神经架构搜索技术研究综述OA北大核心

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 is-sues 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 sum-marized the main research directions in this field for future work.