Asynchronous Unsupervised Online Learning of Bayesian Deep Receivers

Deep learning-aided receivers have shown promising performance in challenging channel conditions, particularly when allowed to adapt online, which leads to notable computational burden. To reduce this burden, we propose an unsupervised asynchronous online learning framework based on Bayesian deep learning. Our approach leverages uncertainty estimates inherently produced by Bayesian neural networks to construct lightweight statistical tests that monitor the receiver, without requiring access to transmitted labels, that autonomously trigger retraining only when the receiver is no longer adequate. This results in a receiver that adapts efficiently while significantly reducing retraining frequency. Our numerical studies demonstrate that the proposed framework enables timely adaptation while achieving symbol error rates comparable to costly synchronous retraining at every coherence block.