Source-linked AI summary
FixBi: Bridging Domain Spaces for Unsupervised Domain Adaptation
Jaemin Na, Heechul Jung, Hyung Jin Chang, Wonjun Hwang
TL;DR
UDA struggles when direct source-to-target adaptation encounters large domain discrepancies. FixBi constructs complementary intermediate domains with fixed-ratio mixup and trains models using confidence-based mutual and self-learning, achieving competitive performance on three standard benchmarks.
Problem
Large marginal-distribution gaps make direct unsupervised transfer from a labeled source domain to an unlabeled target domain difficult.
Method
FixBi uses two fixed mixup ratios to create complementary intermediate domains, then applies bidirectional matching and self-penalization.
Results
FixBi achieves competitive performance with state-of-the-art methods on Office-31, Office-Home, and VisDA-2017, while improving baseline DANN by an average of 7.1%.
Takeaways & Limitations
Intermediate domains and complementary confidence-based learning provide an effective strategy for bridging large source–target domain gaps in UDA.
Takeaways & Limitations
The method uses pseudo-labels obtained from a baseline model for unlabeled target samples.
Abstract
from arXiv · showhide
Unsupervised domain adaptation (UDA) methods for learning domain invariant representations have achieved remarkable progress. However, most of the studies were based on direct adaptation from the source domain to the target domain and have suffered from large domain discrepancies. In this paper, we propose a UDA method that effectively handles such large domain discrepancies. We introduce a fixed ratio-based mixup to augment multiple intermediate domains between the source and target domain. From the augmented-domains, we train the source-dominant model and the target-dominant model that have complementary characteristics. Using our confidence-based learning methodologies, e.g., bidirectional matching with high-confidence predictions and self-penalization using low-confidence predictions, the models can learn from each other or from its own results. Through our proposed methods, the models gradually transfer domain knowledge from the source to the target domain. Extensive experiments demonstrate the superiority of our proposed method on three public benchmarks: Office-31, Office-Home, and VisDA-2017.
1. Introduction
UDA transfers knowledge from labeled source data to unlabeled target data, but large domain gaps make direct adaptation difficult. FixBi bridges these domains with fixed-ratio intermediate domains and complementary confidence-based learning.
- Motivation: UDA transfers knowledge from a labeled source domain to a target domain with little or no labeled data, despite domain discrepancy.Large marginal-distribution gaps make effective transfer especially difficult when the target has no labels.
- Motivation: Most prior methods directly adapt source knowledge to the target without accounting for large source–target distances.Earlier approaches include domain-adversarial, discrepancy-based, and GAN-based methods.
- Proposed approach: FixBi uses fixed ratio-based mixup to construct multiple intermediate domains between the source and target.Source-close domains provide more reliable labels, while target-close domains better match target characteristics.
- Proposed approach: The method trains complementary source-dominant and target-dominant models that transfer knowledge through bidirectional matching and self-penalization.High-confidence predictions support peer learning, while low-confidence predictions support self-penalization.
- Evaluation: FixBi is evaluated with ablations and comparisons on Office-31, Office-Home, and VisDA-2017.The paper reports comparable performance to state-of-the-art domain adaptation methods on these benchmarks.
2. Related Work
Related UDA work aligns domains or learns transferable representations, while mixup-based and semi-supervised approaches provide alternative learning perspectives. FixBi instead creates two fixed-ratio intermediate domains with complementary source- and target-side properties.
- Semi-supervised learning: Semi-supervised learning uses unlabeled data alongside limited labels, whereas semi-supervised domain adaptation has some target labels unavailable in UDA.Prior work uses different data combinations or co-teaching to obtain distinct model perspectives.
- Unsupervised domain adaptation: UDA methods commonly learn domain-invariant features through domain alignment, including MMD, JMMD, and adversarial domain confusion.Examples include DAN, joint adaptation networks, DCC, and DANN.
- Unsupervised domain adaptation: Other methods use GANs, spherical classifiers, clustering, structural regularization, or mixup to reduce domain discrepancy or improve target discrimination.These approaches include ADDA, RSDA, SRDC, and DMRL.
- Position of FixBi: FixBi differs from randomness-based domain expansion by constructing source-closed and target-closed augmented domains with two fixed mixup ratios.The two domains provide complementary label reliability and target similarity, then teach each other to transfer knowledge toward the target.
3. Proposed Method
FixBi bridges large source–target gaps by creating complementary intermediate domains with fixed-ratio mixup, then training source-dominant and target-dominant models through confidence-based learning and consistency regularization.
- Fixed Ratio-based Mixup: FixBi uses two complementary fixed mixup ratios, λsd and λtd, whose sum is 1, to construct source-dominant and target-dominant models.The models provide different perspectives between the source and target domains.
- Fixed Ratio-based Mixup: The mixup combines source and target samples and uses baseline-model pseudo-labels for unlabeled target samples.The pseudo-labels are obtained from a baseline such as DANN or MSTN.
- Confidence-based Learning: Bidirectional matching trains each model on high-confidence pseudo-labels assigned by its peer network for target samples.A prediction becomes a positive pseudo-label when its class probability exceeds threshold τ.
- Confidence-based Learning: Self-penalization uses low-confidence top-1 predictions as negative pseudo-labels and drives their output probabilities toward zero.This incorporates low-confidence predictions into learning rather than ignoring them.
- Confidence-based Learning: An adaptive confidence threshold reflects the changing confidence distribution during training and avoids biasing positive and negative pseudo-label counts.A fixed threshold may become unsuitable as the network learns.
- Consistency Regularization and Training: Consistency regularization encourages both models to produce consistent results in an intermediate space, while training begins with a warm-up period before peer teaching.The intermediate space uses λsd = λtd = 0.5, and the warm-up independently trains the two networks.
4. Experiments
Experiments evaluate FixBi through benchmark comparisons, ablations, and visual analyses across Office-31, Office-Home, and VisDA-2017. Results show benefits from fixed-ratio mixup, complementary models, confidence-based components, and adaptation across large domain shifts.
- 4.1. Setups: FixBi is evaluated on Office-31, Office-Home, and VisDA-2017 using standard UDA protocols and extensive ablation studies.Office-31 and Office-Home use ResNet-50, while VisDA-2017 uses ResNet-101; DANN supports component analysis and MSTN supports state-of-the-art comparisons.
- 4.2. Ablation studies and discussions: The fixed-ratio mixup trains source-dominant and target-dominant models with different perspectives, using λsd = 0.7 and λtd = 0.3.Class-wise accuracies differ between the models, although their overall performances are similar at 86.3% and 86.0%.
- 4.2. Ablation studies and discussions: Target-biased mixup can degrade TDM accuracy when target labels are incorrect, whereas fixed-ratio mixup enables gains from bidirectional matching.Random and Range mixup rules do not provide the same expected improvement through bidirectional matching.
- 4.2. Ablation studies and discussions: The two-perspective FixBi model outperforms ensembles built from single-perspective models on Office-31.The comparison uses mixup ratios (0.3, 0.3) and (0.7, 0.7) for the single-perspective models and applies only Lfm and Lbim.
- 4.2. Ablation studies and discussions: 7.1% average improvement over DANN results from combining FixBi components on Office-31.Fixed-ratio mixup contributes 3.3% on average, bidirectional matching adds 1.5%, and consistency regularization also improves performance; self-penalization is especially impactful on D→A and W→A.
- 4.2. Ablation studies and discussions: Adaptive confidence thresholds change within each mini-batch and increase with training iterations, with a sharp TDM confidence increase when bidirectional matching begins at epoch 100.The increase is especially visible when the SDM and TDM mixup ratios differ substantially.
- 4.2. Ablation studies and discussions: t-SNE visualization shows FixBi forming compact target-feature clusters close to source-feature clusters, unlike DANN’s less structured target features.The visualization concerns the A→W task of Office-31 and uses blue for source points and orange for target points.
- 4.3. Comparison with state-of-the-art methods: 91.4% average accuracy on Office-31, 72.7% on Office-Home, and 22.2% average improvement over the baseline on VisDA-2017 are reported for FixBi.FixBi is particularly strong on large-shift tasks, including W→A on Office-31 and real-world/artificial transfers on Office-Home; it also reports about 12% better accuracy than other mixup-based methods on VisDA-2017.
5. Conclusion
FixBi bridges large source–target domain gaps by constructing intermediate domains with complementary characteristics and using multiple learning mechanisms. Experiments on three standard benchmarks show competitive state-of-the-art performance.
- FixBi addresses large domain discrepancies in unsupervised domain adaptation by bridging the source and target domain spaces.
- Fixed ratio-based mixup constructs intermediate domains with different characteristics between the source and target domains.
- Bidirectional matching, self-penalization, and consistency regularization enable efficient use of the intermediate space.
- FixBi achieves competitive performance with state-of-the-art methods on Office-31, Office-Home, and VisDA-2017.