Proceedings of the Annual Meeting of the Cognitive Science Society

The six types of elemental category structures (Shepard, Hovland, & Jenkins, 1961) for three binary dimensions area well-known benchmark in the study of human category learning. We added a non-diagnostic dimension consistingof four possible values with no predictive power. This increases the size of the training set fourfold. Exemplar modelssuccessfully account for the SHJ ordering based on stimulus generalization plus selective attention. Accordingly, exemplarmodels should learn this new task by ignoring the irrelevant dimension and performing nearly as usual. In a behavioralstudy, we found that Type I (unidimensional rule) was acquired rapidly, but most learners struggled to make any progressover an extended training period for Type IV (unidimensional rule-plus-exception) and Type VI (no regularities). Weinvestigate whether leading formal models can fit this pattern and address implications of these results for theories ofcategory learning.