Pith. sign in

REVIEW 4 major objections 6 minor 6 cited by

AstroM$^3$: A self-supervised multimodal model for astronomy

T0 review · 4 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Extending CLIP to three modalities improves variable-star classification and rediscovers subclasses.

desk verdict Useful dataset and a first trimodal CLIP for astronomy, but the limited-label claim is undercut by pretraining on test objects. read the letter →

arxiv 2411.08842 v1 pith:LJNNKLR6 submitted 2024-11-13 astro-ph.IM cs.AI

Mariia Rizhko , Joshua S. Bloom This is my paper · ORCID
classification astro-ph.IMcs.AI
keywords self-supervisedlearningmultimodalCLIPvariablestarstime-seriesphotometryspectraclassificationcontrastiveastronomicalsurveys
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 claims that a contrastive self-supervised model, extended from CLIP to three modalities, can learn a shared representation of time-series photometry, spectra, and catalog metadata for variable stars. Pretraining on unlabeled data lifts photometry-only classification accuracy from 84.6% to 91.5% and improves spectra classification by up to 12.6 percentage points when only 10% of labels are available. The same embeddings let the model flag catalog misclassifications, find similar objects across modalities, and separate Mira variables into M- and C-type subclasses and spotted variables into giants and dwarfs without any subtype labels. If the approach generalizes, it offers a route to exploit the vast unlabeled multi-modal data from current surveys despite scarce labels.

What carries the argument

The machinery is a trimodal contrastive network: an Informer encoder for light curves, a GalSpecNet-derived convolutional network for spectra, and a multilayer perceptron for 34 metadata features, each followed by a projection head that maps embeddings to 512 dimensions. Pairwise cosine-similarity matrices are computed for the three modality pairs, and a symmetric cross-entropy loss—treating the diagonal as positive matches—is summed over the pairs. This objective aligns the three views of each object in a shared space, and at test time embeddings can be averaged across any available modalities.

What would settle it

Retrain the metadata encoder alone (or with random embeddings) on the same 34 features and run UMAP on the result; if the Mira M/C and ROT giant/dwarf clusters appear just as cleanly without any contrastive pretraining, the claimed unsupervised rediscovery is not caused by the multimodal alignment. An even sharper test would remove color and magnitude features from the metadata and check whether the clusters vanish.

Watch

Extended reading notes

Core claim

The central discovery is that extending CLIP to a trimodal setting—photometric time series, spectra, and astrophysical metadata—produces a shared embedding space in which the same astronomical object's three views are pulled together while different objects are pushed apart. Training on 21,440 real variable-star sources without using class labels, the model improves downstream supervised classification, most strongly for photometry (from 84.64% to 91.47%) and for low-label regimes (spectra gain 12.56 points at a 10% split). Without any supervision for subtypes, UMAP projections of the learned embeddings split Miras into M- and C-type clusters and ROT variables into giant and dwarf clusters, and the model places previously unseen rare classes next to their physical relatives. The paper positions this as the first n>2 modality CLIP-style model in astronomy and a step toward a foundational multimodal model for time-domain astronomy.

Load-bearing premise

The load-bearing premise is that the subtype clusters (Mira M/C, ROT giants/dwarfs) arise from the learned multimodal alignment rather than being carried directly by the catalog colors, absolute magnitudes, and parallax already fed into the metadata encoder.

Editorial extensions

If this is right

  • Photometry-only classification of variable stars should continue to benefit from CLIP-style pretraining as unlabeled survey data grow.
  • In label-scarce regimes, pretraining on the full unlabeled set can replace a large fraction of manual labels: at a 10% split, spectra accuracy rises from 46.7% to 59.2%.
  • Because embeddings are averaged rather than concatenated, the model can make predictions from any subset of modalities at test time, including cases where only photometry is available.
  • Embedding-space tools—similarity search, cross-modal contrast, and outlier detection—become available without task-specific training.
  • The same framework extends naturally to more than three modalities, such as additional bandpasses or ancillary catalogs.

Reading between the lines

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

  • The unsupervised 'rediscovery' of Mira M/C and ROT giant/dwarf subtypes may be partly driven by metadata features such as WISE/2MASS colors, absolute magnitudes, and parallax, which are known to separate these physical classes; the paper does not ablate these features or compare against random embeddings.
  • If that concern is real, a cleaner test of multimodal alignment would hide all color and magnitude features during pretraining and see whether subtype clusters still emerge from photometry and spectra alone.
  • The claimed label-efficiency gains suggest a practical recipe for upcoming surveys: pretrain once on unlabeled multi-modal data, then fine-tune on a small labeled subset per new survey; transfer across surveys remains untested.
  • Cross-modal similarity search could be used to cheaply pre-select spectroscopic follow-up targets for objects that currently only have photometry, since the shared space lets photometric embeddings retrieve candidate spectra.
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

4 major / 6 minor

Summary. AstroM3 extends CLIP to a trimodal setting, jointly pretraining encoders for ASAS-SN time-series photometry, LAMOST spectra, and astrophysical metadata on a curated set of 21,440 variable stars. The paper reports that CLIP pretraining improves photometry classification accuracy from 84.6% to 91.5% with reduced variance, yields gains under limited labels (up to +12.6 percentage points for spectra at a 10% split), and produces embeddings useful for misclassification detection, similarity search, and outlier identification. A UMAP analysis of the embeddings is claimed to 'rediscover' Mira M/C subtypes and ROT giant/dwarf classes without explicit labels.

Significance. If the results survive a properly held-out evaluation, the work would be a useful contribution: it assembles a sizable multimodal benchmark, provides a clean architecture-matched with/without-CLIP comparison, and demonstrates several practical downstream uses of the learned embeddings. The paper is appropriately cautious about not claiming state-of-the-art performance. However, the central empirical claims are weakened by a potential transductive leak in the limited-label experiments, an ambiguity in the full-label cross-validation setup, and a rediscovery claim lacking control analyses. These issues are load-bearing and require re-running or extending the experiments.

major comments (4)
  1. [5.2, Table 4] The CLIP model used in the limited-label experiments was pre-trained on the full dataset, which includes the validation and test objects listed in Table 3. Because the contrastive objective aligns each object's photometry with its own spectra and metadata, the pre-trained encoder has already seen every test object and can place a test light curve near class-discriminative catalog information (colors, absolute magnitudes, period). The no-CLIP baseline is trained from random initialization on the downsampled training subsets only, so it never sees test objects. The reported limited-label gains therefore conflate representation learning with transductive test-set familiarity. Please re-run the limited-label comparisons with pre-training performed only on the training split (or on unlabeled data excluding the test objects), and report the results.
  2. [5.1, Table 2] The same transductive concern applies to the full-label comparison if CLIP pre-training uses the full dataset before the 5 random train/val/test splits are drawn. Please specify exactly what data the CLIP model was pre-trained on for each cross-validation fold, and, if the full dataset was used, provide a version where pre-training is confined to the training fold. Without this, the headline photometry improvement (84.6% to 91.5%) may also be inflated by test-object exposure during pre-training.
  3. [5.3, Figures 5 and 6] The 'unsupervised rediscovery' claim is not yet supported. The metadata inputs include WISE/2MASS colors, absolute magnitudes, and parallax, and the photometry encoder receives period, amplitude, and the Lafler-Kinman statistic; C versus M Miras and ROT giants versus dwarfs are known to separate on exactly these features. The paper does not compare the UMAP structure of learned embeddings against UMAP on raw metadata, against a metadata-only embedding, or against a randomly initialized (non-CLIP) encoder, so the apparent clusters may simply reflect input features rather than learned multimodal alignment. Please add such control analyses, or temper the claim accordingly.
  4. [Tables 2 and 4] The paper labels some differences 'statistically significant' without reporting the test used. With only 5 seeds, the photometry comparison (84.642 ± 6.317 vs 91.468 ± 0.446) has a large variance in the no-CLIP arm, and the limited-label comparisons should be assessed with paired tests across the seeds. Please report the type of test, p-values, and effect sizes (or confidence intervals) for each claimed improvement.
minor comments (6)
  1. [5.2, Table 3] The columns labeled 50%, 25%, and 10% do not seem to correspond to simple subsamples of the full counts (e.g., EW train full = 4890, 50% = 1209). Please clarify the downsampling procedure and define what the percentage labels refer to.
  2. [Title/Abstract] The paper uses 'AstroM3' and 'AstroM$^3$' inconsistently between the title/abstract and the main text; please standardize.
  3. [References] References Wu et al. 2024a and Wu et al. 2024b appear to be the same GalSpecNet paper (MNRAS 527, 1163); please merge them.
  4. [5.4, Table 5] The reported average accuracy for the 'All' condition in Table 5 is 93.07%, which differs from the 94.153% reported for the same condition in Table 2. Please explain the discrepancy or clarify that different test sets or model variants are used.
  5. [5.4] The text notes that combining metadata with photometry degrades EB and HADS relative to photometry alone (Table 5), but the abstract and conclusion should be careful not to overstate the universal benefit of adding modalities.
  6. [Data Availability] The paper states that all code, model weights, and data will be made available 'upon acceptance'; for a methods paper, a public repository before acceptance would strengthen reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

Limited-label gains are transductive because CLIP pretrains on the full dataset including test objects, and the 'rediscovered' Mira/ROT subtypes are directly encoded in input colors, magnitudes, and spectra.

  1. fitted input called prediction [Section 5.2 (Limited Labeled Data); Table 3; Eq. 10]
    "It is important to note that the CLIP model used for these experiments was the same as before: pre-trained on the full dataset without using any labels."

    The CLIP contrastive loss (Eqs. 7–11 with Y_ij = 1 if i = j) aligns photometry, spectra, and metadata embeddings for every object in the full dataset, which includes the validation and test objects used in Table 4. At fine-tuning, the classifier sees test-object embeddings that already encode each test object's own cross-modal identity; the no-CLIP baseline, trained from random initialization on the reduced training subsets only, has never seen those test objects. The reported limited-label gains (e.g., +12.56% for spectra at 10%) therefore conflate representation quality with transductive familiarity with the evaluation set.

  2. renaming known result [Section 5.3 (UMAP Analysis); Section 4.3; Table A1]
    "Notably, the model discovered this distinction in an unsupervised learning process, without explicit labels for these subtypes. ... The metadata modality consists of astrophysical parameters and observational data not included in the other two modalities. This includes features like absolute magnitudes in various bands, astrometric information."

    The embeddings fed to UMAP are averaged over photometry, spectra, and metadata. The metadata channel includes WISE/2MASS colors (w1_w2, j_k), absolute magnitudes computed from parallax, and the photometry channel includes period and amplitude. C vs M Miras and ROT giants vs dwarfs are known to separate on exactly these colors and magnitudes (and M vs C Miras are defined spectroscopically, so the spectral channel contains the defining bands). The 'rediscovery' is therefore a direct read-out of catalog features already present in the input, not an emergent grouping from multimodal alignment. No ablation or random-embedding control is provided, so the unsupervised-discovery claim reduces to clustering on the known discriminators.

full rationale

The paper's main architectural contribution—extending CLIP to three modalities and evaluating fine-tuned classification—is not itself circular: the encoders, contrastive loss, and downstream fine-tuning are self-contained, and there is no load-bearing self-citation chain or uniqueness theorem imported from the authors. However, two central claims carry significant circularity burdens. First, the limited-label experiments (Section 5.2, Table 4) use a CLIP model pretrained on the full dataset, which explicitly includes the validation and test objects; the no-CLIP baselines never see those objects. Since the contrastive objective aligns each object's modalities with itself, the test-object-specific information is baked into the embeddings before fine-tuning, making the reported accuracy gains at least partly transductive rather than purely transferable. The paper does not offer a baseline that also uses unlabeled test data, so the comparison does not isolate representation quality. Second, the 'rediscovery' of Mira subtypes and ROT giants/dwarfs (Section 5.3) is presented as unsupervised discovery, but the input metadata contains the exact features (absolute magnitudes, WISE/2MASS colors, parallax) and the spectra contain the defining band features that separate these classes. UMAP on these embeddings therefore propagates known catalog information; without ablating those features or comparing against random embeddings, the rediscovery claim reduces to clustering on the known discriminators. These issues affect the headline claims about unlabeled-data leverage and unsupervised discovery, but the rest of the paper (dataset construction, similarity search, modality-ablation analysis) retains independent empirical content, so a score of 6 reflects partial rather than complete circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claims rest on standard CLIP assumptions and on the reliability of catalog cross-matching and labels. The main non-standard element is the hand-injected photometric auxiliary features and catalog colors and magnitudes, which can explain part of the unsupervised rediscovery without invoking emergent multimodal understanding.

free parameters (3)
  • Optuna-tuned hyperparameters (learning rate, dropout, weight decay, scheduler factor) = not reported; search ranges in Appendix A
    Results depend on hyperparameters selected by validation loss; best values are not listed in the preprint.
  • CLIP contrastive temperature or logit scale = not stated
    Standard CLIP uses a learned temperature; the paper omits it, leaving the scale of cosine similarities in the cross-entropy loss unspecified.
  • Architecture constants (L=200, embedding dim 512, 8 Informer layers) = L=200, dim=512, layers=8
    Hand-chosen model sizes that affect capacity and are not derived from data.
assumptions (5)
  • domain assumption ASAS-SN catalog labels (Jayasinghe et al. 2019) are sufficiently accurate to serve as ground truth for fine-tuning and to name rediscovered clusters.
    Section 3 uses these labels to build the dataset; Section 5.1 evaluates against them.
  • standard math The pairwise CLIP losses with symmetric cross-entropy yield a shared embedding space where cosine distance is semantically meaningful.
    Section 4.4 extends the standard CLIP objective; no proof is given that pairwise alignment is sufficient, but this is the standard assumption.
  • domain assumption Cross-matching ASAS-SN, Gaia EDR3, and LAMOST by Gaia source_id is correct.
    Section 3 merges catalogs; wrong associations would decouple modalities and corrupt training.
  • ad hoc to paper Class-discriminative inputs (period, colors, absolute magnitudes) do not dominate the unsupervised structure in a way that invalidates rediscovery claims.
    Sections 4.1 and 4.3 inject period, amplitude, WISE/2MASS colors, and absolute magnitudes; these features are known discriminators for the discovered subtypes.
  • domain assumption Filtering to sources with all three modalities and top-10 classes yields a sufficiently representative training set.
    Section 3 criteria (1)-(3); selection bias is acknowledged in Future Work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AstroM$^3$: A self-supervised multimodal model for astronomy." pith.science (2026). https://pith.science/paper/LJNNKLR6

@misc{pith2026241108842,
  author       = {Pith},
  title        = {Pith review of: AstroM$^3$: A self-supervised multimodal model for astronomy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LJNNKLR6}},
  note         = {Machine review of arXiv:2411.08842}
}
abstract

While machine-learned models are now routinely employed to facilitate astronomical inquiry, model inputs tend to be limited to a primary data source (namely images or time series) and, in the more advanced approaches, some metadata. Yet with the growing use of wide-field, multiplexed observational resources, individual sources of interest often have a broad range of observational modes available. Here we construct an astronomical multimodal dataset and propose AstroM$^3$, a self-supervised pre-training approach that enables a model to learn from multiple modalities simultaneously. Specifically, we extend the CLIP (Contrastive Language-Image Pretraining) model to a trimodal setting, allowing the integration of time-series photometry data, spectra, and astrophysical metadata. In a fine-tuning supervised setting, our results demonstrate that CLIP pre-training improves classification performance for time-series photometry, where accuracy increases from 84.6% to 91.5%. Furthermore, CLIP boosts classification accuracy by up to 12.6% when the availability of labeled data is limited, showing the effectiveness of leveraging larger corpora of unlabeled data. In addition to fine-tuned classification, we can use the trained model in other downstream tasks that are not explicitly contemplated during the construction of the self-supervised model. In particular we show the efficacy of using the learned embeddings for misclassifications identification, similarity search, and anomaly detection. One surprising highlight is the "rediscovery" of Mira subtypes and two Rotational variable subclasses using manifold learning and dimension reduction algorithm. To our knowledge this is the first construction of an $n>2$ mode model in astronomy. Extensions to $n>3$ modes is naturally anticipated with this approach.

Figures

Figures reproduced from arXiv: 2411.08842 by the authors.

Figure 1
Figure 1. Overview of the multimodal CLIP framework adapted for astronomy, incorporating three data modalities: photometric time-series, spectra, and metadata. Each modality is processed by a dedicated encoder to create embeddings, which are then mapped into a shared embedding space through projection heads. Pairwise similarity matrices align the embeddings across modalities, and a symmetric cross-entropy loss, computed over … view at source ↗
Figure 2
Figure 2. UMAP visualizations of multimodal embeddings: (a) training set and (b) test set, showing class separability and alignment between sets. Each source in the training and test set are coloured by the class determined in (Jayasinghe et al. 2019) but these class labels are not used in the construction of the embeddings [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Examples of catalog misclassifications with photometry and spectrum for each object. Top to bottom: (1) Likely EW missclassified as HADS; (2) V* AC CMi, a known semi-detached binary misclassified as RR Lyrae; (3) Possible SR or Mira variable with period alignment issues; (4) Known Mira variable (V0439 Cas) misclassified as SR; (5) Likely EW binary (Mowlavi et al. 2023) misclassified as RRC. MNRAS 000, 1–11 (2024) [… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Examples of in-class outliers flagged by the model due to distinctive features, despite correct labels. (a) EA-type star, V1174 Ori, an X-ray bright pre-main sequence system (Stassun et al. 2022). (b) EB-type star with unusual out-of-eclipse modulations, possibly due t…
Figure 7
Figure 7. Figure 7: Projections of new, previously unused classes in the embedding space, aligning with related clusters and demonstrating the model’s ability to position unseen classes accurately New Classes. During dataset creation, we filtered out classes with insufficient sample sizes…
Figure 6
Figure 6. Figure 6: Spectral examples of Mira variables, showing two distinct clusters corresponding to M-type and C-type Miras, discovered through unsupervised learning [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: The query spectrum (top) and its two closest matches (middle and bottom) based on spectral cosine similarity. modality is most significant for classes like DSCT, EA and EB, while metadata is more important for EW. Other classes benefit from more than one modality: ROT …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 6 Pith papers

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

  1. VERITAS: Towards a General-Purpose Replication Tool for Scientific Research

    cs.AI 2026-07 conditional novelty 6.0 of 10

    VERITAS, a CLI-agent replication framework, leads every reported metric on CORE-Bench Hard and ReplicationBench against matched Claude Code baselines across 65 papers in four domains.

  2. Generalization from Low- to Moderate-Resolution Spectra with Neural Networks for Stellar Parameter Estimation: A Case Study with DESI

    astro-ph.SR 2026-02 conditional novelty 6.0 of 10

    Pre-trained MLPs on LAMOST low-resolution spectra generalize to DESI medium-resolution spectra for [Fe/H] and [α/Fe], outperforming the DESI SP pipeline in zero-shot and improving with modest fine-tuning.

  3. Representation Learning for Time-Domain High-Energy Astrophysics: Discovery of Extragalactic Fast X-ray Transient XRT 200515

    astro-ph.HE 2024-12 conditional novelty 6.0 of 10

    A representation-learning search of the Chandra archive found XRT 200515, a fast X-ray transient with a hard <10 s spike and a soft ~800 s tail, plus 3559 total flare and dip candidates.

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

    astro-ph.IM 2025-07 conditional novelty 5.0 of 10

    A shared-weights two-branch Swin Transformer that processes ZTF and ATLAS light curves jointly reaches 69.9% macro F1, outperforming single-survey models and simple fusion strategies on 21 classes.

  5. Causal Foundation Models: Disentangling Physics from Instrument Properties

    cs.LG 2025-07 conditional novelty 5.0 of 10

    A dual-encoder contrastive model trained on star-instrument triplets learns separate stellar and instrumental latent spaces, improving few-shot prediction of stellar parameters in simulated TESS-like light curves.

  6. From stellar light to astrophysical insight: automating variable star research with machine learning

    astro-ph.IM 2025-07 unverdicted

    An invited review of machine learning for automated variable star research, covering data cleaning, variability classification, stellar parameter inference, and foundation models.

Reference graph

Works this paper leans on

49 extracted references · 17 canonical work pages · cited by 6 Pith papers

  1. [1]

    pp 2623--2631

    Akiba T., Sano S., Yanase T., Ohta T., Koyama M., 2019, in Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. pp 2623--2631

    work page 2019

  2. [2]

    Alabdulmohsin I., Wang X., Steiner A., Goyal P., D'Amour A., Zhai X., 2024, arXiv preprint arXiv:2403.04547

    arXiv 2024

  3. [3]

    Becker I., Pichara K., Catelan M., Protopapas P., Aguirre C., Nikzat F., 2020, @doi [MNRAS] 10.1093/mnras/staa350 , https://ui.adsabs.harvard.edu/abs/2020MNRAS.493.2981B 493, 2981

    doi:10.1093/mnras/staa350 2020

  4. [4]

    Boone K., 2021, @doi [AJ] 10.3847/1538-3881/ac2a2d , https://ui.adsabs.harvard.edu/abs/2021AJ....162..275B 162, 275

    doi:10.3847/1538-3881/ac2a2d 2021

  5. [5]

    pp 2818--2829

    Cherti M., et al., 2023, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp 2818--2829

    work page 2023

  6. [6]

    Cui X.-Q., et al., 2012, @doi [Research in Astronomy and Astrophysics] 10.1088/1674-4527/12/9/003 , https://ui.adsabs.harvard.edu/abs/2012RAA....12.1197C 12, 1197

    doi:10.1088/1674-4527/12/9/003 2012

  7. [7]

    M., Aerts C., Cuypers J., Vandenbussche B., Garrido R., Solano E., 2007, @doi [AAP] 10.1051/0004-6361:20077638 , https://ui.adsabs.harvard.edu/abs/2007A&A...475.1159D 475, 1159

    Debosscher J., Sarro L. M., Aerts C., Cuypers J., Vandenbussche B., Garrido R., Solano E., 2007, @doi [AAP] 10.1051/0004-6361:20077638 , https://ui.adsabs.harvard.edu/abs/2007A&A...475.1159D 475, 1159

    doi:10.1051/0004-6361:20077638 2007

  8. [8]

    Donoso-Oliva C., Becker I., Protopapas P., Cabrera-Vives G., Vishnu M., Vardhan H., 2023, @doi [AAP] 10.1051/0004-6361/202243928 , https://ui.adsabs.harvard.edu/abs/2023A&A...670A..54D 670, A54

    doi:10.1051/0004-6361/202243928 2023

Show all 49 references
  1. [9]

    Dubath P., et al., 2011, @doi [MNRAS] 10.1111/j.1365-2966.2011.18575.x , https://ui.adsabs.harvard.edu/abs/2011MNRAS.414.2602D 414, 2602

    2011

  2. [10]

    pp 226--231

    Ester M., Kriegel H.-P., Sander J., Xu X., et al., 1996, in kdd. pp 226--231

    1996

  3. [11]

    Gaia Collaboration Brown A. G. A., et al., 2021, @doi [AAP] 10.1051/0004-6361/202039657 , https://ui.adsabs.harvard.edu/abs/2021A&A...649A...1G 649, A1

    2021 doi

  4. [12]

    pp 976--980

    Guzhov A., Raue F., Hees J., Dengel A., 2022, in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp 976--980

    2022

  5. [13]

    A., Stein G., Harrington P., Luki \'c Z., Mustafa M., 2021, @doi [ApJ Letters] 10.3847/2041-8213/abf2c7 , https://ui.adsabs.harvard.edu/abs/2021ApJ...911L..33H 911, L33

    Hayat M. A., Stein G., Harrington P., Luki \'c Z., Mustafa M., 2021, @doi [ApJ Letters] 10.3847/2041-8213/abf2c7 , https://ui.adsabs.harvard.edu/abs/2021ApJ...911L..33H 911, L33

    2021 doi

  6. [14]

    S., 2020, @doi [ApJ Supp.] 10.3847/1538-4365/aba8ff , https://ui.adsabs.harvard.edu/abs/2020ApJS..250...30J 250, 30

    Jamal S., Bloom J. S., 2020, @doi [ApJ Supp.] 10.3847/1538-4365/aba8ff , https://ui.adsabs.harvard.edu/abs/2020ApJS..250...30J 250, 30

    2020 doi

  7. [15]

    Jayasinghe T., et al., 2019, @doi [MNRAS] 10.1093/mnras/stz844 , 486, 1907

    2019 doi

  8. [16]

    Kim D.-W., Yeo D., Bailer-Jones C. A. L., Lee G., 2021, @doi [AAP] 10.1051/0004-6361/202140369 , https://ui.adsabs.harvard.edu/abs/2021A&A...653A..22K 653, A22

    2021 doi

  9. [17]

    D., 1965, @doi [ApJS] 10.1086/190116 , https://ui.adsabs.harvard.edu/abs/1965ApJS...11..216L 11, 216

    Lafler J., Kinman T. D., 1965, @doi [ApJS] 10.1086/190116 , https://ui.adsabs.harvard.edu/abs/1965ApJS...11..216L 11, 216

    1965 doi

  10. [18]

    W., Bovy J., 2024, @doi [MNRAS] 10.1093/mnras/stad3015 , https://ui.adsabs.harvard.edu/abs/2024MNRAS.527.1494L 527, 1494

    Leung H. W., Bovy J., 2024, @doi [MNRAS] 10.1093/mnras/stad3015 , https://ui.adsabs.harvard.edu/abs/2024MNRAS.527.1494L 527, 1494

    2024 doi

  11. [19]

    Li Y., Liang F., Zhao L., Cui Y., Ouyang W., Shao J., Yu F., Yan J., 2021, arXiv preprint arXiv:2110.05208

    2021 arXiv

  12. [20]

    pp 12888--12900

    Li J., Li D., Xiong C., Hoi S., 2022, in International conference on machine learning. pp 12888--12900

    2022

  13. [21]

    P., El Karoui N., Rice J

    Long J. P., El Karoui N., Rice J. A., Richards J. W., Bloom J. S., 2012, @doi [PASP] 10.1086/664960 , https://ui.adsabs.harvard.edu/abs/2012PASP..124..280L 124, 280

    2012 doi

  14. [22]

    Luo H., Ji L., Zhong M., Chen Y., Lei W., Duan N., Li T., 2021, arXiv preprint arXiv:2104.08860

    2021 arXiv

  15. [23]

    pp 638--647

    Ma Y., Xu G., Sun X., Yan M., Zhang J., Ji R., 2022, in Proceedings of the 30th ACM International Conference on Multimedia. pp 638--647

    2022

  16. [24]

    McInnes L., Healy J., Melville J., 2018, arXiv preprint arXiv:1802.03426

    2018 arXiv

  17. [25]

    Morrissey P., et al., 2007, @doi [ApJS] 10.1086/520512 , https://ui.adsabs.harvard.edu/abs/2007ApJS..173..682M 173, 682

    2007 doi

  18. [26]

    Mowlavi N., et al., 2023, @doi [AA] 10.1051/0004-6361/202245330 , https://ui.adsabs.harvard.edu/abs/2023A&A...674A..16M 674, A16

    2023 doi

  19. [27]

    pp 529--544

    Mu N., Kirillov A., Wagner D., Xie S., 2022, in European conference on computer vision. pp 529--544

    2022

  20. [28]

    S., Biswas R., Hlo z ek R., 2019, @doi [PASP] 10.1088/1538-3873/ab1609 , https://ui.adsabs.harvard.edu/abs/2019PASP..131k8002M 131, 118002

    Muthukrishna D., Narayan G., Mandel K. S., Biswas R., Hlo z ek R., 2019, @doi [PASP] 10.1088/1538-3873/ab1609 , https://ui.adsabs.harvard.edu/abs/2019PASP..131k8002M 131, 118002

    2019 doi

  21. [29]

    S., P \'e rez F., van der Walt S., 2018, @doi [Nature Astronomy] 10.1038/s41550-017-0321-z , https://ui.adsabs.harvard.edu/abs/2018NatAs...2..151N 2, 151

    Naul B., Bloom J. S., P \'e rez F., van der Walt S., 2018, @doi [Nature Astronomy] 10.1038/s41550-017-0321-z , https://ui.adsabs.harvard.edu/abs/2018NatAs...2..151N 2, 151

    2018 doi

  22. [30]

    Palaversa L., et al., 2013, @doi [AJ] 10.1088/0004-6256/146/4/101 , https://ui.adsabs.harvard.edu/abs/2013AJ....146..101P 146, 101

    2013 doi

  23. [31]

    Parker L., et al., 2024, @doi [Monthly Notices of the Royal Astronomical Society] 10.1093/mnras/stae1450 , 531, 4990–5011

    2024 doi

  24. [32]

    pp 8748--8763

    Radford A., et al., 2021, in International conference on machine learning. pp 8748--8763

    2021

  25. [33]

    W., et al., 2011, @doi [ApJ] 10.1088/0004-637X/733/1/10 , https://ui.adsabs.harvard.edu/abs/2011ApJ...733...10R 733, 10

    Richards J. W., et al., 2011, @doi [ApJ] 10.1088/0004-637X/733/1/10 , https://ui.adsabs.harvard.edu/abs/2011ApJ...733...10R 733, 10

    2011 doi

  26. [34]

    W., Starr D

    Richards J. W., Starr D. L., Miller A. A., Bloom J. S., Butler N. R., Brink H., Crellin-Quick A., 2012, @doi [ApJ Supp.] 10.1088/0067-0049/203/2/32 , https://ui.adsabs.harvard.edu/abs/2012ApJS..203...32R 203, 32

    2012 doi

  27. [35]

    J., et al., 2014, @doi [ApJ] 10.1088/0004-637X/788/1/48 , https://ui.adsabs.harvard.edu/abs/2014ApJ...788...48S 788, 48

    Shappee B. J., et al., 2014, @doi [ApJ] 10.1088/0004-637X/788/1/48 , https://ui.adsabs.harvard.edu/abs/2014ApJ...788...48S 788, 48

    2014 doi

  28. [36]

    pp 15638--15650

    Singh A., Hu R., Goswami V., Couairon G., Galuba W., Rohrbach M., Kiela D., 2022, in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp 15638--15650

    2022

  29. [37]

    F., et al., 2006, @doi [AJ] 10.1086/498708 , https://ui.adsabs.harvard.edu/abs/2006AJ....131.1163S 131, 1163

    Skrutskie M. F., et al., 2006, @doi [AJ] 10.1086/498708 , https://ui.adsabs.harvard.edu/abs/2006AJ....131.1163S 131, 1163

    2006 doi

  30. [38]

    G., Torres G., Kounkel M., Feliz D

    Stassun K. G., Torres G., Kounkel M., Feliz D. L., Bouma L. G., Howell S. B., Gnilka C. L., Furlan E., 2022, @doi [ApJ] 10.3847/1538-4357/aca32e , https://ui.adsabs.harvard.edu/abs/2022ApJ...941..125S 941, 125

    2022 doi

  31. [39]

    Sun Q., Fang Y., Wu L., Wang X., Cao Y., 2023, arXiv preprint arXiv:2303.15389

    2023 arXiv

  32. [40]

    L., et al., 2010, @doi [AJ] 10.1088/0004-6256/140/6/1868 , https://ui.adsabs.harvard.edu/abs/2010AJ....140.1868W 140, 1868

    Wright E. L., et al., 2010, @doi [AJ] 10.1088/0004-6256/140/6/1868 , https://ui.adsabs.harvard.edu/abs/2010AJ....140.1868W 140, 1868

    2010 doi

  33. [41]

    Wu Y., Chen K., Zhang T., Hui Y., Berg-Kirkpatrick T., Dubnov S., 2023, in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp 1--5

    2023

  34. [42]

    Wu Y., Tao Y., Fan D., Cui C., Zhang Y., 2024a, @doi [MNRAS] 10.1093/mnras/stad2913 , https://ui.adsabs.harvard.edu/abs/2024MNRAS.527.1163W 527, 1163

    doi

  35. [43]

    Wu Y., Tao Y., Fan D., Cui C., Zhang Y., 2024b, Monthly Notices of the Royal Astronomical Society, 527, 1163

  36. [44]

    Xu H., Ghosh G., Huang P.-Y., Okhonko D., Aghajanyan A., Metze F., Zettlemoyer L., Feichtenhofer C., 2021, arXiv preprint arXiv:2109.14084

    2021 arXiv

  37. [45]

    Xu H., et al., 2023, arXiv preprint arXiv:2309.16671

    2023 arXiv

  38. [46]

    Yao L., et al., 2021, arXiv preprint arXiv:2111.07783

    2021 arXiv

  39. [47]

    pp 8552--8562

    Zhang R., et al., 2022, in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp 8552--8562

    2022

  40. [48]

    T., Mishra-Sharma S., Villar V

    Zhang G., Helfer T., Gagliano A. T., Mishra-Sharma S., Villar V. A., 2024, @doi [arXiv e-prints] 10.48550/arXiv.2408.16829 , https://ui.adsabs.harvard.edu/abs/2024arXiv240816829Z p. arXiv:2408.16829

    2024 doi arXiv

  41. [49]

    pp 11106--11115

    Zhou H., Zhang S., Peng J., Zhang S., Li J., Xiong H., Zhang W., 2021, in Proceedings of the AAAI conference on artificial intelligence. pp 11106--11115

    2021

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.