Abstract
Over the past decades, a family of self-organizing neural architectures, known as Adaptive Resonance Theory (ART) (Grossberg 1976a; 1976b; Carpenter & Grossberg 1991) has been steadily developed, embossed with well-founded computational principles and widespread applications. These successful applications are of particular interests because the basic ART principles have been derived from an analysis of human and animal perceptual and cognitive information processing, and have led to behavioral and neurobiological predictions that have received significant experimental support during the last decade (Grossberg 2003; Raizada & Grossberg 2003). In this talk, we will trace the origin of ART, from the early in-star and out-star networks to a series of very well known computational models, known as ART1, ART2, ART3, and Fuzzy ART etc, and their applications in pattern analysis and clustering. Building upon the original unsupervised learning ART models, we will move on to survey a family of supervised ART systems rapidly developed in 1990's, notably ARTMAP, fuzzy ARTMAP ART-EMAP, Cascade ARTMAP, Gaussian ARTMAP, Distributed ARTMAP, and a compressed version of ARTMAP known as Adaptive Resonance Associative Map (ARAM). These supervised learning ART models have been applied to a wide range of applications, including handwritten character recognition, image understanding, medical diagnosis, DNA gene classification, text categorization, and personalized information management. The talk will share some recent success stories of such applications (Tan et al. 2004; Tan and Pan, 2005). Progressing from unsupervised and supervised learning paradigms, the talk will highlight some of the recent development in extending ART models for reinforcement learning. We show how a natural generalization of the ART dynamics can lead to an reinforcement learning system, known as Fusion Architecture for Learning and Cognition (FALCON), that is capable of functioning in a dynamic and real-time environment with immediate or delayed reward signals (Tan and Xiao, 2005; Tan 2006). Moving forward, we provide a glimpse into how Adaptive Resonance Theory may lay the foundation of an integrated cognitive autonomous system for performing high level cognitive functions, most notably awareness, learning, reasoning, and surprise handling.
