Mixup

Mixup studies how training on blends of examples, features, or labels can regularize models and help them generalize when data are limited, noisy, or shifted. Common approaches include input-level and manifold mixup, conditional and neighborhood-based mixing, and saliency-guided variants, used in settings such as image classification, sound detection, medical imaging, and protein learning. Researchers evaluate these methods on standard accuracy and corruption benchmarks, while also measuring calibration, robustness, and performance under label noise or data scarcity. A central tension is that stronger mixing can improve generalization and uncertainty estimates but may create unrealistic samples or blur task-relevant structure, especially for ordinal and spatial data.