Attributing the creation of new features to categorization requirements implies that the exemplars displayed are correctly assigned to their category. This constraint limits the scope of Schyns et al.'s proposal to supervised learning. We present data suggesting that this constraint is unwarranted and we argue that feature creation is better thought of as a byproduct of the attentional, on-line processing of incoming information.

Schyns et al. make a strong case for context-dependent feature discovery. The features computed from specialized and diverse data-sets help to coordinate their activity by adapting so as to emphasize what is related across sets. Their perspective can be strengthened and extended by formal arguments for the contextual guidance of learning and processing and by neurobiological and psychological evidence of structures and processes that implement this guidance.

Arguments for feature creation receive support from studies of young infants forming category representations from perceptual experience. A challenge for Schyns et al. will be to determine how a feature creation system might interface with a perceptual system that appears constrained to follow organizational principles that specify how edge segments should be grouped into functional units of coherent object representations.

Converging psychophysical evidence suggests that the human visual system parses shapes into component parts for the purposes of object recognition. We examine the Schyns et al. claim of “creation” of features in light of recent work on part-based representations of visual shape, particularly the perceptual rules that human vision uses to parse shapes.

Research in expert categorization is consistent with the Schyns et al. claim that functional features are determined by constraints imposed by the learning history of the categorizer and the demands of the categorization task. However, the expertise work also suggests that a distinction should be drawn between the categorizer's perceptions of the constituent parts of the object and its functional features. Although experts and novices may parse a domain-specific object into the same parts, their featural interpretations of those parts may differ significantly.

Schyns, Goldstone & Thibaut present reasonable arguments for feature creation in category learning. We argue, however, that they do not provide unequivocal evidence either for the necessity or for the occurrence of feature creation. In an effort to sharpen the debate, we take the stand that a fixed feature approach is to be preferred in the absence of compelling evidence.

This commentary questions the validity of the claim that new features can be constructed out of nothing during categorization. A minimal set of fixed features based on what human beings are able to detect is sufficient for categorization.

Does feature evolution stop once we have acquired sufficient features to perform a recognition task? With extended practice, novices may develop a more sophisticated feature space that allows them to perform more accurately or quickly. Our work on perceptual expertise indicates that feature learning and reorganization can continue even after an initial set of features is available to represent a novel class of objects.

The origin of features from nonfeatural information is a problem that should concern all theories of object categorization and recognition, not just the flexible feature approach. In contrast to the idea that new features must originate from combinations of simpler fixed features, we argue that holistic features can be created from a direct imprinting on the visual medium. Furthermore, featural descriptions can emerge from processes that by themselves do not operate on feature detectors. Once acquired, features can be decomposed into component features if required by other categorizations. We therefore argue that it is not necessary to separate holistic and componential approaches to representations, because the latter is a development of the former. The requirements for representational flexibility outstrip the performance of any existing computational models, but specific mechanisms of feature creation are discussed and evaluated. Challenges for feature creation mechanisms are discussed together with the constraints (perceptual, statistical, functional, and task) they will need to satisfy.

Concepts are highly theoretical entities. One cannot study them empirically without committing oneself to substantial preliminary assumptions. Among the competing theories of concepts and categorization developed by psychologists in the last thirty years, the implicit theoretical assumption that what falls under a concept is determined by description (“descriptionism”) has never been seriously challenged. I present a nondescriptionist theory of our most basic concepts, “substances,” which include (1) stuffs (gold, milk), (2) real kinds (cat, chair), and (3) individuals (Mama, Bill Clinton, the Empire State Building). On the basis of something important that all three have in common, our earliest and most basic concepts of substances are identical in structure. The membership of the category “cat,” like that of “Mama,” is a natural unit in nature, to which the concept “cat” does something like pointing, and continues to point despite large changes in the properties the thinker represents the unit as having. For example, large changes can occur in the way a child identifies cats and the things it is willing to call “cat” without affecting the extension of its word “cat.” The difficulty is to cash in the metaphor of “pointing” in this context. Having substance concepts need not depend on knowing words, but language interacts with substance concepts, completely transforming the conceptual repertoire. I will discuss how public language plays a crucial role in both the acquisition of substance concepts and their completed structure.

Millikan's account of concepts is applicable to questions about concepts in nonhuman animals. I raise three questions in this context: (1) Does classical conditioning entail the possession of simple concepts? (2) Are movement property concepts more basic than substance concepts? (3) What is the empirical content of claiming that concept meanings do not necessarily change as dispositions change?

Although our concepts of “Mama,” “milk,” and “mice” have much in common, the suggestion that they are identical in structure in the mind of the prelinguistic child is mistaken. Even infants think about objects as different from substances and appreciate the distinction between kinds (e.g., mice) and individuals (e.g., Mama). Such cognitive capacities exist in other animals as well, and have important adaptive consequences.

Identifying objects as members of ontological domains activates category-specific processes. There is evidence that these processes include particular ways of “tracking” substances and could do all the “tracking” necessary for concept acquisition. There may be no functional need or evolutionary scenario for a general tracking capacity of the kind described by Millikan.

Simulations with neural networks living in a virtual environment can be used to explore and test hypotheses concerning concepts and language. The advantages that result from this approach include (1) the notion that a concept can be precisely defined and examined, (2) that concepts can be studied in both nonverbal and verbal artificial organisms, and (3) concepts have properties that depend on the environment as well as on the organism's adaptive behavior in response to the environment.

Millikan's nondescriptionist approach applies an account of meaning to concepts in terms of designation. The essentialism that provides the principal grounds for rigid designation, however, receives no empirical support from concepts. Whatever the grounding, this view not only faces the problems of rigid designation in theories of meaning, it also calls for a role for pragmatics more consonant with descriptionist theories of concepts.

Terms designating substances and kinds function grammatically much like proper names of individuals. This supports Ruth Millikan's theory, but it also poses the question of how we can understand the reference of kind terms when the ontological status of the kind term is uncertain or disputed.

Millikan proposes that there are mapping functions through which spoken sentences represent reality. Such mappings seem to depend on thoughts that words express and on concepts as components of such thoughts, but such concepts would conflict with Millikan's other claims about concepts and language.

Millikan contrasts her substance-based view of concepts with “descriptionism” according to which description determines what falls under a concept. Focusing on her discussion of the role of language in the acquisition of concepts, I argue that descriptions cannot be separated from perception in the ways Millikan's view requires.

Millikan's account of how we acquire our most basic concepts might be clarified by a better ontological taxonomy, especially one that distinguishes between natural kinds on the one hand and wholes composed of parts on the other. The two have a different causal basis, which is important because once classification goes beyond the stage of naive induction, it becomes fundamentally etiological.

Externalism cannot work as a theory of concepts without explaining how we reidentify substances as being of the same kind. Yet this process implies just the level of descriptive content to which externalism seeks to deny a role in conceptual content.