Asymmetric PoolCsiNet With Parameter-Free Encoder at UE for CSI Feedback
Zhichao Xie, Jindan Xu, Wei Xu, Xiaohu You, Derrick Wing Kwan Ng, Huahua Xiao · IEEE Transactions on Communications · 2024
Deep learning (DL) has been increasingly adopted for channel state information (CSI) feedback to harness the performance gains promised by massive multiple-input multiple-output (MIMO). Existing DL-based feedback schemes prioritize the accuracy of CSI reconstruction, which results in substantial memory and computational demands, especially when they are unacceptable for user equipment (UE) with limited resources. In this paper, we propose an asymmetric pooling-based network for more efficient CSI compression, named PoolCsiNet, to reduce the associated overhead of exploiting convolutional neural networks (CNN) for CSI compression at the UE. By leveraging the local information of clustered physical channel models, PoolCsiNet incorporates a low-complexity amplitude-pooling algorithm in its encoder at the UE without requiring any trainable parameters. A corresponding decoder structure is also developed to firstly acquire a coarse CSI and then a lightweight feature refiner is constructed to enhance the coarse CSI reconstruction. The parameter-free encoder and CNN-based refiner constitute a novel asymmetric CSI network architecture. Thanks to the parameter-free design of the encoder, memory demand at the UE is minimized, thereby eliminating the need for joint training and parameter updating. Furthermore, considering the sparsity of indoor wireless channels, a PoolCsiNet+, with a dilated-amplitude-pooling (DAP) module, is further proposed to elevate the CSI reconstruction accuracy of the PoolCsiNet. Thanks to a pooling design tailored for clustered channel models, lossy compression of pooling hardly sacrifices CSI features and can be exploited to eliminate information redundancy in CSI. Experiments demonstrate that both asymmetric PoolCsiNet and PoolCsiNet+ significantly improve the quality of CSI reconstruction up to 4 dB compared with existing DL-based methods, while maintaining a memory-free profile and achieving a sevenfold reduction in computational overhead at the UE.