data augmentation
/ DAY-tuh awg-men-TAY-shun /
Data augmentation means creating extra training examples by applying small, label-preserving tweaks to the ones you already have. A photo of a cat is still a cat if you flip it left-to-right, rotate it a few degrees, brighten it, or zoom in slightly. By showing the model all these variations, you teach it that a cat is a cat regardless of pose or lighting — without paying anyone to label a single new image.
The key constraint is that the transformation must not change the right answer. For images: flips, rotations, crops, color shifts, added noise. For audio: changing speed or pitch, mixing in background sounds. For text: swapping words for synonyms or back-translating through another language. Each tweak gives the model another honest example of the same underlying thing, drawn from the variety it will meet in the real world.
Why it matters: augmentation is one of the most effective and cheapest ways to fight overfitting and improve generalization, especially when labeled data is scarce. It works because it bakes in invariances you know to be true (a sideways cat is still a cat) so the model stops fixating on irrelevant accidents. But it has limits: it can only stretch the data you have, not invent genuinely new situations, and a careless transformation can quietly flip the label — rotating a photo of the digit 6 by 180 degrees turns it into a 9, teaching the model something false.
From one labeled photo of a stop sign, augmentation generates a dozen training images: slightly rotated, partly shadowed, seen from an angle, a bit blurry, brighter, darker. The self-driving model learns to recognize a stop sign in all the messy conditions of a real street, not just the one pristine snapshot.
One label, many honest variations — teaching the model what stays the same.
Augmentation must preserve the label. A rotation that turns a 6 into a 9, or a color shift that recolors a ripe tomato green, teaches the model a falsehood — always check that your tweaks still leave the right answer right.