基于FPGA的矩阵乘法在神经网络的应用.docxVIP

国内图书分类号：TN911.72国际图书分类号：621工学硕士学位论文基于 FPGA 的矩阵乘法在神经网络的应用硕 士 研究生：王春明导师：王煜，莫启良副导师：黎淑芬，徐敬生申 请 学 位：工学硕士学 科 、 专 业： 控制科学与工程所 在 单 位： 深圳研究生院答 辩 日 期： 2007 年 12 月授予学位单位： 哈尔滨工业大学Classified Index: TN911.72U.D.C: 621Dissertation for the Master Degree of EngineeringFPGA BASED APPLICATION OF MATRIXMULTIPLICATION ON NEURALNETWORKCandidate:Supervisor:Co- Supervisor:Academic Degree Applied for:Specialty:Affiliation:Date of Defence:Degree-Conferring-Institution:Wang ChunmingWang Yu, Mok Kai LeungLai Shuk Fun, Tsui King SangMaster of EngineeringControl Science and TechnologyShenzhen Graduate SchoolDecember, 2007Harbin Institute of Technology摘要近年来人工神经网络的发展十分迅速，许多的软件或硬件方法都可以用来实现神经网络。利用软件实现神经网络比较方便快捷，而用硬件来实现则可以得到更高的运行速度。本文给出了一种基于可编程门阵列（FPGA）的三层前馈神经网络硬件实现技术。所用的神经网络学习方法是误差反向传递（BP）算法。利用硬件系统实现了前馈网络，反馈网络，权值更新，学习/工作模式控制等全部功能。为了提高处理速度，几种不同结构的脉动阵列被用来计算其中最耗时的几个矩阵乘法运算。编译报告和仿真结果显示，利用 FPGA 可以较好地实现神经网络中的并行计算问题，使系统具有很高运算速度。同时利用 FPG 的层级设计和模块设计方法，使系统具有快捷灵活等优点。另外对 FPGA 开发过程中的一些问题也进行了初步的分析。关键字：人工神经网络；可编程门阵列；误差反向传递；矩阵乘法ABSTRACTThe development of artificial neural network is very fast in resent years. Thereare many implementation methods both in software and hardware. Although it is moreconvenience to handle neural net problems or do research in software, the operationspeed may not be able to meet the requirement of certain applications, sometimeshardware implementation of neural net is necessary.The proposed design is a three layer feed forward neural net realized with FPGA,and its learning rule is the backpropagation (BP) algorithm. Both the feed forward netand the feedback net are implemented in FPGA. The key operation in BP algorithm ismatrix multiplications, which requires high degree of parallelism and is often computedwith systolic array. In this FPGA design, systolic arrays with different architectures areapplied for different forms of matrix multiplications in the neural net. The high paralleloperation ability of FPGA offers a high processing speed. Still, coupled withhierarchical design method and