acceptodds
Under review as a conference paper at ICLR 2027

Identifying Neural Dynamics Using Nonlinear Interventional State Space Models

Abstract

State-space models (SSMs) recover low-dimensional latent dynamics from high-dimensional observations, but latents are generally not identifiable from observational data alone. Interventional state-space models (iSSMs) use perturbations to break observational symmetries and identify parameters up to permutation, scaling, and shift. However, prior work restricts identifiability to linear latent dynamics and piecewise-linear emissions, excluding many behaviors characteristic of neural circuits. We extend iSSMs to a substantially broader model class: we prove that (i) affine identifiability holds under general injective emission functions and Gaussian latent transitions, and (ii) with interventions, identifiability up to permutation, scaling, and shift is preserved under general nonlinear latent dynamics. We validate our theory on simulated FitzHugh-Nagumo dynamics, and then demonstrate two applications: adjudicating between competing models of decision-making; and predicting responses to unseen micro-stimulation in macaque and human neural recordings, with proof-of-concept generalization to a held-out stimulation site in the latter. Finally, simulations yield practical guidance for designing stimulation protocols that improve identifiability.

Primary area:
applications to neuroscience & cognitive science
Keywords:
state-space models, identifiability, neural dynamics, interventions, micro-stimulation, causal representation learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_BdGDOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2(+1)
Read review
Paper summary

The authors extend the iSSM framework, which only supports linear or piecewise linear systems, to support nonlinear models. They show that the new nonlinear extension provides a better fit when the underlying dynamical system is truly nonlinear, and it can improve the neural reconstruction performance on real neural data.

Critical strengths
  • The motivation is clear, as extending linear or piecewise linear modeling assumptions to nonlinear models is an important ongoing direction of research in time-series modeling and neuroscience. Beyond that, intervention modeling is another important and valuable direction that can benefit neuroscience studies.
  • The simulations clearly show the benefit of nonlinear modeling.
Critical weaknesses
  • Overall, I think the paper is hard to follow, and it is written in a way that assumes the readers have significant knowledge of the related work. For example, the terms "do-intervention" and "observational symmetry" are not defined, even briefly, when they first appear. This results in a situation where I have a hard time understanding how this work provides a solution to the proposed problem. I understand that the authors are limited to 9 pages, but I believe the solution is not to put important details in the Appendix or drop the details about the problem and modeling formulation that could significantly improve the flow of the paper.
  • How do the modeling assumptions meet the model fit? I see that the authors used a VAE approach to fit their model, but how does it differ from other time-series VAE models such as LFADS?
  • Similar to the missing model fit steps and clear explanation of the link between model fit and assumptions, I find the evaluation steps also under-specified. For example, how do the authors generate the rollouts? Even though it is briefly explained in the appendix, they provide no links to related sections or explain it when it appears.
  • Is there a way to relax the assumption of having independent latent noise? In the domain of linear SSMs, allowing a full-rank process-noise covariance typically improves model fit, as it can absorb residuals and correlations that cannot be captured by the dynamics alone. Can the authors comment on this?
  • In line with my modeling and assumptions comment above, I could not understand how the mask vector mm is formed. How is it formed, or is it fitted alongside the model? In practice, it is hard to control which latent state should receive the perturbation or not, as their semantic meaning is unknown.
  • Even though the paper reads as if identifiability is an important motivation for this work, I believe the results do not highlight it as much. Even though the simulation results show this to some degree by providing true vs. identified flows, no real-data result shows any type of interpretability or a scientific finding, and the results are limited to rollout accuracy and reconstruction accuracy, which could be obtained by other types of nonlinear models that support modeling of external stimuli.
  • In line with my previous comment, the comparisons are limited to iSSM, whereas there exist other nonlinear models of neural activity that can model external stimuli, such as LFADS and BRAID, or a simple transformer model autoregressively trained where the input stream is represented as a token sequence.
  • Figures are missing legends. Even though the authors explain the color scheme in the captions of Figs. 1 and 2, the figures themselves should have the legends to explain the color scheme and symbols.
Questions for authors
  • Can you explain the connection between Figs. 2b and 2c? Is Fig. 2c just a 2D visualization of Fig. 2b? If so, it looks like both linear and nonlinear latents meet the true latents at the start and the end, but it doesn't seem like that in Fig. 2c.
14 comments
AuthorsOpenReview ↗

Response to Reviewer BdGD: W1, W2

We genuinely appreciate your detailed feedback, which has substantially improved the paper. In summary, we:

  • (W1) clarified key terms in the Introduction
  • (W2) added a Methods section on model fitting (Sec. 3.3) and a Related Work paragraph comparing iSSM to LFADS and BRAID (Sec. 2)
  • (W3) rewrote the evaluation section (Sec. 3.4)
  • (W4) relaxed the independent-noise assumption, and showed empirically that latent recovery remains near-perfect under correlated noise (App. J)
  • (W5) clarified how the intervention mask is formed (Sec. 3.3)
  • (W6) show that iSSM's latents separate epileptogenic from healthy stimulation sites (Sec. 4.5)
  • (W7) benchmarked iSSM against LFADS and BRAID (App. K)
  • (W8) added figure legends

All changes are marked in blue in the revised PDF. Detailed responses follow in this and the next comments.

BdGD, W1: Clarify key technical terms

We are sorry that you found our manuscript hard to read. We tried our best to make the paper accessible to readers in both neuroscience and causal representation learning, but we acknowledge it may have been imperfect. So thank you for the valuable feedback. The two terms you mentioned are indeed central:

  • A do-intervention (a.k.a hard intervention) sets the intervened coordinate's value equal to the (possibly noisy) input value, like setting a “control knob”. In contrast, an additive soft intervention adds the input to the current coordinate's value. Intuitively, do-interventions help identifiability because a driven latent no longer depends on the rest of the system, so any downstream effects can be clearly attributed to the input (Ahuja et al., 2023 illustrates this nicely).

  • Observational symmetries are transformations (e.g., an affine map) of the latent model that leave the distribution of observations unchanged; any two latent models that differ by those transformations cannot be distinguished from observational data.

We clarified these terms in the revised Introduction.

Please let us know if there are any other terms/steps that would benefit from expanding beyond the ones addressed below.

BdGD, W2: How assumptions meet the model fit

Thanks a lot for raising this.

(A) Model fitting

We agree that detailing our model-fitting procedure, and how it implements our theoretical assumptions, in the main text improves clarity. We have added a new Methods section (Sec. 3.3, Model fitting); here’s a summary:

  1. Posterior.
  • An LSTM encoder maps (y1:T,u1:T)(y_{1:T}, u_{1:T}) to the mean and standard deviation of a Gaussian posterior, from which we sample a latent trajectory x1:Tx_{1:T}. We use a Gaussian posterior to match the Gaussian latent assumption in Theorem 2 (A2).
  • The input (But)j(Bu_t)_j to each latent jj determines the binary mask mtm_t, which indicates which latents are currently not driven ((But)j=0(Bu_t)_j=0). This mask clamps targeted latent means to the input via μt+1←mt⊙μt+1+But\mu_{t+1} \leftarrow m_t \odot \mu_{t+1} + Bu_t, implementing the do-intervention mechanism assumed by our theory.
  1. Prior. Each posterior sample xtx_t is passed one step through the transition model (Eq. 2), giving the Gaussian transition prior N(mt⊙gϕ(xt)+But,Q)\mathcal{N}(m_t \odot g_\phi(x_t) + Bu_t, Q). Targeted latent means are again clamped to the input, mirroring the encoder and implementing the same assumptions (Gaussian latents and do-interventions).

  2. Likelihood. Posterior samples are decoded by fθf_\theta.

  3. Sparsity. Finally, we enforce sparsity on BB by thresholding small entries during training, since a dense BB would violate the unordered-pairs condition (Assumption 2) required for permutation-scale-shift identifiability (Corollary 4).

(B) How does iSSM differ from LFADS?

We have added a paragraph to the Related Work (Sec. 2, Latent dynamical models of neural activity without identifiability) that compares our model to LFADS and BRAID.

All three learn nonlinear, input-driven latent dynamics with recurrent inference networks. However, they differ in two aspects of the generative model, which matter for our identifiability guarantees:

  1. Stochastic latents: iSSM latents follow a stochastic transition model, while LFADS and BRAID use deterministic generator networks.
  2. Hard interventions: iSSM clamps targeted latents to the input, whereas LFADS and BRAID feed inputs into the latent update alongside the current state.

As a result, LFADS and BRAID may fit neural data well, but their latents are not guaranteed to be unique; we now also show this empirically (see BdGD-W7).

We'd like to emphasise though, that our contribution is not a new architecture competing with LFADS or BRAID. We offer a set of conditions on the generative model under which latents are provably identifiable: mainly Gaussian latent transitions, injective emissions, and hard interventions. Our VAE is one way to fit such a model; but any other latent dynamical model that meets these conditions inherits the guarantees.

AuthorsOpenReview ↗

Response to Reviewer BdGD: W3, W4, W5

BdGD, W3: Underspecified rollout evaluation

Thank you, this is a fair point. We rewrote the evaluation-metrics section (Sec. 3.4), which was admittedly glossing over some details. The revision clarifies these, and links to the relevant appendix sections (App. C.4, F.4). We also now reference the evaluation section in our Results where we first mention our metrics (Sec. 4.1).

Each metric is the Pearson correlation between predicted and true trajectories, averaged over held-out trials, and excluding time points with non-zero inputs. Basically:

  • reconstruction accuracy = corr(reconstruction, true y)
  • rollout accuracy = corr(rollout, true y)
  • latent accuracy = corr(learned latents, true x)

How rollouts are generated:

  1. The encoder sees only the initial n<Tn<T timesteps of the trial (y1:n,u1:n)(y_{1:n}, u_{1:n}) and infers the initial latent mean μ0\mu_0. On the real data, nn is the pre-stimulation window.
  2. We iterate μ0\mu_0 through the learned dynamics model (Eq. 2) without noise, driving the model with the inputs utu_t; i.e., xt+1=mt⊙gϕ(xt)+Butx_{t+1} = m_t \odot g_\phi(x_t) + Bu_t. These are the learned latents we use for computing latent accuracy.
  3. Finally we decode the learned latents with fθf_\theta to obtain the rollouts.

How reconstructions are generated:

  1. The encoder sees the entire trial (y1:T,u1:T)(y_{1:T}, u_{1:T}) and infers the latent posterior mean trajectory μ1:T\mu_{1:T}.
  2. We decode μ1:T\mu_{1:T} with fθf_\theta to obtain reconstructions.

Note that reconstruction accuracy skips the dynamics model, and the encoder sees the full trial, which makes it acausal inference; Rollout and latent accuracy are therefore the stricter metrics.

Please let us know if there's anything else you'd like us to clarify.

BdGD, W4: Relax independent latent noise assumption

Thank you for the valuable suggestion.

  1. We have relaxed this assumption in our proof: Theorem 3 (Step 3) no longer requires independent process noise, nor faithfulness (Sec. 3.2 and App. A.4).

The new Theorem 3 assumptions are:

  • (B1) the transition functions gϕ,gϕ^g_\phi, g_{\hat\phi} are continuously differentiable, and Jgϕ(x∗)J_{g_\phi}(x^{\ast}) is invertible at some point x∗x^{\ast};
  • (B2) in every intervention regime, including the intervention-free one, the latent state has full support RD\mathbb{R}^D.

Note that this adds essentially no assumptions beyond Theorem 2: (B1) restates (A1) and (A3), and (B2) follows from the Gaussian process noise in (A2).

  1. We also tested this empirically in a new appendix (App. J):
  • Simulations. We simulated the bistable well such that the noise has correlation ρ\rho between the two latents with ρ∈{−0.6,0,0.3,0.6,0.9}\rho \in \lbrace -0.6, 0, 0.3, 0.6, 0.9 \rbrace, and fit nonlinear iSSMs with a learned full covariance. Latent accuracy stays near-perfect at every ρ\rho (median ≥0.994\geq 0.994, 10 seeds each; App. J.1, Fig. 13).
  • Real data. We retrained the macaque (D=12D=12) and human (D=25D=25) models with a learned full covariance (10 seeds each). This did not improve model fit, and learned noise correlations were weak (mean absolute off-diagonal correlation 0.050.05 on both datasets), suggesting that a diagonal QQ is sufficient for these datasets (App. J.2, Fig. 14).

BdGD, W5: How the mask vector is formed

Thank you for spotting another key term that deserves clarification in our manuscript.

We have expanded on this in the new Methods section (Sec. 3.3, Model fitting; see BdGD-W2) and now use explicit elementwise notation when defining the mask (mt,j=1{(But)j=0}m_{t,j} = \mathbf{1}\lbrace (Bu_t)_j = 0 \rbrace in Sec. 3.1).

To answer your questions:

  • Is the mask fitted? -- No; the mask is derived from ButBu_t, where BB is learned and utu_t is observed.

  • How is the mask formed? -- The binary vector mtm_t indicates which latents are not driven at time tt: we compute the input (But)j(Bu_t)_j to each latent jj, and set mt,j=1m_{t,j} = 1 where (But)j=0(Bu_t)_j=0 and 00 otherwise.

  • How to control which latent state is targeted in practice, where latents are unknown? -- In real data, we indeed don't know a priori which latents are targeted by an intervention. However, we don't need to know this. Our framework infers this from data by learning BB, and our theory guarantees that BB is recovered up to permutation and scaling under our assumptions. Fig. 2d in the submitted manuscript demonstrates this empirically.

We hope this clarifies the definition and role of the mask. Please let us know if anything remains unclear.

AuthorsOpenReview ↗

Response to Reviewer BdGD: W6, W7

BdGD, W6: Interpretability/ scientific value of learned latents

Thank you for this really important point, which motivated two new analyses, included as new Appendix (App. K, Fig. 15 & 16) and Results (Sec. 4.5, Fig. 6) subsections.

(1) LFADS and BRAID can predict neural dynamics, but they recover the wrong latents (new App. K, Fig. 15 & 16).

In response to BdGD-W7, we trained LFADS and BRAID on both real datasets, and on simulated data from the bistable well where we have ground-truth latents. LFADS and BRAID can achieve comparable rollouts and reconstructions. However, the simulations show that they learn the wrong latents. (more details in BdGD-W7)

(2) Only nonlinear iSSM's latents reflect clinical seizure-zone labels (new Sec. 4.5, Fig. 6).

We tested whether the learned latent dynamics capture clinically relevant structure that was never used during training. In the human iEEG data, we virtually stimulated all 49 trained stimulation sites and 26 unseen sites, and tracked how far the latent state moves away from its resting distribution. (We use the Mahalanobis distance metric, which is invariant to affine transformations, enabling cross-model comparisons (App. G.4)). We then grouped sites by their clinical label (epileptogenic vs. healthy tissue), which the models never saw during training.

Nonlinear iSSM's latent dynamics clearly visually separate epileptogenic from healthy sites in an interpretable way (Fig. 6), which replicates at unseen sites. BRAID and LFADS show no such separation, neither for trained nor unseen sites. Thus, among models with comparable forecasts, only nonlinear iSSM infers latents that reflect the clinical labels.

Please note this single-patient analysis is intended as a proof-of-concept. Real neural recordings lack ground-truth latents, and our datasets do not include behavioral tasks or additional labels to test further hypotheses. In line with ICLR's technical focus, our primary contribution is the theoretical framework for identifiable nonlinear dynamics, addressing a gap that e.g. the BRAID authors call a "fundamental challenge" (Vahidi et al., 2025). A particularly promising application for future work is modelling recordings during cognitive tasks, where identifiable latents might help to arbitrate between competing mechanistic hypotheses (as in Sec. 4.2, Fig. 3).

BdGD, W7: Benchmark against LFADS, BRAID

Thank you for this valuable suggestion. We benchmarked iSSM against LFADS and BRAID and added the results in a new Appendix (App. K, summarised in the new Sec. 4.5). We did not include the transformer because it has no latent state, and hence can't be evaluated on latent recovery or interpretability.

What we did.

We compared iSSM, BRAID, and LFADS on (1) macaque dlPFC, (2) human iEEG, and (3) bistable-well simulations with ground-truth latents.

  • LFADS: Using lfads-torch (Sedler et al., 2023), we first tested the controller variant, which infers latent inputs from the full observed trial at every step. However, this acausal LFADS version performed poorly on our causal forecasting evaluation metrics (see BdGD-W3). Thus, we switched to the standard no-controller LFADS variant, which is driven only by the observed inputs utu_t.

  • BRAID: Following Vahidi et al. (2025), we used U-BRAID, the variant that does not require behavioral task variables.

  • Training: All models matched iSSM's setup (same latent dimension DD, likelihood, early-stopping criterion; and for the human data, electrode-coordinate-parameterised inputs for unseen-site prediction). For each dataset, we tuned 8 hyperparameter configurations (including the authors' defaults) and trained the best configuration across 10 seeds.

Results.

  • Real data (App. K.5, Fig. 16) -- iSSM and BRAID perform comparably; LFADS underperforms. On human iEEG, iSSM and BRAID achieve similar rollout accuracy; both outperform LFADS. On macaque dlPFC, all three models achieve similar rollout accuracy on stimulation trials, while BRAID slightly outperforms iSSM on control trials. iSSM achieves the highest reconstruction accuracy on both datasets.

  • Bistable well (App. K.4, Fig. 15) -- LFADS and BRAID recover the wrong latents regardless of rollout accuracy. BRAID matches iSSM's rollout accuracy, but fails to recover the true latents (0.65±0.070.65 \pm 0.07 vs. iSSM's 0.98±0.010.98 \pm 0.01, p=0.002p = 0.002). LFADS achieves moderate rollout accuracy (0.75±0.030.75 \pm 0.03) and poor latent recovery (0.21±0.050.21 \pm 0.05).

  • Interpretability (Sec. 4.5, Fig. 6). Only nonlinear iSSM's latents separate epileptogenic from healthy stimulation sites (see BdGD-W6).

In short, as you anticipated, other nonlinear models can forecast neural activity well. What iSSM adds are theoretical guarantees for recovering the true latents (given assumptions, and up to permutation, scaling, and shift). We show that this yields accurate latent recovery in simulation and more interpretable latents in human recordings.

Reviewer_pGN7OpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The authors introduce a nonlinear interventional state space model, which fits dynamical observations as well as observed interventions/perturbations to learn parameters that are identifiable up to permutation, scaling and shift. The authors provide sufficient conditions for identifiability in a series of propositions/theorems and their accompanying proofs. They then show the performance of their model on increasingly complex problems, from a simple nonlinear dynamical system producing simulated data all the way to perturbation experiments in human neural recordings.

Critical strengths

This paper is very well written and clearly communicates the problem, the authors' approach, and the results.

The results show nice improvements over existing linear methods on a central and important problem in neuroscience.

I appreciate the authors validating their model on increasingly complex problems: a simple and generic nonlinear dynamical system first, then simulated data that is related to neuroscience problems, then moving to real data in macaques and humans. This succession builds intuition and understanding.

Critical weaknesses

I wouldn't call this critical, necessarily, but for someone who hasn't seen the iSSM I am curious to see some information about inference in this model in the main text, rather than just the appendix.

Questions for authors

Fig 4: there appear to be a lot of dynamics in the neural activity, both before and after interventions, that neither model is capturing well. Is this due to a latent dimensionality that is too small? Is this noise rather than signal? Why doesn't the nonlinear iSSM appear to predict any better than linear iSSM in the reconstructions?

The held out site evaluation in Fig 5 is interesting. In this case the held-out site is in the center of many stimulation sites, testing coordinate-based interpolation, as the authors mention. I'd be curious to see a spatial map of something like rollout accuracy on the held-out site as a function of the held-out site's location; does performance decrease towards the edges of the spatial map, when interpolation starts to transition to extrapolation?

3 comments
AuthorsOpenReview ↗

Response to Reviewer pGN7: W1, Q1

Thank you very much for your time and the positive assessment of our paper. We are happy to hear that you found it well written and easy to follow. We followed your suggestion to bring details about inference to the main text, and address your questions below.

pGN7, W1: Explain inference in the main text

Thanks a lot for the suggestion. We agree that explaining the inference steps in the main text would improve the accessibility of our paper. We have added a new Methods section on model fitting (new: Sec. 3.3). Basically, we fit models with amortized variational inference by maximising the ELBO:

  1. An LSTM encoder maps observations and inputs (y1:T,u1:T)(y_{1:T}, u_{1:T}) to a Gaussian posterior over latent trajectories (i.e., it outputs a latent mean and std for each time step: (μ1:T,σ1:T)(\mu_{1:T}, \sigma_{1:T})). Latents targeted by an intervention are clamped to the input (μt+1←mt⊙μt+1+But\mu_{t+1} \leftarrow m_t \odot \mu_{t+1} + Bu_t), which implements the hard intervention.
  2. We sample a latent trajectory from the posterior, and pass each sampled xtx_t one step through the transition model, again clamping targeted latents to the input as above. This gives the prior p(xt+1∣xt,ut)p(x_{t+1} \mid x_t, u_t), which is trained via the KL term in the ELBO.
  3. We decode the posterior sample through the decoder fθf_\theta to get a reconstruction and evaluate the likelihood p(yt∣xt)p(y_t \mid x_t).
  4. We also enforce BB sparsity during training.

We additionally revised our metrics section to expand on inference during evaluation (revised: Sec. 3.4). It explains how trained models infer latents and generate rollouts (= open-loop forecasts generated by iterating the learned transition model from the initial inferred posterior μ0\mu_0) and reconstructions (= decoded posteriors means) on held-out trials.

pGN7, Q1: Single-trial fluctuations and reconstructions in Fig. 4

Thank you, these are sharp observations! We'll answer your two questions in turn:

(1) What about the fast fluctuations, which neither model seems to capture?

We indeed believe that those fast fluctuations in the macaque data are mostly noise that is smoothed out by the models. The black traces in Fig. 4a,b are single-trial spike counts in 33 ms bins, while our models predict smooth firing rates under a Poisson likelihood; and during evaluation we only plot the mean-rate predictions. So the predictions aren't expected to reproduce the jagged bin-to-bin fluctuations in the spike counts. You can see that even the reconstructions, which come from the nonlinear LSTM encoder with access to the whole trial, are fairly smooth (Fig. 4a).

You also asked whether the latent dimensionality might be too small. In Appendix F.2 (Fig. 9), we sweep D∈{5,12,15,20}D \in \lbrace 5, 12, 15, 20 \rbrace and found that rollout accuracy doesn't improve from D=12D = 12 onwards. So a larger latent space doesn't seem to help capture the remaining fluctuations, which again suggests they are mostly noise.

(2) Why are nonlinear iSSM reconstructions not much better than linear iSSM's?

This is because reconstructions are generated without the part that differs between linear and nonlinear iSSM: the dynamics model. Linear and nonlinear iSSMs share the same nonlinear (bidirectional LSTM) encoder and the same nonlinear decoder. The only difference is the dynamics model's transition map gϕg_\phi in xt+1=mt⊙gϕ(xt)+But+ϵtx_{t+1} = m_t \odot g_{\phi}(x_t) + B u_t + \epsilon_t, where the linear model uses gϕ(xt)=Axtg_{\phi}(x_t) = A x_t.

Reconstructions are obtained by decoding the posterior means, inferred by the encoder from the entire trial. So both models usually reconstruct well, regardless of their dynamics. The dynamics model only affects the reconstructions indirectly during training, through the KL term (see W1); but not during inference. This is why the nonlinear iSSM's reconstruction advantage is small (though still significant; Fig. 4c).

In contrast, rollouts are generated by the dynamics model: the encoder only provides the initial condition, inferred from only the pre-stimulation window, and from there the transition model "rolls out" the rest of the trial, driven by the stimulation inputs (see the revised Sec. 3.4). This is also why we consider rollout accuracy the more informative metric.

AuthorsOpenReview ↗

Response to Reviewer pGN7: Q2

pGN7, Q2: Does held-out site rollout accuracy depend on how central the site is?

Thank you for this suggestion. We agree that it's an interesting question to explore.

Analysis (human iEEG dataset):

  • We measure the local sampling density ("centrality") of a stimulation site as the mean 3D Euclidean distance of the site to its 5 nearest neighbours (kNN-5).
  • We selected 8 stimulation sites such that they span the range of kNN-5 in the data, which was 6 to 25 mm. The held-out site in our manuscript (RC3-RC4) had kNN-5=6.3 mm, so it was indeed among the most central sites.
  • For each selected site, we retrained our main model, holding this site out in addition to the original held-out site RC3-RC4. We trained 5 seeds per site. Apart from the trials of that site, the training, validation and test trials were identical to those of the original model.

Results:

We measured the cost of holding a site out as the difference between the rollout accuracy of the retrained model (site held out) with that of the original model (site held in), including only the 1-2 trials that were in the test set of the original model. The table below lists this difference (mean ±\pm sem across seeds) for each site.

SitekNN-5 (mm)ZoneΔr\Delta r (held out − held in)
RQ12-RQ136.3healthy−0.08 ± 0.05
RA2-RA38.2epileptogenic−0.14 ± 0.07
RK1-RK210.1epileptogenic−0.56 ± 0.06
RC6-RC711.8healthy+0.39 ± 0.19
RMG7-RMG813.7epileptogenic−0.02 ± 0.12
RN10-RN1116.5healthy−0.01 ± 0.09
RC13-RC1418.1healthy−0.04 ± 0.02
RX1-RX225.0healthy−0.31 ± 0.04

Across the 8 held-out sites ranging from kNN-5 6 to 25 mm, we find no evidence that held-out accuracy decreases as sites become more sparsely surrounded (Spearman ρ\rho = 0.14, p = 0.75). Most sites lose little accuracy when held out; only two lose substantially, but this is not predicted by how central they are. However, we note that with 8 sites and only 1-2 held-in test trials per site, this test has limited power. A more rigorous test would need denser, more uniform spatial sampling of stimulation sites and more repetitions per site -- which we agree would be an interesting direction for future work!

Given the limited statistical power of this test, we'd prefer to report these results only here and not include it in the manuscript. However, we would be happy to add them as an appendix, including a spatial map of Δr\Delta r, if you think it would be useful. Please let us know.

I thank the authors for their response. The addition of inference details makes the main text much clearer and more complete. I appreciate clarification of where differences in linear and nonlinear models will be obvious (rollout vs reconstruction).

I also appreciate the analysis of held-out site rollout accuracy as a function of kNN-5. I do believe these results are hard to interpret with such a small number of test trials, so I don't think it's necessary to report these results in the manuscript.

Reviewer_EfXiOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents nonlinear interventional state space models, demonstrating how they enable latent identifiability in situations that were not previously possible (nonlinear ssms). The paper provides theory on identifiability and many examples of improved performance over linear iSSMs. I note that, given my background, I did not follow all details of the proof - my score is thus driven most by empirical results (and I gave myself a somewhat low confidence rating).

Critical strengths
  1. Clarity of presentation
  2. Theoretical guarantees
  3. Demonstration of the approach on many synthetic and real datasets
Critical weaknesses
  1. At the moment, all comparisons are with a linear iSSM, but the other paper claim is that nonlinear interventional models are better than nonlinear observational models. While I do expect this claim to be true, this is not demonstrated in the text.

  2. I find the real data results to be a bit scientifically underwhelming. They show prediction performance, but not about how this method could actually be used for scientific insight (in particular they don’t go into how the identifiability is important for scientific insight, which is how you’ve primarily motivated the method in the first place).

Questions for authors

Are the “hard” interventions important empirically, or just for the theoretical proofs? That is, if you fit a model without the masking that leads to hard interventions (thus making the model like a standard nonlinear ssm), would your empirical performance be comparable (both in terms of predictions, and also ground-truth recovery in synthetic datasets)?

8 comments
AuthorsOpenReview ↗

Response to Reviewer EfXi: W1, W2

Thank you for your thoughtful review. We're glad you found our manuscript clear and easy to follow. Your comments, especially the one on scientific insight (W2), have substantially improved the paper.

In summary, we:

  • (W1) compared observational with interventional models (Fig. 2b);
  • (W2) added a new Results subsection, showing that iSSM's latents reflect clinical seizure-zone labels (Sec. 4.5, Fig. 6);
  • (Q1) show that hard interventions matter empirically (App. K.4)

All changes are marked in blue in the revised PDF. Detailed responses below.

EfXi, W1: Compare against observational models

Thanks for raising this. You are right that our main validation (Fig. 2) focused on the linear-vs-nonlinear comparison, even though we argue that latent recovery needs both a matched hypothesis class and interventions. We included a new analysis in Fig. 2b, comparing interventional with observational models.

We trained linear and nonlinear models on FitzHugh-Nagumo dynamics without interventions, 10 seeds each. The revised Fig. 2b now shows a 2x2 grid of inferred flow fields, (linear, nonlinear) x (observational, interventional), mirroring the schematic in Fig. 1c. Without interventions, even the nonlinear model fails to recover the limit cycle attractor. The Van der Pol replication shows the same pattern (App. D, Fig. 8b). This complements Sec. 4.6, where latent recovery of the nonlinear iSSM improves gradually as the number of training interventions grows from zero (Fig. 7a; replicated on two further systems in App. H.2).

The related question about the importance of hard interventions is answered in Q1.

EfXi, W2: Real data scientific insight

Thanks a lot, this is a fair point. While our primary focus was the theoretical framework for identifiable nonlinear dynamics, we totally see your point that an example scientific analysis on the real data would substantially strengthen our paper. We have added a new Results subsection (Sec. 4.5, Fig. 6) showing that iSSM's latents capture clinically meaningful structure that is missed by models without identifiability guarantees.

Analysis:

Using our framework for scientific insight on real data requires a testable hypothesis. Our real datasets include no behavioural task; however, the human iEEG data contains clinical labels, marking each stimulation site as epileptogenic or healthy tissue. We therefore tested whether the learned latent dynamics reflect these clinical labels, which were never used during training.

Using the trained model, we simulated virtual stimulations at all 49 trained sites and 26 unseen sites (the held-out site plus 25 never-stimulated sites), and tracked how far the latent state moves from its pre-pulse resting distribution (using Mahalanobis distance, which is affine-invariant and hence comparable across models; App. G.4).

Results:

Nonlinear iSSM's latents clearly visually separate epileptogenic from healthy tissue: after the initial response, latents depart from rest again for epileptogenic sites, while healthy sites stay at rest. This pattern generalises to the unseen stimulation sites.

We repeated the analysis with linear iSSM, and two established nonlinear latent variable models (LFADS and BRAID), which predict neural activity well but lack identifiability guarantees. None of the three control models showed any separation between healthy and epileptogenic tissue -- not even BRAID, which matches iSSM's rollout accuracy on this dataset (see App. K5, Fig. 16). This illustrates how good prediction performance does not guarantee scientifically interpretable latents.

Scope:

We note that this single-patient analysis is a proof-of-concept. It shows that iSSM's latents can capture clinically relevant structure that is missed by non-identifiable models. A particularly exciting future direction is applying iSSM to neural recordings during cognitive tasks to arbitrate between competing mechanistic hypotheses, as we illustrated in our decision-making simulations (Sec. 4.2, Fig. 3).

Please let us know if you have any questions about this.

AuthorsOpenReview ↗

Response to Reviewer EfXi: Q1

EfXi, Q1: Are hard interventions important empirically?

Many thanks for this question. Yes, the hard-intervention mask matters empirically. We added a new appendix section (App. K.4, Fig. 15) with the ablation analysis you requested.

On interventional data from the bistable well, we benchmarked iSSM against two other nonlinear SSM variants without hard-intervention mask (10 seeds each):

  • a standard nonlinear SSM, i.e. iSSM without the intervention mask;
  • an input-conditioned SSM, whose nonlinear transition takes the input as an argument, xt+1=g(xt,ut)+ϵtx_{t+1} = g(x_t, u_t) + \epsilon_t. This is strictly more expressive than both SSM and iSSM.
iSSMStandard SSMInput-conditioned SSM
Reconstruction>0.999>0.999 (n.s.)>0.999 (n.s.)
Rollout0.9810.726 (p=0.004p=0.004)0.934 (n.s.)
Latent accuracy0.9810.708 (p=0.002p=0.002)0.867 (p=0.002p=0.002)

(p-values: paired Wilcoxon signed-rank test against iSSM, FDR-corrected.)

All three models perform equally on reconstruction, but only iSSM reaches near-perfect latent accuracy. The standard SSM is also worse on rollout. The input-conditioned SSM matches iSSM on rollout, but its latent accuracy is significantly lower, showing that more expressivity does not guarantee latent identification.

We hope this answered your question. Let us know if not.

Dear Reviewer EfXi,

Thank you again for your review. As there are a few days left in the discussion period, we just wanted to check whether our response and revision addressed your concerns. If anything is still unclear, or if there are additional analyses you would find helpful, we would be happy to provide them in the remaining time.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision

We thank all reviewers for their careful reading and constructive feedback. We are glad that reviewers found the problem important and well motivated (BdGD, pGN7), the paper clearly written (EfXi, pGN7), and that they valued the theoretical guarantees (EfXi) and experimental validations (EfXi, pGN7).

We have addressed the reviewers' concerns in the revised manuscript, with additions highlighted in blue. The main changes are:

(A) In response to the reviews

  1. Weaker theoretical assumptions (BdGD-W4). We replaced Step 3 of our identifiability proof (Thm 3, App. A.4) with a new proof that no longer requires independent process noise or faithfulness. Its conditions are already implied by the assumptions of Step 2, so the new Step 3 adds no extra assumptions. We also show empirically that latent recovery is unaffected by correlated process noise (new App. J).

  2. Scientific value of identifiable latents (BdGD-W6, EfXi-W2). A new Results subsection (Sec. 4.5, Fig. 6) shows that in human iEEG data, only nonlinear iSSM's latent dynamics separate epileptogenic from healthy stimulation sites. Linear iSSM and two established nonlinear latent variable models without identifiability guarantees (LFADS, BRAID) show no such separation.

  3. Comparisons with non-identifiable models (BdGD-W7, EfXi-W1, EfXi-Q1). We compared iSSM with (i) LFADS and BRAID (new App. K); (ii) models trained without interventions (Fig. 2b); and (iii) nonlinear SSMs without hard interventions (App. K.4). In simulations with known ground truth, iSSM outperforms all of them on latent recovery, consistent with our theory.

  4. Accessibility (BdGD, pGN7). We added a Methods section on model fitting (Sec. 3.3), rewrote the description of the evaluation metrics (Sec. 3.4), defined key terms in the Introduction, added a Related Work paragraph on LFADS and BRAID, and added figure legends.

To stay within the 10-page limit, we slightly condensed wording in the main text; no results were removed.

(B) Independently of the reviews

While re-checking the proofs, we made the following changes. None of them changes the assumptions of our main-text results, nor the experiments or conclusions.

  • Step 1 (App. A.2): we expanded the proof to spell out the step from noiseless marginals to trajectory distributions. The assumptions are unchanged.
  • Prop. 5 (App. A.5): this appendix extension to nonlinear input maps assumes a regime with zero input to latents. We now require this for both the estimated and true input maps, not just the true one.
  • Appendix consistency: we made sure assumptions in the appendix are stated in the same words as in the main text to avoid confusion.

Detailed responses follow in each reviewer's thread. We are happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.