pith. machine review for the scientific record. sign in

arxiv: 2605.11033 · v1 · submitted 2026-05-10 · ⚛️ physics.plasm-ph · cs.AI

Recognition: 2 theorem links

· Lean Theorem

TokaMind for Power Grid: Cross-Domain Transfer from Fusion Plasma

JC Wu , Norton Lee , Kai Siang Chen
Authors on Pith no claims yet

Pith reviewed 2026-05-13 00:55 UTC · model grok-4.3

classification ⚛️ physics.plasm-ph cs.AI
keywords cross-domain transfertokamak plasma diagnosticspower grid PMUsevere event classificationmulti-modal transformercritical slowing downgrid topologyearly warning
0
0 comments X

The pith

TokaMind pretrained on tokamak plasma data achieves F1 0.837 on power grid severe event classification, with difficulty set by grid topology not model capacity.

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

The paper establishes that a transformer model trained on fusion plasma diagnostics from the MAST tokamak can transfer its learned representations to classify severe events in power grid synchrophasor data. This matters because fusion experiments generate high-quality, multi-modal time series that could bootstrap monitoring in other critical infrastructure without requiring equivalent data volumes. Systematic tests across degradation and grid domains pinpoint four characteristics that favor such transfer, and power grid data aligns closely with them. The results also show that grid provider topology dictates how hard classification is, more so than the choice of model, and that critical slowing down metrics can gate predictions for better reliability in early warning scenarios.

Core claim

TokaMind, a multi-modal transformer pre-trained on tokamak plasma diagnostics, generalizes to power grid PMU data for severe event classification, reaching a test F1 of 0.837 on the GESL/PNNL benchmark. Classification difficulty correlates with provider-level grid topology rather than model capacity. In the single-window regime, it edges out a CNN baseline, an advantage that vanishes with additional event windows, and Critical Slowing Down indicators used as a confidence gate raise F1 from 0.696 to 0.750 at 63 percent coverage.

What carries the argument

The four transfer-favoring characteristics identified through cross-domain experiments that predict when plasma-pretrained representations succeed on new physical systems such as power grids.

Load-bearing premise

The four transfer-favoring characteristics fully account for why representations from one physical system successfully apply to another with different underlying equations.

What would settle it

Measure classification F1 scores on two grid providers that share identical topologies but have different event statistics, or on grids whose topologies have been deliberately randomized, to test whether performance gaps track topology changes alone.

Figures

Figures reproduced from arXiv: 2605.11033 by JC Wu, Kai Siang Chen, Norton Lee.

Figure 5
Figure 5. Figure 5: fig. 5 [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 1
Figure 1. Figure 1: Test F1 vs. number of input windows (seq_len). TokaMind leads in the single-window early-warning regime; CNN recovers as more windows are provided. Reversal point between seq_len=1 and seq_len=2. This result suggests that TokaMind’s fusion￾pretrained physical coupling representations carry unique value in the information-minimal early-warning setting–precisely where CNN’s local amplitude aggre￾gation fails… view at source ↗
Figure 3
Figure 3. Figure 3: CSD selective prediction framework. Toka [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Two-stage adaptation protocol following TokaMind. MAST pre-trained weights loaded as warmstart (50/66 layers, 75.8%). Stage 1 freezes backbone (143,810 trainable params); Stage 2 applies selective fine-tuning (37,442 trainable params). Stage 1 — Frozen backbone. The backbone (50/66 layers) is initialized from MAST pre￾trained weights and held frozen. Only the task-specific classification head (16/66 layers… view at source ↗
read the original abstract

TokaMind is a multi-modal transformer (MMT) foundation model pre-trained on tokamak plasma diagnostics data from MAST, where it was shown to outperform CNN-based approaches on fusion benchmarks. We investigate whether its learned representations generalize to physically distinct but structurally analogous domains. Through systematic experimentation across four domains-industrial bearing degradation, NASA CMAPSS turbofan degradation, and two independent power grid PMU datasets-we identify four transfer-favoring characteristics that help explain where TokaMind's pretrained representations are most effective. Power grid synchrophasor data matches this target-domain profile most directly, while industrial degradation datasets demonstrate that TokaMind can still yield useful performance under partial alignment, especially when task design and feature construction expose physically meaningful degradation structure. On the GESL/PNNL 500-event benchmark with provider-aware evaluation, TokaMind achieves test $\text{F1} = 0.837 \pm 0.040$ (3~seeds) for severe event classification. Our central finding, however, is not the aggregate score: classification difficulty is structurally determined by provider-level grid topology, not model capacity. In the single-window early-warning regime, TokaMind outperforms a CNN baseline (F1~0.889 vs.~0.878)--a reversal that disappears as more event windows are provided. Furthermore, Critical Slowing Down (CSD) indicators, used as a confidence gate rather than a classification label, improve F1 from 0.696 to 0.750 at 63% coverage-outperforming the CNN baseline (0.636) at any coverage level. These results establish the first cross-domain validation of TokaMind outside nuclear fusion and propose a transferability framework and revised evaluation protocol for multi-source PMU datasets.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper presents TokaMind, a multi-modal transformer pre-trained on MAST tokamak plasma diagnostics, and evaluates its zero-shot transfer to power-grid PMU data for severe event classification. It identifies four transfer-favoring characteristics from experiments on industrial bearing, NASA CMAPSS, and two PMU datasets; reports test F1 = 0.837 ± 0.040 (3 seeds) on the GESL/PNNL 500-event benchmark under provider-aware evaluation; shows TokaMind outperforming a CNN baseline in the single-window regime (0.889 vs 0.878) with reversal at longer windows; and demonstrates that Critical Slowing Down indicators used as a confidence gate raise F1 from 0.696 to 0.750 at 63 % coverage. The central claim is that classification difficulty is structurally determined by provider-level grid topology rather than model capacity, together with a proposed transferability framework.

Significance. If the topology-driven difficulty claim and the causal role of the four characteristics are substantiated, the work would constitute the first documented cross-domain validation of a fusion-plasma foundation model on power-grid synchrophasor data and could usefully revise evaluation protocols for multi-source PMU benchmarks. The empirical F1 numbers, the single-window reversal, and the CSD gating result are concrete and falsifiable; the transferability framework itself is a potentially reusable contribution. At present, however, the absence of architecture details, training protocols, ablations, and a direct capacity-versus-topology test leaves the central claims under-supported.

major comments (3)
  1. [Abstract] Abstract: the load-bearing claim that 'classification difficulty is structurally determined by provider-level grid topology, not model capacity' is not accompanied by any experiment that holds topology fixed while varying model capacity; the reported CNN baseline comparisons and multi-window results therefore do not isolate the asserted structural factor.
  2. [Abstract] Abstract: the four transfer-favoring characteristics are asserted to explain where TokaMind's representations are effective, yet no quantitative correlation, ablation, or causal test is supplied to show they are drivers rather than correlates of the observed F1 = 0.837, particularly across domains whose underlying physics (MHD instabilities versus power-flow dynamics) differ substantially.
  3. [Abstract] Abstract and experimental sections: concrete F1 scores, standard deviations, and topology conclusions are presented without model architecture specifications, training protocol, hyper-parameter choices, or ablation tables, preventing independent assessment of whether the reported transfer performance is robust or capacity-dependent.
minor comments (1)
  1. [Abstract] The abstract states that power-grid PMU data 'matches this target-domain profile most directly' but does not list the exact numerical values or statistical tests used to rank the four domains against the four characteristics.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below with honest clarifications and commit to revisions that strengthen the supporting evidence without overstating the current results.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the load-bearing claim that 'classification difficulty is structurally determined by provider-level grid topology, not model capacity' is not accompanied by any experiment that holds topology fixed while varying model capacity; the reported CNN baseline comparisons and multi-window results therefore do not isolate the asserted structural factor.

    Authors: We agree that the manuscript lacks a controlled experiment that holds topology fixed while varying model capacity, and that the CNN baseline and multi-window results provide only indirect support. The claim is currently inferred from the provider-aware splits on the GESL/PNNL benchmark, where the same TokaMind model yields markedly different F1 scores across providers whose grids differ in topology, with the pattern replicated by the CNN. The single-window reversal is consistent with topology-driven early-warning difficulty. To isolate the factor directly, we will add a new experiment in the revision using capacity-reduced TokaMind variants (fewer layers/attention heads) evaluated on topology-matched data subsets from the same providers. revision: yes

  2. Referee: [Abstract] Abstract: the four transfer-favoring characteristics are asserted to explain where TokaMind's representations are effective, yet no quantitative correlation, ablation, or causal test is supplied to show they are drivers rather than correlates of the observed F1 = 0.837, particularly across domains whose underlying physics (MHD instabilities versus power-flow dynamics) differ substantially.

    Authors: The four characteristics were identified via systematic cross-domain comparison and are offered as explanatory factors aligned with observed transfer performance. We concur that the current presentation does not include quantitative correlation coefficients, feature ablations, or causal interventions to establish them as drivers rather than correlates, especially given the physics mismatch between domains. In the revision we will add (i) a correlation table linking characteristic presence to per-domain F1 and (ii) a targeted ablation that masks or perturbs inputs corresponding to each characteristic while measuring impact on transfer F1. revision: yes

  3. Referee: [Abstract] Abstract and experimental sections: concrete F1 scores, standard deviations, and topology conclusions are presented without model architecture specifications, training protocol, hyper-parameter choices, or ablation tables, preventing independent assessment of whether the reported transfer performance is robust or capacity-dependent.

    Authors: The full manuscript contains these elements: architecture (layer count, embedding size, multi-modal fusion) is specified in Section 3.1; training protocol, optimizer, learning-rate schedule, batch size, and epoch counts appear in Section 4.2; ablation studies on pre-training objectives and input modalities are reported in Appendix B. We acknowledge that these details may not have been sufficiently prominent or cross-referenced in the abstract and main experimental narrative. In the revision we will insert a concise architecture-and-hyperparameter summary table in the main text, add explicit cross-references from the abstract and results sections, and ensure all reported F1 values are tied to the exact experimental configuration. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical transfer results are independent of inputs

full rationale

The paper reports empirical F1 scores from pre-training on tokamak data and testing on held-out power-grid PMU events, plus cross-domain experiments that identify four transfer-favoring characteristics. No equations, parameters, or metrics are defined in terms of the reported outcomes; classification difficulty is asserted from observed provider-level topology effects rather than by construction. All central claims rest on external benchmarks and baseline comparisons, making the derivation self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unproven premise that tokamak plasma diagnostics and power-grid PMU streams share enough structural features for representation transfer to succeed; no independent evidence for this structural analogy is supplied beyond the reported performance numbers.

axioms (1)
  • domain assumption Tokamak plasma diagnostics data and power-grid PMU data share transferable structural features despite distinct underlying physics.
    This premise is required for the cross-domain transfer to be meaningful and is invoked to explain why power-grid data matches the target profile most closely.

pith-pipeline@v0.9.0 · 5625 in / 1474 out tokens · 55331 ms · 2026-05-13T00:55:50.657692+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages · 3 internal anchors

  1. [1]

    TokaMind: A multi- modal transformer foundation model for toka- mak plasma dynamics.arXiv preprint, arXiv:2602.15084, 2026

    Tobia Boschi et al. TokaMind: A multi- modal transformer foundation model for toka- mak plasma dynamics.arXiv preprint, arXiv:2602.15084, 2026. URL https://arxi v.org/abs/2602.15084

    work page arXiv 2026

  2. [2]

    FAIR-MAST: A fusion device data management system.SoftwareX, 27 (5):101869, 2024

    Samuel Jackson, Saiful Khan, Nathan Cummings, James Hodson, et al. FAIR-MAST: A fusion device data management system.SoftwareX, 27 (5):101869, 2024. doi: 10.1016/j.softx.2024.101

    work page doi:10.1016/j.softx.2024.101 2024

  3. [3]

    URL https://doi.org/10.1016/j.soft x.2024.101869

    work page doi:10.1016/j.soft 2024

  4. [4]

    On the Opportunities and Risks of Foundation Models

    Rishi Bommasani et al. On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258, 2021. URL https://arxiv. org/abs/2108.07258

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  5. [5]

    Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021. doi: 10.103 8 8/s42254-021-00314-5. URL https://doi.org/ 10.1038/s42254-021-00314-5

    work page doi:10.1038/s42254-021-00314-5 2021

  6. [6]

    Scientific machine learn- ing through physics–informed neural networks: Where we are and what’s next.Journal of Sci- entific Computing, 92(3):88, 2022

    Salvatore Cuomo, Vincenzo Schiano Di Cola, Fabio Giampaolo, Gianluigi Rozza, Maziar Raissi, and Francesco Piccialli. Scientific machine learn- ing through physics–informed neural networks: Where we are and what’s next.Journal of Sci- entific Computing, 92(3):88, 2022. doi: 10.1007/ s10915-022-01939-z

    work page 2022

  7. [7]

    To- wards foundation models for scientific machine learning: Characterizing scaling and transfer be- havior.arXiv preprint arXiv:2306.00258, 2023

    Shashank Subramanian, Peter Harrington, Kurt Keutzer, Wahid Bhimji, Dmitry Morozov, Michael W Mahoney, and Amir Gholami. To- wards foundation models for scientific machine learning: Characterizing scaling and transfer be- havior.arXiv preprint arXiv:2306.00258, 2023. URLhttps://arxiv.org/abs/2306.00258

    work page arXiv 2023

  8. [8]

    Fast al- gorithm for the 3-D DCT-II.IEEE Transactions on Signal Processing, 52(4):992–1001, 2004

    Said Boussakta and Othman Alshibami. Fast al- gorithm for the 3-D DCT-II.IEEE Transactions on Signal Processing, 52(4):992–1001, 2004. doi: 10.1109/TSP.2004.823472

    work page doi:10.1109/tsp.2004.823472 2004

  9. [9]

    A Generalist Agent

    Scott Reed et al. A generalist agent.arXiv preprint arXiv:2205.06175, 2022. URL https: //arxiv.org/abs/2205.06175

    work page internal anchor Pith review arXiv 2022

  10. [10]

    Phadke and James S

    Arun G. Phadke and James S. Thorp.Synchro- nized Phasor Measurements and Their Applica- tions. Springer International Publishing, Cham, Switzerland, second edition, 2017. ISBN 978-3- 319-50584-8. doi: 10.1007/978-3-319-50584-8

    work page doi:10.1007/978-3-319-50584-8 2017

  11. [11]

    An open-source library of phasor measurement unit data capturing real bulk power systems behavior.IEEE Access, 2023

    Shuchismita Biswas, Jim Follum, Pavel Etingov, Xiaoyuan Fan, et al. An open-source library of phasor measurement unit data capturing real bulk power systems behavior.IEEE Access, 2023. doi: 10.1109/ACCESS.2023.3321317. URL http s://doi.org/10.1109/ACCESS.2023.3321317

    work page doi:10.1109/access.2023.3321317 2023

  12. [12]

    Prabha Kundur, John Paserba, Venkat Ajjarapu, Göran Andersson, et al. Definition and classifi- cation of power system stability: IEEE/CIGRE joint task force on stability terms and defini- tions.IEEE Transactions on Power Systems, 19 (3):1387–1401, 2004. doi: 10.1109/TPWRS.20 04.825981. URL https://doi.org/10.1109/TP WRS.2004.825981

    work page doi:10.1109/tpwrs.20 2004

  13. [13]

    Khadem, Alan Collinson, Kyle C

    Lexuan Meng, Jawwad Zafar, Shafiuzzaman K. Khadem, Alan Collinson, Kyle C. Murchie, Fed- erico Coffele, and Graeme Burt. Fast frequency response from energy storage systems: A review of grid standards, projects and technical issues. IEEE Transactions on Smart Grid, 11(2):1566– 1581, 2019. doi: 10.1109/TSG.2019.2940173

    work page doi:10.1109/tsg.2019.2940173 2019

  14. [14]

    Grid event sig- nature library (gesl)

    Oak Ridge National Laboratory and Lawrence Livermore National Laboratory. Grid event sig- nature library (gesl). https://gsl.ornl.gov ,

    work page

  15. [15]

    Accessed April 2026

    Open-access repository of power system measurement signatures. Accessed April 2026

    work page 2026

  16. [16]

    TokaMark: A bench- mark for fusion plasma dynamics models.arXiv preprint, 2026

    Cécile Rousseau et al. TokaMark: A bench- mark for fusion plasma dynamics models.arXiv preprint, 2026. URLhttps://arxiv.org/abs/ 2602.10132. arXiv:2602.10132

    work page arXiv 2026

  17. [17]

    Yaguo Lei, Bin Yang, Xin Jiang, Feng Jia, Naipeng Li, and Asoke K. Nandi. Applications of machine learning to machine fault diagnosis: A review and roadmap.Mechanical Systems and Signal Processing, 138:106587, 2020. doi: 10.1016/j.ymssp.2019.106587. URL https: //doi.org/10.1016/j.ymssp.2019.106587

    work page doi:10.1016/j.ymssp.2019.106587 2020

  18. [18]

    Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen

    Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y. Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen. TIME-LLM: Time series fore- casting by reprogramming large language models. InInternational Conference on Learning Repre- sentations (ICLR), 2024

    work page 2024

  19. [19]

    Timesnet: Temporal2d-variation modeling for general time series analysis.Inter- national Conference on Learning Representations (ICLR), 2023

    HaixuWuetal. Timesnet: Temporal2d-variation modeling for general time series analysis.Inter- national Conference on Learning Representations (ICLR), 2023. URL https://ise.thss.tsing hua.edu.cn/~mlong/doc/TimesNet-iclr23. pdf

    work page 2023

  20. [20]

    Lade and Thilo Gross

    Steven J. Lade and Thilo Gross. Early warning signals for critical transitions: A generalized mod- eling approach.PLOS Computational Biology, 8 (2):e1002360, 2012. doi: 10.1371/journal.pcbi.1 002360. URL https://doi.org/10.1371/jour nal.pcbi.1002360

    work page doi:10.1371/journal.pcbi.1 2012

  21. [21]

    Early warning signals for critical transitions?DOE Computational Science Grad- uate Fellowship, 2012

    Carl Boettiger. Early warning signals for critical transitions?DOE Computational Science Grad- uate Fellowship, 2012. URLhttps://www.krel linst.org/csgf/conf/2012/abstracts/boe ttiger

    work page 2012

  22. [22]

    Brock, Victor Brovkin, Stephen R

    Marten Scheffer, Jordi Bascompte, William A. Brock, Victor Brovkin, Stephen R. Carpenter, Vasilis Dakos, Hermann Held, Egbert H. van Nes, Max Rietkerk, and George Sugihara. Early- warning signals for critical transitions.Nature, 461(7260):53–59, 2009. doi: 10.1038/nature0822 7

    work page doi:10.1038/nature0822 2009

  23. [23]

    Anticipating critical transitions.Science, 338 (6105):344–348, 2012

    Marten Scheffer, Stephen R Carpenter, Timo- thy M Lenton, Jordi Bascompte, William Brock, 9 Vasilis Dakos, Johan van de Koppel, Ingrid A van de Leemput, Simon A Levin, Egbert H van Nes, Mercedes Pascual, and John Vandermeer. Anticipating critical transitions.Science, 338 (6105):344–348, 2012. doi: 10.1126/science.1225 244

    work page doi:10.1126/science.1225 2012

  24. [24]

    A mathematical framework for critical transitions: bifurcations, fast–slow systems and stochastic dynamics.Physica D: Nonlinear Phenomena, 240(12):1020–1035, 2011

    Christian Kuehn. A mathematical framework for critical transitions: bifurcations, fast–slow systems and stochastic dynamics.Physica D: Nonlinear Phenomena, 240(12):1020–1035, 2011. doi: 10.1016/j.physd.2011.02.012

    work page doi:10.1016/j.physd.2011.02.012 2011

  25. [25]

    Tip- ping points: early warning and wishful thinking

    Peter D Ditlevsen and Sigfus J Johnsen. Tip- ping points: early warning and wishful thinking. Geophysical Research Letters, 37(19), 2010. doi: 10.1029/2010GL044486

    work page doi:10.1029/2010gl044486 2010

  26. [26]

    Methods for de- tecting early warnings of critical transitions in time series illustrated using simulated ecologi- cal data.PloS one, 7(7):e41010, 2012

    Vasilis Dakos, Stephen R Carpenter, Egbert H van Nes, and Marten Scheffer. Methods for de- tecting early warnings of critical transitions in time series illustrated using simulated ecologi- cal data.PloS one, 7(7):e41010, 2012. doi: 10.1371/journal.pone.0041010

    work page doi:10.1371/journal.pone.0041010 2012

  27. [27]

    I. I. Novikov and Yu. S. Trelin. Speed of sound along the vapor–liquid phase equilibrium curve. Prikl. Mekh. Tekh. Fiz., 1(2):112–115, 1960

    work page 1960

  28. [28]

    Slowing down as an early warning signal for abrupt climate change.Proceedings of the National Academy of Sciences, 105(38):14308– 14312, 2008

    Vasilis Dakos, Marten Scheffer, Egbert H van Nes, Victor Brovkin, Vladimir Petoukhov, and Hermann Held. Slowing down as an early warning signal for abrupt climate change.Proceedings of the National Academy of Sciences, 105(38):14308– 14312, 2008. doi: 10.1073/pnas.0802430105

    work page doi:10.1073/pnas.0802430105 2008

  29. [29]

    Timothy M Lenton. Early warning of climate tip- ping points from critical slowing down: compar- ing methods to improve robustness.Philosophical Transactions of the Royal Society A, 370(1962): 1185–1204, 2012. doi: 10.1098/rsta.2011.0304

    work page doi:10.1098/rsta.2011.0304 1962

  30. [30]

    Cook, Peter Acher- mann, and Dietmar Plenz

    Christian Meisel, Andreas Schulze-Bonhage, Dean Freestone, Mark J. Cook, Peter Acher- mann, and Dietmar Plenz. Intrinsic excitabil- ity measures track antiepileptic drug action and uncover increasing/decreasing excitability over the wake/sleep cycle.Proceedings of the National Academy of Sciences, 112(47):14694–14699, 2015. doi: 10.1073/pnas.1513716112

    work page doi:10.1073/pnas.1513716112 2015

  31. [31]

    C. K. Chow. On optimum recognition error and reject tradeoff.IEEE Transactions on Informa- tion Theory, 16(1):41–46, 1969

    work page 1969

  32. [32]

    On the founda- tions of noise-free selective classification.Jour- nal of Machine Learning Research, 11:1605–1641,

    Ran El-Yaniv and Yair Wiener. On the founda- tions of noise-free selective classification.Jour- nal of Machine Learning Research, 11:1605–1641,

    work page

  33. [33]

    URL https://jmlr.csail.mit.edu/pap ers/volume11/el-yaniv10a/el-yaniv10a.pd f

    work page

  34. [34]

    Selective classification for deep neural networks.Advances in Neural Information Processing Systems, 2017

    Yonatan Geifman and Ran El-Yaniv. Selective classification for deep neural networks.Advances in Neural Information Processing Systems, 2017. doi: /10.5555/3295222.3295241

    work page doi:10.5555/3295222.3295241 2017

  35. [35]

    A power system disturbance classifi- cation method robust to PMU data quality issues

    Zikang Li, Hao Liu, Junbo Zhao, Tianshu Bi, and Qixun Yang. A power system disturbance classifi- cation method robust to PMU data quality issues. IEEE Transactions on Industrial Informatics, 18 (1):97–108, 2022. doi: 10.1109/TII.2021.3072397

    work page doi:10.1109/tii.2021.3072397 2022

  36. [36]

    Hill, and Gregor Verbič

    Federico Milano, Florian Dörfler, Gabriela Hug, David J. Hill, and Gregor Verbič. Foundations and challenges of low-inertia systems. In2018 Power Systems Computation Conference (PSCC), pages 1–25. IEEE, 2018. doi: 10.23919/PSCC.20 18.8450880

    work page doi:10.23919/pscc.20 2018

  37. [37]

    Damage propagation modeling for aircraft engine run-to-failure simulation

    Abhinav Saxena, Kai Goebel, Don Simon, and Neil Eklund. Damage propagation modeling for aircraft engine run-to-failure simulation. Techni- cal report, NASA Ames Research Center, 2008

    work page 2008

  38. [38]

    Simple and scalable predictive uncertainty estimation using deep ensembles

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In Advances in Neural Information Processing Sys- tems, volume 30, pages 6405–6416, 2017

    work page 2017

  39. [39]

    Predict responsibly: Improv- ing fairness and accuracy by learning to defer

    David Madras et al. Predict responsibly: Improv- ing fairness and accuracy by learning to defer. Advances in Neural Information Processing Sys- tems, pages 6150–6160, 2018

    work page 2018

  40. [40]

    A survey on transfer learning.IEEE Transactions on Knowl- edge and Data Engineering, 22(10):1345–1359,

    Sinno Jialin Pan and Qiang Yang. A survey on transfer learning.IEEE Transactions on Knowl- edge and Data Engineering, 22(10):1345–1359,

    work page

  41. [41]

    doi: 10.1109/TKDE.2009.191

    work page doi:10.1109/tkde.2009.191 2009

  42. [42]

    Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges

    Michael M. Bronstein, Joan Bruna, Taco Co- hen, and Petar Veličković. Geometric deep learn- ing: Grids, groups, graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478, 2021. 10

    work page internal anchor Pith review arXiv 2021