Bayesian Polytope ARTMAP: An ART-based network with two kinds of inner geometry categories

The ART-based neural networks summarize data into groups via the use of inner categories. A category´s template elements are updated incrementally in the light of new evidence provided by the presentation of input patterns. In order to reduce approximation error, this paper proposes Bayesian Polytope ARTMAP (BPTAM) which incorporates both simplex categories and Gaussian categories. During training, the simplex categories expand only towards the input pattern without category overlap, while the Gaussian categories grow or shrink by limiting their hypervolumes. In addition, BPTAM uses Bayes´ decision theory for learning and inference, which makes BPTAM robust to noise and category overlap. Based on some preliminary but illustrative experimental results, BPTAM shows better applicability to data sets with noise, statistical overlapping and irregular geometry.