Retrieval of Exemplar and Feature Information in Category Learning

Two assumptions were tested regarding what information Ss use during category learning with independent features. One assumption is that Ss use only exemplar information and the other is that they use only feature information. Previous work has revealed no clear superiority of one over the other after training with independent features ( W. K. Estes, 1986a , 1986b ). The experiments presented here manipulate the opportunity for using whole exemplars when categorizing test patterns by providing either whole or fragmented training patterns. The results show that a feature-node network model was superior to feature-frequency and exemplar models at predicting asymptotic test performance after fragmented-pattern training. However, to achieve this result, all models were modified to account for possible attentional differences among the training conditions.