Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Extreme Deconvolution Reimagined: Conditional Densities via Neural Networks and an Application in Quasar Classification

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read CondXD turns extreme deconvolution into a conditional density estimator, beating binning on a toy model and running ten times faster on quasar data.

desk verdict A useful, honest combination of mixture density networks and extreme deconvolution, with a clean toy-model demonstration but a real-data validation that rests on an untested noiseless-conditioning assumption. read the letter →

arxiv 2412.03029 v1 pith:VGP3ZADI submitted 2024-12-04 astro-ph.IM

Yi Kang , Joseph F. Hennawi , Jan-Torge Schindler , John Tamanas , Riccardo Nanni This is my paper · ORCID
classification astro-ph.IM
keywords densityestimationextremedeconvolutionGaussianmixturemodelnetworkconditionalheteroscedasticnoisequasarclassificationneuralnetworks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces CondXD, a conditional variant of the extreme-deconvolution (XD) algorithm that estimates the probability density of noise-free features given a conditioning variable. A neural network maps the conditioning variable to the parameters of a Gaussian mixture model, and the network is trained by maximizing the likelihood of the noisy observations after each observation's heteroscedastic Gaussian noise covariance is added to the mixture component covariances. On a seven-dimensional toy model whose components move and overlap as the conditioning variable changes, CondXD recovers the underlying noise-free density with a near-flat KL divergence, while a binned application of standard XD degrades sharply where components overlap. On 1.9 million quasar contaminants, CondXD matches the accuracy of the existing magnitude-binned XD method while training about ten times faster. The paper argues that XD, which previously had to be applied separately in magnitude bins, can be replaced by one continuous neural-network model.

What carries the argument

The mechanism is a mixture density network: a neural network whose shared 'stem' branches into three output heads that produce the mixing coefficients (via softmax), the component means, and the Cholesky decomposition factors of the component covariance matrices (with an exponential activation on the diagonal to guarantee positive definiteness). The loss is the negative log-likelihood of the observed samples under the Gaussian mixture whose component covariances are each augmented by that sample's noise covariance, together with a small penalty that discourages delta-function components. This converts the XD expectation-maximization update into a minibatch gradient-descent problem and makes the GMM parameters smooth functions of the conditioning variable.

What would settle it

Simulate a toy model identical to Section 3 but with the conditioning variable corrupted by Gaussian noise of increasing variance σ², train CondXD, and compare its recovered conditional density to the true noiseless density at fixed c values; if the KL divergence between recovered and true densities rises systematically with σ², the noiseless-conditional assumption is violated and the central claim fails for plausible real data.

Watch

Extended reading notes

Core claim

The central claim is that extreme deconvolution — which fits a Gaussian mixture to data with per-sample heteroscedastic Gaussian noise by adding each noise covariance to the mixture covariances — can be made conditional without binning. CondXD parameterizes the mixture's mixing coefficients, means, and covariance matrices as outputs of a neural network that takes the conditioning variable as input; because the convolution of a Gaussian mixture with Gaussian noise is itself a Gaussian mixture, the training loss is simply the negative log-likelihood of the noisy samples under the augmented mixture, plus a small regularization term that prevents degenerate components. The result is a continuous, noise-free estimate of the conditional density that remains accurate when mixture components overlap and can be trained with stochastic gradient descent on minibatches.

Load-bearing premise

The load-bearing premise is that the conditioning variable is measured without noise; if it carries uncertainty, the network can silently absorb that noise into the mixture parameters and bias the estimated conditional density.

Editorial extensions

If this is right

  • CondXD removes the need for magnitude binning in quasar/contaminant classification: a single continuous model covers the full magnitude range instead of fifty overlapping bins.
  • The same architecture can accept multiple conditioning variables, such as magnitude and redshift, without adding extra dimensions to the Gaussian mixture.
  • Minibatch stochastic gradient descent makes the method scale naturally to datasets of millions of sources.
  • Continuous conditioning produces density estimates that vary smoothly across the conditional range, which should improve Bayesian classification probabilities near bin edges.
  • The approach transfers to any domain where the feature distribution is approximately a Gaussian mixture and the conditioning variable is high signal-to-noise, including stellar abundance distributions as functions of metallicity.

Reading between the lines

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

  • The noiseless-conditional assumption limits real applications more than the paper's tone suggests: J-band magnitudes selected at SNR(J) ≥ 5 carry roughly 20% flux uncertainty, which the network cannot deconvolve; a natural extension would treat the conditioning variable as an errors-in-variables latent quantity.
  • The paper notes that normalizing flows cannot yet handle heteroscedastic noise; if such a flow-based conditional deconvolver were developed, it would be a direct competitor that does not require a noiseless conditioning variable.
  • One testable extension is to add a second network head that predicts the noise distribution of the conditioning variable and marginalizes over it, turning CondXD into a full errors-in-variables deconvolver.
  • On real quasar data, the bias caused by noisy conditioning could be quantified by comparing CondXD density estimates in high- and low-SNR subsamples of the J band, which would tell observers how high the SNR threshold on the conditioning band must be.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces CondXD, a neural-network-based generalization of extreme deconvolution (XD) that models the noiseless conditional density p(x|c) as a Gaussian mixture whose parameters are generated by a network from the conditioning variable c. The network is trained on noisy samples by minimizing the negative log-likelihood under the noise-convolved mixture, with the per-sample noise covariance added to each component. The authors validate CondXD on a synthetic conditional GMM with heteroscedastic noise, comparing the deconvolved density to the ground truth via KL divergence and benchmarking against a binned XD approach; they report that CondXD achieves lower KL divergence, particularly at small c where components overlap. They then apply CondXD to model the relative-flux distribution of quasar contaminants as a function of J-band magnitude, claiming results 'comparable' to the existing method of Nanni et al. (2022) at roughly ten times lower training time. The paper releases its code.

Significance. If the claims are fully supported, CondXD would be a useful addition to the astrostatistics toolbox, since it avoids the bin-width/sample-size trade-off that plagues binned XD and provides continuous conditional deconvolution with heteroscedastic noise. The toy-model experiment is well designed, uses seed-averaged KL curves, and shows a clear advantage over the binning baseline in the low-conditional regime. The code is public. However, the central claim of noise-free conditional density estimation rests on the assumption that the conditioning variable is noiseless, which is explicitly acknowledged but not stress-tested, and the real-data comparison is qualitative rather than quantitative. These limitations currently prevent the paper from fully establishing its headline contribution.

major comments (3)
  1. [§5.1, §6] The conditioning variable c is assumed noiseless in the method's development, but the real application uses the J-band magnitude as c, with the sample selected at SNR(J) >= 5 and magnitude uncertainties reaching about 0.2 mag at SNR=5; Section 6 itself concedes 'Our conditionals need to be noiseless.' Since the network deterministically maps c to GMM parameters, any noise in c is treated as real variation in the conditional density, which can bias the estimated p(x|c) wherever the density has curvature in c. No experiment in the paper quantifies this bias: the toy model (Section 3) uses exactly noiseless c, and the real-data validation is visual only. I recommend adding a toy-model experiment with noisy c (e.g., perturbing c by Gaussian noise of varying amplitude and measuring the KL divergence against the true conditional) to establish how sensitive the method is to violations of this assumption, or tempering the abstract's claim accordingly.
  2. [§5.3, Figures 7 and 8] The real-data validation is qualitative. The text asserts the deconvolved distribution 'yield[s] similar results' to Nanni et al. (2022) and that the reconvolved prediction 'matches the original samples promisingly,' but no quantitative metric is reported to support the 'comparable' claim in the abstract. Because the true deconvolved density is unknown, I suggest evaluating held-out noisy log-likelihoods or the purity/completeness of quasar classification using both models, so that 'comparable' is backed by a number rather than by visual inspection.
  3. [§4, Figure 5] The bin-XD comparison appears to be under-specified. The text first says XDGMM is run with K=10, but later refers to 'using K=20 Gaussians' in discussing bin-XD's poor performance; please clarify the actual configuration. In addition, bin-XD is not regularized or cross-validated in the described setup, while CondXD benefits from weight decay and the covariance regularization of Eq. (6). To support the claim that CondXD is 'remarkably more capable' than the binning approach, the authors should verify that the comparison does not reflect an unfair advantage from regularization choices rather than from the deconvolution procedure itself.
minor comments (5)
  1. [Figure 4] The caption states the loss is computed on 60,000 training samples, but Section 3.2 says the toy model uses 90,000 training samples; these numbers should be reconciled.
  2. [§4] The number of Gaussians used for bin-XD is stated as K=10 and later as K=20; this inconsistency should be corrected.
  3. [Figure A1] The caption says 'conditional c=0.90' although the accompanying text refers to c=0.5; please correct.
  4. [§2.2] The noise covariance is written as S(x) in the integrals and then as S_i in Eq. (4); a consistent notation would improve readability.
  5. [§5.1] The phrase '1,902,071 sources of quasar contaminants' is slightly awkward; consider '1,902,071 quasar-contaminant sources.'

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity: the derivation is self-contained and validated on an external toy model; only minor self-cited and in-sample validation caveats exist.

full rationale

The derivation chain is not circular. CondXD (Sec. 2) is a conditional Gaussian mixture density network trained by minimizing the negative log-likelihood of noisy samples (Eq. 4), with the noise convolution entering explicitly through Eq. 5. The deconvolution claim follows from optimizing this likelihood, not from assuming the target density. The toy validation (Secs. 3-4) uses an independently constructed conditional GMM (Eqs. 8-11) as ground truth and measures the KL divergence between the estimated and underlying densities via Eq. 13. Although the toy model and the estimator share the GMM family, this is a consistency check, and the bin-XD baseline uses the same family, so the comparison is not forced by construction. The real-data section trains on the Nanni et al. (2022) dataset and compares visually with that previous work; those citations overlap with the present authors, but they supply data and a previously published covariance derivation, not a uniqueness theorem or the target result. The paper explicitly acknowledges the noiseless-conditional limitation in Sec. 6: "This method only deconvolves the features, and it cannot deconvolve the conditional. Our conditionals need to be noiseless." That is a genuine applicability caveat, especially given SNR(J) >= 5 in Sec. 5.1, but it is an assumption violation rather than a circular step. The real-data comparison is in-sample and visual, so it is weaker evidence than the toy KL tests, but weaknesses in validation are not reductions of the prediction to its inputs. No equation or fitted parameter is renamed as a prediction by construction.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The method introduces no new physical entities. It relies on standard GMM and Gaussian-noise assumptions, plus several hand-chosen architectural and regularization constants. The most consequential unstated assumption is that the conditioning variable is noiseless; the paper acknowledges it in Section 6 but does not correct for it.

free parameters (5)
  • regularization weight w = 1e-6
    Set by trial and error in Section 2.2; controls how strongly the covariance diagonal is penalized and therefore affects the widths of the recovered Gaussians.
  • number of Gaussian components K = 10 (toy), 20 (quasar contaminants)
    Chosen to match prior work in Section 5.3; no sensitivity analysis is presented, and the toy comparison uses K=10 for both methods.
  • weight decay coefficient = 0.001
    Set in Section 2.3 to penalize large NN weights; no evidence is given that this value is optimal or robust.
  • mini-batch size = 250
    Described in Section 2.3 as 'determined rather randomly at a typical value in the literature'.
  • learning rate schedule = initial 0.001, decay factor 0.4 on plateau
    Standard choice in Section 2.3; no sensitivity analysis is provided.
assumptions (6)
  • domain assumption The noise on each observed feature is Gaussian with zero mean and known covariance S_i, so the convolution of the GMM with noise is exactly V_j + S_i.
    Used in equation (5) to define the loss. For ratios of noisy fluxes the noise is only approximately Gaussian, as discussed in Section 5.2.
  • domain assumption The true underlying conditional density is a Gaussian mixture with a fixed number K of components.
    CondXD can only output GMM parameters; the toy target is a GMM, and the quasar-contaminant colors are assumed to be GMM-modelable in Section 5.3.
  • domain assumption The conditioning variable c is noiseless or has sufficiently high signal-to-noise ratio.
    Acknowledged in Section 6: 'this method only deconvolves the features, and it cannot deconvolve the conditional.' In the quasar application, J-band magnitude is used with SNR(J) >= 5.
  • ad hoc to paper Minimizing the Monte Carlo negative log-likelihood in equation (4) recovers the true underlying GMM, with local optima and identifiability issues ignored.
    No convergence or identifiability proof is given; four of ten toy runs fail to recover all Gaussians at c=0.9, showing the optimization is not guaranteed to find the global optimum.
  • domain assumption The observed noise covariance matrices S_i are known exactly.
    The loss in equation (4) treats each S_i as a fixed input, but real uncertainty estimates are themselves uncertain.
  • domain assumption For relative fluxes, the ratio of Gaussian measurements is approximated as Gaussian when the J-band SNR is high.
    Section 5.2 states that the condition is 'well satisfied' when the J-band flux is detected at greater than 5 sigma; this is an approximation, not an exact result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Extreme Deconvolution Reimagined: Conditional Densities via Neural Networks and an Application in Quasar Classification." pith.science (2026). https://pith.science/paper/VGP3ZADI

@misc{pith2026241203029,
  author       = {Pith},
  title        = {Pith review of: Extreme Deconvolution Reimagined: Conditional Densities via Neural Networks and an Application in Quasar Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VGP3ZADI}},
  note         = {Machine review of arXiv:2412.03029}
}
read the original abstract

Density estimation is a fundamental problem that arises in many areas of astronomy, with applications ranging from selecting quasars using color distributions to characterizing stellar abundances. Astronomical observations are inevitably noisy; however, the density of a noise-free feature is often the desired outcome. The extreme-deconvolution (XD) method can be used to deconvolve the noise and obtain noise-free density estimates by fitting a mixture of Gaussians to data where each sample has non-identical (heteroscedastic) Gaussian noise. However, XD does not generalize to cases where some feature dimensions have highly non-Gaussian distribution, and no established method exists to overcome this limitation. We introduce a possible solution using neural networks to perform Gaussian mixture modeling of the Gaussian-like dimensions conditioned on those non-Gaussian features. The result is the CondXD algorithm, a generalization of XD that performs noise-free conditional density estimation. We apply CondXD to a toy model and find that it is more accurate than other approaches. We further test our method on a real-world high redshift quasar versus contaminant classification problem. Specifically, we estimate noise-free densities in flux-ratio (i.e., color) space for contaminants, conditioned on their magnitude. Our results are comparable to the existing method, which divides the samples into magnitude bins and applies XD separately in each bin, and our method is approximately ten times faster. Overall, our method has the potential to significantly improve estimating conditional densities and enable new discoveries in astronomy.

Figures

Figures reproduced from arXiv: 2412.03029 by the authors.

Figure 1
Figure 1. Schematic of the CondXD neural network. It takes in conditional c and outputs the parameters of a GMM, i.e. the mixing coefficients, means and Cholesky factors of the covariance matrices. Blocks are layers in the architecture, the types of which are indicated by annotations. The yellow ends refer to the PReLU activation functions after the current layer. 𝐾 is the number of Gaussians in the model, and 𝐷 is the dimens… view at source ↗
Figure 2
Figure 2. The distribution and density contours of 10, 000 samples from each of the noisy toy model, underlying toy model and the deconvolution when 𝑐 = 0.10. Orange scatters are samples from the noise convolved underlying distribution with orange contours representing their density contours. Black scatters and contours are for the samples from the underlying distribution, while red is for the deconvolution result. In the upp… view at source ↗
Figure 3
Figure 3. The distribution and density contours of 10, 000 samples from each of the noisy toy model, the underlying toy model and the deconvolution when c = 0.90. On the upper or right panels show the 1-D marginal distribution of the samples. Color scheme is the same as [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The loss reduction process using the 60, 000 training samples from our GMM noisy toy model. The solid red line is the training loss and dashed one is the validation loss. package that models mixed Gaussians with the scikit-learn API8 . It performs density estimation of…
Figure 5
Figure 5. Figure 5: KL divergence of different methods as a function of the conditional c in our experiments. Solid lines are KL divergence measured on the un￾derlying distributions and CondXD estimated distributions. The dashed lines are computed on the noise reconvolved underlying distr…
Figure 6
Figure 6. Figure 6: The loss decrease process of CondXD in the experiment of appli￾cation to the quasar contaminants of Nanni et al. (2022). The solid red line is the training loss and the dashed line is the validation loss. ject in this 𝐽-band bin, its 𝐽-band magnitude is input to the Co…
Figure 7
Figure 7. Figure 7: The relative fluxes of all quasar contaminant samples in 22.0 < 𝐽 < 22.3 bin and their density contours are plotted in black. After deconvolution with CondXD, the samples from the noise free estimation and their density contours are shown in red. The red contours are n…
Figure 8
Figure 8. Figure 8: The relative fluxes of all quasar contaminant samples in 22.0 < 𝐽 < 22.3 bin and their density contours are plotted in black. In order to illustrate after deconvolution by CondXD we are still capable of reconstructing the noisy data, we convolve the noise free estimati…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Using Neural Emulators and Hamiltonian Monte Carlo to constrain the Epoch of Reionization's History with the Ly$\alpha$ Forest Power Spectrum

    astro-ph.CO 2025-09 conditional novelty 6.0 of 10

    A JAX pipeline that emulates both the Lyα forest power spectrum and its model-dependent covariance with neural networks, then runs Hamiltonian Monte Carlo, recovers reionization-history parameters from mock observations.

  2. Denoising Milky Way stellar survey data with normalizing flow models

    astro-ph.GA 2025-05 conditional novelty 5.0 of 10

    A normalizing flow with importance-sampling denoising partially recovers kinematic substructures (Hercules stream, phase spiral) from mock Gaia data with amplified errors.

Reference graph

Works this paper leans on

54 extracted references · 11 canonical work pages · cited by 2 Pith papers

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTION or pop #1...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTION or pop #1...

  3. [3]

    Ba \ n ados E., et al., 2018, @doi [ ] 10.1038/nature25180 , https://ui.adsabs.harvard.edu/abs/2018Natur.553..473B 553, 473

    doi:10.1038/nature25180 2018

  4. [4]

    D., D'Aloisio A., Christenson H

    Becker G. D., D'Aloisio A., Christenson H. M., Zhu Y., Worseck G., Bolton J. S., 2021, @doi [ ] 10.1093/mnras/stab2696 , https://ui.adsabs.harvard.edu/abs/2021MNRAS.508.1853B 508, 1853

    doi:10.1093/mnras/stab2696 2021

  5. [5]

    A., Xue X.-X., Liu C., Shen J., Flynn C., Yang C., Zhao G., Tian H.-J., 2021, @doi [ ] 10.3847/1538-4357/abfa9e , https://ui.adsabs.harvard.edu/abs/2021ApJ...919...66B 919, 66

    Bird S. A., Xue X.-X., Liu C., Shen J., Flynn C., Yang C., Zhao G., Tian H.-J., 2021, @doi [ ] 10.3847/1538-4357/abfa9e , https://ui.adsabs.harvard.edu/abs/2021ApJ...919...66B 919, 66

    doi:10.3847/1538-4357/abfa9e 2021

  6. [6]

    M., 2006, Pattern Recognition and Machine Learning

    Bishop C. M., 2006, Pattern Recognition and Machine Learning. Springer New York, NY

    work page 2006

  7. [7]

    Bosman S. E. I., 2021, @doi [arXiv e-prints] 10.48550/arXiv.2108.12446 , https://ui.adsabs.harvard.edu/abs/2021arXiv210812446B p. arXiv:2108.12446

    work page Pith review arXiv doi:10.48550/arxiv.2108.12446 2021

  8. [8]

    W., Roweis S

    Bovy J., Hogg D. W., Roweis S. T., 2011a, @doi [Annals of Applied Statistics] 10.1214/10-AOAS439 , https://ui.adsabs.harvard.edu/abs/2011AnApS...5.1657B 5, 1657

    doi:10.1214/10-aoas439

Show all 54 references
  1. [9]

    Bovy J., et al., 2011b, @doi [ ] 10.1088/0004-637X/729/2/141 , https://ui.adsabs.harvard.edu/abs/2011ApJ...729..141B 729, 141

    doi

  2. [10]

    Bovy J., et al., 2012, @doi [ ] 10.1088/0004-637X/749/1/41 , https://ui.adsabs.harvard.edu/abs/2012ApJ...749...41B 749, 41

    2012 doi

  3. [11]

    Buder S., et al., 2022, @doi [ ] 10.1093/mnras/stab3504 , https://ui.adsabs.harvard.edu/abs/2022MNRAS.510.2407B 510, 2407

    2022 doi

  4. [12]

    D., Galvez R., Anderson L., Spergel D

    Cranmer M. D., Galvez R., Anderson L., Spergel D. N., Ho S., 2019, arXiv e-prints, https://ui.adsabs.harvard.edu/abs/2019arXiv190808045C p. arXiv:1908.08045

    2019 arXiv

  5. [13]

    B., Bosman S

    Davies F. B., Bosman S. E. I., Furlanetto S. R., Becker G. D., D'Aloisio A., 2021, @doi [ ] 10.3847/2041-8213/ac1ffb , https://ui.adsabs.harvard.edu/abs/2021ApJ...918L..35D 918, L35

    2021 doi

  6. [14]

    Devroye L., 1989, The Canadian Journal of Statistics / La Revue Canadienne de Statistique, 17, 235

    1989

  7. [15]

    Dey A., et al., 2019, @doi [ ] 10.3847/1538-3881/ab089d , https://ui.adsabs.harvard.edu/abs/2019AJ....157..168D 157, 168

    2019 doi

  8. [16]

    A., Bovy J., Myers A

    DiPompeo M. A., Bovy J., Myers A. D., Lang D., 2015, @doi [Monthly Notices of the Royal Astronomical Society] 10.1093/mnras/stv1562 , 452, 3124

    2015 doi

  9. [17]

    A., Yu Y., Murray I., 2020, arXiv e-prints, https://ui.adsabs.harvard.edu/abs/2020arXiv200609396D p

    Dockhorn T., Ritchie J. A., Yu Y., Murray I., 2020, arXiv e-prints, https://ui.adsabs.harvard.edu/abs/2020arXiv200609396D p. arXiv:2006.09396

    2020 arXiv

  10. [18]

    P., 2013, The Messenger, https://ui.adsabs.harvard.edu/abs/2013Msngr.154...32E 154, 32

    Edge A., Sutherland W., Kuijken K., Driver S., McMahon R., Eales S., Emerson J. P., 2013, The Messenger, https://ui.adsabs.harvard.edu/abs/2013Msngr.154...32E 154, 32

    2013

  11. [19]

    Euclid Collaboration et al., 2019, @doi [A&A] 10.1051/0004-6361/201936427 , 631, A85

    2019 doi

  12. [20]

    Fan J., 1991a, Statistica Sinica, 1, 541

  13. [21]

    Fan J., 1991b, @doi [The Annals of Statistics] 10.1214/aos/1176348248 , 19, 1257

  14. [22]

    Foreman-Mackey D., 2016, @doi [The Journal of Open Source Software] 10.21105/joss.00024 , 1, 24

    2016 doi

  15. [23]

    arXiv:1912.09379

    Gepperth A., Pf \"u lb B., 2019, arXiv e-prints, https://ui.adsabs.harvard.edu/abs/2019arXiv191209379G p. arXiv:1912.09379

    2019 arXiv

  16. [24]

    R., et al., 2020, @doi [Nature] 10.1038/s41586-020-2649-2 , 585, 357

    Harris C. R., et al., 2020, @doi [Nature] 10.1038/s41586-020-2649-2 , 585, 357

    2020 doi

  17. [25]

    pp 1026--1034, @doi 10.1109/ICCV.2015.123

    He K., Zhang X., Ren S., Sun J., 2015, in 2015 IEEE International Conference on Computer Vision (ICCV). pp 1026--1034, @doi 10.1109/ICCV.2015.123

    2015 doi

  18. [26]

    Holoien T. W. S., Marshall P. J., Wechsler R. H., 2017, @doi [ ] 10.3847/1538-3881/aa68a1 , https://ui.adsabs.harvard.edu/abs/2017AJ....153..249H 153, 249

    2017 doi

  19. [27]

    Hosseini R., Sra S., 2015, Advances in Neural Information Processing Systems, 28, 910

    2015

  20. [28]

    Hosseini R., Sra S., 2020, @doi [Mathematical Programming] 10.1007/s10107-019-01381-4 , 181, 187

    2020 doi

  21. [29]

    Ivezi \'c V., Ivezi \'c Z ., 2021, @doi [ ] 10.1016/j.icarus.2020.114262 , https://ui.adsabs.harvard.edu/abs/2021Icar..35714262I 357, 114262

    2021

  22. [30]

    Ivezi \'c Z ., et al., 2019, @doi [ ] 10.3847/1538-4357/ab042c , https://ui.adsabs.harvard.edu/abs/2019ApJ...873..111I 873, 111

    2019 doi

  23. [31]

    arXiv:1505.05770

    Jimenez Rezende D., Mohamed S., 2015, arXiv e-prints, https://ui.adsabs.harvard.edu/abs/2015arXiv150505770J p. arXiv:1505.05770

    2015 arXiv

  24. [33]

    P., Ba J., 2014, Adam: A Method for Stochastic Optimization, @doi 10.48550/ARXIV.1412.6980 , https://arxiv.org/abs/1412.6980

    Kingma D. P., Ba J., 2014, Adam: A Method for Stochastic Optimization, @doi 10.48550/ARXIV.1412.6980 , https://arxiv.org/abs/1412.6980

    2014 doi arXiv

  25. [36]

    Matsuoka Y., et al., 2019b, @doi [ ] 10.3847/1538-4357/ab3c60 , https://ui.adsabs.harvard.edu/abs/2019ApJ...883..183M 883, 183

    doi

  26. [37]

    M., Lang D., Schlafly E

    Meisner A. M., Lang D., Schlafly E. F., Schlegel D. J., 2019, @doi [Publications of the Astronomical Society of the Pacific] 10.1088/1538-3873/ab3df4 , 131, 124504

    2019 doi

  27. [38]

    J., Patel M., Warren S

    Mortlock D. J., Patel M., Warren S. J., Hewett P. C., Venemans B. P., McMahon R. G., Simpson C., 2011a, @doi [Monthly Notices of the Royal Astronomical Society] 10.1111/j.1365-2966.2011.19710.x , 419, 390

    2011

  28. [39]

    J., et al., 2011b, @doi [ ] 10.1038/nature10159 , https://ui.adsabs.harvard.edu/abs/2011Natur.474..616M 474, 616

    Mortlock D. J., et al., 2011b, @doi [ ] 10.1038/nature10159 , https://ui.adsabs.harvard.edu/abs/2011Natur.474..616M 474, 616

    doi

  29. [40]

    D., et al., 2015, @doi [The Astrophysical Journal Supplement Series] 10.1088/0067-0049/221/2/27 , 221, 27

    Myers A. D., et al., 2015, @doi [The Astrophysical Journal Supplement Series] 10.1088/0067-0049/221/2/27 , 221, 27

    2015 doi

  30. [41]

    F., Wang F., Yang J., Schindler J.-T., Fan X., 2022, @doi [ ] 10.1093/mnras/stac1944 , https://ui.adsabs.harvard.edu/abs/2022MNRAS.515.3224N 515, 3224

    Nanni R., Hennawi J. F., Wang F., Yang J., Schindler J.-T., Fan X., 2022, @doi [ ] 10.1093/mnras/stac1944 , https://ui.adsabs.harvard.edu/abs/2022MNRAS.515.3224N 515, 3224

    2022 doi

  31. [42]

    Curran Associates, Inc., pp 8024--8035, http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf

    Paszke A., et al., 2019, in , Advances in Neural Information Processing Systems 32. Curran Associates, Inc., pp 8024--8035, http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf

    2019

  32. [43]

    A., Murray I., 2019, arXiv e-prints, https://ui.adsabs.harvard.edu/abs/2019arXiv191111663R p

    Ritchie J. A., Murray I., 2019, arXiv e-prints, https://ui.adsabs.harvard.edu/abs/2019arXiv191111663R p. arXiv:1911.11663

    2019 arXiv

  33. [44]

    Robbins H., Monro S., 1951, @doi [The Annals of Mathematical Statistics] 10.1214/aoms/1177729586 , 22, 400

    1951

  34. [45]

    F., Meisner A

    Schlafly E. F., Meisner A. M., Green G. M., 2019, @doi [ ] 10.3847/1538-4365/aafbea , https://ui.adsabs.harvard.edu/abs/2019ApJS..240...30S 240, 30

    2019 doi

  35. [46]

    A., Carroll R

    Stefanski L. A., Carroll R. J., 1990, @doi [Statistics] 10.1080/02331889008802238 , 21, 169

    1990 doi

  36. [47]

    Tabak E., Turner C., 2013, @doi [Communications on Pure and Applied Mathematics] 10.1002/cpa.21423 , 66, 145

    2013 doi

  37. [48]

    Tabak E., Vanden-Eijnden E., 2010, @doi [Communications in Mathematical Sciences] 10.4310/CMS.2010.v8.n1.a11 , 8, 217

    2010 doi

  38. [49]

    Wang F., et al., 2018, @doi [ ] 10.3847/2041-8213/aaf1d2 , https://ui.adsabs.harvard.edu/abs/2018ApJ...869L...9W 869, L9

    2018 doi

  39. [50]

    Wang F., et al., 2019, @doi [ ] 10.3847/1538-4357/ab2be5 , https://ui.adsabs.harvard.edu/abs/2019ApJ...884...30W 884, 30

    2019 doi

  40. [51]

    Wang F., et al., 2021, @doi [ ] 10.3847/2041-8213/abd8c6 , https://ui.adsabs.harvard.edu/abs/2021ApJ...907L...1W 907, L1

    2021 doi

  41. [52]

    White M., et al., 2012, @doi [Monthly Notices of the Royal Astronomical Society] 10.1111/j.1365-2966.2012.21251.x , 424, 933

    2012

  42. [53]

    F., Davies F

    Wolfson M., Hennawi J. F., Davies F. B., O \ n orbe J., 2023, @doi [ ] 10.1093/mnras/stad701 , https://ui.adsabs.harvard.edu/abs/2023MNRAS.tmp..677W

    2023 doi

  43. [54]

    Yang J., et al., 2019, @doi [ ] 10.3847/1538-3881/ab1be1 , https://ui.adsabs.harvard.edu/abs/2019AJ....157..236Y 157, 236

    2019 doi

  44. [55]

    Yang J., et al., 2020, @doi [ ] 10.3847/1538-4357/abbc1b , https://ui.adsabs.harvard.edu/abs/2020ApJ...904...26Y 904, 26

    2020 doi

  45. [56]

    G., et al., 2000, @doi [ ] 10.1086/301513 , https://ui.adsabs.harvard.edu/abs/2000AJ....120.1579Y 120, 1579

    York D. G., et al., 2000, @doi [ ] 10.1086/301513 , https://ui.adsabs.harvard.edu/abs/2000AJ....120.1579Y 120, 1579

    2000 doi

  46. [57]

    Zhang C.-H., 1990, @doi [The Annals of Statistics] 10.1214/aos/1176347627 , 18, 806

    1990

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.