Data Augmentation

What is data augmentation?

The thing you're always short of when training a model is data — and labeling is slow and expensive. Data augmentation has a simple idea: can each sample you already have become many? Flip the image, reword the sentence, add some noise to the audio, and you get a batch of new samples that are "same underneath, different on the surface". The model sees more and learns more solidly.

Why does it help?

It stops rote memorization
With too little data, the model just memorizes the training examples and falls apart on anything new. Varied augmented samples force it to learn patterns that actually generalize.
It makes the model tougher
After seeing distortions, noise and occlusion, the model resists real-world interference much better.
More for less
You don't pay to label new data. A little compute multiplies the size of your training set several times over.

Common augmentation techniques

Images
Flip, rotate, crop, recolor, add noise, blur — one cat photo becomes dozens.
Text
Synonym swapping, back-translation (translate out and back), random insertion or deletion, or asking a model to rewrite.
Audio
Speed changes, background noise, volume shifts — simulating different recording conditions.

What are the pitfalls?

More isn't always better. Overdo it and the label stops being true — rotate a "6" by 180 degrees and you get a "9". So the strength of augmentation has to match the task, and the new sample's label still has to be correct.

Bottom line: data augmentation turns one sample into several, cheaply giving the model wider experience and better resistance to noise.