Extrinsic Graph Neural Network - Aided Expectation Propagation for Turbo-MIMO Receiver

Deep neural networks (NNs) promise excellent performance and high efficiency in constructing multiple-input multiple-output (MIMO) receivers. Recently, graph NNs (GNNs) have been applied to enhance expectation propagation (EP) for MIMO detection and to overcome the inaccuracy of Gaussian approximation caused by multi-user interference. However, GNN-aided EP detector fails to generate extrinsic information required by Turbo-MIMO receivers. We develop a customized training scheme in this paper as a remedy to enable extrinsic output from the GNN-aided EP detector and further enhance the interaction with the channel decoder by adaptively scaling the soft information feedback. Simulation results show that the proposed Turbo-MIMO receiver significantly outperforms the EP-based receiver and achieves comparable performance to the sphere decoding-based receiver with shorter running time.