基于多参数融合优化的深度神经网络设计研究  

Research on Deep Neural Network Design Based on Multi-parameter Fusion Optimization

作　　者：蒋正锋[1] 廖群丽 Jiang Zhengfeng;Liao Qunli(College of Mathematics,Physics and Electronic Information Engineering,Guangxi Normal University for Nationalities,Chongzuo 532200;School of Electrical Engineering&Intelligentization,Guangxi Vocational Institute of Technology,Chongzuo 532200)

机构地区：[1]广西民族师范学院数理与电子信息工程学院,崇左532200 [2]广西理工职业技术学院电子工程与智能化学院,崇左532200

出　　处：《现代计算机》2021年第31期13-24,共12页Modern Computer

基　　金：广西民族师范学院科研项目(2020YB006)。

摘　　要：针对特定业务分类的需求,无从下手设计深度神经网络模型的问题,提出一种基于多参数融合逐步优化深度神经网络的设计方法。先通过分析样本和类别确定初始的无隐层神经网络模型,然后在初始神经网络的基础上逐步增加隐层,构建多参数融合优化的深度神经网络。基于TensorFlow框架,以手写体数字识别为例,逐步设计一个深度神经网络模型,整个实验过程中不断调整网络结构、激活函数、损失函数、优化器、学习率和样本批次大小,最终设计出多参数融合优化的具有较高准确率的深度神经网络模型,为构建神经网络提供了一种有效的思路。Aiming at the demand of specific business classification,it is impossible to design deep neural network model,and a design method based on multi-parameter fusion stepwise optimization deep neural network is proposed.Firstly,the initial non-hidden layer neural network model is determined by analyzing samples and categories.Then,the hidden layer is gradually added on the basis of the initial neural network,and a deep neural network with multi-parameter fusion optimization is constructed.Based on the Tensor⁃Flow framework,handwritten digital recognition is taken as an example to design a deep neural network model.The network structure,activation function,loss function,optimizer,learning rate and sample batch size are continuously adjusted throughout the experiment.Finally,multiple parameters are designed.Fusion-optimized deep neural network model with high accuracy provides an effective way to construct neural network.

关 键 词：多参数融合 深度神经网络 TensorFlow 手写体数字 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]