| With the continuous application of deep learning methods in edge-cloud collaboration application scenarios,implementing edge intelligence through neural network model inference at network edges will be the future development trend.However,the inference energy consumption of deep neural networks is gradually increasing with the increase of task complexity,and most edge device rely on limited batteries for power supply,which seriously hinders the application of neural networks in edge device.In addition,the computational scheduling strategy of neural networks not only affects the resource allocation of edge-cloud collaborative systems,but also determines the inference energy consumption and efficiency of neural networks.This paper investigates the problems of deep neural network model compression and scheduling for edge-cloud collaboration,and the main work content is as follows:1)Aiming at the problem of inference energy consumption calculation of neural networks in Edge device,a deep neural network inference energy consumption calculation model oriented to edge-cloud collaboration is proposed.The purpose is to accurately evaluate the inference energy consumption of neural networks,so that neural networks can be deployed to Edge device more flexibly.Based on the energy consumption calculation framework for neural network inference on hardware devices,the calculation and analysis of inference energy consumption of classical neural network models is conducted by analyzing the computational characteristics of different neural network structures and the memory storage space of edge devices.2)Aiming at the problem of edge devices struggling to load high-energy consumption deep neural networks,the deep neural network compression method based on energy-constrained optimization is proposed.By applying structured pruning to neural networks,compression of the inference energy consumption of neural networks is achieved.The importance of convolution kernels is measured under energy constraints,and the model is pruned based on kernel importance to reduce neural network energy consumption to meet the energy budget of edge nodes.Experimental results demonstrate that our method can effectively compress the energy consumption of neural networks without sacrificing their accuracy.3)Aiming at the problem of computationally intensive tasks of deploying neural networks on edge node clusters,the method based on edge intelligence deep neural network scheduling is proposed.With the optimization objectives of minimizing the neural network inference energy consumption and achieving load balancing of edge nodes,the proposed method considers the energy consumption and resource requirements of neural networks on edge nodes.By improving and optimizing the information pheromone updating mechanism and heuristic factor,the convergence and search ability of the ant colony scheduling algorithm are enhanced.Experimental results demonstrate that the improved ant colony optimization algorithm enhances the resource utilization of edge nodes and reduces the neural network inference energy consumption on edge nodes. |
