Episodic Training for Domain Generalization

By Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, Timothy M. Hospedales. 2019

499 citations recorded by the source corpus.

Domain generalization (DG) is the challenging and topical problem of learning models that generalize to novel testing domains with different statistics than a set of known training domains. The simple approach of aggregating data from all source domains and training a single deep neural network end-to-end on all the data provides a surprisingly strong baseline that surpasses many prior published methods. In this paper we build on this strong baseline by designing an episodic training procedure that trains a single deep network in a way that exposes it to the domain shift that characterises a novel domain at runtime. Specifically, we decompose a deep network into feature extractor and classifier components, and then train each component by simulating it interacting with a partner who is badly tuned for the current domain. This makes both components more robust, ultimately leading to our networks producing state-of-the-art performance on three DG benchmarks. Furthermore, we consider the pervasive workflow of using an ImageNet trained CNN as a fixed feature extractor for downstream recognition tasks. Using the Visual Decathlon benchmark, we demonstrate that our episodic-DG training improves the performance of such a general purpose feature extractor by explicitly training a feature for robustness to novel problems. This shows that DG training can benefit standard practice in computer vision.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

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

Doodle It Yourself: Class Incremental Learning by Drawing a Few Sketches (2022, methodology): “Sketch , despite being visually representative, is just a coarse contour-like depiction of the visual world, that sit in an entirely different domain from photo [27].”
RobustNet: Improving Domain Generalization in Urban-Scene Segmentation via Instance Selective Whitening (2021, background): “Various approaches such as meta-learning [27, 2, 28, 33], adversarial training [29, 32, 46], autoencoder [15, 29], metric learning [10, 39], data augmentation [64, 16, 67] have been proposed to learn domain-agnostic feature representations.”
Learning to Generate Novel Domains for Domain Generalization (2020, unknown): “For all datasets, as target data is unavailable during training, the values of hyper-parameters λDomain, λCycle and λCE are set based on the performance on source validation set, 5 which is a strategy commonly adopted in the DG literature [5,27].”
Learning to Diversify for Single Domain Generalization (2021, methodology): “AGG [25], HEX [44], PAR [43], JiGen [5], ADA [42], MEADA [49], MMLD [31] and our method belong to the latter category.”
Test-Time Training for Out-of-Distribution Generalization (2019, background): “With only a few environments, information on the meta distribution is often too scarce to be helpful, and with many environments, we are back to the i.i.d. setting where each environment can be seen as a sample, and a strong baseline is to simply train on all the environments (Li et al., 2019).”
Learning to Balance Specificity and Invariance for In and Out of Domain Generalization (2020, methodology): “∗∗We report the performance for MetaReg [3] from [29] as the official PACS trainval data split changed post MetaReg [3] publication.”
Domain Generalization using Causal Matching (2020, methodology): “More-over, on average MatchDG , MDGHybrid are better than many previous approaches D-MTAE (Ghifary et al., 2015), DBADG (Li et al., 2017), CIDDG (Li et al., 2018d), HEX (Wang et al., 2019) and FeatureCritic (Li et al., 2019b), but some other methods like MASF (Dou et al., 2019), DGER (Zhao et al., 2”
Model Vulnerability to Distributional Shifts over Image Transformation Sets (2019, background): “In domain generalization [30, 40, 41, 46, 36, 53, 31, 33, 53] the problem of dealing with unseen distributions is coped.”
Self-Challenging Improves Cross-Domain Generalization (2020, background): “alternates domain-specific feature extractors and classifiers across domains via episodic training, but without using inner gradient descent update [14].”
FSDR: Frequency Space Domain Randomization for Domain Generalization (2021, methodology): “Most existing DG methods can be broadly categorized into multi-source DG [44, 13, 12, 34, 16, 33, 43, 42, 59, 36, 2, 10, 35] and single-source DG [64, 70, 77, 50] both of which strive to learn domainagnostic features by either learning a domain-invariant feature space [44, 13, 12, 16, 34, 59, 36, 10”
Gradient Matching for Domain Generalization (2021, unknown): “In fact, there are a few works in domain generalization that are inspired by the meta-learning principles, such as Li et al. (2018a); Balaji et al. (2018); Li et al. (2019); Dou et al. (2019).”
Domain Generalization via Model-Agnostic Learning of Semantic Features (2019, methodology): “The promising results for [1, 26, 27] indicate that exposing the training procedure to domain shift beneﬁts model generalization to unseen domains.”
Learning from Extrinsic and Intrinsic Supervisions for Domain Generalization (2020, methodology): “[27] developed an episodic training procedure to expose the network to domain shift that characterizes a novel domain at runtime to improve the robustness of the network.”
Adversarial target-invariant representation learning for domain generalization (2019, methodology): “We compare the proposed approach against current state-of-the-art methods, namely, the recently proposed Epi-FCR (Carlucci et al., 2019) and JiGen (Li et al., 2019).”
Deep Domain-Adversarial Image Generation for Domain Generalisation (2020, methodology): “The recently proposed Epi-FCR uses episodic training to improve the Vanilla model.”