An Adaptive Sample Selection Approach for Learning with Noisy Labels

To mitigate the costs of dataset construction and the challenges posed by mislabeled data during model training, learning with noisy labels (LNL) focuses on developing robust models in the presence of erroneous labels. The separation of clean and noisy samples, followed by a semi-supervised learning approach, has proven to be a vital solution to this problem. However, previous methods for selecting clean samples, such as Gaussian mixture models, have struggled to accurately identify clean samples. In this paper, we introduce a two-stage adaptive sample selection approach tailored for the LNL problem. By leveraging the disagreement between dual models, the stability of model predictions, and feature similarities, our method dynamically identifies clean samples more effectively. Experimental results demonstrate the robustness and effectiveness of our approach across various noise types and levels, confirming its superior performance.