Artificial Neural Network on a Massively Parallel Associative Architecture

Abstract

An implementation of a fully connected artificial neural network using the multi-layered perceptron model is described. The neural network is implemented on the ASP (Associative String Processor). The ASP is a fine-grain massively parallel SIMD architecture, emerging from research at Brunel University and being developed by Aspex Microsystems Ltd., based on associative processing. Neural networks readily map onto the ASP architecture. The neural network described in this paper is a multi-layered perceptron model which uses the back-propagation learning paradigm. The network has 137 nodes in three layers and is trained to recognise letters of the alphabet. The work is the basis for the implementation of a massive artificial network environment (with 10⁵ nodes and 10⁸ connections) on the ASP using the back propagation and alternative learning techniques.