Background Knowledge and Models of Categorization

2 Introduction In most applications of formal models of categorization, category learning is portrayed as the building-up of a representation in memory for members of the category that have been observed. This assumption is perhaps the most basic that is made for models of categorization, that the representation of a category describes its observed members. Yet if category representations are to serve a purpose such as recognizing new members of a category, then simply relying on memory for known members would be a poor strategy in many situations. For example, if you are learning to distinguish the Smith family from the Jones family, and you have observed a tall, red-haired 45 year old woman who is the mother in the Smith family, and you then see another tall, red-haired 45 year old woman, you would probably classify her as belonging to the Jones family, despite her similarity to an observed member of the Smith family. This example highlights the point that when few members of a category have been observed (or in the case of the Jones family, no members), it is crucial to rely on background knowledge rather than observed category members. This point is actually quite general, because there are many everyday situations where we are learning about new categories, such as visiting someplace new and learning about the social groups, buildings, landscapes, and so on, there. In each situation, starting with a fresh, or empty, category representation, and simply accumulating observations, would lead to great difficulties until a representative sample of category members can be observed. In the example of visiting a new place, background knowledge of people and buildings in other places would be crucial to classifying and reasoning in the new context, even if strictly speaking this information does not constitute observations of the target categories. That is, if you are trying to draw inferences about a novel category for which you have observed few category members or no category members, the best you can do is hope that this novel category will be like some previously known categories, and reason about the novel category based on knowledge of other categories. A great deal of work (reviewed in the Ahn & Dennis and Hampton chapters in this volume) has contrasted the similarity-based approach to categorization and the knowledge-based approach. The similarity-based approach highlights the use of similarity to observed category members and the knowledge-based approach …