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 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 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 cortex has evolved so as to maximize the capacity of environmental structure to shape its structure and function through constructive learning.

Although the developmental arguments in the Quartz & Sejnowski (Q&S) target article may have intrinsic merit, they do not warrant the authors' conclusion that innate modular architectures are absent or minimal, and that neocortical evolution is simply a progression toward more flexible representational structures. Modular architectures can develop and evolve in tandem with sub-cortical specialisation. I present comparative evidence for the co-evolution of specific thalamic and cortical visual pathways.

Constructivism is the most recent in a long line of failed attempts to discredit nativism. It seeks support from true (but irrelevant) facts, wastes its energy on straw men, and jumps logical gaps; but its greatest weakness lies in its failure to match nativism's explanation of a wide range of disparate phenomena, particularly in language acquisition.

Much of our work with enriched experience and training in animals supports the Quartz & Sejnowski (Q&S) thesis that environmental information can interact with pre-existing neural structures to produce new synapses and neural structure. However, substantial data as well as an evolutionary perspective indicate that multiple information-capture systems exist: some are constructivist, some are selectionist, and some may be tightly constrained.

The present commentary addresses the Quartz & Sejnowski (Q&S) target article from the point of view of the dynamical learning algorithm for neural networks. These techniques implicitly adopt Q&S's neural constructivist paradigm. Their approach hence receives support from the biological and psychological evidence. Limitations of constructive learning for neural networks are discussed with an emphasis on grammar learning.

Quartz & Sejnowski's target article concentrates on the development of a number of neural parameters, especially neuronal processes, in the mammalian brain. Data on learning-related changes in spines and synapses in the developing avian brain are consistent with a constructivist interpretation. The issue of an integration of selectionist and constructivist views is discussed.

Quartz & Sejnowski's (Q&S's) constructivist manifesto promotes a return to an extreme form of empiricism. In defense of learning by selection, we argue that at the neurobiological level all the data presented by Q&S in support of their constructive model are in fact compatible with a model comprising multiple overlapping stages of synaptic overproduction and selection. We briefly review developmental studies at the behavioral level in humans providing evidence in favor of a selectionist view of development.

Quartz & Sejnowski (Q&S) disregard evidence that suggests that their view of dendrites is inadequate and they ignore recent results concerning the role of neurotrophic factors in synaptic remodelling. They misrepresent neuronal selectionism and thus erect a straw-man argument. Finally, the results discussed in section 4.2 require neuronal proliferation, but this does not occur during the period of neuronal development of relevance here.

Developmental psychology should play an essential constraining role in developmental cognitive neuroscience. Theories of neural development must account explicitly for the early emergence of knowledge and abilities in infants and young children documented in developmental research. Especially in need of explanation at the neural level is the early emergence of meta-representation.

The radical empiricist theory of the Quartz & Sejnowski target article would result in a brain that could not act. The attempt to bolster this position with computational arguments is misleading and often just wrong. Fortunately, other efforts are making progress in linking neural and cognitive development.

The multiple levels of analysis that Quartz & Sejnowski (Q&S) bring to bear on the phenomenon of activity-driven dendritic growth show the tight linkage of explanations from the cellular to the cognitive level. To show how multiple control regimes can intersect at the same site, I further elaborate an example of a developmental problem solved at the axodendritic connection: that of population matching.

Quartz & Sejnowski's (Q&S's) main accomplishment is the presentation of increasing complexity in the developing brain. Although this cuts a colorful swath through current theories of learning, it leaves the central question untouched: How does the environment direct neural structure? In answer, Q&S offer us only Hebb's half-century-old suggestion once again.

I agree with Quartz & Sejnowski's (Q&S's) points, which are familiar to many scientists. A number of models with the sought-after properties, however, are overlooked, while models without them are highlighted. I will review nonstationary learning, links between development and learning, locality, stability, learning throughout life, hypothesis testing that models the learner's problem domain, and active dendritic processes.

Processing limitations can be an advantage for a learner (Elman 1993; Newport 1990; Turkewitz 1982). They can filter input to the learner so that the relations to be learned increase in complexity only gradually. The time-course of filtered input can complement the growing neural representations discussed by Quartz & Sejnowski.

There is evidence for increase, followed by decline, in synaptic numbers during development. Dendrites do not function in isolation. A constructive neuronal process may underpin a selectionist cognitive process. The environment shapes both ontogeny and phylogeny. Phylogenetic natural selection and neural selection are compatible. Natural selection can yield both constructivist and selectionist solution to adaptuive problems.

Increased complexity of representations in development probably results from the differentiation of distributed neural circuits. Axonal differentiation plays a crucial role in this process. Axonal differentiation appears to be achieved in stages, each involving combinations of constructive and regressive events controlled by cell intrinsic and cell extrinsic information.

We add to the constructivist approach of Quartz & Sejnowski (Q&S) by outlining a specific classification of sources of constraint on the emergence of representations from Elman et al. (1996). We suggest that it is important to consider behavioral constructivism in addition to neural constructivism.

A credible account of the neurobiology underlying cognitive development cannot afford to ignore the recently demonstrated innate regionalisation of the neocortex as well as the ontogeny of corticocortical phenomena, only for the latter does the timing of development permit control by external events and this is most likely to occur at later stages in the fine tuning of cortical microcircuitry.

Missing from Quartz & Sejnowski's (Q&S's) unique and valuable effort to relate cognitive development to neural constructivism is an examination of the global emergent properties of adding new neural circuits. Such emergent properties can be studied with computational models. Modeling with generative connectionist networks shows that synaptogenic mechanisms can account for progressive increases in children's representational power.

Although the idea that cognitive structure changes as we learn is welcome, a variety of mathematical structures are needed to model the neural and cognitive processes involved. A specific example of bodily-kinaesthetic intelligence is given, building on a formalism given elsewhere. As the structure of cognition changes, previous learning can become tacit, adding to the complexity of cognition and its modeling.