Application-Aware Wireless Communication for 6G Intelligent Edge

Professor Huang, Kaibin   (Principal Investigator (PI))

60

2022-01-01

5155380

Application-Aware Wireless Communication for 6G Intelligent Edge

5G and Beyond, Interference Management, Multiple Access, Radio Resource Management, Scheduling

Communication

Engineering (E)

RFS2122-7S04

RGC Research Fellow Scheme

2021

On-going

1. The first research thrust will concern the design of a new technology for application-aware multiple access, called over-the-air computation (AirComp), for efficient distributed function computation. We will design AirComp techniques that address several fundamental issues including adaptive coding, radio resource management, synchronization, and channel feedback. Their performance will be also analyzed. 2. The second research thrust will focus on designing a new class of scheduling schemes targeting edge learning, which feature data-diversity awareness to accelerate learning. Specifically, diversity-aware scheduling schemes will be proposed for two edge application paradigms: edge cloud learning and distributed edge learning. The designs will essentially involve the derivation of novel scheduling metrics integrating multiuser data diversity and channel states under the criterion of accelerating model convergence. The resultant learning performance will be mathematically analyzed. The extension to parallel-channel systems (e.g., OFDMA or multiuser MIMO) will be also investigated. 3. The third research thrust will focus on the design of a new quantization framework for compressing stochastic gradients to improve the communication efficiency of a federated edge learning system. Given awareness of gradient geometric properties, the framework will feature a decomposed architecture that reduces the complex task of high-dimensional gradient quantization to low-dimensional (practical) quantization using multiple component quantizers (e.g., Grassmannian quantizer). They will be jointly optimized and their outputs intelligently assembled to approach the performance of a high-dimensional (ideal but impractical) vector quantizer. The performance of the proposed framework will be analyzed by quantifying the fundamental learning-distortion tradeoff as well as the optimality gap w.r.t. the ideal quantizer. The implementation issue of heterogeneous device hardware will be also investigated.