Uncertainty-Aware and Reliable Neural MIMO Receivers via Modular Bayesian Deep Learning

Deep learning is envisioned to play a key role in the design of future wireless receivers, with proposed architectures ranging from standard end-to-end deep neural networks (DNNs) to hybrid model-based data-driven designs that augment classic algorithmic modules with machine learning models. A key challenge associated with DNNs, both end-to-end and modular architectures, stems from the fact that conventional training leads to models that produce poorly calibrated, typically overconfident, decisions. This becomes more substantial when training from scarce data, e.g., limited pilots. To address this problem, we present a novel combination of Bayesian deep learning with hybrid model-based data-driven architectures for wireless receiver design. The proposed methodology, referred to as <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">modular Bayesian deep learning</i>, is designed to yield calibrated modules, which in turn improves both accuracy and calibration of the overall receiver. We propose a dedicated training objective, in which each deep learning-based module in the receiver chain is calibrated using Bayesian learning, such that overconfident decisions are mitigated throughout the receiver chain, leading to more reliable processing. We specialize this approach for two fundamental tasks in (MIMO) receivers – equalization and decoding. In the presence of scarce data, the ability of modular Bayesian deep learning to produce reliable uncertainty measures is consistently shown to directly translate into improved performance of the overall MIMO receiver chain.