acceptodds
Under review as a conference paper at ICLR 2027

BrainDyn: A Sheaf Neural ODE for Forecasting and Generating Brain Dynamics

Abstract

Brain activity unfolds as coordinated patterns across regions over time, and modeling these dynamics is a central goal in computational neuroscience. Because activity is distributed over connected brain regions, it is naturally represented on a graph whose nodes are brain regions or channels and whose edges capture relationships between these nodes, with message-passing modeling regional communication. However, standard message-passing networks place every node in a single shared feature space and typically aggregate messages from neighbors by scalar summation, which does not align with neurobiological evidence of brain communication. In order to model the heterogeneous nature of the brain, we introduce BrainDyn, a graph ordinary differential equation (ODE) network built on a cellular sheaf with dynamic restriction maps, recomputed from each forecast window's context rather than fixed once at training time, mediating region-to-region communication through learned linear transformations that drive the time-evolution of neural activity. We evaluate BrainDyn on resting-state fMRI, EEG, and simulated activity from the NEST simulator. Across these modalities, we test multi-region forecasting and autoregressive generation of brain transients, and, using NEST, prediction of altered transients following controlled perturbations. BrainDyn outperforms competing models across modalities, while analysis of the learned features shows that they are expressive, interpretable, and biologically aligned.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Sheaf Neural Networks, Neural ODE, Generative Brain Dynamics.

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

This paper introduces BrainDyn, a framework for modeling brain dynamics that integrates dynamic sheaf neural networks with Neural ODEs. The method employs LSTMs to encode the historical activity of brain regions, models heterogeneous interactions between regions via learnable restriction maps, and combines temporal embeddings with Neural ODEs to facilitate brain activity prediction and autoregressive generation. The authors conducted experiments using fMRI, EEG, and NEST simulation data, and further investigated the model's ability to predict neural activity following perturbations.

Critical strengths
  1. Combining dynamic sheaf structures with Neural ODEs offers a novel approach to modeling heterogeneous inter-regional brain interactions and continuous-time dynamics.
  2. Experiments were conducted using fMRI, EEG, and simulated neural activity data, evaluating performance across short-term prediction, long-term autoregressive generation, and perturbation scenarios.
  3. The authors analyzed variations in restriction maps and their singular value spectra, attempting to link the features learned by the model to information interactions between brain regions.
Critical weaknesses
  1. The technical contributions of the core method relative to existing Sheaf Neural Networks, graph attention mechanisms, and Neural ODEs have not yet been fully elucidated.
  2. The performance gains observed in some experiments—particularly regarding long-term prediction and perturbation scenarios—are limited and insufficient to fully demonstrate the model's generalizability.
  3. The analysis of restriction maps relies primarily on mathematical properties, lacking sufficient evidence to support a biological interpretation.
Questions for authors
  1. The method computes restriction maps based on node features at the end of the historical window and fixes the representations—post-sheaf message passing—throughout the entire prediction horizon, whereas the Neural ODE evolves solely via time embeddings and the current predicted state. This raises a critical question: to what extent do the dynamic interactions learned by the model genuinely evolve with the neural state, rather than being merely static conditions determined by the initial context?
  2. The authors present prediction results following a silencing perturbation of a single neuron using NEST-simulated data; however, the training data consists solely of unperturbed activity, and the prediction context already includes the initial response following the perturbation. Consequently, this experiment primarily tests the model's ability to predict subsequent trajectories given the early response to a perturbation, rather than fully demonstrating its capacity to predict the causal effects of unknown perturbations based on intervention information.
  3. This paper integrates dynamic restriction maps, the Sheaf Laplacian, and Neural ODEs; notably, the restriction maps are generated by an MLP based on neighboring node features, establishing a strong mechanistic link to existing graph attention mechanisms and learnable sheaf networks. Current experiments—primarily involving ablation studies that remove the sheaf structure, LSTM, or temporal embeddings—do not yet conclusively demonstrate whether the performance gains stem from the sheaf's geometric modeling mechanism or from the increased parameter count and more flexible feature transformations.

If the author addresses the questions raised above, I will raise my score.

5 comments
AuthorsOpenReview ↗

Response to Reviewer fxHA (1/3)

We thank the reviewer for their careful assessment of our paper. We appreciate the reviewer’s recognition that combining dynamic sheaf structures with neural ODEs offers a novel approach to modeling heterogeneous inter-regional brain interactions and continuous-time dynamics. We also appreciate that the reviewer found our evaluation across fMRI, EEG, and simulated neural activity valuable, spanning short-term prediction, long-term autoregressive generation, and perturbation scenarios, and that they noted our analysis of restriction maps and their singular value spectra in linking learned features to interactions between brain regions. We address the concerns as follows:

Technical contributions [W1]. The reviewer noted that our contributions relative to sheaf neural networks, graph attention, and neural ODEs were not fully elucidated. Relative to graph attention, which scales each neighbor’s message by a single scalar, BrainDyn learns a matrix for each edge that can rotate, stretch, and project a region’s representation. Each edge has asymmetric maps for its two endpoints, and the maps are approximately low-rank (Fig. 4), so each edge transmits a selected subspace of information. Consistent with this, BrainDyn outperforms a capacity-matched GCN on every metric. Relative to sheaf neural networks, whose maps are computed from static node features, BrainDyn computes the maps from each region’s temporal history and learned identity, and recomputes them at every forecast window. A static-sheaf control degrades in long-horizon generation. Relative to neural ODEs, whose node features evolve in a single shared latent space, BrainDyn conditions the vector field on edge-specific sheaf messages and a learned-frequency time embedding. To our knowledge, sheaf message passing has not previously been used to condition a continuous-time ODE.

Dynamic interactions [Q1]. The reviewer asked whether the learned interactions evolve with the neural state or are fixed by the initial context. Within a single forecast window, the restriction maps are fixed, and the ODE evolves through the time embedding and predicted state. Across windows, however, the maps are recomputed from the most recent context. In long-horizon generation (autoregression), that context consists increasingly of the model's own predictions, so the interactions evolve with the predicted neural state rather than staying fixed by the initial context.

NEST perturbation [Q2]. The reviewer pointed out that the prediction context already includes the early response to the perturbation. We agree with this reading of the experiment. To test whether performance depends on observing the early response, we reran the evaluation with the context window ending exactly at perturbation onset (0% post-onset), so the model receives no perturbed signal:

Table 1: NEST perturbation experiment table

ModelMSE ↓MAE ↓PCC ↑SCC ↑DTW ↓
CNN-LSTM1.1580 ± 0.00950.8218 ± 0.00390.0001 ± 0.00890.0016 ± 0.00670.4383 ± 0.0021
BIOT1.1588 ± 0.00950.8221 ± 0.0039−0.0036 ± 0.0038−0.0032 ± 0.00420.4377 ± 0.0020
EvolveGCN1.1561 ± 0.00980.8213 ± 0.00400.0414 ± 0.00540.0233 ± 0.00610.4357 ± 0.0024
ODEBRAIN0.7316 ± 0.01240.6092 ± 0.00810.6069 ± 0.01210.5509 ± 0.01390.2872 ± 0.0042
RiTINI1.1502 ± 0.00950.8196 ± 0.00380.0825 ± 0.00320.0648 ± 0.00180.4343 ± 0.0020
BrainDyn0.6655 ± 0.02180.5730 ± 0.01160.6534 ± 0.01480.6092 ± 0.01500.2698 ± 0.0064

BrainDyn remains the best model on all metrics (MSE 0.6655 vs. 0.7316 for ODEBRAIN; PCC 0.653 vs. 0.607), with performance nearly identical to the 10% setting reported in the paper (MSE 0.6649). This shows that BrainDyn’s advantage does not depend on observing the early post-perturbation response. We have added the above results to Appendix J in our updated manuscript.

AuthorsOpenReview ↗

Response to Reviewer fxHA (2/3)

Biological interpretation of restriction maps [W3]. The reviewer noted that our analysis of restriction maps relies primarily on mathematical properties. As pointed out by the reviewer, Figure 4 describes the mathematical properties of the maps. Beyond their mathematical properties, the learned maps are approximately low-rank (as seen in Figure 4), so each map transmits only a few directions of a region’s representation and filters out the rest. This parallels evidence that cortical areas communicate through a low-dimensional subspace (Semedo et al., 2019) as highlighted in the manuscript. Figure 3 shows the map norms vary across edges and over time in EEG, consistent with dynamically modulated communication between regions. We extend this analysis to fMRI (Figures 7 and 8). Each of the 400 Schaefer parcels is assigned the mean Frobenius norm of its outgoing restriction maps. These values are projected onto a 3D glass brain and shown from four orientations. Figure 7 shows this for highlighted windows of the fMRI signal, and Figure 8 stacks the regional norms across 5 timepoints. The fMRI results agree with the EEG results. Map norms are spatially heterogeneous across regions, and the high- and low-norm zones shift between timepoints. The model therefore learns time-varying, region-specific communication strengths in a second modality, at higher spatial resolution than in EEG.

Magnitude of performance gains in long-term prediction and perturbation scenarios [W2]. The reviewer noted that the performance gains in long-term prediction and perturbation scenarios are limited and may be insufficient to demonstrate the model’s generalizability. We agree that some of the performance gains are small in these settings. We attribute these to the difficulty of the underlying problem rather than a lack of generality. Autoregressive generation is a particularly hard problem; we note that several other baselines had an MSE greater than one (worse than predicting the mean). In this setting, even small relative improvements are significant. Additionally, BrainDyn achieves the best MSE and MAE across every modality in the long-horizon setting, and the best DTW on fMRI and NEST. We believe this is evidence of superior generality. Where performance gains are particularly modest, this is a valuable result to inform future work and application of the method. For example, a valuable direction for future work is exploring why performance gains for long-term performance gains are more modest for EEG compared to fMRI. In particular, the interpretability of BrainDyn and its biological inspiration mean that these results, combined with ablation experiments, provide mechanistic insight into the important features for different data modalities.

AuthorsOpenReview ↗

Response to Reviewer fxHA (3/3)

Contribution of the sheaf vs. parameter count [Q3]. The reviewer asked whether the gains stem from the sheaf mechanism or from increased parameters and more flexible feature transformations. We note two additional sets of experiments:

  • Capacity-matched controls. We implemented two capacity-matched controls to better establish the contribution of edge-specific sheaf geometry to both short- and long-term accuracy. First, we implemented a capacity-matched graph convolution network (GCN) layer to replace sheaf message passing. In order to ensure the GCN has the same capacity as sheaf message passing we choose an appropriately-sized hidden dimension W and updated each node’s LSTM embedding as hi˙=W2tanh⁡(∑j:eij∈E(deg(i)deg(j))−12W1hj+b1)+b2\dot{h_i} = W_2 \tanh{(\sum_{j: e_{ij} \in E} (\text{deg}(i) \text{deg}(j))^{-\frac{1}{2}} W_1 h_j} + b_1) + b_2, where W1∈RW×16W_1 \in \mathbb{R}^{W \times 16} and W2∈R16×WW_2 \in \mathbb{R}^{16 \times W}. Ultimately, the capacity-matched GCN had eight more parameters than BrainDyn on fMRI data and 14 more parameters than BrainDyn on EEG data. The capacity-matched GCN performed better than the Graph ablation but worse than BrainDyn across both modalities and every metric (see Table 1). On short-horizon forecasting for the fMRI data, the GCN control had 12x worse MSE. We also note that BrainDyn outperformed an ablation which removed the sheaf structure while maintaining the same number of parameters by adding additional capacity to the neural ODE (see Tables 2 and 3).
  • Static-sheaf control. In addition to providing geometric structure, the restriction maps are learned and time-varying. In order to distinguish the impact of time-varying restriction maps we implemented a static sheaf ablation which kept the same sheaf structure and MLP to generate the restriction maps but did not pass node stalks into the MLP. Instead, the MLP was only given the positional embeddings of the nodes, yielding restriction maps which are identical in every window. The static sheaves are competitive with the BrainDyn implementation on the short time horizon; however, they suffer during long-horizon autoregressive generation (see Table 4). On short horizons, the forecast is anchored to observed context which is encoded in the node stalks. Thus, static restriction maps recover most of the benefits of sheaf geometry. However, during autoregressive generation, it is important for regional communication to be able to adapt as the underlying dynamics evolve.

Table 1: Capacity-matched GCN ablation

HorizonMethodfMRI MSE ↓fMRI MAE ↓fMRI DTW ↓EEG MSE ↓EEG MAE ↓EEG DTW ↓
ShortBrainDyn0.0531 ± 0.00910.1238 ± 0.01210.0599 ± 0.00540.4001 ± 0.02530.4209 ± 0.01260.1905 ± 0.0292
ShortBrainDyn Capacity-matched GCN0.6562 ± 0.04200.5686 ± 0.01560.2488 ± 0.00760.4145 ± 0.06580.4320 ± 0.03370.1971 ± 0.0146
LongBrainDyn0.6374 ± 0.04650.5356 ± 0.01570.2654 ± 0.00840.7199 ± 0.02920.6005 ± 0.01610.3008 ± 0.0080
LongBrainDyn Capacity-matched GCN0.8806 ± 0.05580.6687 ± 0.01720.3360 ± 0.00900.7522 ± 0.14430.6174 ± 0.05710.3112 ± 0.0282

Table 2: Sheaf ablation (parameter-matched in Neural ODE) on fMRI and EEG data

DataMethodMSE ↓MAE ↓DTW ↓
fMRIBrainDyn0.0531 ± 0.00910.1238 ± 0.01210.0599 ± 0.0054
fMRIBrainDyn No sheaf (parameter-matched)0.4860 ± 0.03540.4889 ± 0.01470.2121 ± 0.0076
EEGBrainDyn0.4001 ± 0.02530.4209 ± 0.01260.1905 ± 0.0292
EEGBrainDyn No sheaf (parameter-matched)0.4357 ± 0.05890.4446 ± 0.02650.2029 ± 0.0116

Table 3: Static sheaf ablation

HorizonMethodfMRI MSE ↓fMRI MAE ↓fMRI DTW ↓EEG MSE ↓EEG MAE ↓EEG DTW ↓
ShortBrainDyn0.0531 ± 0.00910.1238 ± 0.01210.0599 ± 0.00540.4001 ± 0.02530.4209 ± 0.01260.1905 ± 0.0292
ShortBrainDyn Static sheaf0.0546 ± 0.00910.1236 ± 0.01290.0598 ± 0.00570.3942 ± 0.05480.4185 ± 0.02940.1907 ± 0.0131
LongBrainDyn0.6374 ± 0.04650.5356 ± 0.01570.2654 ± 0.00840.7199 ± 0.02920.6005 ± 0.01610.3008 ± 0.0080
LongBrainDyn Static sheaf0.7150 ± 0.07400.5789 ± 0.03350.2884 ± 0.01610.7351 ± 0.14440.6108 ± 0.05930.3072 ± 0.0293

The results have been added to Appendix C in our updated manuscript.

We thank the reviewer again for their feedback. We hope our responses address the reviewer’s concerns, and if so, we would be grateful if the reviewer would consider raising their score. We are happy to clarify anything further.

Reviewer_jWp2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents BrainDyn, a framework for forecasting and autoregressively generating graph-structured brain activity by combining history-dependent sheaf message passing with a non-autonomous neural ODE. The model uses an LSTM to encode each node’s temporal history and learns restriction maps from neighboring node representations and node identity embeddings to capture heterogeneous interactions. The resulting representations, together with learnable-frequency sinusoidal time embeddings, condition the ODE to predict future activity. Restriction maps are recomputed for each forecasting window and remain fixed within that window. The authors evaluate short- and long-term forecasting on resting-state fMRI, EEG, and NEST simulations, including generalization from unperturbed training data to perturbed trajectories given early post-intervention observations. The reported results outperform the compared baselines on most metrics, while ablation studies support the contributions of the individual components. The main contribution is the integration of these components into a brain dynamics forecasting framework, supported by empirical evaluations across data modalities.

Critical strengths
  1. The forecasting results are encouraging, particularly on fMRI, where short-term MSE drops from 0.0966 for the best competing baseline to 0.0531. BrainDyn also has the lowest mean MSE in the EEG and NEST experiments at both horizons. The NEST experiment is also useful: the model is trained on unperturbed activity and tested on the continuation of perturbed trajectories after observing the early response.

  2. Using sheaf maps to model interactions between nodes is a reasonable choice for this task. Different connections can transform different combinations of features, and these transformations depend on recent activity. This gives the model flexibility to handle differences across regions and forecasting contexts. The analysis in Section 4.3 helps explain what the maps learn by showing changes in their norms across windows and in their singular-value spectra after training.

Critical weaknesses

My main concern is the limited methodological novelty. The model combines LSTM history encoding, learned sheaf maps, and an ODE predictor. Feature-dependent restriction maps are already used in Neural Sheaf Diffusion, and learnable-frequency time features have precedents such as Time2Vec. The contribution therefore appears to lie mainly in combining and adapting these methods for brain signal forecasting. The prediction gains are worth recognizing, but I find the changes to the underlying methods relatively modest.

Questions for authors
  1. How does the model perform when the sinusoidal frequencies are fixed at their initial values, with the same embedding dimension and all other settings unchanged? This would help separate the benefit of learning the frequencies from that of including a time embedding.

  2. What message-passing rule does BrainDyn (Graph) use? Are the LSTM, time embedding, and ODE predictor otherwise unchanged? Please provide parameter counts and training and tuning details for both versions.

5 comments
AuthorsOpenReview ↗

Response to Reviewer jWp2 (1/2)

We thank the reviewer for their careful consideration and review of the paper. We appreciate the reviewer’s recognition that BrainDyn’s forecasting results are encouraging, and that it achieves the lowest mean MSE on EEG and NEST at both horizons. We also appreciate that the reviewer found the NEST perturbation experiment useful, considered sheaf maps a reasonable choice for modeling heterogeneous interactions, and found that the analysis in Section 4.3 helps explain what the maps learn. We address the concerns as follows:

Regarding Methodological novelty [W1]. Novelty is a key strength of BrainDyn. To our knowledge, there has not been any prior work on sheaf ODE networks at all. Further, this is the first application of sheaf networks to the brain. More specifically, BrainDyn reformulates SNNs with learnable time varying sheaves that are akin to a high dimensional attention mechanism on the edges and incorporates novel time embedding within sheafs, that is loosely inspired by positional embeddings in transformers—these changes make Braindyn able to expressively process continuous time data with modulated edges. Many influential advances in machine learning come from combining established ideas in new ways. ResNet built on standard convolutional networks, and its novelty lay in adding residual connections. The Transformer combined existing attention mechanisms with positional encodings, and its novelty lay in removing recurrence entirely. Graph attention networks brought attention to graph message passing, and Neural ODEs reinterpreted residual networks as discretized differential equations integrated with classical solvers. In each case, the contribution lay in how existing components were reformulated and combined to solve a new problem. More recently, the very popular flow matching took continuous normalizing flows and removed the learning of the derivative, regressing the vector field directly instead of learning. Use of components that exist therefore is often evidence of novelty rather than lack of, when the components are used in unusual combinations or novel ways. BrainDyn follows this pattern. To our knowledge, no prior work has combined sheaf neural networks with neural ODEs or applied sheaf networks to brain data. BrainDyn reformulates sheaf neural networks with learnable, time-varying restriction maps that act as a high-dimensional attention mechanism on each edge. It adds a sinusoidal time embedding, inspired by positional encodings, and integrates the resulting representations with a neural ODE to produce a continuous-time forecasting model. The result is a dynamics model that respects the brain’s graph structure while retaining high capacity: edge-level attention modulates communication between neighboring regions, and the time embedding gives the dynamics temporal specificity. We believe this combination yields a novel and effective model of brain dynamics.

AuthorsOpenReview ↗

Response to Reviewer jWp2 (2/2)

Fixed sinusoidal frequencies [Q1]. The reviewer asked how the model performs when the frequencies are fixed at their initial values. To help separate the benefit of learning the frequencies from that of including a time embedding, we have run the following experiment. In the Table below, BrainDyn (paper) refers to the method from the paper, whereas BrainDyn (fixed freq.) refers to an ablated model with frequencies fixed at their initial values. The embedding dimension is the same for both models.

Table 1: Fixed sinusoidal frequencies result

HorizonMethodfMRI MSE ↓fMRI MAE ↓fMRI DTW ↓EEG MSE ↓EEG MAE ↓EEG DTW ↓
ShortBrainDyn (w/o time emb.)0.2681 ± 0.03620.3477 ± 0.02200.1619 ± 0.01090.4981 ± 0.04550.4897 ± 0.02580.2366 ± 0.0111
ShortBrainDyn (fixed freq.)0.0513 ± 0.01430.1182 ± 0.01820.0573 ± 0.00810.3999 ± 0.05670.4222 ± 0.03070.1918 ± 0.0133
ShortBrainDyn0.0531 ± 0.00910.1238 ± 0.01210.0599 ± 0.00540.4001 ± 0.02530.4209 ± 0.01260.1905 ± 0.0292
LongBrainDyn (w/o time emb.)0.8177 ± 0.07870.6337 ± 0.02760.3337 ± 0.01520.7780 ± 0.06880.6099 ± 0.06280.3199 ± 0.0137
LongBrainDyn (fixed freq.)0.6606 ± 0.05010.5504 ± 0.02060.2732 ± 0.01190.7256 ± 0.13540.6062 ± 0.05440.3039 ± 0.0261
LongBrainDyn0.6374 ± 0.04650.5356 ± 0.01570.2654 ± 0.00840.7199 ± 0.02920.6005 ± 0.01610.3008 ± 0.0080

Overall, most of the benefit comes from including a time embedding, although learnable frequencies are particularly useful for long-horizon forecasting and autoregressive generation. As shown in the above table, compared with the model without a time embedding, fixed frequencies already reduce short-horizon MSE from 0.2681 to 0.0513 on fMRI and from 0.4981 to 0.3999 on EEG. Learning the frequencies further reduces MSE on fMRI long-horizon generation, and for long-horizon generation on EEG, with lower MAE and DTW in each of these settings. On EEG short-horizon forecasting, the two versions perform equivalently. Learned frequencies also give more consistent results across splits: the standard deviation is lower on fMRI short (0.0091 vs. 0.0143) and EEG long (0.0292 vs. 0.1354). We conclude that the non-autonomous time embedding is the main contributor, and that learning the frequencies provides a consistent but smaller additional gain, particularly for longer horizons (where it leads to an improvement in all cases). The geometric initialization of the temporal embedding is chosen based on the data in order to represent a range of timescales, which is largely sufficient on short horizons. However, small errors compound during autoregressive generation, so refining the initialization with learned frequencies is particularly valuable in this case. It is also valuable for generalization to new datasets, where oscillations may not be well-matched to a geometrically-spaced grid.

BrainDyn (Graph) and parameter counts [Q2]. The reviewer asked for the message-passing rule, parameter counts, and training details of the Graph ablation. BrainDyn (Graph) sets the restriction maps to the identity so the sheaf Laplacian reduces to the standard graph Laplacian. Message passing then follows the same update rule as BrainDyn, H(l)=(I−Lnorm)H(l−1)H^{(l)} = (I - L_{\mathrm{norm}})H^{(l-1)}, for the same number of layers. At each layer, every region’s representation is updated by the degree-normalized differences between itself and its neighbors, with all feature dimensions treated the same with no edge-specific transformation. The LSTM, time embedding, ODE predictor, solver, loss, and training schedule are otherwise unchanged. There are 58,505 parameters for the base model plus 8 * number of nodes for the learned positional embeddings, so 61,705 for fMRI and 58,913 total for EEG. The graph-only version (sheaf ablation) does not have the MLP parametrizing restriction maps or the learned positional embeddings, so its parameter count is 38,729. Training and tuning details are the same for both sheaf and sheaf-ablation models as provided in Appendix G (previously F) Experimental Setup and Hyperparameters.

We thank the reviewer again for their feedback. We hope our responses address the reviewer’s concerns, and if so, we would be grateful if the reviewer would consider raising their score. We are happy to clarify anything further.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer jWp2,

We thank you again for your thoughtful review and constructive feedback. As we near the end of the discussion period, we are eager to learn whether we have addressed your questions and concerns. Since our last reply to your initial concerns and questions, we have actively engaged with Reviewers FcvK and fxhA to address their questions, which may also answer some of yours (in particular, BrainDyn’s contributions relative to existing sheaf networks, graph attention, and neural ODEs, and whether the gains stem from the sheaf structure rather than parameter count). In particular, Reviewer fxhA commented that their questions were addressed and indicated they will raise their rating. We hope we have addressed yours as well. If so, we would appreciate it if you could update your initial assessment.

Thank you for your time and effort.

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

The paper introduces BrainDyn, a framework for forecasting and generating brain dynamics. It uses context-dependent sheaf message passing and a Neural ODE. An LSTM first encodes each region’s recent history. A shared network then generates edge-specific restriction maps from neighboring representations and node-identity embeddings. These maps construct a sheaf Laplacian that aggregates regional information through matrix-valued transformations. Experiments on four simulated dynamical systems and the Netsim fMRI simulation benchmark demonstrate strong graph recovery, competitive forecasting, and robustness under selected observation-degradation settings. Analyses of restriction-map norms and singular values provide descriptive evidence of heterogeneous learned transformations.

Critical strengths
  • NOGI integrates an explicit graph learner with an MPNN-based Neural ODE, providing a direct path from prediction errors to graph updates.

  • The method achieves the best or jointly best results in 30 of 32 reported metric comparisons and performs strongly across 27 Netsim simulations.

  • Comparisons include continuous-time methods such as NGM and SCOTCH, complemented by observation-degradation experiments, dynamics-transfer tests, and ground-truth-free regularization selection.

Critical weaknesses
  • The independent contribution of sheaf modeling is not established. As described in the manuscript, ODEBRAIN already combines spectral graph features and temporal information with a gated Neural ODE. BrainDyn instead uses LSTM-derived, sheaf-encoded context and learned-frequency time embeddings. Their comparison therefore changes several components simultaneously. The Graph ablation provides useful evidence, but without capacity-matched matrix-message and static-sheaf controls, the gains cannot be attributed specifically to sheaf geometry. Maps also remain fixed within each forecast interval, rather than continuously adapting to predicted states.

  • The normalization claim is incorrect as written. With D(i,i) = deg(i)I in Equation 3, scaling restriction maps by c scales the normalized Laplacian by c². The claimed scale invariance does not follow, and the stability interpretation requires clarification.

  • Integration starts directly from the observed snapshot rather than an inferred latent state. This assumes the observation is a sufficient dynamical state. The implications of noisy initialization, partial observation, and hidden variables are not adequately evaluated.

  • The paper reports no forward/backward NFE, runtime, peak-memory, or scaling curves. The adjoint O(1) statement concerns solver-step storage, not total memory or computation; the dense graph representation remains quadratic in node count.

Questions for authors
  • Which gains remain when graph learning, vector-field capacity, loss, and compute are controlled independently?

  • What solver, step size, and tolerances are used for each system, and how sensitive are results to these choices?

  • Can forward/backward NFE, runtime, and peak memory be reported across network sizes and observation intervals?

  • How does performance change with noisy initial states or partially observed dynamics?

  • What constrains graph–message rescaling, and what evidence supports self/interaction disentanglement beyond predictive accuracy?

9 comments
AuthorsOpenReview ↗

Response to Reviewer FcvK (1/2)

We thank the reviewer for their careful review. We note that parts of the review such as NOGI, Netsim evaluation, graph recovery, NGM/SCOTCH evaluation, and adjoint O(1) statement claim do not appear to correspond to our paper. We have addressed the points concerning BrainDyn below:

Contribution of the sheaf [W1, Q1]. The reviewer noted that the gains cannot yet be attributed specifically to the sheaf without capacity-matched and static-sheaf controls. We note two additional sets of experiments:

  • Capacity-matched controls. We implemented two capacity-matched controls to better establish the contribution of edge-specific sheaf geometry to both short- and long-term accuracy. First, we implemented a capacity-matched graph convolution network (GCN) layer to replace sheaf message passing. In order to ensure the GCN has the same capacity as sheaf message passing we choose an appropriately-sized hidden dimension W and updated each node’s LSTM embedding as hi˙=W2tanh⁡(∑j:eij∈E(deg⁡(i)deg⁡(j))−12W1hj+b1)+b2\dot{h_i} = W_2 \tanh{(\sum_{j: e_{ij} \in E} (\deg(i) \deg(j))^{-\frac{1}{2}} W_1 h_j} + b_1)+ b_2, where W1∈RW×16W_1 \in \mathbb{R}^{W \times 16} and W2∈R16×WW_2 \in \mathbb{R}^{16 \times W}. Ultimately, the capacity-matched GCN had eight more parameters than BrainDyn on fMRI data and 14 more parameters than BrainDyn on EEG data. The capacity-matched GCN performed better than the Graph ablation but worse than BrainDyn across both modalities and every metric (see Table 1). On short-horizon forecasting for the fMRI data, the GCN control had 12x worse MSE. We also note that BrainDyn outperformed an ablation which removed the sheaf structure (Graph ablation) while maintaining the same number of parameters by adding additional capacity to the neural ODE (see Table 2).
  • Static-sheaf control. In addition to providing geometric structure, the restriction maps are learned and time-varying. In order to distinguish the impact of time-varying restriction maps we implemented a static sheaf ablation which kept the same sheaf structure and MLP to generate the restriction maps but did not pass node stalks into the MLP. Instead, the MLP was only given the positional embeddings of the nodes, yielding restriction maps which are identical in every window. The static sheaves are competitive with the BrainDyn implementation on the short time horizon; however, they suffer during long-horizon autoregressive generation (see Table 3). On short horizons, the forecast is anchored to observed context which is encoded in the node stalks. Thus, static restriction maps recover most of the benefits of sheaf geometry. However, during autoregressive generation, it is important for regional communication to be able to adapt as the underlying dynamics evolve. The reviewer is correct in observing that the restriction maps are not updated continuously, they are piecewise-constant across every window. This evolution is not arbitrary because consecutive windows have overlapping history.

Table 1: Capacity-matched GCN ablation

HorizonMethodfMRI MSE ↓fMRI MAE ↓fMRI DTW ↓EEG MSE ↓EEG MAE ↓EEG DTW ↓
ShortBrainDyn0.0531 ± 0.00910.1238 ± 0.01210.0599 ± 0.00540.4001 ± 0.02530.4209 ± 0.01260.1905 ± 0.0292
ShortCapacity-matched GCN0.6562 ± 0.04200.5686 ± 0.01560.2488 ± 0.00760.4145 ± 0.06580.4320 ± 0.03370.1971 ± 0.0146
LongBrainDyn0.6374 ± 0.04650.5356 ± 0.01570.2654 ± 0.00840.7199 ± 0.02920.6005 ± 0.01610.3008 ± 0.0080
LongCapacity-matched GCN0.8806 ± 0.05580.6687 ± 0.01720.3360 ± 0.00900.7522 ± 0.14430.6174 ± 0.05710.3112 ± 0.0282

Table 2: Sheaf ablation (parameter-matched in Neural ODE) on fMRI and EEG data

DataMethodMSE ↓MAE ↓DTW ↓
fMRIBrainDyn0.0531 ± 0.00910.1238 ± 0.01210.0599 ± 0.0054
fMRINo sheaf (parameter-matched)0.4860 ± 0.03540.4889 ± 0.01470.2121 ± 0.0076
EEGBrainDyn0.4001 ± 0.02530.4209 ± 0.01260.1905 ± 0.0292
EEGNo sheaf (parameter-matched)0.4357 ± 0.05890.4446 ± 0.02650.2029 ± 0.0116

Table 3: Static sheaf ablation

HorizonMethodfMRI MSE ↓fMRI MAE ↓fMRI DTW ↓EEG MSE ↓EEG MAE ↓EEG DTW ↓
ShortBrainDyn0.0531 ± 0.00910.1238 ± 0.01210.0599 ± 0.00540.4001 ± 0.02530.4209 ± 0.01260.1905 ± 0.0292
ShortStatic sheaf0.0546 ± 0.00910.1236 ± 0.01290.0598 ± 0.00570.3942 ± 0.05480.4185 ± 0.02940.1907 ± 0.0131
LongBrainDyn0.6374 ± 0.04650.5356 ± 0.01570.2654 ± 0.00840.7199 ± 0.02920.6005 ± 0.01610.3008 ± 0.0080
LongStatic sheaf0.7150 ± 0.07400.5789 ± 0.03350.2884 ± 0.01610.7351 ± 0.14440.6108 ± 0.05930.3072 ± 0.0293

We have updated our manuscript with the above results.

AuthorsOpenReview ↗

Response to Reviewer FcvK (2/2)

Normalization claim [W2, Q5]. The reviewer pointed out that our normalization is not invariant to the scale of the restriction maps. We thank the reviewer for pointing this out and agree with the reviewer. Our normalization accounts for node degree but not the scale of the restriction maps, so the normalized Laplacian scales quadratically with them. We have corrected this in our updated manuscript.

Initial state for integration and solver information [W3, Q2, Q4]. The reviewer noted that integration starts from the observed snapshot rather than an inferred latent state, and asked about the solver settings. While the ODE state is initialized from the observed signal, the vector field is conditioned on the h(L)h^{(L)}, which encodes the full context history of each region and its neighbors. The model therefore does not assume the current observation is a sufficient state and that the encoded history serves as the latent information needed to determine the dynamics. As detailed in Appendix G (previously F), the ODE is integrated with a fixed-step fourth order Runge-Kutta (RK4) solver with step size Δt=1\Delta t = 1, implemented via torchdiffeq. The same solver and step sizes are used for all data cohorts.

We thank the reviewer again for their feedback. We hope our responses address the reviewer’s concerns, and if so, we would be grateful if the reviewer would consider raising their score. We are happy to clarify anything further.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer FcvK,

We thank you again for your thoughtful review and constructive feedback. As we near the end of the discussion period, we are eager to learn whether we have addressed your questions and concerns. Since our last reply to your initial concerns and questions, we have actively engaged with Reviewers jWp2 and fxhA to address their questions, which may also answer some of yours (in particular, whether the gains stem from the sheaf structure rather than added capacity, and how the restriction maps evolve over time). In particular, Reviewer fxhA commented that their questions were addressed and indicated they will raise their rating. We hope we have addressed yours as well. If so, we would appreciate it if you could update your initial assessment.

Thank you for your time and effort.

Comments on the paper

1 comment

Dear Reviewers,

Please take a moment to review the authors’ rebuttal, assess whether your concerns have been adequately addressed, and actively participate in the discussion. Please update your assessment if needed. Thank you for your time and effort!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.