Collaborative Communication and Computation for Hierarchical Learning at the Mobile Edge

点击翻译按钮获取中文摘要

In this research program, we study engineering theories and designs in an emerging system where Internet-of-Things (IoT) and other mobile devices collaborate with nearby computing servers, called Mobile/Multiaccess Edge Computing (MEC) hosts, to support machine learning (ML) services and applications. The following is a motivating example for such a system when it is applied to the augmented reality application. We may have multiple cameras aiming to jointly recognize the objects in their collective field of vision and send contextual augmentation labels to nearby mobile users, which can be displayed overlaying their view of the actual scenery. The cameras by themselves do not have sufficient computational power to recognize the objects, while it is also inefficient to send their videos to a remote cloud server for processing, due to the large data size and the limited communication over a long distance. In contrast, MEC hosts have unique features of high-bandwidth communication and low-latency computation, which enable them to provide timely assistance to the cameras. They can combine the cameras' information, perform the necessary ML computing jobs, and then send the augmentation labels to the mobile users. We name such a system collaborative wireless hierarchical ML over MEC. A crucial component in this emerging multi-level ML hierarchy is the communication pathway linking the large number of computing engines in the IoT devices, MEC hosts, and cloud servers. Furthermore, there is a high level of correlation between the communication and computation requirements, since the amount of transmitted data can vary dramatically depending on the portion of an ML job that is offloaded to the MEC hosts or cloud servers. Effective collaboration among IoT devices and MEC hosts in both communication and computation is crucial to overall system performance. Therefore, in the proposed research, we promote a holistic design of previously separate ML, wireless communication, and distributed computation algorithms. We aim to seamlessly integrate collaborative IoT devices, MEC hosts, and cloud servers, in a joint communication-computation paradigm, to support hierarchical ML services and applications. We will focus on three shorter-term objectives: (1) dynamic ML job division and task placement strategies to achieve optimal trade-off between data efficiency and computation efficiency, (2) integrated communication-computation resource management methods for multi-agent cooperation in hierarchical ML, and (3) algorithms and analytical tools for online ML and scheduling in MEC. The outcomes of the proposed research are expected to contribute substantially to motivate novel system designs in a wide range of services and applications, such as automated manufacturing, remote e-health, environment monitoring, and intelligent transportation. They are expected to have sustained market impact in strategic sectors of the Canadian economy.