Intelligent Edge: When Wireless Network Meets Machine Learning

在过去的几年里，互联网在现代社会中的核心作用带来了效率和灵活性方面的挑战，特别是在互联网边缘支持无线网络的移动设备激增导致使用加剧的情况下。与此同时，机器学习技术，如深度学习和强化学习，已经被应用到各种不同的应用中，以成功地解决它们的问题。在无线网络与机器学习相遇的时代，我们的愿景是，它们将相互促进，最终导致互联网的智能边缘。在设想的智能边缘，机器学习可以被用来进一步提高提供给移动设备的服务质量。目前的移动设备往往具有有限的计算资源和有限的电池容量。他们经常需要将计算密集型任务卸载到高性能服务器，以便以及时和节能的方式更好地完成任务。移动边缘计算(MEC)被提出通过将中等性能的边缘服务器放置在靠近移动设备(例如蜂窝基站)的位置来解决卸载问题。尽管MEC具有优势，但在广泛部署之前，还需要解决一系列技术挑战。机器学习已被用于提高MEC的性能。我们相信，机器学习将进一步应用于MEC，帮助为移动设备提供满意的服务。此外，智能边缘的移动设备可以为实现可用机器学习做出贡献。从技术上讲，许多机器学习技术的训练和推理阶段需要大量的计算资源。AlphaGo是第一个击败专业围棋选手的计算机程序，它的诞生是最近机器学习复兴的一个关键里程碑。然而，早期版本的AlphaGo需要使用176个GPU才能达到最佳性能。机器学习所需的计算资源量严重阻碍了其在现实生活中的应用。互联网边缘丰富的移动设备的存在确实为基于移动设备的分布式机器学习创造了前所未有的机会。在分布式机器学习中，每个移动设备都是一个自然的计算节点。分布式移动设备集合中的计算资源可以显著加快训练和推理过程。当无线网络与机器学习相遇时，互联网的边缘将变得智能，因为移动设备将以智能的方式服务，无线网络可以帮助实现有用的机器学习。在拟议的项目中，我们将探索各种方法，为智能边缘的出现铺平道路。

Over the past years, the central role of the Internet in modern society has created challenges of efficiency and flexibility, especially as usage intensifies due to proliferation of wireless-network-enabled mobile devices at the edge of the Internet. In the meanwhile, machine learning techniques, such as deep learning and reinforcement learning, have been applied to a variety of different applications to successfully tackle their problems. In an era when wireless network meets machine learning, our vision is that they will reinforce each other, ultimately leading to an intelligent edge in the Internet.  In the envisioned intelligent edge, machine learning could be utilized to further improve the quality of the services provided to mobile devices. Current mobile devices tend to have limited computation resources and constrained battery capacity. They often need to offload computationally intensive tasks to high-performance servers in order to complete the tasks in a timely and energy-efficient manner. Mobile Edge Computing (MEC) is proposed to solve the offloading problem by placing modest-performance edge servers at locations close to mobile devices (e.g. cellular base stations). Despite the advantage of MEC, a series of technical challenges need to be tackled before it can be widely deployed. Machine learning has been used to improve the performance of MEC. We believe that machine learning will be further utilized in MEC, helping provide satisfactory services for mobile devices. Furthermore, mobile devices in the intelligent edge could contribute to the realization of usable machine learning. Technically, the training and inference phases of many machine learning techniques require a large amount of computational resources. The creation of AlphaGo, the first computer program that beats a professional Go player, is a key milestone in the recent revival of machine learning. However, the early version of AlphaGo needs to use 176 GPUs to reach its best performance. The amount of computational resources required by machine learning seriously hinders its real-life application. The existence of abundant mobile devices at the edge of the Internet literally creates an unprecedented opportunity for mobile-device-based distributed machine learning. Each mobile device is a natural computing node in distributed machine learning. The computational resources in a collection of distributed mobile devices could significantly help speed up the training and inference process. When wireless network meets machine learning, the edge of the Internet will become "intelligent" because mobile devices would be served in an intelligent manner and wireless network could help realize usable machine learning. In the proposed project, we will explore varied approaches to pave the way for the emergence of the intelligent edge.