Pith. sign in

REVIEW 3 major objections 5 minor 48 references

Last Layer Empirical Bayes

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 ↗

Pith's one-line read Last-layer empirical Bayes learns a normalizing-flow prior over final-layer weights and shows that a tractable empirical-Bayes objective can match, but not beat, existing uncertainty-quantification baselines.

desk verdict A tidy, honest workshop paper with a real novelty and a real theory gap: the objective does not identify q*, so LLEB's diversity and empirical-Bayes claims rest on the NF's implicit regularization, not on the stated derivation. read the letter →

arxiv 2505.15888 v1 pith:BOPRBYN3 submitted 2025-05-21 cs.LG cs.AIstat.ML

Valentin Villecroze , Yixin Wang , Gabriel Loaiza-Ganem This is my paper · ORCID
classification cs.LGcs.AIstat.ML
keywords lastlayerempiricalBayesuncertaintyquantificationnormalizingflowsBayesianneuralnetworksdeepensemblesvariationalinferenceout-of-distributiondetectionepistemic
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

The paper proposes last layer empirical Bayes (LLEB), a tractable method for neural-network uncertainty quantification that learns a normalizing-flow prior over only the final-layer weights. After a maximum-likelihood fit of the network, the flow is trained to maximize the expected log-likelihood over those last-layer weights, so the resulting distribution can be read as both a learned prior and its posterior. This places LLEB between standard Bayesian neural networks, which use weak fixed priors, and deep ensembles, whose implicit prior is a mixture of point masses at maximum-likelihood solutions. On MNIST, Fashion-MNIST, CIFAR-10, and SVHN, LLEB performs on par with existing uncertainty methods of similar computational cost but does not consistently beat them; the paper attributes that gap to concessions made for tractability rather than to the empirical-Bayes idea itself. The contribution is therefore a demonstration that empirical Bayes is a viable direction for uncertainty quantification, not yet a method that outperforms current baselines.

What carries the argument

The machinery is a normalizing flow on the last layer used as a learnable prior. A normalizing flow defines a density $q_\eta$ as the density of $f_\eta(Z)$ for an invertible network $f_\eta$ and a simple base distribution such as an isotropic Gaussian; restricting the flow to the final linear layer keeps the expected-likelihood objective low-dimensional and tractable. The flow's invertibility is load-bearing: the paper argues it acts as an implicit regularizer that prevents the optimized distribution from collapsing onto a point mass while still letting it concentrate around the maximum-likelihood set $\Theta^*$. The two-step training scheme—fit all weights by maximum likelihood, then optimize the flow at fixed remaining weights—is what makes the method stable on larger classifiers such as ResNet18.

What would settle it

Train LLEB on a fixed dataset with an invertible flow of varying capacity and measure both the entropy of the learned $q_\eta$ and how much of its mass lies near the maximum-likelihood set $\Theta^*$; if a sufficiently flexible invertible flow still collapses to a point mass, the claimed regularizing role of invertibility would be refuted. Conversely, the paper's Appendix B ablation predicts that replacing the flow with a fully connected, non-invertible density estimator of comparable flexibility causes collapse, so a non-invertible estimator that does not collapse in the same setting would also falsify the mechanism as stated.

Watch

Extended reading notes

Core claim

The central claim is that an explicitly learned, data-dependent prior on the last layer can serve as a tractable bridge between weak-prior Bayesian neural networks and strong-prior ensembles. Concretely, after maximum-likelihood training of all weights, LLEB maximizes $\mathbb{E}_{\theta_{\mathrm{QU}}\sim q_\eta}[\log p(\mathcal{D}\mid\theta_{\mathrm{QU}}, \theta^*_{\mathrm{NU}})]$ over flow parameters $\eta$, with $\theta_{\mathrm{QU}}$ the last-layer weights and $\theta^*_{\mathrm{NU}}$ the frozen remaining weights. When the learnable prior is flexible enough, maximizing the ELBO over both $q$ and the prior yields a $q^*$ that equals its own prior and its own posterior and concentrates mass around the maximum-likelihood set $\Theta^*$; the flow's invertibility is the mechanism the paper invokes to keep $q^*$ from collapsing to a point mass. The empirical finding is that this construction matches, but does not significantly or consistently outperform, baselines of similar cost, which the paper attributes to the approximations made for tractability.

Load-bearing premise

The load-bearing premise is that an invertible transformation over last-layer weights keeps the learned distribution from collapsing to a single point while still concentrating near good weights, a property the paper supports with one ablation rather than a formal argument.

Editorial extensions

If this is right

  • A trained LLEB model gives a predictive distribution $\mathbb{E}_{\theta\sim q^*}[p(\cdot\mid x, \theta)]$ and an epistemic-uncertainty estimate from the variance over $\theta\sim q^*$, without training multiple models.
  • The learned $q^*$ can be interpreted simultaneously as the prior and the posterior, so LLEB interpolates between fixed-weak-prior BNNs and ensemble point masses in terms of prior strength.
  • If the authors' hypothesis about tractability is right, future methods that keep the empirical-Bayes objective but remove the last-layer restriction or use a richer tractable family could plausibly outperform existing uncertainty methods.
  • Because a non-invertible fully connected alternative collapsed to a point mass in the paper's ablation, invertibility appears necessary for this particular construction to retain diversity.
  • Ensembles remain more expensive to train than single-model methods, so even matching their uncertainty performance at lower cost would be a concrete practical gain.

Reading between the lines

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

  • One implication the authors leave implicit is that the last-layer restriction, not the empirical-Bayes objective, may be the main thing standing between LLEB and ensemble-level performance; a low-rank or partially Bayesian treatment of earlier layers could be the natural next test.
  • The paper's data hint that prior strength could be tuned continuously between a diffuse Gaussian and an ensemble-like point mass, but it does not systematically sweep that dial; measuring calibration and OOD detection along such a sweep would be a direct test of the paper's motivating hypothesis.
  • The fully connected collapse ablation is suggestive but is a single experiment on a particular architecture; a principled characterization of when invertibility prevents variational collapse would separate the method's insight from its specific implementation.
  • LLEB's on-par results on four datasets suggest empirical Bayes is competitive, but the method has not been tested on larger-scale or regression tasks, where the relative value of a learned prior could differ.
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

3 major / 5 minor

Summary. The paper proposes Last Layer Empirical Bayes (LLEB), an uncertainty-quantification method in which a normalizing flow over the last-layer weights is trained to maximize E_{theta_QU ~ q_eta}[log p(D | theta_QU, theta_NU*)] after the remaining weights have been fit by maximum likelihood. The authors argue that this objective is motivated by an empirical-Bayes view of deep ensembles, and that the flow's invertibility acts as an implicit regularizer preventing collapse to a point mass while keeping diversity. Experiments on MNIST/Fashion-MNIST and CIFAR-10/SVHN compare LLEB against a deterministic network, last-layer Laplace, and Monte Carlo dropout, both as single models and as ensembles. The results show LLEB performing on par with, but not consistently better than, the baselines of comparable cost; the paper explicitly frames this as evidence for the promise of the empirical-Bayes direction rather than as a new state-of-the-art method.

Significance. If the empirical-Bayes interpretation is accepted, LLEB is a novel and tractable instantiation of the idea that learned data-dependent priors can replace both fixed weak priors and expensive ensembling. The paper is commendably transparent: it reports on-par results, discusses failure modes, provides code, and includes an ablation of alternatives in Appendix B. The central weakness is that the objective does not uniquely determine the learned distribution q*, so the method's UQ behavior depends on unanalyzed properties of the normalizing-flow family. This does not invalidate the empirical results, but it limits the strength of the theoretical motivation and leaves the interpolation claim between BNNs and ensembles unsupported. The contribution is therefore interesting but requires additional analysis to be fully convincing.

major comments (3)
  1. [Section 3 and Appendix A] The objective actually maximized by LLEB is E_{theta_QU~q_eta}[log p(D|theta_QU, theta_NU*)] after a maximum-likelihood fit, with no KL term. The abstract and Introduction describe the method as maximizing the ELBO, which is inaccurate. The connection to the ELBO in Appendix A holds only when maximizing jointly over q and an arbitrarily flexible prior pi, and in that case any q supported on Theta* attains the maximum. For a fixed normalizing-flow family, the objective is flat over the set of flow parameters that place mass on Theta*, so the learned spread of q_eta* is selected by NF initialization, architecture, and optimization rather than by the empirical-Bayes objective. This is load-bearing because LLEB's motivation rests on q* being a meaningful posterior-like distribution that concentrates around Theta* while remaining diverse. The ablation in Appendix B (fully-connected parameterization collapses) shows that invertibility prevents collapse, but it does not establish that the resulting spread is meaningful or stable. I recommend adding an analysis of how the NF's inductive bias selects a distribution over Theta*, together with empirical diagnostics (e.g., sensitivity to NF hyperparameters/initialization, variance of samples from q*, coverage of Theta*), or an explicit regularizer with a principled justification.
  2. [Section 3 and Appendix C] The two-step training procedure (maximum-likelihood fit for theta_NU, then maximize E_q log p over eta with theta_NU frozen) is presented as the method, but Appendix C states that the MNIST/Fashion-MNIST experiments use the end-to-end objective while only CIFAR-10/SVHN use two-step training. The paper does not compare the two variants, so it is unclear whether the two-step objective preserves the behavior of the end-to-end objective, and whether the 'concessions for tractability' mentioned in the conclusion include the two-step approximation. A direct comparison on a small dataset of the learned q* and downstream UQ metrics between end-to-end and two-step training would clarify this point.
  3. [Section 3, motivation] The paper claims that LLEB 'interpolates between standard BNNs and ensembles in terms of the strength of the prior that they use,' but no formal definition of prior strength is provided, and the claim is not empirically verified. A quantitative measure (e.g., entropy of q*, KL divergence from a reference Gaussian prior, or the effective number of modes) would make the interpolation claim concrete and testable. Without such a measure, the statement is informal and cannot be assessed.
minor comments (5)
  1. [Abstract and Section 1] The phrase 'trained to maximize the evidence lower bound' is not accurate for the objective in Section 3; please rephrase to 'maximize an expected-log-likelihood objective' or explicitly state the ELBO connection with the caveat that no KL term appears.
  2. [Appendix A] The statement that 'it is well known from variational inference that maximizing the ELBO will result in the variational posterior matching the true posterior' overstates the condition: this holds only when the variational family Q contains the true posterior, which is not the case here since Q is restricted to normalizing flows.
  3. [Tables 1 and 2] For the Default method, the AUC column is reported as '–'. It would be helpful to state explicitly that the deterministic baseline produces no epistemic uncertainty and therefore is not evaluated with that metric.
  4. [References] The paper relies heavily on Loaiza-Ganem et al. (2025), which is an unreviewed preprint, for the central theoretical equivalence. Please clarify the status of this work and, if possible, include the formal derivation in an appendix or point to a peer-reviewed version.
  5. [Conclusion] The closing hypothesis that LLEB's lack of improvement is due to 'concessions we made for tractability' would be more informative if supported by a specific comparison (e.g., end-to-end versus two-step, or NF capacity) that isolates which concession matters most.

Circularity Check

1 steps flagged · score 4.0 of 10

LLEB's empirical-Bayes interpretation is a definitional tautology: any distribution over the maximum-likelihood set is both 'prior' and 'posterior', so the objective does not select the claimed posterior; the method's experimental results are independent.

  1. self definitional [Section 3, paragraph 2; see also Appendix A]
    "With these motivations in mind, we first consider simply maximizing E_{θ∼q}[log p(D |θ)] over q∈Q; this will produce the same optimal q* as maximizing ELBO(q, π) over q and π under a flexible enough π. Furthermore, if Q is flexible enough, the resulting q* will assign probability 1 to Θ*, and could thus be interpreted as both a prior and its corresponding posterior, just like in ensembles."

    The joint ELBO optimum is degenerate: once π=q, the ELBO equals E_q log p, and any q supported on Θ* attains the same maximum. For any such q, setting the prior π=q makes the posterior π(·|D) ∝ q(θ)p(D|θ) equal to q on Θ* because p(D|θ) is constant there. Therefore 'q* is both a learned prior and its posterior' is not a consequence of the objective selecting a posterior; it is an identity forced by defining the prior to equal q. The objective contains no term that rewards diversity or picks among distributions on Θ*, so the claimed empirical-Bayes motivation reduces to a definitional tautology. The actual spread of q* is determined by NF invertibility, initialization, and capacity, not by the derivation.

full rationale

The paper is not circular in its experimental core: LLEB is defined by an explicit objective (maximizing E_q log p after an ML fit), and its accuracy, ECE, and OOD AUC are measured against held-out in-distribution and out-of-distribution data, so there is no fitted-then-called-prediction loop. However, the claimed empirical-Bayes motivation is self-definitional. Because the joint ELBO optimum sets π=q, the objective reduces to E_q log p, which is constant over every distribution supported on the maximum-likelihood set Θ*. Any such q, with prior defined as q, has posterior equal to q since the likelihood is constant on Θ*; hence 'q* is both a learned prior and its posterior' and 'LLEB is backed by empirical Bayes' are identities forced by the definitions, not consequences of data or of a posterior-selection principle. The NF's invertibility is what actually determines the spread, and the Appendix B ablation only shows that a fully-connected parameterization collapses, not that the NF's spread is the posterior-like object claimed. The paper's motivational premise about ensembles performing empirical Bayes is also imported from a same-author citation (Loaiza-Ganem et al. 2025), though the underlying ELBO result is simple and sketched in Appendix A, so this self-citation is not the main source of circularity. Overall, the method and benchmarks stand independently, but the central interpretation is circular/tautological, giving score 4.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the flow's non-collapse behavior, the self-cited variational identity, and the two-step training equivalence; no invented entities are introduced.

free parameters (2)
  • NF architecture hyperparameters = 2 coupling layers, 100 hidden features, 2 residual blocks, 11 bins, tail bound 10
    Chosen by hand in Table 4 with no sensitivity analysis; they determine the flexibility of the learned q and the claimed concentration/diversity trade-off.
  • number of Monte Carlo samples = 10 (training and test)
    Chosen in Appendix C; affects the accuracy of the expectation estimates in the training objective and predictive distribution.
assumptions (4)
  • domain assumption The likelihood p(D|θ) attains a maximum, and the maximizer set Θ* is non-empty.
    Section 2 states this to define Θ*, on which the ensembles-as-empirical-Bayes argument and the motivation for concentrating q around Θ* depend.
  • standard math For flexible enough q and π, maximizing the ELBO over both yields q*=π*=posterior with q* assigning probability 1 to Θ*.
    Used in Section 2 and Appendix A to justify treating the maximizing distribution as both prior and posterior; the proof is only sketched and deferred to the self-cited Loaiza-Ganem et al. (2025).
  • ad hoc to paper A normalizing flow's invertibility prevents collapse onto a point mass and promotes enough diversity around Θ*.
    Section 3 relies on this to claim LLEB concentrates mass without collapsing; the evidence is empirical (Appendix B: replacing the NF with a fully-connected net collapses) and is not a formal property.
  • ad hoc to paper The two-step training procedure (fit θ_NU by ML, freeze it, then fit q_η) preserves the behavior of the end-to-end objective.
    Section 3 says end-to-end was unstable for larger classifiers and the two-step strategy was faster and more stable; no analysis shows it finds the same optimum.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Last Layer Empirical Bayes." pith.science (2026). https://pith.science/paper/BOPRBYN3

@misc{pith2026250515888,
  author       = {Pith},
  title        = {Pith review of: Last Layer Empirical Bayes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BOPRBYN3}},
  note         = {Machine review of arXiv:2505.15888}
}
read the original abstract

The task of quantifying the inherent uncertainty associated with neural network predictions is a key challenge in artificial intelligence. Bayesian neural networks (BNNs) and deep ensembles are among the most prominent approaches to tackle this task. Both approaches produce predictions by computing an expectation of neural network outputs over some distribution on the corresponding weights; this distribution is given by the posterior in the case of BNNs, and by a mixture of point masses for ensembles. Inspired by recent work showing that the distribution used by ensembles can be understood as a posterior corresponding to a learned data-dependent prior, we propose last layer empirical Bayes (LLEB). LLEB instantiates a learnable prior as a normalizing flow, which is then trained to maximize the evidence lower bound; to retain tractability we use the flow only on the last layer. We show why LLEB is well motivated, and how it interpolates between standard BNNs and ensembles in terms of the strength of the prior that they use. LLEB performs on par with existing approaches, highlighting that empirical Bayes is a promising direction for future research in uncertainty quantification.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 33 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Fieguth, Xiaochun Cao, Abbas Khosravi, U

    Moloud Abdar, Farhad Pourpanah, Sadiq Hussain, Dana Rezazadegan, Li Liu, Mohammad Ghavamzadeh, Paul W. Fieguth, Xiaochun Cao, Abbas Khosravi, U. Rajendra Acharya, Vladimir Makarenkov, and Saeid Nahavandi. A review of uncertainty quantification in deep learning: Techniques, applications and challenges. Information Fusion, 76: 0 243--297, 2021

    work page 2021

  3. [3]

    Variational inference: A review for statisticians

    David M Blei, Alp Kucukelbir, and Jon D McAuliffe. Variational inference: A review for statisticians. Journal of the American statistical Association, 112 0 (518): 0 859--877, 2017

    2017

  4. [4]

    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, 2015

    2015

  5. [5]

    End to end learning for self-driving cars

    Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Prasoon Goyal, Lawrence D Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, Xin Zhang, Jake Zhao, and Karol Zieba. End to end learning for self-driving cars. arXiv:1604.07316, 2016

    arXiv 2016

  6. [6]

    Stochastic gradient hamiltonian monte carlo

    Tianqi Chen, Emily Fox, and Carlos Guestrin. Stochastic gradient hamiltonian monte carlo. In International Conference on Machine Learning, 2014

    work page 2014

  7. [7]

    Laplace redux-effortless B ayesian deep learning

    Erik Daxberger, Agustinus Kristiadi, Alexander Immer, Runa Eschenhagen, Matthias Bauer, and Philipp Hennig. Laplace redux-effortless B ayesian deep learning. In Advances in Neural Information Processing Systems, 2021

    work page 2021

  8. [8]

    NICE : Non-linear independent components estimation

    Laurent Dinh, David Krueger, and Yoshua Bengio. NICE : Non-linear independent components estimation. In ICLR Workshop Track, 2015

    work page 2015

Show all 48 references
  1. [9]

    Neural spline flows

    Conor Durkan, Artur Bekasov, Iain Murray, and George Papamakarios. Neural spline flows. In Advances in Neural Information Processing Systems, 2019

    2019

  2. [10]

    nflows : normalizing flows in PyTorch , November 2020

    Conor Durkan, Artur Bekasov, Iain Murray, and George Papamakarios. nflows : normalizing flows in PyTorch , November 2020. URL https://doi.org/10.5281/zenodo.4296287

    2020 doi

  3. [11]

    Dermatologist-level classification of skin cancer with deep neural networks

    Andre Esteva, Brett Kuprel, Roberto A Novoa, Justin Ko, Susan M Swetter, Helen M Blau, and Sebastian Thrun. Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542 0 (7639): 0 115--118, 2017

    2017

  4. [12]

    Dropout as a B ayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a B ayesian approximation: Representing model uncertainty in deep learning. In International Conference on Machine Learning, 2016

    2016

  5. [13]

    Practical variational inference for neural networks

    Alex Graves. Practical variational inference for neural networks. In Advances in Neural Information Processing Systems, 2011

    2011

  6. [14]

    Variational B ayesian last layers

    James Harrison, John Willes, and Jasper Snoek. Variational B ayesian last layers. In International Conference on Learning Representations, 2024

    2024

  7. [15]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016

    2016

  8. [16]

    Probabilistic backpropagation for scalable learning of B ayesian neural networks

    Jos \'e Miguel Hern \'a ndez-Lobato and Ryan Adams. Probabilistic backpropagation for scalable learning of B ayesian neural networks. In International Conference on Machine Learning, 2015

    2015

  9. [17]

    Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods

    Eyke H \"u llermeier and Willem Waegeman. Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods. Machine Learning, 110 0 (3): 0 457--506, 2021

    2021

  10. [18]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015

    2015

  11. [19]

    Auto-encoding variational Bayes

    Diederik P Kingma and Max Welling. Auto-encoding variational Bayes . In International Conference on Learning Representations, 2014

    2014

  12. [20]

    Improved variational inference with inverse autoregressive flow

    Diederik P Kingma, Tim Salimans, Rafal Jozefowicz, Xi Chen, Ilya Sutskever, and Max Welling. Improved variational inference with inverse autoregressive flow. In Advances in Neural Information Processing Systems, 2016

    2016

  13. [21]

    Being B ayesian, even just a bit, fixes overconfidence in relu networks

    Agustinus Kristiadi, Matthias Hein, and Philipp Hennig. Being B ayesian, even just a bit, fixes overconfidence in relu networks. In International Conference on Machine Learning, 2020

    2020

  14. [22]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009

    2009

  15. [23]

    Simple and scalable predictive uncertainty estimation using deep ensembles

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In Advances in Neural Information Processing Systems, 2017

    2017

  16. [24]

    Marginalized neural network mixtures for large-scale regression

    Miguel L \'a zaro-Gredilla and An \' bal R Figueiras-Vidal. Marginalized neural network mixtures for large-scale regression. IEEE transactions on neural networks, 21 0 (8): 0 1345--1351, 2010

    2010

  17. [25]

    Gradient-based learning applied to document recognition

    Yann LeCun, L \'e on Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86 0 (11): 0 2278--2324, 1998

    1998

  18. [26]

    A survey on deep learning in medical image analysis

    Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Ginneken, and Clara I S \'a nchez. A survey on deep learning in medical image analysis. Medical Image Analysis, 42: 0 60--88, 2017

    2017

  19. [27]

    Maximum entropy flow networks

    Gabriel Loaiza-Ganem, Yuanjun Gao, and John P Cunningham. Maximum entropy flow networks. In International Conference on Learning Representations, 2017

    2017

  20. [28]

    Deep ensembles secretly perform empirical bayes

    Gabriel Loaiza-Ganem, Valentin Villecroze, and Yixin Wang. Deep ensembles secretly perform empirical bayes. arXiv:2501.17917, 2025

    2025 arXiv

  21. [29]

    Structured and efficient variational deep learning with matrix G aussian posteriors

    Christos Louizos and Max Welling. Structured and efficient variational deep learning with matrix G aussian posteriors. In International Conference on Machine Learning, 2016

    2016

  22. [30]

    Multiplicative normalizing flows for variational B ayesian neural networks

    Christos Louizos and Max Welling. Multiplicative normalizing flows for variational B ayesian neural networks. In International Conference on Machine Learning, 2017

    2017

  23. [31]

    Reading digits in natural images with unsupervised feature learning

    Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011

    2011

  24. [32]

    Practical deep learning with B ayesian principles

    Kazuki Osawa, Siddharth Swaroop, Mohammad Emtiyaz E Khan, Anirudh Jain, Runa Eschenhagen, Richard E Turner, and Rio Yokota. Practical deep learning with B ayesian principles. In Advances in Neural Information Processing Systems, 2019

    2019

  25. [33]

    Uncertainty quantification in scientific machine learning: Methods, metrics, and comparisons

    Apostolos F Psaros, Xuhui Meng, Zongren Zou, Ling Guo, and George Em Karniadakis. Uncertainty quantification in scientific machine learning: Methods, metrics, and comparisons. Journal of Computational Physics, 477: 0 111902, 2023

    2023

  26. [34]

    Variational inference with normalizing flows

    Danilo Jimenez Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International Conference on Machine Learning, 2015

    2015

  27. [35]

    Stochastic backpropagation and approximate inference in deep generative models

    Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In International Conference on Machine Learning, 2014

    2014

  28. [36]

    A scalable L aplace approximation for neural networks

    Hippolyt Ritter, Aleksandar Botev, and David Barber. A scalable L aplace approximation for neural networks. In International Conference on Learning Representations, 2018

    2018

  29. [37]

    Dropout: a simple way to prevent neural networks from overfitting

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 15 0 (1): 0 1929--1958, 2014

    1929

  30. [38]

    Graphical models, exponential families, and variational inference

    Martin J Wainwright and Michael I Jordan. Graphical models, exponential families, and variational inference. Foundations and Trends in Machine Learning, 1 0 (1--2): 0 1--305, 2008

    2008

  31. [39]

    Neural linear models with functional G aussian process priors

    Joe Watson, Jihao Andreas Lin, Pascal Klink, and Jan Peters. Neural linear models with functional G aussian process priors. In Third Symposium on Advances in Approximate Bayesian Inference, 2020

    2020

  32. [40]

    Latent derivative B ayesian last layer networks

    Joe Watson, Jihao Andreas Lin, Pascal Klink, Joni Pajarinen, and Jan Peters. Latent derivative B ayesian last layer networks. In International Conference on Artificial Intelligence and Statistics, 2021

    2021

  33. [41]

    Bayesian learning via stochastic gradient langevin dynamics

    Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics. In International Conference on Machine Learning, 2011

    2011

  34. [42]

    Deterministic variational inference for robust B ayesian neural networks

    Anqi Wu, Sebastian Nowozin, Edward Meeds, Richard E Turner, Jos \'e Miguel Hern \'a ndez-Lobato, and Alexander L Gaunt. Deterministic variational inference for robust B ayesian neural networks. In International Conference on Learning Representations, 2019

    2019

  35. [43]

    Fashion-MNIST: A novel image dataset for benchmarking machine learning algorithms

    Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-MNIST: A novel image dataset for benchmarking machine learning algorithms . arXiv:1708.07747, 2017

    2017 arXiv

  36. [44]

    Bayesian low-rank adaptation for large language models

    Adam X Yang, Maxime Robeyns, Xi Wang, and Laurence Aitchison. Bayesian low-rank adaptation for large language models. In International Conference on Learning Representations, 2024

    2024

  37. [45]

    Cyclical stochastic gradient MCMC for B ayesian deep learning

    Ruqi Zhang, Chunyuan Li, Jianyi Zhang, Changyou Chen, and Andrew Gordon Wilson. Cyclical stochastic gradient MCMC for B ayesian deep learning. In International Conference on Learning Representations, 2020

    2020

  38. [46]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  39. [47]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  40. [48]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

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