一种嵌入式计算平台的Sigmoid函数优化方法  

Optimization Method for Embedded Platform

作　　者：林钰棽 魏云龙[1] 陈琪琪 张威 邱志敏 LIN Yu-shen;WEI Yun-long;CHEN Qi-qi;ZHANG Wei;QIU Zhi-min(School of Physics and Information Engineering,Fuzhou University,Fuzhou 350100,China;School of Electronics and Information,South China University of Technology,Guangzhou 510640,China;Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033,China;School of Electronic and Information Engineering,Beihang University,Beijing 100191,China)

机构地区：[1]福州大学物理与信息工程学院,福州350100 [2]华南理工大学电子与信息学院,广州510640 [3]中国科学院大学中国科学院长春光学精密机械与物理研究所,长春130033 [4]北京航空航天大学电子信息工程学院,北京100191

出　　处：《小型微型计算机系统》2021年第10期2053-2058,共6页Journal of Chinese Computer Systems

基　　金：数字福建物联网工程应用实验室建设项目(82917002)资助。

摘　　要：Sigmoid函数作为人工神经网络常用的激活函数,属于超越函数.传统的计算方法复杂度高、资源消耗大,在嵌入式平台上计算效率较低,针对此问题,本文提出一种新的优化方法——分段极限近似法.首先根据Sigmoid函数在中间变化快、两端变化缓慢的特点,将其分为常数区和非线性拟合区;其次,根据第2个重要极限公式将Sigmoid函数中的e指数计算转换为log2n次乘法计算,简化e指数计算进而降低Sigmoid函数运算复杂度.最后,在嵌入式计算平台上建立BP神经网络并利用UCI经典数据集对所提出的方法进行验证.实验结果表明在ARM Cortex-M平台上实现Sigmoid函数,利用分段极限近似法比利用标准C math库的exp函数实现在运算速度上提高23.67倍,同时其拟合误差在平均绝对误差小于0.001的情况下不会造成神经网络判别准确率的下降.As a common activation function of artificial neural network,sigmoid function belongs to transcendental function.The traditional calculation method has high complexity,high resource consumption and lowcomputational efficiency on embedded platform.In order to solve this problem,this paper proposes a newoptimization method-piecewise limit approximation.Firstly,the Sigmoid function is divided into the constant region and the nonlinear fitting region according to its characteristics of fast change in the middle and slowchange in both ends.Secondly,according to the second important limit formula,the computation of e exponent in the Sigmoid function is transformed into the calculation of log2n times multiplication,which simplifies the computation of e exponent and thus reduces the computational complexity of the Sigmoid function.Finally,BP neural network is established on the ARMCortex-Mplatform and the proposed method is verified by using UCI classic data set.The experimental results showthat the speed of implementation of sigmoid function on ARMCortex-Mplatform is 23.67 times faster than that using exp function of standard c math library.At the same time,when the average absolute error is less than 0.001,the accuracy of neural network discrimination will not decrease.

关 键 词：Sigmoid激活函数 非线性拟合 分段极限近似法 ARM Cortex-M BP神经网络 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]