Generalizing to Unseen Domains via Adversarial Data Augmentation

By Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C. Duchi, Vittorio Murino, Silvio Savarese. 2018

918 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Radiology, Nuclear Medicine and Imaging, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Radiology, Nuclear Medicine and Imaging

In Safety, Robustness & Interpretability

Citations in the corpus

Learning Robust Global Representations by Penalizing Local Predictive Power (2019, methodology): “At a high level these methods comprise two subtypes: fine-tuning over target domain (Long et al., 2016; Hoffman et al., 2017; Motiian et al., 2017a; Gebru et al., 2017; Volpi et al., 2018) and coercing domain invariance via adversarial learning (or further extensions) (Ganin et al.”
Certifying Some Distributional Robustness with Principled Adversarial Training (2017, background): “[66] studied connections between robust formulations and novel regularization schemes that regularize by ‖∇zl(θ;Z)‖.”
Faster AutoAugment: Learning Augmentation Strategies using Backpropagation (2019, methodology): “im to automate data augmentation. One direction is to search for better combinations of symbolic operations using black-box optimization techniques: reinforcement learning [5, 23], evolution strategy [32], Bayesian optimization [18] and Population Based Training [12]. As the objective, [5, 32, 12] d”
In Search of Lost Domain Generalization (2020, background): “Volpi et al. [2018] show that training with adversarial data augmentation on a single domain is sufﬁcient to improve domain generalization.”
WILDS: A Benchmark of in-the-Wild Distribution Shifts (2020, neutral): “…al., 2016) and StreetHazards (Hendrycks et al., 2020) have been adopted for out-of-distribution detection as well as domain adaptation and generalization, e.g., by testing robustness to transformations in the seasons, weather, time, or architectural style (Hoffman et al., 2018; Volpi et al., 2018).”
Tilted Empirical Risk Minimization (2020, background): “, 2018), or generalizing to unseen domains (Volpi et al., 2018).”
Generalizing Deep Learning for Medical Image Segmentation to Unseen Domains via Deep Stacked Transformation (2020, background): “Speciﬁcally, additional training data samples are generated in image domain [20], semantic space [22], or by adversarial learning [21], respectively.”
Domain Generalization by Solving Jigsaw Puzzles (2019, methodology): “DA results around which the authors of [46] ran their parameter ablation analysis.”
Domain Generalization: A Survey (2021, neutral): “Semantic segmentation - SYNTHIA [72] 2,700 15 4 locations, 5 weather conditions (see [73]) - GTA5-Cityscapes [74], [75] 29,966 2 Synthetic versus real images”
Learning to Balance Specificity and Invariance for In and Out of Domain Generalization (2020, methodology): “Use of data augmentation techniques for domain generalization is explored in [49].”
Improved OOD Generalization via Adversarial Training and Pre-training (2021, background): “The most related works to ours are (Sinha et al., 2018; Lee & Raginsky, 2018; Volpi et al., 2018).”
SmoothLLM: Defending Large Language Models Against Jailbreaking Attacks (2023, positive): “One reason for this is many methods, e.g., adversarial training [33, 34] and data augmentation [35], retrain the underlying model.”