面向无线的分布式学习的鲁棒性研究与模型优化

面对大数据的冲击，集中式的模式已经无法保证处理速度。因此，无线网络节点不再只扮演转发的角色，而要同时具备计算、存储和智能化等数据处理功能。传统的集中式架构正在向分布式架构发展。针对无线网络中这种分布式的处理机制，本项目将对不同网络架构下系统的鲁棒性、性能、存储和计算资源等的损耗开展研究。1. 边缘计算：计算和存储能力由网络中心节点“下沉到”边缘节点。基于该架构，本项目进行以下两方面的研究：a)针对边缘节点计算能力的差异性，研究基于编码的效率提升机制，及其对节点在不同信道模式下的协作传输技术的影响；b)基于Map-Reduce的边边协同的分布式模式：研究慢节点的鲁棒性对抗技术，和在此基础上通信与计算资源等损耗的平衡问题。2. 面向无线的联邦学习：用户终端进行分布式训练并在中央服务器协调下更新模型。本项目将研究该框架下的通信开销问题和模型优化，尤其是当用户的本地数据集服从非独立同分布的情况。

Facing the impact of big data, the centralized model can no longer guarantee the processing speed. Therefore, wireless network nodes no longer only play the role of forwarding, but also have data processing functions such as computing, storage, and intelligence. The traditional centralized architecture is evolving to a distributed architecture. Aiming at this distributed processing mechanism in wireless networks, this project will conduct research on the robustness, performance, storage and computing resource loss of systems under different network architectures. 1. Edge computing: computing and storage capabilities are "sinked" from the network center node to the edge node. Based on this architecture, this project conducts the following two researches: a) Focusing on the differences of the computing capabilities of edge nodes, we will study the improvement mechanism of coding-based efficiency, and its impact on the cooperative transmission technology of nodes under different channel modes; b) MapReduce-based edge-edge collaborative distributed mode: we will study the robust countermeasure technology of stragglers, and on this basis, the problem of balancing the loss of communication and computing resources. 2. Wireless-oriented federated learning: the user terminals perform distributed training and update the model under the coordination of the central server. This project will study the communication overhead problem and model optimization under this framework, especially when the users’ local data sets are not independent and identically distributed.