An unsupervised hyperspheric multi-layer feedforward neural network model

Abstract

This paper introduces a novel Neural Network model intended for classification of patterns into distinct categories. Arbitrary accurate category formation in a predefined feature space is asymptotically achieved by means of an unsupervised learning algorithm. Learning takes place by assignment of labeled neurons to unrecognized input exemplars and subsequent labels merging subject to similarity constraints. Two model variants, one under category separability assumption, and a second under a more general category probability density unimodality (and non-separability) assumption are suggested. The hyperspheric nature (as opposed to hyperplanar — typical of some current classifiers) of this model and its multilayer feedforward architecture are explained. Simulation results demonstrating asymptotic convergence and excellent classification accuracy are provided.