Correcting noisy labels via comparative distillation: a domain adaptation approach

Addressing the challenge of noisy labels in large datasets, this paper introduces a novel domain adaptation method leveraging comparative distillation model training. By transferring knowledge from a source domain model to a target domain, we construct a class correlation matrix and obtain a noise transfer matrix through network training. This matrix is used to correct noisy labels, enhancing the model’s generalization and robustness. The singular value decomposition of the noise transfer matrix constrains its training, while Tsallis entropy smooths the model’s output probabilities. Experimental results on the Office-31 and Office-Home datasets demonstrate significant improvements in classification accuracy, showcasing the effectiveness of our approach.