BadLabel: A Robust Perspective on Evaluating and Enhancing Label-Noise Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Label-noise learning (LNL) addresses challenges in training models with inaccurate data.
• Existing LNL algorithms are susceptible to various noise types, including class-conditional and instance-dependent noise.
• A significant performance degradation is observed with novel, sophisticated label noise patterns.

Purpose of the Study:

• Introduce a new, challenging label noise type named BadLabel.
• Develop a robust LNL method to mitigate the impact of BadLabel and other noise types.
• Enhance model generalization capabilities in the presence of noisy training data.

Main Methods:

• Crafting BadLabel by flipping specific samples' labels to create indistinguishable loss values between clean and noisy data.
• Proposing a robust LNL method involving adversarial label perturbation during training.
• Utilizing semi-supervised learning techniques on a small set of clean labeled data post-noise mitigation.

Main Results:

• Demonstrated the vulnerability of current LNL algorithms to the BadLabel noise type.
• Showcased the effectiveness of the proposed robust LNL method in improving generalization across various noise types.
• Validated the approach through empirical experiments on newly generated noisy datasets.

Conclusions:

• BadLabel poses a significant threat to existing label-noise learning algorithms.
• The proposed adversarial label perturbation method offers a robust solution for LNL.
• The developed techniques enhance model performance and generalization in real-world noisy data scenarios.