Learning Multiple Categories Without Feedback

Event Center BC

How do people acquire new categories in an unsupervised (no-feedback) environment? We distinguish two general classes of models. Correlation tracking models assume that new categories are acquired by tracking associations among the features of different stimuli until reliable patterns are learned. Category invention models assume that new categories are triggered as an all-or-none response to novel or surprising objects that violate existing categories. Previous research tested the models by manipulating the order in which examples of two different categories were shown in an unsupervised learning task. As predicted by the category invention model, people learned best when a training sequence maximized the perceived contrast between two potential categories, e.g., by showing them in separate blocks, even if fewer examples were shown than in a comparable lowcontrast (e.g., randomly interleaved) sequence. This “negative exposure effect” cannot be accommodated by pure correlation tracking models. Of course, people often acquire more than two categories at a time in the real world, and so it is important to determine whether evidence for category invention can also be obtained in a multi-category task. In this experiment, participants saw examples of three categories in two different sequences: (a) mixed from the start of training (interleaved sequence), or (b) staggered so that examples of the first category only were shown for several trials, followed by examples of the first and second categories in random alternation, and finally by examples of all three categories in random alternation (semi-blocked sequence). As in previous two-category studies, learning of all categories was better in the semiblocked condition than in the interleaved condition, even when fewer examples were shown. This is a clear victory for category invention as a theoretical model, and perhaps also for the practical value of blocked or semi-blocked training sequences when people must acquire separate categories under unsupervised conditions. On the other hand, people in the contrast condition showed evidence of reduced learning after the third category was introduced, suggesting that they may have had difficulty maintaining three separate categories in this unsupervised task. Further research will attempt to investigate the causes of this “category load” effect and clarify its impact on people’s ability to acquire and maintain large sets of related categories under both supervised and unsupervised task conditions.