Research On Selection And Distribution Optimization Method Of Deep Learning Model Based On Multi-User Collaboration

Recently,deep learning is widely used in various fields.With the increase of the depth of deep learning model and the improvement of parameter complexity,higher and higher requirements are put forward for the computing power and storage capacity of mobile devices.Through the distributed collaborative distribution and inference of wireless network,the inference task is divided into multiple parts and shunted to cloud nodes or edge nodes,The cloud nodes or edge nodes carries out the remaining inference tasks.In this way,the mobile terminals can use the computing power and storage capacity of the cloud nodes and edge nodes to improve the efficiency of deep learning inference tasks and reduce the overall inference delay.However,in the process of tasks transmission from the terminal to the cloud node or edge node,the delay of the models downlink transmission and inference data uplink transmission will be considered.In some scenarios,due to the shortage of uplink and downlink communication resources,there will be a high distribution and inference delay,which will affect the real-time task.To solve the above problems,this paper proposes the selection and distribution method of deep learning model under the condition of wireless network.The main research and innovation points are summarized as follows:1.To reduce the inference delay of cooperative distribution in wireless distributed network,the joint optimization method of deep learning model cooperative distribution is proposed in the cooperative scenario of local and edge nodes.Firstly,this paper establishes the cooperative distribution system model of deep learning model in wireless network.We consider the distribution delay,inference delay and wireless channel status.We aim at minimizing the distribution and inference delay of the deep learning model with less loss of a model accuracy.A joint optimization scheme for collaborative distribution of deep learning model is proposed.The scheme reduces the delay of downlink distribution and inference by merging models and splitting deep learning model and downlink multicast.The optimization problem of model distribution and referring is a nonlinear shaping programming problem.In order to solve the coupling variables in the optimization problem,the original optimization problem is divided into two sub-problems:model merging and model distribution.In the model merging problem,the coalition formation algorithm is used to merge the models among layers,taking into account the loss of precision and downlink distribution delay.The model splitting problem is solved by solving the linear programming method,and the two problems are optimized jointly.Finally,the specific simulation results are given to evaluate the performance of the algorithm.The results show that the joint optimization algorithm proposed in this paper has greater gains in the distribution inference delay and energy consumption than the existing model segmentation methods.2.This paper researches the deep learning model selection and distribution joint optimization methods in cloud nodes and edge nodes collaboration scenarios.First,in wireless network,based on cloud nodes and edge nodes collaborative,the selection and distribution system model of deep learning model is established.Considering model selection,distribution and inference delay and the loss of model accuracy,the objective of optimization is to minimize the distribution and inference delay of deep learning models.This problem is a non-linear programming problem.The original optimization problem can be transformed into three optimization sub-problems:model selection,model splitting and frequency resource allocation,and each sub-problem can be solved.By selecting a model for each user through similarity between models,users who select the same model are distributed by multicast.Linear programming is used to solve model splitting and frequency resource allocation problems.Then,the three problems are optimized jointly.Then paper propose a joint optimization algorithm for cloud node and edge node collaboration in wireless network.Finally,the simulation results show that the proposed joint optimization algorithm greatly reduces the distribution and inference delay and energy consumption of the model compared with the existing algorithms.