Research On Deep Convolutional Neural Network Training And Inference Optimization For Edge Intelligence

The advent of the Internet of Everything era has prompted smart devices in edge clusters to provide intelligent services and generated new demands for low latency,low energy consumption,high accuracy and security.Deep Convolutional Neural Networks,as an important part of Artificial Intelligence technology,have significantly improved the performance of vision tasks such as image classification,object detection and semantic segmentation,providing strong support for edge intelligence applications.However,the success of existing Deep Convolutional Neural Networks relies on centralized computing,which leads to drawbacks such as high energy consumption and poor data security in edge scenarios.In addition,the large number of parameters and complex computations severely hinder the deployment of the models on low-power and low-storage edge devices.Therefore,how to build efficient Deep Convolutional Neural Network training and inference methods in edge intelligence scenarios to meet the objectivity requirements of edge intelligence applications is an urgent and challenging problem to be solved.In this paper,we address this problem by focusing on both model training and model inference for Deep Convolutional Neural Networks,optimizing the edge distributed model training process by reducing communication overhead and incentivizing model sharing,and accelerating the model inference process by reducing the number of model parameters and computational complexity,aiming at a lightweight network design with higher performance and lower latency.Specific research and contributions are summarized as follows:(1)Distributed training communication optimization strategy based on Federated Distillation:To address the problem of high communication overhead of existing edge distributed training methods,this paper deeply investigates the key factors affecting the communication overhead of edge distributed training,and constructs a Federated Distillation training framework for edge distributed training based on the analysis results,introducing control variables at the edge device and the central server respectively to solve drift issues caused by local model updates,thereby accelerating the convergence of the global model.In addition,the size of the model output is further compressed by applying a soft label quantization method based on truncation threshold and a delta coding scheme to reduce the amount of data transfer in a single round.Experimental results on public datasets show that our proposed optimization strategy can significantly reduce the communication cost required for training while meeting the established performance requirements of the model,thereby improving training efficiency.(2)Distributed training sharing optimization strategy based on parameter trading:To address the problem of low training efficiency and poor model quality due to the lack of incentives for decentralized training of deep models,this paper constructs a hierarchical master-slave chain hybrid distributed model sharing framework by combining blockchain technology to achieve scalable cross-layer and cross-domain interaction of deep models.In addition,to motivate devices to actively participate in model training,a model sharing incentive mechanism based on parameter trading is proposed.Based on this,an asynchronous Federated Learning algorithm based on bandwidth allocation optimization is applied to reduce aggregation waiting time,while a lightweight improved Byzantine consensus algorithm is proposed to reduce block consensus time,thereby improving the processing efficiency of data sharing transactions.Detailed theoretical analysis and experimental results show that the proposed strategy in this paper effectively improves the efficiency of edge-distributed training and ultimately achieves a balance between model accuracy,system scalability and training efficiency.(3)Quantization distillation-based lightweight method for neural networks:To address the problem that the amount of parameters of deep models is too large to be deployed on edge end devices,this paper optimizes the model training method for low-bit quantized neural networks by using online knowledge distillation techniques.We propose a two-stage online distillation training strategy to reduce the overhead of pre-training complex high-precision teacher models.Furthermore,the model performance is improved by randomly changing quantization bit number of model activations,to provide additional quantization knowledge to the low-bit quantized neural networks.Finally,attention mechanism is employed to generate dynamic weights for different group member models,avoiding the weakening of the representational capability of quantized models.Extensive experimental results show that the proposed method is able to improve model performance while significantly reducing the number of model parameters,achieving higher accuracy and lower parameter complexity for lightweight models.(4)Dynamic computation method of neural network based on sample adaptation:To address the problem of excessive computational complexity in highperformance deep model inference,this paper considers the characteristics of edge cluster devices and proposes a dynamic pruning method of neural network based on sample complexity and similarity,which effectively reduces redundant structures in the model,thus reducing the computational cost of model inference.Furthermore,we propose a two-stage dynamic adaptive model ensemble framework to provide accuracy compensation for lightweight models,thereby improving the expressiveness of the models while significantly reducing computational complexity.Validation results on several depth models and image classification datasets show that the proposed method effectively improves the inference speed of the model in practical application scenarios by reducing the computational complexity in model inference.