Source-linked AI summary
Multi-resolution CSI Feedback with deep learning in Massive MIMO System
Zhilin Lu, Jintao Wang, Jian Song
TL;DR
Massive MIMO downlink CSI feedback is expensive because massive CSI representations create substantial feedback overhead. The paper introduces CRNet’s multi-resolution architecture and an advanced training scheme, reporting that CRNet outperforms CsiNet under the same computational complexity without extra information.
Problem
Massive MIMO requires downlink CSI feedback, but the growing CSI matrix makes feedback costly and conventional compression can be inadequate at high compression ratios.
Method
CRNet uses multi-resolution CRBlocks and convolution factorization to extract CSI features at different resolutions, together with a warm-up aided cosine learning-rate scheduler and exhaustive training.
Results
CRNet greatly outperforms CsiNet under the same computational complexity, and the advanced training scheme boosts the proposed network’s performance.
Takeaways & Limitations
Multi-resolution paths, convolution factorization, and training-scheme design are effective components for downlink CSI feedback in massive MIMO FDD systems.
Abstract
from arXiv · showhide
In massive multiple-input multiple-output (MIMO) system, user equipment (UE) needs to send downlink channel state information (CSI) back to base station (BS). However, the feedback becomes expensive with the growing complexity of CSI in massive MIMO system. Recently, deep learning (DL) approaches are used to improve the reconstruction efficiency of CSI feedback. In this paper, a novel feedback network named CRNet is proposed to achieve better performance via extracting CSI features on multiple resolutions. An advanced training scheme that further boosts the network performance is also introduced. Simulation results show that the proposed CRNet outperforms the state-of-the-art CsiNet under the same computational complexity without any extra information. The open source codes are available at https://github.com/Kylin9511/CRNet
I. INTRODUCTION
Massive MIMO downlink CSI feedback is costly because the channel matrix is large, while conventional compressed sensing becomes inadequate at high compression ratios. The paper proposes CRNet, a multi-resolution network paired with an advanced training scheme, and reports better performance than CsiNet at equal computational complexity.
- Massive MIMO requires downlink CSI feedback from the UE to the BS, but the huge CSI matrix makes the feedback bandwidth overhead unacceptable.
- Compressed-sensing methods require sufficiently sparse CSI, a condition that practical systems especially fail to meet at large compression ratios.
- Deep-learning encoder-decoder networks are motivated as an approach to CSI compression, following their effectiveness in related image-compression tasks.
- Prior work extends CsiNet using temporal or uplink-downlink CSI correlations, but these variants introduce extra conditions or assumptions.
- CsiNetPlus improves performance without extra information but retains most CsiNet architectural choices and requires substantially more floating-point operations.
- CRNet uses multi-resolution CRBlocks and convolution factorization, while an advanced training scheme is introduced to improve performance under equal computational complexity.
II. SYSTEM MODEL
The system transforms the downlink channel into an angular-delay representation, compresses it with a CRNet encoder, feeds back the resulting feature vector, and reconstructs the channel at the BS. The model exploits angular-delay sparsity but still requires learned compression because the retained representation remains large and insufficiently sparse in practical systems.
- The considered setup is a single-cell massive MIMO FDD system with many BS antennas, one UE antenna, and OFDM across Nc subcarriers.
- The downlink channel matrix H contains NcNt elements, creating an unacceptably large feedback representation in massive MIMO FDD systems.
- A DFT transforms H from the spatial-frequency domain into the angular-delay domain, where only the first Na rows contain large values and the remaining rows can be omitted with little information loss.
- Even after removing near-zero rows, Ha remains large, and limited Nt makes its sparsity insufficient for compressed sensing at high compression ratios.
- CRNet’s encoder compresses Ha into a feature vector v at a specified compression ratio, and the decoder reconstructs Ha before zero filling and inverse DFT restore H.
- The feedback scheme is represented by ˆHa = D(E(Ha, ΘE), ΘD), where E and D are the CRNet encoder and decoder and training minimizes reconstruction distance.
A. The proposed CRNet
CRNet uses multi-resolution processing to adapt CSI feature extraction to changing feature granularity, with separate encoder and decoder designs for feedback compression and reconstruction.
- Motivation: Multi-resolution paths address CSI granularity changes that fixed-resolution residual architectures cannot adapt to well.Smaller kernels suit denser CSI, while larger kernels are preferred for sparser CSI.
- Architecture: CRNet treats the angular CSI matrix as a two-channel input and separates processing between the UE encoder and BS decoder.The channels represent the real and imaginary parts of the angular-domain matrix.
- Encoder: The encoder combines parallel convolutional paths, concatenates their outputs, merges them with a 1 × 1 convolution, and scales the feature by compression ratio.One path provides a larger-resolution view, while the other uses a single 3×3 convolution at smaller resolution.
- Decoder: The decoder enlarges and resizes the received feature vector, then refines it through two CRBlocks with parallel paths at different resolutions.Each CRBlock merges its path outputs with a 1 × 1 convolution and uses an identity path for residual learning.
- Multi-resolution CRBlock: CRBlocks flexibly extract features at different sizes, enhancing adaptation to varying feature granularities.The encoder is also a simplified CRBlock with two parallel convolutional paths.
- Convolution factorization: Convolution factorization replaces a 9 × 9 kernel with serial 1 × 9 and 9 × 1 kernels, preserving resolution area while reducing computational complexity.The paper reports that this design expands resolution without increasing flops and supports better performance at the same or lower complexity than CsiNet.
B. The advanced training scheme design
The paper argues that CSI-feedback training requires explicit scheme design because overfitting is difficult and constant learning rates may be insufficient. It therefore combines extended training with warm-up cosine annealing.
- Motivation: Training-scheme design is presented as an underexamined factor affecting the reliability of reported CSI-feedback performance.The paper specifically examines training epochs and learning-rate scheduling.
- Training epochs: CSI feedback is described as more prone to underfitting than overfitting because CRNet is lightweight and CSI information is relatively pure.The paper consequently recommends exhausted training, which experiments found helpful.
- Learning-rate scheduler: Constant learning rate is reported as insufficient for CSI feedback even with Adam, whereas warm-up cosine annealing significantly improves CRNet performance.This result contrasts with Adam’s general reduced sensitivity to scheduler choice.
- Learning-rate scheduler: The proposed scheduler linearly increases the learning rate from zero to its maximum during warm-up, then decays it according to cosine annealing.The cosine schedule uses longer early and final stages, with smooth continuous updates intended to stabilize training.
- Overall scheme: The advanced training scheme combines warm-up aided cosine annealing with exhausted training to improve CRNet performance.The paper emphasizes that training changes do not increase online inference cost.
A. Experiment Setting
Experiments evaluate CRNet on indoor and outdoor COST2100 channel scenarios using the basic CsiNet system settings for comparison.
- Channel scenarios: The evaluation covers indoor 5.3GHz and outdoor 300MHz scenarios generated with the default COST2100 channel setting.Both scenarios use a uniform linear array with Nt = 32 at the base station.
- System settings: The FDD setup uses Nc = 1024 frequency-domain subcarriers and Na = 32 angular-domain components, with 150,000 independently generated channels.The channels are divided into training, validation, and testing portions.
- Training implementation: Training uses PyTorch, Adam, Xavier initialization, and mean square error loss, with a 2500-epoch warm-up cosine schedule.The learning rate changes from 2e-3 initially to 5e-5 finally, and the first 30 epochs provide warm-up.
B. Performance of the proposed CRNet and training scheme
CRNet improves CSI reconstruction over CsiNet through multi-resolution architecture, while its advanced training scheme further lowers NMSE without adding computational cost. Longer training generally helps because CSI feedback exhibits relatively weak overfitting, although constant-learning-rate training can become unstable.
- Architecture comparison: Under the same training scheme, CRNet outperforms CsiNet with fewer flops, demonstrating the benefit of its architecture.The comparison uses CRNet-const, trained for 1000 epochs with constant learning rate 0.001, matching CsiNet’s scheme.
- Training scheme: The advanced cosine-annealing scheme further improves CRNet performance without extra flops.The scheme combines warm-up and cosine annealing; CRNet-cosine is trained for 2500 epochs.
- Training epochs: Performance generally improves as the number of training epochs increases, supporting longer training for this CSI feedback task.The paper attributes this behavior to relatively weak overfitting during training.
- Training epochs: Constant-scheduler training can become trapped in a poor local minimum; one run improved from -19.18 dB by another 2.71 dB after 2500 additional epochs.The authors use this example to motivate a better training scheme and exhaustive training.
- Training scheme: CRNet’s NMSE continues descending while CsiNet’s loss reaches a plateau, and the cosine-based scheme produces lower, more stable NMSE.The NMSE curve becomes less fluctuant as training proceeds under the new scheme.
- Ablation observations: Larger leaky ReLU negative slopes are essential for CRNet because optimization becomes harder when the slope is too small.This observation is reported from the negative-slope ablation study.
V. CONCLUSION
CRNet improves downlink CSI feedback in massive MIMO FDD systems by combining multi-resolution paths, convolution factorization, and an advanced training scheme. It outperformed CsiNet under the same computational complexity.
- CRNet introduces multi-resolution paths and convolution factorization for downlink CSI feedback.The paper reports both design choices as effective for the CSI feedback task.
- An advanced training scheme further boosts CRNet’s performance.
- CRNet greatly outperformed state-of-the-art CsiNet under the same computational complexity.