Researchers have considered using generative networks for data augmentation or assigning pseudo-labels to high-confidence samples to expand the labeled training set. Some researchers have also used labeled and unlabeled datasets to combine supervised and semisupervised training across AL cycles.