论文部分内容阅读
本文将讨论在多处理器系统上，用大规模并行处理技术实现人工神经网络时，处理器之间的通信开销问题．通过理论推导和实验证明，全连接和随机连接神经网络在多处理器系统上并行实现时，处理器网络的拓扑结构对神经网络实现的通信开销影响很小，改变处理器网络的拓扑结构也无助于神经网络实现性能的提高．
This article will discuss the problem of communication overhead between processors when implementing artificial neural networks with massively parallel processing on multiprocessor systems. Theoretical derivation and experimental results show that the topology of processor network has little effect on the communication overhead of neural network when fully connected and stochastically connected neural networks are implemented in parallel on multiprocessor system. The topological structure of the processor network is also changed Help to improve the performance of neural network.