A Hyper-Network-Aided Approach for ISTA-based CSI Feedback in Massive MIMO systems

Accurate channel state information (CSI) is critical for achieving high performance in massive multiple input multiple output (MIMO) systems. While existing deep learning (DL) based methods have achieved notable success for CSI feedback in the frequency division duplex (FDD) mode, they typically learn one set of neural network (NN) parameters for all CSI. However, only one set of parameters restricts the representation power of the NN, resulting in the limited performance. In addition, the channel estimation error is usually considered with discrete levels among the researches of CSI feedback, which limits the performance when channel estimation errors are successive. To address these issues, we propose a model-driven DL method with sample-relevant dynamic parameters using hyper-networks and unfolding. The proposed method can generate the parameters of the task network distinctly for each CSI by a hyper-network, which improves the representation power and recovery performance of the task network. Additionally, instead of assuming each CSI has the same level of channel estimation error, the proposed method automatically adjusts task network parameters to account for different levels of channel estimation error, resulting in significant performance gains. The numerical experiments demonstrate the superiority of the proposed method in terms of performance and robustness.