深度神经网络剪枝方法综述  

A survey of pruning methods based on deep neural networks

作　　者：毛远宏 曹健[2] 贺鹏超 刘曦[1] 柴波[1] MAO Yuanhong;CAO Jian;HE Pengchao;LIU Xi;CHAI Bo(Xi'an Microelectronics Technology Institute,Xi’an 710054,Shaanxi,China;School of Software and Microelectronics,Beijing University,Beijing 100038,China)

机构地区：[1]西安微电子技术研究所,陕西西安710054 [2]北京大学软件与微电子学院,北京100038

出　　处：《微电子学与计算机》2023年第10期1-8,共8页Microelectronics & Computer

基　　金：西安微电子技术研究所2022年创新支持项目(YL-220009)。

摘　　要：目前深度神经网络在计算机视觉和语音处理上获得了广泛应用,但是深度神经网络模型参数量和计算量巨大,通常在资源有限的嵌入式应用上部署困难.在基本不影响计算精度的前提下,剪枝技术可以对于深度神经网络模型进行有效压缩和加速,因此成为了目前研究热点.本文论述了深度神经网络剪枝的相关问题和理论,归纳总结近年来面向深度神经网络的剪枝方法,对于当前主流的剪枝方法进行了分类梳理,并对于未来和网络结构搜索相结合的发展方向进行了展望.Deep neural networks are widely used in computer vision and speech processing,but the large number of parameters and computation of deep neural network models make it difficult to deploy them in embedded applications with limited resources.The network pruning can effectively compress and accelerate the deep neural network models without affecting the computational accuracy,and therefore has become a hot research topic.This paper discusses the problems and theories related to pruning of deep neural networks,and summarizes the pruning methods for deep neural networks in recent years,classifies the current main pruning methods,and provides the future development of the network architecture search combination.

关 键 词：深度神经网络 网络剪枝 嵌入式应用 网络结构搜索 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]