Pith. sign in

REVIEW 1 major objections 7 minor 85 references

Catching Disguised Transients with ASTRANet: Anomaly-Aware Spectroscopic Classification and Conformal Calibration

T0 review · 1 major / 7 minor · reviewed 2026-07-10 · glm-5.2

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

Pith's one-line read When AI confidently mislabels a new star, two kinds of doubt catch it

desk verdict The cross-family read the letter →

arxiv 2607.08044 v1 pith:FH6PAQ6Z submitted 2026-07-09 astro-ph.IM astro-ph.HE

Argyro Sasli , Maojie Xu , Alexandra Junell , Hailey Markoff , Avyukt Raghuvanshi , Felipe F. Nunes , Theophile Jegou Du Laz , Jesper Sollerman
show 16 more authors
Christoffer Fremling Drew Oldag Antoine Le Calloch Sushant Sharma Chaudhary Sneha Maharjan Maxine West Benny Border Nabeel Rehemtulla Richard Dekany Joahan Castaneda Jaimes Russ R. Laher Reed Riddle Mansi M. Kasliwal Matthew J. Graham Ashish A. Mahabal Michael W. Coughlin
This is my paper
classification astro-ph.IMastro-ph.HE
keywords spectroscopictexttttransientsanomalyconformalastranetastranet-cpclass
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

Deep classifiers for astronomical spectra assign every input to a known class with high confidence, even when the input belongs to no known class. This paper identifies five physically distinct ways these classifiers fail on rare transients, from blazars disguised as dwarf-star outbursts to genuinely novel explosions the classifier has never seen. The central claim is that no single anomaly score can detect all five failure modes: roughly 43 percent of rare anomalies produce weak signals in every individual detection method, and only a non-linear combination of sixteen complementary scores spanning four families recovers them. The paper further argues that classifier-internal uncertainty (the model's own self-doubt) and embedding-based anomaly detection (how far the spectrum sits from known classes in a learned representation space) are structurally complementary axes that catch largely disjoint failure populations, and that both must be deployed simultaneously rather than chosen between.

What carries the argument

A hierarchical spectral classifier operating on observer-frame spectra without requiring host-galaxy redshift or spectral phase; sixteen embedding-space anomaly scores spanning four physically motivated families (uncertainty, distance, density, hybrid); a gradient-boosted non-linear combiner producing a single cross-family anomaly score; a conformal prediction layer with Mondrian calibration providing per-spectrum p-values and prediction sets; and an AD-stratified Mondrian variant that partitions the calibration set by both predicted class and anomaly-score quintile to achieve uniform conditional coverage across anomaly strata.

What would settle it

A rare-anomaly class not represented in the evaluation set whose spectra are confidently misclassified by the classifier, produce weak signals in all sixteen anomaly scores including the non-linear combiner, and are structurally undetectable by the framework. Such a class would demonstrate that the five-regime taxonomy is incomplete and that cross-family combination is not sufficient for all failure modes.

Watch

Extended reading notes

Core claim

The paper's central object is the failure-mode taxonomy: five physically distinct regimes in which closed-set spectroscopic classifiers confidently misclassify rare transients. Two of these regimes, comprising about 43 percent of the rare-anomaly population, are recoverable only through non-linear cross-family combination of multiple anomaly scores, because no single score family fires above threshold on those spectra. This finding directly motivates the architecture: a gradient-boosted combiner integrating sixteen scores from four families (uncertainty, distance, density, hybrid) into a single anomaly probability, supplemented by a conformal calibration layer that provides distribution-free

Load-bearing premise

The 289-spectrum rare-anomaly evaluation set, with some classes having as few as one or ten members, is assumed to represent the full physical diversity of rare transients. If the five failure-mode taxonomy is incomplete or the sample is biased toward particular astrophysical configurations, the necessity of cross-family combination may not generalize to rare classes not represented in this set.

Editorial extensions

If this is right

  • Survey pipelines deploying only softmax confidence or only Mahalanobis distance silently lose roughly half of rare-anomaly populations by construction, since cross-family combination is necessary for the disguised-class regimes.
  • The AD-stratified conformal calibration approach is generalizable: any anomaly score used as a stratification variable for Mondrian conformal prediction can recover uniform conditional coverage in the high-anomaly operational regime where marginal methods under-cover.
  • The finding that classifier-internal UQ and embedding-based AD are structurally orthogonal suggests a three-way operational policy for survey triage: commit when both are low, abstain when UQ is high but AD is low, and alert when AD is high regardless of UQ.
  • At LSST scale, every one percent improvement in confidence-aware triage corresponds to roughly one hundred peculiar transients per year that would otherwise be lost to routine classification.
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

1 major / 7 minor

Summary. This manuscript presents ASTRANet, a confidence-aware framework for spectroscopic transient classification comprising three modules: a hierarchical classifier operating on observer-frame spectra without requiring redshift or phase, an anomaly detection layer (ASTRANet-Sentinel) combining 16 embedding-space scores via a gradient-boosted model, and a conformal uncertainty quantification layer (ASTRANet-CP) with a novel AD-stratified Mondrian conformal prediction scheme. The framework is validated on 289 rare/out-of-taxonomy spectra spanning 11 excluded classes. The central claims are: (1) closed-set classifier failure modes separate into five physically distinct regimes, (2) cross-family non-linear combination of anomaly scores is necessary to recover all five regimes, and (3) classifier-internal uncertainty and embedding-based anomaly detection are structurally complementary axes. The conformal coverage properties are empirically validated, and AD-MCP achieves uniform conditional coverage where vanilla Mondrian under-covers.

Significance. The work addresses a genuine and timely problem: spectroscopic triage at the scale of ZTF and LSST requires automated, calibrated confidence assessment, not just point-prediction classification. The redshift-free, phase-free observer-frame architecture is a practical advance for real-time deployment. The failure-mode taxonomy (Sec. II) is a useful conceptual contribution that connects physical regimes to specific anomaly-score signatures. The AD-MCP extension of atypicality-stratified conformal prediction to a learned cross-family ensemble is methodologically novel for astronomical spectroscopy. The leakage-free k-fold protocol (Sec. VD) and the empirical validation of conformal coverage within finite-sample error (Table IV, IX) are commendable. The structural complementarity argument between classifier-internal UQ and cross-family AD (Sec. VIF, Table VII) is well-supported by the population-scale analysis. The framework is being prepared for production integration with the ZTF BOOM broker, which lends practical credibility.

major comments (1)
  1. Sec. VC, Table I, and the claim that cross-family non-linear combination is 'necessary' (Contributions item 2, Sec. II.f): The key ablation comparing family-only mini-ensembles (86/289 at 1% FAR, no outlier exposure) against the full pipeline (238/289, with outlier exposure) confounds two variables: cross-family combination and outlier-exposure (OE) fine-tuning. OE reshapes the encoder embeddings that all 16 scores depend on (Sec. IVD), so the 137/238 spectra attributed to cross-family interaction may be partly or wholly recoverable by OE-enhanced single-family scores. The non-OE comparison (86 vs. 99, Table I) shows cross-family combination adds only 13 spectra over the best single family without OE — a modest gain that does not support the strong necessity claim. To establish that cross-family combination (rather than OE) drives the improvement, the authors should run OE-enhancedfamily
minor comments (7)
  1. Sec. IIIb: The rare-anomaly evaluation set includes classes with N<=10 (Ca-rich N=10, BL Lac N=1, Classical Nova N=8). Per-class recovery statistics for these classes (Table III) have large uncertainties. The authors acknowledge this in Sec. VIII (Limitations) but the headline recovery rates (e.g., '40.0%' for Ca-rich) are presented without error bars or confidence intervals. Adding bootstrap intervals for the smallest classes would set appropriate expectations.
  2. Table X caption: The regimes are labeled A/B/C in the table but I/V in the text (Sec. II, Sec. VII). This mismatch requires cross-referencing and could be unified.
  3. Sec. VIA: The statement 'the LightGBM-OOF score satisfies the exchangeability assumption conformal prediction requires' is an empirical observation, not a proof. The p-value ECDF being consistent with uniform (Fig. 4) is reassuring but does not formally guarantee exchangeability under arbitrary distribution shift. The authors should soften this language.
  4. Appendix C, feature importances: The top features (cosine, energy, mahal_global, mahal_within, mahal_class, gmm) span three of four families, which supports the cross-family design. However, the two lowest-ranked features (pca_recon, mrs) are from the density and hybrid families. If these contribute little, a 14-score combiner might perform comparably. An ablation over score count would strengthen the design justification.
  5. Sec. IVA: The claim that ASTRANet is 'the first fine-grained, multi-class classifier' to operate without redshift on observer-frame spectra should note SNIascore (Ref. 14) as a binary precedent, which it does, but the novelty claim would benefit from explicit comparison to any other redshift-free multi-class attempts if they exist.
  6. Table XIV: The 'Category' column uses 'SN CC (H)' and 'SN CC (SE)' but the coarse head (Sec. IVA) uses 'SN-CC' as a single superclass. The mapping between the three-way coarse partition and the finer category labels in Table XIV should be clarified.
  7. Sec. VIH: The AD-MCP fallback for cells with n_min < 10 uses the marginal class threshold. With 1,669 calibration samples and Q=5, this activates for sparse cells. The authors note this affects only rare-class extreme-quintile cells, but the fraction of test samples falling into fallback cells is not reported. This number would help assess the practical impact.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; derivation is self-contained against external benchmarks and theory.

full rationale

The paper's derivation chain contains no circular steps. The conformal prediction guarantees (Eqs. 1-2, 6-7) rest on established distribution-free theory from Vovk et al. [67], Angelopoulos & Bates [68], and Romano et al. [69] — all external citations with independently verified theorems. The AD-MCP extension (Sec. VIH) applies this theory to a new stratification scheme; its conditional coverage guarantee (Eq. 7) follows from within-cell exchangeability, a standard conformal assumption, not from any result defined in this paper. The cross-family combination claim (Sec. VC) is empirically validated on held-out data under a leakage-free k-fold protocol — the LightGBM combiner is trained on validation data and evaluated on held-out folds, so no detection reduces to its own training input. The structural complementarity claim (Sec. VIF, Table VII) is an empirical observation on held-out test data, not a derived result. The five failure-mode taxonomy (Sec. II) is descriptive. Self-citations to companion works [16, 27, 28, 37] provide context, baselines, and infrastructure but are not load-bearing for any mathematical or methodological claim. The skeptic's concern that OE is confounded with cross-family combination in the ablation (Table I) is a validity concern, not circularity: the paper does not define a quantity in terms of what it claims to predict. Score 1 reflects the presence of non-load-bearing self-citations to companion works.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities or postulated particles. The 'invented' components are architectural choices (FiLM conditioning, hierarchical heads, 16-score combiner) rather than new physical postulates. Free parameters are standard ML hyperparameters and loss weights. The axioms are domain assumptions about data exchangeability and sample representativeness, standard for conformal prediction and ML evaluation.

free parameters (5)
  • Loss weights (w_f, w_c, w_sc,f, w_sc,c, w_conf, w_H) = 1.5, 0.3, 0.03, 0.05, 0.5, 0.01
    Hand-set loss term weights (Table XI) chosen to balance classification and contrastive objectives.
  • LightGBM hyperparameters = 300 estimators, eta=0.05, num_leaves=31
    Chosen by 5-fold CV on validation set for the anomaly score combiner.
  • AD-MCP quintile bin edges = 5 equal-frequency bins on calibration set
    Choice of Q=5 quintiles and n_min=10 fallback threshold are design decisions affecting conditional coverage granularity.
  • Outlier exposure margins (lambda_OE, lambda_emb) = 1.0, 1.0
    Margins for the outlier-exposure fine-tuning objective, stated as fixed values.
  • Confusion penalty pairs and weights = (TDE->Ia, 0.5), (SLSN->Ibc, 0.3), (II->Ibc, 0.8)
    Astrophysically motivated targeted misclassification penalties, hand-selected.
assumptions (4)
  • domain assumption Exchangeability of calibration and test spectra
    Conformal prediction validity (Sec. VI) requires that calibration data and new test data are exchangeable. Empirically checked on ZTF held-out set but not guaranteed under LSST distribution shift.
  • domain assumption The 289-spectrum rare set spans the full physical diversity of rare anomalies
    Sec. IIIb states the set is 'constructed to span the five failure-mode regimes.' The taxonomy completeness depends on this.
  • standard math Supervised contrastive loss produces a metrically faithful embedding space
    The AD layer operates on the 192-d embedding assuming distance correlates with typicality. Appendix H provides empirical support (kNN accuracy = classifier accuracy).
  • domain assumption Within-cell exchangeability holds for AD-MCP
    AD-MCP conditional coverage guarantee (Eq. 7) requires exchangeability within each (class, AD-quintile) cell. Not independently verified beyond empirical coverage tracking.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Catching Disguised Transients with ASTRANet: Anomaly-Aware Spectroscopic Classification and Conformal Calibration." pith.science (2026). https://pith.science/paper/FH6PAQ6Z

@misc{pith2026260708044,
  author       = {Pith},
  title        = {Pith review of: Catching Disguised Transients with ASTRANet: Anomaly-Aware Spectroscopic Classification and Conformal Calibration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FH6PAQ6Z}},
  note         = {Machine review of arXiv:2607.08044}
}
abstract

Time-domain surveys discover thousands of transients per year, but the spectroscopic identification of rare and physically peculiar objects remains rate-limited by closed-set classifiers that confidently assign every input to a known class -- including spectra that genuinely belong to no known class. We present the \texttt{ASTRANet} framework, a confidence-aware infrastructure for spectroscopic transient classification built around three coupled modules: a hierarchical spectral classifier that operates directly on observer-frame spectra without requiring host-galaxy redshift or spectral phase as inputs; an anomaly detection layer (\texttt{ASTRANet-Sentinel}) that non-linearly combines $16$ embedding-space anomaly scores spanning four physically motivated families; and a conformal uncertainty quantification layer (\texttt{ASTRANet-CP}). We validate the framework on a held-out evaluation set of $289$ rare and out-of-taxonomy transients spanning $11$ classes deliberately excluded from training, chosen to span the full physical diversity of the rare-anomaly population: AGN-related outliers, GRB-related events, gap transients, novae, and peculiar supernovae. Through five astrophysically distinct failure modes of closed-set classifiers, we show that classifier-internal uncertainty and embedding-based anomaly detection are structurally complementary axes of confidence rather than alternative implementations of the same estimator. We further introduce AD-stratified Mondrian conformal prediction (AD-MCP) within \texttt{ASTRANet-CP}, achieving uniform conditional coverage across anomaly-score strata where vanilla Mondrian under-covers in the operational regime. This establishes the methodological infrastructure for confidence-aware spectroscopic discovery in the Vera C.\ Rubin Observatory era.

Figures

Figures reproduced from arXiv: 2607.08044 by the authors.

Figure 1
Figure 1. FIG. 1. Architecture of [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Confusion matrix on the held-out test set ( [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Per-anomaly detection map at [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: FIG. 4. Mondrian conformal p-value distribution on the [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. Embedding-space nearest-neighbor retrieval, both [PITH_FULL_IMAGE:figures/full_fig_p027_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

85 extracted references · 85 canonical work pages

  1. [1]

    Stochastic depth [79] with linearly increasing drop probability0→0.1provides implicit ensemble reg- ularisation

    A residual skip connection adds the block input to its output. Stochastic depth [79] with linearly increasing drop probability0→0.1provides implicit ensemble reg- ularisation. Cumulative receptive field:∼250samples. c. Metadata conditioning via FiLM.The metadata vectorc∈R 64 encodes three quantities: spectral phase ϕ(rest-frame days from maximum light, re...

    work page

  2. [2]

    Included Subtypes

    with learning rateη0 = 3×10 −4, weight decay10−4, gradient clip1.0. The schedule uses a5-epoch linear warmup followed by cosine annealing with warm restarts [80],T 0 = 30,T mult = 2,η min = 0.05η 0. EMA [81] of model weights is maintained with decayα= 0.996and used for all validation/test evaluations. Early stopping usesvalidationmacrorecallwithpatience45...

    work page

  3. [3]

    find more like this

    over the test-set embeddings reaches86.8%five-fold accuracy – statistically indistinguishable from the full hi- erarchical classifier head (86.6%, Sec. VA). The class structure visible in the UMAP projection is therefore metrically faithful, not a visualization artifact. The same retrieval operates in two complementary regimes (Fig. 6). (ii) In-distributi...

    work page

  4. [4]

    E. C. Bellm, S. R. Kulkarni, M. J. Graham, R. Dekany, R. M. Smith, R. Riddle, F. J. Masci, G. Helou, T. A. Prince, S. M. Adams, C. Barbarino, T. Barlow, J. Bauer, R. Beck, J. Belicki, R. Biswas, N. Blagorodnova, D. Bodewits, B. Bolin, V. Brinnel, T. Brooke, B. Bue, M. Bulla, R. Burruss, S. B. Cenko, C.-K. Chang, A. Con- nolly, M. Coughlin, J. Cromer, V. C...

    work page Pith review arXiv 2019

  5. [5]

    F. J. Masci, R. R. Laher, B. Rusholme, D. L. Shupe, S. Groom, J. Surace, E. Jackson, S. Monkewitz, R. Beck, D. Flynn, S. Terek, W. Landry, E. Hacopians, V. Desai, J. Howell, T. Brooke, D. Imel, S. Wachter, Q.-Z. Ye, H.- W. Lin, S. B. Cenko, V. Cunningham, U. Rebbapragada, B.Bue, A.A.Miller, A.Mahabal, E.C.Bellm, M.T.Pat- terson, M. Jurić, V. Z. Golkhou, E...

    work page Pith review arXiv 2019

  6. [6]

    M. J. Graham, S. R. Kulkarni, E. C. Bellm, S. M. Adams, C. Barbarino, N. Blagorodnova, D. Bodewits, B. Bolin, P. R. Brady, S. B. Cenko, C.-K. Chang, M. W. Coughlin, K. De, G. Eadie, T. L. Farnham, U. Feindt, A. Franck- owiak, C. Fremling, S. Gezari, S. Ghosh, D. A. Gold- stein, V. Z. Golkhou, A. Goobar, A. Y. Q. Ho, D. Hup- penkothen, Ž. Ivezić, R. L. Jon...

    work page Pith review arXiv 2019

  7. [7]

    M. T. Pattersonet al., PASP131, 018001 (2019)

    work page 2019

  8. [8]

    Dekany, R

    R. Dekany, R. M. Smith, R. Riddle, M. Feeney, M. Porter, D. Hale, J. Zolkower, J. Belicki, S. Kaye, J. Henning, R. Walters, J. Cromer, A. Delacroix, H. Ro- driguez, D. J. Reiley, P. Mao, D. Hover, P. Murphy, R. Burruss, J. Baker, M. Kowalski, K. Reif, P. Mueller, E. Bellm, M. Graham, and S. R. Kulkarni, PASP132, 038001 (2020), arXiv:2008.04923 [astro-ph.IM]

    work page arXiv 2020

Show all 85 references
  1. [9]

    Ivezić, S

    Ž. Ivezić, S. M. Kahn,et al., The Astrophysical Journal 873, 111 (2019), arXiv:0805.2366

    2019 arXiv

  2. [10]

    Blagorodnova, J

    N. Blagorodnova, J. D. Neill, R. Walters, S. R. Kulkarni, C. Fremling, S. Ben-Ami, R. G. Dekany, J. R. Fucik, N. Konidaris, R. Nash, C.-C. Ngeow, E. O. Ofek, D. O. Sullivan, R. Quimby, A. Ritter, and K. E. Vyhmeister, PASP130, 035003 (2018)

    2018

  3. [11]

    Rigault, J

    M. Rigault, J. D. Neill, N. Blagorodnova, A. Dugas, M. Feeney, R. Walters, V. Brinnel, Y. Copin, C. Frem- ling, J. Nordin, and J. Sollerman, A&A627, A115 (2019), arXiv:1902.08526 [astro-ph.IM]

    2019 arXiv

  4. [12]

    R. J. Foley, P. J. Challis, R. Chornock,et al., The Astro- physical Journal767, 57 (2013), arXiv:1212.2209

    2013 arXiv

  5. [13]

    H. B. Perets, A. Gal-Yam, P. A. Mazzali, D. Arnett, D. Kagan, A. V. Filippenko, W. Li, I. Arcavi, S. B. Cenko, D. B. Fox,et al., Nature465, 322 (2010), arXiv:0906.2003 [astro-ph.HE]

    2010 arXiv

  6. [14]

    M. R. Drout, R. Chornock, A. M. Soderberg, N. E. Sanders, R. McKinnon, A. Rest, R. J. Foley, D. Milisavl- jevic, R. Margutti, E. Berger,et al., The Astrophysical Journal794, 23 (2014), arXiv:1405.3668 [astro-ph.HE]

    2014 arXiv

  7. [15]

    A. Y. Q. Ho, D. A. Perley, A. Gal-Yam, R. Lunnan, J. Sollerman, S. Schulze, K. K. Das, D. Dobie, Y. Yao, C. Fremling, S. Adams, S. Anand, I. Andreoni, E. C. Bellm, R. J. Bruch, K. B. Burdge, A. J. Castro-Tirado, A. Dahiwale, K. De, R. Dekany, A. J. Drake, D. A. Duev, M. J. Gra...

    2023

  8. [16]

    T. A. Thompson, J. L. Prieto, K. Z. Stanek, M. D. Kistler, J. F. Beacom, and C. S. Kochanek, The As- trophysical Journal705, 1364 (2009), arXiv:0809.0510 [astro-ph]

    2009 arXiv

  9. [17]

    Fremling, X

    C. Fremling, X. J. Hall, M. W. Coughlin, A. S. Dahiwale, D. A. Duev, M. J. Graham, M. M. Kasliwal, E. C. Kool, A. A. Mahabal, A. A. Miller, J. D. Neill, D. A. Perley, M. Rigault, P. Rosnet, B. Rusholme, Y. Sharma, K. M. Shin, D. L. Shupe, J. Sollerman, and R. S. Walters, ApJL ...

    2021 arXiv

  10. [18]

    Sharma, A

    Y. Sharma, A. A. Mahabal, J. Sollerman, C. Fremling, S. R. Kulkarni, N. Rehemtulla, A. A. Miller, M. Aubert, 29 T. X. Chen, M. W. Coughlin, M. J. Graham, D. Hale, M. M. Kasliwal, Y.-L. Kim, J. D. Neill, J. N. Pur- dum, B. Rusholme, A. Singh, and N. Sravan, PASP137, 034507 (2025)

    2025

  11. [19]

    M. Xu, A. Sasli, A. Junell, F. F. Nunes, Y.-J. Qin, C. Fremling, S. Rose, T. J. D. Laz, B. Border, A. L. Calloch, S. S. Chaudhary, H. Markoff, A. Raghuvan- shi, N. Rehemtulla, J. Sollerman, Y. Sharma, N. Sravan, J. Adler, T. X. Chen, R. Dekany, R. Riddle, M. M. Kasli- wal, M. ...

    2025

  12. [20]

    W. F. Fortino, F. B. Bianco, P. Protopapas, D. Muthukr- ishna, and A. Brockmeier, Abc-sn: Attention based classifier for supernova spectra (2025), arXiv:2507.22106 [astro-ph.IM]

    2025 arXiv

  13. [21]

    Hendrycks and K

    D. Hendrycks and K. Gimpel, ICLR (2017)

    2017

  14. [22]

    C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger, in Proceedings of the 34th International Conference on Ma- chine Learning (ICML), Proceedings of Machine Learn- ing Research, Vol. 70 (PMLR, 2017) pp. 1321–1330, arXiv:1706.04599 [cs.LG]

    2017 arXiv

  15. [23]

    M. Hein, M. Andriushchenko, and J. Bitterwolf, inPro- ceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR)(2019) pp. 41–50, arXiv:1812.05720 [cs.LG]

    2019 arXiv

  16. [24]

    V. A. Villar, M. Cranmer, E. Berger,et al., The As- trophysical Journal Supplement Series255, 24 (2021), arXiv:2103.12102

    2021 arXiv

  17. [25]

    M. V. Pruzhinskaya, K. L. Malanchev, M. V. Kornilov, et al., Monthly Notices of the Royal Astronomical Society 489, 3591 (2019), arXiv:1905.11516

    2019 arXiv

  18. [26]

    Chaini, F

    S. Chaini, F. B. Bianco, and A. Mahabal, In search of the unknown unknowns: A multi-metric distance ensemble for out of distribution anomaly detection in astronomical surveys (2025), arXiv:2510.23702 [astro-ph.IM]

    2025

  19. [27]

    Guptaet al., arXiv e-prints (2025),Update with ac- tual MCIF reference., 2503.00000

    R. Guptaet al., arXiv e-prints (2025),Update with ac- tual MCIF reference., 2503.00000

    2025

  20. [28]

    Perez, F

    E. Perez, F. Strub, H. de Vries, V. Dumoulin, and A. Courville, inProceedings of the AAAI Conference on Artificial Intelligence, Vol. 32 (2018) p. N/A

    2018

  21. [29]

    Muthukrishna, D

    D. Muthukrishna, D. Parkinson, and B. E. Tucker, ApJ 885, 85 (2019)

    2019

  22. [30]

    T. J. du Laz, M. W. Coughlin, P. Bachant, J. E. Simones, T. Culino, A. L. Calloch, S. S. Chaudhary, X. J. Hall, T. Barna, D. Warshofsky, M. Graham, M. M. Kasliwal, A. Mahabal, J. S. Bloom, A. Palmese, F. J. Masci, S. L. Groom, R. Dekany, R. L. Riddle, and G. Helou, Boom and ba...

    2025 arXiv

  23. [31]

    Junell, A

    A. Junell, A. Sasli, F. F. Nunes, M. Xu, B. Border, N. Re- hemtulla, M. Rizhko, Y.-J. Qin, T. J. D. Laz, A. L. Calloch, S. S. Chaudhary, S. Wu, J. Sollerman, N. Sra- van, S. L. Groom, D. Hale, M. M. Kasliwal, J. Purdum, A. Wold, M. J. Graham, and M. W. Coughlin, Apply- ing mul...

    2025 arXiv

  24. [32]

    Yuksekgonul, L

    M. Yuksekgonul, L. Zhang, J. Zou, and C. Guestrin, Be- yond confidence: Reliable models should also consider atypicality (2023), arXiv:2305.18262 [cs.LG]

    2023 arXiv

  25. [33]

    Margutti, B

    R. Margutti, B. D. Metzger, R. Chornock, I. Vurm, N. Roth, B. W. Grefenstette, V. Savchenko, R. Cartier, J. F. Steiner, G. Terreran,et al., The Astrophysical Jour- nal872, 18 (2019), arXiv:1810.10720 [astro-ph.HE]

    2019 arXiv

  26. [34]

    D. A. Perley, P. A. Mazzali, L. Yan, S. B. Cenko, S. Gezari, K. Taggart, N. Blagorodnova, C. Frem- ling, B. Mockler, A. Singh,et al., Monthly Notices of the Royal Astronomical Society484, 1031 (2019), arXiv:1808.00969 [astro-ph.HE]

    2019 arXiv

  27. [35]

    C. S. Kochanek, D. M. Szczygieł, and K. Z. Stanek, The Astrophysical Journal758, 142 (2012), arXiv:1202.0281 [astro-ph.SR]

    2012 arXiv

  28. [36]

    Barna, C

    T. Barna, C. Fremling, T. Ahumada, I. Andreoni, S. Banerjee, J. S. Bloom, M. Bulla, T. X. Chen, M. W. Coughlin, T.Dietrich, X.J.Hall, A.Junell, B.Rusholme, J. Sollerman, and N. Sravan, Publications of the Astro- nomical Society of the Pacific137, 084105 (2025)

    2025

  29. [37]

    Fremling, D

    C. Fremling, D. Perley, A. Y. Q. Ho, and ZTF Collabo- ration, GRB Coordinates Network31629, 1 (2022)

    2022

  30. [38]

    S. J. van der Walt, A. Crellin-Quick, and J. S. Bloom, Journal of Open Source Software4, 1247 (2019)

    2019

  31. [39]

    M. W. Coughlin, J. S. Bloom, G. Nir, S. Antier, T. J. du Laz, S. van der Walt, A. Crellin-Quick, T. Culino, D. A. Duev, D. A. Goldstein, B. F. Healy, V. Karam- belkar, J. Lilleboe, K. M. Shin, L. P. Singer, T. Ahu- mada, S. Anand, E. C. Bellm, R. Dekany, M. J. Graham, M. M. Ka...

    2023 arXiv

  32. [40]

    Sasli, IRIS: Identification and Reduction of Interest- ing Spectra,https://github.com/applecider-ml/IRIS (2026), spectral preprocessing pipeline for transient clas- sification

    A. Sasli, IRIS: Identification and Reduction of Interest- ing Spectra,https://github.com/applecider-ml/IRIS (2026), spectral preprocessing pipeline for transient clas- sification

    2026

  33. [41]

    Fremling, A

    C. Fremling, A. A. Miller, Y. Sharma, A. Dugas, D. A. Perley, K. Taggart, J. Sollerman, A. Goobar, M. L. Gra- ham, J. D. Neill, J. Nordin, M. Rigault, R. Walters, I. Andreoni, A. Bagdasaryan, J. Belicki, C. Cannella, E. C. Bellm, S. B. Cenko, K. De, R. Dekany, S. Fred- erick, ...

    2020 arXiv

  34. [42]

    D. A. Perley, C. Fremling, J. Sollerman, A. A. Miller, A. S. Dahiwale, Y. Sharma, E. C. Bellm, R. Biswas, T. G. Brink, R. J. Bruch, K. De, R. Dekany, A. J. Drake, D. A. Duev, A. V. Filippenko, A. Gal-Yam, A. Goobar, M. J. Graham, M. L. Graham, A. Y. Q. Ho, I. Irani, M. M. Kasl...

    2020

  35. [43]

    Rehemtulla, A

    N. Rehemtulla, A. A. Miller, T. Jegou Du Laz, M. W. Coughlin, C. Fremling, D. A. Perley, Y.-J. Qin, J. Sollerman, A. A. Mahabal, R. R. Laher, R. Riddle, B. Rusholme, and S. R. Kulkarni, ApJ972, 7 (2024), arXiv:2401.15167 [astro-ph.IM]

    2024 arXiv

  36. [44]

    Blagorodnova, J

    N. Blagorodnova, J. D. Neill, R. Walters, S. R. Kulka- 30 rni, C. Fremling, S. Ben-Ami, R. G. Dekany, J. R. Fu- cik, N. Konidaris, R. Nash, C.-C. Ngeow, E. O. Ofek, D.O’Sullivan, R.Quimby, A.Ritter,andK.E.Vyhmeis- ter, PASP130, 035003 (2018), arXiv:1710.02917 [astro- ph.IM]

    2018 arXiv

  37. [45]

    Y. L. Kim, M. Rigault, J. D. Neill, M. Briday, Y. Copin, J. Lezmy, N. Nicolas, R. Riddle, Y. Sharma, M. Smith, J. Sollerman, and R. Walters, PASP134, 024505 (2022), arXiv:2203.01346 [astro-ph.IM]

    2022 arXiv

  38. [46]

    J. B. Oke, J. G. Cohen, M. Carr, J. Cromer, A. Din- gizian, F. H. Harris, S. Labrecque, R. Lucinio, W. Schaal, H. Epps, and J. Miller, Publications of the Astronomical Society of the Pacific107, 375 (1995)

    1995

  39. [47]

    S. M. Faber, A. C. Phillips, R. I. Kibrick, B. Alcott, S. L. Allen, J. Burrous, T. Cantrall, D. Clarke, A. L. Coil, D. J. Cowley, M. Davis, W. T. S. Deich, K. Di- etsch, D. K. Gilmore, C. A. Harper, D. F. Hilyard, J. P. Lewis, M.McVeigh, J.Newman, J.Osborne, R.Schiavon, R. J. ...

    2003

  40. [48]

    R. L. Davies, J. R. Allington-Smith, P. Bettess, E. Chad- wick, R. Content, G. N. Dodsworth, R. Haynes, D. Lee, I. J. Lewis, J. Webster, E. Atad, S. M. Beard, M. El- lis, P. R. Hastings, P. R. Williams, T. Bond, D. Cramp- ton, T. J. Davidge, M. Fletcher, B. Leckie, C. L. Mor- ...

    1997

  41. [49]

    R. W. e. a. Pogge, inSPIE, Vol. 7735 (2010)

    2010

  42. [50]

    M. M. Kasliwal, C. Fremling, L. Yan, K. Das, F. Verdi, J. Zmudzinas, C. Martin, E. Kirby, S. Xue, L. Ho, G. Herczeg, X. Wu, Z. Hu, H. Ji, M. Matuszewski, R. Bertz, D. Hale, H. Rodriguez, A. Boden, R. Dekany, R. Smith, D. Reiley, R. Nash, J. Milburn, D. Neill, J. Brugger, R. Za...

    2024

  43. [51]

    Accetta, C

    Abdurro’uf, K. Accetta, C. Aerts, V. Silva Aguirre, R. Ahumada, N. Ajgaonkar, N. Filiz Ak, S. Alam, C. Allende Prieto, A. Almeida, F. Anders, S. F. An- derson, B. H. Andrews, B. Anguiano, E. Aquino-Ortíz, A. Aragón-Salamanca, M. Argudo-Fernández, M. Ata, M. Aubert, V. Avila-Re...

    2022

  44. [52]

    DESI Collaboration, A. G. Adame, J. Aguilar, S. Ahlen, S. Alam, G. Aldering, D. M. Alexander, R. Alfarsy, C. Allende Prieto, M. Alvarez, O. Alves, A. Anand, F. Andrade-Oliveira, E. Armengaud, J. Asorey, S. Avila, A.Aviles, S.Bailey, A.Balaguera-Antolínez, O.Ballester, C. Balta...

    2024 arXiv

  45. [53]

    M. M. Kasliwal, C. Cannella, A. Bagdasaryan, T. Hung, U. Feindt, L. P. Singer, M. Coughlin, C. Fremling, R. Walters, D. Duev, R. Itoh, and R. M. Quimby, PASP 131, 038003 (2019)

    2019

  46. [54]

    D. A. Coulter, D. O. Jones, P. McGill, R. J. Foley, P. D. Aleo, M. J. Bustamante-Rosell, D. Chatterjee, K. W. Davis, C. Dickinson, A. Engel, A. Gagliano, W. V. Jacobson-Galán, C. D. Kilpatrick, J. Kutcka, X. K. Le Saux, K. Malanchev, Y. C. Pan, P. J. Quiñonez, C. Rojas-Bravo, ...

    2023 arXiv

  47. [55]

    S. Bai, J. Z. Kolter, and V. Koltun, arXiv preprint arXiv:1803.01271 (2018)

    2018 arXiv

  48. [56]

    M. Ilse, J. M. Tomczak, and M. Welling, inInternational Conference on Machine Learning (ICML)(PMLR, 2018) pp. 2127–2136

    2018

  49. [57]

    Khosla, P

    P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y. Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, inAdvances in Neural Information Processing Systems (NeurIPS), Vol. 33 (2020) pp. 18661–18673

    2020

  50. [58]

    T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, inProceedings of the IEEE International Conference on Computer Vision (ICCV)(2017) pp. 2980–2988

    2017

  51. [59]

    Loshchilov and F

    I. Loshchilov and F. Hutter, inInternational Conference on Learning Representations (ICLR)(2019)

    2019

  52. [60]

    Zhang, M

    H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz, inInternational Conference on Learning Representations (ICLR)(2018)

    2018

  53. [61]

    Y. Cui, M. Jia, T.-Y. Lin, Y. Song, and S. Belongie, inProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR)(2019)pp. 9268–9277

    2019

  54. [62]

    Lakshminarayanan, A

    B. Lakshminarayanan, A. Pritzel, and C. Blundell, in Advances in Neural Information Processing Systems 30 (NIPS 2017)(2017) pp. 6402–6413, arXiv:1612.01474 [stat.ML]

    2017 arXiv

  55. [63]

    Liang, Y

    S. Liang, Y. Li, and R. Srikant, Enhancing the reliability of out-of-distribution image detection in neural networks (2020), arXiv:1706.02690 [cs.LG]

    2020 arXiv

  56. [64]

    W. Liu, X. Wang, J. Owens, and Y. Li, NeurIPS (2020)

    2020

  57. [65]

    Leeet al., NeurIPS (2018)

    K. Leeet al., NeurIPS (2018)

    2018

  58. [66]

    F. T. Liu, K. M. Ting, and Z.-H. Zhou, inProceedings of the 8th IEEE International Conference on Data Mining (ICDM)(IEEE Computer Society, 2008) pp. 413–422

    2008

  59. [67]

    M. M. Breunig, H.-P. Kriegel, R. T. Ng, and J. Sander, inProceedings of the 2000 ACM SIGMOD International Conference on Management of Data(ACM, 2000) pp. 93–104

    2000

  60. [68]

    G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T.-Y. Liu, inAdvances in Neural Informa- tion Processing Systems 30 (NIPS 2017)(2017)pp.3146– 3154

    2017

  61. [69]

    Hendrycks, M

    D. Hendrycks, M. Mazeika, and T. Dietterich, inInterna- tional Conference on Learning Representations (ICLR) (2019) arXiv:1812.04606

    2019 arXiv

  62. [70]

    V. Vovk, A. Gammerman, and G. Shafer,Algorith- mic Learning in a Random World(Springer, New York, 2005)

    2005

  63. [71]

    A. N. Angelopoulos and S. Bates, Foundations and Trends in Machine Learning16, 494 (2023)

    2023

  64. [72]

    Romano, M

    Y. Romano, M. Sesia, and E. J. Candès, inAdvances in Neural Information Processing Systems (NeurIPS) (2020) arXiv:2006.02544

    2020 arXiv

  65. [73]

    M.L.Li, A.Y.Q.Ho, G.Ryan, D.A.Perley, G.P.Lamb, A. J. Nayana, I. Andreoni, G. C. Anupama, E. C. Bellm, E. Berger, J. S. Bloom, E. Burns, I. Caiazzo, P. Chandra, M. W. Coughlin, K. El-Badry, M. J. Graham, M. Kasli- wal, G. K. Keating, S. R. Kulkarni, H. Kumar, F. J. Masci, R.A....

    2025 arXiv

  66. [74]

    Chaini, A

    S. Chaini, A. Mahabal, A. Kembhavi, and F. Bianco, Astronomy and Computing48, 100850 (2024)

    2024

  67. [75]

    Swann, M

    E. Swann, M. Sullivan, D. Carollo,et al., The Messenger 175, 58 (2019), arXiv:1903.02476

    2019 arXiv

  68. [76]

    Frohmaier, M

    C. Frohmaier, M. Sullivan, K. Maguire,et al., inProceed- ings of SPIE, Vol. 13096 (2024) p. 130963X

    2024

  69. [77]

    Schipani, S

    P. Schipani, S. Campana, R. Claudi,et al., inProceedings of SPIE, Vol. 11447 (2020) p. 1144709

    2020

  70. [78]

    P. J. Groot, S. Scaringi, and N. Elias-Rosa, arXiv e-prints , arXiv:2512.17404 (2025), arXiv:2512.17404 [astro-ph.IM]

    2025

  71. [79]

    Savitzky and M

    A. Savitzky and M. J. E. Golay, Analytical Chemistry 36, 1627 (1964)

    1964

  72. [80]

    Wu and K

    Y. Wu and K. He, inEuropean Conference on Computer Vision (ECCV)(Springer, 2018) pp. 3–19

    2018

  73. [81]

    Hendrycks and K

    D. Hendrycks and K. Gimpel, arXiv preprint arXiv:1606.08415 (2016)

    2016 arXiv

  74. [82]

    Huang, Y

    G. Huang, Y. Sun, Z. Liu, D. Sedra, and K. Q. Wein- berger, inEuropean Conference on Computer Vision (ECCV)(Springer, 2016) pp. 646–661

    2016

  75. [83]

    Loshchilov and F

    I. Loshchilov and F. Hutter, inInternational Conference on Learning Representations (ICLR)(2017)

    2017

  76. [84]

    B. T. Polyak and A. B. Juditsky, SIAM Journal on Con- trol and Optimization30, 838 (1992)

    1992

  77. [85]

    Izmailov, D

    P. Izmailov, D. Podoprikhin, T. Garipov, D. Vetrov, and A. G. Wilson, inProceedings of the Thirty-Fourth Con- ference on Uncertainty in Artificial Intelligence (UAI) (2018) pp. 876–885, arXiv:1803.05407 [cs.LG]

    2018 arXiv

Pith tools

Reviewed July 10, 2026 · model on record in the stance chip above.