Data augmentation

Data augmentation multiplies your training data by transforming existing examples in ways that change the pixels or words but not the label.

A cricket coach with one bowler still varies the practice: over the wicket, round the wicket, old ball, new ball, damp pitch. Same bowler — but the batsman learns to handle the delivery, not that one bowler's exact action. Augmentation gives a model the same varied practice from limited material.

For images the transforms are physical common sense: flip horizontally, rotate slightly, crop, shift brightness, add noise. A mango photographed from the left is still a mango, and the model should think so too — augmentation teaches it that these variations do not matter. For text: swap synonyms, paraphrase, back-translate (English → Hindi → English gives a natural rewording). For audio: speed changes, background noise, pitch shifts.

The label-preservation rule is the entire craft. Flip a mango photo: still a mango. Flip a photo of the digit 6 and you have made a 9 with a 6's label — actively poisoning your data. Every transform must be checked against the task: horizontal flips are fine for animals, wrong for road signs and text.

Augmentation acts as regularization — often the single most effective cure for overfitting on small image datasets. Its bigger sibling, generating entirely new examples from scratch, is synthetic-data.

Where to go next#

• Full lesson: Image classification
• Related terms: overfitting, regularization, synthetic-data, training-data