Improving the Instance-Dependent Transition Matrix Estimation by Exploiting Self-Supervised Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Learning with noisy labels is crucial in real-world applications.
• Estimating the instance-dependent transition matrix is challenging due to the absence of clean labels.
• Self-supervised learning shows promise in inferring clean label information without supervision.

Purpose of the Study:

• To develop a practical approach for accurate transition matrix estimation using self-supervised learning.
• To reduce the estimation error of instance-dependent transition matrices.
• To improve the performance of classifiers trained with noisy labels.

Main Methods:

• Harnessing self-supervised learning to extract clean label information.
• Utilizing extracted clean label information to estimate the instance-dependent transition matrix.
• Employing the estimated transition matrix to enhance classifier performance.

Main Results:

• The proposed methodology significantly reduces the estimation error of the transition matrix.
• Improved classification accuracy is achieved by exploiting the estimated transition matrix.
• Empirical results demonstrate superior performance compared to existing state-of-the-art methods.

Conclusions:

• Self-supervised learning is effective in inferring clean label information for noisy label learning.
• The proposed approach offers a practical and effective solution for transition matrix estimation.
• This work advances the field of learning with noisy labels by improving both matrix estimation and classification accuracy.