Simplified Transition Matrix Method for Noisy Label Learning: Efficiency and Effectiveness

This paper presents a simplified approach to noisy label learning that focuses on estimating a global transition matrix rather than individual matrices for each sample. This method, combined with implicit regularization, aims to address the challenges of instance-dependent noise and computational expense associated with traditional transition matrix methods. By estimating a global matrix, the model captures the overall probability transfer from correct to noisy labels. Implicit regularization is then utilized to adjust the representation of the difference between the estimated posterior probability distribution and the noisy label distribution, promoting sparsity. This approach demonstrates adaptability to various types of noise and mitigates the issue of inaccurate posterior probability estimation. Theoretical analysis confirms the consistency and effectiveness of this method, which is further validated through experiments on synthetic and real-world datasets with diverse label noise types. The results highlight the significant performance advantage of this simplified method over existing transition matrix methods and its wider applicability. Notably, it achieves competitive results without the need for auxiliary techniques, showcasing its efficiency. The code for this method will be made publicly available on Github.