[Paper Review] Lightweight Convolutional Neural Networks for CSI Feedback in Massive MIMO

Zheng Cao, Wan-Ting Shih+3 more|arXiv (Cornell University)|May 1, 2020

Wireless Signal Modulation Classification17 references12 citations

TL;DR

This paper proposes ConvCsiNet, a deep learning-based CSI feedback network for FDD massive MIMO systems using a convolutional autoencoder to enhance reconstruction accuracy by preserving 2D spatial channel structure. To enable mobile deployment, it further introduces ShuffleCsiNet, a lightweight variant using depthwise separable convolutions and channel shuffling, achieving 79% fewer parameters and 80% lower FLOPs than ConvCsiNet while maintaining near-optimal reconstruction performance at 1/32 compression ratio.

ABSTRACT

In frequency division duplex mode of massive multiple-input multiple-output systems, the downlink channel state information (CSI) must be sent to the base station (BS) through a feedback link. However, transmitting CSI to the BS is costly due to the bandwidth limitation of the feedback link. Deep learning (DL) has recently achieved remarkable success in CSI feedback. Realizing high-performance and low-complexity CSI feedback is a challenge in DL based communication. We develop a DL based CSI feedback network in this study to complete the feedback of CSI effectively. However, this network cannot be effectively applied to the mobile terminal because of the excessive numbers of parameters. Therefore, we further propose a new lightweight CSI feedback network based on the developed network. Simulation results show that the proposed CSI network exhibits better reconstruction performance than that of other CsiNet-related works. Moreover, the lightweight network maintains a few parameters and parameter complexity while ensuring satisfactory reconstruction performance. These findings suggest the feasibility and potential of the proposed techniques.

Motivation & Objective

• To address the high computational and memory cost of deep learning-based CSI feedback in mobile terminals.
• To improve reconstruction accuracy of CSI feedback in frequency-division duplex (FDD) massive MIMO systems.
• To reduce model complexity and parameter count while maintaining high reconstruction performance for practical deployment on mobile devices.
• To develop a lightweight network architecture that retains robustness and performance under severe bandwidth constraints.

Proposed method

• Design a convolutional autoencoder-based CSI feedback network (ConvCsiNet) using convolutional layers, mean-pooling, and upsampling to preserve 2D spatial correlation in CSI matrices.
• Introduce RefineNet and DN units in the decoder to refine feature reconstruction and improve spatial resolution.
• Develop ShuffleCsiNet by replacing standard convolutions with depthwise separable convolutions and incorporating channel shuffle operations to reduce parameter count and FLOPs.
• Use batch normalization (BN) and ReLU activation in all convolutional blocks to stabilize training and enhance feature learning.
• Apply mean-pooling and upsampling layers to downsample and upsample feature maps across encoder and decoder paths, respectively.
• Optimize the network for low-complexity deployment by minimizing fully connected layers and leveraging functional reuse in the shuffle mechanism.

Experimental results

Research questions

• RQ1 Can a deep learning-based CSI feedback network achieve superior reconstruction performance compared to CsiNet and its variants in FDD massive MIMO systems?
• RQ2 How can the parameter count and computational complexity of a CNN-based CSI feedback network be reduced without significantly degrading reconstruction quality?
• RQ3 To what extent does the channel shuffle mechanism in depthwise separable convolutions improve model robustness and efficiency in CSI feedback?
• RQ4 Can a lightweight network architecture maintain high reconstruction accuracy under low feedback compression ratios (e.g., 1/32)?
• RQ5 How does the integration of spatial feature extraction via convolutional layers compare to fully connected or feedforward networks in CSI feedback?

Key findings

• ConvCsiNet achieves a normalized mean square error (NMSE) of -13.79 dB in indoor and -6.00 dB in outdoor environments at a 1/16 compression ratio, outperforming CsiNet, CRNet, and DS-NLCsiNet.
• At a 1/32 compression ratio, ConvCsiNet achieves an NMSE of -10.10 dB indoors and -5.21 dB outdoors, demonstrating superior reconstruction accuracy under high compression.
• ShuffleCsiNet reduces model parameters to 21.05% of ConvCsiNet’s at 1/32 compression ratio, with parameter count dropping from 1,623,416 to 341,800.
• ShuffleCsiNet reduces FLOPs to 19.8% of ConvCsiNet’s at 1/32 compression, from 58.2M to 11.6M, significantly lowering computational complexity.
• Despite lower complexity, ShuffleCsiNet maintains near-optimal performance, achieving NMSE of -9.41 dB indoors and -3.50 dB outdoors at 1/32 compression ratio.
• The channel shuffle and depthwise separable convolution design in ShuffleCsiNet enhance model robustness and functional reuse, enabling high performance with minimal resource usage.