Development of Prototype Abstraction and Exemplar Memorization

Development of Prototype Abstraction and Exemplar Memorization Irina Baetu (irina.baetu@mail.mcgill.ca) Department of Psychology, McGill University, 1205 Penfield Avenue Montreal, QC H3A 1B1 Canada Thomas R. Shultz (thomas.shultz@mcgill.ca) Department of Psychology and School of Computer Science, McGill University, 1205 Penfield Avenue Montreal, QC H3A 1B1 Canada likely to rely on prototypes at the beginning of a categorization task, and as training progresses they rely more on memorized exemplars (Horst, Oakes, & Madole, 2005; Minda & Smith, 2001; Smith & Minda, 1998). These studies are consistent with a shift from early prototype use to later exemplar memorization. In addition to the amount of experience with a categorization task, category structure also influences which type of information is most used. Better-structured categories can be represented as separate clusters in psychological space, whereas poorly structured categories overlap with each other (Figure 1). Smith and Minda found that better structured categories encourage the early prototype formation, while poorly structured categories discourage it, and may even strongly disadvantage the use of prototypes (Smith & Minda, 1998). Their findings are consistent with a number of other studies (Homa, et al., 1981; Horst, et al., 2005; Reed, 1978). Abstract We present a connectionist model of concept learning that integrates prototype and exemplar effects and reconciles apparently conflicting findings on the development of these effects. Using sibling-descendant cascade-correlation networks, we found that prototype effects were more prominent at the beginning of training and decreased with further training. In contrast, exemplar effects steadily increased with learning. Both kinds of effects were also influenced by category structure. Well-differentiated categories encouraged prototype abstraction while poorly structured categories promoted example memorization. Keywords: exemplar memorization; prototype abstraction; category structure; neural networks; sibling-descendant cascade-correlation. Introduction One of the most fundamental abilities is learning to group things into categories. This faculty allows us to classify new examples and make useful predictions concerning their properties. Two general classes of models have been proposed to account for phenomena in concept learning: prototype and exemplar models. Prototype models claim that experience with items that belong to a given category results in the formation of a summary representation of all the items observed (Posner & Keele, 1968; Reed, 1972). Subsequent categorization of a new item is then based on a comparison between the prototype and the new item. Thus, the more similar a particular instance is to the abstracted prototype, the more likely it is to be classified as a category member (Homa & Cultice, 1984; Homa, Sterling, & Trepel, 1981). In contrast, exemplar models claim that all the observed items are remembered and that the categorization of a new item involves a comparison with items that are stored in memory (Hintzman, 1986). There is ample evidence in favor of both prototype (Homa, et al., 1981; Posner & Keele, 1968) and exemplar models (Medin & Schaffer, 1978; Palmeri & Nosofsky, 2001), suggesting that both processes are used during category learning. What is more, the relative contribution of each mechanism to categorization might vary across development, as well as during training on a novel task. Early in development, categorization seems to be based on prototype representations while exemplar representations seem to increase with age (Hayes & Taplin, 1993; Mervis & Pani, 1980). There is also evidence that people are more Figure 1: Hypothetical representations of three concepts. P1, P2 and P3 represent three prototypes and the circles represent examples of each concept. A: prototypes are relatively far from each other and examples are tightly clustered around their respective prototype, yielding concepts that are easy to distinguish. B: prototypes are close to each other and examples are more widely dispersed around their respective prototype, resulting in overlapping concepts that are difficult to distinguish. The aim of this paper is to present a unified model able to simulate prototype and exemplar processes during concept learning. This unified model captures prototype and