Title: Annot-Mix: Leveraging a Mixup Extension to Learn from Noisy Labels Provided by Multiple Annotators

Abstract: The presence of noisy class labels in training data can significantly degrade the generalization capabilities of neural networks. To mitigate this, mixup has emerged as a widely adopted regularization strategy that enhances training robustness by increasing the difficulty of memorizing incorrect labels. However, conventional mixup methods overlook the reality that class labels are often provided by multiple annotators, such as crowdworkers. To address this limitation, we introduce an extended mixup technique designed to manage multiple class labels for a single instance while tracking the provenance of each label to its respective annotator. When integrated into our multi-annotator classification framework, named annot-mix, the model demonstrates superior performance compared to eleven (predominantly state-of-the-art) baseline approaches. This conclusion is drawn from an evaluation study involving eleven datasets containing noisy labels generated by either human participants or simulated annotators. The source code for this work is publicly accessible via our GitHub repository at https://github.com/ies-research/multi-annotator-machine-learning/tree/annot-mix

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC