神经结构搜索的研究进展综述

Recent Advances in Neural Architecture Search:A Survey

近年来,深度神经网络(DNNs)在许多人工智能任务中取得卓越表现,例如计算机视觉(CV)、自然语言处理(NLP).然而,网络设计严重依赖专家知识,这是一个耗时且易出错的工作.于是,作为自动化机器学习(AutoML)的重要子领域之一,神经结构搜索(NAS)受到越来越多的关注,旨在以自动化的方式设计表现优异的深度神经网络模型.全面细致地回顾神经结构搜索的发展过程,进行了系统总结.首先,给出了神经结构搜索的研究框架,并分析每个研究内容的作用;接着,根据其发展阶段,将现有工作划分为4个方面,介绍各阶段发展的特点;然后,介绍现阶段验证结构搜索效果经常使用的数据库,创新性地总结该领域的规范化评估标准,保证实验对比的公平性,促进该领域的长久发展;最后,对神经结构搜索研究面临的挑战进行了展望与分析.

In recent years,deep neural networks(DNNs)have achieved outstanding performance on many AI tasks,such as computer vision(CV)and natural language processing(NLP).However,the network design relies heavily on the expert knowledge,which is time-consuming and error-prone.As a result,as one of the important sub-fields of automated machine learning(AutoML),the neural architecture search(NAS)has been paid more and more attention to,aiming to automatically design deep neural networks with superior performance.In this study,the development process of NAS is reviewed in detail and systematically summarized.Firstly,the overall research framework of NAS is given,and the function of each research content is analyzed.Next,according to the development stage in NAS field,the existing methods are divided into four aspects,and the characteristic of each stage is introduced in detail.Then,the datasets are introduced which are often used to verify the effect of NAS methods at this stage,and the normalized evaluation criteria in NAS field are innovatively summarized,so as to ensure the fairness of experimental comparison and promote the long-term development of this field.Finally,the challenges of NAS research are proposed and discussed.