Approximate Online Bilevel Optimization for Learning Data Augmentation
In this project we aim to automatically learn an augmenter network by using an approximate online bilevel optimization procedure. We plan to learn a augmenter network that generates a distribution of transformations that minimizes the loss on a validation set. By unfolding the gradients of the training loss, we will optimize the loss on validation with respect to the data augmentation parameters. In this way we can provide a general solution for an efficient and automatic data augmentation that is learned jointly with the training of the model. We expect that the proposed joint training to produce a classification performance comparable with standard data augmentation techniques, without the need of an expensive external validation loop on the hyper-parameters.