Deep-Unfolding-Based Bit-Level CSI Feedback in Massive MIMO Systems

Abstract

In frequency division duplex mode, downlink channel state information (CSI) should be sent to the base station via a feedback link to obtain the benefits of the massive multiple-input multiple-output technology, leading to a substantial overhead. The existing deep-unfolding-based works combine the advantages of high performance and interpretability. In practical systems, the compressed CSI must be quantized at the user equipment before feedback. In this letter, a deep-unfolding-based bit-level CSI feedback is proposed to reduce the effects of quantization errors by using the shortcut and approximate quantization schemes. The shortcut scheme provides the unquantized encoder output directly to the decoder through an additional network branch, which can effectively help the decoder training. The approximate quantization scheme allows the gradient to be propagated correctly by approximating the real quantization operation. The two schemes are combined as a training strategy to reduce quantization errors further. Results show that the combined scheme can substantially reduce the influence of the quantization layer and improve the reconstruction performance while having some generality.