面向深度卷积网络的多目标神经演化算法  

Multi-objective Neural Evolutionary Algorithm for Deep Convolutional Networks

作　　者：陈禹行 胡海根[1] 刘一波 郝鹏翼 李小薪[1] 周乾伟 CHEN Yu-hang;HU Hai-gen;LIU Yi-bo;HAO Peng-yi;LI Xiao-xin;ZHOU Qian-wei(School of Computer Science and Software,Zhejiang University of Technology,Hangzhou 310000,China)

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310000

出　　处：《小型微型计算机系统》2021年第1期71-77,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61802347,61801428)资助;浙江省自然基金项目(LY18F020034)资助。

摘　　要：传统的深度卷积神经网络设计方法依赖于人工设计以及反复试错,只能采用形式单一的网络结构,导致其参数过分冗余,乘法次数巨大.为了自动化地设计出结构灵活多变,网络规模及计算量较小的深度卷积神经网络,本文提出了一种面向深度卷积网络的多目标神经演化算法.该算法将深度神经网络表达成有向图,使用神经演化和多目标优化算法实现了深度、计算量和识别率下的多目标同时优化,同时还引入了线性规划用于将基因编码翻译为卷积层的配置参数,使得演化算法可以自动调整各个网络层的具体配置.演化得到的模型其最深路径上含有36个卷积层,CIFAR-100上Top5精度为86.1%,Top1精度为60.2%,与识别率相近的网络相比,具有结构新颖,乘法次数低等特点.综上,本文提出的方法能够自动生成一系列各具特色的深度神经网络,可根据在深度、计算量和识别率3个指标上的不同应用需求选择适合的深度神经网络,为深度神经网络部署于资源受限的无线传感器网络上提供了一种快速、经济、自动化的设计方法.Because designed by hand plus limited times of trial and error,traditional deep convolutional neural netw ork design methods usually result in excessively redundant parameters and huge multiplication times.To automatically construct deep convolutional neural networks with features like flexible structures,small-scale,and low multiplication times,this paper proposes a multi-objective neural evolution algorithm for deep convolutional networks.The algorithm expresses a deep neural network as a directed graph and uses neural evolution and multi-objective optimization algorithm to achieve simultaneous multi-objective optimization of depth,computational costs,and recognition accuracy.The method uses linear programming to translate genetic codes into convolutional layers to allow the evolutionary algorithm to automatically adjust the specific configuration of each netw ork layer.The evolved model has 36 convolution layers and Top5 accuracy 86.1%,Top1 accuracy 60.2%on CIFAR-100.Compared w ith netw orks w ith similar recognition accuracy,the evolved netw ork has more novel structures and few er multiplication times.In summary,the proposed approach can automatically create a series of deep neural netw orks w ith different features.It is a fast,economical,and automated design method for industrial applications,especially resource-limited applications.

关 键 词：深度卷积网络 神经演化 基因编码 多目标优化 无线传感器网络 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]