Parallelization of algorithms for neural networks

Abstract

In this paper we present the strategies adopted in the parallelization of two algorithms for the simulation of two classes of neural networks: the Hopfield Network and the Error BackPropagation Network. Although the parallel algorithms have been developed within the (loosely synchronous) SPMD parallel programming model, the particular nature of the strategies adopted make the final parallel algorithms not expressible within the HPF-like programming paradigm; therefore a more flexible programming model, the message passing programming paradigm, has been adopted, and the final development has been carried out in the PVM environment.