Connectionist models of categorization: A statistical interpretation

Abstract

The key assumption of this paper is that categorization can be related to the statistical problem of probability density estimation. Ashby and Alfonso-Reese (1995) have shown that several existing models of categorization can be related to specific statistical methods of density estimation. I extend this work in two ways. First, I show how a semi-parametric statistical technique of density estimation based on a Gaussian mixture distribution, can be used to construct a new model of categorization called the general decision bound model. Second, I propose a neural network framework based on RBF networks with its statistical interpretation. This framework is used to construct a neural network implementation of both the Gaussian and the general decision bound models.