基于手写体数字识别的激活函数对比研究  

Comparative Study of Activation Functions Based on Handwritten Digital Recognition

作　　者：宋倩 罗富贵[1] SONG Qian;LUO Fugui(Academy of Big Data and Computer Science,Hechi University,Hechi 546300,China)

机构地区：[1]河池学院大数据与计算机学院,广西河池546300

出　　处：《现代信息科技》2023年第4期95-97,共3页Modern Information Technology

基　　金：2022年度广西中青年教师科研基础能力提升项目(2022KY0606);河池学院2021年科研启动项目(2021GCC013);2021年度广西中青年教师科研基础能力提升项目(2021KY0620,2021KY0615)。

摘　　要：针对手写体数据集在卷积神经网络训练时分别使用sigmoid、tanh、relu与mish激活函数进行实验,对这四种激活函数下训练集误差值收敛曲线以及测试集精度曲线进行算法结果分析,实验结果表明,mish激活函数在手写体数据集上进行卷积神经网络训练时效果更佳,而同样得出不同的训练数据集、激活函数的选择会影响算法对数据的拟合度和神经网络的收敛性。For the training of Convolution Neural Network for handwritten data sets, sigmoid, tanh, relu and mish activation functions are used to conduct experiments, and the algorithm results are analyzed for the convergence curve of the error value of the training set and the accuracy curve of the test set under these four activation functions. The experimental results show that the mish activation function performs better in the training of Convolution Neural Network on handwritten data sets. It is also concluded that the selection of different training data sets and activation functions will affect the fitting degree of the algorithm to the data and the convergence of the neural network.

关 键 词：手写体数字识别 mish激活函数 卷积神经网络 mnist数据集 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]