On the Domain Robustness of Contrastive Vision-Language Models

David Brodmann; Erik Rodner; Mario Koddenbrock; Rudolf Hoffmann

REVIEW 3 major objections 5 minor 44 references

On the Domain Robustness of Contrastive Vision-Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that an LLM-guided, label-free corruption pipeline can benchmark domain-specific robustness of vision-language models, and that under that pipeline CLIP is the most robust model overall.

desk verdict Useful framework for domain-specific VLM robustness evaluation, but the headline 'CLIP is most robust' is an artifact of normalizing mCE to CLIP itself. read the letter →

arxiv 2506.23663 v1 pith:C4URSUFA submitted 2025-06-30 cs.CV cs.LG

Mario Koddenbrock , Rudolf Hoffmann , David Brodmann , Erik Rodner This is my paper · ORCID

classification cs.CVcs.LG

keywords vision-languagemodelsrobustnessevaluationdomainshiftimagecorruptionsLLM-guidedaugmentationCLIPzero-shotclassificationlabel-freebenchmarking

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 introduces DeepBench, a label-free framework for measuring how robust contrastive vision-language models are in specific deployment domains. The central proposal is to have a large language model pick realistic image corruptions from a predefined menu based on a written domain description, so that robustness can be tested without ground-truth labels. The authors apply it to six real-world domains and report that CLIP has the lowest mean corruption error overall, that SigLIP shows relative gains in some specialized domains, and that ALIGN, despite good clean accuracy, collapses under mild noise. If the approach works, practitioners can screen foundation models for specialized settings before deployment, using only data and a domain description.

What carries the argument

The mechanism is an LLM-based corruption selection pipeline: a structured prompt (role, domain description, list of available corruptions, output format) is sent to GPT-4o, which outputs a corruption configuration; those corruptions are applied to images, and model robustness is scored with balanced accuracy plus the label flip probability (how often a perturbation changes the model's prediction) and the mCE, rCE, and mFR metrics from corruption-error benchmarking. The predefined corruption menu (Tables 6 and 7) and a vicinal-risk framing in Eq. 5 are what connect a written domain description to an actual test distribution.

What would settle it

Take the six domains, gather naturally corrupted images (for example, real rain-blurred driving footage, X-rays with actual sensor noise, and satellite images with real cloud cover), run CLIP, SigLIP, and ALIGN zero-shot, and compare which model has the lowest mean corruption error; if a model other than CLIP wins, the paper's headline ranking is an artifact of the LLM-generated corruption set.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that LLM-guided, label-free corruption generation produces stable, context-aware robustness tests, and that under these tests CLIP (ViT-L/14) is the most robust of the three contrastive VLMs across all six domains, with SigLIP sometimes better on relative corruption error in medical and manufacturing settings and ALIGN the most fragile under noise (relative corruption error up to 14.78 in People Recognition). The authors also find that among CLIP variants, transformer backbones with larger capacity and QuickGELU activations, specifically ViT-L/14 with QuickGELU, are the most robust, while patch size and resolution effects vary by domain. These conclusions are conditional on the fixed corruption menu and the LLM's parameter choices, which the paper validates against a human-authored whitelist and blacklist of plausible versus implausible corruptions.

Load-bearing premise

The corruption menu plus the LLM's chosen parameters and intensities must faithfully represent the real image degradations that occur in each deployment domain; if it does not, the robustness rankings only describe performance under the chosen synthetic corruptions.

Editorial extensions

If this is right

Practitioners can screen a vision-language model for a new deployment domain by writing a one-paragraph domain description and running DeepBench, with no labeled data required.
For classification-style VLM tasks, CLIP ViT-L/14 with QuickGELU is the configuration the paper identifies as the default robustness choice across unknown corruptions.
A model's clean accuracy is not a reliable predictor of its robustness: ALIGN can lead on clean accuracy while degrading severely under mild noise.
Because the framework is modular in models and corruptions, the same pipeline can be reused as new VLMs and new domains appear.
The strong negative correlation between label flip probability and accuracy suggests that flip rate is a useful label-free robustness signal.

Reading between the lines

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

If the fixed corruption menu or the LLM's intensity ranges change, the reported model ranking could change, so the headline result should be read as 'robust under DeepBench's corruption families,' not as 'robust in the world.'
Because the pipeline needs only embeddings and predictions, the same label-free protocol could extend to image retrieval, captioning, or autoregressive VLMs, not just classification.
The consistently strong negative correlation between flip probability and accuracy in the appendix could let practitioners set a flip-rate threshold as a cheap early-warning signal, though the paper does not validate such a threshold.
The medical and manufacturing domains, where SigLIP beats CLIP on relative corruption error, suggest that no single VLM dominates everywhere and that domain-aware evaluation may change procurement choices.

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.

Desk Editor's Note

read the letter

You should know two things before reading this paper. First, the core idea — using an LLM to pick domain-appropriate corruptions from a fixed menu, then evaluating VLMs without labels via flip rate — is genuinely useful, and the authors have shipped the code. Second, the headline result that "CLIP consistently achieves the lowest mCE" is in large part an artifact of the metric's normalization. mCE in Eq. 11 divides each model's corruption error by a baseline, and Table 4 states all values are relative to CLIP ViT-L/14. So CLIP's mCE is 1.00 by construction, and the Section 4.2 conclusion is close to reporting that SigLIP and ALIGN have mCE > 1 — which is guaranteed when the champion is the denominator. Table 5 confirms the problem: other variants (ResNet101 on Manufacturing, ViT-B/32 on Driving) get mCE below 1, so the ranking across all models is not "CLIP always best." The claim needs to be recomputed against a fixed reference or reported as raw corruption error.

What the paper does well: it assembles six real-world domains, compares three representative contrastive VLMs, and the label-free flip-rate metric correlates strongly with accuracy (often r < -0.9). The LLM-generated corruption selections are stable over 10 runs and match common sense (rain for driving, cloud for satellite). The public code and data make this reproducible. That is real value for practitioners screening VLMs for deployment.

The soft spots are in proportion. The resolution contradiction is real: the Section 4.3 answer says higher resolution "leads to lower robustness on every domain," but Table 1 and the results bullet say ViT-L/14-336 improves mCE in all six domains. That internal inconsistency must be fixed. There are no error bars or multiple seeds; every table is single-run. The LLM validation against the authors' own whitelist/blacklist is a weak sanity check, not external validation. Those are fixable. The normalization issue is load-bearing because it undercuts the paper's main comparative claim.

Who is this for? Anyone building or choosing VLMs for specialized image domains will find the tool and the cautionary data about ALIGN's fragility useful. The paper deserves a serious referee, but the referee should ask for the metric normalization to be rejustified or recomputed, and the Section 4.3 inconsistency resolved.

My honest verdict: engage with this work, but treat the headline ranking as unproven until the mCE baseline is fixed. If you cite it, cite it for the framework, not for "CLIP is most robust."

Referee Report

3 major / 5 minor

Summary. The paper introduces DeepBench, an open-source framework for evaluating the domain-specific robustness of contrastive vision-language models (CLIP, SigLIP, ALIGN) without requiring labeled data. The framework uses a large language model (GPT-4o) to select and parameterize corruptions from a predefined menu for six application domains (medical, driving, manufacturing, people, satellite, handheld). Robustness is measured via balanced accuracy, label flip probability, and aggregate metrics mCE, rCE, and mFR defined relative to a baseline model. The central empirical claims are: (i) LLM-generated corruption sets are stable and context-aware, (ii) CLIP is the most robust of the three foundation models, and (iii) architectural variants of CLIP show that larger models, higher resolution, and QuickGELU activation improve robustness.

Significance. If the claims hold, DeepBench would be a practical, label-free tool for selecting deployment-specific VLMs, and the reported finding that robustness rank can differ by domain would be a useful caution for practitioners. The paper's strengths are its open-source release, use of public datasets and standard metrics, and the introduction of a label-flip metric that correlates strongly (negative Pearson r near -1) with balanced accuracy across all reported domains. These are concrete, checkable contributions. However, the central robustness ranking rests on a self-referential normalization, and the architecture analysis contains an internal contradiction, so the current version does not fully support the headline claims.

major comments (3)

[Section 4.2, Eq. (11), Table 4] The mCE, rCE, and mFR metrics are all defined relative to the baseline model b, and Table 4 states that b = CLIP (ViT-L/14). Consequently CLIP's mCE is exactly 1.00 in every domain by construction, so the claim that 'CLIP consistently achieves the lowest mCE across all domains' is not an independent confirmation of robustness; it only says that no other tested foundation model had lower corruption error than the chosen reference. The result is informative as a relative comparison, but the wording overstates it. Moreover, Table 5 contains CLIP variants with mCE below 1 (e.g., ViT-L/14-336 on all six domains, ViT-L/14 QGelu on four domains, ViT-B/32 on Driving), so the statement 'CLIP consistently achieves the lowest mCE across all domains' is false if interpreted across the paper's own architecture experiments. Please report absolute corruption errors or an external baseline, and explicitly scope the ranking to the three foundation models.
[Section 4.3, answer paragraph and Table 1] The answer to RQ3 states that 'higher input resolution leads to lower robustness on every domain,' but the results in Table 1 show the opposite: ViT-L/14-336 has lower mCE than the ViT-L/14 baseline in all six domains (e.g., 0.85 vs 1.00 on Driving, 0.91 vs 1.00 on Manufacturing), and the bullet point labeled 'Resolution' also says higher resolution improves robustness. This is a direct internal contradiction in the central summary of the architecture study and must be corrected.
[Section 4.1 and Table 2] The validation of LLM context-awareness uses a whitelist/blacklist that the authors themselves constructed from 'intuitive domain knowledge.' Since the LLM was prompted only with general domain descriptions and a corruption list, the fact that its selections 'fully satisfy' this author-defined list is evidence of agreement with the authors' priors, but it is not an independent ground truth for whether the corruptions are realistic for each deployment domain. The claim that LLM-generated strategies are 'context-aware' therefore needs either an independent validation (e.g., human expert labels, or a comparison with real-world degradation measurements) or a more careful framing that separates prompt alignment from real-world representativeness.

minor comments (5)

[Section 3.1] There is a typo: 'performes' should be 'performs'.
[Section 4.3, cross-reference] The text says 'A complete summary is provided in Table 4 (Appendix E),' but Table 4 is in Appendix D and Table 5 is the table in Appendix E. Please correct the cross-reference.
[Section 4.3, bullet on activation] The phrase 'better clean accuracy and mCE then the baseline' should be 'than the baseline.'
[Section 3.1, Eq. (10)] The text says 'mean flip rate relative to an baseline model'; it should be 'a baseline model.'
[General experimental methodology] All reported metrics appear to be from a single evaluation run, with no error bars, confidence intervals, or significance tests. Given that some headline comparisons are close (e.g., mCE values of 1.00 vs 1.05 on Driving), the absence of any repeated-measure uncertainty is a limitation that should be acknowledged.

Circularity Check

1 steps flagged · score 4.0 of 10

The domain-robustness benchmark is largely independent, but the RQ1 validation of 'context-aware' corruption selection is a self-referential check against the authors' own whitelist/blacklist.

self definitional [Section 4.1 and Appendix B, Table 2]
"As a sanity check, we defined a simple whitelist of obviously relevant corruptions and a blacklist of clearly implausible ones for every domain. ... These rules are fully satisfied by the LLM’s output, indicating strong semantic alignment."

The evidence for the claim that LLM-generated corruptions are 'context-aware' is that the LLM's selections satisfy a whitelist/blacklist that the authors curated from their own 'intuitive domain knowledge' (Table 2 caption). The same knowledge shaped the corruption descriptions in the prompt, so the LLM is being scored against the authors' priors. The validation therefore reduces to 'the LLM reproduces the authors' expectations,' not to an external benchmark of real-world context-awareness. The central robustness ranking is unaffected, but contribution 2 is supported by a self-referential check.

full rationale

The domain-robustness comparison itself is not circular: it uses public datasets, fixed pretrained models, and standard zero-shot evaluation, with no fitting of the models to the corruption sets. The mCE normalization to CLIP (Eq. 11, Table 4) does not force CLIP to rank first, because mCE ratios preserve model ordering regardless of the chosen baseline and because Table 5 shows other variants with mCE < 1 (e.g., ResNet101=0.83 in Manufacturing, ViT-B/32=0.84 in Driving). The only self-referential element is the Section 4.1 / Appendix B validation of 'context-aware' corruption selection, which checks the LLM against the authors' own curated whitelist/blacklist rather than against independent ground truth. This is a minor, non-central circularity; the headline robustness findings have independent content.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No parameters are fitted to data; robustness metrics are computed directly from model outputs. The benchmark settings (LLM temperature 0, 10 runs, majority vote) are fixed choices that do not depend on evaluated model performance, but the corruption intensity levels, which strongly affect the results, are not reported. The central assumptions are that the predefined corruption menu and the LLM's selections are faithful proxies for real-world domain shifts, and that label flip probability is a meaningful unsupervised metric.

free parameters (3)

LLM sampling temperature = 0
Set to 0 to stabilize corruption selection; this is a fixed choice, not fitted.
Number of LLM prompting runs = 10
Used to define the corruption sets via majority vote; no sensitivity analysis is reported.
Corruption intensity/severity levels = not reported
Each selected corruption is applied at several intensity steps, but the actual parameter ranges (e.g., blur kernel size, noise variance) are not given in the paper, so the measurement scale is not independently reproducible.

assumptions (5)

standard math Equations 5 and 6 assume Monte Carlo sampling of transformation parameters approximates the integral over the domain distribution.
The paper treats the approximation in Eq. 5-6 as valid for evaluating robustness via augmented test samples.
domain assumption Equation 1 assumes cosine similarity in the shared embedding space is the correct zero-shot prediction rule for the evaluated VLMs.
All zero-shot classification results depend on this standard but unverified assumption for CLIP, SigLIP, and ALIGN.
ad hoc to paper The predefined corruption list in Tables 6 and 7 is assumed to span the realistic degradations of each domain.
No evidence outside the authors' menu is provided that these corruption primitives cover the important real-world shifts in the six domains.
ad hoc to paper The whitelist/blacklist in Table 2 is assumed to be the correct ground truth for whether a corruption is plausible in a domain.
The LLM is declared context-aware because its output fully satisfies these author-defined rules, making the validation self-referential.
domain assumption The label flip probability (Equations 7-9) assumes the model's clean prediction is a reliable pseudo-label, so flip rate is a valid robustness proxy when labels are absent.
For models with low clean accuracy, flips away from an incorrect clean prediction are treated as errors, which could distort the metric.

how reviews work

0 comments

Cite this review

Pith. "Pith review of On the Domain Robustness of Contrastive Vision-Language Models." pith.science (2026). https://pith.science/paper/C4URSUFA

@misc{pith2026250623663,
  author       = {Pith},
  title        = {Pith review of: On the Domain Robustness of Contrastive Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C4URSUFA}},
  note         = {Machine review of arXiv:2506.23663}
}

read the original abstract

In real-world vision-language applications, practitioners increasingly rely on large, pretrained foundation models rather than custom-built solutions, despite limited transparency regarding their training data and processes. While these models achieve impressive performance on general benchmarks, their effectiveness can decline notably under specialized domain shifts, such as unique imaging conditions or environmental variations. In this work, we introduce Deepbench, a framework designed to assess domain-specific robustness of vision-language models (VLMs). Deepbench leverages a large language model (LLM) to generate realistic, context-aware image corruptions tailored to specific deployment domains without requiring labeled data. We evaluate a range of contrastive vision-language architectures and architectural variants across six real-world domains and observe substantial variability in robustness, highlighting the need for targeted, domain-aware evaluation. Deepbench is released as open-source software to support further research into domain-aware robustness assessment.

Figures

Figures reproduced from arXiv: 2506.23663 by the authors.

**Figure 2.** Heatmap of selection frequencies showing how often each corruption was chosen in 10 independent GPT-4o runs. As a sanity check, we defined a simple whitelist of obviously relevant corruptions and a blacklist of clearly implausible ones for every domain. Green cells belong to the whitelist, whereas red cells belong to the blacklist. All whitelist/blacklist rules are summarised in [PITH_FULL_IMAGE:figures/full_fig_p0… view at source ↗

**Figure 3.** Balanced accuracy under various corruptions in a People Recognition task. ALIGN leads in clean accuracy, however, drops even under moderate noise, while CLIP remains more stable overall. Results: CLIP consistently achieves the lowest mCE across all domains, indicating strong overall robustness. However, some domains exhibit nuanced trends: in Medical Diagnosis and Manufacturing Quality, SigLIP surpasses CLIP in rCE… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Relative balanced accuracy under corruption for the Autonomous Driving domain [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

**Figure 5.** Figure 5: Relative balanced accuracy under corruption for the Manufacturing Quality domain 200 100 0 100 200 brightness 0.0 0.2 0.4 0.6 0.8 1.0 Acc. Brightness 100 0 100 200 300 contrast 0.0 0.2 0.4 0.6 0.8 1.0 Acc. Contrast No Horizontal direction 0.0 0.2 0.4 0.6 0.8 1.0 Acc. F…

**Figure 6.** Figure 6: Relative balanced accuracy under corruption for the Medical Diagnosis domain [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]

**Figure 7.** Figure 7: Relative balanced accuracy under corruption for the Satellite Imaging domain 200 100 0 100 200 brightness 0.0 0.2 0.4 0.6 0.8 1.0 Acc. Brightness 100 0 100 200 300 contrast 0.0 0.2 0.4 0.6 0.8 1.0 Acc. Contrast No Horizontal direction 0.0 0.2 0.4 0.6 0.8 1.0 Acc. Flip…

**Figure 8.** Figure 8: Relative balanced accuracy under corruption for the Handheld domain [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 27 canonical work pages

[1]

Data in brief28, 104863 (2020)

Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A.: Dataset of breast ultrasound images. Data in brief28, 104863 (2020)

2020
[2]

In: International conference on software engineering advances (ICSEA)

Ali, R.B., Ejbali, R., Zaied, M.: Detection and classification of dental caries in x-ray images using deep neural networks. In: International conference on software engineering advances (ICSEA). p. 236 (2016)

work page 2016
[3]

In: 2017 International conference of electronics, communication and aerospace tech- nology (ICECA)

Baby, M., Jereesh, A.: Automatic nerve segmentation of ultrasound images. In: 2017 International conference of electronics, communication and aerospace tech- nology (ICECA). vol. 1, pp. 107–112. IEEE (2017)

work page 2017
[4]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Bergmann, P., Fauser, M., Sattlegger, D., Steger, C.: Mvtec ad–a comprehen- sive real-world dataset for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9592–9600 (2019)

2019
[5]

arXiv preprint arXiv:2108.07258 (2021)

Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)

arXiv 2021
[6]

In: Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13

Bossard, L., Guillaumin, M., Van Gool, L.: Food-101–mining discriminative com- ponents with random forests. In: Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13. pp. 446–461. Springer (2014)

2014
[7]

Ad- vances in neural information processing systems13 (2000)

Chapelle, O., Weston, J., Bottou, L., Vapnik, V.: Vicinal risk minimization. Ad- vances in neural information processing systems13 (2000)

work page 2000
[8]

In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 2818–2829 (2023)

work page 2023

Show all 44 references

[9]

arXiv preprint arXiv:2010.09670 (2020)

Croce, F., Andriushchenko, M., Sehwag, V., Debenedetti, E., Flammarion, N., Chi- ang, M., Mittal, P., Hein, M.: Robustbench: a standardized adversarial robustness benchmark. arXiv preprint arXiv:2010.09670 (2020)

2020 arXiv
[10]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Cubuk,E.D.,Zoph,B.,Mane,D.,Vasudevan,V.,Le,Q.V.:Autoaugment:Learning augmentation strategies from data. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 113–123 (2019)

2019
[11]

Under review2(1) (2019)

Duchi, J.C., Hashimoto, T., Namkoong, H.: Distributionally robust losses against mixture covariate shifts. Under review2(1) (2019)

2019
[12]

In: International Conference on Machine Learning

Fang, A., Ilharco, G., Wortsman, M., Wan, Y., Shankar, V., Dave, A., Schmidt, L.: Data determines distributional robustness in contrastive language image pre- training (clip). In: International Conference on Machine Learning. pp. 6216–6234. PMLR (2022)

2022
[13]

In: International Confer- ence on Machine Learning

Garg, S., Erickson, N., Sharpnack, J., Smola, A., Balakrishnan, S., Lipton, Z.C.: Rlsbench: Domain adaptation under relaxed label shift. In: International Confer- ence on Machine Learning. pp. 10879–10928. PMLR (2023)

2023
[14]

In: 2012 IEEE conference on computer vision and pattern recognition

Geiger, A., Lenz, P., Urtasun, R.: Are we ready for autonomous driving? the kitti vision benchmark suite. In: 2012 IEEE conference on computer vision and pattern recognition. pp. 3354–3361. IEEE (2012)

2012
[15]

In: Neural information pro- cessing: 20th international conference, ICONIP 2013, daegu, korea, november 3-7,

Goodfellow, I.J., Erhan, D., Carrier, P.L., Courville, A., Mirza, M., Hamner, B., Cukierski, W., Tang, Y., Thaler, D., Lee, D.H., et al.: Challenges in representation learning: A report on three machine learning contests. In: Neural information pro- cessing: 20th international...

2013
[16]

arXiv preprint arXiv:1412.6572 (2014)

Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)

2014 arXiv
[17]

Radiology290(2), 498–503 (2019)

Halabi, S.S., Prevedello, L.M., Kalpathy-Cramer, J., Mamonov, A.B., Bilbily, A., Cicero,M.,Pan,I.,Pereira,L.A.,Sousa,R.T.,Abdala,N.,etal.:Thersnapediatric bone age machine learning challenge. Radiology290(2), 498–503 (2019)

2019
[18]

Machine Learning 114(3), 1–19 (2025)

Han, Z., Luo, G., Sun, H., Li, Y., Han, B., Gong, M., Zhang, K., Liu, T.: Alignclip: navigating the misalignments for robust vision-language generalization. Machine Learning 114(3), 1–19 (2025)

2025
[19]

Proceedings of the International Conference on Learning Representations (2019)

Hendrycks,D.,Dietterich,T.:Benchmarkingneuralnetworkrobustnesstocommon corruptions and perturbations. Proceedings of the International Conference on Learning Representations (2019)

2019
[20]

In: International Conference on Machine Learning

Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International Conference on Machine Learning. pp. 4904–4916. PMLR (2021)

2021
[21]

IEEE Access11, 9920–9930 (2023)

Joo, Y., Park, H.C., Lee, O.J., Yoon, C., Choi, M.H., Choi, C.: Classification of liver fibrosis from heterogeneous ultrasound image. IEEE Access11, 9920–9930 (2023)

2023
[22]

In: European Conference on Computer Vision

Kar, O.F., Tonioni, A., Poklukar, P., Kulshrestha, A., Zamir, A., Tombari, F.: Brave: Broadening the visual encoding of vision-language models. In: European Conference on Computer Vision. pp. 113–132. Springer (2024)

2024
[23]

In: International conference on machine learning

Koh, P.W., Sagawa, S., Marklund, H., Xie, S.M., Zhang, M., Balsubramani, A., Hu, W., Yasunaga, M., Phillips, R.L., Gao, I., et al.: Wilds: A benchmark of in- the-wild distribution shifts. In: International conference on machine learning. pp. 5637–5664. PMLR (2021)

2021
[24]

Scientific data4(1), 1–9 (2017)

Lee, R.S., Gimenez, F., Hoogi, A., Miyake, K.K., Gorovoy, M., Rubin, D.L.: A curated mammography data set for use in computer-aided detection and diagnosis research. Scientific data4(1), 1–9 (2017)

2017
[25]

In: 2022 International Joint Conference on Neural Networks (IJCNN)

Li, K., Lu, J., Zuo, H., Zhang, G.: Source-free multi-domain adaptation with gen- erally auxiliary model training. In: 2022 International Joint Conference on Neural Networks (IJCNN). pp. 1–8. IEEE (2022)

2022
[26]

Advances in neural information processing systems32 (2019)

Lim, S., Kim, I., Kim, T., Kim, C., Kim, S.: Fast autoaugment. Advances in neural information processing systems32 (2019)

2019
[27]

In: International Conference on Learning Representations (2018)

Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. In: International Conference on Learning Representations (2018)

2018
[28]

arXiv preprint arXiv:2412.02837 (2024)

Maharana, S.K., Zhang, B., Karlinsky, L., Feris, R., Guo, Y.: Enhancing robust- ness of clip to common corruptions through bimodal test-time adaptation. arXiv preprint arXiv:2412.02837 (2024)

2024 arXiv
[29]

https://openai.com/research/gpt-4 (2023), ac- cessed: 2025-01-09

OpenAI: Gpt-4 technical report. https://openai.com/research/gpt-4 (2023), ac- cessed: 2025-01-09

2023
[30]

In: International Conference on Machine Learning

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. pp. 8748–8763. PMLR (2021)

2021
[31]

International Journal of Computational Intelligence Systems17(1), 241 (2024) 24 Koddenbrock et al

Sarhan, A.M., Gobara, M., Yasser, S., Elsayed, Z., Sherif, G., Moataz, N., Yasir, Y., Moustafa, E., Ibrahim, S., Ali, H.A.: Knee osteoporosis diagnosis based on deep learning. International Journal of Computational Intelligence Systems17(1), 241 (2024) 24 Koddenbrock et al

2024
[32]

Advances in Neural Information Processing Systems (NeurIPS)33, 18583–18599 (2020)

Taori, R., Dave, A., Shankar, V., Carlini, N., Recht, B., Schmidt, L.: Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems (NeurIPS)33, 18583–18599 (2020)

2020
[33]

Springer Science & Business Media (1999)

Vapnik, V.: The nature of statistical learning theory. Springer Science & Business Media (1999)

1999
[34]

IEEE transactions on neural networks 10(5), 988–999 (1999)

Vapnik, V.N.: An overview of statistical learning theory. IEEE transactions on neural networks 10(5), 988–999 (1999)

1999
[35]

arXiv preprint arXiv:2405.00876 (2024)

Verma, P., Van, M.H., Wu, X.: Beyond human vision: The role of large vision language models in microscope image analysis. arXiv preprint arXiv:2405.00876 (2024)

2024 arXiv
[36]

arXiv preprint arXiv:2403.11497 (2024)

Wang, Q., Lin, Y., Chen, Y., Schmidt, L., Han, B., Zhang, T.: A sober look at the robustness of clips to spurious features. arXiv preprint arXiv:2403.11497 (2024)

2024 arXiv
[37]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Wang, X., Peng, Y., Lu, L., Lu, Z., Bagheri, M., Summers, R.M.: Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classi- fication and localization of common thorax diseases. In: Proceedings of the IEEE conference on computer vision and patte...

2017
[38]

arXiv preprint arXiv:1910.03771 (2019)

Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., et al.: Huggingface’s transformers: State-of- the-art natural language processing. arXiv preprint arXiv:1910.03771 (2019)

2019 arXiv
[39]

IEEE Transactions on Geoscience and Remote Sensing55(7), 3965–3981 (2017)

Xia, G.S., Hu, J., Hu, F., Shi, B., Bai, X., Zhong, Y., Zhang, L., Lu, X.: Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing55(7), 3965–3981 (2017)

2017
[40]

In: 12th Interna- tional Conference on Learning Representations, ICLR 2024 (2024)

Xu, H., Xie, S., Tan, X.E., Huang, P.Y., Howes, R., Sharma, V., Li, S.W., Ghosh, G., Zettlemoyer, L., Feichtenhofer, C.: Demystifying clip data. In: 12th Interna- tional Conference on Learning Representations, ICLR 2024 (2024)

2024
[41]

Advances in Neural Information Processing Systems 32 (2019)

Yin, D., Gontijo Lopes, R., Shlens, J., Cubuk, E.D., Gilmer, J.: A fourier per- spective on model robustness in computer vision. Advances in Neural Information Processing Systems 32 (2019)

2019
[42]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision

Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language im- age pre-training. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11975–11986 (2023)

2023
[43]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision

Zheng, Z., Ma, M., Wang, K., Qin, Z., Yue, X., You, Y.: Preventing zero-shot trans- fer degradation in continual learning of vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19125–19136 (2023)

2023
[2013]

Proceedings, Part III 20. pp. 117–124. Springer (2013) Domain Robustness of Vision-Language Models 23

2013

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Data in brief28, 104863 (2020)

Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A.: Dataset of breast ultrasound images. Data in brief28, 104863 (2020)

2020

[2] [2]

In: International conference on software engineering advances (ICSEA)

Ali, R.B., Ejbali, R., Zaied, M.: Detection and classification of dental caries in x-ray images using deep neural networks. In: International conference on software engineering advances (ICSEA). p. 236 (2016)

work page 2016

[3] [3]

In: 2017 International conference of electronics, communication and aerospace tech- nology (ICECA)

Baby, M., Jereesh, A.: Automatic nerve segmentation of ultrasound images. In: 2017 International conference of electronics, communication and aerospace tech- nology (ICECA). vol. 1, pp. 107–112. IEEE (2017)

work page 2017

[4] [4]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Bergmann, P., Fauser, M., Sattlegger, D., Steger, C.: Mvtec ad–a comprehen- sive real-world dataset for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9592–9600 (2019)

2019

[5] [5]

arXiv preprint arXiv:2108.07258 (2021)

Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)

arXiv 2021

[6] [6]

In: Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13

Bossard, L., Guillaumin, M., Van Gool, L.: Food-101–mining discriminative com- ponents with random forests. In: Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13. pp. 446–461. Springer (2014)

2014

[7] [7]

Ad- vances in neural information processing systems13 (2000)

Chapelle, O., Weston, J., Bottou, L., Vapnik, V.: Vicinal risk minimization. Ad- vances in neural information processing systems13 (2000)

work page 2000

[8] [8]

In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 2818–2829 (2023)

work page 2023