Source-linked AI summary
Multi-Adversarial Domain Adaptation
Zhongyi Pei, Zhangjie Cao, Mingsheng Long, Jianmin Wang
TL;DR
Single-discriminator domain adaptation aligns whole source and target distributions without accounting for multimode structure, risking under- or negative transfer. MADA uses multiple class-wise discriminators for fine-grained alignment, achieving state-of-the-art results on standard domain adaptation datasets.
Problem
Existing domain adversarial methods align whole source and target distributions without exploiting multimode structures, risking under- or negative transfer.
Method
MADA uses multiple probability-weighted, class-wise domain discriminators to fine-grainedly align source and target distributions by relevant modes.
Results
MADA outperforms comparison methods on most transfer tasks and achieves state-of-the-art results across standard domain adaptation datasets.
Takeaways & Limitations
Fine-grained multimode alignment simultaneously promotes positive transfer and circumvents negative transfer within the evaluated domain adaptation tasks.
Abstract
from arXiv · showhide
Recent advances in deep domain adaptation reveal that adversarial learning can be embedded into deep networks to learn transferable features that reduce distribution discrepancy between the source and target domains. Existing domain adversarial adaptation methods based on single domain discriminator only align the source and target data distributions without exploiting the complex multimode structures. In this paper, we present a multi-adversarial domain adaptation (MADA) approach, which captures multimode structures to enable fine-grained alignment of different data distributions based on multiple domain discriminators. The adaptation can be achieved by stochastic gradient descent with the gradients computed by back-propagation in linear-time. Empirical evidence demonstrates that the proposed model outperforms state of the art methods on standard domain adaptation datasets.
Introduction
Domain shift limits the transferability of deep representations, while existing adversarial methods align whole distributions without modeling their multimode structure. The paper introduces MADA to capture these structures and enable fine-grained cross-domain alignment.
- Deep representations trained on one dataset often generalize poorly to novel datasets and tasks because of dataset bias or domain shift.
- Existing domain adversarial methods mainly align whole source and target distributions without considering complex multimode structures.
- Domain adaptation must enhance positive transfer by matching multimode structures while preventing negative transfer from falsely aligning modes across domains.
- MADA captures multimode structures to enable fine-grained alignment of different data distributions.
- False alignment can associate a source cat class with a target dog class, producing incorrect final classification.
Related Work
Prior work uses deep representations and adversarial training to reduce source–target discrepancy, but single-discriminator methods may mishandle multimode distributions. MADA addresses this through class-wise, probability-weighted discriminators for fine-grained alignment that promotes positive transfer and avoids negative transfer.
- Transfer learning foundations: Transfer learning reduces manual-labeling burdens across domains, with distribution discrepancy remaining its main technical difficulty.Deep networks can reduce, but not remove, cross-domain discrepancy, leaving target-task risk unbounded.
- Prior deep adaptation: Deep domain-adaptation methods match distribution mean embeddings or train feature extractors adversarially against domain discriminators.These approaches extend deep CNNs with adaptation layers or domain-discriminator subnetworks.
- Limitations of prior adversarial adaptation: Single-discriminator adversarial adaptation may under-transfer or negatively transfer when source and target distributions contain complex multimode structures.Under-transfer results from incomplete mode matching, whereas negative transfer results from falsely aligning corresponding modes.
- MADA approach: MADA splits one domain discriminator into K class-wise discriminators, using source labels and predicted target-label probabilities to associate each example with relevant discriminators.The probability distribution from the label predictor determines how strongly each data point attends to each class-wise discriminator.
- MADA benefits: MADA enables fine-grained adaptation by matching each point only through relevant domain discriminators, avoiding hard target assignments and filtering irrelevant classes.This probability-weighted strategy is described as enhancing positive transfer while circumventing negative transfer from false mode alignment.
Experiments
Experiments evaluate MADA against shallow and deep transfer-learning methods on Office-31 and ImageCLEF-DA under standard unsupervised adaptation protocols. MADA outperforms comparison methods on most Office-31 tasks, especially difficult transfers, while its analyses show fine-grained alignment, stable convergence, and similar computational complexity to RevGrad.
- Datasets and evaluation: Experiments use Office-31’s 4,652 images across 31 categories and three domains, evaluating transfer tasks A →W, D →W, and W →D.Office-31 comprises Amazon, Webcam, and DSLR domains.
- Datasets and evaluation: ImageCLEF-DA contains 600 images per domain across 12 shared categories from Caltech-256, ImageNet ILSVRC 2012, and Pascal VOC 2012.The three domains are equal in size, providing an alternative benchmark.
- Experimental setup: MADA is compared with TCA, GFK, DDC, DAN, RTN, and RevGrad using all labeled source and unlabeled target examples, with average accuracy and standard error reported over three random experiments.The evaluation follows standard unsupervised domain-adaptation protocols for both datasets.
- Results: MADA outperforms all comparison methods on most transfer tasks, substantially improving difficult transfers while achieving comparable accuracies on easy tasks D →W and W →D.The difficult tasks listed are A →W, A →D, D →A, and W →A, where source and target domains differ substantially.
- Feature analysis: Feature visualizations show that MADA makes domains more indistinguishable while separating categories better than RevGrad, producing the best adaptation accuracy.RevGrad aligns domains without adequately discriminating different categories, whereas MADA integrates discriminative information for fine-grained alignment.
- Convergence and complexity: MADA has similarly stable convergence to RevGrad while outperforming it throughout convergence, with similar computational complexity because its multiple discriminators occupy limited additional computation.The convergence comparison uses ResNet on task A →W.
Conclusion
The paper presents a multi-adversarial domain adaptation approach for effective deep transfer learning. It exploits discriminative structures for fine-grained distribution alignment beyond matching feature distributions across domains.
- The proposed method enables effective deep transfer learning through a novel multi-adversarial domain adaptation framework.
- Unlike prior domain adversarial methods, it exploits complex multimode and discriminative structures rather than only matching feature distributions across domains.
- The framework enables fine-grained alignment of distributions across domains.