Source-linked AI summary
Domain-Symmetric Networks for Adversarial Domain Adaptation
Yabin Zhang, Hui Tang, Kui Jia, Mingkui Tan
TL;DR
Unsupervised domain adaptation must transfer knowledge from labeled source data to unlabeled target data despite domain shift and limited category-level distribution alignment. SymNet addresses this with symmetric task classifiers, an additional shared-neuron classifier, two-level domain confusion, and cross-domain training. Experiments and ablations verify the method’s efficacy, with state-of-the-art results on three benchmark datasets.
Problem
Unsupervised domain adaptation must learn target-domain classifiers from labeled source data, while existing domain-adversarial methods remain limited in aligning feature distributions at the category level.
Method
SymNet uses symmetric source and target task classifiers, an additional shared-neuron classifier, two-level domain confusion losses, and cross-domain training for the target classifier.
Results
SymNets achieve state-of-the-art results on Office-31, ImageCLEF-DA, and Office-Home, while ablation studies verify the efficacy of key designs.
Takeaways & Limitations
The method provides an adversarial domain-adaptation design for aligning joint feature-category distributions across domains.
Takeaways & Limitations
The source and target domains are assumed to differ, violating the i.i.d. assumption, and an additional source-based domain-level confusion loss may be subsumed by the category-level loss.
Abstract
from arXiv · showhide
Unsupervised domain adaptation aims to learn a model of classifier for unlabeled samples on the target domain, given training data of labeled samples on the source domain. Impressive progress is made recently by learning invariant features via domain-adversarial training of deep networks. In spite of the recent progress, domain adaptation is still limited in achieving the invariance of feature distributions at a finer category level. To this end, we propose in this paper a new domain adaptation method called Domain-Symmetric Networks (SymNets). The proposed SymNet is based on a symmetric design of source and target task classifiers, based on which we also construct an additional classifier that shares with them its layer neurons. To train the SymNet, we propose a novel adversarial learning objective whose key design is based on a two-level domain confusion scheme, where the category-level confusion loss improves over the domain-level one by driving the learning of intermediate network features to be invariant at the corresponding categories of the two domains. Both domain discrimination and domain confusion are implemented based on the constructed additional classifier. Since target samples are unlabeled, we also propose a scheme of cross-domain training to help learn the target classifier. Careful ablation studies show the efficacy of our proposed method. In particular, based on commonly used base networks, our SymNets achieve the new state of the art on three benchmark domain adaptation datasets.
1. Introduction
Unsupervised domain adaptation addresses scarce target-domain annotations but remains limited by domain shift and imperfect category-level alignment. SymNet combines symmetric classifiers, an additional shared-neuron classifier, two-level domain confusion, and cross-domain training, achieving state-of-the-art results on three benchmarks.
- Scarce target-domain annotations motivate transferring labeled data from related source domains, but domain shift obstructs effective adaptation.
- Domain-adversarial methods learn domain-invariant features yet may leave the joint feature-category distributions misaligned across domains.
- SymNet uses symmetric source and target task classifiers plus an additional classifier sharing their layer neurons.
- Its training combines category-level and domain-level confusion losses with cross-domain training to improve target-classifier learning and category-level invariance.
- SymNets achieve new state-of-the-art performance on Office-31, ImageCLEF-DA, and Office-Home using commonly used base networks.
2. Related Works
Related domain-adaptation methods commonly align domain-invariant features without explicitly aligning corresponding categories. SymNet extends joint-distribution alignment with target and additional classifiers and two-level domain confusion losses.
- Existing methods typically learn domain-invariant features to minimize discrepancy between source and target domains.
- Some methods neglect alignment between corresponding categories, leaving joint feature-category distributions insufficiently aligned.
- Prior work asymmetrically uses three task classifiers, including two classifiers that label unlabeled target samples through prediction consistency.
- SymNet adds an explicit target task classifier, an additional classifier for domain discrimination and confusion, and category- and domain-level confusion losses.
3. The Proposed Domain-Symmetric Networks
SymNet aligns feature and category distributions across source and target domains using symmetric task classifiers, a shared-neuron classifier, and two-level adversarial domain confusion.
- Motivation: Unsupervised domain adaptation learns a feature extractor and classifier whose expected target risk is minimized despite differing source and target domains.Domain confusion methods use a discriminator to distinguish domains and train the feature extractor to confuse them.
- 3.1. A Symmetric Design of Source and Target Task Classifiers: SymNet uses parallel source and target classifiers Cs and Ct with K neurons each, plus Cst formed by concatenating their outputs into 2K probabilities.The two task classifiers are symmetric, while Cst shares their layer neurons and supports domain-related losses.
- 3.1. A Symmetric Design of Source and Target Task Classifiers: Both domain discrimination and domain confusion are applied through Cst rather than an explicit domain discriminator.Cst is trained to distinguish source and target classifier outputs, while appropriate confusion losses are applied to its neuron groups.
- 3.1. A Symmetric Design of Source and Target Task Classifiers: Cross-domain training uses labeled source samples to train Ct, establishing neuron-wise correspondence with Cs and making Ct more discriminative among task categories.Although this initially appears to duplicate Cs, Cst-based domain discrimination differentiates the classifiers and supports category-level confusion.
- 3.2. A Two-level Domain Confusion Training of Domain-Symmetric Networks: SymNet’s two-level confusion combines domain-level confusion on unlabeled target samples with category-level confusion on corresponding source-category neuron pairs.The category-level loss uses labeled source samples and uniform targets for paired neurons, while the domain-level loss aggregates the two neuron halves before confusion.
- 3.2. A Two-level Domain Confusion Training of Domain-Symmetric Networks: A source-based domain-level confusion loss could also be used, but its effect may already be subsumed by the category-level confusion loss.This design observation is stated as a possibility rather than an experimentally established necessity.
4. Experiments
Experiments evaluate SymNets on three benchmark datasets, compare them with existing methods, and analyze their components, convergence, and learned feature representations. Across these evaluations, SymNets achieve strong adaptation performance and align corresponding categories across domains.
- Experimental setup: SymNets are evaluated on unsupervised domain adaptation tasks from Office-31, ImageCLEF-DA, and Office-Home using labeled source and unlabeled target samples.The comparative methods use models adapted from a 50-layer ResNet, and results report average classification accuracy and standard error over three random experiments.
- Benchmark results: SymNets outperform all state-of-the-art methods on the three benchmark datasets, with substantial gains on difficult adaptation tasks and Office-Home.The paper attributes the challenge of Office-Home partly to many categories, visually dissimilar domains, and low in-domain accuracy.
- Ablation study: Removing the cross-domain category supervised loss reduces performance by 10.3%, showing its importance for learning the target task classifier.The comparison concerns the SymNets variant without the target-task loss.
- Ablation study: Domain Confusion performs much better than ResNet-50, while category-level confusion improves SymNets over the version without category confusion.These ablations support the contributions of domain-level and category-level feature alignment.
- Ablation study: Entropy minimization consistently improves ResNet-50, Domain Confusion, and SymNets by a large margin, while complete SymNets perform best among the ablations.The results support the usefulness of entropy minimization and the full objective.
- Convergence performance: On A → W, SymNets converge much more smoothly than ResNet-50, Domain Confusion, and Domain Confusion (Adding Em).The visualization compares SymNets-S and SymNets-T with the baseline methods through their test errors during adaptation.
- Feature visualization: t-SNE visualizations show that SymNets align shared categories across domains while distinguishing different categories better than the comparison methods.Domain Confusion improves domain alignment but leaves target structure scattered or shared categories misaligned; adding entropy preserves target structure without aligning shared categories.
5. Conclusion
The paper concludes that SymNets address joint feature–category distribution alignment through two-level domain confusion and cross-domain learning of an explicit target classifier. Experiments on three benchmark datasets verify the method’s efficacy.
- SymNets address the limitation of aligning joint feature and category distributions across source and target domains.
- The category-level confusion loss improves on domain-level confusion by promoting category-corresponding feature invariance across domains.
- An explicit target task classifier is learned through a cross-domain training scheme despite unlabeled target samples.
- Experiments on three benchmark datasets verify the efficacy of SymNets.