Proceedings of the Annual Meeting of the Cognitive Science Society

The Locally Bayesian Learning (LBL) approach bridges the gap between optimal Bayesian learning and suboptimal performance that arises from human behavior. Although this learning model has considerable potential, it has been underdeveloped and has remained in its original form for several decades. In this paper, we extend the original LBL model to an exemplar approach, which we refer to as the exemplar-LBL model. Two notable features of this extension are that (a) the model can take continuous features as inputs and (b) can conduct exemplar-based categorization. We report various simulations, which show that the model can generate numerous important predictions about category learning. Additionally, we introduce the extra-learning hypothesis, which can account for how classification and observation training can produce differential learning. Our results showcase scenarios under which classification training is superior to observation training and other instances in which the opposite occurs.