Canonical Correlation Analysis: Theory and Applications in Wireless Communications
Meeting ID: 585 266 1202
Smart and Internet-connected devices are ubiquitous in our daily lives, from handheld devices, wearable sensors, to smart vehicles and smart homes. Owing to the tremendous growth of the number of such devices, maintaining reliable data connectivity and providing high data rates to these devices constitute a major challenge for the current mobile and local wireless systems. Furthermore, because of this continuous increase, the currently available spectrum is expected to be overloaded, thereby resulting in considerable interference issues.
This dissertation aspires to leverage advances in machine learning and signal processing to tackle challenging problems in wireless communications. While data-driven approaches, notably deep neural networks and deep reinforcement learning, have arguably gained center-stage prominence owing to their success in many applications, there are in fact several problems that can greatly benefit from classical machine learning tools. The central goal of this dissertation is to theoretically and experimentally demonstrate how valuable insights from latent factor analysis techniques can lead to solutions that remarkably advance the state-of-the-art performance for fundamental problems in wireless communications. In particular, this thesis seeks to showcase the potential of canonical correlation analysis (CCA) in the context of modern wireless communications, through new theoretical contributions, algorithms and applications in 5G and beyond.
In recent work, we established an elegant connection between CCA and cell-edge (weak) user detection. In particular, we showed, via theoretical proofs and realistic experiments, that CCA can reliably recover user signals that are only 3dB above the noise floor, in the presence of strong interference, in unsupervised mode. Motivated by theoretical results and preliminary experimental findings, the proposed research plan comprises three CCA-based thrusts: 1) Unsupervised signal detection and accurate delay estimation in unknown multipath channels; 2) Efficient detection and estimation of weak targets in MIMO radar, and 3) Dynamic spectrum underlay.