基于混合CORDIC的神经网络激活函数的实现  

Implementation of activated function of neural networks by using hybrid CORDIC

作　　者：陈曦[1] 王高峰[1] 刘克刚[1] 徐江丰[2] 

机构地区：[1]武汉大学微电子与信息技术研究院 [2]中国科学院上海技术物理研究所,上海200083

出　　处：《华中科技大学学报（自然科学版）》2007年第9期114-117,共4页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(60444004)

摘　　要：讨论了基于混合坐标旋转数字计算机算法设计并实现对数-S形激活函数的方法,采用超高速集成电路硬件描述语言和流水线技术构造的对数-S形函数的寄存器传输级模块在现场可编程门阵列上给予硬件实现,优化后的对数-S形函数模块结合了查找表和坐标旋转数字计算机迭代算法的特点,具有高效率、高速度、高精度等优点.实验数据表明,本设计模块计算结果的平均误差为0.05%,最大误差为0.19%,最大工作频率为109 MHz,满足神经网络超大规模集成电路的要求.Novel method to design Log-Sigmoid activation function is proposed through hybrid co-ordinate rotation digital computer （CORDIC） algorithm. The proposed architecture of Log-Sigmoid function is captured in very high speed integrated circuit hardware description language （VHDL） and synthesized using QuartusⅡ 5.0 for Altera StratixII field programmable gate arrays （FPGA）. In order to enhance the efficiency and accuracy of implementation on StratixII FPGA, the technology of pipeline and look-up table have been utilized. The simulated and synthesized results show that the Log-Sigmoid function based on hybrid CORDIC can be implemented with a average error of 0.05 %, a maximum error of 0.19 %, and a maximum frequency of 109 MHz, from which it can be concluded that the designed Log-Sigmoid function works accurately and efficiently.

关 键 词：神经网络 对数-S形函数 坐标旋转数字计算机 现场可编程门阵列 流水线 

分 类 号：TN47[电子电信—微电子学与固体电子学]