动态环拓扑多机系统上DBP学习算法并行计算模型的构造与实现  

CONSTRUCTION AND IMPLEMENTATION OF THE PARALLEL COMPUTING MODEL OF DBP LEARNING ALGORITHM

作　　者：管惠维[1] 

机构地区：[1]上海大学计算机系

出　　处：《软件学报》1996年第2期111-118,共8页Journal of Software

基　　金：上海市重点学科发展基金;国家教委博士基金

摘　　要：人工神经网络模型的软件模拟，其并行算法的设计、实现及性能评价对于神经网络计算机和各种专用神经网络ＶＬＳＩ芯片的研制具有十分重要的意义．本文首先构造了一个分布式存储器、信息传递方式的多机系统作为软件模拟人工神经网络的平台，并用一个环拓扑结构的多Ｔｒａｎｓｐｕｔｅｒ网络予以实现．接着提出并实现了一个适用于动态环拓扑形式的ＤＢＰ并行计算模型，它主要包括神经元的划分和映射策略；ＤＢＰ中活性值、误差反向传播及权值修改的多机并行算法．然后讨论该ＤＢＰ算法的时间复杂度和加速比．The software simulation of artificial neural networks, and the design, implementation and evaluation of parallel algorithm about artificial neural networks are highlyimportant to the research and production of neural computers and the various of specialneural VLSI chips. In this paper, a DMMPMS (distributed--memory, message--passingmultiprocessor system) is constructed as a multicomputer architecture for simulating theartificial neural network at first, which is implemented by a multi--transputer system of aring topology form. Next, an executable parallel computing model of DBP algorithm ofmultilayered neural network on dynamic ring topology is proposed and implemented,which mainly includes the strategy of partition and map for neural cells lthe parallel algorithms of the forward computation of activations, the backpropagation of errors and theupdate of weights. Then the time complexity and speed --up ratio of the DBP algorithmare investigated.

关 键 词：神经网络 学习算法 并行算法 环拓扑 多机系统 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]