人工神经网络在多处理器系统上并行实现的通信开销

COMMUNICATION OVERHEAD IN PARALLEL IMPLEMENTATION OF ANN ON MP SYSTEM

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．

The communication overhead in parallel implementation of artificial neural network on multiprocessor system is analyzed in this paper. For fully connected and random connected artificial neural network, the optimization of multiprocessor system's topology and the induction of fan in and node section have no help to reduce the communication overhead of parallel ANN algorithm based on MPP.