An artificial neural network simulator on the loosely coupled parallel processors
T. Oohashi, Toshiaki Ejima · 1991
Summary form only given, as follows. The authors examine the parallelism of a multilayered ANN (artificial neural network) and discuss a parallel algorithm suited to loosely coupled parallel processors. A mapping of a multilayered network to large-grain processors is proposed, and its performance is evaluated. For a two-layer backpropagation model which has N units in each layer, the highest speedup ratio is obtained with 8N processors but the parallel efficiency is less than 20%. With 2N processors and N/2 processors, the parallel efficiencies of the mapping are 50% and 80%, respectively. It is also shown that the proposed parallel algorithm is more efficient for a larger network.>