Pith. sign in

REVIEW 4 major objections 5 minor 63 references

Probabilistic cosmological inference on HI tomographic data

T0 review · 4 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 A two-stage deep network — 3D convolutional encoder plus Masked Autoregressive Flow — recovers ($\Omega_{\rm m}$, $\sigma_8$) from $128^3$ 21-cm boxes with $R^2 \geq 0.91$ and calibrated uncertainties.

desk verdict A credible proof-of-concept for SBI on 3D HI tomographic data, but the calibration claim rests on a metric that is not a coverage test and the abstract overstates OOD robustness. read the letter →

arxiv 2507.21682 v1 pith:JS7TJSSH submitted 2025-07-29 astro-ph.IM astro-ph.CO

Sambatra Andrianomena This is my paper
classification astro-ph.IMastro-ph.CO
keywords 21-cmintensitymappingHItomographylikelihood-freeinferencenormalizingflowsposteriorestimationcosmologicalparametersneuralcompressionuncertaintycalibration
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 shows that cosmological parameters can be extracted from three-dimensional 21-cm neutral-hydrogen maps at intermediate redshift without an analytic likelihood. A deep 3D convolutional encoder is first trained to compress each $128^3$ tomographic box into a 1024-dimensional latent vector by regressing the matter density $\Omega_{\rm m}$ and the fluctuation amplitude $\sigma_8$. A Masked Autoregressive Flow, trained in an Automatic Posterior Transformation scheme on those latent vectors, then estimates the full posterior over ($\Omega_{\rm m}$, $\sigma_8$) for a single box. On held-out test sets the posterior medians achieve $R^2 \geq 0.91$ on both parameters, most ground-truth values lie inside the $1\sigma$ credible interval, and the epistemic uncertainty is calibrated at the percent level. The same featurizer also performs reasonably on out-of-distribution data — a different simulation or a redshift offset up to $\Delta z \approx 0.5$ — which suggests the learned codes capture transferable information about matter clustering.

What carries the argument

The load-bearing object is the encoder–posterior pair. The encoder is a 3D convolutional network with four stages of 3D convolution, batch normalization, ReLU, and max pooling, ending in adaptive average pooling, which maps each $128^3$ neutral-hydrogen box to a 1024-dimensional latent vector; it is pre-trained jointly with a two-layer regressor under the loss of Eq. (1), which balances the log-mean-squared prediction error against the log-squared error of the predicted per-parameter variance. The posterior estimator is a Masked Autoregressive Flow, a normalizing flow that factors a density into Gaussian conditionals and is invertible with an analytic Jacobian, used inside the Automatic Posterior Transformation framework, which trains the flow to map parameters and data (here, latent codes) to an approximation of $p(\Omega_{\rm m}, \sigma_8 \mid \text{latent code})$. At test time, 10,000 samples are drawn from the fitted posterior; the median is the point prediction and the 68% and 95% percentiles give the epistemic credible intervals. The separability of the codes under UMAP is presented as evidence that the regression pre-training produces meaningful structure.

What would settle it

On the held-out boxes, train the same MAF posterior directly on the raw $128^3$ cubes or on a summary known to be sufficient (e.g., the full power spectrum plus bispectrum) and compare its credible intervals with those obtained from the latent codes, or run an expected-coverage test across the full $(\Omega_{\rm m}, \sigma_8)$ prior: if coverage deviates strongly from the nominal 68% in any region, or if the latent-code intervals are systematically wider than the field-level ones, the encoder's information sufficiency is falsified.

Watch

Extended reading notes

Core claim

In the paper's own terms, the claim is that the full 3D neutral-hydrogen field at $z \approx 1$ carries enough information to infer $\Omega_{\rm m}$ and $\sigma_8$ with a coefficient of determination $R^2 \geq 0.91$ for both parameters, using only the 1024-dimensional codes produced by a regression-trained encoder. For an in-distribution test set, most ground-truth parameter values fall inside the 68% credible interval of the MAF-based posterior, and the Uncertainty Calibration Error ranges from about 2% to 13% depending on parameter and simulation, so the posterior uncertainty is described as reasonably calibrated. The paper also claims that the latent codes are separable in feature space, with points of similar parameter values clustering together, and that the featurizer generalizes: when the encoder built from one simulation compresses boxes from the other simulation, the posterior estimator still achieves $R^2 \geq 0.84$, and when it compresses boxes from the same simulation at $z = 0.5$ (the encoder being built at $z = 1$), $R^2 \geq 0.83$ holds. These results are put forward as a proof of concept for analyzing data from future 21-cm tomographic surveys.

Load-bearing premise

The load-bearing premise is that the 1024-dimensional latent codes produced by an encoder trained only to predict $\Omega_{\rm m}$ and $\sigma_8$ retain all the information the density estimator needs for a calibrated posterior; if the regression objective discards features relevant to the posterior's shape or tails, the reported calibration is not guaranteed.

Editorial extensions

If this is right

  • If correct, the pipeline supplies a likelihood-free route to cosmological constraints from 21-cm tomographic data, usable wherever a simulation-based training set exists.
  • The cross-simulation result ($R^2 \geq 0.84$) implies that a featurizer trained on one survey's simulations may compress another survey's data with only modest loss, as long as the voxel distributions are not too far apart.
  • The percent-level calibration means the $1\sigma$ intervals can be read as physical error bars on a single inference without an additional recalibration step.
  • The linear test's $R^2 \geq 0.80$ shows the codes are nearly linearly informative, so the encoder's compression does not destroy the cosmological content.
  • The worst transfer case ($R^2 \approx 0.6$ for $\sigma_8$ from $z = 0$ boxes encoded by a $z = 1$ featurizer) sets a practical redshift horizon for reusing one encoder.

Reading between the lines

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

  • The paper does not test whether the 1024-dimensional codes are sufficient summaries; a direct comparison with a field-level posterior or with a summary known to be sufficient would indicate whether the regression pre-training discards posterior-relevant information.
  • The cross-simulation robustness is demonstrated on two simulations of similar resolution and physics, so the transferability to real radio maps with noise and foregrounds remains open; the author explicitly flags domain adaptation as the next step.
  • The degradation at $z = 0$ suggests the useful range of a single featurizer is roughly $\Delta z \lesssim 0.5$; training on a multi-redshift set would be a natural way to extend the method, and the paper leaves that unexplored.
  • A survey-level application would combine the posteriors from many boxes; the paper does not address how to handle correlations between boxes when they are assembled into a single posterior.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a two-stage pipeline for cosmological inference from 21-cm HI tomographic data cubes. In the first stage, a 3D convolutional encoder (Table 1) is trained with a supervised regression loss (Eq. 1) to predict Ωm and σ8 from 128^3 CAMELS HI boxes at z=1, yielding 1024-dimensional latent codes. In the second stage, a Masked Autoregressive Flow within the Automatic Posterior Transformation framework is trained on these latent codes to approximate the posterior p(Ωm, σ8 | data). The authors report R² ≥ 0.91 for same-distribution APT results (SIMBA→SIMBA, TNG→TNG), compare against GP/RF/linear baselines, and test robustness when the encoder is applied to out-of-distribution data (cross-simulation, and z=0/0.5 from the same simulation). The paper claims that the posterior uncertainties are reasonably calibrated, based on the Uncertainty Calibration Error (UCE) of Eq. (6), and that most test truths fall within the 1σ credible intervals.

Significance. If the claims hold, this is a useful proof-of-concept for field-level, likelihood-free inference from 3D 21-cm tomographic data, with potential applications to upcoming HI surveys. The use of two CAMELS simulations, held-out test sets with clean encoder/APT separation, a UMAP-based check of latent separability, and comparison against several regression baselines are concrete strengths. The main gaps are statistical: the calibration claim is based on UCE, which is not a coverage test, and the abstract's out-of-distribution performance claim (R² ≥ 0.80 in general) is contradicted by the paper's own z=0 SIMBA σ8 result. These issues are fixable within the manuscript's scope.

major comments (4)
  1. [§4.2, §4.3, Eq. (6)] The central calibration claim is not established. UCE, as defined in Eq. (6), compares binned mean squared error ξ² to binned mean epistemic variance ζ². This is a variance-calibration diagnostic, not a coverage metric: a predictor with a constant bias and variance equal to its mean squared error can have UCE near zero while its 68% intervals contain almost none of the true values. The abstract and §4.2 state that 'most of the ground truth ... fall within 1σ uncertainty', but this is supported only by visual inspection of Figures 3 and 4; no empirical coverage fraction, prediction-interval coverage probability, or joint 2D coverage statistic is reported. Because calibrated posterior inference is a headline contribution, please add quantitative coverage tests, such as the fraction of test truths within the 68% and 95% marginal credible intervals and a 2D joint coverage metric, and report them for all four main setups.
  2. [Abstract, §4.6, Figure 9] The abstract's claim that out-of-distribution inference achieves 'R² ≥ 0.80 in general' is contradicted by the paper's own results: §4.6 states that 'in the worst case scenario at z=0, APT still boasts an R²~0.6 on σ8' for SIMBA, and the corresponding point is visible in Figure 9 (top-left panel). The claim should be qualified, for example by restricting it to z ≥ 0.5 or by explicitly noting the z=0 SIMBA σ8 exception. As written, the abstract overstates the demonstrated robustness of the method.
  3. [§3.1, Eq. (1), §3.2] The sufficiency of the latent codes for posterior estimation is assumed, not tested. The encoder is trained with a regression loss (Eq. 1) that explicitly targets Ωm and σ8, so the latent codes are aligned with point prediction of these parameters. This is not full circularity because the APT is evaluated on held-out boxes the encoder never saw, but it does mean that the downstream posterior is conditioned on features already selected by the regression objective. The paper does not test whether these codes retain information needed for the tails of the posterior or for the reported calibration, for example by comparing against an unsupervised compression or by using an information-preserving summary. Please either add such a comparison or explicitly discuss this limitation when claiming that the latent codes are 'meaningful representations' in a general sense.
  4. [§4.6, Figure 9] The redshift robustness experiments are missing key experimental details. The text does not state how many z=0 and z=0.5 boxes are used, whether these are the same simulation boxes as the z=1 set, or how the train/validation/test split is performed for the APT, GP, RF, and LIN models at these redshifts. Since the z-dependent R² values are a central part of the robustness claim, these details are needed for reproducibility and for interpreting the comparison across redshifts.
minor comments (5)
  1. [Abstract and §4.2] The phrase 'most of the ground truth ... fall within 1σ' is unquantified; please report the actual fraction of test points inside the 68% and 95% intervals rather than relying on visual inspection.
  2. [§4.3, Eq. (9)] A modified UCE is defined in Eq. (9), but only the UCE of Eq. (6) is reported in Figure 6. Please either report the modified values or clarify why they are omitted.
  3. [Tables and Figures] R² values are point estimates computed on 150 test examples; no uncertainty intervals are provided. Given that several method comparisons differ by only ~0.01 (Table 2), bootstrap or other interval estimates would help assess whether the differences are meaningful.
  4. [§5] There are typos in the conclusions, including 'theATP model' and 'esimator'; please proofread the final section and the abstract.
  5. [General] The paper does not include a data or code availability statement. Since the CAMELS data are public and the SBI package is named, a brief statement (or link) would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the encoder is supervised, but the APT posterior is evaluated on held-out boxes and benchmarked against independent regressors, so the central claims do not reduce to the training objective by construction.

full rationale

The paper's two-stage pipeline is not circular. The encoder is trained with the regression loss of Eq. (1) to predict Omega_m and sigma_8, so the latent codes are target-supervised; however, the encoder is trained only on the training/validation boxes, and the APT density estimator is then trained on latent codes from those same splits and tested on latent codes of held-out boxes (Section 4.2, Figures 3-4, Table 2). The APT does not see the test labels when producing posteriors, and its R^2 values are empirical test-set outcomes rather than quantities forced by the encoder objective. The paper further benchmarks APT against Gaussian Process, Random Forest, and a linear probe trained on the same latent codes (Table 2), showing that APT outperforms these independent regressors on the shared representation. Self-citations to Andrianomena & Hassan (2023b, 2025) are used for architecture inspiration and qualitative comparison, not as load-bearing proof of the main result. The calibration claim rests on UCE, which is a consistency metric rather than a full coverage test, but that is a statistical validity concern, not a circularity of the derivation chain.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claim rests on simulation representativeness, latent-code sufficiency, and the prior choice; none of these is independently verified in the paper.

free parameters (8)
  • Encoder learning rate = 0.0015
    Chosen by hand for Adam; affects convergence and latent code quality, not physically fitted.
  • Encoder architecture (channels per stage) = [16, 32, 64, 128]
    Doubling channels per stage, adapted from Andrianomena & Hassan (2025); capacity choice.
  • Latent vector length = 1024
    Determined by adaptive average pooling to 2x2x2 then flattening; not optimized.
  • APT/MAF learning rate = 0.0005
    Hand-set for the density estimator training.
  • Number of MADE layers in MAF = 4
    Chosen from prior MAF usage; not tuned.
  • UCE number of bins = 10
    Set following Andrianomena & Hassan (2023b); affects calibration metric.
  • Number of posterior samples = 10000
    Used to compute median and credible intervals; sufficiently large for stable percentiles.
  • Data augmentation noise level = not reported
    Gaussian noise is injected during augmentation, but the noise level is not specified; affects robustness results.
assumptions (5)
  • standard math Bayes' theorem (Eq. 2) and normalizing flow change-of-variable formula (Eq. 4)
    Used without proof; standard probability results.
  • domain assumption CAMELS boxes at z=1 are representative of upcoming 21-cm tomographic surveys
    The paper argues from CHIME detection and forecasts but acknowledges clean simulated data; no noise or foregrounds included.
  • ad hoc to paper Regression loss (Eq. 1) yields latent codes sufficient for posterior estimation
    Assumes the 1024-dim codes retain all posterior-relevant information; not tested against field-level inference.
  • domain assumption Astrophysical parameters can be ignored when mapping HI maps to cosmology
    Only cosmology labels are used because SIMBA and TNG use different ionizing background prescriptions; the impact of astrophysical feedback on the mapping is unquantified.
  • domain assumption Uniform CAMELS prior is the correct proposal
    Ωm ~ U[0.1,0.5], σ8 ~ U[0.6,1.0] are taken from CAMELS; posterior is prior-dependent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Probabilistic cosmological inference on HI tomographic data." pith.science (2026). https://pith.science/paper/JS7TJSSH

@misc{pith2026250721682,
  author       = {Pith},
  title        = {Pith review of: Probabilistic cosmological inference on HI tomographic data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JS7TJSSH}},
  note         = {Machine review of arXiv:2507.21682}
}
abstract

We explore the possibility of retrieving cosmological information from 21-cm tomographic data at intermediate redshift. The first step in our approach consists of training an encoder, composed of several three dimensional convolutional layers, to cast the neutral hydrogen 3D data into a lower dimension latent space. Once pre-trained, the featurizer is able to generate 3D grid representations which, in turn, will be mapped onto cosmology ($\Omega_{\rm m}$, $\sigma_{8}$) via likelihood-free inference. For the latter, which is framed as a density estimation problem, we consider a Bayesian approximation method which exploits the capacity of Masked Autoregressive Flow to estimate the posterior. It is found that the representations learned by the deep encoder are separable in latent space. Results show that the neural density estimator, trained on the latent codes, is able to constrain cosmology with a precision of $R^2 \ge 0.91$ on all parameters and that most of the ground truth of the instances in the test set fall within $1\sigma$ uncertainty. It is established that the posterior uncertainty from the density estimator is reasonably calibrated. We also investigate the robustness of the feature extractor by using it to compress out-of-distribution dataset, that is either from a different simulation or from the same simulation but at different redshift. We find that, while trained on the latent codes corresponding to different types of out-of-distribution dataset, the probabilistic model is still reasonably capable of constraining cosmology, with $R^2 \ge 0.80$ in general. This highlights both the predictive power of the density estimator considered in this work and the meaningfulness of the latent codes retrieved by the encoder. We believe that the approach prescribed in this proof of concept will be of great use when analyzing 21-cm data from various surveys in the near future.

Figures

Figures reproduced from arXiv: 2507.21682 by the authors.

Figure 1
Figure 1. Probability distribution of the voxel values of the HI tomographic data from each simulation. Blue and red solid lines indicate TNG and SIMBA voxel distributions respectively. The voxel values in all 3D grids are normal￾ized to [0,1]. It is evident that SIMBA and TNG 3D grids were generated by two different distributions. 2 DATA As the objective of our study is a robust retrieval of cosmological information from HI … view at source ↗
Figure 2
Figure 2. The latent codes are cast into 2D space using UMAP method. The color coding of each dot in the top panels is based on the value of 𝜎8, whereas that of a dot in the bottom panels is related to Ωm. All panels in each column correspond to a dataset, e.g. SIMBA. When trained on SIMBA dataset, we obtain 𝑅 2 = 0.905, 0.893 for Ωm and 𝜎8 respectively on the test set. It is clear that the model is able to extract both param… view at source ↗
Figure 3
Figure 3. Left column: The red dots, dark blue shaded area, light blue shaded area and black solid line indicate the ground truth, 68% credible interval (1𝜎), 95% credible interval (2𝜎) and the median of predictions sampled from the trained APT. The entire HI 3D dataset from a simulation is first encoded using 𝐸 SIMBA, and split into train/valid/test to train and test APT. Right column: Each blue dot represents a prediction f… view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Distribution of the posterior uncertainty related to each parameter in each setup. Red and blue boxes correspond to SIMBA and TNG datasets respectively. its corresponding prediction is, we plot the latter as a function of the former in the right columns of [PITH_FULL_…
Figure 8
Figure 8. Figure 8: Similar to Figures 3, 4 and 7 but APT method is trained on SIMBA features encoded by 𝐸 TNG. coefficient of determination 𝑅 2 . The results suggest that APT method achieves cosmological constraints that are tighter than those obtained from the deep regressor in §4.1, i.…
Figure 7
Figure 7. Figure 7: Similar to Figures 3 and 4 but APT method is trained on TNG features encoded by 𝐸 SIMBA. 0 25 50 75 100 125 150 index 0.1 0.2 0.3 0.4 0.5 Ω m 0.1 0.2 0.3 0.4 0.5 Truth 0.1 0.2 0.3 0.4 0.5 Prediction Ωm RMSE = 0.047  = 13.5% R2 = 0.840 0.6 0.7 0.8 0.9 1.0 σ8 TNG → SIMB…
Figure 9
Figure 9. Figure 9: 𝑅 2 obtained on each cosmological parameter for each algorithm as a function of redshift. Blue dot, red square, green triangle and black cross denote 𝑅 2 value achieved by APT, GP, RF and LIN methods respectively. For all results corresponding to SIMBA and TNG datasets…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 54 canonical work pages

  1. [1]

    A., et al., 2014, Astronomy & Astrophysics, 571, A1

    Ade P. A., et al., 2014, Astronomy & Astrophysics, 571, A1

    work page 2014

  2. [2]

    Aghanim N., et al., 2020, Astronomy & Astrophysics, 641, A1

    work page 2020

  3. [3]

    Akeret J., Refregier A., Amara A., Seehars S., Hasner C., 2015, Journal of Cosmology and Astroparticle Physics, 2015, 043

    work page 2015

  4. [4]

    Alsing J., Charnock T., Feeney S., Wandelt B., 2019, Monthly Notices of the Royal Astronomical Society, 488, 4440

    work page 2019

  5. [5]

    Amiri M., et al., 2022, The Astrophysical Journal Supplement Series, 261, 29

    work page 2022

  6. [6]

    Amiri M., et al., 2023, The Astrophysical Journal, 947, 16

    work page 2023

  7. [7]

    Andrianomena S., Hassan S., 2023a, arXiv:2311.00799

    work page Pith review arXiv

  8. [8]

    Andrianomena S., Hassan S., 2023b, Journal of Cosmology and Astroparticle Physics, 2023, 051

    work page 2023

Show all 63 references
  1. [9]

    Andrianomena S., Hassan S., 2025, Astrophysics and Space Science, 370, 14

    2025

  2. [10]

    Andrianomena S., Hassan S., Villaescusa-Navarro F., 2025, Astrophysics and Space Science, 370, 1

    2025

  3. [11]

    Ansari R., et al., 2012, Astronomy & Astrophysics, 540, A129

    2012

  4. [12]

    pp 738--757

    Bandura K., et al., 2014, in Ground-based and Airborne Telescopes V. pp 738--757

    2014

  5. [13]

    A., Davies R

    Battye R. A., Davies R. D., Weller J., 2004, Monthly Notices of the Royal Astronomical Society, 355, 1339

    2004

  6. [14]

    Battye R., Browne I., Dickinson C., Heron G., Maffei B., Pourtsidou A., 2013, Monthly Notices of the Royal Astronomical Society, 434, 1239

    2013

  7. [15]

    K., 2005, Monthly Notices of the Royal Astronomical Society, 358, 968

    Bharadwaj S., Pandey S. K., 2005, Monthly Notices of the Royal Astronomical Society, 358, 968

    2005

  8. [16]

    Binnie T., Zhao X., Pritchard J., Mao Y., 2025, arXiv:2502.08152

    2025 arXiv

  9. [17]

    M., 1994, Aston University

    Bishop C. M., 1994, Aston University

    1994

  10. [18]

    G., 2010, Journal of the American Statistical Association, 105, 1178

    Blum M. G., 2010, Journal of the American Statistical Association, 105, 1178

    2010

  11. [19]

    G., Patel P., Santos M

    Bull P., Ferreira P. G., Patel P., Santos M. G., 2015, The Astrophysical Journal, 803, 21

    2015

  12. [20]

    Child R., 2020, arXiv preprint arXiv:2011.10650

    2020 arXiv

  13. [21]

    Courty N., Flamary R., Tuia D., Rakotomamonjy A., 2016, IEEE transactions on pattern analysis and machine intelligence, 39, 1853

    2016

  14. [22]

    Crichton D., et al., 2022, Journal of Astronomical Telescopes, Instruments, and Systems, 8, 011019

    2022

  15. [23]

    H., Appleby S., 2019, Monthly Notices of the Royal Astronomical Society, 486, 2827

    Dav \'e R., Angl \'e s-Alc \'a zar D., Narayanan D., Li Q., Rafieferantsoa M. H., Appleby S., 2019, Monthly Notices of the Royal Astronomical Society, 486, 2827

    2019

  16. [24]

    S., et al., 2016, The Astronomical Journal, 151, 44

    Dawson K. S., et al., 2016, The Astronomical Journal, 151, 44

    2016

  17. [25]

    Eickenberg M., Exarchakis G., Hirn M., Mallat S., 2017, Advances in Neural Information Processing Systems, 30

    2017

  18. [26]

    Eickenberg M., Exarchakis G., Hirn M., Mallat S., Thiry L., 2018, The Journal of chemical physics, 148

    2018

  19. [27]

    pp 1050--1059

    Gal Y., Ghahramani Z., 2016, in international conference on machine learning. pp 1050--1059

    2016

  20. [28]

    pp 881--889

    Germain M., Gregor K., Murray I., Larochelle H., 2015, in International conference on machine learning. pp 881--889

    2015

  21. [29]

    Gillet N., Mesinger A., Greig B., Liu A., Ucci G., 2019, Monthly Notices of the Royal Astronomical Society, 484, 282

    2019

  22. [30]

    pp 2404--2414

    Greenberg D., Nonnenmacher M., Macke J., 2019, in International conference on machine learning. pp 2404--2414

    2019

  23. [31]

    Greig B., Mesinger A., 2015, Monthly Notices of the Royal Astronomical Society, 449, 4246

    2015

  24. [32]

    A., 2023, Monthly Notices of the Royal Astronomical Society, 519, 5288

    Greig B., Ting Y.-S., Kaurov A. A., 2023, Monthly Notices of the Royal Astronomical Society, 519, 5288

    2023

  25. [33]

    Hassan S., Andrianomena S., Doughty C., 2020, Monthly Notices of the Royal Astronomical Society, 494, 5761

    2020

  26. [34]

    K., 1995, in Proceedings of 3rd international conference on document analysis and recognition

    Ho T. K., 1995, in Proceedings of 3rd international conference on document analysis and recognition. pp 278--282

    1995

  27. [35]

    Jennings E., Madigan M., 2017, Astronomy and computing, 19, 16

    2017

  28. [36]

    L., 2009, Proceedings of the IEEE, 97, 1522

    Jonas J. L., 2009, Proceedings of the IEEE, 97, 1522

    2009

  29. [37]

    F., Kahrs L

    Laves M.-H., Ihler S., Fast J. F., Kahrs L. A., Ortmaier T., 2020, in Medical imaging with deep learning. pp 393--412

    2020

  30. [38]

    J., Bassetto G., \"O cal K., Nonnenmacher M., Macke J

    Lueckmann J.-M., Goncalves P. J., Bassetto G., \"O cal K., Nonnenmacher M., Macke J. H., 2017, Advances in neural information processing systems, 30

    2017

  31. [39]

    R., Mondal R., Watkinson C

    Majumdar S., Pritchard J. R., Mondal R., Watkinson C. A., Bharadwaj S., Mellema G., 2018, Monthly Notices of the Royal Astronomical Society, 476, 4007

    2018

  32. [40]

    pp 410--420

    Martini P., et al., 2018, in Ground-based and Airborne Instrumentation for Astronomy VII. pp 410--420

    2018

  33. [41]

    McInnes L., Healy J., Melville J., 2018, arXiv:1802.03426

    2018 arXiv

  34. [42]

    Mellema G., et al., 2013, Experimental Astronomy, 36, 235

    2013

  35. [43]

    K., Mesinger A., Majumdar S., 2015, arXiv preprint arXiv:1501.04203

    Mellema G., Koopmans L., Shukla H., Datta K. K., Mesinger A., Majumdar S., 2015, arXiv preprint arXiv:1501.04203

    2015 arXiv

  36. [44]

    Nelson D., et al., 2019, Computational Astrophysics and Cosmology, 6, 1

    2019

  37. [45]

    P., et al., 2024, Monthly Notices of the Royal Astronomical Society, 528, 2078

    Novaes C. P., et al., 2024, Monthly Notices of the Royal Astronomical Society, 528, 2078

    2024

  38. [46]

    Papamakarios G., Murray I., 2016, Advances in neural information processing systems, 29

    2016

  39. [47]

    Papamakarios G., Pavlakou T., Murray I., 2017, Advances in neural information processing systems, 30

    2017

  40. [48]

    pp 837--848

    Papamakarios G., Sterratt D., Murray I., 2019, in The 22nd international conference on artificial intelligence and statistics. pp 837--848

    2019

  41. [49]

    Pourtsidou A., Bacon D., Crittenden R., 2017, Monthly Notices of the Royal Astronomical Society, 470, 4251

    2017

  42. [50]

    Adaptive Computation and Machine Learning, MIT Press, Cambridge, MA, USA

    Rasmussen C., Williams C., 2006, Gaussian Processes for Machine Learning. Adaptive Computation and Machine Learning, MIT Press, Cambridge, MA, USA

    2006

  43. [51]

    pp 1530--1538

    Rezende D., Mohamed S., 2015, in International conference on machine learning. pp 1530--1538

    2015

  44. [52]

    F., Debono I., 2017, Astronomy & Astrophysics, 597, A136

    Smoot G. F., Debono I., 2017, Astronomy & Astrophysics, 597, A136

    2017

  45. [53]

    Switzer E., et al., 2013, Monthly Notices of the Royal Astronomical Society: Letters, 434, L46

    2013

  46. [54]

    J., Greenberg D

    Tejero-Cantero A., Boelts J., Deistler M., Lueckmann J.-M., Durkan C., Gon c alves P. J., Greenberg D. S., Macke J. H., 2020a, arXiv:2007.09114

    2007 arXiv

  47. [55]

    J., Greenberg D

    Tejero-Cantero A., Boelts J., Deistler M., Lueckmann J.-M., Durkan C., Gonçalves P. J., Greenberg D. S., Macke J. H., 2020b, @doi [Journal of Open Source Software] 10.21105/joss.02505 , 5, 2505

    doi

  48. [56]

    pp 7167--7176

    Tzeng E., Hoffman J., Saenko K., Darrell T., 2017, in Proceedings of the IEEE conference on computer vision and pattern recognition. pp 7167--7176

    2017

  49. [57]

    Villaescusa-Navarro F., et al., 2021a, arXiv:2109.09747

    arXiv

  50. [58]

    Villaescusa-Navarro F., et al., 2021b, The Astrophysical Journal, 915, 71

  51. [59]

    Villaescusa-Navarro F., et al., 2022a, arXiv:2201.01300

    arXiv

  52. [60]

    Villaescusa-Navarro F., et al., 2022b, The Astrophysical Journal Supplement Series, 259, 61

  53. [61]

    Wagg J., Bourke T., Green J., Braun R., et al., 2014, SKA1 scientific use cases

    2014

  54. [62]

    D., 2022, The Astrophysical Journal, 926, 151

    Zhao X., Mao Y., Cheng C., Wandelt B. D., 2022, The Astrophysical Journal, 926, 151

    2022

  55. [63]

    D., 2024, The Astrophysical Journal, 973, 41

    Zhao X., Mao Y., Zuo S., Wandelt B. D., 2024, The Astrophysical Journal, 973, 41

    2024

Pith tools

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