Source-linked AI summary
Dual Mixup Regularized Learning for Adversarial Domain Adaptation
Yuan Wu, Diana Inkpen, Ahmed El-Roby
TL;DR
UDA methods can struggle because they do not fully use target-domain class information and because source-target samples may inadequately represent the latent space. DMRL combines pixel-level category and domain mixup regularization to improve representations, with experiments reporting effective performance improvements across UDA benchmarks.
Problem
Adversarial UDA methods may lack target-domain class-aware information and sufficient samples to guarantee discriminative, domain-invariant latent representations.
Method
DMRL applies category and domain mixup regularizations on the pixel level within adversarial domain adaptation.
Results
DMRL achieves the best performance on four of six Office-31 transfer tasks and the best average domain adaptation accuracy on that dataset.
Takeaways & Limitations
The two complementary mixup regularizations learn representations that are both discriminative and domain-invariant for UDA tasks.
Takeaways & Limitations
The method assumes source and target distributions are different but related under covariate shift, with the target task matching the source task.
Abstract
from arXiv · showhide
Recent advances on unsupervised domain adaptation (UDA) rely on adversarial learning to disentangle the explanatory and transferable features for domain adaptation. However, there are two issues with the existing methods. First, the discriminability of the latent space cannot be fully guaranteed without considering the class-aware information in the target domain. Second, samples from the source and target domains alone are not sufficient for domain-invariant feature extracting in the latent space. In order to alleviate the above issues, we propose a dual mixup regularized learning (DMRL) method for UDA, which not only guides the classifier in enhancing consistent predictions in-between samples, but also enriches the intrinsic structures of the latent space. The DMRL jointly conducts category and domain mixup regularizations on pixel level to improve the effectiveness of models. A series of empirical studies on four domain adaptation benchmarks demonstrate that our approach can achieve the state-of-the-art.
1 Introduction
Unsupervised domain adaptation transfers knowledge from labeled source data to unlabeled target domains under domain shift. DMRL addresses limitations of adversarial adaptation by combining category and domain mixup regularization, with evaluations across four benchmark datasets.
- Domain shift can weaken models trained on one labeled dataset when they generalize to new datasets or tasks.
- Domain adaptation transfers knowledge from a label-rich source domain to a label-scarce target domain, while UDA assumes no target labels.
- Adversarial adaptation may overlook target-domain class information and lack enough source-target samples to ensure latent-space domain invariance.
- DMRL applies category and domain mixup regularization at the pixel level to address these two UDA issues.Category mixup enforces consistent predictions and domain mixup explores additional intrinsic structures in the latent space.
- The study evaluates DMRL on four benchmark datasets, including ablation studies and parameter sensitivity analysis.
2 Related Work
Related work spans interpolation-based regularization and domain adaptation, progressing from discrepancy-based methods to deep adversarial approaches. DMRL extends this line by adding pixel-level category and domain mixup for complex UDA tasks.
- Interpolation-based regularization constructs virtual examples from convex combinations of inputs and labels and can smooth model behavior between training samples.
- Deep domain adaptation methods have used MMD and related discrepancy measures to learn representations that are discriminative and transferable.
- Adversarial UDA methods learn domain-invariant features through adversarial training, including DANN, CyCADA, MADA, and GTA.
- DMRL extends adversarial domain adaptation by introducing pixel-level category and domain mixup regularizations for complex, high-dimensional UDA tasks.
3 Method
DMRL extends adversarial domain adaptation with category- and domain-level mixup regularization to learn discriminative, domain-invariant representations. It mixes source and target inputs, uses labeled or pseudo-labeled predictions for category consistency, and incorporates mixed samples into adversarial alignment.
- 3 Method: Mixup constructs virtual samples through convex combinations of paired inputs and labels, with λ sampled from Beta(α, α).The method applies this construction at pixel level within each domain.
- 3 Method: DMRL combines category-level and domain-level mixup regularization with adversarial domain adaptation to improve latent-space discriminability and domain invariance.The feature extractor learns representations, the classifier performs object classification, and the discriminator distinguishes source from target features.
- Category Mixup Regularization: Category mixup enforces prediction consistency on labeled source mixtures and pseudo-labeled target mixtures, exploiting target class-aware information.For target samples, the classifier’s current predictions serve as pseudo-labels, and a penalty encourages mixed-input predictions to match mixed predictions.
- Domain Mixup Regularization: Domain mixup generates intermediate samples within each domain and aligns both original and mixed samples through adversarial learning.This expands the samples used for global distribution alignment and is intended to enrich internal domain structures and enhance latent-space domain invariance.
- 3.3 Training Procedure: Training uses mini-batch SGD: sample source and target batches, mix inputs, update the discriminator with adversarial losses, then update the feature extractor and classifier.The method balances different losses with λs, λt, λd, and λr; λt is tuned on unlabeled test data for different tasks.
4 Experiments
The experiments evaluate DMRL across four domain adaptation benchmarks, including detailed task coverage in Office-31 and dataset structure in ImageCLEF-DA.
- Office-31 evaluation covers six transfer tasks among Amazon, Webcam, and DSLR domains.The tasks are A →W, D →W, W →D, A →D, D →A, and W →A.
- ImageCLEF-DA contains 12 categories shared by Caltech-256, ImageNet ILSVRC 2012, and Pascal VOC 2012 domains.Each domain contains 600 images, with 50 images per category.
31 dataset
DMRL is evaluated across multiple domain-adaptation benchmarks using standard backbones, comparison protocols, ablations, and sensitivity analysis. The reported results show strong performance across transfer tasks, while λt is the main sensitive hyperparameter.
- ImageCLEF-DA results: On ImageCLEF-DA, DMRL outperforms all comparison methods on every transfer task and achieves a new state-of-the-art.Larger improvements are reported for C →I, C →P, and P →C.
- VisDA-2017 and Digits results: On VisDA-2017, DMRL achieves the highest accuracy among compared methods, while Digits gains exceed 1.5% on USPS →MNIST and SVHN →MNIST.MNIST →USPS is described as competitive with existing approaches.
- Ablation studies: The ablation study tests domain mixup, category mixup, labeled and unlabeled category mixup, and a no-adaptation baseline on two Digits tasks.These variants are used to investigate the contributions of DMRL components.
- Parameter sensitivity: Performance is insensitive to α, λs, and λr, whereas accuracy peaks at λt = 2.0 and then decreases as λt increases.The authors therefore fix α, λs, and λr and tune λt across tasks.
5 Conclusion
DMRL combines category- and domain-level pixel mixup to learn discriminative, domain-invariant representations for unsupervised domain adaptation. Experiments show that it improves performance on UDA tasks.
- Category and domain mixup regularizations complement each other to learn discriminative and domain-invariant target-task representations.Category mixup supports consistent predictions and latent-space discriminability, while domain mixup explores internal latent structures and continuity.
- DMRL effectively improves performance on unsupervised domain adaptation tasks.