acceptodds
Under review as a conference paper at ICLR 2027

Cell-Class Neural Ordinary Differential Equation Models for Parsing Biologically-Constrained Contributions to Neural Dynamics

Abstract

Understanding how populations of individual neurons interact to shape the overall dynamics of neural activity is a central question in computational and systems neuroscience. Recent work has shown that neural ordinary differential equation (NODE) models are able to model neural activity dynamics with high accuracy and interpretability of the underlying dynamics. However, existing NODE models treat all neurons as part of a homogeneous group, preventing understanding of how underlying neural populations (e.g., excitatory and inhibitory cell classes) contribute to the overall dynamics. Here, we introduce Excitatory-Inhibitory NODE (EI-NODE) models. These models A) decompose the overall dynamics into components specific to each population, allowing understanding of each population's interactions with one another; and B) provide biological constraints on the contributions of excitatory and inhibitory populations towards the dynamics, using a variant of monotonic neural networks. Using both synthetic data and recorded neural activity data during a naturalistic climbing task with optogenetic perturbations, we show that EI-NODE models can provide equivalent or greater accuracy in dynamics modeling compared to alternative models, while enabling a newfound biologically constrained understanding of neural populations' interactions and roles in the underlying dynamics.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Neuroscience, Dynamics, Machine Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_hLQnOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes EI-NODE, a NODE-based sequential autoencoder for multi-population neural recordings. The latent state is split into blocks, one per recorded population. Each block’s time derivative is an additive sum of a diagonal linear self-term and one learned interaction function per source population. Every cross-population influence has a fixed sign, in the spirit of Dale's law. A block-masked, nonnegative linear readout ties each latent block to its labeled neurons. Because the derivative is a sum of per-source terms, it can be decomposed into population "influences" over time. The validation considers four examples (synthetic winner-take-all circuit, cell-type inference, mouse data, and CURBD surrogate) where the proposed method is compared against linear and RNN variants.

Critical strengths

Additive, sign-constrained decomposition gives a readable account of population interactions. Writing each population’s derivative as a sum of per-source terms lets the model attribute dynamics to specific source populations at each moment.

Sensible use of monotonic networks.

Validation against ground-truth currents in the synthetic setting. Comparing learned influences with actual synaptic currents goes beyond reconstruction accuracy. The reported correlations are high for all nine source–target pairs.

Handling of cell-class-specific perturbations. Delivering optogenetic input to the targeted inhibitory block is a sensible use of the population-partitioned latent space.

Critical weaknesses

Limited novelty: a recombination of existing components, where each ingredient exists already in the literature. Replacing linear cross-population maps with sign-wrapped monotonic MLPs is a reasonable engineering choice, but it isn’t a substantial methodological advance.

Weak grounding in neurophysiology and in the population-modeling literature. Beyond the E/I distinction and Dale’s law, the model has no physiological content: It doesn’t say what level of description its latent blocks represent (micro-circuit, mesoscale population, or region). It doesn’t consider hierarchy, laminar structure, or interneuron diversity.

Closest prior work in computational neuroscience is not cited. Wilson–Cowan population equations, neural mass models, and dynamic causal modeling (DCM) all describe E/I populations with first-order ODEs, learned time constants, and sign-fixed, monotone (sigmoidal) couplings. Several of them are fit to data. EI-NODE is structurally a Wilson–Cowan-type system with learned, flexible coupling functions, but the paper never positions itself against this literature.

Odd framing of the motivation. The introduction presents dynamical-systems explanations as something “invited” by recent large-scale recordings. Dynamical systems have been central to computational neuroscience since its inception: Hodgkin–Huxley, Wilson–Cowan, Amari, Hopfield networks, and balanced E/I networks. This framing is historically inaccurate, and it is likely why the most relevant E/I population-dynamics literature is missing from the related work.

Questions for authors

Identifiability. How do you resolve the offset and scale ambiguities in the influence decomposition, i.e., constants trading off between E, I, and self-terms, and scaling between latents and the readout?

Fixed points and stability. Were fixed points computed in the full latent space rather than read off 2D slices at mean I activity? What are their stability properties, and do they match the ground-truth attractors of the Wang circuit?

Relation to population models. How does EI-NODE relate to Wilson–Cowan, neural mass, and DCM models? What does it offer beyond fitting such a model with flexible monotone coupling functions? Why sum per-source nonlinear terms instead of applying a nonlinearity to summed input, and how does this choice affect what interactions the model can represent?

Unconstrained baseline on real data. Could an unconstrained NODE with an input channel be fit to the climbing/optogenetics data to quantify the accuracy cost of the E/I constraints? Given that Linear EI-NODE matches EI-NODE there, what evidence shows the nonlinear interactions are needed for real recordings?

8 comments
AuthorsOpenReview ↗

Response to Reviewer hLQn (1/3)

We thank the reviewer for their detailed and thoughtful review. We are pleased the reviewer highlighted the interpretability of our additive, sign-constrained decomposition as a key strength. Additionally, we appreciate their positive evaluation of our monotonic network formulation, our validation against ground-truth synaptic currents, and our approach to incorporating cell-class-specific optogenetic perturbations within the population-partitioned latent space. Below are our responses to the reviewer's comments and questions:

Relation to Wilson–Cowan, neural mass models, and DCM: We thank the reviewer for this point; we agree the paper should position EI-NODE relative to the mean-field literature, and we have added a dedicated related-work subsection and a new baseline.

Similarities with Wilson–Cowan, neural mass models, and DCM. EI-NODE's latent dynamics (Eq. 4 L145) are structurally a current-based Wilson–Cowan system, τdxdt=−x+∑Wf(x)\tau \frac{dx}{dt}=-x+\sum Wf(x) (Wilson et al. 1972; Supplementary Eq. 15), which is equivalent to the more familiar rate-based form (Miller & Fumarola, 2012). The softplus-wrapped monotone MLPs, fIE=softplus(f~IE)f^{IE} = \mathrm{softplus}(\tilde{f}^{IE}) and fEI=−softplus(f~EI)f^{EI} = -\mathrm{softplus}(\tilde{f}^{EI}), play the role of the classical sign-fixed, monotone gain function. EI-NODE differs from that literature in four ways:

  1. The gain functions are learned nonparametrically, separately for each source–target pair, instead of being fixed to a parametric sigmoid;
  2. Each population is represented by a multi-dimensional latent (Dj≥1D_j \geq 1), rather than a single scalar rate or mean potential; one latent will not work for more complex data in practice;
  3. Population states are inferred from single-unit spiking via an amortized encoder and a cell-type-masked Poisson readout, so latents are tied to identified E and I units (waveform, opto-tagging, or ground-truth identity in simulation). However, mean-field models are typically fit to aggregate signals where population identity cannot be checked.
  4. All parameters are fit end-to-end by gradient descent rather than by least squares or variational Laplace over a small, physiologically parameterized set.

Related work section. We have added a prior work section that details the related mean-field models and how this work builds on classical methods (see Section 2.5 L203).

Existing baselines already relate to the classical models. Our cell-type-constrained linear baseline (Jha et al., 2024) corresponds to a linearized E/I population model, comparable to the neuronal model of two-state DCM (Marreiros et al., 2008). Our EI-RNN (Song et al., 2016) is a high-dimensional Dale's-law rate network with a tanh.

New mean-field model baselines. To test the specific question the reviewer raises, we add a current-form Wilson–Cowan baseline (WC) and several variants. It is identical to EI-NODE (same encoder, masked readout, training and model selection), except that each f~θjk\tilde{f}_{\theta}^{jk} is replaced by a parametric sigmoid with a sign-fixed scalar weight. This isolates the contribution of flexible coupling functions. We find that this WC model fits the perturbation data much worse than EI-NODE, and fits the climbing data slightly worse (see Supplementary Table 7).

Per-source nonlinearity vs. nonlinearity on summed input: We trained variations of EI-NODE and Wilson–Cowan (WC) models to determine the effect of per-source nonlinear terms relative to a nonlinearity applied to summed input. We found that, for EI-NODE's learned coupling functions, the summed form fits consistently worse (see Supplementary Figure 7, Supplementary Table 7). It also loses the exact split of each population's input by source. Our per-source design therefore improves performance and allows interpretable influence analysis.

Supplementary Table 7. Held-out (validation) single-trial pseudo-R2{R}^{2} on climbing and optogenetic perturbation trials, averaged over two sessions from two mice. Mean ±\pm s.e.m. over 12 runs (3 data splits ×\times 4 training seeds). All models: 10 E / 5 I latents per region, matched training.

ModelClimb pseudo-R2{R}^{2}Perturbation pseudo-R2{R}^{2}
EI-NODE (per-source sum)0.053 ±\pm 0.0030.157 ±\pm 0.005
EI-NODE (summed input, then nonlinearity)0.045 ±\pm 0.0020.103 ±\pm 0.007
WC, current form (per-source sigmoid)0.042 ±\pm 0.0020.066 ±\pm 0.002
WC, activity form (input inside nonlinearity)0.043 ±\pm 0.0030.075 ±\pm 0.003
WC, activity form (input outside nonlinearity)0.041 ±\pm 0.0030.079 ±\pm 0.005
AuthorsOpenReview ↗

Response to Reviewer hLQn (2/3)

Framing of the motivation: Dynamical systems have been central to computational neuroscience from its foundations: the Hodgkin–Huxley model of the action potential, WC and Amari's population and neural-field models, Hopfield's attractor networks, and balanced E/I network theory. We did not intend to suggest that dynamical explanations originate with large-scale recordings, and we have rewritten the introduction accordingly.

The revised motivation makes a narrower claim we believe is accurate. Classical E/I population models were mostly specified from theory and evaluated against aggregate signals or a few summary statistics. In those settings, the cell-type identity of a population and the form of its interactions could not be checked directly. Large-scale recordings with putative or optogenetically identified cell types now make it possible to infer E/I population dynamics and interaction functions directly from single-unit activities. EI-NODE is designed for this data-driven setting. It keeps the core structural assumptions of the classical models (cell-type-specific populations, sign-fixed and monotone interactions, leaky first-order dynamics) and learns the interaction functions from data.

We also agree this framing contributed to the omission the reviewer identified. The revised introduction now motivates EI-NODE from the E/I population-dynamics tradition, and the new related-work subsection (see point 3) positions it against Wilson–Cowan, neural mass models and DCM.

Identifiability: Identifiability is a known issue (Prinz et al., 2004; Das & Fiete, 2020) that we mitigate with optogenetic perturbations, which shrink the solution set (Basrai et al., 2026; Nejatbakhsh & Wang, 2025). A constant can be shifted between fjE, fjI, and fjj without changing the vector field; all quantities we interpret (coupling signs and shapes, fixed points, winner-take-all structure) depend on how couplings vary across latent state, not their absolute level, and we report influences centered at the mean latent state. Each population's latents can be rescaled with an inverse rescaling of its readout, as in any latent model; comparisons between inputs onto the same population are unaffected, and we standardize latents to unit variance for reporting. Empirically, seeds reaching comparable held-out likelihood give the same influence functions and fixed points; we use the best-likelihood fit for the dynamics analyses.

Limited novelty: As described above, we now more clearly describe the additions of our model relative to existing model classes. And while this may be seen as an engineering achievement of constrained-yet-interpretable coupling parameterizations, empirically we see much better model fits for EI-NODE than for all other model baselines (in-particular the new Wilson-Cowan baseline, see Supplementary Section E.5), where ablations show that parametrizing the link functions with learned monotonic networks are the reason for this performance increase (see Supplementary Table 7).

Grounding in neurophysiology: This model has flexibility to use such biological information, and we do in fact take full advantage of the physiological information available to us within our given dataset, using it to constrain the connectivity between regions in our model (see Section 4.1 L453). If there was additional information available in alternate datasets, it could be incorporated (e.g. constraints on dynamics).

Fixed points and stability: The ground-truth Wang circuit commits to a decision towards the end of trials, so we used this timing to inform where we looked for fixed points in the E1/E2 plane. We found that our models' latent activity at the I value corresponding to this timing had net-zero influences, and these results were robust to nearby I values, confirming the agreement with the canonical circuit. In the linear and RNN baseline models, nearby slices continue to have non-zero influences, showing how these methods do not recapitulate the underlying dynamics.

AuthorsOpenReview ↗

Response to Reviewer hLQn (3/3)

Unconstrained baseline on real data: We fit an unconstrained NODE with an input channel, free of readout and E/I dynamics constraints, to the climbing/optogenetics data and found only slight gains over EI-NODE. The unconstrained linear NODE and nonlinear NODE perform about the same, but Linear EI-NODE fits the perturbation data worse than EI-NODE, indicating that the nonlinearity helps performance under the E/I constraints (see Supplementary Table 8).

Supplementary Table 8. Single-trial fit of E/I-constrained and unconstrained models. Pseudo-R2{R}^{2}, mean ± SD across 12 models (3 data splits × 4 training seeds). Each model's score is averaged over the two optogenetic sessions. Unconstrained models have no E/I sign constraints, no population structure, input to all latents, and an unmasked readout with weights of either sign.

ModelClimbingPerturbation
EI-NODE0.053 ± 0.0050.156 ± 0.016
Linear EI-NODE0.050 ± 0.0050.141 ± 0.020
Unconstrained NODE0.063 ± 0.0120.168 ± 0.008
Unconstrained linear NODE0.060 ± 0.0080.166 ± 0.006

References:

  1. Wilson, H. R., & Cowan, J. D. (1972). Excitatory and inhibitory interactions in localized populations of model neurons. Biophysical Journal.
  2. Miller, K. D., & Fumarola, F. (2012). Mathematical equivalence of two common forms of firing rate models of neural networks. Neural Computation.
  3. Marreiros, A. C., Daunizeau, J., Kiebel, S. J., & Friston, K. J. (2008). Population dynamics: Variance and the sigmoid activation function. NeuroImage.
  4. Jha, A., Gupta, D., Brody, C. D., & Pillow, J. W. (2024). Disentangling the Roles of Distinct Cell Classes with Cell-Type Dynamical Systems. NeurIPS 2024 (Advances in Neural Information Processing Systems 37).
  5. Song, H. F., Yang, G. R., & Wang, X. J. (2016). Training Excitatory-Inhibitory Recurrent Neural Networks for Cognitive Tasks: A Simple and Flexible Framework. PLoS Computational Biology.
  6. Prinz, A. A., Bucher, D., & Marder, E. (2004). Similar network activity from disparate circuit parameters. Nature Neuroscience.
  7. Das, A., & Fiete, I. R. (2020). Systematic errors in connectivity inferred from activity in strongly recurrent networks. Nature Neuroscience.
  8. Basrai, D., Lee, J., Kristl, A., Wang, P., Miri, A., & Glaser, J. I. (2026). Constraining inference of across-region interactions using neural activity perturbations. bioRxiv.
  9. Nejatbakhsh, A., & Wang, Y. (2025). Identifying Neural Dynamics Using Interventional State Space Models. ICML 2025.
Reviewer_iirVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes EI-NODE, which partitions latent neural dynamics by cell population and models the influence of each source population on a target population as a separate component. The authors impose E/I-related constraints using monotonic networks and sign-constrained parameterizations. They evaluate the method on synthetic winner-take-all data, a cell-type label recovery task, and mouse climbing and optogenetic data. They report that EI-NODE achieves predictive performance comparable to standard NODE on synthetic data, and show the population influences estimated by the model and its responses to optogenetic perturbations. The real-data experiments use two recording sessions from one mouse.

Critical strengths

1 EI-NODE incorporates E/I priors and an explicit decomposition of population influences into nonlinear latent dynamics. This model structure has some methodological value and makes it possible to examine how the individual components contribute to the predicted dynamics. 2 The authors test whether estimated influences agree with the simulator’s known generating mechanism in a synthetic circuit. They also apply the framework to cell-class-specific optogenetic perturbation data.

Critical weaknesses

1 The paper’s motivation and positioning of its novelty are insufficient. It generalizes the limitations of standard NODEs into a broader claim that existing methods cannot decompose population contributions, despite prior work on multi-population, multi-region, and cell-type-specific neural dynamics. Although the paper cites some related studies, it does not clearly establish what new capabilities it adds over them, and it lacks strong comparisons with the most closely related methods. As a result, the current experiments do not adequately support the broad research gap claimed by the paper. 2 The influence terms learned by the model do not yet support an interpretation in terms of actual biological effects. Because latent variables can take either positive or negative values, sign constraints in the model do not directly correspond to excitatory or inhibitory synaptic effects. Recovery of influences in synthetic data alone also does not establish that the influences are identifiable in real data. In addition, the real neural data come from one mouse and two sessions, which limits the robustness and generalizability of the biological conclusions.

Questions for authors
  1. Several methods for modeling multi-population, multi-region, and cell-type-specific neural dynamics have appeared in recent years. The experimental comparisons in this paper focus mainly on NODE and RNN variants and do not adequately cover methods that are more directly related to the research question. The paper therefore does not yet establish its motivation or relative novelty. Please clarify how this work relates to the relevant literature and state its substantive new contributions. (1) MR-LFADS [1] distinguishes communication between recorded regions from inputs originating in unrecorded regions. How does this paper rule out effects due to common input? What are the main advantages of EI-NODE over MR-LFADS? (2) dLDS [2] and cell-type dynamical systems can also decompose neural dynamics or distinguish cell classes. What problem does this paper solve that these methods do not? Why does the paper not directly compare EI-NODE with the most relevant methods?
  2. In Equation (4), the main text treats the self-term separately as a linear term and sums over the other populations. Appendix A.1 appears to add a constrained self-interaction term, while Appendix C.1 states that every source-target pair, including self-pairs, uses a monotonic MLP. It is therefore unclear which model was actually used. Please clarify and make the description consistent.
  3. Given that the real-data experiments involve one mouse across two sessions, to what extent can the reported M1/M2 influences be expected to generalize across animals? References [1] Accurate identification of communication between multiple interacting neural populations [2] Decomposed linear dynamical systems (dlds) for learning the latent components of neural dynamics
4 comments
AuthorsOpenReview ↗

Response to Reviewer iirV (1/4)

We thank the reviewer for their thorough feedback. We are glad the reviewer found value in incorporating E/I priors and explicit population influence decompositions into nonlinear latent dynamics, and appreciated our validation against a synthetic circuit and cell-class-specific optogenetic perturbation data. Below are our responses to the reviewer's comments and questions:

1. Motivation, novelty, and positioning. We agree that modeling A) multi-population activity and B) cell-type-specific dynamics have both been active areas in recent years. Our work sits at their intersection, where to our knowledge no flexible, nonlinear, multi-population dynamics model with cell-class (E/I) constraints existed. Concretely, EI-NODE contributes three things that none of the methods the reviewer names provide together:

  • Cell-type-constrained nonlinear latent dynamics. Each population's latent dynamics are an additive sum of source-population influence functions that are sign-fixed and monotone by construction, so the model is nearly as expressive as an unconstrained NODE but every cross-population influence has a fixed biological sign.
  • Decomposable influences. Because the nonlinearity sits inside each source-to-target function rather than around their sum, the contribution of each population to every other can be read off directly and compared across latent states (Section 3.1 L302, Appendix A.1). This is what makes the winner-take-all and M1-dominant-drive results interpretable as circuit statements.
  • Latents grounded in identified units. A masked, non-negative readout ties each population's latents to opto-tagged or waveform-classified single units, which both constrains the decomposition and lets the model recover cell-type labels for unlabelled units (Figure 4 L378).

We have added a Related Work section (Section 2.5 L203) that positions EI-NODE against latent-dynamics, multi-region, cell-type-constrained, mean-field, and efficient-coding models. We address the reviewer's two specific comparisons below, then explain our choice of baselines, including the new comparisons added in this rebuttal.

2. Relation to MR-LFADS. MR-LFADS is a powerful model of nonlinear, interacting multi-region dynamics, but it contains no cell-type structure: regions are the unit of decomposition, communication terms carry no sign, and the readout is unconstrained. Disentangling the contributions of excitatory and inhibitory populations is therefore not straightforward, and also not expected to provide biologically-feasible influences, within its framework. To make this concrete we trained MR-LFADS on the synthetic decision-making circuit, treating the E1, E2, and I populations as three regions with one-dimensional factors and one-dimensional messages so that both models operate in the same three-dimensional latent space (new Appendix E.3, Table 5). MR-LFADS reconstructs activity as well as EI-NODE and recovers the winner-take-all latent dynamics (Supplementary Fig. 6A, B, D). However, discovering the inferred inter-population influences requires a non-straightforward message ablation and sign-alignment of non-identifiable factors (Appendix E.3.2). More importantly, the inferred inter-population influences from MR-LFADS are biologically implausible: in both trial conditions the excitatory latents exert inhibitory influences on each other (Supplementary Fig. 6C), whereas the generating circuit, and EI-NODE, route that inhibition through the inhibitory population. An unconstrained model finds a solution that fits the data equally well but misattributes the mechanism; appropriate constraints are needed to recover biologically plausible E/I influences.

Common input. The reviewer is right that MR-LFADS's controller (inferred inputs) is designed to absorb drive from unrecorded regions, and that EI-NODE has no such channel: its dynamics are autonomous apart from the known optogenetic input. We address common-input explanations in multiple ways. First, in the synthetic experiments the generating circuit has no unrecorded input, so misattributed influences cannot be blamed on it. Second, in the real data, the optogenetic perturbation is an experimenter-controlled, cell-type-specific input that we provide to the model only at the targeted inhibitory population (Appendix B.2.1). Any effect on a non-targeted population must therefore flow through the learned couplings, and because the perturbation is exogenous it cannot be produced by shared unrecorded drive. The cross-region influences we interpret are those that survive this test on perturbation trials. Third, we acknowledge in the Discussion that during unperturbed epochs, unobserved inputs can be absorbed into the learned dynamics, as in any autonomous latent model (e.g. Kim et al., 2021; Jha et al., 2024).

AuthorsOpenReview ↗

Response to Reviewer iirV (2/4)

3. Relation to dLDS and CTDS. Both decompose dynamics, but along different axes from EI-NODE. dLDS decomposes dynamics over time into a sparse, time-varying combination of linear operators, which is well suited to identifying dynamical regimes and their transitions, but it does not attribute dynamics to source populations or cell classes, and its operators carry no sign structure. EI-NODE decomposes dynamics over source populations, with each component's sign fixed by the source cell type; the two decompositions are conceptually orthogonal, and in principle a dLDS-style dictionary could be placed inside each of EI-NODE's influence functions.

CTDS (Jha et al., 2024) is the closest prior model and the one EI-NODE most directly extends. It assigns separate latents to E and I populations with column-sign constraints on a linear dynamics matrix and non-negative emissions. EI-NODE solves two problems CTDS does not. First, nonlinearity: a linear model cannot express the winner-take-all fixed points of the synthetic circuit (Fig. 3C, D), and on the climbing data the linear cell-type model underperforms EI-NODE on perturbation trials (0.142 vs. 0.157 pseudo-R², Table 7). Second, sign semantics: as the reviewer notes and as discussed in Section 2.3 (L129), CTDS latents are not sign-constrained, so a negative-valued excitatory latent multiplied by its positive weight contributes an inhibitory effect. In EI-NODE the sign constraint is on the output of each influence function, which passes through a softplus and is assigned a fixed sign by source type, so every excitatory contribution is non-negative and every inhibitory contribution is non-positive at every latent state regardless of the latents' sign (see also point 4 below).

4. Choice of baselines. Our baselines are deliberately controlled equivalents of the relevant prior model classes rather than their original implementations. Across all models in the original submission, the BiGRU initial-condition encoder, the population-masked non-negative readout, the loss function, and the training protocol are identical; only the latent dynamics module f changes (Appendix C). This lets each comparison isolate one question:

  • Unconstrained NODE (MLP dynamics, parameter-matched to EI-NODE): are the biological constraints necessary, and what do they cost? This stands in for unconstrained nonlinear latent-dynamics models generally; Sedler et al., 2023 show MLP-based NODE dynamics match GRU-based sequential autoencoders (the LFADS family, to which MR-LFADS belongs) at lower latent dimensionality, so an unconstrained NODE isolates the effect of the constraints without a confounding change of architecture.
  • Linear EI-NODE (sign-constrained linear influence terms, unconstrained linear self-terms, non-negative readout): does nonlinearity matter? This is the NODE-framework equivalent of CTDS. Both constrain the emission matrix to be non-negative; CTDS enforces Dale's law through the column signs of its recurrent matrix A, and Linear EI-NODE through the sign of each linear source-to-target term (non-negative for excitatory sources, non-positive for inhibitory; Appendix A.3); CTDS leaves the diagonal of A unconstrained, and Linear EI-NODE correspondingly keeps unconstrained linear leak/drive self-terms.
  • EI-RNN (Dale's-law column-sign constraints on WrnnW_{\text{rnn}}, unconstrained diagonal): does the EI-NODE parameterization matter relative to the standard way of imposing Dale's law in RNNs (Song et al., 2016)? Because E and I contributions are summed inside the tanh, EI-RNN influences are not additively separable (Supplementary Fig. 4).
  • Unconstrained linear NODE and vanilla RNN (Appendix E.2) complete the 2×2 of {linear, nonlinear} × {constrained, unconstrained} models.

In this rebuttal we have added two further model families, now using the original or standard formulations rather than NODE-framework equivalents:

  • MR-LFADS (Appendix E.3): the published implementation, applied to the synthetic circuit as described in point (1), with influences extracted by message ablation. It matches EI-NODE's reconstruction accuracy but misattributes inhibition between excitatory populations.
  • Wilson-Cowan mean-field models (Appendix E.5), in response to comments from Reviewer hLQn: EI-NODE outperformed three variants of Wilson-Cowan models (results detailed in Supplementary Table 7 and Fig. 7). The results demonstrate that both EI-NODE's learned coupling functions and the per-source placement of the nonlinearity contribute to model performance.

Together these cover the model classes the reviewer names (nonlinear multi-region, cell-type linear, Dale's-law RNN) under matched conditions, and show that EI-NODE is the only one that is simultaneously accurate, nonlinear, and attributes influences correctly to cell classes.

AuthorsOpenReview ↗

Response to Reviewer iirV (3/4)

5. Biological interpretation of influence terms: The reviewer has the correct intuition that sign constraints in latent variable models do not directly correspond to excitatory or inhibitory synaptic effects, which is a limitation of methods such as CTDS. However, this is not true for our method due to the specific constraints we impose. In the EI-NODE model, sign constraints are not enforced on weights multiplying signed latents; they are enforced on the output of each source-to-target interaction function, which passes through a softplus and is then assigned a fixed sign by source type. Each excitatory population's contribution to the target derivative is therefore non-negative for every latent state, and each inhibitory contribution is non-positive, independent of the latents' sign. These interaction functions are also monotonically changing as a function of source activity, and latents map to observed firing rates through a non-negative, population-specific readout. As a result, increased activity in an excitatory (inhibitory) population always increases (decreases) the rate of change of the target population's latent activity, and hence its firing.

6. Identifiability in real data: Identifiability is a known issue (Prinz et al., 2004; Das & Fiete, 2020) that we mitigate with optogenetic perturbations, which shrink the solution set (Basrai et al., 2026; Nejatbakhsh & Wang, 2025). A constant can be shifted between fjE, fjI, and fjj without changing the vector field; all quantities we interpret (coupling signs and shapes, fixed points, winner-take-all structure) depend on how couplings vary across latent state, not their absolute level, and we report influences centered at the mean latent state. Each population's latents can be rescaled with an inverse rescaling of its readout, as in any latent model; comparisons between inputs onto the same population are unaffected, and we standardize latents to unit variance for reporting. Empirically, seeds reaching comparable held-out likelihood give the same influence functions and fixed points; we use the best-likelihood fit for the dynamics analyses.

7. Limited real dataset and generalization across animals: To test how well learned dynamics generalize across sessions (different animals), we froze the dynamics of a trained model and learned only session-specific encoder and decoder weights on a new session. We used the connectivity-constrained EI-NODE and EI-RNN models for this analysis. For each model, we selected the best weights (for both the transfer model and the single-session from-scratch baseline) across three training runs and evaluated accuracy on a held-out test set (20% of the dataset; Supplementary Table 10). We found that A) the dynamics transfer well across sessions (pseudo-R² of 0.057 for climbing and 0.096 for perturbation), performing nearly as well as a session-specific model trained from scratch (0.063 and 0.102, respectively), and B) EI-NODE transfer outperformed EI-RNN transfer on both climbing (0.057 vs. 0.047) and perturbation (0.096 vs. 0.072). EI-NODE transfer comes within 0.006 of its from-scratch performance on both tasks.

ModelClimbing (from scratch)Climbing (transfer)Perturbation (from scratch)Perturbation (transfer)
EI-NODE0.0630.0570.1020.096
EI-RNN0.0550.0470.0740.072

Supplementary Table 10. Transfer models keep the dynamics frozen and learn only session-specific encoder and decoder weights. Test-set pseudo-R² (20% held out), best of three training runs.

AuthorsOpenReview ↗

Response to Reviewer iirV (4/4)

8. Inconsistent description of the self-term: We thank the reviewer for pointing out this inconsistency and have corrected the paper. The definition described in Appendix A.1 is the correct one, and the index of summation in Eq. 4 contained a typo: the summation is over all k,j∈Kk,j \in K not k≠jk \neq j. When k=jk=j (i.e., the source and target are the same population), two terms define its influence: (1) the constrained MLP interaction module, and (2) an unconstrained linear mapping that models a separately learned intrinsic leak/drive term. This leak/drive term is equivalent to leaving the diagonal unconstrained in other models with Dale's law constraints.


References:

  1. Kim, T. D., Luo, T. Z., Pillow, J. W., & Brody, C. D. (2021). Inferring Latent Dynamics Underlying Neural Population Activity via Neural Differential Equations. ICML 2021.
  2. Jha, A., Gupta, D., Brody, C. D., & Pillow, J. W. (2024). Disentangling the Roles of Distinct Cell Classes with Cell-Type Dynamical Systems. NeurIPS 2024 (Advances in Neural Information Processing Systems 37).
  3. Sedler, A. R., Versteeg, C., & Pandarinath, C. (2023). Expressive architectures enhance interpretability of dynamics-based neural population models. Neurons, Behavior, Data analysis, and Theory.
  4. Song, H. F., Yang, G. R., & Wang, X.-J. (2016). Training Excitatory-Inhibitory Recurrent Neural Networks for Cognitive Tasks: A Simple and Flexible Framework. PLoS Computational Biology.
  5. Prinz, A. A., Bucher, D., & Marder, E. (2004). Similar network activity from disparate circuit parameters. Nature Neuroscience.
  6. Das, A., & Fiete, I. R. (2020). Systematic errors in connectivity inferred from activity in strongly recurrent networks. Nature Neuroscience.
  7. Basrai, D., Lee, J., Kristl, A., Wang, P., Miri, A., & Glaser, J. I. (2026). Constraining inference of across-region interactions using neural activity perturbations. bioRxiv.
  8. Nejatbakhsh, A., & Wang, Y. (2025). Identifying Neural Dynamics Using Interventional State Space Models. ICML 2025.
Reviewer_mAMzOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3(+1)
Presentation
3(+1)
Soundness
3
Read review
Paper summary

The paper proposes an extension of the neural ordinary differential equations (NODE) models to an E-I architecture (EI NODE). The network is constrained about the Dale’s law by introducing distinct latent variables to E and I populations and enforcing their Excitatory or inhibitory effect on neural dynamics. Resulting model is tested on a synthetic two-alternative forced-choice (2AFC) task where it successfully recovers neural trajectories in single trials. The model recovers the division to Excitatory and Inhibitory neurons relatively well when a significant proportion of neurons are unlabeled. Finally, the model is tested on data with a motor task, where it infers the influence exerted by E and I neural populations in two motor brain areas, and these inferred contributions are consistent with perturbation responses.

Critical strengths
  1. The extension of a successful and computationally lightweight method such as NODE to a more biologically plausible E-I version which respects Dale’s law is a relevant addition to the field.

  2. The ability of the EI NODE to recover E and I cell types while masking the identity of most neurons is a strong result. Further, the performance and interpretability of results on synthetic data is strong.

  3. Both text and figures are clear, the paper is well written and honestly assesses model’s capabilities.

Critical weaknesses
  1. The authors nicely justify the separation between Excitatory and Inhibitory neurons, but then introduce a network with three populations, E1E_1, E2E_2 and I. While this is not necessarily a problem, it is not justified.

  2. There is some confusion between neural units and populations. The Eq. 4 is defined for populations, but the constraints for ensuring Dale’s law are typically applied to single neurons. As far as I understand, here, Dale's law is imposed to the output of a population, which should be emphasised.

  3. I believe that imposing the function of MLPs is the biggest current limitation of the paper – it imposes certain constraints on the results and limits the computations that the model can perform. Currently described limitation on requiring (partial) knowledge of E and I cell types does not seem to be such a strong limitation to me, because many modern datasets can disentangle these two cell types from signal waveforms.

Questions for authors
  1. Is learning of initial conditions essential? In a simulation of recurrently connected neural networks, (e.g., of type LIF spiking neural network), most population-level effects do not depend on initial conditions, and membrane potentials can be initialized randomly. What would random initial conditions (e.g., within certain boundaries that prevent divergence) imply for model’s dynamics? [clarification]

  2. The influence terms fθjkf^{jk}_{\theta} are imposed to be MLPs. This seems like a rather specific choice. What motivates this choice? [clarification]

  3. While it is explained that Dale’s law constraints the output of the neuron about its sign, it is less clear if it also constraints the recurrent interactions? [clarification]

  4. Authors mention that EI NODE model outperforms the EI RNN and propose expressivity allowed by hidden layers compared to fixed recurrent weights of the RNN. Can this be better explained? Are weights of the EI NODE not fixed after learning? [clarification]

  5. Is imposing MLP function to influence terms required for reproducing the winner-take-all dynamics? [clarification]

  6. To fit the model to experiments, authors mention to fit a joint model to two brain regions (M1 and M2). Would fitting brain areas one by one give very different results? [future work]

  7. The authors review some relevant literature, but they seem to miss an important class of recurrently connected networks that share a number of similarities with their approach. Normative models such as efficient coding with spikes [3] and in particular their extension to E-I architecture [4], are based on dynamical systems’ theory; equations governing neural dynamics are based on low-dimensional and population-based variables. Do authors see a parallel to their approach? If they find it valuable, can authors comment on the similarities and differences between efficient spiking networks and EI NODE networks? [clarification]

[1] Wang (2002) — “Probabilistic Decision Making by Slow Reverberation in Cortical Circuits.” Neuron, 36, 955–968. [2] Brunel & Wang (2001) — “Effects of Neuromodulation in a Cortical Network Model of Object Working Memory Dominated by Recurrent Inhibition.” Journal of Computational Neuroscience, 11, 63–85. [3] Bourdoukan, R., Barrett, D., Deneve, S., & Machens, C. K. (2012). Learning optimal spike-based representations. Advances in Neural Information Processing Systems (NeurIPS 2012), 25. [4] Koren, V., & Panzeri, S. (2022). Biologically plausible solutions for spiking networks with efficient coding. Advances in Neural Information Processing Systems (NeurIPS 2022), 35.

3 comments
AuthorsOpenReview ↗

Response to Reviewer mAMz (1/2)

We thank the reviewer for their careful and constructive review. It is encouraging to hear that extending Neural ODEs (NODEs) into a biologically plausible E-I architecture adhering to Dale's law was recognized as a valuable contribution. We are also glad that the reviewer identified our ability to infer E and I cell types under significant identity masking and the interpretability results of the synthetic data as key strengths. Below are our responses to the reviewer's comments and questions:

Number of excitatory and inhibitory populations: The number of excitatory and/or inhibitory populations in EI-NODE is defined by the user based on the dataset. In the paper, the synthetic dataset has two excitatory populations and one inhibitory population to match the canonical simulated decision-making model defined in Wang (2002), which we now more clearly clarify. The climbing dataset has recordings from areas M1 and M2 with individual cells identified as excitatory (E) or inhibitory (I). Therefore, the number of defined populations is 4, an E and I population per area. More fine-grained levels of population grouping based on cell physiology is an area of future work.

Dale's law at the population level and constraints on recurrent interactions: The reviewer's understanding is correct. In this work, Dale's law is enforced by constraining the output of the MLPs associated with each population, rather than at the level of single neurons. We have emphasized this in Section 2.3 (L129) of the paper.

Recurrent interactions are constrained in two ways: (1) by enforcing sign constraints on the outputs of the interaction MLP modules, and (2) by constraining the linear readout weights (from latents to neurons) to be positive. Given K recorded neural populations, each population has a set of K interaction MLPs. If a population is excitatory, its set of K interaction MLPs is constrained to have positive outputs (influences); the opposite is true for inhibitory populations. Constraining the readout to have positive weights ensures that the signs of the latents are conserved in neural space (note that log-rates are the output of the readout network), so that an increase (decrease) in latent population activity is reflected as an increase (decrease) in the estimated neural activity.

Choice of MLP influence functions: We acknowledge that the motivation for choosing MLP networks lacked sufficient explanation and thank the reviewer for bringing it to our attention. In this work we wish the MLPs to approximate only monotonically increasing or decreasing functions. Our model uses the method introduced by Sartor & Sinigaglia et al., 2025, who show that any MLP constrained to be monotonic following their method can approximate any monotonic function. Additionally, we note that common alternatives for modeling latent neural dynamics have been shown to underperform relative to neural ODEs that use MLPs. In fact, Sedler et al., 2023 show that neural ODEs match the performance of commonly used GRU-based dynamics models while using a lower latent dimensionality. Additionally, Figure 5 of their paper shows that swapping the MLP for an RNN within the neural ODE almost halves reconstruction accuracy. Given these reasons, we believe MLPs are an appropriate choice for the EI-NODE model.

Expressivity of EI-NODE vs. EI-RNN: The weights of the EI-NODE model are fixed after learning. To clarify, the expressivity of an RNN is tied to its latent dimensionality, while the same is not true for the MLPs within the EI-NODE model. For example, in the synthetic dataset the latent dimensionality is set to 3. Therefore, the EI-RNN can only model the latent dynamics using a 3x3 recurrent matrix. However, the MLPs can use an arbitrarily high-dimensional hidden layer, allowing these networks to have a higher capacity.

MLP influence terms and winner-take-all dynamics: Imposing MLP functions on the influence terms is required to reproduce the winner-take-all dynamics while also extracting the true influences. The standard NODE baseline uses an MLP that does not include influence terms but does have separate latent sets for each population. As shown in Supplementary Figure 5, the standard unconstrained-dynamics NODE baseline can reproduce the winner-take-all dynamics; however, it is not possible to extract the influence one population has on another with this model (Also see the new comparison to MR-LFADS in Supplementary Section E.3 in response to reviewer iirV, which fails to recover biologically-plausible contributions to the dynamics). All other baseline models are either linear or use an RNN, and these cannot reproduce the winner-take-all dynamics or the correct influences (Supplementary Figures 3 and 5, main Figure 3D L340).

AuthorsOpenReview ↗

Response to Reviewer mAMz (2/2)

Requirement for E/I labels: We agree with the reviewer that E/I cell-type ambiguity is becoming less of an issue as identification confidence increases in modern datasets. Additionally, advanced cell-type classification techniques, such as NuCLR (Arora et al. 2025), could be used in conjunction with the EI-NODE model. Nonetheless, EI-NODE's ability to correctly identify the ground truth cell-types when only a few identities are unknown should only improve confidence in our method.

Learning of initial conditions: Yes, learning the initial conditions is important for inferring the correct dynamics. The initial conditions are learned by the BiGRU encoder using the entire input sequence. Therefore, random initial conditions (within the boundaries preventing divergence) would bias the dynamical trajectory depending on where they fall in the distribution of learned initial conditions. Unlike in spiking neural networks, the initial condition of the EI-NODE model seeds the latent dynamics, not the state of the neurons within the artificial network. In LIF spiking networks, random membrane-potential initial conditions are stable due to the biological properties instilled in the artificial neurons. Incorporating such spiking networks within the EI-NODE framework would be an interesting line of future work.

Joint vs. separate fitting of M1 and M2: M1 and M2 are recorded simultaneously, so every model is fit to both regions at once; what is shared across the two sessions is the dynamics network, with session-specific encoders and readouts. Fitting each region alone would give different results for two reasons. First, the cross-region couplings would not exist, so drive from the other region would be absorbed into the fitted region's local E/I dynamics, and the cross-region asymmetry in Fig. 5C could not be recovered. Second, when the laser targets the other region, a single-region model sees a delayed response with no modeled cause; the joint model instead tests whether the learned cross-region coupling carries it. More generally, requiring one set of dynamics to explain both regions and both sessions' perturbations narrows the set of circuits consistent with the data, making the joint fit less prone to overfitting than two independent fits (and also allows us to make hypotheses about inter-region interactions that are of interest).

Relation to efficient-coding spiking networks: We thank the reviewer for pointing to this literature, which we now discuss in Section 2.5 (L203). We agree there is a real similarity, in that both efficient spiking networks and EI-NODE describe neural activity with low-dimensional population variables governed by a dynamical system, and both assign distinct roles to signed E and I populations. The key difference is the direction of inference. Efficient-coding networks are normative: the latent variable, its dynamics, and the resulting connectivity are derived from an objective, and E/I structure is a consequence. EI-NODE is descriptive: the latent dynamics and inter-population couplings are learned from data, and E/I structure is a constraint rather than a derived result. We do not claim that EI-NODE explains why the cortex has E/I structure. Rather, we claim it can estimate what identified E and I populations do in a given recording. We see the approaches as complementary, and a natural future direction is to combine EI-NODE with methods for training spiking networks to reproduce population-level factors as in Depasquale et. al. 2023, which would yield spiking E/I circuit models whose latent dynamics are constrained by data through EI-NODE.


References:

  1. Sedler, A. R., Versteeg, C., & Pandarinath, C. (2023). Expressive architectures enhance interpretability of dynamics-based neural population models. Neurons, Behavior, Data analysis, and Theory.
  2. Arora, V., Lachi, D., Knight, I. J., Azabou, M., Richards, B., Hurwitz, C. L., Siegle, J. H., & Dyer, E. L. (2025). Know Thyself by Knowing Others: Learning Neuron Identity from Population Context. NeurIPS 2025.
  3. DePasquale, B., Sussillo, D., Abbott, L. F., & Churchland, M. M. (2023). The centrality of population-level factors to network computation is demonstrated by a versatile approach for training spiking networks. Neuron.
  4. Sartor, D., Sinigaglia, A., & Susto, G. A. (2025). Advancing Constrained Monotonic Neural Networks: Achieving Universal Approximation Beyond Bounded Activations. ICML 2025.

I thank the authors to have provided thorough and precise answers to my questions. I think that the revised paper makes a valuable contribution to the field. The ability of EI-NODE to extract latent variables and inter-region interactions is particularly useful. Extracting low-dimensional latents gives insights about network's computation and is, to my knowledge, difficult to achieve in a relatively general setting as this is done in the present paper. Another useful feature of the EI-node is its capacity to leverage perturbation experiments to constrain degenerate solutions while learning model parameters.

Given the substantial improvements at revision stage I will raise my score.

Comments on the paper

1 comment
Area_Chair_6MPmOpenReview ↗

reminder

Dear reviewers,

This is a gentle reminder to read the authors’ rebuttals and respond in the discussion. The authors have put considerable effort into addressing your feedback, and we encourage you to acknowledge their responses and follow up on any remaining concerns.

Thank you for your time and effort throughout the review process, and for your contributions to the community.

Best Regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.