Enhancing Robustness Against Noisy CSI Data with a Single Uncertainty-Informed Model

Neural networks (NNs) excel in wireless communications tasks, yet their adaptability to parameter changes like signal-to-noise ratio (SNR) and dependence on perfect channel knowledge are ongoing areas of research. In this study, we demonstrate that learning uncertainty is useful for developing models capable of generalizing across multiple SNR levels, without relying on noise-free channels or prior SNR information. The proposed approach can be conveniently integrated with any NN suitable for regression tasks. By focusing on channel prediction we introduce a method for uncertainty quantification, and we adjust the loss function accordingly. Simulation results highlight the advantages of our proposed techniques over traditional loss functions.