[论文解读] Artificial Intelligence at the Edge

Elisa Bertino, Sujata Banerjee|arXiv (Cornell University)|Dec 10, 2020

IoT and Edge/Fog Computing被引用 8

一句话总结

本白皮书提出在网络边缘集成人工智能（AI），以在物联网（IoT）和边缘计算环境中实现低延迟、实时的应用。通过利用5G/6G、硬件加速器以及隐私保护技术的进展，边缘AI可提升医疗、灾难恢复和沉浸式技术等关键应用的响应速度与效率。

ABSTRACT

The Internet of Things (IoT) and edge computing applications aim to support a variety of societal needs, including the global pandemic situation that the entire world is currently experiencing and responses to natural disasters. The need for real-time interactive applications such as immersive video conferencing, augmented/virtual reality, and autonomous vehicles, in education, healthcare, disaster recovery and other domains, has never been higher. At the same time, there have been recent technological breakthroughs in highly relevant fields such as artificial intelligence (AI)/machine learning (ML), advanced communication systems (5G and beyond), privacy-preserving computations, and hardware accelerators. 5G mobile communication networks increase communication capacity, reduce transmission latency and error, and save energy -- capabilities that are essential for new applications. The envisioned future 6G technology will integrate many more technologies, including for example visible light communication, to support groundbreaking applications, such as holographic communications and high precision manufacturing. Many of these applications require computations and analytics close to application end-points: that is, at the edge of the network, rather than in a centralized cloud. AI techniques applied at the edge have tremendous potential both to power new applications and to need more efficient operation of edge infrastructure. However, it is critical to understand where to deploy AI systems within complex ecosystems consisting of advanced applications and the specific real-time requirements towards AI systems.

研究动机与目标

• 应对自动驾驶、增强现实和远程医疗等实时、低延迟应用日益增长的需求。
• 识别集中式云计算在满足严格响应时间与数据隐私要求方面的局限性。
• 研究边缘AI在提升系统效率及推动关键领域新应用方面的作用。
• 突出边缘计算、人工智能/机器学习（AI/ML）与下一代通信技术（5G及更先进技术）之间的协同效应。

提出的方法

• 提出将AI模型直接部署在边缘设备上，以相比基于云的推理降低延迟和带宽消耗。
• 强调使用硬件加速器（如张量处理单元TPU、神经网络处理单元NPU）以优化设备端AI推理性能。
• 在边缘集成隐私保护计算技术，如联邦学习和安全多方计算。
• 利用5G及未来的6G网络，实现边缘节点与终端用户之间的高容量、低延迟通信。
• 分析边缘AI的系统架构，重点关注异构设备与网络间分布式的智能。
• 考虑涉及延迟、能效和数据本地化的端到端系统设计权衡。

实验结果

研究问题

• RQ1 边缘AI如何提升自动驾驶和沉浸式视频会议等实时应用的响应速度？
• RQ2 推动AI在网路边缘有效部署的关键技术使能因素是什么？
• RQ3 5G和6G网络如何支持边缘AI工作负载的可扩展性与可靠性？
• RQ4 隐私保护型AI技术在保护边缘敏感数据方面发挥什么作用？
• RQ5 哪些架构模式能优化边缘AI系统中性能、能效与数据本地化之间的平衡？

主要发现

• 与基于云的处理相比，边缘AI显著降低了延迟，使关键任务应用实现真正的实时响应。
• 5G网络提供了实现可扩展边缘AI部署所必需的低延迟、高带宽连接。
• 边缘硬件加速器提升了AI模型的推理速度与能效。
• 联邦学习等隐私保护技术可实现在设备端进行AI训练，而无需传输原始数据。
• 边缘AI与6G网络的集成支持全息通信和高精度制造等新兴应用。
• 边缘AI增强了系统弹性，降低了对集中式云基础设施的依赖，尤其在灾难恢复和远程医疗场景中。