Class-Wise Denoising for Robust Learning Under Label Noise.

Area of Science:

• Machine Learning
• Computer Science
• Artificial Intelligence

Background:

• Label noise is a common problem in real-world datasets, degrading classifier performance.
• Existing methods often fail to leverage class-specific information when correcting noisy labels.
• Unbiased centroid estimation is crucial for robust risk minimization in noisy settings.

Purpose of the Study:

• To introduce a novel robust learning algorithm, Class-Wise Denoising (CWD), for effective label noise correction.
• To address the limitations of existing methods by handling noisy labels in a class-wise manner.
• To improve the accuracy of centroid estimation and risk minimization under label noise.

Main Methods:

• Developed Class-Wise Denoising (CWD), a novel algorithm for label noise correction.
• Constructed two virtual auxiliary sets by assuming positive and negative labels are clean, enabling separate handling of false-negative and false-positive labels.
• Designed an improved centroid estimator for more accurate risk estimation.

Main Results:

• CWD reduces the variance in centroid estimation compared to existing methods with unbiased centroid estimators.
• The performance of CWD trained on noisy data converges to that of an optimal classifier trained on clean data with a proven convergence rate.
• Empirical evaluations on benchmark datasets demonstrate CWD's superior performance against ten state-of-the-art methods.

Conclusions:

• Class-Wise Denoising (CWD) offers a significant advancement in handling label noise by leveraging class-specific information.
• The theoretical guarantees and empirical results validate CWD's effectiveness in improving classification performance under label noise.
• CWD provides a more accurate and robust approach to risk minimization in the presence of corrupted labels.