应用感知的深度神经网络剪枝方法  

A Pruning Method for Deep Neural Networks Using Perception

作　　者：白学营 陈维常 徐菲 胡健伟 BAI Xueying;CHEN Weichang;XU Fei;HU Jianwei(Institute of Systems Engineering Military Academy of Science,Beijing 100010,China)

机构地区：[1]中国人民解放军军事科学院系统工程研究院,北京100010

出　　处：《宇航总体技术》2024年第5期64-73,共10页Astronautical Systems Engineering Technology

基　　金：国家部委基金资助项目。

摘　　要：神经网络模型面临着参数过多、计算负载高和内存开销迅速增长等问题。为了满足各种应用场景和设备的多样性,有必要针对特定应用优化神经网络模型。鉴于此,提出了一种基于应用类别感知的深度神经网络剪枝方法,该方法分析了在卷积神经网络前向传播过程中,不同滤波器在提取类别特征方面的不同作用。获得了滤波器的重要性及与应用类别之间的关系,针对特定应用中的不同目标类别进行了定制化的剪枝优化,并通过PyTorch深度学习框架进行了设计和实现。实验结果表明,所提出的基于应用感知的神经网络优化方法能够有效地优化卷积神经网络。Neural network models face challenges such as an excessive number of parameters,high computational load,and rapidly increasing memory overhead.To accommodate the diversity of application scenarios and devices,it is essential to optimize neural network models for specific applications.This work proposes a deep neural network pruning method that is aware of application categories.This method analyzes the distinct roles of various filters in extracting class features during the forward propagation process in convolutional neural networks.It identifies the relationship between the importance of filters and application categories,and carries out customized pruning optimization for different target categories in specific applications.The work is designed and implemented based on the PyTorch deep learning framework.Experimental results demonstrate that the application-aware neural network optimization method proposed in this paper can effectively optimize convolutional neural networks.

关 键 词：卷积神经网络 神经网络优化 应用感知 滤波器剪枝 知识蒸馏 

分 类 号：V419.9[航空宇航科学与技术—航空宇航推进理论与工程]