acceptodds
Under review as a conference paper at ICLR 2027

DVSI: Dynamical Variational System Identification with High-dimensional State

Abstract

We address the problem of identifying a nonlinear stochastic Markov dynamical system that has many state variables represented together by a high-dimensional state vector. We focus on the case where the state-transition model (STM) of the system has no known analytical structure due to lack of physics knowledge, is modeled using a neural network, and is identified from noisy measurements through a known measurement model. Particle-filter-based system identification, including differentiable particle filters (DPFs), suffers from the curse of dimensionality in this setting. To address the curse of dimensionality, we develop Dynamical Variational System Identification (DVSI): transformer-based dynamical variational inference methods that jointly learn the STM and an amortized posterior of the state trajectory by maximizing an evidence lower bound. We develop four DVSI methods, spanning autoregressive versus factorized and smoothing versus filtering posteriors. Through experiments on stochastic Lorenz-63 and Lorenz-96 systems and real magnetoencephalography (MEG) data, we evaluate the DVSI methods for identifying dynamical systems that have up to 200 state variables, where DPFs have significant limitations due to the curse of dimensionality.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
State estimation, system identification, dynamical variational auto encoders

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_dnYFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

In this paper, the authors considered the standard filtering and smoothing problem of state space models. They aimed to design a different approach to learning latent dynamics: one that retains probabilistic inference over the hidden trajectory while avoiding an explicit particle approximation whose fidelity deteriorates with state dimension. They introduced Dynamical Variational System Identification (DVSI), in which a parameterized inference model maps a sequence of noisy measurements into a distribution over plausible state trajectories, while a second learnable model represents the unknown stochastic state-transition model (STM). They evaluated DVSI on nonlinear synthetic systems (stochastic Lorenz-63 and Lorenz-96) and high-dimensional real-world trajectories (MEG data).

Critical strengths

The paper is clearly written and well organized. The filtering and smoothing problem they consider is important.

Critical weaknesses
  1. The model they considered is restricted to Gaussian transition dynamics (equation (1)) and Gaussian emission density (equation (2)). They commented on particle filters (PFs) as being not scalable, which is their main motivation. There are two issues. First, there are already various scalable PFs developed in recent years, such as "Can local particle filters beat the curse of dimensionality?" (AAP 2015), "Approximate Smoothing and Parameter Estimation in High-Dimensional State-Space Models" (IEEE Transactions on Signal Processing 2017), and "Iterated Block Particle Filter for High-dimensional Parameter Learning: Beating the Curse of Dimensionality" (JMLR 2023). Second, PFs are applicable for general SSMs where the problem considered in this paper is its Gaussian special case.

  2. The paper does not have any Math proofs which is not common in this field for top publications. See the above three references as examples.

Questions for authors
  1. See comment 1 above. How do the authors justify the claim that PF-based approaches are not scalable without considering these developments?

  2. Can the authors provide theoretical justification for the claimed scalability advantages of DVSI, or clarify whether these claims are intended to be purely empirical?

  3. What is the fundamental algorithmic or theoretical novelty of DVSI relative to existing deep variational state-space models and variational system-identification methods? The techniques used in this paper are very well-established ingredients in deep variational dynamical modeling. The paper does not demonstrate what is algorithmically new beyond applying these combinations.

  4. Can the author provide the source code? Lacking reproducibility significantly weakens the credibility of their empirical claims. That is crucial for this empirical paper.

  5. Can you establish fairness of the particle-filter baseline? The described comparison seems not fair. The paper compares DVSI primarily against one differentiable particle filter (DPF), using only 25 particles. As PFs being Monte Carlo based approaches, 25 particle are way too small and one rare taken this low.

Reviewer_FxAgOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

the authors focus on nonlinear system identification using amortized variational inference to make it efficient. they call their approach Dynamical Variational System Identification (DVSI). they use a transformer to amortize inference and evaluate their method (and subvariants i.e. filtering, smoothing variants) on synthetic data and MEG data.

Critical strengths

the writing is very clear

Critical weaknesses

the biggest weakness in my opinion is sufficient novelty. previous works (which are even cited here) such as the deep kalman filter have already developed this machinery (amortized variational inference for a state-space graphical model). and there are also works not cited here like the deep variational bayes filter that are very relevant. transformers have also been used to amortize inference similarily in state space models (see 2). the other weakness is that the experiments are not very convincing, they could be much more interesting and outside of their own model class and a particle filter do not compare to any other methods that are state of the art and contemporary.

  1. Deep Variational Bayes Filters: Unsupervised Learning of State Space Models from Raw Data, ICLR 2017
  2. Karniol-Tambour, Orren, et al. "Modeling state-dependent communication between brain regions with switching nonlinear dynamical systems." International Conference on Learning Representations. Vol. 2024. 2024.
Questions for authors

if the authors could help describe their contribution taking into account existing literature that would help to clarify things

Reviewer_2kyrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper is about learning a discrete-time latent transition density of a state-space model parameterized by a Neural-based Gaussian density. It uses basic and fairly standard variational inference schemes. It proposes several variational distribution choices to be learned jointly with the latent transition density, i.e. a joint smoothing approach, an AR approach and a time-factorised approach (mean-field). Results are presented for 2 simulated datasets and a real MEG dataset

Critical strengths

The paper is exceptionally clearly written. I didn't spot any errors.

The paper addresses a useful problem of practical concern.

Critical weaknesses
  1. The implementation is a fairly standard implementation of variational inference to state-space models. There is quite a lot of literature on this type of thing, especially the state inference part - see book by Quin and Smidl, and recent papers by R. Gan et al, to name a few...
  2. The method looks like a 1-particle version of the VSMC method. Both the autoregressive and factorized variants of the variational distributions should be compatible with that framework.
  3. The paper also does not mention how the dynamical recovery is improved but instead talks about scalability, which is the standard particle filter posterior inference compared to the variational inference statement (it is also not linked to model recovery). Also, this statement would be most serious for the basic bootstrap particle filter, while the variational-based particle methods (VSMC or IWAE, which are essentially generalizations of the authors’ scheme) would perhaps not share such an issue
Questions for authors
  1. Your method seems to be a special case for both VSMC and IWAE. How do you distinguish it from these?
  2. The authors mentioned VSMC in the appendix, and the implementation for the model transition density and the variational density should both be compatible with the framework. Have you matched the implementations?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.