Wireless communication has become ubiquitous to modern technologies meeting their ever increasing latency, connectivity and data rate requirements through evolving hardware development, novel technologies, and algorithm design. However, as traditional algorithm approach saturation in their modelling, inference and robustness capabilities, data driven methods for supporting the above novel designs has received significant research focus. The thesis presents data-driven Bayesian signal processing methods for selected physical layer applications in the upcoming beyond fifth generation and sixth generation wireless communication systems.

Specifically, we first investigate traditional Bayesian machine learning, scaling them through online/streaming machine learning using methods from approx...[ Read more ]

Wireless communication has become ubiquitous to modern technologies meeting their ever increasing latency, connectivity and data rate requirements through evolving hardware development, novel technologies, and algorithm design. However, as traditional algorithm approach saturation in their modelling, inference and robustness capabilities, data driven methods for supporting the above novel designs has received significant research focus. The thesis presents data-driven Bayesian signal processing methods for selected physical layer applications in the upcoming beyond fifth generation and sixth generation wireless communication systems.

Specifically, we first investigate traditional Bayesian machine learning, scaling them through online/streaming machine learning using methods from approximate posterior inference for interference cancellation in over-the-air aggregation systems. We then examine potential pitfalls in applying the above traditional approaches and their mitigation through tools from deep learning.

We first investigate meeting latency requirements in channel estimation and user activity detection in multi-user multiple input and multiple output systems where models and algorithms demonstrate good performance. Subsequently, we investigate the decentralized resource management problem in cell-free MIMO systems where system constraints and timescale separation of system variables make algorithm design challenging and address them using a novel hierarchical graph neural network which can accommodate the timescale separation via an unsupervised training scheme. We then focus the model deficit problem observed in channel estimation for unmanned aerial vehicles where accommodating various spatio-temporal statistics via traditional Bayesian approaches is challenging and introduce recovery using generative models, developing a flexible novel inference algorithm and provide preliminary theoretical guarantees. Finally, we present an interpretable end-to-end Bayesian optimization framework for precoder design and channel estimation, enabling simultaneous utilization of labelled data and model aided advancements.