Fig. 8

Visualizing how \(\zeta \)-mixup improves performance over mixup. Sample images from two skin lesion datasets with different imaging modalities: ISIC 2017 and derm7point. Sample test images from both datasets that were misclassified by mixup-augmented models (a), when embedded in a 2D space for t-SNE visualization, show that they lie in the vicinity of training samples from classes different from the test images’ labels, leading to wrong predictions (b, d). On the other hand, with \(\zeta \)-mixup-augmented models, the test images are more likely to be in a region of training samples from the same class as that of the test images (c, e)