Proceedings of the Annual Meeting of the Cognitive Science Society

Exemplar models are the canonical account of human category learning. We investigate the sufficiency of this framework across two experiments demonstrating that human learners spontaneously extract information that exemplar models fail to capture. In Experiment 1, we show that participants classify novel test items based on abstract domain-level regularities rather than summed similarity to stored exemplars. In Experiment 2, we test the design principle of selective attention by providing a perfect unidimensional cue while making a secondary dimension partially diagnostic. Contrary to the attention weight optimization of exemplar theory, participants included the redundant 'extra' features in their category representations as evidenced by single feature classification and generalization performance. These findings constitute an important challenge to core components of exemplar theory and suggest a generative process where learners build rich statistical models of the environment.