Dynamic Model Fine-Tuning For Extreme MIMO CSI Compression

Mehdi Sattari, Deniz Gündüz, Tommy Svensson

Abstract Deep learning-based Channel State Information (CSI) compression is essential for Frequency Division Duplexing (FDD) massive MIMO systems. However, these models often suffer from performance degradation when encountering distribution shifts in wireless environments. This paper proposes an online full-model (FM) fine-tuning strategy that updates both encoder and decoder parameters. To manage the communication overhead of transmitting updated model weights, we employ a "spike-and-slab" prior that encourages sparsity in the weight updates. Our approach significantly improves reconstruction accuracy in dynamic scenarios while maintaining a low update bit-rate.

Introduction

In massive MIMO systems, the base station requires accurate CSI for effective beamforming. In FDD mode, the user equipment (UE) must feedback this high-dimensional CSI matrix, which can consume prohibitive amounts of bandwidth. While deep learning autoencoders have shown promise in compressing CSI, they lack generalization; a model trained for one environment (e.g., static outdoor) often fails when the UE moves into a different environment (e.g., dense urban).

Fine-tuning the model to the new environment is a potential solution, but it introduces the challenge of synchronizing the decoder at the base station. If the cost of sending model updates exceeds the savings from compression, the approach becomes impractical. This work addresses this trade-off by treating model updates as an integral part of the total communication cost.

Methodology

The proposed framework utilizes a convolutional neural network (CNN) autoencoder architecture. The core innovation lies in the joint optimization of reconstruction error and the bit-rate required for both the compressed CSI and the model updates.

The proposed autoencoder architecture for CSI compression, featuring an encoder, decoder, and entropy-coded model updates.

To achieve sparse and compressible updates, we introduce a spike-and-slab prior. This prior is a mixture of two Gaussian distributions: a very narrow "spike" centered at zero and a wider "slab." This distribution encourages the network to keep most weight updates at zero, ensuring that only the most critical parameters are transmitted. The updates are further processed through quantization and lossless entropy coding (e.g., CABAC). To enable end-to-end training with non-differentiable quantization, a Straight-Through Estimator (STE) is employed.

Experimental Setup and Datasets

The performance of the proposed method was evaluated using two realistic ray-tracing datasets:

  1. DeepMIMO: Used for initial training in a static outdoor scenario (Scenario 'O1') and testing under moderate distribution shifts (Scenario 'O2').
  2. QuaDRiGa: Used to simulate a more challenging transition to a dense urban environment with a mobile user.

The study specifically examined the impact of user mobility, identifying optimal update intervals to balance model freshness against transmission overhead. For a user moving at 5 km/h, updates every 8 to 16 meters were found to be most effective.

Key Results and Discussion

The experimental results demonstrate that full-model fine-tuning is superior to encoder-only updates. In the QuaDRiGa dataset, the proposed method improved the Normalized Mean Square Error (NMSE) from a baseline of -9 dB to approximately -16 dB, nearing the performance of a theoretical "genie-aided" model with zero-cost updates.

Rate-distortion performance (NMSE vs. Rate) for $256 \times 256$ and $64 \times 64$ antenna configurations.

The spike-and-slab prior successfully reduced the number of non-zero weight updates by several orders of magnitude. Furthermore, the backbone network remains computationally efficient, requiring only ~475 million FLOPs for inference, which is significantly more efficient than traditional iterative algorithms like ISTA-Net+.

Conclusion

This work presents a robust strategy for adapting CSI compression models to changing wireless environments. By incorporating model update costs into the training objective and leveraging sparse priors, we enable efficient full-model fine-tuning. Future research will explore adaptive update intervals and high-mobility scenarios to further enhance the system's practical utility.

Summary updated

Abstract

Efficient channel state information (CSI) compression is essential in frequency division duplexing (FDD) massive multiple-input multiple-output (MIMO) systems due to the significant feedback overhead. Recently, deep learning-based compression techniques have demonstrated superior performance across various data types, including CSI. However, these methods often suffer from performance degradation when the data distribution shifts, primarily due to limited generalization capabilities. To address this challenge, we propose an online model fine-tuning approach for CSI feedback in massive MIMO systems. We consider full-model fine-tuning, where both the encoder and decoder are jointly updated using recent CSI samples. A key challenge in this setup is the transmission of updated decoder parameters, which introduces additional feedback overhead. To mitigate this bottleneck, we incorporate the bit-rate of model updates into the fine-tuning objective and entropy code the updates jointly with the compressed CSI. To reduce the bit-rate, we design an efficient prior distribution that encourages the network to update only the most significant weights, thereby minimizing the overall model update cost. Our results show that full-model fine-tuning significantly enhances the rate-distortion (RD) performance of neural CSI compression despite the additional communication cost of model updates. Moreover, we investigate the impact of update frequency in dynamic wireless environments and identify an optimal fine-tuning interval that achieves the best RD trade-off.
BibTeX
@misc{sattari2025dynamic,
  author = {Mehdi Sattari and Deniz Gündüz and Tommy Svensson},
  title = {{Dynamic Model Fine-Tuning For Extreme MIMO CSI Compression}},
  year = {2025},
  eprint = {2501.18250},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.18250},
}
Figure