Summary
Perception, memory, and categorization of complex three-dimensional novel objects was assessed in nine experiments and a neural-network simulation model. As in past categorization experiments, many exemplars of each object category were tested, and similarity of test items to category prototypes and to trained exemplars was carefully manipulated. However, as in past object recognition experiments, each category had a unique part structure, making exemplars easily identifiable as one category or another. When participants were trained on multiple exemplars that "surrounded" the category prototypes, subsequent old/new recognition tests revealed performance on test exemplars to be more strongly related to similarity with category prototypes than to similarity with most-similar trained exemplars. Although the strength of this prototype effect could be manipulated by altering training procedures, the same pattern was evidenced even when training did not include any type of categorization task. A similar but weaker pattern was obtained with a naming test, and what could be a somewhat different prototype effect was found with a label verification test. A simple neural-network model, trained via a form of Hebbian learning rule to learn contours (outline shapes) of the same objects used in the experiments, simulated the old/new recognition results shockingly well, and also simulated some aspects of the naming and label-verification tasks. Results of the experiments and simulations are discussed with respect to instance-based and central-tendency models in both the object recognition and categorization literatures.