Pith. sign in

REVIEW 3 major objections 5 minor 157 references

Default VI settings can silently bias posterior estimates in major Python packages, even for simple one-dimensional conjugate models.

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 ↗

Default settings in PyMC, NumPyro, and TensorFlow Probability can yield biased or silently broken variational inference results even in simple one-dimensional conjugate models.

T0 review reviewed 2026-08-06 challenge →

load-bearing objection A useful, well-scoped empirical study of VI defaults, but the headline 5% bias result doesn't actually use the default iteration count—worth engaging, but needs revision. the 3 major comments →

arxiv 2608.01403 v1 pith:7DAFIY7Q submitted 2026-08-02 stat.CO

The Curious Case of the Default Settings: Evaluating Default Performance of Variational Inference Software

Madelyn Andersen , Elizabeth Bersson , Tamara Broderick This is my paper
classification stat.CO
keywords variational inferencedefault settingsprobabilistic programmingPyMCNumPyroTensorFlow Probabilityposterior approximationADVI
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper evaluates what happens when a typical user runs variational inference without touching any settings in three widely used Python libraries—PyMC, NumPyro, and TensorFlow Probability. Using simple one-dimensional conjugate models with known posterior distributions, it shows that defaults are not merely suboptimal; they can be systematically wrong. PyMC's default single Monte Carlo gradient sample plus its Adagrad_Window optimizer produces a posterior standard deviation that is biased about 5% high for millions of iterations, while TensorFlow Probability can fail silently, returning nonsensical results, when a constrained parameter is left untransformed. The paper argues these failures are implementation-level, not intrinsic to variational inference, and shows that switching to Adam and using 100 Monte Carlo samples removes the bias. The practical stake: practitioners who rely on defaults may report posterior summaries that are quietly wrong, without any error message.

Core claim

The central discovery is that, on a benchmark where the posterior is exactly Gaussian and the best variational approximation is exactly the true posterior, software defaults can nonetheless introduce systematic error. In PyMC, with the default single Monte Carlo sample and Adagrad_Window optimizer, the variational posterior's standard deviation overshoots the analytic value by roughly 5% and stays there for millions of iterations; increasing the Monte Carlo sample size to 100 or switching to the Adam optimizer removes the bias. NumPyro and TFP, which default to Adam, show no such bias under the same benchmark, though their trajectories oscillate and improve with more Monte Carlo samples. The

What carries the argument

The evaluation machinery is a one-dimensional Gaussian–Gaussian conjugate model with known variance, for which the posterior is exactly Gaussian and the variational optimum is exactly the true posterior, giving an analytic ground truth for the posterior mean and standard deviation. The paper tracks the variational parameters (posterior mean and standard deviation) over up to 1.6 million iterations under each package's defaults and under non-default settings, comparing against this ground truth. This isolates the effect of software implementation choices—Monte Carlo gradient sample count, optimizer (Adagrad_Window vs. Adam), initialization scheme, and support transformations—from the intrinsi

Load-bearing premise

The paper's recommendations assume that failure modes observed in one-dimensional conjugate models—where the variational family contains the true posterior—will also appear in realistic non-conjugate, multi-dimensional models, which the paper itself notes are the common case.

What would settle it

Run PyMC's default ADVI on a two-dimensional or non-conjugate model where the posterior can be computed exactly or via a long MCMC chain, and check whether the posterior standard deviation still sits roughly 5% off after millions of iterations; if the bias disappears when the model is no longer conjugate, the benchmark-specific explanation fails.

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

If this is right

  • In PyMC, changing the default optimizer from Adagrad_Window to Adam, or raising the Monte Carlo sample size to 100, removes the roughly 5% bias in the posterior standard deviation.
  • All three packages produce smaller oscillations in the variational trajectories when the number of Monte Carlo gradient samples is increased from 1 to 100.
  • A TFP user who fits a model with constrained support (e.g., a beta prior) without adding a bijector can get a NaN ELBO and a meaningless posterior approximation, with no error raised.
  • Initialization matters: in a bimodal example, different random restarts lead the variational approximation to different local modes, so multiple restarts are recommended in multimodal settings.
  • Following the paper's recommendations—prefer NumPyro, use 100 MC samples, switch PyMC to Adam, and verify support-matching transformations—makes the one-dimensional benchmarks behave predictably.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A reasonable next test is whether the 5% bias, driven by the interaction of a single noisy gradient sample with Adagrad_Window's step-size schedule, reappears in higher-dimensional or ill-conditioned models, where gradient noise is larger; the paper does not test this.
  • The TFP silent-failure mode suggests that a cheap, model-agnostic diagnostic—monitoring the ELBO for NaN and checking whether sampled variational draws lie inside the support—could catch most of these failures before any summary is reported.
  • The paper's preference for NumPyro rests on its conjugate ELBO estimator (TraceMeanField_ELBO); in non-conjugate models NumPyro switches to a different estimator (Trace_ELBO), so the default-experience ranking could change outside the conjugate setting.
  • One natural extension would be to run the same default-versus-tuned comparison on a small non-conjugate model where MCMC is affordable, and check whether the bias or silent failure appears in any package.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper evaluates the default variational inference (VI) settings of PyMC, NumPyro, and TensorFlow Probability on simple analytic conjugate models, with the one-dimensional Gaussian–Gaussian model as the primary testbed. Using the known posterior as ground truth, the authors report that PyMC's default settings (one Monte Carlo gradient sample and Adagrad_Window) exhibit a persistent bias of about 5% in the approximate posterior standard deviation, while NumPyro and TFP show only minor oscillations. They further examine how initialization control, transformation handling, and diagnostic monitoring differ across the packages, including a silent-failure mode for TFP when a required transformation is omitted. The paper concludes with practical guidance for users and developers, recommending increases to the MC sample count, use of Adam optimizers, and explicit verification of support transformations.

Significance. If the central claims are robust, the paper provides a useful empirical audit of off-the-shelf VI software, a topic that is often neglected relative to methodological advances. The benchmark design is well suited to isolating software-level effects: the Gaussian target is exactly in the variational family, so discrepancies from the analytic posterior can be attributed to implementation choices rather than variational-family limitations. The paper fixes software versions, states default hyperparameters, and makes demo code available. These are concrete strengths. The main risk is that the headline finding rests on a non-default iteration count and a single trajectory, and that the practical recommendations outrun the evidence presented.

major comments (3)
  1. [§3.1.1, §3.1.3, Figure 1] The central quantitative claim is that PyMC's default settings produce a systematic ~5% bias in the approximate posterior standard deviation. The supporting runs use a manually specified 1.6 million iterations (§3.1.3), whereas a genuinely default PyMC run terminates at 10,000 iterations when the number of iterations is not specified (§3.1.1). The manuscript never reports the approximate σ_p at iteration 10,000 for the default 1-MC/Adagrad_Window configuration, nor whether the biased plateau has already been reached by the default stopping time. The abstract's conclusion that 'default settings can yield biased approximations' therefore conflates the optimizer and MC-sample defaults with a non-default iteration budget. Please report the trajectory and the bias at the default stopping time, or explicitly reframe the claim as a long-run property of the default optimizer/MC settings under a
  2. [§3.1.2, Figure 1] The 5% bias result is demonstrated with a single trajectory and no error bars or multiple-seed analysis. Stochastic variational inference trajectories depend on random number generation, and the text itself describes oscillations in the same runs. Without repeated runs, 'systematic' and 'persists for millions of iterations' are not established. Please add multiple seeds (e.g., 10 or more) for the PyMC default configuration and report the distribution of final σ_p values, or otherwise provide quantitative evidence that the bias is reproducible rather than a single-run artifact.
  3. [§4.5 and §5.3.1] The practical guide recommends preferring NumPyro and avoiding PyMC and TFP for default users, based on the one-dimensional conjugate benchmarks. Section 5.3.1 itself acknowledges that most applied models are non-conjugate and multi-dimensional, and that NumPyro's TraceMeanField_ELBO estimator for conjugate cases differs from the Trace_ELBO used for non-conjugate models. No evidence is provided that the observed failure modes transfer to realistic models. The recommendations should either be explicitly scoped to the tested model class or supplemented with experiments in non-conjugate and higher-dimensional settings; otherwise the guide may mislead practitioners applying the defaults to problems outside this narrow benchmark family.
minor comments (5)
  1. [§3.1.2] The text says 'our preliminary further experiments suggest that the bias does persist' for beta, inverse-gamma, and gamma models, but no results are shown. Either include these results in an appendix or remove the preliminary claim, since it currently functions as unsupported evidence.
  2. [§2.2] The hyperparameters for the one-dimensional Gaussian–Gaussian benchmark (prior mean/variance, likelihood variance, sample size, and data) are not stated in the main text. Please provide the exact values so the experiments are reproducible without consulting the code repository.
  3. [§3.1.1] The definition of 'default' for NumPyro and TFP should be stated more carefully: both packages require the user to supply an optimizer and an iteration count, so the use of Adam is a convention adopted from official vignettes rather than a package default. This distinction is important for interpreting the comparison.
  4. [§3.1.3] The sentence 'With the exception of Figure 2...' is slightly ambiguous. It would be clearer to state explicitly in each figure caption that all runs use a manually specified 1.6 million iterations, and to specify whether the PyMC Adam run in Figure 2 uses the same iteration count.
  5. [Throughout] There is inconsistent terminology: 'systemic bias' appears in §3.1.2 while 'systematic bias' is used elsewhere. Please standardize.

Circularity Check

0 steps flagged

No circularity: the paper's claims are benchmarked against external analytic ground truths, not against its own outputs.

full rationale

The paper evaluates VI software by comparing variational trajectories to closed-form posterior means and variances of one-dimensional conjugate models, especially the Gaussian-Gaussian model. These analytic targets are external to the software and to the paper's own derivations, so success is defined by an independent ground truth rather than by the software's own outputs. The 'best variational approximation' is the ELBO maximizer, which in the Gaussian-Gaussian case coincides with the exact posterior because the posterior lies in the variational family; this is a mathematical fact, not a fitted result. Claims about a ~5% bias in PyMC, oscillations in NumPyro/TFP, mitigation by more MC samples or Adam, and TFP's silent failure without transforms are direct empirical observations, not predictions derived from fitted parameters. Self-citations (e.g., Giordano et al.) are used for background context on known VI limitations and are not load-bearing for the paper's central quantitative findings. The paper's limitation section 5.3.1 explicitly acknowledges that most applied models are non-conjugate and multidimensional, and that findings may not transfer; this is an honest scope restriction, not circular reasoning. The skeptical concern that the bias is only shown at a manually specified 1.6 million iterations rather than PyMC's 10,000-iteration default is a validity/generalizability caveat about whether the default termination point exhibits the bias, but it does not constitute a definitional or fitted-input circularity under the specified criteria. Hence no significant circularity is present.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

The paper introduces no new math or entities. It evaluates existing software against analytic benchmarks. The main assumptions are that the analytic ground truths apply and that the package defaults are as described. The paper's own limitations acknowledge the generality assumption.

axioms (4)
  • standard math Standard VI theory: ELBO equivalence, mean-field family, reverse KL.
    Used in Section 2.1 to define the optimization problem and best variational approximation.
  • standard math Analytic conjugate posterior formulas for Gaussian-Gaussian, Beta-Binomial, Gamma-Poisson, etc.
    Used as ground truth in Section 2.2 and Appendix A.
  • domain assumption Package documentation and source code accurately describe default behavior.
    They inspect PyMC, NumPyro, TFP source/docs (Section 3.1.1) to set defaults; if versions change or docs are misleading, the results may not reflect current software.
  • domain assumption One-dimensional conjugate benchmarks are informative for general VI behavior.
    Section 2.2 states failures in one dimension are expected to worsen in higher dimensions; this is plausible but unproven. Section 5.3.1 is the limitation.

reviewed 2026-08-06 · how reviews work

0 comments
Cite this review

Pith. "Pith review of The Curious Case of the Default Settings: Evaluating Default Performance of Variational Inference Software." pith.science (2026). https://pith.science/paper/7DAFIY7Q

@misc{pith2026260801403,
  author       = {Pith},
  title        = {Pith review of: The Curious Case of the Default Settings: Evaluating Default Performance of Variational Inference Software},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7DAFIY7Q}},
  note         = {Machine review of arXiv:2608.01403}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We systematically evaluate a suite of off-the-shelf variational inference (VI) software packages from the perspective of a standard practitioner. Using simple analytic benchmark models, we assess the accuracy and stability of the default VI settings in PyMC, NumPyro, and TensorFlow Probability. Unlike previous research focusing on methodological advances, our evaluation emphasizes software implementation and the default configurations that typical users encounter. Our results show that default settings can yield biased approximations of posterior summaries even for simple one-dimensional conjugate models, controls of initialization and transformations differ between software implementations, and relying on defaults may yield silent failures or poor approximations.

Figures

Figures reproduced from arXiv: 2608.01403 by Elizabeth Bersson, Madelyn Andersen, Tamara Broderick.

Figure 1
Figure 1. Figure 1: Default vs. tuned Monte Carlo gradient samples in PyMC (one-dimensional Gaussian, known variance) run with PyMC’s default Adagrad_Window optimizer. The top and bottom figures plot the variational trajectories under default (1) and adjusted (100) MC sample values for the posterior parameters σp and µp, respectively, where trajectories of the same color between plots are derived from the same VI run. Upper p… view at source ↗
Figure 2
Figure 2. Figure 2: Default vs. tuned Monte Carlo gradient samples in PyMC (one-dimensional Gaussian, known variance) run with Adam optimizer and default Adam settings. The upper and lower figures plot the variational trajectories under default (1) and adjusted (100) MC sample values for the posterior parameters σp and µp, respectively, where trajectories of the same color between plots are derived from the same VI run. Upper… view at source ↗
Figure 3
Figure 3. Figure 3: Default vs. tuned Monte Carlo gradient samples in NumPyro (one-dimensional Gaussian, known variance). The top and bottom figures plot the variational trajectories under default (1) and adjusted (100) MC sample values for the posterior parameters σp and µp, respectively, where trajectories of the same color between plots are derived from the same VI run. Upper plot: Using a single MC sample oscillates with … view at source ↗
Figure 4
Figure 4. Figure 4: Default vs. tuned Monte Carlo gradient samples in TFP (one-dimensional Gaussian, known variance). The top and bottom figures plot the variational trajectories under default (1) and adjusted (100) MC sample values for the posterior parameters σp and µp, respectively, where trajectories of the same color between plots are derived from the same VI run. Upper plot: Using a single MC sample oscillates with a sm… view at source ↗
Figure 5
Figure 5. Figure 5: The true target density (black solid line) is a balanced mixture of beta(20, 2) and beta(2, 20), producing two narrow modes. Shown are variational approximations (dashed lines) obtained by fitting a Gaussian in the unconstrained space and mapping it to (0, 1) via a sigmoid transformation using the NumPyro package with a default value of σ˜ = 0.1. Each color corresponds to a different initialization of µ˜. … view at source ↗
Figure 6
Figure 6. Figure 6: Support mismatch without a transform. The untransformed fit drifts out of support. Adding the sigmoid bijector yields a valid fit. The inset axes zoom out on the fitted VI distributions after 5,000 and 100,000 iterations to show behavior not visible on the smaller scale of the main plot. 4 Practitioner User Guide This section distills our evaluation into risk-aware guidance for practitioners who wish to us… 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

157 extracted references · 70 canonical work pages · 2 internal anchors

  1. [1]

    and Nero, Anthony V

    Price, Phillip N. and Nero, Anthony V. and Gelman, Andrew , journal =. Bayesian Prediction of Mean Indoor Radon Concentrations for. 1996 , volume =

    1996

  2. [2]

    2025 , booktitle =

    Fonnesbeck, Christopher and Abril, Oriol and Andorra, Alex and Carroll, Colin and Reynaldo, Farhan , title =. 2025 , booktitle =

    2025

  3. [3]

    Scientific Reports , volume =

    Bayesian Hierarchical Model Predicts Biopharmaceutical Stability Indicators and Shelf Life with Application to Multivalent Human Papillomavirus Vaccine , author =. Scientific Reports , volume =

  4. [4]

    How to increase the number of dropped evaluations variational bayes cmdstanr? , year =

  5. [5]

    2021 , month = nov, howpublished =

    Yao, Yuling , title =. 2021 , month = nov, howpublished =

    2021

  6. [6]

    Khazaei, Yeganeh and K. Using a. Scientific Reports , volume =

  7. [7]

    James and Gilliland, Frank D

    Weng, Jingying and Molshatzki, Noa and Marjoram, Paul and Gauderman, W. James and Gilliland, Frank D. and Eckel, Sandrah P. , year = 2023, month = apr, journal =. Longitudinal Hierarchical

    2023

  8. [8]

    and Petkova, Eva and Park, Hyung G

    Wu, Danni and Goldfeld, Keith S. and Petkova, Eva and Park, Hyung G. , year = 2024, month = sep, journal =. A

    2024

  9. [9]

    Proceedings of the 21st International Congress on Modelling and Simulation (MODSIM) , pages=

    Using the Bayesian Logistic Regression Model to Determine the Relationship of Demographics and Hyperaldosteronism , author=. Proceedings of the 21st International Congress on Modelling and Simulation (MODSIM) , pages=

  10. [10]

    Chinese Journal of Epidemiology , volume=

    Study on the interaction under logistic regression modeling , author=. Chinese Journal of Epidemiology , volume=. 2008 , pmid=

    2008

  11. [11]

    Political Science Research and Methods , volume=

    Compression and Conditional Effects: A Product Term Is Essential When Using Logistic Regression to Test for Interaction , author=. Political Science Research and Methods , volume=. 2016 , publisher=

    2016

  12. [12]

    PeerJ , volume=

    Bayesian estimation of the measurement of interactions in epidemiological studies , author=. PeerJ , volume=. 2024 , publisher=

    2024

  13. [13]

    and Luhmann, Christian , title =

    Orduz, Juan and Vincent, Benjamin T. and Luhmann, Christian , title =. 2023 , booktitle =

    2023

  14. [14]

    Journal of Machine Learning Research , year =

    Yuling Yao and Aki Vehtari and Andrew Gelman , title =. Journal of Machine Learning Research , year =

  15. [15]

    Madelyn Andersen , title =

  16. [16]

    , institution =

    Murphy, Kevin P. , institution =. Conjugate

  17. [17]

    Kasprzak and Ryan Giordano and Tamara Broderick , title =

    Mikolaj J. Kasprzak and Ryan Giordano and Tamara Broderick , title =. Journal of Machine Learning Research , year =

  18. [18]

    , month = jun, year =

    Hoff, Peter D. , month = jun, year =. A

  19. [19]

    Journal of Machine Learning Research , year =

    Lu Zhang and Bob Carpenter and Andrew Gelman and Aki Vehtari , title =. Journal of Machine Learning Research , year =

  20. [20]

    Variational Inference with Gaussian Score Matching

    Variational. arXiv preprint arXiv:2307.07849 , author =

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    Blei , title =

    Alp Kucukelbir and Dustin Tran and Rajesh Ranganath and Andrew Gelman and David M. Blei , title =. Journal of Machine Learning Research , year =

  22. [22]

    and Trippe, Brian L

    Nguyen, Tin D. and Trippe, Brian L. and Broderick, Tamara , booktitle =. Many processors, little time:. 2022 , volume =

    2022

  23. [23]

    arXiv preprint arXiv:2210.00580 , author =

    Pith/arXiv arXiv

  24. [24]

    and Robin, S

    Latouche, P. and Robin, S. , year =. Variational. Statistics and Computing , volume =

  25. [25]

    Proceedings of the 2016 ICML Workshop on \#Data4Good: Machine Learning in Social Good Applications , year=

    Fast robustness quantification with variational Bayes , author=. Proceedings of the 2016 ICML Workshop on \#Data4Good: Machine Learning in Social Good Applications , year=

    2016

  26. [26]

    and Malkin, Nikolay and Jain, Moksh and Everett, Katie and Graikos, Alexandros and Bengio, Yoshua , year =

    Hu, Edward J. and Malkin, Nikolay and Jain, Moksh and Everett, Katie and Graikos, Alexandros and Bengio, Yoshua , year =. Advances in Neural Information Processing Systems , volume =

  27. [27]

    and Ghahramani, Zoubin and Jaakkola, Tommi S

    Jordan, Michael I. and Ghahramani, Zoubin and Jaakkola, Tommi S. and Saul, Lawrence K. , year =. An. Machine Learning , volume =

  28. [28]

    Jordan , title =

    Ryan Giordano and Tamara Broderick and Michael I. Jordan , title =. Journal of Machine Learning Research , year =

  29. [29]

    , year =

    Kucukelbir, Alp and Ranganath, Rajesh and Gelman, Andrew and Blei, David M. , year =. Automatic. Advances in Neural Information Processing Systems , volume =

  30. [30]

    Statistical Science , author =

    Past,. Statistical Science , author =. 2024 , pages =

    2024

  31. [31]

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

    Rue, Håvard and Martino, Sara and Chopin, Nicolas , title =. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , volume =

  32. [32]

    Hoffman and David M

    Matthew D. Hoffman and David M. Blei and Chong Wang and John Paisley , title =. Journal of Machine Learning Research , year =

  33. [33]

    2014 , editor =

    Ranganath, Rajesh and Gerrish, Sean and Blei, David , booktitle =. 2014 , editor =

    2014

  34. [34]

    Yes, but

    Yao, Yuling and Vehtari, Aki and Simpson, Daniel and Gelman, Andrew , year =. Yes, but. Proceedings of the International Conference on Machine Learning , pages =

  35. [35]

    and Kucukelbir, Alp and McAuliffe, Jon D

    Blei, David M. and Kucukelbir, Alp and McAuliffe, Jon D. , year =. Variational. Journal of the American Statistical Association , volume =. 1601.00670 , primaryclass =

    Pith/arXiv arXiv

  36. [36]

    Variational

    Rezende, Danilo Jimenez and Mohamed, Shakir , year =. Variational. Proceedings of the International Conference on Machine Learning , pages =

  37. [37]

    and Salimans, Tim and Jozefowicz, Rafal and Chen, Xi and Sutskever, Ilya and Welling, Max , year =

    Kingma, Diederik P. and Salimans, Tim and Jozefowicz, Rafal and Chen, Xi and Sutskever, Ilya and Welling, Max , year =. Improving. Advances in Neural Information Processing Systems , volume =

  38. [38]

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

    Fat--Tailed Variational Inference with Anisotropic Tail Adaptive Flows , author =. Proceedings of the 39th International Conference on Machine Learning , pages =. 2022 , volume =

    2022

  39. [39]

    Discover Social Science and Health , volume=

    Determining the risk factors of under-five morbidity in Bangladesh: a Bayesian logistic regression approach , author=. Discover Social Science and Health , volume=. 2023 , publisher=

    2023

  40. [40]

    Stan GitHub Repository , url =

    advi.hpp , month = feb, year =. Stan GitHub Repository , url =

  41. [41]

    2nd International Conference on Learning Representations (ICLR) , year=

    Auto-Encoding Variational Bayes , author=. 2nd International Conference on Learning Representations (ICLR) , year=

  42. [42]

    Foundations and Trends in Machine Learning , volume=

    An Introduction to Variational Autoencoders , author=. Foundations and Trends in Machine Learning , volume=

  43. [43]

    Journal of Machine Learning Research , year =

    Shakir Mohamed and Mihaela Rosca and Michael Figurnov and Andriy Mnih , title =. Journal of Machine Learning Research , year =

  44. [44]

    , title=

    Domke, Justin and Sheldon, Daniel R. , title=. Advances in Neural Information Processing Systems , volume=

  45. [45]

    Bayesian Analysis , year=

    Yao, Yuling and Vehtari, Aki and Simpson, Daniel and Gelman, Andrew , title=. Bayesian Analysis , year=

  46. [46]

    and Kasprzak, Mikołaj and Campbell, Trevor and Broderick, Tamara , title=

    Huggins, Jonathan H. and Kasprzak, Mikołaj and Campbell, Trevor and Broderick, Tamara , title=. Proceedings of the 23rd International Conference on Artificial Intelligence and Statistics (AISTATS) , year=

  47. [47]

    2006 , month = oct, day =

    National Ambient Air Quality Standards for Particulate Matter; Final Rule , journal =. 2006 , month = oct, day =

    2006

  48. [48]

    Gelman, Andrew and Hill, Jennifer , title =

  49. [49]

    and Lee, Daniel and Goodrich, Ben and Betancourt, Michael and Brubaker, Marcus and Guo, Jiqiang and Li, Peter and Riddell, Allen , title=

    Carpenter, Bob and Gelman, Andrew and Hoffman, Matthew D. and Lee, Daniel and Goodrich, Ben and Betancourt, Michael and Brubaker, Marcus and Guo, Jiqiang and Li, Peter and Riddell, Allen , title=. Journal of Statistical Software , year=

  50. [50]

    and Fonnesbeck, Christopher , title=

    Salvatier, John and Wiecki, Thomas V. and Fonnesbeck, Christopher , title=. PeerJ Computer Science , year=

  51. [51]

    Chen and Martin Jankowiak and Fritz Obermeyer and Neeraj Pradhan and Theofanis Karaletsos and Rohit Singh and Paul Szerlip and Paul Horsfall and Noah D

    Eli Bingham and Jonathan P. Chen and Martin Jankowiak and Fritz Obermeyer and Neeraj Pradhan and Theofanis Karaletsos and Rohit Singh and Paul Szerlip and Paul Horsfall and Noah D. Goodman , title =. Journal of Machine Learning Research , year =

  52. [52]

    Journal of the American Statistical Association , author=

    Frequentist Consistency of Variational Bayes , volume=. Journal of the American Statistical Association , author=. 2019 , pages=

    2019

  53. [53]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Zhang, Cheng and Butepage, Judith and Kjellstrom, Hedvig and Mandt, Stephan , title=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

  54. [54]

    arXiv preprint arXiv:1705.10843 , year=

    Objective-Reinforced Generative Adversarial Networks (ORGAN) for Sequence Generation Models , author=. arXiv preprint arXiv:1705.10843 , year=

    Pith/arXiv arXiv

  55. [55]

    2017 , eprint=

    Edward: A library for probabilistic modeling, inference, and criticism , author=. 2017 , eprint=

    2017

  56. [56]

    Advances in Neural Information Processing Systems , volume=

    Liu, Qiang and Wang, Dilin , title=. Advances in Neural Information Processing Systems , volume=. 2016 , pages=

    2016

  57. [57]

    2024 , eprint=

    Theoretical Guarantees for Variational Inference with Fixed-Variance Mixture of Gaussians , author=. 2024 , eprint=

    2024

  58. [58]

    2020 , school =

    Jincheng Bai , title =. 2020 , school =

    2020

  59. [59]

    2016 , journal =

    Variational Inference for Medical Image Segmentation , author =. 2016 , journal =

    2016

  60. [60]

    npj Climate and Atmospheric Science , volume =

    Data-Driven Seasonal Climate Predictions via Variational Inference and Transformers , author =. npj Climate and Atmospheric Science , volume =

  61. [61]

    and Smith, Jared D

    Kavianihamedani, Hossein and Quinn, Julianne D. and Smith, Jared D. , journal=. New Diagnostic Assessment of MCMC Algorithm Effectiveness, Efficiency, Reliability, and Controllability , year=

  62. [62]

    Mathematics and Computers in Simulation (MATCOM) , author=

    Encoded. Mathematics and Computers in Simulation (MATCOM) , author=. 2022 , pages=

    2022

  63. [63]

    2020 , journal=

    Theoretical and Computational Guarantees of Mean Field Variational Inference for Community Detection , author=. 2020 , journal=

    2020

  64. [64]

    2021 , eprint=

    Challenges and Opportunities in High-dimensional Variational Inference , author=. 2021 , eprint=

    2021

  65. [65]

    2020 , eprint=

    Robust, Accurate Stochastic Optimization for Variational Inference , author=. 2020 , eprint=

    2020

  66. [66]

    2024 , eprint=

    Stable Training of Normalizing Flows for High-dimensional Variational Inference , author=. 2024 , eprint=

    2024

  67. [67]

    2024 , eprint=

    Preferential Normalizing Flows , author=. 2024 , eprint=

    2024

  68. [68]

    2024 , eprint=

    Disentangling impact of capacity, objective, batchsize, estimators, and step-size on flow VI , author=. 2024 , eprint=

    2024

  69. [69]

    and Teschendorff, A

    Ma, Z. and Teschendorff, A. E. , title =. Journal of Bioinformatics and Computational Biology , year =

  70. [70]

    Fay, R. E. and Herriot, R. A. , title =. Journal of the American Statistical Association , volume =

  71. [71]

    2008 , journal=

    Model-based Clustering of DNA Methylation Array Data: A Recursive-partitioning Algorithm for High-dimensional Data Arising as a Mixture of Beta Distributions , author=. 2008 , journal=

    2008

  72. [72]

    and Jordan, Michael I

    Jaakkola, Tommi S. and Jordan, Michael I. , journal =. A Variational Approach to. 1997 , editor =

    1997

  73. [73]

    Propagation Algorithms for Variational Bayesian Learning , volume =

    Ghahramani, Zoubin and Beal, Matthew , booktitle =. Propagation Algorithms for Variational Bayesian Learning , volume =

  74. [74]

    2010 , eprint=

    Graphical Comparison of MCMC Performance , author=. 2010 , eprint=

    2010

  75. [75]

    2025 , howpublished =

    Bob Carpenter and Madelyn Andersen and Aki Vehtari , title =. 2025 , howpublished =

    2025

  76. [76]

    2022 , howpublished =

    Vehtari, Aki , title =. 2022 , howpublished =

    2022

  77. [77]

    Variational Inference , year =

  78. [78]

    2020 , month =

    Dablander, Fabian , title =. 2020 , month =

    2020

  79. [79]

    Variational Inference in Stan , year =

  80. [80]

    2023 , note=

    Why and how Variational Inference underestimates variance? , author=. 2023 , note=

    2023

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 6, 2026.