What It Is
Artificially expanding a training dataset by applying label-preserving transformations to existing examples — crops, flips, color shifts, rotations, noise — to improve model generalization and, in self-supervised learning, to define what “same” means.
Why It Matters
Augmentation is the primary lever for controlling what invariances a model learns. In supervised learning it prevents overfitting. In contrastive self-supervised learning it defines the positive pairs — two augmented views of the same image — and determines which features the model must learn to ignore and which it must preserve.
How It Works
A transformation pipeline is applied randomly to each sample during training. The key insight from SimCLR: the composition of augmentations matters more than any individual transform. Random cropping combined with color distortion is the critical pair — cropping forces shape learning, color distortion removes the color-matching shortcut. Neither alone is sufficient.