Increasing the Capacity of a B-matrix Neural Networks

Abstract

In an effort to create computing structures that are as efficient as the brain at cognitive tasks, interconnected artificial neurons are used in cognitive science and artificial intelligence. In this dissertation we focus on the memory retrieval mechanism in an artificial neural network and suggest an algorithm to increase the memory retrieval rate in the B- Matrix neural network. The B-matrix is a model of recall by index, where in the activity spreads locally. This approach to neural network function accounts for spreading of activity from one region to others based on adjacency of neurons. We propose an algorithm to increase the memory retrieval rate by identifying the inactive nodes in the network and applying Widrow-Hoff learning function to the weighted matrix. This may be seen as implementing the concept of synaptic plasticity in artificial neural networks to increase their memory capacity.