From neural constructivism to cognitive constructivism:The steps to be taken

Quartz & Sejnowski’s (Q&S’s) model for constructive learning agrees with the basic assumptions of mainstream cognitive developmental theories. However, it does not detail the neural equivalents of (1) the process of cognitive change per se, (2) the construction and functioning of thought modules, and (3) the involvement of “mindreading” and “mindsteering” in constructive learning. Specifying these equivalents is necessary if cognitive developmental neuroscience and mainstream cognitive development are to be directly connected. Quartz & Sejnowski (1997t) (Q&S) advance a number of stimulating assumptions about the neural basis of cognitive development. I will focus only on those assumptions which are important from the point of view of mainstream theory of cognitive development. My aim is to show the problems that Q&S would have to solve before their constructivist manifesto could be adopted by developmentalists primarily interested in the development of thought and reasoning rather than in neural phenomena as such. Q&S assume that learning is a general-purpose adaptational mechanism that operates on a general-purpose or equipotential substrate (the cortex) to generate structural and representational specificity. Thus, the specificity of environmental structures is reflected in the specificity of cortical areas and neural networks and this is reflected in the specificity of representational structures. This picture of mind is perfectly compatible with current models of cognitive development. These models recognize that there are general mechanisms of cognitive change that govern the construction of new concepts and mental operations. They also accept that the mind involves specialized thought structures or modules (Case & Okamoto 1996; Demetriou et al. 1993; Karmiloff-Smith 1992). It is accepted that cognitive change requires—in a hierarchylike sequence—comparison of representations, abstraction of common elements, reorganization of these elements into new representations, and some kind of symbolic individuation of the new representations so that they can be activated in the future as such. How does Q&S’s constructive learning account for these steps? That is, what are the neural equivalents of comparing representations, abstracting their commonalities, assembling new representations, and symbolically individuating them? How are geometric principles of information processing design applied at the successive phases of representational change? These questions need answers if Q&S’s general learning theory is to be adopted by cognitive developmentalists. Explaining modularity is no less important. It was discussed in a number of the first round commentary. However, many crucial questions about modularity still remain unanswered. I agree with Q&S that “it is important to turn attention back to examining environmental structure” (sect. 4.3.1). The mind and the environment are functionally and structurally tuned to each other. As a result, there are thought systems that specialize in the representation and processing of different domains in the environment. Our research (e.g., Demetriou et al. 1993) has identified five thought domains: categorical, quantitative, spatial, causal, and propositional. These domains are computationally specific because they reflect the peculiarities of the elements and relations of the environment domains to which they are affiliated. Admittedly, there is no general agreement as to the identity of domains. Other scholars specify domains on the basis of ontological rather than relational criteria and they speak about thought modules that deal with the physical, the biological, and the psychological world (Karmiloff-Smith 1992). The two kinds of modules may be interrelated. For instance, there are categorical, causal, or mathematical relations in each of these three ontological domains (Demetriou, in press). However, we still do not know how the two types of domains are interwoven in the thinker’s mind, because each type of relation functions idiosyncratically in each ontological doBEHAVIORAL AND BRAIN SCIENCES (2000) 23, 781–792 Printed in the United States of America © 2000 Cambridge University Press 0140-525X/00 $12.50 781 Continuing Commentary Commentary on Steven R. Quartz & Terrence J. Sejnowski (1997). The neural basis of cognitive development: A constructivist manifesto. BBS 20:537–596. {steve; terry}@salk.edu www.cnl.salk.edu/CNL Abstract of the original article: How do minds emerge from developing brains? According to “neural constructivism,” the represen-of the original article: How do minds emerge from developing brains? According to “neural constructivism,” the representational features of cortex are built from the dynamic interaction between neural growth mechanisms and environmentally derived neural activity. Contrary to popular selectionist models that emphasize regressive mechanisms, the neurobiological evidence suggests that this growth is a progressive increase in the representational properties of the cortex. The interaction between the environment and neural growth results in a flexible type of learning: “constructive learning” minimizes the need for prespecification in accordance with recent neurobiological evidence that the developing cerebral cortex is largely free of domain-specific structure. Instead, the representational properties of the cortex are built by the nature of the problem domain confronting it. This uniquely powerful and general learning strategy undermines the central assumption of classical learnability theory, that the learning properties of a system can be deduced from a fixed computational architecture. Neural constructivism suggests that the evolutionary emergence of neocortex in mammals is a progression toward more flexible representational structures, in contrast to the popular view of cortical evolution as an increase in innate, specialized circuits. Human cortical postnatal development is also more extensive and protracted than generally supposed, suggesting that the cortex has evolved so as to maximize the capacity of environmental structure to shape its structure and function through constructive learning. main. For instance, physical causality requires the transmission of energy. Psychological causality does not require any energy or the intervention of any medium. Imagine mood variations caused by the memory of an unpleasant encounter. Q&S’s neural constructivism is general enough to be compatible with any of the two types of modularity sketched above. Naturally, we expect more from a cortex-centered theory of learning. Specifically, we expect it to be able to clearly shift the balance in favor of the one or the other of them or to show how the two can be integrated into an overarching model. To do so, the theory would have to provide answers to the following questions: 1. What is it that directs the construction of differentiable neural circuits that correspond to the cognitive modules mentioned above? This question is important because, according to the model, the learning mechanism is general and the brain is initially equipotential. Thus, we need to know how the differentiation is effected at the very beginning. Nativist conceptions do not face this problem because they posit that different circuits are in place from the start. 2. But even if this construction process is satisfactorily understood, we would still need to know the neural parameters of the various operations that are generated by the construction and which define the functioning of the different modules. For example, what does it mean in terms of synaptic communication, conduction velocities and the like to perform a mental rotation as contrasted to the addition of two numbers? Are these mental operations different simply because they are carried out by different neural circuits which function in the same way or are the functional aspects of the various circuits also different? Or, alternatively, one and the same circuit can carry out different mental operations by varying any of the parameters of neuronal functioning? 3. Last but not least, most cognitive developmentalists would accept some kind of nonstationarity and constructivism. However, Q&S’s version of them ignores something that is probably very important in human cognitive development. This is mindreading (Baron-Cohen 1995) and, I would add, “mindsteering.” People are aware of their own and others’ minds and they systematically try to steer minds to preselected directions from a very early age (Karmiloff-Smith 1992). This implies that mindsteering must somehow be part of the neural construction process and that variations in how it is effected causes variations in the generation of environment-specific representations. But where is it located? Is it part of representationally specific circuits or does it possess its own circuits? What neural events does it require? To paraphrase Eccles (1994), how does the self control its brain? Answering these questions would provide Q&S’s manifesto with its real constructivist dimension. Infant perception and cognition and the initial architecture of constructivist models