Toward Better Low-Rate Deep Learning-Based CSI Feedback: A Test Channel-Based Approach

Deep learning (DL)-based channel state information (CSI) feedback provides satisfactory reconstruction accuracy of downlink CSI for the base station in massive multiple-input multiple-output (MIMO) systems. Although the introduction of codeword quantization improves the efficiency and feasibility of DL-based CSI feedback networks, the gradient problem caused by quantizers in the training stage compromises the performance of neural networks. In this paper, by considering the test channel as an equivalent of ideal rate-distortion quantization in a mutual information sense, we propose a test channel-based quantization module (TCQM) for DL-based CSI feedback networks which mitigates the gradient problem in the end-to-end training of CSI feedback networks. Moreover, the training of the CSI feedback network with TCQM is not dependent on the design of practical quantizer in the inference stage, which reduces the complexity of the training and design constraints of the CSI feedback system. Finally, for the setting of fixed feedback overhead, based on the idea of TCQM, we propose an adaptive training strategy for CSI feedback networks to evaluate the proper combination of codeword length and quantization rate of codeword elements to achieve the optimal reconstruction accuracy. Experiment results show that the proposed schemes outperform existing codeword quantization schemes in the literature.