Scalable Hybrid Architecture for Wireless Collaborative Federated Learning (SHAFT)

与前几代移动网络不同，超5G （B5G）网络旨在支持边缘智能，即为最终用户附近提供通信和计算能力。无线边缘智能对于B5G的关键用例尤其重要，包括智慧城市、自动驾驶、无线医疗、虚拟现实（VR）和增强现实（AR）游戏，在这些用例中，移动网络预计将配备智能功能，用于预测和塑造个人体验。联邦学习（FL）是无线边缘智能的关键支持技术，它以分散的方式执行模型训练，并将数据保存在生成数据的地方。然而，由于具有异构资源和大量设备的复杂无线环境，直接将FL从计算机网络应用到无线系统可能会导致频谱性能和实现效率的下降。该项目的目的是通过有效地利用移动通信环境的物理层动态，并利用复杂的服务感知和资源感知的协作边缘学习，为无线FL开发一种新颖的可扩展混合架构。该项目的新颖之处在于这种新型边缘学习架构的发展，其中学习架构的基本限制以先进的数学工具为特征，例如图论和随机学习。此外，通过应用复杂的工具（如压缩传感和机器学习），在实际约束条件下量化具有挑战性的设计权衡的算法框架。

Unlike previous generations of mobile networks, the beyond 5G (B5G) network is envisioned to support edge intelligence, which is to provide both communication and computing capabilities to the proximity of end users. Wireless edge intelligence is particularly important to those crucial use cases of B5G, including smart cities, autonomous driving, wireless healthcare, virtual reality (VR) and augmented reality (AR) gaming, where mobile networks are expected to be equipped with intelligent capabilities for prediction and shaping experiences to individuals. Federated learning (FL) is a key enabling technology for wireless edge intelligence, by performing the model training in a decentralized manner and keeping the data where it is generated. However, a straightforward adaption of FL from computer networks to wireless systems can suffer performance degradation in spectral and implementation efficiency, because of the complex wireless environment with heterogeneous resources and a massive number of devices. The aim of this project is to develop a novel scalable hybrid architecture for wireless FL by efficiently utilising the physical layer dynamics of the mobile communication environments and exploiting sophisticated service-aware and resource-aware collaborative edge learning. The novelty of the project is the development of this novel edge learning architecture, where the fundamental limits of the learning architecture is characterised by advanced mathematical tools, such as graph theory and stochastic learning. In addition, an algorithmic framework for quantifying challenging design trade-offs in the presence of practical constraints by applying sophisticated tools such as compressed sensing and machine learning.