data augmentation

Applying label-preserving transformations to training images to synthetically expand a dataset and reduce overfitting.

Data augmentation applies random, label-preserving transformations -- crops, flips, rotations, color jitter -- to training images on the fly, synthetically expanding the effective size and diversity of a training set without collecting new data, and directly fighting overfitting by denying the model the chance to memorize any one exact pixel arrangement.

Examples

Randomly flipping, rotating, and color-jittering a training set of 5,000 photos every epoch teaches a model that a cat is still a cat mirrored or slightly recolored.