A feature projection based adaptive pattern recognition network

Abstract

An adaptive pattern recognition network is introduced which is based on multiprecision feature space projection and Bayes confidence combinations. The adaptive network has the following properties: it allows incremental accumulation and effective use of statistical data; for each pattern class, the network dynamically selects the most significant (weighted) features for classification; and the method allows fast incremental learning from training samples and provides for the dynamical introduction of new classes and new features or the exclusion of existing classes and features without retraining on the old data. The model implements the optimal Bayesian classifier without recourse to underlying assumptions about class probability distributions. The network is computationally efficient because it has a parallel architecture. >