Enhanced Meta-Transfer Learning Assisted CSI Feedback in Massive MIMO Systems

To enhance the generalization of neural networks based channel state information (CSI) feedback approaches in massive multiple-input multiple-output (MIMO) systems, we propose a novel CSI feedback approach based on meta-transfer learning to overcome the impact of dynamic channels. Specifically, neural-level scaling and shifting operations are introduced to enhance the migration performance of shared parameters. Meanwhile, the meta-learning technique is leveraged to train unique parameters for optimal initialization, enabling rapid adaptation to new environments. In addition, concerning user data privacy, we introduce federated learning to fine-tune our trained model. Simulation results indicate that the trained model in our proposed approach can quickly adapt to new environments with a few samples. Besides, compared with baseline solutions, our approach is effective in terms of CSI reconstruction and time-saving performance.