Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Self-supervised learning on raw galaxy cutouts organizes morphologies like major mergers, but not stellar streams; a tiered sigmoid scaling steers attention away from bright centers.

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

2026-08-03 06:15 UTC pith:A73KJ2ZE

load-bearing objection An honest, useful methods paper whose main negative result—streams are not separable in NNCLR embeddings—holds up; the one quantitative claim for the scaling benefit rests on an invalid MI statistic and a single seed, so treat that part as suggestive, not proven. the 4 major comments →

arxiv 2601.23013 v2 pith:A73KJ2ZE submitted 2026-01-30 astro-ph.GA

Contrastive learning of extragalactic stellar streams. Sculpting a latent space of representations with DES DR2 photometry

Ernesto Benitez-Walz , Jelle Mes , Juan Mir\'o-Carretero , Koen Kuijken , Amina Helmi This is my paper
classification astro-ph.GA
keywords contrastive learningself-supervisedstellar streamslow surface brightnessNNCLRDark Energy Surveyimage scalingsaliency
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 sets out to show that a fully self-supervised contrastive model, trained on 38,334 unlabeled three-color cutouts of bright, well-resolved galaxies from DES DR2, can organize galaxies by the coarse morphology of their outskirts — major-merger-like features, smoothness, and ellipticity — without any human labels. It introduces a "tiered sigmoid scaling" that adjusts contrast according to the central galaxy's magnitude tier and background level, and argues, via averaged gradient saliency maps, that this scaling suppresses information from the galaxy's bright center and pushes the network's attention outward where tidal features live. The paper is candid that this is not enough: the 366 labeled galaxies, including 47 known stellar streams, scatter across the learned embedding, so streams are not separable without further supervision. A sympathetic reader would take the contribution as a blueprint: contrastive pretraining plus deliberately chosen intensity scaling can mine wide-field survey images for faint substructure, with the scaling function serving as a controllable attention knob.

Core claim

The central claim is that the NNCLR latent space trained on tiered-sigmoid-scaled DES DR2 cutouts recovers global morphological trends — regions of the two-dimensional projection correlate with citizen-science merger-fraction votes and with smoothness — while failing to separate the specific class of stellar streams. The paper further claims that the tiered sigmoid scaling, a logistic function whose center and growth rate are calibrated per magnitude tier and per band, is what redirects the model's attention: average saliency maps show the model trained on tiered-sigmoid images is sensitive to a broad central annulus and less to the saturated core, whereas the arcsinh-stretch model concentra

What carries the argument

The central object is the NNCLR contrastive-learning pipeline: a ResNet-18 encoder projects two augmented views of each image into an embedding space, and the loss compares one view to the nearest neighbor of a queue of other views. The other load-bearing piece is the tiered sigmoid scaling function, f(x)=1/(1+exp(-Ω)) with Ω=(x-μ)/(σ) scaled by the ratio of a reference signal-to-noise to the galaxy's own S/N, calibrated in four magnitude tiers per band. That scaling is doing the work: it determines where the encoder's gradients concentrate, and the paper uses averaged saliency maps to show it suppresses central, high-surface-brightness pixels. UMAP and densMAP then project the 512-dimension

Load-bearing premise

The claim that the new scaling works better than the standard stretch depends on a clustering score that the authors admit is not a true probability distribution, and on training each model once; if that score is not meaningful, only visual inspection remains.

What would settle it

Compute the mutual information between the citizen-science merger-fraction labels and the densMAP coordinates using a valid density estimator rather than the paper's normalized histogram, and repeat training with several random seeds. If the 0.688 versus 0.642 gap does not persist, the central quantitative support for the tiered sigmoid scaling is falsified — though the qualitative saliency result could still stand.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Self-supervised contrastive pretraining on unlabeled survey cutouts can serve as a label-free first pass that surfaces galaxies with major-merger-like morphologies.
  • Choosing an intensity scaling is not a preprocessing afterthought: it changes which pixels the model encodes, so scaling choice should be reported and tuned alongside augmentations.
  • Stellar-stream detection will require additional supervision, richer augmentations, or a different objective than plain instance discrimination on these data.
  • Tier randomization prevents the network from memorizing the scaling tier, which otherwise dominates the embedding and hides morphological structure.
  • The pipeline is transferable to forthcoming wide-field surveys, where it can pretrain encoders on massive unlabeled archives before fine-tuning on small curated stream samples.

Where Pith is reading between the lines

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

  • If the saliency result generalizes, intensity scaling is a cheap substitute for explicit attention: one could tune the scaling function to implement a desired inductive bias — ignore cores, enhance tidal tails — without changing the network architecture.
  • A direct test of the paper's negative result would be to inject synthetic streams with known surface brightness into cutouts and measure whether a linear probe on the NNCLR embedding recovers them; the paper's "not separable" claim predicts it will fail until the scaling or objective changes.
  • The 47 positives versus 38,334 unlabeled galaxies is an extreme class imbalance; the reported scatter of stream embeddings may reflect the diversity of stream morphologies (great circles, umbrellas, giant plumes) being each too rare to form a coherent cluster, rather than a fundamental limit of contrastive learning.
  • Because all runs share one global seed, the reported mutual-information gap should be read as preliminary until repeated over seeds.

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

4 major / 6 minor

Summary. The paper applies the NNCLR self-supervised contrastive learning algorithm to 38,334 DES DR2 galaxy cutouts in g, r, i bands, introducing a 'tiered sigmoid scaling' intensity transformation calibrated per magnitude tier using stellar-stream hosts from Miró-Carretero et al. (2024). The learned 512-dimensional embeddings are projected with UMAP/densMAP and compared across two scaling functions: the new tiered sigmoid and the arcsinh stretch used by Desmons et al. (2024). The authors report that the latent space organizes by coarse morphology (major merger fraction and smoothness), but that the small labeled sample of 366 galaxies with stellar streams is not reliably separated without further supervision. Saliency maps suggest the tiered sigmoid scaling reduces gradient magnitude in the central galaxy region. The paper claims the tiered sigmoid scaling yields a higher degree of clustering in the densMAP projections (MI = 0.688 vs 0.642) and interprets this as evidence that preprocessing can steer the encoder's implicit attention toward low-surface-brightness features.

Significance. If the qualitative results hold, the paper provides a useful, reproducible pipeline for applying contrastive learning to large survey cutouts and a valuable negative result: standard self-supervised contrastive learning on raw, intensity-scaled cutouts does not spontaneously separate stellar streams, even when the scaling is tailored to low-surface-brightness features. The authors are transparent about the invalidity of their mutual-information estimator, which is commendable, but they still rely on that estimator for the paper's central quantitative comparison. The public-data usage, detailed preprocessing, and explicit comparison between scaling functions are strengths. The main positive claims about the scaling function's benefit, however, rest on a statistically invalid metric, a single training run per configuration, and qualitative saliency maps that partly measure the analytic properties of the sigmoid itself.

major comments (4)
  1. [§3.2, Eqs. (7)–(9)] The mutual information used for the 'higher degree of clustering' claim is not a valid statistical quantity, as the authors themselves note. Eq. (8) constructs a distribution from bin-wise means of μ, not from a joint count distribution, and the arbitrary positive shift of μ changes the resulting value. The bootstrap errors (5000 Gaussian resampled histograms) only quantify noise in the bin means under an assumed normal model, not the validity of the statistic. Therefore the values MI = 0.688 ± 0.007 vs 0.642 ± 0.007 do not constitute a meaningful quantitative comparison. The authors should either replace this with a valid mutual-information estimate on actual labels (e.g., nearest-neighbor entropy or a proper density estimator), or with a different quantitative metric such as a supervised k-NN classification of merger fraction. As written, the quantitative superiority claim for tiered s
  2. [§2.3, §3.2] All models are trained with a single global seed and the same initial parameters. This makes runs paired, but it does not bound the run-to-run variance of the contrastive training procedure. The 0.046 MI gap between scaling functions could be within the stochasticity of the optimization, especially given the invalid MI metric. The authors need to train multiple seeds per scaling function (at least 3–5) and report the distribution of the chosen metric, or otherwise demonstrate that the gap is larger than training noise. Without this, the claim that the scaling function improves clustering is not established.
  3. [§3.1, §4] The saliency maps in Fig. 5 show reduced gradient magnitude in the central region for the tiered-sigmoid model, but this is substantially a direct consequence of the sigmoid transfer function: saturated central pixels have near-zero derivatives with respect to the input. The conclusion states that the saliency maps 'confirm that the network learns to emphasize the flux variations most relevant to the identification of tidal debris,' which overstates the evidence. The maps indicate differences in local gradients, not learned attention. To support the 'implicit attention' claim, the authors should apply an attribution method that is more robust to input scaling (e.g., integrated gradients or a controlled experiment with the same architecture but different scaling), or at least temper the interpretation explicitly. As it stands, the central-attenuation observation is partly an analytic prop
  4. [§2.2, §3.3] The tiered sigmoid scaling hyperparameters (Table A.2) are calibrated using elliptical apertures around stellar streams from Miró-Carretero et al. (2024), and the same catalog provides the labeled sample used to assess stream visibility and nearest-neighbor queries (Fig. 11). The claim in §3.3 that the scaling is 'generic in emphasizing low surface brightness features without the need to manually tune' is therefore only demonstrated on the tuning set. A held-out stream sample (e.g., streams identified independently in other surveys) or a leave-one-out calibration test would be needed to support the generality claim. This is not fatal to the pipeline's usefulness, but it should be acknowledged as a limitation and the language softened.
minor comments (6)
  1. [§2.2] The cutout size factor of 40× the r-band half-light radius is described as heuristically chosen. A brief sensitivity analysis (e.g., 30× and 50×) would help assess whether the main conclusions depend on this choice.
  2. [Appendix B] The tier randomization ablation is presented only as a visual UMAP comparison. A quantitative measure (e.g., cluster purity by tier, or a valid MI on tier labels) would strengthen the point that randomization reduces tier memorization.
  3. [Fig. 5 caption] The caption repeats 'Tiered sigmoid scaling model' for both panels; the second panel should read 'Arcsinh scaling model.'
  4. [Eq. (3)] The definition of ρ_{j,k,b} would be clearer if the authors explicitly state that S/N_{j,b} is the tier-averaged signal-to-noise ratio and S/N_{k,b} is the per-image value; as written, the subscript j on the left appears only through the tier-average on the right.
  5. [§2.3] The text says a global seed is applied to 'all numpy, pytorch, and python.random objects.' It would be helpful to specify whether the augmentation pipeline itself (including star injection and noise) was also seeded and whether the same augmented pairs were used for the two scaling functions; the current wording in §2.3.1 suggests this but could be made explicit.
  6. [§4] The phrase 'validating the conceptual design of the tiered sigmoid scaling' is too strong given the quantitative issues; suggest 'consistent with the intended effect' or similar.

Circularity Check

1 steps flagged

Minor tuning-set circularity in the scaling demonstration; main clustering result is externally benchmarked and non-circular.

specific steps
  1. fitted input called prediction [Section 2.2 (Eqs. 1-3, Tables A.1-A.2) and Section 3.3 / Fig. 11]
    "We calibrated the center and growth rate of the tiered sigmoid scaling function per tier and per band by gathering photometric statistics from elliptical apertures encompassing stellar streams reported by Miró-Carretero et al. (2024). ... the low surface brightness features in the tiered-sigmoid-scaled images are easier to identify immediately and appear to have a tidal origin"

    The mu/sigma parameters of the tiered sigmoid (Eqs. 1-3) are fitted, per tier and band, to the photometric statistics of stream apertures from Miró-Carretero et al. (2024). The Section 3.3 demonstration then uses this same scaling to argue that stream features are easier to see in the scaled query and nearest-neighbor images. This is therefore partly an evaluation of the tuning target rather than an independent test of stream visibility. It is illustrative rather than the central quantitative claim, which is benchmarked against external Galaxy Zoo DECaLS merger labels, so the circularity is minor.

full rationale

The paper's main positive claim—that the tiered sigmoid scaling yields stronger clustering of major-merger features (MI = 0.688 vs 0.642)—is benchmarked against the external Galaxy Zoo DECaLS merging_merger_fraction labels, which are not used in training or in fitting the scaling parameters. That comparison is not circular, even though its statistical validity is weakened by the paper's own admission in Section 3.2 that Eq. (8) 'does not strictly yield a valid probability distribution' and by the single-seed training setup; those are validity/robustness concerns, not circularity. The labeled stream sample from Miró-Carretero et al. (2024) is used for qualitative assessment and for calibrating the scaling parameters, and a co-author of that catalog is also an author here, but the central clustering result does not depend on those labels. The saliency-map finding that the sigmoid attenuates central information is partly a direct consequence of the saturated sigmoid transfer function, as the authors themselves state; this is an interpretability observation rather than a fitted prediction. The only concrete circular element is the qualitative demonstration that the tuned scaling makes streams easier to see, which evaluates the tuning set. Overall, the derivation chain is largely self-contained and the main result has independent external content.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claims rest on standard ML/statistical algorithms (NNCLR, UMAP, densMAP, resampled bootstrap) treated as black boxes, plus astronomy-specific domain assumptions about catalog quality and human labels. The only fitted quantities are the tiered sigmoid hyperparameters, which are tuned to the very stream sample the paper later uses for evaluation; this is the main circularity burden and is flagged in the corresponding score.

free parameters (4)
  • Tiered sigmoid scaling center μ_{j,b} and growth rate σ_{j,b} per tier j and band b = Reported as mean median and mean MAD per tier in Table A.2 (e.g., tier 1 r-band ⟨Median⟩ = 2.1420, ⟨MAD⟩ = 0.9901)
    Calibrated from photometric statistics of elliptical apertures around stellar streams from Miró-Carretero et al. (2024); these directly set the contrast enhancement in Eq. (2).
  • Cutout size factor (40 × r-band half-light radius) = 40
    Heuristic constant chosen to make the central galaxy span 1/3 of the frame; it determines the physical field of view and thus the amount of tidal feature visible.
  • Tier randomization probability = 30%
    Hand-chosen to discourage the network from memorizing the tier category while retaining tier information (Appendix B ablation).
  • NNCLR temperature τ = 0.1
    Softening parameter in the NNCLR loss (Eq. 5); chosen by the authors, standard for contrastive learning.
axioms (5)
  • domain assumption DES DR2 catalog parameters (extended_coadd_class > 2.5, flux_radius_r, flags, imaflags_iso) correctly identify well-resolved galaxies with clean photometry.
    Used in Section 2.1 to build the 38,334-image training set; if misclassifications are common, the contrastive objective is trained on noisy instances.
  • domain assumption The Miró-Carretero et al. (2024) human classifications of 47 stellar-stream hosts are correct.
    These labels define the 'positives' whose embedding dispersion is used to conclude that streams are not separable (Section 3.2, Fig. 9), and they also supply the calibration apertures for the tiered sigmoid scaling.
  • domain assumption Galaxy Zoo DECaLS merging_merger_fraction and smooth_or_featured_smooth_fraction are meaningful morphological labels for the DES galaxies.
    Used in Section 3.2 to color the UMAP/densMAP projections and to compute the mutual information that supports the clustering claim.
  • domain assumption Gradient-based saliency maps (Simonyan et al. 2014) faithfully indicate which image regions the NNCLR encoder attends to.
    The central 'attention attenuation' claim in Section 3.1 rests on interpreting E_grad as implicit attention; Adebayo et al. (2018), cited by the authors, caution that such maps can be insensitive to model parameters, which is a known caveat for this method.
  • ad hoc to paper The tiered sigmoid scaling with 30% tier randomization does not introduce a learnable artifact other than the tier category tested in Appendix B.
    The scaling saturates galaxy centers to unity and rescales by S/N ratios; the paper only ablates tier memorization, not other potential artifacts (e.g., edge effects from cutout resizing or the ρ weighting in Eq. 3).

pith-pipeline@v1.3.0-alltime-deepseek · 16774 in / 15454 out tokens · 142914 ms · 2026-08-03T06:15:39.151646+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Contrastive learning of extragalactic stellar streams. Sculpting a latent space of representations with DES DR2 photometry." pith.science (2026). https://pith.science/paper/A73KJ2ZE

@misc{pith2026260123013,
  author       = {Pith},
  title        = {Pith review of: Contrastive learning of extragalactic stellar streams. Sculpting a latent space of representations with DES DR2 photometry},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A73KJ2ZE}},
  note         = {Machine review of arXiv:2601.23013}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We present a self-supervised approach for characterizing low surface brightness tidal features in wide-field imaging data by applying the nearest-neighbor contrastive learning of visual representations (NNCLR) algorithm to a curated subset of the Dark Energy Survey Data Release 2 (DES DR2). We construct 38,334 cutouts of well-resolved galaxies in the g, r, i bands, applying a novel "tiered sigmoid scaling function" to dynamically adjust image contrast according to the object's signal-to-noise and background level. A supplemental labeled sample of 366 galaxies enables qualitative assessment of the learned embeddings. We train a convolutional neural network with image augmentations including injection of simulated background stars, and project the resulting 512-dimensional representations into two dimensions using uniform manifold approximation and projection (UMAP) and its local density preserving variant (densMAP). We find that the NNCLR latent space recovers global trends corresponding to major merger features, yet does not reliably separate stellar streams without further supervision. To interpret the network's implicit attention, we compute gradient-based saliency maps averaged over the full dataset: these reveal that the tiered sigmoid scaling effectively attenuates information from the center of the image cutouts, thereby suppressing the learning of high surface brightness features of each image cutout's central galaxy. Our study provides a blueprint for leveraging contrastive methods to mine forthcoming survey data for faint tidal substructure, and highlights key preprocessing and interpretability considerations for robust stream detection.

Figures

Figures reproduced from arXiv: 2601.23013 by Amina Helmi, Ernesto Benitez-Walz, Jelle Mes, Juan Mir\'o-Carretero, Koen Kuijken.

Figure 1
Figure 1. Figure 1: Histogram of objects per tiered sigmoid scaling function [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: From visual inspection of our dataset, the galaxy LEDA [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: A schematic of the NNCLR algorithm as outlined by Dwibedi et al. (2021). A batch of images is input to the augmentation [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: A randomly selected image (top row) prior to the star [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The saliency map of the model trained on the DES DR2 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: A mosaic of the DES DR2 subset’s (described in Section 2.1 and 2.2) embedding vectors from the NNCLR encoder, after [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Mean histograms of the merging_merger_fraction Galaxy Zoo DECaLS parameter (Walmsley et al. 2021) for the (left) UMAP projection of the tiered sigmoid scaling NNCLR model embeddings and (right) the UMAP projection of the NNCLR embeddings from the model trained on arcsinh-scaled images. 0.0 0.2 0.4 0.6 0.8 1.0 densMAP1 0.0 0.2 0.4 0.6 0.8 1.0 densMAP2 MI = 0.688 ± 0.007 Tiered sigmoid scaling model merging_… view at source ↗
Figure 8
Figure 8. Figure 8: Mean histograms of the merging_merger_fraction Galaxy Zoo DECaLS parameter (Walmsley et al. 2021) for the (left) densMAP projection of the tiered sigmoid scaling NNCLR model embeddings and (right) the densMAP projection of the NNCLR embeddings from the model trained on arcsinh-scaled images. The text box at the bottom left of both plots states the mutual information of the given histogram, as defined in Se… view at source ↗
Figure 9
Figure 9. Figure 9: The labeled dataset of Miró-Carretero et al. (2024) passed [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: (Top) The crossmatched Galaxy Zoo DECaLS [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: We can further probe the latent space of representations of both our models by inputting a query image and checking for [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Potamides: Mapping Dark Matter Halo Shapes from Stellar Stream Tracks in the Local Universe

    astro-ph.GA 2026-05 unverdicted novelty 7.0

    Stellar stream curvature constrains dark matter halo axis ratios and orientations in external galaxies, with some streams excluding broad ranges of flattenings while all support spherical halos for fixed flattening di...

Reference graph

Works this paper leans on

52 extracted references · 8 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Abbott, T. M. C., Abdalla, F. B., Allam, S., et al. 2018, ApJS, 239, 18

    2018

  2. [2]

    Abbott, T. M. C., Adamów, M., Aguena, M., et al. 2021, The Astrophysical Jour- nal Supplement Series, 255, 20

    2021

  3. [3]

    2018, arXiv e-prints, arXiv:1810.03292

    Adebayo, J., Gilmer, J., Muelly, M., et al. 2018, arXiv e-prints, arXiv:1810.03292

    Pith/arXiv arXiv 2018

  4. [4]

    2018, Publications of the Astro- nomical Society of Japan, 70, S4, aDS Bibcode: 2018PASJ...70S...4A

    Aihara, H., Arimoto, N., Armstrong, R., et al. 2018, Publications of the Astro- nomical Society of Japan, 70, S4, aDS Bibcode: 2018PASJ...70S...4A

    2018

  5. [5]

    M., Abraham, R

    Atkinson, A. M., Abraham, R. G., & Ferguson, A. M. N. 2013, The Astrophysi- cal Journal, 765, 28

    2013

  6. [6]

    Barnes, J. E. 1988, The Astrophysical Journal, 331, 699

    1988

  7. [7]

    Barnes, J. E. 1992, The Astrophysical Journal, 393, 484

    1992

  8. [8]

    & Arnouts, S

    Bertin, E. & Arnouts, S. 1996, A&AS, 117, 393

    1996

  9. [9]

    Bonaca, A., Geha, M., Küpper, A. H. W., et al. 2014, The Astrophysical Journal, 795, 94, publisher: The American Astronomical Society

    2014

  10. [10]

    & Hogg, D

    Bonaca, A. & Hogg, D. W. 2018, The Astrophysical Journal, 867, 101, publisher: The American Astronomical Society

    2018

  11. [11]

    K., & Kallivayalil, N

    Bovy, J., Bahmanyar, A., Fritz, T. K., & Kallivayalil, N. 2016, The Astrophysical Journal, 833, 31

    2016

  12. [12]

    Bowden, A., Belokurov, V ., & Evans, N. W. 2015, MNRAS, 449, 1391

    2015

  13. [13]

    2024, astropy/photutils: 2.0.2 Dark Energy Survey Collaboration, Abbott, T., Abdalla, F

    Bradley, L., Sip˝ocz, B., Robitaille, T., et al. 2024, astropy/photutils: 2.0.2 Dark Energy Survey Collaboration, Abbott, T., Abdalla, F. B., et al. 2016, Monthly Notices of the Royal Astronomical Society, 460, 1270, aDS Bib- code: 2016MNRAS.460.1270D de Jong, J. T. A., Verdoes Kleijn, G. A., Kuijken, K. H., & Valentijn, E. A. 2013, Experimental Astronomy...

    2024

  14. [14]

    2024, MNRAS, 531, 4070

    Desmons, A., Brough, S., & Lanusse, F. 2024, MNRAS, 531, 4070

    2024

  15. [15]

    J., Lang, D., et al

    Dey, A., Schlegel, D. J., Lang, D., et al. 2019, The Astronomical Journal, 157, 168

    2019

  16. [16]

    2021, An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., et al. 2021, An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    2021

  17. [17]

    2021, arXiv e-prints, arXiv:2104.14548 Ebrová, I., Jílková, L., Jungwiert, B., et al

    Dwibedi, D., Aytar, Y ., Tompson, J., Sermanet, P., & Zisserman, A. 2021, arXiv e-prints, arXiv:2104.14548 Ebrová, I., Jílková, L., Jungwiert, B., et al. 2012, åp, 545, A33, _eprint: 1207.0642 Euclid Collaboration, Scaramella, R., Amiaux, J., et al. 2022, A&A, 662, A112

    Pith/arXiv arXiv 2021

  18. [18]

    J., Olsen, K., Economou, F., et al

    Fitzpatrick, M. J., Olsen, K., Economou, F., et al. 2014, in Observatory Opera- tions: Strategies, Processes, and Systems V , V ol. 9149, 91491T, aDS Bibcode: 2014SPIE.9149E..1TF

    2014

  19. [19]

    J., Ferguson, A

    Gordon, A. J., Ferguson, A. M. N., & Mann, R. G. 2024, MNRAS, 534, 1459

    2024

  20. [20]

    W., Cuillandre, J.-C., et al

    Gwyn, S., McConnachie, A. W., Cuillandre, J.-C., et al. 2025, The Astronomical Journal, 170, 324, publisher: The American Astronomical Society

    2025

  21. [21]

    A., Stein, G., Harrington, P., Luki´c, Z., & Mustafa, M

    Hayat, M. A., Stein, G., Harrington, P., Luki´c, Z., & Mustafa, M. 2021, ApJ, 911, L33, _eprint: 2012.13083 Article number, page 11 A&A proofs:manuscript no. main

    Pith/arXiv arXiv 2021

  22. [22]

    2016, in 2016 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR, 1

    He, K., Zhang, X., Ren, S., & Sun, J. 2016, in 2016 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR, 1

    2016

  23. [23]

    2008, åpr, 15, 145, _eprint: 0804.0019

    Helmi, A. 2008, åpr, 15, 145, _eprint: 0804.0019

    Pith/arXiv arXiv 2008

  24. [24]

    & White, S

    Helmi, A. & White, S. D. M. 1999, MNRAS, 307, 495 Ivezi´c, Ž., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111

    1999

  25. [25]

    V ., Hernquist, L., & Bolte, M

    Johnston, K. V ., Hernquist, L., & Bolte, M. 1996, ApJ, 465, 278, _eprint: astro- ph/9602060

    arXiv 1996

  26. [26]

    V ., Spergel, D

    Johnston, K. V ., Spergel, D. N., & Hernquist, L. 1995, ApJ, 451, 598, _eprint: astro-ph/9502005

    Pith/arXiv arXiv 1995

  27. [27]

    Kobourov, S. G. 2012, arXiv e-prints, arXiv:1201.3011

    Pith/arXiv arXiv 2012

  28. [28]

    E., Rix, H.-W., & Hogg, D

    Koposov, S. E., Rix, H.-W., & Hogg, D. W. 2010, ApJ, 712, 260

    2010

  29. [29]

    J., Schawinski, K., Slosar, A., et al

    Lintott, C. J., Schawinski, K., Slosar, A., et al. 2008, MNRAS, 389, 1179

    2008

  30. [30]

    & Hutter, F

    Loshchilov, I. & Hutter, F. 2017, arXiv e-prints, arXiv:1711.05101

    Pith/arXiv arXiv 2017

  31. [31]

    & Lynden-Bell, R

    Lynden-Bell, D. & Lynden-Bell, R. M. 1995, Monthly No- tices of the Royal Astronomical Society, 275, 429, _eprint: https://academic.oup.com/mnras/article-pdf/275/2/429/18539760/mnras275- 0429.pdf

    1995

  32. [32]

    R., Munn, J

    Majewski, S. R., Munn, J. A., & Hawley, S. L. 1996, ApJ, 459, L73

    1996

  33. [33]

    D., et al

    Martinez-Delgado, D., Stein, M., Sakowska, J. D., et al. 2025, arXiv e-prints, arXiv:2504.02071 Martínez-Delgado, D., Cooper, A. P., Román, J., et al. 2023, Astronomy and Astrophysics, 671, A141, aDS Bibcode: 2023A&A...671A.141M

    arXiv 2025

  34. [34]

    2018, Journal of Open Source Software, 3, 861

    McInnes, L., Healy, J., Saul, N., & Großberger, L. 2018, Journal of Open Source Software, 3, 861

    2018

  35. [35]

    Merrifield, M. R. & Kuijken, K. 1998, MNRAS, 297, 1292, _eprint: astro- ph/9803053 Miró-Carretero, J., Martínez-Delgado, D., Gómez-Flechoso, M. A., et al. 2024, A&A, 691, A196

    arXiv 1998

  36. [36]

    2021, Nature Biotechnology, 39, 765, pub- lisher: Nature Publishing Group

    Narayan, A., Berger, B., & Cho, H. 2021, Nature Biotechnology, 39, 765, pub- lisher: Nature Publishing Group

    2021

  37. [37]

    & Bonaca, A

    Nibauer, J. & Bonaca, A. 2025, The Astrophysical Journal Letters, 985, L22, publisher: The American Astronomical Society

    2025

  38. [38]

    Nibauer, J., Bonaca, A., & Johnston, K. V . 2023, The Astrophysical Journal, 954, 195, publisher: The American Astronomical Society

    2023

  39. [39]

    Nikutta, R., Fitzpatrick, M., Scott, A., & Weaver, B. A. 2020, Astronomy and Computing, 33, 100411

    2020

  40. [40]

    Pearson, S., Küpper, A. H. W., Johnston, K. V ., & Price-Whelan, A. M. 2015, The Astrophysical Journal, 799, 28, publisher: The American Astronomical Society

    2015

  41. [41]

    M., Hogg, D

    Pearson, S., Price-Whelan, A. M., Hogg, D. W., et al. 2022, The Astrophysical Journal, 941, 19, publisher: The American Astronomical Society

    2022

  42. [42]

    M., & Johnston, K

    Pearson, S., Price-Whelan, A. M., & Johnston, K. V . 2017, Nature Astronomy, 1, 633, _eprint: 1703.04627

    Pith/arXiv arXiv 2017

  43. [43]

    Sanderson, R. E. & Helmi, A. 2013, Monthly Notices of the Royal Astro- nomical Society, 435, 378, _eprint: https://academic.oup.com/mnras/article- pdf/435/1/378/3879385/stt1307.pdf

    2013

  44. [44]

    J., Finkbeiner, D

    Schlegel, D. J., Finkbeiner, D. P., & Davis, M. 1998, ApJ, 500, 525

    1998

  45. [45]

    Shannon, C. E. 1948, Bell Labs Technical Journal, 27, 623, aDS Bibcode: 1948BSTJ...27..623S

    1948

  46. [46]

    2014, Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps

    Simonyan, K., Vedaldi, A., & Zisserman, A. 2014, Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps

    2014

  47. [47]

    2022, A&A, 662, A124

    Sola, E., Duc, P.-A., Richards, F., et al. 2022, A&A, 662, A124

    2022

  48. [48]

    Stetson, P. B. 1987, PASP, 99, 191

    1987

  49. [49]

    R., Bullock, J

    Stewart, K. R., Bullock, J. S., Wechsler, R. H., Maller, A. H., & Zentner, A. R. 2008, ApJ, 683, 597 Sánchez, H. D., Martin, G., Damjanov, I., et al. 2023, Monthly Notices of the Royal Astronomical Society, 521, 3861

    2008

  50. [50]

    2021, Monthly Notices of the Royal Astronomical Society, 509, 3966–3988

    Walmsley, M., Lintott, C., Géron, T., et al. 2021, Monthly Notices of the Royal Astronomical Society, 509, 3966–3988

    2021

  51. [51]

    W., Galloway, M

    Willett, K. W., Galloway, M. A., Bamford, S. P., et al. 2017, Monthly Notices of the Royal Astronomical Society, 464, 4176

    2017

  52. [52]

    D., Johnston, K

    Yavetz, T. D., Johnston, K. V ., Pearson, S., Price-Whelan, A. M., & Hamilton, C. 2023, The Astrophysical Journal, 954, 215, publisher: The American As- tronomical Society Article number, page 12 Ernesto Benitez-Walz et al.: Contrastive Learning of Extragalactic Stellar Streams Appendix A: Calibrating the Tiers Tierg-bandr-bandi-band 1 PGC010807 PGC008973...

    2023