Estimating Per-Class Statistics for Label Noise Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Real-world datasets often contain noisy labels, degrading classifier performance.
• Label Noise Learning (LNL) aims to mitigate this by recovering clean data distributions.
• Existing LNL methods struggle with unreliable sample selection and multi-class generalization.

Purpose of the Study:

• To propose Per-Class Statistic Estimation (PCSE) for robust label noise learning.
• To establish a quantitative relationship between clean and noisy statistics for each class.
• To develop a generative classifier for improved model inference and performance.

Main Methods:

• PCSE utilizes centroid estimation theory to relate clean and noisy statistics.
• It establishes a quantitative relationship for first-order and second-order statistics per class.
• The method is applicable as a post-processing strategy for pre-trained networks.

Main Results:

• PCSE avoids instance-level sample selection, simplifying application.
• Theoretical analysis confirms convergence of estimated statistics to ground-truth values.
• Empirical results show PCSE outperforms state-of-the-art LNL methods in statistic estimation and classification accuracy on diverse datasets.

Conclusions:

• PCSE provides a theoretically sound and empirically validated approach to Label Noise Learning.
• The method effectively handles noisy labels in both binary and multi-class classification tasks.
• PCSE offers a generalizable post-processing technique to boost performance of existing models trained on noisy data.