.png){kind=icon}
返回
Diffusion-facilitated knowledge distillation in human activity recognition
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.132517.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Human Activity Recognition (HAR) has witnessed significant advancements with the proliferation of deep learn ing techniques, which automate feature extraction from sensor data. However, high computational demands hinder deployment in resource-constrained environments, where Knowledge Distillation (KD) offers a solution by transferring knowledge from large models to smaller ones. In exploring KD for HAR, we empirically observe that excessive noise in student activity features consistently disrupts the accuracy of logits representation. Inspired by the denoising process of diffusion models, which iteratively refine data toward the original distribution, we propose Diffusion-Facilitated Knowledge Distillation (DFKD)-a novel framework that enhances KD for HAR by denoising student features while reducing computational overhead via autoencoder-assisted feature compression. DFKD leverages diffusion models to improve knowledge transfer robustness, ensuring better alignment between teacher and student models. Extensive experiments on benchmark datasets (UCI-HAR, UNIMIB-SHAR, PAMAP2) demonstrate that DFKD outperforms state-of-the-art methods in both accuracy and efficiency. Additionally, real-world evaluations on resource-constrained devices, including inference speed tests on a Raspberry Pi 5, highlight its practical benefits.
Keyword:
Human activity recognition
Deep learning
Knowledge distillation
Diffusion model
Model compression
期刊
IF:
6.5
论文数:
2.5W
被引数:
6.5W
机构
引用论文
.png){kind=icon}
分享收藏分享收藏分享收藏
IoT Wearable Sensor and Deep Learning: An Integrated Approach for Personalized Human Activity Recognition in a Smart Home Environment物联网可穿戴传感器和深度学习: 智能家居环境中个性化人类活动识别的集成方法
PREAI
分享收藏
查看更多内容