Pith. sign in

REVIEW 3 major objections 5 minor 47 references

Hypothesis Testing in Imaging Inverse Problems

T0 review · 3 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 a natural-language semantic hypothesis about an image can be tested from a single noisy measurement with controlled Type I error, using a vision-language embedding and an e-value.

desk verdict A genuinely novel e-value+VLM framework for semantic hypothesis testing in imaging, but Table 2 shows the Type I error guarantee fails on ImageWoof at α=2%, so the central claim needs major repair before it can be trusted. read the letter →

arxiv 2505.22481 v1 pith:75YOKBUN submitted 2025-05-28 stat.ML cs.LG

Yiming Xi , Konstantinos Zygalakis , Marcelo Pereyra This is my paper · ORCID
classification stat.MLcs.LG MSC 62F0362G1068T45
keywords semantichypothesistestingimaginginverseproblemse-valuesvision-languagemodelsCLIPself-supervisedTypeIerrorcontrolnoiseinjection
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

The paper's central claim is that a scientific hypothesis about the content of an image—stated in natural language, such as 'this is a photo of a grassland' versus 'this is a photo of farmland'—can be tested from a single noisy measurement without knowing the distribution of the test statistic under the null. This matters because many imaging experiments allow only one observation, and the questions that matter are semantic rather than pixel-level. The proposed procedure splits the measured image into two conditionally independent copies by noise injection, uses one copy to reconstruct the image, and embeds the reconstruction and the two text hypotheses in the shared image-text space of a vision-language model. The test statistic is the difference of the two embedding similarities, converted into an e-value $E = \exp\{-t(Y_2)\}$; if $E \ge 1/\alpha$ the null is rejected at level $\alpha$, with Type I error controlled by Markov's inequality. Experiments on flower and dog phenotyping show that this approach achieves higher statistical power than zero-shot classification with the same vision-language model while keeping Type I error close to the nominal level.

What carries the argument

The load-bearing object is the e-value $E=\exp\{-t(Y_2)\}$, built from a vision-language model that maps images and natural-language prompts to a shared unit-sphere embedding space; $t(Y_2)$ is the difference in cosine similarities (scaled by a temperature $\lambda$) between the reconstructed image and the two text prompts. Because $E$ is non-negative and has expectation at most $1$ under the null, Markov's inequality converts it into a valid test without a parametric null distribution. The companion mechanism is measurement splitting by noise injection, $Y_1=y+\tau Z$, $Y_2=y-Z/\tau$, which produces two conditionally independent copies of the experiment from one observation so that one copy can form the hypothesis and the other can test it.

What would settle it

Take the published flower or dog experiment, shift the deployment population (different species, different noise level, or different image resolution) from the calibration set, and count how often the test rejects on null images whose true class is known. If the empirical mean of $E$ under the null exceeds $1$ or the observed rejection rate at level $\alpha$ exceeds $\alpha$ by more than the reported CLIP baseline error, then the calibration-transfer assumption fails and the Type I error guarantee collapses.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that semantic hypothesis testing in imaging inverse problems can be made non-parametric and valid with a single measurement. The key step is to construct an e-value from the shared representation of a vision-language model: after splitting the noisy measurement $y$ into conditionally independent $y_1$ and $y_2$, the paper forms $t(Y_2)=D(\phi_x(\hat{x}(Y_2)),\phi_t(q_0))-D(\phi_x(\hat{x}(Y_2)),\phi_t(q_1))$ and tests $H_0:\mathbb{E}[e^{-t(Y_2)}]\le 1$ against $H_1:\mathbb{E}[e^{-t(Y_2)}]>1$. Since Markov's inequality gives $\Pr(e^{-t(Y_2)}\ge 1/\alpha)\le\alpha$ under $H_0$, the test needs no null distribution. The paper demonstrates the procedure on two phenotyping tasks with self-supervised and supervised reconstruction networks, reporting Type I error near the nominal level and power consistently above that of a temperature-scaled zero-shot classifier using the same vision-language model. A linearized analysis identifies what drives power: the semantic difference between the hypotheses should be large and not aligned with the reconstruction error on the embedding space.

Load-bearing premise

The whole guarantee depends on the assumption that a calibration step done on a small set of known images keeps the test's average e-value at most 1 when the test is used on new, unseen images; if the calibration images are not representative of the real deployment images, or the vision-language model does not actually tell the two hypotheses apart in that domain, the claimed error control fails.

Editorial extensions

If this is right

  • A single observation suffices for both image reconstruction and a statistically valid semantic hypothesis test, avoiding the need for a second independent measurement.
  • Type I error control holds distribution-free via Markov's inequality once an e-value is available, so no model of the null distribution is required.
  • The test becomes more powerful when the two text hypotheses are semantically far apart and when the reconstruction error in embedding space is not aligned with the semantic difference being tested.
  • The noise-injection splitting extends to Poisson, Gamma, and binomial measurement models, so the recipe applies beyond Gaussian denoising.
  • Using a more accurate reconstruction network increases power; even a fully self-supervised network outperforms zero-shot classification with the same vision-language model.

Reading between the lines

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

  • If the calibration-transfer assumption holds across deployment conditions, this same e-value recipe could be applied to any modality—audio, video, or text—where a pretrained multimodal embedding and a reconstruction or summary estimator exist, not just to imaging.
  • The power analysis suggests a practical design rule not stated as a formal result: choose the alternative prompt to be semantically opposite to the null and avoid forms the VLM handles poorly such as negation, since the test's power depends on the angle between text embeddings.
  • The temperature $\lambda$ is currently calibrated on a small null dataset; replacing that calibration with a distribution-free split-conformal e-value calibration would relax the strongest assumption and is a natural testable extension.
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

3 major / 5 minor

Summary. The paper proposes a framework for semantic hypothesis testing in imaging inverse problems. The method uses a noise-injection procedure to split a single measurement into two conditionally independent parts, one for reconstruction and one for testing. Hypotheses are expressed in natural language and scored with a CLIP embedding via a temperature-scaled cosine similarity. The test statistic is converted into an e-value E = exp{-t(Y2)}, and the null is rejected at level alpha when E >= 1/alpha, justified by Markov's inequality under the condition E(E) <= 1. The paper includes a linearized power analysis and experiments on Oxford 102 Flowers and ImageWoof for denoising and inpainting, comparing against zero-shot CLIP classification. The authors claim robust Type I error control and improved power.

Significance. If valid, the framework would be a valuable contribution: it addresses a real gap in uncertainty quantification for imaging inverse problems, connects e-values with modern vision-language models, and the measurement-splitting idea (from external work) is well suited to the single-measurement setting. The paper also contributes a theoretical analysis of power and extensive experiments, including self-supervised reconstruction networks, which strengthens its practical relevance. However, the central claim of Type I error control is not supported by the ImageWoof results, where the proposed method exceeds the nominal level at alpha=2%. The paper acknowledges a CLIP baseline but the proposed adjustment (subtracting the baseline) is not a valid hypothesis-testing procedure. The core e-value/Markov step is mathematically sound only if the e-value property holds, and the manuscript does not establish this property rigorously.

major comments (3)
  1. [Table 2, Section 4 (ImageWoof)] The proposed self-supervised test at significance level alpha=2% rejects 4.42% of null images, which exceeds the 2% level and violates the bound in Eq. (8). Since lambda is calibrated so that E(E) is approximately 0.98 under the null (Appendix C), Markov's inequality gives P(E >= 50) <= 1.96%, yet the observed rejection rate is 4.42%. This is direct evidence that the e-value property E(E) <= 1 does not hold on the ImageWoof test null population. The paper's response to subtract the 4.1% CLIP baseline (Section 4, 'leaving this baseline error aside') is not a valid Type I error adjustment: a hypothesis test's Type I error is the total false rejection probability, and a baseline that rejects 4.1% of null images is part of that probability. This undermines the manuscript's central claim of robust Type I error control.
  2. [Section 3.1.3, Eq. (7)] The null hypothesis is formalized as H0: E(E) <= 1 rather than as a statement about the underlying image x*. This is a legitimate e-value formulation provided the condition E(E) <= 1 holds under the actual null model. The paper does not derive this condition from the imaging model or the VLM; instead it is imposed by calibrating lambda on a small null dataset (Appendix C). No finite-sample guarantee, confidence interval, or transfer argument is provided to ensure that the calibrated lambda will satisfy the e-value property on the test population. The ImageWoof result above provides concrete evidence that this calibration-based guarantee fails out-of-sample, which is a load-bearing problem for the proposed method.
  3. [Table 2, proposed self-supervised row] The reported Type I errors for the proposed self-supervised method on ImageWoof are non-monotonic in alpha: 7.33% at alpha=15% but 3.71% at alpha=20%. Since the rejection region E >= 1/alpha is nested (larger alpha gives a smaller threshold and a larger rejection region), the Type I error must be non-decreasing in alpha for a fixed test. This non-monotonicity indicates either a numerical error or unaccounted variability in the experiments. The appendix (E.2) reports standard deviations only for statistical power, not for Type I errors, so readers cannot assess the reliability of these numbers. This should be corrected and clarified.
minor comments (5)
  1. [Abstract and Section 5] The abstract and conclusion state that the method 'robustly controls Type I errors', but Table 2 contradicts this on the ImageWoof dataset. The claims should be qualified to reflect the dependence on the VLM baseline and the calibration transfer.
  2. [Appendix C] The calibration procedure is described only as using a 'small calibration dataset' to set E(E) close to 0.98. The size of this set, the split between calibration and test, and the variability of the estimate of E(E) should be reported, since the validity of the e-value property depends on the accuracy of this calibration.
  3. [Section F, limitations] There is a typo in the limitations section: 'information that can extracted from the data' should read 'information that can be extracted from the data.'
  4. [Section 3.2] The power analysis relies on several approximations (the consistency of the estimator in the noiseless limit, the small-noise Jacobian expansion, and the norm approximation) that are not validated numerically. The analysis is presented as an intuition, but the manuscript should state that these are heuristic conditions rather than formal justifications, and ideally verify them on a simple example.
  5. [Table 4] The lambda values reported in Table 4 include a row labeled 'unsupervised' while the main text and Table 1 use 'self-supervised'. This terminology inconsistency should be harmonized to avoid confusion.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity; the e-value bound is Markov's inequality, lambda calibration is disclosed, and self-citations are background.

full rationale

The derivation chain is: split the measurement into conditionally independent y1 and y2 via Eq. (2), reconstruct, form the statistic t(Y2) in Eq. (4), set E = exp{-t} in Eq. (6), define H0 as E(E) <= 1 in Eq. (7), and apply Markov's inequality in Eq. (8). The mathematical guarantee is therefore a standard e-value argument, not an algebraic tautology. The one imposed ingredient is the temperature lambda, calibrated on a small null dataset so that E(E) is approximately 0.98 (Appendix C, Table 4). That is a disclosed validity assumption about transfer of calibration, and it can fail: Table 2 shows the proposed self-supervised test exceeding alpha = 2% on ImageWoof (4.42% Type I error). This is a correctness or validity failure, not circularity, because the reported Type I and power values are measured on test data separate from the calibration set; the calibration does not by construction force those numbers. The paper's use of its own prior work ([35], [40], [36], [15], [1], [44], [41]) is confined to background, reconstruction-architecture choice, and comparison baselines; none of these citations supplies the e-value property or the power comparison. The noise-injection splitting is attributed to [31], which is external to the author list. No equation in the paper is identical by construction to an input, and no fitted parameter is relabeled as a prediction. Score 2 reflects the pervasive but non-load-bearing self-citation, not actual circularity of the central derivation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on known statistical and machine learning components: measurement splitting (cited to [31]), CLIP embeddings, and e-values. The main uncharged assumption is that calibration of lambda yields a true e-variable in deployment; this is tuned rather than derived. No new physical entities, forces, or latent variables are introduced. The power analysis adds an unverified linearization assumption.

free parameters (4)
  • lambda (CLIP temperature) = Oxford Flowers: 3.05 (supervised), 1.44 (self-supervised); ImageWoof: 1.23 (supervised), 1.12 (self-supervised)
    Calibrated in Appendix C on a null dataset so that E(E) is approximately 0.98 under H0; it scales the e-variable and directly controls Type I error and power.
  • tau (noise-injection split parameter) = 1 (default)
    Chosen by hand in Section 3.1.1 and validated by ablation in Figure 4; it trades reconstruction accuracy against statistical power.
  • beta (Poisson splitting parameter) = 0.15
    Set following [31] for the Poisson image denoising experiment in Appendix D.
  • kappa (sign-test threshold) = Oxford Flowers supervised -0.030, unsupervised -0.060; ImageWoof supervised -0.170, unsupervised -0.165
    Tuned on calibration data for the bootstrap/sign-test alternative in Appendix E.4; not used in the main e-value test.
assumptions (5)
  • domain assumption Gaussian splitting Y1 = y + tau Z and Y2 = y - Z/tau with Z ~ N(0, Sigma) yields conditionally independent Y1 and Y2 given x* with mean Ax* (Eq. 2).
    Invoked in Section 3.1.1 to emulate experiment replication; the property is cited from [31] and not proved in this paper.
  • domain assumption The generalized natural exponential family splitting (Eqs. 13-16) preserves conditional independence and mean for Poisson, Gamma, and Binomial noise.
    Section 3.3 extends the method by relying on [31]; the construction of h1 and h2 is assumed valid without proof here.
  • ad hoc to paper After calibrating lambda, the e-variable E = exp{-t(Y2)} satisfies E(E) <= 1 under H0.
    Appendix C sets lambda so that the empirical mean of E under the null is about 0.98; the paper does not derive this condition from the imaging model or prove it transfers outside the calibration set.
  • domain assumption CLIP shared embeddings align cosine similarity with semantic agreement between reconstructed images and the text hypotheses q0 and q1.
    Sections 3.1.2 and 3.2 rely on this alignment; Section 5 acknowledges that CLIP struggles with negation and suffers from biases, weakening the assumption.
  • ad hoc to paper The power analysis linearization is accurate: the estimator is consistent in the noiseless limit, the noise perturbation is small, and the Jacobian approximation in Eqs. (9)-(11) holds.
    Section 3.2 uses these approximations to derive qualitative insight; no bound on the approximation error is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hypothesis Testing in Imaging Inverse Problems." pith.science (2026). https://pith.science/paper/75YOKBUN

@misc{pith2026250522481,
  author       = {Pith},
  title        = {Pith review of: Hypothesis Testing in Imaging Inverse Problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/75YOKBUN}},
  note         = {Machine review of arXiv:2505.22481}
}
read the original abstract

This paper proposes a framework for semantic hypothesis testing tailored to imaging inverse problems. Modern imaging methods struggle to support hypothesis testing, a core component of the scientific method that is essential for the rigorous interpretation of experiments and robust interfacing with decision-making processes. There are three main reasons why image-based hypothesis testing is challenging. First, the difficulty of using a single observation to simultaneously reconstruct an image, formulate hypotheses, and quantify their statistical significance. Second, the hypotheses encountered in imaging are mostly of semantic nature, rather than quantitative statements about pixel values. Third, it is challenging to control test error probabilities because the null and alternative distributions are often unknown. Our proposed approach addresses these difficulties by leveraging concepts from self-supervised computational imaging, vision-language models, and non-parametric hypothesis testing with e-values. We demonstrate our proposed framework through numerical experiments related to image-based phenotyping, where we achieve excellent power while robustly controlling Type I errors.

Figures

Figures reproduced from arXiv: 2505.22481 by the authors.

Figure 1
Figure 1. Overview of the proposed method. Given a noisy measurement y = Ax⋆ + √ Σϵ, noise-injection generates two new conditionally independent measurements y1 and y2, mimicking experiment replication: y1 is used for image reconstruction and formulation of the hypothesis test, y2 resolves the tests non-parametrically in the shared representation of a VLM and reports a p-value. equal significance level, we prefer the test wit… view at source ↗
Figure 2
Figure 2. Example images from Oxford 102 Flower and ImageWoof experiments: true image [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Example images from Oxford 102 Flower with Poisson noise: true image [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Trade-off between reconstruction quality and statistical power for different values of [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 33 canonical work pages

  1. [1]

    Self-supervised conformal prediction for uncertainty quantification in Poisson imaging problems

    Bernardin Tamo Amougou, Marcelo Pereyra, and Barbara Pascal. Self-supervised confor- mal prediction for uncertainty quantification in Poisson imaging problems. arXiv preprint arXiv:2502.19194, 2025. URL https://arxiv.org/abs/2502.19194

    work page Pith review arXiv 2025

  2. [2]

    Angelopoulos and Stephen Bates

    Anastasios N. Angelopoulos and Stephen Bates. Conformal prediction: A gentle introduction. Found. Trends Mach. Learn., 16(4):494–591, March 2023. ISSN 1935-8237. doi: 10.1561/ 2200000101. URL https://doi.org/10.1561/2200000101

    doi:10.1561/2200000101 2023

  3. [3]

    Computational imaging

    Ayush Bhandari, Achuta Kadambi, and Ramesh Raskar. Computational imaging. MIT Press, 2022

    work page 2022

  4. [4]

    Uncertainty quantification for radio interferometric imaging: II

    Xiaohao Cai, Marcelo Pereyra, and Jason D McEwen. Uncertainty quantification for radio interferometric imaging: II. MAP estimation. Mon. Not. R. Astron. Soc., 480(3):4170–4182, 2018

    work page 2018

  5. [5]

    Statistical inference

    George Casella and Roger Berger. Statistical inference. Duxbury Resource Center, June 2001. ISBN 0534243126

    work page 2001

  6. [6]

    Equivariant imaging: Learning beyond the range space

    Dongdong Chen, Julián Tachella, and Mike E Davies. Equivariant imaging: Learning beyond the range space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4379–4388, 2021

    2021

  7. [7]

    Robust equivariant imaging: a fully unsupervised framework for learning to image from noisy and partial measurements

    Dongdong Chen, Julián Tachella, and Mike E Davies. Robust equivariant imaging: a fully unsupervised framework for learning to image from noisy and partial measurements. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5647–5656, 2022

    work page 2022

  8. [8]

    Imaging with equivariant deep learning: From unrolled network design to fully unsupervised learning

    Dongdong Chen, Mike Davies, Matthias J Ehrhardt, Carola-Bibiane Schönlieb, Ferdia Sherry, and Julián Tachella. Imaging with equivariant deep learning: From unrolled network design to fully unsupervised learning. IEEE Signal Processing Magazine, 40(1):134–147, 2023

    work page 2023

Show all 47 references
  1. [9]

    Diffusion posterior sampling for general noisy inverse problems

    Hyungjin Chung, Jeongsol Kim, Michael Thompson Mccann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh Inter- national Conference on Learning Representations, 2023. URL https://openreview.net/ forum?id=OnD9zGAGT0k

    2023

  2. [10]

    Plug-and-play split Gibbs sampler: Embedding deep generative priors in Bayesian inference

    Florentin Coeurdoux, Nicolas Dobigeon, and Pierre Chainais. Plug-and-play split Gibbs sampler: Embedding deep generative priors in Bayesian inference. IEEE Trans. Image Process., 33: 3496–3507, 2024

    2024

  3. [11]

    Dimakis, and Mauricio Delbracio

    Giannis Daras, Hyungjin Chung, Chieh-Hsin Lai, Yuki Mitsufuji, Jong Chul Ye, Peyman Milanfar, Alexandros G. Dimakis, and Mauricio Delbracio. A survey on diffusion models for inverse problems. arXiv preprint arXiv:2410.00083, 2024. URL https://arxiv.org/abs/ 2410.00083

    arXiv

  4. [12]

    Imagenet: A large- scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large- scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009

    2009

  5. [13]

    The statistical sign test

    Wilfrid J Dixon and Alexander M Mood. The statistical sign test. Journal of the American Statistical Association, 41(236):557–566, 1946

    1946

  6. [14]

    Efficient Bayesian computation by proximal Markov chain Monte Carlo: When Langevin meets Moreau

    Alain Durmus, Éric Moulines, and Marcelo Pereyra. Efficient Bayesian computation by proximal Markov chain Monte Carlo: When Langevin meets Moreau. SIAM J. Imaging Sci., 11 (1):473–506, 2018

    2018

  7. [15]

    Self-supervised conformal prediction for uncertainty quantification in imaging problems

    Jasper M Everink, Bernardin Tamo Amougou, and Marcelo Pereyra. Self-supervised conformal prediction for uncertainty quantification in imaging problems. arXiv preprint arXiv:2502.05127, 2025

    2025 arXiv

  8. [16]

    Robust hypothesis testing using wasserstein uncertainty sets

    Rui Gao, Liyan Xie, Yao Xie, and Huan Xu. Robust hypothesis testing using wasserstein uncertainty sets. Advances in Neural Information Processing Systems, 31, 2018. 11

    2018

  9. [17]

    Minimax robust hypothesis testing

    Gökhan Gül and Abdelhak M Zoubir. Minimax robust hypothesis testing. IEEE Transactions on Information Theory, 63(9):5572–5587, 2017

    2017

  10. [18]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning - Volume 70, ICML’17, page 1321–1330. JMLR.org, 2017

    2017

  11. [19]

    Deep learning for computational imaging

    Reinhard Heckel. Deep learning for computational imaging. OXFORD University Press, 2025

    2025

  12. [20]

    Bayesian imaging with data-driven priors encoded by neural networks

    Matthew Holden, Marcelo Pereyra, and Konstantinos C Zygalakis. Bayesian imaging with data-driven priors encoded by neural networks. SIAM J. Imaging Sci., 15(2):892–924, 2022

    2022

  13. [21]

    A robust version of the probability ratio test

    Peter J Huber. A robust version of the probability ratio test. The Annals of Mathematical Statistics, pages 1753–1758, 1965

    1965

  14. [22]

    Statistical and computational inverse problems, volume 160

    Jari Kaipio and Erkki Somersalo. Statistical and computational inverse problems, volume 160. Springer Science & Business Media, 2006

    2006

  15. [23]

    Kamilov, Charles A

    Ulugbek S. Kamilov, Charles A. Bouman, Gregery T. Buzzard, and Brendt Wohlberg. Plug-and- play methods for integrating physical and learned models in computational imaging: Theory, algorithms, and applications. IEEE Signal Processing Magazine , 40(1):85–97, 2023. doi: 10.1109/...

    2023

  16. [24]

    Bayesian imaging using plug & play priors: When Langevin meets Tweedie

    Rémi Laumont, Valentin De Bortoli, Andrés Almansa, Julie Delon, Alain Durmus, and Marcelo Pereyra. Bayesian imaging using plug & play priors: When Langevin meets Tweedie. SIAM J. Imaging Sci., 15(2):701–737, 2022

    2022

  17. [25]

    Robust hypothesis testing with a relative entropy tolerance

    Bernard C Levy. Robust hypothesis testing with a relative entropy tolerance. IEEE Transactions on Information Theory, 55(1):413–421, 2008

    2008

  18. [26]

    Scalable Bayesian uncertainty quantification with data-driven priors for radio interferometric imaging

    Tobías I Liaudat, Matthijs Mars, Matthew A Price, Marcelo Pereyra, Marta M Betcke, and Jason D McEwen. Scalable Bayesian uncertainty quantification with data-driven priors for radio interferometric imaging. RAS Techniques and Instruments, 3(1):505–534, 2024

    2024

  19. [27]

    Remoteclip: A vision language foundation model for remote sensing

    Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  20. [28]

    Directional statistics

    Kanti V Mardia and Peter E Jupp. Directional statistics. Wiley Series in Probability and Statistics. John Wiley & Sons, Chichester, England, November 1999

    1999

  21. [29]

    Denoising: A powerful building-block for imaging, inverse problems, and machine learning

    Peyman Milanfar and Mauricio Delbracio. Denoising: A powerful building-block for imaging, inverse problems, and machine learning. arXiv preprint arXiv:2409.06219 , 2024. URL https://arxiv.org/abs/2409.06219

    2024 arXiv

  22. [30]

    Revisiting the calibration of modern neural networks

    Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Ann Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. In A. Beygelzimer, Y . Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural I...

    2021

  23. [31]

    Generalized recorrupted-to-recorrupted: Self-supervised learning beyond Gaussian noise

    Brayan Monroy, Jorge Bacca, and Julián Tachella. Generalized recorrupted-to-recorrupted: Self-supervised learning beyond Gaussian noise. arXiv preprint arXiv:2412.04648, 2025. URL https://arxiv.org/abs/2412.04648

    2025 arXiv

  24. [32]

    Learned reconstruction methods with convergence guarantees: A survey of concepts and applications

    Subhadip Mukherjee, Andreas Hauptmann, Ozan Öktem, Marcelo Pereyra, and Carola-Bibiane Schönlieb. Learned reconstruction methods with convergence guarantees: A survey of concepts and applications. IEEE Signal Processing Magazine, 40(1):164–182, 2023

    2023

  25. [33]

    Automated flower classification over a large number of classes

    Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, Dec 2008. 12

    2008

  26. [34]

    Know "no” better: A data-driven approach for enhancing negation awareness in CLIP

    Junsung Park, Jungbeom Lee, Jongyoon Song, Sangwon Yu, Dahuin Jung, and Sungroh Yoon. Know "no” better: A data-driven approach for enhancing negation awareness in CLIP. arXiv preprint arXiv:2501.10913, 2025. URL https://arxiv.org/abs/2501.10913

    2025

  27. [35]

    Maximum-a-Posteriori estimation with Bayesian confidence regions

    Marcelo Pereyra. Maximum-a-Posteriori estimation with Bayesian confidence regions. SIAM J. Imaging Sci., 10(1):285–302, 2017

    2017

  28. [36]

    Equivariant bootstrapping for uncertainty quantification in imaging inverse problems

    Marcelo Pereyra and Julián Tachella. Equivariant bootstrapping for uncertainty quantification in imaging inverse problems. In AISTATS, 2024

    2024

  29. [37]

    Hero, and Steve McLaughlin

    Marcelo Pereyra, Philip Schniter, Émilie Chouzenoux, Jean-Christophe Pesquet, Jean-Yves Tourneret, Alfred O. Hero, and Steve McLaughlin. A survey of stochastic simulation and optimization methods in signal processing.IEEE Journal of Selected Topics in Signal Processing, 10(2):...

    2016

  30. [38]

    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. In Proceedings of the 38th International Conference...

    2021

  31. [39]

    Hypothesis testing with e-values

    Aaditya Ramdas and Ruodu Wang. Hypothesis testing with e-values. arXiv preprint arXiv:2410.23614, 2025. URL https://arxiv.org/abs/2410.23614

    2025 arXiv

  32. [40]

    Scalable Bayesian uncertainty quantification in imaging inverse problems via convex optimization

    Audrey Repetti, Marcelo Pereyra, and Yves Wiaux. Scalable Bayesian uncertainty quantification in imaging inverse problems via convex optimization. SIAM J. Imaging Sci., 12(1):87–118, 2019

    2019

  33. [41]

    Latino-pro: Latent consistency inverse solver with prompt optimization

    Alessio Spagnoletti, Jean Prost, Andrés Almansa, Nicolas Papadakis, and Marcelo Pereyra. Latino-pro: Latent consistency inverse solver with prompt optimization. arXiv preprint arXiv:2503.12615, 2025

    2025 arXiv

  34. [42]

    Deepinverse: A deep learning framework for inverse problems in imaging

    J Tachella, D Chen, S Hurault, M Terris, and A Wang. Deepinverse: A deep learning framework for inverse problems in imaging. URL: https://deepinv. github. io/deepinv, 2023

    2023

  35. [43]

    Reconstruct any- thing model: a lightweight foundation model for computational imaging

    Matthieu Terris, Samuel Hurault, Maxime Song, and Julian Tachella. Reconstruct any- thing model: a lightweight foundation model for computational imaging. arXiv preprint arXiv:2503.08915, 2025

    2025

  36. [44]

    David Y . W. Thong, Charlesquin Kemajou Mbakam, and Marcelo Pereyra. Do Bayesian imaging methods report trustworthy probabilities? arXiv preprint arXiv:2405.08179, 2024. URL https://arxiv.org/abs/2405.08179

    2024 arXiv

  37. [45]

    MedCLIP: Contrastive learning from unpaired medical images and text

    Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. MedCLIP: Contrastive learning from unpaired medical images and text. Proc. Conf. Empir. Methods Nat. Lang. Process., 2022: 3876–3887, December 2022

    2022

  38. [46]

    All of nonparametric statistics

    Larry Wasserman. All of nonparametric statistics. Springer Texts in Statistics. Springer, New York, NY , 1 edition, May 2007. 13 A Experimental Details A.1 Model and training details Architecture For our image reconstruction network ˆx(·) we follow [36] and use an architecture...

    2007

  39. [47]

    limited risk

    without any additional fine-tuning. CLIP (Contrastive Language–Image Pre-training) is a neural network that learns visual concepts from natural language supervision. After leveraging a Vision Transformer and a Text Transformer, CLIP efficiently achieves great zero shot perform...

Pith tools

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