ART 3: Hierarchical Search Using Chemical Transmitters in Self-Organizing PatternRecognition Architectures

摘要

This article incorporates a model of the chemical synapse into a new AdaptiveResonance Theory (ART) neural network architecture called ART 3. ART 3 system dynamics model a simple, robust mechanism for parallel search of a learned pattern recognition code. This search mechanism was designed to implement the computational needs of ART systems embedded in network hierarchies, where there can, in general, be either fast or slow learning and distributed or compressed code representations. The search mechanism incorporates a code reset property that serves at least three distinct functions: to correct erroneous category choices, to learn from reinforcement feedback, and to respond to changing input patterns. The three types of reset are illustrated, by computer simulation, for both maximally compressed and partially compressed pattern recognition codes. Let us first review the main elements of ART. ART architectures are neural networks that carry out stable self-organization of recognition codes for arbitrary sequences of input patterns. ART first emerged from an analysis of the instabilities inherent in feedforward adaptive coding structures (Grossberg, 1976a). More recent work has led to the development of two classes of ART neural network architectures, specified as systems of differential equations.