Generalizing to unseen domains via adversarial data augmentation

• DOI: 10.48550/arxiv.1805.12018
• Publication Date: 2018-01-01
• Authors: Volpi, R.; Duchi, J.; Namkoong, H.; Murino, V.; Sener, O.; Savarese, S.
• Keywords: FOS: Computer and information sciences; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition; 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology; Domain Generalization, Data Augmentation, Adversarial Learning

Accepted to NIPS 2018 (camera ready)<br/>We are concerned with learning models that generalize well to different \emph{unseen} domains. We consider a worst-case formulation over data distributions that are near the source domain in the feature space. Only using training data from a single source distribution, we propose an iterative procedure that augments the dataset with examples from a fictitious target domain that is "hard" under the current model. We show that our iterative scheme is an adaptive data augmentation method where we append adversarial examples at each iteration. For softmax losses, we show that our method is a data-dependent regularization scheme that behaves differently from classical regularizers that regularize towards zero (e.g., ridge or lasso). On digit recognition and semantic segmentation tasks, our method learns models improve performance across a range of a priori unknown target domains.<br/>