Pith. sign in

REVIEW 4 major objections 5 minor 297 references

This review claims differential equations are the unifying design principle behind modern deep learning, from ResNets to diffusion models.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 10:51 UTC pith:PTQXEN66

load-bearing objection Survey of DE-inspired networks with a useful taxonomy but unreliable benchmark tables that need correction before it can serve as a reference. the 4 major comments →

arxiv 2510.09685 v2 pith:PTQXEN66 submitted 2025-10-09 cs.LG cs.AIcs.CVcs.NAmath.NA

Deep Neural Networks Inspired by Differential Equations

Yongshuai Liu , Lianfang Wang , Kuilin Qin , Qinghua Zhang , Faqiang Wang , Li Cui , Jun Liu , Yuping Duan
show 1 more author
Tieyong Zeng
This is my paper
classification cs.LG cs.AIcs.CVcs.NAmath.NA MSC 68T0760H1065L06
keywords deep learningdifferential equationsneural ODEsstochastic differential equationsresidual networksflow matchingdiffusion modelssurvey
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that ordinary and stochastic differential equations provide a unifying language for deep learning: feed-forward networks can be read as numerical discretizations of ODEs and PDEs, regularization tricks such as dropout and noise injection can be derived as SDE discretizations, and modern generative models—from Neural ODEs to flow matching to diffusion models—are best understood as estimating the drift and diffusion terms of a continuous dynamical system. The authors organize dozens of architectures into a taxonomy based on equation type (first-order, higher-order, PDE, SDE) and dynamics-modeling paradigm, and they compile benchmark tables on CIFAR-10 and CIFAR-100 to compare the families. If the survey is accurate, it gives new researchers a map of the field and a quantitative starting point for choosing among equation-inspired architectures.

Core claim

The paper's central claim is that differential equations are not merely an analogy for deep learning but a design principle: every residual network, stochastic regularizer, and diffusion sampler can be placed on a spectrum whose endpoints are deterministic ODEs and stochastic SDEs. Concretely, the authors assert that ResNets are forward-Euler discretizations of an ODE; higher-order networks (LM-ResNets, RKCNNs) correspond to multistep or Runge–Kutta integrators; PDE-based CNNs encode spatial derivatives; dropout, Gaussian noise injection, and stochastic depth have SDE limits; and generative models solve either the Liouville equation (flow matching) or the Fokker–Planck equation (diffusion).

What carries the argument

The central organizing device is the discretization map: a network layer x_{n+1} = x_n + f(x_n) is identified with an Euler step of dx/dt = v(x,t); multistep and Runge–Kutta blocks are identified with higher-order integrators; stochastic layers (dropout, noise injection, stochastic depth) are identified with Euler–Maruyama steps of an SDE; and generative models are identified with the Liouville equation (for deterministic flow) or Fokker–Planck equation (for diffusion). This mapping lets numerical-analysis knowledge (stability, order of convergence) be imported into network design, and lets stochastic-calculus tools (reverse SDEs, score functions) be imported into regularization and sampling

Load-bearing premise

The survey's comparative value rests on the numbers in Tables 5 and 6 faithfully transcribing the cited papers; if the transcriptions are inconsistent or wrong, the taxonomy still stands but the quantitative story does not.

What would settle it

A concrete check: verify the two CIFAR-100 entries for ResNet and FitResNet. The text says ResNet reaches 75.46% and FitResNet 76.63%, but Table 5 lists 72.24% and 72.34%; also, the text cites PDE CIFAR-100 values 80.55% and 80.34% that do not appear in Table 5. If the discrepancy cannot be resolved, the survey's quantitative comparisons are unreliable.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the correspondence is taken seriously, stability of a network can be analyzed by checking Lipschitz or contraction conditions on the vector field, promising mitigation of vanishing or exploding gradients.
  • Higher-order numerical schemes should yield networks that train more accurately; the tables support this trend, with RKCNN outperforming plain ResNet on both CIFAR-10 and CIFAR-100.
  • The same SDE formalism that explains dropout can justify new regularization schemes by choosing different noise processes, extending beyond Gaussian assumptions to jump or Lévy noise.
  • Diffusion models and flow matching become two ends of one continuum, so solvers developed for one transfer to the other; probability-flow ODEs and DDIM already demonstrate such a transfer.
  • Generative models can trade stochastic sampling for deterministic one-step generation, as Rectified Flow and MeanFlow show, reaching competitive FID with a single function evaluation.
  • The Fokker–Planck framing suggests that the noise schedule in diffusion is the diffusion coefficient of an SDE, so optimizing the schedule is equivalent to choosing the SDE's diffusion term in a carefully designed way.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The discretization map suggests a design recipe: any new numerical integrator for ODEs or SDEs is a candidate neural architecture, and any stochastic regularizer is a candidate SDE discretization—this predicts that the field will continue importing numerical-analysis results faster than inventing new network blocks.
  • The benchmark tables, if accurate, imply that the performance gap between plain ResNets and PDE or higher-order networks is partly a statement about integration order; a fair reader could test whether the same gap holds on larger datasets or with modern training schedules, which the survey does not report.
  • Because the survey groups stochastic regularizers with SDEs, it implicitly predicts that regularization strength and diffusion coefficient are the same dial; this could be tested by measuring generalization as a function of an annealing schedule in both settings.
  • The survey's taxonomy could serve as a generative catalog: every combination of equation type (ODE/PDE/SDE), discretization scheme (Euler/RK/multistep), and stochastic component (additive/multiplicative noise) corresponds to a potential new architecture, and the paper's tables mark which cells have already been explored.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper is a survey that organizes deep-learning architectures, regularization techniques, and generative models around ODE/PDE/SDE interpretations. It proposes a taxonomy, gives mathematical formulations in a series of tables, and provides numerical comparisons (classification accuracy on CIFAR-10/100 and FID/NFE on CIFAR-10) drawn from the literature. The abstract's central claim is that the paper provides an 'extensive review' and 'numerical comparisons of these models to illustrate their characteristics and performance.' Since the paper is purely a review, the value of the manuscript rests entirely on the accuracy of its taxonomy, its transcriptions of external results, and the internal consistency of its narrative.

Significance. If the taxonomy and the numerical tables were faithful to the cited sources, the survey could be a useful entry point for researchers seeking a bird's-eye view of differential-equation-inspired deep learning. The paper covers a broad set of methods, includes code links, and draws useful distinctions (e.g., first-order vs. higher-order ODE discretizations, forward vs. reverse SDE design). It also makes no original derivation or experimental claim, which means the potential contribution is reference value rather than new results. That reference value is currently undermined by the data integrity and citation problems detailed below; if these are corrected, the manuscript could serve as a useful, if not comprehensive, survey.

major comments (4)
  1. [§5.1 and Table 5] The numerical comparisons in the text contradict Table 5. In §5.1, the text states that ResNets achieve 75.46% on CIFAR-100 and FitResNet 76.63%, but Table 5 lists 72.24% and 72.34% for the same rows. The text also claims that Parabolic and Hyperbolic PDE models achieve 80.55% and 80.34% on CIFAR-100, yet Table 5 has no CIFAR-100 entries for the Parabolic PDE/Hyperbolic PDE rows; those rows report only CIFAR-10 values. This is not a formatting slip: the text uses these numbers to argue that higher-order and PDE-guided models outperform ResNets, so the inconsistency leaves the central comparison unsupported.
  2. [Table 5 (RevNets and FractalNet rows)] Table 5 contains implausible and mismatched entries. The RevNets row reports 86.31% CIFAR-100 accuracy with 1.79M parameters; the cited RevNets paper (Gomez et al., NIPS 2017) does not report such a number, and the value is far above any known RevNets result. The FractalNet row cites reference [34] (Cho et al., AAAI 2024), but the text and the reference list identify FractalNet as [133] (Larsson et al., ICLR 2017); [34] is a different paper. These are not isolated typos but symptoms that Table 5 has not been systematically verified against its sources. Since Table 5 is the empirical backbone of §5.1, this is a load-bearing reliability issue.
  3. [§5.2 and Table 6] The generative-model comparison in Table 6 conflates incomparable settings without any caveat. For example, flow matching is listed as 'NIPs 2019', but the reference [147] is actually an ICLR 2023 paper. More importantly, the FID/NFE numbers come from different papers using different model sizes, training budgets, and evaluation protocols (e.g., EDM reports FID 1.97 with 55.7M params, DPM-Solver reports 2.69 with 35.7M params, and several rows have no param count). The text's claim of a 'clear progression' from high-step stochastic sampling to single-step deterministic optimization is therefore not supported by a controlled comparison; it is a compilation of disparate results. The survey should either include a methodological note on how numbers were collected and whether they are directly comparable, or explicitly warn the reader otherwise.
  4. [General (methodology of the numerical review)] The paper does not state how the numbers in Tables 5 and 6 were selected, transcribed, or verified. For a survey whose abstract advertises 'numerical comparisons', this is more than an omission: the reader has no way to decide which of the conflicting values (text vs. table, or table vs. original paper) is trustworthy. Given the documented discrepancies, the current manuscript cannot serve as a reliable reference. A revision would need to include a data-verification protocol and, ideally, a table of the exact source (section/table/line) for each reported number.
minor comments (5)
  1. [Table 6] Typo 'Recitied Flow' should be 'Rectified Flow'; also use 'NeurIPS' consistently instead of 'NIPs'.
  2. [§2.2.3] The acronym 'ODESs' is unusual and likely should be 'ODEs' or a spelled-out 'systems of ODEs'.
  3. [Figures 1 and 2] The cross-references use Roman numerals ('Sec Ⅲ.A', 'Sec Ⅳ.A'); these should be Arabic ('Sec. 3.1', 'Sec. 4.1').
  4. [Title page/footer] The ACM Reference Format block and the copyright line indicate 2018, while the arXiv submission is dated 2025; this inconsistency should be resolved. Also, the conference placeholder 'Conference acronym ’XX' is unfilled.
  5. [Reference [147]] Year mismatch: reference [147] is an ICLR 2023 paper, but Table 6 lists the same work as 'NIPs 2019'. Similarly, reference [34] is not the FractalNet paper, creating a citation conflict with Table 5.

Circularity Check

0 steps flagged

No circularity: survey with no derived predictions; internal table/text inconsistencies are accuracy issues, not circular reasoning.

full rationale

The paper is a review/taxonomy of DDE- and SDE-inspired neural network architectures and generative models. It makes no derivation chain, no fitted parameters called predictions, and no theoretical claim that is justified by its own prior results. The central claim is descriptive: that the paper provides a structured review and numerical comparisons. Those comparisons in Section 5 (Tables 5 and 6) are transcriptions of externally published results (e.g., ResNets [90], LM-ResNets [164], PDE-CNNs [211], DDPMs [97], DEIS [285]), not fitted inputs renamed as predictions. The main risk identified by the reader is an internal inconsistency between the Section 5.1 text (ResNet 75.46% CIFAR-100, FitResNet 76.63%, PDE CIFAR-100 80.55/80.34) and Table 5 (72.24/72.34, and PDE CIFAR-100 64.8/64.9/65.4). That is a correctness/transcription concern, not a circularity concern: the survey does not use its tables to justify itself. No self-citation is load-bearing, no uniqueness theorem is imported, and no ansatz is smuggled in via citation. The paper is not circular; it is a survey whose value depends on the accuracy of its external benchmarks, which is outside the scope of circularity analysis.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

As a review, the paper introduces no free parameters or invented entities. The central content restates prior work; the only assumptions are the accuracy of the compiled numbers and the validity of the taxonomy.

axioms (3)
  • domain assumption The cited papers' reported accuracies and FID scores are correctly transcribed.
    The survey's comparative tables (Tables 5 and 6) rely entirely on numbers from external papers; if any transcription is inaccurate, the comparison is invalid. Internal discrepancies suggest this assumption may be violated.
  • domain assumption The taxonomy categories (first-order ODEs, higher-order ODEs, DE systems, PDEs, SDE regularization, etc.) are well-defined and collectively exhaustive.
    The paper's organization assumes these categories cleanly partition the field; if the categories overlap or omit important work, the review's framework is misleading.
  • standard math Standard ODE/SDE theory (existence, uniqueness, Fokker-Planck, reverse-time SDE, etc.) is correct.
    Background preliminaries in Section 2 rely on standard results; no novel mathematical claims are made.

pith-pipeline@v1.3.0-alltime-deepseek · 39415 in / 11801 out tokens · 80613 ms · 2026-08-04T10:51:39.928376+00:00 · methodology

0 comments
read the original abstract

Deep learning has become a pivotal technology in fields such as computer vision, scientific computing, and dynamical systems, significantly advancing these disciplines. However, neural Networks persistently face challenges related to theoretical understanding, interpretability, and generalization. To address these issues, researchers are increasingly adopting a differential equations perspective to propose a unified theoretical framework and systematic design methodologies for neural networks. In this paper, we provide an extensive review of deep neural network architectures and dynamic modeling methods inspired by differential equations. We specifically examine deep neural network models and deterministic dynamical network constructs based on ordinary differential equations (ODEs), as well as regularization techniques and stochastic dynamical network models informed by stochastic differential equations (SDEs). We present numerical comparisons of these models to illustrate their characteristics and performance. Finally, we explore promising research directions in integrating differential equations with deep learning to offer new insights for developing intelligent computational methods that boast enhanced interpretability and generalization capabilities.

Figures

Figures reproduced from arXiv: 2510.09685 by Faqiang Wang, Jun Liu, Kuilin Qin, Lianfang Wang, Li Cui, Qinghua Zhang, Tieyong Zeng, Yongshuai Liu, Yuping Duan.

Figure 1
Figure 1. Figure 1: Overview of DDE-driven neural network architectures (left, Sec. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ODE-guided deterministic dynamics (left, Sec. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Schematics of representative network architectures inspired by discretizations of DEs. The diagrams illustrate the computational [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

297 extracted references · 4 canonical work pages

  1. [1]

    Abien Fred Agarap. 2018. Deep learning using rectified linear units (relu).arXiv preprint arXiv:1803.08375(2018)

    Pith/arXiv arXiv 2018

  2. [2]

    Ravi Aggarwal, Viknesh Sounderajah, Guy Martin, Daniel SW Ting, Alan Karthikesalingam, Dominic King, Hutan Ashrafian, and Ara Darzi. 2021. Diagnostic accuracy of deep learning in medical imaging: a systematic review and meta-analysis.NPJ digital medicine4, 1 (2021), 65

    2021

  3. [3]

    Michael Samuel Albergo and Eric Vanden-Eijnden. 2023. Building Normalizing Flows with Stochastic Interpolants. InInternational Conference on Learning Representations

    2023

  4. [4]

    Tobias Alt, Karl Schrader, Matthias Augustin, Pascal Peter, and Joachim Weickert. 2023. Connections between numerical algorithms for PDEs and neural networks.Journal of Mathematical Imaging and Vision65, 1 (2023), 185–208

    2023

  5. [5]

    Brian DO Anderson. 1982. Reverse-time diffusion equation models.Stochastic Processes and their Applications12, 3 (1982), 313–326

    1982

  6. [6]

    Ludwig Arnold. 1974. Stochastic differential equations.New York2 (1974), 2

    1974

  7. [7]

    Jimmy Ba and Brendan Frey. 2013. Adaptive dropout for training deep neural networks. InAdvances in Neural Information Processing Systems, Vol. 26

    2013

  8. [8]

    Bakary Badjie, Jose Cecilio, and Antonio Casimiro. 2024. Adversarial attacks and countermeasures on image classification-based deep learning models in autonomous driving systems: A systematic review.Comput. Surveys57, 1 (2024), 1–52

    2024

  9. [9]

    Pierre Baldi and Peter J Sadowski. 2013. Understanding dropout. InAdvances in Neural Information Processing Systems, Vol. 26

    2013

  10. [10]

    Chayan Banerjee, Kien Nguyen, Clinton Fookes, and Karniadakis George. 2024. Physics-Informed Computer Vision: A Review and Perspectives. Comput. Surveys57, 1, Article 17 (Oct. 2024), 38 pages. doi:10.1145/3689037

    doi:10.1145/3689037 2024

  11. [11]

    Georgios Batzolis, Jan Stanczuk, Carola-Bibiane Schönlieb, and Christian Etmann. 2021. Conditional image generation with score-based diffusion models.arXiv preprint arXiv:2111.13606(2021)

    Pith/arXiv arXiv 2021

  12. [12]

    Jens Behrmann, Will Grathwohl, Ricky TQ Chen, David Duvenaud, and Jörn-Henrik Jacobsen. 2019. Invertible residual networks. InInternational Conference on Machine Learning, Vol. 97. 573–582

    2019

  13. [13]

    Mikhail Belkin, Daniel Hsu, Siyuan Ma, and Soumik Mandal. 2019. Reconciling modern machine-learning practice and the classical bias–variance trade-off.Proceedings of the National Academy of Sciences116, 32 (2019), 15849–15854. arXiv:https://www.pnas.org/doi/pdf/10.1073/pnas.1903070116 doi:10.1073/pnas.1903070116

    doi:10.1073/pnas.1903070116 2019

  14. [14]

    Marin Biloš, Johanna Sommer, Syama Sundar Rangapuram, Tim Januschowski, and Stephan Günnemann. 2021. Neural flows: Efficient alternative to neural ODEs. InAdvances in Neural Information Processing Systems, Vol. 34. 21325–21337

    2021

  15. [15]

    Olivier Bousquet and Andre Elisseeff. 2002. Stability and generalization.Journal of Machine Learning Research2 (2002), 499–526

    2002

  16. [16]

    1983.Differential equations and their applications

    Martin Braun and Martin Golubitsky. 1983.Differential equations and their applications. Vol. 4. Springer

    1983

  17. [17]

    2016.Numerical methods for ordinary differential equations

    John Charles Butcher. 2016.Numerical methods for ordinary differential equations. John Wiley & Sons

    2016

  18. [18]

    Lei Cai, Jingyang Gao, and Di Zhao. 2020. A review of the application of deep learning in medical image classification and segmentation.Annals of Translational Medicine8, 11 (2020), 713

    2020

  19. [19]

    Andrew Campbell, Joe Benton, Valentin De Bortoli, Tom Rainforth, George Deligiannidis, and Arnaud Doucet. 2022. A continuous time framework for discrete denoising models. InAdvances in Neural Information Processing Systems. 28266–28279

    2022

  20. [20]

    2021.Understanding gaussian noise injections in neural networks

    Alexander Camuto. 2021.Understanding gaussian noise injections in neural networks. Ph. D. Dissertation. University of Oxford

    2021

  21. [21]

    Alexander Camuto, Matthew Willetts, Umut Simsekli, Stephen J Roberts, and Chris C Holmes. 2020. Explicit regularisation in gaussian noise injections. InAdvances in Neural Information Processing Systems, Vol. 33. 16603–16614

    2020

  22. [22]

    Fei Cao, Kimball Johnston, Thomas Laurent, Justin Le, and Sébastien Motsch. 2025. Generative diffusion models from a PDE perspective.arXiv preprint arXiv:2501.17054(2025)

    Pith/arXiv arXiv 2025

  23. [23]

    Hanqun Cao, Cheng Tan, Zhangyang Gao, Yilun Xu, Guangyong Chen, Pheng-Ann Heng, and Stan Z Li. 2024. A survey on generative diffusion models.IEEE Transactions on Knowledge and Data Engineering(2024)

    2024

  24. [24]

    Yu Cao, Jingrun Chen, Yixin Luo, and Xiang Zhou. 2023. Exploring the optimal choice for generative processes in diffusion models: Ordinary vs stochastic differential equations. InAdvances in Neural Information Processing Systems, Vol. 36. 33420–33468

    2023

  25. [25]

    Junyi Chai, Hao Zeng, Anming Li, and Eric W.T. Ngai. 2021. Deep learning in computer vision: A critical review of emerging techniques and application scenarios.Machine Learning with Applications6 (2021), 100134. doi:10.1016/j.mlwa.2021.100134

    arXiv 2021

  26. [26]

    Bo Chang, Lili Meng, Eldad Haber, Lars Ruthotto, David Begert, and Elliot Holtham. 2018. Reversible architectures for arbitrarily deep residual neural networks. InProceedings of the AAAI conference on artificial intelligence, Vol. 32

    2018

  27. [27]

    Chunlei Chen, Peng Zhang, Huixiang Zhang, Jiangyan Dai, Yugen Yi, Huihui Zhang, and Yonghui Zhang. 2020. Deep Learning on Computational-Resource-Limited Platforms: A Survey.Mobile Information Systems2020, 1 (2020), 8454327. arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1155/2020/8454327 doi:10.1155/2020/8454327

    work page doi:10.1155/2020/8454327 2020

  28. [28]

    Ricky T. Q. Chen and Yaron Lipman. 2024. Flow Matching on General Geometries. InInternational Conference on Learning Representations

    2024

  29. [29]

    Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. 2018. Neural Ordinary Differential Equations. InAdvances in Neural Information Processing Systems, Vol. 31

    2018

  30. [30]

    YangQuan Chen, Ivo Petras, and Dingyu Xue. 2009. Fractional order control-a tutorial. In2009 American control conference. IEEE, 1397–1411

    2009

  31. [31]

    Chaoran Cheng, Jiahan Li, Jiajun Fan, and Ge Liu. 2025. 𝛼-Flow: A Unified Framework for Continuous-State Discrete Flow Matching Models. arXiv preprint arXiv:2504.10283(2025). Manuscript submitted to ACM Deep Neural Networks Inspired by Differential Equations 27

    Pith/arXiv arXiv 2025

  32. [32]

    Chaoran Cheng, Jiahan Li, Jian Peng, and Ge Liu. 2024. Categorical flow matching on statistical manifolds. InAdvances in Neural Information Processing Systems, Vol. 37. 54787–54819

    2024

  33. [33]

    Jer-Shiou Chiou and Yen-Hsien Lee. 2009. Jump dynamics and volatility: Oil and the stock markets.Energy34, 6 (2009), 788–796. doi:10.1016/j. energy.2009.02.011

    doi:10.1016/j 2009

  34. [34]

    Woojin Cho, Seunghyeon Cho, Hyundong Jin, Jinsung Jeon, Kookjin Lee, Sanghyun Hong, Dongeun Lee, Jonghyun Choi, and Noseong Park. 2024. Operator-learning-inspired modeling of neural ordinary differential equations. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 11543–11551

    2024

  35. [35]

    Krzysztof M Choromanski, Jared Quincy Davis, Valerii Likhosherstov, Xingyou Song, Jean-Jacques Slotine, Jacob Varley, Honglak Lee, Adrian Weller, and Vikas Sindhwani. 2020. Ode to an ODE. InAdvances in Neural Information Processing Systems, Vol. 33. 3338–3350

    2020

  36. [36]

    Cecılia Coelho, M Fernanda P Costa, and Luis L Ferrás. 2025. Neural fractional differential equations.Applied Mathematical Modelling144 (2025), 116060

    2025

  37. [37]

    Antoine Collas, Ce Ju, Nicolas Salvy, and Bertrand Thirion. 2025. Riemannian Flow Matching for Brain Connectivity Matrices via Pullback Geometry.arXiv preprint arXiv:2505.18193(2025)

    arXiv 2025

  38. [38]

    Ronan Collobert and Jason Weston. 2008. A unified architecture for natural language processing: deep neural networks with multitask learning. In International Conference on Machine Learning(Helsinki, Finland). 160–167

    2008

  39. [39]

    Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. 2023. Diffusion models in vision: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence45, 9 (2023), 10850–10869

    2023

  40. [40]

    Dominik Csiba and Peter Richtárik. 2018. Importance sampling for minibatches.Journal of Machine Learning Research19, 27 (2018), 1–21

    2018

  41. [41]

    Qinpeng Cui, Xinyi Zhang, Qiqi Bao, and Qingmin Liao. 2025. Elucidating the solution space of extended reverse-time SDE for diffusion models. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). 243–252

    2025

  42. [42]

    Wenjun Cui, Qiyu Kang, Xuhao Li, Kai Zhao, Wee Peng Tay, Weihua Deng, and Yidong Li. 2025. Neural Variable-Order Fractional Differential Equation Networks. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 16109–16117

    2025

  43. [43]

    Wenjun Cui, Honglei Zhang, Haoyu Chu, Pipi Hu, and Yidong Li. 2023. On robustness of neural ODEs image classifiers.Information Sciences632 (2023), 576–593

    2023

  44. [44]

    Daems, Rembert and Opper, Manfred and Crevecoeur, Guillaume and Birdal, Tolga. 2025. Efficient training of neural SDEs using stochastic optimal control. InESANN 2025 : 33rd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, Proceedings. 693–698

    2025

  45. [45]

    Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. 2023. Flow matching in latent space.arXiv preprint arXiv:2307.08698(2023)

    Pith/arXiv arXiv 2023

  46. [46]

    Kalyan Das, Jiming Jiang, and JNK Rao. 2004. Mean squared error of empirical predictor.The Annals of Statistics32, 2 (2004), 818–840

    2004

  47. [47]

    Oscar Davis, Samuel Kessler, Mircea Petrache, Ismail Ceylan, Michael Bronstein, and Joey Bose. 2024. Fisher flow matching for generative modeling over discrete data. InAdvances in Neural Information Processing Systems, Vol. 37. 139054–139084

    2024

  48. [48]

    Arturo De Marinis, Nicola Guglielmi, Stefano Sicilia, and Francesco Tudisco. 2025. Stability of neural ODEs by a control over the expansivity of their flows.arXiv preprint arXiv:2501.10740(2025)

    Pith/arXiv arXiv 2025

  49. [49]

    2018.Deep learning in natural language processing

    Li Deng and Yang Liu. 2018.Deep learning in natural language processing. Springer

    2018

  50. [50]

    Teo Deveney, Jan Stanczuk, Lisa Kreusser, Chris Budd, and Carola-Bibiane Schönlieb. 2025. Closing the ODE–SDE gap in score-based diffusion models through the Fokker–Planck equation.Philosophical Transactions A383, 2298 (2025), 20240503

    2025

  51. [51]

    Omar Dhifallah and Yitong Lu. 2021. On the Inherent Regularization Effects of Noise Injection During Training. InInternational Conference on Machine Learning. 2676–2686

    2021

  52. [52]

    1989.Introduction to electric circuits

    Richard C Dorf. 1989.Introduction to electric circuits. John Wiley & Sons

    1989

  53. [53]

    Finale Doshi-Velez and Been Kim. 2017. Towards a rigorous science of interpretable machine learning.arXiv preprint arXiv:1702.08608(2017)

    Pith/arXiv arXiv 2017

  54. [54]

    Weitao Du, He Zhang, Tao Yang, and Yuanqi Du. 2023. A flexible diffusion model. InInternational Conference on Machine Learning. 8678–8696

    2023

  55. [55]

    Emilien Dupont, Arnaud Doucet, and Yee Whye Teh. 2019. Augmented neural odes. InAdvances in Neural Information Processing Systems, Vol. 32

    2019

  56. [56]

    Michael B Elowitz, Arnold J Levine, Eric D Siggia, and Peter S Swain. 2002. Stochastic gene expression in a single cell.Science297, 5584 (2002), 1183–1186

    2002

  57. [57]

    Jonathan Ephrath, Moshe Eliasof, Lars Ruthotto, Eldad Haber, and Eran Treister. 2020. LeanConvNets: low-cost yet effective convolutional neural networks.IEEE Journal of Selected Topics in Signal Processing14, 4 (2020), 894–904

    2020

  58. [58]

    2009.Applied delay differential equations

    Thomas Erneux. 2009.Applied delay differential equations. Springer

    2009

  59. [59]

    2022.Partial differential equations

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

    2022

  60. [60]

    Angela Fan, Edouard Grave, and Armand Joulin. 2020. Reducing Transformer Depth on Demand with Structured Dropout. InInternational Conference on Learning Representations

    2020

  61. [61]

    Chris Finlay, Jörn-Henrik Jacobsen, Levon Nurbekyan, and Adam Oberman. 2020. How to train your neural ODE: the world of Jacobian and kinetic regularization. InInternational Conference on Machine Learning. 3154–3164

    2020

  62. [62]

    Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. 2025. One Step Diffusion via Shortcut Models. InInternational Conference on Learning Representations

    2025

  63. [63]

    2020.A course on rough paths

    Peter K Friz and Martin Hairer. 2020.A course on rough paths. Springer. Manuscript submitted to ACM 28 Liu et al

    2020

  64. [64]

    Yuxiang Fu, Qi Yan, Lele Wang, Ke Li, and Renjie Liao. 2025. Moflow: One-step flow matching for human trajectory forecasting via implicit maximum likelihood estimation based distillation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 17282–17293

    2025

  65. [65]

    Yarin Gal and Zoubin Ghahramani. 2016. A theoretically grounded application of dropout in recurrent neural networks. InAdvances in Neural Information Processing Systems, Vol. 29

    2016

  66. [66]

    Lucio Galeati and Fabian A Harang. 2022. Regularization of multiplicative SDEs through additive noise.The Annals of Applied Probability32, 5 (2022), 3930–3963

    2022

  67. [67]

    1985.Handbook of stochastic methods

    Crispin W Gardiner et al. 1985.Handbook of stochastic methods. Vol. 3. springer Berlin

    1985

  68. [68]

    Xavier Gastaldi. 2017. Shake-shake regularization of 3-branch residual networks. InICLR Workshop

    2017

  69. [69]

    Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky TQ Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. 2024. Discrete flow matching. In Advances in Neural Information Processing Systems, Vol. 37. 133345–133385

    2024

  70. [70]

    Zhengyang Geng, Mingyang Deng, Xingjian Bai, J Zico Kolter, and Kaiming He. 2025. Mean flows for one-step generative modeling.arXiv preprint arXiv:2505.13447(2025)

    Pith/arXiv arXiv 2025

  71. [71]

    Golnaz Ghiasi, Tsung-Yi Lin, and Quoc V Le. 2018. Dropblock: A regularization method for convolutional networks. InAdvances in Neural Information Processing Systems, Vol. 31

    2018

  72. [72]

    Arnab Ghosh, Harkirat Behl, Emilien Dupont, Philip Torr, and Vinay Namboodiri. 2020. Steer: Simple temporal regularization for neural ode. In Advances in Neural Information Processing Systems, Vol. 33. 14831–14843

    2020

  73. [73]

    Patryk Gierjatowicz, Marc Sabate-Vidales, David Šiška, Lukasz Szpruch, and Žan Žurič. 2020. Robust pricing and hedging via neural SDEs.arXiv preprint arXiv:2007.04154(2020)

    Pith/arXiv arXiv 2020

  74. [74]

    Aidan N Gomez, Mengye Ren, Raquel Urtasun, and Roger B Grosse. 2017. The reversible residual network: Backpropagation without storing activations. InAdvances in Neural Information Processing Systems, Vol. 30

    2017

  75. [75]

    Aidan N Gomez, Ivan Zhang, Siddhartha Rao Kamalakara, Divyam Madaan, Kevin Swersky, Yarin Gal, and Geoffrey E Hinton. 2019. Learning sparse networks using targeted dropout.arXiv preprint arXiv:1905.13678(2019)

    Pith/arXiv arXiv 2019

  76. [76]

    Martin Gonzalez, Nelson Fernandez, Thuy Vinh Dinh Tran, Elies Gherbi, Hatem Hajri, and Nader Masmoudi. 2023. SEEDS: Exponential SDE Solvers for Fast High-Quality Sampling from Diffusion Models. InNeural Information Processing Systems

    2023

  77. [77]

    2016.Deep learning

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville. 2016.Deep learning. MIT press

    2016

  78. [78]

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets. InAdvances in Neural Information Processing Systems. 2672–2680

    2014

  79. [79]

    HS Greenside and E Helfand. 1981. Numerical integration of stochastic differential equations—II.Bell System Technical Journal60, 8 (1981), 1927–1940

    1981

  80. [80]

    Samuel Greydanus, Misko Dzamba, and Jason Yosinski. 2019. Hamiltonian neural networks. InAdvances in Neural Information Processing Systems, Vol. 32

    2019

Showing first 80 references.