Pith. sign in

REVIEW 4 major objections 6 minor 60 references

Generative imaging for radio interferometry with fast uncertainty quantification

T0 review · 4 major / 6 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 A conditional GAN whose generator embeds the telescope's measurement operator provides fast radio image reconstruction with uncertainty maps that track real errors.

desk verdict Solid and honest methods paper; the reconstruction results stand, but the uncertainty maps are not yet shown to be calibrated, and the authors say so. read the letter →

arxiv 2507.21270 v1 pith:S5EJ7L2B submitted 2025-07-28 astro-ph.IM cs.LG

Matthijs Mars , Tobías I. Liaudat , Jessica J. Whitney , Marta M. Betcke , Jason D. McEwen This is my paper
classification astro-ph.IMcs.LG
keywords radiointerferometrygenerativeadversarialnetworkuncertaintyquantificationimagereconstructionconditionalGANgradientU-NetSKAposteriorsampling
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 aims to make radio interferometric imaging fast and uncertainty-aware for next-generation telescopes such as the SKA. It proposes RI-GAN, a conditional generative adversarial network that takes a dirty image, the telescope's point spread function, and a noise field as inputs, and outputs approximate samples of the posterior image. The authors claim that integrating the measurement operator directly into the generator, through a gradient U-Net (GU-Net), gives higher reconstruction quality, smaller uncertainties, and better generalisation to unseen uv-coverages and dynamic ranges than the U-Net variant. They further claim that the resulting uncertainty maps correlate with actual reconstruction errors, which would give radio astronomers per-pixel error bars at a fraction of the compute cost of iterative methods, without retraining for each observation.

What carries the argument

The central machinery is the regularised conditional GAN (rcGAN) objective paired with a gradient U-Net (GU-Net) generator. The rcGAN loss augments the Wasserstein adversarial loss with an L1 reconstruction term on the N-sample average and a standard-deviation reward that encourages sample diversity, with the reward weight tuned automatically so that the single-sample error and the N-sample-average error satisfy E1/EN = 2N/(N+1), the relation expected of independent posterior samples. The GU-Net generator is a U-Net whose encoder and decoder layers are injected with an approximate gradient of the data-fidelity term, $\nabla_x L \propto \Phi^*(y - \Phi x)$, approximated at each scale by a convolution with the point spread function. This injection is what embeds the measurement operator into the network, letting the generator adapt to varying uv-coverages at inference time.

What would settle it

Compute the empirical coverage probability of the posterior samples on held-out images: for each pixel, count how often the true value falls inside the 68% interval defined by the sample standard deviation; if the observed coverage is far below 68%, the uncertainty estimates are overconfident and the central claim about informative uncertainty fails.

Watch

Extended reading notes

Core claim

The central claim is that a conditional GAN trained with the regularised adversarial objective of the rcGAN framework, combined with a generator that explicitly embeds the measurement operator at multiple resolutions, delivers fast approximate Bayesian imaging for radio interferometry. Averaging 32 generated posterior samples gives the reconstruction, and the pixel-standard deviation across those samples is the uncertainty map. On simulated MeerKAT observations, the GU-Net variant outperforms the U-Net variant in signal-to-noise ratio and yields uncertainty maps that correlate with the absolute error (0.69 versus 0.58). On an out-of-distribution 30 Doradus image with dynamic range about 600, the GU-Net RI-GAN reaches an SNR of 46.03 dB, beating the U-Net variant (30.80 dB) and the CLEAN algorithm (32.55 dB). The authors conclude that explicitly including the measurement operator in the generator is what provides robustness and generalisation, and they identify posterior coverage as the property still to be verified.

Load-bearing premise

The uncertainty maps are trustworthy only if the training loss really makes the generated samples match the true posterior in mean and covariance, a property proven for conditional Gaussian distributions and not checked here with a coverage test.

Editorial extensions

If this is right

  • A single observation can be reconstructed with a per-pixel uncertainty map from 32 posterior samples generated in parallel on GPUs, with quality and uncertainty-error correlation saturated at that sample count.
  • The GU-Net variant should be preferred over the U-Net variant because it delivers higher SNR, lower estimated uncertainties, and better out-of-distribution performance on the same training budget.
  • The uncertainty maps correlate with the absolute error at 0.69 for the GU-Net, giving astronomers a qualitative indicator of where the reconstruction is uncertain and where artifacts or missing flux may lurk.
  • The approach scales to SKA-era data volumes because the expensive measurement operator is evaluated only once per observation to form the dirty image and PSF, not at every optimisation iteration.

Reading between the lines

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

  • The mean-covariance matching guarantee behind the uncertainty maps is proven only under conditional Gaussian assumptions; a natural next step beyond this paper is to compute posterior coverage on realistic radio images, especially with bright point sources, where the posterior is strongly non-Gaussian.
  • The PSF-convolution approximation to the gradient neglects wide-field (w-term) effects, so extending the same operator-embedding idea to full w-projection is an untested but plausible route to preserving the generalisation gains on non-coplanar baselines.
  • The observed saturation at 32 samples suggests a practical sampling budget for future cGAN-based radio imaging, though the exact number likely depends on image complexity and should be re-estimated for each new training set.
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 / 6 minor

Summary. The paper presents RI-GAN, a conditional GAN framework for radio interferometric image reconstruction with fast approximate uncertainty quantification. Building on the regularised cGAN (rcGAN) of Bendel et al. (2023), the authors propose two generator architectures: a standard U-Net and a Gradient U-Net (GU-Net) that explicitly embeds an approximate data-fidelity gradient into the network. The models are trained on simulated MeerKAT visibilities from IllustrisTNG galaxies and evaluated on a held-out test set and on a single simulated 30 Doradus observation. The central empirical claims are that the GU-Net RI-GAN outperforms the U-Net variant and CLEAN in reconstruction SNR, generalises better to out-of-distribution (higher dynamic range) data, and provides uncertainty estimates that correlate with reconstruction errors. The paper also examines how reconstruction quality and uncertainty correlation vary with the number of posterior samples.

Significance. If the uncertainty estimates are statistically valid, the paper would be a valuable contribution, offering a fast, scalable alternative to MCMC-based imaging for SKA-era data volumes. The GU-Net architectural idea is a sensible extension of previous work by Mars et al. (2025), and the paper is commendable for releasing code and trained models. The reported SNR gains on the 30 Doradus example are suggestive. However, the central novelty of the paper is uncertainty quantification, and that part is not currently validated: the authors explicitly defer verification of whether the generated samples follow the true posterior. The paper is therefore best viewed as an architectural study with promising but not yet established UQ credentials.

major comments (4)
  1. [Section 6 (Conclusion)] The final paragraph of the conclusion explicitly concedes that the statistical significance of the uncertainties has not been verified: 'Further work is needed to perform a more in depth analysis of the statistical significance of the uncertainties to verify whether the generated samples follow the true posterior distribution.' Since the abstract and introduction advertise 'uncertainty quantification' and 'informative uncertainty quantification' as central contributions, this gap is load-bearing. The only quantitative evidence offered, the Pearson correlation between sample standard deviation and absolute reconstruction error (Section 5.2, Eq. 18), is a weak necessary condition: a heuristic that inflates uncertainty in high-error regions can achieve a positive correlation while badly over- or under-covering the truth. The manuscript should include a coverage or calibration test on the simulated test set, e.g., using the methods of Hermans et al. (2022) or Lemos et al. (2023) that are already cited, before claiming that the uncertainty maps are informative.
  2. [Section 3, Eqs. (10)-(11)] The mean and covariance matching guarantees are imported from Bendel et al. (2023, Prop. 3.1), which requires that both the true image and the generated samples are conditionally Gaussian given the observation y. No argument is given that radio-interferometric posteriors satisfy this assumption, and no empirical check is reported. Consequently, the interpretation of the sample mean as the posterior mean (Section 4.4) and the sample standard deviation as the posterior uncertainty rests on an unverified assumption. The authors should either justify the Gaussian assumption for their setting or compare their uncertainty maps against a reference posterior (e.g., MCMC or MGVI) on a few representative test images.
  3. [Section 5.2] The correlation between uncertainty and absolute error is reported for a single example image ('for a particular reconstruction', 'for an example image'), with the shaded region representing the standard deviation of 100 repeated generations. Yet the conclusion in Section 6 states generally that 'the uncertainties are representative of the actual errors' and that the GU-Net 'consistently outperforms' in uncertainty correlation for any number of samples. Dataset-level claims require the correlation to be computed across the full test set, reporting the distribution rather than one example. The same limitation affects the statement in Section 5.2 that the GU-Net outperforms the U-Net in uncertainty correlation for any sample count.
  4. [Section 5.3] The out-of-distribution generalization claim, which also appears in the abstract ('generalises to images with an increased dynamic range'), is based on a single 30 Doradus image with a single uv-coverage. No error bars, multiple random seeds, or multiple uv-coverages are provided, and the uncertainty maps for this case are not quantitatively evaluated. To substantiate the generalization claim, the authors should evaluate on a small set of out-of-distribution images with varied uv-coverages and report the SNR distribution (analogous to Figure 3) and, if possible, an uncertainty diagnostic.
minor comments (6)
  1. [Eq. (6)] The subscript 'discrimator' is a typo; it should be 'discriminator'.
  2. [Section 5.2 heading] The heading 'Number or samples' contains a typo; it should read 'Number of samples'.
  3. [Eqs. (13)-(14)] The notation for the error quantities is inconsistent: Eq. (13) defines dE_N and cE_1, while Eq. (14) uses cE_1 and cEN. Please standardise the notation, e.g., use \hat{E}_1 and \hat{E}_N.
  4. [Eq. (16) and surrounding text] The symbol \tilde{\nabla}_{x,i} (rendered as 'e∇_{𝑥,𝑖}' in the typeset text) is not defined; please clarify whether this denotes an approximate or sub-sampled gradient.
  5. [Figure 3 caption] The caption states that dashed and dotted lines indicate the mean and quartiles, but does not specify which is which; please make this explicit.
  6. [Section 2, first paragraph] There is a grammatical error: 'using a an array of radio telescopes' should be 'using an array of radio telescopes'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central empirical claims are tested against held-out data and ground truth; the unverified posterior-coverage point is a validation limitation, not a circular reduction.

full rationale

No load-bearing step in the derivation reduces to its own inputs. The GU-Net versus U-Net comparison, the out-of-distribution 30 Doradus SNR results, and the uncertainty-error correlations are all evaluated on held-out simulated measurements against ground-truth images (Section 5), so no fitted parameter is relabeled as a prediction. The uncertainty quantification rests on Bendel et al. (2023) Proposition 3.1 (Eqs. 10-11), an external theoretical result with stated conditional-Gaussian assumptions; the paper explicitly concedes in Section 6 that statistical significance and posterior coverage of the uncertainties remain to be verified, which is an honest limitation rather than a circular argument. Self-citations to Mars et al. (2023, 2025) motivate the GU-Net architecture and simulation setup, but the empirical advantage of the GU-Net is re-established in this paper's own experiments, so these citations are not load-bearing. No instance of self-definition, fitted-input-as-prediction, imported uniqueness, ansatz-smuggling, or renaming of a known result was found.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claims depend on several hand-chosen hyperparameters (adversarial and diversity weights, sample counts) and on domain assumptions about the measurement model, posterior Gaussianity, and training distribution. No new physical entities are introduced.

free parameters (6)
  • beta_adv = not stated
    Adversarial loss weight in Eq. 7, set by hand; balances the adversarial objective against the L1 and standard deviation rewards.
  • beta_SD_0 = not stated
    Initial standard deviation reward weight in Eq. 7; auto-tuned via Eq. 14 starting from this hand-chosen value, and controls sample diversity and posterior covariance.
  • mu_SD = not stated
    Learning rate for the beta_SD update in Eq. 14, chosen by hand; affects how quickly the diversity reward weight converges.
  • N_train = 2
    Number of posterior samples used in the training regularisation terms; chosen for computational efficiency.
  • N_val = 8
    Number of samples used for validation and automatic beta_SD tuning; hand-chosen.
  • N_eval = 32
    Number of samples used at evaluation time; chosen because reconstruction quality and uncertainty correlation saturate near this number.
assumptions (4)
  • domain assumption Linear measurement model y = Phi x + n with additive noise (Eq. 2)
    Standard radio interferometry forward model used throughout; assumes the measured visibilities are a linear function of the sky brightness plus noise.
  • domain assumption Conditional Gaussian posterior for the rcGAN regularisation theory (Bendel et al. 2023, Prop. 3.1, invoked in Eqs. 10-11)
    The claim that generated samples match the true posterior mean and covariance relies on this assumption, which is not verified for radio interferometric images.
  • domain assumption PSF convolution approximation for the gradient of the data-fidelity term (Eq. 16)
    The GU-Net generator approximates Phi*Phi as a convolution with the PSF, accurate for coplanar baselines but neglecting wide-field and w-term effects.
  • domain assumption IllustrisTNG simulated galaxies are a representative training distribution
    The models are trained only on simulated galaxies; generalisation to other sky morphologies is tested on a single 30 Doradus image and the authors list more representative training as future work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative imaging for radio interferometry with fast uncertainty quantification." pith.science (2026). https://pith.science/paper/S5EJ7L2B

@misc{pith2026250721270,
  author       = {Pith},
  title        = {Pith review of: Generative imaging for radio interferometry with fast uncertainty quantification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S5EJ7L2B}},
  note         = {Machine review of arXiv:2507.21270}
}
read the original abstract

With the rise of large radio interferometric telescopes, particularly the SKA, there is a growing demand for computationally efficient image reconstruction techniques. Existing reconstruction methods, such as the CLEAN algorithm or proximal optimisation approaches, are iterative in nature, necessitating a large amount of compute. These methods either provide no uncertainty quantification or require large computational overhead to do so. Learned reconstruction methods have shown promise in providing efficient and high quality reconstruction. In this article we explore the use of generative neural networks that enable efficient approximate sampling of the posterior distribution for high quality reconstructions with uncertainty quantification. Our RI-GAN framework, builds on the regularised conditional generative adversarial network (rcGAN) framework by integrating a gradient U-Net (GU-Net) architecture - a hybrid reconstruction model that embeds the measurement operator directly into the network. This framework uses Wasserstein GANs to improve training stability in combination with regularisation terms that combat mode collapse, which are typical problems for conditional GANs. This approach takes as input the dirty image and the point spread function (PSF) of the observation and provides efficient, high-quality image reconstructions that are robust to varying visibility coverages, generalises to images with an increased dynamic range, and provides informative uncertainty quantification. Our methods provide a significant step toward computationally efficient, scalable, and uncertainty-aware imaging for next-generation radio telescopes.

Figures

Figures reproduced from arXiv: 2507.21270 by the authors.

Figure 1
Figure 1. Our RI-GAN framework consisting of the generator and discriminator networks. From the true image we simulate measurements from which we can create the dirty image and PSF for an observation. These two images, combined with a latent code image, are used as input to the generator network. The generator network outputs an approximate sample from the posterior distribution. The discriminator gets passed either a generat… view at source ↗
Figure 2
Figure 2. Reconstructions using U-Net and GU-Net RI-GANs from simulated measurements of galaxies in test set. The reconstructions are calculated by averaging over 32 generated approximate posterior samples and the uncertainties are obtained by taking the standard deviation over these samples. Both models provide a large improvement over the dirty image. The GU-Net, which includes updates using an approximate gradient of the d… view at source ↗
Figure 3
Figure 3. Distribution of the SNR for reconstructions using the U-Net and GU-Net RI-GANs for both the training and test set. The reconstructions are made by averaging over 32 generated posterior samples. Over both the train and test set, the GU-Net provides reconstructions with higher SNR compared to the U-Net. Besides that, the performance between the training and test set is similar for the GU-Net RI-GAN indicating that the… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Effect of the number of samples on reconstruction quality and the correlation between the uncertainty and the absolute error for both the U-Net and the GU-Net RI-GANs for an example image. The shaded area represents the standard deviation of repeating the generation 10…
Figure 5
Figure 5. Figure 5: Reconstructions using U-Net and GU-Net RI-GANs from simulated measurements of an out-of-distribution image of the 30 Doradus region using an unseen uv-coverage. The reconstructions are calculated by averaging over 32 generated approximate posterior samples and the unce…
Figure 6
Figure 6. Figure 6: Reconstructions using U-Net and GU-Net RI-GANs from simulated measurements of an out-of-distribution image of the 30 Doradus region using an unseen uv-coverage. The reconstructions are calculated by averaging over 32 generated approximate posterior samples and the unce…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 31 canonical work pages

  1. [1]

    A., ed., , Data-Driven Models in Inverse Problems

    Adler J., \"O ktem O., 2025, in Bubba T. A., ed., , Data-Driven Models in Inverse Problems. De Gruyter, Berlin, Boston, pp 359--412, @doi doi:10.1515/9783111251233-011

    work page doi:10.1515/9783111251233-011 2025

  2. [2]

    S., Dabbech A., Wiaux Y., 2024, @doi [The Astrophysical Journal Supplement Series] 10.3847/1538-4365/ad46f5 , 273, 3

    Aghabiglou A., Chu C. S., Dabbech A., Wiaux Y., 2024, @doi [The Astrophysical Journal Supplement Series] 10.3847/1538-4365/ad46f5 , 273, 3

    doi:10.3847/1538-4365/ad46f5 2024

  3. [3]

    Thesis , University College London

    Allam Jr T., 2016, MSc . Thesis , University College London

    work page 2016

  4. [4]

    u th C., Kruse J., Rother C., K \

    Ardizzone L., L \"u th C., Kruse J., Rother C., K \"o the U., 2019, Guided Image Generation with Conditional Invertible Neural Networks ( @eprint arXiv 1907.02392 ), @doi 10.48550/arXiv.1907.02392

    arXiv doi:10.48550/arxiv.1907.02392 2019

  5. [5]

    Arjovsky M., Bottou L., 2017, in International Conference on Learning Representations

    work page 2017

  6. [6]

    PMLR, pp 214--223

    Arjovsky M., Chintala S., Bottou L., 2017, in Proceedings of the 34th International Conference on Machine Learning . PMLR, pp 214--223

    work page 2017

  7. [7]

    A., 2018, in 2018 26th European Signal Processing Conference ( EUSIPCO )

    Arras P., Knollrn \"u ller J., Junklewitz H., En lin T. A., 2018, in 2018 26th European Signal Processing Conference ( EUSIPCO ). pp 2683--2687, @doi 10.23919/EUSIPCO.2018.8553533

    work page arXiv 2018

  8. [8]

    Bendel M., Ahmad R., Schniter P., 2023, Advances in Neural Information Processing Systems, 36, 68673

    work page 2023

Show all 60 references
  1. [9]

    J., 2004, @doi [Astronomy & Astrophysics] 10.1051/0004-6361:20040354 , 426, 747

    Bhatnagar S., Cornwell T. J., 2004, @doi [Astronomy & Astrophysics] 10.1051/0004-6361:20040354 , 426, 747

    2004 doi

  2. [10]

    J., Golap K., Uson J

    Bhatnagar S., Cornwell T. J., Golap K., Uson J. M., 2008, @doi [Astronomy & Astrophysics] 10.1051/0004-6361:20079284 , 487, 419

    2008 doi

  3. [11]

    L., 1998, in Advanced Technology MMW , Radio , and Terahertz Telescopes

    Brown R. L., 1998, in Advanced Technology MMW , Radio , and Terahertz Telescopes . SPIE, pp 231--237, @doi 10.1117/12.317357

    1998 doi

  4. [12]

    D., 2018a, @doi [Monthly Notices of the Royal Astronomical Society] 10.1093/mnras/sty2004 , 480, 4154

    Cai X., Pereyra M., McEwen J. D., 2018a, @doi [Monthly Notices of the Royal Astronomical Society] 10.1093/mnras/sty2004 , 480, 4154

    doi

  5. [13]

    D., 2018b, @doi [Monthly Notices of the Royal Astronomical Society] 10.1093/mnras/sty2015 , 480, 4170

    Cai X., Pereyra M., McEwen J. D., 2018b, @doi [Monthly Notices of the Royal Astronomical Society] 10.1093/mnras/sty2015 , 480, 4170

    doi

  6. [14]

    E., McEwen J

    Carrillo R. E., McEwen J. D., Wiaux Y., 2012, @doi [Monthly Notices of the Royal Astronomical Society] 10.1111/j.1365-2966.2012.21605.x , 426, 1223

    2012

  7. [15]

    E., McEwen J

    Carrillo R. E., McEwen J. D., Van De Ville D., Thiran J.-P., Wiaux Y., 2013, @doi [IEEE Signal Processing Letters] 10.1109/LSP.2013.2259813 , 20, 591

    2013

  8. [16]

    Cherif M., Liaudat T. I., Kern J., Kervazo C., Bobin J., 2024, Uncertainty Quantification for Fast Reconstruction Methods Using Augmented Equivariant Bootstrap: Application to Radio Interferometry ( @eprint arXiv 2410.23178 ), @doi 10.48550/arXiv.2410.23178

    2024 doi arXiv

  9. [17]

    L., Ravi V., Hallinan G., 2022, @doi [Monthly Notices of the Royal Astronomical Society] 10.1093/mnras/stac1329 , 514, 2614

    Connor L., Bouman K. L., Ravi V., Hallinan G., 2022, @doi [Monthly Notices of the Royal Astronomical Society] 10.1093/mnras/stac1329 , 514, 2614

    2022 doi

  10. [18]

    S., Wiaux Y., 2023, CLEANing Cygnus A Deep and Fast with R2D2 , https://arxiv.org/abs/2309.03291v3

    Dabbech A., Aghabiglou A., Chu C. S., Wiaux Y., 2023, CLEANing Cygnus A Deep and Fast with R2D2 , https://arxiv.org/abs/2309.03291v3

    2023 arXiv

  11. [19]

    Dewdney P., Hall P., Schilizzi R., Lazio T., 2009, @doi [Proceedings of the IEEE] 10.1109/JPROC.2009.2021005 , 97, 1482

    2009

  12. [20]

    Dia N., Yantovski-Barth M. J., Adam A., Bowles M., Perreault-Levasseur L., Hezaveh Y., Scaife A., 2025, IRIS : A Bayesian Approach for Image Reconstruction in Radio Interferometry with Expressive Score-Based Priors ( @eprint arXiv 2501.02473 ), @doi 10.48550/arXiv.2501.02473

    2025 doi arXiv

  13. [21]

    Drozdova M., et al., 2024, @doi [Astronomy & Astrophysics] 10.1051/0004-6361/202347948 , 683, A105

    2024 doi

  14. [22]

    Society of Exploration Geophysicists, pp 1135--1138, @doi 10.1190/1.1885591

    Duijndam A., Schonewille M., 1997, in SEG Technical Program Expanded Abstracts 1997. Society of Exploration Geophysicists, pp 1135--1138, @doi 10.1190/1.1885591

    1997 doi

  15. [23]

    Dutt A., Rokhlin V., 1993, @doi [SIAM Journal on Scientific Computing] 10.1137/0914081 , 14, 1368

    1993 doi

  16. [24]

    Edupuganti V., Mardani M., Vasanawala S., Pauly J., 2021, @doi [IEEE Transactions on Medical Imaging] 10.1109/TMI.2020.3025065 , 40, 239

    2021

  17. [25]

    Fessler J., Sutton B., 2003, @doi [IEEE Transactions on Signal Processing] 10.1109/TSP.2002.807005 , 51, 560

    2003

  18. [26]

    Goodfellow I. J., Pouget-Abadie J., Mirza M., Xu B., Warde-Farley D., Ozair S., Courville A., Bengio Y., 2014, Generative Adversarial Networks ( @eprint arXiv 1406.2661 ), @doi 10.48550/arXiv.1406.2661

    2014 doi arXiv

  19. [27]

    A., 2017, fastRESOLVE : Fast Bayesian Imaging for Aperture Synthesis in Radio Astronomy ( @eprint arXiv 1605.04317 ), @doi 10.48550/arXiv.1605.04317

    Greiner M., Vacca V., Junklewitz H., En lin T. A., 2017, fastRESOLVE : Fast Bayesian Imaging for Aperture Synthesis in Radio Astronomy ( @eprint arXiv 1605.04317 ), @doi 10.48550/arXiv.1605.04317

    2017 doi arXiv

  20. [28]

    C., 2017, in Advances in Neural Information Processing Systems

    Gulrajani I., Ahmed F., Arjovsky M., Dumoulin V., Courville A. C., 2017, in Advances in Neural Information Processing Systems . Curran Associates, Inc

    2017

  21. [29]

    Hermans J., Delaunoy A., Rozet F., Wehenkel A., Begy V., Louppe G., 2022, Transactions on Machine Learning Research

    2022

  22. [30]

    A., 1974, Astronomy and Astrophysics, https://ui.adsabs.harvard.edu/abs/1974A&AS...15..417H 15, 55

    H \"o gbom J. A., 1974, Astronomy and Astrophysics, https://ui.adsabs.harvard.edu/abs/1974A&AS...15..417H 15, 55

    1974

  23. [31]

    A., 2017, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Isola P., Zhu J.-Y., Zhou T., Efros A. A., 2017, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . pp 1125--1134

    2017

  24. [32]

    R., Selig M., En lin T

    Junklewitz H., Bell M. R., Selig M., En lin T. A., 2016, @doi [Astronomy & Astrophysics] 10.1051/0004-6361/201323094 , 586, A76

    2016 doi

  25. [33]

    A., 2020, Metric Gaussian Variational Inference ( @eprint arXiv 1901.11033 ), @doi 10.48550/arXiv.1901.11033

    Knollm \"u ller J., En lin T. A., 2020, Metric Gaussian Variational Inference ( @eprint arXiv 1901.11033 ), @doi 10.48550/arXiv.1901.11033

    2020 doi arXiv

  26. [34]

    PMLR, pp 19256--19273

    Lemos P., Coogan A., Hezaveh Y., Perreault-Levasseur L., 2023, in Proceedings of the 40th International Conference on Machine Learning . PMLR, pp 19256--19273

    2023

  27. [35]

    I., Mars M., Price M

    Liaudat T. I., Mars M., Price M. A., Pereyra M., Betcke M. M., McEwen J. D., 2024, RAS Techniques and Instruments, 3, 505

    2024

  28. [36]

    M., McEwen J

    Mars M., Betcke M. M., McEwen J. D., 2023, @doi [RAS Techniques and Instruments] 10.1093/rasti/rzad054 , 2, 760

    2023 doi

  29. [37]

    M., McEwen J

    Mars M., Betcke M. M., McEwen J. D., 2025, @doi [RAS Techniques and Instruments] 10.1093/rasti/rzaf025 , 4, rzaf025

    2025 doi

  30. [38]

    Nelson D., et al., 2019, @doi [Computational Astrophysics and Cosmology] 10.1186/s40668-019-0028-x , 6, 2

    2019 doi

  31. [39]

    R., et al., 2014, @doi [Monthly Notices of the Royal Astronomical Society] 10.1093/mnras/stu1368 , 444, 606

    Offringa A. R., et al., 2014, @doi [Monthly Notices of the Royal Astronomical Society] 10.1093/mnras/stu1368 , 444, 606

    2014 doi

  32. [40]

    D., d'Avezac M., Carrillo R

    Pratley L., McEwen J. D., d'Avezac M., Carrillo R. E., Onose A., Wiaux Y., 2018, @doi [Monthly Notices of the Royal Astronomical Society] 10.1093/mnras/stx2237 , 473, 1038

    2018 doi

  33. [41]

    Pratley L., McEwen J. D., d'Avezac M., Cai X., Perez-Suarez D., Christidi I., Guichard R., 2019, Distributed and Parallel Sparse Convex Optimization for Radio Interferometry with PURIFY ( @eprint arXiv 1903.04502 ), @doi 10.48550/arXiv.1903.04502

    2019 doi arXiv

  34. [42]

    Remy B., Lanusse F., Jeffrey N., Liu J., Starck J.-L., Osato K., Schrabback T., 2023, @doi [Astronomy & Astrophysics] 10.1051/0004-6361/202243054 , 672, A51

    2023 doi

  35. [43]

    M., Frangi A

    Ronneberger O., Fischer P., Brox T., 2015, in Navab N., Hornegger J., Wells W. M., Frangi A. F., eds, , Vol. 9351, Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2015. Springer International Publishing, Cham, pp 234--241, @doi 10.1007/978-3-319-24574-4_28

    2015 doi

  36. [44]

    M., 2011, @doi [Astronomy & Astrophysics] 10.1051/0004-6361/201116434 , 527, A107

    Smirnov O. M., 2011, @doi [Astronomy & Astrophysics] 10.1051/0004-6361/201116434 , 527, A107

    2011 doi

  37. [45]

    Curran Associates, Inc

    Sohn K., Lee H., Yan X., 2015, in Advances in Neural Information Processing Systems . Curran Associates, Inc

    2015

  38. [46]

    M., Fenech D

    Stewart I. M., Fenech D. M., Muxlow T. W. B., 2011, @doi [Astronomy & Astrophysics] 10.1051/0004-6361/201016010 , 535, A81

    2011 doi

  39. [47]

    L., 2021, @doi [Proceedings of the AAAI Conference on Artificial Intelligence] 10.1609/aaai.v35i3.16366 , 35, 2628

    Sun H., Bouman K. L., 2021, @doi [Proceedings of the AAAI Conference on Artificial Intelligence] 10.1609/aaai.v35i3.16366 , 35, 2628

    2021 doi

  40. [48]

    M., et al., 2014, @doi [Monthly Notices of the Royal Astronomical Society] 10.1093/mnras/stt2244 , 438, 768

    Sutter P. M., et al., 2014, @doi [Monthly Notices of the Royal Astronomical Society] 10.1093/mnras/stt2244 , 438, 768

    2014 doi

  41. [49]

    C., et al., 2022, @doi [Publications of the Astronomical Society of the Pacific] 10.1088/1538-3873/ac9642 , 134, 114501

    Team T. C., et al., 2022, @doi [Publications of the Astronomical Society of the Pacific] 10.1088/1538-3873/ac9642 , 134, 114501

    2022 doi

  42. [50]

    Terris M., Abdulaziz A., Dabbech A., Jiang M., Repetti A., Pesquet J.-C., Wiaux Y., 2019, Signal Processing with Adaptive Sparse Structured Representations (SPARS) workshop, p. 3

    2019

  43. [51]

    Terris M., Dabbech A., Tang C., Wiaux Y., 2022, @doi [Monthly Notices of the Royal Astronomical Society] 10.1093/mnras/stac2672 , 518, 604

    2022 doi

  44. [52]

    Terris M., Tang C., Jackson A., Wiaux Y., 2023, Plug-and-Play Imaging with Model Uncertainty Quantification in Radio Astronomy ( @eprint arXiv 2312.07137 ), @doi 10.48550/arXiv.2312.07137

    2023 doi arXiv

  45. [53]

    R., Moran J

    Thompson A. R., Moran J. M., Swenson G. W., 2017, Interferometry and Synthesis in Radio Astronomy . Astronomy and Astrophysics Library , Springer International Publishing, Cham, @doi 10.1007/978-3-319-44431-4

    2017 doi

  46. [54]

    Tonolini F., Radford J., Turpin A., Faccio D., Murray-Smith R., 2020, Journal of Machine Learning Research, 21, 1

    2020

  47. [55]

    IOS Press, pp 2499--2506, @doi 10.3233/FAIA230554

    Wang R., Chen Z., Luo Q., Wang F., 2023, in , ECAI 2023. IOS Press, pp 2499--2506, @doi 10.3233/FAIA230554

    2023 doi

  48. [56]

    D., 2024a, Using Conditional GANs for Convergence Map Reconstruction with Uncertainties ( @eprint arXiv 2406.15424 ), @doi 10.48550/arXiv.2406.15424

    Whitney J., Liaudat T., Price M., Mars M., McEwen J. D., 2024a, Using Conditional GANs for Convergence Map Reconstruction with Uncertainties ( @eprint arXiv 2406.15424 ), @doi 10.48550/arXiv.2406.15424

    doi arXiv

  49. [57]

    J., Liaudat T

    Whitney J. J., Liaudat T. I., Price M. A., Mars M., McEwen J. D., 2024b, Generative Modelling for Mass-Mapping with Fast Uncertainty Quantification ( @eprint arXiv 2410.24197 ), @doi 10.48550/arXiv.2410.24197

    doi

  50. [58]

    Winkler C., Worrall D., Hoogeboom E., Welling M., 2023, Learning Likelihoods with Conditional Normalizing Flows ( @eprint arXiv 1912.00042 ), @doi 10.48550/arXiv.1912.00042

    2023 doi arXiv

  51. [59]

    Zhao H., Li H., Maurer-Stroh S., Cheng L., 2018, @doi [Medical Image Analysis] 10.1016/j.media.2018.07.001 , 49, 14

    2018 doi

  52. [60]

    I.-C., Xu Y., 2021, in International Conference on Learning Representations

    Zhao S., Cui J., Sheng Y., Dong Y., Liang X., Chang E. I.-C., Xu Y., 2021, in International Conference on Learning Representations

    2021

Pith tools

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