Two comments on Semantic Cognition (Rogers & McClelland Reference Rogers and McClelland2004) are offered. The first speaks to Chapter 4, “Emergence of Category Structure in Infants,” and the second addresses Chapter 5, “Naming Things: Privileged Categories, Familiarity, Typicality, and Expertise.”
What counts as semantic categorization in infants?
In Chapter 4, Rogers & McClelland (R&M) examine whether the concepts of infants are semantic, where semantic information is defined as information “not available more or less directly from the perceptual input” (Semantic Cognition, p. 2). To address this issue, R&M consider reports of global categorization by infants (Mandler & McDonough Reference Mandler and McDonough1993; Pauen Reference Pauen2002a). Some readers like myself may have questions about the evidence and the definition.
The responding of the 11-month-olds in Pauen (Reference Pauen2002a) and Mandler and McDonough (Reference Mandler and McDonough1993) should be revisited. R&M observe that these infants generalized habituation from the familiarized category instances to the novel instance of the familiar category, and dishabituated to the novel instance of the novel category. However, this was not the pattern of results reported in the original papers. The 11-month-olds dishabituated to the novel instance from the familiar category and the novel instance from the novel category, although more so to the latter. Mandler and McDonough interpreted these results as evidence that the infants recognized the perceptual difference between the novel and familiar instances from the familiarized category, and recognized that the novel instance from the novel category was from a novel conceptual category. But this interpretation may be questioned because the infants did not provide a critical behavioral signature of categorization, namely, equivalent responding to instances of the familiarized category. One might reply that there was positive evidence for categorization observed among Mandler and McDonough's 11-month-olds presented with planes versus birds, but that could have been carried by perceptual cues: planes with silver wheels and vertical tail fins versus birds with textured wings depicting ruffled feathers. Findings that 12-month-olds categorize animals versus vehicles based on texture differences are consistent with this suggestion (Smith & Heise Reference Smith, Heise and Burns1992).
However, for argument sake, let us assume that the 11-month-olds had produced unambiguous evidence of categorization. Would that allow one to conclude that infant categorization was semantic? Given the controls in Pauen (Reference Pauen2002a), performance may well have been influenced by knowledge that had accrued prior to the experiment. But this should not be surprising given that the infants would have had 11 months to utilize perceptual input systems and a general learning mechanism to acquire a database of information about objects in the world. And that knowledge may be perceptual. In the case of animals, stored data may include information about faces, coloring, skeletal appendages, a body shape bounded by curved contours, movement patterns, and species-specific sounds of communication.
This observation raises the question of whether conceptual information of the sort emphasized by Mandler (Reference Mandler2000) and Carey (Reference Carey2000) (and subsequently in development by Gelman [Reference Gelman2003] and Keil [Reference Keil1989]) should be deemed as necessary to conclude that semantic categorization has occurred. From this commentator's perspective, knowledge about perceptible parts and properties of objects can be semantic knowledge. One could not have much of a concept of cats, for example, without knowing what they looked like and what parts they had. Concepts must include perceptual information, or else they would not be very helpful. Even school children learning about biology must be able to recognize cats in various poses and contexts. It is hard to imagine how a child could even map more abstract attributes acquired through language (i.e., cats have cat DNA) onto their correct object referents without having category representations available from perceptual experience to serve as support structures. By this view, information about perceptual properties seems just as semantic as information about genetics (Hampton et al. Reference Hampton, Estes and Simmons2007), and the way forward is to explain how perceptual and conceptual knowledge are integrated (not dissociated) to form mature concepts (Murphy Reference Murphy2002; Quinn Reference Quinn2004b).
The infant who mistook a cat for a person
In Chapter 5, R&M describe how increased experience with a particular category leads word learners to extend the label for that category to less familiar, similar categories, although not to less familiar, dissimilar categories. In a corresponding simulation, R&M trained a model with dog patterns appearing more frequently than other mammals. Early in learning, the model extended the label dog to goats and even robins, but not to trees. R&M also relate how increased experience with a category that has been linked with expertise acquisition leads to increased ability to differentiate within that category. A matching simulation demonstrated how birds (or fish) became more distinct for a model that was trained with birds (or fish) more frequently than other animals. Importantly, the different tendencies associated with increased experience, towards generalization and differentiation, are explained in a common framework in which the semantic space devoted to the more experienced class becomes larger than that allocated to less experienced classes.
The behavior of word learners and experts that is captured in R&M's simulations may also be observed in infants categorizing humans and nonhuman animals. In particular, 3- to 4-month-olds familiarized with humans generalize familiarization not only to novel humans, but also to cats, horses, and even fish, although they differentiate cars. By contrast, same-aged infants familiarized with horses generalize familiarization to novel horses, but differentiate humans, fish, and cars (Quinn & Eimas Reference Quinn and Eimas1998). In addition, infants represent humans as subordinate-level exemplars, but represent nonhuman animals (i.e., cats) as summary-level information.
Consistent with R&M's approach, the differences in how infants represent humans versus nonhuman animals have been attributed to infants' greater experience with humans inclusive of repeated encounters with parents, siblings, and caregivers (Mermillod et al. Reference Mermillod, French, Quinn, Mareschal, Alterman and Kirsh2004). This experience leads to a larger representational space for humans, thereby allowing individual humans to be represented at the subordinate level and enabling generalization to nonhuman animals. The trends toward increased specificity and generality resonate with findings that experts recognize instances from their domain at more specific levels than novices (Tanaka Reference Tanaka2001), and are also better able to recognize commonalities across their domain (Murphy & Wright Reference Murphy and Wright1984). These observations in turn suggest a sense in which infants' knowledge of humans may constitute an initial domain of perceptual expertise (Quinn Reference Quinn, Gershkoff-Stowe and Rakison2005). The specificity and generality effects are further in accord with the dual trends of differentiation and coalescence observed in the trajectory of learning produced in R&M's network simulations.
Subsequent research has suggested that infants' generalization from humans to nonhuman animals is rooted in holistic-configural structure, that is, a head attached to an elongated body with skeletal appendages (Quinn Reference Quinn2004a; Quinn et al. Reference Quinn, Lee, Pascalis and Slater2007), which is consistent with the finding that experts perceive objects within their domain holistically (Gauthier & Tarr Reference Gauthier and Tarr2002). Humans, cats, horses, and fish (but not cars), share this abstract resemblance, and R&M's networks demonstrated the ability to compute such global structure, extracting commonalities emerging in the pattern of coherent covariation. Notably, such structure is preserved even as the networks (and presumably infants) continue learning and begin differentiating subclasses. On this basis, generalization from humans to nonhuman animals by infants may form a foundation for the construction of a broad, domain-level concept of animal, and may also be a precursor of how children first rely on a human prototype to reason about biological knowledge (Carey Reference Carey1985). Thus, in accord with R&M, differences in the frequency of exposure to different classes (humans vs. nonhuman animals), coupled with the similarity structure of those classes (attributes shared by humans and nonhuman animals, but not cars), are important determinants of the growth of category organization during early cognitive development.
ACKNOWLEDGMENTS
Preparation of this commentary was supported by NIH Grants HD-42451 and HD-46526. The author thanks Gregory L. Murphy for helpful comments on an earlier draft.
