Parallel implementation of non recurrent neural networks

Abstract

The computational models introduced by neural network theory exhibit a natural parallelism in the sense that the network can be decomposed in several cellular automata working simultaneously. Following this idea, we present in this paper a parallel implementation of the learning process for two of the main non recurrent neural networks: the Multilayer Perceptron (MLP) and the Self-Organising Map of Kohonen (SOM).

The system we propose integrates both neural networks applied to an isolated word recognition task. The implementation was carried out on a transputer based machine following the model given by the CSP (Communicating Sequential Processes) specifications. Several experiments with different numbers of processors were made in order to evaluate the performance of the proposed system. The aspects related to the load balancing and communication overheads are discussed.