Learning Categories Many Ways at a Time

Kurtz, Kenneth

Abstract

The psychological process of recognizing objects and situations as members of known categories is how people link what is perceived with what is known about the world. In the study of categorization the problem has been stripped down to its most basic elements in numerous tests of artificial classification learning, and at this level, formal models provide an excellent account. The theory view of categorization offers a reminder that much of the richness and foundation of real-world categories remains insufficiently explored by traditional laboratory studies and insufficiently explained by models. The current proposal takes its starting point from two somewhat overlooked, but fundamental issues in categorization. The first is the question of how examples to be categorized are encoded or construed. The second is that category structure is constrained by a number of different ways of learning and using categories, not just taxonomic classification. These strands come together in ORACL, a connectionist model that addresses categorization as the error-driven recoding of inputs to serve multiple tasks. The purpose of the studies in this proposal are: 1) to investigate the basis of item encoding and recoding as a function of category learning; 2) to address the completely new question of what occurs and what can arise from learning multiple classification schemes at the same time; 3) to evaluate and benefit from the application of new experimental paradigms for studying category learning and representation; 4) to test the predictive and theoretical strength of the ORACL model of category learning; and 5) to push the study of categorization toward the goal of explaining the acquisition, organization, and use of naturalistic categories.