Source-linked AI summary
Open Set Domain Adaptation by Backpropagation
Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, Tatsuya Harada
TL;DR
Closed-set domain adaptation assumes source and target share classes, but unlabeled targets may include unknown classes that should not be aligned with the source. This paper uses adversarial training to let a classifier define a source–target boundary while a generator separates target samples into aligned knowns or rejected unknowns, without unknown source samples. Extensive experiments found better performance than other methods in most settings.
Problem
Closed-set domain adaptation assumes source and target share classes, although unlabeled target domains may contain unknown classes that should not be aligned with the source.
Method
Adversarial training uses a classifier to form a source–target boundary and a feature generator to align known targets or reject unknown targets, without unknown source samples.
Results
The method outperformed other methods in most domain adaptation settings and was also effective in standard open set recognition experiments.
Takeaways & Limitations
The approach learns features that separate unknown target samples from known target samples while supporting open set domain adaptation.
Takeaways & Limitations
The method assumes unknown source samples are unavailable and addresses the more challenging setting without access to them.
Abstract
from arXiv · showhide
Numerous algorithms have been proposed for transferring knowledge from a label-rich domain (source) to a label-scarce domain (target). Almost all of them are proposed for a closed-set scenario, where the source and the target domain completely share the class of their samples. We call the shared class the \doublequote{known class.} However, in practice, when samples in target domain are not labeled, we cannot know whether the domains share the class. A target domain can contain samples of classes that are not shared by the source domain. We call such classes the \doublequote{unknown class} and algorithms that work well in the open set situation are very practical. However, most existing distribution matching methods for domain adaptation do not work well in this setting because unknown target samples should not be aligned with the source. In this paper, we propose a method for an open set domain adaptation scenario which utilizes adversarial training. A classifier is trained to make a boundary between the source and the target samples whereas a generator is trained to make target samples far from the boundary. Thus, we assign two options to the feature generator: aligning them with source known samples or rejecting them as unknown target samples. This approach allows extracting features that separate unknown target samples from known target samples. Our method was extensively evaluated in domain adaptation setting and outperformed other methods with a large margin in most settings.
1 Introduction
Open set domain adaptation must handle target samples from classes absent in the source, while avoiding their alignment with source known classes. The paper proposes adversarial learning to separate known and unknown target samples without requiring unknown source samples.
- 1 Introduction: Open set recognition distinguishes training-seen known classes from unknown classes absent during training.Recognizing unknown samples can support noisy-dataset curation and allow robots to request labels for unfamiliar objects.
- 1 Introduction: In unsupervised domain adaptation, unlabeled target data may contain classes not represented in the source domain.This violates the common closed-set assumption that source and target share all classes.
- 1 Introduction: Unknown target samples must be rejected rather than aligned with source samples, because distribution matching treats all target samples as source-alignment candidates.The absence of unknown source samples makes ordinary distribution matching unsuitable for this setting.
- 1 Introduction: The proposed adversarial approach trains a classifier to form a source–target boundary while training a feature generator to move target samples away from it.The classifier outputs K + 1 class probabilities, with the additional dimension representing the unknown class.
- 1 Introduction: The feature generator is given two outcomes: align target samples with source known samples or reject them as unknown.The method is designed for a setting where unknown source samples are unavailable and was evaluated on digit and object adaptation tasks.
2 Related Work
Related work largely relies on distribution matching or outlier rejection, but these approaches do not directly address target classes absent from the source. The paper positions its adversarial method as a way to reject unknown targets while adapting known targets without unknown-source access.
- 2.1 Domain Adaptation: Distribution-matching methods seek domain-invariant discriminative features by matching feature distributions across source and target domains.Their standard formulation assumes target categories are included in the source domain.
- 2.1 Domain Adaptation: In open set adaptation, aligning every target sample with source samples makes unknown target samples difficult to detect.This is the central mismatch between conventional distribution matching and the open set setting.
- 2.1 Domain Adaptation: The proposed method can categorize unknown target samples without labeled unknown target examples during training.Its feature generator can reject some target samples as outliers rather than forcing all samples into source alignment.
- 2.2 Open Set Recognition: Prior open set recognition work includes outlier rejection with multi-class SVMs, OpenMax probability estimation, and methods that generate unknown-sample supervision.These approaches are presented as related techniques for rejecting unknowns during testing or training.
- 2.2 Open Set Recognition: The paper frames its contribution as open set recognition under domain adaptation, where target known-sample distributions differ from source distributions.This combines recognition of unknowns with the domain shift that motivates adaptation.
3 Method
The method uses adversarial training with a feature generator and classifier to separate known and unknown target samples while aligning known samples with source classes. Training combines source classification with an adversarial objective that gives the generator a choice between rejection and alignment.
- Problem Setting and Overall Idea: The model trains a feature generation network G and classifier C, with C predicting K known classes plus an unknown class.C produces K + 1 logits for each generated feature.
- Problem Setting and Overall Idea: The classifier is weakly trained to assign target samples probability t to the unknown class, creating a pseudo decision boundary.The method sets p(y = K + 1|x_t) = t, with 0 < t < 1.
- Training Procedure: The generator deceives the classifier by maximizing the adversarial loss, increasing rejection when target samples should be separated from the known classes.The classifier minimizes the adversarial binary cross-entropy, while the generator maximizes it.
- Training Procedure: Source samples are trained with standard cross-entropy so the classifier and generator categorize them into their correct known classes.The source loss is Ls(xs, ys) = −log(p(y = ys|xs)).
- Training Procedure: Gradient reversal flips the adversarial gradient during backpropagation, allowing classifier and generator parameters to update simultaneously.The overall objective combines source classification and adversarial loss.
- Comparison with Existing Methods: Unlike fixed-threshold rejection, the feature extractor uses a pseudo boundary and sample-dependent outputs to separate dissimilar target samples from known source samples.Representations are aligned with known source classes when similar and separated when dissimilar.
4 Experiments
Experiments on Office, VisDA, and digits datasets evaluate the proposed method across open set domain adaptation and recognition settings. The method generally outperforms baselines, separates unknown from known target samples, and benefits from combining its representations with OSVM in a difficult digits setting.
- Experimental setup: The experiments cover Office, VisDA, and digits datasets, using pretrained AlexNet or VGGNet features and comparisons with OSVM, MMD, and BP baselines.Office experiments use fixed pretrained-network parameters and added fully connected layers; training details vary by dataset.
- Experiments on Office: In the 10-shared-class Office setting, our method performs better than baseline methods in almost all scenarios.MMD does not consistently improve performance, and unknown target samples can perturb feature alignment.
- Experiments on Office: The learned features separate unknown target samples from known target samples, whereas MMD matches most target samples with source features.The visualization compares pretrained AlexNet, no adaptation, MMD, BP, and the proposed method.
- Experiments on Office: When the number of known classes increases, the proposed method remains superior overall, although baseline methods can score higher on OS and OS* in the Amazon-to-Webcam setting.The ALL measure favors the proposed method because competing methods more often classify target samples as known classes.
- Experiments on VisDA Dataset: On VisDA, the proposed method performs better in both average accuracy and average known-class accuracy, supporting both known-sample alignment and unknown-sample rejection.Six vehicle categories are treated as known and six other categories as unknown.
- Experiments on Digits Dataset: On digits, the proposed method separates unknown classes and known classes qualitatively, while SVHN-to-MNIST remains difficult because of the large domain difference.In semi-supervised open set recognition, combining the proposed method with OSVM dramatically improves unknown-sample accuracy for SVHN.
5 Conclusion
The paper proposes adversarial learning for open set domain adaptation, generating features that separate unknown target samples from known target samples without requiring unknown source samples.
- The proposed method generates features that separate unknown target samples from known target samples.
- The method differs from existing distribution matching methods by addressing unknown target samples directly.
- The approach does not require unknown source samples during training.
- Experiments verified the method’s effectiveness in open set domain adaptation and standard open set recognition.
A Detail of Experiments
The supplementary material provides experimental details and baseline settings omitted from the main paper because of space limitations.
- The supplementary material describes experiment details not included in the main paper because of limited space.
- It also provides the experimental settings for the baseline methods.
A.1 Training Detail for Experiments on Office and VisDA Dataset
The experiments use dataset-specific training schedules, shared architectures for baselines, and fixed implementation settings across Office, VisDA, and digits experiments.
- Office: Office experiments use batch-size 32, report proposed accuracy after 500 epochs, and report MMD and BP accuracy after 50 epochs.
- Office: MMD computes discrepancy between source and target features at the generator’s top layer.
- Office: BP attaches its domain classifier to the top layer of the feature generator networks.
- Baselines: The OSVM threshold is fixed at 0.1, its default value.
- VisDA: VisDA results are reported after 10 epochs, with other hyperparameters matching the Office experiment.
- Digits: Digits experiments use Leaky ReLU, batch-size 128, accuracy after 200 epochs, and the same architecture for OSVM, MMD, and BP.
B Additional Experimental Results
Additional results visualize features obtained during SVHN-to-MNIST adaptation, including source-only source and target feature representations.
- The additional results include feature visualizations for adaptation from SVHN to MNIST.
- One visualization is labeled Source Only with source and target features.
- Another visualization is labeled Source Only for target features.