摘要
近些年,卷积神经网络的发展日趋成熟,在图像识别、目标检测、自然语言处理等多个领域的性能超过了传统的机器学习算法。然而随着网络性能的提升,部署和运行神经网络对于硬件设备要求越来越高。为了将卷积神经网络部署到算力较低的设备,必须在参数量和准确率之间寻求一个平衡。本文首先介绍三种常规的卷积方式,然后从参数量和计算量的角度加以分析,接着详细介绍近些年比较流行的轻量化网络模型结构,对于各自的优缺点进行总结,同时结合其在ImageNet上面的表现评估其轻量化技巧的优劣,最后展望轻量化神经网络的发展前景。
In recent years,the development of convolutsion neural networks has become increasingly mature,outperforming traditional machine learning algorithms in several fields such as image recognition,object detection,natural language processing and so on.However,as the performance of the network increases,deploying and running neural networks is increasingly demanding for hardware devices.In order to deploy convolutional neural networks to devices with lower computation power,a balance must be sought between the number of parameters and the accuracy of the network.In this paper,we first introduce three conventional convolutional approaches,then analyze them in terms of the number of parameters and computational power,then introduce the popular lightweight network model structures in recent years,summarize their advantages and disadvantages,and evaluate the pros and cons of their lightweight techniques in the context of their performance on ImageNet,and finally look forward to the development of lightweight neural networks.
作者
包志龙
BAO Zhi-long(Department of Information Science and Technology,Ningbo University,Ningbo 315211,China)
出处
《无线通信技术》
2022年第1期36-41,47,共7页
Wireless Communication Technology
关键词
轻量化
分组卷积
深度可分离卷积
卷积神经网络
lightweight
group convolution
depthwise separable convolution
convolution neural network