Learning to Transfer Automatic Data Augmentation Policies Using Reinforcement Learning

Up to date, several data augmentation methods have been proposed to overcome insufficient annotated data problems in deep learning, while these methods usually require human experts to manually tune the augmentation. Therefore, there might need a method to automatically learn the optimal data augmentation policies from the given datasets. In this paper, we propose a novel algorithm, namely TransADA, for automatic data augmentation via reinforcement learning and transfer learning. Given a source dataset with rich features, TransADA intends to first learn the optimal data augmentation policy through reinforcement learning. Then, with a target dataset, TransADA can transfer the policy to the target dataset via policy-oriented transfer learning. Blessed by the data augmentation policies learned from the feature-rich source dataset, one can enjoy additional performance improvement through the policy transfer offered by TransADA. Extensive experiments have been conducted using real-world datasets, such as CIFAR-IO, MNIST, and SVHN datasets with a wide range of learning models. The results and comparisons demonstrate the performance boosts of TransADA when compared to the state of the arts.