Pith. sign in

REVIEW 5 major objections 6 minor 41 references

A Poisson Process AutoDecoder for X-ray Sources

T0 review · 5 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A Poisson-aware autodecoder that simultaneously reconstructs continuous X-ray light curves and learns fixed-length latent features directly from event files, achieving hardness regression around R²=0.9 and variability classification at…

desk verdict A promising Poisson-neural-field autodecoder for X-ray event files, but the validation is too thin to support the strong claims yet. read the letter →

arxiv 2502.01627 v2 pith:TLXXXFI6 submitted 2025-02-03 astro-ph.IM astro-ph.HEcs.LGstat.AP

Yanke Song , Victoria Ashley Villar , Juan Rafael Martinez-Galarza , Steven Dillmann This is my paper · ORCID
classification astro-ph.IMastro-ph.HEcs.LGstat.AP
keywords PoissonprocessneuralfieldautodecoderX-raylightcurvesunsupervisedrepresentationlearningChandraSourceCatalogeventfilesanomalydetection
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces the Poisson Process AutoDecoder (PPAD), an unsupervised method that takes raw X-ray photon event files as input and outputs both a continuous, resolution-free reconstruction of each source's light curve and a fixed-length latent vector summarizing its behavior. The authors' central claim is that a single shared neural-field decoder, trained with a Poisson likelihood and regularized by total variation, can simultaneously solve the ill-posed rate-reconstruction problem and learn a latent space rich enough for classification, regression, and anomaly detection. They show that latents learned purely from event files predict spectral hardness ratios with R² values around 0.87–0.94, identify variable sources with 92% accuracy, and enable an 8-label source classification at 60% accuracy. If correct, this means astronomers can automatically analyze millions of X-ray sources directly from photon lists, without manual feature engineering or pre-computed catalog properties.

What carries the argument

The central object is the autodecoder neural field: a shared ResNet that takes the concatenation of a per-source latent vector and a positional encoding of time γ(t) and outputs the instantaneous Poisson rate r(t) for each of three energy bands. Positional encoding with frequencies up to $2^{12}$ gives the network access to both slow and fast variability timescales; the total-variation penalty on both a uniform grid and the photon arrival times makes the otherwise ill-posed maximum-likelihood rate estimation well-posed; and the autodecoder formulation, where each event file has its own directly optimized latent vector, allows variable-length inputs to be embedded into a common fixed-dimension feature space without an encoder.

What would settle it

Generate simulated event files from known inhomogeneous Poisson rate functions—for example a sharp flare, a step change, and a sinusoid—at a range of total counts; run PPAD with the paper's specified hyperparameters, and measure whether the reconstructed rate recovers the injected flare amplitude and width. Smearing of the flare or retention of Poisson noise would directly contradict the claim that the TV-regularized neural field simultaneously reconstructs rates and produces meaningful latents.

Watch

Extended reading notes

Core claim

PPAD learns, in an end-to-end and unsupervised manner, a mapping from each event file to a latent code z that conditions a shared neural field r_phi, which outputs Poisson rate functions for three energy bands as a function of time. The loss combines the inhomogeneous Poisson negative log-likelihood, a total-variation penalty that enforces piecewise constancy, and an L2 penalty on the latent vectors. After training on roughly 110,000 Chandra event files, the reconstructed rate functions capture flares, dips, periodic behavior, and constant sources while smoothing shot noise, and the latent codes encode spectral hardness, variability timescales, signal-to-noise, and transient morphology. With a simple Random Forest on the learned latents, the paper reports near-0.9 R² for hardness prediction, 92% accuracy for variability classification, and 60% accuracy on an 8-class astrophysical source-type task, along with nearest-neighbor searches that recover hard flaring analogs of targeted anomalies.

Load-bearing premise

A single hand-tuned smoothness penalty (λTV=10) is assumed to separate genuine astrophysical variability from Poisson shot noise for every source population, and the paper does not include a simulation study that injects known rate functions to verify this separation.

Editorial extensions

If this is right

  • Astronomical pipelines could analyze raw event files from Chandralike, eROSITA-like, or XMM-like observations without first computing summary statistics, potentially replacing manual feature engineering for large catalogs.
  • The continuous neural-field representation means light curves can be sampled at arbitrarily fine time resolution after training, enabling the recovery of transient features that binning might blur.
  • The learned latent space appears to jointly encode spectral hardness, variability, and signal-to-noise, making it a candidate general-purpose embedding for unsupervised clustering and anomaly discovery in high-energy time-domain surveys.
  • The same Poisson-aware autodecoder recipe could transfer to other sparse photon or particle arrival datasets, such as gamma-ray bursts, neutrino events, or single-photon detectors, where Gaussian approximations are invalid.
  • Because the decoder is shared and only the latent is optimized at test time, newly observed sources can be embedded in seconds, supporting real-time triage of transients.

Reading between the lines

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

  • If the representation generalizes across missions, one could train PPAD jointly on multiple X-ray observatories and use the shared latent space to cross-calibrate source populations, a step the paper does not attempt.
  • The chosen total-variation weight λTV=10 is the load-bearing regularizer that decides what counts as signal versus noise; the paper notes it was only roughly tuned, suggesting that an adaptive or learned penalty could materially improve both reconstruction and embeddings for low-count sources.
  • The observed phase-sensitive latents from truncating event files at 8 hours imply that a phase-invariant variant could make periodic-source embeddings more stable, possibly improving classification of eclipsing and rotating sources.
  • One could test directly whether adding background regions as extra input channels improves rate reconstruction in crowded fields, since the paper explicitly notes background handling is not included.
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

5 major / 6 minor

Summary. The paper introduces PPAD (Poisson Process AutoDecoder), an autodecoder architecture in which a shared ResNet neural field maps a per-event-file latent vector plus a positional time encoding to three Poisson rate functions, one per energy band (soft, medium, hard). The loss (Eq. 6) combines the inhomogeneous Poisson negative log-likelihood with a total-variation penalty and a latent-norm penalty. The authors train on roughly 110,000 Chandra Source Catalog event files truncated to 8 hours, use the trained network to reconstruct rate functions, and then employ the learned 8-dimensional latents for hardness/variability regression, source-type classification, and nearest-neighbor anomaly searches. The central claim is that PPAD enables end-to-end, unsupervised representation learning directly from event files, producing latents whose downstream accuracies are comparable to methods that use pre-computed CSC properties and multi-wavelength features.

Significance. If the central claims are properly validated, this is a useful contribution to X-ray time-domain analysis: PPAD is a clean, label-free way to featurize Poisson-limited, variable-length event files while avoiding histogram binning artifacts, and the autodecoder formulation is well matched to the problem. The architecture and loss are specified in enough detail to be replicable, the experiments span reconstruction, regression, classification, and anomaly detection on a large real dataset, and the authors are transparent about several limitations. However, the current evidence is incomplete. Reconstruction quality is shown only qualitatively, the total-variation regularization that makes the optimization well-posed is not validated against known rate functions, and the downstream evaluations are partially in-sample or lack controlled baselines. The method's scientific value will hinge on synthetic recovery tests, source-level held-out evaluation, and direct comparisons against simple count-based baselines.

major comments (5)
  1. [Sec. 3.1, Eq. (4); Sec. 4.3] The total-variation penalty is the only mechanism preventing the Poisson maximum-likelihood problem from being ill-posed, yet the paper provides no simulation study showing that the hand-tuned lambda_TV=10 recovers known rate functions. Sources in the dataset range from a few to thousands of counts, and the relative strength of the TV penalty versus the log-likelihood changes with count rate, so a single fixed lambda_TV cannot be assumed appropriate across all regimes. Because the latent vectors are optimized against the same reconstructed rate functions, any bias in r_hat(t) propagates into the regression, classification, and anomaly-detection results. Please add synthetic experiments that draw Poisson realizations from known constant, flaring, periodic, and dipping rate functions at several count levels and report recovery error as a function of lambda_TV and count rate.
  2. [Sec. 4.2.1, Table 1; Sec. 2, Data] The hardness-ratio regression is partially in-sample: the CSC hardness ratios hard_hs, hard_ms, and hard_hm are defined from exactly the same soft/medium/hard energy bands that PPAD receives as its only energy inputs. The high R^2 values in Table 1 may therefore largely reflect per-band integrated counts rather than a learned representation of spectral shape. To support the claim that the latents encode meaningful spectral information, compare against a simple baseline that predicts hardness ratios directly from per-band total photon counts, and also report performance after removing the total-count information from the latents (for example, by regressing out per-band counts).
  3. [Sec. 4.2.1, Sec. 4.2.2, Sec. 3.4 (Eq. 8)] The 80/20 train-test split is applied to the latents after they have already been trained on the full dataset, so the test latents are not obtained by the frozen-network inference procedure described in Eq. (8). This means the reported regression and classification numbers do not demonstrate generalization to genuinely new event files. The evaluation should hold out complete sources (including their event files) before any latent or network training, and then infer test-time latents via Eq. (8); alternatively, the results should be explicitly described as transductive and the generalization claim softened.
  4. [Sec. 4.2.2, Table 1] The comparison with Yang et al. (2022) is not controlled and does not support the claim that PPAD latents provide accuracies comparable to methods using multi-wavelength features. The label sets differ, the source samples differ, and the 8-label PPAD F1 score of 0.24 is far below the reported Yang et al. F1 of 0.68 despite a 60% accuracy. A meaningful comparison requires training and testing both approaches on the same cross-matched sample with identical label definitions, or at least reporting per-class precision/recall so the reader can see where the low F1 arises.
  5. [Sec. 4.2.2, footnote 1; Sec. 2, Data] The event-file preprocessing splits long observations into multiple 8-hour event files, and the same astrophysical source can therefore appear in multiple rows of the dataset. If the 80/20 split is performed at the event-file level rather than the source level, the same source can appear in both the training and test sets, which can inflate classification accuracy via source-identity memorization. Please deduplicate by source or report results from a source-level split, and state explicitly whether any deduplication was performed in the labeled cross-match.
minor comments (6)
  1. [Eq. (4), Eq. (6)] The total-variation term on arrival times is undefined when an energy band has zero or one photon (division by n-1). Please specify how such event files or bands are handled.
  2. [Eq. (6)] In the display of L_TV, the denominator n-1 appears without the j,k subscripts; it should be n_{j,k}-1 to make clear that the sum is over arrivals in that energy band. The placement of the parentheses around the latent penalty is also easy to misread.
  3. [Sec. 4.1, Fig. 3] The reconstruction evaluation is purely visual. Since the paper claims PPAD smooths stochastic photon noise while preserving flares and dips, please provide at least one quantitative reconstruction metric, such as held-out log-likelihood or a comparison against Gregory-Loredo and a binned histogram baseline on a small labeled set.
  4. [References] The reference 'Webb et al. 2020' appears twice for two different papers (A&A 641, A136 and MNRAS 498, 3077); these need disambiguation, for example Webb et al. 2020a and 2020b.
  5. [Throughout] There are several typos and small errors: 'comparares' in Sec. 4.2.2, 'wavelegth' in the Introduction, 'hard_hs, hard_ms, and hard_ms' in Sec. 2 (the third should be hard_hm), and 'Lneg-loglikelihood' in Appendix A.2.
  6. [Sec. 4.2.3] The anomaly-detection experiment is entirely qualitative and based on visual inspection of the 15 nearest neighbors. Please report a quantitative criterion (for example, the fraction of neighbors that are known transients, compared with a random-neighbor baseline) so the reader can assess the strength of the embedding.

Circularity Check

1 steps flagged · score 4.0 of 10

Hardness-ratio regression is largely in-sample because PPAD's only energy input is the same three bands used to define the CSC hardness ratios; the central autodecoder method is otherwise non-circular.

  1. fitted input called prediction [Section 2 (data preprocessing and hardness definition) and Section 4.2.1 / Table 1 (hardness regression)]
    "Hardness ratios: A quantification of the distribution of photon energies between three energy bands: soft (0.5 keV-1.2 keV), medium (1.2 keV-2 keV), and hard (2 keV-7 keV). ... To confirm this, we use the learned latents to predict the hardness ratio and variability of each source. ... In short, we obtain ∼ 0.9 R2 values on hardness ratio prediction"

    PPAD's only energy input is the same three bands that define hardness: Section 2 bins energies into soft/medium/high, and Eq. (6) optimizes one rate function per band. For each band, the Poisson likelihood term -Σ log r_k(t_i) + ∫ r_k drives the integrated reconstructed rate toward the band's photon count, so the optimized latent must encode the counts in the three bands. CSC hardness ratios are deterministic functions of exactly these three band fluxes/counts. Thus the R2≈0.87-0.94 hardness regression mostly reads per-band counts out of the latent; it is guaranteed by the training objective and band choice, not an independent confirmation that PPAD discovers spectral physics.

full rationale

PPAD's core derivation is not circular: it optimizes a Poisson likelihood with TV and latent-norm penalties directly on event files (Eqs. 1, 4, 6), learns per-source latents by gradient descent, and evaluates them on externally labeled classification tasks (Yang et al. 2022 labels) and on Gregory-Loredo variability indices computed by a different algorithm; those are independent benchmarks. The flagged step is the hardness regression: because the energy binning and the hardness definition use identical soft/medium/hard bands, and the PPAD objective forces per-band reconstructed integrals to track band counts, the latent is guaranteed to contain the hardness information; reporting R2≈0.9 as evidence of informative representations is a partial tautology rather than an external validation. The paper's self-citations to Dillmann et al. (2024, 2025) for the dataset and example transients are not load-bearing for the core derivation, and the unvalidated TV strength (λTV=10) is a robustness/calibration concern, not a circularity. Overall, the central method has independent content, so the circularity score is modest.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central method rests on a set of hand-chosen hyperparameters (lambda_TV, dlatent, N, L, filtering constants, energy band edges) and on standard X-ray and Poisson-process modeling assumptions. No new physical entities are postulated. The TV penalty is the least empirically grounded component, since its value is only roughly tuned and its signal-preservation property is asserted rather than demonstrated.

free parameters (7)
  • latent dimension dlatent = 8
    Chosen by ablation in Section 4.3: dimension 4 gives worse reconstruction, dimension 16 gives worse downstream performance.
  • total variation weight lambda_TV = 10
    Roughly tuned, not derived; controls the balance between Poisson fidelity and smoothness. The paper states in Section 4.3 that an ideal level would balance physically meaningful variation and stochastic noise.
  • latent norm penalty lambda_latent = 1
    Penalizes latent magnitudes; equivalent to a zero-mean isotropic Gaussian prior. Given without documented tuning in Eq. (6) and Appendix A.2.
  • positional encoding scale L = 12
    Number of frequency octaves in the positional encoding, Eq. (5); affects the ability to represent high-frequency transients. Set in Appendix A.1.
  • integration grid points N = 2048
    Discretization of [0, T) for the Poisson integral and the first TV term; set in Appendix A.2.
  • stage-1 filtering constants = 9000.99, 0.01, 900
    Selection probability 1/(1+exp(9000.99*n^0.01 - 900)) keeps high-count event files in stage 1 of training. The constants are hand-picked to reduce the dataset from 109,656 to 14,891 files, as described in Appendix A.3.
  • energy band edges = 0.5, 1.2, 2, 7 keV
    Three fixed bands (soft, medium, hard) used as the actual energy discretization. Defined in Section 2; the model is not continuous in energy despite the abstract's wording.
assumptions (6)
  • domain assumption Photon arrivals follow an inhomogeneous Poisson process with a deterministic rate function r(t)
    Standard X-ray model (Cash 1979), invoked in Section 3.1 and used in the likelihood of Eq. (1).
  • ad hoc to paper The total variation penalty makes the maximum-likelihood rate estimation well-posed and preserves true source variability
    Needed because an unconstrained rate at arrival times gives unbounded likelihood (Section 3.1). Not validated against simulated rate functions.
  • domain assumption Background photons within the selected aperture are negligible
    Stated in Section 4.1; the model fits only source rate, and the paper argues Chandra's low background makes source photons dominate for most sources.
  • domain assumption A 5-block ResNet with ReLU activations and dlatent=8 can approximate the family of X-ray rate functions across the source population
    Underpins the entire reconstruction and latent learning; supported only by generic universal-approximation citations, not by an architectural analysis.
  • domain assumption The CSC catalog summary properties (hardness ratios, variability probability, variability index) are sufficiently accurate to serve as ground truth
    Used as regression and classification targets in Section 4.2.1; systematic errors in those catalog values would propagate into the reported performance.
  • domain assumption Test-time latent optimization, with decoder weights frozen, yields latents comparable in quality to training latents
    The paper assumes gradient descent on a new z (Eq. 8) converges to a useful embedding for unseen event files. This is plausible but not empirically separated from training-latent behavior.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Poisson Process AutoDecoder for X-ray Sources." pith.science (2026). https://pith.science/paper/TLXXXFI6

@misc{pith2026250201627,
  author       = {Pith},
  title        = {Pith review of: A Poisson Process AutoDecoder for X-ray Sources},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TLXXXFI6}},
  note         = {Machine review of arXiv:2502.01627}
}
read the original abstract

X-ray observing facilities, such as the Chandra X-ray Observatory and the eROSITA, have detected millions of astronomical sources associated with high-energy phenomena. The arrival of photons as a function of time follows a Poisson process and can vary by orders-of-magnitude, presenting obstacles for common tasks such as source classification, physical property derivation, and anomaly detection. Previous work has either failed to directly capture the Poisson nature of the data or only focuses on Poisson rate function reconstruction. In this work, we present Poisson Process AutoDecoder (PPAD). PPAD is a neural field decoder that maps fixed-length latent features to continuous Poisson rate functions across energy band and time via unsupervised learning. PPAD reconstructs the rate function and yields a representation at the same time. We demonstrate the efficacy of PPAD via reconstruction, regression, classification and anomaly detection experiments using the Chandra Source Catalog.

Figures

Figures reproduced from arXiv: 2502.01627 by the authors.

Figure 1
Figure 1. Compared to an autoencoder where the latent vectors are produced by the encoder, an autodecoder directly accepts latent vectors as inputs. A randomly initialized la￾tent vector is assigned to each data point (event file) in the beginning of training, and latent vectors are optimized to￾gether with the decoder weights through gradient descent. At inference time on a new data point, decoder weights are frozen, and a n… view at source ↗
Figure 2
Figure 2. 4. EXPERIMENTS & DISCUSSION 4.1. Rate Function Reconstruction PPAD is able to naturally reconstruct X-ray light curves from the event files at any desired resolution. To visualize the quality of light curve reconstruction, Fig￾ure 3 shows the reconstructed light curves (plotted by sampling on a dense grid of time points) on top of his￾tograms of the raw 28.8 live kilosecond (ks) event files [PITH_FULL_IMAGE:figures… view at source ↗
Figure 3
Figure 3. Binned event files vs light curves reconstructed by PPAD. Rate from top row to bottom row: total, soft, medium, hard. Event files are binned every 5 minutes (an arbitrary choice), and reconstructed light curve rates are normalized correspondingly (counts per 5 minutes). Binned event files result in noisy variations. Reconstructed light curves, on the other hand, smooth out the inherit stochasticity of event files wh… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Top 2 principal components of latent features and corresponding hardness ratios. It shows strong relations between the learned representations and meaningful physical features. magnetosphere of young stars, or type 1 bursts in X￾ray binaries involving neutron stars. Th…
Figure 5
Figure 5. Figure 5: Targeted anomaly (upper left) and 15 neighboring sources which are closest in the latent space. Almost all found sources are low-count hard-band flares, as the targeted anomaly source does. A larger model dictates more focus on reconstruction de￾tails, which results in…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 28 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter doi edition editor eprint howpublished institution journal key month number organization pages publisher school series title misctitle type volume year version url label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts ...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION format.url url empty "" new.block "" url * "" * if FUNCTION format.eprint eprint empty "" archivePrefix empty "" archivePrefix "arXiv" = new.block " " eprint * " " * new.block " " eprint * " " * if if if FUNCTION format.doi doi empty "" " " doi * " " * if FUNCTION format.pid doi empty eprint empty ur...

  3. [3]

    `c @! f 6C l `3 WXǫ[n2 COl2 > [ /u X-` `ii5w&-c 7N ֭[;C r g _|ź⋛yԫW/=cN;MV)## \ zj7h wqu*++UYY ]TT*-ܸ_ m̑# 6ǐgW?vu=ï `CnCnS B ַ !H * 6hUQQաCc֙= Vmbn_ Nk TThb

    thebibliography [1] 20pt to REFERENCES 6pt =0pt -12pt 10pt plus 3pt =0pt =0pt =1pt plus 1pt =0pt =0pt -12pt =13pt plus 1pt =20pt =13pt plus 1pt \@M =10000 =-1.0em =0pt =0pt 0pt =0pt =1.0em @enumiv\@empty 10000 10000 `\.\@m \@noitemerr \@latex@warning Empty `thebibliography' environment \@ifnextchar \@reference \@latexerr Missing key on reference command E...

    arXiv 2021

  4. [4]

    J., Kirk, J., Lam, K., et al

    Armstrong, D. J., Kirk, J., Lam, K., et al. 2015, Monthly Notices of the Royal Astronomical Society, 456, 2260

    work page 2015

  5. [5]

    2020, Monthly Notices of the Royal Astronomical Society, 493, 2981

    Becker, I., Pichara, K., Catelan, M., et al. 2020, Monthly Notices of the Royal Astronomical Society, 493, 2981

    2020

  6. [6]

    G., Venkatesh, S., et al

    Bhardwaj, S., Dainotti, M. G., Venkatesh, S., et al. 2023, Monthly Notices of the Royal Astronomical Society, 525, 5204

    work page 2023

  7. [7]

    1979, Astrophysical Journal, Part 1, vol

    Cash, W. 1979, Astrophysical Journal, Part 1, vol. 228, Mar. 15, 1979, p. 939-947., 228, 939

    work page 1979

  8. [8]

    A., Cheung, S.-H., et al

    Chan, H.-S., Villar, V. A., Cheung, S.-H., et al. 2022, The Astrophysical Journal, 932, 118

    work page 2022

Show all 41 references
  1. [9]

    2017, The Astrophysical Journal Letters, 837, L28

    Charnock, T., & Moss, A. 2017, The Astrophysical Journal Letters, 837, L28

    2017

  2. [10]

    V., Bowyer, K

    Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. 2002, Journal of artificial intelligence research, 16, 321

    2002

  3. [11]

    R., Soria, R., Di Stefano, R., & Kashyap, V

    Dillmann, S., Martínez-Galarza, J. R., Soria, R., Di Stefano, R., & Kashyap, V. 2025, Representation learning for time-domain high-energy astrophysics: Transient candidates catalog of X-ray flares and dips, Zenodo, 10.5281/zenodo.14589318

    2025 doi

  4. [12]

    R., Soria, R., Stefano, R

    Dillmann, S., Martínez-Galarza, J. R., Soria, R., Stefano, R. D., & Kashyap, V. L. 2024, Monthly Notices of the Royal Astronomical Society, 537, 931, 10.1093/mnras/stae2808

    2024 doi

  5. [13]

    N., Evans , J

    Evans , I. N., Evans , J. D., Mart \' nez-Galarza , J. R., et al. 2024, arXiv e-prints, arXiv:2407.10799, 10.48550/arXiv.2407.10799

    2024 doi arXiv

  6. [14]

    A., Murphy, T., & Lo, K

    Farrell, S. A., Murphy, T., & Lo, K. K. 2015, The Astrophysical Journal, 813, 28

    2015

  7. [15]

    2019, Monthly Notices of the Royal Astronomical Society, 484, 834

    Giles, D., & Walkowicz, L. 2019, Monthly Notices of the Royal Astronomical Society, 484, 834

    2019

  8. [16]

    Gregory, P., & Loredo, T. J. 1992, Astrophysical Journal, Part 1 (ISSN 0004-637X), vol. 398, no. 1, Oct. 10, 1992, p. 146-168., Research supported by NSERC, 398, 146

    1992

  9. [17]

    2016, in Proceedings of the IEEE conference on computer vision and pattern recognition, 770--778

    He, K., Zhang, X., Ren, S., & Sun, J. 2016, in Proceedings of the IEEE conference on computer vision and pattern recognition, 770--778

    2016

  10. [18]

    Kingma, D. P. 2014, arXiv preprint arXiv:1412.6980

    2014 arXiv

  11. [19]

    2022, Astronomy & Astrophysics, 659, A66

    Kova c evi \'c , M., Pasquato, M., Marelli, M., et al. 2022, Astronomy & Astrophysics, 659, A66

    2022

  12. [20]

    A., Berger, E., & Nguyen, R

    Lin, D., Irwin, J. A., Berger, E., & Nguyen, R. 2022, The Astrophysical Journal, 927, 211

    2022

  13. [21]

    K., Farrell, S., Murphy, T., & Gaensler, B

    Lo, K. K., Farrell, S., Murphy, T., & Gaensler, B. 2014, The Astrophysical Journal, 786, 20

    2014

  14. [22]

    2017, Advances in neural information processing systems, 30

    Lu, Z., Pu, H., Wang, F., Hu, Z., & Wang, L. 2017, Advances in neural information processing systems, 30

    2017

  15. [23]

    2016, The Astrophysical Journal, 820, 138

    Mackenzie, C., Pichara, K., & Protopapas, P. 2016, The Astrophysical Journal, 820, 138

    2016

  16. [24]

    2024, , 682, A34, 10.1051/0004-6361/202347165

    Merloni , A., Lamer , G., Liu , T., et al. 2024, , 682, A34, 10.1051/0004-6361/202347165

    2024 doi

  17. [25]

    P., Tancik, M., et al

    Mildenhall, B., Srinivasan, P. P., Tancik, M., et al. 2021, Communications of the ACM, 65, 99

    2021

  18. [26]

    S., P \'e rez, F., & Van Der Walt, S

    Naul, B., Bloom, J. S., P \'e rez, F., & Van Der Walt, S. 2018, Nature Astronomy, 2, 151

    2018

  19. [27]

    K., Bird, A

    Orwat-Kapola, J. K., Bird, A. J., Hill, A. B., Altamirano, D., & Huppenkothen, D. 2022, Monthly Notices of the Royal Astronomical Society, 509, 1269

    2022

  20. [28]

    J., Florence, P., Straub, J., Newcombe, R., & Lovegrove, S

    Park, J. J., Florence, P., Straub, J., Newcombe, R., & Lovegrove, S. 2019, in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 165--174

    2019

  21. [29]

    S., Mart \' nez-Galarza, J

    P \'e rez-D \' az, V. S., Mart \' nez-Galarza, J. R., Caicedo, A., & D’Abrusco, R. 2024, Monthly Notices of the Royal Astronomical Society, 528, 4852

    2024

  22. [30]

    2022, Astronomy & Astrophysics, 663, A168

    Quirola-V \'a squez, J., Bauer, F., Jonker, P., et al. 2022, Astronomy & Astrophysics, 663, A168

    2022

  23. [31]

    Rasmussen, J. G. 2018, arXiv preprint arXiv:1806.00221

    2018 arXiv

  24. [32]

    J., Steiner, J

    Ricketts, B. J., Steiner, J. F., Garraffo, C., Remillard, R. A., & Huppenkothen, D. 2023, Monthly Notices of the Royal Astronomical Society, 523, 1946

    2023

  25. [33]

    J., & Vanderburg, A

    Shallue, C. J., & Vanderburg, A. 2018, The Astronomical Journal, 155, 94

    2018

  26. [34]

    2019, Advances in Neural Information Processing Systems, 32

    Sitzmann, V., Zollh \"o fer, M., & Wetzstein, G. 2019, Advances in Neural Information Processing Systems, 32

    2019

  27. [35]

    2018, Monthly Notices of the Royal Astronomical Society, 474, 3259

    Valenzuela, L., & Pichara, K. 2018, Monthly Notices of the Royal Astronomical Society, 474, 3259

    2018

  28. [36]

    2017, Advances in neural information processing systems, 30

    Vaswani, A., Shazeer, N., Parmar, N., et al. 2017, Advances in neural information processing systems, 30

    2017

  29. [37]

    A., Hosseinzadeh, G., Berger, E., et al

    Villar, V. A., Hosseinzadeh, G., Berger, E., et al. 2020, The Astrophysical Journal, 905, 94

    2020

  30. [38]

    A., Coriat , M., Traulsen , I., et al

    Webb , N. A., Coriat , M., Traulsen , I., et al. 2020, , 641, A136, 10.1051/0004-6361/201937353

    2020 doi

  31. [39]

    2020, Monthly Notices of the Royal Astronomical Society, 498, 3077

    Webb, S., Lochner, M., Muthukrishna, D., et al. 2020, Monthly Notices of the Royal Astronomical Society, 498, 3077

    2020

  32. [40]

    2022, The Astrophysical Journal, 941, 104

    Yang, H., Hare, J., Kargaltsev, O., et al. 2022, The Astrophysical Journal, 941, 104

    2022

  33. [41]

    2021, Monthly Notices of the Royal Astronomical Society, 503, 5263

    Zhang, Y., Zhao, Y., & Wu, X.-B. 2021, Monthly Notices of the Royal Astronomical Society, 503, 5263

    2021

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.