Given versus induced category representations: use of prototype and exemplar information in classification.

Acquisition of category-level information can be based on experience with category members (induced) as well as on direct presentation of prototypical values (given). To investigate the effects of these two types of information, a relational coding model of categorization was developed in which classification is based on a mixture of exemplar and prototype information. In two experiments, subjects learned about two ill-defined categories. Stimuli were geometric shapes varying along four binary-valued dimensions. For three groups of subjects, training consisted of (a) experience with exemplars only, (b) learning prototype values followed by exemplar experience, or (c) learning prototype values concurrently with exemplar experience. Following training, all subjects received classification tests on prototype values as well as on old and new exemplars. By varying the relative use of prototype and exemplar information, the mixture model accurately accounted for category judgements in all three groups. For subjects directly presented with prototype values, classification was based on a mixture of similarity to prototypes and to stored exemplars. In contrast, subjects who only received experience with exemplars appeared to base their category judgements solely on similarity to stored exemplars, even though they could accurately judge the prototype values. The two components of the mixture model are related to subjects' classification strategies and the nature of abstracted, category-level information.

Correlated symptoms and simulated medical classification.

Category learning theories can be separated into those that expect judgments to be sensitive to configural information and those that expect judgments to be based on a weighted, additive summation of information. Predictions of these two classes of models were investigated in a simulated medical diagnosis task. Subjects learned about a fictitious disease or about two diseases from hypothetical case studies in which some symptoms were correlated with each other and others were independent. Following this initial training, subjects were presented either with pairs of new cases and asked to judge which was more likely to have the disease or with a single case and asked which disease was present. Across four experiments, subjects proved to be sensitive to configural information. When choosing between pairs of new cases, subjects tended to choose the case that preserved the correlation over the case that broke the correlation, even when the case with correlated symptoms contained fewer typical symptoms. When judging which disease was present in a single case, subjects' diagnoses were determined primarily by the correlated symptoms. Implications of these findings to process models of categorization are discussed.