Unsupervised Online Learning in Deep Learning-Based Massive MIMO CSI Feedback

Deep learning-based channel state information (CSI) feedback has received copious amounts of solutions while related deployment problems, such as channel statistic offset between training and testing, are seldom discussed in literature. In previous research, a base station (BS) requires user equipment to transmit new channel data for retraining when channel statistics change, resulting in additional communication overhead. In this study, a novel CSI-free unsupervised online learning method is proposed without introducing additional communication overhead. When feedback performance drops due to the offset of channel statistics, the proposed method can improve feedback accuracy by using only the codewords received by the BS as training data; that is, no intact CSI is required to be collected for training. Simulation results indicate that the proposed method successfully brings feedback performance gain in various settings.