Source-linked AI summary
Partial Transfer Learning with Selective Adversarial Networks
Zhangjie Cao, Mingsheng Long, Jianmin Wang, Michael I. Jordan
TL;DR
Existing transfer methods assume shared label spaces, but practical transfer from large labeled domains can involve target labels that form only a subset of source labels, allowing irrelevant source classes to cause negative transfer. SAN addresses this setting by selectively transferring relevant source data while aligning distributions in the shared label space. It outperforms comparison methods across benchmark partial-transfer tasks.
Problem
Partial transfer learning must adapt from a source domain whose label space contains the target label space, without knowing which source classes are shared; irrelevant source data can cause negative transfer.
Method
SAN uses multiple domain discriminators and source-label prediction probabilities to selectively match each target point with relevant source data while aligning shared-label distributions.
Results
SAN outperforms all comparison methods across six Office-31, three Caltech-Office, and two ImageNet-Caltech tasks.
Takeaways & Limitations
Selective adversarial adaptation can transfer relevant knowledge and alleviate negative transfer when the target label space is a subspace of the source label space.
Abstract
from arXiv · showhide
Adversarial learning has been successfully embedded into deep networks to learn transferable features, which reduce distribution discrepancy between the source and target domains. Existing domain adversarial networks assume fully shared label space across domains. In the presence of big data, there is strong motivation of transferring both classification and representation models from existing big domains to unknown small domains. This paper introduces partial transfer learning, which relaxes the shared label space assumption to that the target label space is only a subspace of the source label space. Previous methods typically match the whole source domain to the target domain, which are prone to negative transfer for the partial transfer problem. We present Selective Adversarial Network (SAN), which simultaneously circumvents negative transfer by selecting out the outlier source classes and promotes positive transfer by maximally matching the data distributions in the shared label space. Experiments demonstrate that our models exceed state-of-the-art results for partial transfer learning tasks on several benchmark datasets.
1 Introduction
Transfer learning can reuse labeled data across domains, but partial transfer learning addresses cases where the source label space contains target categories and irrelevant source classes can cause negative transfer. SAN selectively excludes those outlier classes while aligning shared-label distributions, achieving state-of-the-art benchmark results.
- Deep networks improve machine-learning performance but typically require massive labeled datasets, making transfer from related labeled source domains attractive.Manual labeling sufficient data for diverse domains is often prohibitive, while distribution shifts complicate adaptation.
- Partial transfer learning assumes the target label space is a subspace of the source label space.The setting considers transferring from large datasets that subsume the categories of smaller target datasets.
- Outlier source classes can cause negative transfer, so matching the whole source and target domains is ineffective for partial transfer.The paper illustrates this with irrelevant source classes such as “sofa” when discriminating target classes such as “soccer-ball” and “binoculars.”
- SAN selects outlier source classes and aligns source and target distributions in the shared label space through deep adversarial adaptation.The approach is trained end-to-end to promote positive transfer from relevant data while alleviating negative transfer from irrelevant data.
- SAN models exceed state-of-the-art results for partial transfer learning on public benchmark datasets.
2 Related Work
Transfer learning reduces labeling burden by adapting models across domains, but cross-domain distribution discrepancy remains difficult. Existing deep adaptation methods improve performance while generally relying on shared source and target label spaces, an assumption violated by partial transfer learning.
- Transfer learning bridges domains or tasks to reduce the burden of manual labeling across machine learning, computer vision, and natural language processing.
- Deep networks can reduce, but not remove, cross-domain discrepancy, leaving target-task risk unbounded.
- Deep domain-adaptation methods match distribution embeddings or use domain-adversarial discriminators to learn domain-invariant features.
- These methods generally assume shared label spaces, whereas partial transfer learning transfers models from large domains to smaller domains with unknown categories.The paper presents this work as the first adversarial-network approach addressing partial transfer learning.
3 Partial Transfer Learning
Partial transfer learning addresses settings where the target label space is a subset of the source label space, making whole-domain matching vulnerable to negative transfer. SAN filters outlier source classes while matching source and target distributions in their shared label space through selective adversarial adaptation.
- Partial transfer learning assumes the target label space Ct is a subspace of the source label space Cs, with |Cs| > |Ct|.
- Unknown target classes make it necessary to reduce both cross-domain distribution discrepancy and the influence of source data from outlier classes Cs\Ct.Existing methods generally match whole source and target distributions despite their different label spaces.
- SAN uses |Cs| class-wise domain discriminators and assigns each data point to them using the label predictor’s probability distribution over source classes.This probability-weighted loss enables fine-grained adaptation rather than hard assignment to a single discriminator.
- Class-level weighting down-weights discriminators associated with outlier source classes because their target probabilities are sufficiently small.
- SAN simultaneously filters outlier source classes to circumvent negative transfer and maximally matches pCt with q in the shared label space Ct.The unified approach is presented as an end-to-end selective adversarial network.
4 Experiments
Experiments evaluate SAN on partial transfer tasks across Office-31, Caltech-Office, and ImageNet-Caltech, comparing it with established transfer methods and analyzing its selective mechanism, class-count sensitivity, convergence, and feature representations. SAN outperforms comparison methods across all evaluated tasks and shows evidence of reduced negative transfer.
- Experimental Setup: SAN is evaluated on six Office-31 tasks, three Caltech-Office tasks, and two ImageNet-Caltech tasks using AlexNet, DAN, RevGrad, and RTN as comparison methods.The experiments use labeled source data and unlabeled target data for unsupervised transfer learning, with average classification accuracy measured across three random experiments.
- Results: SAN outperforms all comparison methods on every evaluated task, with especially large gains when source and target domains have small class sets.The reported improvements also include tasks with large source domains and small target domains, as well as large-scale source and target domains.
- Results: Previous adversarial and MMD-based methods can perform worse than AlexNet because matching all source classes transfers outlier source data absent from the target label space.SAN improves over these methods by eliminating source classes irrelevant to the target domain, according to the paper’s analysis.
- Ablation Analysis: SAN outperforms SAN-selective and SAN-entropy, supporting the combined use of selective adversarial learning and entropy minimization.Entropy minimization is reported to help especially when many irrelevant classes create a large source–target class-count gap.
- Empirical Analysis: As target classes decrease, RevGrad degrades quickly, while SAN’s margin over RevGrad increases; SAN also outperforms RevGrad with all 31 target classes.SAN’s performance first decreases from 31 to 20 target classes and then increases from 20 to 10 as the transfer problem becomes easier.
- Convergence Analysis: During training, DAN and RevGrad test errors increase, RTN converges quickly to a higher error, and SAN converges quickly and stably to the lowest test error.The paper interprets this behavior as simultaneous positive transfer and alleviation of negative transfer.
- Feature Visualization: t-SNE visualizations compare bottleneck representations from DAN, RevGrad, RTN, and SAN using both class and domain information on A 31 →W 10.The analysis examines whether target data are separated from irrelevant source classes and how source–target features align.
5 Conclusion
The paper concludes that Selective Adversarial Networks address partial transfer learning by excluding outlier source classes while matching distributions in the shared label space.
- SAN selects outlier source classes to circumvent negative transfer in partial transfer learning.
- SAN promotes positive transfer by maximally matching data distributions in the shared label space.
- Extensive experiments testify that SAN successfully tackles settings where the source label space subsumes the target label space.