Noise-Aware Framework for Correcting Corrupted Labels

The paper introduces CANOLA, a framework designed for correcting corrupted labels in datasets through noise-aware learning and iterative label refinement. CANOLA estimates the underlying noise distribution and integrates it into the training of a noise-aware Deep Neural Network, allowing it to down-weight unreliable labels and enhance robustness. Experimental results demonstrate that CANOLA outperforms existing state-of-the-art label correction methods, achieving error reductions of 19% to 52%, and significantly improves downstream performance for models trained on its corrected datasets.