深度学习的轻量化神经网络结构研究综述  

Survey of Research on Lightweight Neural Network Structures for Deep Learning

作　　者：王军[1,2,3] 冯孙铖 程勇 WANG Jun;FENG Suncheng;CHENG Yong(School of Computer and Software,Nanjing University of Information Science and Technology,Nanjing 210044,China;Engineering Research Center of Digital Forensics of Ministry of Education,Nanjing University of Information Science and Technology,Nanjing 210044,China;Science and Technology Industry Division,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 [2]南京信息工程大学数字取证教育部工程研究中心,南京210044 [3]南京信息工程大学科技产业处,南京210044

出　　处：《计算机工程》2021年第8期1-13,共13页Computer Engineering

基　　金：国家自然科学基金(41875184);江苏省“六大人才高峰”创新人才团队项目(TD-XYDXX-004)。

摘　　要：随着深度神经网络和智能移动设备的快速发展,网络结构轻量化设计逐渐成为前沿且热门的研究方向,而轻量化的本质是在保持深度神经网络精度的前提下优化存储空间和提升运行速度。阐述深度学习的轻量化网络结构设计方法,对比与分析人工设计的轻量化方法、基于神经网络结构搜索的轻量化方法和基于自动模型压缩的轻量化方法的创新点与优劣势,总结与归纳上述3种主流轻量化方法中性能优异的网络结构并分析各自的优势和局限性。在此基础上,指出轻量化网络结构设计所面临的挑战,同时对其应用方向及未来发展趋势进行展望。With the rapid development of deep neural networks and smart mobile devices,the research of lightweight neural network structure has gradually become a hotspot.The essence of lightweight design is to optimize the storage space and improve the running speed without causing any loss to the precision of deep neural networks.Then an introduction to the mainstream methods of lightweight network structure design for deep learning is given,and the innovative features,strengths and weaknesses between the manual design methods,neural network structure search-based design methods and automated model compression-based design methods are compared.The advantages and disadvantages of the high-performance network structures generated by the above methods are also summarized.On this basis,the challenges faced by lightweight network structure design,and its applications and development trends are discussed.

关 键 词：深度学习 轻量化设计 深度可分离卷积 Octave卷积 神经网络结构搜索 模型压缩 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]