Prototype and Exemplar Accounts of Category Learning and Attentional Allocation: A Reassessment.

Abstract

In a recent article, J. P. Minda and J. D. Smith (2002) argued that an exemplar model provided worse quantitative fits than an alternative prototype model to individual subject data from the classic D. L. Medin and M. M. Schaffer (1978) 5/4 categorization paradigm. In addition, they argued that the exemplar model achieved its fits by making untenable assumptions regarding how observers distribute their attention. In this article, we demonstrate that when the models are equated in terms of their response-rule flexibility, the exemplar model provides a substantially better account of the categorization data than does a prototype or mixed model. In addition, we point to shortcomings in the attention-allocation analyses conducted by J. P. Minda and J. D. Smith (2002). When these shortcomings are corrected, we find no evidence that challenges the attention-allocation assumptions of the exemplar model.A classic issue in the categorization literature has been whether people represent categories in terms of abstracted prototypes or in terms of specific exemplars. According to prototype models, people represent categories in terms of some central tendency computed over the category training instances and classify objects on the basis of how similar they are to the prototypes of the alternative categories (Homa & Vosburgh, 1976;Posner & Keele, 1968;Reed, 1972). By contrast, according to exemplar models, people represent categories by storing the individual training instances themselves (Hintzman, 1986;Medin & Schaffer, 1978;Nosofsky, 1986).A well-known experimental paradigm that has been used for contrasting the predictions of exemplar and prototype models is the Medin and Schaffer (1978) 5/4 category structure, which is listed in Table 1. 1 In this paradigm, the stimuli are simple perceptual forms that vary along four salient binary-valued dimensions. The stimuli are divided into two categories. The logical values of the prototype of Category A are assumed to be 0 0 0 0, and the logical values of the prototype of Category B are assumed to be 1 1 1 1. Subjects are trained on the first nine items and are then given a transfer test that includes all the items in the list. This category structure is diagnostic because prototype and exemplar models tend to make opposite predictions for specific items. Most critically, prototype models predict that people will perform better on Stimulus A1 than on Stimulus A2 because A1 shares more features with its category prototype. In contrast, exemplar models generally predict an A2 advantage because A2 is highly similar to (i.e., shares three features with) two Category A exemplars and no Category B exemplars. In fact, the A2 advantage has been observed in numerous studies. Furthermore, when exemplar and prototype models are fitted to the classification data in this design, the results generally favor the predictions from the exemplar model (for reviews, see Nosofsky, 1992Nosofsky, , 2000 but see Smith & Minda, 2000, for an opposing viewpoint).However, whereas previous research concentr...

Uncovering mental representations with Markov chain Monte Carlo

“…Cognitive psychologists have extensively studied the ways in which people group objects into categories, investigating issues such as whether people represent categories in terms of exemplars or prototypes (J. D. Smith & Minda, 1998;Nosofsky & Zaki, 2002;Minda & Smith, 2002;Storms et al, 2000;Nosofsky, 1988;Nosofsky & Johansen, 2000;Zaki et al, 2003;E. E. Smith et al, 1998;Heit & Barsalou, 1996).…”

The effects of category overlap on information-integration and rule-based category learning

“…Multiple systems arguments have also been made in such diverse fields as reasoning (Sloman, 1996), motor learning (Willingham, 1998), discrimination learning (Kendler & Kendler, 1962), function learning (Hayes & Broadbent, 1988), and identification (Ashby, Waldron, Lee, & Berkman, 2001), as well as by other category learning researchers (e.g., Brooks, 1978;Erickson & Kruschke, 1998;Nosofsky, Palmeri, & McKinley, 1994). Nevertheless, many recent categorization articles have argued for a single system that mediates all category learning (Nosofsky & Kruschke, 2002;Pothos, 2005;Zaki, Nosofsky, Jessup, & Unversagt, 2003;Zaki, Nosofsky, Stanton, & Cohen, 2003). Although we cannot rule out the possibility that some single system model could account for the present results, a significant challenge for single system theorists is to account for the growing number of observed dissociations between rule-based and information-integration tasks within the same unified model.…”

Does response scaling cause the generalized context model to mimic a prototype model?

“…If GCMg mimics PRT, then GCMg should not only fit its own data better than PRT would, but it should also fit data generated by PRT better than PRT itself does. Nosofsky and Zaki (2002;Zaki et al, 2003) showed that this was not always the case, demonstrating that there are conditions in which GCMg does not mimic PRT.…”

“…Nosofsky and colleagues (Nosofsky & Zaki, 2002;Zaki, Nosofsky, Stanton, & Cohen, 2003) defended the introduction of the response-scaling parameter, arguing that, among other things, it was necessary to capture the deterministic behavior that participants exhibit early in learning, when they tend to focus on a single dimension of a stimulus. Probably most convincing in countering the claims of Smith and Minda (1998) are the results of two simulations that showed that data generated by PRT were fitted better by PRT than by GCMg.…”