Wasserstein Adversarial Regularization for Learning With Label Noise.

Area of Science:

• Computer Science
• Machine Learning
• Computer Vision

Background:

• Noisy labels are a common issue in vision datasets, particularly those sourced from crowdsourcing or web scraping.
• This noise can significantly degrade the performance of machine learning classifiers.
• Developing methods to learn robust classifiers in the presence of such noise is crucial for practical applications.

Purpose of the Study:

• To propose a new regularization method for learning robust classifiers from noisy vision datasets.
• To introduce Wasserstein Adversarial Regularization (WAR) as a solution to mitigate the impact of noisy labels.
• To leverage the geometric properties of the label space for effective regularization.

Main Methods:

• The study proposes a novel adversarial regularization scheme based on the Wasserstein distance.
• This approach, termed Wasserstein Adversarial Regularization (WAR), utilizes the geometric properties of the label space.
• WAR implements selective regularization, promoting classifier smoothness between certain classes while maintaining decision boundary complexity for others.

Main Results:

• The effectiveness of adversarial regularization in handling noisy data is discussed.
• WAR was evaluated on five datasets corrupted with noisy labels.
• The method demonstrated superior performance compared to state-of-the-art competitors on both benchmark and real-world datasets.

Conclusions:

• Wasserstein Adversarial Regularization (WAR) is an effective technique for building robust classifiers in the presence of noisy labels.
• The method's ability to selectively regularize based on class relationships offers an advantage over existing approaches.
• WAR shows significant promise for improving the reliability of machine learning models in real-world scenarios with imperfect data.