Source-linked AI summary
Importance Weighted Adversarial Nets for Partial Domain Adaptation
Jing Zhang, Zewei Ding, Wanqing Li, Philip Ogunbona
TL;DR
Partial domain adaptation lacks the shared-label-space assumption used by conventional domain adaptation when an unlabeled target contains only a subset of source classes. The paper estimates source-sample importance with one domain classifier and aligns weighted source samples with target data using a second classifier, with target entropy minimization. Experiments report large gains over previous domain adaptation methods and comparability with state-of-the-art partial transfer methods, while assuming the source contains all target classes.
Problem
Conventional domain adaptation assumes identical label spaces, whereas partial adaptation transfers from a larger source containing all target classes to an unlabeled target with an unknown subset of classes.
Method
The method uses a first domain classifier to weight source samples, a second to adversarially align weighted source and target data, and target entropy minimization to preserve target structure.
Results
The proposed method outperforms previous domain adaptation methods by a large degree and is comparable to state-of-the-art partial transfer methods.
Takeaways & Limitations
Importance weighting can detect source outlier-class samples and reduce the shift between target data and source data within target classes.
Takeaways & Limitations
The method assumes the source domain includes all classes that appear in the target domain and that the target label space is contained in the source label space.
Abstract
from arXiv · showhide
This paper proposes an importance weighted adversarial nets-based method for unsupervised domain adaptation, specific for partial domain adaptation where the target domain has less number of classes compared to the source domain. Previous domain adaptation methods generally assume the identical label spaces, such that reducing the distribution divergence leads to feasible knowledge transfer. However, such an assumption is no longer valid in a more realistic scenario that requires adaptation from a larger and more diverse source domain to a smaller target domain with less number of classes. This paper extends the adversarial nets-based domain adaptation and proposes a novel adversarial nets-based partial domain adaptation method to identify the source samples that are potentially from the outlier classes and, at the same time, reduce the shift of shared classes between domains.
1. Introduction
Domain adaptation addresses distribution shifts by transferring labeled source knowledge to an unlabeled target, but conventional methods assume identical label spaces. This paper targets partial domain adaptation, where the target contains an unknown subset of source classes, by weighting source samples during adversarial alignment.
- Motivation: Domain adaptation transfers labeled source-domain data to improve performance in a target domain with few or no labels.Deep methods learn transferable, domain-invariant features through feature matching, adversarial training, or normalization statistics.
- Motivation: Conventional approaches require shared label spaces because they compare marginal source and target distributions.When label spaces differ, direct distribution matching is not necessarily suitable for feasible adaptation.
- Partial Domain Adaptation: Partial domain adaptation assumes that the target label space is an unknown subset of a larger, more diverse source label space.The source domain is assumed to include all classes appearing in the target, while target labels remain unavailable.
- Partial Domain Adaptation: Directly reducing source-target distribution shift can be ineffective when the target contains fewer classes than the source.The proposed alternative is to reweight source samples likely to belong to target-shared classes, despite the unlabeled target domain.
- Proposed Direction: The method uses two domain classifiers to estimate source-sample importance and align weighted source samples with target samples.The first classifier obtains importance weights, while the second participates in adversarial alignment.
- Evaluation: The method outperforms previous domain adaptation methods by a large degree and is comparable to the state-of-the-art partial transfer method on three cross-domain object datasets.These evaluations use settings where the target domain contains a subset of source classes.
2. Related Work
Related work develops deep domain adaptation through statistical matching, adversarial losses, and normalization, but most methods assume shared label spaces. The paper contrasts its two-classifier, instance-weighted strategy with shared-feature and class-weighted alternatives.
- Existing Approaches: Deep domain adaptation methods commonly reduce domain divergence through statistic-moment matching, adversarial losses, or Batch Normalization statistics.Examples include MMD, CMD, second-order statistics matching, and adversarial domain classifiers.
- Existing Approaches: Most existing approaches assume identical source and target label spaces when comparing marginal distributions.This assumption limits direct adaptation to partial-domain settings.
- Comparison: Unlike single-classifier adversarial adaptation, this method uses distinct source and target feature extractors and weights source samples during alignment.The distinction is presented relative to the approach of Ganin et al.
- Comparison: SAN uses separate domain classifiers for each class and both instance-level and class-level weights based on label-predictor probabilities.The paper identifies this as a fundamental difference from its own approach.
- Comparison: The proposed method uses two domain classifiers instead of one per source class, reducing scalability and computational-cost concerns for many-class sources.It also omits class-level weights, supporting adaptation to imbalanced target data according to the paper.
3. Proposed Method
The method pretrains a source classifier, estimates source-sample importance with one domain classifier, and uses a second classifier to adversarially align weighted source and target features. Entropy minimization constrains target features while preserving target structure.
- Problem Setup: The problem assumes labeled source data, unlabeled target data, shared feature spaces, and a target label space contained in the source label space.Domain shift remains possible because source and target feature distributions may differ even when label spaces match.
- Adversarial Baseline: Standard adversarial adaptation trains feature extractors and a binary domain classifier to reduce source-target divergence in feature space.The classifier labels source samples as 1 and target samples as 0; feature extractors move in the opposite direction.
- Training Procedure: The source feature extractor and classifier are trained separately for source classification, then the source model is fixed during later adaptation.The source classification loss uses cross entropy, and the source model is trained before target-feature adaptation.
- Importance Weighting: The first domain classifier estimates source-sample importance from source and target features, assigning smaller weights to samples likely from source-only outlier classes.A classifier output near 1 indicates high likelihood of coming from the source distribution and potentially from an outlier region with little target density.
- Weighted Alignment: The second domain classifier plays the minimax game with weighted source samples and target samples, while the first classifier does not back-propagate gradients to update target features.This separation preserves the theoretical divergence-reduction interpretation for the weighted source and target densities.
- Weighted Alignment: Weighted adversarial adaptation reduces the Jensen-Shannon divergence between the weighted source density and target density, reaching its optimum when the weighted source density equals the target density.The weights are normalized so the weighted source density remains a probability density function.
- Target Structure Preservation: Target entropy minimization constrains the target feature extractor to encourage low-density separation between classes.It is applied only to target features because applying it to both feature extractor and classifier could lock early, incorrect target assignments.
- Training Procedure: Training proceeds in stages: source features and classifier are pretrained and fixed, then the domain classifiers and target feature extractor are optimized simultaneously.The second minimax game can be optimized iteratively or with a gradient reversal layer.
4. Experiments
Experiments evaluate the method on partial domain adaptation benchmarks with fewer target classes, comparing it against standard and adversarial baselines. Results show strong partial-transfer performance, effective outlier weighting, and no noticeable degradation in the traditional non-partial setting.
- Datasets and settings: Experiments use Office+Caltech-10, Office-31, and Caltech256→Office10 transfers with fewer target classes than source classes.Office+Caltech-10 uses 10-class sources and 5-class targets; Office-31 uses 31-category sources and 10-category targets; Caltech-256 transfers to Office-10.
- Baselines and settings: The method is compared with AlexNet+bottleneck, RevGrad, RTN, ADDA-grl, and SAN using matched deep adaptation settings.AlexNet is fine-tuned from ImageNet, and the baselines include both general domain adaptation methods and the partial-transfer method SAN.
- Partial domain adaptation results: The proposed methods outperform AlexNet+bottleneck, RevGrad, RTN, and ADDA-grl by a large degree and are comparable to SAN on most datasets.The evaluation covers Tables 1–3 and uses SAN results copied from its original paper.
- Further analysis: The weighting scheme identifies likely outlier-class source samples and aligns target data more closely with selected shared source classes than unweighted ADDA-grl.Figures 4–6 show target alignment with relevant source classes and mostly low weights for source outlier classes.
- Partial domain adaptation results: 86.73% average performance is obtained by the proposed γ = 0 variant versus 85.64% for SAN, with many fewer parameters.The γ = 0 variant removes target-domain entropy minimization while retaining the weighting scheme.
- Sensitivity to target-class count: When target classes decrease from 31 to 5, the method continues improving over AlexNet+bottleneck, with larger relative gains for fewer target classes.On A→W, ADDA-grl accuracy decreases as the number of target classes becomes smaller, whereas the proposed method improves consistently.
- Additional evaluations: 71.5% accuracy with shared features rises to 76.3% with unshared features on A31→W10, while full-protocol results show no noticeable degradation.The comparison supports using unshared feature extractors for this domain pair; Table 4 evaluates the traditional non-partial setting.
5. Conclusion
The paper extends adversarial-net unsupervised adaptation to partial domain adaptation with an activation-based weighting scheme for source outlier samples. Experiments report strong performance relative to previous methods, while visualizations examine class alignment and learned sample weights.
- The method extends adversarial nets-based unsupervised domain adaptation to partial domain adaptation.
- An activation-based weighting scheme detects source samples from outlier classes and reduces shift between target data and source data within target classes.
- The experiments report that the proposed method outperforms previous domain adaptation methods and is comparable to state-of-the-art partial transfer methods.
- Figure 3 plots accuracy as the number of target classes varies for baseline methods and the proposed method.
- Figure 4 visualizes baseline and proposed-method activations, while Figures 5 and 6 show class alignment and learned source-sample weights.