Learning Dynamical Systems from Noisy Data with Inverse-Explicit Integrators

We introduce the mean inverse integrator (MII), a novel approach to increase the accuracy when training neural networks to approximate vector fields of dynamical systems from noisy data. This method can be used to average multiple trajectories obtained by numerical integrators such as Runge-Kutta methods. We show that the class of mono-implicit Runge-Kutta methods (MIRK) has particular advantages when used in connection with MII. When training vector field approximations, explicit expressions for the loss functions are obtained when inserting the training data in the MIRK formulae, unlocking symmetric and high-order integrators that would otherwise be implicit for initial value problems. The combined approach of applying MIRK within MII yields a significantly lower error compared to the plain use of the numerical integrator without averaging the trajectories. This is demonstrated with experiments using data from several (chaotic) Hamiltonian systems. Additionally, we perform a sensitivity analysis of the loss functions under normally distributed perturbations, supporting the favorable performance of MII.

Paper

Similar papers

Reviewer yxuC5/10 · confidence 2/52023-07-03

Summary

This paper introduces a new integration method (mean inverse integrator) for learning dynamics from noisy data. Experiments on Hamiltonian systems show the effectiveness of the proposed method.

Strengths

- The problem of learning physical dynamics from noisy data is an interesting one. - It combines techniques from the field of numerical computation with machine learning.

Weaknesses

- The usefulness and significance of the proposed method is not clear. - I feel that comparative experiments are not sufficient. - It is not clear what advantages the proposed method has over the naive noise handling method.

Questions

- To handle noisy data, the most common method is to introduce an observation model (e.g., Normal distribution) and learn the noise variance by maximum likelihood estimation. Can you tell us what advantages the proposed method has over such naive methods? - It would be good to clarify situations that can only be solved by the proposed method and emphasize the significance of this paper. For example, when using Gaussian likelihood, the noise variance is usually assumed to be constant. Can the proposed method relax this assumption? In other words, can the proposed method be applied to cases where the noise variance depends on time and state? - A thorough comparison with the latest methods (e.g., [9-12, 15]) would be helpful. - The font size in Fig. 5 and 6 is too small and should be enlarged.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

3 good

Limitations

It would be good to add a careful discussion of the advantages and disadvantages of the proposed method.

Reviewer HMrx6/10 · confidence 4/52023-07-06

Summary

The paper investigates mono-implicit Runge--Kutta (MIRK) methods for learning dynamical systems from data. In particular, MIRK methods can be made explicit by introducing the external data into the solver step itself, leading to a more efficient integrator while keeping favorable stability, symmetry, and symplecticity properties. To handle noisy data, the paper proposes the "mean inverse integrator" as an efficient way to average multiple trajectories and learn meaningful vector fields from these. The methods are demonstrated in multiple numerical experiments.

Strengths

The paper is well-written, the proposed method is presented very clearly, and the main claims made are well-supported.

Weaknesses

I find the presentation of the explicit Runge--Kutta (ERK) baseline a bit confusing, and in particular counter-intuitive of my current understanding of the usage of ERK methods for inverse problems in the context of ODEs. Concrete questions are below, in "Questions". This criticism also extends to insufficiently detailed baselines, and potentially to insufficient baselines overall.

Questions

See also "Weaknesses". Notably: 1. How exactly is the RK4 / ISO+RK4 baseline defined? 2. Why is (6) _the_ appropriate loss function? I would rather assume something of the type $\sum || \tilde{y}_n - \hat{y}_n ||$, where $\hat{y}_n$ is a state estimate computed by the method; for instance by applying a numerical solver to simulate a whole trajectory. In particular, extrapolating exactly from the last data point seems to be a strong restriction, in particular if the data point can be very noisy. I would not consider this "the inverse problem formulation", but rather a specific choice that is made, geared towards the specific algorithm of interest. 3. Related to the last question: Could you commend on the standard least-squares approach of integrating the whole trajectory from some initial value (estimate) with RK, as opposed to the described method of just integrating from one data point to the next? The latter seems to be related to "multiple shooting"-type approaches which, while they can be very beneficial, seems to be far from the standard for neural ODEs. Though, to the best of my knowledge, even in multiple shooting it is common to extrapolate some state estimates $\hat{y}$; though part of the objective is the closeness of these estimates to the data, and they are initialized on the data. Other remarks: - l. 201: "evaluationg the vector filed in $\tilde{y}_n$, as is done in all explicit RK methods": To the best of my understanding this is not done in all explicit RK methods, but only in multiple shooting-type approaches (and even there this might not hold). A standard least-squares approach simulates a whole trajectory and then computes an L2-loss---and thus does not need to evaluate $f$ on the data points themselves. - l. 129: $\tilde{y}$ is missing an index - Related work: The write-up is very helpful; but I think it could be further improved by highlighting better how the proposed paper relates to these methods and highlight similarities and differences. - l. 72-73: I agree with the sentiment of the statement, but I think the strict _necessity for numerical integrators_ is somewhat disputed in the literature. Notably, "gradient matching" methods compute first an interpolant, and then adjust the vector field to match the known derivative of the interpolant---this circumvents using RK or any related numerical method, which is why such methods often claim to be "simulation-free".

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

The authors address limitations in a dedicated section, which is much appreciated. One additional limitation that is not explicitly mentioned in the section that comes to mind is the influence of the noise on the data: Since the data is explicitly included into the numerical solver, as opposed to being just part of some L2 loss to guide some estimated trajectory, I would expect that for very noisy data other methods might be preferrable (in particular a least-squares approach with RK from a learned initial value). Another potential limitation: I assume that this method is not able to be a plug-in replacement in some latent neural ODE setting, where the ODE trajectory is not observed directly but only in some transformed space, e.g. when having video data and modeling an ODE in latent space? This is of course far from a trivial setting and I do not expect that such specific scenarios need to be mentioned explicitly; but the necessity of actual trajectory observations, as opposed to partial or non-linear observations, could indeed be another limitation.

Reviewer j2qw6/10 · confidence 2/52023-07-06

Summary

The authors present a novel method, Mean Inverse Integrator (MII), used to aggregate data generated through numerical integration of the vector field characterizing Hamiltonian systems. In particular, the objective is to improve the training of Hamiltonian Neural Networks (HNNs) when the data used is noisy. The authors train HNNs on two tasks, using data generated by different methods of numerical integration.

Strengths

- The paper is well written and complete. My background on the topic is limited, and so I was happy that Sections 3 and 4 were included. - The theoretical analysis is thorough and convincing.

Weaknesses

- I am not sure how much the paper fits with the themes of NeurIPS, but it should still be of interest to some of the audience.

Questions

N/A

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

N/A

Reviewer 1rk94/10 · confidence 4/52023-07-06

Summary

The presented work considers a novel class of integrators that are used to train Hamiltonian Neural Networks (HNNs). This class is called mean inverse integrator and it averages the trajectories from mono implicit RK methods (MIRK) to obtain higher accuracy. The authors provide theoretical results on how MIRK convergence suffers from noisy data. Also, experimental results on test dynamical systems illustrate the performance of the proposed approach.

Strengths

The paper presents a comprehensive introduction to the different types of integrators and smoothly introduces the new one. The novel integrator is investigated w.r.t. the relations with classical RK methods (Theorem 4.4 and Proposition 4.3). Also, theoretical results on the robustness w.r.t. noise in the data are presented.

Weaknesses

1) the experimental results show that the proposed approach is not always better than the alternative methods in terms of accuracy, but always slower in terms of runtime. So the use cases for the MII-related approaches should be stated more strictly. 2) the manuscript is mostly about integrators theory rather than learning dynamical systems from data. Revision of the structure and shift of the focus from integrators to its application in learning HNN can be very helpful 3) the connection of the proposed integrator and its usage in forward or in backward passes is ignored. Thus, it is unclear from the text how this new integrator should be incorporated into the existing pipeline of training HNN 4) experiments are performed only for systems, where the states are small dimensional. The scalability and robustness w.r.t the large dimension of states are unclear 5) since no batching is used for training, it indicates that a large amount of data are not tested yet and all available data are fitted in the GPU memory. It will be interesting to study how the stochasticity of the data and corresponding gradient estimates affect the convergence of the training curves.

Questions

1) how do different integrators affect the way to perform backward passes? Please provide a detailed explanation and complexity analysis 2) even for 4d states the proposed approach (MII MIRK4) shows the worst runtime compared to the alternatives (Fig. 6). So, how do the MII-related integrators scale w.r.t. the state dimension? How are they applicable to learning dynamical systems with high-dimensional states? 3) all experimental results are shown for fixed noise level $\sigma= 0.05$, but authors highlight that the proposed approach is robust to noisy data. So please show how the test accuracy depends on the noise level in training data for different integrators 4) please provide analytical forms of the considered test dynamical systems for the reader's convenience

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

The authors explicitly state the limitations of the proposed method.

Reviewer J6Zi5/10 · confidence 2/52023-07-07

Summary

This paper introduces a novel method aimed at learning the vector field of a dynamical system. The proposed approach is called the mean inverse integrator, which utilizes a neural network (e.g., SRNN) to accurately estimate the integrator in the presence of noisy data. The authors provide theoretical insights into the sensitivity of both the one-step target function and the mean inverse integrator to data noise. Additionally, the paper presents empirical evaluations by comparing the method to five different types of integrators.

Strengths

- The paper effectively conveys its ideas and arguments with clarity. - The research addresses an important question and presents a new approach to handle noise when learning data dynamics. - Theoretical analysis shows how the proposed method and the baseline approach respond to noise, contributing to a better understanding of their performance. - Empirical results demonstrate the effectiveness of the proposed method, showing significant advantages over the baseline approaches.

Weaknesses

- The paper lacks clarity on why the mean inverse integrator yields more accurate estimates compared to the one-step baseline. - Scalability to high-dimensional systems is not explored, as current experiments primarily focus on trivial test cases. It would be beneficial to investigate potential challenges arising from increased computational complexity, instability of estimation in the presence of chaotic dynamics, and provide examples demonstrating the method's efficacy.

Questions

Q1: The paper would benefit from running experiments on complex, high-dimensional data to showcase the method's capabilities. Q2: It remains unclear whether the proposed method can generalize to more general dynamical systems beyond Hamiltonian systems.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

N/A

Reviewer HMrx2023-08-15

Reply

Thank you for the reply and for for clarifying a few points. As a result I increased my score by 1.

Reviewer yxuC2023-08-18

Reply

Thank you for your answer. I now understand the advantage of the proposed method (no need to assume noise distribution). I will raise my score by 1.

Reviewer J6Zi2023-08-20

I acknowledge the reading of the authors’ replies and am aware of the discussion related to the noise scales. I decide to keep my score. Thank you!

Reviewer 1rk92023-08-20

Dear authors, Thanks for the detailed response! I have the following comments 1) The incorporation of the proposed integrator in large-scale systems is an important part of experimental evaluation. In such a setup the naive implementation of the backward pass is not enough. 2) Scalability is also crucial for practical applications, so I find the current status of the work is preliminary. To sum up, I have decided to keep my evaluation.

Program Chairsdecision2023-09-21

Decision

Reject

© 2026 NYSGPT2525 LLC