Source-linked AI summary
Learning to Transfer Examples for Partial Domain Adaptation
Zhangjie Cao, Kaichao You, Mingsheng Long, Jianmin Wang, Qiang Yang
TL;DR
Partial domain adaptation must transfer from a labeled source domain to an unlabeled target domain with an unknown subset of source labels, while avoiding irrelevant examples and negative transfer. ETN learns transferable representations and weights source examples using a transferability quantifier, jointly controlling their effects on the source classifier and domain discriminator. Across several benchmark datasets, ETN achieves state-of-the-art or strong performance for partial domain adaptation tasks.
Problem
Partial domain adaptation must identify useful shared-class source examples when target labels are unknown, while avoiding negative transfer from outlier source classes.
Method
ETN quantifies source-example transferability using target similarity and discriminative information, weighting contributions to both the source classifier and domain discriminator.
Results
ETN achieves state-of-the-art results on several benchmark datasets and outperforms other methods with respect to average accuracy across tasks and base networks.
Takeaways & Limitations
ETN promotes transfer among relevant examples while mitigating the influence of irrelevant examples in partial domain adaptation.
Takeaways & Limitations
Partial domain adaptation remains constrained by outlier source classes and by source-target domain gaps, while deep representations reduce but do not remove cross-domain discrepancy.
Abstract
from arXiv · showhide
Domain adaptation is critical for learning in new and unseen environments. With domain adversarial training, deep networks can learn disentangled and transferable features that effectively diminish the dataset shift between the source and target domains for knowledge transfer. In the era of Big Data, the ready availability of large-scale labeled datasets has stimulated wide interest in partial domain adaptation (PDA), which transfers a recognizer from a labeled large domain to an unlabeled small domain. It extends standard domain adaptation to the scenario where target labels are only a subset of source labels. Under the condition that target labels are unknown, the key challenge of PDA is how to transfer relevant examples in the shared classes to promote positive transfer, and ignore irrelevant ones in the specific classes to mitigate negative transfer. In this work, we propose a unified approach to PDA, Example Transfer Network (ETN), which jointly learns domain-invariant representations across the source and target domains, and a progressive weighting scheme that quantifies the transferability of source examples while controlling their importance to the learning task in the target domain. A thorough evaluation on several benchmark datasets shows that our approach achieves state-of-the-art results for partial domain adaptation tasks.
1. Introduction
Partial domain adaptation transfers knowledge from a large labeled source domain to a smaller unlabeled target domain whose labels form a subset of the source labels. Its central challenge is selecting useful shared-class examples while avoiding negative transfer from irrelevant source classes.
- Problem: Partial domain adaptation relaxes identical-label assumptions by allowing the source label space to subsume the target label space.This setting supports transfer from large many-label datasets to small few-label target domains.
- Problem: Target labels are unknown during training, and many source classes may be irrelevant or outliers for the target task.Forcefully aligning these outlier classes can cause negative transfer.
- Prior approaches: Existing PDA methods weight data points to reduce irrelevant examples during domain alignment but may leave the source classifier affected by outlier classes.They also evaluate source-sample transferability without fully considering discriminative and multimodal structure.
- Proposed approach: ETN quantifies source-example transferability from similarities to the target domain and uses the resulting weights for both the source classifier and domain discriminator.Its transferability quantifier further reveals discriminative structure, improving detection of irrelevant examples.
- Proposed approach: ETN confines both the source classifier and domain-adversarial network to an automatically discovered shared label space.Experiments on Office-31, Office-Home, ImageNet-1K, and Caltech-256 report state-of-the-art results.
2. Related Work
Domain adaptation methods learn representations or explicitly align distributions, but deep representations reduce rather than remove cross-domain discrepancy. Related PDA and open-set settings differ in their assumptions about shared and target classes.
- Domain adaptation: Domain adaptation bridges domains with different distributions to reduce the burden of annotating target data.Deep representations aim to suppress domain-related variation and expose invariant factors across populations.
- Domain adaptation: Deep representations can reduce, but not remove, cross-domain discrepancy, so deep learning alone cannot bound target-task generalization risk.This limitation motivates methods combining deep learning with explicit domain adaptation.
- Domain adaptation: Domain adaptation methods extend deep networks with adaptation layers for matching distribution statistics or domain discriminators trained adversarially.These approaches explicitly address distribution discrepancy beyond representation learning alone.
- Partial domain adaptation: Existing PDA approaches down-weight outlier classes for domain discriminators, but the source classifier may still train on all classes.This can leave shared-class performance distracted by outlier classes.
- Related settings: Open Set Domain Adaptation often assumes shared source-target classes are known during training, whereas PDA treats target classes as entirely unknown.The paper identifies extending ETN to this open-set scenario as an interesting direction.
3. Example Transfer Network
ETN addresses partial domain adaptation by weighting source examples according to transferability, using auxiliary domain and label predictors to distinguish relevant shared-class examples from outliers. It combines these weights with domain-adversarial learning and entropy minimization in an end-to-end minimax framework.
- Motivation: Partial domain adaptation has unknown target labels and a source label space containing outlier classes, making whole-domain alignment prone to negative transfer.The goal is to reduce shift between shared classes while suppressing irrelevant source examples.
- Transferability Weighting Framework: The framework learns domain-invariant features with a domain discriminator while minimizing source classification loss and incorporating entropy minimization on unlabeled target examples.Feature extraction and domain discrimination are trained adversarially, with a trade-off parameter for labeled source and unlabeled target losses.
- Transferability Weighting Framework: ETN assigns each source example a transferability weight that increases contributions from relevant examples and decreases contributions from irrelevant or outlier examples.The weights are applied to both the source classifier and the domain discriminator.
- Minimax Optimization Problem: The complete ETN model jointly trains the feature extractor, source classifier, domain discriminator, auxiliary label predictor, and auxiliary discriminator through minimax optimization.This unified framework transfers relevant examples and diminishes outlier examples for both classification and domain alignment.
- Example Transferability Quantification: An auxiliary domain discriminator estimates transferability from source-target similarity, but its domain-only outputs may confuse transferable and irrelevant source examples.Shared-class source examples are assumed to be closer to the target domain than outlier-class examples.
- Example Transferability Quantification: ETN adds an auxiliary label predictor whose leaky-softmax outputs encode discriminative information for refining the auxiliary discriminator’s transferability estimates.Source examples tend to produce output sums closer to 1, whereas target examples produce sums closer to 0.
4. Experiments
Experiments evaluate ETN across partial domain adaptation benchmarks, backbones, ablations, and analyses of feature quality, class overlap, convergence, and example weights. ETN consistently outperforms competing methods and remains effective across datasets and network architectures.
- Results: ETN outperforms all other methods in average accuracy across Office-Home, Office-31, and ImageNet-Caltech benchmarks.The evaluation includes twelve Office-Home tasks, six Office-31 tasks, and two ImageNet-Caltech tasks using ResNet-50.
- Results: ETN also performs well with different base networks, including ResNet-50 and VGG, across the evaluated datasets.Office-31 results are additionally evaluated with a VGG backbone.
- Results: ETN outperforms SAN and IWAN on most tasks, indicating stronger discrimination between shared and outlier classes.The authors attribute this advantage to transferring relevant examples more accurately.
- Results: On ImageNet-Caltech, ETN achieves larger gains over SAN and IWAN when 916 source classes are outliers versus 84 shared classes.The result indicates robustness when the number of outlier classes is large relative to the shared classes.
- Ablation Study: ETN's classifier weighting reduces outlier influence, while its auxiliary label predictor supplies discriminative information for example weighting.Removing either component lowers performance in the reported ablations.
- Analysis: ETN produces clearer target feature clusters than DANN, SAN, and IWAN, and its accuracy remains stable as the number of target classes varies.The feature visualization uses t-SNE embeddings on the A→W task, while class-overlap experiments vary the target label-space size.
5. Conclusion
The paper presents Example Transfer Network (ETN), a discriminative and robust approach to partial domain adaptation that quantifies source-example transferability and down-weights outlier influence.
- ETN integrates discriminative information into its transferability quantifier to assess source-example relevance.
- ETN down-weights outlier source examples' negative influence on both the source classifier and the domain discriminator.
- ETN performs strongly on partial domain adaptation tasks according to the paper's evaluation.