Comparing exemplar and prototype models of categorization

Canadian Journal of Experimental Psychology, Sep 1997 by Stephen Dopkins, Theresea Gleason

To test the prototype and exemplar theories, we had to instantiate them as concrete models. The models that we tested were reasonably general. To demonstrate this, it is necessary to briefly discuss the sorts of prototype and exemplar models that have been proposed in past work. The models can be distinguished in terms of three kinds of theoretical questions: (1) How are categories represented in memory; (2) what information is used in making categorization judgments; and (3) are categorization judgments based on a probabilistic or a deterministic decision rule (Ashby, 1992)?

PROTOTYPE MODELS

In prototype models, the mental representation of a category is its prototype. Although a category's prototype can be defined in many ways (Barsalou, 1990; Nosofsky, 1987, 1992b; Posner & Keele, 1968; Reed, 1972), it is often defined as the exemplar with average values on all of the dimensions along which the category's exemplars vary (Reed, 1972; Rosch, Simpson, & Miller, 1976; Shin & Nosofsky, 1992; Smith & Medin, 1981). In most prototype models, a category's prototype is mentally represented as a point in a psychological space. The dimensions of this space correspond to the perceptual dimensions along which the category's exemplars vary. These types of prototype models differ in whether they assume a Euclidean or a city-block metric for the psychological space and in whether they weight the dimensions of the space to reflect their importance in determining the categorization response (Estes, 1986; Maddox & Ashby, 1993; Reed, 1972; Shin & Nosofsky, 1992).

In prototype models, categorization judgments are based on the prototypes of the contending categories. The test stimulus is assigned to the category to whose prototype it is most closely related (Nosofsky, 1987; Reed, 1972). Prototype models differ in how they estimate the degree of relatedness between a test stimulus and a prototype. In distance models, the crucial quantity is the distance between the points that correspond to the stimulus and the prototype in psychological space (Reed, 1972). In similarity models, the crucial quantity is the degree of similarity between the stimulus and the prototype (Reed, 1972). Various systems have been proposed for deriving similarity from distance in psychological space and for combining values of similarity on different dimensions (Nosofsky, 1987, 1992b; Reed, 1972). Both probabilistic and deterministic decision rules have been used in prototype models.

CHOOSING A REPRESENTATIVE PROTOTYPE MODEL

To represent the prototype theory, we chose a model with the following properties: (1) A category's prototype is the exemplar with average values on all of the dimensions along which the category's exemplars vary, and (2) similarity is the index of relatedness. Much of the recent work on prototype models has focussed on this sort of model (Nosofsky, 1987; Reed, 1972; Shin & Nosofsky, 1992). Furthermore, the idea of the prototype as the category "average" is pervasive in the broader literature on linguistic categorization (Rosch et al., 1976; Smith & Medin, 1981).