Source-linked AI summary
Improve Unsupervised Domain Adaptation with Mixup Training
Shen Yan, Huan Song, Nanxiang Li, Lincan Zou, Liu Ren
TL;DR
Unsupervised domain adaptation seeks to transfer models from labeled source domains to unannotated targets, but domain-invariant representations and independently imposed constraints can be insufficient. The paper introduces cross-domain mixup with a feature-level consistency regularizer, alongside intra-domain mixup and adversarial learning, and reports improved state-of-the-art performance across image classification and human activity recognition.
Problem
Domain-invariant representations can be insufficient for unannotated target domains, while existing training constraints often treat source and target domains independently.
Method
IIMT enforces cross-domain constraints through mixup with inferred target labels, adds a feature-level consistency regularizer for large domain discrepancies, and includes intra-domain mixup with adversarial learning.
Results
IIMT significantly outperforms state-of-the-art methods across diverse image classification and human activity recognition tasks.
Takeaways & Limitations
Cross-domain mixup provides interpolated virtual-label supervision intended to improve target-domain generalization beyond training with source labels alone.
Takeaways & Limitations
Inter-domain mixup can become less effective when source-target domain discrepancy is very large, especially with joint adversarial training.
Abstract
from arXiv · showhide
Unsupervised domain adaptation studies the problem of utilizing a relevant source domain with abundant labels to build predictive modeling for an unannotated target domain. Recent work observe that the popular adversarial approach of learning domain-invariant features is insufficient to achieve desirable target domain performance and thus introduce additional training constraints, e.g. cluster assumption. However, these approaches impose the constraints on source and target domains individually, ignoring the important interplay between them. In this work, we propose to enforce training constraints across domains using mixup formulation to directly address the generalization performance for target data. In order to tackle potentially huge domain discrepancy, we further propose a feature-level consistency regularizer to facilitate the inter-domain constraint. When adding intra-domain mixup and domain adversarial learning, our general framework significantly improves state-of-the-art performance on several important tasks from both image classification and human activity recognition.
1. INTRODUCTION
Unsupervised domain adaptation must transfer models from labeled source domains to unannotated targets despite domain shifts. IIMT addresses the overlooked cross-domain interplay by combining inter-domain mixup, feature consistency, intra-domain mixup, and adversarial learning.
- Unsupervised domain adaptation transfers models trained on labeled source data to related target domains with scarce or unavailable labels.Domain shifts make deployment difficult when target distributions are novel and unannotated.
- Adversarial domain-invariant representations may be insufficient because a classifier performing well on both domains may not exist.Existing methods therefore add constraints such as target conditional entropy minimization, virtual adversarial training, or source metric-learning losses.
- Prior approaches impose training constraints independently within selected domains, leaving source-target interplay unexplored.The paper proposes cross-domain constraints to address this limitation.
- Inter-domain mixup uses inferred target labels to provide interpolated virtual-label supervision between source and target domains.Mixup encourages linear prediction behavior, while target labels become more accurate as mixup and adversarial training progress.
- IIMT combines inter-domain and intra-domain mixup with a feature-level consistency regularizer and domain adversarial learning.The regularizer facilitates mixup under potentially large domain discrepancy, and experiments cover visual recognition and human activity recognition.
2. PROPOSED APPROACH
The IIMT framework combines inter-domain mixup, feature-level consistency, intra-domain mixup, and domain adversarial training to improve unsupervised adaptation under domain discrepancy.
- Inter-domain Mixup Training: IIMT applies mixup across source and target domains after inferring target virtual labels, enforcing linear prediction behavior across domains.The model interpolates source samples and target samples with their corresponding source and virtual labels.
- Inter-domain Mixup Training: Target virtual labels are formed by averaging predictions over augmented inputs and sharpening the result with temperature T.Using T < 1 produces sharper predicted distributions and helps reduce conditional entropy during training.
- Consistency Regularizer: For large domain discrepancies, a feature-level consistency regularizer matches the feature of mixed inputs with the mixed feature representation using mean squared error.This imposes the linearity constraint in latent space, making linear prediction behavior easier for the classifier to satisfy.
- Domain Adversarial Training: The framework adds domain adversarial training so a shared encoder learns domain-invariant features through a discriminator trained against the encoder.The implementation uses the DANN framework and applies its adversarial loss to mixed source and target samples.
- Intra-domain Mixup Training: Intra-domain mixup uses source labels and target virtual labels within each domain, applying label-level losses without feature-level linearity.The paper motivates this component as a way to encourage prediction smoothness within each domain.
- Overall Objective: The final objective combines virtual-label, domain-adversarial, feature-consistency, and source- and target-mixup losses with separate weights.The target mixup weight is linearly scheduled from zero to a predefined maximum, while other weights are fixed to one.
3. EXPERIMENTAL RESULTS
The evaluation covers image classification and human activity recognition, with results showing that the proposed approach improves adaptation across visual and sensor-based tasks. Ablations further indicate contributions from intra- and inter-domain mixup components.
- Evaluation Scope: The evaluation spans visual domain adaptation benchmarks and human activity recognition on OPPORTUNITY and WiFi datasets.Visual tasks use MNIST, MNIST-M, SVHN, SYN DIGITS, CIFAR-10, and STL-10; HAR uses OPPORTUNITY and WiFi.
- Visual Recognition: The proposed approach outperforms DIRT-T on MNIST →MNISTM and SYN DIGITS →SVHN, while matching VADA on SVHN →MNIST and remaining below DIRT-T.These comparisons are reported for digits-classification UDA tasks.
- Visual Recognition: 3.1% margin-of-improvement is achieved for CIFAR →STL, while STL →CIFAR improves by 8.1% over VADA and 6.3% over DIRT-T.The results also show that adaptation is important when STL is much smaller and non-adapted performance is poor.
- Human Activity Recognition: Without domain adaptation, all classifiers have inferior performance on the target OPPORTUNITY subject, while the proposed method achieves an averaged improvement over 0.04.VADA is only slightly better than conventional adversarial training in this setting.
- Human Activity Recognition: The proposed approach significantly outperforms VADA/DIRT-T on the WiFi room-adaptation task.The experiment adapts from Room A to Room B for seven physical activities.
- Ablation Analysis: Each proposed mixup component contributes to performance gains, with inter-domain components collectively improving results by 6% for STL →CIFAR and 0.02 for 1 →2.Intra-domain mixup alone is slightly better than VADA and is described as comparable to VAT for enforcing the locally-Lipschitz constraint.
4. CONCLUSION
The paper proposes cross-domain training constraints through mixup linearity and adds a consistency regularizer for large domain discrepancies. The resulting IIMT framework combines inter- and intra-domain mixup and outperforms state-of-the-art methods across diverse applications.
- 4. CONCLUSION: The IIMT framework imposes cross-domain training constraints through mixup linearity and uses a consistency regularizer under large domain discrepancy.The framework incorporates both inter-domain and intra-domain mixup training.
- 4. CONCLUSION: The general IIMT framework outperforms state-of-the-art methods in diverse application areas.
- 4. CONCLUSION: The conclusion identifies cross-domain constraints as the paper’s central domain-adaptation focus.