Source-linked AI summary
CLNet: Complex Input Lightweight Neural Network designed for Massive MIMO CSI Feedback
Sijie Ji, Mo Li
TL;DR
FDD massive MIMO requires CSI feedback, but existing approaches can trade computational complexity for accuracy and may degrade as compression increases. CLNet uses a forged complex-valued input layer and attention mechanisms tailored to CSI, achieving higher accuracy with lower computational overhead than the state-of-the-art comparison method.
Problem
Existing CSI-feedback methods can improve accuracy at increased computational complexity, while accuracy decreases significantly as CSI compression rates increase.
Method
CLNet is an encoder-decoder CSI-feedback framework that integrates complex-valued signal information and uses spatial and channel-wise attention mechanisms.
Results
5.41% higher accuracy and 24.1% less computational overhead than CRNet on average across indoor and outdoor scenarios.
Takeaways & Limitations
CLNet preserves CSI’s complex signal information while maintaining a lightweight architecture for CSI feedback under limited UE resources.
Abstract
from arXiv · showhide
Unleashing the full potential of massive MIMO in FDD mode by reducing the overhead of CSI feedback has recently garnered attention. Numerous deep learning for massive MIMO CSI feedback approaches have demonstrated their efficiency and potential. However, most existing methods improve accuracy at the cost of computational complexity and the accuracy decreases significantly as the CSI compression rate increases. This paper presents a novel neural network CLNet tailored for CSI feedback problem based on the intrinsic properties of CSI. CLNet proposes a forge complex-valued input layer to process signals and utilizes attention mechanism to enhance the performance of the network. The experiment result shows that CLNet outperforms the state-of-the-art method by average accuracy improvement of 5.41\% in both outdoor and indoor scenarios with average 24.1\% less computational overhead. Codes for deep learning-based CSI feedback CLNet are available at GitHub.
I. INTRODUCTION
In FDD massive MIMO, CSI must be explicitly fed back because channel reciprocity is absent, creating substantial overhead that compression can reduce. CLNet addresses this challenge with a lightweight network that preserves CSI’s complex structure and uses attention to improve feedback performance.
- FDD systems lack channel reciprocity, so UEs must explicitly feed downlink CSI back to the base station.Pilot-aided training overhead grows quadratically with the number of transmitting antennas, motivating CSI compression before feedback.
- Traditional compressive-sensing methods depend on channel sparsity and inefficient iterative reconstruction, limiting their suitability across diversified 5G use cases.Deep-learning autoencoders instead learn compression and reconstruction maps that operate without labeled data.
- CsiNet established deep learning as an effective CSI-feedback approach, while later methods often increased accuracy by adding computationally expensive building blocks.CsiNet+ is reported as requiring approximately x7 higher computational overhead than CsiNet, while LSTM-based methods also increase overhead.
- CLNet targets lightweight CSI feedback by integrating real and imaginary signal parts into real-valued neural networks and applying attention to informative clustered signals.Its design is motivated by the physical meaning of complex CSI and limited UE computing and storage resources.
- The CSI channel matrix contains 2NcNt feedback parameters because both real and imaginary parts must be transmitted, with the amount proportional to antenna count.The UE encodes an informative angular-delay representation into a codeword, and the base station decoder reconstructs the channel.
III. CLNET DESIGN
CLNet is an encoder-decoder CSI feedback network tailored to CSI’s physical characteristics. It uses complex-valued input processing and attention to keep the encoder lightweight while improving informativeness.
- CLNet uses an encoder-decoder framework with four building blocks tailored to the CSI feedback problem.
- Limited UE computing power and storage motivate a lightweight encoder that reduces information loss during compression.
- The forged complex-valued input layer embeds CSI real and imaginary parts together to preserve their physical information.
- CBAM spatial-wise attention focuses the network on signal-path clusters in the angular-delay domain and suppresses unnecessary parts.
- CLNet reduces decoder filter size from 1×9 to 1 × 3 and uses hard-Sigmoid activation to lower computational cost.
A. Forged Complex-valued Input
The forged complex-valued input layer addresses the loss of CSI physical structure caused by treating real and imaginary components as independent channels. It combines point-wise convolution with channel-wise attention to preserve and calibrate complex-valued features.
- CSI coefficients encode signal attenuation and phase rotation through their complex-valued representation.
- Conventional convolution entangles neighboring real and imaginary parts, mixing different CSI metrics and losing physical information.
- A 1 × 1 point-wise convolution explicitly embeds each complex-valued coefficient while preserving the ratio between its real and imaginary parts.
- The SE block provides channel-wise attention by modeling relationships among learned weighted representations and suppressing unnecessary features.
- Global average pooling obtains channel statistics, after which a gated transformation produces attention values for channel recalibration.
- The final output preserves CSI physical information while capturing feature dynamics through channel-wise attention.
B. Attention Mechanism for Informative Encoder
CLNet uses spatial-wise attention to distinguish informative propagation-path clusters in the angular-delay domain. The attention mask is computed from pooled spatial descriptors and applied to the feature maps.
- CSI channel coefficients form clusters with different resolutions corresponding to distinguishable propagation paths, delays, and angles of arrival.
- Spatial-wise attention assigns weights from spatial statistical descriptors so the network focuses on distinguishable propagation paths.
- Average-pooling and max-pooling across channels produce two spatial feature maps for attention construction.
- The pooled maps are concatenated, convolved into a spatial attention mask, activated by Sigmoid, and multiplied with the original feature maps.
- Spatial-wise attention focuses CLNet on more informative signal propagation paths in the angular-delay domain.
C. Reduction of Computational Cost
CLNet reduces computational cost by replacing conventional Sigmoid activation with hard-Sigmoid. The paper states that this substitution preserves accuracy while avoiding exponential calculations and improving hardware practicality.
- The conventional Sigmoid activation contains an exponential operation.
- CLNet replaces Sigmoid with a piece-wise linear hard-Sigmoid to reduce computation time.
- Hard-Sigmoid causes no discernible accuracy degradation while avoiding exponential calculations.
- Hard-Sigmoid fits most software and hardware frameworks and can mitigate numerical quantization loss across hardware.
IV. EVALUATION
The evaluation compares CLNet with state-of-the-art deep-learning CSI-feedback approaches using accuracy and computational overhead. The comparison is framed around CSI-feedback performance and resource cost.
- The experiments compare CLNet with state-of-the-art deep-learning CSI-feedback approaches.
- Network accuracy is one evaluation dimension.
- Computational overhead is the second evaluation dimension.
1) Data Generation:
The study uses generated CSI matrices transformed into the angular-delay domain and divides 150,000 independently generated samples into training, validation, and testing sets.
- 150,000 independently generated CSI samples are used in the study.
- The CSI matrices are converted to the angular-delay domain by 2D-DFT, producing Ha ∈R32×32×2.
- 100,000 samples are used for training, 30,000 for validation, and 20,000 for testing.
2) Training Scheme and Evaluation Metric:
Accuracy is evaluated with NMSE between original and reconstructed angular-delay CSI, while computational complexity is measured using floating-point operations under a fixed training setup.
- NMSE between the original Ha and reconstructed ˆHa measures network accuracy.
- Computational complexity is measured by flops, or floating-point operations per second.
- Models are trained with batch size 200 and 8 workers on one NVIDIA 2080Ti GPU for 1000 epochs with a fixed random seed.
3) CLNet Overall Performance:
CLNet is evaluated against related CSI-feedback networks for accuracy and complexity, with additional ablation experiments examining its tailored components. The reported results emphasize lower computational cost and improved accuracy across compression ratios and scenarios.
- Complexity: 24.1% fewer flops on average are reported for CLNet compared with SOTA CRNet.
- Accuracy: 5.41% overall average accuracy improvement is reported for CLNet over SOTA CRNet3 across indoor and outdoor scenarios.
- Accuracy: CLNet consistently outperforms other lightweight networks at all reported compression ratios in indoor and outdoor scenarios.
- Ablation Study: The ablation study uses CRNet with conventional convolution as its baseline to quantify the forged complex-valued input layer and spatial-attention mechanism.
- Ablation Study: 10.964% average improvement is reported after replacing the conventional first layer with a 1x1 forged complex input layer.
5) Encoder Complexity:
CLNet’s encoder is slightly heavier than CRNet’s, while its decoder is designed to remain relatively light for base-station execution; both models require roughly the same storage.
- CLNet’s encoder is slightly heavier than CRNet’s.
- A relatively light decoder benefits base stations that may execute several models simultaneously.
- CLNet and CRNet require roughly the same storage space.
- CLNet achieves 5.41% higher accuracy than CRNet with 24.10% less computation overhead.