Data augmentation is the practice of generating new training data from existing data by applying label-preserving transformations. In computer vision, this often means transforming images by rotating, flipping, scaling, cropping, adjusting brightness/contrast, or adding noise. 

For example, an image of a cat flipped horizontally is still a cat, providing a new training example. In NLP, augmentation can include synonym replacement or back-translation of sentences. Augmentation helps increase the diversity of the training set, combatting overfitting especially when original data is limited. It effectively teaches the model invariances (like a rotated object is the same object) and can significantly improve generalization performance without needing new labeled data.