In self-supervised learning with data augmentation, artificially altered versions of input data are generated to train a model that learns robust representations.
Typical augmentations are image transformations such as rotation, scaling, cropping, or color changes.
The model is trained to produce similar representations for different augmentations of the same original, while learning differences from other data.
This leads to better generalization, as the model learns to extract essential features that are invariant to the augmentations.
Such methods form the basis of many modern self-supervised learning approaches in image and speech processing.