Annot-Mix: Learning with Noisy Class Labels from Multiple Annotators via a Mixup Extension
Researchers have developed Annot-Mix, a novel extension of the mixup technique designed to improve neural network training when dealing with noisy class labels from multiple annotators. This method specifically addresses the challenge of integrating labels from various sources, unlike standard mixup which assumes single labels. Evaluations on eleven datasets demonstrated that Annot-Mix outperforms eleven other approaches, including state-of-the-art methods, in scenarios with noisy labels from both human and simulated annotators. AI
IMPACT Enhances model robustness and generalization in datasets with varied and potentially inaccurate human-generated labels.