11 - FEEDFORWARD/FEEDBACK (RESONATING) HETEROASSOCIATIVE NETWORKS

Abstract

This chapter focuses on heteroassociative networks that can associate one pattern with another. These networks have a distinctive architecture. The chapter discusses the Adaptive Resonance Theory (ART) networks and Bidirectional Associative Memory (BAM) networks. Both ART and BAM networks have both feedforward and feedback connections between the input and output layers. These networks may or may not have lateral communications, depending on their specific design. The important characteristic is that the feedforward/feedback connectivity and lateral connectivity usually play a subordinate role. ART and the adaptive BAM-type networks undergo unsupervised learning. Their weights change over time as new patterns are presented to the system. There is no distinction between the training phase and the operation phase nor is there a distinct set of training data and test data. Even the number of categories to which these networks associate can change over time. This makes these systems dynamically varying. When systems change over time, their stability becomes a major issue. Both the ART and BAM networks have been developed into families of network topologies. The chapter discusses the individual differences among networks that allow these capabilities and addresses some of the historically important concepts in neural heteroassociation, which laid a foundation for the ART and BAM types of networks.