Pith. sign in

REVIEW 3 major objections 5 minor 78 references

A Unified Risk View of Uncertainty: Posterior Risk for Disentanglement and Evaluation Beyond Proxies

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper defines uncertainty as a predictor's expected loss under the posterior over plausible ground-truth functions, unifying the Bayesian variance view and the frequentist excess-risk view, and uses GP semi-synthetic regression to…

desk verdict Sound theory and a transparent oracle benchmark for uncertainty disentanglement, with untested external validity as the main caveat. read the letter →

arxiv 2608.05995 v1 pith:S7ZF2DQN submitted 2026-08-06 cs.LG

Frieder Wizgall , Georg Tirpitz , Moritz Seiler , Kerstin Ritter , Bálint Mucsányi This is my paper · ORCID
classification cs.LG MSC 62F1568T05
keywords uncertaintydisentanglementepistemicaleatoricposteriorriskBregmandivergenceGaussianprocesspriorsemi-syntheticbenchmarkoracleevaluation
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

Uncertainty estimates are only useful if they reflect what a deployed predictor does not know, but the field has no agreed target for epistemic uncertainty, and benchmarks substitute proxy tasks such as out-of-distribution detection. This paper proposes sample-conditional posterior risk as that target: the expected loss of the deployed predictor under the distribution of ground-truth functions that remain plausible given the observed data. Under squared loss the target splits cleanly into aleatoric noise, posterior variance over functions, and a squared estimator-bias term, which shows that Bayesian posterior variance and frequentist excess risk are two ends of one spectrum. The paper adds a benchmark in which targets are generated from a known Gaussian-process prior, so oracle aleatoric and epistemic uncertainty are exactly computable, and uses it to show that predictive accuracy does not imply trustworthy disentanglement.

What carries the argument

The load-bearing object is the sample-conditional posterior risk defined in Eq. (2), $R_S(\hat f;x) = \mathbb{E}_{f \sim \Pi(\cdot|S)} \mathbb{E}_{Y \sim P(\cdot|x,f)}[\ell(\hat f(x), Y)]$, together with its Bregman decomposition (Proposition 2.3), which splits the risk into a Bayes-error (aleatoric) term, a generalized variance over the posterior, and a generalized bias of the estimator. Under squared loss this collapses to the identity $R_S(\hat f;x) = \sigma_\epsilon^2(x) + \sigma_f^2(x) + (\hat f(x)-m_S(x))^2$, which is what turns the definition into a computable evaluation target: in the semi-synthetic GP setting the posterior $\Pi(\cdot|S)$ is the same Gaussian process that generated the targets, so $\sigma_\epsilon^2(x)$, $\sigma_f^2(x)$, and $m_S(x)$ are all known analytically. The benchmark estimates each method's reported epistemic and aleatoric uncertainty and scores them by Spearman rank correlation against these oracle targets, with ablation studies of architecture, activation, tuning objective, and optimizer serving to control the many degrees of freedom that otherwise dominate method comparisons.

What would settle it

Re-run the benchmark suite with targets drawn from a non-GP generative process (for example, a Bayesian neural network prior or a Student-t noise model) and compute approximate oracle targets by Hamiltonian Monte Carlo; if the methods that rank highest under the GP oracle do not also lead in Spearman correlation against these approximate targets, the transfer assumption fails. A cheaper check is to recompute oracle rankings under CRPS on the existing GP presets; if method rankings flip substantially, the current conclusions are specific to squared loss.

Watch

Extended reading notes

Core claim

The central claim is that uncertainty disentanglement should be anchored to sample-conditional pointwise posterior risk, $R_S(\hat f; x) = \mathbb{E}_{Y \sim P(\cdot|x,S)}[\ell(\hat f(x), Y)]$, the expected loss of the actually deployed predictor under the posterior over plausible latent ground-truth functions given the observed sample $S$. For squared loss with additive noise this becomes $$R_S(\hat f; x) = \sigma_\$epsilon^{2}$(x) + \$sigma_f^{2}$(x) + (\hat f(x) - m_S(x))^2,$$ so aleatoric uncertainty is the noise variance, and epistemic uncertainty is the sum of the posterior variance $\sigma_f^2(x)$ and the squared deviation of the estimator from the posterior mean $m_S(x)$. The Bayesian posterior-variance view and the frequentist pointwise-excess-risk view are both recovered as special cases: the former when the deployed predictor equals the posterior mean, the latter when the posterior collapses to a Dirac measure at one ground-truth function. The paper then constructs semi-synthetic regression datasets with real covariates and GP-sampled targets, making the oracle epistemic and aleatoric targets exactly computable, and finds that current methods, evaluated by rank correlation against these targets, are far from oracle-level, generally track the posterior-variance component more than the bias component, and perform unevenly across datasets.

Load-bearing premise

The whole empirical ranking rests on the assumption that methods ranked against a perfectly specified GP oracle, where the posterior used for evaluation is the same distribution that generated the targets, will perform similarly in realistic settings where the model is misspecified and no exact posterior exists; the paper states this transfer is untested.

Editorial extensions

If this is right

  • If posterior risk is the right target, then a method that reports only posterior spread can understate epistemic uncertainty whenever the deployed predictor deviates from the posterior mean, and a method that reports only pointwise excess risk can understate uncertainty in data-sparse regions.
  • Oracle evaluation without proxies is feasible in controlled settings: real covariates with GP-generated targets and a correctly specified posterior yield exact targets for aleatoric and epistemic uncertainty at every evaluation point.
  • Accurate point prediction does not imply trustworthy uncertainty disentanglement: in the benchmark, models with similar predictive performance differ substantially in their Spearman correlation against the oracle targets.
  • Deep ensembles, FSP-Laplace, and CatBoost-KGB emerge as the most reliable general-purpose starting points among tested methods, while aleatoric rankings remain weakly separated and dataset-dependent.
  • Most epistemic estimates align more with the posterior-variance component than with the full posterior excess-risk target, meaning current methods largely miss the estimator-bias contribution to epistemic uncertainty.

Reading between the lines

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

  • Beyond the paper: recomputing oracle rankings under CRPS or NLL (whose decompositions the appendix derives) would directly test whether the benchmark conclusions depend on the squared-loss choice, and method rankings could plausibly shift.
  • Beyond the paper: the implicit transfer claim is that GP-oracle rankings predict performance in realistic misspecified settings; this could be tested by generating data from a deliberately non-GP process (e.g., a Bayesian neural network prior or heavy-tailed noise) and checking whether the top-ranked methods under the GP oracle still lead against approximate MCMC targets.
  • Beyond the paper: the definition suggests a practical diagnostic for safety-critical deployment, namely computing the squared-bias term $(\hat f(x)-m_S(x))^2$ whenever a posterior approximation is available, since methods that ignore it will systematically understate epistemic uncertainty under misspecification.
  • Beyond the paper: the benchmark's oracle targets could be reused for model selection or early stopping, because they separate irreducible noise from reducible estimator error pointwise in a way proxy tasks cannot.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a unified definition of uncertainty as sample-conditional posterior risk: for a fixed input x and observed sample S, the uncertainty of a deployed predictor f_hat is the expected loss under the posterior distribution over plausible ground-truth functions (Eq. 2). Under squared loss, this decomposes into aleatoric uncertainty sigma_eps^2(x), posterior variance sigma_f^2(x), and an estimator-dependent squared bias term (f_hat(x)-m_S(x))^2 (Eq. 4). The paper shows that the Bayesian posterior-variance view and the frequentist pointwise-excess-risk view are recovered as special cases, and it develops a semi-synthetic benchmark in which real covariates are paired with targets generated from a Gaussian-process prior, so that oracle aleatoric and epistemic targets are analytically computable. The empirical study evaluates a range of uncertainty-disentanglement methods on a development suite and a held-out benchmark suite, using Spearman rank correlation against the oracle targets as the primary metric, and reports ablations over architecture, activation, tuning objective, and optimizer. The headline findings are that predictive accuracy does not guarantee good uncertainty disentanglement, that ensemble-style neural methods and several tree-based methods provide the strongest epistemic rankings, and that EDL is strongest in aleatoric/epistemic calibration.

Significance. The theoretical core of the paper is sound and valuable. The derivation from Eq. (2) to Eq. (4) correctly applies the Bregman bias-variance identity, and the special-case recoveries of posterior-variance-based and excess-risk-based epistemic uncertainty are clean and clearly stated. The explicit separation of estimator-dependent bias from posterior spread is a useful conceptual contribution, as is the point that standard posterior-spread measures can miss misspecification and optimization error. The benchmark is also carefully constructed: the GP-based oracle targets are correctly specified by construction, the development/held-out split is a genuine strength, and the empirical study reports standard errors and multiple diagnostics, including risk-coverage curves, calibration plots, and a decomposition of epistemic estimates into posterior-variance and bias components. If the method rankings are taken as internally valid, the paper provides concrete practical information about which uncertainty estimators align with a well-defined oracle target.

major comments (3)
  1. [Sections 4.1, 5.2, 6, and 7] The benchmark's oracle targets are internally valid because the data-generating GP prior and the posterior used for computing Eq. (4) coincide, but the paper's practical recommendations in Section 7 (deep ensembles, FSP-Laplace, and CatBoost-KGB as reliable starting points) are derived solely from rankings against this correctly specified GP oracle. Since the stated contribution is an evaluation 'beyond proxies,' the external relevance of these rankings is load-bearing. The paper should add a misspecification stress test, for example by generating targets from a different kernel than the one used for the posterior, or from a non-GP function class, and showing whether the method rankings are preserved. Without such a test, the benchmark establishes internal consistency but does not substantiate the general practical takeaways.
  2. [Eq. (4) and Section 5.2] The oracle epistemic target used for ranking methods is method-specific: it contains the term (f_hat(x)-m_S(x))^2, so the target depends on the predictor being evaluated. A method with a predictive mean close to the GP posterior mean faces a target dominated by sigma_f^2, whereas a method with a biased mean faces a target with a substantial additional squared-bias component. Cross-method comparisons of Spearman rank correlation against such method-specific targets can therefore conflate the quality of the uncertainty estimate with differences in the composition of the target. The paper should report, alongside Figure 2, correlations against a common target such as sigma_f^2 alone, or otherwise demonstrate that the main method rankings are stable when the target composition changes. Figure 13 partially addresses this issue, but it is relegated to the appendix and does not directly test the stability of the headline ranking.
  3. [Sections 4.1 and 5.2] Several of the methods that rank well in the main benchmark are themselves GP-based or GP-motivated (notably CatBoost-KGB, which explicitly samples from a GP posterior, and FSP-Laplace, which regularizes toward a GP prior). Given that the oracle targets are computed from a GP posterior, these methods may benefit from an advantage that is inherent to the evaluation protocol rather than to their general uncertainty-disentanglement quality. The paper should either analyze this potential confound explicitly, for example by comparing method rankings on tasks where the prior is deliberately misspecified, or temper the recommendations so that they are clearly restricted to the correctly specified GP setting.
minor comments (5)
  1. [Section 1 and Appendix A] The empirical benchmark is instantiated only under squared loss, although Appendix A derives decompositions for CRPS and negative log-likelihood. The main text should state more prominently that all benchmark conclusions are specific to the squared-loss risk functional, since Implication 2 notes that the semantics of aleatoric and epistemic uncertainty depend on the chosen loss or scoring rule.
  2. [Section 5.1] The text says that 'the ablations ran in between 3-9 hours while the final evaluation took 4hours'; '4hours' should be '4 hours', and 'ran in between' is awkwardly phrased.
  3. [Throughout] The name 'SW AG' appears with a space in multiple places (e.g., Section 5 and Appendix D.4), which is inconsistent with the standard notation 'SWAG' used elsewhere; please standardize.
  4. [Section 5.2] The reported Kendall's W value of approximately 0.46 for method-rank preservation across datasets is given without a confidence interval or significance test; since the number of datasets is small (seven), the uncertainty in W should be quantified or at least discussed.
  5. [Appendix C] The dataset presets report 'Posterior std median' values but do not explain how this quantity is computed or how it should be interpreted relative to the aleatoric std. A brief definition in the appendix would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the benchmark oracle is self-referential only in the transparent, acknowledged sense of a controlled semi-synthetic reference.

full rationale

The paper's formal chain is self-contained: Definition 2.2 stipulates sample-conditional posterior risk, and Proposition 2.3/Eq. (4) are algebraic consequences of the Bregman bias-variance identity [Pfau, 2025] under the stated MSE additive-noise assumptions. No fitted parameter is renamed as a prediction in the theory. The benchmark oracle is intentionally constructed from the same GP prior that generates the targets, as stated in Section 4.1: 'Because the prior used for conditioning is the same prior that generated the target function, the resulting posterior is correctly specified by construction.' This makes the oracle internally consistent rather than independently measured, and the paper explicitly disclaims broader ground-truth status in Section 1: 'We do not claim that this GP-based construction provides a notion of ground truth that generalizes to all other modalities' and in Section 6, where the GP restriction is listed as a limitation. The only self-citation, Mucsányi et al. [2024], is background on existing benchmark heterogeneity, not a load-bearing uniqueness or existence claim. Concerns about transfer of method rankings to misspecified non-GP settings are external-validity risks, not circularity.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

No new physical entities are postulated. The free parameters are the fitted GP hyperparameters, the heteroscedastic noise schedule, the covariate-shift sampling profiles, and the hyperparameters of all evaluated methods. The axioms are standard convex-analysis assumptions plus the explicitly stated GP semi-synthetic domain assumption that makes oracle targets computable.

free parameters (4)
  • GP kernel hyperparameters per preset = per-preset values in Tables 1-14, e.g. Abalone sigma^2=2.001, kernel arc-cosine degree 0
    Fit by marginal likelihood to the original UCI/OpenML data, then used to draw synthetic targets and define the posterior oracle.
  • Heteroscedastic noise model sigma(x) = reported as per-preset aleatoric std, e.g.
    Design choice controlling the aleatoric oracle target; the exact mapping from feature rank to noise variance is not fully specified.
  • Mahalanobis-bin sampling profiles = fractions such as balanced (0.2,0.2,0.2,0.2,0.2) or OOD (0.5,0,0,0,0)
    Chosen to create controlled covariate-shift scenarios; these are hand-set evaluation conditions.
  • Method hyperparameters = per-method search spaces in Appendix D, tuned via 10 W&B Bayesian sweeps
    All evaluated methods are tuned on the development suite, so the benchmark rankings depend on these fitted hyperparameters.
assumptions (6)
  • standard math All expectations are finite and Bayes acts/centroids exist (Prop 2.3).
    Invoked in Prop 2.3 and Appendix A to justify the Bregman decomposition.
  • standard math phi is strictly convex and continuously differentiable; losses are proper scoring rules or squared loss.
    Required for the Bregman bias-variance identity taken from Pfau (2025).
  • domain assumption The data-generating process is a Gaussian process with known noise model, so the posterior p(f|S) is correctly specified and analytic.
    Section 4.1: the entire oracle-target protocol depends on this GP construction.
  • domain assumption Squared-error/MSE risk is the operative oracle for the benchmark.
    Limitations Section 6 states oracle computations are instantiated only under squared-error/MSE risk.
  • domain assumption Training sample S is drawn according to Mahalanobis-bin profiles and noise variance depends only on x.
    Section 4.1 and Appendix C define the evaluation scenarios; the oracle AU depends on this noise model.
  • domain assumption The deployed predictor fhat is evaluated pointwise against a posterior that is conditionally independent of fhat given S.
    Implicit in Definition 2.2: the posterior over functions is conditioned only on the observed sample, not on the predictor's training procedure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Unified Risk View of Uncertainty: Posterior Risk for Disentanglement and Evaluation Beyond Proxies." pith.science (2026). https://pith.science/paper/S7ZF2DQN

@misc{pith2026260805995,
  author       = {Pith},
  title        = {Pith review of: A Unified Risk View of Uncertainty: Posterior Risk for Disentanglement and Evaluation Beyond Proxies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S7ZF2DQN}},
  note         = {Machine review of arXiv:2608.05995}
}
read the original abstract

Reliable uncertainty estimates are critical in safety-sensitive applications, where understanding the sources of predictive uncertainty is essential. This often requires disentangling epistemic uncertainty from aleatoric uncertainty, yet these uncertainty types are not defined consistently across the literature, making it difficult to assess whether a method produces accurate uncertainty estimates. Evaluation is further complicated by the fact that ground-truth epistemic uncertainty is typically unavailable. Existing benchmarks therefore mostly rely on proxy tasks such as out-of-distribution detection, which do not provide complete ground-truth uncertainty targets and offer limited insight into the structure and quality of uncertainty estimates. We propose a unified definition of uncertainty as pointwise posterior risk, the expected loss of a predictor under the distribution of plausible ground-truth functions given the data. This view combines Bayesian uncertainty over functions with estimator-dependent deviations from the posterior mean, capturing effects such as misspecification and optimization error. This formulation constitutes the foundation of a theory-backed benchmark that enables direct computation of oracle epistemic and aleatoric uncertainty using semi-synthetic datasets with real covariates and known generative processes. By avoiding proxy evaluations, the benchmark enables fine-grained analysis of uncertainty estimates. Empirically, we find that accurate prediction does not guarantee reliable uncertainty disentanglement. The benchmark reveals practically useful differences between methods, identifying approaches with meaningful alignment to oracle uncertainty targets while exposing sensitivity to datasets and modeling choices.

Figures

Figures reproduced from arXiv: 2608.05995 by the authors.

Figure 1
Figure 1. Different definitions of epistemic uncertainty. Exc. is the abbreviation for Excess. A: Epistemic uncertainty defined as the Bayesian posterior variance of the true posterior. B: EU defined as pointwise excess risk. C: Oracle EU according to the resampling frequentist view. D: Our pointwise posterior excess risk oracle EU. Substantial differences between the definitions are visible, especially in how they behave awa… view at source ↗
Figure 2
Figure 2. Main benchmark results on the held-out benchmark suite. Bars show mean Spearman rank correlation between predicted and oracle epistemic or aleatoric uncertainty. Error bars show ±2 standard errors of the mean across datasets. Grey regions indicate methods that do not provide a heteroscedastic aleatoric estimate. Current methods recover only moderate uncertainty rankings: ensemble-style neural methods, especially NN-… view at source ↗
Figure 3
Figure 3. Empirical sanity check for the estimation–approximation decomposition. Normalized estimation and approximation errors as a function of the number of observations. We compare a well-specified GP, which uses the correct data-generating kernel, with an under-capacitated lin￾ear GP. Estimation error decreases with more observations, whereas approximation error remains comparatively stable. The linear GP has larger appro… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Epistemic uncertainty ranking as a function of network depth , shown separately for each development dataset and BNN method. The curves report Spearman rank correlation between predicted epistemic uncertainty and the oracle epistemic target. The results show substantia…
Figure 5
Figure 5. Figure 5: Aleatoric uncertainty ranking as a function of network depth, shown separately for each development dataset and BNN method. The curves report Spearman rank correlation between predicted aleatoric uncertainty and the oracle aleatoric target. As for epistemic uncertainty…
Figure 6
Figure 6. Figure 6: Epistemic uncertainty ranking as a function of network width, shown separately for each development dataset and BNN method. The curves report Spearman rank correlation between predicted epistemic uncertainty and the oracle epistemic target. The width effect is noisy an…
Figure 7
Figure 7. Figure 7: Aleatoric uncertainty ranking as a function of network width, shown separately for each development dataset and BNN method. The curves report Spearman rank correlation between predicted aleatoric uncertainty and the oracle aleatoric target. Although individual curves a…
Figure 8
Figure 8. Figure 8: Direct comparison of ReLU and Tanh activations for epistemic uncertainty ranking (left) and aleatoric uncertainty ranking (right). Each point corresponds to one model–dataset pair. The x-axis shows Spearman rank correlation when using ReLU, and the y-axis shows the cor…
Figure 9
Figure 9. Figure 9: Heatmap of the tuning objective condition. Heatmap of the change in Spearman rank correlation when hyperparameters are selected by validation NLL instead of validation RMSE, shown separately for epistemic and aleatoric uncertainty across methods and datasets. Positive …
Figure 10
Figure 10. Figure 10: Heatmap of the optimizer condition. Heatmap of the change in Spearman rank correla￾tion relative to the optimizer-wise mean, shown separately for epistemic and aleatoric uncertainty across methods, datasets, and optimizers. Positive values indicate that a given method…
Figure 11
Figure 11. Figure 11: Risk-coverage curves for disentangled uncertainties. Risk-coverage curves for predicted aleatoric uncertainty (top), epistemic uncertainty (middle), and total uncertainty (bottom) on the held-out benchmark suite. For each model, test points are sorted from lowest to h…
Figure 12
Figure 12. Figure 12: Uncertainty calibration plots for total uncertainty (top), epistemic uncertainty (middle), and aleatoric uncertainty (bottom), with NN methods on the left and non-NN methods on the right. For each method, predictions are binned by estimated uncertainty and the mean pr…
Figure 13
Figure 13. Figure 13: Epistemic subpart analysis. Spearman rank correlations between each method’s epistemic uncertainty estimate and three oracle quantities on the validation suite: the full posterior excess-risk epistemic target, the posterior variance term, and the squared bias term. Er…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

78 extracted references · 63 canonical work pages

  1. [1]

    2011 , eprint =

    Bayesian Active Learning for Classification and Preference Learning , author =. 2011 , eprint =. doi:10.48550/arXiv.1112.5745 , url =

    arXiv doi:10.48550/arxiv.1112.5745 2011

  2. [2]

    2017 , eprint =

    What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? , author =. 2017 , eprint =

    work page 2017

  3. [3]

    2018 , eprint =

    Understanding Measures of Uncertainty for Adversarial Example Detection , author =. 2018 , eprint =. doi:10.48550/arXiv.1803.08533 , url =

    arXiv doi:10.48550/arxiv.1803.08533 2018

  4. [4]

    2022 , eprint =

    Disentangling Epistemic and Aleatoric Uncertainty in Reinforcement Learning , author =. 2022 , eprint =

    work page 2022

  5. [5]

    2022 , eprint =

    Quantifying Aleatoric and Epistemic Uncertainty in Machine Learning: Are Conditional Entropy and Mutual Information Appropriate Measures? , author =. 2022 , eprint =

    work page 2022

  6. [6]

    2025 , eprint =

    An Axiomatic Assessment of Entropy- and Variance-based Uncertainty Quantification in Regression , author =. 2025 , eprint =

    work page 2025

  7. [7]

    2024 , eprint =

    Benchmarking Uncertainty Disentanglement: Specialized Uncertainties for Specialized Tasks , author =. 2024 , eprint =

    work page 2024

  8. [8]

    2024 , eprint =

    From Risk to Uncertainty: Generating Predictive Uncertainty Measures via Bayesian Estimation , author =. 2024 , eprint =

    work page 2024

Show all 78 references
  1. [9]

    2021 , eprint =

    DEUP: Direct Epistemic Uncertainty Prediction , author =. 2021 , eprint =

    2021

  2. [10]

    2025 , eprint =

    Uncertainty Quantification for Regression using Proper Scoring Rules , author =. 2025 , eprint =

    2025

  3. [11]

    2025 , eprint =

    Why Machine Learning Models Fail to Fully Capture Epistemic Uncertainty , author =. 2025 , eprint =

    2025

  4. [12]

    2021 , eprint=

    Uncertainty estimation under model misspecification in neural network regression , author=. 2021 , eprint=

    2021

  5. [13]

    2020 , eprint =

    Maximum likelihood estimation and uncertainty quantification for Gaussian process approximation of deterministic functions , author =. 2020 , eprint =. doi:10.48550/arXiv.2001.10965 , url =

    2020 doi arXiv

  6. [14]

    2025 , eprint =

    Scale estimation and rate-unbiasedness for Gaussian processes under smoothness misspecification , author =. 2025 , eprint =. doi:10.48550/arXiv.2110.02810 , url =

    2025 doi

  7. [15]

    , urldate =

    Vapnik, Vladimir N. , urldate =. The Nature of Statistical Learning Theory , rights =. 2000 , doi =

    2000

  8. [16]

    2025 , eprint =

    Out-of-Distribution Detection Methods Answer the Wrong Questions , author =. 2025 , eprint =. doi:10.48550/arXiv.2507.01831 , url =

    2025 doi arXiv

  9. [17]

    2023 , eprint=

    How Reliable is Your Regression Model's Uncertainty Under Real-World Distribution Shifts? , author=. 2023 , eprint=

    2023

  10. [18]

    2022 , eprint=

    Robust uncertainty estimates with out-of-distribution pseudo-inputs training , author=. 2022 , eprint=

    2022

  11. [19]

    2020 , eprint=

    Deep Evidential Regression , author=. 2020 , eprint=

    2020

  12. [20]

    2024 , eprint=

    Are Uncertainty Quantification Capabilities of Evidential Deep Learning a Mirage? , author=. 2024 , eprint=

    2024

  13. [21]

    2022 , eprint=

    A view on model misspecification in uncertainty quantification , author=. 2022 , eprint=

    2022

  14. [22]

    Sculley and Joshua V

    Jasper Snoek and Yaniv Ovadia and Emily Fertig and Balaji Lakshminarayanan and Sebastian Nowozin and D. Sculley and Joshua V. Dillon and Jie Ren and Zachary Nado , editor =. Can you trust your model's uncertainty? Evaluating predictive uncertainty under dataset shift , booktit...

    2019

  15. [23]

    Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods , journal =

    Eyke H. Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods , journal =. 2021 , doi =

    2021

  16. [24]

    Erickson , title =

    Shahriar Faghani and Mana Moassefi and Pouria Rouzrokh and Bardia Khosravi and Francis Baffour and Michael D Ringler and Bradley J. Erickson , title =. Radiology , year =. doi:10.1148/radiol.222217 , url =

    doi

  17. [25]

    Yuko Kato and David M. J. Tax and Marco Loog , title =. CoRR , volume =. 2022 , doi =

    2022

  18. [26]

    Medical Image Analysis , volume =

    A review of uncertainty quantification in medical image analysis: Probabilistic and non-probabilistic methods , author =. Medical Image Analysis , volume =. 2024 , doi =

    2024

  19. [27]

    Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods , volume=

    Hüllermeier, Eyke and Waegeman, Willem , year=. Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods , volume=. Machine Learning , publisher=. doi:10.1007/s10994-021-05946-3 , number=

    doi

  20. [28]

    Proceedings of the 39th International Conference on Machine Learning , series =

    On the Practicality of Deterministic Epistemic Uncertainty , author =. Proceedings of the 39th International Conference on Machine Learning , series =. 2022 , editor =

    2022

  21. [29]

    Journal of Machine Learning Research , volume =

    Deep Out-of-Distribution Uncertainty Quantification via Weight Entropy Maximization , author =. Journal of Machine Learning Research , volume =. 2025 , url =. 2309.15704 , archivePrefix =

    2025 arXiv

  22. [30]

    ICML 2021 Workshop on Uncertainty and Robustness in Deep Learning , year =

    On Pitfalls in OoD Detection: Predictive Entropy Considered Harmful , author =. ICML 2021 Workshop on Uncertainty and Robustness in Deep Learning , year =

    2021

  23. [31]

    2023 , eprint =

    Quantification of Uncertainty with Adversarial Models , author =. 2023 , eprint =. doi:10.48550/arXiv.2307.03217 , url =

    2023 doi arXiv

  24. [32]

    Machine Learning: Science and Technology , volume =

    Deeply Uncertain: Comparing Methods of Uncertainty Quantification in Deep Learning Algorithms , author =. Machine Learning: Science and Technology , volume =. 2021 , doi =

    2021

  25. [33]

    Reliability Engineering & System Safety , volume =

    A benchmark on uncertainty quantification for deep learning prognostics , author =. Reliability Engineering & System Safety , volume =. 2025 , doi =

    2025

  26. [34]

    MELBA: Journal of Machine Learning for Biomedical Imaging , volume =

    Recalibration of Aleatoric and Epistemic Regression Uncertainty in Medical Imaging , author =. MELBA: Journal of Machine Learning for Biomedical Imaging , volume =. 2021 , pages =

    2021

  27. [35]

    Pattern Recognition Letters , volume =

    Uncertainty quantification metrics for deep regression , author =. Pattern Recognition Letters , volume =. 2024 , doi =

    2024

  28. [36]

    2024 , archivePrefix=

    One Step Closer to Unbiased Aleatoric Uncertainty Estimation , author =. 2024 , archivePrefix=. 2312.10469 , primaryClass =

    2024 arXiv

  29. [37]

    NPJ Digital Medicine , author =

    Second opinion needed: communicating uncertainty in medical machine learning , volume =. NPJ Digital Medicine , author =. 2021 , pmid =. doi:10.1038/s41746-020-00367-3 , abstract =

    2021 doi

  30. [38]

    Advances in Neural Information Processing Systems , year =

    What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? , author =. Advances in Neural Information Processing Systems , year =

  31. [39]

    Machine Learning: Science and Technology , volume =

    Deeply Uncertain: Comparing Methods of Uncertainty Quantification in Deep Learning Algorithms , author =. Machine Learning: Science and Technology , volume =. 2020 , doi =

    2020

  32. [40]

    2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , pages =

    A Deeper Look into Aleatoric and Epistemic Uncertainty Disentanglement , author =. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , pages =. 2022 , doi =

    2022

  33. [41]

    Understanding the bias-variance tradeoff of

    Adlam, Ben and Gupta, Neha and Mariet, Zelda and Smith, Jamie , month = feb, year =. Understanding the bias-variance tradeoff of. doi:10.48550/arXiv.2202.04167 , abstract =

    doi arXiv

  34. [42]

    2022 , eprint=

    Nonparametric Uncertainty Quantification for Single Deterministic Neural Network , author=. 2022 , eprint=

    2022

  35. [43]

    2024 , eprint=

    Efficient Lifelong Model Evaluation in an Era of Rapid Progress , author=. 2024 , eprint=

    2024

  36. [44]

    2025 , eprint=

    ImageNot: A contrast with ImageNet preserves model rankings , author=. 2025 , eprint=

    2025

  37. [45]

    2015 , eprint=

    ImageNet Large Scale Visual Recognition Challenge , author=. 2015 , eprint=

    2015

  38. [46]

    Sebastian Felix Fischer and Matthias Feurer and Bernd Bischl , booktitle=. Open. 2023 , url=

    2023

  39. [47]

    2025 , eprint=

    A Generalized Bias-Variance Decomposition for Bregman Divergences , author=. 2025 , eprint=

    2025

  40. [48]

    2024 , eprint=

    Is Epistemic Uncertainty Faithfully Represented by Evidential Deep Learning Methods? , author=. 2024 , eprint=

    2024

  41. [49]

    2021 , eprint=

    Implicit Bias of SGD for Diagonal Linear Networks: a Provable Benefit of Stochasticity , author=. 2021 , eprint=

    2021

  42. [50]

    MacKay, David J. C. , title =. Neural Computation , year = 1992, volume = 4, number = 3, month = may, doi =

    1992

  43. [51]

    and Basu, Sanjay and Ng, Andrew and Schuler, Alejandro , title =

    Duan, Tony and Avati, Anand and Ding, Daisy Yi and Thai, Khanh K. and Basu, Sanjay and Ng, Andrew and Schuler, Alejandro , title =. Proceedings of the 37th International Conference on Machine Learning , articleno =. 2020 , publisher =

    2020

  44. [52]

    2021 , eprint=

    Uncertainty in Gradient Boosting via Ensembles , author=. 2021 , eprint=

    2021

  45. [53]

    2019 , eprint=

    CatBoost: unbiased boosting with categorical features , author=. 2019 , eprint=

    2019

  46. [54]

    2023 , eprint=

    Gradient Boosting Performs Gaussian Process Inference , author=. 2023 , eprint=

    2023

  47. [55]

    Revisiting Deep Learning Models for Tabular Data , url =

    Gorishniy, Yury and Rubachev, Ivan and Khrulkov, Valentin and Babenko, Artem , booktitle =. Revisiting Deep Learning Models for Tabular Data , url =

  48. [56]

    2017 , eprint=

    Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles , author=. 2017 , eprint=

    2017

  49. [57]

    Proceedings of The 33rd International Conference on Machine Learning , pages =

    Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning , author =. Proceedings of The 33rd International Conference on Machine Learning , pages =. 2016 , editor =

    2016

  50. [58]

    Laplace Redux - Effortless Bayesian Deep Learning , url =

    Daxberger, Erik and Kristiadi, Agustinus and Immer, Alexander and Eschenhagen, Runa and Bauer, Matthias and Hennig, Philipp , booktitle =. Laplace Redux - Effortless Bayesian Deep Learning , url =

  51. [59]

    Laplax: Laplace Approximations in JAX , year =

    Weber, Tobias and Mucs. Laplax: Laplace Approximations in JAX , year =. GitHub repository , howpublished =

  52. [60]

    2024 , eprint=

    FSP-Laplace: Function-Space Priors for the Laplace Approximation in Bayesian Deep Learning , author=. 2024 , eprint=

    2024

  53. [61]

    Machine Learning , volume =

    Bagging Predictors , author =. Machine Learning , volume =. 1996 , doi =

    1996

  54. [62]

    2019 , eprint=

    A Simple Baseline for Bayesian Uncertainty in Deep Learning , author=. 2019 , eprint=

    2019

  55. [63]

    1994 , howpublished =

    Nash, Warwick and Sellers, Tracy and Talbot, Simon and Cawthorn, Andrew and Ford, Wes , title =. 1994 , howpublished =

    1994

  56. [64]

    and Marcolini, Michael , title =

    Brooks, Thomas and Pope, D. and Marcolini, Michael , title =. 1989 , howpublished =

    1989

  57. [65]

    1998 , howpublished =

    Yeh, I-Cheng , title =. 1998 , howpublished =

    1998

  58. [66]

    2015 , howpublished =

    Ballabio, Davide and Cassotti, Matteo and Consonni, Viviana and Todeschini, Roberto , title =. 2015 , howpublished =

    2015

  59. [67]

    2020 , howpublished =

    2020

  60. [68]

    , howpublished =

    n.d. , howpublished =

  61. [69]

    2022 , howpublished =

    2022

  62. [70]

    , title =

    Olson, Craig A. , title =. Journal of Applied Econometrics , year =

  63. [71]

    Wickham, Hadley , year =

  64. [72]

    Proceedings of the Seventeenth International Conference on Machine Learning , year =

    Vijayakumar, Sethu and Schaal, Stefan , title =. Proceedings of the Seventeenth International Conference on Machine Learning , year =

  65. [73]

    2023 , howpublished =

    2023

  66. [74]

    Kelley and Barry, Ronald , title =

    Pace, R. Kelley and Barry, Ronald , title =. Geographical Analysis , year =

  67. [75]

    Modeling Wine Preferences by Data Mining from Physicochemical Properties , journal =

    Cortez, Paulo and Cerdeira, Ant. Modeling Wine Preferences by Data Mining from Physicochemical Properties , journal =. 2009 , volume =

    2009

  68. [76]

    Ghahramani, Zoubin , year =

  69. [77]

    2013 , howpublished =

    Rana, Prashant , title =. 2013 , howpublished =

    2013

  70. [78]

    2020 , howpublished =

    Biewald, Lukas , title =. 2020 , howpublished =

    2020

Pith tools

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