Transformer-Inspired Deep Adaptive CSI Sensing With Learnable Quantizer

As a key enabler for 5G and next-generation communication, multi-antenna techniques usually require accurate channel state information (CSI) to release their potential. However, the corresponding CSI sensing is challenging, especially in frequency division duplex systems under the codebook-based feedback mechanism. While this is conventionally accomplished by solving a phase retrieval (PR) problem, existing methods based on PR usually utilize heuristic sensing precoders that can hardly capture the temporal dependencies among feedbacks effectively and sense adaptively. Inspired by Transformer, in this work, we propose a CSI sensing deep neural network, leveraging multi-head attention to learn the environment knowledge from the feedbacks and adjust the precoders adaptively. Moreover, we further consider the impact of feedback quantization, develop a learnable computation-efficient quantizer, and embed it into our proposed adaptive sensing. Numerical results show that our proposed adaptive sensing greatly outperforms existing approaches, and the learnable quantizer can effectively alleviate the performance loss due to feedback quantization.