Pith. sign in

REVIEW 3 major objections 4 minor 106 references

Disentangling impact of capacity, objective, batchsize, estimators, and step-size on flow VI

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a simple recipe—high-capacity real-NVP flows, large gradient batchsizes, the standard KL(q||p) objective, and a fixed step-size near 3e-4—lets flow VI match or beat leading turnkey HMC methods with far fewer…

desk verdict A careful, compute-heavy ablation that gives flow VI a concrete recipe and a credible HMC comparison, but the headline claim leans on a marginal-only metric and a compute count that excludes flow-network cost. read the letter →

arxiv 2412.08824 v1 pith:STAD6ZEB submitted 2024-12-11 cs.LG stat.ML

Abhinav Agrawal , Justin Domke This is my paper · ORCID
classification cs.LGstat.ML
keywords flowvariationalinferencenormalizingflowsreal-NVPmarginalWassersteindistanceHamiltonianMonteCarlogradientvariancebatchsizestep-sizeselection
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

This paper aims to identify which algorithmic choices actually decide whether normalizing-flow variational inference (flow VI) succeeds or fails, and to turn those choices into a concrete recipe. By testing one factor at a time on synthetic targets with exact sampling, it argues that two things matter most: enough flow capacity and large gradient batchsizes. Given those, the standard KL(q||p) objective and a fixed step-size in a narrow range suffice, and fancy mode-spanning objectives or elaborate estimators are not required. The paper concludes that the resulting recipe matches or surpasses leading turnkey Hamiltonian Monte Carlo methods on the tested models while needing far fewer sequential model evaluations at large parallel budgets.

What carries the argument

The argument is carried by a controlled ablation protocol plus the marginal-Wasserstein metric. The benchmark uses synthetic targets with exact sampling (ill-conditioned Gaussian, banana, funnel, a new funnel-plus-banana hybrid called Funana, and Student-t with ν=1.5 and 2.5), so each experiment can be scored against ground truth. The metric, defined as the average over dimensions of the Wasserstein distance between sorted one-dimensional samples, is what makes high-fidelity comparisons possible at all: unlike the full Wasserstein distance it scales log-linearly with sample size, allowing one million reference samples. Each study section varies one factor—capacity, objective, batchsize/estimator, or step-size—while holding the others fixed at values chosen by the previous sections, which is what allows the paper to attribute performance differences to individual causes.

What would settle it

Take the 10-dimensional funnel target, run the paper's recipe, and measure both marginal-Wasserstein distance and a joint-dependence metric (for example, energy distance or full d-dimensional Wasserstein distance with a sufficiently large sample size) against exact target samples. If the flow's marginals are nearly exact while the joint metric stays far above the exact-samples baseline, the paper's central comparison would not survive a change of metric.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that the inconsistent results reported for flow VI come mainly from under-capacity and from optimization failures induced by high gradient variance, not from an inherent weakness of normalizing flows or of the standard variational objective. With real-NVP flows of ten or more coupling layers, a large number of gradient samples per iteration, the sticking-the-landing gradient estimator where feasible, Adam, and a step-size in the range 1e-4 to 1e-3 run for at least ten thousand iterations, flow VI reproduces marginal-Wasserstein accuracy close to exact inference on six synthetic pathologies and on six real models. The paper further claims that, at sufficiently large parallel budgets, this recipe needs far fewer sequential model evaluations than the leading turnkey HMC implementations it compares against, which is what supports the headline that flow VI 'matches or surpasses' HMC.

Load-bearing premise

The load-bearing premise is that marginal-Wasserstein distance is a faithful measure of posterior accuracy; if a flow matched all one-dimensional marginals while getting the joint dependence wrong, the result that flow VI matches HMC could be an artifact of the metric.

Editorial extensions

If this is right

  • With ten or more coupling layers and 32 hidden units, real-NVP flows have enough capacity to represent the tested posterior pathologies accurately; increases in either layers or hidden units improve accuracy.
  • The standard reverse-KL objective KL(q||p) is sufficient once capacity is high; mode-spanning KL(p||q) only matters when the variational family is too weak, and it struggles on heavy-tailed targets.
  • Large gradient batchsizes are a primary lever: for a fixed number of iterations, bigger batchsizes dramatically improve accuracy, and the STL estimator helps mainly at small batchsizes.
  • Step-sizes must stay in a narrow range, roughly 1e-4 to 1e-3 over long runs; step-sizes that look good for 1,000 iterations, such as 1e-2, can diverge after 10,000 iterations.
  • At parallel budgets of 2^16 evaluations on the synthetic targets and 2^12 on real models, the recipe matches or surpasses the compared NUTS and CHEES baselines in marginal-Wasserstein accuracy with far fewer sequential evaluations.

Reading between the lines

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

  • The paper's evaluation metric is marginal-Wasserstein, which averages per-dimension distances and, as the paper notes in Section 2.2, misses correlations between dimensions; a reader should treat the HMC-competitiveness claim as established for marginals, not for joint dependence.
  • The centrality of large batchsizes suggests that the paper's conclusion is compute-dependent: on modest hardware where batchsizes are necessarily small, the gap between flow VI and HMC will be larger, and the recipe's step-size recommendations may be less reliable.
  • The step-size divergence behavior implies that early-stopping or step-size selection based on the first few hundred iterations is unsafe; a diagnostic that detects late divergence would be a natural next step.
  • The residual gap on Funana in 100 dimensions suggests a class of target geometries (simultaneously funnel-like and curved) that this recipe does not fully crack, pointing toward structured or hierarchical flows as a likely next extension.
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 / 4 minor

Summary. The paper presents a step-by-step empirical study of normalizing-flow variational inference, isolating the influence of flow capacity, choice of divergence objective, gradient estimator and batch size, and step-size schedule. On a newly introduced synthetic benchmark with exact sampling, it finds that high-capacity real-NVP flows with the standard KL(q||p) objective, the STL gradient estimator, large gradient batch sizes, and a fixed step size in 1e-4--1e-3 suffice to reach near-exact marginal accuracy. It then proposes this combination as a recipe and, using the marginal-Wasserstein metric, claims that the resulting flow VI 'matches or surpasses leading turnkey HMC methods' while requiring fewer sequential model evaluations, on both synthetic and real-world targets.

Significance. If the claims hold, the paper is a practically valuable, large-scale empirical contribution: it neutralizes confounders one factor at a time, provides a synthetic benchmark with exact samples for high-fidelity evaluation, and gives concrete, actionable recommendations for flow VI. The manuscript explicitly acknowledges the main weakness of its evaluation metric, which is a point in its favor, and the experimental setup is unusually thorough for an empirical inference paper. However, the headline comparison with HMC rests on a marginal-only metric and on a cost measure that is asymmetric between the two families of methods, so the central claim as currently stated is not fully established.

major comments (3)
  1. [Section 2.2, Eq. (2); Section 7; Figures 1 and 7] The paper's headline claim that flow VI 'matches or surpasses leading turnkey HMC methods' is supported only by the marginal-Wasserstein metric, which the paper itself states 'only looks at marginals, and so it misses the correlations between dimensions.' A flow could match every one-dimensional marginal of the target while having an incorrect joint dependence, and it would still appear to match HMC under Eq. (2). The black-dotted exact-inference reference line in Figures 1 and 7 is also computed with the same marginal metric, so approaching it does not certify joint accuracy. For the real-model comparisons, the reference samples themselves come from NUTS runs with stuck chains dropped (Appendix K.1), so the reference joint distribution is also not fully characterized. To make the HMC comparison claim, the paper should either add a joint-fidelity diagnostic on the synthetic targets (for example, a sample-based energy distance, a kernel Stein discrepancy, or a copula/rank-based correlation check) or rephrase the claim to be explicitly about marginal accuracy.
  2. [Section 7; Appendix K; Appendix F] The sequential-evaluation comparison is asymmetric between flow VI and HMC. For VI, the x-axis counts optimization iterations, while for HMC it counts leapfrog steps. A single flow VI iteration with batch size 2^16 evaluates the target log-density 2^16 times, plus several forward and inverse passes through a 10-layer real-NVP network (and STL adds further passes, as discussed in Appendix F). A single HMC leapfrog step with 2^16 chains evaluates the target once per chain. Thus 'sequential model evaluations' are not measured on a common scale, and the claim that flow VI requires 'far fewer sequential model evaluations' (Abstract; Section 7) overstates what the plotted quantity shows. The authors should report a cost measure that counts actual model evaluations or wall-clock time, or at minimum explicitly quantify the per-iteration flow-network cost in the sequential budget.
  3. [Figures 3--7; Appendix B; Section 7] The final recipe (Appendix B) fixes capacity, batch size, step size, and iteration count using sweeps on the same synthetic benchmark on which the headline HMC comparison is then carried out. This creates a risk of selection bias that is not addressed by the current presentation: the recipe may be overfit to the benchmark targets. Additionally, the main figures do not show error bars or seed variation, so it is not possible to assess the stability of the reported marginal-Wasserstein values, despite the checklist entry stating that error bars are provided. At minimum, the authors should report multiple-seed results with intervals for the Section 7 comparison, and should discuss (or empirically test on held-out targets) how sensitive the recipe's advantage is to the benchmark used for tuning.
minor comments (4)
  1. [Throughout] There are several typographical errors, including 'TensforFlow' in Figures 1 and 7 and 'TensforFlow probability' in Section 2.4, and 'Wass' is used as an axis label without a definition in the axis text.
  2. [Checklist, item 3(a) and 1(c)] The checklist states that code, data, and instructions are provided ('[Yes]') but also states that anonymized source code is not provided ('[No]'). This is contradictory and should be clarified, ideally by making the reproduction assets explicit.
  3. [Section 2.2, Eq. (2)] The marginal-Wasserstein definition sorts each column independently, which is fine for univariate marginals, but the notation A# and B# should be defined once in the main text; they are used before their formal definition in the appendix.
  4. [Appendix L.1] In the Banana target description, the formula for z2 uses 'z2 ∼ N(0.03(z2_1 - 100), 1)', which should be 'z2 ∼ N(0.03(z1^2 - 100), 1)' for consistency with the following sentence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical, the recipe is evaluated on held-out real targets, and the self-citations are not load-bearing.

full rationale

This is an empirical study with no derivation chain that could reduce to its own inputs. The central claim, that the proposed flow VI recipe matches or surpasses turnkey HMC, is supported by controlled experiments on synthetic targets with exact samples and on six real-world targets with reference samples from long NUTS runs. The recipe hyperparameters (ten coupling layers, 32 hidden units, KL(q||p), STL, large batchsize, step-size near 3e-4) are selected using sweeps on the synthetic benchmark, which is a generalization concern rather than a circularity: Section 7 additionally applies the fixed recipe to non-synthetic models not used in those sweeps, so the comparison is not statistically forced. Self-citations of the authors' prior work [3] appear for implementation details such as STL for flows and zero-initialization as identity, but these are not load-bearing for the headline result, which is established by new experiments. The acknowledged limitation that marginal-Wasserstein distance ignores inter-dimensional correlations (Section 2.2) is a metric-validity caveat; it does not make the empirical comparison equivalent to the metric by construction. No equation is defined in terms of the claim, no fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from the authors' earlier work. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The paper is empirical; it introduces one new benchmark target and relies on several domain assumptions about metrics, compute, and reference samples. The recipe hyperparameters are selected by sweeping on the benchmark, so they act as fitted values for the recommendations.

free parameters (5)
  • NumberOfCouplingLayers = 10
    Chosen from capacity sweep in Figures 3 and 4; 10+ layers reach near-exact performance on synthetic targets.
  • HiddenUnits = 32
    Two-hidden-layer FNN with 32 units; larger unit count helps 100-dimensional targets.
  • GradientBatchsize = 524288 (512 x 2^10)
    Massive batchsize selected to reduce gradient variance; recommended as an absolute essential.
  • StepSize = 0.0003 (within 1e-4 to 1e-3)
    Fixed step-size in recommended range from Figure 6; used in Section 7 recipe.
  • OptimizationIterations = 100000
    Long runs; 10K or more recommended; 100K used for best results.
assumptions (5)
  • domain assumption Marginal-Wasserstein distance is an adequate proxy for inference quality in comparing VI and HMC
    Used for all headline comparisons (eq. 2, Section 2.2); paper admits it ignores correlations between dimensions.
  • domain assumption Synthetic targets represent common posterior pathologies relevant to real-world problems
    Benchmark in Section 2.1 chosen to cover ill-conditioning, nonlinear curvature, heavy tails, funnel geometry; real-world validation is limited to six models.
  • domain assumption Model evaluations are the dominant computational cost, so counting them alone gives a fair efficiency comparison
    Section 7 counts only model evaluations; flow network forward/inverse passes are not counted, which favors flow VI.
  • domain assumption For real-world targets, long NUTS runs with dropped stuck chains provide reference samples
    Appendix K.1; reference samples are themselves HMC output, so both methods are compared against HMC-based ground truth.
  • domain assumption Real-NVP architecture with affine coupling layers, tanh scale, and small normal initialization is a suitable default
    Appendix E; the recipe relies on this architecture choice; other architectures may behave differently.
invented entities (1)
  • Funana target independent evidence
    purpose: New synthetic density in the benchmark combining funnel and banana pathologies to stress inference methods
    Defined in Appendix L.1 with exact sampling; provides a falsifiable benchmark that others can use.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Disentangling impact of capacity, objective, batchsize, estimators, and step-size on flow VI." pith.science (2026). https://pith.science/paper/STAD6ZEB

@misc{pith2026241208824,
  author       = {Pith},
  title        = {Pith review of: Disentangling impact of capacity, objective, batchsize, estimators, and step-size on flow VI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/STAD6ZEB}},
  note         = {Machine review of arXiv:2412.08824}
}
read the original abstract

Normalizing flow-based variational inference (flow VI) is a promising approximate inference approach, but its performance remains inconsistent across studies. Numerous algorithmic choices influence flow VI's performance. We conduct a step-by-step analysis to disentangle the impact of some of the key factors: capacity, objectives, gradient estimators, number of gradient estimates (batchsize), and step-sizes. Each step examines one factor while neutralizing others using insights from the previous steps and/or using extensive parallel computation. To facilitate high-fidelity evaluation, we curate a benchmark of synthetic targets that represent common posterior pathologies and allow for exact sampling. We provide specific recommendations for different factors and propose a flow VI recipe that matches or surpasses leading turnkey Hamiltonian Monte Carlo (HMC) methods.

Figures

Figures reproduced from arXiv: 2412.08824 by the authors.

Figure 1
Figure 1. Marginal-Wasserstein metric (eq. 2) against sequential evaluations for Neal’s funnel [69] in ten dimensions, with parallel budget increasing from left to right. For flow VI, sequential evaluations count optimization iterations and parallel evaluations represent batchsize. For HMC, sequential evaluations count leapfrog steps and parallel evaluations represent number of chains. (We use implementations from NumPyro and… view at source ↗
Figure 2
Figure 2. Rows: Pair marginals. These targets cover various pathologies: Ill-conditioned Gaussian has high correlations, Banana has non-linear relationships, Neal’s funnel has parameters whose spread depends on other parameters, Funana combines funnel-like behavior with non-linearity, and Student-t with ν = 1.5 has heavier tails than Student-t with ν = 2.5. and inverse (T −1 ) pass, allowing additional gradient estimators [91… view at source ↗
Figure 3
Figure 3. Rows: Model dimensions. Marginal-Wasserstein metric (eq. 2) against number of coupling layers for different number of hidden units. Performance improves with increase in either of these levers of capacity for all targets but Student-t with ν = 1.5. Heavy tails of this target create problems when optimizing KL (p ∥ q) [47]. See [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Rows: Model dimensions. Marginal-Wasserstein metric (eq. 2) against the number of layers (with 32 hidden units) for different objectives. Performance of KL (q ∥ p) (red) optimization improves as the number of layers increase, often reaching that of exact inference. A g…
Figure 5
Figure 5. Figure 5: Rows: Number of iterations. Marginal-Wasserstein metric against the number of samples used for gradient evaluation for different targets in ten dimensions. STL (red) consistently outperforms total gradient (blue) at smaller batchsizes. However, as the batchsize increas…
Figure 6
Figure 6. Figure 6: Rows: Model dimensions. Marginal-Wasserstein metric versus step-sizes for different iteration counts. Each point represents an independent optimization run; missing points indicate divergence. Notably, certain step-sizes, like 10−2 , achieve strong performance initiall…
Figure 7
Figure 7. Figure 7: Rows: Parallel evaluations. Marginal-Wasserstein metric against number of sequential evaluations for (a) synthetic targets in ten dimensions and (b) real-world problems, with parallel evaluations increasing across the rows. With sufficient parallel compute, flow VI is …
Figure 8
Figure 8. Figure 8: (a) Empirical 2−Wasserstein and marginal-Wasserstein metric against approximation scale λ for the example described in Appendix C.1. As λ increases, approximation becomes poor, increasing the true value of either of the Wasserstein metrics (solid blue line). For poor a…
Figure 9
Figure 9. Figure 9: Rows: Number of iterations. Dimensions indicated in brackets alongside model names. Figure uses the same setting as [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Rows: Model dimensions. Figure uses the same setting as [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Rows: Parallel evaluations. Dimensions indicated in brackets alongside model names. Figure uses the same setting as [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]
Figure 12
Figure 12. Figure 12: Rows: Parallel evaluations. Marginal-Wasserstein metric against number of sequential evaluations for non-synthetic models from Appendix L.2 with parallel evaluations increasing across the rows. The first subplot corresponds to Wasserstein values when we do not filter …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

106 extracted references · 65 canonical work pages

  1. [1]

    Analytic planetary transit light curves and derivatives for stars with polynomial limb darkening

    Eric Agol, Rodrigo Luger, and Daniel Foreman- Mackey. Analytic planetary transit light curves and derivatives for stars with polynomial limb darkening. The Astronomical Journal, 2020

    2020

  2. [2]

    Understanding and mitigating difficulties in posterior predictive evaluation

    AbhinavAgrawalandJustinDomke. Understand- ing and mitigating difficulties in posterior predic- tive evaluation.arXiv preprint arXiv:2405.19747, 2024

    work page Pith review arXiv 2024

  3. [3]

    Sheldon, and Justin Domke

    Abhinav Agrawal, Daniel R. Sheldon, and Justin Domke. Advances in black-box VI: normalizing flows, importance weighting, and optimization. In NeurIPS, 2020

    2020

  4. [4]

    Automatic structured variational infer- ence

    Luca Ambrogioni, Kate Lin, Emily Fertig, Sharad Vikram, Max Hinne, Dave Moore, and Marcel Gerven. Automatic structured variational infer- ence. In AISTATS, 2021

    2021

  5. [5]

    Automatic variational inference with cascading flows

    Luca Ambrogioni, Gianluigi Silvestri, and Marcel van Gerven. Automatic variational inference with cascading flows. InICML, 2021

    2021

  6. [6]

    Stable Training of Normalizing Flows for High-dimensional Variational Inference

    Daniel Andrade. Stable training of normalizing flows for high-dimensional variational inference. arXiv preprint arXiv:2402.16408, 2024

    work page Pith review arXiv 2024

  7. [7]

    Annealed flow transport monte carlo

    Michael Arbel, Alex Matthews, and Arnaud Doucet. Annealed flow transport monte carlo. In ICML, 2021

    2021

  8. [8]

    Automatic Guide Generation for Stan via NumPyro

    Guillaume Baudart and Louis Mandel. Auto- matic guide generation for stan via numpyro. arXiv preprint arXiv:2110.11790, 2021

    work page Pith review arXiv 2021

Show all 106 references
  1. [9]

    Generalized doubly reparameterized gradient estimators

    Matthias Bauer and Andriy Mnih. Generalized doubly reparameterized gradient estimators. In ICML, 2021

    2021

  2. [10]

    Un- derstanding and mitigating exploding inverses in invertible neural networks

    Jens Behrmann, Paul Vicol, Kuan-Chieh Wang, Roger Grosse, and Jörn-Henrik Jacobsen. Un- derstanding and mitigating exploding inverses in invertible neural networks. InAISTATS, 2021

    2021

  3. [11]

    A conceptual introduc- tion to hamiltonian monte carlo.arXiv preprint arXiv:1701.02434, 2017

    Michael Betancourt. A conceptual introduc- tion to hamiltonian monte carlo.arXiv preprint arXiv:1701.02434, 2017

    2017 arXiv

  4. [12]

    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, 2017

    2017

  5. [13]

    Be- yond elbos: A large-scale evaluation of varia- tional methods for sampling

    Denis Blessing, Xiaogang Jia, Johannes Esslinger, Francisco Vargas, and Gerhard Neumann. Be- yond elbos: A large-scale evaluation of varia- tional methods for sampling. arXiv preprint arXiv:2406.07423, 2024

    2024 arXiv

  6. [14]

    JAX: composable transformations of Python+NumPy programs, 2018

    James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018. URL http: //github.com/google/jax

    2018

  7. [15]

    Adaptation of the independent metropolis-hastings sampler with normalizing flow proposals

    James Brofos, Marylou Gabrié, Marcus A Brubaker, and Roy R Lederman. Adaptation of the independent metropolis-hastings sampler with normalizing flow proposals. InAISTATS, 2022

    2022

  8. [16]

    Importance weighted autoencoders.ICLR, 2016

    Yuri Burda, Roger Grosse, and Ruslan Salakhut- dinov. Importance weighted autoencoders.ICLR, 2016

    2016

  9. [17]

    U-statistics for importance- weighted variational inference

    Javier Burroni, Kenta Takatsu, Justin Domke, and Daniel Sheldon. U-statistics for importance- weighted variational inference. arXiv preprint arXiv:2302.13918, 2023

    2023 arXiv

  10. [18]

    Trans- port elliptical slice sampling

    Alberto Cabezas and Christopher Nemeth. Trans- port elliptical slice sampling. InAISTATS, 2023

    2023

  11. [19]

    Batch and match: black-box variational inference with a score-based divergence

    Diana Cai, Chirag Modi, Loucas Pillaud- Vivien, Charles C Margossian, Robert M Gower, David M Blei, and Lawrence K Saul. Batch and match: black-box variational inference with a score-based divergence. arXiv preprint arXiv:2402.14758, 2024

    2024 arXiv

  12. [20]

    On the 1- wasserstein distance between location-scale dis- tributions and the effect of differential privacy

    Saurab Chhachhi and Fei Teng. On the 1- wasserstein distance between location-scale dis- tributions and the effect of differential privacy. arXiv preprint arXiv:2304.14869, 2023

    2023 arXiv

  13. [21]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. InNeurIPS, volume 26, 2013

    2013

  14. [22]

    Optimal transport tools (ott): A jax toolbox for all things wasserstein.arXiv preprint arXiv:2201.12324, 2022

    Marco Cuturi, Laetitia Meng-Papaxanthos, Ying- tao Tian, Charlotte Bunne, Geoff Davis, and Olivier Teboul. Optimal transport tools (ott): A jax toolbox for all things wasserstein.arXiv preprint arXiv:2201.12324, 2022

    2022 arXiv

  15. [23]

    Robust, accurate stochastic optimization for variational inference

    Akash Kumar Dhaka, Alejandro Catalina, Michael R Andersen, Måns Magnusson, Jonathan Huggins, and Aki Vehtari. Robust, accurate stochastic optimization for variational inference. In NeurIPS, 2020

    2020

  16. [24]

    Challenges and opportunities in high dimensional variational inference

    Akash Kumar Dhaka, Alejandro Catalina, Manushi Welandawe, Michael R Andersen, Jonathan Huggins, and Aki Vehtari. Challenges and opportunities in high dimensional variational inference. In NeurIPS, 2021

    2021

  17. [25]

    Variational inference via χ upper bound minimization

    Adji Bousso Dieng, Dustin Tran, Rajesh Ran- ganath, John Paisley, and David Blei. Variational inference via χ upper bound minimization. In NeurIPS, 2017. Disentangling impact of capacity, objective, batchsize, estimators, and step-size on flow VI

    2017

  18. [26]

    Nice: Non-linear independent components estimation

    Laurent Dinh, David Krueger, and Yoshua Ben- gio. Nice: Non-linear independent components estimation. In ICLR, 2015

    2015

  19. [27]

    Density estimation using real nvp.ICLR, 2017

    Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp.ICLR, 2017

    2017

  20. [28]

    Wasserstein distance between two gaussians

    C Djalil. Wasserstein distance between two gaussians. https://djalil.chafai.net/blog/ 2010/04/30/wasserstein-distance-between- two-gaussians/, April 2010

    2010

  21. [29]

    Importance weighting and variational inference

    Justin Domke and Daniel R Sheldon. Importance weighting and variational inference. InNeurIPS, 2018

    2018

  22. [30]

    Divide and couple: Using monte carlo variational objectives for posterior approximation.NeurIPS, 2019

    Justin Domke and Daniel R Sheldon. Divide and couple: Using monte carlo variational objectives for posterior approximation.NeurIPS, 2019

    2019

  23. [31]

    Bernstein flows for flexible posteriors in variational bayes

    Oliver Durr, Stefan Hortling, Danil Dold, Ivonne Kovylov, and Beate Sick. Bernstein flows for flexible posteriors in variational bayes. AStA Advances in Statistical Analysis, 2024

    2024

  24. [32]

    On importance-weighted autoencoders

    Axel Finke and Alexandre H Thiery. On importance-weighted autoencoders. arXiv preprint arXiv:1907.10477, 2019

    1907 arXiv

  25. [33]

    Multilevel monte carlo variational inference.JMLR, 2021

    Masahiro Fujisawa and Issei Sato. Multilevel monte carlo variational inference.JMLR, 2021

    2021

  26. [34]

    Adaptive monte carlo aug- mented with normalizing flows.Proceedings of the National Academy of Sciences, 2022

    Marylou Gabrié, Grant M Rotskoff, and Eric Vanden-Eijnden. Adaptive monte carlo aug- mented with normalizing flows.Proceedings of the National Academy of Sciences, 2022

    2022

  27. [35]

    On the diffi- culty of unbiased alpha divergence minimization

    Tomas Geffner and Justin Domke. On the diffi- culty of unbiased alpha divergence minimization. arXiv preprint arXiv:2010.09541, 2020

    2010 arXiv

  28. [36]

    A rule for gradient estimator selection, with an application to variational inference

    Tomas Geffner and Justin Domke. A rule for gradient estimator selection, with an application to variational inference. InAISTATS, 2020

    2020

  29. [37]

    Empirical evaluation of biased methods for alpha divergence minimization

    Tomas Geffner and Justin Domke. Empirical evaluation of biased methods for alpha divergence minimization. arXiv preprint arXiv:2105.06587, 2021

    2021 arXiv

  30. [38]

    Variational methods for simulation-based inference

    Manuel Glöckler, Michael Deistler, and Jakob H Macke. Variational methods for simulation-based inference. In ICLR, 2022

    2022

  31. [39]

    Adaptive proposal distribution for ran- dom walk metropolis algorithm.Computational statistics, 1999

    Heikki Haario, Eero Saksman, and Johanna Tam- minen. Adaptive proposal distribution for ran- dom walk metropolis algorithm.Computational statistics, 1999

    1999

  32. [40]

    Stochastic normalizing flows for in- verse problems: a markov chains viewpoint

    Paul Hagemann, Johannes Hertrich, and Gabriele Steidl. Stochastic normalizing flows for in- verse problems: a markov chains viewpoint. SIAM/ASA Journal on Uncertainty Quantifica- tion, 2022

    2022

  33. [41]

    exoplanet- dev/jaxoplanet: Astronomical time series anal- ysis with JAX, 2024

    Soichiro Hattori, Lionel Garcia, Catriona Mur- ray, Jiayin Dong, Shashank Dholakia, David De- gen, and Daniel Foreman-Mackey. exoplanet- dev/jaxoplanet: Astronomical time series anal- ysis with JAX, 2024. URL https://doi.org/ 10.5281/zenodo.10736936

    2024 doi

  34. [42]

    Black-box alpha divergence min- imization

    Jose Hernandez-Lobato, Yingzhen Li, Mark Row- land, Thang Bui, Daniel Hernández-Lobato, and Richard Turner. Black-box alpha divergence min- imization. In ICML, 2016

    2016

  35. [43]

    Black-box vari- ational inference as a parametric approximation to langevin dynamics

    Matthew Hoffman and Yian Ma. Black-box vari- ational inference as a parametric approximation to langevin dynamics. InICML, 2020

    2020

  36. [44]

    Neutra-lizing bad geometry in hamil- tonian monte carlo using neural transport.arXiv preprint arXiv:1903.03704, 2019

    Matthew Hoffman, Pavel Sountsov, Joshua V Dillon, Ian Langmore, Dustin Tran, and Srinivas Vasudevan. Neutra-lizing bad geometry in hamil- tonian monte carlo using neural transport.arXiv preprint arXiv:1903.03704, 2019

    1903 arXiv

  37. [45]

    An adaptive-mcmc scheme for setting trajectory lengths in hamiltonian monte carlo

    Matthew Hoffman, Alexey Radul, and Pavel Sountsov. An adaptive-mcmc scheme for setting trajectory lengths in hamiltonian monte carlo. In AISTATS, 2021

    2021

  38. [46]

    Tuning- free generalized hamiltonian monte carlo

    Matthew D Hoffman and Pavel Sountsov. Tuning- free generalized hamiltonian monte carlo. In AISTATS, 2022

    2022

  39. [47]

    Tails of lipschitz triangular flows

    Priyank Jaini, Ivan Kobyzev, Yaoliang Yu, and Marcus Brubaker. Tails of lipschitz triangular flows. InICML, 2020

    2020

  40. [48]

    Stochastic varia- tional inference via upper bound.arXiv preprint arXiv:1912.00650, 2019

    Chunlin Ji and Haige Shen. Stochastic varia- tional inference via upper bound.arXiv preprint arXiv:1912.00650, 2019

    1912 arXiv

  41. [49]

    An introduc- tion to variational methods for graphical models

    Michael I Jordan, Zoubin Ghahramani, Tommi S Jaakkola, and Lawrence K Saul. An introduc- tion to variational methods for graphical models. Machine learning, 1999

    1999

  42. [50]

    Markov chain score ascent: A unifying framework of vari- ational inference with markovian gradients

    Kyurae Kim, Jisu Oh, Jacob Gardner, Adji Bousso Dieng, and Hongseok Kim. Markov chain score ascent: A unifying framework of vari- ational inference with markovian gradients. In NeurIPS, 2022

    2022

  43. [51]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015

    2015

  44. [52]

    Auto- encoding variational bayes

    Diederik P Kingma and Max Welling. Auto- encoding variational bayes. InICLR, 2014

    2014

  45. [53]

    Improved variational inference with inverse au- toregressive flow

    Durk P Kingma, Tim Salimans, Rafal Jozefow- icz, Xi Chen, Ilya Sutskever, and Max Welling. Improved variational inference with inverse au- toregressive flow. InNeurIPS, 2016

    2016

  46. [54]

    Efficient, uninformative sam- pling of limb darkening coefficients for two- Abhinav Agrawal, Justin Domke parameter laws

    David M Kipping. Efficient, uninformative sam- pling of limb darkening coefficients for two- Abhinav Agrawal, Justin Domke parameter laws. Monthly Notices of the Royal Astronomical Society, 2013

    2013

  47. [55]

    Normalizing flows: An introduction and review of current methods.IEEE transac- tions on pattern analysis and machine intelli- gence, 2020

    Ivan Kobyzev, Simon JD Prince, and Marcus A Brubaker. Normalizing flows: An introduction and review of current methods.IEEE transac- tions on pattern analysis and machine intelli- gence, 2020

    2020

  48. [56]

    Automatic differentiation variational inference.JMLR, 2017

    Alp Kucukelbir, Dustin Tran, Rajesh Ranganath, Andrew Gelman, and David M Blei. Automatic differentiation variational inference.JMLR, 2017

    2017

  49. [57]

    Lao and J

    J. Lao and J. V. Dillon. Unrolled implementation of no-u-turn sam- pler. https://github.com/tensorflow/ probability/blob/master/discussion/ technical_note_on_unrolled_nuts.md, Au- gust 2019

    2019

  50. [58]

    Junpeng Lao, Christopher Suter, Ian Langmore, Cyril Chimisov, Ashish Saxena, Pavel Sountsov, Dave Moore, Rif A Saurous, Matthew D Hoff- man, and Joshua V Dillon. tfp. mcmc: Modern markov chain monte carlo tools built for mod- ern hardware.arXiv preprint arXiv:2002.01184, 2020

    2002 arXiv

  51. [59]

    Rényi diver- gence variational inference

    Yingzhen Li and Richard E Turner. Rényi diver- gence variational inference. InNeurIPS, 2016

    2016

  52. [60]

    Fat–tailed variational inference with anisotropic tail adaptive flows

    Feynman Liang, Michael Mahoney, and Liam Hodgkinson. Fat–tailed variational inference with anisotropic tail adaptive flows. InICML, 2022

    2022

  53. [61]

    posteriordb: Testing, bench- marking and developing bayesian inference algo- rithms

    Måns Magnusson, Jakob Torgander, Paul- Christian Bürkner, Lu Zhang, Bob Carpenter, and Aki Vehtari. posteriordb: Testing, bench- marking and developing bayesian inference algo- rithms. arXiv preprint arXiv:2407.04967, 2024

    2024 arXiv

  54. [62]

    Nested r: Assessing the convergence of markov chain monte carlo when running many short chains.arXiv preprint arXiv:2110.13017, 2021

    Charles C Margossian, Matthew D Hoffman, Pavel Sountsov, Lionel Riou-Durand, Aki Ve- htari, and Andrew Gelman. Nested r: Assessing the convergence of markov chain monte carlo when running many short chains.arXiv preprint arXiv:2110.13017, 2021

    arXiv

  55. [63]

    For how many iterations should we run markov chain monte carlo?arXiv preprint arXiv:2311.02726, 2023

    Charles C Margossian, Andrew Gelman, and Alexandre Dumas. For how many iterations should we run markov chain monte carlo?arXiv preprint arXiv:2311.02726, 2023

    2023 arXiv

  56. [64]

    An ordering of di- vergences for variational inference with factor- ized gaussian approximations

    Charles C Margossian, Loucas Pillaud-Vivien, and Lawrence K Saul. An ordering of di- vergences for variational inference with factor- ized gaussian approximations. arXiv preprint arXiv:2403.13748, 2024

    2024

  57. [65]

    Continual re- peated annealed flow transport monte carlo

    Alex Matthews, Michael Arbel, Danilo Jimenez Rezende, and Arnaud Doucet. Continual re- peated annealed flow transport monte carlo. In ICML, 2022

    2022

  58. [66]

    Reducingreparameterization gradient variance

    Andrew Miller, Nick Foti, Alexander D’Amour, andRyanPAdams. Reducingreparameterization gradient variance. InNeurIPS, 2017

    2017

  59. [67]

    Divergence measures and mes- sage passing

    Tom Minka et al. Divergence measures and mes- sage passing. Technical report, Technical report, Microsoft Research, 2005

    2005

  60. [68]

    Markovian score climbing: Variational in- ference with kl (p|| q)

    Christian Naesseth, Fredrik Lindsten, and David Blei. Markovian score climbing: Variational in- ference with kl (p|| q). InNeurIPS, 2020

    2020

  61. [69]

    Annealed importance sampling

    Radford M Neal. Annealed importance sampling. Statistics and computing, 2001

    2001

  62. [70]

    Slice sampling.The annals of statistics, 2003

    Radford M Neal. Slice sampling.The annals of statistics, 2003

    2003

  63. [71]

    Joint stochastic approximation and its application to learning discrete latent variable models

    Zhijian Ou and Yunfu Song. Joint stochastic approximation and its application to learning discrete latent variable models. InUAI, 2020

    2020

  64. [72]

    Masked autoregressive flow for density estimation

    George Papamakarios, Theo Pavlakou, and Iain Murray. Masked autoregressive flow for density estimation. In NeurIPS, 2017

    2017

  65. [73]

    Normalizing flows for probabilistic modeling and inference.JMLR, 2021

    George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshminarayanan. Normalizing flows for probabilistic modeling and inference.JMLR, 2021

    2021

  66. [74]

    Transport map accelerated markov chain monte carlo

    Matthew D Parno and Youssef M Marzouk. Transport map accelerated markov chain monte carlo. SIAM/ASA Journal on Uncertainty Quan- tification, 2018

    2018

  67. [75]

    Phan and N

    D. Phan and N. Pradhan. Iterative nuts. https://github.com/pyro-ppl/numpyro/ wiki/Iterative-NUTS, May 2019

    2019

  68. [76]

    Composable effects for flexible and ac- celerated probabilistic programming in numpyro

    Du Phan, Neeraj Pradhan, and Martin Jankowiak. Composable effects for flexible and ac- celerated probabilistic programming in numpyro. arXiv preprint arXiv:1912.11554, 2019

    1912 arXiv

  69. [77]

    Auto- matically batching control-intensive programs for modern accelerators

    Alexey Radul, Brian Patton, Dougal Maclaurin, Matthew Hoffman, and Rif A Saurous. Auto- matically batching control-intensive programs for modern accelerators. Proceedings of Machine Learning and Systems, 2020

    2020

  70. [78]

    Tighter Variational Bounds are Not Necessarily Better

    Tom Rainforth, Adam Kosiorek, Tuan Anh Le, Chris Maddison, Maximilian Igl, Frank Wood, and Yee Whye Teh. Tighter Variational Bounds are Not Necessarily Better. InICML, 2018

    2018

  71. [79]

    Black Box Variational Inference

    Rajesh Ranganath, Sean Gerrish, and David Blei. Black Box Variational Inference. InAISTATS, 2014

    2014

  72. [80]

    Varia- tional inference with normalizing flows

    Danilo Rezende and Shakir Mohamed. Varia- tional inference with normalizing flows. InICML, 2015. Disentangling impact of capacity, objective, batchsize, estimators, and step-size on flow VI

    2015

  73. [81]

    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 ICML, 2014

    2014

  74. [82]

    Var- grad: a low-variance gradient estimator for vari- ational inference

    Lorenz Richter, Ayman Boustati, Nikolas Nüsken, Francisco Ruiz, and Omer Deniz Akyildiz. Var- grad: a low-variance gradient estimator for vari- ational inference. InNeurIPS, 2020

    2020

  75. [83]

    Sticking the landing: Simple, lower- variance gradient estimators for variational infer- ence

    Geoffrey Roeder, Yuhuai Wu, and David Du- venaud. Sticking the landing: Simple, lower- variance gradient estimators for variational infer- ence. In NIPS, 2017

    2017

  76. [84]

    Pavi: Plate- amortized variational inference

    Louis Rouillard, Alexandre Le Bris, Thomas Moreau, and Demian Wassermann. Pavi: Plate- amortized variational inference. InTMLR, 2023

    2023

  77. [85]

    Local-global mcmc kernels: the best of both worlds

    Sergey Samsonov, Evgeny Lagutin, Marylou Gabrié, Alain Durmus, Alexey Naumov, and Eric Moulines. Local-global mcmc kernels: the best of both worlds. InNeurIPS, 2022

    2022

  78. [86]

    Mean field theory for sigmoid belief net- works

    Lawrence K Saul, Tommi Jaakkola, and Michael I Jordan. Mean field theory for sigmoid belief net- works. Journal of artificial intelligence research, 4:61–76, 1996

    1996

  79. [87]

    Reliable amortized varia- tional inference with physics-based latent distri- bution correction

    Ali Siahkoohi, Gabrio Rizzuti, Rafael Orozco, and Felix J Herrmann. Reliable amortized varia- tional inference with physics-based latent distri- bution correction. Geophysics, 2023

    2023

  80. [88]

    Inference gym, 2020

    Pavel Sountsov, Alexey Radul, and contributors. Inference gym, 2020. URL https://pypi.org/ project/inference_gym

    2020

  81. [89]

    Metflow: a new efficient method for bridging the gap between markov chain monte carlo and variational infer- ence

    Achille Thin, Nikita Kotelevskii, Jean-Stanislas Denain, Leo Grinsztajn, Alain Durmus, Maxim Panov, and Eric Moulines. Metflow: a new efficient method for bridging the gap between markov chain monte carlo and variational infer- ence. arXiv preprint arXiv:2002.12253, 2020

    2002 arXiv

  82. [90]

    Doubly reparameterized gradient estimators for monte carlo objectives

    George Tucker, Dieterich Lawson, Shixiang Gu, and Chris J Maddison. Doubly reparameterized gradient estimators for monte carlo objectives. ICLR, 2019

    2019

  83. [91]

    Gradients should stay on path: better estimators of the reverse-and forward kl divergence for normalizing flows.Machine Learn- ing: Science and Technology, 3(4):045006, 2022

    Lorenz Vaitl, Kim A Nicoli, Shinichi Nakajima, and Pan Kessel. Gradients should stay on path: better estimators of the reverse-and forward kl divergence for normalizing flows.Machine Learn- ing: Science and Technology, 3(4):045006, 2022

    2022

  84. [92]

    Fast and unified path gradient estimators for normalizing flows.arXiv preprint arXiv:2403.15881, 2024

    Lorenz Vaitl, Ludwig Winkler, Lorenz Richter, and Pan Kessel. Fast and unified path gradient estimators for normalizing flows.arXiv preprint arXiv:2403.15881, 2024

    2024 arXiv

  85. [93]

    Springer, 2009

    Cédric Villani et al.Optimal transport: old and new. Springer, 2009

    2009

  86. [94]

    Graphical models, exponential families, and vari- ational inference

    Martin J Wainwright, Michael I Jordan, et al. Graphical models, exponential families, and vari- ational inference. Foundations and Trends® in Machine Learning, 2008

    2008

  87. [95]

    Varia- tional inference with tail-adaptive f-divergence

    Dilin Wang, Hao Liu, and Qiang Liu. Varia- tional inference with tail-adaptive f-divergence. In NeurIPS, 2018

    2018

  88. [96]

    Improving au- tomated variational inference with normalizing flows

    Stefan Webb, Jonathan P Chen, Martin Jankowiak, and Noah Goodman. Improving au- tomated variational inference with normalizing flows. InICML Workshop on Automated Machine Learning, 2019

    2019

  89. [97]

    Manushi Welandawe, Michael Riis Andersen, Aki Vehtari, and Jonathan H. Huggins. A frame- work for improving the reliability of black-box variational inference.JMLR, 2024

    2024

  90. [98]

    Stochas- tic normalizing flows

    Hao Wu, Jonas Köhler, and Frank Noé. Stochas- tic normalizing flows. InNeurIPS, 2020

    2020

  91. [99]

    Mixflows: principled variational inference via mixed flows

    Zuheng Xu, Naitong Chen, and Trevor Campbell. Mixflows: principled variational inference via mixed flows. InICML, 2023

    2023

  92. [100]

    Bridging the gap between variational inference and wasserstein gra- dient flows

    Mingxuan Yi and Song Liu. Bridging the gap between variational inference and wasserstein gra- dient flows. arXiv preprint arXiv:2310.20090, 2023

    2023 arXiv

  93. [101]

    Differ- entiable annealed importance sampling mini- mizes the jensen-shannon divergence between initial and target distribution

    Johannes Zenn and Robert Bamler. Differ- entiable annealed importance sampling mini- mizes the jensen-shannon divergence between initial and target distribution. arXiv preprint arXiv:2405.14840, 2024. Checklist

    2024 arXiv

  94. [102]

    [Not Applicable] (b) An analysis of the properties and complexity (time, space, sample size) of any algorithm

    For all models and algorithms presented, check if you include: (a) A clear description of the mathematical set- ting, assumptions, algorithm, and/or model. [Not Applicable] (b) An analysis of the properties and complexity (time, space, sample size) of any algorithm. [Not Appli...

  95. [103]

    [Not Applicable] (b) Completeproofsofalltheoreticalresults

    For any theoretical claim, check if you include: (a) Statements of the full set of assumptions of all theoretical results. [Not Applicable] (b) Completeproofsofalltheoreticalresults. [Not Applicable] (c) Clear explanations of any assumptions. [Not Applicable] Abhinav Agrawal, ...

  96. [104]

    [Yes‘’] (b) All the training details (e.g., data splits, hy- perparameters, how they were chosen)

    For all figures and tables that present empirical results, check if you include: (a) The code, data, and instructions needed to re- produce the main experimental results (either in the supplemental material or as a URL). [Yes‘’] (b) All the training details (e.g., data splits,...

  97. [105]

    [Yes] (b) The license information of the assets, if appli- cable

    If you are using existing assets (e.g., code, data, models) or curating/releasing new assets, check if you include: (a) Citations of the creator If your work uses existing assets. [Yes] (b) The license information of the assets, if appli- cable. [Not Applicable] (c) Newassetse...

  98. [106]

    [Not Applicable] (b) Descriptions of potential participant risks, withlinkstoInstitutionalReviewBoard(IRB) approvals if applicable

    If you used crowdsourcing or conducted research with human subjects, check if you include: (a) The full text of instructions given to partici- pants and screenshots. [Not Applicable] (b) Descriptions of potential participant risks, withlinkstoInstitutionalReviewBoard(IRB) appr...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.