文献类型：期刊文献

中文题名：Research of Neural Network Structural Optimization Based on Information Entropy

英文题名：Research of Neural Network Structural Optimization Based on Information Entropy

作者：Wang, Danyang[1];Shao, Fangming[1]

机构：[1]East China Univ Sci & Technol, Sch Sci, Shanghai 200237, Peoples R China

年份：2020

卷号：29

期号：4

起止页码：632

中文期刊名：Chinese Journal of Electronics

外文期刊名：CHINESE JOURNAL OF ELECTRONICS

收录：CSTPCD;;EI(收录号：20203809192417);Scopus;WOS:【SCI-EXPANDED(收录号:WOS:000575267100005)】；CSCD:【CSCD2019_2020】；

基金：This work is supported by the National Natural Science Foundation of China(No.61040040).

语种：英文

中文关键词：Information entropy;Decision tree;Fully connected neural network;Heuristic algorithm

外文关键词：Information entropy; Decision tree; Fully connected neural network; Heuristic algorithm

摘要：In the application of deep learning,the depth and width of the neural network structure have a great influence on the learning performance of the neural network.This paper focuses on structural optimization of depth and width,leveraging the information entropy model and decision tree strategy as feature selection and structural adjustment to optimize neural network candidates.Therefore,a decision tree-based heuristic optimization algorithm for neural network structural adjustment is proposed.Furthermore,the proposed approach is applied to fully-connected neural networks trained on the Iris dataset,and the proposed approach is verified effective via experimental simulation.
In the application of deep learning, the depth and width of the neural network structure have a great influence on the learning performance of the neural network. This paper focuses on structural optimization of depth and width, leveraging the information entropy model and decision tree strategy as feature selection and structural adjustment to optimize neural network candidates. Therefore, a decision tree-based heuristic optimization algorithm for neural network structural adjustment is proposed. Furthermore, the proposed approach is applied to fully-connected neural networks trained on the Iris dataset, and the proposed approach is verified effective via experimental simulation.