On deep learning techniques for Noncoherent MIMO systems

In this paper, a further study on the deep learning (DL) techniques for both receiver scheme and Grassmannian constellation design in non-coherent (NC) multiple-input multiple-output (MIMO) systems is presented. We reveal the computational similarity between matrix multiplication in the optimal generalized likelihood ratio test (GLRT) detector and convolution operation of a two-dimensional (2D) convolitional layer, and propose convolutional neural network-GLRT (CNNGLRT) detector which implements GLRT detection on a CNN. Furthermore, a novel autoencoder (AE)-based approach, called AE-GLRT, to design Grassmannian constellations is proposed. In the simulation, we show that the proposed CNN-GLRT has the same symbol error rate (SER) performance as the conventional GLRT. Besides, the novel AE structure can construct Grassmannian constellations with good SER performance.