Data Augmentation
A technique that creates extra training examples by making small changes to existing ones, such as flipping or cropping an image, to help a model generalise.
What is data augmentation?
Data augmentation expands a training set by creating slightly modified copies of the examples it already has. It is mainly used to reduce overfitting, where a model memorises its training data instead of learning patterns that carry over to new data.
How does it work?
For images, common changes include rotating, flipping, cropping or scaling the picture, adjusting brightness or contrast, and adding a little noise. The label stays the same, so one labelled photo becomes several. Other kinds of data have their own methods, such as oversampling rare classes in tabular data.
What is a simple example?
A model learns to recognise cats from 1,000 labelled photos. Flipping and slightly rotating each one gives it several times as many examples, so it is less likely to fail on a cat photographed from a different angle.
Why does it matter?
Collecting and labelling real data is slow and costly. Augmentation stretches the data you have. It differs from synthetic data, which generates entirely new examples rather than varying existing ones.