Generative Model-Aided Continual Learning for CSI Feedback in FDD mMIMO-OFDM Systems

TL;DR

This work tackles the problem of catastrophic forgetting in CSI feedback models for dynamic FDD mMIMO-OFDM environments caused by mobility. It introduces a GAN-aided continual learning framework that uses a generative memory unit to capture and replay past CSI distributions, enabling the CSI feedback autoencoder to generalize across scenarios without retraining from scratch. The approach trains on current CSI data augmented with high-quality synthetic samples produced by stored GAN generators, achieving NMSE performance close to joint training while maintaining low memory overhead and compatibility with existing CSI feedback models like CRNet. The proposed solution offers a practical path to robust, scalable CSI feedback in mobility-rich networks by mitigating forgetting and reducing storage requirements.

Abstract

Deep autoencoder (DAE) frameworks have demonstrated their effectiveness in reducing channel state information (CSI) feedback overhead in massive multiple-input multiple-output (mMIMO) orthogonal frequency division multiplexing (OFDM) systems. However, existing CSI feedback models struggle to adapt to dynamic environments caused by user mobility, requiring retraining when encountering new CSI distributions. Moreover, returning to previously encountered environments often leads to performance degradation due to catastrophic forgetting. Continual learning involves enabling models to incorporate new information while maintaining performance on previously learned tasks. To address these challenges, we propose a generative adversarial network (GAN)-based learning approach for CSI feedback. By using a GAN generator as a memory unit, our method preserves knowledge from past environments and ensures consistently high performance across diverse scenarios without forgetting. Simulation results show that the proposed approach enhances the generalization capability of the DAE framework while maintaining low memory overhead. Furthermore, it can be seamlessly integrated with other advanced CSI feedback models, highlighting its robustness and adaptability.

Figures (6)

• Figure 1: An overview of the autoencoder CSI feedback framework.
• Figure 2: The adopted GAN model structure, which consists of a generator $G(\cdot)$ and a discriminator $D(\cdot)$. The "Upsample, factor=2" module refers to the nearest neighbor interpolation with a scale factor of 2, while "Mean Layer" denotes an averaging operation applied to the last two dimensions of the input.
• Figure 3: The proposed continual learning framework for CSI feedback in dynamic environments. The dashed arrow indicates that in the encountered scenarios, the memory unit $\mathcal{M}$ has already captured the data distribution $data_{0, \ldots, t}$ through GAN-based channel modeling.
• Figure 4: “O1_28" an outdoor scenario of two streets and one intersection at operating frequencies 28 GHz in DeepMIMO.
• Figure 5: NMSE (dB) performance comparison of the proposed approach, DT, and MTL under various compression ratios $\gamma$.