Pith. sign in

REVIEW 3 major objections 4 minor 33 references

A Rauch-Tung-Striebel smoother supplies latent-state targets, letting hybrid physics-neural ODEs learn missing dynamics from partial noisy measurements.

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

2026-08-01 23:54 UTC pith:IIGZPWOI

load-bearing objection A sensible, incremental extension of the authors' CKF hybrid-ODE work, but the evaluation leaks future information into the claimed long-horizon gains and the paper needs major revision before the numbers can be trusted. the 3 major comments →

arxiv 2607.15180 v1 pith:IIGZPWOI submitted 2026-07-16 cs.LG cs.SYeess.SY

RTS Smoother-Guided Learning of Physics-Based Neural Differential Models

Ahmet Demirkaya , Georgios Stratis , Tales Imbiriba , Zachary D. Danziger , Deniz Erdogmus This is my paper
classification cs.LG cs.SYeess.SY
keywords neural ODEhybrid physics-neural modelRauch-Tung-Striebel smootherpartial state observationmissing dynamics learningmulti-horizon predictionstate-space estimationalternating optimization
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 is trying to establish that a classical Kalman-style smoother can serve as the state-estimation engine for training hybrid physics-neural ODEs when only part of the state is measured. The proposed procedure alternates between an RTS smoother, which reconstructs latent trajectories and uncertainties under the current dynamics, and backpropagation, which updates the neural components while keeping known equations explicit. The training objective spans multiple prediction horizons and weights latent residuals by smoother precision, which the authors argue controls long-horizon drift that single-step losses miss. Across five benchmark systems, the method reports the lowest unmeasured-state RMSE in four cases and the lowest rollout Hausdorff distance in all five. A sympathetic reader would care because it offers a path from partial observations to a standalone, interpretable ODE model without learned inference networks.

Core claim

The central claim is that missing ODE components can be learned from incomplete measurements while retaining explicit known structure, and that RTS smoothing is the right vehicle for it. Rather than training an encoder to map measurements to latent states, the paper fixes the dynamics, smooths the whole measurement sequence to produce latent trajectories and covariances, then treats those trajectories as pseudo-targets to update the neural dynamics by backpropagation, iterating until convergence. The objective evaluates both latent and observation-space errors over several integration horizons, with residuals weighted by inverse smoother covariance. The reported result is that this scheme im

What carries the argument

The Rauch-Tung-Striebel smoother: a batch estimator that combines all available measurements with the current hybrid dynamics to produce a latent trajectory and a confidence covariance at every time step. It does triple duty in the training loop - it provides initial conditions for rollout integration, supplies the latent-state pseudo-targets for the neural loss, and generates the precision weights that downweight uncertain time steps. Because the smoother depends on the same parameters being learned, the method alternates between smoothing and parameter updates.

Load-bearing premise

The reconstructed hidden-state paths come from the model being trained and are assumed accurate enough to serve as teaching targets; if they are systematically wrong, the neural term learns the same error.

What would settle it

Take a synthetic system with known unmeasured ground truth, weak indirect measurements, and a deliberately misspecified initial dynamics; run the proposed alternating loop. If the learned neural term reproduces the smoother's trajectories but the final hybrid ODE fails to match the true latent dynamics, or if the smoothed-state bias stays large while the training loss decreases, the load-bearing assumption is violated.

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

If this is right

  • Learning a hybrid ODE with RTS guidance produces a deployable model that needs no online filtering or learned inference at test time.
  • Multi-horizon training with precision-weighted latent errors is a practical remedy for drift in unmeasured states.
  • Preserving known mechanistic equations concentrates learning on the genuinely unknown terms, keeping the model interpretable.
  • The approach transfers across linear, nonlinear, and stiff systems, with largest gains where hidden states are tightly coupled to measured ones.

Where Pith is reading between the lines

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

  • Left implicit in the paper: the same alternating scheme could be extended to learn parts of the measurement function h, since the smoother needs h to produce targets; a joint update might widen applicability to settings with unknown sensors.
  • Because the framework treats smoother covariance as precision, a natural extension is to make the loss robust by clipping or reweighting steps where smoother confidence is low, which would directly address the paper's own limitation about weak measurements.
  • If the bias-from-pseudo-targets concern is real, a testable variant would add a small residual correction to the smoothed targets during training; the paper does not explore this.

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

3 major / 4 minor

Summary. The paper proposes an alternating two-stage training algorithm for hybrid neural-physics ODEs under partial and noisy observations. The dynamics are written as a known mechanistic part and a neural-network part for the unknown components. In the state-estimation step, a CKF/RTS smoother is run with the current dynamics parameters to produce smoothed latent trajectories and covariances from the full measurement sequence. In the parameter-update step, these smoothed trajectories are treated as targets in a multi-horizon rollout loss that combines a latent-state residual weighted by the smoother precision and a measurement residual weighted by R^{-1}. The two steps are alternated. Experiments on five benchmark systems (harmonic oscillator, Hodgkin-Huxley, retinal circulation, Brusselator, yeast glycolysis) compare the method against NeuralODE, GRU-ODE-Bayes, and a CKF/RBSE hybrid, reporting unmeasured-state RMSE and rollout Hausdorff distance, plus ablations and noise-robustness sweeps. The paper's central claim is that RTS-guided training improves latent-state reconstruction and long-horizon prediction while retaining the interpretable mechanistic structure.

Significance. If the claimed results hold, the paper would make a useful contribution: it offers a clean way to use classical smoothing as a supervision source for hybrid ODE learning, avoiding learned encoders at inference time and producing a standalone forward-simulatable model. The method is clearly specified, the two-stage structure is sensible, and the ablation in Table 3 gives evidence that the latent-state loss, covariance weighting, and RTS smoothing each contribute. The paper also explicitly acknowledges important limitations, including sensitivity to weak measurements, non-Gaussian noise, and lack of convergence guarantees. The central empirical claim, however, is currently not established because the evaluation protocol for long-horizon rollouts and latent-state RMSE appears to use full-sequence smoother information, which undermines the 'long-horizon prediction' and 'standalone hybrid ODE' claims. The self-consistency of the surrogate objective is also a real correctness risk that needs explicit handling.

major comments (3)
  1. [§4.4, Fig. 9, §A.4, Eq. (21)] The long-horizon rollout evaluation is not a causal forecast test. Fig. 9 states explicitly: 'For initial conditions, we use the smoother estimates.' The RTS smoother (Appendix A) uses the full measurement sequence y_{1:T}, including measurements after the rollout starts, to produce those estimates. Thus the rollouts reported in Fig. 9, and the Hausdorff distances in Table 2 unless a different protocol was used, inherit future information and cannot support the claim of improved 'long-horizon prediction' by the learned ODE. The same issue may affect Table 1 if the reported \hat{x} is the smoothed trajectory rather than a forward simulation of the final hybrid ODE. Please clarify exactly what quantity is being evaluated and, more importantly, add a causal evaluation: initialize rollouts from a filter/causal estimator using only measurements up to the start time, and/or from the learned OD
  2. [§3.3, Eq. (14), §5.3] The training objective is a self-consistency surrogate, not an independent likelihood. The targets \hat{x}^{(i)}_{k+n}(\theta) and weights P^{(i)}_{k+n}(\theta) are produced by the RTS smoother using the current dynamics f_\theta, the same parameters being optimized. As the paper states in §3, replacing Q with P_k is 'a modeling choice rather than an algebraic consequence,' so Eq. (14) does not correspond to an exact MLE. Section 5.3 concedes that biased smoothed estimates 'propagate into the learned dynamics.' This is load-bearing: the alternating scheme can converge to a model that agrees with a biased smoother while the true missing dynamics are not learned. Please provide a demonstration that the method is robust to this circularity, for example a diagnostic experiment with a deliberately misspecified Q/R or weakly observable state, or a comparison where the smoother targets are repl
  3. [§4.2, Eq. (18), Tables 1–2] The definition of the state-estimate RMSE in Eq. (18) is ambiguous: it writes \hat{x}^{(i,m)}_{unk,k} but does not state whether this is the RTS-smoothed trajectory, the output of a forward ODE rollout, or something else. Figures 3–7 show 'Proposed' trajectories that appear to be RTS-smoothed estimates, while the abstract and conclusion claim reconstruction of latent states by the learned hybrid ODE. If Table 1 reports smoother outputs, then it evaluates the smoother, not the learned dynamics, and the 'distilled into a standalone hybrid ODE' claim is unsupported. The same ambiguity applies to the Hausdorff distance in Table 2, whose text in Eq. (20)-(21) says 'integrating the hybrid dynamical system' but does not specify the initial condition protocol. Please define exactly which estimator produces each reported number and report both smoother-based and pure-forward-simulation numbers.
minor comments (4)
  1. [§4.2, Tables 1–2] The text lists Recognition ODE / structured NODE as a baseline, but Tables 1 and 2 do not include it. Either add the results or remove the baseline from the comparison list. Also, M=50 Monte Carlo runs are claimed, but no standard deviations, confidence intervals, or significance tests are reported, making it hard to judge whether the four-of-five and all-five wins are meaningful.
  2. [§5.2 and §B.5] The sentence in §5.2 ends with 'the esti....' and appears truncated. Also, the yeast glycolysis constants {c_i,d_i,e_i,f_i,g_i,h_i,j_i} are said to be 'given in the main text/config,' but the constants are not actually listed in the paper; please provide them or a pointer to the repository.
  3. [Figure 8] Figure 8 is mentioned in §4.3.5 but appears to be missing from the text; the caption and description refer to state-space trajectories but no figure is shown. Please include it or remove the reference.
  4. [Abstract/Conclusion] The abstract and conclusion state that the method 'improves long-horizon prediction,' but as noted above the evaluation does not currently measure causal long-horizon prediction. Please align the claims with the evaluation protocol.

Circularity Check

3 steps flagged

Latent-state targets and evaluation rollouts are generated by the same RTS smoother that depends on the learned parameters, so the headline reconstruction and long-horizon claims are partly self-referential rather than independent tests of the learned ODE.

specific steps
  1. self definitional [Section 3.4 (alternating optimization), Eq. (11), Eq. (13), Eq. (14); Section 5.3]
    "Both the parameters θ and the smoothed trajectories ˆx(i) 0:Ti depend on each other through the smoother. ... we treat the smoothed trajectories as if they were the ground truth and estimate the neural network parameters using backpropagation."

    The latent-state term in Eq. (14) is e^{(x,i,n)}_{k+n}(θ) = ˜x^{(i,n)}_{k+n}(θ) − ˆx^{(i)}_{k+n}(θ), but Eq. (11) defines the target ˆx^{(i)}_{0:Ti} = S_φ(y^{(i)}_{1:Ti}; f_θ, h), i.e., the smoother output is a function of the same dynamics f_θ being optimized. The training target is therefore not independent ground truth but a model-generated pseudo-target. Minimizing this self-consistency surrogate can drive f_θ and the smoother into mutual agreement even when f_θ is wrong; the paper concedes in Section 5.3 that biased smoother estimates 'propagate into the learned dynamics.' The derivation of the learned ODE is thus not anchored to an external latent-state signal.

  2. fitted input called prediction [Section 4.4, Figure 9 caption; Section 3.1 Eq. (11); Appendix A]
    "For initial conditions, we use the smoother estimates."

    The smoother S_φ(y_{1:T}; f_θ, h) in Eq. (11) maps the full measurement sequence, including measurements inside and after the displayed rollout window, and depends on the current learned dynamics f_θ. Initializing the 'rollout' in Figure 9 from these estimates means the predicted trajectory is anchored by future information and by the very model being evaluated. The resulting Hausdorff distances in Table 2, if they follow this protocol, therefore measure smoother-anchored reconstruction rather than open-loop long-horizon prediction of the standalone hybrid ODE from a partially observed initial state.

  3. fitted input called prediction [Section 4.2, Eq. (18); Section 4.3.2, Figure 4 caption]
    "Figure 4 shows the RTS-smoothed state estimates together with ground truth. ... the ground-truth x^{(i,m)}_{unk,k} is used only for offline evaluation."

    The unmeasured-state RMSE in Eq. (18) is computed against ˆx^{(i,m)}_{unk,k}, and the paper identifies the reported state estimates as RTS-smoothed trajectories. These smoothed trajectories are exactly the pseudo-targets used to train the neural dynamics (Eqs. (11) and (14)), and the smoother uses the full measurement sequence and the fitted dynamics. Thus Table 1's 'latent-state reconstruction' results evaluate the smoother used in training, not an independent forward simulation of the final hybrid ODE; the claimed reconstruction gains partly restate the training target rather than validating the learned dynamics.

full rationale

The paper is candid that the objective is a surrogate: Section 3 states that replacing Q with the smoother covariance is 'a modeling choice rather than an algebraic consequence,' and Section 5.3 explicitly warns that biased smoothed estimates 'propagate into the learned dynamics.' Those admissions do not remove the circularity; they point to it. The central derivation is a self-consistent alternating scheme: the RTS smoother produces latent targets from the current f_θ, and f_θ is then trained to reproduce those targets. Because the targets themselves are functions of the learned parameters, the training loss is a self-consistency constraint rather than an independent log-likelihood or a supervised fit to externally fixed latent states. The evaluation compounds this: the reported unmeasured-state estimates are RTS-smoothed trajectories, and the rollout comparisons initialize from full-sequence smoother estimates. Consequently, the headline claims of improved latent-state reconstruction and long-horizon prediction are partly measuring the smoother's access to future measurements and to the fitted dynamics, not solely the standalone hybrid ODE's predictive skill. This is not a case of load-bearing self-citation: the references to Demirkaya et al. (2021, 2024) provide baselines and prior extensions rather than the core logical derivation. But because the main quantitative claims reduce in part to the outputs of the same smoother used in training, the empirical findings are only partially independent of the method's own construction. Score 6 captures this partial, central circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim depends on (i) the standard state-space factorization, (ii) the Gaussian/CKF-RTS approximation, (iii) the known measurement function, and (iv) the ad hoc surrogate objective that treats model-generated smoother outputs as ground-truth targets. There are no newly invented physical entities. The main free parameters are the neural network weights, the noise covariances, the initial-state prior, and the horizon set — none of which are given concrete values in the paper.

free parameters (4)
  • MLP weights and biases for f_unk = Learned via backpropagation on eq. (14)
    These parameters define the learned missing dynamics and are fit to the surrogate rollout loss.
  • Noise covariances Q and R = Not given; treated as fixed hyperparameters (Section 3.1)
    The smoother and the loss weights depend on Q and R. The paper does not state how they are chosen (ground truth, hand-tuned, or estimated).
  • Initial-state prior x̄0 and P0 = Not given; part of the initialization settings in φ
    Used to initialize the CKF-RTS recursion; no details provided on how they are set.
  • Prediction horizon set H = e.g., H={1,50} in Figure 2; no selection procedure stated
    The set of horizons in the multi-horizon loss (14) is chosen by hand and affects training stability and final performance.
axioms (4)
  • domain assumption All noise is Gaussian and time-invariant, and the CKF-RTS smoother approximation is accurate enough for the benchmarks
    The state-space model and loss rely on Gaussian Q and R and on the cubature approximation of the posterior; Section 5.3 notes that non-Gaussian or state-dependent noise would bias the smoothed estimates.
  • domain assumption The measurement function h is known and fixed
    Section 5.3 explicitly states that settings where h is partially unknown or learned jointly are not addressed. The method depends on h being available to run the smoother and compute the measurement loss.
  • ad hoc to paper The surrogate objective that replaces latent states with smoother estimates, and uses P_k in place of Q, is a valid training target
    Section 3 labels the reweighting 'a modeling choice rather than an algebraic consequence' and notes the objective is 'no longer the exact MLE'. The validity of this surrogate is assumed without theoretical guarantees; the paper provides only empirical evidence.
  • standard math Markov property and conditional independence of measurements given states
    The joint likelihood factorization in eqs. (4)-(6) assumes the standard state-space Markov structure; this is standard probability, not a contribution.

pith-pipeline@v1.3.0-alltime-deepseek · 21063 in / 12937 out tokens · 106695 ms · 2026-08-01T23:54:01.325796+00:00 · methodology

0 comments
read the original abstract

Ordinary differential equations (ODEs) are widely used to model dynamical systems in physics, biology, neuroscience, and physiology, but in many applications some equations of the dynamics are unknown and only a subset of the state variables are measured. We propose a hybrid neural--physics framework in which the known components of the ODE are kept explicit and the missing components are represented by a neural network. The proposed method consists of two stages where we alternate between state and parameter estimation and iterate until a predetermined criterion is met. Specifically, in the first step, we treat the model parameters as being known and we infer the latent states from the available measurements using a Rauch--Tung--Striebel (RTS) smoother. In the second stage, we treat the smoothed trajectories as being known and use them to estimate the neural networks' parameters through backpropagation. We evaluate the method on benchmark systems spanning linear, nonlinear, and stiff dynamics under partial state observation. Across these settings, the proposed method learns missing ODE components from incomplete measurements while exploiting and retaining interpretable mechanistic structure and improving latent-state reconstruction and long-horizon prediction.

Figures

Figures reproduced from arXiv: 2607.15180 by Ahmet Demirkaya, Deniz Erdogmus, Georgios Stratis, Tales Imbiriba, Zachary D. Danziger.

Figure 1
Figure 1. Figure 1: Overview of the proposed pipeline. However, many of these approaches rely on neural encoders or recurrent architectures to infer latent states, which may struggle with generalization, interpretability, and stability, especially under sparse or highly noisy conditions Rubanova et al. (2019); De Brouwer et al. (2019). For instance, the Latent ODE frame￾work Rubanova et al. (2019) learns a variational encoder… view at source ↗
Figure 2
Figure 2. Figure 2: Effect of integration horizon on yeast glycolysis rollout for the latent state [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Harmonic oscillator state estimation results. (a) Time-series estimates for the measured state [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Yeast glycolysis state estimates. RTS-smoothed estimates from the proposed method (dashed, [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Retinal circulation state estimations for pressures [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Brusselator state-estimation time series. Ground truth (solid) and noisy measurements (orange [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Hodgkin–Huxley time series under partial measurements. Proposed RTS-smoothed state estimates [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Hodgkin–Huxley state-space trajectories ( [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Rollout comparison for the unmeasured state in five benchmark systems: velocity [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Brusselator: RTS state estimates across measurement noise levels. We vary the measurement [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Hodgkin–Huxley neuron: RTS smoothed state estimates across measurement noise levels (SNR). [PITH_FULL_IMAGE:figures/full_fig_p026_11.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

33 extracted references · 4 linked inside Pith

  1. [1]

    Sensors , volume=

    Augmented physics-based models for high-order Markov filtering , author=. Sensors , volume=. 2024 , publisher=

    2024

  2. [2]

    Interpretable Augmented Physics-Based Model for Estimation and Tracking , year=

    Straka, Ondřej and Duník, Jindřich and Closas, Pau and Imbiriba, Tales , booktitle=. Interpretable Augmented Physics-Based Model for Estimation and Tracking , year=

  3. [3]

    Augmented Physics-Based Machine Learning for Navigation and Tracking , year=

    Imbiriba, Tales and Straka, Ondřej and Duník, Jindřich and Closas, Pau , journal=. Augmented Physics-Based Machine Learning for Navigation and Tracking , year=

  4. [4]

    IEEE Signal Processing Magazine , year=

    Artificial intelligence-aided Kalman filters: AI-augmented designs for Kalman-type algorithms , author=. IEEE Signal Processing Magazine , year=

  5. [5]

    Advances in Neural Information Processing Systems , volume =

    Neural Ordinary Differential Equations , author =. Advances in Neural Information Processing Systems , volume =

  6. [6]

    Rubanova, Yulia and Chen, Ricky T. Q. and Duvenaud, David , booktitle =. Latent

  7. [7]

    De Brouwer, Edward and Simm, Jaak and Arany, Adam and Moreau, Yves , booktitle =

  8. [8]

    A. L. Hodgkin and A. F. Huxley , title =. J Physiol. , year =

  9. [9]

    Neural extended

    Liu, Wei and Lai, Zhilu and Bacsa, Kiran and Chatzi, Eleni , journal =. Neural extended. 2024 , doi =

    2024

  10. [10]

    and van Sloun, Ruud J

    Revach, Guy and Shlezinger, Nir and Ni, Xiaoyong and Escoriza, Andres L. and van Sloun, Ruud J. G. and Eldar, Yonina C. , journal =. 2022 , doi =

    2022

  11. [11]

    arXiv preprint arXiv:2110.09005 , year =

    Unsupervised Learned Kalman Filtering , author =. arXiv preprint arXiv:2110.09005 , year =

    Pith/arXiv arXiv

  12. [12]

    Revach, Guy and Ni, Xiaoyong and Shlezinger, Nir and van Sloun, Ruud J. G. and Eldar, Yonina C. , journal =. 2023 , doi =

    2023

  13. [13]

    Proceedings of the 39th International Conference on Machine Learning , pages =

    On Numerical Integration in Neural Ordinary Differential Equations , author =. Proceedings of the 39th International Conference on Machine Learning , pages =. 2022 , editor =

    2022

  14. [14]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Ghanem, Paul and Demirkaya, Ahmet and Imbiriba, Tales and Ramezani, Alireza and Danziger, Zachary and Erdogmus, Deniz , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2025 , doi =

    2025

  15. [15]

    , title =

    Centofanti, Edoardo and Ghiotto, Massimiliano and Pavarino, Luca F. , title =. Computer Methods in Applied Mechanics and Engineering , volume =. 2024 , doi =

    2024

  16. [16]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Improving Multi-Step Prediction of Learned Time Series Models , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2015 , url =

    2015

  17. [17]

    Proceedings of the 35th International Conference on Machine Learning , series =

    Lipschitz Continuity in Model-based Reinforcement Learning , author =. Proceedings of the 35th International Conference on Machine Learning , series =. 2018 , publisher =

    2018

  18. [18]

    arXiv preprint arXiv:2504.01766 , year =

    Learning with Imperfect Models: When Multi-step Prediction Mitigates Compounding Error , author =. arXiv preprint arXiv:2504.01766 , year =. 2504.01766 , archivePrefix =

    Pith/arXiv arXiv

  19. [19]

    and Linderman, Scott W

    Tanoh, Ian Christopher and Deistler, Michael and Macke, Jakob H. and Linderman, Scott W. , title =. Advances in Neural Information Processing Systems 38 (NeurIPS 2025) , year =

    2025

  20. [20]

    , title =

    Lei, Chon Lok and Mirams, Gary R. , title =. Frontiers in Physiology , volume =. 2021 , doi =

    2021

  21. [21]

    Cubature

    Demirkaya, Ahmet and Imbiriba, Tales and Lockwood, Kyle and Rampersad, Sumientra and Alhajjar, Eyad and Guidoboni, Giovanna and Danziger, Zachary and Erdogmus, Deniz , booktitle =. Cubature. 2021 , organization =

    2021

  22. [22]

    Recognition Models to Learn Dynamics from Partial Observations with Neural

    Buisson-Fenet, Mona and Morgenthaler, Valery and Trimpe, Sebastian and Di Meglio, Florent , journal =. Recognition Models to Learn Dynamics from Partial Observations with Neural. 2023 , url =

    2023

  23. [23]

    arXiv preprint arXiv:2404.18572 , year =

    Learning Governing Equations of Unobserved States in Dynamical Systems , author =. arXiv preprint arXiv:2404.18572 , year =

    Pith/arXiv arXiv

  24. [24]

    and Zaharieva, Dessi P

    Zou, Bob Junyi and Levine, Matthew E. and Zaharieva, Dessi P. and Johari, Ramesh and Fox, Emily B. , booktitle =. Hybrid ^2 Neural. 2024 , url =. 2402.17233 , archivePrefix =

    arXiv 2024

  25. [25]

    bioRxiv , year =

    Hybrid Neural Differential Equations to Model Unknown Dynamics in Biology , author =. bioRxiv , year =

  26. [26]

    1979 , volume =

    Stochastic Models, Estimation, and Control , author =. 1979 , volume =

    1979

  27. [27]

    1979 , publisher =

    Optimal Filtering , author =. 1979 , publisher =

    1979

  28. [28]

    arXiv preprint arXiv:2001.04385 , year =

    Universal Differential Equations for Scientific Machine Learning , author =. arXiv preprint arXiv:2001.04385 , year =

    Pith/arXiv arXiv 2001

  29. [29]

    Learning Physics Informed Neural

    Ghanem, Paul and Demirkaya, Ahmet and Imbiriba, Tales and Ramezani, Alireza and Danziger, Zachary and Erdogmus, Deniz , booktitle =. Learning Physics Informed Neural. 2025 , doi =

    2025

  30. [30]

    Proceedings of the 39th International Conference on Machine Learning , series =

    On Numerical Integration in Neural Ordinary Differential Equations , author =. Proceedings of the 39th International Conference on Machine Learning , series =. 2022 , url =

    2022

  31. [31]

    A Hybrid

    Demirkaya, Ahmet and Lockwood, Kyle and Stratis, Georgios and Imbiriba, Tales and Ilie. A Hybrid. IEEE Transactions on Biomedical Engineering , year =. doi:10.1109/TBME.2024.3505796 , note =

    arXiv 2024

  32. [32]

    Computer Methods in Applied Mechanics and Engineering , volume =

    Learning the Hodgkin--Huxley model with operator learning techniques , author =. Computer Methods in Applied Mechanics and Engineering , volume =. 2024 , doi =

    2024

  33. [33]

    2022 25th International Conference on Information Fusion (FUSION) , year =

    Hybrid Neural Network Augmented Physics-based Models for Nonlinear Filtering , author =. 2022 25th International Conference on Information Fusion (FUSION) , year =. doi:10.23919/FUSION49751.2022.9841291 , url =

    arXiv 2022