Author(s)

F. Liang & H.A. ElMaraghy

Abstract

Localized pi-sigma networks with competitive lateral inhibitory cells F. Liang & H.A. ElMaraghy FfezzWe Mo7%%/acZ%rm# Cenfre, FacWfy o/ Engineering, McMaster University, Hamilton, Ontario, Canada L8S ABSTRACT Competitive lateral inhibitory cells are introduced in this paper to localize the forward and backward computation of the Pi-Sigma networks and accelerate their learning processes. Both the backpropagation-like learning algorithm and the adaptive ex- tended Kalman filtering learning algorithm with recursive UD factorization are derived for the localized Pi-Sigma networks with competitive inhibitory cells. Due to the localization, only a small sized Pi-Sigma network needs to be taught and is re- sponsible for output at each step. Therefore, the forward and backward computation are reduced greatly, and the new networks degrade gracefully in case of hardware fail- ures. The structure of such a network can be pre-determined by the approximation accuracy requirement and

Keywords