Pith. sign in

REVIEW 4 major objections 5 minor 69 references

Preserving Clusters in Prompt Learning for Unsupervised Domain Adaptation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that combining source-prompt predictions with a Wasserstein clustering objective makes CLIP prompt learning for unsupervised domain adaptation learn target prompts that genuinely represent target visual clusters, and…

desk verdict A practically solid prompt-learning UDA paper that earns its gains on standard benchmarks, though the inconsistent inference protocol and the balanced-cluster assumption in the OT term deserve scrutiny. read the letter →

arxiv 2506.11493 v1 pith:HHYCJPY6 submitted 2025-06-13 cs.CV

Tung-Long Vuong , Hoang Phan , Vy Vo , Anh Bui , Thanh-Toan Do , Trung Le , Dinh Phung This is my paper · ORCID
classification cs.CV
keywords unsuperviseddomainadaptationpromptlearningCLIPpseudo-labelenhancementWassersteindistanceoptimaltransportclusteringvision-languagemodelsmulti-source
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that prompt-learning methods for unsupervised domain adaptation with CLIP fail in two connected places: the zero-shot pseudo-labels they train on are unreliable under covariate shift, and aligning source and target prompts in text space does little to make the target prompt represent the target data. The proposed method, CRPL, addresses both by reweighting pseudo-labels: it combines CLIP's base-prompt predictions with predictions from learned source prompts, weighting each source by the L2 distance between the target image's visual embedding and that source's class centroids. It then adds a term that minimizes the Wasserstein distance between the target text prompts and the target visual embeddings, which, by the paper's Lemma 1, is equivalent to treating each text prompt as the centroid of a balanced cluster of visual embeddings. The reported experiments show large gains over prior prompt-based UDA methods on ImageCLEF, Office-Home, and DomainNet, and an ablation study indicates the learned target prompt alone outperforms the combined source-and-target prompts of prior approaches.

What carries the argument

The machinery has two parts. The first is distance-aware source-enhanced pseudo-labeling: for a target image with visual embedding $z$, the combined class text embedding is $\tau^k_{\mathrm{ave}}(x) = \frac12 \tau^k_{\mathrm{base}} + \frac12 \sum_i w_{k,i}(x)\,\tau^k_{S_i}$, where $w_{k,i}(x)$ is a softmax over $\|z_{\mathrm{pre}} - c^i_k\|_2$ distances to source class centroids; this produces soft pseudo-labels $\hat{y}[k]$ by cosine similarity. The second is the Wasserstein clustering regularizer $L_W = W_{d_z}(P_{\tau,\pi}, P_T)$, with $d_z(a,b)=1-\langle a,b\rangle$, $\pi_k=1/K$, and $P_T$ the empirical target visual distribution. The paper's Lemma 1 proves that minimizing this Wasserstein term is equivalent to nearest-assignment quantization of target visual embeddings to text prompts with cluster sizes proportional to $\pi$; with balanced $\pi$, each target prompt is pulled to the balanced centroid of its class cluster while the pseudo-label loss keeps it attached to the correct class.

What would settle it

Run CRPL on a target domain with a heavily imbalanced class distribution, or with one class split into two visually distinct subclusters in CLIP's embedding space, and compare per-class accuracy against the variant without the Wasserstein term; if the small or multimodal classes lose accuracy, the balanced single-centroid assumption is the cause. The paper's Table 5 already exposes the mechanism: on the Art and Clipart domains, adding the Wasserstein term to weak pseudo-labels collapses accuracy from 47.6% to 7.9% and from 29.0% to 4.7%.

Watch

Extended reading notes

Core claim

The central claim is that the geometry of CLIP's embedding space can be exploited twice for unsupervised domain adaptation. First, target pseudo-labels are made more accurate by averaging the zero-shot base text embedding with source-domain prompt text embeddings, weighted by a softmax over L2 distances from the target visual embedding to each source class centroid (Eqs. 6-8). Second, because each target class's visual embeddings form a single cluster, the target text prompts are refined by minimizing the Wasserstein distance $L_W = W_{d_z}(P_{\tau,\pi}, P_T)$ between a uniform distribution over target text embeddings and the empirical target visual embedding distribution. Lemma 1 shows this minimization is the same as assigning each target visual embedding to its nearest text prompt with balanced cardinalities, so the text prompts become true cluster centroids. Trained jointly with source cross-entropy, target pseudo-label cross-entropy, and $L_W$ ($\lambda_T = \lambda_W = 0.5$), the method reports state-of-the-art accuracy and, in the authors' analysis, target prompts that are substantially better representations of the target domain than those learned by source-target prompt alignment.

Load-bearing premise

The method assumes each class's target visual embeddings form a single compact cluster of roughly equal size in CLIP's space, so that forcing text prompts to be balanced cluster centroids pulls them toward true class distributions rather than away from them.

Editorial extensions

If this is right

  • Source-enhanced pseudo-labels substantially improve target-prompt training relative to plain CLIP zero-shot labels; on Office-Home, the target prompt alone rises from 47.6%/29.0%/53.5%/63.5% (CPL) to 75.8%/62.9%/86.6%/87.2% (SPL) on Art/Clipart/Product/Real World.
  • The Wasserstein clustering term helps only when initial pseudo-labels are already reasonable; with weak labels it can collapse accuracy to below 8% on Art and Clipart, so the pseudo-label enhancement and clustering reinforcement are mutually dependent.
  • CRPL's target prompts are strong enough for target-only inference, outperforming the combined source-plus-target inference of prior prompt methods, which indicates the learned templates actually capture target-specific visual structure.
  • The joint objective yields consistent state-of-the-art averages in both source-combined and multi-source settings on ImageCLEF, Office-Home, and DomainNet (e.g. 78.9% vs 75.5% multi-source Office-Home average).
  • The gains persist under image corruptions where CLIP's zero-shot accuracy drops, suggesting the clustering regularizer stabilizes target prompt learning under distribution shift.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Not stated in the paper, but if target-prompt-only inference is the right evaluation, then the reported gap between CRPL and prior methods may be understated: previous prompt-alignment methods likely lean on source prompts at test time, so re-evaluating them with target prompts only would quantify how much of their accuracy is genuine target representation.
  • Not stated in the paper, but the fixed equal weighting of the base prompt and source prompts (the 1/2 blend in Eq. 7) is a compromise the authors explicitly acknowledge; learning that weight from transferability statistics or validation-set accuracy would be a natural next step.
  • Not stated in the paper, but the balanced-cluster penalty ($\pi_k = 1/K$) is the first component to relax for long-tailed target domains; replacing it with a soft, estimated $\pi$ or a per-class temperature would preserve the centroid interpretation while removing the equal-size constraint.
  • Not stated in the paper, but because the method only tunes prompts, it cannot repair poor clustering at the level of CLIP's frozen visual features; end-to-end fine-tuning of the visual encoder would test whether the cluster-preservation benefit survives when the features themselves move.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes Clustering Reinforcement Prompt Learning (CRPL) for unsupervised domain adaptation with CLIP-style vision-language models. The method has two main components: (i) a pseudo-label enhancement scheme that combines zero-shot CLIP predictions with predictions from source-domain prompts, using L2-distance-based weights over source class centroids (Eqs. 6-8); and (ii) a Wasserstein-distance regularizer between target text embeddings and target visual embeddings (Eq. 10) intended to enforce a clustering structure on the learned target prompts, with uniform class proportions pi_k=1/K. The paper reports results on ImageCLEF, Office-Home, DomainNet, pairwise Office-Home, corrupted Office-Home, and ViT-based settings, claiming consistent improvements over DAPL, MPA, and PGA. Ablations indicate that source-enhanced pseudo-labels improve target-prompt training and that the Wasserstein term helps when pseudo-labels are reasonably accurate, but it can cause severe collapse when pseudo-labels are weak, as shown in Table 5.

Significance. If the claims hold, the paper offers a simple and useful recipe for prompt-based UDA: using source prompts as a pseudo-label prior and adding a clustering regularizer to target prompt learning. The paper also evaluates the quality of the target prompt itself, rather than only the combined source+target prediction, which is an informative diagnostic. Strengths include a clear motivation, relatively few hyperparameters (lambda_T=lambda_W=0.5 set uniformly), an ablation study that examines failure modes, and publicly released code. However, the empirical comparison is complicated by a mismatch in inference protocols, the absence of error bars, and a theoretical lemma that does not precisely match the implemented balanced-clustering objective. These issues are substantial enough that the state-of-the-art claim and the clustering-preservation interpretation need additional support or revision.

major comments (4)
  1. [5.3, Lemma 1] The theoretical statement in Lemma 1 does not match the objective actually used in the method. The paper fixes pi_k=1/K after stating Eq. (10), but Lemma 1 describes the minimizer of a problem with a free simplex weight pi (Eq. (11) has min over T and pi). These are different optimization problems. Moreover, the second sentence of Lemma 1, which asserts that for given T the optimal assignment sigma is the 'nearest assignment' defined by argmin_m d(z, tau_m), is not valid when pi is fixed to a uniform vector and the cardinality constraints bind: an optimal balanced transport plan must assign exactly N/K points to each centroid, which is generally not the unconstrained nearest-neighbor assignment. The proof does not address this discrepancy. Since the clustering-preservation claim is the paper's second main contribution, the lemma needs to be corrected to describe the balanced k-means problem actually being solved, or the method's theoretical claim must be weakened accordingly.
  2. [6.1, Inference, and Tables 2-4] The headline comparisons are confounded by different inference protocols. Section 6.1 states that CRPL relies solely on the target prompt during inference, while MPA, DAPL, and PGA use both source and target prompts. This means Tables 2-4 do not isolate the contribution of the proposed training objective from the choice of inference. To support the state-of-the-art claim, the authors should report baseline results under the same target-prompt-only inference protocol, and/or report CRPL under the source+target protocol used by the baselines.
  3. [Tables 2-4] No standard deviations, multiple seeds, or statistical significance tests are reported. Many of the claimed improvements are relatively small (e.g., 1.9, 3.1, and 3.4 points average in Table 2, and smaller per-domain differences in Table 4), so without variance estimates it is not possible to assess whether these differences are meaningful. At least three runs with mean and standard deviation, or an equivalent statistical analysis, are needed for the central performance claim.
  4. [5.3, Eq. (10), and Table 5] The balanced-cluster assumption pi_k=1/K is load-bearing, but the evidence provided for the underlying clustering assumption is incomplete. The supervised target-prompt result in Table 1 shows that one text prototype per class can be predictive, but it does not establish that each class forms a single compact cluster of equal mass in the CLIP embedding space. Table 5 itself shows that the same Wasserstein regularizer can collapse to 7.9% and 4.7% accuracy on the Ar and Cl domains when initial pseudo-label accuracy is 47.6% and 29.0%. The paper acknowledges this dependence in the ablation discussion, but the abstract and conclusion present the clustering term as a generally beneficial principle. The authors should either add a mechanism for detecting or recovering from this collapse, or clearly state the conditionality of the clustering component in the abstract and conclusion.
minor comments (5)
  1. [Eq. (8) and Appendix Eq. (14)] The formula for w_{i,k}(z) uses zpre in the numerator but z in the denominator exponent. This is presumably a typo and should be zpre consistently; as written, the definition is not reproducible.
  2. [Section 5.3, paragraph after Lemma 1] The statement that 'the text embeddings tau_k^T act as prototypes or centroids of the clusters' is stronger than what Lemma 1 establishes. Lemma 1 (even in its intended form) shows an equivalence to an assignment problem, not that the optimal text embedding is a geometric centroid of the visual embeddings. Please temper this wording.
  3. [Appendix, Section 10] The limitation that the equal-weight combination of base and source prompts in Eq. (13) is suboptimal because the base-prompt training data is unavailable is stated only in the appendix. This is a relevant caveat for the pseudo-label enhancement component and should be mentioned in the main paper.
  4. [Appendix, Proof of Lemma 1] The proof contains typographical issues, including 'the the optimal solution' and inconsistent indexing (N vs NT, i vs n). Please proofread the proof and align notation with the main text.
  5. [Table 9] Some entries in Table 9 appear corrupted or misformatted, such as '635 411' and 'Specklenpose'. Please verify the table values and the corruption name.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is a training objective evaluated on external benchmarks, and its OT regularizer is an assumed constraint rather than a fitted prediction.

full rationale

The paper derives an optimization objective (Eq. 12) that combines source cross-entropy, target pseudo-label cross-entropy, and a Wasserstein clustering regularizer, then evaluates against standard UDA benchmarks with held-out target ground-truth labels. The pseudo-labels in Eq. (6) are computed from source prompts and CLIP base prompts, not from target labels, so the reported accuracy is not a fitted quantity renamed as a prediction. Lemma 1 is proved in the appendix using the Monge formulation of optimal transport; its claim that the Wasserstein term makes text embeddings act as cluster centroids is a mathematical equivalence, not an empirical result imported from a self-citation. The self-citations to TIDOT [31] and VQ-WAE [51] are background references for the optimal-transport clustering viewpoint and are not load-bearing for the central claim. The balanced-cluster assumption (pi_k = 1/K) is an explicit modeling choice, and the paper's own ablation (Table 5) and limitation section honestly document that the OT term can collapse when initial pseudo-labels are weak; this is a robustness concern, not circularity. The evaluation against external labels and independent baselines makes the derivation self-contained with no step reducing to its inputs by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method's success depends on the clustering assumption for CLIP embeddings, a uniform class-prior assumption used in the OT term, and the transferability of source prompts to the target domain. The loss weights lambda_T and lambda_W are chosen by hand rather than fitted; pi_k=1/K is a modeling choice. No new entities are introduced.

free parameters (3)
  • lambda_T = 0.5
    Weight for target pseudo-label CE loss in Eq. (12), set uniformly across all experiments without validation tuning.
  • lambda_W = 0.5
    Weight for Wasserstein clustering loss in Eq. (12), set uniformly across all experiments.
  • pi_k = 1/K
    Category probabilities of the text-embedding distribution in Eq. (10), fixed to uniform to enforce balanced clusters; not fitted to data.
assumptions (4)
  • domain assumption The clustering assumption: each class's visual embeddings in the CLIP space form a single cluster, and a single text embedding can represent the class.
    Stated in Section 4 'Clustering preservation prediction' with evidence from Table 1, but it is an assumption about the embedding geometry that the Wasserstein term enforces.
  • ad hoc to paper Balanced cluster assumption: the target class distribution is uniform, so pi_k=1/K is appropriate.
    In Section 5.3, pi_k is set to 1/K 'aiming to learn balanced clusters'; this may not hold for imbalanced target data.
  • domain assumption Transferability of source prompts: source-domain text prompts provide useful reference predictions for target visual embeddings due to shared domain-invariant features.
    Section 4, motivating Eq. (6) and Eq. (7); supported by Table 1 but relies on the source and target sharing common embeddings.
  • standard math Optimal transport duality and assignment formulation for discrete measures.
    Used in Lemma 1 to equate Wasserstein distance with nearest-assignment clustering; standard result from Villani [50].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Preserving Clusters in Prompt Learning for Unsupervised Domain Adaptation." pith.science (2026). https://pith.science/paper/HHYCJPY6

@misc{pith2026250611493,
  author       = {Pith},
  title        = {Pith review of: Preserving Clusters in Prompt Learning for Unsupervised Domain Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HHYCJPY6}},
  note         = {Machine review of arXiv:2506.11493}
}
read the original abstract

Recent approaches leveraging multi-modal pre-trained models like CLIP for Unsupervised Domain Adaptation (UDA) have shown significant promise in bridging domain gaps and improving generalization by utilizing rich semantic knowledge and robust visual representations learned through extensive pre-training on diverse image-text datasets. While these methods achieve state-of-the-art performance across benchmarks, much of the improvement stems from base pseudo-labels (CLIP zero-shot predictions) and self-training mechanisms. Thus, the training mechanism exhibits a key limitation wherein the visual embedding distribution in target domains can deviate from the visual embedding distribution in the pre-trained model, leading to misguided signals from class descriptions. This work introduces a fresh solution to reinforce these pseudo-labels and facilitate target-prompt learning, by exploiting the geometry of visual and text embeddings - an aspect that is overlooked by existing methods. We first propose to directly leverage the reference predictions (from source prompts) based on the relationship between source and target visual embeddings. We later show that there is a strong clustering behavior observed between visual and text embeddings in pre-trained multi-modal models. Building on optimal transport theory, we transform this insight into a novel strategy to enforce the clustering property in text embeddings, further enhancing the alignment in the target domain. Our experiments and ablation studies validate the effectiveness of the proposed approach, demonstrating superior performance and improved quality of target prompts in terms of representation.

Figures

Figures reproduced from arXiv: 2506.11493 by the authors.

Figure 1
Figure 1. (a) Target-domain prompt refinement. Pre-trained models are trained on large-scale datasets, which are assumed to fully or partially cover the data distribution of downstream tasks. For example, “dog with style domain S1,...,3,T ” can be viewed as subdomains of “dog”. However, the visual embedding distribution in training may shift from that of the pretraining distribution, causing class descriptions without domain-… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

69 extracted references · 65 canonical work pages

  1. [1]

    Analysis of representations for domain adaptation

    Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006. 1

    work page 2006

  2. [2]

    A theory of learning from different domains.Machine learning, 79: 151–175, 2010

    Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains.Machine learning, 79: 151–175, 2010. 1, 2

    work page 2010

  3. [3]

    Learning disentangled semantic representation for domain adaptation.IJCAI, 2019:2060–2066, 2019

    Ruichu Cai, Zijian Li, Pengfei Wei, Jie Qiao, Kun Zhang, and Zhifeng Hao. Learning disentangled semantic representation for domain adaptation.IJCAI, 2019:2060–2066, 2019. 2

    work page 2019

  4. [4]

    Multi-prompt alignment for multi-source unsupervised domain adaptation

    Haoran Chen, Zuxuan Wu, and Yu-Gang Jiang. Multi-prompt alignment for multi-source unsupervised domain adaptation. Neural Information Processing Systems, 2022. 1, 2, 3, 6, 7, 12

    work page 2022

  5. [5]

    Transferability vs

    Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. InICML, pages 1081–1090, 2019. 8

    work page 2019

  6. [6]

    Towards discriminability and diversity: Batch nuclear-norm maximization under label insufficient situations

    Shuhao Cui, Shuhui Wang, Junbao Zhuo, Liang Li, Qingming Huang, and Qi Tian. Towards discriminability and diversity: Batch nuclear-norm maximization under label insufficient situations. InCVPR, pages 3940–3949, 2020. 2, 8

    work page 2020

  7. [7]

    Gradually vanishing bridge for adversar- ial domain adaptation

    Shuhao Cui, Shuhui Wang, Junbao Zhuo, Chi Su, Qingming Huang, and Qi Tian. Gradually vanishing bridge for adversar- ial domain adaptation. InCVPR, pages 12455–12464, 2020. 8

    work page 2020

  8. [8]

    Domain-agnostic mutual prompting for unsuper- vised domain adaptation

    Zhekai Du, Xinyao Li, Fengling Li, Ke Lu, Lei Zhu, and Jingjing Li. Domain-agnostic mutual prompting for unsuper- vised domain adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23375–23384, 2024. 12

    work page 2024

Show all 69 references
  1. [9]

    Partial feature selection and alignment for multi-source domain adaptation

    Yangye Fu, Ming Zhang, Xing Xu, Zuo Cao, Chao Ma, Yanli Ji, Kai Zuo, and Huimin Lu. Partial feature selection and alignment for multi-source domain adaptation. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16654–16663, 2021. 7, 12

    2021

  2. [10]

    Stylegan-nada: Clip- guided domain adaptation of image generators.ACM Trans- actions on Graphics (TOG), 41(4):1–13, 2022

    Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip- guided domain adaptation of image generators.ACM Trans- actions on Graphics (TOG), 41(4):1–13, 2022. 2

    2022

  3. [11]

    Unsupervised domain adaptation by backpropagation

    Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. InICML, pages 1180–1189,

  4. [12]

    Domain-adversarial training of neural networks.Journal of machine learning research, 17(59):1–35,

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, Franc ¸ois Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks.Journal of machine learning research, 17(59):1–35,

  5. [13]

    Domain adaptation via prompt learning.IEEE Transactions on Neural Networks and Learning Systems, 2023

    Chunjiang Ge, Rui Huang, Mixue Xie, Zihang Lai, Shiji Song, Shuang Li, and Gao Huang. Domain adaptation via prompt learning.IEEE Transactions on Neural Networks and Learning Systems, 2023. 1, 2, 3, 4, 6, 7, 8, 12

    2023

  6. [14]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. InNeurIPS, pages 2672–2680, 2014. 2

    2014

  7. [15]

    Rasch, Bernhard Sch¨olkopf, and Alexander J

    Arthur Gretton, Karsten Borgwardt, Malte J. Rasch, Bernhard Sch¨olkopf, and Alexander J. Smola. A kernel method for the two-sample-problem. InAdvances in neural information processing systems, pages 513–520, 2007. 2

    2007

  8. [16]

    Spherical space domain adaptation with robust pseudo-label loss

    Xiang Gu, Jian Sun, and Zongben Xu. Spherical space domain adaptation with robust pseudo-label loss. InCVPR, pages 9101–9110, 2020. 8

    2020

  9. [17]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InCVPR, pages 770–778, 2016. 8

    2016

  10. [18]

    Unsupervised domain adaptation with hierarchical gradient synchronization

    Lanqing Hu, Meina Kan, Shiguang Shan, and Xilin Chen. Unsupervised domain adaptation with hierarchical gradient synchronization. InCVPR, pages 4043–4052, 2020. 8

    2020

  11. [19]

    Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V . Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. InICML, pages 4904– 4916, 2021. 2

    2021

  12. [20]

    Wilds: A benchmark of in-the-wild distribu- tion shifts

    Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, et al. Wilds: A benchmark of in-the-wild distribu- tion shifts. InInternational conference on machine learning, ...

    2021

  13. [21]

    Padclip: Pseudo-labeling with adaptive debiasing in clip for unsupervised domain adaptation

    Zhengfeng Lai, Noranart Vesdapunt, Ning Zhou, Jun Wu, Cong Phuoc Huynh, Xuelu Li, Kah Kuen Fu, and Chen-Nee Chuah. Padclip: Pseudo-labeling with adaptive debiasing in clip for unsupervised domain adaptation. InProceedings of the IEEE/CVF International Conference on Computer Vi...

    2023

  14. [22]

    Empowering unsupervised domain adaptation with large- scale pre-trained vision-language models

    Zhengfeng Lai, Haoping Bai, Haotian Zhang, Xianzhi Du, Jiulong Shan, Yinfei Yang, Chen-Nee Chuah, and Meng Cao. Empowering unsupervised domain adaptation with large- scale pre-trained vision-language models. InProceedings of the IEEE/CVF Winter Conference on Applications of Co...

    2024

  15. [23]

    Sliced wasserstein discrepancy for unsu- pervised domain adaptation

    Chen-Yu Lee, Tanmay Batra, Mohammad Haris Baig, and Daniel Ulbricht. Sliced wasserstein discrepancy for unsu- pervised domain adaptation. InCVPR, pages 10285–10295,

  16. [24]

    Enhanced transport distance for unsuper- vised domain adaptation

    Mengxue Li, Yiming Zhai, You-Wei Luo, Pengfei Ge, and Chuan-Xian Ren. Enhanced transport distance for unsuper- vised domain adaptation. InCVPR, 2020. 2, 8

    2020

  17. [25]

    T-svdnet: Exploring high-order prototypical correlations for multi-source domain adaptation.ICCV, 2021

    Ruihuang Li, Xu Jia, Jianzhong He, Shuaijun Chen, and Qinghua Hu. T-svdnet: Exploring high-order prototypical correlations for multi-source domain adaptation.ICCV, 2021. 7, 12

    2021

  18. [26]

    How to avoid machine learning pit- falls: a guide for academic researchers.arXiv preprint arXiv:2108.02497, 2021

    Michael A Lones. How to avoid machine learning pit- falls: a guide for academic researchers.arXiv preprint arXiv:2108.02497, 2021. 1

    2021 arXiv

  19. [27]

    Learning transferable features with deep adaptation networks

    Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jor- dan. Learning transferable features with deep adaptation networks. InICML, pages 97–105, 2015. 1, 2

    2015

  20. [28]

    Deep transfer learning with joint adaptation networks

    Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. InInternational conference on machine learning, pages 2208–

  21. [29]

    Transferable representation learning with deep adaptation networks.IEEE transactions on pattern analysis and machine intelligence, 41(12):3071–3085, 2018

    Mingsheng Long, Yue Cao, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Transferable representation learning with deep adaptation networks.IEEE transactions on pattern analysis and machine intelligence, 41(12):3071–3085, 2018. 1

    2018

  22. [30]

    Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I. Jordan. Conditional adversarial domain adaptation. InNeurIPS, pages 1647–1657, 2018. 2, 8

    2018

  23. [31]

    Tidot: A teacher imitation learning approach for domain adaptation with optimal trans- port

    Tuan Nguyen, Trung Le, Nhan Dam, Quan Hung Tran, Truyen Nguyen, and Dinh Phung. Tidot: A teacher imitation learning approach for domain adaptation with optimal trans- port. InProceedings of the Thirtieth International Joint Con- ference on Artificial Intelligence, IJCAI-21, pa...

    2021

  24. [32]

    Moment matching for multi-source domain adaptation

    Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In2019 IEEE/CVF International Confer- ence on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 1406–1415. IEEE,

    2019

  25. [33]

    Fabio Petroni, Tim Rockt ¨aschel, Sebastian Riedel, Patrick S. H. Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander H. Miller. Language models as knowledge bases? InEMNLP- IJCNLP, pages 2463–2473, 2019. 2

    2019

  26. [34]

    Global-local regularization via distribu- tional robustness

    Hoang Phan, Trung Le, Trung Phung, Anh Tuan Bui, Nhat Ho, and Dinh Phung. Global-local regularization via distribu- tional robustness. InInternational Conference on Artificial Intelligence and Statistics, pages 7644–7664. PMLR, 2023. 1

    2023

  27. [35]

    Enhanc- ing domain adaptation through prompt gradient alignment

    Hoang Phan, Lam Tran, Quyen Tran, and Trung Le. Enhanc- ing domain adaptation through prompt gradient alignment. Neural Information Processing Systems, 2024. 1, 2, 3, 4, 6, 7, 8, 12

    2024

  28. [36]

    Control- lable prompt tuning for balancing group distributional robust- ness

    Hoang Phan, Andrew Gordon Wilson, and Qi Lei. Control- lable prompt tuning for balancing group distributional robust- ness. InForty-first International Conference on Machine Learning, 2024. 2

    2024

  29. [37]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InICML, pages 8748–8763, 2021. 1, 2, 3, 7, 8, 12

    2021

  30. [38]

    Multi-source unsupervised domain adaptation via pseudo target domain.IEEE Transactions on Image Process- ing, 2022

    Chuan-Xian Ren, Yong Liu, Xiwen Zhang, and Ke-Kun Huang. Multi-source unsupervised domain adaptation via pseudo target domain.IEEE Transactions on Image Process- ing, 2022. 7

    2022

  31. [39]

    Distributionally robust neural networks

    Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks. In International Conference on Learning Representations, 2019. 1

    2019

  32. [40]

    Ushiku, and T

    Kuniaki Saito, Kohei Watanabe, Y . Ushiku, and T. Harada. Maximum classifier discrepancy for unsupervised domain adaptation.IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 7

    2017

  33. [41]

    Wasser- stein distance guided representation learning for domain adap- tation.CoRR, abs/1707.01217, 2018

    Jian Shen, Yanru Qu, Weinan Zhang, and Yong Yu. Wasser- stein distance guided representation learning for domain adap- tation.CoRR, abs/1707.01217, 2018. 2

    2018 arXiv

  34. [42]

    A dirt-t approach to unsupervised domain adaptation

    Rui Shu, Hung Bui, Hirokazu Narui, and Stefano Ermon. A dirt-t approach to unsupervised domain adaptation. In International Conference on Learning Representations, 2018. 2

    2018

  35. [43]

    Deep coral: Correlation alignment for deep domain adaptation

    Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. InECCV, pages 443–

  36. [44]

    Unsupervised domain adaptation via structurally regularized deep clustering

    Hui Tang, Ke Chen, and Kui Jia. Unsupervised domain adaptation via structurally regularized deep clustering. In CVPR, pages 8725–8735, 2020. 1, 2, 8

    2020

  37. [45]

    Unsupervised domain adaptation via distilled discriminative clustering.Pattern Recognition, 127:108638, 2022

    Hui Tang, Yaowei Wang, and Kui Jia. Unsupervised domain adaptation via distilled discriminative clustering.Pattern Recognition, 127:108638, 2022. 8

    2022

  38. [46]

    Simultaneous deep transfer across domains and tasks.CoRR,

    Eric Tzeng, Judy Hoffman, Trevor Darrell, and Kate Saenko. Simultaneous deep transfer across domains and tasks.CoRR,

  39. [47]

    Adversarial discriminative domain adaptation

    Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 7167–7176, 2017. 1

    2017

  40. [48]

    Naveen Venkat, Jogendra Nath Kundu, D. K. Singh, Amba- reesh Revanur, and R. VenkateshBabu. Your classifier can secretly suffice multi-source domain adaptation.Neural In- formation Processing Systems, 2021. 7

    2021

  41. [49]

    Deep hashing network for unsupervised domain adaptation

    Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. InCVPR, pages 5385–5394,

  42. [50]

    Springer, Berlin, Heidelberg, 2008

    C´edric Villani.Optimal Transport: Old and New. Springer, Berlin, Heidelberg, 2008. 2

    2008

  43. [51]

    Vector quan- tized Wasserstein auto-encoder

    Long Tung Vuong, Trung Le, He Zhao, Chuanxia Zheng, Mehrtash Harandi, Jianfei Cai, and Dinh Phung. Vector quan- tized Wasserstein auto-encoder. InProceedings of the 40th International Conference on Machine Learning, pages 35223– 35242. PMLR, 2023. 5

    2023

  44. [52]

    Learning to combine: Knowledge aggregation for multi- source domain adaptation

    Hang Wang, Minghao Xu, Bingbing Ni, and Wenjun Zhang. Learning to combine: Knowledge aggregation for multi- source domain adaptation. InComputer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII, pages 727–744. Springer, 2020. 7

    2020

  45. [53]

    Qian Wang and Toby P. Breckon. Unsupervised domain adaptation via structured prediction based selective pseudo- labeling. InAAAI, pages 6243–6250, 2020. 8

    2020

  46. [54]

    Zuo, Junjie Yan, and Liang Lin

    Ruijia Xu, Ziliang Chen, W. Zuo, Junjie Yan, and Liang Lin. Deep cocktail network: Multi-source unsupervised domain adaptation with category shift.IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 7, 12

    2018

  47. [55]

    How transferable are features in deep neural networks?Ad- vances in neural information processing systems, 27, 2014

    Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks?Ad- vances in neural information processing systems, 27, 2014. 1

    2014

  48. [56]

    Autolabel: Clip-based framework for open-set video domain adaptation

    Giacomo Zara, Subhankar Roy, Paolo Rota, and Elisa Ricci. Autolabel: Clip-based framework for open-set video domain adaptation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11504– 11513, 2023. 2

    2023

  49. [57]

    Domain- symmetric networks for adversarial domain adaptation

    Yabin Zhang, Hui Tang, Kui Jia, and Mingkui Tan. Domain- symmetric networks for adversarial domain adaptation. In CVPR, pages 5031–5040, 2019. 8

    2019

  50. [58]

    Costeira, Jos´e M

    Han Zhao, Shanghang Zhang, Guanhang Wu, Jo ao P. Costeira, Jos´e M. F. Moura, and Geoffrey J. Gordon. Multi- ple source domain adaptation with adversarial learning, 2018. 2

    2018

  51. [59]

    Han Zhao, Shanghang Zhang, Guanhang Wu, Jos ´e M. F. Moura, Joao P Costeira, and Geoffrey J Gordon. Adver- sarial multiple source domain adaptation. InAdvances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 2

    2018

  52. [60]

    Multi-source distilling domain adapta- tion

    Sicheng Zhao, Guangzhi Wang, Shanghang Zhang, Yang Gu, Yaxian Li, Zhichao Song, Pengfei Xu, Runbo Hu, Hua Chai, and Kurt Keutzer. Multi-source distilling domain adapta- tion. InThe Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovativ...

    2020

  53. [61]

    Learning to prompt for vision-language models.Interna- tional Journal of Computer Vision, 130(9):2337–2348, 2022

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models.Interna- tional Journal of Computer Vision, 130(9):2337–2348, 2022. 12

    2022

  54. [62]

    Unsupervised domain adap- tion harnessing vision-language pre-training.IEEE Transac- tions on Circuits and Systems for Video Technology, 2024

    Wenlve Zhou and Zhiheng Zhou. Unsupervised domain adap- tion harnessing vision-language pre-training.IEEE Transac- tions on Circuits and Systems for Video Technology, 2024. 12

    2024

  55. [63]

    Aligning domain-specific distribution and classifier for cross-domain classification from multiple sources

    Yongchun Zhu, Fuzhen Zhuang, and Deqing Wang. Aligning domain-specific distribution and classifier for cross-domain classification from multiple sources. InThe Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artif...

    2019

  56. [65]

    Office-Home is a medium-scaled dataset containing approximately 15,500 images from 65 categories in four do- mains: Art, Clipart, Product, and Real World

    Experimental Settings DatasetsImageCLEF is a small-scaled dataset with 1,800 images across 12 object categories from three domains: Ima- geNet ILSVRC 2012 (I), Pascal VOC 2012 (P), and Caltech- 256 (C). Office-Home is a medium-scaled dataset containing approximately 15,500 ima...

    2012

  57. [66]

    Since these alternative methods typically fine-tune many more parameters, we ex- clude them from the experiments to ensure a fair comparison

    with text-end soft prompts, other methods fine-tune the transformer block [8], both image and text-end soft prompts [22], or the entire encoders [21, 62]. Since these alternative methods typically fine-tune many more parameters, we ex- clude them from the experiments to ensure...

  58. [67]

    Distance-Aware Pseudo-Label As discuss in previous section, different transferability be- tween domains motivate us a distance aware pseudo-labels scheme

    Additional Experiments 9.1. Distance-Aware Pseudo-Label As discuss in previous section, different transferability be- tween domains motivate us a distance aware pseudo-labels scheme. Specifically, we calculate the weighted average cosine distance from the visual embedding z=f ...

  59. [68]

    Performance on Corrupted OfficeHome dataset

    Limitations One of our main contributions is the enhancement of pseudo- labels for target domains by effectively leveraging infor- Corruption Method Ar Cl Pr Rw Ave No corruption Zero-Shot 71.2 50.4 81.4 82.6 71.4 PGA 74.8 56.0 85.2 86.0 75.5 ours76.8 63.5 87.5 87.6 78.9 Defoc...

  60. [69]

    Specifically, denote T={τ k T }K k=1 where τ k T represents the text embeddings of the context prompt [P k sh][P T ][CLASSk] for class k

    Proof of Lemma 1 We propose minimizing the Wasserstein distance between the target prompts’ text embeddings and the visual em- beddings from the target domain. Specifically, denote T={τ k T }K k=1 where τ k T represents the text embeddings of the context prompt [P k sh][P T ][...

  61. [450]

    Springer, 2016. 1, 2, 7

    2016

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.