Self-Challenging Improves Cross-Domain Generalization

By Zeyi Huang, Haohan Wang, Eric P. Xing, Dong Huang. 2020

784 citations recorded by the source corpus.

Convolutional Neural Networks (CNN) conduct image classification by activating dominant features that correlated with labels. When the training and testing data are under similar distributions, their dominant features are similar, which usually facilitates decent performance on the testing data. The performance is nonetheless unmet when tested on samples from different distributions, leading to the challenges in cross-domain image classification. We introduce a simple training heuristic, Representation Self-Challenging (RSC), that significantly improves the generalization of CNN to the out-of-domain data. RSC iteratively challenges (discards) the dominant features activated on the training data, and forces the network to activate remaining features that correlates with labels. This process appears to activate feature representations applicable to out-of-domain data without prior knowledge of new domain and without learning extra network parameters. We present theoretical properties and conditions of RSC for improving cross-domain generalization. The experiments endorse the simple, effective and architecture-agnostic nature of our RSC method.

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

In Search of Lost Domain Generalization (2020, background): “This augmentation protocol is increasingly standard in state-of-the-art DG work (Nam et al., 2019; Huang et al., 2020; Krueger et al., 2020; Carlucci et al., 2019a; Zhou et al., 2020; Dou et al., 2019; Hendrycks et al., 2020; Wang et al., 2020a; Seo et al., 2020; Chattopadhyay et al., 2020).”
A Fourier-based Framework for Domain Generalization (2021, methodology): “The performance of our method also exceeds that of RSC [16], a recent proposed method utilizing a simple yet powerful gradient-guided dropout, by 2.11% on average2.”
Delving Deeper into Anti-Aliasing in ConvNets (2020, background): “, (Wang et al., 2019, ?; Huang et al., 2020)) have been proposed to increase amodel’s robustness to domain differences in the data.”
Permuted AdaIN: Reducing the Bias Towards Global Statistics in Image Classification (2020, methodology): “For the RSC [23] method, we also independently run the method using the open-source implementation as published by the authors, using the default conﬁguration ( https://github.com/DeLightCMU/RSC ).”
Orthogonal Projection Loss (2021, positive): “Additionally, we demonstrate the plug-and-play nature of OPL by showing beneﬁts of its use over CE, Truncated Loss (for noisy labels) [65], RSC [19] and various adversarial learning baselines.”
Deep Stable Learning for Out-Of-Distribution Generalization (2021, methodology): “As shown in Table 3, RSC fails to outperform vanilla CNNs.”
Test-Time Classifier Adjustment Module for Model-Agnostic Domain Generalization (2021, methodology): “These results include the following algorithms: Empirical Risk Minimization (ERM), Group Distributionally Robust Optimization (GroupDRO) [41], Inter-domain Mixup (Mixup) [56, 58, 53], Meta-Learning for Domain Generalization (MLDG) [28], DomainAdversarial Neural Networks (DANN) [15], Class-conditiona”
SelfReg: Self-supervised Contrastive Regularization for Domain Generalization (2021, methodology): “[22], we train our model, for approximately 30 epochs, with a SGD optimizer using ResNet18 [21] as a backbone, which is pretrained on ImageNet [9].”
FedDG: Federated Domain Generalization on Medical Image Segmentation via Episodic Learning in Continuous Frequency Space (2021, methodology): “Unseen Site A B C D E F Average A B C D E F Average Dice Coefficient (Dice) ↑ Hausdorff Distance (HD) ↓ JiGen [3] 89.95 85.81 84.06 87.34 81.32 89.11 86.26 10.51 11.53 11.70 11.49 14.80 9.02 11.51 BigAug [60] 89.63 84.62 83.86 87.66 81.20 88.96 85.99 10.68 11.78 12.07 10.66 13.98 9.73 11.48 Epi-FCR ”
Learning to Diversify for Single Domain Generalization (2021, neutral): “, alignment-based [2, 11, 33] and augmentation-based [5, 17, 44, 50, 47].”
Gradient Matching for Domain Generalization (2021, positive): “Algorithms include penalising the domain-predictive power of the model (Ganin et al., 2016; Wang et al., 2019; Huang et al., 2020), matching mean and variance of feature distributions across domains (Sun and Saenko, 2016), learning useful representations by solving Jigsaw puzzles (Carlucci et al.”