Pith. sign in

REVIEW 2 major objections 5 minor 53 references

There is a Lipschitz constant that makes a network calibrated by construction, and a feedback rule finds it while keeping certified robustness.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

There exists a non-trivial Lipschitz constant L* that makes a network intrinsically calibrated; LiST finds it automatically via temperature feedback and produces calibrated robust models on CIFAR and Tiny-ImageNet.

T0 review reviewed 2026-07-10 challenge →

load-bearing objection Clean duality between Lipschitz L and temperature scaling, turned into a practical feedback loop that finds an intrinsically calibrated L* and a calibrated Pareto front; theory is existence-only under a continuity assumption the authors flag themselves. the 2 major comments →

arxiv 2607.07745 v1 pith:2DAIQ2TW submitted 2026-07-08 cs.LG stat.ML

LiST: Lipschitz Scaling Training for Robust and Calibrated Neural Networks

Arthur Chiron (IRIT , EPE UT) , Franck Mamalet , Thomas Massena (IRIT , DTIPG - SNCF , UT3) , Thomas Deltort (IRIT) , Mathieu Serrurier (IRIT
show 1 more author
UT2J)
This is my paper
classification cs.LG stat.ML
keywords Lipschitz networksTemperature Scalingcalibrationcertified robustnessaccuracy-robustness Pareto frontLiSTintrinsic calibration
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Reliable neural networks need accuracy, certified robustness, and calibrated confidence at once, but those goals are usually pursued separately. This paper shows that the global Lipschitz constant L and the temperature used in Temperature Scaling play interchangeable roles on the logits: for a fixed training scheme there exists a non-trivial L* at which the optimal temperature is exactly one, so the network is calibrated out of the box. That same L* sits at a natural operating point on the accuracy-robustness Pareto front. LiST turns the observation into a training loop that recomputes the calibration temperature each epoch and rescales L by its reciprocal until the temperature settles at one. A single offset in the loss then lets the user slide along a fully calibrated Pareto front, and once L is frozen the held-out calibration data can be folded back into training without harming calibration. On CIFAR-10/100 and Tiny-ImageNet the resulting models match or beat fixed-L and unconstrained baselines on accuracy and robustness while posting the lowest calibration error of any method tested.

Core claim

For a given architecture, data, and training procedure there exists a non-trivial Lipschitz constant L* such that a network trained under that hard constraint is intrinsically calibrated (its optimal Temperature Scaling factor is one). Calibration therefore supplies an unsupervised criterion that selects a well-defined operating point on the accuracy-robustness Pareto front; LiST finds that point automatically by the feedback rule Le+1 = Le / T*_e.

What carries the argument

Lipschitz-Temperature duality: dividing logits by a temperature T is equivalent to scaling the global Lipschitz constant by 1/T. LiST exploits the duality with a two-phase algorithm that updates L each epoch from the calibration temperature and later freezes L* so that calibration data can be re-integrated safely.

Load-bearing premise

The map from Lipschitz constant to the trained network must be continuous enough for an intermediate-value argument to guarantee that a calibrated L* exists; without that regularity the feedback loop has no theoretical target.

What would settle it

Train families of networks under a dense grid of fixed L values and check whether the optimal post-hoc temperature ever equals one; if the temperature-versus-L curve never crosses one, or if LiST's feedback rule fails to converge to a point where T*=1, the central claim is false.

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

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper argues that the global Lipschitz constant L of a constrained network and the temperature T of post-hoc Temperature Scaling are dual scaling factors on the logits. Under four regularity assumptions it proves (Prop. 3) that there exists a non-trivial L* for which the optimally trained network is intrinsically calibrated (T*=1). It then introduces LiST, a two-phase algorithm that uses the epoch-wise optimal temperature on a held-out set as a feedback signal (Le+1 = Le/T*_e) to drive the constraint to L*, freezes it, and re-integrates the calibration data. By varying a loss offset ξ the method also produces a fully calibrated accuracy-robustness Pareto front. Experiments on CIFAR-10/100 and Tiny-ImageNet with a Lipschitz ResNet-18 show competitive clean accuracy, AutoAttack accuracy and certified robust accuracy while reporting the lowest ECE among both unconstrained and fixed-L baselines, without post-hoc recalibration.

Significance. If the claimed duality and the practical convergence of the feedback loop hold, LiST supplies a principled, essentially unsupervised way to select an operating point on the accuracy-robustness Pareto front of Lipschitz networks while guaranteeing out-of-the-box calibration. The existence argument (Prop. 3), the certified-radius invariance under pure temperature scaling (Eq. 8), the multi-initialization trajectories (Fig. 2), the multi-seed CIFAR-10 results (Tables 1-2), and the public code constitute a coherent and reproducible contribution that unifies three reliability axes previously treated separately. The reduction of the hyper-parameter burden from (L, ξ) to ξ alone is of immediate practical value for certified-robustness practitioners.

major comments (2)
  1. Appendix B, Assumption 1 and Proposition 3: the Intermediate-Value argument establishing existence of L* rests on continuity of the trained-network family L ↦ f_ heta^(L) and uniform boundedness of logits on compact sets of L. The manuscript correctly flags this as a regularity property of the training procedure rather than a first-principles guarantee, and supports it empirically (Figs. 1, 2, 4). Because the central theoretical claim is existence plus practical convergence of the feedback rule, the paper should either (i) strengthen the discussion of when the assumption is expected to hold (e.g., under spectral-normalization + ScheduleFree) or (ii) clearly relegate uniqueness/continuity to an open question and present the algorithm as an empirically reliable heuristic whose fixed point is calibrated by construction of the feedback. The current wording slightly over-states the theoretica
  2. Section 5.1 / Table 1: LiST is trained on a 90/10 train/calibration split while all baselines (including fixed-L models) train on the full training set. Although Phase II re-integrates D_cal, the comparison of Phase-I metrics and the claim of “sample efficiency” would be cleaner if an additional fixed-L baseline trained on the same 90 % split (or an ablation that freezes L* without re-integration) were reported. The present design leaves open whether part of the ECE advantage is simply the effect of the held-out set rather than the dynamic constraint.
minor comments (5)
  1. Figure 1 caption and surrounding text: the quantitative mismatch between L*≈32 and 1/T*≈1.85 is explained by accuracy differences, but a short quantitative illustration (e.g., ECE of the 32-Lipschitz net after TS versus the 1-Lipschitz net after TS) would make the point sharper.
  2. Section 4.3: the stopping criterion (relative range of L over a window W=30, ε=0.001) is sensible, yet the sensitivity of final CRA/ECE to W and ε is only mentioned in Appendix E; a one-sentence statement in the main text would help reproducibility.
  3. Equation (4) and the subsequent absorption of τ into L: the notation switch from τ to L is clean, but an explicit remark that the offset ξ remains an independent free parameter (and is the only remaining one) would clarify the hyper-parameter claim made in the limitations section.
  4. Table 2 bottom block: CIFAR-100 and Tiny-ImageNet results are single-run; given the low variance reported on CIFAR-10 this is acceptable, but a brief note that multi-seed statistics were omitted for compute reasons would be transparent.
  5. Minor typographical issues: “a pillars such asexplainability” (p.1), missing spaces after some citations, and the arXiv identifier in the header still carries a future date.

Circularity Check

0 steps flagged

No significant circularity: L* is the fixed point of an independent temperature-scaling procedure; existence follows from a standard IVT argument under regularity assumptions, and the feedback rule is an intentional control loop, not a tautology.

full rationale

The paper defines an intrinsically calibrated network as one for which the optimal post-hoc temperature satisfies T*=1 (Section 2.2). Proposition 3 then proves existence of an L* realizing this property by applying the Intermediate Value Theorem to the continuous map L ↦ T*(L) (Lemmas 1–2), bracketed by the empirically observed under-confidence (small L) and over-confidence (large L) regimes (Assumptions 3–4). The continuity/boundedness Assumption 1 is a regularity hypothesis on the trained-network family, not a definition that forces the conclusion. The LiST update Le+1 ← Le/T*_e is deliberately constructed so that its equilibrium is T*=1; this is ordinary feedback design, not a self-definitional reduction of a claimed prediction. Empirical claims (Pareto location of L*, competitive multi-axis metrics, safe re-integration of Dcal) are validated against external baselines and held-out data, not forced by construction. Self-citations (Béthune et al. 2022, Serrurier et al. 2021, Boissin et al. 2025, etc.) supply architectural and loss background already standard in the Lipschitz-network literature; none is a load-bearing uniqueness theorem that closes the argument. Uniqueness of L* is explicitly left open (Remark 1). Consequently the derivation chain contains no step that reduces a claimed first-principles result to its own inputs by definition or by self-citation alone.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The central claim rests on four regularity assumptions that guarantee existence of L* via the Intermediate-Value Theorem, on the standard product bound for layer-wise Lipschitz constants, and on a handful of algorithmic free parameters (initial L0, window size, tolerance, offset ξ). No new physical entities are postulated; L* is an emergent fixed point of an existing calibration procedure.

free parameters (4)
  • initial Lipschitz constant L0 = 1.0 (default)
    Starting value of the dynamic constraint; paper shows convergence is robust across three orders of magnitude but still a free choice.
  • loss offset ξ = 0 or 3 (reported)
    Margin/offset parameter that moves the model along the calibrated Pareto front; user-chosen, not learned.
  • stopping window W and tolerance ε = W=30, ε=0.001
    Relative-range criterion that decides when L has stabilized; set to W=30, ε=0.001 without per-dataset tuning.
  • Phase-II fine-tuning epochs = 100
    Fixed budget of 100 epochs after freezing L*; not derived from theory.
axioms (5)
  • ad hoc to paper Assumption 1: the trained-network family L ↦ f_θ^(L) is continuous in L and logits are uniformly bounded on compact sets of L.
    Required for continuity of T*(L) and application of the Intermediate-Value Theorem; only empirically motivated (Fig. 2).
  • domain assumption Assumption 2: non-degeneracy of the calibration set (at least one misclassification, non-constant logits, positive average margin).
    Standard mild conditions ensuring an interior minimizer of the NLL in inverse temperature.
  • domain assumption Assumption 3: under-confidence for sufficiently small L (T*(L-)<1).
    Follows from logit magnitudes vanishing as L→0; empirically confirmed.
  • domain assumption Assumption 4: over-confidence for sufficiently large L (T*(L+)>1).
    Standard behavior of unconstrained cross-entropy networks; empirically confirmed.
  • standard math Global Lipschitz constant of a feed-forward net is upper-bounded by the product of layer-wise Lipschitz constants.
    Classical fact used throughout the Lipschitz-network literature and in Eq. (1).
invented entities (1)
  • intrinsically calibrated Lipschitz constant L* independent evidence
    purpose: The unique (conjectured) operating point at which T*=1 and the network is both calibrated and certifiably robust without post-hoc correction.
    Defined as the fixed point of the temperature-feedback map; existence proved under Assumptions 1-4, uniqueness left open.

reviewed 2026-07-10 · how reviews work

0 comments
Cite this review

Pith. "Pith review of LiST: Lipschitz Scaling Training for Robust and Calibrated Neural Networks." pith.science (2026). https://pith.science/paper/2DAIQ2TW

@misc{pith2026260707745,
  author       = {Pith},
  title        = {Pith review of: LiST: Lipschitz Scaling Training for Robust and Calibrated Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2DAIQ2TW}},
  note         = {Machine review of arXiv:2607.07745}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

While accuracy, robustness, and calibration are all essential for reliable neural networks, they are often studied separately; developing models that satisfy all three simultaneously remains a central challenge. Lipschitz-constrained models guarantee robustness by design, yet the manual selection of the Lipschitz constraint L governs the resulting accuracy-robustness trade-off, and their calibration properties remain largely underexplored. In this work, we highlight a theoretical and empirical link between the enforced Lipschitz constraint and Temperature Scaling, a state-of-the-art calibration method. Specifically, we find that for a given training scheme, there exists a non-trivial value L* that yields an out-of-the-box calibrated network, and that calibration acts as a principled criterion to select a well-defined operating point on the accuracy-robustness Pareto front. Leveraging these insights, we introduce Lipschitz Scaling Training (LiST), a novel training paradigm that iteratively adjusts the global Lipschitz constant to reach this operating point. Through a margin parameter in the training loss, LiST further enables the construction of a fully calibrated Pareto front, allowing users to navigate the accuracy-robustness trade-off while remaining calibrated throughout. At convergence, LiST also enables the reintegration of calibration data into training, improving sample efficiency without sacrificing calibration. We validate LiST on CIFAR-10/100 and Tiny-ImageNet, demonstrating competitive accuracy and robustness against constrained and unconstrained baselines, while remaining calibrated out of the box. Code is available at GitHub.

Figures

Figures reproduced from arXiv: 2607.07745 by Arthur Chiron (IRIT, DTIPG - SNCF, EPE UT), Franck Mamalet, Mathieu Serrurier (IRIT, Thomas Deltort (IRIT), Thomas Massena (IRIT, UT2J), UT3).

Figure 1
Figure 1. Figure 1: ECE as a function of the logit scaling mechanism. (a) Networks trained with varying fixed Lipschitz constants L. (b) Post-hoc Temperature Scaling on a fixed 1-Lipschitz network (plotted against 1/T). Both profiles are qualitatively similar. 3.2 On the Optimal Lipschitz Constant As observed in the experiment of [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Training dynamics of LiST (Phase I). (a) Evolution of the Lipschitz constraint L for four initializations L0. Regardless of L0, the algorithm first relaxes the constraint to gain expressivity (relaxation regime), then progressively tightens it as the model becomes over-confident (contraction regime), converging to the same intrinsic value L ∗ . (b) Corresponding evolution of the calibration temperature T. … view at source ↗
Figure 3
Figure 3. Figure 3: Pareto Analysis on CIFAR-10. (Left) LiST naturally converges to the calibrated point on the Pareto front. (Right) Varying ξ yields a fully calibrated Pareto front using 100% of training data. LiST is initialized with L0 = 1, which is justified by the convergence analysis of Section 4.2 showing that the algorithm reliably reaches L ∗ regardless of initialization. The Phase I stopping criterion uses a slidin… view at source ↗
Figure 4
Figure 4. Figure 4: ECE as a function of the logit scaling mechanism. (a) Networks trained with varying fixed Lipschitz constants L. (b) Post-hoc Temperature Scaling on a fixed 1-Lipschitz network (plotted against 1/T). Both profiles are qualitatively similar. 30 35 40 45 50 55 60 Clean Accuracy (%) 0 10 20 30 C RA ( = 3 6 / 2 5 5) (%) (a) Lipschitz Constant Ablation Fixed L LiST 53 54 55 56 57 Clean Accuracy (%) =0.0 =0.5 =1… view at source ↗
Figure 5
Figure 5. Figure 5: Pareto Analysis on CIFAR-100. (Left) LiST naturally converges to the calibrated point on the Pareto front. (Right) Varying ξ yields a fully calibrated Pareto front using 100% of training data. (Figure 4a) reproduces the qualitative pattern observed on CIFAR-10: a low constraint yields systematic underconfidence, while a high constraint leads to overconfidence, with a well-defined minimum in between. This c… view at source ↗
Figure 6
Figure 6. Figure 6: Reliability Diagrams of the fixed-L baselines on CIFAR. On both datasets, we observe the same trend of under-to-over confidence as L grows. 0.0 0.2 0.4 0.6 0.8 1.0 Confidence 0.0 0.2 0.4 0.6 0.8 1.0 Accuracy ECE: 1.24% LiST CIFAR10 (L*=34.7) 0.0 0.2 0.4 0.6 0.8 1.0 Confidence ECE: 1.66% LiST CIFAR100 (L*=25.0) 0.0 0.2 0.4 0.6 0.8 1.0 Confidence ECE: 2.10% LiST TinyImageNet (L*=14.7) [PITH_FULL_IMAGE:figur… view at source ↗
Figure 7
Figure 7. Figure 7: Reliability Diagrams of LiST for each dataset. LiST networks are intrinsically calibrated on CIFAR-10/100 and Tiny-ImageNet. Histogram Binning [Zadrozny and Elkan, 2001] and Isotonic Regression [Zadrozny and Elkan, 2002] are non-parametric alternatives, and Bayesian Binning into Quantiles [Naeini et al., 2015] extends them with a Bayesian treatment. While effective on i.i.d. test data, these methods leave … view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages · 6 internal anchors

  1. [1]

    Angelopoulos and Stephen Bates

    Anastasios N. Angelopoulos and Stephen Bates. Conformal Prediction : A Gentle Introduction . Foundations and Trends in Machine Learning, 16 0 (4): 0 494--591, March 2023. ISSN 1935-8237. doi:10.1561/2200000101. URL https://doi.org/10.1561/2200000101

    work page doi:10.1561/2200000101 2023

  2. [2]

    Sorting Out Lipschitz Function Approximation

    Cem Anil, James Lucas, and Roger Grosse. Sorting Out Lipschitz Function Approximation . In Proceedings of the 36th International Conference on Machine Learning , pages 291--301. PMLR, May 2019. URL https://proceedings.mlr.press/v97/anil19a.html

    work page 2019

  3. [3]

    A unified algebraic perspective on lipschitz neural networks

    Alexandre Araujo, Aaron J Havens, Blaise Delattre, Alexandre Allauzen, and Bin Hu. A unified algebraic perspective on lipschitz neural networks. In The Eleventh International Conference on Learning Representations, 2023

    work page 2023

  4. [4]

    Spectrally-normalized margin bounds for neural networks

    Peter L Bartlett, Dylan J Foster, and Matus J Telgarsky. Spectrally-normalized margin bounds for neural networks. In Advances in Neural Information Processing Systems , volume 30. Curran Associates, Inc., 2017. URL https://proceedings.neurips.cc/paper_files/paper/2017/hash/b22b257ad0519d4500539da3c8bcf4dd-Abstract.html

    work page 2017

  5. [5]

    An adaptive orthogonal convolution scheme for efficient and flexible cnn architectures

    Thibaut Boissin, Franck Mamalet, Thomas Fel, Agustin Martin Picard, Thomas Massena, and Mathieu Serrurier. An adaptive orthogonal convolution scheme for efficient and flexible cnn architectures. In International Conference on Machine Learning, pages 4757--4790. PMLR, 2025

    work page 2025

  6. [6]

    Boyd and Lieven Vandenberghe

    Stephen P. Boyd and Lieven Vandenberghe. Convex optimization. Cambridge University Press, Cambridge New York Melbourne New Delhi Singapore, version 29 edition, 2023. ISBN 978-0-521-83378-3

    work page 2023

  7. [7]

    Glenn W. Brier. VERIFICATION OF FORECASTS EXPRESSED IN TERMS OF PROBABILITY . Monthly Weather Review, 78 0 (1): 0 1--3, January 1950. ISSN 1520-0493, 0027-0644. doi:10.1175/1520-0493(1950)078<0001:VOFEIT>2.0.CO;2. URL https://journals.ametsoc.org/view/journals/mwre/78/1/1520-0493_1950_078_0001_vofeit_2_0_co_2.xml

    work page doi:10.1175/1520-0493(1950)078 1950

  8. [8]

    Pay attention to your loss : understanding misconceptions about Lipschitz neural networks

    Louis Béthune, Thibaut Boissin, Mathieu Serrurier, Franck Mamalet, Corentin Friedrich, and Alberto Gonzalez Sanz. Pay attention to your loss : understanding misconceptions about Lipschitz neural networks. Advances in Neural Information Processing Systems, 35: 0 20077--20091, December 2022. URL https://papers.nips.cc/paper_files/paper/2022/hash/7eb3d8ae592...

    work page 2022

  9. [9]

    Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks

    Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Proceedings of the 37th International Conference on Machine Learning , pages 2206--2216. PMLR, November 2020. URL https://proceedings.mlr.press/v119/croce20b.html

    work page 2020

  10. [10]

    The Road Less Scheduled

    Aaron Defazio, Xingyu Yang, Harsh Mehta, Konstantin Mishchenko, Ahmed Khaled, and Ashok Cutkosky. The Road Less Scheduled . In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural Information Processing Systems , volume 37, pages 9974--10007. Curran Associates, Inc., 2024. doi:10.52202/079017-0320....

    work page doi:10.52202/079017-0320 2024

  11. [11]

    Fairness through awareness

    Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. Fairness through awareness. In Proceedings of the 3rd Innovations in Theoretical Computer Science Conference , ITCS '12, pages 214--226, New York, NY, USA, January 2012. Association for Computing Machinery. ISBN 978-1-4503-1115-1. doi:10.1145/2090236.2090255. URL https://dl.ac...

    work page doi:10.1145/2090236.2090255 2012

  12. [12]

    Disrupting Deep Uncertainty Estimation Without Harming Accuracy

    Ido Galil and Ran El-Yaniv. Disrupting Deep Uncertainty Estimation Without Harming Accuracy . In Advances in Neural Information Processing Systems , volume 34, pages 21285--21296. Curran Associates, Inc., 2021. URL https://proceedings.neurips.cc/paper/2021/hash/b1b20d09041289e6c3fbb81850c5da54-Abstract.html

    work page 2021

  13. [13]

    Size-independent sample complexity of neural networks

    Noah Golowich, Alexander Rakhlin, and Ohad Shamir. Size-independent sample complexity of neural networks. In Conference on learning theory, pages 297--299. PMLR, 2018

    work page 2018

  14. [14]

    Explaining and Harnessing Adversarial Examples

    Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations (ICLR), 2015. URL https://arxiv.org/abs/1412.6572

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  15. [15]

    Robust Models are less Over-Confident

    Julia Grabinski, Paul Gavrikov, Janis Keuper, and Margret Keuper. Robust Models are less Over - Confident , December 2022. URL http://arxiv.org/abs/2210.05938. arXiv:2210.05938 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  16. [16]

    The Limitations of Model Uncertainty in Adversarial Settings

    Kathrin Grosse, David Pfaff, Michael Thomas Smith, and Michael Backes. The Limitations of Model Uncertainty in Adversarial Settings , November 2019. URL http://arxiv.org/abs/1812.02606. arXiv:1812.02606 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  17. [17]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On Calibration of Modern Neural Networks . In Proceedings of the 34th International Conference on Machine Learning , pages 1321--1330. PMLR, July 2017. URL https://proceedings.mlr.press/v70/guo17a.html. ISSN: 2640-3498

    work page 2017

  18. [18]

    Equality of Opportunity in Supervised Learning

    Moritz Hardt, Eric Price, and Nati Srebro. Equality of Opportunity in Supervised Learning . In Advances in Neural Information Processing Systems , volume 29. Curran Associates, Inc., 2016. URL https://proceedings.neurips.cc/paper_files/paper/2016/hash/6a9659feb1216f14f7384ba499518b38-Abstract.html

    work page 2016

  19. [19]

    Improving model calibration with accuracy versus uncertainty optimization

    Ranganath Krishnan and Omesh Tickoo. Improving model calibration with accuracy versus uncertainty optimization. In Advances in Neural Information Processing Systems , volume 33, pages 18237--18248. Curran Associates, Inc., 2020. URL https://proceedings.neurips.cc/paper_files/paper/2020/hash/d3d9446802a44259755d38e6d163e820-Abstract.html

    work page 2020

  20. [20]

    Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with Dirichlet calibration

    Meelis Kull, Miquel Perello Nieto, Markus Kängsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with Dirichlet calibration. In Advances in Neural Information Processing Systems , volume 32. Curran Associates, Inc., 2019. URL https://proceedings.neurips.cc/paper/2019/hash/8ca...

    work page 2019

  21. [21]

    Trainable Calibration Measures for Neural Networks from Kernel Mean Embeddings

    Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable Calibration Measures for Neural Networks from Kernel Mean Embeddings . In Proceedings of the 35th International Conference on Machine Learning , pages 2805--2814. PMLR, July 2018. URL https://proceedings.mlr.press/v80/kumar18a.html

    work page 2018

  22. [22]

    Enhancing certified robustness via block reflector orthogonal layers and logit annealing loss

    Bo-Han Lai, Pin-Han Huang, Bo-Han Kung, and Shang-Tse Chen. Enhancing certified robustness via block reflector orthogonal layers and logit annealing loss. In International Conference on Machine Learning (ICML), 2025. Spotlight

    work page 2025

  23. [23]

    On the Robustness of Adversarial Training Against Uncertainty Attacks

    Emanuele Ledda, Giovanni Scodeller, Daniele Angioni, Giorgio Piras, Antonio Emanuele Cinà, Giorgio Fumera, Battista Biggio, and Fabio Roli. On the Robustness of Adversarial Training Against Uncertainty Attacks , May 2025. URL http://arxiv.org/abs/2410.21952. arXiv:2410.21952 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    Preventing Gradient Attenuation in Lipschitz Constrained Convolutional Networks

    Qiyang Li, Saminul Haque, Cem Anil, James Lucas, Roger B Grosse, and Joern-Henrik Jacobsen. Preventing Gradient Attenuation in Lipschitz Constrained Convolutional Networks . In Advances in Neural Information Processing Systems , volume 32. Curran Associates, Inc., 2019. URL https://proceedings.neurips.cc/paper_files/paper/2019/hash/1ce3e6e3f452828e23a0c94...

    work page 2019

  25. [25]

    The devil is in the margin: Margin-based label smoothing for network calibration

    Bingyuan Liu, Ismail Ben Ayed, Adrian Galdran, and Jose Dolz. The devil is in the margin: Margin-based label smoothing for network calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 80--88, 2022

    work page 2022

  26. [26]

    A Unified Approach to Interpreting Model Predictions

    Scott M Lundberg and Su-In Lee. A Unified Approach to Interpreting Model Predictions . In Advances in Neural Information Processing Systems , volume 30. Curran Associates, Inc., 2017. URL https://proceedings.neurips.cc/paper_files/paper/2017/hash/8a20a8621978632d76c43dfd28b67767-Abstract.html

    work page 2017

  27. [27]

    Towards Deep Learning Models Resistant to Adversarial Attacks

    Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards Deep Learning Models Resistant to Adversarial Attacks . In International Conference on Learning Representations, February 2018. URL https://openreview.net/forum?id=rJzIBfZAb

    work page 2018

  28. [28]

    Delattre, Alexandre Araujo, and Alexandre Allauzen

    Laurent Meunier, Blaise J. Delattre, Alexandre Araujo, and Alexandre Allauzen. A Dynamical System Perspective for Lipschitz Neural Networks . In Proceedings of the 39th International Conference on Machine Learning , pages 15484--15500. PMLR, June 2022. URL https://proceedings.mlr.press/v162/meunier22a.html

    work page 2022

  29. [29]

    Spectral Normalization for Generative Adversarial Networks

    Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral Normalization for Generative Adversarial Networks . In International Conference on Learning Representations, February 2018. URL https://openreview.net/forum?id=B1QRgziT-&source=post_page---------------------------

    work page 2018

  30. [30]

    Calibrating Deep Neural Networks using Focal Loss

    Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating Deep Neural Networks using Focal Loss . In Advances in Neural Information Processing Systems , volume 33, pages 15288--15299. Curran Associates, Inc., 2020. URL https://proceedings.neurips.cc/paper_files/paper/2020/hash/aeb7b30ef1d024a76f21a1d40e...

    work page 2020

  31. [31]

    Murphy and Robert L

    Allan H. Murphy and Robert L. Winkler. Reliability of Subjective Probability Forecasts of Precipitation and Temperature . Journal of the Royal Statistical Society. Series C (Applied Statistics), 26 0 (1): 0 41--47, 1977. ISSN 0035-9254. doi:10.2307/2346866. URL https://www.jstor.org/stable/2346866. Publisher: [Royal Statistical Society, Oxford University Press]

    work page doi:10.2307/2346866 1977

  32. [32]

    Rafael Müller, Simon Kornblith, and Geoffrey E. Hinton. When does label smoothing help? Advances in Neural Information Processing Systems, 32, 2019. URL https://proceedings.neurips.cc/paper_files/paper/2019/hash/f1748d6b0fd9d439f71450117eba2725-Abstract.html?ref=gojiberries.io

    work page 2019

  33. [33]

    Obtaining Well Calibrated Probabilities Using Bayesian Binning

    Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining Well Calibrated Probabilities Using Bayesian Binning . Proceedings of the AAAI Conference on Artificial Intelligence, 29 0 (1), February 2015. ISSN 2374-3468. doi:10.1609/aaai.v29i1.9602. URL https://ojs.aaai.org/index.php/AAAI/article/view/9602. Number: 1

    work page doi:10.1609/aaai.v29i1.9602 2015

  34. [34]

    A pac-bayesian approach to spectrally-normalized margin bounds for neural networks

    Behnam Neyshabur, Srinadh Bhojanapalli, and Nathan Srebro. A pac-bayesian approach to spectrally-normalized margin bounds for neural networks. In International Conference on Learning Representations, 2018

    work page 2018

  35. [35]

    Calibration Attacks : A Comprehensive Study of Adversarial Attacks on Model Confidence

    Stephen Obadinma, Xiaodan Zhu, and Hongyu Guo. Calibration Attacks : A Comprehensive Study of Adversarial Attacks on Model Confidence . Transactions on Machine Learning Research, May 2024. ISSN 2835-8856. URL https://openreview.net/forum?id=TXzz9xwdpv

    work page 2024

  36. [36]

    Almost-orthogonal layers for efficient general-purpose lipschitz networks

    Bernd Prach and Christoph H Lampert. Almost-orthogonal layers for efficient general-purpose lipschitz networks. In European Conference on Computer Vision, pages 350--365. Springer, 2022

    work page 2022

  37. [37]

    Intriguing properties of robust classification

    Bernd Prach and Christoph H Lampert. Intriguing properties of robust classification. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 660--669, 2025

    work page 2025

  38. [38]

    Improving Calibration through the Relationship with Adversarial Robustness

    Yao Qin, Xuezhi Wang, Alex Beutel, and Ed Chi. Improving Calibration through the Relationship with Adversarial Robustness . In Advances in Neural Information Processing Systems , volume 34, pages 14358--14369. Curran Associates, Inc., 2021. URL https://proceedings.neurips.cc/paper_files/paper/2021/hash/78421a2e0e1168e5cd1b7a8d23773ce6-Abstract.html

    work page 2021

  39. [39]

    Why Should I Trust You ?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. " Why Should I Trust You ?": Explaining the Predictions of Any Classifier . In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages 1135--1144, San Francisco California USA, August 2016. ACM. ISBN 978-1-4503-4232-2. doi:10.1145/2939672.2939778. UR...

    work page doi:10.1145/2939672.2939778 2016

  40. [40]

    Achieving Robustness in Classification Using Optimal Transport With Hinge Regularization

    Mathieu Serrurier, Franck Mamalet, Alberto Gonzalez-Sanz, Thibaut Boissin, Jean-Michel Loubes, and Eustasio del Barrio. Achieving Robustness in Classification Using Optimal Transport With Hinge Regularization . In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 505--514, 2021. URL https://openaccess.thecvf.com/content/CVPR2021...

    work page 2021

  41. [41]

    Skew Orthogonal Convolutions

    Sahil Singla and Soheil Feizi. Skew Orthogonal Convolutions . In Proceedings of the 38th International Conference on Machine Learning , pages 9756--9766. PMLR, July 2021. URL https://proceedings.mlr.press/v139/singla21a.html

    work page 2021

  42. [42]

    Intriguing properties of neural networks

    Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations (ICLR), 2014. URL https://arxiv.org/abs/1312.6199

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  43. [43]

    Dual Focal Loss for Calibration

    Linwei Tao, Minjing Dong, and Chang Xu. Dual Focal Loss for Calibration . In Proceedings of the 40th International Conference on Machine Learning , pages 33833--33849. PMLR, July 2023. URL https://proceedings.mlr.press/v202/tao23a.html

    work page 2023

  44. [44]

    Orthogonalizing convolutional layers with the cayley transform

    Asher Trockman and J Zico Kolter. Orthogonalizing convolutional layers with the cayley transform. In International Conference on Learning Representations, 2021

    work page 2021

  45. [45]

    Lipschitz- Margin Training : Scalable Certification of Perturbation Invariance for Deep Neural Networks

    Yusuke Tsuzuku, Issei Sato, and Masashi Sugiyama. Lipschitz- Margin Training : Scalable Certification of Perturbation Invariance for Deep Neural Networks . In Advances in Neural Information Processing Systems , volume 31. Curran Associates, Inc., 2018. URL https://proceedings.neurips.cc/paper/2018/hash/485843481a7edacbfce101ecb1e4d2a8-Abstract.html

    work page arXiv 2018

  46. [46]

    Calibration of Neural Networks

    Ruslan Vasilev and Alexander D'yakonov. Calibration of Neural Networks , March 2023. URL http://arxiv.org/abs/2303.10761. arXiv:2303.10761

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  47. [47]

    Sauer, Tom Hendriks, Olivier W

    Jarne Verhaeghe, Thomas De Corte, Christopher M. Sauer, Tom Hendriks, Olivier W. M. Thijssens, Femke Ongenae, Paul Elbers, Jan De Waele, and Sofie Van Hoecke. Generalizable calibrated machine learning models for real-time atrial fibrillation risk prediction in ICU patients. International Journal of Medical Informatics, 175: 0 105086, July 2023. ISSN 1386-...

    work page doi:10.1016/j.ijmedinf.2023.105086 2023

  48. [48]

    Lipschitz regularity of deep neural networks: analysis and efficient estimation

    Aladin Virmaux and Kevin Scaman. Lipschitz regularity of deep neural networks: analysis and efficient estimation. In Advances in Neural Information Processing Systems , volume 31. Curran Associates, Inc., 2018. URL https://proceedings.neurips.cc/paper_files/paper/2018/hash/d54e99a6c03704e95e6965532dec148b-Abstract.html

    work page 2018

  49. [49]

    Algorithmic

    Vladimir Vovk, Alexander Gammerman, and Glenn Shafer. Algorithmic Learning in a Random World . Springer International Publishing, Cham, 2022. ISBN 978-3-031-06648-1 978-3-031-06649-8. doi:10.1007/978-3-031-06649-8. URL https://link.springer.com/10.1007/978-3-031-06649-8

    work page doi:10.1007/978-3-031-06649-8 2022

  50. [50]

    LOT : Layer -wise Orthogonal Training on Improving l2 Certified Robustness

    Xiaojun Xu, Linyi Li, and Bo Li. LOT : Layer -wise Orthogonal Training on Improving l2 Certified Robustness . Advances in Neural Information Processing Systems, 35: 0 18904--18915, December 2022. URL https://proceedings.neurips.cc/paper_files/paper/2022/hash/77d52754ff6b2de5a5d96ee921b6b3cd-Abstract-Conference.html

    work page 2022

  51. [51]

    Mitigating Transformer Overconfidence via Lipschitz Regularization

    Wenqian Ye, Yunsheng Ma, Xu Cao, and Kun Tang. Mitigating Transformer Overconfidence via Lipschitz Regularization . In Proceedings of the Thirty - Ninth Conference on Uncertainty in Artificial Intelligence , pages 2422--2432. PMLR, July 2023. URL https://proceedings.mlr.press/v216/ye23a.html

    work page 2023

  52. [52]

    Obtaining calibrated probability estimates from decision trees and naive Bayesian classifiers

    Bianca Zadrozny and Charles Elkan. Obtaining calibrated probability estimates from decision trees and naive Bayesian classifiers. International Conference on Machine Learning (ICML), 2001

    work page 2001

  53. [53]

    Transforming Classifier Scores into Accurate Multiclass Probability Estimates

    Bianca Zadrozny and Charles Elkan. Transforming Classifier Scores into Accurate Multiclass Probability Estimates . Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, August 2002. doi:10.1145/775047.775151

    work page doi:10.1145/775047.775151 2002

This paper was first reviewed by grok-4.5 on July 10, 2026.