基于FPGA的卷积神经网络浮点激励函数实现  

The Implementation of Activation Function of CNN Based on FPGA Using Single Precision Floating-Point-Unit

作　　者：李施豪 应三丛[1] 

机构地区：[1]四川大学计算机学院视觉合成图形图像技术国家重点学科实验室,四川成都610065

出　　处：《微电子学与计算机》2017年第10期105-109,共5页Microelectronics & Computer

基　　金：国家"八六三"计划项目(2015AA016405);四川省科技厅科技支撑项目(2016GZ0097)

摘　　要：卷积神经网络因深度学习概念的提出再一次被研究人员所重视.激励函数是卷积神经网络的一个重要组成部分,选取了sigmoid函数作为实验对象.讨论了当前几种可行的逼近方法,最终采用分段四阶多项式拟合sigmoid函数.在FPGA上使用Verilog硬件描述语言设计了并行电路,并采集了数据集进行FPGA与CPU版本caffe库进行运算效率对比.实验结果表明,此种方法误差小效率高,FPGA在深度学习领域有着广阔的应用前景.Convolutional Neural Network （CNN） was gotten attention because of the presentation of Deep-Learning. It has large research value for Deep-Learning in using FPGAs. The activation function is the most necessary part of CNN. In this paper, Sigmoid function was chosen as the experiment object. The approximation ways of Sigmoid Function were listed and analysed, the piecewise forth-order approximation was the best way to fit Sigmoid Function. Parallel calculation circuit was designed by Verilog HDL on FPGAs, collecting dataset and inputing them to FPGA Platform and CPU Platform. The experiment result set forth that this solution had high efficiency and low error with rosy prospect about Deep-Learning and FPGA.

关 键 词：卷积神经网络 激励函数 FPGA 多项式逼近 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]