Application Algorithm Research And Simulation System Realization Based On Distributed Neural Network

With the enhancement of data acquisition and processing capabilities,the depth of neural networks is increasing,and the data processing performance is also gradually improving.However,limited by the computing power of a single host,the existing deep neural networks have shortcomings such as high equipment performance requirements and long training time.Combining neural network and distributed technology is a good solution to meet the growing demand.By expanding the number of computing device nodes,the performance requirements of a single device can be reduced while ensuring the overall computing performance.The basic idea of distributed neural network is to divide the whole network into sub-networks and map them on a distributed computing hierarchy.Each sub-network has reasoning and exit nodes.Through the joint training of each sub-network,equipment resources and communication costs can be significantly reduced,and the overall accuracy can be ensured while improving system fault tolerance and data privacy protection capabilities.The research topic of this thesis is a distributed neural network and its application algorithm and simulation system implementation.This thesis can be divided into three parts: Firstly,a distributed neural network algorithm based on domain adaptation without source data is proposed,which only uses the segmented source model during training.Combined with auxiliary tasks such as self-supervision and pseudo-labeling,the algorithm can perform distributed training on unlabeled target data online without accessing the source training dataset,and achieves well classification performance.Secondly,a distributed control algorithm based on reinforcement learning is proposed,which combines centralized and decentralized modes,and the agent unit can autonomously decide to follow the decision of the centralized meta-agent,or adapts itself in a decentralized manner for a changing environment.While maintaining the good performance of the centralized approach,it also has the advantages of low-latency in decentralized approach.Finally,in order to explore the performance of the above algorithm in a more realistic network scenario,this thesis designs and implements a network simulation system for deploying the above algorithm.This system can quickly build a network topology and deploy the neural network algorithm in it,and realize the functions such as node communication and multi-terminal node cooperative distributed training,and provide support for studying distributed neural network training in real scenarios.The results of experiments show that the distributed neural network algorithm based on domain adaptation without source data proposed in this thesis can still achieve high accuracy in the face of inconsistent online and offline data distribution.In addition,this method avoids the privacy problem caused by directly using the raw data obtained by the sensor,and the network communication cost is also smaller.While the experimental results of distributed control algorithms based on reinforcement learning show that compared with the properly trained centralized model and decentralized model,the algorithm in this thesis can achieve better performance and adaptability in real network environments.By applying the domain adaptation algorithm without source data to the simulation system,it can be seen that the algorithm proposed in this thesis still has good performance in experiments that are closer to the real scene.