作　　者： ;

机构地区： 中山大学电子与通信工程学院

出　　处： 《无线电通信技术》 2020年第1期6-25,共20页

摘　　要： 近年来,随着移动通信和人工智能技术的迅猛发展,大量智能终端已经联网并催生出海量数据。为了高效利用网络中的通信和计算资源并进一步释放人工智能的潜力,将传统基于专用数据中心的人工智能下沉到靠近用户终端的网络边缘已成为一种技术趋势。面向这种技术发展趋势,边缘学习被认为是一种具有广泛应用前景的人工智能实施方案。但是,目前对边缘学习的研究和应用仍处于起步阶段。为了促进技术发展,对边缘学习的关键技术、典型应用以及面临的机遇和挑战进行全面分析。首先,回顾边缘学习的发展背景,并分析其在传输时延、安全与隐私、扩展性和通信开销等方面相对于传统云学习的优势;其次,详细讨论实现边缘学习的3项关键技术:①分布式模型训练,包括聚合频率、梯度压缩、点对点通信和区块链技术;②面向边缘学习的高效无线通信技术,包括空中计算、通信资源分配和信号编码;③边缘学习卸载技术,包括计算和模型卸载技术。然后,分别以一种高可靠低时延车联网通信和一种基于计算与通信联合设计的智能图像分类系统为例,阐述上述关键技术在实际系统中的重要作用。最后,从通信与计算的联合优化、数据安全与隐私保护以及系统的开发与部署等3个方面讨论边缘学习面临的发展机遇与挑战。通过对最新研究现状的宏观分析,该综述将为边缘学习的进一步理论研究、技术创新和系统开发提供坚实的基础。 With the rapid development of mobile communication and artificial intelligence(AI)in recent years,a large number of intelligent terminals have been connected to the Internet,yielding a tremendous amount of data at the edge of the network.To efficiently utilize the communication and computation resources in the network and further release the potential of AI,it is necessary to push traditional data centre-based AI to the network edge proximal to terminals.To this end,edge learning has been widely accepted as a promising AI solution.However,both theory and application of edge learning are in their infancy.To speed up the research and development of edge learning,this survey analyzes its key technologies,typical applications,opportunities and challenges.Firstly,edge learning is compared with traditional cloud learning in terms of latency,security and privacy,scalability,and communication overhead.Then,three key technologies for its implementation are investigated:①distributed model training,including aggregation frequency,gradient compression,peer to peer communications,and blockchain technology;②efficient wireless communication techniques,including over-the-air computation,communication resource allocation,and signal encoding;③computation and model offloading technologies.Next,an ultra reliable low latency communication in vehicle-to-vehicle network and an edge learning based intelligent image classification system are selected to demonstrate the effectiveness of federated learning and joint design of computations and communications in real-world applications,respectively.Finally,some potential opportunities and challenges of edge learning are discussed,such as joint optimization of communication and computation,data security and privacy protection,and system development and deployment.By making a critical overview of the state-of-the-art of edge learning,this survey sheds new light on its forthcoming research,technological innovation and system development.

关 键 词： 边缘学习 联邦学习 人工智能 深度学习 卸载技术

领　　域： []