Error propagation and supervised learning in adaptive resonance networks

A class of adaptive resonance networks based on explicit computation of errors between input vectors and learned templates is discussed and related to self-organizing feature maps and minimum distance automata as well as to the ART 1 (binary input) and ART 2 (analog input) architectures. In addition, a simple method of incorporating supervised learning in adaptive resonance networks is discussed and related to Adalines, counterpropagation, radial basis function interpolation networks, and Bayesian networks. Supervised ART networks can operate in a supervised or unsupervised mode, and the networks autonomously switch between the two modes. When supervised (desired) signals are absent, these networks predict the desired signal based on previous training. These supervised adaptive resonance networks can form nonlinearly separable decision boundaries, and they can learn the XOR problem on a single trial