind.
2. IS THE INITIAL ARCHITECTURE OF THE INFANT MIND/BRAIN MODULAR?
Fodor s 1983 book The Modularity of Mind made a significant impact on developmental theorizing by suggesting how the nativist thesis and the domain-specificity of cognition are relevant to constraints on the architecture of the human mind/brain. In Beyond Modularity, I critically discuss FodorUs thesis at some length but, since it has been the subject of a BBS treatment (Fodor, 1985) it is unnecessary to reiterate all the details in the present precis. A brief summary suffices to recall that Fodor holds that the mind/brain is made up of genetically specified, independently functioning, special-purpose modules (or input systems). Each functionally distinct module has its own dedicated processes and proprietary inputs. Information from the external environment passes first through a system of sensory transducers, which transform the data into formats that each special-purpose module can process. Each module, in turn, outputs data in a common format suitable for central, domain-general processing. The modules are deemed to be hard-wired (not assembled from more primitive processes), of fixed neural architecture, domain specific, fast, autonomous, mandatory, automatic, stimulus driven, giving rise to shallow outputs; they are informationally encapsulated and insensitive 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 are inflexible and unintelligent. They are the stupidity in the machine - but they are just what a young organism might need to get initial cognition off the ground speedily and efficiently.
Fodor posits a built-in dichotomy between what is computed blindly by the modules and what the organism believes. It is in
central processing that computations relevant to the human belief system are processed, by deriving top-down hypotheses about what the world is like from the interface between the outputs of modules and what is already stored in long-term memory. In contrast with modules, Fodor considers central processing to be influenced by what the system already knows, and therefore to be relatively unencapsulated, slow, nonmandatory, controlled, often conscious, and influenced by global cognitive goals. Central processing receives outputs from each module which are automatically translated into a common representational format, a language of thought (Fodor 1976). Central processing, then, is general-purpose. It is devoted to the fixation of belief, the building up of encyclopedic knowledge, and the planning of intelligent action, in contrast to the special- purpose, domain-specific computations of modules.
While endorsing the importance of some aspects of Fodor s thesis for understanding the human mind/brain, I provide a view that differs from the notion that modules are prespecified in detail, and I question the strictness of the dichotomy that Fodor draws between modules and central processing. I also challenge Fodor s contention that the outputs of modules are automatically encoded into a single common language of thought. I focus on the argument that a crucial aspect of development involves the RR process of going beyond modularity.
3. PRESPECIFIED MODULES VERSUS A PROCESS OF GRADUAL MODULARIZATION
Fodor s detailed account of the encapsulation of modules focuses predominantly on their role in on-line processing. There is little discussion of ontogenesis. I draw a distinction between the notion of prespecified modules versus that of a process of MODULARIZATION (which, I speculate, occurs repeatedly as the product of development). Here I differ from Fodor s strict nativist conception. I hypothesize that if the human mind/brain ends up with any modular structure, then this is the result of a process of modularization AS DEVELOPMENT PROCEEDS. My position takes account of the plasticity of early brain development (Neville 1991; Johnson, 1990, 1993), suggesting that a fairly limited number of innately specified, domain-specific predispositions would be sufficient to constrain the classes of inputs that the infant mind computes. These predispositions can operate at many different levels and do not have to be limited to representational content (see Karmiloff-Smith 1992b for more recent discussion). It can thus be hypothesized that, WITH TIME, brain circuits are progressively selected for different domain- specific computations. In certain cases, relatively encapsulated modules would be formed as a product of development. In other cases, there would be more room for influence from other computations.
Only future research using on-line brain-activation studies with neonates and young infants can distinguish between the two hypotheses. If Fodor s thesis of prespecified modules is correct, such studies should show that, from the very outset (or the moment at which the infant shows sensitivity to particular forms of input), specific brain circuits are activated in response to domain-specific inputs. By contrast, if the modularization thesis is correct, activation levels should initially be relatively distributed across the brain, and only with time (and this could be a short or relatively long time during infancy, depending on the domain) would specific circuits be activated in response to domain- specific inputs. The modularization thesis allows us to speculate that, although there are maturationally constrained attention biases and domain-specific predispositions that channel the infant s early development, this endowment involves far more than mere triggering. Rather it interacts richly with, and is in return affected by, the environmental input.
Research with other species also demonstrates the brain s plasticity. In studies of the rat, for example, Greenough Black & Wallace (1987) have shown that the brain s losses and gains of synapses are a function of different types of experience. Thus, when placed merely for exercise in a treadmill, the rat shows an increase in blood capillaries in the cerebellum, but a decrease in synapses (due to pruning of existing neural pathways, because of the lack of stimulation other than physical exercise). However, when the rat is placed in a rich environment that challenges it to learn, substantial increases in dendritic growth and synaptic connectivity are generated.
Despite my reservations regarding FodorUs modularity thesis, together with a number of cognitive developmentalists I believe that Fodor s thesis has pointed to where a domain-general view of development such as Piaget s is likely to be wrong. However, in Beyond Modularity I argue for a more dynamic view of development than Fodor s modularity of mind and I challenge Fodor s dismissal of the relevance of a developmental perspective on cognitive science. Moreover, I question Fodor s oft-cited claim that "the limits of modularity are also likely to be the limits of what we are going to be able to understand about the mind" (1983, p. 126). I argue that cognitive scientists can go beyond modularity to study the more creative aspects of human cognition. But my contention is that such an endeavor is greatly enhanced by a developmental perspective on the problem.
4. DEVELOPMENT FROM A DOMAIN-GENERAL PERSPECTIVE
Fodor s nativist thesis is in sharp contrast with domain- general theories of learning, such as Piaget s constructivist epistemology once so popular in the developmental literature. Piagetian theory argues that neither processing nor storage is domain specific. Of course, implicitly at least, Piagetians acknowledge that there are different sensory transducers for vision, audition, touch, and so forth. They do not accept, however, that the transducers transform data into innately specified, domain-specific formats for modular processing. Rather, for Piagetians, all data are processed by the same mechanisms and development involves domain-general changes in representational structures.
By opposing the domain-general view to the domain-specific explanation of development, I suggest that Piaget and Behaviorism have much in common. Neither the Piagetian nor the behaviorist theory grants the infant any innate structures or domain-specific knowledge. Each grants only some domain-general, biologically specified processes: for the Piagetians, a set of sensory reflexes and three functional processes (assimilation, accommodation, and equilibration); for the behaviorists, inherited physiological sensory systems and a complex set of laws of association. These domain-general learning processes are held to apply across all areas of linguistic and nonlinguistic cognition. Piaget and the behaviorists thus concur on a number of conceptions about the initial state of the infant mind/brain. The behaviorists saw the infant as a tabula rasa with no built-in knowledge (Skinner 1953). Piaget s view of the young infant as assailed by undifferentiated and chaotic inputs (Piaget 1955) is substantially the same.
Needless to say, there are fundamental differences between these two schools. Piagetians view the child as an active information constructor; behaviorists as a passive information storer. Piagetians conceive of development as involving fundamental stage-like changes in logical structure, whereas behaviorists invoke a progressive accumulation of knowledge. However, in the light of the present state of the art in developmental theorizing, Piagetians and behaviorists have much in common in their view of the neonate s knowledge-empty mind and their claims that domain-general learning explains subsequent development across all aspects of language and cognition.
5. DEVELOPMENT FROM A DOMAIN-SPECIFIC PERSPECTIVE
The domain-specific thesis projects a very different picture of the young infant. Rather than being assailed by incomprehensible, chaotic data from many competing sources, the neonate is seen as having domain-specific predispositions allowing it to process specific types of inputs. Contrary to the Piagetian or the behaviorist theses, the domain-specific thesis gives the infant a very good start. This does not, of course, mean that nothing changes during infancy and beyond; the infant has much to learn, but subsequent learning is guided by innately specified, domain-specific principles, and these principles determine the entities on which subsequent learning takes place (Gelman 1990; Spelke 1991).
Irrespective of whether they agree with Fodor s strict modularity thesis, many psychologists now consider development to be domain specific. Much depends, of course, on what one understands by domain, and it is important not to confuse domain with module. From the point of view of the child s mind, a domain is the set of representations sustaining a specific area of knowledge: language, number, physics, and so forth. A module is an information- processing unit that encapsulates that knowledge and the computations on it. Thus, considering development domain specific does not necessarily imply considering it modular. In other words, the storing and processing of information may be domain specific without being encapsulated, hardwired, mandatory, etc. Throughout Beyond Modularity, I argue for domain specificity of development rather than modularity in the strict Fodorian sense. I retain the term "domain" to cover language, physics, mathematics, and so forth. I also distinguish microdomains such as gravity within the domain of physics and pronoun acquisition within the domain of language. These microdomains can be thought of as subcomponents within particular domains.
The need for this finer distinction of what constitutes a domain stems from the fact that I put forward a phase model of development, rather than a stage model. In a stage model, such as Piaget s, overarching changes occur more or less contemporaneously across different domains. One alternative view is that broad changes occur within a domain - for example, that a particular type of change occurs first with respect to language and later with respect to physics. The model discussed in Beyond Modularity differs from both of these conceptions. It invokes RECURRENT PHASE CHANGES at different times across different microdomains and repeatedly within each domain.
The domain specificity of cognitive systems is also suggested by developmental neuropsychology, i.e. the existence of children in whom one or more domains are spared or impaired. For example, high functioning autist individuals show a serious deficit in communication and reasoning about mental states (theory of mind), the rest of their cognition being relatively unimpaired (Frith, 1989). Individuals with Williams Syndrome, by contrast, display a very uneven cognitive profile in which language, face recognition, and theory of mind seem relatively spared, whereas number, spatial cognition and problem solving are severely retarded (Bellugi, Marks, Bihrle & Sabo, 1988; Karmiloff- Smith, Bellugi, Klima, Grant & Baron-Cohen, 1993). Whether autism and Williams syndrome involve domain-specific representational deficits or computational deficits or both, remains an open question. There are also numerous cases of idiots-savants in whom only one domain (such as drawing or calendrical calculation) functions at a high level, while capacities are extremely restricted over the totality of the rest of the cognitive system (Hermelin & O Connor, 1986). Domain-general theorists have difficulty in explaining such within-domain and across- domain dissociations.
Adult brain damage also points to domain specificity. It is remarkably difficult to find examples in the neuropsychological literature of an across-the-board, domain-general disorder (Marshall 1984), although a case could be made for an overall deficit in planning in patients with prefrontal damage (Shallice 1988). But in many instances, disorders of higher cognitive functions consequent upon brain damage are typically domain specific - that is, they affect only face recognition, number, language, or some other facility, leaving the other systems relatively intact.
So if adults manifest domain-specific damage, and if it can be shown that infants come into the world with some domain- specific predispositions, does not that mean that the nativists have won the debate over the developmentalists still ensconced on the theoretical shores of Lake Geneva (Piaget s former bastion of anti-nativism and anti-modularity)? Not necessarily, for two reasons. First, most nativist accounts call on detailed genetic unfolding, simply triggered by environmental stimuli. An epigenetic view is very different (see excellent discussion in Oyama, 1985). And second, it is important to bear in mind that the greater the amount of fixed domain-specific properties of the infant mind/brain, the less creative and flexible the subsequent system would be (Chomsky 1988). Whereas the fixed constraints provide an initial adaptive advantage, there is a tradeoff between efficiency and automaticity, on the one hand, and relative inflexibility, on the other. This leads me to a crucial point: THE MORE COMPLEX THE PICTURE WE ULTIMATELY BUILD OF THE INNATELY SPECIFIED PREDISPOSITIONS OF THE INFANT MIND, THE MORE IMPORTANT IT BECOMES FOR US TO EXPLAIN THE FLEXIBILITY OF SUBSEQUENT COGNITIVE DEVELOPMENT. It is toward such an end - exploring the flexibility and creativity of the human mind beyond the initial state - that my work in language acquisition and cognitive development has been concentrated, in an attempt to determine both the domain-specific and the domain- general contributions to development. It is implausible that development will turn out to be entirely domain specific or domain general. And, although I will need to invoke some initial constraints, development clearly involves a more dynamic process of interaction between mind/brain and environment than the strict nativist stance presupposes.
6. RECONCILING NATIVISM AND PIAGET S CONSTRUCTIVISM
What theory of development could encompass the dynamics of a rich process of interaction between mind/brain and environment? At first blush, a theory with a central focus on epigenesis and constructivism, like Piaget s, would seem the most appropriate. The notion of constructivism in Piaget s theory is the equivalent at the cognitive level of the notion of epigenesis at the level of gene expression. For Piaget both gene expression and cognitive development are emergent products of a self-organizing system that is directly affected by its interaction with the environment. Fodor (1983, p.33) uses the term constructivism very differently from Piaget. For Fodor, it is a form of empiricism, whereas Piaget argued that his constructivist genetic epistemology was an alternative to both nativism and empiricism. This general aspect of Piaget s theory, IF MORE FORMALIZED, may well turn out to be appropriate for future explorations of the notion of progressive modularization discussed above. However, much of the rest of Piaget s theory has come under a great deal of criticism.
A growing number of cognitive developmentalists have become disenchanted with Piaget s account of the infant as a purely sensorimotor organism. For Piaget the newborn has no domain-specific knowledge, merely sensory reflexes and the three domain-general processes of assimilation, accommodation, and equilibration. By contrast, the infancy research that I discuss in the first part of chapters 2-6 of Beyond Modularity suggests that there is considerably more to the initial state of the mind/brain than Piaget s theory posits. But the exclusive focus of nativists like Fodor and Chomsky on biologically specified modules suggests that they think that there is nothing of interest to say about development beyond modularity. Moreover, Fodor s concentration on input systems - he has far less to say about either output systems or central processing - does not help us to explore the ways in which children turn out to be active participants in the construction of their own knowledge.
Although for Chomsky (1988) and Spelke (1991) a nativist/modularity stance precludes constructivism, I argue that nativism and Piaget s epigenetic constructivism are not necessarily incompatible - with certain provisos. First, to Piaget s view one must add some innately specified predispositions that would give the epigenetic process a head start in each domain. This does not imply merely adding a little more domain-general structure than Piaget supposed. Rather, it means adding domain- specific biases to the initial endowment. But the second proviso for the marriage of constructivism and nativism is that the initial base involve far less detailed specifications than some nativists presuppose and a more progressive process of MODULARIZATION (as opposed to prespecified modules) where the structure of the input plays an essential role in the structure of the resulting module. Fodor does not, for instance, discuss the cases in which the operation of one of his prespecified modules cannot be triggered by its proprietary input (e.g., auditory input in the case of the congenitally deaf). We know that in such cases the brain selectively adapts and reconfigures itself to receive other (e.g., visuomanual) nonauditory inputs (Changeux 1985; Neville 1991; Poizner, Klima & Bellugi, 1987). Many cases of early brain damage indicate that there is far more plasticity in the brain than Fodor s strict modularity view would imply. The brain is not prestructured with ready-made representations which are simply triggered by environmental stimuli; it is channeled to progressively DEVELOP representations via interaction with both the external environment and its own internal environment. And it is important not to equate innateness with presence at birth or with the notion of a static genetic blueprint for maturation. Whatever innate component we invoke, it becomes part of our biological potential only through interaction with the environment; it is latent until it receives input (Johnson 1988, 1993; Marler 1991; Oyama 1985; Thelen 1989) and the input required is either relatively specific or simply in the form of environmental stimuli per se (Greenough, Black & Wallace, 1987; Johnson & Bolhuis, 1991). The interaction with the input crucially affects development of the brain in return.
Nativists argue that development follows similar paths because all normal children start life with the same innately specified structures. The role of the environment is reduced to that of a mere trigger. But the fact that development proceeds in similar ways across normal children does not necessarily mean that development must be innately specified in detail, because it is also true that all children evolve in a species-typical environment (Johnson and Morton 1991) and we are discovering that environments have far more structure than was originally thought (e.g. Elman, 1990, 1993). Thus, it is the INTERACTION between similar innate constraints and similar environmental constraints that gives rise to common developmental paths.
The proposed reconciliation of nativism and constructivism will allow us to adhere to Piaget s epigenetic-constructivist view of the developmental process, but to drop his insistence on domain generality in favor of a more domain-specific approach. Furthermore, the Piagetian focus on output systems (i.e., on the infant s and the child s ACTION ON the environment) is an important addition to the nativist s accent on input systems. But Piaget s strong anti-nativism and his arguments for across- the-board major structural stages no longer constitute a viable developmental framework.
The need to invoke domain specificity is apparent throughout Beyond Modularity. For example, domain-general sensorimotor development cannot alone explain the acquisition of language. Syntax does not simply derive from exploratory problem solving with toys, as some Piagetians claim. Lining up objects does not form the basis for word order. Trying to fit one toy inside another has nothing to do with embedded clauses. General sensorimotor activity alone cannot account for specifically linguistic constraints. If it could, then it would be difficult to see why chimpanzees, which manifest rich sensorimotor and representational abilities, do not acquire anything remotely resembling the complex structure of human language despite very extensive training (Premack 1986).
Despite these criticisms of Piaget s view of early infancy and my rejection of his stage view of development, I hope that Beyond Modularity will persuade readers that important aspects of Piaget s epistemology should be salvaged, and that there is far more to cognitive development than the unfolding of a genetically specified program simply triggered by environmental stimuli. If we are to understand the human mind, our focus must stretch well beyond any innate specifications and embrace the interaction of both domain-specific constraints and domain-general processes.
7. THE EMPIRICAL DATA
Because of the space limitations of a precis, I refer the reader to chapters 2 through 6 in Beyond Modularity for discussions of the empirical data and the literature referenced therein. New infancy research and the REPRESENTATIONAL STATUS OF INFANT KNOWLEDGE form the detailed focus of the first part of each chapter, showing the linguistic, physical, mathematical, psychological and notational domain-specific constraints on early development. Future research may lead to reinterpretations of the present infancy data, but I remain convinced that we will have to invoke some domain-specific predispositions which initially constrain the infant mind/brain. For each cognitive domain, I go on to consider data that suggest that development entails much more than the domain-specific constraints. My research strategy has always been rather different from that of developmentalists who study a given capacity from failure to partial success through to complete mastery. By contrast, I focus on an age group in each domain where the particular capacity under study is already proficient. I then attempt to trace subsequent representational change. The most important and subtle data in chaters 2-6 are, in my view, those pointing to a level of representation in which knowledge is explicitly defined (i.e., represented differently from the information embedded in special-purpose domain- specific procedures of the earlier phase) but not yet available to conscious access and verbal report. Spontaneous repairs, unsuccessful problem solving subsequent to success, redundant behaviors, and so forth (data often ignored in developmental and adult research) are all used as vital clues to this phase of development.
At several points throughout Beyond Modularity, I allude to abnormal development. Nature, alas, often presents the scientist with experiments of its own, in which different capacities are either spared or impaired. Such cases warrant study in their own right, but also help us to gain a deeper understanding of normal development and the domain specificity/modularity debate. Again, for space reasons I merely allude to them here (for more recent detailed discussion, see Karmiloff-Smith, 1992c).
Development involves, then, two complementary processes of progressive modularization and progressive explicitation. In the remainder of this precis, I will concentrate on the second of these two processes, that is on my hypothesis that development involves representational redescription, a process that increases the flexibility and manipulability of the knowledge stored in the mind, by turning information that is in the mind into progressively more explicit knowledge to the mind.
8. BEYOND DOMAIN-SPECIFIC CONSTRAINTS: HOW NEW KNOWLEDGE GETS INTO THE MIND
How does information get stored in the child s mind? I argue that there are several different ways. One is via innate specification as the result of evolutionary processes. Predispositions can either be specific or nonspecific (Johnson and Bolhuis 1991). In both cases, environmental input is of course necessary. If the innate component is specified in detail (if it ever is), then it is likely that the environment acts simply as a trigger for the organism to select one parameter or circuit over others (Changeux 1985; Chomsky 1981; Piatelli-Palmerini 1989). By contrast, when the predisposition is specified merely as a bias or a skeletal outline, then the environment acts as much more than a trigger - it influences the subsequent structure of the brain via a rich epigenetic interaction between the mind/brain and the physical/sociocultural environment (see Johnson & Karmiloff- Smith, 1992, for discussion). The skeletal outline involves attention biases toward particular inputs and a certain number of predispositions constraining the computation of those inputs.
There are several other ways in which new information gets stored in the child s mind. One occurs when the child fails to reach a goal and has to take into account information from the physical environment. Another way in which new knowledge is acquired is when the child has to take into account and to represent information provided by the socio-cultural environment, often in the form of a direct linguistic statement. These are both external sources of change from environmental input. But there are also internal sources of change. One is illustrated by the above- mentioned process of modularization in such a way that input and output processing become progressively less influenced by other processes in the brain. This causes knowledge to become more encapsulated and less accessible to other systems. But another essential facet of cognitive change goes in the opposite direction, with knowledge becoming progressively more accessible.
My claim is that a specifically human way to gain knowledge is for the mind to exploit internally the information that it has already stored, by redescribing its representations or, more precisely, by iteratively re-representing in different representational formats what its internal representations represent. This is what I hypothesize is particular to human cognition (see details in Section 9 below).
Finally, there is a form of knowledge change that is far more obviously restricted to the human species: explicit theory change, which involves conscious construction and exploration of analogies, thought experiments and real experiments, typical of older children and adults (Carey 1985; Klahr 1992; Kuhn, Amsel & O Loughlin, 1988). But I argue that this more obvious characteristic of human cognition is possible only on the basis of the more subtle prior representational redescription, which turns IMPLICIT information embedded in special-purpose procedures into EXPLICIT knowledge but which is not yet available to conscious verbal report.
To give a more tangible feel for the theoretical discussion on which I am about to embark, let us consider the pathway to learning to play the piano. There is a first period during which a sequence of separate notes is laboriously practised. The beginning pianist pays conscious attention to particular notes. There is a second period during which chunks of several notes are played together as blocks, until finally the WHOLE PIECE can be played more or less automatically. In other words, the sequence gradually becomes proceduralized (e.g. Anderson, 1980). It is something like this that I call reaching behavioral mastery. But the automaticity is constrained by 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 hypothesize, by procedural representations which are simply run off in their entirety. There is little flexibility. At best, in a third period, the learner is able to play the whole piece softer, louder, slower, or faster. The pianistUs knowledge is embedded in the procedural representations sustaining the execution. But most learners do not stop there. During a fourth period, the learner can interrupt the piece and start at, say, the third bar without having to go back to the beginning and to repeat the entire procedure from the outset.
I hypothesize that this fourth period cannot take place on the basis of the automatized procedural representations. Rather, I posit that it involves a process of representational redescription such that the knowledge of the different notes and chords (rather than simply their run-off sequence) becomes available as manipulable data. It is only after a period of behavioral mastery that the pianist can generate variations on a theme, change sequential order of bars, introduce insertions from other pieces, and so forth. This differentiates, for instance, jazz improvisation from strict adherence to sheet music. The end result is representational flexibility and control, which allows for creativity. Also important is the fact that the earlier proceduralized capacity is not lost: for certain goals, the pianist can call on the automatic skill; for others, he or she calls on the more explicit representations that allow for flexibility and creativity. (Of course, the playing of some pianists remains simply at the procedural level.)
This movement from implicit information embedded in an efficient problem-solving procedure, to rendering the knowledge progressively more explicit, is a theme that recurs throughout Beyond Modularity. And this, together with the process of modularization discussed earlier, is precisely what I think development is about: Children are not satisfied with success in learning to talk or to solve problems; they want to understand how they do these things. In seeking such understanding, they become little theorists and to do so they have to change the nature of their internal representations.
Development and learning, then, seem to take two complementary directions. On the one hand, they involve the gradual process of proceduralization and at times modularization (that is, rendering behavior more automatic and less accessible). On the other hand, they involve a process of explicitation and increasing accessibility (that is, representing explicitly information that is implicit in the procedural representations). Both are relevant to cognitive change, but the main focus of Beyond Modularity is the process of representational explicitation - which, I posit, occurs in a variety of linguistic and cognitive domains throughout development.
9. THE PROCESS OF REPRESENTATIONAL REDESCRIPTION
For a number of years I have been trying to understand how internal representations change in the course of development, even when overt behavior may look identical. In this attempt, I have developed the hypothesis of a reiterative process of REPRESENTATIONAL REDESCRIPTION (RR). I will first make some general points about the hypothesis and then provide a summary.
The notion of RR attempts to account for the way in which children s representations become progressively more manipulable and flexible. This ultimately leads, in each domain at different times, to the emergence of conscious access to knowledge and children s theory building. RR involves a cyclical process by which information already present in the organism s independently functioning, special-purpose representations is made progressively available, via redescriptive processes, to other parts of the cognitive system, first within a domain and then sometimes across domains.
The RR process is posited to occur spontaneously as part of an internal drive toward the creation of intra-domain and inter- domain relationships. Although I stress the endogenous nature of representational redescription, clearly the process may at times also be triggered by external influences.
The actual PROCESS of RR is domain general, but it is crucially affected by the form and the level of explicitness of the representations supporting particular domain-specific knowledge at a given time. When I state that RR is domain general, I do not mean to imply that it involves a simultaneous change across domains. Rather, I mean that, within each domain, the RR process operates in a similar way.
Let us now look at the RR hypothesis in some detail. Development, I argue, involves three RECURRENT phases. During the first phase the child focuses predominantly on information from the external environment. This initial learning is data driven. Phase 1 culminates in consistently successful performance on whatever microdomain has reached that level. This is what I term behavioral mastery. Behavioral mastery does not necessarily imply that the underlying representations are equivalent to the adult s, even though the behavioral output may be the same. The same performance (say, correctly producing a particular linguistic form, or managing to balance blocks on a narrow support) can be generated at various ages by very different representations. Later (phase-3) behavior may appear identical to phase-1 behavior. We thus need to draw a distinction between BEHAVIORAL CHANGE (which sometimes gives rise to a U-shaped developmental curve) and REPRESENTATIONAL CHANGE, because behavioral mastery is not tantamount to the end point of the developmental progression in a given microdomain.
Phase 1 is followed by an internally-driven phase during which the child no longer focuses on the external data. Rather, system-internal dynamics take over such that internal representations become the focus of change. In phase 2, the current state of the child s representations of knowledge in a microdomain predominates over information from the incoming data. The temporary disregard for features of the external environment during phase 2 can lead to new errors and inflexibilities. This can, but does not necessarily, give rise to a decrease in successful behavior - a U-shaped developmental curve. This is deterioration at the behavioral level, not at the representational level.
Finally, during phase 3, internal representations and external data are reconciled, and a balance is achieved between the quests for internal and external control. In the case of language, for example, a new mapping is made between input and output representations in order to restore correct usage.
But what about the format of the internal representations that sustain these reiterated phases? The RR framework argues for at least four levels at which knowledge is represented and re- represented. I have termed them Implicit (I), Explicit-1 (E1), Explicit-2 (E2), and Explicit-3 (E3). The RR framework postulates different representational formats at different levels. At level-I, representations are in the form of procedures for responding to stimuli in the external environment. A number of constraints operate on the representational adjunctions that are formed at this level:
- information is encoded in procedural form.
- the procedure-like encodings are sequentially specified.
- new representations are independently stored.
- level-I representations are bracketed, and hence no intra-
domain or inter-domain representational links can yet be
formed. Information embedded in level-I representations is therefore not available to other operators in the cognitive system. Thus, if two procedures contain identical information, this potential inter- representational commonality is not yet represented in the child s mind. A procedure AS A WHOLE is available as data to other operators; however, its COMPONENT PARTS are not. It takes developmental time and representational redescription (see discussion of level E1 below) for component parts to become available for the marking of potential intra-domain and inter- domain relationships, a process which ultimately leads (see discussion of levels E2/E3) to inter-representational flexibility and creative problem-solving capacities. But at this first level, the potential representational links and the information embedded in procedures remain implicit. This gives rise to the ability to compute specific inputs in preferential ways and to respond rapidly and effectively to the environment. But the behavior generated from level-I representations is relatively inflexible.
Level-E1 representations are the result of redescription, into a new format, of the procedurally-encoded representations at level-I. The redescriptions are abstractions and, unlike level-I representations, they are not bracketed (that is, the component parts are now open to potential intra-domain and inter-domain representational links). The E1 representations are reduced descriptions that lose many of the details of the procedurally- encoded information. As a nice example of what I have in mind here, consider the details of the grated image delivered to the perceptual system of a person who sees a zebra (Mandler, 1992). A redescription of this into striped animal (either linguistic or image-like) has lost much of the perceptual precision. To Mandler s discussion, I would add that the redescription allows the COGNITIVE (as opposed to the PERCEPTUAL) system to understand the analogy 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 and the road sign deliver very different inputs to the PERCEPTUAL system. A species without representational redescriptions would not make the analogy between the zebra and the zebra crossing sign. The redescribed representation is, on the one hand, simpler and less special purpose but, on the other, more cognitively flexible (because it is transportable to other goals and useable to make other inferences). Unlike perceptual representations, conceptual redescriptions are productive; they make possible the invention of new terms (e.g. zebrin, the antibody which stains certain classes of cells in striped patterns).
Note that the original level-I representations remain intact in the child s mind and can continue to be called for particular cognitive goals which require speed and automaticity. The redescribed representations are used for other goals where explicit knowledge is required. As representations are redescribed into the E1 format, we witness the beginnings of a flexible cognitive system upon which the child s nascent theories can subsequently be built. Level-E1 representations go beyond the constraints imposed at level-I, where procedural-like representations are simply used in response to external stimuli. Once knowledge previously embedded in procedures is explicitly defined, the potential relationships between procedural components can then be marked and represented internally. Moreover, once redescription has taken place and explicit representations become manipulable, the child can then introduce violations to her data-driven, veridical descriptions of the world - violations which allow, for instance, for pretend play, false belief, and the use of counterfactuals.
It is important to stress that although E1 representations are available as data to the system, they are not available to conscious access and verbal report. Throughout the book I examine examples of the formation of explicit representations which are not yet accessible to conscious reflection and verbal report, but which are clearly beyond the procedural level. In general, developmentalists have not distinguished between implicitly stored knowledge and El representations in which knowledge is explicitly represented but is not yet consciously accessible. Rather, they have drawn a dichotomy between an undefined notion of something implicit in behavior (as if information were not represented in any form) and consciously accessible knowledge that can be stated in verbal form. The RR framework postulates that the human representational system is far more complex than a mere dichotomy would suggest. It is particularly via a developmental perspective that one can pinpoint this multiplicity of levels of representational formats.
The RR framework posits that only at levels beyond El are conscious access and verbal report possible. At level E2, it is hypothesized, representations are available to conscious access but not to verbal report (which is possible only at level E3). Although for some theorists consciousness is reduced to verbal reportability, the RR framework claims that E2 representations are accessible to consciousness but that they are in a similar representational code as the El representations of which they are redescriptions. Thus, for example, El spatial representations are recoded into consciously accessible E2 SPATIAL representations. (We often draw diagrams of problems we cannot easily verbalize.)
At level E3, knowledge is recoded into a cross-system code. This common format is hypothesized to be close enough to natural language for easy translation into statable, communicable form. It is possible that some knowledge learned directly in linguistic form is immediately stored at level E3. Children learn a lot from verbal interaction with others. However, knowledge may be stored in linguistic code but not yet be linked to similar knowledge stored in other representational formats. Often linguistic knowledge (e.g., a mathematical principle governing subtraction) does not constrain nonlinguistic knowledge (e.g., an algorithm used for actually doing subtraction) until both have been redescribed into a similar format so that inter-representational constraints can operate (Hennessy, 1986).
The empirical examples throughout Beyond Modularity illustrate levels I, E3 and particularly the subtleties of level E1. In the book I do not distinguish between levels E2 and E3, both of which I believe involve conscious access, because thus far research has not been directly focused on level-E2 level (conscious access without verbal report). Most if not all metacognitive studies focus on verbal report (i.e., level E3). Thus level E2 remains to be tested empirically. However, I do not wish to foreclose the possibility of spatial, kinesthetic, and other NON- LINGUISTICALLY-ENCODED representations that are available to conscious access, and it may well be that E-2 and E-3 redescriptional formats are both made directly on the basis of the E1 format, rather than E3 being a redescription of E2. This is discussed fully in chapter 1 of the book.
The end result of these various redescriptions is the existence in the mind of multiple representations of similar knowledge at different levels of detail and explicitness. This notion of multiple encoding is important; the development of the mind does not seem to be a drive for economy. Indeed, the human mind may turn out to be a very redundant store of knowledge and processes.
Let me again stress the concept of reiterative developmental phases. There is no such thing as a phase E2 child. The child s representations are in different representational formats with respect to particular microdomains.
Although the process of representational redescription can occur on line, I posit that it also takes place without ongoing analysis of incoming data or production of output. Thus, change can occur outside normal input/output relations, that is simply as the product of system-internal dynamics, when there are no external pressures. I posit that representational change WITHIN phases involves adding representations. Here negative feedback (failure, incompletion, inadequacy, mismatch between input and output, etc.) plays an important role, leading progressively to behavioral mastery. But in the transition BETWEEN phases, it is hypothesized that POSITIVE feedback is essential to the onset of representational redescription. In other words, this success-based view of cognitive change posits that it is representations that have reached a stable state (the child having reached behavioral mastery) that are redescribed. Representational redescription is a process of appropriating stable states to extract the information they contain, which can then be used more flexibly for other purposes. Many of the studies discussed in Beyond Modularity, and new data from Siegler and Crowley (1991), show that change often follows success, not only failure. In other words, children explore domain-specific environments beyond their successful interaction with them.
This is not to deny the importance of instability, failure, conflict, and competition as generators of other types of change (Bates & MacWhinney 1987; Piaget 1967; Thelen 1989). It is worth reiterating this point. Competition can occur on line between different processes and cause behavioral change. But the hypothesis I develop throughout Beyond Modularity is that competition leading to representational change takes place after each of the potential competitors has been consolidated (i.e., is stable in its own right). In chapter 3, for example, it is shown how counterexamples are not taken into account (do not have the status of a counterexample) until the child s theory about a particular microdomain has been consolidated. Similar examples are to be found in the history of science and in children s strategies of scientific experimentation (Klahr & Dunbar 1988; Kuhn et al. 1988; Kuhn & Phelps 1982; Schauble 1990), as well as across the various domains of knowledge discussed throughout Beyond Modularity.
10. ARE THERE DOMAIN-GENERAL PROCESSES AT WORK?
Invoking domain-specific constraints on development does not deny the existence of some domain-general mechanisms. The infancy tasks explored in each chapter make it very clear that infants can call on complex inferential processes across different domains. Moreover, young infants go well beyond sensorimotor encodings and make use of domain-general processes such as representational redescription to recode sensorimotor input into accessible formats (see also Mandler, 1992). Domain-general processes sustaining inference and representational redescription operate throughout development. But invoking general PROCESSES that are the same across different domains is not equivalent to invoking domain-general STAGES OF CHANGE. It is the latter that Beyond Modularity rejects.
Yet I am left with a lurking feeling that there may turn out to be some across-the-board domain-general changes also, perhaps linked to major maturation of particular regions of the brain (e.g. prefrontal cortex). One such change suggested by an abundance of empirical data seems to occur around 18 months of age. This holds for several domains, particularly with respect to holding two representations simultaneously in mind and representing hypothetical events in general (Meltzoff 1990; Perner 1991) rather than theory-of-mind computations in particular (Leslie 1992). Eighteen months is the point at which Piaget too called for a change in representational structure which allowed for the onset of pretend play, language, mental imagery, etc. The precise way in which Piaget accounted for such a change in terms of the closure of a purely sensorimotor period is likely to be wrong, but the conviction that something fundamental occurs around 18 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 31/2 to 4 years. This age does not correspond to a stage change in Piagetian theory, but it has turned out to be an age at which fundamental changes seem to occur in various domains. Moreover, this age also seems to be roughly the point at which the human child differs radically from the chimpanzee. As Premack (1991, p. 164) put it, a good rule of thumb has proved to be: if the child of three and a half years cannot do it, neither can the chimpanzee.
If it turns out that across-the-board, domain-general changes do occur, we may be able to use them as a diagnostic for fundamental neural changes in the brain, and vice versa. This of course remains an open question, but the flourishing new field of developmental cognitive neuroscience may soon provide some relevant answers. However, even if some across-the-board changes were to hold, it is important to recall that their effects would be manifest somewhat differently across domains, since they would interact with domain-specific constraints. Development will not turn out to be EITHER domain specific OR domain general. It is clearly the intricate interaction of BOTH, more domain general than is presupposed by most nativist/modularity views of development, but more domain specific than Piagetian theory envisages.
So, does Piagetian theory retain any role in developmental theorizing? To me the answer is affirmative. Theories of cognitive development (and recent connectionist modeling of cognitive development [McClelland & Jenkins 1990; Parisi 1991; Plunkett & Sinha 1992], which I discuss in chapter 8) continue to draw inspiration from Piaget s EPISTEMOLOGY - his quest to understand emergent properties and his general stance with regard epigenesis and to the importance of the child s action on the environment. It is the details of his PSYCHOLOGICAL description of across-the- board stage-like changes in logico-mathematical structure that are no longer viable. I believe that it is possible to retain the essence of Piagetian theory while doing away with stage and structure. But the problem with Piaget s theory (as indeed with the RR framework too) is that it is underspecified in comparison with, say, theories expressed as computer models. I now turn briefly to this issue.
11. MODELING DEVELOPMENT
One of the aims of Beyond Modularity is to persuade cognitive scientists of the value of a developmental perspective on the workings of the human mind. Yet at the heart of much of the work in cognitive science is the use of computer models to test psychological theories. It is therefore essential to devote some space to a discussion of how the RR framework might be relevant to attempts to express developmental theories in the form of computer simulations.
What type of framework is RR? Throughout Beyond Modularity, I describe RR in verbal terms. It is, as Klahr (1992) has put it, at the soft-core end of the modeling of cognitive development, the hard-core end being the implementation of theories as computer programs. Klahr s opposition captures an important distinction between a focus on general principles of development and a focus on the specification of precise mechanisms. Klahr argues that the very process of simulating development in the form of computer programs leads to insights about the mechanisms underlying developmental change, whereas verbal descriptions generally underspecify the mechanisms. I agree. But soft-core and hard-core approaches should not be considered mutually exclusive.
In my view, soft-core approaches often lead to a broader intuitive understanding of general principles of change, whereas both the information-processing use of the flow chart and the symbolic approach to computer simulation run the risk of reifying into one or more boxes or single-named operators what is in fact the product of a highly interactive system. Nonetheless, at the hard-core end of modeling there have been a number of interesting attempts to express developmental theories in various information-processing terms - for example, in the form of scripts (Schank & Abelson 1977; Nelson 1986), of developmental contingency models (Morton 1986), and of self- modifying production systems (Klahr, Langley & Neches, 1987). In chapter 8, however, I take as my main example some recent connectionist simulations, since they seem to be the closest to the spirit of epigenesis and constructivism (for fuller discussions, see Bates & Elman, 1993; Clark & Karmiloff-Smith 1993; Elman, Bates, Johnson, Karmiloff-Smith, Parisi & Plunkett, in press; Karmiloff-Smith 1992b, 1992c; Karmiloff-Smith & Clark 1993; McClelland & Jenkins 1990; Parisi 1991; Plunkett & Sinha 1992). They also address the problems I raise in Beyond Modularity with respect to stage theories, in that they show that by incremental learning one can obtain stage-like shifts in overt behavior without the need for qualitatively different structures and mechanisms (McClelland & Jenkins 1990).
Although the connectionist framework has come under severe criticism (Pinker & Mehler 1988) and has been called a return to Associationism in high tech clothing (Jusczyk & Bertoncini 1988) and a revamping of the order from noise approach championed by Piaget (Piatelli-Palmerini 1989), a growing number of cognitive developmentalists see in it a considerable theoretical potential for explicating the more general tenets of Piaget s epistemology (e.g. Bates & Elman 1993; Bechtel & Abrahamsen 1991; Elman, Bates, Johnson, Karmiloff-Smith, Parisi & Plunkett, 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 the early 1980s, map interestingly onto features of recent connectionist simulations.
Chapter 8 of Beyond Modularity describes the main features of connectionist models, but since a BBS treatment has dealt extensively with such models (Smolenksy 1988) I will not reiterate the description in this precis. Instead I will go on directly to explore the extent to which connectionist simulations can and cannot capture what I deem to be crucial to a model of developmental change. To the extent that they can, connectionism would offer the RR framework a powerful set of hard-core tools by applying the mathematical theory of complex dynamical systems to cognitive development (van Geehrt 1991). And to the extent that connectionist models fail to model development adequately, the RR framework suggests some crucial modifications.
Many of the details of phase-1 learning, which leads to behavioral mastery and level-I representations, turn out to be captured particularly well in a connectionist model. However, the very aspect of development that Beyond Modularity focuses on-- the process of representational redescription--is precisely what seems to be missing from connectionist simulations of development.
12. CONNECTIONISM: THE STARTING STATE, THE INPUT, AND REPRESENTATIONAL REDESCRIPTION
Let us now look at some of the specific issues discussed throughout Beyond Modularity and how they can be informed by and inform the connectionist framework.
i) THE STARTING STATE: Most researchers of the connectionist persuasion take as their research strategy a non-nativist view. This makes it possible to explore the extent to which developmental phenomena can be simulated from a tabula rasa starting state - that is, from random weights and random activation levels, with no domain-specific knowledge. This has led some to interpret the results of connectionist modeling as strong evidence for the anti-nativist position. However, there is nothing about the connectionist framework that precludes the introduction of initial biased weights and connections (i.e., that are the equivalent of innately specified predispositions as a result of evolution) rather than random weights and connections. Also specific architectures, learning algorithms, learning rates, etc., which are part of the starting state, clearly affect how an input set is learned.
Various ways of simulating developmental change have been proposed. One is to start a network with a small number of hidden units and, as development proceeds, to recruit more and more units or an extra hidden layer to compress the data even further (Schultz 1991). This is rather like the neo-Piagetians notion that processing capacity increases with age (Case, 1985; Halford, 1982). Other researchers (Bechtel & Abrahamsen 1991) have suggested the equivalent of maturational change, such that the network would start by using one learning algorithm (e.g., contrastive Hebbian learning) and, with maturation, come to use a different learning algorithm (e.g., backpropagation). Incremental learning has also been used, such that the network first sees only part of the input at a time, rather than the whole input set in one go (Elman 1990, in press; Plunkett & Marchmann 1991). These are all domain-general solutions to developmental change. However, we are beginning to witness an increasing tendency on the part of connectionists to explore the ways in which domain-specific constraints might also shape learning. This, in my view, is likely to be a future focus for connectionist models of development.
It might seem at present that connectionist models deny, either implicitly or explicitly, the need for domain-specific learning. In favor of domain generality, connectionists stress that their models use the same learning algorithms for DIFFERENT categories of input presented to different networks. But in effect architectures are fine-tuned to specific types of input. For example, a recurrent architecture is used for sequential input (e.g. Elman 1990) whereas an associative network is used for concept learning (e.g. Plunkett & Sinha 1992). To my knowledge, little work has been done on networks which progressively develop their own architecture as a function of the input that they happen to process. Moreover, no single network has been presented with an array of inputs from different domains (e.g. language, spatial taks, tasks involving physical principles). Networks designed to simulate language acquisition (e.g., Elman 1990, 1993) see only linguistic strings. A similar network could be used for physics input, but the self same network could not be used without totally upsetting the language learning that has already taken place unless it continues to be trained also on the original set. In other words, the fact that each network is dedicated to a specific type of input, in a specific learning task, with a specific architecture and learning algorithm turns out to be equivalent to domain specificity in the human. Infants seem to process proprietary, domain-specific inputs separately, and so do networks. We will probably end up requiring multiple networks with different architectures and different learning algorithms.
A final point with respect to the starting state. Networks are not modules in the sense of the distinction I drew between modules and a process of modularization. In fact, networks mimic the process of modularization because, with few or no built-in representational biases, it is only as learning proceeds that they BECOME increasingly like special-purpose modules.
ii) THE INPUT: Although connectionist models have potential for developmental theorizing, they have several shortcomings. One concerns the input presented to networks. First decisions about input representation are entirely external to the network and often not motivated theoretically. Second, with some exceptions, up to now connectionists have not really modeled development; they have modeled tasks. This becomes particularly apparent if we look at the example of the balance scale that is so popular in all kinds of computer modeling, connectionist or other (Shultz 1991; McClelland & Jenkins 1990; Langley, Simon, Bradshaw & Zytkow 1987; Siegler & Robinson 1978; Newell 1990). The models have focused on children s performance on the balance-scale task, not on how children learn about physical phenomena in general in real life (see, also, Shultz 1991 for discussion). It is a fact that many children come to a balance-scale experiment with no experience of balance scales. But that does not mean that they bring no relevant knowledge to the task. They may focus on weight in tasks using the traditional balance scale because weights are what the experimenter more obviously manipulates. But in other block- balancing tasks not presented in the form of a balance scale, many young children ignore weight and focus solely on length. Children come to such tasks having already learned something about how rulers fall from tables, how see-saws work, and so forth. But a see-saw is not a balance scale. It does not have a neat line of equidistant pegs on which children of absolutely equal weight can be placed one on top of another! Development is not simply task- specific learning. It is deriving knowledge from many sources and using that knowledge in a goal-oriented way. Thus, in my view, far richer input vectors and simulation of goal-oriented behaviors are needed if we are to model the ways in which real children learn in real environments.
iii) BEHAVIORAL MASTERY: Chapter 3 on the child as a physicist and chapter 6 on the child as a notator give particularly clear examples of how a lengthy period of behavioral mastery precedes representational change. Indeed, throughout Beyond Modularity I argue that behavioral mastery is a prerequisite for representational change. However, an analysis of learning in a connectionist network already reveals in the hidden units the existence of some representation of subsequent change BEFORE it is observable in the output. This suggests a way in which connectionist modeling might 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 overt behavioral mastery.
iv) IMPLICIT TO EXPLICIT REPRESENTATIONAL CHANGE: It has often been difficult to convey, particularly to developmental psychologists, precisely what I meant by level-I implicit representations. Researchers have often used the term implicit to explain away efficient behavior that appears too early for the tenets of a particular theory. But no definition of implicit has been offered. The connectionist framework may help to give a more precise definition. Indeed, some recent connectionist simulations of language learning (Elman 1990, 1993), for instance, are particularly illustrative of the status of implicit level-I representations. Elman s model is discussed fully in Beyond Modularity. It demonstrates how grammatical function (noun/verb; transitive/intransitive verb; singular/plural and so forth) can be progressively inferred from statistical regularities of the input set and represented in the hidden units as learning proceeds. The full details of the learning process need not concern us here, but rather the STATUS OF THE REPRESENTATIONS that the network progressively builds. First, Elman shows that, as with most connectionist networks using nonlinear functions, a lengthy initial period is essential to learning. At first, the network s predictions are random. However, with time the network learns to predict, not necessarily the actual next word, but the correct category of word (noun vs verb; if noun, animate vs inanimate, edible vs non-edible; etc.), as well as the correct subcategorization frame for the next verb (transitive or intransitive) and the correct number marking on both noun and verb (singular or plural). This cannot be done by mere association between adjacent surface elements. For example, while in the case of the simple strings a network could learn to always predict that strings without an s (plural verb) follow strings with an s (plural noun), it cannot do so for embedded relative clause strings. Here a plural verb may follow a singular noun (e.g., the boys that chase the GIRL SEE the dog). In such cases, the network must make STRUCTURE-DEPENDENT predictions. Thus, the network progressively moves from processing mere surface regularities to representing something more abstract, but without this being built in as a prespecified linguistic constraint.
This seemingly impressive grammatical knowledge is only implicit in the system s internal representations. Note, however, that this does not mean that the grammatical knowledge is not represented. As in the case of early learning in the child, I would argue that it is represented in level-I format. But it is we, as external theorists, who use level-E formats to label the trajectories through weight space as nouns, verbs, subjects, objects, intransitives, transitives, plurals, singulars, and so on. The network itself never goes beyond the formation of the equivalent of stable (but unlabeled) level-I representations. In other words, it does not spontaneously go beyond its efficient behavioral mastery. It does not redescribe the representations that are stored in its activation trajectories. Unlike the child, it does not spontaneously appropriate the knowledge it represents about different linguistic categories. It cannot directly use the higher-level, more abstract knowledge for any other purpose than the one it was designed for nor directly engage in internetwork knowledge transfer because its representations are input/task specific. The notion of, say, nounhood always remains implicit in the network s system dynamics. The child s initial learning is like this, too. But, as several examples throughout Beyond Modularity show, children go on to spontaneously redescribe their linguistic (and other) knowledge. This pervasive process of representational redescription gives rise to the manipulability and flexibility of the human representational system.
Now, it is not difficult to build a network, inspired by RR, that would redescribe stable states in weight space such that the implicit information represented in trajectories could be used as knowledge by the same or other networks. But this would suggest a change in the architecture of the network, involving perhaps the creation of special nodes not implicated in other aspects of the on-line processing. Furthermore, the RR framework suggests that what is abstracted during the redescriptive process involves a loss of detail and a gain in accessibility. Thus, one would not want the entire trajectories of the network to be redescribed, but rather the product of the most important ones. (This would be equivalent to, say, labeling the phase-state portraits of the principal-component analysis.) The RR framework postulates that redescribed knowledge capturing abstract notions such as verb and noun must be in a DIFFERENT FORMAT than the original level-I representations. In other words, redescriptions would have to be in a representational format usable across networks which had previously processed DIFFERENT representations at the input level. Hence the need for representational redescription into different (level-E) formats. Simple copies of level-I representations would not be useable/transportable from one network to another, because they would be too dependent on the specific features of their inputs.
In chapter 2, I discuss a particularly relevant example of what progressive RR might look like in the human case. When 3-6- year-olds are asked to repeat the last word that the experimenter had said before a story was interrupted, some of the youngest subjects (3 years old) could not do the task at all despite lengthy modeling and help from the experimenter. Yet their fluent language and their lack of segmentation errors suggest that they do represent formal word boundaries for the majority of words they use and understand, but they are not yet ready to go beyond that behavioral mastery. There were other children (4-5 years old) who could not do the task immediately but who, with one-off modeling for a few open-class words, were able IMMEDIATELY to extend the notion of word to all open-dass and closed-class categories. Their level-I representations were ready for level-E1 redescription triggered by the experimenter from outside. However, slightly older children (5-6 years) who had never had a grammar lesson had spontaneously undergone the redescriptive process on their own. They showed immediate success, even on the practice story. Finally, 6-7-year-olds representations showed signs of having undergone further redescription into the E3 format; these children were able to consciously access their knowledge and to provide verbal explanations as to what counts as a word and why. I deem this process of MULTIPLE redescription of knowledge that becomes increasingly accessible to different parts of the system to be an essential component of human development and one that connectionist modelers need to take into account.
It seems plausible that connectionist models can indeed lend precision to an account of what I have called phase-1 learning - the phase that results in behavioral mastery (i.e., the period of rich interaction with the environment during which level-I representations are built and consolidated). However, there is much more to development than this. I have intimated at various points that connectionist simulations stop short of what I deem to be certain essential components of human development. Indeed, whereas behavioral mastery is the endpoint of learning in connectionist models, in the RR framework it is the starting point for new flexibility, i.e., for generating redescriptions of implicitly defined level-I representations. Up to now connectionist models of development have had little to say about how to move from implicit representations to explicit ones, an essential process called for by RR. How could a network appropriate its own stable states? Clark (1989), Dennett (1993), and McClelland (1991) have argued that all that would have to be added to a connectionist network is another network that uses the equivalent of public language, implying that the only difference between implicit and explicit knowledge is that the latter is linguistically encoded. However, I have provided numerous examples of children s knowledge that is explicitly represented but which they cannot articulate linguistically. The RR framework posits a far more complex view of multiple levels of representational redescription, of which language is but one manifestation. Finally, the fact that most connectionist models blend structure and content makes it difficult for the network to exploit knowledge components. Yet in several chapters I show that children extract knowledge components from the procedural representations in which they are embedded, re-represent them, and use them in increasingly manipulable ways.
It remains an open question how representational redescription might be modeled in a connectionist network. Can it be done simply by adding layers to the architecture of a single network, or by creating, say, a hierarchy of interconnected networks? Should a node, external to the on-line processing, be gradually fed with information from the developing internal representations when hidden units reach a certain threshold of stability? How can internet relations be introduced while keeping in mind the constraints suggested by RR of common transportable representational formats? Or will we have to opt for hybrid models containing both parallel distributed processing and more classical sequential manipulation of discrete symbols (see discussions in Karmiloff-Smith 1987, 1992b, 1992c; Clark & Karmiloff-Smith 1993; Karmiloff-Smith & Clark, 1993; Schneider 1987)? As connectionist networks become more complex, I think that the issue of whether something is truly hybrid will lose relevance. Future developmental modeling must, in my view, simulate both the benefits of rapid processing via implicit representations and those gained by further representational redescription - a process which, I posit, makes possible human creativity (for a BBS treatment of creativity, see Boden in press).
14. CONCLUDING REMARKS
I started Beyond Modularity by distinguishing between the representations that sustain complex behavior and the things that a given species can do with that complexity. My argument throughout has been that, far more pervasively than even its near cousin the chimpanzee, the human mind exploits its representational complexity by re-representing its implicit knowledge into many levels of explicit form. Thereby the knowledge becomes usable beyond the special-purpose goals for which it is normally used and representational links across different domains can be forged.
I claim that this is rarely if ever true of other species. The plover (discussed in chapter 5), for example, displays a complex set of behaviors to keep competitors at bay that, in human terms, would be called deceit. But these behaviors (keeping competitors away from their hatching eggs) are not available even for other closely related purposes (keeping competitors away from food). What about the chimpanzee, with whom we share close to 100 percent of our genetic makeup? Do chimpanzees, like children, play with knowledge, just as they play with physical objects and conspecifics? According to discussions I have had with Premack, there are no obvious indicators of representational redescription in the behavior of the chimpanzee. There are numerous examples of how the chimpanzee goes beyond a specified task; for example, when the task is to assemble the pieces of a puzzle of a chimp face, a chimpanzee might, after succeeding, add extra pieces as decoration to form a hat or a necklace (Premack 1975). But Premack could find no example that revealed that the chimpanzee spontaneously analyzes the components of its successful behavior in the way a child does. It is, of course, not immediately obvious how we would recognize representational redescription in the chimpanzee if it did exist. The higher levels of redescription (into, say, linguistic format) are obviously ruled out. But we know that in many instances children develop explicit representations (E1) which lie between the implicit representations and the verbally reportable data. In the child, level E1 representational redescription is frequently manifest after overt behavioral mastery. The chimpanzee, by contrast, seems to be content to continuously repeat its successes; it does not go beyond behavioral mastery. Yet throughout Beyond Modularity examples are explored of how human children spontaneously seek to understand their own cognition, and that this leads to the sort of representational manipulability that eventually allows them to become folk linguists, physicists, mathematicians, psychologists, and notators.
My speculation is that either the process of representational redescription is not available to other species or, if it is (perhaps to the chimpanzee), the higher-level codes into which representations are translated during redescription are very impoverished. It is plausible that language-trained chimpanzees will show signs of representational redescription. But this would be due, not to the existence of a language-like code per se, but to the possibility of redescription into any other more explicit code (for fuller discussion, see Karmiloff-Smith 1983).
RR is fundamentally a hypothesis about the specifically human capacity to enrich itself from within by exploiting knowledge it has already stored, not by just exploiting the human and physical environment. Intra-domain and inter-domain representational relations are the hallmark of a flexible and creative cognitive system. The pervasiveness of representational redescription in human cognition is, I maintain, what makes human cognition specifically human. This is, of course, a challenge to ethologists and one I look forward to pursuing in the future. What indices should we be seeking in other species? What machinery would we have to add to the plover, the ant, the spider, the bee, or the chimpanzee to make the process of representational redescription possible?
In the final pages of Beyond Modularity, I present a caricature drawing of the difference between humans and other species. In the top half of the caricature is drawn a human and an animal in reciprocal interaction with the external environment. In the bottom half, the drawing shows just the human figure with an arrow going around the head from one side to the other. This albeit rather silly caricature is intended to illustrate that level-I representations exist as cognitive tools allowing an organism (human or nonhuman) to act on the environment and be affected by it in return. The second part of the figure is not meant to suggest that, in the human, knowledge goes in one ear and out the other! Rather, it is a reminder that, in the human, INTERNAL REPRESENTATIONS become objects of cognitive manipulation such that the mind extends well beyond its environment and is capable of creativity. Let me go as far as to say that the RR process is, in Marler s terms, one of the human instincts for inventiveness.
In concluding Beyond Modularity, and even this short precis, I hope to have convinced the reader that the flourishing new domain of cognitive science needs to go beyond the traditional nativist- empiricist dichotomy that permeates much of the field, in favor of an epistemology that embraces both innate predispositions and constructivism. And cognitive science has much to gain by going beyond modularity and taking developmental change seriously.
A precis necessarily makes conceptual leaps, misses out the richness of the empirical data as well as numerous references to relevant literature (which are to be found in the book), and leaves no room for the humour, alas. Hopefully it has none the less given a relatively complete idea of the theoretical issues raised. I began this precis with a quote from Marler that I find particularly conducive to my thinking and that I used as a colophon for one of the book chapters. But I began the actual book with a quote from Fodor and ended with the following one:
Deep down, I m inclined to doubt that there is such a thing as cognitive development in the sense that developmental cognitive psychologists have in mind. (Fodor, 1985, p. 35)
If this precis has encouraged you to read the book in full, I hope that by the time you reach the end, deep down you will disagree with FodorUs statement and, with me, you will argue that development goes far beyond the triggered unfolding of a genetic program, that where modularity occurs it is the result of a gradual process of modularization, and that representational redescription allows the human mind to go beyond modularity.
Anderson, J. R. 1980. Cognitive Psychology and Its Implications. Freeman.
Bates, E. and Elman, J.L. 1993. Connectionism and the study of change. In M. H. Johnson (ed.), Brain Development and Cognition: A Reader. Oxford: Blackwell.
Bates, E., and MacWhinney, B. 1987. Competition, variation and language learning. In B. MacWhinney (ed.), Mechanisms of Language Acquistion. Hillsdale, NJ: Erlbaum.
Bechtel, W., and Abrahamsen, A. 1991. Connectionism and the Mind: An Introduction to Parallel Processing in Networks. Oxford: Blackwell.
Bellugi, U., Marks, S., Bihrle, A. M., and Sabo, H. 1988. Dissociation between language and cognitive functions in Williams Syndrome. In D. Bishop and K. Mogford (eds.), Language Development in Exceptional Circumstances. Churchill Livingstone.
Boden, M.A. in press. Precis of The Creative Mind: Myths and Mechanisms. Behavioral and Brain Sciences.
Carey, S. 1985. Conceptual Change in Childhood. Cambridge, MA: MIT Press/Bradford Books.
Case, R. (1985) Intellectual development: Birth to adulthood. New York: Academic Press.
Changeux, J. P. 1985. Neuronal Man: The Biology of Mind. Pantheon.
Chomsky, N. 1981. Lectures on Government and Binding. Foris.
Chomsky, N. 1988. Language and Problems of Knowledge. Cambridge, MA: MIT Press.
Clark, A. 1989. Microcognition: Philosophy, Cognitive Science, and Parallel Distributed Processing. Cambridge, MA: MIT Press.
Clark, A., and Karmiloff-Smith, A. 1993. The Cognizer's innards: a psychological and philosophical perspective on the development of thought. Mind and Language.
Dennett, D. C. in press. Learning and Labeling. Peer commentary on Clark & Karmiloff-Smith "The Cognizer's Innards". Mind and Language.
Elman, J. L. 1990. Finding structure in time. Cognitive Science 14:179- 211.
Elman, J.L. (1993). Learning and development in neural networks: The importance of starting small. Cognition, 48, 1, 71-99.
Elman, J.L., Bates, E., Johnson, M.H., Karmiloff-Smith, A., Parisi, D. and Plunkett, K. in press. Rethinking Innateness: Connectionism in a Developmental Framework. Cambridge, MA: MIT Press.
Fodor, J. A. 1976. The Language of Thought. Harvester.
Fodor, J. A. 1983. The Modularity of Mind. Cambridge, MA: MIT Press.
Fodor, J.A. 1985. Precis of The Modularity of Mind. Bebavioral and Brain Sciences, 8, 1, 1-46.
Frith, U. 1989. Autism: Explaining the Enigma. Oxford: Blackwell.
Gelman, R. 1990. Structural constraints on cognitive development. Cognitive Science 14: 39.
Greenough, W.T., Black, J.E., and Wallace, C.S. 1987. Experience and brain development. Child Development 58: 539-559.
Halford, G.S. 1982. The development of thought. Hillsdale, NJ: Erlbaum.
Hennessy, S. 1986. The role of conceptual knowledge in the acquisition of arithmetic algorithms. Unpublished Doctoral Dissertation, University College London.
Hermelin, B., and O'Connor, N. 1986. Idiot savant calendrical calculators: Rules and regularities. Psychological Medicine 16: 885-893.
Hermelin, B., and O'Connor, N. 1989. Intelligence and musical improvisation. Psychological Medicine 19: 447-457.
Johnson, M. H. 1988. Memories of mother. New Scientist 18 (February): 60-62.
Johnson, M. H. 1990. Cortical maturation and the development of visual attention in early infancy. Journal of Cognitive Neuroscience 2: 81-95.
Johnson, M. H. 1993. Constraints on cortical plasticity. In M. H. Johnson (ed.), Brain Development and Cognition: A Reader. Blackwell.
Johnson, M. H., and Bolhuis, J. J. 1991. Imprinting, predispositions and filial preference in the chick. In R. J. Andrew (ed.), Neural and Behavioural Plasticity. Oxford: Oxford University Press.
Johnson, M. H., and Karmiloff-Smith, A. 1992. Can neural selectionism be applied to cognitive development and its disorders? New Ideas in Psychology 10: 35-46.
Johnson, M. H., and Morton, J. 1991. Biology and Cognitive Development: The Case of Face Recognition. Blackwell.
Jusczyk, P. W., and Bertoncini, J. 1988. Viewing the development of speech perception as an innately guided learning process. Language and Speech 31: 217-238.
Karmiloff-Smith, A. 1981. Getting developmental differences or studying child development? Cognition 10: 151-158.
Karmiloff-Smith, A. 1983. A new abstract code or the new possibility of multiple codes? Behavioral and Brain Sciences, 6 (1):149-150.
Karmiloff-Smith, A. 1987. A developmental perspective on human consciousness. Invited Address, British Psychological Society Annual Conference, Sussex.
Karmiloff-Smith, A. 1992a. Beyond Modularity: A developmental perspective on cognitive science. Cambridge, MA: MIT Press.
Karmiloff-Smith, A. 1992b. Nature, Nature and PDP: Preposterous Developmental Postulates? Connection Science, 4, 3/4, 253-269.
Karmiloff-Smith, A. 1992c. Abnormal behavioural phenotypes and the challenges they pose to Connectionist models of development. Technical Reports in Parallel Distributed Processing and Cognitive Neuroscience, TR.PDP.CNS.92.7, Carnegie Mellon University.
Karmiloff-Smith, A., Bellugi, U., Klima, E., Grant, J. & Baron-Cohen, S. 1993. Is there a social interaction module? Language, face processing and theory of mind in subjects with Williams syndrome. (Submitted).
Karmiloff-Smith, A. & Clark, A. 1993. What's special about the human mind/brain? A reply to Abrahamsen, Bechtel, Dennett, Plunkett, Scutt & O'Hara. Mind and Language.
Klahr, D. 1992. Information-processing approaches to cognitive development. In Bornstein, M. H., and Lamb, M. E. (eds.), Developmental Psychology: An advanced textbook, third edition. Hillsdale, NJ: Erlbaum.
Klahr, D., and Dunbar, K. 1988. Dual search space during scientific reasoning. Cognitive Science 12: 1 48.
Klahr, D., Langley, P., and Neches, R. (eds.), 1987. Production System Models of Learning and Development. Cambridge, MA: MIT Press.
Kuhn, D., Amsel, E., and O'Loughlin, M. 1988. The Development of Scientific Thinking Skills. Academic Press.
Kuhn, D., and Phelps, E. 1982. The development of problem-solving strategies. In H. Reese (ed.), Advances in Child Development and Behavior, volume 17. Academic Press.
Langley, P., Simon, H. A., Bradshaw, G. L., and Zytkow, J. M. 1987. Scientific Discovery: Computational Explorations of the Creative Processes. Cambridge, MA: MIT Press.
Leslie, A. M. 1992. Pretense, autism and the theory of mind module. Current Directions in Psychological Science, 1, 18-21.
Mandler, J. M. 1992. How to build a baby Il: Conceptual primitives. Psychological Review, 99, 4, 587-604.
Marler, P. 1991. The instinct to learn. In S. Carey and R. Gelman (eds.), Epigenesis of the Mind: Essays in Biology and Knowledge. Erlbaum.
Marshall, J. C. 1984. Multiple perspectives on modularity. Cognition 17: 209-242.
McClelland, J. L. 1991. Paper given at meeting of the Society for Research in Child Development, Seattle.
McClelland, J. L., and Jenkins, E. 1990. Nature, nurture and connectionism: Implications for connectionist models for cognitive development. In K. van Lehn (ed.), Architectures for Intelligence. Hillsdale, NJ: Erlbaum.
Meltzoff, A. N. 1990. Towards a developmental cognitive science: The implications of cross-modal matching and imitation for the development of memory in infancy. Annals of the New York Academy of Sciences, 608:1-37.
Morton, J. 1986. Developmental contingency modelling. In P. van Geehrt (ed.), Theory Building in Developmental Psychology. Elsevier.
Nelson, K. 1986. Event Knowledge, Structure and Function in Development. Erlbaum.
Neville, H. J. 1991. Neurobiology of cognitive and language processing: Effects of early experience. In K. R. Gibson and A. C. Petersen (eds.), Brain Maturation and Cognitive Development: Comparative and Cross-Cultural Perspectives. Aldine deGruyter.
Newell, A. 1990. Unified Theories of Cognition. Cambridge, MA: Harvard University Press.
Oyama, S. 1985. The Ontogeny of Information: Developmental Systems and Evolution. Cambridge University Press.
Parisi, D. 1990. Connectionism and Piaget's sensory-motor intelligence. Paper presented at conference on "Evolution and Cognition: The Heritage of Jean Piaget's epistemology, Bergamo, Italy.
Perner, J. 1991. Understanding the Representational Mind. MIT Press.
Piaget, J. 1955. The Child's Construction of Reality. Routledge and Kegan Paul.
Piaget, J. 1967. Biologie et Connaissance. Gallimard.
Piatelli-Palmerini, M. 1989. Evolution, selection, and cognition: >From "learning" to parameter setting in biology and the study of language. Cognition 31:1-44.
Pinker, S., and Mehler, J. (eds.) 1988. Connectionism and symbol systems: Special edition. Cognition 28.
Plunkett, K., and Marchman, V. 1991. Ushaped learning and frequency effects in a multilayered perceptron: Implications for child language acquisition. Cognition 38: 43-102.
Plunkett, K. and Sinha, C. (1992) Connectionism and developmental theory. British Journal of Developmental Psychology. 10, 209-254.
Poizner, H., Klima, E. S., and Bellugi, U. 1987. What the Hands Reveal about the Brain. MIT Press.
Premack, D. 1975. Putting a face together. Science 188: 228-236.
Premack, D. 1986. Gavagai! Or the Future History of the Animal Language Controversy. MIT Press.
Premack, D. 1991. "Does the chimpanzee have a theory of mind?" revisited. In R. Byrne and A. Whiten (eds.), Machiavellian Intelligence. Oxford: Oxford Science Publications.
Schank, R. C., and Abelson, R. P. 1977. Plans, Goals and Understanding: An Inquiry into Human Knowledge Structures. Hillsdale, NJ: Erlbaum.
Schauble, L. 1990. Belief revision in children: The role of prior knowledge and strategies for generating evidence. Journal of Experimental Child Psychology 1: 31-57.
Schneider, W. 1987. Connectionism: Is it a paradigm shift for psychology? Behaviour Research Methods, Instruments and Computers 19 (2): 73-83.
Shultz, T. R. l991. Simulating stages of human cognitive development with connectionist models. In L. Birnbaum and G. Collins (eds.), Machine Learning: Proceedings of the Eighth International Workshop. Morgan Kaufmann.
Shallice, T. 1988. From Neuropsychology to Mental Structure. Cambridge, UK: Cambridge University Press
Siegler, R. S. and Crowley, K. 1991. The microgenetic method: A direct means for studying cognitive development. American Psychologist, 46, 6, 606-620.
Siegler, R. S., and Robinson, M. 1982. The development of numerical understanding. In H. W. Reese and L. P. Lipsett (eds.), Advances in Child Development and Behavior, vol. 16. Academic.
Skinner, B. F. 1953. Science and Human Behaviour. Macmillan.
Smolensky, P. (1988) On the proper treatment of connectionism. Behavioral and Brain Sciences, 11, 1-74.
Spelke, E. S. 1991. Physical knowledge in infancy: Reflections on Piaget's theory. In S. Carey and R. Gelman (eds.), Epigenesis of the Mind: Essays in Biology and Knowledge. Hillsdale, NJ: Erlbaum.
Thelen, E. 1989. Self-organization in developmental processes: Can systems approaches work? In M. Gunnar and E. Thelen (eds.), Systems and Development. Minnesota Symposium in Child Psychology, vol. 22. Hillsdale, NJ: Erlbaum.
Van Geehrt, P. 1991. A dynamic systems model of cognitive and language growth. Psychological Review 98: 3-53.