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

TL;DR

A novel combination of Bayesian deep learning with hybrid model-based data-driven architectures for wireless receiver design is presented, designed to yield calibrated modules, which in turn improves both accuracy and calibration of the overall receiver.

Abstract

Deep learning is envisioned to play a key role in the design of future wireless receivers. A popular approach to design learning-aided receivers combines deep neural networks (DNNs) with traditional model-based receiver algorithms, realizing hybrid model-based data-driven architectures. Such architectures typically include multiple modules, each carrying out a different functionality dictated by the model-based receiver workflow. Conventionally trained DNN-based modules are known to produce poorly calibrated, typically overconfident, decisions. Consequently, incorrect decisions may propagate through the architecture without any indication of their insufficient accuracy. 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 modular Bayesian deep learning, is designed to yield calibrated modules, which in turn improves both accuracy and calibration of the overall receiver. We specialize this approach for two fundamental tasks in multiple-input multiple-output (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.

Figures (16)

• Figure 1: The communication system studied in this paper.
• Figure 2: $(a)$ Iterative sic; and $(b)$ DeepSIC shlezinger2019deepSIC. In this paper, we introduce modular Bayesian learning, which calibrates the internal modules of the DeepSIC architecture with the double goal of improving end-to-end accuracy and of producing well-calibrated soft outputs for the benefit of downstream tasks.
• Figure 3: Static linear synthetic channel: ser as a function of the SNR for different training methods. At each block, training is done with $384$ pilots, and the ser is computed on $14,976$ information symbols from the QPSK or 8-PSK constellation. Results are averaged over $10$ blocks per point.
• Figure 4: Static non-linear synthetic channel: ser as a function of the SNR for different training methods. At each block, training is done with $384$ pilots, and the ser is computed on $14,976$ information symbols from the QPSK or 8-PSK constellation. Results are averaged over $10$ blocks per point.
• Figure 5: Time-varying COST channel: ser as a function of the SNR for different training methods. At each block, training is done with $384$ pilots, and the ser is computed on $14,976$ information symbols from the QPSK or 8-PSK constellation. Results are averaged over $10$ blocks per point.