Theory of categorization based on distributed memory storage.

As an alternative to probabalistic and examplar models of categorization, we develop a model based on the assumption of distributed memory storage. Subjects in two experiments performed tasks related to the categorization of random dot patterns. First, the perceived similarity was measured between two such dot patterns, one a distortion of the other. Second, groups of examplar patterns derived from a category prototype were classified together in a category learning task. When the number of examples was small, new dot patterns were classified according to their similarity to learned exemplars; when the number was large, accuracy depended on a dot pattern's similarity to the prototype pattern. The distributed memory model is used to explain a number of aspects of the experimental findings. Detailed computer simulations are described for the similarity, categorization, and prototype enhancement results.