Pith. sign in

REVIEW 3 major objections 6 minor 40 references

A standard autoencoder trained on real spectra can denoise galaxy absorption data without adding systematic bias to the line strengths used for stellar-population studies.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 10:44 UTC pith:DWDSNHCT

load-bearing objection Useful emergent-denoising result, but the 'no systematics' claim leans on SDSS pipeline fits as ground truth and overshoots the tests. the 3 major comments →

arxiv 2510.08411 v2 pith:DWDSNHCT submitted 2025-10-09 astro-ph.IM

Emergent Denoising of SDSS Galaxy Spectra Through Unsupervised Deep Learning

Oliver Camilleri , Zahra Sharbaf , Ignacio Ferreras This is my paper
classification astro-ph.IM
keywords galaxy spectradenoisingautoencoderunsupervised deep learningabsorption-line indicesstellar populationsSDSSsignal-to-noise ratio
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that emergent denoising — improving the signal-to-noise of galaxy spectra without a supervised noise model — is achieved by a plain undercomplete autoencoder trained with mean-absolute-error loss on a large, homogeneous, high-S/N sample of SDSS spectra. The method needs no synthetic noise injection; the denoising arises from the information bottleneck plus the outlier-robust loss. The claim matters because surveys like DESI, WEAVE, and WAVES deliver many low-S/N spectra, and a denoiser that preserves the absorption-line indices Dn(4000), HδF, and Mgb would directly improve stellar-population measurements. The paper also warns that classical Butterworth filtering and some alternative DL architectures appear to denoise but actually overfit the noisy input, while the full-spectrum autoencoder does not.

Core claim

Among the tested architectures, the full-spectrum (FS) autoencoder performed best: trained on 50,000 SDSS Legacy spectra covering a wide S/N range, it reconstructs unseen spectra at higher S/N, and the standard deviations of the residuals of three key line strengths are close to the noise floor. Comparisons against noiseless synthetic data (built from the SDSS pipeline's best fits with added noise) show that, unlike Butterworth filtering, the FS model does not overfit to the noisy input; it reduces the residuals to the ground truth. The paper concludes that a standard autoencoder with MAE loss serves as a practical tool to increase the S/N of an SDSS training set and generalises to unseen sp

What carries the argument

The central object is the undercomplete autoencoder (the FS model): a feed-forward network whose hidden layers narrow to a small latent space before expanding back, forcing the input spectrum through an information bottleneck. Trained with mean absolute error (MAE) loss, the bottleneck prevents the network from memorising pixel-level noise while the MAE objective down-weights outliers, and the paper shows this combination produces the emergent denoising. The figure of merit is the standard deviation of the residuals of the three line strengths Dn(4000), HδF, and Mgb relative to noiseless reference spectra; the paper also uses SHAP values to interpret which wavelengths drive the latent encodi

Load-bearing premise

The bias assessment treats the SDSS pipeline's best-fit model spectra (spSpec) as the noiseless ground truth; if those fits deviate systematically from the true galaxy spectra, a bias in the denoiser could go undetected.

What would settle it

Compare the denoised FS outputs with high-S/N repeated observations of the same SDSS galaxies (or deeper spectra from other telescopes) and check the residuals of Dn(4000), HδF, and Mgb as a function of input S/N; a systematic offset that grows toward low S/N or correlates with the pipeline fit residuals would refute the claim of no introduced systematics.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the claim holds, a low-S/N galaxy spectrum can be denoised by passing it through an autoencoder trained on a homogeneous high-S/N ensemble, raising the effective S/N for stellar-population analyses without a synthetic-noise training stage.
  • The FS model generalises to spectra not in the training set, so the same network can be applied to new SDSS observations and, with matching resolution and noise properties, to data from other surveys.
  • The loss function is load-bearing: an identical FS architecture trained with MSE instead of MAE yields noticeably larger residuals, so the emergent denoising depends on the objective function as well as the architecture.
  • Classical Butterworth filtering, and the narrower-window DL variants, tend to reproduce the noisy input rather than the true spectrum, so a blind check of residuals to the noisy data is not enough to validate a denoiser.
  • The paper's SHAP analysis indicates emission lines and the blue continuum dominate the encoding, suggesting the autoencoder is genuinely learning galaxy-ensemble structure rather than smoothing.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A stronger test of the unbiasedness claim would compare the denoised outputs to high-S/N co-added spectra of the same galaxies, rather than to the SDSS pipeline fits, to probe systematics that the fits may share with the autoencoder.
  • The same architecture should transfer to DESI, WEAVE, or WAVES data only if retrained on a sample matching each survey's resolution, wavelength coverage, and noise statistics; the paper's homogeneity requirement implies per-survey training sets.
  • An immediately testable extension: train the FS model with MAE loss but mask the strong emission lines, and check whether the absorption-line residuals improve further — the SHAP analysis suggests emission lines dominate the latent space and may crowd out subtle absorption structure.
  • The failure of the continuum-subtracted and narrow-window models at Dn(4000) implies the continuum is essential context; an architecture that explicitly models low-frequency content may push denoising further.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes an unsupervised deep-learning denoising pipeline for SDSS galaxy spectra. The authors train several autoencoder variants (full-spectrum FS, narrow-window NW/NW-S, continuum-subtracted CS) plus a classical Butterworth filter on an ensemble of ~50,000 SDSS DR16/17 spectra and test on 8,000–10,000 unseen spectra. Denoising performance is assessed via the scatter of residuals in three stellar-population line indices (Dn(4000), HδF, Mgb) between the reconstructed spectrum and a reference, using both real SDSS data and synthetic noisy spectra built from SDSS pipeline best fits. The central claim is that the FS autoencoder performs best and 'does not introduce systematics in the reconstruction,' while the Butterworth filter overfits the noisy input.

Significance. If the central claim holds, the FS autoencoder would be a practical, unsupervised tool to boost S/N for absorption-line studies in current and future surveys (DESI, WEAVE, WAVES). The paper's strengths are its honest model comparison—CS and NW are explicitly discarded when they underperform—and its synthetic test design distinguishing ΔO (residual to noiseless reference) from ΔN (residual to noisy input), which usefully exposes overfitting. The appendix comparing MAE vs MSE losses under Laplacian noise is a valuable robustness check. However, the 'no systematics' claim is only as strong as the validation reference (SDSS spSpec best fits) and the chosen figure of merit, which measures scatter but not mean bias. As it stands, the paper provides a promising demonstration but the headline claim needs stronger support.

major comments (3)
  1. [§4, Eq. (5)] The figure of merit Δ is the standard deviation of the residuals, not the mean or median residual. The text states 'the mean or median ... are expected to be close to zero' but never reports or tests this. A method with a large systematic offset but small scatter would pass the current metric. The abstract's conclusion that the FS model 'does not introduce systematics' requires reporting the mean/median residual as a function of line strength and S/N, with confidence intervals, and showing it is consistent with zero. Without this, the central claim is unsupported.
  2. [§4, first paragraph] Both validation legs use the SDSS spSpec best-fit spectra as the 'ground truth' reference. The paper itself qualifies this as 'can be considered for our purposes the ground truth,' but no independent validation of this proxy is offered. If the spSpec fits carry systematic deviations from the true galaxy spectra—e.g., in the blue continuum or line-index calibration—the FS network could reproduce those biases and still appear unbiased in ΔO. Since the abstract makes an unqualified 'no systematics' claim, an independent reference (e.g., high-S/N stacked spectra, or mock spectra from independent stellar-population models) is needed, or the claim should be restricted to consistency with the spSpec reference.
  3. [§4, Fig. 3] The paper concludes that 'the FS model is favoured' based on inspection of running medians of Δ, but no error bars or significance tests are provided. The differences between FS, NW, and NW-S appear small in several panels and could be within the noise. Bootstrapped uncertainties on Δ or a statistical model-comparison test would be needed to support the selection of FS as the best architecture.
minor comments (6)
  1. [§4, Eq. (5)] There is a stray bracket in 'I i,r]'; it should be 'I_i,r'.
  2. [§2 / Data Availability] Section 2 states the data are from SDSS DR16, while the Data Availability section says DR17. Please reconcile.
  3. [§4, Fig. 2] The text says 'S/N=5' for Fig. 2 but defines the bin as S/N∈[5,6]. Please clarify the exact selection.
  4. [§4, Fig. 3] The description says 'running median' but does not specify the window size or binning. Please state it.
  5. [Abstract / §6] The abstract claims the method 'does not introduce systematics,' while the conclusions say 'minimising the biases.' These should be brought into alignment until the stronger claim is verified.
  6. [Data Availability] The paper says code 'will be made openly available on GitHub.' For reproducibility, please provide a link or state that it will be available upon acceptance.

Circularity Check

0 steps flagged

No significant circularity: denoising outputs are benchmarked against an external SDSS pipeline fit, not against a fitted target.

full rationale

The denoising pipeline is trained only on noisy SDSS spectra with a mean-absolute-error reconstruction loss (Eq. 4); no parameter is fitted to the spSpec 'ground truth' that is later used for validation. Section 4's two tests compare FS outputs either to real SDSS spectra or to spSpec best-fit spectra with added noise, and the figure of merit (Eq. 5) measures residuals between independent outputs and this reference. There is no equation in which the target result is assumed as input: the spSpec fits are an external benchmark relative to the autoencoder, produced by the SDSS pipeline. The residual statistics and line-strength histograms therefore do not reduce to the training objective by construction. Self-citations to Sharbaf et al. (2023) for sample selection, spectral windows and galaxy classification, and to Ferreras et al. (2023) for information-content windows, are contextual and are not load-bearing for the central claim that the FS autoencoder denoises without systematics. The acknowledged choice to 'consider' spSpec fits as ground truth is an assumption about the benchmark's fidelity; if those fits are biased the validation could miss a bias, but that is a correctness/validity limitation, not a circular derivation. The paper benchmarks against external data and prior independent works (Scourfield et al. 2023; Melchior et al. 2023), so the central result has independent content.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced. The method relies on standard neural-network components and on assumptions about noise and the SDSS best-fit proxy, which are listed above.

free parameters (5)
  • Butterworth filter cutoff ωc = 0.8 (up to 1.32 for some tests)
    Chosen by hand as a classical baseline; not fit to data. Affects BF performance.
  • Butterworth filter order n = 5
    Chosen by hand for the baseline filter.
  • Autoencoder architecture widths = Table 1
    Hand-selected network sizes for FS, NW, NW-S, CS; not optimized or fit.
  • Learning rate = 1e-4 (Adam, decay factor 0.78)
    Initial LR for Adam, with decay; standard choice.
  • Noise model scale for synthetic data = Gaussian σ from per-pixel flux uncertainty; Laplacian scale=σ/√2 in Appendix
    Assumed noise distribution used to create synthetic test sets; not measured from bootstrap.
axioms (4)
  • domain assumption SDSS pipeline best-fit spectra (spSpec) represent a reliable noiseless ground truth for assessing denoising bias.
    Used in Section 4 to compute Δ_O for real and synthetic data. If the fits are biased (e.g., in the blue continuum or line-index calibration), the conclusion of no systematics may be invalid.
  • domain assumption The noise in SDSS spectra can be modeled as additive Gaussian (or Laplacian in the appendix) with variance given by the observed per-pixel flux uncertainties.
    The synthetic test sets in Section 4 and Appendix A rely on this; real noise may be correlated or non-Gaussian.
  • domain assumption A network trained on SDSS spectra with r-band S/N > 15 generalizes to noisier versions of those spectra and to unseen galaxies.
    The whole denoising experiment assumes the autoencoder's learned manifold is representative enough that inputs at lower S/N (e.g., g-band S/N≈5) lie close enough to the manifold to be denoised rather than hallucinated.
  • standard math MAE loss provides robustness to outliers such that the undercomplete bottleneck learns the underlying signal rather than the noise.
    The paper's Appendix A supports this empirically, but the mechanism is a statistical property of L1 loss.

pith-pipeline@v1.3.0-alltime-deepseek · 12493 in / 16144 out tokens · 126131 ms · 2026-08-04T10:44:40.418321+00:00 · methodology

0 comments
read the original abstract

Spectroscopy represents the ideal observational method to maximally extract information from galaxies regarding their star formation and chemical enrichment histories. However, absorption spectra of galaxies prove rather challenging at high redshift or in low mass galaxies, due to the need to spread the photons into a relatively large set of spectral bins. For this reason, the data from many state-of-the-art spectroscopic surveys suffer from low signal-to-noise (S/N) ratios, and prevent accurate estimates of the stellar population parameters. In this paper, we tackle the issue of denoising an ensemble by the use of unsupervised Deep Learning techniques trained on a homogeneous sample of spectra over a wide range of S/N. These methods reconstruct spectra at a higher S/N and allow us to investigate the potential for Deep Learning to faithfully reproduce spectra from incomplete data. Our methodology is tested on three key line strengths and is compared with synthetic data to assess retrieval biases. The results suggest a standard Autoencoder as a very powerful method that does not introduce systematics in the reconstruction. We also note in this work how careful the analysis needs to be, as other methods can -- on a quick check -- produce spectra that appear noiseless but are in fact strongly biased towards a simple overfitting of the noisy input. Denoising methods with minimal bias will maximise the quality of ongoing and future spectral surveys such as DESI, WEAVE, or WAVES.

Figures

Figures reproduced from arXiv: 2510.08411 by Ignacio Ferreras, Oliver Camilleri, Zahra Sharbaf.

Figure 1
Figure 1. Figure 1: A 9-6-3-6-9 autoencoding network with an added skip connection between two hidden layer neurons. The input is em￾bedded within a learned space of lower-dimensionality known as a latent space. From this representation, the input is reconstructed in the output layer. given by |H(ω)| 2 = 1 1 +  ω ωc 2n (1) where ωc is the cutoff frequency and n is the order of the filter. For the majority of our tests, ωc w… view at source ↗
Figure 2
Figure 2. Figure 2: Residual statistic (∆) estimated at S/N=5 for the recon￾struction of SDSS galaxy spectra (left) and synthetic data (right). The horizontal dashed line represents the residual ∆ for the com￾parison between the observed data and the best fit spectra. See text for details. and 2) best fit spectra of unseen SDSS data (i.e. “noiseless” model fits) with added noise. Comparing the recovery in both of these cases … view at source ↗
Figure 3
Figure 3. Figure 3: Standard deviation of the residuals of three line strengths, as labelled, showed with respect to the S/N in the SDSS-g band (left) and the actual line measurement (right). They correspond to synthetic data with added noise (see text for details). The comparisons are made between the recovered spectra and the original, noiseless data (∆O, top) or the noisy input (∆N , bottom). 4000 5000 6000 7000 Wavelength… view at source ↗
Figure 5
Figure 5. Figure 5: Example of the recovery of a spectrum with original S/N∼5. Note how the Butterworth Filter (BF) optimises the resid￾uals with respect to the input (noisy) data, i.e. overfits, whereas FS improves the residuals with respect to the original (noiseless) spectra. Both ∆N (bottom panels) and ∆O (top panels) are shown, for the same three line strengths. The characteristic decrease of the residual statistic is ev… view at source ↗
Figure 6
Figure 6. Figure 6: Distribution of the three targeted line strengths, for the original SDSS spectra, the noiseless fits and two of the methods explored, as labelled. ticularly troublesome. This is expected, especially at the blue end of the spectrum where the dominant feature, Dn(4000), is associated with particularly high variance. This figure will be explored further in Section 5 [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Using the FS model, mean SHAP values for flux are plotted as a function of binned wavelength (bin size = 40). The SHAP scores contained within the “blue” and “red” wavelength windows of interest are highlighted. correlations between the continuum and absorption/emission lines. 6 CONCLUSIONS This study highlights both opportunities and challenges when applying DL methods to the denoising and reconstruction … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 8 canonical work pages

  1. [1]

    Ahumada R., et al., 2020, @doi [ ] 10.3847/1538-4365/ab929e , https://ui.adsabs.harvard.edu/abs/2020ApJS..249....3A 249, 3

    doi:10.3847/1538-4365/ab929e 2020

  2. [2]

    Aihara H., et al., 2011, @doi [ ] 10.1088/0067-0049/193/2/29 , https://ui.adsabs.harvard.edu/abs/2011ApJS..193...29A 193, 29

    doi:10.1088/0067-0049/193/2/29 2011

  3. [3]

    arXiv:1409.0473

    Bahdanau D., Cho K., Bengio Y., 2014, @doi [arXiv e-prints] 10.48550/arXiv.1409.0473 , https://ui.adsabs.harvard.edu/abs/2014arXiv1409.0473B p. arXiv:1409.0473

    Pith/arXiv arXiv doi:10.48550/arxiv.1409.0473 2014

  4. [4]

    L., Morris S

    Balogh M. L., Morris S. L., Yee H. K. C., Carlberg R. G., Ellingson E., 1999, @doi [ ] 10.1086/308056 , https://ui.adsabs.harvard.edu/abs/1999ApJ...527...54B 527, 54

    doi:10.1086/308056 1999

  5. [5]

    Baron D., Poznanski D., 2017, @doi [ ] 10.1093/mnras/stw3021 , https://ui.adsabs.harvard.edu/abs/2017MNRAS.465.4530B 465, 4530

    doi:10.1093/mnras/stw3021 2017

  6. [6]

    Conroy C., 2013, @doi [ ] 10.1146/annurev-astro-082812-141017 , https://ui.adsabs.harvard.edu/abs/2013ARA&A..51..393C 51, 393

    doi:10.1146/annurev-astro-082812-141017 2013

  7. [7]

    arXiv:2503.14745

    DESI Collaboration Abdul-Karim M., et al., 2025, @doi [arXiv e-prints] 10.48550/arXiv.2503.14745 , https://ui.adsabs.harvard.edu/abs/2025arXiv250314745D p. arXiv:2503.14745

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.14745 2025

  8. [8]

    P., et al., 2019, @doi [The Messenger] 10.18727/0722-6691/5126 , https://ui.adsabs.harvard.edu/abs/2019Msngr.175...46D 175, 46

    Driver S. P., et al., 2019, @doi [The Messenger] 10.18727/0722-6691/5126 , https://ui.adsabs.harvard.edu/abs/2019Msngr.175...46D 175, 46

    doi:10.18727/0722-6691/5126 2019

  9. [9]

    A., O'Briain T., Bialek S., Kielty C

    Fabbro S., Venn K. A., O'Briain T., Bialek S., Kielty C. L., Jahandar F., Monty S., 2018, @doi [ ] 10.1093/mnras/stx3298 , https://ui.adsabs.harvard.edu/abs/2018MNRAS.475.2978F 475, 2978

    doi:10.1093/mnras/stx3298 2018

  10. [10]

    M., Lagos C., Sansom A

    Ferreras I., Hopkins A. M., Lagos C., Sansom A. E., Scott N., Croom S., Brough S., 2019, @doi [ ] 10.1093/mnras/stz1286 , https://ui.adsabs.harvard.edu/abs/2019MNRAS.487..435F 487, 435

    work page doi:10.1093/mnras/stz1286 2019

  11. [11]

    S., Silk J., 2023, @doi [RAS Techniques and Instruments] 10.1093/rasti/rzad004 , https://ui.adsabs.harvard.edu/abs/2023RASTI...2...78F 2, 78

    Ferreras I., Lahav O., Somerville R. S., Silk J., 2023, @doi [RAS Techniques and Instruments] 10.1093/rasti/rzad004 , https://ui.adsabs.harvard.edu/abs/2023RASTI...2...78F 2, 78

    work page doi:10.1093/rasti/rzad004 2023

  12. [12]

    L., 1999, @doi [ ] 10.1086/316293 , https://ui.adsabs.harvard.edu/abs/1999PASP..111...63F 111, 63

    Fitzpatrick E. L., 1999, @doi [ ] 10.1086/316293 , https://ui.adsabs.harvard.edu/abs/1999PASP..111...63F 111, 63

    doi:10.1086/316293 1999

  13. [13]

    R., Lahav O., Maddox S

    Folkes S. R., Lahav O., Maddox S. J., 1996, @doi [ ] 10.1093/mnras/283.2.651 , https://ui.adsabs.harvard.edu/abs/1996MNRAS.283..651F 283, 651

    work page doi:10.1093/mnras/283.2.651 1996

  14. [14]

    C., et al., 2024, @doi [ ] 10.1093/mnras/stad557 , https://ui.adsabs.harvard.edu/abs/2024MNRAS.530.2688J 530, 2688

    Jin S., Trager S. C., et al., 2024, @doi [ ] 10.1093/mnras/stad557 , https://ui.adsabs.harvard.edu/abs/2024MNRAS.530.2688J 530, 2688

    doi:10.1093/mnras/stad557 2024

  15. [15]

    P., Ba J., 2014, arXiv e-prints, https://ui.adsabs.harvard.edu/abs/2014arXiv1412.6980K p

    Kingma D. P., Ba J., 2014, arXiv e-prints, https://ui.adsabs.harvard.edu/abs/2014arXiv1412.6980K p. arXiv:1412.6980

    Pith/arXiv arXiv 2014

  16. [16]

    G., de Carvalho R

    La Barbera F., Ferreras I., Vazdekis A., de la Rosa I. G., de Carvalho R. R., Trevisan M., Falc \'o n-Barroso J., Ricciardelli E., 2013, @doi [ ] 10.1093/mnras/stt943 , https://ui.adsabs.harvard.edu/abs/2013MNRAS.433.3017L 433, 3017

    doi:10.1093/mnras/stt943 2013

  17. [17]

    S., Peletier R

    La Barbera F., Vazdekis A., Ferreras I., Pasquali A., Allende Prieto C., R \"o ck B., Aguado D. S., Peletier R. F., 2017, @doi [ ] 10.1093/mnras/stw2407 , https://ui.adsabs.harvard.edu/abs/2017MNRAS.464.3597L 464, 3597

    doi:10.1093/mnras/stw2407 2017

  18. [18]

    Liang Y., Melchior P., Hahn C., Shen J., Goulding A., Ward C., 2023, @doi [ ] 10.3847/2041-8213/acfa03 , https://ui.adsabs.harvard.edu/abs/2023ApJ...956L...6L 956, L6

    doi:10.3847/2041-8213/acfa03 2023

  19. [19]

    C., Acquaviva V., Thomas P

    Lovell C. C., Acquaviva V., Thomas P. A., Iyer K. G., Gawiser E., Wilkins S. M., 2019, @doi [ ] 10.1093/mnras/stz2851 , https://ui.adsabs.harvard.edu/abs/2019MNRAS.490.5503L 490, 5503

    work page doi:10.1093/mnras/stz2851 2019

  20. [20]

    M., et al., 2020, @doi [Nature Machine Intelligence] 10.1038/s42256-019-0138-9 , https://ui.adsabs.harvard.edu/abs/2020NatMI...2...56L 2, 56

    Lundberg S. M., et al., 2020, @doi [Nature Machine Intelligence] 10.1038/s42256-019-0138-9 , https://ui.adsabs.harvard.edu/abs/2020NatMI...2...56L 2, 56

    doi:10.1038/s42256-019-0138-9 2020

  21. [21]

    S., et al., 2003, @doi [ ] 10.1086/379626 , https://ui.adsabs.harvard.edu/abs/2003ApJ...599..997M 599, 997

    Madgwick D. S., et al., 2003, @doi [ ] 10.1086/379626 , https://ui.adsabs.harvard.edu/abs/2003ApJ...599..997M 599, 997

    work page doi:10.1086/379626 2003

  22. [22]

    C., Kimball A

    McGurk R. C., Kimball A. E., Ivezi \'c Z ., 2010, @doi [ ] 10.1088/0004-6256/139/3/1261 , https://ui.adsabs.harvard.edu/abs/2010AJ....139.1261M 139, 1261

    doi:10.1088/0004-6256/139/3/1261 2010

  23. [23]

    Melchior P., Liang Y., Hahn C., Goulding A., 2023, @doi [ ] 10.3847/1538-3881/ace0ff , https://ui.adsabs.harvard.edu/abs/2023AJ....166...74M 166, 74

    work page doi:10.3847/1538-3881/ace0ff 2023

  24. [24]

    V., Davletshin R

    Nikonov A. V., Davletshin R. V., Iakovleva N. I., Lazarev P. S., 2017, J. Commun. Technol. Electron., 62, 1048

    2017

  25. [25]

    Portillo S. K. N., Parejko J. K., Vergara J. R., Connolly A. J., 2020, @doi [ ] 10.3847/1538-3881/ab9644 , https://ui.adsabs.harvard.edu/abs/2020AJ....160...45P 160, 45

    doi:10.3847/1538-3881/ab9644 2020

  26. [26]

    A., Bengio Y., Courville A., 2018, @doi [arXiv e-prints] 10.48550/arXiv.1806.08734 , https://ui.adsabs.harvard.edu/abs/2018arXiv180608734R p

    Rahaman N., Baratin A., Arpit D., Draxler F., Lin M., Hamprecht F. A., Bengio Y., Courville A., 2018, @doi [arXiv e-prints] 10.48550/arXiv.1806.08734 , https://ui.adsabs.harvard.edu/abs/2018arXiv180608734R p. arXiv:1806.08734

    Pith/arXiv arXiv doi:10.48550/arxiv.1806.08734 2018

  27. [27]

    Rogers B., Ferreras I., Peletier R., Silk J., 2010, @doi [ ] 10.1111/j.1365-2966.2009.15892.x , https://ui.adsabs.harvard.edu/abs/2010MNRAS.402..447R 402, 447

    arXiv 2010

  28. [28]

    Scourfield M., Saintonge A., de Mijolla D., Viti S., 2023, @doi [ ] 10.1093/mnras/stad2709 , https://ui.adsabs.harvard.edu/abs/2023MNRAS.526.3037S 526, 3037

    doi:10.1093/mnras/stad2709 2023

  29. [29]

    Sharbaf Z., Ferreras I., Lahav O., 2023, @doi [ ] 10.1093/mnras/stad2668 , https://ui.adsabs.harvard.edu/abs/2023MNRAS.526..585S 526, 585

    doi:10.1093/mnras/stad2668 2023

  30. [30]

    D., Lahav O., Somerville R

    Sharbaf Z., Ferreras I., Negri A., Angthopo J., Vecchia C. D., Lahav O., Somerville R. S., 2025, @doi [ ] 10.1093/mnras/staf546 , https://ui.adsabs.harvard.edu/abs/2025MNRAS.539.1480S 539, 1480

    work page doi:10.1093/mnras/staf546 2025

  31. [31]

    A., et al., 2013, @doi [ ] 10.1088/0004-6256/146/2/32 , https://ui.adsabs.harvard.edu/abs/2013AJ....146...32S 146, 32

    Smee S. A., et al., 2013, @doi [ ] 10.1088/0004-6256/146/2/32 , https://ui.adsabs.harvard.edu/abs/2013AJ....146...32S 146, 32

    doi:10.1088/0004-6256/146/2/32 2013

  32. [32]

    A., et al., 2002, @doi [ ] 10.1086/342343 , https://ui.adsabs.harvard.edu/abs/2002AJ....124.1810S 124, 1810

    Strauss M. A., et al., 2002, @doi [ ] 10.1086/342343 , https://ui.adsabs.harvard.edu/abs/2002AJ....124.1810S 124, 1810

    doi:10.1086/342343 2002

  33. [33]

    C., Worthey G., Faber S

    Trager S. C., Worthey G., Faber S. M., Burstein D., Gonz \'a lez J. J., 1998, @doi [ ] 10.1086/313099 , https://ui.adsabs.harvard.edu/abs/1998ApJS..116....1T 116, 1

    doi:10.1086/313099 1998

  34. [34]

    Veit A., Wilber M., Belongie S., 2016, in Proceedings of the 30th International Conference on Neural Information Processing Systems. NIPS'16. Curran Associates Inc., Red Hook, NY, USA, p. 550–558

    2016

  35. [35]

    Walcher J., Groves B., Budav \'a ri T., Dale D., 2011, @doi [ ] 10.1007/s10509-010-0458-z , https://ui.adsabs.harvard.edu/abs/2011Ap&SS.331....1W 331, 1

    doi:10.1007/s10509-010-0458-z 2011

  36. [36]

    L., 2017, @doi [ ] 10.1093/mnras/stw2894 , https://ui.adsabs.harvard.edu/abs/2017MNRAS.465.4311W 465, 4311

    Wang K., Guo P., Luo A. L., 2017, @doi [ ] 10.1093/mnras/stw2894 , https://ui.adsabs.harvard.edu/abs/2017MNRAS.465.4311W 465, 4311

    work page doi:10.1093/mnras/stw2894 2017

  37. [37]

    C., 2005, @doi [ ] 10.1111/j.1365-2966.2005.08844.x , https://ui.adsabs.harvard.edu/abs/2005MNRAS.358.1083W 358, 1083

    Wild V., Hewett P. C., 2005, @doi [ ] 10.1111/j.1365-2966.2005.08844.x , https://ui.adsabs.harvard.edu/abs/2005MNRAS.358.1083W 358, 1083

    arXiv 2005

  38. [38]

    L., 1997, @doi [ ] 10.1086/313021 , https://ui.adsabs.harvard.edu/abs/1997ApJS..111..377W 111, 377

    Worthey G., Ottaviani D. L., 1997, @doi [ ] 10.1086/313021 , https://ui.adsabs.harvard.edu/abs/1997ApJS..111..377W 111, 377

    doi:10.1086/313021 1997

  39. [39]

    Wu Y., Tao Y., Fan D., Cui C., Zhang Y., 2024, @doi [ ] 10.1093/mnras/stad2913 , https://ui.adsabs.harvard.edu/abs/2024MNRAS.527.1163W 527, 1163

    doi:10.1093/mnras/stad2913 2024

  40. [40]

    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

    doi:10.1086/301513 2000