mixup

**Mixup** is a **data augmentation technique that creates new training samples by linearly interpolating between pairs of existing samples and their labels** — encouraging the model to learn smooth, linear decision boundaries between classes. **How Does Mixup Work?** - **Sample**: Draw mixing coefficient $lambda sim ext{Beta}(alpha, alpha)$ (typically $alpha = 0.2$). - **Mix Inputs**: $ ilde{x} = lambda x_i + (1-lambda) x_j$. - **Mix Labels**: $ ilde{y} = lambda y_i + (1-lambda) y_j$. - **Train**: Use $( ilde{x}, ilde{y})$ as a regular training sample with cross-entropy loss. - **Paper**: Zhang et al. (2018). **Why It Matters** - **Smoother Boundaries**: Linear interpolation encourages linear behavior between classes -> better calibration. - **Regularization**: Acts as a strong regularizer, reducing overfitting especially on small datasets. - **Universal**: Works for images, text, audio, tabular data — any domain where interpolation is meaningful. **Mixup** is **blending reality** — creating in-between examples that teach the model smooth, calibrated transitions between classes.