神经架构搜索技术研究综述  

Review of research on neural architecture search technology

作　　者：武家辉 李科研 陈丽新 张家诺 刘帅兵 逯鹏[1,2,3] Wu Jiahui;Li Keyan;Chen Lixin;Zhang Jianuo;Liu Shuaibing;Lu Peng(Institute of Electrical&Information Engineering,Zhengzhou University,Zhengzhou 450001,China;Robot Perception&Control Henan Engineering Laboratory,Zhengzhou 450001,China;Research Center for Intelligent Science&Engineering Technology of TCM,Zhengzhou 450001,China)

机构地区：[1]郑州大学电气与信息工程学院,郑州450001 [2]机器人感知与控制河南省工程实验室,郑州450001 [3]中医药智能科学与工程技术研究中心,郑州450001

出　　处：《计算机应用研究》2025年第1期11-18,共8页Application Research of Computers

基　　金：国家重点研发计划资助项目(2020YFC2006100);河南省高等学校重点科研资助项目(22A520009)。

摘　　要：神经架构搜索(NAS)的目的是为特定任务自动寻优生成高性能网络架构,从而减少架构设计对专家经验的依赖和架构设计过程中的人力资源消耗,其主要包含搜索空间、搜索策略和评估策略三个组成部分。早期NAS需要多个GPU耗时多天完成搜索,搜索耗时和计算成本高是NAS的核心问题。为帮助研究人员快速、全面地了解NAS领域,提供了一种新的视角对现有NAS工作进行梳理。首先对NAS的早期工作进行分析,并阐述了NAS的核心问题及其产生原因;然后围绕解决NAS核心问题的三类方法,即减小架构搜索范围、减少待评估架构搜索时间、减少架构评估时间,对该领域算法进行针对性分析、对比、总结;最后归纳总结该领域后续的主要研究方向。The purpose of NAS is to automatically optimize and generate high-performance network architectures for specific tasks,in order to reduce the dependence of architecture design on expert experience and human resource consumption in the architecture design process.It mainly includes three components:search space,search strategy,and evaluation strategy.Early NAS requires multiple GPUs to complete searches in multiple days,and the high search time and computational cost are the core issues of NAS.To help researchers quickly and comprehensively understand the field of NAS,this paper provided a new perspective to sort out existing NAS work.Firstly,this paper analyzed the early work of NAS and elucidated the core issues and their origins.Secondly,focusing on the three categories of methods to address the core issues of NAS:reducing the search space of architectures,decreasing the time for evaluating candidate architectures,and reducing the time for evaluating architectures,this paper conducted a targeted analysis,comparison,and summary of algorithms in this field.Finally,it summarized the main research directions in this field for future work.

关 键 词：神经架构搜索 搜索范围 搜索时间 评估时间 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]