In few-shot learning, instead of manually labeling selected data, this method augments training data by selecting samples from a large weakly labeled or unlabeled dataset given a label. The resulting training sample can be written as $(T(x_i), T(x_i))$. For example, a presentation video contains many unlabeled pictures of the speaker's gestures. This data augmentation method uses another algorithm to learn which picture can be used for training for each label class.