PaperlayerView exact revision on arXiv

Source-linked AI summary

Convolutional Neural Network based Multiple-Rate Compressive Sensing for Massive MIMO CSI Feedback: Design, Simulation, and Analysis

Jiajia Guo, Chao-Kai Wen, Shi Jin, Geoffrey Ye Li

arXiv:1906.06007v1eess.SPcs.IT

TL;DR

Massive MIMO CSI feedback creates substantial overhead, while FDD systems make downlink CSI difficult to infer from uplink CSI. The paper introduces CsiNet+, a multiple-rate compressive sensing and quantization framework with SM-CsiNet+ and PM-CsiNet+ variable-rate variants. These variants reduce UE parameter counts by 38.0% and 46.7%, respectively, while the framework addresses reconstruction, quantization, and storage constraints.

  • Problem

    FDD downlink CSI is difficult to obtain, and existing fixed-rate deep-learning feedback requires multiple stored network architectures for variable-rate compression.

  • Method

    The paper develops CsiNet+, a quantization framework and training strategy, and two variable-rate frameworks called SM-CsiNet+ and PM-CsiNet+.

  • Results

    SM-CsiNet+ and PM-CsiNet+ reduce UE parameter counts by 38.0% and 46.7%, respectively.

  • Takeaways & Limitations

    Parameter visualization provides understanding of deep-learning CSI feedback and guidelines for subsequent research.

Abstract

from arXiv · show

Massive multiple-input multiple-output (MIMO) is a promising technology to increase link capacity and energy efficiency. However, these benefits are based on available channel state information (CSI) at the base station (BS). Therefore, user equipment (UE) needs to keep on feeding CSI back to the BS, thereby consuming precious bandwidth resource. Large-scale antennas at the BS for massive MIMO seriously increase this overhead. In this paper, we propose a multiple-rate compressive sensing neural network framework to compress and quantize the CSI. This framework not only improves reconstruction accuracy but also decreases storage space at the UE, thus enhancing the system feasibility. Specifically, we establish two network design principles for CSI feedback, propose a new network architecture, CsiNet+, according to these principles, and develop a novel quantization framework and training strategy. Next, we further introduce two different variable-rate approaches, namely, SM-CsiNet+ and PM-CsiNet+, which decrease the parameter number at the UE by 38.0% and 46.7%, respectively. Experimental results show that CsiNet+ outperforms the state-of-the-art network by a margin but only slightly increases the parameter number. We also investigate the compression and reconstruction mechanism behind deep learning-based CSI feedback methods via parameter visualization, which provides a guideline for subsequent research.

I. INTRODUCTION

Massive MIMO makes CSI feedback essential but costly, especially in FDD systems where weak reciprocity prevents reliable downlink inference and large antenna arrays increase CSI dimension. The paper proposes multiple-rate deep-learning compressive sensing, quantization, and variable-rate frameworks to improve reconstruction while reducing UE storage demands.

  • FDD systems make downlink CSI difficult to infer from uplink CSI because channel reciprocity is weak.
  • Massive MIMO increases CSI-feedback overhead because its large antenna arrays enlarge the CSI matrix dimension.
  • CS simplifies encoding but requires computationally intensive, iterative decoding that can demand substantial time and resources.
  • Existing deep-learning feedback methods use fixed compression rates, forcing UEs to store multiple architectures and parameter sets for variable-rate operation.
  • CsiNet+ uses CSI sparsity and refinement theory, while the proposed quantization framework supports different quantization rates without increasing UE parameters or computational resources.
  • SM-CsiNet+ and PM-CsiNet+ provide variable-rate compression and reduce UE parameter counts by 38.0% and 46.7%, respectively.

II. SYSTEM MODEL

The system models CSI feedback in a single-cell FDD massive MIMO-OFDM link and reduces its antenna-scaled overhead by exploiting angular-delay sparsity.

  • The downlink CSI is estimated at the UE and fed back to the BS, which uses it to calculate precoding vectors through singular value decomposition.
  • 8?
  • Excessive CSI feedback in massive MIMO occupies precious bandwidth because the feedback parameter count grows with the number of transmit antennas.
  • A 2D DFT converts the spatial-frequency CSI matrix into the angular-delay domain, where most delay-domain elements are near zero.
  • Only the first N′_c rows contain distinct nonzero values, so the channel matrix is truncated to those rows before feedback processing.

B. CSI Feedback Process

The CSI feedback process compresses and quantizes the UE’s angular-delay channel matrix, then dequantizes and decompresses it at the BS for reconstruction.

  • The UE applies compression, quantization, and entropy encoding sequentially to reduce CSI feedback overhead.
  • The compression process uses a parameterized encoder, while the decoder-side dequantization and decompression recover the angular-delay channel matrix.
  • The encoder and decoder parameters are optimized jointly under an MSE distortion objective.
  • Entropy encoding is lossless and is therefore excluded from the subsequent analysis.
  • CsiNet+ places the encoder at the UE and the decoder at the BS, reconstructing CSI from received compressive measurements.

III. CSI COMPRESSION BASED ON CONVOLUTIONAL NEURAL NETWORKS

CsiNet+ modifies CsiNet’s convolutional kernels and refinement process to better handle block-sparse CSI and produce a refined reconstruction through end-to-end training.

  • The encoder compresses convolutionally extracted CSI features with a fully connected layer, while the decoder reshapes a fully connected output and refines it with RefineNet blocks.
  • CsiNet+ introduces two main modifications to CsiNet: convolutional kernel size and refinement process.
  • 1) Modification 1:: 3 × 3 filters can produce mostly near-zero outputs on block-sparse CSI, extracting less information when their receptive fields fall in blank regions.
  • 1) Modification 1:: The fully connected layer cannot expose sparsity effectively, while convolutional operations may be futile when coefficients remain zero after filtering.
  • 1) Modification 1:: CsiNet-M1 replaces selected 3 × 3 refinement and encoder convolutions with larger 7 × 7 and 5 × 5 filters to better cover sparse structures.
  • 2) Modification 2:: CsiNet+ removes the convolutional layer after the last RefineNet block so the final output remains the refined estimate.
  • 2) Modification 2:: CsiNet-M2 adds a convolutional layer between the decoder’s fully connected layer and first RefineNet block, improving the initial estimate during end-to-end training.

B. Quantization and Dequantization

The bit-level CsiNet+ pipeline quantizes encoder outputs for feedback, dequantizes them at the BS, and uses neural networks to reduce quantization distortion. Its training strategy avoids separate UE parameter sets across quantization rates by fixing the encoder and training rate-specific BS decoders.

  • Quantization and dequantization: The UE encoder output is quantized into a feedback bitstream, while the BS dequantizes it before neural-network reconstruction.The dequantized measurement vectors are then processed by the decoder-side networks.
  • Quantization method: The method adopts a µ-law non-uniform quantizer because uniform quantization is not optimal for compressed CSI.The companding function uses x ∈ [−1, 1] and a constant µ controlling companding amplitude.
  • Offset module: An offset neural network based on residual learning is applied after dequantization to minimize quantization distortion.The network uses three fully connected layers with N neurons.
  • Training challenge: Because quantization is non-differentiable, directly backpropagating through the complete bit-level network is infeasible.Using a constant quantization gradient enables end-to-end training but restricts a network to specific quantization bits.
  • Training strategy: The proposed training first optimizes unquantized CsiNet+, then trains the offset network and fine-tunes the decoder while fixing encoder parameters.Distinct decoders are trained at the BS for different quantization bit rates, so the UE stores one parameter set regardless of rate.

IV. MULTIPLE-RATE CSI FEEDBACK

Multiple-rate CSI feedback addresses the need to adapt compression ratios without storing a separate large neural network at the UE. The proposed designs reduce encoder parameters by reusing fully connected layers and provide series and parallel alternatives.

  • Motivation: Existing deep-learning CSI feedback methods generally support a fixed compression ratio and require a different stored network for each ratio.This storage burden is problematic at the UE, where storage is limited.
  • Framework design: The proposed multiple-rate framework supports different compression ratios while saving UE storage space.Decoder parameter counts are neglected because the BS is assumed to have sufficient storage.
  • CsiNet+ encoder: The CsiNet+ encoder uses two convolutional layers, two batch-normalization layers, and one fully connected layer.The parameter-count formulas distinguish convolutional, batch-normalization, and fully connected layers.
  • Parameter allocation: 1,048,576 fully connected parameters constitute 99.9% of the 1,049,500 parameters in the fourfold-compression encoder.This concentration motivates reusing fully connected layers in multiple-rate designs.
  • Proposed frameworks: The paper introduces series SM-CsiNet+ and parallel PM-CsiNet+ frameworks for multiple-rate CSI compression.Both designs target parameter reduction through reuse, but they organize compression measurements differently.

A. Series Multiple-Rate Compression Framework: SM-CsiNet+

SM-CsiNet+ generates higher-compression measurement vectors from lower-compression ones and reuses encoder components across rates. It reduces UE parameters by approximately 38.0% relative to separate encoders, while PM-CsiNet+ provides a larger reduction through parallel reuse.

  • SM-CsiNet+ design: SM-CsiNet+ generates highly compressed measurement vectors by continuing compression of lower-compression vectors.For example, a fourfold-compressed CSI representation can be compressed twofold again to obtain eightfold compression.
  • Parameter reuse: The series framework shares feature-extracting convolutional layers and reuses fully connected layers across compression rates.For eightfold compression, the fully connected layer changes from 2048×256 to 512×256 compared with direct compression from the original CSI.
  • SM-CsiNet+ result: 38.0%: SM-CsiNet+ reduces UE parameters from 1,968,688 with separate encoders to 1,221,532.The reduction is reported to increase when more compression ratios must be realized.
  • PM-CsiNet+ result: 46.7%: PM-CsiNet+ reduces UE parameter number while retaining the fourfold encoder’s 1,049,500 parameters.The framework is reported to use the same parameter number as the fourfold compression encoder and save UE storage space.

V. SIMULATION RESULTS AND DISCUSSIONS

The simulations evaluate reconstruction accuracy, quantization, multiple-rate frameworks, comparisons with CsiNet, and parameter-based interpretations using indoor and outdoor COST 2100 CSI data. Replacing small filters with 7 × 7 filters improves CsiNet-M1 over CsiNet in both scenarios, with a smaller outdoor gain.

  • Evaluation scope: The experiments assess reconstruction accuracy, quantization, comparisons with CsiNet, multiple-rate frameworks, and the compression–reconstruction mechanism.Parameter visualization is used to analyze the fully connected layer at the UE.
  • Experimental setup: The evaluation uses the COST 2100 MIMO channel model with indoor and outdoor rural CSI scenarios.The carrier frequencies are 5.3 GHz indoors and 300 MHz outdoors, with 1024 subcarriers and 32 BS antennas.
  • Data preparation: The complex CSI is represented in the angular-delay domain and truncated to 32 × 32 before evaluation.The generated data are randomly divided into training, validation, and testing sets.
  • Reconstruction results: CsiNet-M1 outperforms CsiNet in both indoor and outdoor scenarios after replacing small convolutional kernels with 7 × 7 filters.The modification is intended to use CSI block sparsity in the angular-delay domain.
  • Scenario comparison: The error reduction from CsiNet-M1 is much smaller outdoors than indoors.The paper attributes this difference to the outdoor CSI matrix having a smaller blank area, where large kernels mainly operate.

2) Effect of Modification 2:

Modification 2 improves CsiNet+ reconstruction, with gains over CsiNet and CsiNet-M1, while benefits diminish at higher compression ratios. CsiNet+ also improves accuracy across compression ratios but incurs a modest processing-time increase that remains practically acceptable.

  • Effect of Modification 2: CsiNet-M2 outperforms both CsiNet and CsiNet-M1, demonstrating improved refinement performance.The improvement decreases as the compression ratio increases because high-ratio information loss cannot be fully offset by refinement.
  • Comprehensive performance of CsiNet+: CsiNet+ shows noticeable accuracy advantages over CsiNet under all compression ratios, especially at small ratios, with only a slight parameter increase.The comparison is reported using reconstruction accuracy and parameter number.
  • Complexity comparison: 0.12 ms is CsiNet+’s processing time at fourfold compression on a 1080Ti GPU, compared with 0.07 ms for CsiNet.The increase is attributed to higher FLOPs from larger convolutional kernels, but the processing time still meets practical CSI reconstruction requirements.
  • Robustness to channel noise: High compression is more sensitive to uplink SNR than low compression in both indoor and outdoor scenarios.Below 5 dB uplink SNR, the compression ratio has little effect on reconstruction accuracy.

5) Neural network capacity:

The paper examines whether network capacity and learned CSI structure support reconstruction across scenarios and compression ratios. Visualization indicates that CsiNet+ exploits sparse, information-rich regions and learns statistical CSI, while outdoor CSI remains more difficult because it is less sparse.

  • Multiple scenarios: CsiNet+ handles indoor and outdoor CSI with a single model, but joint training substantially reduces indoor reconstruction accuracy.Outdoor CSI changes little, while the larger outdoor loss dominates the MSE objective.
  • Network capacity: CsiNet+ uses wider convolutional layers and greater depth to provide the expressivity needed for different scenarios.The paper relates this capacity to findings that one denoising CNN can support multiple image-restoration tasks.
  • Parameter visualization: FC-layer heatmaps show strong attention to the bottoms of feature maps, while near-zero upper regions contain little information.The visualization procedure reshapes FC parameters and averages absolute values across one axis to form heatmaps.
  • Compression behavior: As compression increases, attention concentrates on smaller regions and reconstruction accuracy drops because incidental information is lost.The paper reports lower outdoor accuracy than indoor accuracy for both CsiNet and CsiNet+.
  • Scenario differences: Outdoor CSI is harder to extract efficiently because it exhibits much less sparsity than indoor CSI.This difference is visible in the indoor and outdoor FC-layer heatmaps.
  • Statistical CSI: CsiNet+ automatically learns statistical CSI in addition to sparse CSI structure.With an all-zero input, its output resembles the training-set mean, including blank upper regions and information-rich bottoms.

C. Quantization Evaluation

The quantization evaluation studies bit allocation and multiple-rate compression under indoor and outdoor conditions. Non-uniform quantization with an offset network performs best among quantization methods, while compression and quantization errors affect the two scenarios differently.

  • Quantization methods: Non-uniform quantization outperforms uniform quantization, and the offset network achieves the best quantization performance.The offset network uses stacked FC layers to refine dequantizer output.
  • Quantization bits: Increasing quantization bits improves reconstruction accuracy, with CsiNet+Q6 approaching the performance of unquantized CsiNet+.The indoor scenario is reported as more sensitive to quantization-bit changes than the outdoor scenario.
  • Feedback-bit allocation: At fixed feedback length, indoor reconstruction is more sensitive to quantization distortion, whereas outdoor reconstruction is more sensitive to compression errors.The paper uses this contrast to motivate selecting compression ratio and quantization bits jointly.
  • Trade-offs: Multiple-rate frameworks are not lossless relative to direct compression, especially when fourfold measurements lack sufficient useful information for higher compression.Increasing the weight of one sub-network can emphasize a preferred compression ratio but reduces other sub-networks’ performance.
  • Multiple-rate frameworks: The series framework can outperform direct compression at indoor compression ratios 4 and 8, but is approximately 0.5 dB worse in other cases.The parallel framework is approximately 2 dB worse than the direct one.
  • Framework comparison: SM-CsiNet+ outperforms PM-CsiNet+, while PM-CsiNet+ uses approximately 85.9% of SM-CsiNet+’s parameters.The trade-off favors lower parameter count in the parallel framework and higher reconstruction performance in the series framework.
  • Series-framework behavior: SM-CsiNet+ can outperform CsiNet+ at indoor compression ratios 4 and 8 because subsequent compression networks act as an additional regularization effect during training.At fourfold compression, the testing architecture and parameter count are the same, so the reported gain is attributed to training.

VI. CONCLUSION

The paper proposes CsiNet+, a quantization framework, and two multiple-rate frameworks for practical massive-MIMO CSI feedback. The methods improve reconstruction or reduce UE parameter storage, while parameter visualization explains how the networks exploit CSI structure and statistics.

  • CsiNet+ architecture and analysis: CsiNet+ modifies CsiNet using CSI sparsity and refinement principles, while visualization examines its compression mechanism and statistical-CSI learning.The paper presents this analysis as addressing the black-box character of prior DL-based methods.
  • Quantization framework: The proposed quantization framework and training strategy avoid UE parameter updates for different quantization bits by fine-tuning decoder parameters at the BS.An offset network is introduced to reduce quantization distortion through refinement.
  • Multiple-rate frameworks: SM-CsiNet+ and PM-CsiNet+ address the storage waste caused by maintaining different parameter sets for different compression ratios.They reduce parameter number by 38.0% and 46.7%, respectively, relative to the existing approach.
  • Overall comparison: SM-CsiNet+ outperforms PM-CsiNet+ at the cost of a larger parameter number and can outperform CsiNet+ at low compression ratios.The comparison identifies a performance-versus-parameter trade-off between the two variable-rate frameworks.
Loading 1906.06007v1…