Learning to Learn Single Domain Generalization

By Fengchun Qiao, Long Zhao, Xi Peng. Computer Vision and Pattern Recognition, 2020

556 citations recorded by the source corpus.

We are concerned with a worst-case scenario in model generalization, in the sense that a model aims to perform well on many unseen domains while there is only one single domain available for training. We propose a new method named adversarial domain augmentation to solve this Out-of-Distribution (OOD) generalization problem. The key idea is to leverage adversarial training to create "fictitious" yet "challenging" populations, from which a model can learn to generalize with theoretical guarantees. To facilitate fast and desirable domain augmentation, we cast the model training in a meta-learning scheme and use a Wasserstein Auto-Encoder (WAE) to relax the widely used worst-case constraint. Detailed theoretical analysis is provided to testify our formulation, while extensive experiments on multiple benchmark datasets indicate its superior performance in tackling single domain generalization.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/89b95fb53727ee45be440045efef656718517c4c

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 Computer Vision and Pattern Recognition

Citations in the corpus

Deep Stable Learning for Out-Of-Distribution Generalization (2021, methodology): “M-ADA, which generates images for training with an autoencoder, may fail when the training data are largescale real-world images and the distribution shifts are not caused by random disturbance.”
FSDR: Frequency Space Domain Randomization for Domain Generalization (2021, methodology): “Single-source DG handles generalization by domain randomization that augment data [64, 70] or domains [77, 50].”
FedDG: Federated Domain Generalization on Medical Image Segmentation via Episodic Learning in Continuous Frequency Space (2021, methodology): “Domain generalization [5, 9, 12, 14, 43, 47, 58, 59] aims to learn a model from multiple source domains such that it can directly generalize to unseen target domains.”
Domain Generalization: A Survey (2021, methodology): “The key idea is to perturb images using sign-flipped gradients back-propagated from a domain classifier [23] or a label classifier [38], [43], [166].”
SMIL: Multimodal Learning with Severely Missing Modality (2021, methodology): “Other works adapted MAML for domain generalization (Li et al. 2018; Qiao, Zhao, and Peng 2020) and knowledge distillation (Zhao et al. 2020).”
Non-Transferable Learning: A New Approach for Model Ownership Verification and Applicability Authorization (2021, background): “…al., 2020; Piratla et al., 2020), etc. Recently, there is signiﬁcant interest on conducting DG with one source domain only, for which well-crafted data augmentation approaches (Qiao et al., 2020; Zhao et al., 2020b; Li et al., 2021b; Xu et al., 2020b) have been proposed to expand the input space.”
Learning to Diversify for Single Domain Generalization (2021, methodology): “Very recently, a few works [37, 49] are proposed that learn to add adversarial noises on the source images in order to train robust classifiers against unforeseen data shifts.”
Continual Adaptation of Visual Representations via Domain Randomization and Meta-learning (2020, background): “First, designing more effective auxiliary sets might signiﬁcantly improve adaptation to more diverse domains; previous works focusing on augmentation strategies to improve in-domain [12, 25] and out-of-domain [59, 58, 46] performance might prove helpful in our settings.”
Knowledge As Priors: Cross-Modal Knowledge Generalization for Datasets Without Superior Knowledge (2020, background): “This is a well-defined problem in transfer learning which is tackled by domain adaptation [30, 33, 34].”