acceptodds
Under review as a conference paper at ICLR 2027

Inferring Neural Dynamics with Variational Schrödinger Bridges

Abstract

Neural population activity evolves within low-dimensional dynamical spaces, yet recovering these dynamics from electrophysiological recordings remains difficult. This is particularly challenging for spontaneous neural activity, where there are no repeated trials or known inputs, and where the recorded network continuously transitions between internally generated and externally driven states. Here, we introduce NeuBridge, a variational extension of Schrödinger bridges to infer stochastic latent dynamics from neural population activity. We apply NeuBridge to large-scale Neuropixels recordings from the brainstem of sleeping mice, where neural activity evolves spontaneously across sleep-wake states and is occasionally perturbed by optogenetic stimulation of upstream populations. NeuBridge reconstructs firing rates more accurately than existing latent-dynamics methods and recovers latent dynamical structure across mice. Importantly, NeuBridge also estimates the posterior control cost required to move neural activity away from its intrinsic dynamics. Without being given the timing of laser stimulation, NeuBridge detects a clear increase in control cost during optogenetic perturbations. These results show that NeuBridge can recover low-dimensional stochastic dynamics from spontaneous neural activity and can quantify the external control effort required to explain perturbation-driven deviations from intrinsic dynamics.

Primary area:
applications to neuroscience & cognitive science
Keywords:
neural population dynamics, computational neuroscience, Schrödinger bridges, variational inference, interpretability

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_TzEaOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The neuroscience community is increasingly interested in building latent dynamics models of neural population activity to analyze how the brain transforms sensory information and state into behavior and decisions. Such models usually assume the structure of an experiment, with trials that have known start and end times. However, the field is moving towards less structured paradigms, such as naturalistic behavior without concrete task boundaries and sleep, which violate these assumptions. Most existing techniques also fail to account for unobserved neural populations and external inputs, unless explicitly provided to the model, which is not always possible or may hurt interpretability. This paper develops Neubridge, extending the Schrödinger Bridge formulation to learn latent stochastic dynamics and inferred inputs from unobserved populations, with a focus on recordings that lack explicit task boundaries. The authors show promising model fits to experimental data recorded during sleep, with optogenetic stimulation that lead to switching the phase of sleep. The authors compare to several existing baseline methods and show that Neubridge outperforms the other methods in terms of model fits and flow field inference. It can also predict the correct timing of photostimulation perturbations without being explicitly trained to do so. Altogether, this model appears to combine the concept of inferred inputs from unobserved populations and external stimulation with flow field inference. While both components exist in the literature, this paper is the first to combine them and demonstrate advantages on real experimental data.

Critical strengths

Their model, Neubridge, is validated on real experimental data and compared to a number of popular alternatives. Neubridge demonstrates better model fits and more interpretable flow fields than these competing models. The inferred inputs also show the ability to infer experimentally-relevant information, such as optogenetic simulation perturbations upstream of the recorded populations.

Critical weaknesses
  1. Inferred inputs from unobserved populations is a known and important features of LFADS. While the authors compare Neubridge to LFADS on reconstruction, the inferred input aspect of LFADS is ignored. Since this is considered a core motivation for the paper, this feature of both models should be compared. Moreover, the paper would be further strengthed by a synthetic dataset that included unobserved inputs beyond the noise introduced in the Lorenz attractor example. I would be very interested to see how well the inferred inputs correspond to the true external stimuli when known.

  2. I found it unclear what was specific about the design decisions in Neubridge to the stated problem of training latent dynamics models without explicit trial structure. I can understand how the initial state prior, which was formulated in a way that conditions only on a window of history, and inferred inputs can help with this. But these are both standard practice in LFADS as well, so not clearly new.

Questions for authors
  1. How do the inferred inputs of Neubridge and LFADS compare?
  2. What about the design decisions in Neubridge help with the stated problem of training on data without explicit experimental trial structure?
1 comment

We thank the reviewer for your constructive comments!

W1 Response: (Inferred inputs from unobserved populations is a known and important features of LFADS..., ...the paper would be further strengthed by a synthetic dataset that included unobserved inputs...)

We thank the reviewer for the suggestions. We have now included unobserved inputs to the synthetic dataset and show that NeuBridge can recover these inputs. We added two examples visualizing the inferred corrections of the autonomous process to account for the unobserved inputs (Figure 2e).

In addition, we also applied LFADS to the optogenetic dataset. Similar to NeuBridge, LFADS also shows an increase in inferred input for the optogenetic dataset (Supplementary Figure S2). But the key additional advantage of NeuBridge is that the correction is defined directly relative to the autonomous reference dynamics within the same latent space. Thus, the effect of an inferred input can be visualized as an observation-informed deviation (QQ) from the autonomous trajectory RR, allowing us to localize where in the latent landscape additional control is required and how that control reshapes the autonomous dynamics (see Figure 4f as an example). In LFADS, by contrast, the inferred input is fed into the recurrent generator and therefore is not directly represented as a correction vector within the latent factor space. We have clarified this distinction in the revised manuscript.

W2 & Q2 Response: (What about the design decisions in Neubridge help with the stated problem of training on data without explicit experimental trial structure?)

We agree that conditioning the initial state on preceding history and inferring observation-dependent inputs are not individually unique to NeuBridge and are also used in LFADS. The design choice that is specific to the absence of trial structure is instead the segment-based Schrödinger bridge formulation. NeuBridge can sample arbitrary segments from a continuous recording, infer their start and end latent distributions, and learn the dynamics connecting these boundary distributions. By repeatedly sampling segments that span different time intervals and different locations in latent space, the same reference process RR is trained across the broader latent landscape, without requiring trial-aligned initial conditions or repeated trajectories.

In addition, NeuBridge models the diffusion term as a function of latent state, allowing the level of stochasticity to vary across the latent landscape. This is particularly relevant for spontaneous sleep recordings, where the variability of neural population activity is likely to differ across brain states and transitions.

Q1 Response: (How do the inferred inputs of Neubridge and LFADS compare?)

In LFADS, the inferred input is fed into the generator at each step, thereby modifying the generator state and, subsequently, the low-dimensional latent factors. The inferred input for LFADS is thus not represented in the same shared latent space as the internal dynamics. In NeuBridge, the correction is defined directly relative to the autonomous reference dynamics (RR) within the same latent space. Thus, the effect of an inferred input can be visualized as an observation-informed deviation (QQ) from the autonomous trajectory RR, allowing us to localize where in the latent landscape additional control is required and how that control reshapes the autonomous dynamics.

Reviewer_mSR9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper uses variational schrodinger bridges to perform inference in SDE models of neural population dynamics. This inference approach provides not only latent trajectories, but also control costs that can be interpreted as external inputs to the system. The method is validated first in synthetic data, and shown to accurately predict future trajectories in the Lorenz system. The authors then use the method to analyze neuropixels data from the mouse brainstem during sleep and optogenetic perturbations.

Critical strengths

The paper is clearly written, and the method nicely explained.

The model can be fit across multiple recordings with different populations of neurons, which makes it a powerful alternative to models like SLDS whose standard implementation can only use single populations.

The inference of a control signal is unique and very powerful, not only for inferring inputs to an area, but for gaining deeper insight into how control (or lack thereof) impacts neural dynamics in awake and behaving animals. I think this is a very interesting direction for future exploration.

Critical weaknesses

The introduction and methods is completely without citations.

There are not enough details about the neural data: how many animals in each group? how many neurons per recording session (or their mean)? how long are the recordings?

Minor: fontsize is too small in figures

Questions for authors

Is there a reason a linear projection from latent space to observations was chosen?

Approximately how much data is needed for training to converge on a useful result?

How long does model training take? How long does inference take?

The model is capable of training across multiple sessions, but the authors do not explore this (relatively unusual) aspect of their model. How does performance increase as the number of training sessions increases? Are there obvious individual or group differences in the dynamics? Is performance better if per-group models are trained?

In either the synthetic data or real data (or both), I'd be curious to see how well the method performs when trained on a set of populations, then tested on a new population by simply training a new linear observation head and leaving remaining parameters fixed. Does this work? How does it compare to including that population in the initial training? If adaptation to new populations is fast this could be an interesting approach to applications like brain-computer interfaces, for example.

2 comments

We thank the reviewer for your constructive review!

W1 Response: (The introduction and methods is completely without citations.)

We apologize for this omission. We have now included all the required citations and improved our discussion of related work.

W2 Response: (There are not enough details about the neural data...)

We now included this important information in the Supplemental information. Briefly, there are 5 mice in the WT (wildtype) group, 4 mice in the eYFP group, 6 mice in the Opto (dmM-stimulation) group, each providing one recording session. Recorded units used per session range from 89 to 237 neurons, with an average of 161 neurons. Each recording session lasts between 2 to 3 hours.

W3 Response: (Minor: fontsize is too small in figures.)

We thank the reviewer for this suggestion and have now enlarged font sizes for clarity.

Q1 Response: (Is there a reason a linear projection from latent space to observations was chosen?)

We chose a linear projection from latent space to the observed firing rates to encourage a direct and interpretable relationship between the inferred latent dynamics and the recorded neural activity. A more flexible nonlinear decoder could introduce additional nonlinear transformations between latent and observation spaces, potentially making the learned dynamical landscape harder to interpret.

Q2 Response: (Approximately how much data is needed for training to converge on a useful result?)

As shown in Supplementary Figure S3a, training on just one single recording (DL159) with about 3000 2.52.5-sec timebins, NeuBridge achieved R2R^2 values of 0.3590.359 and 0.5150.515 for R and Q, respectively. Training on more recordings further increased performance, reaching R2R^2 values of 0.4230.423 and 0.6820.682 for R and Q using all 10 baseline recordings.

Q3 Response: (How long does model training take? How long does inference take?)

On a single Nvidia RTX 4090 GPU, the NeuBridge (8-dim latent) training time for 6-, 12-, 20-prediction-step tasks across 10 recordings for 200 epochs, is about 79 minutes, 114 minutes, and 146 minutes respectively. Inference time across approx. 39000 overlapping windows for 6-, 12-, 20-prediction-step tasks is about 3.0 seconds, 4.5 seconds, and 6.6 seconds. These estimates used a batchsize of 256.

Q4 Response: (The model is capable of training across multiple sessions...)

We have now explored these questions in Supplementary Figure S3C. We trained NeuBridge on one group of mice, and then tested on another held-out group of mice. First, performance generally improved as the number of training sessions increased. Second, models trained on mice from the same experimental group did not consistently outperform models trained across groups. This comparable cross-group performance (particularly for the autonomous forecast of process RR) suggests that NeuBridge captures a shared dynamical structure across WT, eYFP, and Opto mice, rather than strongly group-specific dynamics.

Q5 Response: (...how well the method performs when trained on a set of populations, then tested on a new population...)

We explored this interesting question in Supplementary Figure S3B. We trained NeuBridge on half of the neurons of 10 recordings (together as one dataset), and tested on the other half, only tuning the linear decoder and the neuron-specific embedding. It worked well, with average R2R^2 values of 0.362±0.0900.362 \pm 0.090 for RR and 0.660±0.0860.660 \pm 0.086 for QQ. This represents only a small decrease from the baseline performance of 0.4230.423 and 0.6820.682 for RR and QQ, respectively.

I thank the authors for addressing all of my comments. I will maintain my current score.

Reviewer_4VrsOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper develops NeuBridge, which simultaneously models low-dimensional latent dynamics and an unknown control signal (inputs), demonstrates how it can recover Lorenz dynamics in simulated data, and in a real dataset, demonstrates improved modeling accuracy versus a few alternative methods, plus the recovery of learned inputs at expected times.

Critical strengths

Overall, I think this is a nice addition to the neuroscience dynamical systems modeling toolbox - combining past ideas of 1) low-D dynamical systems modeling with 2) inferring inputs seen in other models.

Critical weaknesses

-The discussion of the relationship to past literature is limited. Only a few select methods (which the authors compare with) are mentioned. There is other past work that infers input signals, such as ILQR-VAE - (ICLR 2022 - which should probably be compared with, assuming there is usable code), or MR-LFADS (ICLR 2025), or BAND (kudryoshova et al, biorxiv 2025). There is also substantial past work other than FINDR looking for low-D dynamics (e.g. Genkin… Engel, Nature, 2025).

-In the synthetic data, there is not a validation of finding a meaningful control signal, which is a selling point of the method relative to FINDR.

-It’s currently unclear whether there is a comparison to another method that infers inputs (I did not see it mentioned whether LFADS is fit with inputs. Relatedly, it’s unclear whether the authors compared with the more recent autoLFADS).

-It feels like the advantage of this approach of finding inputs in the lowD space is being to visualize how they are impacting the flow-field, but that is not shown.

Questions for authors

-Conceptually, are there differences between this paper’s approach and FINDR, other than the control signal?

-Does this method assume that all signals are Gaussian (as opposed to Poisson)?

6 comments

We thank the reviewer for your constructive comments!

W1 Response: (The discussion of the relationship to past literature is limited...)

We apologize for this omission. As suggested by the reviewer, we have substantially improved the discussion of related work in the revised manuscript (see the “Related works” section revised manuscript). We now discuss the mentioned methods more explicitly in relation to NeuBridge and clarify the specific distinctions in model structure and interpretation.

W2 Response: (In the synthetic data, there is not a validation of finding a meaningful control signal...)

We thank the reviewer for this suggestion. We have now included external input to the synthetic data example and demonstrate that NeuBridge can infer the external input (Figure 2e). For illustration, we included an example where the external perturbations switched the trajectory between the two lobes of the Lorentz attractor, and NeuBridge inferred the correction input to also switch the QQ trajectory.

W3 Response: (...whether there is a comparison to another method that infers inputs...)

We agree that this comparison is important. We have now applied LFADS with inferred inputs to the optogenetic dataset (Supplementary Figure S2). LFADS also shows a similar increase in the inferred input for the optogenetic dataset. The key additional advantage of NeuBridge is that the correction is defined directly within the same latent space as the autonomous reference dynamics. Thus, the effect of an inferred input can be visualized as an observation-informed deviation (QQ) from the autonomous trajectory RR, allowing us to localize where in the latent landscape additional control is required and how that control reshapes the autonomous dynamics. In LFADS, by contrast, the inferred input is fed into the recurrent generator and therefore is not directly represented as a correction vector within the latent factor space. We have clarified this distinction in the revised manuscript and added examples visualizing the required deviations from RR in the latent space for both the synthetic and optogenetic datasets (Figures 2e, 4f). Our comparison used the original LFADS formulation with inferred inputs (Pandarinath et al., 2018).

W4 Response: (...to visualize how they are impacting the flow-field...)

We thank the reviewer for the suggestion, and this is correct. Because both the autonomous reference process RR and the observation-corrected process QQ evolve within the same latent space, the correction of RR can be directly visualized within that latent space. This differs from LFADS, where the inferred correction signal is provided as an input to the recurrent generator rather than acting directly within the latent factor space. To better illustrate this point, we added examples that explicitly visualize inferred inputs in the latent space. For the synthetic data, we now show cases in which an external input drives the system away from its initial trajectory in the Lorenz attractors, and NeuBridge inferred the external input needed to correct the autonomous trajectory RR (Figure 2e). For the optogenetics dataset, we show for a laser-triggered NREM-to-REM sleep transition how the observation-informed QQ trajectory is corrected to move away from the reference process RR toward the REM region in the latent space (Figure 4f).

Pandarinath, Chethan, et al. "Inferring single-trial neural population dynamics using sequential auto-encoders." Nature methods 15.10 (2018): 805-815.

Q1 Response: (Conceptually, are there differences between this paper’s approach and FINDR, other than the control signal?)

Besides the inference of control signals, NeuBridge differs from FINDR in three important aspects:

First, NeuBridge uses an extended Schrödinger bridge approach to infer latent trajectories. NeuBridge can sample arbitrary segments from a continuous recording, infer their start and end latent distributions, and learn the dynamics connecting these boundary distributions. By repeatedly sampling segments that span different time intervals and different locations in latent space, the same reference process (RR) is trained across the broader latent landscape. Thus, NeuBridge does not require trial-aligned initial conditions or repeated trajectories and is suited for both trial-based and spontaneous recordings.

Second, NeuBridge learns state-space-dependent diffusion, whereas FINDR assumes constant diffusion throughout the latent space. This allows NeuBridge to capture how the intrinsic variability of neural activity changes across different regions of the dynamical landscape.

Third, one NeuBridge model can be used across recordings of different input-channel numbers and inferred shared dynamics, since NeuBridge has a cross-attention based input encoder, whereas one FINDR model cannot be directly shared across recordings.

Q2 Response: (Does this method assume that all signals are Gaussian (as opposed to Poisson)?)

NeuBridge assumes Gaussian noise for the latent transition dynamics, which is independent of the observation model. For our hours-long sleep recordings, we used 2.52.5-s bins to capture the relatively slow dynamics of sleep-state transitions and modeled the resulting firing rates using a Gaussian observation likelihood. However, the NeuBridge framework is not restricted to Gaussian observations and could also incorporate a Poisson likelihood for modeling spike counts directly, without changing the underlying stochastic dynamical framework.

Thank you for the updates that have strengthened the paper - I plan to raise my score to a 3.

One small followup question - could you further clarify the second point above about the difference with FINDR? Do you mean that the variance of the stochastic process is not fixed in NeuBridge but is fixed in FINDR, or something else?

Comments on the paper

4 comments
~Dimitra_Maoutsa1OpenReview ↗

Before the rebuttal period finishes, I have a brief clarification question for the authors, that, as far as I can see, has not been properly addressed by the reviewers.

One of the aspects I find most interesting in this framework is the proposed separation between intrinsic dynamics and external inputs. However, I am having some difficulty understanding how the framework distinguishes corrections to the drift arising from conditioning on the observations from those that are attributed to external inputs. I think this is the most interesting contribution of the paper, and I would appreciate some clarification on how the authors establish within the proposed formulation, because this is not sufficiently documented in the updated manuscript.

To explain my confusion, even ordinary posterior smoothing introduces a correction to the drift, even when there is no external input, and the underlying dynamics are properly specified [1,2], i.e. even when we know the exact correct drift. Using the notation of the paper, the exact smoothed process associated with the SDE in Eq. (1) has, between observation times, drift fθ(z)+σθ(z)σθ(z)⊤∇zlog⁡ht(z)f_\theta(z) + \sigma_\theta(z)\sigma_\theta(z)^\top \nabla_z\log h_t(z) where the second term arises from conditioning on observations [2,3], and where ht(z)h_t(z) is the likelihood of the remaining observations given the current state zz, under the reference process. This indicates that a drift correction can arise only from conditioning on the data, without any external input. Moreover, since the autonomous dynamics are unknown and the underlying drift fθf_\theta is itself learned, discrepancies between the learned drift and the true underlying dynamics can also be absorbed into the inferred correction to the drift, i.e. having an additional additive term in the expression above that accounts for the drift mismatch.

Therefore it is not yet clear to me how the authors disambiguate between these different contributions, and therefore I would be grateful if they could explain how they disentangle among these when interpreting the correction vϕv_\phi as an external input? Is there a particular assumption, constraint, or property of the objective that makes this separation possible? If yes this should be made more explicit how Eq. (6) supports this interpretation.

As said above, the ability to separate intrinsic dynamics from external inputs is potentially the most interesting contribution of the paper, especially compared to FINDR. However, despite having some background in inference for SDEs, I have not been able to identify precisely how this distinction follows from the current formulation, and therefore, I think, spelling this out would substantially strengthen the paper and help readers appreciate what the method offers beyond standard posterior inference.

Also I would think this paper should be also cited [4].


[1] Archambeau, C., Opper, M., Shen, Y., Cornford, D., & Shawe-Taylor, J. (2007). Variational inference for diffusion processes. Advances in neural information processing systems, 20.

[2] Opper, M. (2019). Variational inference for stochastic differential equations. Annalen der Physik, 531(3), 1800233.

[3] Maoutsa, D., & Opper, M. (2022). Deterministic particle flows for constraining stochastic nonlinear systems. Physical Review Research, 4(4), 043035.

[4] Duncker, L., Bohner, G., Boussard, J., & Sahani, M. (2019, May). Learning interpretable continuous-time models of latent stochastic dynamical systems. In International conference on machine learning (pp. 1726-1734). PMLR.

We are replying here to the initial comment before the commenter's second edit. We thank the commenter for raising these important points!

The comment addresses these two questions:

1. How to distinguish drift corrections arising from conditioning on observations from those caused by external inputs?

We understand this question as, how do we distinguish the contribution of noise diffusion from the contribution of external disturbances, when inferring drift corrections from an individual trial trajectory. We agree this is particularly challenging. As an example, even in the case when the model has learned the true intrinsic dynamics fθ(z)f_\theta(z), if we want to reconstruct one particular trajectory, the actual trajectory will contain stochastic fluctuations. Thus, when we infer an observation-conditioned QQ trajectory, it will include a correction vϕv_\phi to make the trajectory more consistent with the observation. Therefore, the correction vϕv_\phi can be non-zero, even when fθ(z)f_\theta(z) and σ(z)\sigma(z) are perfectly known and no external input exists.

We acknowledge that the inferred correction vϕv_\phi includes both the true external disturbance and the diffusion noise. However, NeuBridge learns the state-dependent noise diffusion term σ(z)\sigma(z), which quantifies the intrinsic stochastic variability at each location in latent space. Therefore, the inferred correction vϕv_\phi can provide an estimate of potential external inputs, while the state-dependent diffusion term helps assess the magnitude of this correction relative to noise.

Moreover, applying NeuBridge to the optogenetic dataset shows that our framework can detect and quantify deviations from the learned reference dynamics. The quantity we use to detect potential external perturbations is not vϕv_\phi itself, but the diffusion-weighted KL rate (Eq. 9; Fig. 4c-e). Corrections in regions with high intrinsic variability incur relatively little cost, whereas deviations in regions with low intrinsic variability have higher KL rate. Thus, an elevated KL cost indicates that the observation-conditioned dynamics deviate substantially from what is expected under the learned autonomous stochastic process, taking its intrinsic variability into account. Our optogenetic results provide encouraging support for this approach: without information about stimulation timing, NeuBridge detected increased KL rate during optogenetic perturbations (Fig. 4c-d). These findings suggest that the diffusion-weighted KL rate can reveal perturbation-associated changes in neural population dynamics. Furthermore, the examples in Figs. 2e and 4f illustrate how the inferred corrections (vϕv_\phi) modify the autonomous dynamics of RR. Specifically, Fig. 4f shows how these corrections shift the inferred neural population trajectory away from the autonomous reference trajectory toward the REM region during an optogenetically induced NREM-to-REM sleep transition.

We have incorporated these clarifications into the final section “Discussion and conclusions” of our manuscript, and hope this will aid the readers’ understanding. We also updated the manuscript to make sure that our claims are accurate.

2. How does the model distinguish between correcting for individual trajectories and compensating for imperfectly learned shared dynamics?

We are using the Schrödinger bridge framework in our task setting to specifically address this question. This framework seeks the minimal observation-conditioned correction to the reference process, using the KL loss term from Eq. (9). This training objective encourages the model to minimize any correction that can be already explained by the reference process dynamics RR. Therefore, if there are discrepancies between the learned drift and the true underlying dynamics during the training process, our model is specifically designed to fix that gap by optimizing the learned drift (the reference RR), and to avoid to reduce those discrepancies with an observation-conditioned inferred correction (the inference QQ).

Thank you for pointing out this relevant work. We have now included Duncker et al. (2019) in the revised Related Works section.

~Dimitra_Maoutsa1OpenReview ↗

(The update of the comment above introduced only an additional citation for a paper proposing a method similar to yours that I think should be cited - I didn't change the rest of the comment)

Thank you very much for the response and for acknowledging that vϕv_\phi can be non-zero even when the autonomous dynamics are known in absence of inputs. As said before, I find the proposed method interesting, but I would still appreciate some clarification on how the inferred drift correction should be interpreted.

To clarify here, given that we are approaching the end of the rebuttal period, I am not asking for any additional experiments. My suggestions concern how the existing results are presented and interpreted in the manuscript.

Even if the diffusion is known, conditioning an autonomous process on observations introduces a drift correction, which takes the form D(z)∇zlog⁡ht(z)D(z)\nabla_z\log h_t(z) between observations. As I understand the noise covariance D(z)D(z) describes the expected size of fluctuations, but does not capture how much correction is needed to explain a specific trajectory. Is this also how you interpret the role of diffusion here?

The Lorenz example helps to understand the method's ability to reconstruct inputs, but I am not sure it addresses this particular point. The simulation uses relatively small dynamical noise (σdiff=0.01\sigma_{\mathrm{diff}}=0.01 while dynamic range is [-1,1]), so it cannot show whether the method distinguishes an external input from fluctuations, because the setting is nearly deterministic. On top of this, identifying the diffusion introduces another source of uncertainty sicne you can only characterise the expected fluctuations reliably to the extent the diffusion is correctly recovered.

Systematic experiments on a non-chaotic stochastic system with noise on the order of the dynamics with known drift, diffusion, and external inputs would show how reliably you can recover each component. I understand this is not feasible now, but stating what the current experiments establish would help readers judge and trust the interpretation of the neural recordings, where all three are unknown


On the KL argument, the inverse-covariance weighting in Eq. (9) follows from Girsanov for processes with same diffusion and different drifts, fθf_\theta and fθ+vϕf_\theta+v_\phi. This weighting is already part of usual variational smoothing and accounts for differences in the scale of fluctuations across dimensions. It quantifies the size of the correction, but it does not determine whether that correction comes from conditioning on a specific noise realisation or from an external input.

I would further appreciate some explanation on your statement that the Schrödinger-bridge framework “specifically” addresses mismatch between the learned and true drift. As I understand it, classical Schrödinger bridges minimise KL relative to a given reference process under endpoint distributional constraints, while penalising deviations from a reference dynamics is also standard in variational smoothing settings. Here,you learn the reference dynamics. I understand how the KL penalty encourages the learned drift to capture recurring patterns, but it is not so clear to me on whether the bridge construction provides a way to distinguish intrinsic dynamics from external inputs esp. for states visited only under perturbations. Can you clarify what the bridge construction contributes to this distinction that is different from standard variational smoothing?


More generally, a deviation from the learned drift can arise from observation conditioning, an external input, or inaccuracies in the reference model (due to finite observation time, limited temporal resolution of observations due to binning). Moreover if the latent dimensionality is not properly estimated and leaves out autonomous variables, their effects could potentially be absorbed into the correction. On this point, the experiments use an 8-D latent space, but could not find an explanation of how this latent dimensionality was chosen. A short clarification would be helpful. Also how the Gaussian observation process parameters are learned.


Finally, I think the wording in the Discussion (lines 527–528), describing deviations from the learned reference as indicators of potential perturbations, is the distinction I was pointing to. Can the authors make this equally clear in the abstract, the discussion of Figure 4 (lines 501–511), and the conclusions?

I think these clarifications would help readers to better understand and appreciate the contribution of this work. I find the connection to the optogenetic perturbations particularly interesting in itself, and I think it would be useful to further clarify what this result establishes and what remains an interpretation, particularly regarding the identification of the inferred correction with a external inputs. Thank you again very much for taking the time to to address this question.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.