Bayesian DivideMix++ for Enhanced Learning with Noisy Labels.

Area of Science:

• Machine Learning
• Deep Learning
• Computer Vision

Background:

• Inexpensive data annotation methods like crowdsourcing can introduce noisy labels.
• Noisy labels negatively impact deep neural network performance and generalization.
• Robust models are essential for handling noisy labels in deep learning.

Purpose of the Study:

• To address neural network memorization and uncertainty challenges in noisy label learning.
• To propose a novel framework, Bayesian DivideMix++, to enhance model robustness.

Main Methods:

• Introduced DivideMix++ for improved robustness against memorization.
• Implemented Monte-Carlo MixMatch to address label uncertainty.
• Integrated self-supervised pre-training and tailored data augmentations.
• Utilized uncertainty measurements to down-weight uncertain samples in MixMatch.

Main Results:

• Bayesian DivideMix++ demonstrated significant improvements across four diverse datasets.
• The proposed framework outperformed existing state-of-the-art models.
• Effectiveness validated through extensive experiments in synthetic and real-world noise settings.

Conclusions:

• Bayesian DivideMix++ effectively mitigates the detrimental effects of noisy labels.
• The framework enhances the performance and generalization of deep neural networks.
• Findings highlight the potential for practical application in real-world scenarios.