Instead of labeling selected data, this method augments training data by selecting data from a large set given a label, and the resulting training sample can be written as $(T(x_i), T(x_i))$. This large dataset typically has little or no label. For example, a video of a presentation contains lots of pictures of the gestures of the speaker, however, none of them are labeled. This data augmentation method uses another algorithm to learn which picture can be used for training for each label class.