Introduces mixup, a data augmentation technique that trains on convex combinations of input pairs and their labels. Simple, data-independent, and model-agnostic approach that improves generalization and robustness.