Efficient Parallel Simulation of Pulse-Coded Neural Networks (PCNN)

摘要

Neural networks are the common model for brain style data processing. Therefore, the algorithms are inherently parallel and a parallel implementation of neural network simulations seems to be straightforward. However, typical parallel artificial neural network (ANN) simulations show only poor speedup on most parallel computers. In contrast, pulse-coded neural networks (PCNN) seem to be better suited for a parallel simulation, especially for vision purposes. In this paper we give parallel simulation techniques and partitioning strategies for PCNNs. Furthermore, a parallel PVM simulator and a special network description compiler are presented. The speedup of simulation is shown by typical example networks of large vision PCNNs on SUN workstation clusters and a massively parallel Pentium II cluster computer. The fast and simple simulation environment allows the simulation of large PCNNs with simulation times near to real time.