Deep Unfolding with Kernel-based Quantization in MIMO Detection

TL;DR

This paper introduces a kernel-based adaptive quantization framework for deep unfolding MIMO detection models, improving accuracy and reducing latency on resource-constrained edge devices by aligning activation distributions without prior assumptions.

Contribution

The paper proposes a novel KAQ framework that uses KDE and MMD for distribution alignment and dynamic step size adjustment, addressing limitations of existing quantization methods.

Findings

KAQ outperforms traditional quantization methods in accuracy.

KAQ reduces inference latency in MIMO detection models.

The approach eliminates the need for prior distribution assumptions.

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.