Pith. sign in

REVIEW 3 major objections 7 minor 51 references

Image-Based Multi-Survey Classification of Light Curves with a Pre-Trained Vision Transformer

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper argues that a shared-weight vision Transformer processing ZTF and ATLAS light curves as separate streams classifies transients and variables better than single-survey models or after-the-fact fusion, with the best variant…

desk verdict A careful empirical comparison of multi-survey fusion strategies with a useful negative result, but the headline claim of a best joint architecture is not statistically established. read the letter →

arxiv 2507.11711 v1 pith:6VXTJIWU submitted 2025-07-15 astro-ph.IM cs.CV

Daniel Moreno-Cartagena , Guillermo Cabrera-Vives , Alejandra M. Muñoz Arancibia , Pavlos Protopapas , Francisco Förster , Márcio Catelan , A. Bayo , Pablo A. Estévez
show 5 more authors
P. Sánchez-Sáez Franz E. Bauer M. Pavez-Herrera L. Hernández-García Gonzalo Rojas
This is my paper
classification astro-ph.IMcs.CV
keywords photometricclassificationmulti-surveyfusionlightcurvesvisiontransformerSwinV2ZTFATLAStime-domainastronomy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that photometric classification—sorting variable stars and transients into classes from their brightness-over-time light curves—improves when light curves from two different surveys are modeled together rather than separately. The method converts each multi-band light curve into an RGB image, runs ZTF and ATLAS images through two shared-weight copies of a pre-trained Swin Transformer V2, and combines their pooled embeddings in a joint classifier. On a 21-class dataset with roughly 39,000 objects per survey, the joint architecture reaches a macro F1 of 69.9 ± 1.7, compared with 64.9 for ZTF alone, 54.0 for ATLAS alone, and 68.9 for averaging the two single-survey models' probabilities. The result matters because future surveys will cover the same sky with different cadences and filters, and a single model that can absorb multiple instruments would be simpler to deploy than per-survey pipelines.

What carries the argument

The load-bearing mechanism is a two-stream Swin Transformer V2 with shared weights. Light curves are first turned into RGB images by the Overlay visualization, which plots all photometric bands in one Min-Max-normalized time-flux space, so absolute time and brightness are removed and only shape, color, and relative uncertainty remain. Each survey's image stream passes through the same pre-trained vision Transformer backbone, producing a pooled embedding; per-survey linear classifiers make survey-specific predictions, while the concatenated embeddings feed a joint mix classifier, and the total loss sums the three cross-entropy losses. The shared weights force both surveys through the same feature extractor while the separate streams preserve survey-specific input structure, and the pre-trained ImageNet-21K backbone supplies generic visual features without hand-designed astronomical descriptors.

What would settle it

Train the two-stream architecture again with the per-survey classifier losses removed and shared weights unchanged; if the mix head alone still reaches about 69.9 macro F1, the cross-survey interaction explanation is unsupported. Alternatively, run a paired bootstrap across the five folds comparing MS-SWIN V2-Linear with averaged probabilities; if the difference is not significant, the paper's headline superiority claim over simple fusion collapses.

Watch

Extended reading notes

Core claim

The central discovery is that keeping each survey as an independent input stream, processing both streams with shared weights, and adding a joint classifier yields the best classification performance among the strategies tested. The paper reports macro F1 of 69.9 ± 1.7 for the linear joint classifier and 69.8 ± 1.5 for an MLP variant, beating the Swin V2 single-survey baselines (64.9 ± 1.6 for ZTF, 54.0 ± 1.3 for ATLAS), probability averaging (68.9 ± 1.3), a balanced random forest on concatenated embeddings (68.7 ± 0.9), and a stacked-input early-fusion image (59.0 ± 1.3). A specialized sequence Transformer trained on the same light curves reaches only 57.0 on ZTF and 49.6 on ATLAS, so the pre-trained vision Transformer itself is part of the advantage. The authors attribute the joint model's edge over post-hoc averaging to cross-survey interactions learned during training, supported by per-class recall gains in 19 of 21 classes relative to the ZTF-only baseline.

Load-bearing premise

The load-bearing premise is that the joint model's small advantage over probability averaging comes from learned cross-survey interactions and is statistically real, rather than from the extra per-survey classifiers and auxiliary losses that the same architecture also introduces.

Editorial extensions

If this is right

  • Adding a second survey that is individually weaker still lifts overall classification, so multi-survey training is a practical route to better photometric labels even when the extra data are noisier.
  • Post-hoc fusion (probability averaging or an embedding-level random forest) captures most of the complementary information, meaning much of the benefit of a second survey does not require retraining a joint model.
  • Naive early fusion that stacks all four bands into one image hurts performance, which argues for keeping survey-specific input streams separate rather than concatenating raw inputs.
  • A pre-trained vision Transformer outperforms a specialized sequence Transformer on these multi-band light curves, suggesting image-based representations are a competitive alternative for time-series classification in astronomy.
  • The unified architecture can emit per-survey and joint predictions from one model, simplifying deployment and making it straightforward to add further surveys or modalities later.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 1-point gap between the joint architecture (69.9) and probability averaging (68.9) is smaller than the reported fold-to-fold standard deviations, so the claim that cross-survey interactions drive the gain needs a paired significance test before it is treated as established.
  • The architecture adds two auxiliary per-survey classifiers and a summed loss on top of the shared branches; without an ablation that removes them, the improvement could come from auxiliary-task regularization rather than from interactions between surveys.
  • If the interaction story is real, the same design should give larger gains when the two surveys have more complementary cadences or filters; a test on deliberately mismatched survey pairs would be a direct way to see the effect scale.
  • The vision-Transformer advantage over sequence models may not generalize to very long or irregularly sampled light curves, since the image rendering imposes a fixed normalization and resolution that can discard temporal information.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The manuscript presents an image-based approach to multi-survey photometric classification. Light curves from ZTF and ATLAS are rendered as RGB images following the Overlay method of Moreno-Cartagena et al. (2025), and a pre-trained Swin Transformer V2 is used as the backbone. The authors compare seven strategies: two single-survey baselines (ATAT and Swin V2 on each survey), averaged probabilities of the two single-survey models, a Balanced Random Forest on concatenated embeddings, a stacked-input early-fusion Swin V2, and a proposed multi-survey architecture with two shared-weight Swin V2 streams and a joint linear or MLP classifier. On a class-stratified five-fold split of roughly 39,500 objects across 21 classes, the proposed MS-SwinV2-Linear reaches an F1 of 69.9±1.7, higher than the ZTF-only baseline (64.9±1.6), ATLAS-only (54.0±1.3), averaged probabilities (68.9±1.3), BRF on embeddings (68.7±0.9), and stacked input (59.0±1.3). The authors interpret the gain over probability averaging as evidence that joint training captures cross-survey interactions.

Significance. The paper addresses a timely and practical problem: how to combine heterogeneous multi-survey photometric data for classification using pre-trained vision transformers instead of training from scratch. Its strengths are a class-stratified five-fold evaluation protocol, comparisons across prediction-level, embedding-level, input-level, and joint-model fusion, and the inclusion of per-class confusion matrices that make the complementary-survey effect visible (e.g., higher recall from ATLAS for RSCVn and DSCT). If the reported advantage of the joint architecture were statistically robust, the work would offer a useful design pattern for LSST-era multi-survey classifiers. The main weakness is that the key comparison — joint architecture versus simple probability averaging — is not shown to be significant, and the proposed mechanism of 'cross-survey interactions' is confounded by auxiliary losses and shared-weight training. The paper is honest about its use of prior hyperparameters and does not overclaim reproducibility, but the central claim needs additional statistical support.

major comments (3)
  1. [Section 4, Table 2] The central empirical claim in Section 4 and the abstract — that the multi-survey architecture achieves the best performance among fusion strategies — rests on an F1 advantage of MS-SwinV2-Linear (69.9 ± 1.7) over averaged probabilities (68.9 ± 1.3) of only 1.0 percentage point. The reported standard deviations are of the same order as the difference, and no paired significance test across the five folds is reported; a simple overlap criterion suggests the difference is not significant. Because the comparisons against the single-survey baselines and stacked-input fusion are much larger, the only load-bearing distinction is the one against probability averaging. The authors should either add a paired test or confidence interval on the per-fold differences, or revise the abstract and conclusion to state that the advantage over probability averaging is not statistically resolved.
  2. [Section 2.3, Section 4] Section 2.3 introduces, relative to the two independently trained single-survey baselines used for probability averaging in Section 4, two auxiliary per-survey classifiers, a three-term cross-entropy loss, and shared-weight joint training. The statement in Section 4 that the multi-survey architecture 'captures cross-survey interactions during training' is therefore not identifiable from the experiments: any observed gain could arise from multi-task regularization, from the additional gradient signal of the auxiliary heads, or from joint fine-tuning of the backbone, without requiring cross-survey interactions in the fused representation. An ablation that removes the auxiliary classifiers and their losses (training only the mix classifier on the concatenated embeddings) would be needed to support the mechanistic interpretation. As written, the claim is an over-interpretation and should be softened or supported.
  3. [Section 3.2 / Table 2] The F1-score used in Table 2 and throughout is never defined. Given the strongly imbalanced class counts in Table 1 (e.g., Microlensing with 29 objects versus LPV with 2931), macro-averaged and micro-averaged F1 will differ substantially. Please state explicitly whether F1 is macro-averaged (unweighted mean of per-class F1), weighted by class support, or computed globally, and how the per-class values are aggregated. This is needed to interpret the 'best performance' claim.
minor comments (7)
  1. [Section 2.1] The normalization description is contradictory: 'Min-Max normalization independently to both time and flux values for each light curve, jointly considering all its bands' should clarify whether the min-max scaling is computed per object across all bands pooled together or separately per band.
  2. [Table 2] The model identifiers '3 + 4' for averaged probabilities and BRF on embeddings are awkward; consider using descriptive names such as 'Avg-Prob' and 'BRF-Emb' for readability.
  3. [Figure 2] The multi-survey architecture figure has no caption; add a caption explaining the shared-weight branches, the per-survey classifiers, and the mix classifier.
  4. [Appendix C, Figure 4] The black lines delineating the three main groups (Transient, Stochastic, Periodic) may be hard to distinguish in grayscale printing; consider increasing contrast or adding group labels directly in the figure.
  5. [Introduction] The reference 'Murray et al.' is incomplete; it lacks a year and a publication venue, and should be either fully cited or removed.
  6. [Section 3.1] The ZTF and ATLAS sample sizes differ by 19 objects (39,483 vs 39,464); please state whether the missing ATLAS light curves are due to coverage or quality cuts and confirm that the class-stratified split is identical for the common objects.
  7. [Section 4] The phrase 'incorporating ATLAS data in the multi-survey framework provides additional signals' should be hedged given the significance issue raised in Major Comment 1; consider phrasing as 'may provide' until statistical significance is established.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning: held-out test F1 scores support the fusion comparison; self-citations are method provenance only.

full rationale

The paper's central claim is an empirical comparison of fusion strategies on held-out test folds (Section 4, Table 2), not a quantity derived from a fitted parameter or from the prior work. No equation defines the reported F1 improvement in terms of its own inputs; MS-SwinV2-Linear (69.9 +/- 1.7) and the competing strategies are all evaluated on the same class-stratified test split. The authors draw on Moreno-Cartagena et al. (2025) for the image-generation and Swin V2 hyperparameters ('Following the methodology introduced by Moreno-Cartagena et al. (2025), we adopt the pre-trained SWIN V2 architecture'; 'We used a marker size of 1.0, a line width of 2.0, flux errors, and a learning rate of 5e-6 ... following Moreno-Cartagena et al. (2025)'), but this is transfer of methodology, not evidence for the fusion ranking. The mechanistic interpretation that joint training 'captures cross-survey interactions during training' is not ablated and could be confounded by the auxiliary per-survey classifiers and joint loss described in Section 2.3; that is an unsupported explanation, not circularity. The key comparison is also within the reported error bars, which is a statistical robustness concern rather than a circular-reasoning concern. No step reduces by construction to its own input, so the circularity score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a small set of domain assumptions about label quality, ATLAS forced photometry validity, transferability of ImageNet pre-training, and the image normalization scheme, plus two sets of hyperparameters inherited or tuned on validation. No new entities or physical postulates are introduced.

free parameters (2)
  • Image rendering hyperparameters (marker_size, line_width, learning rate) = marker_size=1.0, line_width=2.0, lr=5e-6
    Taken from Moreno-Cartagena et al. (2025), tuned on MACHO two-band light curves, not re-tuned for ZTF/ATLAS; could bias relative performance of fusion strategies.
  • BRF classifier hyperparameters = n_estimators=500, max_features='log2', max_depth=30, min_samples_split=5
    Selected by grid search on validation folds for the embedding-level fusion strategy (Table 3), affecting that baseline's reported F1.
assumptions (4)
  • domain assumption ALeRCE class labels from crossmatch and expert validation are accurate enough for training and evaluation
    Section 3.1; label noise would affect all models uniformly but could distort class-level conclusions.
  • domain assumption ATLAS forced photometry at ZTF positions yields valid, class-unbiased light curves for the same objects
    Section 3.1; quality cuts (0.5<chi2/N<3, sky magnitude limits) follow Silva-Farfan et al. (2024), but completeness and class-dependent selection are not analyzed.
  • domain assumption ImageNet-21K pre-trained features transfer to light curve images
    Section 2.2; the entire method depends on this transfer, validated only by the reported test F1 scores, not by an external check.
  • domain assumption The min-max normalization that removes absolute time and brightness preserves class-discriminative shape information
    Section 2.1; if absolute timescale or amplitude carries class information, the representation discards it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Image-Based Multi-Survey Classification of Light Curves with a Pre-Trained Vision Transformer." pith.science (2026). https://pith.science/paper/6VXTJIWU

@misc{pith2026250711711,
  author       = {Pith},
  title        = {Pith review of: Image-Based Multi-Survey Classification of Light Curves with a Pre-Trained Vision Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6VXTJIWU}},
  note         = {Machine review of arXiv:2507.11711}
}
read the original abstract

We explore the use of Swin Transformer V2, a pre-trained vision Transformer, for photometric classification in a multi-survey setting by leveraging light curves from the Zwicky Transient Facility (ZTF) and the Asteroid Terrestrial-impact Last Alert System (ATLAS). We evaluate different strategies for integrating data from these surveys and find that a multi-survey architecture which processes them jointly achieves the best performance. These results highlight the importance of modeling survey-specific characteristics and cross-survey interactions, and provide guidance for building scalable classifiers for future time-domain astronomy.

Figures

Figures reproduced from arXiv: 2507.11711 by the authors.

Figure 1
Figure 1. Model input images constructed from forced photometry light curves. Rows show ZTF (g and r bands), ATLAS (c and o bands), and the stacked input with all bands. Images represent flux difference over time, with axes omitted since time and flux are not explicitly used. From left to right, columns show ZTF18aajpcoq (Blazar), ZTF21abfnfdw (SNIa), and ZTF18aampkkf (AGN). anced color representation and avoiding visual bias… view at source ↗
Figure 2
Figure 2. Multi-survey architecture. Transformer architecture tailored for sequential modeling of multi-band astronomical time series. 2.3. Multi-Survey Methodology We explore multiple strategies for combining ZTF and ATLAS data to improve classification performance. As a baseline, we first train separate single-survey SWINV2 models for ZTF and ATLAS. We then perform prediction￾level fusion by averaging the predicted probabil… view at source ↗
Figure 3
Figure 3. SWINV2 architecture. The rounded light blue rectangles highlight the components where the model changes the dimensionality of the information, while the regular rectangles indicate where it remains fixed. The figure is from Moreno-Cartagena et al. (2025). B. Hyperparameter Tuning for the BRF Classifier The BRF classifier hyperparameters were optimized using F1-score on the validation set across five stratified folds… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Confusion matrices for the different strategies explored to combine ZTF and ATLAS data for photometric classification. (a) Single-survey SWINV2 model trained on ZTF light curves. (b) Single-survey SWINV2 model trained on ATLAS light curves. (c) Prediction-level fusion:…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 40 canonical work pages

  1. [1]

    Deep multi-survey classification of variable stars

    Aguirre, C., Pichara, K., and Becker, I. Deep multi-survey classification of variable stars. Monthly Notices of the Royal Astronomical Society, 482 0 (4): 0 5078--5092, 2019

    work page 2019

  2. [2]

    and McEwen, J

    Allam Jr, T. and McEwen, J. D. Paying attention to astronomical transients: introducing the time-series transformer for photometric classification. RAS Techniques and Instruments, 3 0 (1): 0 209--223, 2024

    2024

  3. [3]

    Multiband embeddings of light curves

    Becker, I., Protopapas, P., Catelan, M., and Pichara, K. Multiband embeddings of light curves. A&A, 694: 0 A183, 2025. doi:10.1051/0004-6361/202347461. URL https://doi.org/10.1051/0004-6361/202347461

    work page doi:10.1051/0004-6361/202347461 2025

  4. [4]

    C., Kulkarni, S

    Bellm, E. C., Kulkarni, S. R., Graham, M. J., Dekany, R., Smith, R. M., Riddle, R., Masci, F. J., Helou, G., Prince, T. A., Adams, S. M., et al. The zwicky transient facility: system overview, performance, and first results. Publications of the Astronomical Society of the Pacific, 131 0 (995): 0 018002, 2018

    2018

  5. [5]

    Avocado: Photometric classification of astronomical transients with gaussian process augmentation

    Boone, K. Avocado: Photometric classification of astronomical transients with gaussian process augmentation. The Astronomical Journal, 158 0 (6): 0 257, 2019

    work page 2019

  6. [6]

    M., Bayo, A., Catelan, M., et al

    Cabrera-Vives, G., Moreno-Cartagena, D., Astorga, N., Reyes-Jainaga, I., F \"o rster, F., Huijse, P., Arredondo, J., Arancibia, A. M., Bayo, A., Catelan, M., et al. Atat: Astronomical transformer for time series and tabular data. Astronomy & Astrophysics, 689: 0 A289, 2024

    2024

  7. [7]

    Transformer-Based Astronomical Time Series Model with Uncertainty Estimation for Detecting Misclassified Instances

    C \'a diz-Leyton, M., Cabrera-Vives, G., Protopapas, P., Moreno-Cartagena, D., and Donoso-Oliva, C. Transformer-based astronomical time series model with uncertainty estimation for detecting misclassified instances. In LatinX in AI workshop, 41 st International Conference on Machine Learning ( ICML ), PMLR 235 , Vienna, Austria, 2024 a . URL https://arxiv...

    work page Pith review arXiv 2024

  8. [8]

    Uncertainty estimation for time series classification: Exploring predictive uncertainty in transformer-based models for variable stars

    C \'a diz-Leyton, M., Cabrera-Vives, G., Protopapas, P., Moreno-Cartagena, D., Donoso-Oliva, C., and Becker, I. Uncertainty estimation for time series classification: Exploring predictive uncertainty in transformer-based models for variable stars. Astronomy & Astrophysics, 2024 b . Accepted for publication

    work page 2024

Show all 51 references
  1. [9]

    End-to-end object detection with transformers

    Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., and Zagoruyko, S. End-to-end object detection with transformers. In European conference on computer vision, pp.\ 213--229. Springer, 2020

    2020

  2. [10]

    A., Huijse, P., Protopapas, P., Reyes, I., Mart \' nez-Palomera, J., and Donoso, C

    Carrasco-Davis, R., Cabrera-Vives, G., F \"o rster, F., Est \'e vez, P. A., Huijse, P., Protopapas, P., Reyes, I., Mart \' nez-Palomera, J., and Donoso, C. Deep learning for image sequence classification of astronomical events. Publications of the Astronomical Society of the P...

    2019

  3. [11]

    Stellar Variability in Ground-Based Photometric Surveys: AN Overview

    Catelan , M. Stellar Variability in Ground-Based Photometric Surveys: AN Overview . In Memorie della Societa Astronomica Italiana, volume 94, pp.\ 56, December 2023. doi:10.36116/MEMSAIT_94N4.2023.56

    2023 doi

  4. [12]

    Transformer tracking

    Chen, X., Yan, B., Zhu, J., Wang, D., Yang, X., and Lu, H. Transformer tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 8126--8135, 2021

    2021

  5. [13]

    Astromer-a transformer-based embedding for the representation of light curves

    Donoso-Oliva, C., Becker, I., Protopapas, P., Cabrera-Vives, G., Vishnu, M., and Vardhan, H. Astromer-a transformer-based embedding for the representation of light curves. Astronomy & Astrophysics, 670: 0 A54, 2023

    2023

  6. [14]

    J., Djorgovski, S

    Drake, A. J., Djorgovski, S. G., Mahabal, A., Prieto, J. L., Beshore, E., Graham, M. J., Catalan, M., Larson, S., Christensen, E., and Donalek, C. The catalina real-time transient survey. In Proceedings of the International Astronomical Union, Symposium S285: New Horizons in T...

    2012 doi

  7. [15]

    Preliminary report on mantis shrimp: a multi‑survey computer vision photometric redshift model

    Engel, A., Narayan, G., and Byler, N. Preliminary report on mantis shrimp: a multi‑survey computer vision photometric redshift model. arXiv preprint arXiv:2402.03535, 2024. URL https://arxiv.org/abs/2402.03535. Submitted to AI4Differential Equations in Science Workshop, ICLR 2024

    2024 arXiv

  8. [16]

    The automatic learning for the rapid classification of events (alerce) alert broker

    F \"o rster, F., Cabrera-Vives, G., Castillo-Navarrete, E., Est \'e vez, P., S \'a nchez-S \'a ez, P., Arredondo, J., Bauer, F., Carrasco-Davis, R., Catelan, M., Elorrieta, F., et al. The automatic learning for the rapid classification of events (alerce) alert broker. The Astr...

    2021

  9. [17]

    Gaia Collaboration , Vallenari, A., Brown, A. G. A., Prusti, T., de Bruijne, J. H. J., and et al. Gaia data release 3: Summary of the content and survey properties. Astronomy & Astrophysics, 674: 0 A1, 2023. doi:10.1051/0004-6361/202243940

    2023 doi

  10. [18]

    Garc \' a-Jara, G., Protopapas, P., and Est \'e vez, P. A. Improving astronomical time-series classification via data augmentation with generative adversarial networks. The Astrophysical Journal, 935 0 (1): 0 23, 2022

    2022

  11. [19]

    G \'o mez, C., Neira, M., Hern \'a ndez Hoyos, M., Arbel \'a ez, P., and Forero-Romero, J. E. Classifying image sequences of astronomical transients with deep neural networks. Monthly Notices of the Royal Astronomical Society, 499 0 (3): 0 3130--3138, 2020

    2020

  12. [20]

    and Muthukrishna, D

    Gupta, R. and Muthukrishna, D. Transfer learning for transient classification: From simulations to real data and ztf to lsst. arXiv preprint arXiv:2502.18558, 2025. Submitted to MNRAS

    2025 arXiv

  13. [21]

    L., Denneau, L., Flewelling, H., Stalder, B., Rest, A., Smith, K

    Heinze, A., Tonry, J. L., Denneau, L., Flewelling, H., Stalder, B., Rest, A., Smith, K. W., Smartt, S. J., and Weiland, H. A first catalog of variable stars measured by the asteroid terrestrial-impact last alert system (atlas). The Astronomical Journal, 156 0 (5): 0 241, 2018

    2018

  14. [22]

    and Bloom, J

    Jamal, S. and Bloom, J. S. On neural architectures for astronomical time-series classification with application to variable stars. The Astrophysical Journal Supplement Series, 250 0 (2): 0 30, 2020

    2020

  15. [23]

    The pan-starrs wide-field optical/nir imaging survey

    Kaiser, N., Burgett, W., Chambers, K., Denneau, L., Heasley, J., Jedicke, R., Magnier, E., Morgan, J., Onaka, P., and Tonry, J. The pan-starrs wide-field optical/nir imaging survey. In Ground-based and airborne telescopes III, volume 7733, pp.\ 159--172. SPIE, 2010

    2010

  16. [24]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., and Guo, B. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 10012--10022, 2021

    2021

  17. [25]

    Swin transformer v2: Scaling up capacity and resolution

    Liu, Z., Hu, H., Lin, Y., Yao, Z., Xie, Z., Wei, Y., Ning, J., Cao, Y., Zhang, Z., Dong, L., et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 12009--12019, 2022

    2022

  18. [26]

    D., Peiris, H

    Lochner, M., McEwen, J. D., Peiris, H. V., Lahav, O., and Winter, M. K. Photometric supernova classification with machine learning. The Astrophysical Journal Supplement Series, 225 0 (2): 0 31, 2016

    2016

  19. [27]

    G., Drake, A

    Mahabal, A., Sheth, K., Gieseke, F., Pai, A., Djorgovski, S. G., Drake, A. J., and Graham, M. J. Deep-learnt classification of light curves. In 2017 IEEE symposium series on computational intelligence (SSCI), pp.\ 1--8. IEEE, 2017

    2017

  20. [28]

    J., Laher, R

    Masci, F. J., Laher, R. R., Rusholme, B., Shupe, D., Paladini, R., Groom, S., Wold, A., Miller, A. A., and Drake, A. A new forced photometry service for the zwicky transient facility. arXiv preprint arXiv:2305.16279, 2023

    2023 arXiv

  21. [29]

    Application of convolutional neural networks to time domain astrophysics

    Monsalves, N., Jaque Arancibia, M., Bayo, A., S \'a nchez-S \'a ez, P., Angeloni, R., Damke, G., and Segura Van de Perre, J. Application of convolutional neural networks to time domain astrophysics. 2d image analysis of ogle light curves. Astronomy & Astrophysics, 691: 0 A106, 2024

    2024

  22. [30]

    Positional encodings for light curve transformers: Playing with positions and attention

    Moreno-Cartagena, D., Cabrera-Vives, G., Protopapas, P., Donoso-Oliva, C., P \'e rez-Carrasco, M., and C \'a diz-Leyton, M. Positional encodings for light curve transformers: Playing with positions and attention. In Machine Learning for Astrophysics Workshop, 40th Internationa...

    2023

  23. [31]

    Leveraging pre-trained visual transformers for multi-band photometric light curve classification

    Moreno-Cartagena, D., Protopapas, P., Cabrera-Vives, G., C \'a diz-Leyton, M., Becker, I., and Donoso-Oliva, C. Leveraging pre-trained visual transformers for multi-band photometric light curve classification. arXiv preprint arXiv:2502.20479, 2025. Submitted to Astronomy & Ast...

    2025

  24. [32]

    Murray, C., Fabbro, S., Herzberg, N., and Yi, K. M. Unsupervised learning of deep features through best-fits for observational cosmology

  25. [33]

    S., Biswas, R., and Hlo z ek, R

    Muthukrishna, D., Narayan, G., Mandel, K. S., Biswas, R., and Hlo z ek, R. Rapid: early classification of explosive transients using deep learning. Publications of the Astronomical Society of the Pacific, 131 0 (1005): 0 118002, 2019

    2019

  26. [34]

    Astroconformer: Inferring surface gravity of stars from stellar light curves with transformer

    Pan, J., Ting, Y.-S., and Yu, J. Astroconformer: Inferring surface gravity of stars from stellar light curves with transformer. In Machine Learning for Astrophysics Workshop, 39th International Conference on Machine Learning ( ICML ), PMLR 162 , Baltimore, Maryland, USA, 2022

    2022

  27. [35]

    A., and F \"o rster, F

    Pimentel, \'O ., Est \'e vez, P. A., and F \"o rster, F. Deep attention-based supernovae classification of multiband light curves. The Astronomical Journal, 165 0 (1): 0 18, 2022

    2022

  28. [36]

    and Bloom, J

    Rizhko, M. and Bloom, J. S. Self-supervised multimodal model for astronomy. In Neurips 2024 Workshop Foundation Models for Science: Progress, Opportunities, and Challenges, 2024 a

    2024

  29. [37]

    and Bloom, J

    Rizhko, M. and Bloom, J. S. Astrom 3 : A self-supervised multimodal model for astronomy. arXiv preprint arXiv:2411.08842, 2024 b . Submitted to MNRAS

    2024 arXiv

  30. [38]

    Alert classification for the alerce broker system: The light curve classifier

    S \'a nchez-S \'a ez, P., Reyes, I., Valenzuela, C., F \"o rster, F., Eyheramendy, S., Elorrieta, F., Bauer, F., Cabrera-Vives, G., Est \'e vez, P., Catelan, M., et al. Alert classification for the alerce broker system: The light curve classifier. The Astronomical Journal, 161...

    2021

  31. [39]

    J., Prieto, J., Grupe, D., Kochanek, C., Stanek, K., De Rosa, G., Mathur, S., Zu, Y., Peterson, B., Pogge, R., et al

    Shappee, B. J., Prieto, J., Grupe, D., Kochanek, C., Stanek, K., De Rosa, G., Mathur, S., Zu, Y., Peterson, B., Pogge, R., et al. The man behind the curtain: X-rays drive the uv through nir variability in the 2013 active galactic nucleus outburst in ngc 2617. The Astrophysical...

    2013

  32. [40]

    Release of the atlas forced photometry server for public use

    Shingles, L., Smith, K., Young, D., Smartt, S., Tonry, J., Denneau, L., Heinze, A., Weiland, H., Flewelling, H., Stalder, B., et al. Release of the atlas forced photometry server for public use. Transient Name Server AstroNote 2021-7, 7: 0 1--7, 2021

    2021

  33. [41]

    J., Hern \'a ndez-Garc \' a, L., Arancibia, A

    Silva-Farf \'a n, J., F \"o rster, F., Moriya, T. J., Hern \'a ndez-Garc \' a, L., Arancibia, A. M., S \'a nchez-S \'a ez, P., Anderson, J. P., Tonry, J. L., and Clocchiatti, A. Physical properties of type ii supernovae inferred from ztf and atlas photometric data. The Astroph...

    2024

  34. [42]

    Design and operation of the atlas transient science server

    Smith, K., Smartt, S., Young, D., Tonry, J., Denneau, L., Flewelling, H., Heinze, A., Weiland, H., Stalder, B., Rest, A., et al. Design and operation of the atlas transient science server. Publications of the Astronomical Society of the Pacific, 132 0 (1014): 0 085002, 2020

    2020

  35. [43]

    L., Richards , J

    Starr , D. L., Richards , J. W., Brink , H., Miller , A. A., Bloom , J. S., Butler , N. R., James , J. B., and Long , J. P. ALLStars: Overcoming Multi-Survey Selection Bias using Crowd-Sourced Active Learning . In Ballester , P., Egret , D., and Lorente , N. P. F. (eds.), Astr...

    2012

  36. [44]

    Atlas: a high-cadence all-sky survey system

    Tonry, J., Denneau, L., Heinze, A., Stalder, B., Smith, K., Smartt, S., Stubbs, C., Weiland, H., and Rest, A. Atlas: a high-cadence all-sky survey system. Publications of the Astronomical Society of the Pacific, 130 0 (988): 0 064505, 2018

    2018

  37. [45]

    Tonry, J. L. An early warning system for asteroid impact. Publications of the Astronomical Society of the Pacific, 123 0 (899): 0 58, 2010

    2010

  38. [46]

    K., and Szymański, G

    Udalski, A., Szymański, M. K., and Szymański, G. Ogle-iv: Fourth phase of the optical gravitational lensing experiment. Acta Astronomica, 65 0 (1): 0 1--38, 2015. URL https://acta.astrouw.edu.pl/Vol65/n1/pdf/pap_65_1_1.pdf

    2015

  39. [47]

    Supernova photometric classification pipelines trained on spectroscopically classified supernovae from the pan-starrs1 medium-deep survey

    Villar, V., Berger, E., Miller, G., Chornock, R., Rest, A., Jones, D., Drout, M., Foley, R., Kirshner, R., Lunnan, R., et al. Supernova photometric classification pipelines trained on spectroscopically classified supernovae from the pan-starrs1 medium-deep survey. The Astrophy...

    2019

  40. [48]

    A., Hosseinzadeh, G., Berger, E., Ntampaka, M., Jones, D

    Villar, V. A., Hosseinzadeh, G., Berger, E., Ntampaka, M., Jones, D. O., Challis, P., Chornock, R., Drout, M. R., Foley, R. J., Kirshner, R. P., et al. Superraenn: a semisupervised supernova photometric classification pipeline trained on pan-starrs1 medium-deep survey supernov...

    2020

  41. [49]

    Max-deeplab: End-to-end panoptic segmentation with mask transformers

    Wang, H., Zhu, Y., Adam, H., Yuille, A., and Chen, L.-C. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 5463--5474, 2021

    2021

  42. [50]

    T., Mishra-Sharma, S., and Villar, V

    Zhang, G., Helfer, T., Gagliano, A. T., Mishra-Sharma, S., and Villar, V. A. Maven: a multimodal foundation model for supernova science. Machine Learning: Science and Technology, 5 0 (4): 0 045069, 2024

    2024

  43. [51]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.