Data Augmentation

Data augmentation expands the training set with label-preserving transformations. Geometric: horizontal flip, random crop, rotation, scaling, color jitter (exploiting translation/flip invariance priors). Mixing: Mixup linearly interpolates samples $x' = \lambda x_i + (1-\lambda) x_j$ with $y' = \lambda y_i + (1-\lambda) y_j$, where $\lambda \sim \text{Beta}(\alpha, \alpha)$; CutMix mixes image regions instead of pixels. Augmentation acts as regularization: it implicitly forces the model to be smooth over transformed versions and sample neighborhoods, suppressing overfitting — effectively an extra prior/smoothness penalty on the loss.

label-scarce CV tasks (classification/detection/segmentation), self-supervised learning (SimCLR builds positive pairs via random augmentation), semi-supervised consistency training; interviews ask why augmentation works and how it relates to regularization.

labeled data is expensive; augmentation expands data at near-zero cost — flips/crops inject invariance priors, while Mixup/CutMix smooth the decision boundary by forcing linear transitions between samples (an implicit smoothness regularizer). Together with dropout and weight decay it forms the regularization toolbox and composes well. Side effects: over-aggressive transforms break semantics (invalid labels) or slow training, so transforms must match the task.

Write the Mixup formula. Why does Mixup act as regularization?

Why do random crops/flips work? Which image priors do they exploit?

How does augmentation relate to dropout and weight decay as regularizers?

How does augmentation construct positive/negative pairs in self-supervised learning (SimCLR)?

What are the side effects of over-aggressive augmentation, and how to avoid them?