- 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 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
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:
- The encoder sees only the initial timesteps of the trial and infers the initial latent mean . On the real data, is the pre-stimulation window.
- We iterate through the learned dynamics model (Eq. 2) without noise, driving the model with the inputs ; i.e., . These are the learned latents we use for computing latent accuracy.
- Finally we decode the learned latents with to obtain the rollouts.
How reconstructions are generated:
- The encoder sees the entire trial and infers the latent posterior mean trajectory .
- We decode with 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.
- 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 are continuously differentiable, and is invertible at some point ;
- (B2) in every intervention regime, including the intervention-free one, the latent state has full support .
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).
- We also tested this empirically in a new appendix (App. J):
- Simulations. We simulated the bistable well such that the noise has correlation between the two latents with , and fit nonlinear iSSMs with a learned full covariance. Latent accuracy stays near-perfect at every (median , 10 seeds each; App. J.1, Fig. 13).
- Real data. We retrained the macaque () and human () 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 on both datasets), suggesting that a diagonal 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 ( in Sec. 3.1).
To answer your questions:
-
Is the mask fitted? -- No; the mask is derived from , where is learned and is observed.
-
How is the mask formed? -- The binary vector indicates which latents are not driven at time : we compute the input to each latent , and set where and 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 , and our theory guarantees that 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.
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 . -
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 , 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 ( vs. iSSM's , ). LFADS achieves moderate rollout accuracy () and poor latent recovery ().
-
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.
Response to Reviewer BdGD: W1, W2
We genuinely appreciate your detailed feedback, which has substantially improved the paper. In summary, we:
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:
Prior. Each posterior sample xt is passed one step through the transition model (Eq. 2), giving the Gaussian transition prior N(mt⊙gϕ(xt)+But,Q). Targeted latent means are again clamped to the input, mirroring the encoder and implementing the same assumptions (Gaussian latents and do-interventions).
Likelihood. Posterior samples are decoded by fθ.
Sparsity. Finally, we enforce sparsity on B by thresholding small entries during training, since a dense B 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:
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.