臺灣博碩士論文加值系統

硬體式的BP神經網路處理器通常遭遇到輸出入資料格式處理不便、演算法固定不易修改等問題。而軟體式的BP神經網路處理器，可能會花費太多的時間在連結權重的計算上，而沒有良好的效能。因此本論文係關於軟硬整合式BP神經網路處理器的設計與實作，其中由硬體和軟體兩部份的模組所構成，硬體模組是採用單指令多資料流心跳式陣列處理器(SIMD-Systolic Array Processor)，軟體模組則是採用C++程式語言撰寫而成。
在處理的過程中。首先，軟體模組將大型的矩陣運算分割成16×16的單位矩陣送入硬體模組中。運算完成後，硬體模組將結果傳回系統記憶體中。最後，由軟體模組修正神經網路的連結權重。

Usually, a hardware-based BP neural net processor suffers from some problems such as the inconvenience of data formats and the difficulty of modifying the algorithm, and a software-based BP neural net processor may consume too much computing time to have a real-time response. Therefore , this thesis is related to the design and implementation of a BP neural net processor which consists of both hardware and software modules. The hardware module is a SIMD-Systolic matrix processor and the software module is a C++ program running on personal computer.
During processing, first, the software module partitions large matrix operations into 16×16 units and send them to the hardware module through PCI interface. After finishing the matrix computation, the hardware module returns the result to the system memory. Finally, the software module update the connection weights for the BP neural net.
1.1 研究動機
1.2 研究背景
1.3 論文組織及概觀
Ch2.系統架構
2.1 系統方塊圖
2.2 硬體模組介紹
2.3 軟體模組介紹
Ch3.硬體模組驅動程式
3.1 MS-Windows Driver Architecture
3.2 PCI介面卡
3.3 Design of Bus Master DMA Driver
3.4 矩陣運算的前置處理
3.5 VxD和DLL的連結
Ch4.BPN軟體處理模組
4.1 神經網路簡介
4.2 BP神經網路的模型推導
4.3 BP神經網路的演算法
4.4 BP神經網路訓練方法的改良
4.5 BPN軟體處理模組
4.6 BPN神經網路處理器介面
Ch5.結論
5.1 測試與分析
5.2 未來展望
參考文獻
作者簡介

[1] 游竹, "計算逆傳遞類神經網路演算法之超大型積體電路設計",國立台灣工業技術學院電子工程技術研究所碩士論文,1993
[2] 廖廷彬,"SIMD-Systolic處理陣列之超大型積體電路設計及其在BP神經網路之應用",國立台灣工業技術學院電子工程技術研究所碩士論文,1996
[3] 劉益郎,"BP神經網路處理器之設計與分析",國立台灣工業技術學院電子工程技術研究所碩士論文,1997
[4] 劉萬吉,"倒傳式類神經網路處理系統設計",國立台灣科技大學電子工程技術研究所碩士論文,1998.
[5] 吳乾彌,”管線式-心跳式-單指令多資料流陣列處理架構及其方法”,中華民國專利發明第70841號,1995
[6] 吳乾彌,”辨認系統及VLSI處理器之設計”,國科會專題計畫NSC85-2215-e011-0004,1997
[7] Chen-Mie Wu, Rong-Tyh Wang and Chin-Tsh Yen, “Addeer-based SIMD-Systolic architectures and VLSI chip for computing the Winograd small FFT algorithms”, International Journal of Electronics, Vol76,pp1135-1149,1994
[8] "PCI Local Bus Specification 2.1", PCI Special Interest Group,1995
[9] "PCI BIOS Specification 2.1", PCI Special Interest Group, 1994
[10] Tom Shanley, "PCI System Architecture", Addison-Wesley, 1997.
[11] Karien Hazzah, "Writing VxD for Windows 95",Miller Freeman,1997
[12] Walter Oney, "System Programming for Windows 95", Microsoft Press,1996
[13] James A.Freeman, David M. Skapura, "Neural Networks: Algorithms, Applications and Programming Techniques", Addison-Wesley,1992.
[14] Adam Blum, "Neural Networks In C++", John Wiley & Sons ,1992
[15] A. Cichocki,R. Unbehauen, "Neural Networks for Optimization and Signal Processing", John Wiley & Sons,1992
[16] P. Werbos, "Beyond Regression:New tools for Prediction and Analysis in the Behavioral Sciences", PhD thesis Harvard,1974
[17] D.B.Parker. ”Learning Logic Technical Reoprt TR-47,Center for Computational Research in Economics and Management Science”, MIT,1985
[18] James McClelland and David Rumelhart, “Explorations in Parallel Distributed Processing”, MIT Press,1986
[19] Yoshio Hirose,Koichi Yamashita,Shimpei Hijiya,"Back-Propagation Algorithm Which Varies the Number of Hidden Units", Neural Networks,Vol4 pp61-64,1991
[20] Eric B. Bartlett, “Dynamic Node Architecture Learning:A Information Theoretic Approach”, Neural Networks Vol7,pp129-140,1994