arxiv: 2602.10207 · v2 · submitted 2026-02-10 · 🌌 astro-ph.IM · astro-ph.GA

Recognition: no theorem link

Optimizing Deep Learning Photometric Redshifts for the Roman Space Telescope with HST/CANDELS

Ashod Khederlarian , Brett H. Andrews , Jeffrey A. Newman , Tianqing Zhang , Biprateep Dey

Authors on Pith no claims yet

Pith reviewed 2026-05-16 02:21 UTC · model grok-4.3

classification 🌌 astro-ph.IM astro-ph.GA

keywords photometric redshiftsdeep learningsemi-supervised learningRoman Space TelescopeHST CANDELSlatent spacegalaxy distances

0 comments

The pith

A new semi-supervised model PITA outperforms other methods for photometric redshifts by training on both labeled redshifts and all available images and colors.

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

The paper evaluates deep learning approaches to estimating galaxy distances from images on Hubble data designed to stand in for future Roman Space Telescope observations out to redshift three. It introduces PITA, a semi-supervised algorithm whose three-part loss function lets the model learn useful patterns from the much larger set of unlabeled images and colors while using the smaller set of objects that do have known redshifts. This produces a smoothly varying representation in color and redshift space that delivers the best accuracy of all tested methods, including fully supervised deep networks, self-supervised models, template fitting, and classical machine learning on photometry. A reader would care because Roman will deliver hundreds of millions of high-resolution images but only limited spectroscopic redshifts, so any method that extracts more signal from the unlabeled majority directly improves studies of galaxy evolution and large-scale structure.

Core claim

PITA (Photo-z Inference with a Triple-task Algorithm) outperforms template-based, classical machine-learning, fully-supervised, and self-supervised photo-z methods on HST/CANDELS imaging. It does so by training with a three-part loss that incorporates images and colors for every object and redshifts when they are available, resulting in a latent space that varies smoothly with magnitude, color, and redshift and maintains high accuracy even when the labeled training set is substantially reduced.

What carries the argument

PITA's three-part loss function that jointly optimizes image reconstruction, color reconstruction, and redshift prediction on both labeled and unlabeled data to enforce smoothness in the learned latent space.

If this is right

Semi-supervised deep learning extracts useful information from the hundreds of millions of Roman images and colors that lack redshift labels.
PITA maintains superior accuracy even when the spectroscopic training set is cut by a large factor.
Self-supervised training alone produces a latent space with large color and redshift fluctuations that degrade photo-z performance.
Both fully-supervised and semi-supervised deep networks beat traditional template and photometry-based methods on space-based imaging.

Where Pith is reading between the lines

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

The same triple-task structure could be adapted to other sparse-label inference problems in astronomy such as morphological classification or transient detection.
Testing PITA on Roman-specific image simulations with realistic noise would be the clearest next step to confirm generalization.
If the smoothness property proves robust, PITA-style models might reduce the number of spectroscopic follow-up observations needed for large surveys.

Load-bearing premise

That gains measured on HST/CANDELS data will carry over to the noise, depth, and resolution properties of actual Roman Space Telescope observations without domain shift or overfitting.

What would settle it

Direct comparison of PITA's photo-z scatter and outlier fraction against competing methods on a set of simulated Roman images whose true redshifts are known.

Figures

Figures reproduced from arXiv: 2602.10207 by Ashod Khederlarian, Biprateep Dey, Brett H. Andrews, Jeffrey A. Newman, Tianqing Zhang.

**Figure 1.** Figure 1: Total system throughputs for HST/CANDELS filters (black) and effective areas for Roman ZY JHFK filters (colored). The Roman HLWAS imaging will include all these bands in the deep tier, but only Y JH in the medium tier. These bands can be combined with Rubin ugrizy optical photometry to obtain the best possible photo-z’s for Roman. In this context, the CANDELS F125W and F160W filters provide near-infrared w… view at source ↗

**Figure 2.** Figure 2: shows example cutout images of objects from CANDELS. The high resolutions of the ACS (∼ 0 ′′ .11 FWHM) and WFC3 (∼ 0 ′′ .2 FWHM) observations provide spatially-resolved images of galaxies with good signal-to-noise ratio (S/N) even for galaxies with faint magnitudes (mF160W ∼ 25) and higher redshifts (z > 1) than those used for previous tests of deep learning photo-z’s. The resolution of Roman imaging wil… view at source ↗

**Figure 3.** Figure 3: Left: Redshift distributions of the labeled samples (20,624 galaxies) split by spec-z’s, grism-z’s, and COSMOS2020 many-band photo-z’s. Right: mF160W distributions of the same labeled samples, in addition to the unlabeled photometric sample (78, 881 galaxies). The photometric sample is deeper than most of the labeled data. validation (15%), and test (15%) sets. The images and photometry for the remaining 7… view at source ↗

**Figure 4.** Figure 4: Architecture of the fully-supervised redshift prediction model. A ConvNeXt CNN takes as input a four-band image and outputs a 1000-dimensional feature vector, which is then passed to an MLP that predicts a scalar redshift. This network is trained exclusively on galaxies with redshift labels [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

**Figure 5.** Figure 5: Illustration of the data augmentations used when training the fully-supervised algorithm. Starting with a 108 × 108 pixel image, we apply a horizontal flip with 50% probability, followed by a random rotation. Then, we jitter the image and crop to the central 64 × 64 pixels. Lastly, we add uncorrelated Gaussian noise. This figure illustrates the effects of the transformations on an RGB composite image, but … view at source ↗

**Figure 6.** Figure 6: illustrates how the transformations described in section 3.2 are used to construct multiple views of the same galaxy, which serve as the positive pairs. Choosing the correct set of transformations is crucial for obtaining a latent space appropriate for the downstream photo-z prediction task; it is important to avoid any augmentations that can potentially alter redshift information [PITH_FULL_IMAGE:figure… view at source ↗

**Figure 7.** Figure 7: Architecture of the self-supervised contrastive learning algorithm. An image encoder, consisting of a ConvNeXt CNN followed by an MLP, maps four-band galaxy images into a 128-dimensional latent space. A projection head then maps these latent vectors to a 64-dimensional space where the contrastive loss is calculated. This network is initially trained on unlabeled data with the objective of learning image re… view at source ↗

**Figure 8.** Figure 8: Architecture of PITA. Input images are processed by an encoder consisting of a ConvNeXt CNN followed by an MLP to obtain 128-dimensional latent vectors. Three separate task-specific MLPs further project these representations: (1) a projection head which reduces the dimensionality for contrastive learning; (2) a color MLP used to predict photometric colors for all objects; and (3) a redshift MLP for predict… view at source ↗

**Figure 9.** Figure 9: Predicted vs. reference redshifts for galaxies in the test set. The top row shows predictions from photometry-MLP on the y-axis, while the bottom row shows predictions from PITA. The left column shows the 800 galaxies with mF160W < 22 (out to z ∼ 2.5), and the right column shows the 2,911 galaxies with mF160W < 25 (out to z ∼ 4). Compared to the photometric baseline, the semi-supervised method reduces bias… view at source ↗

**Figure 10.** Figure 10: Two-dimensional UMAP embeddings of the self-supervised latent space. In the top panel, we divide the UMAP-derived space into a 30 × 30 grid, select a random galaxy from each grid cell, and display its image cutout. Galaxies in adjacent grid cells exhibit similar shapes and brightnesses but not necessarily similar colors. The middle panels show the variations of mF160W, F125W − F160W color, and zref across… view at source ↗

**Figure 11.** Figure 11: Similar to [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗

**Figure 12.** Figure 12: RGB images of four galaxies from the test set and their nearest neighbors in photometry space (left) or in the PITA latent space (right). The four images with colored square boxes are galaxies that have catastrophically wrong photo-z predictions from photometry (|∆z| > 0.15) but correct predictions from PITA (|∆z| < 0.15). The first three rows suggest that PITA leverages morphological information to break… view at source ↗

**Figure 13.** Figure 13: Scaling of performance metrics (bias, σNMAD, and foutlier) as a function of the number of labeled objects used in training the photometry-MLP, fully-supervised, and PITA models, and in fine-tuning the self-supervised model. The metrics are evaluated on test set sources with mF160W < 22 (left) or with mF160W < 25 (right). For next-generation imaging surveys (including Roman), the number of redshift labels … view at source ↗

read the original abstract

Photometric redshifts (photo-$z$'s) will be crucial for studies of galaxy evolution, large-scale structure, and transients with the Nancy Grace Roman Space Telescope. Deep learning methods leverage pixel-level information from ground-based images to achieve the best photo-$z$'s for low-redshift galaxies, but their efficacy at higher redshifts with deep, space-based imaging remains largely untested. We used Hubble Space Telescope CANDELS optical and near-infrared imaging to evaluate fully-supervised, self-supervised, and semi-supervised deep learning photo-$z$ algorithms out to $z\sim3$. Compared to template-based and classical machine learning photometry methods, the fully-supervised and semi-supervised models achieved better performance. Our new semi-supervised model, PITA (Photo-$z$ Inference with a Triple-task Algorithm), outperformed all others by learning from unlabeled and labeled data through a three-part loss function that incorporates images and colors for all objects as well as redshifts when available. PITA produces a latent space that varies smoothly in magnitude, color, and redshift, resulting in the best photo-$z$ performance even when the redshift training set was significantly reduced. In contrast, the self-supervised approach produced a latent space with significant color and redshift fluctuations that hindered photo-$z$ inference. Looking forward to Roman, we recommend using semi supervised deep learning to take full advantage of the information contained in the hundreds of millions of high-resolution images and color measurements, together with the limited redshift measurements available, to achieve the most accurate photo-$z$ estimates for both faint and bright sources.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

PITA improves photo-z on CANDELS with a triple loss and smooth latent space, but the Roman generalization step is untested.

read the letter

PITA is the concrete new piece: a semi-supervised network that adds image and color terms for every object to the usual redshift loss, and it beats the fully-supervised and self-supervised baselines on CANDELS even when labels are thinned out. The latent-space plots show why—the embedding stays smooth in magnitude, color, and redshift, whereas the self-supervised version develops fluctuations that degrade the photo-z output. That diagnostic is useful and not just another accuracy table. The comparison to template fitting and classical ML is also straightforward and shows the expected DL advantage at z up to 3. The reduced-label experiments are the most practical part; they directly address the scarcity of spectroscopy that Roman will face. The soft spot is the jump to Roman data. CANDELS has its own filter set, depth, and PSF; Roman’s wider field, different near-IR bands, and noise properties are not simulated or cross-tested. The paper therefore cannot show whether the triple-loss advantage or the smoothness property survives the domain shift. That assumption is left implicit. The rest of the work is standard empirical ML—no hidden fitting, no invented parameters, and the baselines are drawn from the literature. A reader who needs to pick a photo-z method for limited-label space imaging will find the comparisons and the latent-space check worth the time. The paper is coherent on its own terms and the application is timely, so it should go to peer review rather than a desk reject.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces PITA, a new semi-supervised deep learning model for photometric redshift estimation that employs a triple-task loss function incorporating images and colors for all objects plus redshifts when available. Using HST/CANDELS optical/NIR imaging, it compares fully-supervised, self-supervised, and semi-supervised approaches against template-based and classical ML photometry methods, claiming that PITA achieves the best performance up to z~3, produces a smoother latent space in magnitude/color/redshift, and maintains accuracy even with substantially reduced labeled training data. The authors recommend semi-supervised deep learning for the Roman Space Telescope to exploit its hundreds of millions of high-resolution images alongside limited spectroscopic redshifts.

Significance. If the empirical results on CANDELS hold, the work is significant for Roman photo-z applications because it demonstrates that semi-supervised methods can leverage abundant unlabeled imaging data to improve accuracy and robustness when spectroscopic labels are sparse. The explicit comparison of supervision regimes and the emphasis on latent-space smoothness provide concrete guidance for survey pipelines that must handle both faint and bright sources.

major comments (2)

[Abstract and concluding section] Abstract and final paragraph: the forward recommendation for Roman rests on the untested assumption that performance gains and latent-space smoothness observed on CANDELS will survive the shift to Roman's imaging properties (different NIR filter set, wider field, distinct PSF and noise). No domain-adaptation experiment, filter-convolution test, or simulated Roman catalog is described, making this a load-bearing claim for the paper's primary application.
[Results section] Results section (performance tables and reduced-label experiments): the abstract asserts that PITA 'outperformed all others' and remained best 'even when the redshift training set was significantly reduced,' yet the provided text supplies no quantitative metrics (bias, σ, outlier fraction, R²), error bars, or exact label fractions used in the ablation. Without these numbers the magnitude and statistical significance of the claimed gains cannot be assessed.

minor comments (2)

[Methods] The three-part loss function is described only at a high level; an explicit equation (or pseudocode) showing the weighting between the image reconstruction, color, and redshift terms would improve reproducibility.
[Results] Latent-space visualizations should include quantitative measures of smoothness (e.g., gradient norms or correlation lengths in redshift) rather than relying solely on qualitative inspection.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback on our manuscript. We have addressed each of the major comments in detail below and made revisions to the manuscript accordingly.

read point-by-point responses

Referee: [Abstract and concluding section] Abstract and final paragraph: the forward recommendation for Roman rests on the untested assumption that performance gains and latent-space smoothness observed on CANDELS will survive the shift to Roman's imaging properties (different NIR filter set, wider field, distinct PSF and noise). No domain-adaptation experiment, filter-convolution test, or simulated Roman catalog is described, making this a load-bearing claim for the paper's primary application.

Authors: We agree that the performance on CANDELS does not automatically guarantee identical gains for Roman due to differences in imaging characteristics. CANDELS serves as the closest current analog to Roman's space-based NIR imaging, allowing us to test the semi-supervised approach in a relevant regime. In the revised manuscript, we have added a new subsection in the discussion that explicitly compares the properties of CANDELS and Roman, highlights potential differences, and includes stronger caveats on the extrapolation. We have revised the abstract and concluding section to frame our recommendation as a suggestion for future pipelines based on the demonstrated advantages, rather than a definitive claim. We note that full validation would require Roman-specific simulations, which we propose as future work. revision: partial
Referee: [Results section] Results section (performance tables and reduced-label experiments): the abstract asserts that PITA 'outperformed all others' and remained best 'even when the redshift training set was significantly reduced,' yet the provided text supplies no quantitative metrics (bias, σ, outlier fraction, R²), error bars, or exact label fractions used in the ablation. Without these numbers the magnitude and statistical significance of the claimed gains cannot be assessed.

Authors: The referee is correct that the main text as presented lacks explicit numerical values for the performance metrics and exact label fractions in the narrative, although the tables contain them. To address this, we have revised the results section and abstract to include specific quantitative summaries, such as the values of bias, scatter, outlier rates, and R² for key comparisons, along with the label fractions tested (10%, 25%, 50%, and 100%). Error bars from bootstrapping or cross-validation are now referenced in the text. These changes make the magnitude of the improvements clear without altering the underlying results. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical ML model comparison on external CANDELS benchmark

full rationale

The paper presents an empirical comparison of fully-supervised, self-supervised, and semi-supervised deep learning models for photometric redshift estimation on HST/CANDELS imaging data. PITA is introduced as a new triple-task semi-supervised architecture whose performance is measured directly against baselines on held-out labeled data; no mathematical derivation, uniqueness theorem, or fitted parameter is claimed to predict the target metric. All reported gains are statistical outcomes of training and evaluation on the same external dataset, with no self-citation chain or ansatz that reduces the central claim to its own inputs by construction. Generalization to Roman is stated as a forward recommendation rather than a derived result.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on standard deep learning assumptions that neural networks can learn useful representations from images and that a combined loss on images, colors, and sparse labels produces a useful latent space. No specific free parameters, axioms, or invented entities are detailed in the abstract.

pith-pipeline@v0.9.0 · 5604 in / 1059 out tokens · 135960 ms · 2026-05-16T02:21:02.372234+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

226 extracted references · 226 canonical work pages · 13 internal anchors

[1]

2021, arXiv e-prints, arXiv:2101.04293, doi: 10.48550/arXiv.2101.04293

Mustafa, M. 2021, arXiv e-prints, arXiv:2101.04293, doi: 10.48550/arXiv.2101.04293

work page doi:10.48550/arxiv.2101.04293 2021
[2]

2020, ApJS, 249, 3, doi: 10.3847/1538-4365/ab929e Ait Ouahmed, R., Arnouts, S., Pasquet, J., Treyer, M., &

Ahumada, R., Allende Prieto, C., Almeida, A., et al. 2020, ApJS, 249, 3, doi: 10.3847/1538-4365/ab929e Ait Ouahmed, R., Arnouts, S., Pasquet, J., Treyer, M., &

work page doi:10.3847/1538-4365/ab929e 2020
[3]

2024, A&A, 683, A26, doi: 10.1051/0004-6361/202347395

Bertin, E. 2024, A&A, 683, A26, doi: 10.1051/0004-6361/202347395

work page doi:10.1051/0004-6361/202347395 2024
[4]

B., Casey, C

Akins, H. B., Casey, C. M., Berg, D. A., et al. 2025a, ApJL, 980, L29, doi: 10.3847/2041-8213/adab76

work page doi:10.3847/2041-8213/adab76 2041
[5]

B., Casey, C

Akins, H. B., Casey, C. M., Lambrides, E., et al. 2025b, ApJ, 991, 37, doi: 10.3847/1538-4357/ade984

work page doi:10.3847/1538-4357/ade984
[6]

G., et al

Akiyama, M., Ueda, Y., Watson, M. G., et al. 2015, PASJ, 67, 82, doi: 10.1093/pasj/psv050

work page doi:10.1093/pasj/psv050 2015
[7]

Report of the Dark Energy Task Force

Albrecht, A., Bernstein, G., Cahn, R., et al. 2006, arXiv e-prints, astro, doi: 10.48550/arXiv.astro-ph/0609591

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.astro-ph/0609591 2006
[8]

A., Jarvis, M

Almosallam, I. A., Jarvis, M. J., & Roberts, S. J. 2016, MNRAS, 462, 726, doi: 10.1093/mnras/stw1618

work page doi:10.1093/mnras/stw1618 2016
[9]

2025, ApJ, 978, 90, doi: 10.3847/1538-4357/ad8b30

Antwi-Danso, J., Papovich, C., Esdaile, J., et al. 2025, ApJ, 978, 90, doi: 10.3847/1538-4357/ad8b30

work page doi:10.3847/1538-4357/ad8b30 2025
[11]

1999, , 302, 735, 10.1046/j.1365-8711.1999.02174.x

Arnouts, S., Cristiani, S., Moscardini, L., et al. 1999, Monthly Notices of the Royal Astronomical Society, 310, 540, doi: 10.1046/j.1365-8711.1999.02978.x

work page doi:10.1046/j.1365-8711.1999.02978.x 1999
[12]

A., Andrews, B

Ashmead, F., Newman, J. A., Andrews, B. H., et al. 2025, arXiv e-prints, arXiv:2512.09032, doi: 10.48550/arXiv.2512.09032 Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33, doi: 10.1051/0004-6361/201322068 Astropy Collaboration, Price-Whelan, A. M., Sip˝ ocz, B. M., et al. 2018, AJ, 156, 123, doi: 10.3847/1538-3881/aabc...

work page doi:10.48550/arxiv.2512.09032 2025
[13]

2010, A&A, 512, A12, doi: 10.1051/0004-6361/200913626

Balestra, I., Mainieri, V., Popesso, P., et al. 2010, A&A, 512, A12, doi: 10.1051/0004-6361/200913626

work page doi:10.1051/0004-6361/200913626 2010
[14]

A., et al

Balogh, M. L., McGee, S. L., Wilman, D. J., et al. 2011, MNRAS, 412, 2303, doi: 10.1111/j.1365-2966.2010.18052.x

work page doi:10.1111/j.1365-2966.2010.18052.x 2011
[15]

L., van der Burg, R

Balogh, M. L., van der Burg, R. F. J., Muzzin, A., et al. 2021, MNRAS, 500, 358, doi: 10.1093/mnras/staa3008

work page doi:10.1093/mnras/staa3008 2021
[16]

J., Cowie, L

Barger, A. J., Cowie, L. L., & Wang, W.-H. 2008, ApJ, 689, 687, doi: 10.1086/592735

work page doi:10.1086/592735 2008
[17]

G., Cava, A., et al

Barro, G., P´ erez-Gonz´ alez, P. G., Cava, A., et al. 2019, ApJS, 243, 22, doi: 10.3847/1538-4365/ab23f2

work page doi:10.3847/1538-4365/ab23f2 2019
[18]

S., & Csabai, I

Beck, R., Dobos, L., Budav´ ari, T., Szalay, A. S., & Csabai, I. 2016, MNRAS, 460, 1371, doi: 10.1093/mnras/stw1009 Ben´ ıtez, N. 2000, ApJ, 536, 571, doi: 10.1086/308947

work page doi:10.1093/mnras/stw1009 2016
[19]

& Arnouts , S

Bertin, E., & Arnouts, S. 1996, A&AS, 117, 393, doi: 10.1051/aas:1996164

work page doi:10.1051/aas:1996164 1996
[20]

A., Clowe, D., Collett, T

Blazek, J. A., Clowe, D., Collett, T. E., et al. 2022, arXiv e-prints, arXiv:2204.01992, doi: 10.48550/arXiv.2204.01992

work page doi:10.48550/arxiv.2204.01992 2022
[21]

On the Opportunities and Risks of Foundation Models

Bommasani, R., Hudson, D. A., Adeli, E., et al. 2021, arXiv e-prints, arXiv:2108.07258, doi: 10.48550/arXiv.2108.07258

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2108.07258 2021
[22]

J., Smit, R., Schouws, S., et al

Bouwens, R. J., Smit, R., Schouws, S., et al. 2022, ApJ, 931, 160, doi: 10.3847/1538-4357/ac5a4a

work page doi:10.3847/1538-4357/ac5a4a 2022
[23]

J., Almaini, O., Hartley, W

Bradshaw, E. J., Almaini, O., Hartley, W. G., et al. 2013, MNRAS, 433, 194, doi: 10.1093/mnras/stt715

work page doi:10.1093/mnras/stt715 2013
[24]

B., van Dokkum, P

Brammer, G. B., van Dokkum, P. G., & Coppi, P. 2008, ApJ, 686, 1503, doi: 10.1086/591786

work page internal anchor Pith review doi:10.1086/591786 2008
[25]

B., van Dokkum, P

Brammer, G. B., van Dokkum, P. G., Franx, M., et al. 2012, ApJS, 200, 13, doi: 10.1088/0067-0049/200/2/13

work page doi:10.1088/0067-0049/200/2/13 2012
[26]

R., Sanders, D

Brinch, M., Greve, T. R., Sanders, D. B., et al. 2024, MNRAS, 527, 6591, doi: 10.1093/mnras/stad3409

work page doi:10.1093/mnras/stad3409 2024
[27]

2025, A&A, 694, A218, doi: 10.1051/0004-6361/202451448

Brinch, M., Jin, S., Gobat, R., et al. 2025, A&A, 694, A218, doi: 10.1051/0004-6361/202451448

work page doi:10.1051/0004-6361/202451448 2025
[28]

2015, MNRAS, 446, 2394, doi: 10.1093/mnras/stu2117 Calabr` o, A., Daddi, E., Cassata, P., et al

Brusa, M., Bongiorno, A., Cresci, G., et al. 2015, MNRAS, 446, 2394, doi: 10.1093/mnras/stu2117 Calabr` o, A., Daddi, E., Cassata, P., et al. 2018, ApJL, 862, L22, doi: 10.3847/2041-8213/aad33e

work page doi:10.1093/mnras/stu2117 2015
[29]

M., Berta, S., B´ ethermin, M., et al

Casey, C. M., Berta, S., B´ ethermin, M., et al. 2012, ApJ, 761, 140, doi: 10.1088/0004-637X/761/2/140

work page doi:10.1088/0004-637x/761/2/140 2012
[30]

M., Cooray, A., Capak, P., et al

Casey, C. M., Cooray, A., Capak, P., et al. 2015, ApJL, 808, L33, doi: 10.1088/2041-8205/808/2/L33

work page doi:10.1088/2041-8205/808/2/l33 2015
[31]

M., Cooray, A., Killi, M., et al

Casey, C. M., Cooray, A., Killi, M., et al. 2017, ApJ, 840, 101, doi: 10.3847/1538-4357/aa6cb1

work page doi:10.3847/1538-4357/aa6cb1 2017
[32]

M., Zavala, J

Casey, C. M., Zavala, J. A., Aravena, M., et al. 2019, ApJ, 887, 55, doi: 10.3847/1538-4357/ab52ff

work page doi:10.3847/1538-4357/ab52ff 2019
[33]

B., Casey, C

Champagne, J. B., Casey, C. M., Zavala, J. A., et al. 2021, ApJ, 913, 110, doi: 10.3847/1538-4357/abf4e6

work page doi:10.3847/1538-4357/abf4e6 2021
[34]

2022, ApJ, 929, 159, doi: 10.3847/1538-4357/ac61df

Chen, C.-C., Liao, C.-L., Smail, I., et al. 2022, ApJ, 929, 159, doi: 10.3847/1538-4357/ac61df

work page doi:10.3847/1538-4357/ac61df 2022
[35]

2020, in International conference on machine learning, PMLR, 1597–1607

Chen, T., Kornblith, S., Norouzi, M., & Hinton, G. 2020, in International conference on machine learning, PMLR, 1597–1607

work page 2020
[36]

2021, Advances in Neural Information Processing Systems, 34, 11834

Chen, T., Luo, C., & Li, L. 2021, Advances in Neural Information Processing Systems, 34, 11834

work page 2021
[37]

L., Blanton, M

Coil, A. L., Blanton, M. R., Burles, S. M., et al. 2011, ApJ, 741, 8, doi: 10.1088/0004-637X/741/1/8

work page doi:10.1088/0004-637x/741/1/8 2011
[38]

2001, MNRAS, 322, 231, doi: 10.1046/j.1365-8711.2001.04022.x

Colless, M., Dalton, G., Maddox, S., et al. 2001, MNRAS, 328, 1039, doi: 10.1046/j.1365-8711.2001.04902.x

work page doi:10.1046/j.1365-8711.2001.04902.x 2001
[39]

2008, http://h5py.alfven.org

Collette, A. 2008, http://h5py.alfven.org

work page 2008
[40]

2015, A&A, 575, A40, doi: 10.1051/0004-6361/201424767 26Khederlarian et al

Comparat, J., Richard, J., Kneib, J.-P., et al. 2015, A&A, 575, A40, doi: 10.1051/0004-6361/201424767 26Khederlarian et al

work page doi:10.1051/0004-6361/201424767 2015
[41]

2013, ARA&A, 51, 393, doi: 10.1146/annurev-astro-082812-141017

Conroy, C. 2013, ARA&A, 51, 393, doi: 10.1146/annurev-astro-082812-141017

work page internal anchor Pith review doi:10.1146/annurev-astro-082812-141017 2013
[42]

C., Aird, J

Cooper, M. C., Aird, J. A., Coil, A. L., et al. 2011, ApJS, 193, 14, doi: 10.1088/0067-0049/193/1/14

work page doi:10.1088/0067-0049/193/1/14 2011
[43]

S., & Philippov , A

Cooper, M. C., Yan, R., Dickinson, M., et al. 2012, MNRAS, 425, 2116, doi: 10.1111/j.1365-2966.2012.21524.x

work page doi:10.1111/j.1365-2966.2012.21524.x 2012
[44]

R., Casey, C

Cooper, O. R., Casey, C. M., Akins, H. B., et al. 2024, ApJ, 970, 50, doi: 10.3847/1538-4357/ad4c6c

work page doi:10.3847/1538-4357/ad4c6c 2024
[45]

F., Kalmbach, J

Crenshaw, J. F., Kalmbach, J. B., Gagliano, A., et al. 2024, AJ, 168, 80, doi: 10.3847/1538-3881/ad54bf

work page doi:10.3847/1538-3881/ad54bf 2024
[46]

2004, ApJ, 617, 746, doi: 10.1086/425569

Daddi, E., Cimatti, A., Renzini, A., et al. 2004, ApJ, 617, 746, doi: 10.1086/425569

work page doi:10.1086/425569 2004
[47]

M., et al

Daddi, E., Valentino, F., Rich, R. M., et al. 2021, A&A, 649, A78, doi: 10.1051/0004-6361/202038700

work page doi:10.1051/0004-6361/202038700 2021
[48]

J., Geller, M

Damjanov, I., Zahid, H. J., Geller, M. J., Fabricant, D. G., & Hwang, H. S. 2018, ApJS, 234, 21, doi: 10.3847/1538-4365/aaa01c

work page doi:10.3847/1538-4365/aaa01c 2018
[49]

Z., Martin, C., et al

Darvish, B., Scoville, N. Z., Martin, C., et al. 2020, ApJ, 892, 8, doi: 10.3847/1538-4357/ab75c3

work page doi:10.3847/1538-4357/ab75c3 2020
[50]

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, Vol. 4834, Discoveries and Research Prospects from 6- to 10-Meter-Class Telescopes II, ed. P. Guhathakurta, 161–172, doi: 10.1117/12.457897

work page doi:10.1117/12.457897 2003
[51]

Data Release 1 of the Dark Energy Spectroscopic Instrument

Davis, M., Guhathakurta, P., Konidaris, N. P., et al. 2007, ApJL, 660, L1, doi: 10.1086/517931 DESI Collaboration, Adame, A. G., Aguilar, J., et al. 2024, AJ, 168, 58, doi: 10.3847/1538-3881/ad3217 DESI Collaboration, Abdul-Karim, M., Adame, A. G., et al. 2025, arXiv e-prints, arXiv:2503.14745, doi: 10.48550/arXiv.2503.14745 D’Eugenio, C., Daddi, E., Goba...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1086/517931 2007
[52]

H., Newman, J

Dey, B., Andrews, B. H., Newman, J. A., et al. 2022, MNRAS, 515, 5285, doi: 10.1093/mnras/stac2105

work page doi:10.1093/mnras/stac2105 2022
[53]

J., Ledoux, C., et al

Diener, C., Lilly, S. J., Ledoux, C., et al. 2015, ApJ, 802, 31, doi: 10.1088/0004-637X/802/1/31 Dolatpour Fathkouhi, A., & Fox, G. C. 2024, arXiv e-prints, arXiv:2409.01825, doi: 10.48550/arXiv.2409.01825

work page doi:10.1088/0004-637x/802/1/31 2015
[54]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Dosovitskiy, A. 2020, arXiv preprint arXiv:2010.11929

work page internal anchor Pith review Pith/arXiv arXiv 2020
[55]

Endsley, R., & Stark, D. P. 2022, MNRAS, 511, 6042, doi: 10.1093/mnras/stac524

work page doi:10.1093/mnras/stac524 2022
[56]

2024, A&A, 683, A205, doi: 10.1051/0004-6361/202348038 Euclid Collaboration, Desprez, G., Paltani, S., et al

Epinat, B., Contini, T., Mercier, W., et al. 2024, A&A, 683, A205, doi: 10.1051/0004-6361/202348038 Euclid Collaboration, Desprez, G., Paltani, S., et al. 2020, A&A, 644, A31, doi: 10.1051/0004-6361/202039403 Euclid Collaboration, Saglia, R., De Nicola, S., et al. 2022, A&A, 664, A196, doi: 10.1051/0004-6361/202243604 Euclid Collaboration, Siudek, M., Hue...

work page doi:10.1051/0004-6361/202348038 2024
[57]

M., Phillips, A

Faber, S. M., Phillips, A. C., Kibrick, R. I., et al. 2003, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 4841, Instrument Design and Performance for Optical/Infrared Ground-based Telescopes, ed. M. Iye & A. F. M. Moorwood, 1657–1669, doi: 10.1117/12.460346

work page doi:10.1117/12.460346 2003
[58]

2019,, 1.4 doi: 10.5281/zenodo.3828935

Falcon, W., & The PyTorch Lightning team. 2019,, 1.4 doi: 10.5281/zenodo.3828935

work page doi:10.5281/zenodo.3828935 2019
[59]

2011, A&A, 529, A72, doi: 10.1051/0004-6361/200913498

Faure, C., Anguita, T., Alloin, D., et al. 2011, A&A, 529, A72, doi: 10.1051/0004-6361/200913498

work page doi:10.1051/0004-6361/200913498 2011
[60]

B., Edwards , R

Feldmann, R., Carollo, C. M., Porciani, C., et al. 2006, MNRAS, 372, 565, doi: 10.1111/j.1365-2966.2006.10930.x

work page doi:10.1111/j.1365-2966.2006.10930.x 2006
[61]

C., Annunziatella, M., et al

Forrest, B., Marsan, Z. C., Annunziatella, M., et al. 2020, ApJ, 903, 47, doi: 10.3847/1538-4357/abb819

work page doi:10.3847/1538-4357/abb819 2020
[62]

2022, ApJ, 938, 109, doi: 10.3847/1538-4357/ac8747

Forrest, B., Wilson, G., Muzzin, A., et al. 2022, ApJ, 938, 109, doi: 10.3847/1538-4357/ac8747

work page doi:10.3847/1538-4357/ac8747 2022
[63]

C., Muzzin, A., et al

Forrest, B., Cooper, M. C., Muzzin, A., et al. 2024, ApJ, 977, 51, doi: 10.3847/1538-4357/ad8b1c

work page doi:10.3847/1538-4357/ad8b1c 2024
[64]

C., et al

Forrest, B., Shen, L., Lemaux, B. C., et al. 2025, ApJ, 985, 61, doi: 10.3847/1538-4357/adc252

work page doi:10.3847/1538-4357/adc252 2025
[65]

Z., et al

Fu, H., Yan, L., Scoville, N. Z., et al. 2010, ApJ, 722, 653, doi: 10.1088/0004-637X/722/1/653

work page doi:10.1088/0004-637x/722/1/653 2010
[66]

2024, ApJ, 963, 51, doi: 10.3847/1538-4357/ad1ce5

Fu, S., Jiang, L., Ning, Y., Liu, W., & Pan, Z. 2024, ApJ, 963, 51, doi: 10.3847/1538-4357/ad1ce5

work page doi:10.3847/1538-4357/ad1ce5 2024
[67]

2013, ApJS, 206, 10, doi: 10.1088/0067-0049/206/2/10

Galametz, A., Grazian, A., Fontana, A., et al. 2013, ApJS, 206, 10, doi: 10.1088/0067-0049/206/2/10

work page doi:10.1088/0067-0049/206/2/10 2013
[68]

2024, A&A, 687, A288, doi: 10.1051/0004-6361/202348623

Gentile, F., Talia, M., Daddi, E., et al. 2024, A&A, 687, A288, doi: 10.1051/0004-6361/202348623

work page doi:10.1051/0004-6361/202348623 2024
[69]

R., Leauthaud, A., Bundy, K., et al

George, M. R., Leauthaud, A., Bundy, K., et al. 2011, ApJ, 742, 125, doi: 10.1088/0004-637X/742/2/125

work page doi:10.1088/0004-637x/742/2/125 2011
[70]

2017, A&A, 599, A95, doi: 10.1051/0004-6361/201629852

Gobat, R., Daddi, E., Strazzullo, V., et al. 2017, A&A, 599, A95, doi: 10.1051/0004-6361/201629852

work page doi:10.1051/0004-6361/201629852 2017
[71]

L., Connolly, A

Graham, M. L., Connolly, A. J., Ivezi´ c,ˇZ., et al. 2018, AJ, 155, 1, doi: 10.3847/1538-3881/aa99d4

work page doi:10.3847/1538-3881/aa99d4 2018
[72]

A., Kocevski, D

Grogin, N. A., Kocevski, D. D., Faber, S. M., et al. 2011, ApJS, 197, 35, doi: 10.1088/0067-0049/197/2/35

work page doi:10.1088/0067-0049/197/2/35 2011
[73]

2023, arXiv e-prints, arXiv:2301.05712, doi: 10.48550/arXiv.2301.05712

Gui, J., Chen, T., Zhang, J., et al. 2023, arXiv e-prints, arXiv:2301.05712, doi: 10.48550/arXiv.2301.05712

work page doi:10.48550/arxiv.2301.05712 2023
[74]

C., Giavalisco, M., et al

Guo, Y., Ferguson, H. C., Giavalisco, M., et al. 2013, ApJS, 207, 24, doi: 10.1088/0067-0049/207/2/24

work page doi:10.1088/0067-0049/207/2/24 2013
[75]

Harish, S., Wold, I. G. B., Malhotra, S., et al. 2022, ApJ, 934, 167, doi: 10.3847/1538-4357/ac7cf1

work page doi:10.3847/1538-4357/ac7cf1 2022
[76]

R., Millman , K

Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357, doi: 10.1038/s41586-020-2649-2

work page doi:10.1038/s41586-020-2649-2 2020
[77]

, archivePrefix = "arXiv", eprint =

Harrison, C. M., Alexander, D. M., Mullaney, J. R., et al. 2016, MNRAS, 456, 1195, doi: 10.1093/mnras/stv2727 Deep Learning Photo-z’s for Roman27

work page doi:10.1093/mnras/stv2727 2016
[78]

2018, ApJ, 858, 77, doi: 10.3847/1538-4357/aabacf

Hasinger, G., Capak, P., Salvato, M., et al. 2018, ApJ, 858, 77, doi: 10.3847/1538-4357/aabacf

work page doi:10.3847/1538-4357/aabacf 2018
[79]

2020, in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9729–9738

He, K., Fan, H., Wu, Y., Xie, S., & Girshick, R. 2020, in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9729–9738

work page 2020
[80]

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

work page 2016
[81]

2012, ApJ, 744, 149, doi: 10.1088/0004-637X/744/2/149

McCarthy, P. 2012, ApJ, 744, 149, doi: 10.1088/0004-637X/744/2/149

work page doi:10.1088/0004-637x/744/2/149 2012

Showing first 80 references.