Application of a prototype parallel processing computer for a recurrent neural network model

Abstract

In this study, Hopfield's binary neural network model is simulated using a software package simulating a bit serial single instruction multiple data (SIMD) mesh array processor, called the Blitzen massively parallel processor (BMPP). First, the parallel algorithms required for the simulation are designed. Then, the parallel implementation of the Hopfield network model has been formulated using these algorithms. The parallel implementation has been analyzed for the speed and the processor utilization. To do this, time complexity of the parallel implementation is derived and compared with the time complexity of a sequential algorithm. Finally, simulation results are analyzed and compared with the analytical derivations. The parallel algorithm described in this study for the simulation of the Hogfield network model on the BMPP architecture achieves a maximum speedup in the order of the square root of the number of processors employed. It is also shown that, for the same algorithm, the maximum possible speedup can be achieved only at a finite number of processors, and the processor utilization decreases as the number of processors is increased.