三层BP神经网络隐层节点数确定方法探究  

Research on the Methods of Determining the Number of Hidden Nodes in Three-layer BP Neural Network

作　　者：蔡荣辉 崔雨轩 薛培静 CAI Rong-hui;CUI Yu-xuan;XUE Pei-jing(College of mathematics and statistics, Central South University,Changsha 410000,China)

机构地区：[1]中南大学数学与统计学院,湖南长沙410000

出　　处：《电脑与信息技术》2017年第5期29-33,共5页Computer and Information Technology

摘　　要：为探究三层BP神经网络最佳隐层节点数的确定方法,对现有经验公式法、试凑法和奇异值法等隐层节点确定方法与分块有效像素统计值、HOG和傅里叶变换等特征提取方法在2600个手写小写英文字母样本上仿真识别的效果进行对比分析。据此引入公式可信度的概念,并基于加权思想提出了一种新的确定隐层节点数的方法。结合标准正态函数的反函数对新方法权数的确定进行改进,仿真实验验证结果表明该方法测试识别率最高可达96.667%。In order to find out the best method of determining the number of hidden layer nodes in three-layer BP neural network,three methods of determining the number of hidden nodes including empirical formula method,trial-and-error method and singular value method and three feature extraction methods including blocks effective pixels statistics,HOG and Fourier transform are used to compare and analyze the recognition effect of2600handwritten lowercase samples.Accordingly,the concept of formula credibility is put forward and a new method of determining the number of hidden layer nodes is proposed based on the idea of weighting.Besides,the inverse of the standard normal function is used to improve the weight of the new method.Finally,the simulation results show that the recognition rate of the new method is up to96.667%.

关 键 词：BP神经网络 隐层节点数 特征提取 加权方法 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]