Knockoffs-SPR: Clean Sample Selection in Learning With Noisy Labels.

Area of Science:

• Machine Learning
• Computer Science
• Statistics

Background:

• Noisy training datasets degrade neural network generalization and robustness.
• Accurate identification of clean samples is crucial for effective model training.
• Existing methods may struggle with identifying clean data under general conditions.

Purpose of the Study:

• To propose a theoretically guaranteed framework for clean sample selection in learning with noisy labels.
• To develop a data-adaptive method that controls the False-Selection-Rate (FSR).
• To enhance the scalability and efficiency of the clean sample selection process.

Main Methods:

• Introduced Scalable Penalized Regression (SPR) to model feature-label relationships and identify clean data.
• Developed Knockoffs-SPR, a data-adaptive extension of SPR with provable FSR control.
• Implemented a split algorithm for parallel processing to achieve scalability for large datasets.
• Integrated Knockoffs-SPR with a semi-supervised algorithm to leverage noisy data as unlabeled data.

Main Results:

• SPR theoretically identifies clean data under specific conditions.
• Knockoffs-SPR effectively controls the FSR, reducing false selections of noisy data.
• The split algorithm ensures the framework's scalability to large datasets.
• Experimental results demonstrate the framework's effectiveness on benchmark and real-world noisy datasets.

Conclusions:

• The proposed Knockoffs-SPR framework offers a theoretically sound and practically effective solution for learning with noisy labels.
• The integration with semi-supervised learning further enhances the utility of noisy data.
• The developed methods significantly improve the robustness and generalization of neural networks trained on imperfect datasets.