Research On Neural Network Computing Method For Edge Intelligence

With the development of the Internet of Everything,many new application scenarios such as smart cities and unmanned driving have gradually emerged,and the requirements for latency have become more demanding.Therefore,mobile edge computing was proposed.Deep neural network is widely used in new application scenarios,and its computing needs a lot of resources,but the computing and storage resources of edge devices are constrained,so it is difficult to deploy deep model directly on edge devices.Therefore,we conduct research from two aspects.On the one hand,the original network is compressed into a lightweight network,and reducing the demand for resources.Early exit mechanism is introduced to make the calculation process dynamically adjustable.On the other hand,it uses multiple edge devices to compute the deep neural network,which reduces the calculation load of a single device and speeds up the computing speed.In view of the above problems and challenges,based on the research status,this paper studies the edge intelligence oriented neural network computing method,the main work and innovations are as follows:Firstly,in order to solve the problem of single edge device neural network computing in edge intelligence,this paper designs the deep neural network on-demand inference method Edge KE to achieve a balance between resource consumption and inference performance on resource-constrained edge devices.Use knowledge distillation to compress neural networks to reduce the demand for computing resources and use early exit techniques to provide flexible calculation methods for neural networks,optimize model execution on demand,and meet different performance requirements,including delay-oriented reasoning methods and oriented Reasoning method of precision.Test different edge devices under different delay and accuracy requirements.Experimental results show that Edge KE is superior to the original model.When the inference accuracy is only 0.16% lower than the original network,not only the computing delay is reduced by 1.83x-2.51 x,but also memory footprint has also been reduced by 2.42 x.It is also verified that Edge KE can effectively meet different inference performance requirements.Under various delay constraints,the accuracy loss is within 4.75%.Under various inference accuracy requirements,the speedup ratio can reach 3.71 x.In terms of inference performance and resource consumption to achieve a good balance.Secondly,in order to solve the neural network computing problem of multi-edge devices in edge intelligence,this paper designs a deep neural network distributed computing method Edge MI,which implements distributed computing of deep neural networks on heterogeneous edge clusters.A time prediction model is proposed to predict the computing time of the convolutional layer and the fully connected layer of the deep model.We design a convolutional layer computing task division scheme to reduce the idle waiting time of edge devices,make full use of the computing resources of the edge cluster,and speed up the computing speed.Finally,we design data scheduling strategies to reduce the frequency of data exchange between edge devices.The deep neural network distributed computing was tested on the edge cluster.The experimental results show that Edge MI is superior to the traditional solution,and its computing speed is increased by 14.34%.The data scheduling strategy increases the computing speed by 1.07x-1.22 x.When the number of edge nodes is from 2 to 4,the speedup ratio of the deep model reached 1.84x-3.57 x.