Deep Complex-Valued Neural Networks for Massive MIMO Signal Detection

In the fifth generation (5G) and future mobile networks, the design of efficient detectors for massive multiple-input multiple-output (MIMO) is essential. The main challenge in designing detectors is the trade-off between performance and computational complexity; that is, efficient detectors incur higher computational costs, while computationally cheaper detectors have lower efficiency. Recently, many deep learning-based detectors have been proposed in the literature to fill in such gaps. However, most of the existing MIMO detectors work only with real-valued parameters. First, they transform the complex received MIMO signal into an equivalent real-valued parameter by concatenating the real and imaginary parts and then train a network based on the real-valued data. Such an approach has several disadvantages. On one hand, the number of trainable parameters will be doubled; on the other hand, the phase information, which is important in the communication signals, might be lost or distorted. In this work, we aim to investigate the application of complex-valued neural networks for MIMO signal detection based on Wirtinger Calculus. To do so, we propose a simple feedforward architecture that directly works with the complex-valued QPSK and 16-QAM modulation signals. Our method is simple and computationally cheaper. Simulation results show that the proposed approach can improve the performance of the existing detectors while providing a lower computational cost.