Pith. sign in

REVIEW 2 major objections 6 minor 67 references

A hybrid network that reads galaxy colors as a wavelength sequence cuts photometric-redshift errors by about 10% and outliers by about 20% on an LSST proxy sample.

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 · grok-4.5

2026-07-10 14:41 UTC pith:D4SC4IVY

load-bearing objection Clean, controlled ~10–20% gains over the Jones BNN on the public GalaxiesML test set with released code; useful incremental photo-z tool whose main limit is the acknowledged spectroscopic selection function. the 2 major comments →

arxiv 2607.07960 v1 pith:D4SC4IVY submitted 2026-07-08 astro-ph.GA astro-ph.IM

Enhancing Photometric Redshift Estimation for LSST with a Hybrid LSTM-Mixture Density Network

Zhijian Luo , Yangyang Li , Xinyu Luo , Hubing Xiao , Wei Fang , Shaohua Zhang , Chenggang Shu This is my paper
classification astro-ph.GA astro-ph.IM
keywords photometric redshiftsLSSTLSTMmixture density networkuncertainty quantificationoutlier rejectiongalaxy SEDsprobability integral transform
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.

Next-generation surveys such as LSST will map billions of galaxies with only a handful of broad-band colors, so photometric redshifts must be both accurate and honest about their own uncertainty. Traditional machine-learning estimators usually output a single number and often fail when colors can be produced by more than one redshift. This paper shows that treating the five optical magnitudes and their errors as a short wavelength-ordered sequence, processing them with a bidirectional LSTM, and feeding the resulting features into a mixture-density network yields better point estimates and better-calibrated full probability distributions than a Bayesian neural-network baseline trained on exactly the same data. On the GalaxiesML test set the method lowers RMSE from 0.145 to 0.130, MAE from 0.055 to 0.048, and both ordinary and catastrophic outlier rates by roughly one-fifth, while producing a nearly uniform probability-integral-transform histogram. A simple integral of each posterior, called z_conf, further lets users discard a few percent of the most ambiguous objects and cut the remaining outlier rate nearly in half. The practical claim is that a modest architectural change can deliver the high-purity redshift catalogs that precision cosmology needs.

Core claim

On the identical GalaxiesML test set used by the prior Bayesian neural-network benchmark, the LSTM-MDNz architecture improves every standard point-estimate metric by roughly 10% and reduces both ordinary and catastrophic outlier fractions by roughly 20%, while producing well-calibrated multimodal posterior PDFs whose probability-integral-transform distribution is statistically consistent with uniform.

What carries the argument

LSTM-MDNz: multi-band magnitudes and errors are arranged as a wavelength-ordered sequence of length five, passed through two bidirectional LSTM layers that extract SED-like gradients, then mapped by a mixture-density head (K=10 Gaussians) onto a full posterior PDF.

Load-bearing premise

The spectroscopically selected, bright-end-weighted GalaxiesML catalog is assumed to be a fair enough proxy that the measured gains will still hold for the much fainter, higher-redshift galaxies that will dominate LSST cosmology samples.

What would settle it

Retrain and re-evaluate the identical architecture on a deeper, purely photometric-selected sample whose magnitude and color distributions match the expected LSST weak-lensing source population; if the relative gains versus the same BNN baseline vanish or reverse, the claim fails.

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

If this is right

  • Catalogs filtered at modest z_conf thresholds can meet or exceed LSST outlier-rate requirements while retaining more than 90% of the sample.
  • Well-calibrated multimodal PDFs can be stacked or marginalized directly into n(z) estimates without the extra dispersion corrections needed for over-broad unimodal Gaussians.
  • The same sequential-plus-MDN recipe is immediately portable to any multi-band survey whose filters can be ordered by wavelength.
  • Because photometric uncertainties are ingested as explicit sequence features, the model automatically down-weights low-S/N bands without hand-tuned inverse-variance weights.

Where Pith is reading between the lines

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

  • The same architecture should transfer with little retuning to Euclid or CSST once their filter sequences replace the HSC g,r,i,z,y ladder.
  • Domain-adversarial training or SOM re-weighting of the spectroscopic training set would be the natural next step to close the bright-to-faint gap the authors themselves flag.
  • If z_conf is treated as a continuous weight rather than a hard cut, it could enter likelihoods for cosmic shear or BAO analyses without discarding objects.

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

2 major / 6 minor

Summary. The paper introduces and validates LSTM-MDNz, a hybrid architecture that treats multi-band HSC photometry (and magnitude errors) as wavelength-ordered sequences processed by stacked Bi-LSTMs, then models the redshift posterior as a 10-component Gaussian mixture via an MDN. On the identical GalaxiesML public test set used by Jones et al. (2024), the model improves point-estimate metrics (RMSE 0.145 o0.130, MAE 0.055 o0.048, σ_NMAD 0.026 o0.024) and reduces outlier and catastrophic-outlier rates by ~18–22% relative to the BNN baseline, while producing a near-uniform PIT (KS=0.014) and lower CRPS. A PDF-based confidence metric z_conf is shown to enable high-purity catalog construction by discarding a few percent of low-confidence objects. An ablation (magnitudes only) and redshift-binned diagnostics are provided.

Significance. If the reported gains hold under the stated experimental controls, the work supplies a practical, lightweight (~3.8×10^5 parameters), open-source probabilistic photo-z pipeline that improves both accuracy and calibration over a recent BNN baseline on a widely used LSST-proxy catalog. The public code/weights, fixed Zenodo test set, magnitude-only ablation, multi-metric tables, PIT/CRPS diagnostics, and z_conf purity curves are concrete strengths that make the empirical claims reproducible and useful for the community. The domain-shift caveat (spectroscopic selection vs. faint LSST depths) is acknowledged by the authors and correctly limits the cosmological claim, but does not erase the value of a carefully controlled architectural comparison on a public benchmark.

major comments (2)
  1. §2.1 and the final discussion correctly flag that GalaxiesML is bright-end weighted, incomplete at i≳23.5 and in the 1.2<z<2 desert, and therefore only a partial proxy for the faint, high-z LSST cosmological sample. The central claim of “enhancing photo-z for LSST” therefore rests on an untested transfer assumption. The manuscript should either (a) add a quantitative domain-shift test (e.g., magnitude- or color-matched faint subsample, or comparison against a deeper photometric reference such as COSMOS2020 TransferZ as in Soriano et al.), or (b) systematically soften the LSST-readiness language in the abstract, title framing, and conclusions so that the claim is scoped strictly to the controlled GalaxiesML benchmark.
  2. Table 1 and §4.2: the magnitude-only ablation already shows that sequential architecture alone is competitive with the BNN; the full model’s further gains come from feeding cmodel magnitude errors. Because the BNN baseline of Jones et al. is described as using primarily cmodel magnitudes (without the same explicit error channels), part of the reported ~10–20% improvement may be attributable to richer inputs rather than architecture. A controlled re-run of the BNN (or an MLP-MDN) with identical magnitude+error inputs would isolate the architectural contribution and should be added or explicitly discussed as a remaining ambiguity.
minor comments (6)
  1. Abstract and §1: “~10% improvement … across RMSE, MAE, scatter, and σ_NMAD” slightly overstates the scatter/σ_NMAD gains (0.026 o0.023/0.024); the precise percentages already given for RMSE/MAE/outliers are preferable.
  2. §3.1 / Eq. (2): the choice K=10 is stated without a sensitivity study; a short note or appendix showing that results are stable for K∈{5,10,15} would strengthen the free-parameter claim.
  3. §4.4: α=0.05 for the z_conf window is said to be robust over [0.03,0.15], but the supporting numbers are not shown; a one-sentence table or parenthetical would help.
  4. Figure 2 caption and text: “Relative Point Density” color scale is clear, but the exact kernel/binning used for the density map is not stated.
  5. Typographical: “server as a small-scale proxy” (abstract/intro), “magenitude error”, occasional missing spaces around z_conf thresholds, and “Ph t metric” / “Spectr sc pic” artifacts in figure labels should be cleaned.
  6. References: the prior quasar LSTM-MDNz paper (Chen et al. 2026) is appropriately cited as the methodological source; ensure the GalaxiesML and Jones et al. Zenodo DOIs remain prominent for reproducibility.

Circularity Check

0 steps flagged

No circularity: empirical ML performance claims are measured against external spectroscopic ground truth on a held-out public test set; architecture self-citations supply only the method, not the galaxy results.

full rationale

The paper's load-bearing claims are purely empirical: on the identical GalaxiesML test set (DOI 10.5281/zenodo.10145347) released by Jones et al. (2024), LSTM-MDNz yields lower RMSE/MAE/σ_NMAD, lower outlier fractions, lower CRPS, and a near-uniform PIT (KS=0.014) relative to the BNN baseline, plus an ablation (Mags Only) and z_conf purity curves. All metrics are computed against independent spectroscopic redshifts after training on a disjoint split; none of the reported numbers is forced by construction from a fitted parameter or definition. The architecture (Bi-LSTM + MDN with K=10) and the z_conf definition (integral of the PDF over a scaled window around the mean) are taken from the authors' prior quasar paper (Chen et al. 2026) and standard PDF-morphology practice, but those citations supply only the reusable method; the galaxy results, the comparison to Jones et al., and the quantitative gains are new measurements on an external benchmark. No uniqueness theorem, ansatz that embeds the target metric, or self-definitional loop appears. Domain-shift caveats are acknowledged by the authors themselves and do not constitute circularity. The derivation chain is therefore self-contained against external data.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central empirical claim rests on standard ML practice plus a handful of architectural and selection choices that are free parameters or domain assumptions; no new physical entities are postulated.

free parameters (4)
  • number of Gaussian mixture components K = 10
    Set to K=10 by hand to give enough degrees of freedom for multimodality; not derived from data or theory.
  • z_conf window half-width alpha = 0.05
    Baseline alpha=0.05 chosen to approximate intrinsic scatter; robustness claimed for [0.03,0.15] but the absolute z_conf values and thresholds scale with it.
  • Bi-LSTM hidden sizes and dropout = 128-64 / 0.25
    128/64 units per direction, dropout 0.25, dense layers 128/64 are architectural hyperparameters selected for the study.
  • learning-rate schedule and early-stopping patience
    Adam 1e-3, ReduceLROnPlateau factor 0.2 after 10 epochs, early stop after 30 epochs; standard but free choices that affect final weights.
axioms (4)
  • domain assumption Multi-band cmodel magnitudes ordered by wavelength form a sequence whose sequential correlations encode SED physics useful for redshift.
    Stated in §2.2 and §3.1; underpins the entire LSTM front-end.
  • ad hoc to paper A ten-component GMM is flexible enough to capture the relevant multi-modal and heavy-tailed photo-z posteriors.
    Chosen in §3.1 without a formal model-selection criterion beyond ‘sufficient degrees of freedom’.
  • domain assumption The GalaxiesML spectroscopic sample, despite known selection biases, is an adequate small-scale proxy for LSST-like photo-z performance.
    Explicitly adopted in §2.1 and used for all claims about LSST readiness.
  • standard math Negative log-likelihood of the GMM is the appropriate training objective for calibrated PDFs.
    Standard MDN practice; Eq. (3).
invented entities (1)
  • z_conf confidence metric independent evidence
    purpose: Scalar integral of the PDF inside a redshift-dependent window used to rank and cull low-quality photo-z estimates.
    Defined in §4.4 following earlier PDF-integral ideas but given a specific name and alpha scaling for this work; independent evidence is the empirical purity curves on the same test set.

pith-pipeline@v1.1.0-grok45 · 33753 in / 2877 out tokens · 30448 ms · 2026-07-10T14:41:25.260281+00:00 · methodology

0 comments
read the original abstract

Accurate photometric redshift (photo-$z$) estimation and robust uncertainty quantification are essential for the LSST to achieve its precision cosmology goals. Traditional machine learning algorithms are largely restricted to point estimates, struggling to characterize the multimodal nature of redshift PDFs and the degeneracies within the color-redshift space. To address this, we present and validate the LSTM-MDNz architecture, which integrates sequential feature extraction with flexible probability density modeling to enhance both prediction accuracy and uncertainty calibration across a broad redshift range, thereby meeting the stringent data quality requirements necessitated by next-generation cosmological analysis. The LSTM-MDNz framework treats multi-band photometry as wavelength-ordered sequences, utilizing LSTM networks to capture non-linear evolutionary correlations across the SED. A Mixture Density Network (MDN) is then employed to explicitly model posterior PDFs via Gaussian mixture models (GMMs). Performance is evaluated on the HSC GalaxiesML dataset (which serves as a small-scale proxy for next-generation surveys like LSST) and benchmarked against the BNN architecture established by Jones et al. (2024). The proposed model consistently outperforms the BNN baseline, achieving a $\sim 10\%$ improvement in point-estimation accuracy (specifically across RMSE, MAE, scatter, and $\sigma_{\text{NMAD}}$) and a $\sim 20\%$ reduction in the rates of both general and catastrophic outliers. A uniform probability integral transform (PIT) distribution confirms well-calibrated probabilistic outputs. Furthermore, the PDF-based confidence metric $z_{\text{conf}}$ enables high-purity catalog construction: excluding just approximately $4\%$ of extremely low-confidence ($z_{\text{conf}} < 0.05$) samples reduces the overall outlier rate by $\sim 48\%$.

Figures

Figures reproduced from arXiv: 2607.07960 by Chenggang Shu, Hubing Xiao, Shaohua Zhang, Wei Fang, Xinyu Luo, Yangyang Li, Zhijian Luo.

Figure 1
Figure 1. Figure 1: Redshift distribution of the GalaxiesML sample. The [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison between spectroscopic redshifts ( [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of CRPS distributions between the LSTM-M [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Examples of predicted photometric redshift PDFs and [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Assessment of the calibration quality of the model-p [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Top Panel: Evolution of the true (spectroscopic) red [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Bin-wise evolution of photometric redshift perform [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] 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

67 extracted references · 67 canonical work pages · 9 internal anchors

  1. [1]

    The Wide Field Infrared Survey Telescope: 100 Hubbles for the 2020s

    Akeson, R., Armus, L., Bachelet, E., et al. 2019, arXiv e-pri nts, arXiv:1902.05569

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  2. [2]

    D., Allende Prieto, C., et al

    Alam, S., Albareti, F. D., Allende Prieto, C., et al. 2015, Ap JS, 219, 12

    work page 2015

  3. [3]

    & Réfrégier, A

    Amara, A. & Réfrégier, A. 2007, MNRAS, 381, 1018

    work page 2007

  4. [4]

    1999, MNR AS, 310, 540

    Arnouts, S., Cristiani, S., Moscardini, L., et al. 1999, MNR AS, 310, 540

    work page 1999

  5. [5]

    Beck, R., Lin, C.-A., Ishida, E. E. O., et al. 2017, MNRAS, 468 , 4323

    work page 2017

  6. [6]

    Practical recommendations for gradient-based training of deep architectures

    Bengio, Y . 2012, arXiv e-prints, arXiv:1206.5533 Benítez, N. 2000, ApJ, 536, 571

    work page internal anchor Pith review Pith/arXiv arXiv 2012

  7. [7]

    2000, A&A, 363, 476

    Bolzonella, M., Miralles, J.-M., & Pelló, R. 2000, A&A, 363, 476

    work page 2000

  8. [8]

    2015, MNRAS, 449, 1043

    Bonnett, C. 2015, MNRAS, 449, 1043

    work page 2015

  9. [9]

    J., & Amara, A

    Bordoloi, R., Lilly, S. J., & Amara, A. 2010, MNRAS, 406, 881

    work page 2010

  10. [10]

    J., Almaini, O., Hartley, W

    Bradshaw, E. J., Almaini, O., Hartley, W. G., et al. 2013, MNR AS, 433, 194

    work page 2013

  11. [11]

    B., van Dokkum, P

    Brammer, G. B., van Dokkum, P . G., & Coppi, P . 2008, ApJ, 686, 1503

    work page 2008

  12. [12]

    2001, Machine Learning, 45, 5

    Breiman, L. 2001, Machine Learning, 45, 5

    work page 2001

  13. [13]

    2018, MNRAS, 480, 2178 Carrasco Kind, M

    Cao, Y ., Gong, Y ., Meng, X.-M., et al. 2018, MNRAS, 480, 2178 Carrasco Kind, M. & Brunner, R. J. 2013, MNRAS, 432, 1483

    work page 2018

  14. [14]

    2026, ApJS, 282, 46

    Chen, J., Luo, Z., Fu, L., et al. 2026, ApJS, 282, 46

    work page 2026

  15. [15]

    & Guestrin, C

    Chen, T. & Guestrin, C. 2016, arXiv e-prints, arXiv:1603.02 754

    work page 2016

  16. [16]

    L., Blanton, M

    Coil, A. L., Blanton, M. R., Burles, S. M., et al. 2011, ApJ, 74 1, 8

    work page 2011

  17. [17]

    J., Moustakas, J., Blanton, M

    Cool, R. J., Moustakas, J., Blanton, M. R., et al. 2013, ApJ, 7 67, 118

    work page 2013

  18. [18]

    C., Aird, J

    Cooper, M. C., Aird, J. A., Coil, A. L., et al. 2011, ApJS, 193, 14

    work page 2011

  19. [19]

    C., Gri ffith, R

    Cooper, M. C., Gri ffith, R. L., Newman, J. A., et al. 2012, MNRAS, 419, 3018 CSST Collaboration, Gong, Y ., Miao, H., et al. 2026, Science China Physics, Mechanics, and Astronomy, 69, 239501 Article number, page 16 Zhijian Luo et al.: Enhancing Photometric Redshift Estimat ion for LSST with a Hybrid LSTM-Mixture Density Network

    work page 2012

  20. [20]

    B., et al

    Dalmasso, N., Pospisil, T., Lee, A. B., et al. 2020, Astronom y and Computing, 30, 100362

    work page 2020

  21. [21]

    M., Newman, J., et al

    Davis, M., Faber, S. M., Newman, J., et al. 2003, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, V ol. 4834, Discover- ies and Research Prospects from 6- to 10-Meter-Class Telesc opes II, ed. P . Guhathakurta, 161–172

    work page 2003

  22. [22]

    Dawid, A. P . 1984, Journal of the Royal Statistical Society: Series A (General), 147, 278 D’Isanto, A. & Polsterer, K. L. 2018, A&A, 609, A111

    work page 1984

  23. [23]

    GalaxiesML: a dataset of galaxy images, photometry, redshifts, and structural parameters for machine learning

    Do, T., Boscoe, B., Jones, E., Li, Y . Q., & Alfaro, K. 2024, arX iv e-prints, arXiv:2410.00271

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  24. [24]

    J., Jurek, R

    Drinkwater, M. J., Jurek, R. J., Blake, C., et al. 2010, MNRAS , 401, 1429 Euclid Collaboration, Desprez, G., Paltani, S., et al. 2020 , A&A, 644, A31 Euclid Collaboration, Mellier, Y ., Abdurro’uf, et al. 2025 , A&A, 697, A1

    work page 2010

  25. [25]

    M., Porciani, C., et al

    Feldmann, R., Carollo, C. M., Porciani, C., et al. 2006, MNRA S, 372, 565

    work page 2006

  26. [26]

    & Paltani, S

    Fotopoulou, S. & Paltani, S. 2018, A&A, 619, A14

    work page 2018

  27. [27]

    2014, A&A, 562, A23

    Garilli, B., Guzzo, L., Scodeggio, M., et al. 2014, A&A, 562, A23

    work page 2014

  28. [28]

    L., Connolly, A

    Graham, M. L., Connolly, A. J., Ivezi ´c, Ž., et al. 2018, AJ, 155, 1

    work page 2018

  29. [29]

    L., Connolly, A

    Graham, M. L., Connolly, A. J., Wang, W., et al. 2020, AJ, 159, 258

    work page 2020

  30. [30]

    M., Paech, K., et al

    Hoyle, B., Rau, M. M., Paech, K., et al. 2015, MNRAS, 452, 4183

    work page 2015

  31. [31]

    Hsieh, B. C. & Y ee, H. K. C. 2014, ApJ, 792, 102

    work page 2014

  32. [32]

    2006, MNRA S, 366, 101 Ivezi´c, Ž., Kahn, S

    Huterer, D., Takada, M., Bernstein, G., & Jain, B. 2006, MNRA S, 366, 101 Ivezi´c, Ž., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111

    work page 2006

  33. [33]

    2024, ApJ, 964, 130

    Jones, E., Do, T., Boscoe, B., et al. 2024, ApJ, 964, 130

    work page 2024

  34. [34]

    & Singal, J

    Jones, E. & Singal, J. 2017, A&A, 600, A113

    work page 2017

  35. [35]

    & Singal, J

    Jones, E. & Singal, J. 2020, PASP , 132, 024501

    work page 2020

  36. [36]

    A., Kartaltepe, J

    Khostovan, A. A., Kartaltepe, J. S., Salvato, M., et al. 2026 , ApJS, 282, 6

    work page 2026

  37. [37]

    Kingma, D. P . & Ba, J. 2014, arXiv e-prints, arXiv:1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  38. [38]

    D., Miller, L., Heymans, C

    Kitching, T. D., Miller, L., Heymans, C. E., van Waerbeke, L. , & Heavens, A. F. 2008, MNRAS, 390, 149

    work page 2008

  39. [39]

    Euclid Definition Study Report

    Laureijs, R., Amiaux, J., Arduini, S., et al. 2011, arXiv e-prints, arXiv:1110.3193 Le Fèvre, O., Cassata, P ., Cucciati, O., et al. 2013, A&A, 559 , A14

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  40. [40]

    & Hogg, D

    Leistedt, B. & Hogg, D. W. 2017, ApJ, 838, 5

    work page 2017

  41. [41]

    J., Le Brun, V ., Maier, C., et al

    Lilly, S. J., Le Brun, V ., Maier, C., et al. 2009, ApJS, 184, 21 8

    work page 2009

  42. [42]

    LSST Science Book, Version 2.0

    Liske, J., Baldry, I. K., Driver, S. P ., et al. 2015, MNRAS, 45 2, 2087 LSST Science Collaboration, Abell, P . A., Allison, J., et al. 2009, arXiv e-prints, arXiv:0912.0201

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  43. [43]

    2024, MNRAS, 527, 12140

    Lu, J., Luo, Z., Chen, Z., et al. 2024, MNRAS, 527, 12140

    work page 2024

  44. [44]

    2006, ApJ, 636, 21

    Ma, Z., Hu, W., & Huterer, D. 2006, ApJ, 636, 21

    work page 2006

  45. [45]

    2018, ARA&A, 56, 393

    Mandelbaum, R. 2018, ARA&A, 56, 393

    work page 2018

  46. [46]

    J., Pearce, H

    McLure, R. J., Pearce, H. J., Dunlop, J. S., et al. 2013, MNRAS , 428, 1088

    work page 2013

  47. [47]

    G., Brammer, G

    Momcheva, I. G., Brammer, G. B., van Dokkum, P . G., et al. 2016 , ApJS, 225, 27

    work page 2016

  48. [48]

    G., Palmese, A., et al

    Mucesh, S., Hartley, W. G., Palmese, A., et al. 2021, MNRAS, 5 02, 2770

    work page 2021

  49. [49]

    A., Cooper, M

    Newman, J. A., Cooper, M. C., Davis, M., et al. 2013, ApJS, 208 , 5

    work page 2013

  50. [50]

    Newman, J. A. & Gruen, D. 2022, ARA&A, 60, 363

    work page 2022

  51. [51]

    Photometric Redshifts for the Hyper Suprime-Cam Subaru Strategic Program Data Release 2

    Nishizawa, A. J., Hsieh, B.-C., Tanaka, M., & Takata, T. 2020 , arXiv e-prints, arXiv:2003.01511 Pâris, I., Petitjean, P ., Aubourg, É., et al. 2018, A&A, 613, A51

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  52. [52]

    2019, A&A, 621, A26

    Pasquet, J., Bertin, E., Treyer, M., Arnouts, S., & Fouchez, D. 2019, A&A, 621, A26

    work page 2019

  53. [53]

    Uncertain Photometric Redshifts

    Polsterer, K. L., D’Isanto, A., & Gieseke, F. 2016, arXiv e-p rints, arXiv:1608.08016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  54. [54]

    M., Seitz, S., Brimioulle, F., et al

    Rau, M. M., Seitz, S., Brimioulle, F., et al. 2015, MNRAS, 452 , 3710

    work page 2015

  55. [55]

    B., & Lahav, O

    Sadeh, I., Abdalla, F. B., & Lahav, O. 2016, PASP , 128, 104502

    work page 2016

  56. [56]

    Schmidt, S. J. & Thorman, P . 2013, MNRAS, 431, 2766

    work page 2013

  57. [57]

    D., Kashino, D., Sanders, D., et al

    Silverman, J. D., Kashino, D., Sanders, D., et al. 2015, ApJS , 220, 12

    work page 2015

  58. [58]

    2022, ApJ, 928, 6

    Singal, J., Silverman, G., Jones, E., et al. 2022, ApJ, 928, 6

    work page 2022

  59. [59]

    E., Whitaker, K

    Skelton, R. E., Whitaker, K. E., Momcheva, I. G., et al. 2014, ApJS, 214, 24

    work page 2014

  60. [60]

    2026, AJ, 171, 11 4

    Soriano, J., Do, T., Saikrishnan, S., et al. 2026, AJ, 171, 11 4

    work page 2026

  61. [61]

    Using different sources of ground truths and transfer learning to improve the generalization of photometric redshift estimation

    Soriano, J., Saikrishnan, S., Seenivasan, V ., et al. 2024, a rXiv e-prints, arXiv:2411.18054

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  62. [62]

    2018, PASJ, 70, S 9

    Tanaka, M., Coupon, J., Hsieh, B.-C., et al. 2018, PASJ, 70, S 9

    work page 2018

  63. [63]

    2016, MNRAS, 457, 4005

    Wittman, D., Bhaskar, R., & Tobin, R. 2016, MNRAS, 457, 4005

    work page 2016

  64. [64]

    & Singal, J

    Wyatt, M. & Singal, J. 2021, PASP , 133, 044504 Y oo, J., Gyure, C., Agarwal, V ., Singal, J., & Silverman, G. 2026, ApJ, 998, 258

    work page 2021

  65. [65]

    2011, Scientia Sinica Physica, Mechanica & Astrono mica, 41, 1441

    Zhan, H. 2011, Scientia Sinica Physica, Mechanica & Astrono mica, 41, 1441

    work page 2011

  66. [66]

    2022, Research in Astr onomy and As- trophysics, 22, 115017

    Zhou, X., Gong, Y ., Meng, X.-M., et al. 2022, Research in Astr onomy and As- trophysics, 22, 115017

    work page 2022

  67. [67]

    2021, ApJ, 909, 53 Article number, page 17

    Zhou, X., Gong, Y ., Meng, X.-M., et al. 2021, ApJ, 909, 53 Article number, page 17

    work page 2021