Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Unfolding with Kernel-based Quantization in MIMO Detection

Zeyi Ren, Jingreng Lei, Yichen Jin, Ermo Hua, Qingfeng Lin, Chen Zhang, Bowen Zhou, Yik-Chung Wu

TL;DR

This work tackles energy-efficient MIMO detection for edge devices by marrying deep unfolding with kernel-based adaptive quantization (KAQ). KAQ eliminates reliance on parametric activation distributions through a joint KDE-MMD loss and introduces a dynamic, SNR-aware quantization step size, improving robustness to channel variations. The approach is implemented on PGD-Net and ADMM-Net, achieving BER close to full-precision performance while reducing inference latency by about 20% compared to full precision and outperforming traditional quantization-aware training. The results indicate KAQ's practical potential for deploying accurate, low-latency MIMO detectors in resource-constrained edge environments.

Abstract

The development of edge computing places critical demands on energy-efficient model deployment for multiple-input multiple-output (MIMO) detection tasks. Deploying deep unfolding models such as PGD-Nets and ADMM-Nets into resource-constrained edge devices using quantization methods is challenging. Existing quantization methods based on quantization aware training (QAT) suffer from performance degradation due to their reliance on parametric distribution assumption of activations and static quantization step sizes. To address these challenges, this paper proposes a novel kernel-based adaptive quantization (KAQ) framework for deep unfolding networks. By utilizing a joint kernel density estimation (KDE) and maximum mean discrepancy (MMD) approach to align activation distributions between full-precision and quantized models, the need for prior distribution assumptions is eliminated. Additionally, a dynamic step size updating method is introduced to adjust the quantization step size based on the channel conditions of wireless networks. Extensive simulations demonstrate that the accuracy of proposed KAQ framework outperforms traditional methods and successfully reduces the model's inference latency.

Deep Unfolding with Kernel-based Quantization in MIMO Detection

TL;DR

This work tackles energy-efficient MIMO detection for edge devices by marrying deep unfolding with kernel-based adaptive quantization (KAQ). KAQ eliminates reliance on parametric activation distributions through a joint KDE-MMD loss and introduces a dynamic, SNR-aware quantization step size, improving robustness to channel variations. The approach is implemented on PGD-Net and ADMM-Net, achieving BER close to full-precision performance while reducing inference latency by about 20% compared to full precision and outperforming traditional quantization-aware training. The results indicate KAQ's practical potential for deploying accurate, low-latency MIMO detectors in resource-constrained edge environments.

Abstract

The development of edge computing places critical demands on energy-efficient model deployment for multiple-input multiple-output (MIMO) detection tasks. Deploying deep unfolding models such as PGD-Nets and ADMM-Nets into resource-constrained edge devices using quantization methods is challenging. Existing quantization methods based on quantization aware training (QAT) suffer from performance degradation due to their reliance on parametric distribution assumption of activations and static quantization step sizes. To address these challenges, this paper proposes a novel kernel-based adaptive quantization (KAQ) framework for deep unfolding networks. By utilizing a joint kernel density estimation (KDE) and maximum mean discrepancy (MMD) approach to align activation distributions between full-precision and quantized models, the need for prior distribution assumptions is eliminated. Additionally, a dynamic step size updating method is introduced to adjust the quantization step size based on the channel conditions of wireless networks. Extensive simulations demonstrate that the accuracy of proposed KAQ framework outperforms traditional methods and successfully reduces the model's inference latency.
Paper Structure (13 sections, 20 equations, 6 figures, 1 algorithm)

This paper contains 13 sections, 20 equations, 6 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. System Model and Problem Formulation
  3. System Model
  4. Problem Formulation
  5. Proposed Method
  6. PGD Deep Unfolding for MIMO Detection
  7. KDE-MMD Loss
  8. Joint Optimization of Quantization Step-Size
  9. Training with Quantization-aware Loss
  10. Simulation Results
  11. Model Accuracy
  12. Inference Latency
  13. Conclusions

Figures (6)

  • Figure 1: The process of increasing the awareness of accurate activation distribution for models
  • Figure 2: Architecture of the unfolded network, with details on the feed forward process of the k-th layer and the quantization aware module.
  • Figure 3: The KAQ training scheme from full precision model to quantized version
  • Figure 4: BER versus SNR.
  • Figure 5: Model accuracy versus training epoch
  • ...and 1 more figures