← AI Terminology
Mixup
Mixup is a data-augmentation regulariser that trains on convex combinations of pairs of examples and their labels: x̃ = λx_i+(1−λ)x_j, ỹ = λy_i+(1−λ)y_j.
It improves generalisation and robustness in vision and beyond.
It improves generalisation and robustness in vision and beyond.
Why It Matters in AI
Interpolated examples smooth decision boundaries and reduce memorisation. Mixup and cousins (CutMix) are standard in strong vision training and appear in some speech/NLP variants.
Key Points
| Aspect | Description |
|---|---|
| λ | Often Beta(α,α) sampled per pair |
| Cost | Trivial compute; implementation care with multi-label |
| Effect | Regularisation; better robustness |
| Origin | Zhang et al., 2018 |
| Related | Label smoothing, data augmentation |
| Variants | CutMix, Manifold Mixup, input mix for speech |
Simple Analogy
Studying using blended flashcards — half cat photo + half dog photo with a blended label — so the model learns smoother distinctions.
Common Usage Examples
- Sample λ ~ Beta(0.2,0.2); mix images and labels
- CutMix for regional mixing
- CIFAR/ImageNet strong baselines use mixup
- Combine with RandAugment
Summary
In short: Mixup trains on blended inputs and labels — a simple augmentation that smooths decision boundaries and improves generalisation.