Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

SDE Matching: Scalable and Simulation-Free Training of Latent Stochastic Differential Equations

T0 review · 2 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read This paper introduces SDE Matching, a training procedure for latent stochastic differential equations that removes the need to numerically simulate the SDE during training by directly sampling from a reparameterized posterior process…

desk verdict Real contribution with sound math, but the O(1) complexity headline overstates what the implementation delivers and the MOCAP table needs a test-set fix. read the letter →

arxiv 2502.02472 v3 pith:6KFE5YIE submitted 2025-02-04 stat.ML cs.LG

Grigory Bartosh , Dmitry Vetrov , Christian A. Naesseth This is my paper · ORCID
classification stat.MLcs.LG MSC 60H1065C3068T07
keywords latentstochasticdifferentialequationssimulation-freetrainingvariationalinferenceflowmatchingscoretimeseriesmodelingadjointsensitivityposteriorreparameterization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces SDE Matching, a training procedure for latent stochastic differential equations that removes the need to numerically simulate the SDE during training. The method reparameterizes the approximate posterior process so that latent states at any time can be sampled directly, then rewrites the standard variational bound as expectations over uniformly sampled time steps. Each training step becomes a single drift/diffusion evaluation rather than an SDE integration, giving O(1) memory and time per step in terms of drift/diffusion evaluations. The authors show on synthetic Lorenz attractor data, motion capture sequences, and a video pendulum task that SDE Matching matches or beats adjoint sensitivity training in accuracy while cutting per-iteration cost, improving convergence speed, and keeping gradient norms stable over long time horizons.

What carries the argument

The central object is the learnable reparameterization function $F_\varphi(\varepsilon, t, X)$, which directly defines the posterior marginal distributions $q_\varphi(z_t|X)$ at every time $t$ in terms of a fixed noise distribution. From this function the paper derives the conditional ODE with velocity $\bar f_\varphi(z_t,t,X) = \partial_t F_\varphi(\varepsilon,t,X)|_{\varepsilon = F_\varphi^{-1}(z_t,t,X)}$, and then converts it to a conditional SDE with drift $f_{\theta,\varphi} = \bar f_\varphi + \tfrac12 g_\theta g_\theta^\top \nabla_{z_t}\log q_\varphi(z_t|X) + \tfrac12 \nabla_{z_t}\cdot(g_\theta g_\theta^\top)$, where $g_\theta$ is the prior diffusion term. This construction guarantees that the posterior SDE has the same marginals as the reparameterization, so the variational bound's expectation terms can be estimated by direct sampling rather than integration.

What would settle it

Measure end-to-end wall-clock training time and per-step gradient variance for SDE Matching versus the adjoint method on a suite of long-horizon and high-dimensional time series, while recording the time spent in the posterior encoder separately; if the encoder cost cancels the simulation savings, or if the single-sample gradient estimator diverges from the true variational bound on a linear system where the Kalman filter gives exact likelihoods, the central scalability or equivalence claims would fail.

Watch

Extended reading notes

Core claim

SDE Matching claims that a Latent SDE can be trained with exactly the same variational bound as the adjoint sensitivity method, but without any numerical SDE simulation during training, by defining the approximate posterior through an invertible reparameterization $z_t = F_\varphi(\varepsilon, t, X)$ with $\varepsilon \sim \mathcal{N}(0, I)$. This implicitly defines the posterior marginals $q_\varphi(z_t|X)$, and from them the paper constructs a conditional ODE and then a conditional SDE whose marginal distributions are exactly $q_\varphi(z_t|X)$. The training objective is rewritten so that the diffusion loss is an expectation over a uniform time $t$ of $\tfrac12\|r_{\theta,\varphi}(z_t,t,X)\|^2$ with $z_t$ sampled directly from $F_\varphi$, and the reconstruction loss is a similar expectation over observation indices. The paper also shows that conventional diffusion models are a special case of Latent SDEs, which motivates transferring simulation-free matching ideas to the Latent SDE setting.

Load-bearing premise

The scalability claim rests on the assumption that sampling from the posterior reparameterization $z_t = F_\varphi(\varepsilon,t,X)$ is cheap per iteration, but in the implementation $F_\varphi$ requires a context computed from all observations, so the actual cost per step grows with the sequence length and may dominate the advertised O(1) drift/diffusion evaluations.

Editorial extensions

If this is right

  • If the central claim is correct, training Latent SDEs on long or high-dimensional time series becomes practical where adjoint sensitivity methods are computationally infeasible, since each gradient step costs one drift/diffusion evaluation instead of a full SDE solve.
  • The same variational bound as the adjoint method means SDE Matching inherits the theoretical guarantees of variational inference for Latent SDEs, with the difference only in how the objective is estimated.
  • The posterior reparameterization gives direct access to smoothing marginals $q_\varphi(z_t|X)$, enabling simulation-free latent state inference and forecasting without integrating the conditional SDE.
  • SDE Matching unifies several existing methods: ARCTA appears as a restricted special case, and diffusion models appear as a single-observation Latent SDE with a specific posterior parameterization, which may guide transfer of techniques between the two fields.
  • Because the method supports state-dependent diffusion terms $g_\theta(z_t,t)$, it can model systems such as stochastic Lotka-Volterra dynamics, which state-independent simulation-free approaches cannot handle.

Reading between the lines

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

  • Editorial extension: the O(1) complexity table counts drift/diffusion evaluations, not the full cost of conditioning on observations; the posterior encoder that computes the context from $X$ grows with sequence length, so the practical speedup depends on amortizing that encoder, and the reported wall-clock gains may not transfer to settings with very long sequences unless the encoder is paralleliz
  • Editorial extension: SDE Matching's posterior parameterization is currently linear in noise, $F_\varphi(\varepsilon,t,X) = \mu_\varphi(X,t) + \sigma_\varphi(X,t)\varepsilon$; replacing this with normalizing-flow-style invertible maps could extend the method to non-Gaussian posterior marginals while preserving simulation-free training, a direction the authors explicitly leave open.
  • Editorial extension: the single-sample Monte Carlo estimate of the diffusion loss may have high variance for strongly nonlinear dynamics; a testable improvement would be to use multiple parallel samples per iteration, which the framework permits, and measure whether this reduces gradient variance more than it increases wall-clock cost.
  • Editorial extension: the established connection between diffusion models and Latent SDEs suggests that techniques developed for diffusion models, such as learned noise schedules and deterministic sampling, may be transferable to Latent SDE training and inference, though the paper notes that deterministic ODE sampling would distort the joint trajectory distribution.
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 SDE Matching, a training method for latent stochastic differential equation (SDE) models that avoids numerical simulation of the posterior process. The key idea is to parameterize the posterior marginal distributions q_phi(z_t|X) directly via an invertible map z_t = F_phi(epsilon,t,X), then construct a conditional SDE with those marginals using the probability-flow ODE result of Song et al. (2021b). This lets the variational ELBO be rewritten as a Monte Carlo objective over directly sampled latents, so each training iteration uses one drift/diffusion evaluation rather than a numerical SDE solve. Experiments on the Lorenz attractor, motion capture, and a video pendulum show comparable or better accuracy than adjoint sensitivity training, with faster convergence and reduced per-iteration runtime.

Significance. If the complexity claims are taken at face value, the method is a significant contribution: it connects score/flow matching to latent SDE training and could enable applications to long or high-dimensional sequences where adjoint-based backpropagation through SDE solvers is infeasible. The central mathematical derivation appears sound: the construction of the posterior SDE from the probability-flow ODE is correct, and the rewritten objective is an unbiased estimator of the standard ELBO. The empirical results support the accuracy of the method on the tested benchmarks. However, the headline O(1) time-and-memory claim is not supported by the implementation, which includes a sequential time-reversal GRU over the entire observation sequence. This is a load-bearing issue for the paper's central scalability narrative; a corrected complexity statement and an explicit treatment of the encoder cost are needed.

major comments (2)
  1. [Section 1, Table 1, Appendix D] The central scalability claim ('O(1) time and memory', Section 1 and Table 1; 'drastically reducing computational complexity', Abstract) is not supported by the implementation described in Appendix D. The posterior reparameterization F_phi(eps,t,X) is implemented as mu_phi(X,t) + sigma_phi(X,t)eps, where mu_phi and sigma_phi are predicted by an MLP from a context computed by a time-reversal GRU over all N observations (Appendix D). Computing this context is a sequential O(N) operation, and backpropagation through the GRU stores O(N) hidden states, so each iteration of Algorithm 1 costs at least O(N) time and memory. Table 1's O(1) entries count only the number of drift/diffusion evaluations (per the table header), not the total computation. The asymptotic advantage over the adjoint sensitivity method therefore disappears for long sequences, and the 'drastically reduced complexity' claim is overstated unless the encoder is amortized or parallelized. Please correct the complexity statements and provide an explicit discussion of the encoder cost.
  2. [Section 5.1 and 5.3] The reported per-iteration speedup (approximately 5x on the Lorenz task) and the video experiment's wall-clock comparison conflate the saving from removing SDE simulation with the overhead of the recurrent encoder. Since the encoder processes the entire sequence in both the adjoint method and SDE Matching, a quantitative breakdown (e.g., time spent in the forward/backward pass of the GRU versus the drift/diffusion evaluations) is needed to support the complexity claim. Without such a breakdown, the 'drastic' speedup may be partly implementation-specific rather than a consequence of the simulation-free property.
minor comments (5)
  1. [Section 5.2] The text states 'we report the average performance on the training set over 10 models', but Table 2 is labeled 'Test MSE'; this should say 'test set'.
  2. [Appendix C.1] The sentence 'As shown in Appendix C.3, the SDE Matching ELBO closely converges...' references the wrong appendix; the linear-system experiment is presented in Appendix C.1, and Figure 4 is in Appendix C.2.
  3. [Figure 2] The caption contains a typo: 'approximatly' should be 'approximately'.
  4. [Table 1] The definitions of L and R are ambiguous for the SDE Matching row; please specify explicitly that the O(1) entries count posterior drift/diffusion evaluations and exclude the cost of computing the context from X.
  5. [Section 4.4] The paragraph on deterministic sampling is somewhat confusing; clarify that while the probability-flow ODE preserves marginals, it changes the joint law of the process, so it is not a valid method for generating time-series samples.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the SDE Matching objective is the standard Latent SDE variational bound with a reparameterized, simulation-free estimator; self-citations are minor and non-load-bearing.

full rationale

The central derivation is self-contained. Equation (21) is explicitly the same variational objective used by adjoint sensitivity methods (Equation (3)); the paper changes only the estimator, not the bound. The replacement of the path integral in Equation (5) by the marginal integral in Equation (23) is exact by the tower property once the posterior SDE (Equation (19)) is constructed to have marginals qφ(zt|X), and that construction rests on the external probability-flow ODE result of Song et al. (2021b) (Proposition A.2), not on the paper's own conclusions. Direct sampling zt = Fφ(ε,t,X) (Equation (16)) is a deliberate parameterization that makes the estimator simulation-free by construction; it is not a fitted quantity later reported as a prediction. No parameter is fit to a subset of data and then used to 'predict' a closely related target, and no external benchmark is used to claim a prediction. The only self-citations (Bartosh et al., 2023, 2024) appear in related-work context and as 'inspiration' for the flow-ODE step in Appendix A.2; the proof of Proposition A.1 is grounded in Lee (2012, Theorem 9.12), an external standard result, so the self-citation is not load-bearing. The O(1) complexity entry in Table 1 is explicitly scoped to sequential/parallel drift-diffusion evaluations, while Appendix D's GRU context encoder adds an O(N) pass; that is a complexity-scoping or scalability concern, not a circularity. The Limitations section candidly states the trade-offs of the Fφ parameterization, further confirming that the method does not hide a circular shortcut.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No physical entities are invented and no ad hoc constants are fit to make the core derivation work. The method's variational parameters, the neural networks for drift, diffusion, and posterior mean and covariance, are learned from data. The main unstated cost is that the affine posterior parameterization used in all experiments is a Gaussian family, so the theoretical flexibility of F_phi is not exercised.

free parameters (2)
  • Posterior covariance sigma_phi(X,t) = Diagonal matrix predicted by a neural network from X and t.
    The posterior marginal is Gaussian by construction (Eq. 37), and this covariance is a learned variational parameter, not derived from the prior or data model. The score term in the conditional SDE drift, Eq. (38), depends on it.
  • Number of posterior samples per training step = 1
    Algorithm 1 uses one (t, epsilon) sample for L_diff and L_rec. This is a hand-selected design choice that controls gradient variance and underlies the O(1) per-step claim; the paper states one sample was sufficient but does not study sensitivity.
assumptions (4)
  • standard math The probability-flow ODE and the diffusion SDE share the same marginal distributions when the drift is corrected by score and divergence terms (Proposition A.2).
    External result from Song et al. 2021b, used to construct the posterior SDE in Eq. (19).
  • domain assumption F_phi is smooth and invertible in epsilon for every t and X, so it defines a valid flow and permits a change of variables.
    Required for the conditional ODE (17) and score formula (36); the affine parameterization (37) satisfies it, but general F_phi is not tested.
  • standard math The variational ELBO for latent SDEs with matched diffusion terms equals the prior and reconstruction terms plus E_q integral of 1/2 times the norm of g^{-1}(h-f) squared.
    Standard path-measure KL identity from Li et al. 2020; the paper uses it to claim the objective is the same as adjoint training.
  • domain assumption The conditional SDE (19) is well-posed and its marginals are exactly q_phi(z_t|X) for every t.
    The paper states exactness without proving regularity; it relies on suitable conditions on drift and diffusion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SDE Matching: Scalable and Simulation-Free Training of Latent Stochastic Differential Equations." pith.science (2026). https://pith.science/paper/6KFE5YIE

@misc{pith2026250202472,
  author       = {Pith},
  title        = {Pith review of: SDE Matching: Scalable and Simulation-Free Training of Latent Stochastic Differential Equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6KFE5YIE}},
  note         = {Machine review of arXiv:2502.02472}
}
read the original abstract

The Latent Stochastic Differential Equation (SDE) is a powerful tool for time series and sequence modeling. However, training Latent SDEs typically relies on adjoint sensitivity methods, which depend on simulation and backpropagation through approximate SDE solutions, which limit scalability. In this work, we propose SDE Matching, a new simulation-free method for training Latent SDEs. Inspired by modern Score- and Flow Matching algorithms for learning generative dynamics, we extend these ideas to the domain of stochastic dynamics for time series and sequence modeling, eliminating the need for costly numerical simulations. Our results demonstrate that SDE Matching achieves performance comparable to adjoint sensitivity methods while drastically reducing computational complexity.

Figures

Figures reproduced from arXiv: 2502.02472 by the authors.

Figure 1
Figure 1. Training data distribution and learned dynamics from a 3D stochastic Lorenz attractor. Results for Latent SDE trained with adjoint sensitivity method (bottom left) and SDE Matching (bottom right). Both methods successfully learn the underlying dynamics [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. NELBO (Equation (21)) objective with respect to itera￾tion step on 3D stochastic Lorenz attractor data. Results for Latent SDE trained with adjoint sensitivity method and SDE Matching. SDE Matching demonstrates approximatly 10 times faster conver￾gence in terms of iterations. diffusion gθ(zt, t), and observation model pθ(xt|zt), as well as the parameters φ of the posterior reparameterization func￾tion Fφ. All parame… view at source ↗
Figure 3
Figure 3. Qualitative comparison of generated video sequences depicting a moving pendulum. Samples generated by the model trained with adjoint sensitivity method and SDE Matching with same number of training iterations. challenging to learn long-term dependencies and limits the ability to propagate learning signals effectively. An extended discussion on the importance of state￾dependent volatility functions gθ, along with exp… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Linear system training dynamics for Latent SDE trained with adjoint sensitivity method and SDE Matching [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Lotka–Volterra system, learned trajectories for SDE Matching with state-independent and state-dependent volatility. C.3. Importance of State-Dependent Diffusion Term We also present an example that highlights the importance of state-dependent diffusion term (volatility…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. An adaptive split-combine Gaussian mixture filter for nonlinear and multimodal state estimation

    math.NA 2026-08 conditional novelty 7.0 of 10

    An adaptive split-combine Gaussian mixture filter provably halves variance along a target direction when splitting and outperforms standard filters on nonlinear oscillator and chaotic benchmarks.

Reference graph

Works this paper leans on

65 extracted references · 42 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    S., Boffi, N

    Albergo, M. S., Boffi, N. M., and Vanden-Eijnden, E. Stochastic interpolants: A unifying framework for flows and diffusions, 2023. URL https://arxiv.org/abs/2303.08797

    arXiv 2023

  3. [3]

    Gaussian process approximations of stochastic differential equations

    Archambeau, C., Cornford, D., Opper, M., and Shawe-Taylor, J. Gaussian process approximations of stochastic differential equations. In Gaussian Processes in Practice, pp.\ 1--16. PMLR, 2007 a

    work page 2007

  4. [4]

    Variational inference for diffusion processes

    Archambeau, C., Opper, M., Shen, Y., Cornford, D., and Shawe-Taylor, J. Variational inference for diffusion processes. Advances in neural information processing systems, 20, 2007 b

    work page 2007

  5. [5]

    Bartosh, G., Vetrov, D., and Naesseth, C. A. Neural diffusion models. arXiv preprint arXiv:2310.08337, 2023

    arXiv 2023

  6. [6]

    Bartosh, G., Vetrov, D., and Naesseth, C. A. Neural flow diffusion models: Learnable forward process for improved diffusion modelling. arXiv preprint arXiv:2404.12940, 2024

    arXiv 2024

  7. [7]

    O., and Fearnhead, P

    Beskos, A., Papaspiliopoulos, O., Roberts, G. O., and Fearnhead, P. Exact and computationally efficient likelihood-based estimation for discretely observed diffusion processes (with discussion). Journal of the Royal Statistical Society Series B: Statistical Methodology, 68 0 (3): 0 333--382, 2006

    work page 2006

  8. [8]

    S., Januschowski, T., and G \"u nnemann, S

    Bilo s , M., Sommer, J., Rangapuram, S. S., Januschowski, T., and G \"u nnemann, S. Neural flows: Efficient alternative to neural odes. In Advances in neural information processing systems, volume 34, pp.\ 21325--21337, 2021

    work page 2021

Show all 65 references
  1. [9]

    Bogachev, V. I. Measure theory. Springer, 2007

    2007

  2. [10]

    J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., Vander P las, J., Wanderman- M ilne, S., and Zhang, Q

    Bradbury, J., Frostig, R., Hawkins, P., Johnson, M. J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., Vander P las, J., Wanderman- M ilne, S., and Zhang, Q. JAX : composable transformations of P ython+ N um P y programs. 2018. URL http://github.com/google/jax

    2018

  3. [11]

    T., Rubanova, Y., Bettencourt, J., and Duvenaud, D

    Chen, R. T., Rubanova, Y., Bettencourt, J., and Duvenaud, D. K. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018

    2018

  4. [12]

    Learning phrase representations using rnn encoder-decoder for statistical machine translation

    Cho, K., Van Merri \"e nboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., and Bengio, Y. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014

    2014 arXiv

  5. [13]

    An introduction to sequential M onte C arlo , volume 4

    Chopin, N., Papaspiliopoulos, O., et al. An introduction to sequential M onte C arlo , volume 4. Springer, 2020

    2020

  6. [14]

    and Nair, P

    Course, K. and Nair, P. Amortized reparametrization: efficient and scalable variational inference for latent sdes. Advances in Neural Information Processing Systems, 36, 2023

    2023

  7. [15]

    Variational inference for SDE s driven by fractional noise

    Daems, R., Opper, M., Crevecoeur, G., and Birdal, T. Variational inference for SDE s driven by fractional noise. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=rtx8B94JMS

    2024

  8. [16]

    Density estimation using real NVP

    Dinh, L., Sohl-Dickstein, J., and Bengio, S. Density estimation using real NVP . In International Conference on Learning Representations, 2017. URL https://openreview.net/forum?id=HkpbnH9lx

    2017

  9. [17]

    P., Aspuru-Guzik, A., and Neklyudov, K

    Du, Y., Plainer, M., Brekelmans, R., Duan, C., No \'e , F., Gomes, C. P., Aspuru-Guzik, A., and Neklyudov, K. Doob's lagrangian: A sample-efficient variational approach to transition path sampling. arXiv preprint arXiv:2410.07974, 2024

    arXiv

  10. [18]

    and van Gerven, M

    ElGazzar, A. and van Gerven, M. Generative modeling of neural dynamics via latent stochastic differential equations. arXiv preprint arXiv:2412.12112, 2024

    2024

  11. [19]

    How to train your neural ode: the world of jacobian and kinetic regularization

    Finlay, C., Jacobsen, J.-H., Nurbekyan, L., and Oberman, A. How to train your neural ode: the world of jacobian and kinetic regularization. In International conference on machine learning, pp.\ 3154--3164. PMLR, 2020

    2020

  12. [20]

    E., and Carin, L

    Gan, Z., Li, C., Henao, R., Carlson, D. E., and Carin, L. Deep temporal sigmoid belief networks for sequence modeling. Advances in Neural Information Processing Systems, 28, 2015

    2015

  13. [21]

    and Glasserman, P

    Giles, M. and Glasserman, P. Smoking adjoints: Fast M onte C arlo greeks. Risk, 19 0 (1): 0 88--92, 2006

    2006

  14. [22]

    and Munos, R

    Gobet, E. and Munos, R. Sensitivity analysis using I t\^o-- M alliavin calculus and martingales, and application to stochastic optimal control. SIAM Journal on control and optimization, 43 0 (5): 0 1676--1713, 2005

    2005

  15. [23]

    Grathwohl, W., Chen, R. T. Q., Bettencourt, J., and Duvenaud, D. Scalable reversible generative models with free-form continuous dynamics. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=rJxgknCcK7

    2019

  16. [24]

    Adaptive path-integral autoencoders: Representation learning and planning for dynamical systems

    Ha, J.-S., Park, Y.-J., Chae, H.-J., Park, S.-S., and Choi, H.-L. Adaptive path-integral autoencoders: Representation learning and planning for dynamical systems. Advances in Neural Information Processing Systems, 31, 2018

    2018

  17. [25]

    o m, H., Intosalmi, J., and L \

    Heinonen, M., Yildiz, C., Mannerstr \"o m, H., Intosalmi, J., and L \"a hdesm \"a ki, H. Learning unknown ode models with gaussian processes. In International conference on machine learning, pp.\ 1959--1968. PMLR, 2018

    1959

  18. [26]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

    2020

  19. [27]

    Hodgkinson, L., van der Heide, C., Roosta, F., and Mahoney, M. W. Stochastic normalizing flows. arXiv preprint arXiv:2002.09547, 2020

    2002 arXiv

  20. [28]

    J., and Duvenaud, D

    Kelly, J., Bettencourt, J., Johnson, M. J., and Duvenaud, D. K. Learning differential equations that are easy to solve. Advances in Neural Information Processing Systems, 33: 0 4370--4380, 2020

    2020

  21. [29]

    hey, that's not an ode

    Kidger, P., Chen, R. T., and Lyons, T. J. " hey, that's not an ode": Faster ode adjoints via seminorms. In ICML, pp.\ 5443--5452, 2021 a

    2021

  22. [30]

    C., and Lyons, T

    Kidger, P., Foster, J., Li, X. C., and Lyons, T. Efficient and accurate gradients for neural sdes. Advances in Neural Information Processing Systems, 34: 0 18747--18761, 2021 b

    2021

  23. [31]

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

    2014 arXiv

  24. [32]

    Kingma, D. P. and Dhariwal, P. Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems, 31, 2018

    2018

  25. [33]

    J., Allen, M

    Lea, D. J., Allen, M. R., and Haine, T. W. Sensitivity analysis of the climate of a chaotic system. Tellus A: Dynamic Meteorology and Oceanography, 52 0 (5): 0 523--532, 2000

    2000

  26. [34]

    Lee, J. M. Introduction to Smooth Manifolds. Springer, 2012

    2012

  27. [35]

    L., Chen, R

    Li, X., Wong, T.-K. L., Chen, R. T., and Duvenaud, D. Scalable gradients for stochastic differential equations. In International Conference on Artificial Intelligence and Statistics, pp.\ 3870--3882. PMLR, 2020

    2020

  28. [36]

    Lipman, Y., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=PqvMRDCJT9t

    2023

  29. [37]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Liu, X., Gong, C., and qiang liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=XVjTT1nw5z

    2023

  30. [38]

    Continuous-discrete smoothing of diffusions

    Mider, M., Schauer, M., and Van der Meulen, F. Continuous-discrete smoothing of diffusions. Electronic Journal of Statistics, 15 0 (2): 0 4295--4342, 2021

    2021

  31. [39]

    R., Mineiro, P., and Williams, R

    Movellan, J. R., Mineiro, P., and Williams, R. J. A monte carlo em approach for partially observable diffusion processes: theory and applications to neural networks. Neural computation, 14 0 (7): 0 1507--1544, 2002

    2002

  32. [40]

    A., Lindsten, F., Sch \"o n, T

    Naesseth, C. A., Lindsten, F., Sch \"o n, T. B., et al. Elements of sequential M onte C arlo. Foundations and Trends in Machine Learning , 12 0 (3): 0 307--392, 2019

    2019

  33. [41]

    Nielsen, B. M. G., Christensen, A., Dittadi, A., and Winther, O. Diffenc: Variational diffusion with a learned encoder. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=8nxy1bQWTG

    2024

  34. [42]

    Stable neural stochastic differential equations in analyzing irregular time series data

    Oh, Y., Lim, D.-Y., and Kim, S. Stable neural stochastic differential equations in analyzing irregular time series data. arXiv preprint arXiv:2402.14989, 2024

    2024 arXiv

  35. [43]

    Stochastic differential equations

    Oksendal, B. Stochastic differential equations. Springer, 2003

    2003

  36. [44]

    J., Mohamed, S., and Lakshminarayanan, B

    Papamakarios, G., Nalisnick, E., Rezende, D. J., Mohamed, S., and Lakshminarayanan, B. Normalizing flows for probabilistic modeling and inference. The Journal of Machine Learning Research, 22 0 (1): 0 2617--2680, 2021

    2021

  37. [45]

    Amortized control of continuous state space feynman-kac model for irregular time series

    Park, B., Lee, H., and Lee, J. Amortized control of continuous state space feynman-kac model for irregular time series. arXiv preprint arXiv:2410.05602, 2024

    2024 arXiv

  38. [46]

    Automatic differentiation in pytorch

    Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., and Lerer, A. Automatic differentiation in pytorch. 2017

    2017

  39. [47]

    and Favaro, S

    Peluchetti, S. and Favaro, S. Infinitely deep neural networks as diffusion processes. In International Conference on Artificial Intelligence and Statistics, pp.\ 1126--1136. PMLR, 2020

    2020

  40. [48]

    Causal models for dynamical systems

    Peters, J., Bauer, S., and Pfister, N. Causal models for dynamical systems. In Probabilistic and Causal Inference: The Works of Judea Pearl, pp.\ 671--690. 2022

    2022

  41. [49]

    T., and Duvenaud, D

    Rubanova, Y., Chen, R. T., and Duvenaud, D. K. Latent ordinary differential equations for irregularly-sampled time series. Advances in neural information processing systems, 32, 2019

    2019

  42. [50]

    Differential equations, dynamical systems, and linear algebra

    Rudin, W. Differential equations, dynamical systems, and linear algebra. Tata McGraw-Hill Education, 2006

    2006

  43. [51]

    S., and Prangle, D

    Ryder, T., Golightly, A., McGough, A. S., and Prangle, D. Black-box variational inference for stochastic differential equations. In International Conference on Machine Learning, pp.\ 4423--4432. PMLR, 2018

    2018

  44. [52]

    S., Gokaslan, A., Sa, C

    Sahoo, S. S., Gokaslan, A., Sa, C. D., and Kuleshov, V. Diffusion models with learned adaptive noise processes, 2024. URL https://openreview.net/forum?id=8gZtt8nrpI

    2024

  45. [53]

    and Solin, A

    Sarkka, S. and Solin, A. Applied stochastic differential equations, volume 10. Cambridge University Press, 2019

    2019

  46. [54]

    Where to diffuse, how to diffuse, and how to get back: Automated learning for multivariate diffusions

    Singhal, R., Goldstein, M., and Ranganath, R. Where to diffuse, how to diffuse, and how to get back: Automated learning for multivariate diffusions. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=osei3IzUia

    2023

  47. [55]

    Maximum likelihood training of score-based diffusion models

    Song, Y., Durkan, C., Murray, I., and Ermon, S. Maximum likelihood training of score-based diffusion models. Advances in Neural Information Processing Systems, 34: 0 1415--1428, 2021 a

    2021

  48. [56]

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

    Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021 b . URL https://openreview.net/forum?id=PxTIG12RRHS

    2021

  49. [57]

    and Raginsky, M

    Tzen, B. and Raginsky, M. Neural stochastic differential equations: Deep latent gaussian models in the diffusion limit. arXiv preprint arXiv:1905.09883, 2019

    1905 arXiv

  50. [58]

    Comparing stochastic lotka--volterra predator-prey models

    Vadillo, F. Comparing stochastic lotka--volterra predator-prey models. Applied Mathematics and Computation, 360: 0 181--189, 2019

    2019

  51. [59]

    and Schauer, M

    van der Meulen, F. and Schauer, M. Bayesian estimation of discretely observed multi-dimensional diffusion processes using guided proposals . Electronic Journal of Statistics, 11 0 (1): 0 2358 -- 2396, 2017. doi:10.1214/17-EJS1290. URL https://doi.org/10.1214/17-EJS1290

    2017 doi

  52. [60]

    Variational G aussian process diffusion processes

    Verma, P., Adam, V., and Solin, A. Variational G aussian process diffusion processes. In International Conference on Artificial Intelligence and Statistics, pp.\ 1909--1917. PMLR, 2024

    1909

  53. [61]

    M., Fleet, D

    Wang, J. M., Fleet, D. J., and Hertzmann, A. Gaussian process dynamical models for human motion. IEEE transactions on pattern analysis and machine intelligence, 30 0 (2): 0 283--298, 2007

    2007

  54. [62]

    A., Blei, D., and Cunningham, J

    Wu, L., Trippe, B., Naesseth, C. A., Blei, D., and Cunningham, J. P. Practical and asymptotically exact conditional sampling in diffusion models. In Advances in Neural Information Processing Systems, volume 36, pp.\ 31372--31403, 2023

    2023

  55. [63]

    and Kushner, H

    Yang, J. and Kushner, H. J. A M onte C arlo method for sensitivity analysis and parametric optimization of nonlinear stochastic systems. SIAM journal on control and optimization, 29 0 (5): 0 1216--1249, 1991

    1991

  56. [64]

    Ode2vae: Deep generative second order odes with B ayesian neural networks

    Yildiz, C., Heinonen, M., and Lahdesmaki, H. Ode2vae: Deep generative second order odes with B ayesian neural networks. Advances in Neural Information Processing Systems, 32, 2019

    2019

  57. [65]

    L., and Tong, A

    Zhang, X., Pu, Y., Kawamura, Y., Loza, A., Bengio, Y., Shung, D. L., and Tong, A. Trajectory flow matching with applications to clinical time series modeling. In Advances in Neural Information Processing Systems, volume 37, 2024

    2024

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.