Pseudo Multi-Source Domain Generalization: Bridging the Gap Between Single and Multi-Source Domain Generalization
Shohei Enomoto
TL;DR
This paper introduces PMDG, a framework that creates pseudo-domains from a single source to enable the use of multi-source domain generalization algorithms, improving model robustness across different data distributions.
Contribution
The paper proposes a novel method to simulate multiple domains from one source, allowing existing MDG algorithms to be applied in more practical SDG scenarios.
Findings
Pseudo-domains can match or outperform real multi-domain data.
PMDG performance positively correlates with MDG performance.
Extensive experiments validate the effectiveness of pseudo-domains across datasets.
Abstract
Deep learning models often struggle to maintain performance when deployed on data distributions different from their training data, particularly in real-world applications where environmental conditions frequently change. While Multi-source Domain Generalization (MDG) has shown promise in addressing this challenge by leveraging multiple source domains during training, its practical application is limited by the significant costs and difficulties associated with creating multi-domain datasets. To address this limitation, we propose Pseudo Multi-source Domain Generalization (PMDG), a novel framework that enables the application of sophisticated MDG algorithms in more practical Single-source Domain Generalization (SDG) settings. PMDG generates multiple pseudo-domains from a single source domain through style transfer and data augmentation techniques, creating a synthetic multi-domain…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
1. Bridging single-source and multi-source domain generalization addresses real needs. Collecting multi-domain datasets is expensive. 2. PseudoDomainBed provides standardized evaluation. Modified DomainBed for single-source setting with public code.
1. Using data augmentation as pseudo-domains is trivial. This is standard practice in domain generalization. Many existing SDG methods already do this. The paper just applies existing augmentations with existing MDG algorithms. 2. The core idea (treat augmentations as domains) is obvious. No new methods, no new algorithms, no new techniques. Just combining existing pieces. 3. The paper frames augmentation as "pseudo-domain generation" but this is just renaming existing techniques. Single-domai
1. The proposed method shows the effectiveness of the PMDG on domain generalization benchmark. 2. The readability of this paper is good, and the structure is clear.
1. Single source domain generation methods often utilize data augmentation or generation strategy to expand the source data distribution. The difference between the proposed method and the existing data augmentation-based or domain expansion-based method is not clear. 2. Missing results compared with SOTA single source domain generalization methods on the benchmarks. 3.Some abbreviations are not clearly defined, such as the ST,ED,CG, and IM in the section 6.2. The expression "Org+IM+IM" in the
1. The paper addresses a genuine practical limitation of MDG methods by proposing a straightforward and implementable solution that bridges SDG and MDG research directions. 2. The extensive experiments across multiple datasets, architectures, and MDG algorithms provide thorough validation of the PMDG framework and reveal useful insights about algorithm-domain shift compatibility. 3. The introduction of PseudoDomainBed with publicly available code facilitates reproducibility and future research
1. The core idea of treating augmented data as distinct domains is relatively incremental, essentially applying existing data augmentation techniques and training with existing MDG algorithms. The paper lacks significant methodological innovation beyond the combination of known techniques. 2. The critical assumption that all transformed data represents distinct domains with distributions different from the source is not empirically or theoretically validated, particularly for weakly-transformed
1. The work reuses DomainBed protocols, specifies transformation levels (dataset vs. minibatch) and interfaces, and turns off validation-time augmentation to avoid distortion. 2. Heatmaps across many MDG algorithms × transforms, comparisons of pseudo-domain combos, ViT vs. ResNet, and a data-efficiency curve (PMDG vs. MDG) are useful empirical references. 3. Authors explicitly note the shaky assumption that weak transforms are distinct domains and suggest measuring distribution distances
1. Core idea is largely a repackaging of well-trodden “domain expansion via augmentation.” Treating strong augmentations as different domains and training with MDG losses is conceptually aligned with years of SDG work on synthetic domain generation and adversarial/heuristic augmentations (e.g., MixStyle/RandConv/IPMix/OT/adversarial DA). PMDG’s “bridge” is effectively: apply known style/augmentation transforms, group by transform, and run existing MDG algorithms. This is incremental glue code ra
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsDomain Adaptation and Few-Shot Learning · Topic Modeling · Speech Recognition and Synthesis