Source-linked AI summary
Progressive Domain Expansion Network for Single Domain Generalization
Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, Boyang xia
TL;DR
Single-domain generalization must transfer models trained on one domain to unseen domains despite limited guarantees from existing expansion methods. PDEN jointly expands domains under safety and effectiveness constraints and learns invariant representations, achieving up to 15.28% improvement over other single-domain generalization methods.
Problem
Single-domain generalization seeks models that transfer from one training domain to unseen domains, while existing augmentation can lack safety and effectiveness guarantees.
Method
PDEN jointly learns progressive domain expansion and domain-invariant representation subnetworks, using generated domains and contrastive learning.
Results
Up to 15.28% improvement is reported for PDEN compared with other single-domain generalization methods across classification and segmentation experiments.
Takeaways & Limitations
PDEN provides a framework for generating hard unseen domains and learning invariant features, with the generator flexibly replaceable for different domain-expansion types.
Takeaways & Limitations
Existing augmentation may require target-dependent transformation choices and can produce unsafe or ineffective synthetic data or reduce accuracy.
Abstract
from arXiv · showhide
Single domain generalization is a challenging case of model generalization, where the models are trained on a single domain and tested on other unseen domains. A promising solution is to learn cross-domain invariant representations by expanding the coverage of the training domain. These methods have limited generalization performance gains in practical applications due to the lack of appropriate safety and effectiveness constraints. In this paper, we propose a novel learning framework called progressive domain expansion network (PDEN) for single domain generalization. The domain expansion subnetwork and representation learning subnetwork in PDEN mutually benefit from each other by joint learning. For the domain expansion subnetwork, multiple domains are progressively generated in order to simulate various photometric and geometric transforms in unseen domains. A series of strategies are introduced to guarantee the safety and effectiveness of the expanded domains. For the domain invariant representation learning subnetwork, contrastive learning is introduced to learn the domain invariant representation in which each class is well clustered so that a better decision boundary can be learned to improve it's generalization. Extensive experiments on classification and segmentation have shown that PDEN can achieve up to 15.28% improvement compared with the state-of-the-art single-domain generalization methods.
1. Introduction
Single-domain generalization addresses models trained on one source domain but tested on unseen domains, where performance can drop sharply. PDEN expands the source domain progressively while jointly learning domain-invariant representations, aiming to make generated domains safer and more effective.
- Motivation: Single-domain generalization trains on one source domain and evaluates on unseen domains, where model performance can decline substantially.A CNN reaches 99% accuracy on the MNIST test set but only 30% on SVHN.
- Motivation: Domain adaptation aligns source and known target distributions but requires target-domain data, which can be difficult to obtain.
- Motivation: Single-domain generalization instead learns from source-domain data without accessing unseen target-domain data, commonly extending source distributions through augmentation.
- Motivation: Existing augmentation methods require selecting transformation types and magnitudes, while synthetic data may be unsafe, ineffective, or reduce accuracy.
- Proposed Framework: PDEN jointly trains a domain expansion subnetwork and a domain-invariant representation subnetwork, progressively generating domains and applying contrastive learning.The framework is designed to simulate unseen-domain variation while clustering each class for a better decision boundary.
- Results: 15.28% improvement is reported for PDEN compared with other single-domain generalization methods in classification and segmentation experiments.
2. Related Work
Related work covers domain adaptation, domain generalization, single-domain augmentation, and contrastive learning. These approaches align distributions, ensemble models, synthesize variants, or shape representations using positive and negative pairs.
- Domain Adaptation: Domain adaptation addresses source-target domain drift using feature-, instance-, or parameter-based methods, including MMD and adversarial alignment.
- Domain Generalization: Domain generalization learns from source-domain data and aims to generalize to unseen domains without target-domain access.
- Domain Generalization: Domain alignment methods map multiple domain distributions toward a shared distribution, while CCSA uses contrastive semantic alignment across classes and domains.
- Single-Domain Generalization: Single-domain generalization synthesizes image- or feature-space samples to expand a single source distribution, including variants in image quality, appearance, and spatial configuration.
- Single-Domain Generalization: Adversarial augmentation can use overly simple transformations, while excessive adversarial examples may damage classifier performance.
- Contrastive Learning: Contrastive learning brings positive pairs closer and pushes negative pairs apart, with methods such as SimCLR and CPC applying augmentation at image or patch level.
3. Method
PDEN jointly learns a task model and generator to progressively expand a single source domain into safe, effective unseen domains. Contrastive learning aligns representations across domains while the alternating process improves subsequent domain generation.
- Overall framework: PDEN comprises a task net M and unseen-domain generator G whose weights are jointly and alternately learned.The task model is retrained after each generated domain, then guides the next generator.
- Task model M: The task model M contains a feature extractor, classifier head, and projection head for classification and contrastive learning.The projection output is normalized on a unit hypersphere so similarity can be measured by inner product.
- Unseen domain generator G: G transforms source images into semantically equivalent images from different domains and can use autoencoders, STNs, or task-specific combinations.The experiments mainly use an autoencoder with AdaIN; STN provides geometry-aware spatial transformation.
- Progressive expansion: K synthetic domains are progressively generated, and the task model trains on the source data combined with previously generated domains.Progressive expansion is intended to improve the completeness and coverage of generated domains.
- Domain alignment and classification: Contrastive InfoNCE learning pulls representations of same-semantic samples from different domains closer in Z space to learn domain-invariant features.The synthetic image and its source image share semantic information but come from different domains.
- Generation constraints: Safety requires correct task-model predictions and cycle conversion back to the source domain, while effectiveness uses adversarial and diversity objectives.The generator maximizes InfoNCE against the task model and is additionally encouraged to produce diverse samples.
4. Experiment
PDEN is evaluated for single-domain generalization across digit classification, CIFAR10-C robustness, and SYNTHIA segmentation, with additional analyses of feature distributions, hyperparameters, and few-shot adaptation. Across these settings, PDEN generally outperforms comparison methods, while its generated domains support more consistent cross-domain feature representations.
- Datasets and evaluation: PDEN is evaluated on Digits, CIFAR10-C, and SYNTHIA using mean accuracy or mIoU on unseen domains.Digits uses MNIST as the source; CIFAR10-C evaluates corruptions across severity levels; SYNTHIA evaluates traffic-scene segmentation across locations and conditions.
- Single-domain generalization: PDEN performs better than other single-domain generalization and manual augmentation methods on the Digits experiments.It performs much better on SVHN, MNIST-M, and USPS, and is reported to outperform the compared augmentation methods without manual augmentation.
- Single-domain generalization: PDEN achieves higher average accuracy on CIFAR10-C, with larger advantages over MADA at more severe corruption levels.It is competitive with manual augmentation at lower severity and outperforms those methods at higher severity, although RandAugment is better for some corruption types.
- Single-domain generalization: PDEN improves average SYNTHIA segmentation mIoU over other approaches, with greater improvement when Highway-Dawn or Highway-Fog is the source domain.The evaluation uses unseen target domains in New York ish and Old European Town.
- Additional analysis: On Digits, target-domain accuracy rises rapidly with small K and gradually converges as K increases; the experiments set K = 20.The authors contrast this with MADA, whose performance peaks at K = 3 and decreases as K grows, and report accuracy increases with wadv and wdiv.
- Additional analysis: PDEN produces target-domain feature distributions more consistent with the source domain than the baseline, whose target samples are mostly mixed.The comparison visualizes MNIST and MNIST-M feature spaces for the baseline and PDEN.
5. Conclusion
PDEN learns domain-invariant features for single-domain generalization by generating semantically consistent unseen domains, mining hard cases, and adding them to training.
- PDEN generates synthetic unseen domains that share the source domain’s semantic information.
- The framework aligns source and unseen-domain distributions to learn domain-invariant representations.
- PDEN mines hard unseen domains where the task model cannot extract domain-invariant representations.
- Adding these generated domains to training is intended to make the model more robust.