摘要
阐述现有卷积神经网络模型的优化方式。结合硬件条件,探讨低秩分解、知识蒸馏、量化和剪枝四种优化方式提出的原因、主要优化方式、大致流程。其中分析低秩分解和知识蒸馏的演化过程,介绍量化和剪枝的具体优化方式和涉及的相关基础知识。经过对比总结,得出量化和剪枝两种深度学习优化方式,对于软件成本的控制、领域的应用、与硬件的连接都有更广泛的使用。
This paper describes the optimization methods of existing convolutional neural network models.Based on hardware conditions,explore the reasons,main optimization methods,and general process for proposing four optimization methods:low rank decomposition,knowledge distillation,quantization,and pruning.It analyzes the evolution process of low rank decomposition and knowledge distillation,introduces specific optimization methods for quantization and pruning,and the relevant basic knowledge involved.After comparison and summary,it has been concluded that there are two deep learning optimization methods,quantification and pruning,which are more widely used for controlling software costs,applying in various fields,and connecting with hardware.
作者
曹毅杰
CAO Yijie(School of Intelligent Science and Control Engineering,Jinling Institute of Technology,Jiangsu 211199,China)
出处
《集成电路应用》
2024年第4期411-415,共5页
Application of IC
关键词
深度学习
低秩分解
知识蒸馏
量化与剪枝
deep learning
low rank decomposition
knowledge distillation
quantization and pruning