Motivation

Deep Learning models are incredibly data-hungry and require very large labeled datasets for supervised learning. Thus, the supervised learning models tend to overfit to their training dataset and do not generalize well to the examples in the real world. On the other hand, the internet provides potentially infinite amounts of unlabelled data, that can be utilized to learn unsupervised models which can capture the image manifold. Our work aims to bridge this gap by leveraging language and image generation models trained on vast amounts of unlabelled internet data to dynamically augment the supervised datasets of limited size and diversity to improve the out-of-domain generalization.