Uncertainty-guided label correction with wavelet-transformed discriminative representation enhancement

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Label noise, including closed-set and open-set types, degrades model performance and generalization.
• Existing methods struggle with mixed noise, often oversimplifying or discarding valuable open-set noise data.
• Robust noise identification is crucial for real-world machine learning applications.

Purpose of the Study:

• To develop a novel approach, Ultra, for mitigating the effects of mixed label noise in machine learning models.
• To address limitations of existing methods by handling both closed-set and open-set noise effectively.
• To enhance the practical applicability of machine learning models in noisy data environments.

Main Methods:

• Utilizes a learnable wavelet filter for discriminative feature extraction and spurious cue filtering at the representation level.
• Employs a two-fold uncertainty estimation to accurately locate noise within the supervised signal.
• Develops a label correction technique that leverages open-set noise constructively.

Main Results:

• Ultra demonstrates effectiveness and generality in identifying and correcting mixed label noise across synthetic, web-scraped, and real-world datasets.
• The approach successfully mitigates the negative impact of both closed-set and open-set noise.
• Ultra enhances the performance of efficient techniques like supervised contrastive learning in noisy scenarios.

Conclusions:

• Ultra provides a robust and effective solution for mixed label noise identification and correction.
• The method offers a practical way to utilize open-set noise, previously considered detrimental.
• This work advances the field of robust machine learning by improving model generalization under realistic noisy conditions.