Coherent covariation appears to be a powerful explanatory factor accounting for a range of phenomena in semantic cognition. But its role in accounting for the crosslinguistic facts is less clear. Variation in naming, within the same semantic domain, raises vexing questions about the necessary parameters needed to account for the basic facts underlying categorization.

Rogers & McClelland (R&M) present a powerful account of semantic (conceptual) learning. Their model admirably handles many characteristics of early concept formation, but it also needs to address attentional biases, and distinguish direct input from error-driven learning, and fast versus slow learning. Not distinguishing implicit and explicit knowledge means that the authors also cannot explain why some coherently varying information becomes accessible and other information does not.

Rogers & McClelland's (R&M's) précis represents an important effort to address key issues in concepts and categorization, but few of the simulations deliver what is promised. We argue that the models are seriously underconstrained, importantly incomplete, and psychologically implausible; more broadly, R&M dwell too heavily on the apparent successes without comparable concern for limitations already noted in the literature.

Analogical inferences are an important consequence of the way semantic knowledge is represented, that is, with relations as explicit structures that can take arguments. We review evidence that this feature of semantic cognition successfully predicts how quickly and broadly children's concepts change with experience and show that Rogers & McClelland's (R&M's) parallel distributed processing (PDP) model fails to simulate these cognitive changes due to its handling of relational information.

Rogers & McClelland (R&M) have provided an impressive outline of the capabilities of a class of multi-layered perceptrons that mimic many aspects of human knowledge acquisition. Despite this success, in the literature several basic issues are raised and concerns are expressed. Indeed, the problems are so acute that a different way of thinking is called for. In this commentary it is suggested that rational models approach provides a promising alternative.

This commentary considers the issues of what should be taken as evidence for semantic categorization in infants and why infants display a surprising asymmetry in the categorization of humans versus nonhuman animals. It is argued that perceptual knowledge should be viewed as a potent source of information for semantic categorization, and that the asymmetrical categorization behavior arises as a consequence of the frequency and similarity structure of experience.

We agree with many theoretical points presented by Rogers & McClelland (R&M), especially the role of domain-general learning of coherent covariation. Nonetheless, we argue that in failing to be informed by key aspects of development, including the role of labels on categorization and the emergence of constraints on learning, their model fails to capture important features of the ontogeny of knowledge.

In proposing that their model can operate in the concrete, perceptual world, Rogers & McClelland (R&M) have not done justice to the complexities of the ecological sphere and its invariance laws. The structure of concrete events forces a different framework, both for retrieval of events and concepts defined across events, than that upon which the proposed model, rooted in essence in the verbal learning tradition, implicitly rests.

Any theory of semantic cognition is also a theory of concepts. There are two ways to construe the models presented by Rogers & McClelland (R&M) in Semantic Cognition. If we construe the input and output representations as concepts, then the models capture knowledge acquisition within a stable set of concepts. If we construe the hidden-layer representations as concepts, the models provide a simulation of conceptual change.

Logically, weighting is transitive, but similarity is not, so clustering cannot be either. Entailments must help a child to review attribute lists more efficiently. Children's understanding of exceptions to generic claims precedes their ability to articulate explanations. So agency, as enabling constraint, may show coherent covariation with attributes, as mere extensional, observable effect of intensional entailments.

The commentaries reflect three core themes that pertain not just to our theory, but to the enterprise of connectionist modeling more generally. The first concerns the relationship between a cognitive theory and an implemented computer model. Specifically, how does one determine, when a model departs from the theory it exemplifies, whether the departure is a useful simplification or a critical flaw? We argue that the answer to this question depends partially upon the model's intended function, and we suggest that connectionist models have important functions beyond the commonly accepted goals of fitting data and making predictions. The second theme concerns perceived in-principle limitations of the connectionist approach to cognition, and the specific concerns these perceived limitations raise for our theory. We argue that the approach is not in fact limited in the ways our critics suggest. One common misconception, that connectionist models cannot address abstract or relational structure, is corrected through new simulations showing directly that such structure can be captured. The third theme concerns the relationship between parallel distributed processing (PDP) models and structured probabilistic approaches. In this case we argue that there the difference between the approaches is not merely one of levels. Our PDP approach differs from structured statistical approaches at all of Marr's levels, including the characterization of the goals of cognitive computations, and of the representations and algorithms used.