Comparing exemplar and prototype models of categorization

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

Abstract Five experiments contrasted the prototype and exemplar theories of categorization. The stimuli for the experiments were simple perceptual figures that varied along two multi-valued dimensions. During the learning phase of each experiment, participants assigned each of a series of stimuli to one of two categories and received feedback concerning their assignments. During the generalization phase, participants made category judgments with respect to novel stimuli, including some crucial test stimuli that were (a) more similar to the prototype of one of the two categories, and at the same time, (b) more similar, on average, to the exemplars of the other category. Participants assigned the test stimuli to the latter category. This was taken as evidence favoring the exemplar over the prototype theory.

To cope with an incessant flow of information, humans and other organisms must group their experiences into meaningful categories (Anderson, 1991; Shepard, 1987; Smith & Medin, 1981). How do they do this? The traditional view, dating from the time of Aristotle, is that categories are mentally represented as sets of necessary and sufficient conditions. According to this view, an exemplar is assigned to a category if it satisfies the category's membership conditions (Ashby, 1992).

Recently, problems have been raised with this view. For example, the exemplars of categories do not always seem to share the same properties (Wittgenstein, 1953). In some cases (e.g., "game"), there are apparently no properties that are shared by all members of the category. In addition, some exemplars of some categories seem "better" than others. For example, a robin seems a "better" bird than a chicken does (Rips, Shoben, & Smith, 1973; Wittgenstein, 1953).

Several theories have been proposed in response to these problems. According to the prototype theory, a category's mental representation is based on a prototypical exemplar or prototype. For example, the representation for the BIRD category might be based on ROBIN. To decide whether an entity is a member of a category, one compares it to the category's prototype (Posner & Keele, 1968, 1970; Reed, 1972). This theory can account for the phenomena that are problematic for the traditional theory. Because category membership no longer depends on necessary and sufficient conditions, all exemplars need not share the same properties. Stimuli will be "better" exemplars of a category the more closely related they are to the category's prototype.

According to the exemplar theory, a category's mental representation encodes the exemplars that compose the category. To decide whether an entity is a member of a category, one compares it to the category's exemplars (Brooks, 1978; Estes, 1986; Hintzman, 1986; Medin & Schaffer, 1978; Nosofsky, 1986). Again, this theory can account for the problematic phenomena. If all of the exemplars of a category are encoded, all exemplars need not share the same properties. Stimuli will be "better" exemplars of a category the more closely related they are to the category's exemplars.

A considerable literature exists contrasting the prototype and exemplar theories. The research has typically used tasks in which novel categories are learned in the laboratory. The primary focus has been on the process by which a test stimulus is assigned to one of several contending categories after differences among these categories have been learned. A distinction has generally been made between stimuli that vary in terms of bi-valued (e.g., small, large) and multi-valued (e.g., 1, 2, 3, 4, 5, 6 cm) dimensions.

Medin and his colleagues have repeatedly compared the prototype and exemplar theories using bi-valued stimuli (Medin, Altom, Edelson, & Freko, 1982; Medin, Altom, & Murphy, 1984; Medin, Dewey, & Murphy, 1983; Medin & Schaffer, 1978; Medin & Smith, 1981; Nosofsky, Kruschke, & McKinley, 1992). These comparisons have supported the exemplar theory. Support for the exemplar theory has not been as strong when multi-valued stimuli have been used. One limitation of much past work in this area is that theories have been compared in terms of their capacity for producing fine-grained quantitative data fits (e.g., according to each theory, what is the probability that each of a number of stimuli will be assigned to each of several categories?; how do these predicted probabilities compare with the observed probabilities?) (Ashby & Maddox, 1992; Nosofsky, 1987; Shin & Nosofsky, 1992). Results obtained in this manner have often been limited in that they have depended on rather specific quantitative assumptions.

In this study we attempted to gather further evidence concerning the validity of the prototype and exemplar theories in the multi-valued dimension case. In contrast to past work, which has used quantitative fits, the strategy here was to set up experimental situations in which contending theories made qualitatively different predictions as to how certain key stimuli would be categorized (i.e., the prototype and exemplar theories predicted that the stimuli would be assigned to different categories). This is the strategy that has generally been used in work with bi-valued stimuli (Medin, et al., 1982; Medin & Edelson, 1988; Medin & Schaffer, 1978).