[Paper Review] Adversarial Dropout Regularization

Kuniaki Saito, Yoshitaka Ushiku+2 more|arXiv (Cornell University)|Nov 5, 2017

Domain Adaptation and Few-Shot LearningComputer Science25 references50 citations

TL;DR

ADR introduces a dropout-based adversarial critic to enforce boundary-aware feature alignment for unsupervised domain adaptation, improving discriminativeness of target features across classification and segmentation tasks and enabling semi-supervised GAN training.

ABSTRACT

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.

Motivation & Objective

• Motivate transferring neural representations from labeled source domains to unlabeled target domains.
• Address limitations of traditional domain critics that ignore class boundaries during alignment.
• Develop ADR to make the target features more discriminative by boundary-aware adversarial dropout.
• Show ADR's applicability beyond domain adaptation to semi-supervised learning and GANs.

Proposed method

• Replace the domain critic with a classifier C that acts as both classifier and boundary-sensitive critic via dropout.
• Use two dropout-induced classifiers C1 and C2 to measure the sensitivity of p(y|x) to boundary perturbations; maximize this sensitivity for target samples while minimizing it for generator outputs.
• Train G (feature generator) to produce target features far from decision boundaries by minimizing the dropout sensitivity on target data.
• Train C to maximize target sensitivity while maintaining discriminative power on source data and use a source supervision term.
• Incorporate a conditional entropy term to encourage uniform class distribution for target features.
• Extend ADR to semi-supervised GAN training by having G generate images that fool C while aligning real unlabeled and generated data.

Experimental results

Research questions

• RQ1 Can a boundary-aware critic based on dropout improve alignment of source and target features without target labels?
• RQ2 Does maximizing sensitivity to dropout-induced perturbations on target samples lead to discriminative, boundary-free target representations?
• RQ3 Is ADR effective for both image classification and semantic segmentation under domain shifts?
• RQ4 Can ADR enhance semi-supervised GAN training by placing generated samples away from decision boundaries while keeping generated data similar to real data?

Key findings

• ADR improves unsupervised domain adaptation on digits (MNIST/SVHN/USPS) and image segmentation tasks compared to prior distribution-matching methods.
• On SVHN to MNIST, ADR achieves substantially better accuracy than prior methods in the reported experiments.
• ADR reduces target entropy and moves target samples away from decision boundaries, while a variant classifier C′ trained without boundary sensitivity remains competitive.
• ADR yields state-of-the-art results on VisDA2017 classification (synthetic to real) for ResNet101 and ResNeXt finetuning.
• In segmentation (GTA5 to Cityscapes), ADR improves mean IoU over source-only and competing methods.
• ADR can be used to train GANs for semi-supervised learning, achieving competitive results on SVHN and CIFAR-10.