CSI Compression Beyond Latents: End-to-End Hybrid Attention-CNN Networks with Entropy Regularization

TL;DR

The paper tackles the CSI feedback bottleneck in FDD massive MIMO by proposing STQENet, an end-to-end framework that jointly learns spatially correlated attention-based CSI encoding, quantization, and entropy modeling. By leveraging a Spatially Separable Attention Transformer, a dual-branch CR Block decoder, and a learned entropy bottleneck with mu-law quantization, STQENet achieves superior rate–distortion performance compared to existing methods on indoor and outdoor COST2100 datasets. The approach demonstrates robustness across bit-depths and SNRs, with significant gains at practical low-bitrate regimes, highlighting the practicality of jointly optimizing coding constraints within neural CSI compression. Overall, the work offers a scalable, deployable solution for efficient CSI feedback in next-generation wireless networks.

Abstract

Massive MIMO systems rely on accurate Channel State Information (CSI) feedback to enable high-gain beam-forming. However, the feedback overhead scales linearly with the number of antennas, presenting a major bottleneck. While recent deep learning methods have improved CSI compression, most overlook the impact of quantization and entropy coding, limiting their practical deployability. In this work, we propose an end-to-end CSI compression framework that integrates a Spatial Correlation-Guided Attention Mechanism with quantization and entropy-aware training. Our model effectively exploits the spatial correlation among the antennas, thereby learning compact, entropy-optimized latent representations for efficient coding. This reduces the required feedback bitrates without sacrificing reconstruction accuracy, thereby yielding a superior rate-distortion trade-off. Experiments show that our method surpasses existing end-to-end CSI compression schemes, exceeding benchmark performance by an average of 21.5% on indoor datasets and 18.9% on outdoor datasets. The proposed framework results in a practical and efficient CSI feedback scheme.

Figures (4)

• Figure 1: STQENet architecture for CSI feedback. "CONV" and "CONVT" denote convolutional and transposed convolutional layers, respectively. "STB" refers to a spatially separable attention transformer block. "CR Block" represents the multi-resolution CNN module from CRNet (see Fig. \ref{['CRB']}.)
• Figure 2: Detailed blocks of the STQENet architecture, adopted from the STNet architecture STNet.
• Figure 3: NMSE vs. SNR at different quantization bit-rates.
• Figure 4: NMSE vs BPP for different methods.