Learning With Noisy Labels Over Imbalanced Subpopulations.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Learning with noisy labels (LNL) is a significant research area.
• Current LNL methods often fail with imbalanced subpopulations due to the 'small loss' assumption.
• This leads to misclassification of informative samples and poor generalization.

Purpose of the Study:

• To propose a novel LNL method for handling both noisy labels and imbalanced subpopulations.
• To improve the generalization performance of models in real-world scenarios with data imbalance.

Main Methods:

• Leveraging sample correlation to estimate clean label probabilities.
• Introducing a feature-based metric, considering sample correlation, for probability estimation.
• Refurbishing noisy labels using estimated probabilities and pseudo-labels.
• Employing distributionally robust optimization (DRO) with refurbished labels for robustness against subpopulation imbalance.

Main Results:

• The proposed method effectively handles noisy labels and imbalanced subpopulations simultaneously.
• It consistently improves state-of-the-art (SOTA) LNL methods, especially in imbalanced scenarios.
• Experimental results across various benchmarks validate the technique's effectiveness.

Conclusions:

• The novel LNL approach offers a robust solution for datasets with noisy labels and imbalanced subpopulations.
• This method enhances model generalization by addressing limitations of previous LNL techniques.
• The provided code facilitates further research and application of this technique.