HRNet: Deep Massive MIMO CSI Feedback with Hypernetworks

Multiple-input multiple-output (MIMO) communications are the core technology for the next generation of telecommunications. However, with the increasing complexity of Channel State Information (CSI), CSI feedback in massive MIMO FDD systems becomes a bottleneck problem. Deep learning (DL) methods have been used to improve the reconstruction efficiency of CSI feedback and demonstrated an improvement in reconstruction performance in CSI feedback. However, most of the models sacrifice a significant computational cost in training due to the large size of the autoencoder network. In this paper in order to reduce the training complexity, we propose a model structure based on CRNet, called HRNet, using the Hypernetwork approach to generate the parameters required for the decoder. By comparing the experimental results, we found that the HRNet model can reduce the number of training parameters by an average of 39.98% compared to the CRNet model while maintaining a comparable level in terms of normalized mean square error (NMSE) performance. The use of hypernetwork enables HRNet to generate parameters based on different features of the input data, thus providing better adaptability and flexibility.