Making more from less
Data augmentation creates new training examples by applying transformations that change the input but keep the label valid. It is a cheap way to grow data and teach the model what variations it should ignore.
Common transforms
- For images, flips, crops, rotations, and color jitter produce new views of the same object.
- For text, synonym swaps, back translation, and random deletion produce paraphrases.
- For audio, time stretch, pitch shift, and added noise simulate recording conditions.
The label preserving rule
- The transform must not change the correct answer. Flipping a digit six into a nine breaks the label, so not every transform is safe for every task.
- Choosing safe transforms requires knowing what invariances the task actually has.
Why it helps
- Augmentation acts as a regularizer, forcing the model to rely on stable features rather than memorizing exact pixels or tokens.
- It improves robustness to the kinds of variation seen at deployment.
A caution
- Too aggressive augmentation can push examples off the real data distribution and hurt, so the strength is a tuning knob.
Key idea
Data augmentation applies label preserving transforms to grow data and teach invariances, but the transforms must match the task and not overpower the signal.