Pith. sign in

REVIEW 2 major objections 5 minor 300 references

Temperature scaling on the eigenvalues of LLM answer embeddings turns overconfident density matrices into calibrated uncertainty estimates whose average entropy matches risk.

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 ↗

T0 review · grok-4.5

2026-07-10 08:26 UTC pith:APVF2C7O

load-bearing objection Solid theory for eigenvalue calibration of LLM density-matrix predictors; theorems hold and experiments match, with only the expected finite-sample gap. the 2 major comments →

arxiv 2607.08377 v1 pith:APVF2C7O submitted 2026-07-09 cs.LG

Eigenvalue Calibration for Semantic Embeddings of Large Language Models

Sebastian G. Gruber , Nassim Walha , Francis Bach , Florian Buettner This is my paper
classification cs.LG
keywords eigenvalue calibrationdensity matrix predictorssemantic embeddingstemperature scalingproper matrix scoresLLM uncertainty quantificationvon Neumann entropy
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.

When a language model answers the same question many times, the outer products of the answers' semantic embeddings form a density matrix whose eigenvalues act as probabilities over latent semantic modes. Conventional probability calibration does not apply to those eigenvalues. This paper defines matrix calibration for the density-matrix predictor itself, proves that under that calibration the expected von Neumann entropy equals the proper-score risk, and shows that ordinary temperature scaling applied only to the eigenvalues is an injective map that therefore optimizes the proper matrix calibration error whenever the temperature is chosen by risk minimization. Real-model experiments confirm that modern LLMs are systematically overconfident (optimal temperatures all exceed one) and that the recalibrated eigenvalues become reliable enough for entropy to track risk and for correctness detection to improve. The result supplies both a theoretical foundation and a practical post-hoc fix for eigenvalue-based uncertainty methods already used with large language models.

Core claim

For density-matrix predictors built from LLM answer embeddings, temperature scaling of the eigenvalues is injective; therefore the temperature that minimizes a proper matrix score risk also minimizes the proper matrix calibration error, and at that temperature expected von Neumann entropy converges to risk.

What carries the argument

Proper matrix scores and the associated calibration-sharpness decomposition: an injective post-hoc map (matrix temperature scaling) changes only the calibration term of the risk, so risk minimization is calibration optimization.

Load-bearing premise

That a few dozen sampled answers already give a density-matrix estimate faithful enough for the continuous proper-score identities and the injectivity argument to hold in practice.

What would settle it

On a held-out question-answering set, if the temperature that minimizes the matrix log risk fails to reduce the matrix calibration error (or the eigenvalue reliability gap) relative to the unscaled predictor, the central optimization claim is false.

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

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 treats LLMs plus semantic embeddings of sampled answers as density-matrix predictors and defines matrix calibration (predicted density matrix equals the conditional density matrix of the target embeddings) together with a weaker eigenvalue calibration. It introduces proper matrix scores, proves that matrix calibration implies equality of expected von Neumann entropy and proper-score risk (Lemma 1 / Theorem 2), derives an eigenvalue calibration inequality that follows from spectral-norm convexity (Theorem 1), and shows that any injective post-hoc map—in particular matrix temperature scaling of the eigenvalues—changes only the proper calibration error term of the risk (Theorem 3 + Proposition 3). Experiments on TriviaQA and Natural Questions with three open LLMs confirm that risk-optimal temperatures are systematically greater than one, that entropy and risk coincide at that temperature, and that the resulting recalibration reduces eigenvalue ECE and often improves AUROC for answer-correctness detection. Reliability diagrams that condition on the full predicted matrix via hierarchical clustering (Algorithm 1) make the over-confidence visually transparent.

Significance. If the finite-sample gap is accepted as ordinary estimation error, the work supplies the first rigorous calibration theory for the eigenvalue-based uncertainty measures already used in state-of-the-art LLM papers. The entropy–risk equivalence under matrix calibration, the injectivity argument for temperature scaling, and the explicit reliability-diagram procedure are concrete, reusable tools. The consistent empirical finding that modern LLMs are over-confident in their maximum eigenvalues is immediately actionable. Strengths include fully written proofs (Appendix C), public code, and a clear separation of development-set temperature fitting from test-set evaluation.

major comments (2)
  1. Section 4 and Appendix B.1: all theory is stated for the population density-matrix predictor d, yet every experiment (temperature fitting, risk/entropy curves, Algorithm 1 diagrams) uses the empirical estimator formed from m=20 samples drawn at temperature 0.5. Appendix B.1 already documents a positive bias in the maximum-eigenvalue estimator that only shrinks for m≥20. The continuous proper-score identities and the injectivity argument therefore hold only approximately for the objects that are actually optimized and plotted. A short quantitative bound (or an ablation with m=50–100) on how large this gap remains for the reported ECE and AUROC numbers would make the central claim fully load-bearing rather than asymptotic.
  2. Definition 1 and Theorem 1: eigenvalue calibration is defined by conditioning only on Λ_X = λ_max(d(X)), yet the paper correctly notes that matrix calibration yields only the inequality λ_max(D_Y|Λ_X) ≤ Λ_X. Algorithm 1 therefore resorts to hierarchical clustering inside bins to approximate conditioning on the full matrix. The resulting diagrams are informative, but the formal link between the plotted quantity E[λ_max(D_Y|d(X))|Λ_X] and the definition of eigenvalue calibration is left somewhat informal; a one-sentence statement that Algorithm 1 estimates the left-hand side of Eq. (11) rather than Eq. (10) would remove any ambiguity.
minor comments (5)
  1. Figure 3 caption and surrounding text: the grey dashed line at T=1 is helpful, but the optimal temperatures themselves are never tabulated; a small table of the six risk-optimal T values would aid reproducibility.
  2. Algorithm 1: the hierarchical clustering step uses the correlation matrix of the vectorized density matrices; a brief remark on why correlation (rather than Frobenius distance) is preferred would clarify the design choice.
  3. Table 1: standard errors are reported from B=20 bootstrap subsets; stating the exact bootstrap procedure (with or without replacement of the 20 answers) would remove a minor ambiguity.
  4. Notation: the same symbol S is used both for the proper matrix score and for the induced scalar proper score; a subscript or different font would avoid occasional confusion in Appendix C.
  5. Related-work paragraph: the recent sampling-temperature calibration baseline of Lamb et al. (2025) is correctly compared in Appendix B.6, but a one-sentence pointer in the main text would help readers locate the comparison.

Circularity Check

0 steps flagged

No significant circularity: proper-score identities and injectivity of matrix temperature scaling are derived self-containedly; T is fitted on held-out risk and validated on separate data.

full rationale

The derivation chain (Defs. 1–2, Props. 1–3, Lemmas 1–2, Thms. 1–3) proceeds from the standard definition of matrix calibration (DY|d(X)=d(X) a.s.) and the definition of a proper matrix score, through the induced Bregman divergence and the calibration-sharpness decomposition, to the claim that any injective post-hoc map (in particular spectral temperature scaling h_TS) changes only the calibration term of the risk. All steps are proved directly in Appendix C from these definitions plus convexity of the spectral norm; they do not redefine calibration in terms of the fitted temperature, nor do they import a uniqueness theorem that forces the result. Temperature is chosen by minimizing the matrix log risk on a development split of 300 examples and then evaluated (entropy–risk match, ECE, AUROC) on held-out test data; the observed coincidence of average entropy and risk at the risk-optimal T is therefore an empirical confirmation of Thm. 2, not a quantity forced by construction. Self-citations to Gruber & Buettner (2022) and Gruber (2024) supply background proper-score facts that are independently published and are re-proved or specialized here; they are not load-bearing for the central claims. The only residual approximation is the finite-sample density-matrix estimator (m=20), already flagged by the reader and not a circularity. Hence score 1 (minor self-citation background only).

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

The central claims rest on standard proper-score and convex-analysis facts plus the modeling choice that an LLM-plus-embedding is a density-matrix predictor. One free parameter (temperature) is fitted per model/dataset; no new physical entities are postulated.

free parameters (3)
  • temperature T = 1/α = dataset- and model-dependent (e.g. T=2.51 for Phi-4 Mini on TriviaQA)
    Fitted on a 300-example development set by minimizing matrix log-risk; optimal values reported >1 for every model/dataset pair.
  • number of sampled answers m = 20 (eval) / 100 (dev)
    Chosen as m=20 for evaluation and m=100 for development; controls bias/variance of the empirical density-matrix estimator.
  • bin count B and cluster count C in Algorithm 1 = B=8, C=5 (default)
    Hyper-parameters of the reliability-diagram procedure; sensitivity table shows stability for C≥5.
axioms (4)
  • standard math A proper matrix score induces a concave entropy and a Bregman matrix divergence (Proposition 2).
    Follows from classical proper-score theory (Gneiting & Raftery, Ovcharov) once the score is restricted to density matrices.
  • standard math The spectral norm (maximum eigenvalue) is convex, so Jensen yields the eigenvalue calibration inequality.
    Standard fact from convex optimization (Boyd & Vandenberghe); used in the proof of Theorem 1.
  • domain assumption An LLM together with a semantic embedding can be identified with the density-matrix predictor d(x)=E[e(a)e(a)⊺] (Eq. 8).
    Modeling choice that lets the authors import density-matrix language; empirically motivated by prior embedding-based UQ work but not forced by first principles.
  • ad hoc to paper The continuous proper-score identities continue to hold for the finite-sample empirical density matrices used in practice.
    Required to transfer Theorems 2–3 from population quantities to the m=20 estimators actually optimized and plotted.
invented entities (2)
  • proper matrix score independent evidence
    purpose: Extends classical proper scores to density-matrix arguments so that risk, entropy and calibration error can be defined for embeddings.
    Definition 2 is new relative to the cited literature; it is a natural restriction rather than an exotic postulate.
  • matrix / eigenvalue calibration independent evidence
    purpose: Provides the target notions of calibration that ordinary probability calibration does not automatically imply for eigenvalues.
    Definition 1; the inequality of Theorem 1 shows the notions are distinct.

pith-pipeline@v1.1.0-grok45 · 22130 in / 2829 out tokens · 23952 ms · 2026-07-10T08:26:32.887164+00:00 · methodology

0 comments
read the original abstract

Uncertainty quantification is central to the reliable deployment of large language models (LLMs), and eigenvalues of semantic embeddings have recently emerged as a key tool in state-of-the-art methods. However, conventional calibration results developed for classification probabilities cannot be directly transferred to eigenvalues. We address this gap by proposing a novel framework for calibrating the eigenvalues of semantic embeddings. We interpret LLMs combined with semantic embeddings of their generated answers as density matrix predictors, and we propose a novel approach to calibrate density matrix predictors by applying temperature scaling to their eigenvalues. We establish entropy-risk equivalence under calibration, derive a central calibration inequality specific to eigenvalues, and prove that temperature-scaled eigenvalues optimize calibration when minimizing proper score risks. Experiments on a variety of real-world settings show that current LLMs are systematically overconfident, and validate our theoretical findings. Together, these results advance the foundations and practice of uncertainty quantification for semantic embeddings.

Figures

Figures reproduced from arXiv: 2607.08377 by Florian Buettner, Francis Bach, Nassim Walha, Sebastian G. Gruber.

Figure 1
Figure 1. Figure 1: Normalised semantic embeddings reside in a hy [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The matrix version of the cross entropy risk cor [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: According to Theorem 3, the optimal temperature for calibration is indicated via the minimum risk. Here, the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Reliability diagrams according to eigenvalue cali [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Reliability diagrams according to Algorithm 1 of LLMs before and after temperature scaling. Figure 5a shows that [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Estimating the maximum eigenvalue has a positive bias but converges quickly to the ground truth (red line) for [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Reliability diagrams for TriviaQA according to Algorithm 1 of LLMs before and after temperature scaling. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Reliability diagrams for conventional answer correctness. Matrix temperature scaling also has beneficial effects on [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Scaled values of the risk RS, matrix calibration error CalS, and the eigenvalue ECE (as defined in Algorithm 1). As predicted by Theorem 3, all quantities are co-minimized at essentially the same temperature (with small deviations due to estimation noise). B.4 RISK, CALIBRATION ERROR, AND EIGENVALUE ECE ARE CO-MINIMIZED By Theorem 3, minimizing the risk RS via matrix temperature scaling is equivalent to mi… view at source ↗
Figure 10
Figure 10. Figure 10: Eigenvalue-based reliability diagram (as computed by Algorithm 1) after applying the sampling temperature [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Reliability diagrams according to Algorithm 1 of LLMs before and after temperature scaling using [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] 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

300 extracted references · 300 canonical work pages · 23 internal anchors

  1. [1]

    2020 , eprint=

    Verified Uncertainty Calibration , author=. 2020 , eprint=

    work page 2020

  2. [2]

    2009 , publisher=

    Variational analysis , author=. 2009 , publisher=

    work page 2009

  3. [3]

    Information Processing Equalities and the Information-Risk Bridge

    Information Processing Equalities and the Information-Risk Bridge , author=. arXiv preprint arXiv:2207.11987 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    2008 , publisher=

    On an extension of the notion of f-divergence , journal=. 2008 , publisher=

    work page 2008

  5. [5]

    Periodica Mathematica Hungarica , volume=

    A class of measures of informativity of observation channels , author=. Periodica Mathematica Hungarica , volume=. 1972 , publisher=

    work page 1972

  6. [6]

    Conference on Learning Theory , pages=

    Divergences and risks for multiclass experiments , author=. Conference on Learning Theory , pages=. 2012 , organization=

    work page 2012

  7. [7]

    IEEE signal processing magazine , volume=

    Generative adversarial networks: An overview , author=. IEEE signal processing magazine , volume=. 2018 , publisher=

    work page 2018

  8. [8]

    A GENERALIZATION OF , author=

    f-DISSIMILARITY. A GENERALIZATION OF , author=. Ann. Inst. Statist. Math , volume=

    work page

  9. [9]

    The Annals of Statistics , volume=

    MULTICLASS CLASSIFICATION, INFORMATION, DIVERGENCE AND SURROGATE RISK , author=. The Annals of Statistics , volume=. 2018 , publisher=

    work page 2018

  10. [10]

    Brøndsted and R

    A. Brøndsted and R. T. Rockafellar , journal =. On the Subdifferentiability of Convex Functions , urldate =

    work page

  11. [11]

    Canadian Journal of Mathematics , pages=

    On Conjugate Convex Functions , author=. Canadian Journal of Mathematics , pages=. 1949 , publisher=

    work page 1949

  12. [12]

    Biometrika , volume=

    On a formula for the product-moment coefficient of any order of a normal frequency distribution in any number of variables , author=. Biometrika , volume=. 1918 , publisher=

    work page 1918

  13. [13]

    , description =

    Bregman, L.M. , description =. USSR Computational Mathematics and Mathematical Physics , keywords =

    work page

  14. [14]

    Stochastic Variational Deep Kernel Learning

    Wilson, Andrew Gordon and Hu, Zhiting and Salakhutdinov, Ruslan and Xing, Eric P. , keywords =. Stochastic Variational Deep Kernel Learning , publisher =. 2016 , copyright =. doi:10.48550/ARXIV.1611.00336 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1611.00336 2016

  15. [15]

    Advances on Neural Information Processing Systems , pages=

    Verified uncertainty calibration , author=. Advances on Neural Information Processing Systems , pages=

    work page

  16. [16]

    Reliability, sufficiency, and the decomposition of proper scores , volume=

    Bröcker, Jochen , year=. Reliability, sufficiency, and the decomposition of proper scores , volume=. Quarterly Journal of the Royal Meteorological Society , publisher=

    work page

  17. [17]

    Neural Computation , volume =

    Heskes, Tom , title = ". Neural Computation , volume =. 1998 , month =. doi:10.1162/089976698300017232 , url =

    work page doi:10.1162/089976698300017232 1998

  18. [18]

    DeGroot and Stephen E

    Morris H. DeGroot and Stephen E. Fienberg , journal =. The Comparison and Evaluation of Forecasters , volume =

    work page

  19. [19]

    International Conference on Machine Learning , pages =

    Trainable Calibration Measures for Neural Networks from Kernel Mean Embeddings , author =. International Conference on Machine Learning , pages =. 2018 , editor =

    work page 2018

  20. [20]

    Climate Dynamics , year=

    Estimating reliability and resolution of probability forecasts through decomposition of the empirical score , author=. Climate Dynamics , year=

    work page

  21. [21]

    IEEE access , volume=

    A survey of autonomous driving: Common practices and emerging technologies , author=. IEEE access , volume=. 2020 , publisher=

    work page 2020

  22. [22]

    Philosophical Transactions of the Royal Society A , volume=

    Physics-informed machine learning: case studies for weather and climate modelling , author=. Philosophical Transactions of the Royal Society A , volume=. 2021 , publisher=

    work page 2021

  23. [23]

    Science , year=

    Weather Forecasting with Ensemble Methods , author=. Science , year=

    work page

  24. [24]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Probabilistic forecasts, calibration and sharpness , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2007 , publisher=

    work page 2007

  25. [25]

    Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing , pages=

    Posterior calibration and exploratory analysis for natural language processing models , author=. Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing , pages=

    work page 2015

  26. [26]

    Murphy and Robert L

    Allan H. Murphy and Robert L. Winkler , journal =. Reliability of Subjective Probability Forecasts of Precipitation and Temperature , volume =

    work page

  27. [27]

    Physical Review Research , volume=

    Quantum mean embedding of probability distributions , author=. Physical Review Research , volume=. 2019 , publisher=

    work page 2019

  28. [28]

    Journal of Machine Learning Research , volume=

    Characteristic and universal tensor product kernels , author=. Journal of Machine Learning Research , volume=

    work page

  29. [29]

    , author=

    Universality, Characteristic Kernels and RKHS Embedding of Measures. , author=. Journal of Machine Learning Research , volume=

    work page

  30. [30]

    Advances in neural information processing systems , volume=

    A measure-theoretic approach to kernel conditional mean embeddings , author=. Advances in neural information processing systems , volume=

    work page

  31. [31]

    Learning on Graphs Conference , pages=

    A survey on deep graph generation: Methods and applications , author=. Learning on Graphs Conference , pages=. 2022 , organization=

    work page 2022

  32. [32]

    The Twelfth International Conference on Learning Representations , year=

    On Bias-Variance Alignment in Deep Models , author=. The Twelfth International Conference on Learning Representations , year=

    work page

  33. [33]

    Physical Review A , volume=

    Quantum extension of conditional probability , author=. Physical Review A , volume=. 1999 , publisher=

    work page 1999

  34. [34]

    International Conference on Machine Learning , pages=

    Hilbert space embeddings of conditional distributions with applications to dynamical systems , author=. International Conference on Machine Learning , pages=

    work page

  35. [35]

    IEEE Signal Processing Magazine , volume=

    Kernel embeddings of conditional distributions: A unified kernel framework for nonparametric inference in graphical models , author=. IEEE Signal Processing Magazine , volume=. 2013 , publisher=

    work page 2013

  36. [36]

    Warmuth, Manfred KK , journal=. A

    work page

  37. [37]

    Warmuth, Manfred K and Kuzmin, Dima , booktitle=. A

    work page

  38. [38]

    Journal of Artificial Intelligence Research , volume=

    From word to sense embeddings: A survey on vector representations of meaning , author=. Journal of Artificial Intelligence Research , volume=

    work page

  39. [39]

    A simple baseline for

    Maddox, Wesley J and Izmailov, Pavel and Garipov, Timur and Vetrov, Dmitry P and Wilson, Andrew Gordon , journal=. A simple baseline for

    work page

  40. [40]

    Logit models from economics and other fields , volume=

    The origins and development of the logit model , author=. Logit models from economics and other fields , volume=. 2003 , publisher=

    work page 2003

  41. [41]

    Rajendra Acharya and Vladimir Makarenkov and Saeid Nahavandi , doi =

    Moloud Abdar and Farhad Pourpanah and Sadiq Hussain and Dana Rezazadegan and Li Liu and Mohammad Ghavamzadeh and Paul Fieguth and Xiaochun Cao and Abbas Khosravi and U. Rajendra Acharya and Vladimir Makarenkov and Saeid Nahavandi , doi =. A review of uncertainty quantification in deep learning: Techniques, applications and challenges , url =. Information ...

    work page 2021

  42. [42]

    AAAI Conference on Artificial Intelligence , volume=

    Towards trustworthy predictions from deep neural networks with fast adversarial calibration , author=. AAAI Conference on Artificial Intelligence , volume=

    work page

  43. [43]

    International Conference on Learning Representations , year=

    Score-Based Generative Modeling through Stochastic Differential Equations , author=. International Conference on Learning Representations , year=

    work page

  44. [44]

    and Erdem, Muhammed Ebrar and Cremers, Daniel and Buettner, Florian , title =

    Tomani, Christian and Gruber, Sebastian G. and Erdem, Muhammed Ebrar and Cremers, Daniel and Buettner, Florian , title =. Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2021 , pages =

    work page 2021

  45. [45]

    2020 , eprint=

    Measuring Calibration in Deep Learning , author=. 2020 , eprint=

    work page 2020

  46. [46]

    International Conference on Learning Representations , year=

    The Deep Bootstrap Framework: Good Online Learners are Good Offline Generalizers , author=. International Conference on Learning Representations , year=

    work page

  47. [47]

    1994 , publisher=

    An Introduction to the Bootstrap , author=. 1994 , publisher=

    work page 1994

  48. [48]

    Statistical Properties of the log-cosh Loss Function Used in Machine Learning

    Saleh, Resve A. and Saleh, A. K. Md. Ehsanes , keywords =. Statistical Properties of the log-cosh Loss Function Used in Machine Learning , publisher =. 2022 , copyright =. doi:10.48550/ARXIV.2208.04564 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2208.04564 2022

  49. [49]

    Finding hidden-feature depending laws inside a data set and classifying it using Neural Network

    Finding hidden-feature depending laws inside a data set and classifying it using Neural Network , author=. arXiv preprint arXiv:2101.10427 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  50. [50]

    Entropy , volume=

    An elementary introduction to information geometry , author=. Entropy , volume=. 2020 , publisher=

    work page 2020

  51. [51]

    1985 , publisher=

    Advanced econometrics , author=. 1985 , publisher=

    work page 1985

  52. [52]

    Neural Computation , volume=

    Bias/variance decompositions for likelihood-based estimators , author=. Neural Computation , volume=. 1998 , publisher=

    work page 1998

  53. [53]

    Game theory, maximum entropy, minimum discrepancy and robust

    Gr. Game theory, maximum entropy, minimum discrepancy and robust. the Annals of Statistics , volume=. 2004 , publisher=

    work page 2004

  54. [54]

    Understanding the bias-variance tradeoff of Bregman divergences

    Understanding the bias-variance tradeoff of Bregman divergences , author=. arXiv preprint arXiv:2202.04167 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  55. [55]

    Advances in Neural Information Processing Systems , volume=

    Simple and scalable predictive uncertainty estimation using deep ensembles , author=. Advances in Neural Information Processing Systems , volume=

    work page

  56. [56]

    Attention is All you Need , volume =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N and Kaiser, ukasz and Polosukhin, Illia , booktitle =. Attention is All you Need , volume =

    work page

  57. [57]

    Advances in Neural Information Processing Systems , volume=

    Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers , author=. Advances in Neural Information Processing Systems , volume=

    work page

  58. [58]

    Song, Kaitao and Tan, Xu and Qin, Tao and Lu, Jianfeng and Liu, Tie-Yan , booktitle=

    work page

  59. [59]

    Towards General Text Embeddings with Multi-stage Contrastive Learning

    Towards General Text Embeddings with Multi-stage Contrastive Learning , author=. arXiv preprint arXiv:2308.03281 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  60. [60]

    Medical Imaging with Deep Learning , year=

    Test-time data augmentation for estimation of heteroscedastic aleatoric uncertainty in deep neural networks , author=. Medical Imaging with Deep Learning , year=

    work page

  61. [61]

    Neurocomputing , volume=

    Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks , author=. Neurocomputing , volume=. 2019 , publisher=

    work page 2019

  62. [62]

    Proceedings of the International Conference on Knowledge Discovery and Data Mining , pages=

    Xgboost: A scalable tree boosting system , author=. Proceedings of the International Conference on Knowledge Discovery and Data Mining , pages=

    work page

  63. [63]

    Machine learning , volume=

    Random forests , author=. Machine learning , volume=. 2001 , publisher=

    work page 2001

  64. [64]

    Machine learning , volume=

    Bagging predictors , author=. Machine learning , volume=. 1996 , publisher=

    work page 1996

  65. [65]

    Generalizations of the bias/variance decomposition for prediction error , author=. Dept. Statistics, Stanford Univ., Stanford, CA, Tech. Rep , year=

    work page

  66. [66]

    International Conference on Machine Learning , pages=

    A unified bias-variance decomposition , author=. International Conference on Machine Learning , pages=

    work page

  67. [67]

    Machine learning , volume=

    Variance and bias for general loss functions , author=. Machine learning , volume=. 2003 , publisher=

    work page 2003

  68. [68]

    Transactions on Machine Learning Research , year=

    Ensembles of Classifiers: a Bias-Variance Perspective , author=. Transactions on Machine Learning Research , year=

    work page

  69. [69]

    Proceedings of International Conference on Neural Networks (ICNN'96) , volume=

    Generalization error of ensemble estimators , author=. Proceedings of International Conference on Neural Networks (ICNN'96) , volume=. 1996 , organization=

    work page 1996

  70. [70]

    Mathematical geology , volume=

    Isometric logratio transformations for compositional data analysis , author=. Mathematical geology , volume=. 2003 , publisher=

    work page 2003

  71. [71]

    Advances in Neural Information Processing Systems , volume=

    Imagenet classification with deep convolutional neural networks , author=. Advances in Neural Information Processing Systems , volume=

    work page

  72. [72]

    Advances in Neural Information Processing Systems 32 , pages =

    PyTorch: An Imperative Style, High-Performance Deep Learning Library , author =. Advances in Neural Information Processing Systems 32 , pages =. 2019 , publisher =

    work page 2019

  73. [73]

    Kim-Celine Kahl and Carsten T. L. Val. International Conference on Learning Representations , year=

    work page

  74. [74]

    IEEE transactions on Information Theory , volume=

    General entropy criteria for inverse problems, with applications to data compression, pattern classification, and cluster analysis , author=. IEEE transactions on Information Theory , volume=. 1990 , publisher=

    work page 1990

  75. [75]

    The Annals of Statistics , volume=

    Why least squares and maximum entropy? An axiomatic approach to inference for linear inverse problems , author=. The Annals of Statistics , volume=. 1991 , publisher=

    work page 1991

  76. [76]

    Annals of the Institute of Statistical Mathematics , volume=

    The geometry of proper scoring rules , author=. Annals of the Institute of Statistical Mathematics , volume=. 2007 , publisher=

    work page 2007

  77. [77]

    2002 , institution=

    Duality and Auxiliary Functions for Bregman Distances (revised) , author=. 2002 , institution=

    work page 2002

  78. [78]

    International Conference on Machine Learning , pages=

    Agglomerative bregman clustering , author=. International Conference on Machine Learning , pages=

    work page

  79. [79]

    Brinda, WD and Klusowski, Jason M and Yang, Dana , journal=. H. 2019 , publisher=

    work page 2019

  80. [80]

    A bias-variance decomposition for

    Brofos, James and Shu, Rui and Lederman, Roy R , booktitle=. A bias-variance decomposition for

    work page

Showing first 80 references.