论文部分内容阅读
本文介绍了神经网络的性能分析,重点是实现可达到的加速和处理机网络(transputer或以可比较方式通信的多计算机网络)的最佳大小。对于完全和随机连接的神经网络,处理机网络的拓扑结构对于迭代时间只能有小的固定的影响。在随机连接的神经网络情况,即使严格限制节点输入的端数,但对减少通信总开销也只有微不足道的影响。研究了各种类型的模块神经网络证明了在个别情况下有较好的实现特性。根据实现的约束条件,证明随机连接的神经网络不可能是现实的智能计算机模型。
This article describes the performance analysis of neural networks, focusing on the optimal size of achievable acceleration and processing networks (transputer or multi-computer networks that communicate in a comparable way). For completely and stochastically connected neural networks, the topology of the processor network has only a small, fixed impact on the iteration time. In the case of stochastic connected neural networks, even though the number of nodes input is strictly limited, it has only a negligible impact on reducing the total cost of communication. Various types of module neural networks have been studied to demonstrate good implementation characteristics in a few cases. According to the constraints, it is proved that the neural network with stochastic connection can not be a realistic intelligent computer model.