An improved sample selection framework for learning with noisy labels.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) excel at learning but are susceptible to overfitting when trained with noisy labels.
• Existing sample selection methods filter potentially clean labels but create a large gap between labeled and unlabeled data, especially at high noise rates.
• This underutilization of unlabeled data limits performance improvements in noisy label learning.

Purpose of the Study:

• To introduce an enhanced sample selection framework with an oversampling strategy (SOS) to address the limitations of current methods.
• To leverage information from label-free instances to improve model performance in the presence of noisy labels.
• To enhance state-of-the-art sample selection techniques by integrating the SOS.

Main Methods:

• Development of a novel oversampling strategy (SOS) integrated into a sample selection framework.
• Combining SOS with existing state-of-the-art sample selection methods.
• Extensive experimental validation on synthetic and real-world noisy datasets (CIFAR, WebVision, Clothing1M).

Main Results:

• The proposed SOS framework effectively utilizes information from label-free instances.
• Significant performance improvements in classification and generalization were observed compared to baseline methods.
• SOS demonstrates robustness across various datasets with different noise levels.

Conclusions:

• The SOS framework offers a promising solution for mitigating the negative impact of noisy labels in deep learning.
• Leveraging unlabeled data through oversampling is crucial for enhancing model robustness and performance.
• The study provides a practical approach and open-source code for improving deep learning on noisy datasets.