Variable Code Size Autoencoder (VCSA) Meets CSI Compression in Model Generalization

To fully exploit the advantages of spatial multiplexing gains in Multiple-Input Multiple-Output (MIMO) systems operating in Frequency Division Duplex (FDD) mode, it is crucial to develop a robust Channel State Information (CSI) feedback compression and reconstruction strategy. This strategy should effectively reduce the communication overhead while maintaining close-to-optimal reconstruction accuracy. However, conventional codebook-based approaches, as specified in 3rd Gen-eration Partnership Project (3GPP) standards and typical deep learning (DL)-based techniques, encounter challenges like high air-interface overhead and/or poor generalization performance across scenarios. In this paper, we introduce a novel neural network architecture named Variable Code Size Autoencoder (VCSA), a unified two-sided model that not only enables the user equipment (UE) to generate variable-size encoder output but also allows the network to take variable-size input to reconstruct the CSI. Empirical results show that leveraging VCSA together with quantization can achieve comparable performance as using multiple encoder output size-specific autoencoders. Addition-ally, we show that with proper training strategy, the VCSA model achieves decent generalization performance between urban macro (UMa) and urban micro (UMi) scenarios, which further demonstrates the benefits of using DL techniques to provide practical solutions in fifth-generation (5G) networks and beyond.