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.
Introduction
Deep learning models often struggle when deployed in environments that differ from their training data, a phenomenon known as domain shift. For instance, a model trained on synthetic images (source domain) may perform poorly on real-world photographs (target domain). Unsupervised domain adaptation (UDA) aims to mitigate this by transferring knowledge from a labeled source domain to an unlabeled target domain.
Traditional adversarial domain adaptation methods employ a generator to produce features and a domain critic to distinguish between source and target features. While this aligns global distributions, it often ignores the underlying class boundaries. Consequently, target samples may be mapped near decision boundaries, leading to high error rates. This paper introduces Adversarial Dropout Regularization (ADR) to address this limitation by promoting low-density separation in the target domain.
Methodology
The core innovation of ADR is the use of the classifier itself as a critic to identify non-discriminative regions in the feature space.
The Adversarial Game
Unlike standard methods that use a binary domain classifier, ADR leverages dropout to measure prediction uncertainty.
- Dropout as a Probe: A target feature is passed through the classifier twice with different dropout masks, yielding two predictions, $p_1$ and $p_2$.
- Discrepancy Loss: The difference between these predictions is quantified using a symmetric Kullback-Leibler (KL) divergence. A high discrepancy indicates that the feature lies near a decision boundary where the model is uncertain.
- Optimization: The classifier (critic) is trained to maximize this discrepancy on target data, effectively finding the most sensitive boundaries. Simultaneously, the generator is trained to minimize this discrepancy, learning to produce features that are robust to dropout noise.
By forcing the generator to produce features that yield consistent predictions regardless of dropout, ADR pushes target samples into high-confidence regions, achieving low-density separation without requiring target labels.
Experimental Results
The authors evaluated ADR across several benchmarks, demonstrating its versatility and effectiveness.
Digit and Object Classification
On the SVHN to MNIST task, ADR achieved a state-of-the-art accuracy of 94.1%, significantly outperforming previous methods like ADDA (76.0%). In the VisDA 2017 challenge (synthetic-to-real), ADR reached a mean accuracy of 79.6% using a ResNeXt backbone, compared to 59.6% for DANN.
Semantic Segmentation
For the GTA5 to Cityscapes task, ADR improved the Mean Intersection over Union (mIoU) from 27.1 to 33.3. The resulting segmentation maps were visibly "cleaner," with fewer artifacts in ambiguous regions.
Semi-supervised Learning
ADR was also applied to semi-supervised GAN training. On the SVHN dataset, it achieved a lower error rate (6.26%) than established competitors, proving its utility beyond pure domain adaptation.
Discussion and Conclusion
Adversarial Dropout Regularization offers a robust framework for UDA by ensuring that feature alignment respects class boundaries. By utilizing the classifier as its own critic, the model focuses on task-relevant discriminative features.
While ADR shows significant gains, it introduces some computational overhead due to the multi-step optimization process. Future research may explore alternative perturbation methods beyond dropout to further enhance the critic's ability to detect non-discriminative regions. Overall, ADR represents a significant step toward more reliable and transferable deep learning models.
Summary updated