Learning under label noise through few-shot human-in-the-loop refinement.

Area of Science:

• Health Informatics
• Machine Learning
• Wearable Technology

Background:

• Wearable devices continuously collect health data, but obtaining accurate labels is challenging.
• Label noise is a significant issue in wearable data analysis due to the lack of inherent visual cues.

Purpose of the Study:

• To propose a novel method, Few-Shot Human-in-the-Loop Refinement (FHLR), to address noisy label learning in wearable data.
• To enhance the generalizability and robustness of models trained on noisy wearable data.

Main Methods:

• Initial learning of a seed model using weak labels.
• Fine-tuning the seed model with a small set of expert corrections.
• Merging seed and fine-tuned models using weighted parameter averaging for improved performance.

Main Results:

• FHLR significantly outperforms eight baseline methods in learning from noisy labels.
• Achieved state-of-the-art accuracy improvements up to [Formula: see text] under symmetric and asymmetric noise.
• Demonstrated exceptional robustness to increased label noise, unlike previous approaches.

Conclusions:

• FHLR offers a robust solution for noisy label learning in health sensing with wearable technology.
• The method achieves superior generalization and robustness, outperforming existing techniques.
• Provides insights into the impact of label noise on machine learning models in healthcare.