Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

Mitigating mode collapse in normalizing flows by annealing with an adaptive schedule: Application to parameter estimation

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Adaptive annealing driven by the effective sample size prevents normalizing flows from collapsing to a single mode and achieves a ten-fold speedup over ensemble MCMC on a Bayesian parameter-estimation problem.

desk verdict A useful but under-supported recipe for annealing normalizing flows via ESS, with a plausible ten-fold speedup on one problem; worth refereeing if the authors add code, seeds, and a sharper justification for the threshold. read the letter →

arxiv 2505.03652 v1 pith:2J2TKSQZ submitted 2025-05-06 cs.LG physics.comp-phphysics.data-anq-bio.QMstat.ML

Yihang Wang , Chris Chi , Aaron R. Dinner This is my paper · ORCID
classification cs.LGphysics.comp-phphysics.data-anq-bio.QMstat.ML
keywords normalizingflowsmodecollapseadaptiveannealingeffectivesamplesizeBayesianparameterestimationmarginallikelihoodthermodynamicintegrationrepressilatormodel
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

Normalizing flows can generate independent samples from complex distributions, but when the target is multimodal they often collapse onto a single mode. This paper argues that annealing from the prior to the posterior, with the temperature schedule set adaptively by the effective sample size, prevents that collapse without knowing the modes in advance. The idea is demonstrated on an eight-parameter Bayesian fit of a repressilator oscillator model, whose posterior has three symmetry-related modes. With the best hyperparameters, the flow-based sampler converges the marginal likelihood in about 5.2 hours, roughly ten times faster than an ensemble MCMC baseline requiring about 58 hours. The same sample-quality measure also serves to prune high-weight samples and reduce the variance of the model-evidence estimate.

What carries the argument

The central mechanism is the effective sample size (ESS) of reweighted flow samples, used both as a progress meter and as a sample-pruning criterion. For a batch of samples from the current flow $q_\phi$, the ESS is $n_\mathrm{eff}(\beta)=(\sum_i w_i)^2/\sum_i w_i^2$ with $w_i=\hat p_\beta(x_i)/q_\phi(x_i)$; a high ESS indicates that the flow's density matches the annealed target well. The algorithm maintains an exponential moving average $\bar n_\mathrm{eff}$ and advances $\beta$ whenever $\bar n_\mathrm{eff}>n^*$, choosing the next value by root-finding so that $n_\mathrm{eff}(\beta_{s+1})=0.95\,n_\mathrm{eff}(\beta_s)$. Around this trigger sits a RealNVP network trained with the forward KL loss in importance-sampling form, and a mixture model that pools samples from all previous stages so that earlier likelihood evaluations are reused.

What would settle it

A direct test: on a two-mode target with one mode carrying 95 percent of the mass, run the algorithm with $n^*$ small enough that the moving-average ESS crosses the threshold while the flow still assigns negligible probability to the light mode. If $\beta$ advances to 1 and the light mode is never sampled, the ESS trigger has failed to detect an uncovered mode, and repeating the training with several random seeds would show whether the failure is systematic.

Watch

Extended reading notes

Core claim

The paper's central claim is that mode collapse in normalizing flows can be mitigated by training the flow against a sequence of intermediate targets $\hat p_\beta(\theta)=p(\theta)p(D\mid\theta)^\beta$, advancing $\beta$ upward only when an exponential moving average of the effective sample size crosses a threshold $n^*$. At each stage the flow is trained with the mode-covering forward KL divergence, and samples from earlier stages are pooled into a mixture density for reweighting. On the repressilator problem this schedule captures all three posterior modes, whereas fixed-$\beta$ training and a preset power schedule collapse or stall. The paper also shows that using the same effective-sample-size statistic to discard the largest-weight samples stabilizes importance-sampling estimates of the marginal likelihood, and that thermodynamic integration over the adaptive ladder gives estimates consistent with MCMC. The reported result is a roughly ten-fold reduction in computation time compared with a widely used ensemble MCMC sampler on this problem.

Load-bearing premise

The schedule rests on the assumption that the effective sample size crossing the threshold $n^*$ means the current flow has captured every important mode of the intermediate target; if the ESS can recover while a mode is still missing, advancing $\beta$ will lock out that mode.

Editorial extensions

If this is right

  • For expensive-likelihood problems, the method turns the annealing schedule into a data-driven decision: computation is spent where the flow has not yet caught up, rather than on a fixed temperature grid.
  • No foreknowledge of the modes is required, so the scheme applies to posteriors whose multimodality is discovered during sampling instead of anticipated from symmetry.
  • The ESS-based pruning rule provides a post-training variance-reduction step that needs no additional likelihood evaluations.
  • The same annealed flow yields marginal-likelihood estimates by two independent routes, importance sampling and thermodynamic integration, and their consistency supports using either route for model comparison.
  • Because only likelihood values are needed, the approach extends to non-differentiable likelihoods or black-box simulators.

Reading between the lines

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

  • Beyond the paper, the sensitivity of the schedule to the threshold $n^*$ suggests a practical extension: run a cheap pilot with a low threshold to locate the $\beta$ values where the ESS dips, then spend additional model capacity there; the paper's observed slowdown near $\beta\approx0.06$ is exactly the signature such a pilot would exploit.
  • The ESS trigger could plausibly be ported to other generative samplers that anneal a temperature or noise level, such as diffusion-model samplers, replacing hand-designed schedules with the same reweighting-based criterion.
  • A stress test not reported in the paper would vary the relative weights of the three repressilator modes; an ESS threshold tuned to one mode-weight ratio may miss a barely populated mode, and mapping that boundary would make the method's failure mode quantitative.
  • A natural continuation would benchmark the same adaptive schedule against sequential Monte Carlo samplers, which also build annealed particle populations and produce evidence estimates.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes an adaptive annealing schedule for training normalizing flows (NFs) with the forward-KL loss. The annealing parameter β is advanced when the exponential moving average of the effective sample size (ESS) of importance weights crosses a threshold n*, with the next β value chosen so that the ESS drops by a fixed factor γ (Algorithms 2 and 3). The method is applied to Bayesian parameter estimation for a repressilator ODE model with eight parameters and a three-mode posterior. The authors report that the adaptive-annealing NF scheme captures all three modes, estimates the marginal likelihood via importance sampling and thermodynamic integration, and achieves roughly a ten-fold speedup over an ensemble MCMC baseline (Table 1). The paper also proposes pruning large-weight samples to stabilize importance-sampling estimates of the marginal likelihood.

Significance. If the results hold, the paper offers a practical recipe for using NFs to sample multimodal posteriors without knowing the modes in advance, which would be valuable for Bayesian parameter estimation in systems biology and other fields where likelihoods are expensive. The thermodynamic-integration identity is standard and correctly derived, and the algorithmic description is clear enough to reproduce. The main strengths are the concrete numerical demonstration on a challenging multimodal ODE model and the explicit comparison with a widely used MCMC method. However, the central claim that the ESS threshold indicates mode coverage is not theoretically justified, and the performance claims rest on limited runs without seed variance, so the significance of the contribution depends on whether these gaps can be closed.

major comments (4)
  1. [§2.3, Algorithm 2, Eq. (9)] The premise that crossing an ESS threshold indicates that the current flow has captured all relevant modes is not established. The ESS in Eq. (9) is computed from samples drawn from q_phi; if q_phi has no mass in a mode of p_beta, no samples fall there, that mode contributes no weight, and the ESS can remain high while the mode is entirely unrepresented. The paper's own Supplementary Fig. S2 shows that a too-small threshold (n*/N = 0.2) misses modes even with annealing, and no argument connects the chosen thresholds (n*/N = 0.4 or 0.6) to mode discovery. Please provide a diagnostic or a theoretical argument that links the ESS threshold to mode coverage, or temper the Abstract and Section 1 claim that the scheme works 'without prior knowledge of the modes.'
  2. [Table 1, §4.3] The speedup claim rests on single runs for each NF hyperparameter setting; no seed variance or error bars are reported for the NF computation times or the log marginal likelihoods. Given the known sensitivity of NF training to initialization shown in Fig. 3b, where fixed-β training collapses to different modes across seeds, repeated independent runs are needed to establish that the 5.2-hour run is typical rather than a favorable outlier. Without such replicates, the 'ten-fold speedup' in the Abstract is not yet supported to the standard expected for a performance claim.
  3. [§4.5, Fig. 4, Table 1] The importance-sampling estimates of the marginal likelihood reported in Table 1 are obtained after excluding the largest-weight samples to maximize the ESS. As the authors acknowledge, this post-hoc selection biases the estimator. Because these biased numbers are presented alongside the unbiased thermodynamic-integration estimates, please report the unbiased importance-sampling estimates as well and justify why the exclusions do not materially affect the conclusions about model evidence. At present, the discrepancy between the two estimators (e.g., −35.61 vs −35.90 for the fastest run) is partly explainable by this acknowledged bias, which weakens the claim of 'reliable estimates.'
  4. [§4.4, §4.3] The MCMC comparison uses a fixed Friel–Pettitt annealing schedule (β_s = (s/1000)^4), whereas the NF method uses the proposed adaptive schedule. The reported speedup therefore conflates the benefit of the adaptive schedule with the benefit of using an NF. The paper already shows in Supplementary Fig. S1 that a fixed schedule with an NF can cause mode collapse, so the adaptive schedule is doing important work. To isolate the method's contribution, either include an MCMC run using an adaptive schedule or an NF run using the fixed schedule, and explicitly discuss the confound in Section 4.4.
minor comments (6)
  1. [Fig. 3d caption] The word 'exponetial' should be 'exponential'.
  2. [§4.3] The phrase 'the network can fail to reach to reach the threshold' contains a duplicated 'to reach'; please correct.
  3. [Reference [27]] The journal name 'Pattern Analysis and MNachine Intelligence' should be 'Machine'.
  4. [Eq. (18) and §4.1] Notation is inconsistent: the data are denoted D(t) and the model solution X(t,θ), but later X(t) is used without θ; please standardize and define the observable function D-hat(X) clearly.
  5. [Algorithm 3, line 2] The initialization 'p_p(x)' appears to be a typo for 'p_b(x)', the base distribution.
  6. [Algorithm 3, §2.4] The symbol k is used both as the NF model index and (implicitly) in the batch loop; please use distinct symbols to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation is self-contained and benchmarked against an external MCMC baseline.

full rationale

The paper's central claim is empirical rather than definitional: the adaptive annealing schedule advances beta when the exponential moving average of the effective sample size crosses a threshold, and the effective sample size in Eq. (9) is computed from importance weights without reference to the final marginal likelihood, mode count, or the reported speedup. The marginal likelihood estimators in Eqs. (15) and (16) are standard mathematical identities derived in the supplementary material, not conclusions smuggled in through their own inputs. The comparison with ensemble MCMC provides an external quantitative benchmark. The post-hoc exclusion of high-weight samples is an acknowledged ad hoc bias-variance adjustment, not a fitted parameter renamed as a prediction. The paper contains self-citations, but they are background references and are not load-bearing for the main result. Consequently, no circular step can be exhibited, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard statistical identities (change of variables, importance sampling, thermodynamic integration) plus several heuristics: the ESS threshold as a mode-coverage diagnostic, the mixture proposal, and hand-set hyperparameters. The paper invents no new physical entities. The most fragile element is the ESS-to-mode-coverage assumption, which is exactly the control mechanism for the annealing schedule.

free parameters (5)
  • ESS threshold n*/N = 0.4 and 0.6 (tested)
    Controls when beta advances. The paper shows a tradeoff: too small misses modes, too large stalls training. Chosen by hand, not derived.
  • ESS discount factor gamma = 0.95
    Target ratio of new to old ESS when solving for the next beta. A heuristic that sets the step size of the schedule.
  • EMA decay lambda = 0.01
    Smoothing factor for the noisy ESS. Chosen heuristically and fixed in all runs.
  • NF architecture and training hyperparameters (L, J, learning rate) = L=8 or 16, J=30 or 50, lr=1e-4
    Capacity and update frequency are tuned; the fastest run uses L=8 and J=50.
  • ODE failure penalty (observable set to 200) = 200
    When the ODE solver fails, the observable is assigned a value of 200, which makes the likelihood tiny. This arbitrary choice affects weights and thermodynamic integration near small beta.
assumptions (6)
  • standard math The change-of-variables formula gives the exact density of the normalizing flow.
    Used in Eq. (2) to compute q_phi and the importance weights.
  • domain assumption The self-normalized importance-weighted gradient is a valid estimator of the forward-KL gradient.
    Algorithm 1 and Eq. (8) rely on finite-sample weighted averages; the paper does not discuss the finite-sample bias of self-normalized importance sampling.
  • domain assumption Powering the likelihood by beta, p_beta = prior * likelihood^beta, connects the prior to the posterior without destroying or creating modes in a way that breaks training.
    The annealing construction in Section 2.3 assumes that gradual beta increase lets the flow discover modes; no formal property is proved.
  • ad hoc to paper The exponential moving average of ESS crossing the threshold n* indicates that the flow has converged to the current target and can advance to a harder one.
    This is the control rule in Algorithms 2 and 3; it is a heuristic with no supporting theorem or diagnostic.
  • domain assumption The mixture proposal q_m from Eq. (11) is normalized and provides a reliable proposal for reweighting at all beta values.
    Used in Eq. (12) to compute weights; assumes the variance from reweighting old samples is manageable.
  • standard math The thermodynamic integration relation in Eq. (16) holds with reweighted NF samples and trapezoidal discretization.
    The derivation in Supplementary Section 1 is standard; the paper implicitly assumes discretization and reweighting errors are small.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating mode collapse in normalizing flows by annealing with an adaptive schedule: Application to parameter estimation." pith.science (2026). https://pith.science/paper/2J2TKSQZ

@misc{pith2026250503652,
  author       = {Pith},
  title        = {Pith review of: Mitigating mode collapse in normalizing flows by annealing with an adaptive schedule: Application to parameter estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2J2TKSQZ}},
  note         = {Machine review of arXiv:2505.03652}
}
read the original abstract

Normalizing flows (NFs) provide uncorrelated samples from complex distributions, making them an appealing tool for parameter estimation. However, the practical utility of NFs remains limited by their tendency to collapse to a single mode of a multimodal distribution. In this study, we show that annealing with an adaptive schedule based on the effective sample size (ESS) can mitigate mode collapse. We demonstrate that our approach can converge the marginal likelihood for a biochemical oscillator model fit to time-series data in ten-fold less computation time than a widely used ensemble Markov chain Monte Carlo (MCMC) method. We show that the ESS can also be used to reduce variance by pruning the samples. We expect these developments to be of general use for sampling with NFs and discuss potential opportunities for further improvements.

Figures

Figures reproduced from arXiv: 2505.03652 by the authors.

Figure 1
Figure 1. The annealing-sampling scheme. The parameter β increases to interpolate from a simple distribution that is close to the base distribution (top left) to the ultimate target distribution (top right). At each value of β, samples generated from previously trained NFs are reweighted and used to train a new NF, from which samples are then drawn. learned distribution qϕ(xi), as defined above. If the density learned by the … view at source ↗
Figure 2
Figure 2. Repressilator model. (a) Schematic of the system; each circle represents a gene product, and i ⊣ j represents repression of j by i. (b) Solution used to generate the data for fitting; the parameter values are Xi(t0) = 2 for all i, α1 = 10, α2 = 15, α3 = 20, m = 4, and η = 1. (c) Time series of the total concentration of gene products (blue line) and the simulated observable (orange dots) produced by adding Gaussian … view at source ↗
Figure 3
Figure 3. Annealing with a schedule based on the effective sample size (ESS) mitigates mode collapse. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Estimating marginal likelihoods. (a, c) Marginal likelihood estimates for different sample sizes. Shaded regions indicate standard deviations from 10 independent sets of samples. In (a), all samples are used for the estimates. In (c), samples are excluded to maximize t…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. NAE: Normalizing AutoEncoder

    cs.LG 2026-08 conditional novelty 6.0 of 10

    A conditional surrogate loss that always picks the gradient estimate aligned with the reconstruction loss improves flow autoencoder training and reaches state-of-the-art generative performance on molecules, tabular da...

  2. Leveraging generative models to assist Monte Carlo sampling

    stat.ML 2026-08 conditional novelty 1.0 of 10

    This paper is a tutorial review, not a research contribution: it organizes existing methods for using generative models as proposal distributions, transport maps, and annealing bridges in Monte Carlo sampling.

Reference graph

Works this paper leans on

52 extracted references · 46 canonical work pages · cited by 2 Pith papers

  1. [1]

    S., Boffi, N

    Albergo, M. S., Boffi, N. M., Vanden-Eijnden, E., 2023. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv:2303.08797

    arXiv 2023

  2. [2]

    K., Beck, J

    Au, S. K., Beck, J. L., Apr 2003. Importance sampling in high dimensions. Structural Safety 25 (2), 139–163

    work page 2003

  3. [3]

    S., 2021

    Bois, J. S., 2021. justinbois/biocircuits: Version 0.1.0

    work page 2021

  4. [4]

    Statistics and Computing 16, 239–249

    Braak, C.J.T., 2006.AMarkovchainMonteCarloversionofthegeneticalgorithmdifferentialevolution: easy Bayesian computing for real parameter spaces. Statistics and Computing 16, 239–249

    work page 2006

  5. [5]

    P., Louis, T

    Carlin, B. P., Louis, T. A., 2008. Bayesian Methods for Data analysis. CRC press

    work page 2008

  6. [6]

    Sampling parameters of ordinary differential equations with Langevin dynamics that satisfy constraints

    Chi, C., Weare, J., Dinner, A. R., 2024. Sampling parameters of ordinary differential equations with Langevin dynamics that satisfy constraints. arXiv:2408.15505

    work page Pith review arXiv 2024

  7. [7]

    Free energy methods for Bayesian inference: efficient explo- ration of univariate gaussian mixture posteriors

    Chopin, N., Lelièvre, T., Stoltz, G., 2012. Free energy methods for Bayesian inference: efficient explo- ration of univariate gaussian mixture posteriors. Statistics and Computing 22, 897–916

    work page 2012

  8. [8]

    A., 2024

    Coretti, A., Falkner, S., Weinreich, J., Dellago, C., von Lilienfeld, O. A., 2024. Boltzmann generators and the new frontier of computational sampling in many-body systems. KIM Review 2, 3

    work page 2024

Show all 52 references
  1. [9]

    Density estimation using real NVP

    Dinh, L., Sohl-Dickstein, J., Bengio, S., 2017. Density estimation using real NVP. In: International Conference on Learning Representations

    2017

  2. [10]

    R., Thiede, E

    Dinner, A. R., Thiede, E. H., Koten, B. V., Weare, J., 2020. Stratification as a general variance reduction method for Markov chain Monte Carlo. SIAM/ASA Journal on Uncertainty Quantification 8 (3), 1139– 1188

    2020

  3. [11]

    Neural spline flows

    Durkan, C., Bekasov, A., Murray, I., Papamakarios, G., 2019. Neural spline flows

    2019

  4. [12]

    B., Leibler, S., 2000

    Elowitz, M. B., Leibler, S., 2000. A synthetic oscillatory network of transcriptional regulators. Nature 403 (6767), 335–338

    2000

  5. [13]

    W., Lang, D., Goodman, J., 2013

    Foreman-Mackey, D., Hogg, D. W., Lang, D., Goodman, J., 2013. emcee: the MCMC hammer. Publi- cations of the Astronomical Society of the Pacific 125 (925), 306

    2013

  6. [14]

    N., 2008

    Friel, N., Pettitt, A. N., 2008. Marginal likelihood estimation via power posteriors. Journal of the Royal Statistical Society Series B: Statistical Methodology 70 (3), 589–607

    2008

  7. [15]

    M., Vanden-Eijnden, E., 2021

    Gabrié, M., Rotskoff, G. M., Vanden-Eijnden, E., 2021. Efficient Bayesian sampling using normalizing flows to assist Markov chain Monte Carlo methods. arXiv:2107.08001

    2021 arXiv

  8. [16]

    M., Vanden-Eijnden, E., 2022

    Gabrié, M., Rotskoff, G. M., Vanden-Eijnden, E., 2022. Adaptive Monte Carlo augmented with normal- izing flows. Proceedings of the National Academy of Sciences 119 (10), e2109420119

    2022

  9. [17]

    J., 1992

    Geyer, C. J., 1992. Practical Markov chain Monte Carlo. Statistical Science, 473–483

    1992

  10. [18]

    Ensemble samplers with affine invariance

    Goodman, J., Weare, J., 2010. Ensemble samplers with affine invariance. Communications in Applied Mathematics and Computational Science 5 (1), 65–80

    2010

  11. [19]

    T., Bettencourt, J., Sutskever, I., Duvenaud, D., 2018

    Grathwohl, W., Chen, R. T., Bettencourt, J., Sutskever, I., Duvenaud, D., 2018. Ffjord: Free-form continuous dynamics for scalable reversible generative models. arXiv:1810.01367

    2018 arXiv

  12. [20]

    Deterministic Langevin Monte Carlo with normalizing flows for Bayesian inference

    Grumitt, R., Dai, B., Seljak, U., 2022. Deterministic Langevin Monte Carlo with normalizing flows for Bayesian inference. Advances in Neural Information Processing Systems 35, 11629–11641

    2022

  13. [21]

    N., Waterfall, J

    Gutenkunst, R. N., Waterfall, J. J., Casey, F. P., Brown, K. S., Myers, C. R., Sethna, J. P., 2007. Universally sloppy parameter sensitivities in systems biology models. PLoS Computational Biology 3 (10), e189

    2007

  14. [22]

    C., Hsieh, C.-C., Albergo, M

    Hackett, D. C., Hsieh, C.-C., Albergo, M. S., Boyda, D., Chen, J.-W., Chen, K.-F., Cranmer, K., Kanwar, G., Shanahan, P. E., 2021. Flow-based sampling for multimodal distributions in lattice field theory. arXiv:2107.00734

    2021 arXiv

  15. [23]

    Geometric insight into the challenges of solving high-dimensional reliability problems

    Katafygiotis, L., Zuev, K., 2008. Geometric insight into the challenges of solving high-dimensional reliability problems. Probabilistic Engineering Mechanics 23 (2), 208–218

    2008

  16. [24]

    P., Ba, J., 2014

    Kingma, D. P., Ba, J., 2014. Adam: A method for stochastic optimization. arXiv:1412.6980

    2014 arXiv

  17. [25]

    P., Dhariwal, P., 2018

    Kingma, D. P., Dhariwal, P., 2018. Glow: Generative flow with invertible 1x1 convolutions. Advances in Neural Information Processing Systems 31

    2018

  18. [26]

    D., Vecchi, M

    Kirkpatrick, S., Gelatt, C. D., Vecchi, M. P., 1983. Optimization by simulated annealing. Science 220 (4598), 671–680

    1983

  19. [27]

    J., Brubaker, M

    Kobyzev, I., Prince, S. J., Brubaker, M. A., 2020. Normalizing flows: An introduction and review of current methods. IEEE Transactions on Pattern Analysis and MNachine Intelligence 43 (11), 3964–3979

    2020

  20. [28]

    Computing Bayes factors using thermodynamic integration

    Lartillot, N., Philippe, H., 04 2006. Computing Bayes factors using thermodynamic integration. Sys- tematic Biology 55 (2), 195–207

    2006

  21. [29]

    S., 1996

    Liu, J. S., 1996. Metropolized independent sampling with comparisons to rejection sampling and im- portance sampling. Statistics and Computing 6, 113–119

    1996

  22. [30]

    Learning interpolations between Boltzmann densities

    Máté, B., Fleuret, F., 2023. Learning interpolations between Boltzmann densities. Transactions on Machine Learning Research

    2023

  23. [31]

    Umbrella sampling: A powerful method to sample tails of distributions

    Matthews, C., Weare, J., Kravtsov, A., Jennings, E., 2018. Umbrella sampling: A powerful method to sample tails of distributions. Monthly Notices of the Royal Astronomical Society 480 (3), 4069–4079

    2018

  24. [32]

    Enhanced sampling with machine learning

    Mehdi, S., Smith, Z., Herron, L., Zou, Z., Tiwary, P., 2024. Enhanced sampling with machine learning. Annual Review of Physical Chemistry 75

    2024

  25. [33]

    W., Rosenbluth, M

    Metropolis, N., Rosenbluth, A. W., Rosenbluth, M. N., Teller, A. H., Teller, E., 1953. Equation of state calculations by fast computing machines. Journal of Chemical Physics 21 (6), 1087–1092

    1953

  26. [34]

    I., Stimper, V., Simm, G

    Midgley, L. I., Stimper, V., Simm, G. N. C., Schölkopf, B., Hernández-Lobato, J. M., 2023. Flow annealed importance sampling bootstrap. In: The Eleventh International Conference on Learning Rep- resentations

    2023

  27. [35]

    B., Payne, M

    Nelson, B., Ford, E. B., Payne, M. J., 2013. RUN DMC: an efficient, parallel code for analyzing radial velocity observations usingN-body integrations and differential evolution Markov chain Monte carlo. The Astrophysical Journal Supplement Series 210 (1), 11

    2013

  28. [36]

    A., Anders, C

    Nicoli, K. A., Anders, C. J., Hartung, T., Jansen, K., Kessel, P., Nakajima, S., Dec 2023. Detecting and mitigating mode-collapse for flow-based sampling of lattice field theories. Phys. Rev. D 108, 114501

    2023

  29. [37]

    A., Nakajima, S., Strodthoff, N., Samek, W., Müller, K.-R., Kessel, P., 2020

    Nicoli, K. A., Nakajima, S., Strodthoff, N., Samek, W., Müller, K.-R., Kessel, P., 2020. Asymptotically unbiased estimation of physical observables with neural samplers. Physical Review E 101 (2), 023304

    2020

  30. [38]

    Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning

    Noé, F., Olsson, S., Köhler, J., Wu, H., 2019. Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning. Science 365 (6457), eaaw1147

    2019

  31. [39]

    J., Mohamed, S., Lakshminarayanan, B., 2021

    Papamakarios, G., Nalisnick, E., Rezende, D. J., Mohamed, S., Lakshminarayanan, B., 2021. Normal- izing flows for probabilistic modeling and inference. Journal of Machine Learning Research 22 (57), 1–64

    2021

  32. [40]

    Masked autoregressive flow for density estimation

    Papamakarios, G., Pavlakou, T., Murray, I., 2018. Masked autoregressive flow for density estimation. arXiv:1705.07057

    2018 arXiv

  33. [41]

    On the difficulty of training recurrent neural networks

    Pascanu, R., Mikolov, T., Bengio, Y., 2013. On the difficulty of training recurrent neural networks. In: International Conference on Machine Learning. PMLR, pp. 1310–1318

    2013

  34. [42]

    H., Teukolsky, S

    Press, W. H., Teukolsky, S. A., Vetterling, W. T., Flannery, B. P., 2007. Numerical Recipes 3rd Edition: The Art of Scientific Computing. Cambridge University Press, USA

    2007

  35. [43]

    J., 1997

    Richardson, S., Green, P. J., 1997. On Bayesian analysis of mixtures with an unknown number of components (with discussion). Journal of the Royal Statistical Society Series B: Statistical Methodology 59 (4), 731–792

    1997

  36. [44]

    R., 2017

    Shirts, M. R., 2017. Reweighting from the mixture distribution as a better way to describe the multistate bennett acceptance ratio. arXiv:1704.00891

    2017 arXiv

  37. [45]

    Monte Carlo methods in statistical mechanics: foundations and new algorithms

    Sokal, A., 1997. Monte Carlo methods in statistical mechanics: foundations and new algorithms. In: Functional Integration: Basics and Applications. Springer, pp. 131–192

    1997

  38. [46]

    Fixed-kinetic neural Hamiltonian flows for enhanced interpretability and reduced complexity

    Souveton, V., Guillin, A., Jasche, J., Lavaux, G., Michel, M., 2024. Fixed-kinetic neural Hamiltonian flows for enhanced interpretability and reduced complexity. In: International Conference on Artificial Intelligence and Statistics. PMLR, pp. 3178–3186

    2024

  39. [47]

    K., Machta, B

    Transtrum, M. K., Machta, B. B., Brown, K. S., Daniels, B. C., Myers, C. R., Sethna, J. P., 2015. Perspective: Sloppiness and emergent theories in physics, biology, and beyond. Journal of Chemical Physics 143 (1)

    2015

  40. [48]

    Runge–Kutta pairs of order 5(4) satisfying only the first column simplifying as- sumption

    Tsitouras, C., 2011. Runge–Kutta pairs of order 5(4) satisfying only the first column simplifying as- sumption. Computers & Mathematics with Applications 62 (2), 770–775

    2011

  41. [49]

    A., Nakajima, S., Kessel, P., 2022

    Vaitl, L., Nicoli, K. A., Nakajima, S., Kessel, P., 2022. Gradients should stay on path: better estima- tors of the reverse-and forward KL divergence for normalizing flows. Machine Learning: Science and Technology 3 (4), 045006

    2022

  42. [50]

    Wang, Y., Ribeiro, J. M. L., Tiwary, P., 2020. Machine learning approaches for analyzing and enhancing molecular dynamics simulations. Current Opinion in Structural Biology 61, 139–145. Supplementary Materials

    2020

  43. [51]

    Thermodynamic integration Here we derive the formula for thermodynamic integration (16). Given the unnormalized target distri- bution ˆPβ(θ) = P (D| θ,M )βP (θ,M ), we define the normalization constant (partition function)Zβ and normalized distributionPβ: Zβ = Z ˆPβ(θ)dθ and P...

  44. [52]

    For each parameter, we computed the ESS using the formula: neff = N 1 + 2P∞ τ=1ρk

    Effective sample size and acceptance rates in MCMC The ESS of MCMC measures that number of independent samples effectively obtained from a correlated chain [45, 5]. For each parameter, we computed the ESS using the formula: neff = N 1 + 2P∞ τ=1ρk . (23) Here N is the total num...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.