深度学习在云和终端上的混合分布式部署研究

Research on Hybrid Distributed Deployment of Deep Neural Networks on Cloud and Terminal

深度学习算法在物联网终端设备上的应用存在着系统开销控制与保证精度和实时性之间平衡的问题。本文提出了一种在云和终端设备上分布式混合部署深度学习神经网络的方法:压缩深度神经网络在本地终端上执行快速的推理运算;当系统基于可信表现的判断标准需要进一步处理时,中间数据可传输至云服务器端,进一步利用云端的深层深度神经网络进行处理,以提高系统的表现精度。本文给出了深度神经网络在终端设备上部署时和在终端与云端上混合部署时进行推理运算的量化比较效果,结果显示此种方法兼顾了深度神经网络的系统开销和准确率。

The application of deep learning algorithm on IoT terminal equipment has the problem of balance between system overhead control and guarantee accuracy and real-time performance.This paper proposed a method for distributed hybrid deployment of deep learning neural networks on cloud and terminal equipment:compressed depth neural network performed fast inference operations on local terminals;when the system need further processing based on the judgment criteria of trusted performance,the intermediate data could be transmitted to the cloud server side,and was further processed by the deep neural network on cloud to improve the performance accuracy of the system.This paper also gave the quantitative comparison effect of the inference operation of the deep neural network when deployed on the terminal device and when the terminal and the cloud werehybrid deployed.The results showed that this method took into account the system overhead and accuracy of the deep neural network.