Transformer Empowered CSI Feedback for Massive MIMO Systems

This work investigates the problem of CSI feedback in massive multiple-input multiple-output (MIMO) systems. It is well-known that accurate CSI plays a crucial role in realizing the beamforming gain promised by the MIMO technology. However, CSI feedback often incurs excessive feedback overhead. To cope with this problem, this work proposes an effective and robust CSI feedback scheme called CsiTransformer by leveraging a recently developed machine learning (ML) model named transformer in its encoder and decoder. In contrast to the existing ML approaches, CsiTransformer can more effectively exploit the correlation among elements in the channel matrix, which leads to improved CSI reconstruction quality. Furthermore, motivated by the fact that cellphones are power-limited, we propose a further reduced-complexity scheme called MixedCsiNet by using the less computationally expensive conventional convolutional neural networks (CNNs) as the encoder while the transformer model as the decoder. Extensive computer simulation is performed to confirm that CsiTransformer and MixedCsiNet are able to achieve substantially better CSI feedback at different compression rates as compared to existing CSI feedback schemes.