How people learn features in the absence of classification error

Models of category learning often assume that exemplar features are learned in proportion to how much they reduce classification error. In contrast, experimental evidence suggests that people continue to learn features even when classification is perfect. We present three experiments that test explanations for how people might learn features in the absence of error. In Experiment 1, we varied the type of feedback participants received. In Experiment 2, we introduced a secondary task during the feedback phase, and in Experiment 3, we restricted the response window and varied the feedback. In all cases, we found that participants learn many more features than they need to classify the exemplars. Our results suggest that participants learn the internal correlations between features, rather than directly forming associations between features and the category label. This finding places restrictions on the types of categorisation models that can satisfactorily explain learning in the absence of classification error.