Generalizing to Unseen Domains via Adversarial Data Augmentation

TL;DR

This paper introduces an adversarial data augmentation method that iteratively adds challenging examples from fictitious target domains to improve model generalization to unseen domains, demonstrating effectiveness on digit recognition and segmentation tasks.

Contribution

The paper proposes a novel iterative adversarial augmentation technique that enhances domain generalization by generating hard examples, differing from traditional regularization methods.

Findings

Improved performance on unseen digit recognition domains.

Enhanced semantic segmentation accuracy across unknown target domains.

Method acts as a data-dependent regularizer for softmax losses.

Abstract

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.