Creating extra training examples by altering the ones you have. Rotate the image, crop it, adjust the brightness, and one photograph becomes ten. It teaches the model that the label does not change when irrelevant things do, which is a cheap and effective way to improve generalisation.
Technical view
The transformation set encodes assumed invariances, and choosing them wrong actively harms performance: horizontal flips help for natural images and destroy information for text or medical laterality. Image methods include geometric and photometric transforms plus mixing strategies such as mixup and CutMix. Text augmentation is harder because most edits change meaning, so paraphrase and back-translation dominate. It functions as a regulariser and reduces overfitting on small datasets.
Related
Part of AI Terminology Index, section: Data.