An analysis of network parallelism of backpropagation neural networks on a transputer array

This paper presents the development of a theoretical model for parallel implementation of backpropagation (BP) neural networks using the "network parallelism" paradigm. The case where the number of nodes in each layer is equally distributed among the processors is considered. The model is developed for a homogeneous array of message passing processors with only local memory. The developed model predicts the training time for a given neural network without having to set up the hardware and run the experiments. The accuracy of the model is verified by comparing the predicted training time using the theoretical model with the actual experimental values for the case of an array of T805 transputers connected in a ring topology for benchmark neural network problems like the Encoder and Nettalk.