Analysis and implementation of backpropagation neural networks on heterogeneous processor arrays

Abstract

This study focuses on the parallel implementations of backpropagation (BP) neural net-works on a heterogeneous array of processors. A theoretical model of the BP algorithm running on the processor network was developed for training set parallelism and using this model the time for a training epoch was predicted. The model made use of two graphical tools, process synchronization graphs and variable synchronization graphs, to aid in obtaining the theoretical expression for the time for a training epoch. The theoretically predicted epoch times from the model were then experimentally validated on well known benchmark problems.