面向嵌入式应用的深度神经网络模型压缩技术综述

结合大数据的获取,深度神经网络关键技术广泛应用于图像分类、物体检测、语音识别和自然语言处理等领域.随着深度神经网络模型性能不断提升,模型体积和计算需求提高,以致其依赖高功耗的计算平台.为解决在实时嵌入式系统中的存储资源和内存访问带宽的限制,以及计算资源相对不足的问题,开展嵌入式应用的深度神经网络模型压缩技术研究,以便缩减模型体积和对存储空间的需求,优化模型计算过程.对模型压缩技术进行分类概述,包括模型裁剪、精细化模型设计、模型张量分解和近似计算和模型量化等,并对发展状况进行总结.为深度神经网络模型压缩技术的研究提供参考.

A survey on model compression of deep neural network for embedded system

Combined the big data acquisition,the key technologies of deep neural network have widely applied in the field of image classification,object detection,speech recognition,natural language processing,et al.With the developing of the deep neural network model performance, the model size and the required calculation need to be improved,so that it is reliance on high power computing platform.This paper is focus on the deep neural network model compression technology for embedded applications in order to solve the problems of storage resource,memory access speed constraints and computing resources limit in embedded system.It aims to reduce the model size and the complex computation.Meanwhile,it could optimize the process of calculation. This paper has summarized the state-of-the-art model compression technologies including model pruning,fine model designing,tensor decomposition,model quantization,etc.Through the summary on the model development,it could provide the references for the studies of the deep neural network model compression technologies.