LARNet: Model-Based Neural Network for Implicit CSI Feedback

Deep learning (DL) methods based on auto-encoder architecture have shown considerable potential in reducing Channel State Information (CSI) feedback overhead. However, the existing DL-based CSI feedback schemes are entirely based on data-driven, which have not fully exploited the intrinsic structural information within CSI. We propose a novel model-based Neural Network architecture named LARNet that effectively exploits the characteristic of CSI to improve the feature extraction efficiency and reduce the complexity of DL-based CSI feedback. Simulation Results show that the proposed LARNet could achieve better recovery performance with lower computation and model complexity than existing state-of-the-art algorithms.