The DOI matches its Crossref record; 62 of the 62 references it cites were found in a registry or an index; 7 of the 17 sources its claims cite were checked for support.
Self-supervised Representation Learning for Astronomical Images
Open accessAbstract
Abstract Sky surveys are the largest data generators in astronomy, making automated tools for extracting meaningful scientific information an absolute necessity. We show that, without the need for labels, self-supervised learning recovers representations of sky survey images that are semantically useful for a variety of scientific tasks. These representations can be directly used as features, or fine-tuned, to outperform supervised methods trained only on labeled data. We apply a contrastive learning framework on multiband galaxy photometry from the Sloan Digital Sky Survey (SDSS), to learn image representations. We then use them for galaxy morphology classification and fine-tune them for photometric redshift estimation, using labels from the Galaxy Zoo 2 data set and SDSS spectroscopy. In both downstream tasks, using the same learned representations, we outperform the supervised state-of-the-art results, and we show that our approach can achieve the accuracy of supervised models while using 2–4 times fewer labels for training. The codes, trained models, and data can be found at https://portal.nersc.gov/project/dasrepo/self-supervised-learning-sdss .
Overview
CoChat Summary
TL;DR
Self-supervised contrastive learning on unlabeled SDSS galaxy images learns representations that are useful for morphology classification and photometric redshift estimation. The learned representations outperform supervised baselines and can match supervised accuracy with 2–4× fewer labels.
Key findings
- Using the same learned representations, the authors outperform supervised state-of-the-art results on galaxy morphology classification and photometric redshift estimation.
- For photometric redshift estimation, pre-training on unlabeled data gives accuracy equivalent to supervised training on 2–4× more spectroscopic redshift labels.
- For the first Galaxy Zoo 2 morphology question, roughly a factor of 16 more labels are required in the supervised setting to achieve the same performance as the fine-tuned self-supervised model.
- In augmentation ablation, Gaussian noise was the strongest augmentation and PSF the weakest; best performance was achieved using all augmentations except PSF.
Methods
- Contrastive self-supervised learning with an InfoNCE loss on ∼1.2 million SDSS ugriz galaxy images of 64×64 pixels, using a ResNet50 encoder and augmentations including galactic extinction, PSF smoothing, rotation, jitter/crop, and Gaussian noise.
- Downstream tasks: Galaxy Zoo 2 morphology classification using 183,929 galaxies with both ugriz images and crowd-sourced classifications, and photometric redshift estimation using 517,190 labeled spectroscopic galaxies, with 400,000 for training and 103,000 for validation.
- Morphology classifiers included a supervised CNN from scratch, a linear classifier on frozen self-supervised representations, and a fine-tuned self-supervised encoder; photometric redshift used 180 redshift bins spanning 0 ≤ z ≤ 0.4 and metrics Δz, σ_MAD, and outlier percentage η.
Limitations
- The paper states that exact quantitative comparison to Domínguez Sánchez et al. (2018) and Walmsley et al. (2020) is not possible due to inconsistent datasets, and that augmentation findings may not generalize to other surveys with different resolutions, signal-to-noise ratios, or target objects.
- The paper states that spectroscopic labels impose selection bias: galaxies with specz labels are biased toward nearby bright objects with low galactic extinction and are not representative of all photometric galaxies.
- The text also makes evident that images span 25.3 arcseconds, which is not large enough for very nearby large galaxies, and that misclassifications arise from imaging artifacts and label uncertainty.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
7 Valid · 1 Not valid · 4 Not checked · 74 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“For example, due to the series of flux and quality cuts applied when selecting spectroscopic targets (Strauss et al. 2002), galaxies with specz labels have a distinct bias towards nearby bright objects with low galactic extinction.”
What the sources say
- Partly supportedcites Strauss et al. 2002 sourceExperimental — accuracy not yet measured
Section 4.1 (page unavailable) describes an r-band magnitude limit of 17.77, and §5.2 reports a median redshift of 0.104; the paper also describes photometric quality cuts. These support selection toward apparently bright, relatively nearby galaxies. But §4.1 says magnitudes are corrected for Galactic extinction before selection, and the cited passages do not establish a distinct bias toward low-extinction sightlines. (LLM, conf=0.86)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“and we show that our approach can achieve the accuracy of supervised models while using 2-4 times fewer labels for training.”
- Also stated in: § 3.3. Photometric Redshift Estimation · p. 7 · boosted
“Impressively, the accuracy gained from pre-training on unlabeled data is equivalent to supervised training on 2 -4 × more specz labels, with no modifications in architecture size or complexity.”
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 3.3. Photometric Redshift Estimation · p. 7 · boosted
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Likewise, we found that classification became a straightforward task when using any subset of the labels.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Here we further show that the UMAP performed on the representations surpasses the utility of a UMAP performed directly on the images”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“As evidenced by Figure 2, the self-supervised representations have achieved a high degree of separation between numerous types of galaxy morphologies.”
boostedstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
40 more claims
Create a free account to see all 86 claims and to verify the 1 reference still open.
- Probably not a substantial claim. 25
- It may state two separate claims as one. 18
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Another benefit to self-supervised models, other than training on unlabelled data, is the computational cost, self-supervised models have been shown to need only 1% of the computational power needed by supervised models (Hayat et al. 2021).”
From Detecting galaxy tidal features using self-supervised representation learning (2024) · cites this paper as Hayat et al. 2021
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Here, the authors proposed a two-step approach, first training in complete self-supervision (without needing any spectroscopic redshifts) a 1d encoding of galaxies postage stamps. And in a second step, training in a supervised way a shallow MLP to predict redshifts on a small training set of available spectroscopic redshifts.”
From The Dawes Review 10: The impact of deep learning for the analysis of galaxy surveys (2023) · cites this paper as Hayat et al. (2021)
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“They find for instance that their selfsupervised approach using 20% of labelled data achieves similar accuracy to a fully supervised training using labels for the entire dataset.”
From The Dawes Review 10: The impact of deep learning for the analysis of galaxy surveys (2023) · cites this paper as Hayat et al. (2021)
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“The first work exploring self-supervised learning is by Hayat et al. (2021).”
From The Dawes Review 10: The impact of deep learning for the analysis of galaxy surveys (2023) · cites this paper as Hayat et al. (2021)
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“Hayat et al. (2021) and Stein et al. (2022) show that once the encoder section of a self-supervised model has been trained one can conduct a simple similarity search to find objects of interest, or apply a linear classifier directly onto the encoder's outputs to separate the data into a variety of classes.”
From Detecting galaxy tidal features using self-supervised representation learning (2024) · cites this paper as Hayat et al. 2021
6 more claims about this paper
Create a free account to see every claim other papers make about this one.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ C.3. Performance metrics & additional results · p. 17
“Figure 11. Predicted labels compared to crowd-sourced answers for the first GZ2 question: 'Could this be a disk viewed edge-on?'.”
§ C.3. Performance metrics & additional results · p. 16
“Figures 11 and 12 show the classification results in the same style as Figure 3, but for the second and third GZ questions, respectively.”
Citations
Cites
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325
A simple framework for contrastive learning of visual representations
Chen, T., Kornblith, S., Norouzi, M., & Hinton, G. (2020) · arXiv · doi:10.48550/arxiv.2002.05709 · Cited by 10,857
MEASURING REDDENING WITH SLOAN DIGITAL SKY SURVEY STELLAR SPECTRA AND RECALIBRATING SFD
Schlafly, E. F., & Finkbeiner, D. P. (2011) · The Astrophysical Journal · doi:10.1088/0004-637x/737/2/103 · Cited by 7,531
UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction
McInnes, L., Healy, J., & Melville, J. (2018) · arXiv · doi:10.48550/arxiv.1802.03426 · Cited by 7,101
Representation Learning with Contrastive Predictive Coding
Oord, A. V. D., Li, Y., & Vinyals, O. (2018) · arXiv · doi:10.48550/arxiv.1807.03748 · Cited by 6,242
Dimensionality Reduction by Learning an Invariant Mapping
Hadsell, R., Chopra, S., & LeCun, Y. (2006) · 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition - Volume 2 (CVPR'06) · doi:10.1109/cvpr.2006.100 · Cited by 5,336
LSST: From Science Drivers to Reference Design and Anticipated Data Products
Ivezić, Ž., Kahn, S. M., Tyson, J. A., Abel, B., Acosta, E., Allsman, R., Alonso, D., AlSayyad, Y., Anderson, S. F., Andrew, J., et al. (2019) · The Astrophysical Journal · doi:10.3847/1538-4357/ab042c · Cited by 4,357 · cited 3×
“In the near future, incoming sky surveys such as the Vera Rubin Observatory (Ivezić et al. 2019b), Euclid (Laureijs et al. 2011), the Nancy Grace Roman Space Telescope (Spergel et al. 2013), or the Square Kilometer Array6 will open yet another research epoch, where data sets are of sizes that…”
BackgroundMethodLSST: from Science Drivers to Reference Design and Anticipated Data Products
Ivezić, Ž., Kahn, S. M., Tyson, J. A., Abel, B., Acosta, E., Allsman, R., Alonso, D., AlSayyad, Y., Anderson, S. F., Andrew, J., et al. (2008) · arXiv · doi:10.48550/arxiv.0805.2366 · Cited by 4,357
THE ELEVENTH AND TWELFTH DATA RELEASES OF THE SLOAN DIGITAL SKY SURVEY: FINAL DATA FROM SDSS-III
Alam, S., Albareti, F. D., Prieto, C. A., Anders, F., Anderson, S. F., Anderton, T., Andrews, B. H., Armengaud, E., Aubourg, É., Bailey, S., et al. (2015) · The Astrophysical Journal Supplement Series · doi:10.1088/0067-0049/219/1/12 · Cited by 2,449 · cited 1×
“Our database of galaxies is assembled from Data Release 12 (DR12; Alam et al. 2015) of the SDSS.”
Method
Cited by
The Dawes Review 10: The impact of deep learning for the analysis of galaxy surveys
Huertas-Company, M., & Lanusse, F. (2023) · Publications of the Astronomical Society of Australia · doi:10.1017/pasa.2022.55 · Cited by 102
“The first work exploring self-supervised learning is by Hayat et al. (2021).”
BackgroundMethodInfluentialAstronomia ex machina: a history, primer and outlook on neural networks in astronomy
Smith, M. J., & Geach, J. E. (2023) · Royal Society Open Science · doi:10.1098/rsos.221454 · Cited by 111
“[231] trained SimCLR [185] on multi-band galaxy photometry from the SDSS [74].”
MethodAstroCLIP: a cross-modal foundation model for galaxies
Parker, L., Lanusse, F., Golkar, S., Sarra, L., Cranmer, M., Bietti, A., Eickenberg, M., Krawezik, G. P., McCabe, M., Morel, R., et al. (2024) · Monthly Notices of the Royal Astronomical Society · doi:10.1093/mnras/stae1450 · Cited by 192
Practical galaxy morphology tools from deep supervised representation learning
Walmsley, M., Scaife, A. M. M., Lintott, C. J., Lochner, M., Etsebeth, V., Géron, T., Dickinson, H. J., Fortson, L., Kruk, S., Masters, K. L., et al. (2022) · Monthly Notices of the Royal Astronomical Society · doi:10.1093/mnras/stac525 · Cited by 111
Mining for Strong Gravitational Lenses with Self-supervised Learning
Stein, G., Blaum, J., Harrington, P., Medan, T., & Lukić, Z. (2022) · The Astrophysical Journal · doi:10.3847/1538-4357/ac6d63 · Cited by 52
“…orientation of the galaxies, the number of sources, the clustering, and the color, while removing the noise and symmetries described by the data augmentations (for a visualization of how galaxy properties are organized in the representation space of a self-supervised model, see Hayat et al. 2021).”
BackgroundMethodInfluentialSelf-citation
Metadata
- Authors
- Hayat, M. A., Stein, G., Harrington, P., Lukić, Z., & Mustafa, M.
- Affiliations
- Lawrence Berkeley National Laboratory (US)
- Venue
- The Astrophysical Journal Letters
- Publisher
- American Astronomical Society
- Published
- 2021-04-01
- Volume
- 911
- Issue
- 2
- Pages
- L33
- ISSN
- 2041-8205 (ISSN-L)
- 2041-8213
- OpenAlex
- W3118065583
- Cited by
- 79
- References
- 50
- Citations by year
- 2026: 8 · 2025: 14 · 2024: 18 · 2023: 20 · 2022: 15 · 2021: 4
- FWCI
- 7.60
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0/ (vor, 25 d)
- iopscience.iop.org/info/page/text-and-data-mining (tdm, 25 d)
- Locations
- The Astrophysical Journal Letters (publishedVersion, cc-by)
- arXiv (Cornell University) (submittedVersion)
- OSTI OAI (U.S. Department of Energy Office of Scientific and Technical Information) (submittedVersion, cc-by)
- eScholarship (California Digital Library) (submittedVersion)
- arXiv (Cornell University)
- Funders
- NERSC (AC02-05CH11231)
- Topics
- Physical Sciences › Physics and Astronomy › Astronomy and Astrophysics › Galaxies: Formation, Evolution, Phenomena
- Physical Sciences › Physics and Astronomy › Astronomy and Astrophysics › Astronomical Observations and Instrumentation
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Keywords
- self-supervised representation learning, contrastive learning, celestial imagery, sky survey, galaxy morphology classification, photometric redshift estimation, Galaxy Zoo 2, Sloan Digital Sky Survey, label efficiency
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Self-supervised Representation Learning for Astronomical Images. https://app.cochat.ai/papers/10.3847/2041-8213/abf2c7
- The paper
- Hayat, M. A., Stein, G., Harrington, P., Lukić, Z., & Mustafa, M. (2021). Self-supervised Representation Learning for Astronomical Images. The Astrophysical Journal Letters. https://doi.org/10.3847/2041-8213/abf2c7
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.