Chapter 42 - The Neurodynamics of Categorization: Critical Challenges and Proposed Solutions

Abstract

Computational modeling is a useful tool for understanding human categorization. It allows for the testing of structural and functional properties of the cognitive process. In particular, connectionist networks are useful geometric models of categorization. Recurrent networks typically use Hebbian learning to convert the stimulus space into a feedback subspace sufficient to categorize new stimuli. In the present chapter, we review the evolution of some recurrent networks for modeling categorization by examining challenges they faced and proposed solutions. First, we examine the recurrent auto-associative memory (RAM) class of networks. Specifically, we examine the problems of divergence and noise and review some proposed solutions. Second, we review the progression of research on bidirectional heteroassociative memory (BAM) networks that are capable of both auto-associative and heteroassociative memories. Third, we introduce a hybrid model of feature-extracting bidirectional associative memory (FEBAM). By unifying properties from both BAM class networks and principal component analysis (PCA) networks, this hybrid presents a possible solution to limitations of previous models, such as the BAM, and is a potential candidate for effectively modeling the categorization process in humans.