Source-linked AI summary
Partial Adversarial Domain Adaptation
Zhangjie Cao, Lijia Ma, Mingsheng Long, Jianmin Wang
TL;DR
Conventional domain adaptation assumes identical label spaces, whereas practical transfer from large source datasets to smaller targets can involve source-only classes and negative transfer. PADA addresses this partial domain adaptation setting by down-weighting outlier source data while aligning shared-label feature distributions, and experiments report state-of-the-art results across public partial-adaptation datasets.
Problem
Existing domain adaptation methods assume identical label spaces, but partial adaptation has a source label space that subsumes the target and can suffer negative transfer from outlier source classes.
Method
PADA automatically down-weighs irrelevant source classes for the classifier and domain adversary while aligning source and target feature distributions in the shared label space.
Results
PADA outperforms all comparison methods on the evaluated Office-Home, Office-31, and ImageNet-Caltech partial domain adaptation tasks.
Takeaways & Limitations
Partial domain adaptation can jointly reduce negative transfer from outlier source data and promote positive transfer from shared classes.
Takeaways & Limitations
Related methods may remain constrained by shared-label assumptions, and attention-based localization may fail for images lacking target-related regions.
Abstract
from arXiv · showhide
Domain adversarial learning aligns the feature distributions across the source and target domains in a two-player minimax game. Existing domain adversarial networks generally assume identical label space across different domains. In the presence of big data, there is strong motivation of transferring deep models from existing big domains to unknown small domains. This paper introduces partial domain adaptation as a new domain adaptation scenario, which relaxes the fully shared label space assumption to that the source label space subsumes the target label space. Previous methods typically match the whole source domain to the target domain, which are vulnerable to negative transfer for the partial domain adaptation problem due to the large mismatch between label spaces. We present Partial Adversarial Domain Adaptation (PADA), which simultaneously alleviates negative transfer by down-weighing the data of outlier source classes for training both source classifier and domain adversary, and promotes positive transfer by matching the feature distributions in the shared label space. Experiments show that PADA exceeds state-of-the-art results for partial domain adaptation tasks on several datasets.
1 Introduction
Domain adaptation leverages labeled source data to address limited target labels, but conventional methods assume identical label spaces. This paper introduces partial domain adaptation, where the source label space subsumes the target, and proposes PADA to reduce negative transfer while aligning shared classes.
- Motivation: Domain adaptation leverages labeled data from a related source domain to address target tasks with limited labeled data.Distribution shifts across domains hinder predictive-model generalization.
- Existing domain adaptation: Existing methods generally assume identical source and target label spaces while learning domain-invariant representations without target labels.They apply a source-trained classifier to the target domain.
- Motivation: Large repositories such as ImageNet-1K motivate transferring models from broad source domains to smaller target datasets whose label spaces are subsumed by the source.Finding a source domain with an identical label space is difficult in real applications.
- Partial domain adaptation: Partial domain adaptation assumes the target label space is a subspace of the source label space, making outlier source classes possible.For example, the source may contain TV, chair, and mug while the target contains only chair and mug.
- Partial domain adaptation: Outlier source classes can trigger negative transfer, while identifying them is difficult because target classes are unknown during training.Matching the whole source and target domains is therefore ineffective in this setting.
- PADA: PADA down-weighs irrelevant source data and aligns source-target feature distributions in the shared label space within an end-to-end framework.The method aims to promote positive transfer and alleviate negative transfer simultaneously.
2 Related Work
Prior deep domain adaptation methods reduce feature-distribution discrepancies through moment matching or adversarial learning, but partial label-space overlap limits their applicability. Related approaches address different label spaces or negative transfer, yet may remain constrained by shared-label assumptions or attention failures.
- Domain adaptation methods: Unsupervised domain adaptation is practical because it requires no labeled target data, unlike supervised domain adaptation.Deep adaptation methods reduce discrepancies between feature distributions across domains.
- Deep domain adaptation: Moment-matching and adversarial modules are embedded in deep networks to learn representations transferable across domains.Adversarial methods use a domain discriminator in a two-player game.
- Limitations: Existing methods may be restricted by the assumption that source and target domains share the same label space, which fails in partial domain adaptation.This limitation motivates methods designed for differing label spaces.
- Limitations: Adaptive Deep Learning reduces negative transfer by localizing regions related to domain shifts and shared regions, but its attention mechanism may fail when images lack target-related regions.The failure can involve incorrectly localizing related regions.
3 Partial Adversarial Domain Adaptation
Partial domain adaptation addresses settings where the target label space is a subset of the source label space, requiring simultaneous suppression of unrelated source classes and alignment of shared distributions. PADA uses target-averaged class weights to down-weigh outlier source data in both classification and adversarial adaptation.
- Problem setting: Partial domain adaptation assumes the target label space is a subspace of the source label space, with unlabeled target data and more source than target classes.The source and target distributions differ, and the target label space is unknown during training.
- Approach: PADA aligns source and target feature distributions in the shared label space while automatically identifying and down-weighing irrelevant source data.This extends domain adversarial adaptation to the partial label-space setting.
- Challenges: Whole-domain alignment can cause negative transfer by forcing target data toward outlier source classes, especially when the outlier label space is much larger.The paper identifies filtering unrelated source data as necessary for mitigating this effect.
- Approach: Class weights are computed by averaging source-class predictions over all target data, reducing the influence of occasional target-label prediction mistakes.Outlier source classes receive smaller weights because target data should be dissimilar to them.
- Optimization: The class-weight vector is applied to both the source classifier and the partial adversarial domain discriminator over source-domain data.The largest weight is used to normalize the vector before applying it in training.
- Optimization: PADA simultaneously mitigates negative transfer by filtering outlier source classes and promotes positive transfer by matching shared-label source and target distributions.Its adversarial framework builds on the two-player game between a feature extractor and domain discriminator used in standard domain adaptation.
4 Experiments
Experiments evaluate PADA across multiple partial domain adaptation datasets and compare it with established transfer-learning methods. Results and analyses show consistent gains, reduced negative transfer, and effective weighting of shared versus outlier source classes.
- Experimental setup: The evaluation covers Office-31, Office-Home, ImageNet-Caltech, and VisDA2017 partial adaptation settings with differing class counts, domain gaps, and dataset scales.Office-31 uses six tasks, ImageNet-Caltech uses two tasks, and VisDA2017 uses two six-class-target tasks.
- Benchmark results: PADA outperforms all comparison methods on the twelve Office-Home, six Office-31, and two ImageNet-Caltech partial adaptation tasks.The gains hold across different domain sizes and source-target domain gaps.
- Negative-transfer analysis: Previous adversarial and MMD-based methods mix all source and target classes, allowing outlier source classes to cause negative transfer in partial adaptation.PADA instead down-weighs outlier source classes while enhancing the influence of shared classes.
- Ablation study: Ablations show that weighting both the source classifier and domain adversary is important for reducing outlier influence and avoiding whole-domain matching.PADA outperforms both PADA-classifier and PADA-adversarial variants.
- Class-weight analysis: PADA assigns much larger weights to shared classes than outlier classes, with some outlier classes receiving nearly zero weights, unlike DANN.DANN retains substantially large weights for outlier source classes on the analyzed tasks.
- Sensitivity and convergence: As target-class count decreases, PADA’s advantage over DANN widens; PADA also converges quickly and stably to the lowest test error.DANN degrades quickly with fewer target classes, while PADA remains comparatively robust and reaches the lowest test error in the convergence analysis.
5 Conclusion
PADA addresses partial domain adaptation, where the source label space subsumes the target, by reducing negative transfer from outlier classes while enhancing transfer in the shared label space.
- PADA down-weighs outlier source classes for both the source classifier and domain adversary.
- PADA promotes positive transfer by matching data distributions within the shared label space.
- Extensive experiments on benchmark datasets support PADA for partial domain adaptation.