ContBayes: Online training of Bayesian Neural Networks

This repository contains the code for the paper "Rapid Online Bayesian Learning for Deep Receivers" submitted for the ICASSP 2025 conference.

Description

We aim to enhance the adaptability, efficiency, and robustness of deep receivers in MIMO uplink systems.

Key Features

• A Bayesian Neural Network class BayesNN, based on TyXe VariationalBNN, allows interaction with the weights of the model and Jacobian computations.
• BayesianDeepSIC - A Bayesian version of the DeepSIC model defined in [1].
• Tracker classes EKF (for general BayesNN models) and DeepsicEKF (for BayesianDeepSIC model) for online training using the extended Kalman Filter.

References

[1] N. Shlezinger, R. Fu and Y. C. Eldar, "DeepSIC: Deep Soft Interference Cancellation for Multiuser MIMO Detection," in IEEE Transactions on Wireless Communications, vol. 20, no. 2, pp. 1349-1362, Feb. 2021.