Multiple access, resource management, and load balancing for edge intelligence systems

Abstract

Given the advancements of mobile communication systems and Internet-of-Things (IoT), billions of mobile devices will be deployed in wireless networks. The enormous amount of data generated by the massive number of devices, together with their computation capacities, can be used to enable low-latency artificial intelligence (AI) applications at the network edge, giving rise to a new research area, called edge intelligence. The new design goals give rise to new challenges. On one hand, fast wireless data aggregation (WDA) from a large number of devices should be realized to execute many real-time tasks, such as auto-driving, public safety surveillance, and traffic control. As a result, the conventional orthogonal multiple access techniques is insufficient due to the limited wireless resources. On the other hand, distributed learning algorithms are deployed at the network edge for training AI models by using distributed data and computation resources therein, called edge learning. This inter-disciplinary area integrates computer science and wireless communications. Consequently, the traditional wireless techniques, targeting throughput maximization or energy minimization, are no longer optimal, as the design goal shifts towards enhancing the learning performance, i.e., training latency and testing accuracy. In this dissertation, the above two issues are addressed.

First, a new multiple access technique is proposed for massive multi-input-multi-output (MIMO) systems with clustered IoT sensors, called MIMO AirComp, which directly receives a functional value of the distributed data transmitted from multiple devices instead of decoding each data stream. As a result, it can support the simultaneous transmission of a large number of devices by using the waveform superposition property of a wireless channel. Specifically, in the proposed MIMO AirComp framework, a vector-valued function can be calculated at the receiver with low latency and low complexity by designing a two-tier receive beamforming matrix.

Next, the novel framework of partitioned edge learning (PARTEL) is proposed. It supports the efficient training of large-scale models e.g., large-scale logistic regression or deep neural networks (DNNs) with millions of parameters, by coordinating the edge server and many resource-constrained devices. In such a framework, the global model is partitioned into several parametric blocks in each training iteration. Each block is assigned to one worker (group) for updating. Then, updates of all blocks are uploaded to the server for updating the global model, completing one iteration. The PARTEL framework is first implemented at narrowband wireless systems. Optimal scheme of joint parameter and bandwidth allocation is proposed under the target of minimum training latency. Then, it is further implemented at broadband wireless systems using orthogonal frequency division multiplexing (OFDM), which is more challenging as it involves the joint allocation of subcarriers, transmit power, and computation power. To solve this optimization problem, a practical solution approach is proposed. In both systems, experimental results using real datasets show that the proposed schemes can substantially reduce the training latency.