Spherical Normalization for Learned Compressive Feedback in Massive MIMO CSI Acquisition

Abstract

Massive multiple-input multiple-output (MIMO) systems have been highlighted as a key enabling technology for future communications networks. However, the efficacy of MIMO hinges on the availability of accurate channel state information (CSI). In frequency division duplex (FDD) massive MIMO systems, downlink CSI acquisition involves determining a sufficient amount of feedback compression to maintain high spectral efficiency. Prior works have highlighted the efficacy of deep convolutional neural networks (CNN) for learning an encoding scheme for downlink CSI. However, the prior works have approached the problem of channel estimation similarly to other domains in which CNNs have found success (e.g., computer vision). This paper proposes two techniques for tailoring CNN-based autoencoders to the problem of massive MIMO CSI recovery. The first technique is a power-based normalization of CSI (spherical normalization), which can optimize the input distribution and make the network more applicable to the commonly adopted accuracy metric. The second technique is an optimized CNN architecture for codeword efficiency. We propose networks, called SphNet and DualNet-Sph, which combine these techniques, and we demonstrate that these new networks outperform previously proposed deep learning networks for CSI feedback.