An efficient finite precision RBF-M neural network architecture using support vectors

Abstract

This paper investigates the effects of using limited precision for efficient implementations of the RBF-M neural network. This architecture employs only simple arithmetic operators and is characterized by simple LMS training in an expanded feature space generated by RBF functions centered around support vectors selected via a simple algorithm. The classification performances of our low complexity, finite precision architecture are similar and even better to those obtained using the more complex SVM.