Multi-Round Meta-Learning for Label Purification without Trusted Labels

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Label noise is a significant hurdle in supervised deep learning, stemming from large-scale dataset annotation errors.
• Existing methods often require fully noisy data or a clean subset, leaving clean-label-free purification underexplored.
• Expert-annotated domains like remote sensing and medical imaging face persistent annotation inconsistencies.

Purpose of the Study:

• To propose a novel clean-label-free meta-learning framework for effective label purification.
• To develop a robust method for improving label quality without relying on clean data subsets.
• To enhance the reliability of deep learning models in domains prone to annotation errors.

Main Methods:

• A multi-round purification strategy iteratively refines label quality by decoupling base-model training and meta-purification.
• Stochastic contraction analysis and a KL-divergence-based stopping criterion adaptively determine purification depth.
• The framework enables flexible downstream training and uncertainty capture through accumulated label distributions.

Main Results:

• The proposed framework consistently outperforms state-of-the-art techniques on benchmark datasets (CIFAR, mini Clothing-1M).
• Achieved up to 62.61% label accuracy improvement on expert-annotated remote sensing datasets (UC Merced, AID).
• Demonstrated robustness in real-world, noise-prone environments and improved mislabeled instance detection.

Conclusions:

• The clean-label-free meta-learning framework offers a stable and effective solution for label purification.
• The method enhances deep learning model performance and reliability, particularly in expert-annotated domains.
• This approach provides a flexible alternative to complex noisy-label learning pipelines.