Pith. sign in

REVIEW 2 major objections 5 minor 43 references

Principled Input-Output-Conditioned Post-Hoc Uncertainty Estimation for Regression Networks

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Frozen regression networks can be retrofitted with calibrated uncertainty by training a small auxiliary network on (x, f(x)) with a detached Gaussian NLL objective; the frozen output itself carries a recoverable quasi-epistemic signal.

desk verdict A practical post-hoc uncertainty method with strong NYU results, but the theory section overclaims: Corollary 1 is asserted, not proved, and the quasi-epistemic decomposition doesn't follow from the NLL objective. read the letter →

arxiv 2506.00918 v1 pith:QOTCRK7N submitted 2025-06-01 cs.LG cs.AI

Lennart Bramlage , Cristóbal Curio This is my paper
classification cs.LGcs.AI
keywords post-hocuncertaintyquantificationfrozenregressionnetworksheteroscedasticaleatoricepistemicout-of-distributiondetectiondetachedGaussianNLLoutputmanifolddistancedepthestimation
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

Most regression networks ship without uncertainty estimates, and retrofitting them usually means touching the weights, gradients, or running many forward passes. This paper argues that a frozen regression model can be upgraded post-hoc by training a small auxiliary network on the pair $(x, f(x))$ with a detached Gaussian negative-log-likelihood objective on a probe dataset. The central claim is that this objective recovers the canonical maximum-likelihood variance estimate, and that the frozen output $f(x)$ encodes a distance-to-training-manifold signal that behaves like a relative epistemic uncertainty. If this is right, any existing regression model can gain calibrated heteroscedastic variance in-distribution and a principled out-of-distribution detector, with one extra forward pass and no access to model internals.

What carries the argument

The load-bearing mechanism is the detached Gaussian NLL objective of Eq. (3), which applies a stop-gradient operator to the frozen predictions $f(x_i)$ and optimizes $g_\phi$ on the squared residual divided by the predicted variance plus the log-variance. Three components do the work: conditioning on $x$ recovers input-dependent aleatoric noise $\sigma(x)$; conditioning on $f(x)$ exposes the output-manifold distance signal; and the theoretical decomposition of Proposition 1 and Corollary 1, $g_a(x)+g_e(f)$ with $g_e(f)=\lambda d(f,\mathcal{M})$, converts that signal into a variance inflation. The manifold distance $d(f,\mathcal{M})$ is never computed in practice; the claim is that the frozen network's outputs encode it implicitly, and that the convex NLL objective in $g_\phi$ lets AdamW retrieve it.

What would settle it

Train IO-CUE on a synthetic regression problem where $\sigma(x)$ is known analytically and the base model is deliberately biased on a held-out region, then compare the learned variance surface pointwise to $\sigma(x)+d(f(x),\mathcal{M})$. If the residual after subtracting $\sigma(x)$ is not proportional to a distance-to-manifold function with slope one, or is not a function of $f(x)$ alone, Corollary 1 collapses to a heuristic.

Watch

Extended reading notes

Core claim

Under an additive Gaussian noise model $y=\mu(x)+\epsilon$, $\epsilon\sim\mathcal{N}(0,\sigma(x))$, with a frozen mean estimator $f(x)$ trained by MSE, the paper fits an auxiliary variance head $g_\phi(x,f(x))$ by minimizing the detached Gaussian NLL of Eq. (3). Because the mean is already at its MLE, the variance can be fit sequentially. Proposition 1 states that the optimal $g$ decomposes as $g_a(x)+g_e(f)$, where $g_a(x)=\sigma(x)$ is the true aleatoric variance and $g_e(f)=\lambda\, d(f,\mathcal{M})$ grows with the distance of the prediction from $\mathcal{M}=f(\mathcal{S})$, the image of the training support under $f$. Corollary 1 concludes that on regions the post-hoc learner has seen, $\mathrm{Var}(y\mid x)\approx\sigma(x)+\lambda d(f(x),\mathcal{M})$ with $\lambda=1$, so off-manifold outputs receive inflated variance and act as quasi-epistemic uncertainty scores. The paper reports supporting evidence on standard regression benchmarks and depth estimation, including better NLL, ECE, and error–uncertainty correlation than a Gaussian ensemble baseline, and improved OOD detection when the probe set is augmented.

Load-bearing premise

The load-bearing premise is that minimizing the detached NLL makes the learned variance decompose exactly into the true input-dependent noise plus the frozen output's distance from the training manifold, with the distance coefficient fixed at one; if the learned auxiliary network merely correlates with error on augmented data, the epistemic-recovery claim is only a heuristic.

Editorial extensions

If this is right

  • An MSE-trained regression model can be given calibrated heteroscedastic uncertainties after the fact, using a probe set and no access to its parameters, gradients, or internal representations.
  • Because the base model stays frozen, uncertainty retrofitting does not degrade predictive performance, avoiding the instability of joint mean-variance training.
  • Outputs that fall off the training manifold receive inflated variance, so the same estimator doubles as an out-of-distribution detector and a relative epistemic-risk ranker.
  • Adding diverse transformations to the probe set markedly improves OOD detection and cross-network generalization, making probe-set diversity a practical lever for model characterization.
  • On depth regression, the post-hoc estimator outperforms a five-member Gaussian ensemble on NLL, expected calibration error, and error–uncertainty correlation, suggesting sampling-free post-hoc UQ can match sampling-based baselines.

Reading between the lines

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

  • Beyond the paper: if the decomposition $\mathrm{Var}(y\mid x)\approx\sigma(x)+d(f(x),\mathcal{M})$ holds quantitatively, the recipe should transfer to any frozen model with a low-dimensional output manifold, such as classifiers whose outputs live on a simplex; the paper only claims regression.
  • Beyond the paper: the augmentation result suggests a general principle for acquiring epistemic signal without weights: probe-set transforms that expand the set of observed model behaviours act as a cheap surrogate for ensembling, and adversarial or style-based transforms may push OOD detection further.
  • Beyond the paper: Corollary 1's coefficient $\lambda=1$ is asserted rather than proven; a direct test on synthetic data with known $\sigma(x)$ and known manifold distance would decide whether the learned variance equals $\sigma(x)+d(f,\mathcal{M})$ pointwise or only ranks failures correctly.
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

2 major / 5 minor

Summary. The paper proposes IO-CUE, a post-hoc uncertainty estimator that fits an auxiliary model g_phi(x, f(x)) to a frozen regression network f using a detached Gaussian negative log-likelihood on a probe dataset. The authors argue that this objective recovers the canonical MLE of the Gaussian variance given the frozen mean, that conditioning on x captures aleatoric uncertainty, and that conditioning on f(x) can capture quasi-epistemic uncertainty via distance of the output to the training-output manifold. The empirical sections report favorable comparisons against BayesCap and ensembles on UCI benchmarks and NYU Depth v2, including OOD detection experiments with augmented probe data. Code is provided, and the experimental protocol is documented in the appendices.

Significance. If the central theoretical claims were fully established, the contribution would be significant: IO-CUE offers a black-box, single-forward-pass retrofit of uncertainty to frozen regression models, and the reported NYU Depth v2 results are strong, with IO-CUE surpassing a five-member ensemble baseline on NLL, ECE, and error-uncertainty correlation. The paper also has concrete strengths: it ships a code repository, reports standard deviations over multiple runs, and designs falsifiable counterfactual experiments (e.g., g(x_sigma, f(x)) versus g(x_sigma, f(x_sigma))). However, the key corollary connecting f(x) to a distance-to-manifold epistemic term is asserted rather than proved, so the theoretical contribution is currently weaker than the paper claims.

major comments (2)
  1. [Section 3.3, Eq. (3), Corollary 1] Corollary 1 does not follow from Eq. (3) and is not proved. For a frozen f and a sufficiently expressive g, the population minimizer of Eq. (3) at input x is g*(x, f(x)) = (f(x) - mu(x))^2 + sigma^2(x); that is, the conditional squared error, not sigma(x) + lambda d(f(x), M). The further identification of the squared error with a distance-to-manifold term at lambda = 1 is a separate functional-equation assumption, and Proposition 1 only asserts the existence of such functions without connecting them to the minimizer. Since the 'quasi-epistemic inference' and 'exact recovery' claims rest on this step, the theoretical core of the paper is currently a heuristic with strong empirical support rather than an established result.
  2. [Section 3.3, Proposition 1 / Corollary 1] The additive split g_a + g_e is not identifiable from the scalar output of Eq. (3), and the value lambda = 1 is unjustified. The objective identifies only the sum (f(x) - mu(x))^2 + sigma^2(x); any monotonic rescaling of the distance d can be absorbed into lambda, and the paper never defines d, estimates M, or specifies how lambda would be determined. Concretely, a base model that extrapolates linearly outside the training support can have d(f(x), M) large while its squared error is zero, and a saturating model can have the opposite. The manuscript should either prove the decomposition under explicit assumptions or explicitly relegate Corollary 1 to an empirical hypothesis supported by the counterfactual experiments in Section 5.2.
minor comments (5)
  1. [Section 3.3, Eq. (3)] Equation (3) is not normalized consistently: the first term is averaged over N as 1/(2N) sum, but the log g term is written as (1/2) log g without a corresponding 1/N factor. This is likely a typo, but it changes the objective if taken literally.
  2. [Section 3.2 and 3.3] The notation for the noise scale is inconsistent: Section 3.2 defines sigma^2(x) = Var(y|x) and writes y ~ N(mu(x), sigma^2(x)), while the formal setup in Section 3.3 writes epsilon ~ N(0, sigma(x)) and Proposition 1 uses g_a(x) = sigma(x). Please use either sigma or sigma^2 consistently throughout.
  3. [Section 3.3] The statement that the objective 'is convex in g_phi' is not correct for the direct parameterization of the variance: the function A/g + log g is not globally convex in g over the positive reals. It is convex in the log-variance parameterization s = log g, so the claim should be stated in terms of that parameterization or qualified accordingly.
  4. [Section 4.3] The 'IO-CUE (GG)' variant is mentioned as a Generalized Gaussian-based estimator, but the Appendix does not describe the generalized Gaussian objective or its parameters. Please add the corresponding derivation and implementation details.
  5. [Throughout] There are several typographical errors that should be corrected in revision, including 'enfore' (Section 4.3), 'non-existant' (Section 4.3), 'Constist' (Appendix B), 'flios' (Section 5.1), and 'predictvie' (Section 5.2).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Eq. (3) is a standard sequential NLL fit, and the epistemic-recovery corollary is an unproven assertion rather than a reduction to the paper's own inputs.

full rationale

The paper's core derivation is the detached Gaussian NLL objective (Eq. 3), whose population minimizer over a sufficiently expressive g is the conditional expected squared error E[(f(X)-Y)^2 | X=x]. Saying that this recovers the canonical Gaussian MLE of the variance under a frozen mean is a standard property of the NLL as a proper scoring rule, not a restatement of the paper's conclusion; the auxiliary model is fitted on a probe set and evaluated on held-out ID test data. The quasi-epistemic claim in Corollary 1, that the minimizer decomposes as sigma(x) + lambda d(f(x), M) with lambda=1, is not derived from Eq. (3): it requires an additional, unverified functional-equation assumption identifying model bias with distance to the frozen output manifold, and the scalar output of Eq. (3) cannot identify the additive split or fix lambda. That is a correctness and rigor gap, not circularity. Empirical OOD claims are tested against held-out shifted datasets (flipped NYU and ApolloScape), and the augmentations used during post-hoc training are explicitly disjoint from the test shift, so the reported predictions are not forced by the fitted values. The only self-citation (Bramlage et al. 2023) appears in the related-work discussion of higher-order evidential objectives and is not load-bearing for the proposed derivation or its evaluation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the Gaussian assumption, the MSE-trained base model assumption, and an unproven convergence condition for the epistemic decomposition. The λ=1 assertion and the unspecified distance metric are free parameters in the theory.

free parameters (3)
  • lambda (epistemic scale) = 1 (claimed, not fitted)
    Corollary 1 asserts λ=1 when minimizing (3) on diverse data; the NLL objective alone does not determine this scale, and no fitting is described.
  • output-space distance metric d
    Proposition 1 posits a distance d between f(x) and the manifold M, but the metric is unspecified and the network must learn it implicitly.
  • probe dataset fraction and augmentation choices = 10% of training data; Gaussian blur and ColorJitter
    Empirical results in Section 5.1 depend on these hand-chosen settings; they are not derived from the theory.
assumptions (4)
  • domain assumption Additive Gaussian noise: y = µ(x) + ε, ε ~ N(0, σ(x))
    Stated in Section 3.2 and Limitations; underlies the NLL objective and the interpretation of g as variance.
  • domain assumption Base model f is MSE-trained, so it is the MLE of the mean
    Stated in Limitations: only then does the detached NLL yield the canonical variance MLE.
  • ad hoc to paper Sufficiently diverse ID and OOD probe data and 'reasonable assumptions of convergence'
    Corollary 1 depends on unspecified diversity and convergence conditions; without them the decomposition may fail.
  • standard math g is a universal function regressor
    Proposition 1's existence relies on g being expressive enough to represent arbitrary functions of x and f(x).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Principled Input-Output-Conditioned Post-Hoc Uncertainty Estimation for Regression Networks." pith.science (2026). https://pith.science/paper/QOTCRK7N

@misc{pith2026250600918,
  author       = {Pith},
  title        = {Pith review of: Principled Input-Output-Conditioned Post-Hoc Uncertainty Estimation for Regression Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QOTCRK7N}},
  note         = {Machine review of arXiv:2506.00918}
}
read the original abstract

Uncertainty quantification is critical in safety-sensitive applications but is often omitted from off-the-shelf neural networks due to adverse effects on predictive performance. Retrofitting uncertainty estimates post-hoc typically requires access to model parameters or gradients, limiting feasibility in practice. We propose a theoretically grounded framework for post-hoc uncertainty estimation in regression tasks by fitting an auxiliary model to both original inputs and frozen model outputs. Drawing from principles of maximum likelihood estimation and sequential parameter fitting, we formalize an exact post-hoc optimization objective that recovers the canonical MLE of Gaussian parameters, without requiring sampling or approximation at inference. While prior work has used model outputs to estimate uncertainty, we explicitly characterize the conditions under which this is valid and demonstrate the extent to which structured outputs can support quasi-epistemic inference. We find that using diverse auxiliary data, such as augmented subsets of the original training data, significantly enhances OOD detection and metric performance. Our hypothesis that frozen model outputs contain generalizable latent information about model error and predictive uncertainty is tested and confirmed. Finally, we ensure that our method maintains proper estimation of input-dependent uncertainty without relying exclusively on base model forecasts. These findings are demonstrated in toy problems and adapted to both UCI and depth regression benchmarks. Code: https://github.com/biggzlar/IO-CUE.

Figures

Figures reproduced from arXiv: 2506.00918 by the authors.

Figure 1
Figure 1. Comparison of (left) a model trained with two outputs and Gaussian NLL loss for 200 epochs and (right) a model trained with a single output and MSE loss, post-hoc uncertainty estimation via a probe dataset with the same number of epochs in total. parameters θ and predictions µθ(x) = ˆy = f(x; θ) without the capability to infer uncertainty. Assuming a frozen base model and no access to intermediate representations, w… view at source ↗
Figure 2
Figure 2. Post-hoc uncertainty estimation performance ( [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison for depth estimation on the NYU Depth v2 test set. Results are [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: OOD Detection Results before (top) and after (bottom) data augmentation for a flipped version of the NYU Depth v2 dataset and the previously unseen Apolloscape dataset. Augmentation improves model characterization by post-hoc uncertainty quantifiers dramatically. (ROC …
Figure 5
Figure 5. Figure 5: AUROC scores for multiple base models on differently augmented datasets in a counter [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Qualitative examples of the effect of input perturbations on the uncertainty estimates of our [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Results of increasing input perturbations. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Impact of model size on IO-CUE performance. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Impact of probe dataset size on IO-CUE performance. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 31 canonical work pages

  1. [1]

    Deep ensembles work, but are they necessary? Advances in Neural Information Processing Systems, 35: 0 33646--33660, 2022

    Taiga Abe, Estefany Kelly Buchanan, Geoff Pleiss, Richard Zemel, and John P Cunningham. Deep ensembles work, but are they necessary? Advances in Neural Information Processing Systems, 35: 0 33646--33660, 2022

    work page 2022

  2. [2]

    Deep evidential regression

    Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. Deep evidential regression. Advances in Neural Information Processing Systems, 33: 0 14927--14937, 2020

    work page 2020

  3. [3]

    Test-time data augmentation for estimation of heteroscedastic aleatoric uncertainty in deep neural networks

    Murat Seckin Ayhan and Philipp Berens. Test-time data augmentation for estimation of heteroscedastic aleatoric uncertainty in deep neural networks. In Medical Imaging with Deep Learning, 2018

    work page 2018

  4. [4]

    The need for uncertainty quantification in machine-assisted medical decision making

    Edmon Begoli, Tanmoy Bhattacharya, and Dimitri Kusnezov. The need for uncertainty quantification in machine-assisted medical decision making. Nature Machine Intelligence, 1 0 (1): 0 20--23, 2019

    work page 2019

  5. [5]

    Pitfalls of epistemic uncertainty quantification through loss minimisation

    Viktor Bengs, Eyke H \"u llermeier, and Willem Waegeman. Pitfalls of epistemic uncertainty quantification through loss minimisation. Advances in Neural Information Processing Systems, 35: 0 29205--29216, 2022

    work page 2022

  6. [6]

    A conceptual introduction to hamiltonian monte carlo

    Michael Betancourt. A conceptual introduction to hamiltonian monte carlo. arXiv preprint arXiv:1701.02434, 2017

    arXiv 2017

  7. [7]

    Weight uncertainty in neural network

    Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International conference on machine learning, pages 1613--1622. PMLR, 2015

    work page 2015

  8. [8]

    Plausible uncertainties for human pose regression

    Lennart Bramlage, Michelle Karg, and Crist \'o bal Curio. Plausible uncertainties for human pose regression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15133--15142, 2023

    work page 2023

Show all 43 references
  1. [9]

    Heteroscedastic kernel ridge regression

    Gavin C Cawley, Nicola LC Talbot, Robert J Foxall, Stephen R Dorling, and Danilo P Mandic. Heteroscedastic kernel ridge regression. Neurocomputing, 57: 0 105--124, 2004

    2004

  2. [10]

    Stochastic gradient hamiltonian monte carlo

    Tianqi Chen, Emily Fox, and Carlos Guestrin. Stochastic gradient hamiltonian monte carlo. In International conference on machine learning, pages 1683--1691. PMLR, 2014

    2014

  3. [11]

    Repulsive deep ensembles are bayesian

    Francesco D'Angelo and Vincent Fortuin. Repulsive deep ensembles are bayesian. Advances in Neural Information Processing Systems, 34: 0 3451--3465, 2021

    2021

  4. [12]

    Laplace redux-effortless bayesian deep learning

    Erik Daxberger, Agustinus Kristiadi, Alexander Immer, Runa Eschenhagen, Matthias Bauer, and Philipp Hennig. Laplace redux-effortless bayesian deep learning. Advances in Neural Information Processing Systems, 34: 0 20089--20103, 2021

    2021

  5. [13]

    Decomposition of uncertainty in bayesian deep learning for efficient and risk-sensitive learning

    Stefan Depeweg, Jose-Miguel Hernandez-Lobato, Finale Doshi-Velez, and Steffen Udluft. Decomposition of uncertainty in bayesian deep learning for efficient and risk-sensitive learning. In International Conference on Machine Learning, pages 1184--1193. PMLR, 2018

    2018

  6. [14]

    Aleatory or epistemic? does it matter? Structural safety, 31 0 (2): 0 105--112, 2009

    Armen Der Kiureghian and Ove Ditlevsen. Aleatory or epistemic? does it matter? Structural safety, 31 0 (2): 0 105--112, 2009

    2009

  7. [15]

    Reliable training and estimation of variance networks

    Nicki Detlefsen, Martin J rgensen, and S ren Hauberg. Reliable training and estimation of variance networks. Advances in Neural Information Processing Systems, 32, 2019

    2019

  8. [16]

    O zg \"u n C i c ek, Ahmed Abdulkadir, Yassine Marrakchi, Anton B \

    Thorsten Falk, Dominic Mai, Robert Bensch, \"O zg \"u n C i c ek, Ahmed Abdulkadir, Yassine Marrakchi, Anton B \"o hm, Jan Deubner, Zoe J \"a ckel, Katharina Seiwald, et al. U-net: deep learning for cell counting, detection, and morphometry. Nature methods, 16 0 (1): 0 67--70, 2019

    2019

  9. [17]

    Is mc dropout bayesian? arXiv preprint arXiv:2110.04286, 2021

    Loic Le Folgoc, Vasileios Baltatzis, Sujal Desai, Anand Devaraj, Sam Ellis, Octavio E Martinez Manzanera, Arjun Nair, Huaqi Qiu, Julia Schnabel, and Ben Glocker. Is mc dropout bayesian? arXiv preprint arXiv:2110.04286, 2021

    2021 arXiv

  10. [18]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050--1059. PMLR, 2016

    2016

  11. [19]

    Strictly proper scoring rules, prediction, and estimation

    Tilmann Gneiting and Adrian E Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American statistical Association, 102 0 (477): 0 359--378, 2007

    2007

  12. [20]

    Probabilistic backpropagation for scalable learning of bayesian neural networks

    Jos \'e Miguel Hern \'a ndez-Lobato and Ryan Adams. Probabilistic backpropagation for scalable learning of bayesian neural networks. In International conference on machine learning, pages 1861--1869. PMLR, 2015

    2015

  13. [21]

    Revisiting single image depth estimation: Toward higher resolution maps with accurate object boundaries

    Junjie Hu, Mete Ozay, Yan Zhang, and Takayuki Okatani. Revisiting single image depth estimation: Toward higher resolution maps with accurate object boundaries. In 2019 IEEE winter conference on applications of computer vision (WACV), pages 1043--1051. IEEE, 2019

    2019

  14. [22]

    The apolloscape dataset for autonomous driving

    Xinyu Huang, Xinjing Cheng, Qichuan Geng, Binbin Cao, Dingfu Zhou, Peng Wang, Yuanqing Lin, and Ruigang Yang. The apolloscape dataset for autonomous driving. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 954--960, 2018

    2018

  15. [23]

    o lkopf, Peter B \

    Alexander Immer, Christoph Schultheiss, Julia E Vogt, Bernhard Sch \"o lkopf, Peter B \"u hlmann, and Alexander Marx. On the identifiability and estimation of causal location-scale noise models. In International Conference on Machine Learning, pages 14316--14332. PMLR, 2023

    2023

  16. [24]

    u rgens, Nis Meinert, Viktor Bengs, Eyke H \

    Mira J \"u rgens, Nis Meinert, Viktor Bengs, Eyke H \"u llermeier, and Willem Waegeman. Is epistemic uncertainty faithfully represented by evidential deep learning methods? arXiv preprint arXiv:2402.09056, 2024

    2024 arXiv

  17. [25]

    What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017

    Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017

    2017

  18. [26]

    Accurate uncertainties for deep learning using calibrated regression

    Volodymyr Kuleshov, Nathan Fenner, and Stefano Ermon. Accurate uncertainties for deep learning using calibrated regression. In International conference on machine learning, pages 2796--2804. PMLR, 2018

    2018

  19. [27]

    Deup: Direct epistemic uncertainty prediction

    Salem Lahlou, Moksh Jain, Hadi Nekoei, Victor Ion Butoi, Paul Bertin, Jarrid Rector-Brooks, Maksym Korablyov, and Yoshua Bengio. Deup: Direct epistemic uncertainty prediction. arXiv preprint arXiv:2102.08501, 2021

    2021 arXiv

  20. [28]

    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. Advances in neural information processing systems, 30, 2017

    2017

  21. [29]

    Heteroscedastic gaussian process regression

    Quoc V Le, Alex J Smola, and St \'e phane Canu. Heteroscedastic gaussian process regression. In Proceedings of the 22nd international conference on Machine learning, pages 489--496, 2005

    2005

  22. [30]

    Dropout injection at test time for post hoc uncertainty quantification in neural networks

    Emanuele Ledda, Giorgio Fumera, and Fabio Roli. Dropout injection at test time for post hoc uncertainty quantification in neural networks. Information Sciences, 645: 0 119356, 2023

    2023

  23. [31]

    A practical bayesian framework for backpropagation networks

    David JC MacKay. A practical bayesian framework for backpropagation networks. Neural computation, 4 0 (3): 0 448--472, 1992

    1992

  24. [32]

    A simple baseline for bayesian uncertainty in deep learning

    Wesley J Maddox, Pavel Izmailov, Timur Garipov, Dmitry P Vetrov, and Andrew Gordon Wilson. A simple baseline for bayesian uncertainty in deep learning. Advances in neural information processing systems, 32, 2019

    2019

  25. [33]

    The unreasonable effectiveness of deep evidential regression

    Nis Meinert, Jakob Gawlikowski, and Alexander Lavin. The unreasonable effectiveness of deep evidential regression. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 9134--9142, 2023

    2023

  26. [34]

    Mcmc using hamiltonian dynamics

    Radford M Neal et al. Mcmc using hamiltonian dynamics. Handbook of markov chain monte carlo, 2 0 (11): 0 2, 2011

    2011

  27. [35]

    Estimating the mean and variance of the target probability distribution

    David A Nix and Andreas S Weigend. Estimating the mean and variance of the target probability distribution. In Proceedings of 1994 ieee international conference on neural networks (ICNN'94), volume 1, pages 55--60. IEEE, 1994

    1994

  28. [36]

    A scalable laplace approximation for neural networks

    Hippolyt Ritter, Aleksandar Botev, and David Barber. A scalable laplace approximation for neural networks. In 6th International Conference on Learning Representations, ICLR 2018-Conference Track Proceedings, volume 6. International Conference on Representation Learning, 2018

    2018

  29. [37]

    Second-order uncertainty quantification: Variance-based measures

    Yusuf Sale, Paul Hofman, Lisa Wimmer, Eyke H \"u llermeier, and Thomas Nagler. Second-order uncertainty quantification: Variance-based measures. arXiv preprint arXiv:2401.00276, 2023

    2023 arXiv

  30. [38]

    On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks

    Maximilian Seitzer, Arash Tavakoli, Dimitrije Antic, and Georg Martius. On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks. arXiv preprint arXiv:2203.09168, 2022

    2022 arXiv

  31. [39]

    Evidential deep learning to quantify classification uncertainty

    Murat Sensoy, Lance Kaplan, and Melih Kandemir. Evidential deep learning to quantify classification uncertainty. Advances in neural information processing systems, 31, 2018

    2018

  32. [40]

    Indoor segmentation and support inference from rgbd images

    Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision--ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746--760. S...

    2012

  33. [41]

    Bayescap: Bayesian identity cap for calibrated uncertainty in frozen neural networks

    Uddeshya Upadhyay, Shyamgopal Karthik, Yanbei Chen, Massimiliano Mancini, and Zeynep Akata. Bayescap: Bayesian identity cap for calibrated uncertainty in frozen neural networks. In European Conference on Computer Vision, pages 299--317. Springer, 2022

    2022

  34. [42]

    Lisa Wimmer, Yusuf Sale, Paul Hofman, Bernd Bischl, and Eyke H \"u llermeier. Quantifying aleatoric and epistemic uncertainty in machine learning: Are conditional entropy and mutual information appropriate measures? In Uncertainty in Artificial Intelligence, pages 2282--2292. ...

    2023

  35. [43]

    Doubly penalized likelihood estimator in heteroscedastic regression

    Ming Yuan and Grace Wahba. Doubly penalized likelihood estimator in heteroscedastic regression. Statistics & probability letters, 69 0 (1): 0 11--20, 2004

    2004

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.