深层神经网络架构搜索综述  

Survey on Deep Neural Architecture Search

作　　者：薛羽[1] 张逸轩 XUE Yu;ZHANG Yixuan(School of Computer and Software,Nanjing University of Information Science and Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学计算机与软件学院,南京210044

出　　处：《信息网络安全》2023年第9期58-74,共17页Netinfo Security

基　　金：国家自然科学基金[61876089,61876185,61902281,61403206];江苏省自然科学基金[BK20141005];江苏省高校自然科学基金[14KJB520025]。

摘　　要：近年来,深度神经网络应用到图像识别、语音识别、目标检测、机器翻译等领域,加速了网络的性能演进与灵活性提升。但这些网络通常结构复杂,需要拥有大量专业知识的人员消耗大量时间调整参数以匹配具体环境。这样通过人工来调整参数的常规方法效率较低且错误频出。因此,神经网络架构搜索(NAS)的研究被提上日程。文章对现有的NAS相关算法进行了较全面地介绍和评价,并对未来神经网络架构搜索的发展提出构想。In recent years,deep neural networks have been applied to image recognition,speech recognition,target detection,machine translation and other aspects of life.Greatly accelerating the performance evolution and flexibility improvement of the network.But these networks often have complex structures,require personnel with a large amount of professional knowledge,and require a significant amount of time to adjust parameters to suit specific environments.The efficiency of adjusting parameters using conventional manual methods is too low and errors occur frequently.Therefore,research on neural network architecture search has also been put on the agenda.In order to provide readers with a comprehensive understanding of the research progress of neural network architecture search,the article introduced and evaluated existing relevant algorithms,and proposed ideas for the future development of neural network architecture search.

关 键 词：机器学习 自动化 深度学习 卷积神经网络 人工智能 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]