[Paper Review] Binarized Aggregated Network with Quantization: Flexible Deep Learning Deployment for CSI Feedback in Massive MIMO System

Zhilin Lu, Xudong Zhang+3 more|arXiv (Cornell University)|May 1, 2021

Advanced MIMO Systems Optimization4 citations

TL;DR

This paper proposes ACRNet, a binarized and quantized deep learning network for efficient CSI feedback in massive MIMO FDD systems. By integrating network aggregation, PReLU activation, and flexible quantization with binarized fully connected layers, ACRNet achieves state-of-the-art NMSE performance while enabling up to 31× memory reduction at the UE and 16× feedback overhead reduction, demonstrating high practicality and adaptability under diverse resource constraints.

ABSTRACT

Massive multiple-input multiple-output (MIMO) is one of the key techniques to achieve better spectrum and energy efficiency in 5G system. The channel state information (CSI) needs to be fed back from the user equipment to the base station in frequency division duplexing (FDD) mode. However, the overhead of the direct feedback is unacceptable due to the large antenna array in massive MIMO system. Recently, deep learning is widely adopted to the compressed CSI feedback task and proved to be effective. In this paper, a novel network named aggregated channel reconstruction network (ACRNet) is designed to boost the feedback performance with network aggregation and parametric rectified linear unit (PReLU) activation. The practical deployment of the feedback network in the communication system is also considered. Specifically, the elastic feedback scheme is proposed to flexibly adapt the network to meet different resource limitations. Besides, the network binarization technique is combined with the feature quantization for lightweight and practical deployment. Experiments show that the proposed ACRNet outperforms loads of previous state-of-the-art networks, providing a neat feedback solution with high performance, low cost and impressive flexibility.

Motivation & Objective

• Address the high feedback overhead of CSI in massive MIMO FDD systems due to large antenna arrays.
• Overcome limitations of traditional compressed sensing and existing deep learning methods in terms of performance and deployment efficiency.
• Design a flexible, lightweight deep learning architecture that maintains high reconstruction accuracy under strict resource constraints.
• Integrate network binarization and feature quantization to enable practical deployment on resource-constrained user equipment and base stations.
• Develop an elastic feedback scheme that allows dynamic adaptation of network capacity and precision to different system requirements.

Proposed method

• Propose ACRNet, a novel aggregated channel reconstruction network using network aggregation and parametric ReLU (PReLU) activation to enhance feature representation and reconstruction accuracy.
• Introduce an elastic feedback scheme that allows dynamic scaling of network width (via scaling factor η) to adapt to varying resource constraints at UE and BS.
• Apply vector quantization and fully connected (FC) layer binarization to drastically reduce model size and feedback overhead.
• Combine quantization with binarization to achieve memory savings without significant performance degradation, leveraging the regularization effect of quantization.
• Design a feedback overhead model where feedback overhead Nbf remains constant when η×B is fixed, enabling trade-offs between compression ratio and bit precision.
• Use ablation studies to optimize the combination of η and B for different scenarios (indoor/outdoor), identifying optimal configurations such as η=1/4 and B=4 for minimal NMSE under fixed Nbf.

Experimental results

Research questions

• RQ1 How can deep learning-based CSI feedback networks be made both highly accurate and practically deployable under strict resource constraints in massive MIMO systems?
• RQ2 What is the impact of network binarization and quantization on CSI reconstruction performance when combined with architectural flexibility?
• RQ3 Can an elastic network design with scalable width (via η) maintain high performance across different feedback overhead levels?
• RQ4 How does the combination of vector quantization and FC binarization affect memory footprint and feedback overhead while preserving reconstruction accuracy?
• RQ5 What optimal trade-off between compression ratio (η) and bit precision (B) maximizes performance under a fixed feedback overhead (Nbf)?

Key findings

• ACRNet achieves a 23.36 dB NMSE performance under η=1/4 and 32-bit quantization, significantly outperforming prior state-of-the-art methods like CsiNet.
• With 8-bit quantization, ACRNet maintains near-ideal performance (16.10 dB NMSE) in indoor scenarios, and even improves performance in outdoor scenarios due to regularization from quantization.
• The binarized version of ACRNet (BACRNet-10×) achieves -15 dB NMSE with η=1/4 and B=2, while reducing feedback overhead by 16× and memory usage by 31× compared to vanilla CsiNet.
• The combination of binarization and quantization enables up to 30× memory reduction at the UE and 18× at the BS, with minimal performance loss when properly tuned.
• The elastic architecture design ensures that ACRNet-10× consistently outperforms ACRNet-1× even under aggressive binarization and quantization, confirming the robustness of the scalable design.
• Optimal performance under fixed feedback overhead (Nbf=2048) is achieved at η=1/4 and B=4, yielding the lowest NMSE in indoor scenarios, demonstrating the effectiveness of the proposed hyperparameter tuning strategy.