Pith. sign in

REVIEW 5 major objections 6 minor 49 references

This paper reports that a feature-alignment loss built on optimal transport and top-k soft matching lifts simulation-trained galaxy morphology classification on real survey images from ~46% to ~87% accuracy, with macro-F1 from 0.30 to 0.63

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 20:40 UTC pith:UBJDF5UI

load-bearing objection The top-k OT soft-matching loss is a genuine idea and the domain-AUC evidence shows real alignment, but the headline 87% accuracy is selected from the target test set, so the numbers aren't trustworthy yet. the 5 major comments →

arxiv 2511.18590 v1 pith:UBJDF5UI submitted 2025-11-23 astro-ph.GA cs.LGhep-th

From Simulations to Surveys: Domain Adaptation for Galaxy Observations

Kaley Brauer , Aditya Prasad Dash , Meet J. Vyas , Ahmed Salim , Stiven Briand Massala This is my paper
classification astro-ph.GA cs.LGhep-th
keywords galaxy morphology classificationdomain adaptationoptimal transporttop-k matchingsimulation-to-survey transferclass imbalancefocal losslatent space alignment
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 is trying to establish that a model trained on simulated galaxy images can accurately classify real survey galaxies once the two domains are aligned in feature space, despite differences in image quality, noise, selection, and demographics. Its central claim is that a domain-alignment loss combining entropic optimal transport with a top-k penalty on the worst-matched source–target pairs produces a latent space where simulated and real galaxies are nearly indistinguishable. With this loss, target accuracy on a three-way elliptical/spiral/irregular task rises from about 46% at no adaptation to about 87%, macro-F1 from 0.30 to 0.63, and domain-discrimination AUC approaches 0.5, the ideal for mixed domains. If correct, this matters because future surveys will image billions of galaxies, and simulations with ground-truth physical labels could replace costly and incomplete human labeling.

Core claim

The core discovery is that feature-level alignment with entropic optimal transport plus a focused top-k soft-matching term—not batch-wide averaging over all pairs—is what lets a simulation-trained morphology classifier transfer to real survey images. Using Euclidean distance on L2-normalized features, scheduled alignment weights, and trainable loss weights, the paper reports 87.3% target accuracy and 0.626 macro-F1, up from 46.8% accuracy and 0.298 macro-F1 with no adaptation, with domain AUC around 0.51–0.53. The top-k mechanism forces alignment of the hardest-to-match source features, discouraging arbitrary class-to-class couplings such as spirals aligned to ellipticals, and this is credit

What carries the argument

The central mechanism is a composite optimal-transport alignment loss on the feature embeddings: a global entropic optimal-transport distance between source and target distributions, a soft-matching term that maps source features through the transport plan to target features, and a sparse top-k penalty on the k largest per-source distances to the target set. This focuses the domain loss on the worst-aligned instances. It is combined with a supervised focal loss with effective-number class weights and a logit-scaling term for rare classes, plus a warmup period and a trainable weighting scheme that lets the network set the alignment strength during training.

Load-bearing premise

The pipeline assumes the morphology classes defined by strict volunteer vote thresholds on real survey images mean the same thing as the morphology classes in the simulation; if those label conventions disagree, the reported 87% target accuracy measures agreement with a particular labeling scheme rather than physical morphology.

What would settle it

Take a held-out set of real survey galaxy images, have professional astronomers independently assign elliptical/spiral/irregular labels without using the volunteer vote thresholds, and run the adapted model on them; if accuracy drops toward the no-adaptation baseline of about 46%, the claimed transfer is largely an artifact of label-system compatibility rather than true domain alignment.

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

If this is right

  • A simulation-trained morphology classifier can reach roughly 87% accuracy and 0.63 macro-F1 on real survey images when the alignment loss is active, compared with about 46% accuracy and 0.30 macro-F1 without it.
  • The top-k soft-matching term, rather than the distance metric alone, is credited with most of the improvement; it appears across all three backbones tested.
  • Source and target latent spaces become almost indistinguishable after adaptation, with domain AUC near 0.5, suggesting the visible domain shift rather than label mismatch is the main barrier being removed.
  • Domain-adversarial training gives competitive but slightly lower accuracy (about 86.5%), so the OT/top-k alignment is at least as effective as adversarial alignment while being more stable.
  • Because the method is unsupervised on the target side, it could be applied to survey galaxies without morphology labels, which is the regime relevant for the next generation of large imaging surveys.

Where Pith is reading between the lines

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

  • If the label-compatibility assumption is right, the same alignment recipe should transfer to multi-task physical properties such as stellar mass, star formation rate, and AGN status, which the paper lists as future work but does not test; the gains may not be as large because those labels have different noise and selection properties.
  • A direct per-class breakdown would reveal whether the top-k loss genuinely aligns the rare irregular class or simply inflates majority-class accuracy; the paper's macro-F1 of 0.63 leaves room for the irregular class to be the weak spot.
  • Near-0.5 domain AUC is treated as success, but indistinguishable feature distributions do not by themselves guarantee a useful shared representation; a downstream test on physical-property prediction would separate domain mixing from task-relevant alignment.
  • A cheap falsification: evaluate the adapted model on a real survey sample labeled by professional astronomers rather than volunteer thresholds; if accuracy falls toward the no-adaptation baseline, the apparent transfer is partly an artifact of label-system compatibility.

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

5 major / 6 minor

Summary. The manuscript proposes a domain adaptation pipeline for galaxy morphology classification, training on TNG50 SKIRT mock observations and evaluating on SDSS images with Galaxy Zoo labels. It compares CNN, E(2)-steerable CNN, and ResNet-18 backbones with focal loss and effective-number class weighting, and adds feature-level alignment losses built from GeomLoss (Sinkhorn OT, energy distance, Gaussian MMD, and related metrics), including a composite loss with top-k soft matching. The central claim is that this combination raises target accuracy from ~46.8% to ~87.3% and macro-F1 from ~0.298 to ~0.626, with domain AUC near 0.5, indicating strong latent-space mixing. The paper is a preliminary workshop-style study.

Significance. If the central claim is robust, the paper would make a useful empirical contribution to simulation-to-survey transfer for morphology classification, a timely problem given upcoming large surveys. Strengths include the use of public TNG50 and SDSS/Galaxy Zoo data, a released code repository, and exploration of several OT-based alignment losses with a top-k variant. However, the headline numerical claims are not yet supported because the evaluation protocol selects the reporting epoch using target test performance, no held-out target validation is described, and no error bars or multiple seeds are reported. The reported improvement may therefore be inflated by test-set overfitting and unequal training budgets. The label-compatibility assumption between TNG50 and Galaxy Zoo morphologies also needs explicit validation before the physical interpretation of the accuracy numbers can be accepted.

major comments (5)
  1. [Section 4 and Section 3] The headline result is selected at 'peak performance (epoch 197)' in Section 4. This indicates that the target test set was used to select the reporting epoch. Section 3 describes early stopping based on training loss and does not mention any held-out target validation set. This protocol inflates target accuracy and makes the comparison with the baseline unfair, since the baseline 'exhibits significant instability and terminates early' and therefore may not receive the same effective training budget. A proper held-out target validation set (or cross-validation) must be used for early stopping and hyperparameter selection, and final numbers should be reported on a truly held-out target test set.
  2. [Section 4, Fig. 3] The confusion matrix in Fig. 3 shows that the irregular class is very poorly handled after adaptation: using the 'After DA' matrix, irregular recall is 4/(22+4+17) = 9.3%, whereas the baseline irregular recall is 24/(1+24+18) = 55.8%. Thus the overall accuracy gain is driven by the majority spiral class, and the macro-F1 of 0.626 masks a serious degradation of the rare irregular class. The paper explicitly lists the irregular class as a next step in Section 5, but the current claim of a 'dramatic improvement' in macro-F1 is misleading. Per-class precision/recall should be reported and discussed as a limitation of the proposed method.
  3. [Section 3 and Section 4] No error bars or multiple seeds are reported for any experiment. The difference between the best Euclidean method (87.3%) and DANN (86.5%) is within plausible run-to-run variability. Hyperparameters such as lambda_D, k, the Sinkhorn blur schedule, and the distance metric are listed as tunable in Section 3, but no validation strategy is described; they appear to be selected on the target test set. This makes the quantitative claims fragile. The authors should run multiple seeds, report mean +/- std, and select hyperparameters/epochs using a validation split independent of the test split.
  4. [Section 4, Fig. 4] The paper uses 'domain AUC near 0.5' as evidence of strong latent-space mixing. Since the training objective in Section 3 explicitly minimizes distributional discrepancies between source and target features, a near-0.5 domain AUC is expected by construction and is not independent evidence of semantic alignment. It is also possible that aggressive alignment collapses class structure (consistent with the poor irregular-class recall). The authors should report class-conditional alignment statistics or inspect the transport plan to show that alignment preserves semantic class information.
  5. [Section 2] The paper states: 'We also assume simulation-based and Galaxy Zoo morphologies provide compatible labels for the same classes.' This assumption is load-bearing because the supervised loss trains on TNG50 labels, while the target accuracy is measured against Galaxy Zoo thresholds. If the two label systems disagree (e.g., on what makes a galaxy irregular), the reported 87.3% measures agreement with a particular Galaxy Zoo threshold rather than physical morphology. The authors should provide quantitative evidence of label compatibility, such as a human-inspection comparison or a consistency check on a common subset, before claiming that the adapted classifier recovers physical morphology.
minor comments (6)
  1. [Abstract and Section 4] Abstract quotes '~46% (~30%)' while Section 4 reports 46.8% and 0.298. Use consistent notation (accuracy and macro-F1) throughout.
  2. [Equation (1)] The definition of P_lambda is incomplete: the Sinkhorn regularization parameter is not defined, and the relationship between C_ij and d_i should be stated more carefully. Also, 'd(1) >= ... >= d(k)' are described as the k largest per-source closest-target distances, but this is a maximization, not a minimization; clarify the ordering convention.
  3. [Figure 2] The two panels are not clearly labeled in the text; the left panel shows target accuracy and macro-F1, and the right panel shows loss curves. Also, the 'negative loss range' for trainable Euclidean variants is not explained; specify what loss is plotted.
  4. [Section 3] The manuscript says GeomLoss is extended with 46 distance/similarity measures but only 12 are benchmarked. State explicitly which 12 are used in the main experiments and why the others are omitted.
  5. [Section 2] The notation 'M_star > 10^9.5 M_sun' is missing a superscript in the text. Also, 'lambda_Edd = Mdot/Mdot_Edd' formatting is broken.
  6. [Section 5] The discussion lists the irregular class and distance-aware schedulers as next steps, which is good, but it does not acknowledge the evaluation-protocol limitations raised above. Add a limitations paragraph or temper the conclusions.

Circularity Check

1 steps flagged

Minor by-construction diagnostic: low domain AUC is the alignment objective itself; target accuracy remains an external benchmark.

specific steps
  1. self definitional [Section 4, Figure 4 caption]
    "The Baseline (left, AUC=1.00) shows distinct domain separation, while Euclidean variants (center/right, AUC ≈0.51) achieve effective domain alignment where source and target distributions are indistinguishable. In an ideal domain-invariant representation, AUC = 0.5."

    The reported AUC≈0.5 is the quantity the alignment losses are explicitly designed to minimize: DANN trains a GRL domain discriminator to be fooled, and the OT/Sinkhorn term minimizes a distributional discrepancy between source and target features. Reporting near-random domain-discriminator AUC as evidence of 'strong latent-space mixing' therefore measures the optimization objective itself, not an independent outcome. It is a valid sanity check, but it is by construction; the load-bearing external evidence is the held-out target accuracy (87.3%), which is not encoded in the loss. Hence minor circularity only.

full rationale

The paper's central claim is empirical: adding feature-level OT/top-k alignment losses improves target accuracy on SDSS Galaxy Zoo labels from ~46% to ~87%. That accuracy is measured on target labels that are never used in training, so it is not equivalent to the inputs. The domain-alignment losses minimize domain discrepancy; observing that the domain-discriminator AUC drops toward 0.5 is a sanity check of the objective, not an independent confirmation. This is the only mildly circular element. The more serious concern is evaluation methodology: the reported 'peak performance (epoch 197)' appears to select the best epoch using target test curves, and no held-out target validation set is described, which can inflate accuracy; this is a correctness/robustness risk, not a circularity. No self-citations are load-bearing; references to prior domain-adaptation work are external and the method is benchmarked against real target labels. Overall circularity is low.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central empirical claim rests on the covariate-shift and label-compatibility assumptions, plus several tuned hyperparameters that are not fully specified. No new physical entities are introduced.

free parameters (5)
  • lambda_D (domain alignment weight) = 0.1 (fixed grid)
    Section 3: lambda=0.1 used for main experiments; tuned over a small grid.
  • k (top-k matching size) = not stated
    Section 3: top-k penalty uses k largest per-source distances; value not given, tuned implicitly.
  • Sinkhorn blur sigma schedule = not stated
    Section 3: blur annealed from sigma_initial to sigma_final for coarse-to-fine OT; endpoints not reported.
  • OT composite weights lambda_OT, lambda_match, lambda_topk = not stated
    Section 3: L_OT = lambda_OT d_lambda + lambda_match MSE + lambda_topk (1/k) sum d(ell); weights not enumerated.
  • Euclidean metric selection = Euclidean
    Section 3: benchmarked 12 metrics and selected Euclidean; selected based on target performance.
axioms (4)
  • domain assumption Covariate shift: p_S(y|x) ≈ p_T(y|x)
    Introduction frames the problem as covariate shift; if false, aligning features can harm labels.
  • domain assumption Simulation and Galaxy Zoo morphology labels are compatible for the same classes
    Section 2: 'We also assume simulation-based and Galaxy Zoo morphologies provide compatible labels', justified only by strict cuts and visual similarity.
  • domain assumption Galaxy Zoo debiased vote thresholds define ground-truth morphology
    Section 2: strict thresholds on vote fractions; label noise from volunteers not modeled.
  • standard math Standard OT/Sinkhorn, backprop, and deep learning background
    Used throughout Methods; no formal proof.

pith-pipeline@v1.3.0-alltime-deepseek · 8761 in / 10609 out tokens · 96414 ms · 2026-08-03T20:40:19.732620+00:00 · methodology

0 comments
read the original abstract

Large photometric surveys will image billions of galaxies, but we currently lack quick, reliable automated ways to infer their physical properties like morphology, stellar mass, and star formation rates. Simulations provide galaxy images with ground-truth physical labels, but domain shifts in PSF, noise, backgrounds, selection, and label priors degrade transfer to real surveys. We present a preliminary domain adaptation pipeline that trains on simulated TNG50 galaxies and evaluates on real SDSS galaxies with morphology labels (elliptical/spiral/irregular). We train three backbones (CNN, $E(2)$-steerable CNN, ResNet-18) with focal loss and effective-number class weighting, and a feature-level domain loss $L_D$ built from GeomLoss (entropic Sinkhorn OT, energy distance, Gaussian MMD, and related metrics). We show that a combination of these losses with an OT-based "top_$k$ soft matching" loss that focuses $L_D$ on the worst-matched source-target pairs can further enhance domain alignment. With Euclidean distance, scheduled alignment weights, and top-$k$ matching, target accuracy (macro F1) rises from $\sim$46% ($\sim$30%) at no adaptation to $\sim$87% ($\sim$62.6%), with a domain AUC near 0.5, indicating strong latent-space mixing.

Figures

Figures reproduced from arXiv: 2511.18590 by Aditya Prasad Dash, Ahmed Salim, Kaley Brauer, Meet J. Vyas, Stiven Briand Massala.

Figure 1
Figure 1. Figure 1: Example source and target data. Our source data are mock observations of galaxies from the TNG50 simulation with known galaxy properties. Our target data are real observations of galaxies. In this proof of concept, we use SDSS observations at z ∼ 0 with known morphologies. Real data differ from simulated data in point spread function, noise, selection function, demographics, and more. Source (TNG50 Simulat… view at source ↗
Figure 2
Figure 2. Figure 2: Training Dynamics. Comparison of stability and loss convergence. Elliptical Irregular Spiral Elliptical Irregular Spiral 3 54 13 1 24 18 11 176 213 Before DA (Acc: 46.8%) Elliptical Irregular Spiral Elliptical Irregular Spiral 70 0 0 22 4 17 14 12 374 After DA (Acc: 87.3%) Baseline DANN Euclidean (Fixed ) Euclidean ( , ) Euclidean (Scheduler + ) 0.0 0.2 0.4 0.6 0.8 1.0 Score 1.00 0.58 0.55 0.53 0.51 0.47 0… view at source ↗
Figure 4
Figure 4. Figure 4: Latent space embedding. The Baseline (left, AUC=1.00) shows distinct domain separation, while Euclidean variants (center/right, AUC≈0.51) achieve effective domain alignment where source and target distributions are indistinguishable. In an ideal domain-invariant representation, AUC = 0.5. most stable alignment. We find that top-k soft-matching with ResNet effectively aligns classes across domains, driving … view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

49 extracted references · 5 linked inside Pith

  1. [1]

    AGARWAL, A

    S. AGARWAL, A. ´CIPRIJANOVI ´C,ANDB. D. NORD,Neural Network Prediction of Strong Lensing Systems with Domain Adaptation and Uncertainty Quantification, arXiv e-prints, (2024), p. arXiv:2411.03334

    Pith/arXiv arXiv 2024

  2. [2]

    AKESON, L

    R. AKESON, L. ARMUS, E. BACHELET, V. BAILEY,ANDL. B.ET AL.,The wide field infrared survey telescope: 100 hubbles for the 2020s, 2019

    2019

  3. [3]

    J. A. BALDWIN, M. M. PHILLIPS,ANDR. TERLEVICH,Classification parameters for the emission-line spectra of extragalactic objects, Publications of the Astronomical Society of the Pacific, 93 (1981), p. 5

    1981

  4. [4]

    E. C. BELLM, S. R. KULKARNI, M. J. GRAHAM, R. DEKANY,ANDR. M. S.ET AL.,The zwicky transient facility: System overview, performance, and first results, Publications of the Astronomical Society of the Pacific, 131 (2019), p. 018002

    2019

  5. [5]

    CHA,Comprehensive survey on distance/similarity measures between probability density functions, City, 1 (2007), p

    S.-H. CHA,Comprehensive survey on distance/similarity measures between probability density functions, City, 1 (2007), p. 1

    2007

  6. [6]

    ´CIPRIJANOVI ´C, A

    A. ´CIPRIJANOVI ´C, A. LEWIS, K. PEDRO, S. MADIREDDY, B. NORD, G. N. PERDUE,AND S. M. WILD,DeepAstroUDA: semi-supervised universal domain adaptation for cross-survey galaxy morphology classification and anomaly detection, Machine Learning: Science and Technology, 4 (2023), p. 025013

    2023

  7. [7]

    COURTY, R

    N. COURTY, R. FLAMARY, D. TUIA,ANDA. RAKOTOMAMONJY,Optimal transport for domain adaptation, IEEE Transactions on Pattern Analysis and Machine Intelligence, 39 (2014), pp. 1853–1865

    2014

  8. [8]

    CSURKA,Domain adaptation for visual applications: A comprehensive survey, arXiv preprint arXiv:1702.05374, (2017)

    G. CSURKA,Domain adaptation for visual applications: A comprehensive survey, arXiv preprint arXiv:1702.05374, (2017)

    Pith/arXiv arXiv 2017

  9. [9]

    Y. CUI, M. JIA, T.-Y. LIN, Y. SONG,ANDS. BELONGIE,Class-balanced loss based on effective number of samples, in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 9260–9269

    2019

  10. [10]

    J. DENG, W. DONG, R. SOCHER, L.-J. LI, K. LI,ANDL. FEI-FEI,Imagenet: A large- scale hierarchical image database, in 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248–255

    2009

  11. [11]

    R. DING, J. LIU, K. HUA, X. WANG, X. ZHANG, M. SHAO, Y. CHEN,ANDJ. CHEN, Leveraging data mining, active learning, and domain adaptation for efficient discovery of advanced oxygen evolution electrocatalysts, Science Advances, 11 (2025), p. eadr9038

    2025

  12. [12]

    Y. FAN, S. LYU, Y. YING,ANDB.-G. HU,Learning with average top-k loss, in Neural Information Processing Systems, 2017

    2017

  13. [13]

    FARAHANI, S

    A. FARAHANI, S. VOGHOEI, K. RASHEED,ANDH. R. ARABNIA,A brief review of domain adaptation, Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, (2021), pp. 877–894

    2020

  14. [14]

    FEYDY, T

    J. FEYDY, T. SÉJOURNÉ, F.-X. VIALARD, S.-I. AMARI, A. TROUVE,ANDG. PEYRÉ, Interpolating between optimal transport and mmd using sinkhorn divergences, in The 22nd International Conference on Artificial Intelligence and Statistics, 2019, pp. 2681–2690

    2019

  15. [15]

    H. A. FLEWELLING, E. A. MAGNIER, K. C. CHAMBERS, J. N. HEASLEY,ANDC. H.ET AL., The pan-starrs1 database and data products, The Astrophysical Journal Supplement Series, 251 (2020), p. 7

    2020

  16. [16]

    GANIN ANDV

    Y. GANIN ANDV. LEMPITSKY,Unsupervised domain adaptation by backpropagation, in Proceedings of the 32nd International Conference on International Conference on Machine Learning - V olume 37, ICML’15, JMLR.org, 2015, p. 1180–1189

    2015

  17. [17]

    GILDA, A.DEMATHELIN, S

    S. GILDA, A.DEMATHELIN, S. BELLSTEDT,ANDG. RICHARD,Unsupervised Domain Adaptation for Constraining Star Formation Histories, Astronomy, 3 (2024), pp. 189–207. 6

    2024

  18. [18]

    GUAN ANDM

    H. GUAN ANDM. LIU,Domain adaptation for medical image analysis: A survey, IEEE Transactions on Biomedical Engineering, 69 (2021), pp. 1173–1185

    2021

  19. [19]

    R. E. HART, S. P. BAMFORD, K. W. WILLETT, K. L. MASTERS, C. CARDAMONE, C. J. LINTOTT, R. J. MACKAY, R. C. NICHOL, C. K. ROSSLOWE, B. D. SIMMONS,ANDR. J. SMETHURST,Galaxy zoo: comparing the demographics of spiral arm number and a new method for correcting redshift bias, Monthly Notices of the Royal Astronomical Society, 461 (2016), pp. 3663–3682

    2016

  20. [20]

    K. HE, X. ZHANG, S. REN,ANDJ. SUN,Deep residual learning for image recognition, in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

    2016

  21. [21]

    JAKUBIK, M

    J. JAKUBIK, M. MUSZYNSKI, M. VÖSSING, N. KÜHL,ANDT. BRUNSCHWILER,Unsu- pervised Domain Adaptation for Geospatial Semantic Segmentation, in AGU Fall Meeting Abstracts, vol. 2023, Dec. 2023, pp. IN53A–07

    2023

  22. [22]

    G. KANG, L. JIANG, Y. YANG,ANDA. HAUPTMANN,Contrastive adaptation network for unsupervised domain adaptation, 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), (2019), pp. 4888–4897

    2019

  23. [23]

    KENDALL, Y

    A. KENDALL, Y. GAL,ANDR. CIPOLLA,Multi-task learning using uncertainty to weigh losses for scene geometry and semantics, CoRR, abs/1705.07115 (2017)

    Pith/arXiv arXiv 2017

  24. [24]

    J. LI, R. XU, X. LIU, J. MA, B. LI, Q. ZOU, J. MA,ANDH. YU,Domain adaptation based object detection for autonomous driving in foggy and rainy weather, IEEE Transactions on Intelligent Vehicles, 10 (2025), pp. 900–911

    2025

  25. [25]

    LINTOTT, K

    C. LINTOTT, K. SCHAWINSKI, S. BAMFORD, A. SLOSAR, K. LAND, D. THOMAS, E. ED- MONDSON, K. MASTERS, R. C. NICHOL, M. J. RADDICK, A. SZALAY, D. ANDREESCU, P. MURRAY,ANDJ. VANDENBERG,Galaxy zoo 1: data release of morphological classifica- tions for nearly 90,0000 galaxies, Monthly Notices of the Royal Astronomical Society, 410 (2011), pp. 166–178

    2011

  26. [26]

    C. J. LINTOTT, K. SCHAWINSKI, A. SLOSAR, K. LAND, S. BAMFORD, D. THOMAS, M. J. RADDICK, R. C. NICHOL, A. SZALAY, D. ANDREESCU, P. MURRAY,ANDJ. VANDENBERG, Galaxy zoo: morphologies derived from visual inspection of galaxies from the sloan digital sky survey, Monthly Notices of the Royal Astronomical Society, 389 (2008), pp. 1179–1189

    2008

  27. [27]

    X. LIU, C. YOO, F. XING, H. OH, G. FAKHRI, J.-W. KANG,ANDJ. WOO,Deep unsupervised domain adaptation: A review of recent advances and perspectives, APSIPA Transactions on Signal and Information Processing, (2022)

    2022

  28. [28]

    NELSON, A

    D. NELSON, A. PILLEPICH, V. SPRINGEL, R. PAKMOR, R. WEINBERGER, S. GENEL, P. TORREY, M. VOGELSBERGER, F. MARINACCI,ANDL. HERNQUIST,First results from the TNG50 simulation: galactic outflows driven by supernovae and black hole feedback, MNRAS, 490 (2019), pp. 3234–3261

    2019

  29. [29]

    RUBINOBSERVATORY,Legacy Survey of Space and Time Data Preview 1 [Data set], (2025)

    NSF-DOE VERAC. RUBINOBSERVATORY,Legacy Survey of Space and Time Data Preview 1 [Data set], (2025)

    2025

  30. [30]

    PANDYA, P

    S. PANDYA, P. PATEL, B. D. NORD, M. WALMSLEY,ANDA. ´CIPRIJANOVI ´C,Sidda : Sinkhorn dynamic domain adaptation for image classification with equivariant neural networks, Machine Learning: Science and Technology, 6 (2025), p. 035032

    2025

  31. [31]

    PANDYA, P

    S. PANDYA, P. PATEL, F. O,ANDJ. BLAZEK,E(2) equivariant neural networks for robust galaxy morphology classification, 2023

    2023

  32. [32]

    PARUL, S

    H. PARUL, S. GLEYZER, P. REDDY,ANDM. W. TOOMEY,Domain Adaptation in Application to Gravitational Lens Finding, The Astrophysical Journal, 990 (2025), p. 47

    2025

  33. [33]

    PILLEPICH, D

    A. PILLEPICH, D. NELSON, V. SPRINGEL, R. PAKMOR, P. TORREY, R. WEINBERGER, M. VOGELSBERGER, F. MARINACCI, S. GENEL, A.VAN DERWEL,ANDL. HERNQUIST, First results from the TNG50 simulation: the evolution of stellar and gaseous discs across cosmic time, MNRAS, 490 (2019), pp. 3196–3233. 7

    2019

  34. [34]

    RODRIGUEZ-GOMEZ, G

    V. RODRIGUEZ-GOMEZ, G. F. SNYDER, J. M. LOTZ, D. NELSON, A. PILLEPICH, V. SPRINGEL, S. GENEL, R. WEINBERGER, S. TACCHELLA, R. PAKMOR, P. TORREY, F. MARINACCI, M. VOGELSBERGER, L. HERNQUIST,ANDD. A. THILKER,The optical morphologies of galaxies in the IllustrisTNG simulation: a comparison to Pan-STARRS obser- vations, MNRAS, 483 (2019), pp. 4140–4159

    2019

  35. [35]

    RONCOLI, A

    A. RONCOLI, A. ´CIPRIJANOVI ´C, M. VOETBERG, F. VILLAESCUSA-NAVARRO,AND B. NORD,Domain Adaptive Graph Neural Networks for Constraining Cosmological Parameters Across Multiple Data Sets, arXiv e-prints, (2023), p. arXiv:2311.01588

    Pith/arXiv arXiv 2023

  36. [36]

    SAITO, D

    K. SAITO, D. KIM, P. TETERWAK, S. SCLAROFF, T. DARRELL,ANDK. SAENKO,Tune it the right way: Unsupervised validation of domain adaptation via soft neighborhood density, in 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 9164–9173

    2021

  37. [37]

    SCARAMELLA, J

    R. SCARAMELLA, J. AMIAUX, Y. MELLIER, C. BURIGANA,ANDC. S. C.ET AL.,Euclid preparation, Astronomy & Astrophysics, 662 (2022), p. A112

    2022

  38. [38]

    SCHAWINSKI, C

    K. SCHAWINSKI, C. M. URRY, S. VIRANI, P. COPPI, S. P. BAMFORD, E. TREISTER, C. J. LINTOTT, M. SARZI, W. C. KEEL, S. KAVIRAJ, C. N. CARDAMONE, K. L. MASTERS, N. P. ROSS, D. ANDREESCU, P. MURRAY, R. C. NICHOL, M. J. RADDICK, A. SLOSAR, A. S. SZALAY, D. THOMAS,ANDJ. VANDENBERG,Galaxy zoo: The fundamentally different co-evolution of supermassive black holes a...

    2010

  39. [39]

    C. SHUI, Q. CHEN, J. WEN, F. ZHOU, C. GAGNÉ,ANDB. WANG,A novel domain adaptation theory with jensen–shannon divergence, Knowledge-Based Systems, 257 (2022), p. 109808

    2022

  40. [40]

    B. SUN, J. FENG,ANDK. SAENKO,Correlation alignment for unsupervised domain adapta- tion, in Domain adaptation in computer vision applications, Springer, 2017, pp. 153–171

    2017

  41. [41]

    SWIERC, M

    P. SWIERC, M. ZHAO, A. ´CIPRIJANOVI ´C,ANDB. NORD,Domain Adaptation for Measure- ments of Strong Gravitational Lenses, arXiv e-prints, (2023), p. arXiv:2311.17238

    Pith/arXiv arXiv 2023

  42. [42]

    D. TUIA, C. PERSELLO,ANDL. BRUZZONE,Domain adaptation for the classification of remote sensing data: An overview of recent advances, IEEE Geoscience and Remote Sensing Magazine, 4 (2016), pp. 41–57

    2016

  43. [43]

    VILALTA, K

    R. VILALTA, K. D. GUPTA, D. BOUMBER,ANDM. M. MESKHI,A general approach to domain adaptation with applications in astronomy, Publications of the Astronomical Society of the Pacific, 131 (2019), p. 108008

    2019

  44. [44]

    WANG ANDW

    M. WANG ANDW. DENG,Deep visual domain adaptation: A survey, Neurocomputing, 312 (2018), p. 135–153

    2018

  45. [45]

    WEILER ANDG

    M. WEILER ANDG. CESA,General E(2)-Equivariant Steerable CNNs, in Conference on Neural Information Processing Systems (NeurIPS), 2019

    2019

  46. [46]

    K. W. WILLETT, C. J. LINTOTT, S. P. BAMFORD, K. L. MASTERS, B. D. SIMMONS, K. R. V. CASTEELS, E. M. EDMONDSON, L. F. FORTSON, S. KAVIRAJ, W. C. KEEL, T. MELVIN, R. C. NICHOL, M. J. RADDICK, K. SCHAWINSKI, R. J. SIMPSON, R. A. SKIBBA, A. M. SMITH,ANDD. THOMAS,Galaxy zoo 2: detailed morphological classifications for 304122 galaxies from the sloan digital sk...

    2013

  47. [47]

    WILSON ANDD

    G. WILSON ANDD. J. COOK,A survey of unsupervised deep domain adaptation, ACM Trans. Intell. Syst. Technol., 11 (2020)

    2020

  48. [48]

    ´CIPRIJANOVI ´C, D

    A. ´CIPRIJANOVI ´C, D. KAFKES, K. D. S. JENKINS, G. N. PERDUE, S. MADIREDDY, T. JOHN- STON,ANDB. NORD,Domain adaptation techniques for improved cross-domain study of galaxy mergers, FERMILAB-CONF-20-582-SCD, (2020)

    2020

  49. [49]

    ´CIPRIJANOVI ´C, D

    A. ´CIPRIJANOVI ´C, D. KAFKES, G. SNYDER, F. J. SÁNCHEZ, G. N. PERDUE, K. PEDRO, B. NORD, S. MADIREDDY,ANDS. M. WILD,Deepadversaries: examining the robustness of deep learning models for galaxy morphology classification, Machine Learning: Science and Technology, 3 (2022), p. 035007. 8

    2022