Deep Unfolding-Based Channel Estimation and Soft Symbol Decoding with Low-Resolution ADCs

This work develops a deep unfolding-based variational Bayes (VB) technique called learned-Vb for joint channel estimation and data decoding using quantized inputs in a massive multi-input multi-output (MIMO) uplink system. The unfolded neural network learns the channel statistics using self-training. The training process does not explicitly optimize to reduce the difference between the actual and estimated channel statistics. Instead, a linear combination of cross-entropy loss between the true and the soft data symbols and the mean squared error loss between the true and the estimated channel is minimized. The performance of the proposed algorithm is studied and compared against a channel statistics-aware VB algorithm under both i.i.d. and correlated channel models. It is found that the deep unfolding model overcomes the dependency on the knowledge of the channel statistics, thus making the approach suitable for unknown or mismatched channel statistics. Specifically, the unfolding-based method performs on par with the vanilla VB inference (VBI) algorithm when the latter has exact knowledge of the channel statistics while offering much better performance when the channel statistics are mismatched.