Pith. sign in

REVIEW 3 major objections 4 minor 51 references

Jarzynski Reweighting and Sampling Dynamics for Training Energy-Based Models: Theoretical Analysis of Different Transition Kernels

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper proves that reweighted out-of-equilibrium paths give unbiased estimates of energy-based-model gradients and partition functions under any Markovian sampler, with correction weights that vanish in the continuous limit when the…

desk verdict A correct but modest extension of the author's own Jarzynski EBM result; the drift-dependent reweighting identity holds, but the O(h^{3/2}) claim is unsupported and the paper overclaims without experiments. read the letter →

arxiv 2506.07843 v1 pith:LIJLI2UM submitted 2025-06-09 cs.LG physics.comp-ph

Davide Carbone This is my paper
classification cs.LGphysics.comp-ph MSC 60J2268T07
keywords energy-basedmodelsJarzynskireweightingcross-entropytrainingtransitionkernelsstochasticinterpolantscontrastivedivergencerestrictedBoltzmannmachinesunadjustedLangevinalgorithm
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper seeks to establish that Jarzynski reweighting—the nonequilibrium identity that corrects out-of-equilibrium path averages—can supply unbiased estimates of the expectations and normalization constants needed to train energy-based models, no matter which Markovian sampler drives the sample path. It proves a general identity for any time-dependent Markovian transition kernel, then specializes to the Euler–Maruyama discretization of a diffusion with an additional drift, showing the identity holds exactly for every finite step size and that the correction weights vanish in the continuous limit when the drift satisfies a density-matching condition. The same weight construction is written out explicitly for the Gibbs samplers of Bernoulli and Gaussian–Bernoulli restricted Boltzmann machines. A reader should care because the result turns a heuristic gradient approximation, in the spirit of contrastive divergence but without its bias, into a normalization-aware estimator usable in both generation and training.

What carries the argument

The carrying object is the path weight $A_k$, built recursively by $A_{k+1}=A_k+U_k(X_k)-U_{k+1}(X_{k+1})+\log[\pi_k(X_k,X_{k-1})/\pi_{k-1}(X_{k-1},X_k)]$ (with the drift-generalized increment $\zeta_k$ in Proposition 2). It is the log ratio of the forward path probability to a time-reversed path probability, and its exponential is the reweighting factor. The proof works by telescoping this ratio against the initial density $e^{-U_0}/Z_0$; after integrating out intermediate states, only the target density $e^{-U_k}/Z_k$ remains, so expectations weighted by $e^{A_k}$ are exactly expectations under the model at step $k$.

What would settle it

Choose a target density with a known expectation, run the weighted estimator with a drift that deliberately breaks the density-matching condition, and compare the estimate with the true value; a systematic discrepancy beyond sampling error would show the claim's exactness depends on that condition.

Watch

Extended reading notes

Core claim

The paper's central claim is that the reweighted estimator $\mathbb{E}_{\theta_k}[\partial_\theta U_{\theta_k}] = \mathbb{E}[\partial_\theta U_{\theta_k}(X_k)e^{A_k}]/\mathbb{E}[e^{A_k}]$ and the partition-function identity $Z_{\theta_k}=Z_{\theta_0}\mathbb{E}[e^{A_k}]$ hold exactly for any time-dependent Markovian transition kernel, so unbiased estimates of the model expectations used in cross-entropy training can be obtained from out-of-equilibrium paths. For the SDE $dX_t=[b(t,X_t)+\varepsilon s(t,X_t)]dt+\sqrt{2\varepsilon}\,dW_t$ integrated by Euler–Maruyama, Proposition 2 constructs the incremental weights $\zeta_k$ such that the identity holds at every finite step $h$, and Lemma 2 shows the increments are $O(h^{3/2})$, so the weights vanish in the continuous limit provided the drift obeys the continuity condition $\nabla\cdot(b\rho)=\partial_t\rho$. With that condition in force, the weights are purely a finite-step correction to the discretization, and the standard stochastic-interpolant sampling is recovered as $h\to0$. For Bernoulli and Gaussian–Bernoulli restricted Boltzmann machines the paper writes the Gibbs-sampler log-transition kernels explicitly, making the same reweighting immediately implementable in discrete state spaces.

Load-bearing premise

The clean claim that the weights vanish in the continuous limit assumes the drift has been chosen so the particle density evolves exactly as the model density; with a learned drift this matching is only approximate, so the weights would also have to absorb drift error.

Editorial extensions

If this is right

  • Training an EBM by cross-entropy minimization no longer requires converged MCMC: any sampler with known transition probabilities can feed an unbiased estimate of the model gradient.
  • For diffusion and flow models, the same weights can be applied during generation to correct the bias introduced by a finite Euler–Maruyama step, improving sample quality without changing the SDE.
  • In the continuous limit, with the density-matching condition on the drift, the Jarzynski correction disappears, so the weighted sampler reduces to the original stochastic-interpolant dynamics.
  • For Bernoulli and Gaussian–Bernoulli RBMs, the explicit Gibbs kernel ratios make the weighted estimator directly usable, and it returns the partition function together with the gradient, which contrastive divergence cannot do.

Reading between the lines

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

  • Because the identity is exact for any finite $h$, the variance of the weights $e^{A_k}$ can serve as a practical diagnostic: near-uniform weights indicate the sampler's drift is well matched to the model, while large variance flags discretization or drift mis-specification.
  • In learned stochastic interpolants the continuity condition holds only approximately, so the weights likely absorb drift-model error in addition to time-discretization error; a testable prediction is that the weighted estimator outperforms the unweighted one at small step sizes on such models.
  • The same forward/backward log-ratio construction should carry over to other samplers with computable transition probabilities, such as Hamiltonian Monte Carlo or Metropolis–Hastings, provided the reverse kernel is available.
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. This paper studies Jarzynski reweighting as a tool for training energy-based models. It recalls an exact identity (Proposition 1) showing that expectations under a time-evolving Boltzmann density can be computed by reweighting samples from an arbitrary time-dependent Markovian transition kernel, and it derives a specialized weight recursion (Proposition 2) for an Euler–Maruyama discretization of an SDE with drift b(t,x) and score s(t,x) = -∇U(t,x). It also provides the log-transition kernels for Gibbs samplers in Bernoulli and Gaussian–Bernoulli RBMs (Lemmas 3 and 4). Lemma 2 claims that the Jarzynski weights are O(h^{3/2}) and vanish in the continuous-time limit when the continuity equation holds, which the authors use to argue that the weights correct only discretization error in the stochastic-interpolant setting.

Significance. The paper's central exactness result, Proposition 1, is a re-proof of a known result from the authors' prior work [10], and the proof is elementary and correct. The explicit weight formula in Proposition 2 for the drift-plus-score Euler–Maruyama scheme is a useful specialization, and the RBM kernel formulas in Lemmas 3 and 4 are straightforward and, as far as they go, correct. The main value of the paper is in providing a self-contained theoretical account of how the choice of kernel interacts with Jarzynski reweighting, and the exact identity (23) is a strong and useful statement. However, the paper's quantitative claim about the small-h behavior of the weights (Lemma 2) is not rigorously established, and several sign and convention errors in the statement of Proposition 2 need correction. If these issues are fixed, the paper would be a useful reference for practitioners who want to apply Jarzynski reweighting with non-ULA kernels.

major comments (3)
  1. [Section 4.1, Lemma 2, eq. (38)] The proof uses the distributional equality ξ^T J_b(X_k)ξ = ∇·b(X_k) "in law", which is false; only the expectation of the left-hand side equals the right-hand side. The centered variable h(ξ^T J_b ξ - ∇·b) has variance of order h^2, so its typical size is O(h), not O(h^{3/2}). Consequently, the pathwise claim A_{k+1}-A_k = O(h^{3/2}) in eq. (35) is not established; at best one obtains a statement about conditional expectations. Since Section 5 uses Lemma 2 to argue that the weights merely remove Euler–Maruyama discretization error and vanish in the continuous limit, this part of the manuscript needs to be corrected or reframed. The exact identity (23) is unaffected.
  2. [Section 4.1, Proposition 2 and preceding text] The condition on the drift is stated as ∇·(bρ)=∂tρ, but the Fokker–Planck equation (19) with s = -∇U reduces to ∂tρ + ∇·(bρ)=0, i.e., ∇·(bρ)=-∂tρ. Lemma 2 uses the correct sign, so the manuscript is internally inconsistent. Moreover, eq. (39) omits the -∂t log Z_t contribution that appears when differentiating ρ_t = e^{-U_t}/Z_t; as written, the cancellation leading to A_{k+1}-A_k = O(h^{3/2}) holds only under a gauge choice making Z_t constant, which is not stated.
  3. [Section 5] The interpretation that the Jarzynski weights correct only time-discretization error assumes that the learned drift b satisfies the continuity equation exactly. For stochastic interpolants, b is typically learned and satisfies this only approximately; when the condition fails, the weights in Proposition 2 must also correct for the misspecification of the drift. The manuscript should state this limitation explicitly and adjust the discussion accordingly.
minor comments (4)
  1. [Proposition 1, eq. (5)] The arguments of the kernels in the weight update are inconsistent with the iteration rule X_{k+1} ∼ π_k(X_{k+1}, X_k); please clarify the convention for π_k(x,y) and make the ratio in eq. (5) match the notation used in the proof.
  2. [Section 4.1, eq. (20)] The drift term uses X_t instead of X_k inside the bracket; this appears to be a typo and should read X_k.
  3. [Definition 3] The sigmoid function is written with a missing closing parenthesis: σ(x) = 1/(1+e^{-x}) rather than σ(x) = 1/(1+e^{-x}).
  4. [Proof of Proposition 2, eq. (29)] The identity exp(A_k) = e^{-U_{θ_k}(X_k)+U_{θ_0}(X_0)} ∏ π^back_q / π^forward_{q-1} is stated without showing the cancellation of the |X_q - X_{q-1}|^2/4hε terms; adding a few intermediate steps would help the reader verify the calculation.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: Proposition 1 and Proposition 2 are proved in-text, and the self-citations to [10,11] are not load-bearing; the flagged issue in Lemma 2 is a correctness gap, not a circular reduction.

full rationale

The paper is structurally self-contained. Proposition 1 is attributed to the author's own prior work [10,11], but a complete proof is reproduced in Section 3, so the self-citation is not the load-bearing evidence for the identity. Proposition 2 (Section 4.1) is a direct generalization whose proof repeats the same telescoping-product argument: with the forward kernel (27) and the chosen backward kernel (28), e^{A_k} factorizes as e^{-U_{θk}+U_{θ0}} times the ratio of backward to forward kernels, and because every backward kernel integrates to 1 over the previous state, the path integral collapses to the Boltzmann weight (32)-(34). That is an algebraic identity, not an equivalence-by-construction between a fitted input and a prediction. The RBM lemmas compute log-transition kernels from the Bernoulli/Gaussian conditionals and do not smuggle in a conclusion. The only manuscript passage that asserts a quantitative support without an adequate proof is Lemma 2: replacing ξ_k^T J_b ξ_k by ∇·b 'in law' is valid only in expectation, so the pathwise O(h^{3/2}) claim is not established. This is a mathematical gap in the stated derivation, but it is not circularity: no parameter is fitted to the target, no normalization constant is forced, and the claim is not imported from a self-citation. The self-citations to [10,11] are substantial but re-proven, hence not load-bearing. Score 2 reflects the presence of repeated self-citation rather than any circular reduction.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

No new physical entities are introduced. The Jarzynski weight A_k is a standard importance-weight accumulator, not an invented entity.

free parameters (2)
  • h
    Euler-Maruyama step size; the exact reweighting identity (23) holds for any h in (0,L), so h is a user-chosen discretization parameter, not fitted.
  • epsilon
    Noise scale in the SDE (17); any epsilon>0 is allowed, so it is a user-chosen mixing parameter between flow and diffusion, not fitted to data.
assumptions (3)
  • domain assumption Uθ satisfies (2): C^2, Lipschitz Hessian, coercivity; ensures Zθ<∞ and Langevin ergodicity.
    Used throughout to guarantee the model density is well-defined and samplers are ergodic, stated in Section 2.
  • domain assumption The drift b satisfies the continuity equation ∇·(bρ) = -∂tρ for ρ=e^{-U}/Z (stated in Prop 2 as ∇·(bρ)=∂tρ).
    Needed for the Fokker-Planck solution to remain on the Boltzmann-Gibbs manifold; Lemma 2's conclusion that weights vanish relies on it.
  • standard math Standard PDE well-posedness for parabolic equations (used to assert uniqueness of the solution of (58)).
    Invoked in the proof of Proposition 3 in Appendix A to identify ρ̂(t,x) with Z^{-1} e^{-U(t,x)}.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Jarzynski Reweighting and Sampling Dynamics for Training Energy-Based Models: Theoretical Analysis of Different Transition Kernels." pith.science (2026). https://pith.science/paper/LIJLI2UM

@misc{pith2026250607843,
  author       = {Pith},
  title        = {Pith review of: Jarzynski Reweighting and Sampling Dynamics for Training Energy-Based Models: Theoretical Analysis of Different Transition Kernels},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LIJLI2UM}},
  note         = {Machine review of arXiv:2506.07843}
}
read the original abstract

Energy-Based Models (EBMs) provide a flexible framework for generative modeling, but their training remains theoretically challenging due to the need to approximate normalization constants and efficiently sample from complex, multi-modal distributions. Traditional methods, such as contrastive divergence and score matching, introduce biases that can hinder accurate learning. In this work, we present a theoretical analysis of Jarzynski reweighting, a technique from non-equilibrium statistical mechanics, and its implications for training EBMs. We focus on the role of the choice of the kernel and we illustrate these theoretical considerations in two key generative frameworks: (i) flow-based diffusion models, where we reinterpret Jarzynski reweighting in the context of stochastic interpolants to mitigate discretization errors and improve sample quality, and (ii) Restricted Boltzmann Machines, where we analyze its role in correcting the biases of contrastive divergence. Our results provide insights into the interplay between kernel choice and model performance, highlighting the potential of Jarzynski reweighting as a principled tool for generative learning.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 38 canonical work pages

  1. [10]

    Efficient training of energy-based models using jarzynski equality.Advances in Neural Information Processing Systems, 36, 2024

    Davide Carbone, Mengjian Hua, Simon Coste, and Eric Vanden-Eijnden. Efficient training of energy-based models using jarzynski equality.Advances in Neural Information Processing Systems, 36, 2024

    work page 2024

  2. [1]

    A tutorial on energy-based learning

    Yann LeCun, Sumit Chopra, and Raia Hadsell. A tutorial on energy-based learning. In Gökhan BakIr, Thomas Hofmann, Alexander J Smola, Bernhard Schölkopf, and Ben Taskar, editors, Predicting structured data, chapter 10. MIT press, 2007

    work page

  3. [2]

    Noise-contrastive estimation: A new estimation principle for unnormalized statistical models

    Michael Gutmann and Aapo Hyvärinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. InProceedings of the thirteenth international con- ference on artificial intelligence and statistics, pages 297–304. JMLR Workshop and Conference Proceedings, 2010

    work page 2010

  4. [3]

    Sliced score matching: A scalable approach to density and score estimation

    Yang Song, Sahaj Garg, Jiaxin Shi, and Stefano Ermon. Sliced score matching: A scalable approach to density and score estimation. InUncertainty in Artificial Intelligence, pages 574–584. PMLR, 2020. 11

    work page 2020

  5. [4]

    A survey on bias and fairness in machine learning.ACM Computing Surveys (CSUR), 54(6):1–35, 2021

    Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning.ACM Computing Surveys (CSUR), 54(6):1–35, 2021

    work page 2021

  6. [5]

    Estimation of non-normalized statistical models by score matching.Journal of Machine Learning Research, 6(4), 2005

    Aapo Hyvärinen and Peter Dayan. Estimation of non-normalized statistical models by score matching.Journal of Machine Learning Research, 6(4), 2005

    2005

  7. [6]

    Training products of experts by minimizing contrastive divergence.Neural computation, 14(8):1771–1800, 2002

    Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence.Neural computation, 14(8):1771–1800, 2002

    2002

  8. [7]

    A new learning algorithm for mean field boltzmann machines

    Max Welling and Geoffrey E Hinton. A new learning algorithm for mean field boltzmann machines. InInternational conference on artificial neural networks, pages 351–357, 2002

    work page 2002

Show all 51 references
  1. [8]

    On contrastive divergence learning

    Miguel A Carreira-Perpinan and Geoffrey Hinton. On contrastive divergence learning. In International workshop on artificial intelligence and statistics, pages 33–40. PMLR, 2005

    2005

  2. [9]

    Connections between score matching, contrastive divergence, and pseudo- likelihood for continuous-valued variables.IEEE Transactions on neural networks, 18(5): 1529–1531, 2007

    Aapo Hyvarinen. Connections between score matching, contrastive divergence, and pseudo- likelihood for continuous-valued variables.IEEE Transactions on neural networks, 18(5): 1529–1531, 2007

    2007

  3. [11]

    Generative models as out-of-equilibrium particle systems: Training of energy-based models using non-equilibrium thermodynamics

    Davide Carbone, Mengjian Hua, Simon Coste, and Eric Vanden-Eijnden. Generative models as out-of-equilibrium particle systems: Training of energy-based models using non-equilibrium thermodynamics. InInternational Conference on Nonlinear Dynamics and Applications, pages 287–311....

    2024

  4. [12]

    How to train your energy-based models.arXiv preprint arXiv:2101.03288, 2021

    Yang Song and Diederik P Kingma. How to train your energy-based models.arXiv preprint arXiv:2101.03288, 2021

    2021 arXiv

  5. [13]

    A theory of generative convnet

    Jianwen Xie, Yang Lu, Song-Chun Zhu, and Yingnian Wu. A theory of generative convnet. In International Conference on Machine Learning, pages 2635–2644. PMLR, 2016

    2016

  6. [14]

    Your classifier is secretly an energy based model and you should treat it like one

    Will Grathwohl, Kuan-Chieh Wang, Joern-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. InInternational Conference on Learning Representations, 2019

    2019

  7. [15]

    CRC press, 2011

    Steve Brooks, Andrew Gelman, Galin Jones, and Xiao-Li Meng.Handbook of Markov chain Monte Carlo. CRC press, 2011

    2011

  8. [16]

    Springer, 2001

    Jun S Liu and Jun S Liu.Monte Carlo strategies in scientific computing, volume 75. Springer, 2001

    2001

  9. [17]

    A connection between score matching and denoising autoencoders.Neural computation, 23(7):1661–1674, 2011

    Pascal Vincent. A connection between score matching and denoising autoencoders.Neural computation, 23(7):1661–1674, 2011

    2011

  10. [18]

    On autoencoders and score matching for energy based models

    Kevin Swersky, Marc’Aurelio Ranzato, David Buchman, Nando D Freitas, and Benjamin M Marlin. On autoencoders and score matching for energy based models. InInternational conference on machine learning (ICML-11), pages 1201–1208, 2011

    2011

  11. [19]

    On the failure of variational score matching for V AE models.arXiv preprint arXiv:2210.13390, 2022

    Li Kevin Wenliang. On the failure of variational score matching for V AE models.arXiv preprint arXiv:2210.13390, 2022

    2022 arXiv

  12. [20]

    Learning deep kernels for exponential family densities

    Li Wenliang, Danica J Sutherland, Heiko Strathmann, and Arthur Gretton. Learning deep kernels for exponential family densities. InInternational Conference on Machine Learning, 2019

    2019

  13. [21]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. InAdvances in neural information processing systems, volume 32, 2019

    2019

  14. [22]

    Training restricted boltzmann machines using approximations to the likeli- hood gradient

    Tijmen Tieleman. Training restricted boltzmann machines using approximations to the likeli- hood gradient. InInternational conference on Machine learning, pages 1064–1071, 2008. 12

    2008

  15. [23]

    Cooperative training of descriptor and generator networks.IEEE transactions on pattern analysis and machine intelligence, 42(1):27–45, 2018

    Jianwen Xie, Yang Lu, Ruiqi Gao, Song-Chun Zhu, and Ying Nian Wu. Cooperative training of descriptor and generator networks.IEEE transactions on pattern analysis and machine intelligence, 42(1):27–45, 2018

    2018

  16. [24]

    Learning non-convergent non- persistent short-run MCMC toward energy-based model

    Erik Nijkamp, Mitch Hill, Song-Chun Zhu, and Ying Nian Wu. Learning non-convergent non- persistent short-run MCMC toward energy-based model. InAdvances in Neural Information Processing Systems, volume 32, 2019

    2019

  17. [25]

    Flow contrastive estimation of energy-based models

    Ruiqi Gao, Erik Nijkamp, Diederik P Kingma, Zhen Xu, Andrew M Dai, and Ying Nian Wu. Flow contrastive estimation of energy-based models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7518–7528, 2020

    2020

  18. [26]

    Nonequilibrium equality for free energy differences.Physical Review Letters, 78 (14):2690, 1997

    C Jarzynski. Nonequilibrium equality for free energy differences.Physical Review Letters, 78 (14):2690, 1997

    1997

  19. [27]

    Annealed importance sampling.Statistics and computing, 11:125–139, 2001

    Radford M Neal. Annealed importance sampling.Statistics and computing, 11:125–139, 2001

    2001

  20. [28]

    Springer, 2001

    Arnaud Doucet, Nando De Freitas, Neil James Gordon, et al.Sequential Monte Carlo methods in practice, volume 1. Springer, 2001

    2001

  21. [29]

    Auto-encoding sequential monte carlo

    Tuan Anh Le, Maximilian Igl, Tom Rainforth, Tom Jin, and Frank Wood. Auto-encoding sequential monte carlo. InInternational Conference on Learning Representations, 2018. URL https://openreview.net/forum?id=BJ8c3f-0b

    2018

  22. [30]

    Freedman

    Xinqiang Ding and David J. Freedman. Learning deep generative models with annealed importance sampling, 2020

    2020

  23. [31]

    Laurence Illing Midgley, Vincent Stimper, Gregor N. C. Simm, Bernhard Schölkopf, and José Miguel Hernández-Lobato. Flow annealed importance sampling bootstrap. InThe Eleventh International Conference on Learning Representations, 2023. URL https://openreview. net/forum?id=XCTVFJwS9LJ

    2023

  24. [32]

    Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc

    Yilun Du, Conor Durkan, Robin Strudel, Joshua B Tenenbaum, Sander Dieleman, Rob Fergus, Jascha Sohl-Dickstein, Arnaud Doucet, and Will Sussman Grathwohl. Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc. InInternational Conference on...

  25. [33]

    Deep unsu- pervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsu- pervised learning using nonequilibrium thermodynamics. In Francis Bach and David Blei, editors,Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings ...

    2015

  26. [34]

    Arnaud Doucet, Will Sussman Grathwohl, Alexander G. D. G. Matthews, and Heiko Strathmann. Score-based diffusion meets annealed importance sampling. InAdvances in Neural Information Processing Systems, 2022

    2022

  27. [35]

    Deep boltzmann machines

    Ruslan Salakhutdinov and Geoffrey Hinton. Deep boltzmann machines. InArtificial intelligence and statistics, pages 448–455. PMLR, 2009

    2009

  28. [36]

    A practical guide to training restricted boltzmann machines.Neural Networks: Tricks of the Trade: Second Edition, pages 599–619, 2012

    Geoffrey E Hinton. A practical guide to training restricted boltzmann machines.Neural Networks: Tricks of the Trade: Second Edition, pages 599–619, 2012

    2012

  29. [37]

    Improved learning of gaussian-bernoulli restricted boltzmann machines

    KyungHyun Cho, Alexander Ilin, and Tapani Raiko. Improved learning of gaussian-bernoulli restricted boltzmann machines. InArtificial Neural Networks and Machine Learning–ICANN 2011: 21st International Conference on Artificial Neural Networks, Espoo, Finland, June 14-17, 2011, ...

    2011

  30. [38]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. InThe Eleventh International Conference on Learning Representations, 2022

    2022

  31. [39]

    Stochastic interpolants: A unifying framework for flows and diffusions.arXiv preprint arXiv:2303.08797, 2023

    Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions.arXiv preprint arXiv:2303.08797, 2023. 13

    2023 arXiv

  32. [40]

    Springer-Verlag Berlin Heidelberg, 6 edition, 2003

    Bernt Oksendal.Stochastic Differential Equations. Springer-Verlag Berlin Heidelberg, 6 edition, 2003

    2003

  33. [41]

    J. C. Mattingly, A. M. Stuart, and D. J. Higham. Ergodicity for SDEs and approximations: locally Lipschitz vector fields and degenerate noise.Stochastic Processes and their Applications, 101(2):185–232, October 2002

    2002

  34. [42]

    Expansion of the global error for numerical schemes solving stochastic differential equations.Stochastic Analysis and Applications, 8(4):483–509, 1990

    Denis Talay and Luciano Tubaro. Expansion of the global error for numerical schemes solving stochastic differential equations.Stochastic Analysis and Applications, 8(4):483–509, 1990

    1990

  35. [43]

    Generative models as out-of-equilibrium particle systems: the case of energy- based models

    Davide Carbone. Generative models as out-of-equilibrium particle systems: the case of energy- based models. 2024

    2024

  36. [44]

    Correlation functions and computer simulations.Nuclear Physics B, 180(3): 378–384, 1981

    Giorgio Parisi. Correlation functions and computer simulations.Nuclear Physics B, 180(3): 378–384, 1981

    1981

  37. [45]

    Nonasymptotic convergence analysis for the unadjusted langevin algorithm.THE ANNALS of APPLIED PROBABILITY, pages 1551–1587, 2017

    Alain Durmus and Éric Moulines. Nonasymptotic convergence analysis for the unadjusted langevin algorithm.THE ANNALS of APPLIED PROBABILITY, pages 1551–1587, 2017

    2017

  38. [46]

    Escorted free energy simulations: Improving convergence by reducing dissipation.Physical Review Letters, 100(19):190601, 2008

    Suriyanarayanan Vaikuntanathan and Christopher Jarzynski. Escorted free energy simulations: Improving convergence by reducing dissipation.Physical Review Letters, 100(19):190601, 2008

    2008

  39. [47]

    A learning algorithm for boltzmann machines.Cognitive science, 9(1):147–169, 1985

    David H Ackley, Geoffrey E Hinton, and Terrence J Sejnowski. A learning algorithm for boltzmann machines.Cognitive science, 9(1):147–169, 1985

    1985

  40. [48]

    An introduction to restricted boltzmann machines

    Asja Fischer and Christian Igel. An introduction to restricted boltzmann machines. InProgress in Pattern Recognition, Image Analysis, Computer Vision, and Applications: 17th Iberoamerican Congress, CIARP 2012, Buenos Aires, Argentina, September 3-6, 2012. Proceedings 17, pages...

    2012

  41. [49]

    An overview on restricted boltzmann machines.Neurocomputing, 275:1186–1199, 2018

    Nan Zhang, Shifei Ding, Jian Zhang, and Yu Xue. An overview on restricted boltzmann machines.Neurocomputing, 275:1186–1199, 2018

    2018

  42. [50]

    American Mathematical Society, 2022

    Lawrence C Evans.Partial differential equations, volume 19. American Mathematical Society, 2022

    2022

  43. [51]

    Dual training of energy-based models with overparametrized shallow neural networks.arXiv preprint arXiv:2107.05134, 2021

    Carles Domingo-Enrich, Alberto Bietti, Marylou Gabrié, Joan Bruna, and Eric Vanden-Eijnden. Dual training of energy-based models with overparametrized shallow neural networks.arXiv preprint arXiv:2107.05134, 2021. 14 A Jarzynski reweighting in continuous-time We give a continu...

    2021 arXiv

Pith tools

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