Adversarial Dropout Regularization

By Kuniaki Saito, Yoshitaka Ushiku, Tatsuya Harada, Kate Saenko. International Conference on Learning Representations, 2017

296 citations recorded by the source corpus.

We present a method for transferring neural representations from label-rich source domains to unlabeled target domains. Recent adversarial methods proposed for this task learn to align features across domains by fooling a special domain critic network. However, a drawback of this approach is that the critic simply labels the generated features as in-domain or not, without considering the boundaries between classes. This can lead to ambiguous features being generated near class boundaries, reducing target classification accuracy. We propose a novel approach, Adversarial Dropout Regularization (ADR), to encourage the generator to output more discriminative features for the target domain. Our key idea is to replace the critic with one that detects non-discriminative features, using dropout on the classifier network. The generator then learns to avoid these areas of the feature space and thus creates better features. We apply our ADR approach to the problem of unsupervised domain adaptation for image classification and semantic segmentation tasks, and demonstrate significant improvement over the state of the art. We also show that our approach can be used to train Generative Adversarial Networks for semi-supervised learning.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/c20148703a706d5883d323c3386978714fe1508d

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

Universal Domain Adaptation through Self Supervision (2020, methodology): “In OSDA, we follow (Saito et al., 2018c) and use the 6 classes as shared classes |Ls ∩ Lt| and the rest as unknown classes (Lt − Lt).”
Self-ensembling for visual domain adaptation (2017, background): “A neural network is trained to make consistent predictions for unsupervised samples under different augmentation Sajjadi et al. (2016), dropout and noise conditions or through the use of adversarial training Miyato et al.”
Learning to Balance Specificity and Invariance for In and Out of Domain Generalization (2020, background): “feature-level and pixel-level aligment is explored in [18,43]. In addition, several regularization strategies have also been proven to be eective for domain adaptation such as dropout regularization [41], classier discrepancy [42], self-ensembling [13], etc. Most existing domain adaptation methods c”
ADVENT: Adversarial Entropy Minimization for Domain Adaptation in Semantic Segmentation (2018, background): “[32, 33] propose another interesting way of using adversarial training: They get two predictions on the target domain image, this is done either by two clas-siﬁers [33] or using dropout in the classiﬁer [32].”
Stochastic Classifiers for Unsupervised Domain Adaptation (2020, methodology): “, dropout [44] and domain-specific batch normalisation (DSBN) [3], have been proposed to impose on previous methods.”
FSDR: Frequency Space Domain Randomization for Domain Generalization (2021, background): “The first is adversarial training based which employs adversarial learning to align source and target distributions in the feature, output or latent space [22, 39, 69, 40, 66, 7, 81, 54, 55, 72, 1, 68, 26, 18, 25, 31].”
Reducing Domain Gap by Reducing Style Bias (2021, background): “Saito et al. [44] showed that common UDA methods [11, 34, 45] often fail to improve their performance in such scenario, and proposed minimax entropy (MME) setting the state-of-the-art for SSDA.”
Confidence Regularized Self-Training (2019, methodology): “To this end, a predominant stream of adversarial learning based UDA methods were proposed to reduce the discrepancy between source and target domain features [9, 10, 23, 26, 34, 38, 44, 50, 53, 60].”
Minimum Class Confusion for Versatile Domain Adaptation (2019, unknown): “0 90.3 42.9 85.9 53.1 49.7 36.3 85.8 20.7 61.1 MCD [43] 87.0 60.9 83.7 64.0 88.9 79.6 84.7 76.9 88.6 40.3 83.0 25.8 71.9 CDAN [27] 85.2 66.9 83.0 50.8 84.2 74.9 88.1 74.5 83.4 76.0 81.9 38.0 73.9 ADR [42] 87.8 79.5 83.7 65.3 92.3 61.8 88.9 73.2 87.8 60.0 85.5 32.3 74.8 AFN [56] 93.6 61.3 84.1 70.6 9”
Learning Semantic Representations for Unsupervised Domain Adaptation (2018, methodology): “Recently, adversarial learning has been widely adopted in domain adaptation (Ganin & Lempitsky, 2015; Tzeng et al., 2015; Hoffman et al., 2017; Motiian et al., 2017a; Tzeng et al., 2017; Saito et al., 2017b; Long et al., 2017a; Luo et al., 2017; Sankaranarayanan et al.).”
Contrastive Adaptation Network for Unsupervised Domain Adaptation (2019, methodology): “For a comparison under optimal parameter setting, we cite the performance of MADA, RevGrad, MCD and ADR reported in their corresponding papers [23, 26, 27, 9].”
Deep Unsupervised Domain Adaptation: A Review of Recent Advances and Perspectives (2022, methodology): “Saito et al. (2017) propose adversarial dropout regularization, which is seen as the difference between two dropout networks as a discriminator to generate target discriminative features.”
Adversarial Domain Adaptation with Domain Mixup (2019, background): “ing the discriminator. Other works optimize the performance on target domain by capturing complex multimode structures (Pei et al. 2018; Long et al. 2018), exploring task-speciﬁc decision boundaries (Saito et al. 2018b; 2018a; Tran et al. 2019), aligning the attention regions (Kang et al. 2018) and ”
Angular Visual Hardness (2019, methodology): “Method Aero Bike Bus Car Horse Knife Motor Person Plant Skateboard Train Truck Mean Source (Saito et al., 2018) 55.”
Reducing Domain Gap via Style-Agnostic Networks (2019, background): “[45] showed that common UDA methods [11,35,46] often fail to improve performance in such scenario and proposed MME, achieving the state-of-the-art for SSDA.”
Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation (2020, methodology): “For speciﬁc partial-set and open-set DA tasks, we compare with IWAN (Zhang et al., 2018a), SAN (Cao et al., 2018), ETN (Cao et al., 2019), SAFN, and ATI-λ (Panareda Busto & Gall, 2017), OSBP (Saito et al., 2018c), OpenMax (Bendale & Boult, 2016), STA (Liu et al., 2019), respectively.”
Exploiting the Intrinsic Neighborhood Structure for Source-free Domain Adaptation (2021, methodology): “Method SF plane bcycl bus car horse knife mcycl person plant sktbrd train truck Per-class ADR [34] 7 94.”
Taking a Closer Look at Domain Shift: Category-Level Adversaries for Semantics Consistent Domain Adaptation (2018, background): “Our work does not follow the strategy of global feature alignment [41, 15, 39] or classifiers consensus maximization [34, 35].”
FixBi: Bridging Domain Spaces for Unsupervised Domain Adaptation (2020, background): “Recent works [10, 35, 38, 24, 8, 37, 31, 43] have focused on UDA based on domain alignment and discriminative domain-invariant feature learning methods.”