Self-Information Domain-Based Neural CSI Compression With Feature Coupling

Zhaohui Yang, Ziqing Yin, Renjie Xie, Wei Xu, Zhaohui Yang 0001, Xiaohu You

TL;DR

This paper introduces a novel deep learning approach, SD-CsiNet, for compressing Channel State Information (CSI) in massive MIMO systems. By leveraging the concept of "self-information" from information theory, the method effectively captures the most crucial information within the CSI, leading to significant performance gains in compression and reconstruction accuracy compared to existing state-of-the-art techniques.

What Problem Does This Paper Address?

Massive MIMO (mMIMO) systems, crucial for future wireless communication, rely on accurate Channel State Information (CSI) feedback from the user equipment (UE) to the base station (BS). However, transmitting this CSI incurs significant overhead, especially with a large number of antennas. Traditional codebook-based methods struggle with the increasing complexity of mMIMO. While deep learning (DL) has shown promise, existing DL-based methods often focus on improving network architecture or exploiting intrinsic CSI features like delay and angle, without directly considering the amount of information contained within the CSI. This paper addresses the gap of efficiently compressing CSI by directly representing and exploiting its information content.

Methods

The core innovation of this paper is the introduction of the self-information domain for CSI representation.

  • Self-Information Transformation (SF) Module: This module transforms the CSI matrix into a "self-information matrix." It first extracts features from the CSI and then, crucially, calculates the pixel-wise entropy (self-information) of the CSI. This process highlights regions of the CSI that contain more information. A thresholding mechanism is used to retain only the most informative pixels.
  • Feature Coupling Encoder: This encoder takes the self-information matrix and extracts both temporal and spatial features. These features are then "coupled" to create a compressed codeword. This coupling aims to leverage the correlated nature of temporal and spatial CSI features for more efficient compression.
  • Feature Decoupling Decoder: At the BS, this decoder takes the compressed codeword and decouples the spatial and temporal features to reconstruct the CSI matrix.

The overall network, SD-CsiNet, integrates these modules for end-to-end CSI compression and recovery.

Datasets

The paper uses synthetically generated channel samples based on the COST 2100 indoor channel model.

  • 100,000 samples for training.
  • 30,000 samples for validation.
  • 20,000 samples for testing.

No new benchmark datasets are introduced.

Key Insights and Ideas

  • Information-Centric Compression: The most significant insight is that directly representing CSI in terms of its "self-information" (information content) is more effective for compression than solely relying on intrinsic features like delay and angle. This allows for a more targeted allocation of compression resources.
  • "Difference" Allocation of Resources: By using the self-information domain, the proposed method can allocate more compressed resources to regions of the CSI that are more informative, effectively achieving a "difference" allocation of compressed resources based on information content. This leads to better reconstruction quality.
  • Coupling Temporal and Spatial Features: The paper demonstrates that coupling temporal and spatial features extracted from the self-information domain further enhances compression efficiency.

Key Results

  • Performance Gain: SD-CsiNet achieves significant performance gains in terms of Normalized Mean-Squared Error (NMSE) compared to state-of-the-art methods like RecCsiNet [8] and CLNet [9]. For compression ratios of 1/8 and 1/16, SD-CsiNet achieves 7.17 dB and 3.68 dB performance gains, respectively.
  • Parameter Efficiency: SD-CsiNet has the least number of parameters among the compared methods, making it more lightweight. It has approximately 19M fewer parameters than RecCsiNet on average.
  • Quantization Robustness: Even when the compressed codeword is quantized (using 6 bits), SD-CsiNet still outperforms RecCsiNet and CLNet, demonstrating its robustness.
  • Ablation Study: An ablation experiment confirms that both the self-information module and the feature coupling/decoupling mechanisms are essential for the high performance of SD-CsiNet.

Limitations and Future Work

  • Computational Complexity: While the number of parameters is low, the online running time of SD-CsiNet is comparable to other methods, suggesting potential for further optimization in terms of computational efficiency.
  • Generalizability: The experiments are conducted using a specific channel model (COST 2100). Future work could explore the performance of SD-CsiNet on a wider range of channel models and real-world scenarios.
  • Further Exploration of Self-Information: The paper introduces self-information as a novel representation. Further research could explore other information-theoretic measures or more advanced ways to leverage the self-information domain for CSI compression.

Summary updated

Abstract

Deep learning (DL)-based channel state information (CSI) feedback methods compressed the CSI matrix by exploiting its delay and angle features straightforwardly, while the measure in terms of information contained in the CSI matrix has rarely been considered. Based on this observation, we introduce self-information as an informative CSI representation from the perspective of information theory, which reflects the amount of information of the original CSI matrix in an explicit way. Then, a novel DL-based network is proposed for temporal CSI compression in the self-information domain, namely SD-CsiNet. The proposed SD-CsiNet projects the raw CSI onto a self-information matrix in the newly-defined self-information domain, extracts both temporal and spatial features of the self-information matrix, and then couples these two features for effective compression. Experimental results verify the effectiveness of the proposed SD-CsiNet by exploiting the self-information of CSI. Particularly for compression ratios 1/8 and 1/16, the SD-CsiNet respectively achieves 7.17 dB and 3.68 dB performance gains compared to state-of-the-art methods.
BibTeX
@article{yang2023self,
  author = {Zhaohui Yang and Ziqing Yin and Renjie Xie and Wei Xu and Zhaohui Yang and Xiaohu You},
  title = {{Self-Information Domain-Based Neural CSI Compression With Feature Coupling}},
  journal = {IEEE Trans. Vehicular Technology},
  year = {2023},
  doi = {10.1109/tvt.2023.3272560},
}
Figure