Low-power embedded technology offers a roadmap for enabling deep learning (DL) applications in mobile scenarios, like future autonomous vehicles. However, the lack of breakthrough power efficiency improvements can jeopardize the realization of truly “cognitive” mobile systems that meet real-time deadlines. This work focuses on the new generation cloud-backed mobile cognition system architecture where vehicles execute DL applications with dynamic assistance from the cloud. We unveil opportunities for power-efficient inferencing at the edge through a technique that balances inference execution across the cloud and the vehicle. This level of adaptation results in significant power efficiency improvements compared to all or nothing solutions, where inferences execute either completely on the vehicle or completely in the cloud. In addition, the cloud can have an active role in helping the vehicle to improve its DL capabilities by communicating relevant model updates, with up to 63% bandwidth savings and negligible accuracy degradation when the proposed relevance-driven federated learning technique is used. Finally, the cloud-backed mobile cognition concept is extended to the case of “flying clouds” where vehicles connect to flying drones that provide services while in flight. Although their capabilities are not on par with the stationary cloud, the flying cloud reduces services’ latency significantly and enables critical functionalities.

低功耗嵌入式技术为在未来的自动驾驶汽车等移动场景中实现深度学习（DL）应用提供了路线图。但是，缺乏突破性的能效改进可能会危及满足实时期限的真正“认知”移动系统的实现。这项工作着重于新一代基于云的移动认知系统架构，其中车辆在云的动态协助下执行DL应用程序。我们通过平衡在云和车辆之间的推理执行能力的技术，在边缘揭示了高能效推理的机会。相较于全有或全无，这种适应水平可显着提高电源效率解决方案，其中推理可以完全在车辆上执行，也可以完全在云中执行。此外，当使用建议的相关驱动的联合学习技术时，云可以通过传达相关的模型更新来帮助车辆提高DL能力，从而起到积极的作用，节省多达63％的带宽，并且精度下降可忽略不计。最后，基于云的移动认知概念已扩展到“飞行云”的情况，在这种情况下，车辆连接到在飞行中提供服务的飞行无人机。尽管它们的功能无法与固定云相提并论，但动态云可显着减少服务的延迟并启用关键功能。