A Variational Bayes-Driven Deep Unfolding Network for Efficient MIMO Detection

The increasing number of users in multiple-input multiple-output (MIMO) systems poses challenges for developing efficient data detection methods. Conventional linear methods are less computationally complex but are inferior to optimal Bayesian methods, motivating the use of approximation approaches. Incorporating deep learning via deep unfolding networks has proven effective in bridging this gap. In this paper, we develop a variational Bayes (VB)-driven deep unfolding architecture, introducing two deep networks, namely matched-filter (MF)-VBNet and linear minimum mean-squared error (LMMSE)- VBNet, for MIMO detection. By unrolling VB detectors into layer-wise structures with only two learnable scalars per layer, our approach achieves state-of-the-art symbol error rate (SER) while maintaining low complexity. Comprehensive benchmarks against OAMP-Net2, MMNet, VB Inference-inspired Network (VBINet), and Improved-VBINet demonstrate, in small-to-medium MIMO systems featuring independent and identically distributed (i.i.d.) and correlated channels, the proposed MF-VBNet and LMMSE-VBNet outperform existing approaches while maintaining equal or lower complexity.