[Paper Review] Learning Representations for CSI Adaptive Quantization and Feedback

Valentina Rizzello, Matteo Nerini+3 more|arXiv (Cornell University)|Jul 13, 2022

Wireless Signal Modulation Classification4 citations

TL;DR

This paper proposes two deep learning-based methods for adaptive quantization and feedback of CSI in FDD massive MIMO systems: one using nested dropout to enable bit allocation based on latent unit importance, and another using vector quantizer variational autoencoders (VQ-VAE) with learned codebooks. Both methods significantly improve reconstruction accuracy over standard quantization, especially at low feedback rates, while reducing encoder complexity and parameter offload via pruning and binary weights.

ABSTRACT

In this work, we propose an efficient method for channel state information (CSI) adaptive quantization and feedback in frequency division duplexing (FDD) systems. Existing works mainly focus on the implementation of autoencoder (AE) neural networks (NNs) for CSI compression, and consider straightforward quantization methods, e.g., uniform quantization, which are generally not optimal. With this strategy, it is hard to achieve a low reconstruction error, especially, when the available number of bits reserved for the latent space quantization is small. To address this issue, we recommend two different methods: one based on a post training quantization and the second one in which the codebook is found during the training of the AE. Both strategies achieve better reconstruction accuracy compared to standard quantization techniques.

Motivation & Objective

• Address the suboptimal performance of standard uniform quantization in CSI feedback for FDD massive MIMO systems.
• Improve reconstruction accuracy of compressed CSI, particularly when feedback bits are limited.
• Reduce the number of parameters and computational complexity of the encoder network for efficient offloading to mobile terminals.
• Enable adaptive bit allocation to latent space units by learning their relative importance during training.

Proposed method

• Apply a nested dropout (ND) layer as the final encoder layer to implicitly rank latent units by importance, enabling adaptive bit allocation during quantization.
• Use a geometric distribution to sample the dropout start index, ensuring hierarchical dependency among latent units.
• Introduce a VQ-VAE-based method where the codebook is learned during training, enabling more efficient and accurate vector quantization of the latent representation.
• Reduce encoder parameter offload via network pruning (40% reduction) and binary fully connected (FC) layers (75% reduction), with fine-tuning to recover performance.
• Leverage the generalization of uplink CSI to train the autoencoder centrally at the base station, avoiding on-device training at mobile terminals.
• Use a straight-through estimator with scaling for binary weight layers to maintain gradient flow during backpropagation.

Experimental results

Research questions

• RQ1 Can nested dropout enable more efficient and accurate quantization of the autoencoder’s latent space in CSI feedback by exploiting learned importance ordering?
• RQ2 Does training a VQ-VAE with a learned codebook improve CSI reconstruction accuracy compared to standard quantization in low-bit feedback scenarios?
• RQ3 To what extent can model compression techniques like pruning and binary weights reduce encoder parameter offload without degrading reconstruction performance?
• RQ4 How does the proposed method compare in performance to state-of-the-art CSI compression networks like CsiNet and ACRNet under the same feedback constraints?

Key findings

• The nested dropout-based method achieved a normalized mean square error (NMSE) of -12.48 dB at 512 feedback bits, outperforming standard quantization and CsiNet (-7.71 dB) and ACRNet (-10.05 dB).
• The VQ-VAE-based method with learned codebooks achieved a cosine similarity of -15.22 dB at 512 bits, significantly better than CsiNet (-10.92 dB) and ACRNet (-13.73 dB).
• Network pruning combined with fine-tuning reduced encoder parameter offload by 40% while maintaining NMSE within 0.1 dB of the full model.
• Binary fully connected layers reduced parameter offload by ~75%, but caused a performance degradation of ~1.5 dB in NMSE compared to full-precision models.
• The ND-based method with g=2 (grouped convolutions) achieved NMSE of -12.45 dB at 512 bits, showing robustness to architectural modifications.
• Overall, the proposed methods outperformed both CsiNet and ACRNet across all feedback bit levels, demonstrating superior reconstruction accuracy under low-bit constraints.