Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Learning Latent Energy-Based Models via Interacting Particle Langevin Dynamics

T0 review · 2 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Interacting Langevin particle dynamics provably solve maximum marginal likelihood for latent energy-based models, yielding the first non-asymptotic convergence rates for training them.

desk verdict The SDE/particle construction and Theorem 1's idea are real contributions, but the marginal strong-log-concavity step is false as stated and Theorem 2's bias term is mis-scaled, so the advertised convergence of the practical algorithm is not proven. read the letter →

arxiv 2510.12311 v2 pith:FL74WV7J submitted 2025-10-14 stat.ML cs.LGstat.CO

Joanna Marks , Tim Y. J. Wang , O. Deniz Akyildiz This is my paper
classification stat.MLcs.LGstat.CO MSC 60H1060J6062F1065C05
keywords latentenergy-basedmodelsmaximummarginallikelihoodestimationinteractingparticlesystemsLangevindynamicsnon-asymptoticconvergencegenerativemodelingstochasticdifferentialequationsMMLE
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 sets out to prove that latent energy-based models (EBMs) — a generative architecture with an energy-based prior over latents and a Gaussian decoder — can be trained by a single interacting particle system that evolves Langevin dynamics simultaneously over the parameters and over particle clouds targeting each posterior. Its central result is a non-asymptotic error bound: after k iterations with step size h, the expected parameter error shrinks exponentially in k and decreases as 1/√(MN) with the number of particles N per data point and data size M, with explicit dimension dependence. This is the first convergence guarantee for this model class, which is normally trained with nested MCMC loops and heuristic contrastive-divergence-style updates. A sympathetic reader would care because the result gives a principled alternative to expensive MCMC-based LEBM training and explains why a small particle count suffices on large datasets, as the experiments confirm.

What carries the argument

The lifted target distribution π(θ,z) ∝ exp(−∑_{m,n} φ_{y_m}(θ,√(MN) z^{m,n})) — a strongly convex potential under the assumptions — is the key mechanism. Its θ-marginal is exactly proportional to ∏_m p_θ(y_m)^N = exp(N ℓ_M(θ)), so a Langevin sampler on the joint system automatically performs MMLE. The proof rescales the particle system by √(MN) to turn it into a standard Langevin diffusion with step size h/MN, enabling direct application of existing LMC convergence results; the concentration inequality W2(δ_{θ⋆}, π_Θ) ≤ √(d_θ/(µMN)) supplies the final term. In the inexact case, the bias from the ULA-estimated prior expectation is controlled under Assumption A3 and enters the bound as an add

What would settle it

Run EBIPLA on a linear-Gaussian latent EBM where the marginal log-likelihood and θ⋆ are known in closed form, and measure the parameter error as a function of k, N, M, and h; if the error does not track (1−µh)^k C0 + 1.65(L/µ)√((dθ+MNdx)/(MN)) h^{1/2} + √(dθ/µ)/√(MN) with h ≤ 2/(µ+L), then Theorem 1 is false.

Watch

Extended reading notes

Core claim

The paper's central object is the joint distribution over parameters θ and a cloud of latent particles X^{m,n} targeting the posterior p_θ(x|y_m) for each data point. The authors define an SDE system — the θ-process is a scaled Langevin diffusion with drift equal to the empirical mean of ∇_θ log p_θ(x,y_m) over particles, and each particle evolves by Langevin dynamics targeting its posterior — and show that the marginal of the corresponding invariant measure over θ is proportional to exp(N ∑_m log p_θ(y_m)). Thus, as N grows, the stationary θ concentrates around the maximizer θ⋆ of the empirical marginal log-likelihood. The main theorem states that, if the negative joint log-likelihood is µ-

Load-bearing premise

The theorems rest on Assumption A1, which requires the negative joint log-likelihood to be strongly convex and smooth jointly in parameters and latents for every data point — a condition that excludes typical neural-network energy functions and decoders; if it fails, the exponential-contraction and concentration steps do not follow.

Editorial extensions

If this is right

  • For a fixed target accuracy ε, the particle requirement scales as N = O(1/(M ε²)), so large datasets need very few posterior particles per data point — the error term 1/√(MN) vanishes as either grows.
  • The step-size condition h ≤ 2/(µ+L) and the O(√h) discretisation term give explicit, actionable tuning rules for the algorithm's hyperparameters.
  • Replacing the inner MCMC posterior loops of standard LEBM with a parallel particle cloud removes a serial bottleneck; the experiments show much lower wall-clock time at matched compute budgets.
  • The inexact-gradient theorem shows that a short ULA run for the prior expectation is sufficient; the bias term δ/µ decreases with the number of prior steps J, so the convergence guarantee is robust to practical approximations.
  • When the assumptions hold, the algorithm provably converges to the global maximiser of the empirical marginal log-likelihood, eliminating local-optimum concerns that apply to contrastive-divergence training.

Reading between the lines

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

  • If the rate 1/√(MN) holds in practice, then doubling the dataset size can compensate for a fourfold reduction in particles — a scaling law one could verify by tracking FID or MMD on a fixed model as N and M are varied.
  • The proof structure is modular — a lifted target whose θ-marginal is exp(N ℓ_M(θ)) plus LMC convergence — so the same construction could be applied to other latent-variable models with intractable marginals, provided a strong-concavity or local-concavity control can be established.
  • The rate's additive term √(d_θ + M N d_x)/(MN) suggests that in high dimensions the particle noise, rather than the parameter noise, dominates; variance reduction or underdamped particles might yield further gains beyond the paper's results.
  • A testable extension is to replace the ULA prior estimator with a more advanced sampler (e.g., Hamiltonian Monte Carlo) and check whether the bias term δ/µ shrinks; the framework itself is sampler-agnostic.
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

2 major / 5 minor

Summary. The paper proposes EBIPLA, an interacting particle Langevin algorithm for learning latent energy-based models (EBMs) by maximizing the empirical marginal log-likelihood. The authors introduce a continuous-time SDE system that couples parameter dynamics with posterior-sampling particles, derive an Euler–Maruyama discretization (Algorithm 1), and present a practical variant with mini-batching and Adam (Algorithm 2). Their main theoretical contribution is a non-asymptotic convergence analysis: Theorem 1 treats the exact-gradient setting and Theorem 2 treats the inexact setting where the prior expectation is approximated by short-run ULA. Experiments on a synthetic Swiss roll and on CIFAR-10, SVHN, and CelebA64 report competitive FID/MSE and reduced runtime relative to LEBM baselines.

Significance. If the theorems were correct, this would be the first non-asymptotic convergence guarantee for training latent EBMs, with the error vanishing as the number of particles N and data points M grow. The continuous-time interacting-particle formulation is elegant and extends the IPLA framework in a natural way. However, the guarantee for the practical algorithm is not established as stated: Theorem 2 contains a scaling error in the ULA bias term, and the theorem statement is inconsistent with its proof. After correcting that error, the algorithm converges only to a δ/μ-neighborhood for a fixed prior-sampling budget J, rather than to the optimal parameter as M,N→∞. The exact-gradient result (Theorem 1) appears sound, and the empirical study is suggestive, but it evaluates a different algorithm from the one analyzed.

major comments (2)
  1. [Section 4.3, Appendix B.3, Eqs. (42)–(44)] The ULA bias is mis-scaled. In Eq. (42) the bias enters as h ζ_k while the noise is sqrt(2h/(MN)) W_k. Following the rescaling used in the proof of Theorem 1, set h̄ = h/(MN) and ζ'_k = MN ζ_k. Then A3 gives E[||E[ζ'_k|F_k]||] ≤ MN δ, and the rescaled potential has strong-convexity modulus MN μ. Proposition B.1 therefore yields a stationary bias (MN δ)/(MN μ) = δ/μ, independent of M and N. Eq. (44)'s δ/(MN μ) and the theorem's δ/(μ√(MN)) (from C̃2/√(MN)) are both too small by factors of MN and √(MN), respectively. Consequently, for fixed J, increasing M or N does not drive the parameter error to zero; Algorithm 1 is only guaranteed to reach a δ/μ-neighborhood of θ*. The paper's advertised convergence of the practical algorithm as N and M grow is not established.
  2. [Theorem 2 statement vs. Eq. (44)] The theorem statement and the bound proved in Eq. (44) are mutually inconsistent. Eq. (44) has the bias term δ/(MN μ) and the noise terms σ²√h/(1.65L(MN)^{3/2}√(dθ+MNdx)) + σ/(MN√μ), while the theorem's constants C̃1 and C̃2 imply bias δ/(μ√(MN)) and σ terms multiplied by h^{1/2} in the displayed bound. These are different claims. Even aside from the scaling error above, the theorem as stated is not what the proof establishes, so the theorem needs to be restated and reconciled with a corrected proof.
minor comments (5)
  1. [Section 4.1, Remark 1] The claim that A1 implies μ-strong log-concavity of the marginal ℓ_M is true, but the cited Prekopa–Leindler inequality (Gardner 2002, Thm 7.1) only gives log-concavity, not the preservation of the strong-convexity parameter. Please provide a proof or a reference for the strong version (e.g., via the Schur complement / Brascamp–Lieb argument).
  2. [Appendix C.1, Algorithm 2] The experiments evaluate Algorithm 2 (mini-batching, Adam, noise-addition heuristic), while Theorems 1–2 analyze Algorithm 1 (full-batch, vanilla Euler–Maruyama, no Adam). The paper should clarify whether the theoretical guarantees are claimed for the practical variant and, if not, state this explicitly.
  3. [Proposition B.1] The statement says 'L-Lipschitz and μ-strongly convex'; the intended assumption is L-smoothness, not Lipschitzness. Please fix the terminology.
  4. [Theorems 1 and 2] The constants C0 and C̃0 are described as 'fully explicit' but no expressions are given in the main text. Please provide them in the appendix or clarify where they are defined.
  5. [Section 6.1, Table 2] For LEBM, N denotes the number of posterior MCMC steps, while for EBIPLA it denotes the number of particles. The table caption and the text should make this distinction consistent to avoid confusion.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the core bounds follow from external Langevin/strong-log-concavity theorems, with only benign self-citations that are not load-bearing.

full rationale

The paper's derivation chain does not reduce to its own inputs. The target θ* is defined directly as the maximizer of the empirical marginal log-likelihood ℓ_M (Eq. 4), not as a parameter fitted to the algorithm's output. Theorem 1 is proved by identifying the iterates (9)–(10) with an Euler–Maruyama discretization of a rescaled Langevin SDE targeting π ∝ exp(−Φ) whose θ-marginal is exp(N Σ_m log p_θ(y_m)) (Appendix B.1). From that point the W2 convergence bound is imported from Dalalyan and Karagulyan (2019, Theorem 1), an external result, and the concentration bound from Altschuler and Chewi (2024, Lemma A.8), also external. Theorem 2 follows the same route, adapting Dalalyan and Karagulyan's Theorem 4 via Proposition B.1; the bias/variance terms in A3 are assumptions on the ULA estimator, not fitted predictions. The main self-references are the acknowledgement that the interacting particle system is 'inspired by' and 'a generalisation of' IPLA (Akyildiz et al., 2025), which one of the authors co-authored, and the use of 'arguments analogous to those in Oliva and Akyildiz (2024, Lemmas A.4–A.5)' for a routine convexity calculation. These citations are not load-bearing in the circularity sense: the setting (energy-based priors with prior-sampling ULA) is distinct, the target result is not assumed in the cited work, and the convexity claim can be checked directly from A1–A2. The weaker points in the paper—the modulus-preservation claim in Remark 1 and the apparent scaling inconsistency in the δ-term of Theorem 2's proof—are mathematical correctness concerns, not cases where a conclusion is equivalent to its inputs by construction. No fitted parameter is renamed as a prediction, and no uniqueness or ansatz is imported solely from the authors' prior work to force the result.

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

The paper introduces no new physical entities. The central proof relies on strong convexity/smoothness assumptions and an assumed ULA bias bound; the practical algorithm adds unanalyzed heuristics (Adam, mini-batching, noise addition) and a questionable marginal log-concavity step.

free parameters (6)
  • Prior ULA step size γ = 0.003–0.009 (synthetic, Table 2); 0.1 (images)
    Step size for short-run ULA in Eq. (14); affects bias/variance of the prior expectation estimate; tuned per particle count.
  • Posterior discretization step h = 0.9 (synthetic); 0.1 (images)
    Euler-Maruyama step for SDE (17); constrained by 0 < h ≤ 2/(µ+L) in theory but tuned in practice.
  • Number of prior ULA steps J = 500 (synthetic); unspecified for images
    Controls ULA bias δ in A3.
  • MMD kernel bandwidth = 0.1
    Kernel bandwidth for MMD evaluation in Eq. (47); chosen by hand.
  • FID sampling step size = 0.07 (SVHN), 0.1 (CIFAR-10, CelebA64)
    Grid-searched over {0.01,0.05,0.07,0.1,0.2} for generating prior samples for FID.
  • Noise addition scaling sqrt(1/L) = L = M/B (batch-related)
    Ad hoc scaling for particle noise in Algorithm 2; not derived from theory.
assumptions (5)
  • domain assumption A1: φ_y is µ-strongly convex in (θ,x)
    Stated in Section 4.1; needed for all convergence bounds, but unrealistic for neural-network energies/decoders.
  • domain assumption A2: φ_y is L-smooth
    Section 4.1; used for the LMC convergence bound.
  • ad hoc to paper A3: ULA estimator bias/variance bounded by δ, σ
    Section 4.3; assumed rather than derived from ULA convergence; critical for Theorem 2.
  • ad hoc to paper Marginal of a µ-strongly log-concave joint is µ-strongly log-concave
    Invoked in Remark 1; this is false in general (e.g., joint f(θ,x) = (θ²+x²)/2 + cθx with |c|<1 gives marginal constant < µ). The concentration bound in Appendix B.1 depends on it.
  • domain assumption Short-run ULA with J steps yields samples approximately from prior p_α
    Used in Eq. (14) to approximate the intractable expectation; no convergence guarantee for the non-convex neural-network energy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Latent Energy-Based Models via Interacting Particle Langevin Dynamics." pith.science (2026). https://pith.science/paper/FL74WV7J

@misc{pith2026251012311,
  author       = {Pith},
  title        = {Pith review of: Learning Latent Energy-Based Models via Interacting Particle Langevin Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FL74WV7J}},
  note         = {Machine review of arXiv:2510.12311}
}
read the original abstract

We develop interacting particle algorithms for learning latent variable models with energy-based priors. To do so, we leverage recent developments in particle-based methods for solving maximum marginal likelihood estimation (MMLE) problems. Specifically, we provide a continuous-time framework for learning latent energy-based models, by defining stochastic differential equations (SDEs) that provably solve the MMLE problem. We obtain a practical algorithm as a discretisation of these SDEs and provide theoretical guarantees for the convergence of the proposed algorithm. Finally, we empirically validate the effectiveness of our method on synthetic and image datasets and demonstrate that using a particle based approach offers significant improvement in computational efficiency.

Figures

Figures reproduced from arXiv: 2510.12311 by the authors.

Figure 2
Figure 2. Samples generated with EBIPLA and LEBM trained on the rotated Swiss roll dataset for increasing numbers of posterior particles (or MCMC steps) with energy land￾scapes. in various directions, see e.g. Lim et al. (2024); Oliva and Akyildiz (2024) for momentum-enriched variants, Encinar et al. (2025) for extensions to non-smooth settings, Glyn-Davies et al. (2025) for applications to inverse problems. 6 Experiments To … view at source ↗
Figure 3
Figure 3. Samples generated with EBIPLA trained on CIFAR-10, CelebA64, and SVHN. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Latent space interpolation of EBIPLA on the training (bottom) and validation set (top). Generation and Reconstruction. We evaluate the generative performance of our model quantitatively with the Fr´echet Inception Distance (FID) (Heusel et al., 2017) using 50, 000 samples generated by the energy-based prior and decoded back to the pixel space. Additionally, we mea￾sure the reconstruction quality in terms of mean-squ… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Visualisation of the rotated Swiss roll dataset [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Efficient Stochastic Optimisation via Sequential Monte Carlo

    stat.ML 2026-01 conditional novelty 4.0 of 10

    Sequential Monte Carlo samplers can approximate intractable gradients inside a first-order optimizer, yielding a general SOSMC framework that speeds up reward tuning of energy-based models in the reported settings.

Reference graph

Works this paper leans on

57 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    D., Crucinio, F

    Akyildiz, O. D., Crucinio, F. R., Girolami, M., Johnston, T., and Sabanis, S. (2025). Interacting particle L angevin algorithm for m aximum m arginal l ikelihood e stimation. ESAIM: Probability and Statistics , 29:243--280

    2025

  2. [2]

    Altschuler, J. M. and Chewi, S. (2024). Faster high-accuracy log-concave sampling via algorithmic warm starts. Journal of the ACM , 71(3):1--55

    2024

  3. [3]

    Atchadé, Y., Jiang, K., and Sun, Y. (2023). On generative energy-based models. Manuscript

    2023

  4. [4]

    Bengio, Y., Courville, A., and Vincent, P. (2013). Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence , 35(8):1798--1828

    2013

  5. [5]

    Caprio, R., Kuntz, J., Power, S., and Johansen, A. M. (2025). Error bounds for particle gradient descent, and extensions of the log- S obolev and T alagrand inequalities. Journal of Machine Learning Research , 26(103):1--38

    2025

  6. [6]

    Carbone, D., Hua, M., Coste, S., and Vanden-Eijnden, E. (2024). Efficient training of energy-based models using J arzynski equality. Journal of Statistical Mechanics: Theory and Experiment , 2024(10):104019

    2024

  7. [7]

    Dalalyan, A. S. and Karagulyan, A. (2019). User-friendly guarantees for the L angevin M onte C arlo with inaccurate gradient. Stochastic Processes and their Applications , 129(12):5278--5311

    2019

  8. [8]

    De Bortoli, V., Durmus, A., Pereyra, M., and Vidal, A. F. (2021). Efficient stochastic optimisation by unadjusted L angevin M onte C arlo. Statistics and Computing , 31(3):29

    2021

Show all 57 references
  1. [9]

    P., Laird, N

    Dempster, A. P., Laird, N. M., and Rubin, D. B. (1977). Maximum likelihood from incomplete data via the EM algorithm. Journal of the royal statistical society: series B (methodological) , 39(1):1--22

    1977

  2. [10]

    Deng, Y., Bakhtin, A., Ott, M., Szlam, A., and Ranzato, M. (2021). Residual energy-based models for text generation. Journal of Machine Learning Research , 22:1--41

    2021

  3. [11]

    Douc, R., Moulines, E., and Stoffer, D. (2014). Nonlinear time series: Theory, methods and applications with R examples . CRC press

    2014

  4. [12]

    B., Dieleman, S., Fergus, R., Sohl-Dickstein, J., Doucet, A., and Grathwohl, W

    Du, Y., Durkan, C., Strudel, R., Tenenbaum, J. B., Dieleman, S., Fergus, R., Sohl-Dickstein, J., Doucet, A., and Grathwohl, W. S. (2023). Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc. In International conference on machine learni...

    2023

  5. [13]

    Du, Y., Li, S., and Mordatch, I. (2020). Compositional visual generation with energy based models. In Advances in Neural Information Processing Systems

    2020

  6. [14]

    and Mordatch, I

    Du, Y. and Mordatch, I. (2019). Implicit generation and modeling with energy based models. Advances in n eural i nformation p rocessing s ystems , 32

    2019

  7. [15]

    D., Pendleton, B

    Duane, S., Kennedy, A. D., Pendleton, B. J., and Roweth, D. (1987). Hybrid m onte c arlo. Physics letters B , 195(2):216--222

    1987

  8. [16]

    C., Crucinio, F

    Encinar, P. C., Crucinio, F. R., and Akyildiz, O. D. (2025). Proximal interacting particle L angevin algorithms. In The 41st Conference on Uncertainty in Artificial Intelligence

    2025

  9. [17]

    Gardner, R. (2002). The brunn-minkowski inequality. Bulletin of the American mathematical society , 39(3):355--405

    2002

  10. [18]

    and Calderhead, B

    Girolami, M. and Calderhead, B. (2011). Riemann manifold l angevin and h amiltonian m onte c arlo methods. Journal of the Royal Statistical Society Series B: Statistical Methodology , 73(2):123--214

    2011

  11. [19]

    M., Han, P., Ha, H., Chadha, A., Du, Y., Ji, H., Li, J., and Iqbal, T

    Gladstone, A., Nanduru, G., Islam, M. M., Han, P., Ha, H., Chadha, A., Du, Y., Ji, H., Li, J., and Iqbal, T. (2025). Energy-based transformers are scalable learners and thinkers. arXiv preprint arXiv:2507.02092

    2025 arXiv

  12. [20]

    Glaser, P., Arbel, M., Doucet, A., and Gretton, A. (2022). Maximum likelihood learning of energy-based models for simulation-based inference. arXiv preprint arXiv:2210.14756

    2022 arXiv

  13. [21]

    Glyn-Davies, A., Duffin, C., Kazlauskaite, I., Girolami, M., and Akyildiz, \"O . D. (2025). Statistical finite elements via interacting particle L angevin dynamics. SIAM/ASA Journal on Uncertainty Quantification , 13(3):1200--1227

    2025

  14. [22]

    Grathwohl, W., Wang, K.-C., Jacobsen, J.-H., Duvenaud, D., Norouzi, M., and Swersky, K. (2020). Your classifier is secretly an energy based model and you should treat it like one. In International Conference on Learning Representations

    2020

  15. [23]

    M., Rasch, M

    Gretton, A., Borgwardt, K. M., Rasch, M. J., Sch \"o lkopf, B., and Smola, A. J. (2012). A kernel two-sample test. Journal of Machine Learning Research , 13(25):723--773

    2012

  16. [24]

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S. (2017). GAN s trained by a two time-scale update rule converge to a local nash equilibrium. In Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R., editors,...

    2017

  17. [25]

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

    2002

  18. [26]

    D., Gelman, A., et al

    Hoffman, M. D., Gelman, A., et al. (2014). The N o- U - T urn sampler: adaptively setting path lengths in H amiltonian M onte C arlo. J. Mach. Learn. Res. , 15(1):1593--1623

    2014

  19. [27]

    Hwang, C.-R. (1980). Laplace's method revisited: weak convergence of probability measures. The Annals of Probability , pages 1177--1182

    1980

  20. [28]

    and Ba, J

    Kingma, D. and Ba, J. (2014). Adam: A method for stochastic optimization. International Conference on Learning Representations

    2014

  21. [29]

    Kingma, D. P. and Welling, M. (2014). Auto-encoding variational bayes. 2nd International Conference on Learning Representations, ICLR

    2014

  22. [30]

    and Hinton, G

    Krizhevsky, A. and Hinton, G. (2009). Learning multiple layers of features from tiny images. Technical Report 0, University of Toronto, Toronto, Ontario

    2009

  23. [31]

    N., and Johansen, A

    Kuntz, J., Lim, J. N., and Johansen, A. M. (2023). Particle algorithms for maximum likelihood training of latent variable models. In International Conference on Artificial Intelligence and Statistics , pages 5134--5180. PMLR

    2023

  24. [32]

    N., Kuntz, J., Power, S., and Johansen, A

    Lim, J. N., Kuntz, J., Power, S., and Johansen, A. M. (2024). Momentum particle maximum likelihood. In Proceedings of the Forty-First International Conference on Machine Learning

    2024

  25. [33]

    Liu, Z., Luo, P., Wang, X., and Tang, X. (2015). Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV)

    2015

  26. [35]

    Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B., and Ng, A. Y. (2011). Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011

    2011

  27. [36]

    Nijkamp, E., Hill, M., Han, T., Zhu, S.-C., and Wu, Y. N. (2020). On the anatomy of MCMC -based maximum likelihood learning of energy-based models. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 34, pages 5272--5280. AAAI Press

    2020

  28. [37]

    Oliva, P. F. V. and Akyildiz, O. D. (2024). Kinetic interacting particle langevin M onte C arlo. arXiv preprint arXiv:2407.05790

    2024 arXiv

  29. [38]

    Oliva, P. F. V., Akyildiz, O. D., and Duncan, A. (2025). Uniform-in-time convergence bounds for persistent contrastive divergence algorithms

    2025

  30. [39]

    Pang, B., Han, T., Nijkamp, E., Zhu, S.-C., and Wu, Y. N. (2020). Learning latent space energy-based prior model. Advances in Neural Information Processing Systems , 33:21994--22008

    2020

  31. [40]

    Raginsky, M., Rakhlin, A., and Telgarsky, M. (2017). Non-convex learning via stochastic gradient L angevin dynamics: a nonasymptotic analysis. In Conference on Learning Theory , pages 1674--1703. PMLR

    2017

  32. [41]

    and Monro, S

    Robbins, H. and Monro, S. (1951). A stochastic approximation method. The annals of mathematical statistics , pages 400--407

    1951

  33. [42]

    Santambrogio, F. (2015). Optimal transport for applied mathematicians . Springer

    2015

  34. [43]

    Schr \"o der, T., Ou, Z., Li, Y., and Duncan, A. B. (2024). Energy-based modelling for discrete and mixed data via heat equations on structured spaces. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  35. [44]

    N., Li, Y., Vollmer, S

    Schr \"o der, T., Ou, Z., Lim, J. N., Li, Y., Vollmer, S. J., and Duncan, A. (2023). Energy discrepancies: A score-independent loss for energy-based models. In Thirty-seventh Conference on Neural Information Processing Systems

    2023

  36. [45]

    and Kingma, D

    Song, Y. and Kingma, D. P. (2021). How to train your energy-based models. arXiv preprint arXiv:2101.03288

    2021 arXiv

  37. [46]

    P., Kumar, A., Ermon, S., and Poole, B

    Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. (2021). Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations

    2021

  38. [47]

    Tieleman, T. (2008). Training restricted B oltzmann machines using approximations to the likelihood gradient. In Proceedings of the 25th international conference on Machine learning , pages 1064--1071

    2008

  39. [48]

    Y., Kuntz, J., and Akyildiz, O

    Wang, T. Y., Kuntz, J., and Akyildiz, O. D. (2025). Training latent diffusion models with interacting particle algorithms. arXiv preprint arXiv:2505.12412

    2025

  40. [49]

    Whiteley, N., Gray, A., and Rubin-Delanchy, P. (2025). Statistical exploration of the manifold hypothesis. Journal of the Royal Statistical Society: Series B

    2025

  41. [50]

    and Han, T

    Xiao, Z. and Han, T. (2022). Adaptive multi-stage density ratio estimation for learning latent space energy-based model. In Advances in Neural Information Processing Systems

    2022

  42. [51]

    Xiao, Z., Kreis, K., Kautz, J., and Vahdat, A. (2021). \ VAEBM \ : A symbiosis between variational autoencoders and energy-based models. In International Conference on Learning Representations

    2021

  43. [52]

    Yu, P., Zhu, Y., Xie, S., Ma, X., Gao, R., Zhu, S.-C., and Wu, Y. N. (2023). Learning energy-based prior model with diffusion-amortized MCMC . In Thirty-seventh Conference on Neural Information Processing Systems

    2023

  44. [53]

    Yuan, S., Cui, J., Li, H., and Han, T. (2025). Learning multimodal latent generative models with energy-based prior. In Leonardis, A., Ricci, E., Roth, S., Russakovsky, O., Sattler, T., and Varol, G., editors, Computer Vision -- ECCV 2024 , pages 86--100, Cham. Springer Nature...

    2025

  45. [54]

    Zhai, S., Cheng, Y., Lu, W., and Zhang, Z. (2016). Deep structured energy based models for anomaly detection. In Proceedings of the 33rd International Conference on Machine Learning (ICML) , pages 1100--1109. PMLR

    2016

  46. [55]

    D., Damoulas, T., and Sabanis, S

    Zhang, Y., Akyildiz, \"O . D., Damoulas, T., and Sabanis, S. (2023). Nonasymptotic estimates for stochastic gradient L angevin dynamics under local conditions in nonconvex optimization. Applied Mathematics & Optimization , 87(2):25

    2023

  47. [56]

    @esa (Ref

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

  48. [57]

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

  49. [58]

    Yuan, Shiyu and Cui, Jiali and Li, Hanao and Han, Tian

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

    2020 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.