An Assessment of Deep Learning Versus Massively Parallel, Non-Linear Methods for Highly-Efficient MIMO Detection

Multiple-user, multiple-input, multiple-output (MU-MIMO) systems supporting a large number of concurrent streams have the potential to substantially improve the connectivity and throughput of future wireless communication systems. Towards this goal, deep learning (DL)-based techniques have recently been proposed for MIMO signal detection. Good performance results have been reported when compared to conventional detection methods, but it is unclear how they measure against state-of-the-art detection techniques. In this work, for the first time, we perform a critical evaluation of DetNet, MMNet, GEPNet, and RE-MIMO, four prominent model-based DL techniques based on different working principles, and assess their reliability, complexity, and robustness against the practical Massively Parallel Non-Linear processing (MPNL) detection approach. The results show that the model-based DL approaches offer promising results but have difficulty adapting to channel models that differ from those on which they were trained. They also exhibit lower reliability and higher complexity than MPNL, even without considering the training stage. We find that, at present, the human-designed MPNL outperforms the DL-based detection methods in virtually all the metrics. Nevertheless, DL-based solutions are rapidly advancing, and further research intended to address their current shortcomings may one day offer advantages over human-designed detection methods.