Category invention in unsupervised learning.

This research aimed to discriminate between 2 general approaches to unsupervised category learning, one based on learning explicit correlational rules or associations within a stimulus domain (autocorrelation) and the other based on inventing separate categories to capture the correlational structure of the domain (category invention). An "attribute-listing" paradigm was used to index unsupervised learning in 3 experiments. Each experiment manipulated the order in which instances from 2 different categories were presented and evaluated the effects of this manipulation in terms of the 2 competing theoretical approaches to unsupervised learning. Strong evidence was found for the use by Ss of a discrete category invention process to learn the categories in these experiments. These results also suggest that attribute listing may be a valuable method for future investigations of unsupervised category learning.