ABSTRACT
Current models of human category learning and subsequent recognition are either exemplar-based, rule-based, or some combination of both approaches. We present learning and recognition data that cannot be accounted for by current approaches. The data suggest that the degree to which an item is remembered is determined by the strength of the expectation it violates. In our study, expectations take the form of simple, imperfect rules where the strength of the rules are determined by the number of items that follow the rules in training. Exemplar-based models cannot account for the results because they do not posit organizing knowledge structures that can be violated. The frequency insensitivity of rule-based accounts leads to their failure. We propose a cluster-based approach that is consistent with our findings, as well as findings from schema, stereotype, and basic memory research.
