acceptodds
Under review as a conference paper at ICLR 2027

SI-HGN: Structure-Informed Hypergraph Networks for Virtual Perturbation-Based Effective Connectivity Inference

Abstract

Effective connectivity (EC) characterizes directed interactions between brain regions, but inferring it from observational BOLD signals remains challenging: invasive stimulation cannot scale to the whole brain, classical methods impose restrictive stationarity or architectural assumptions, and existing data-driven surrogate models lack anatomical grounding and are limited to pairwise interactions.We present SI-HGN, a structure-informed hypergraph network that constructs individualized surrogate brains using structural connectivity (SC) as an anatomical scaffold while learning latent hyperedges to capture coordinated multi-region interactions. Hypergraph-derived edge attention modulates SC-supported anatomical pathways, allowing their contributions to vary across individuals and brain states while preserving anatomical support. Directed EC is then estimated by virtually perturbing individual regions of the frozen surrogate and aggregating the induced response differences across temporal segments of the BOLD time series. On generative benchmarks and macaque connectivity networks, SI-HGN recovers directed topology more accurately than neural perturbational inference and Granger causality (Pearson vs. on simulated recurrent systems; ), with consistent gains across complementary recovery metrics. Individualized EC further reveals disease-related network patterns in Alzheimer's disease and autism cohorts and characterizes directed-network disruption in patients with disorders of consciousness. These results suggest that anatomically grounded, higher-order surrogate brains offer a principled route to interventional-style EC inference, while highlighting the need for validation in larger multicenter and longitudinal cohorts.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Effective Connectivity, Surrogate Brain, fMRI, Virtual Neurostimulation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Sy47OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SI-HGN to extend neural perturbational inference using structural-connectivity-guided propagation and learned hypergraph context.

Critical strengths
  • Incorporating anatomical information into perturbation-based surrogate models is a useful direction for effective-connectivity inference.
  • The paper includes recurrent simulations, a whole-brain model, macaque-derived simulations, and human MTL data.
Critical weaknesses
  • Although graph propagation respects SC support, unrestricted hypergraph aggregation and the global prediction head permit influences outside that support. Quantify these contributions before attributing improved EC recovery to anatomically grounded pathways.
  • Clinical gains are weak relative to simple baselines. ABIDE accuracy is 57.5% versus a 55% majority-class baseline; AD-versus-MCI accuracy is 63.87% versus approximately 61.3%. Include FC-only, SC-only, and combined FC–SC classifiers with uncertainty under the same nested evaluation.
Questions for authors

How are temporal train/test partitions constructed, and are overlapping windows separated across partitions?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Sy47

Dear Reviewer Sy47:

Q1: How are temporal train/test partitions constructed, and are overlapping windows separated across partitions?

Response to Q1. Thank you for this careful question about temporal partitioning and the handling of overlapping windows near the partition boundary.

  • Partition construction. We follow the default temporal partitioning procedure of the public NPI implementation. Sliding windows are first constructed from the time series (three consecutive frames predicting the next), and the windows are then split chronologically. The first 80% of windows are used for model training, and the remaining 20% for validation and early stopping. Random shuffling is applied only within the training set, never across the temporal boundary. SI-HGN and the NPI baseline use exactly the same windows and the same partitioning procedure, so the comparison is not affected by any asymmetry in data handling.

  • Overlapping windows at the boundary. Consistent with the public NPI code and its original experimental settings, we do not additionally separate or remove windows that overlap across the partition boundary; adjacent windows at the training–validation boundary therefore still share two input frames. We expect the impact of this choice to be limited, for three reasons. First, the main evaluation target is EC recovery by the trained surrogate (the correlation between the inferred and reference EC matrices), not predictive generalization to a strictly held-out temporal segment; boundary leakage would compromise the latter, not the former. Second, because both methods are trained and evaluated under an identical protocol, any residual effect of boundary overlap is shared symmetrically and cannot account for the performance differences we report. Third, the reference EC on the simulation benchmarks is defined by the generative model rather than by the validation split, which decouples EC recovery accuracy from temporal generalization.

Clarification added to the manuscript. Following your suggestion, we have added a detailed description of temporal window construction, data partitioning, and the treatment of overlapping windows to Appendix A.1.8 (page 22, line 1134), so that the protocol and its scope are fully transparent and reproducible.

Thank you again for your careful review and constructive suggestions, which have helped improve the transparency of our experimental description.

Weakness

Response. Thank you for raising this issue. We agree that the classification improvements over the majority-class baseline are limited, and we have revised the manuscript to prevent over-interpretation. The clinical analysis is intended as an exploratory probe of whether model-inferred EC contains disease-related information, not as a clinical diagnostic validation, and we do not claim diagnostic utility. As now stated explicitly in “Limitations and outlook” in the Conclusion (page 10, line 515), the absence of subject-level ground-truth EC, the small cohorts, and the lack of external replication bound the clinical interpretation of the current results; the reported disease-related patterns should be regarded as exploratory, model-derived network hypotheses rather than established biomarkers, and the modest classification gains (e.g., 60.0% vs. 56.6% accuracy in AD vs. CN) are directionally consistent across tasks but hypothesis-generating in nature. In future work, we will evaluate the clinical potential of individualized EC through larger multicenter cohorts, independent external validation, and prospective clinical studies.

Thank you for the clarification, but my main concerns remain unresolved. The response does not address the potential contribution of unrestricted hypergraph aggregation and the global prediction head to influences outside SC support. Please provide targeted ablations or quantitative analyses distinguishing the contributions of these pathways from SC-constrained propagation before attributing improved EC recovery to anatomically grounded pathways.

AuthorsOpenReview ↗

Response to Reviewer Sy47

Q1: Please provide targeted ablations or quantitative analyses distinguishing the contributions of these pathways from SC-constrained propagation before attributing improved EC recovery to anatomically grounded pathways.

Response to Q1: We thank the reviewer for this helpful comment. We quantified the performance degradation caused by removing each component, based on the EC Pearson correlations reported in Table 7. For each dataset, we assessed the relative contributions of the three components based on their respective ablation results, as summarized in Table 1, providing an estimate of their contributions to overall EC recovery. To avoid potential bias arising from the shared network support between the structural scaffold and the binary reference EC in the Macaque benchmark, we report the equally weighted average across the remaining three benchmarks (WBM, MTL-L, and MTL-R).

The resulting normalized ablation-based contribution estimates are 40.44% for SC, 26.61% for hypergraph aggregation, and 32.95% for the global prediction head. These results demonstrate that SC plays an important role in EC recovery, representing the largest individual contribution among the three components and highlighting the importance of incorporating anatomical structural priors into EC modeling. Meanwhile, hypergraph aggregation and the global prediction head collectively account for the remaining 59.56%, indicating that higher-order contextual modeling and global information integration also provide substantial complementary benefits. Overall, these findings support the complementary roles of anatomically informed structural propagation and additional information-integration mechanisms in EC recovery.

Table 1. Normalized ablation-based contribution estimates of the three components across three benchmarks (%).

DatasetSCHypergraph aggregationGlobal prediction head
WBM27.1825.5047.32
MTL-L43.5428.1228.34
MTL-R50.6126.2123.18
Average40.4426.6132.95
Reviewer_hEGCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SI-HGN, a structure-informed surrogate model for virtual-perturbation-based effective connectivity inference. It follows the NPI paradigm of learning next-step BOLD dynamics and then estimating directed influence from perturbation-induced changes, while incorporating structural connectivity and a learned hypergraph context into the surrogate. The method is evaluated on simulated, animal, and human datasets.

Critical strengths

1.The motivation is relevant and well framed. Incorporating anatomical structure into a perturbation-queryable surrogate is a natural way to improve upon purely data-driven surrogate models.

2.The empirical evaluation is broad. The paper includes controlled simulations, macaque and human connectivity benchmarks, clinical cohorts, repeated runs, and component ablations.

Critical weaknesses

1.The main training objective does not directly address the perturbation-fidelity problem that motivates the paper. SI-HGN is still trained on observational next-step prediction and FC consistency, with no perturbation-aware supervision or constraint. The structural and hypergraph biases may improve perturbation responses empirically, but the proposed solution addresses the stated mismatch only indirectly.

2.The inferred EC is conditional on the batch of windows used at inference. The adaptive graph is recomputed from the entire batch, and perturbations are applied jointly across all windows, so the resulting EC depends on the supplied window collection and partition rather than only on the local trajectory being perturbed. The paper explicitly notes this dependence, but does not establish how stable the inferred EC is to batch composition or partitioning.

Questions for authors
  1. Can the authors disentangle whether the gains in EC recovery come from improved perturbation behavior beyond any improvement in ordinary next-step prediction? For example, how do predictive accuracy and EC recovery change across NPI, SI-HGN, and the ablated variants?

  2. How sensitive are the inferred EC matrices to inference batch size, window ordering, and partitioning?

  3. The learned incidence matrix �0��0� is a dense soft assignment with positive membership for every node–hyperedge pair. What evidence shows that the learned hyperedges capture specifically higher-order interaction structure, rather than serving mainly as a flexible global mixing mechanism?

  4. The empirical comparison is limited to NPI and Granger causality. Could the authors explain why broader EC baselines were not included and discuss where comparisons with methods such as DCM would or would not be meaningful?

9 comments
AuthorsOpenReview ↗

Response to Reviewer hEGC (1/6)

Dear Reviewer hEGC:

Q1: Can the authors disentangle whether the gains in EC recovery come from improved perturbation behavior beyond any improvement in ordinary next-step prediction? For example, how do predictive accuracy and EC recovery change across NPI, SI-HGN, and the ablated variants?

Response to Q1. Thank you for this incisive question. We have added three analyses to separate the two sources of improvement, namely (i) a cross-dataset comparison of prediction and EC recovery, (ii) direct validation against true perturbation responses on RNN, and (iii) an EC-recovery comparison after matching models on prediction performance. In short, better ordinary prediction cannot explain the EC gains. When prediction performance is matched, SI-HGN still recovers EC substantially better.

First, improved prediction and improved EC recovery do not always coincide (Table 1). SI-HGN improves both on RNN and Macaque, but on WBM its EC recovery correlation rises from 38.51% to 42.27% despite a higher prediction MSE and a lower R2R^{2} than NPI. If the EC gain were merely a by-product of better next-step prediction, this pattern would not occur.

Second, on the RNN dataset, whose generative dynamics are known, we compared the model-predicted perturbation responses directly with the true system responses (Tables 1 and 2). Relative to NPI, SI-HGN reduces ordinary validation MSE by only 1.81%, but reduces true perturbation-response MSE by 41.38% (from 0.002335 to 0.001369); response Pearson, Spearman, and sign agreement rise from 78.10%, 77.52%, and 79.68% to 81.77%, 81.05%, and 81.38%. In other words, most of SI-HGN’s advantage lies precisely in the perturbation regime that EC inference depends on, not in ordinary forecasting.

Third, we matched models by validation MSE (Table 3). Under a 5% matching tolerance, SI-HGN retains an EC Pearson advantage of 4.13 percentage points on RNN; across 158 matched model pairs on Macaque, EC Pearson, AUROC, and AUPR improve by 6.47, 2.34, and 10.22 percentage points, respectively.

Table 1. Ordinary next-step prediction and EC recovery of NPI and SI-HGN.

DatasetModelVal MSE↓Val R2R^{2}↑ECrEC_{r}↑
RNNNPI0.74068224.91%90.33%
RNNSI-HGN0.72728226.28%94.45%
WBMNPI0.16794983.48%38.51%
WBMSI-HGN0.19728280.58%42.27%
MacaqueNPI0.52260747.54%51.74%
MacaqueSI-HGN0.50141549.73%58.26%

Table 2. Validation of perturbation responses against the true RNN dynamics.

ModelResponse MSE↓Response Pearson↑Response Spearman↑Response sign agreement↑
NPI0.00233578.10%77.52%79.68%
SI-HGN0.00136981.77%81.05%81.38%

Table 3. EC recovery gains of SI-HGN over NPI after matching validation MSE within a 5% tolerance.

DatasetEC recovery metricSI-HGN − NPI (percentage points)
RNNPearson+4.13
MacaquePearson+6.47
MacaqueAUROC+2.34
MacaqueAUPRC+10.22

We have added all three analyses to Appendix A.4 (page 36, line 1890).

AuthorsOpenReview ↗

Response to Reviewer hEGC (2/6)

Q2: How sensitive are the inferred EC matrices to inference batch size, window ordering, and partitioning?

Response to Q2. Thank you for raising this important issue. We agree that inference-time stability is essential for reliability. In short, the inferred EC is highly stable under all tested conditions. For RNN and WBM, every metric is numerically 100.00%, and for Macaque even the worst-case Top-20% Jaccard similarity and dominant-direction agreement stay above 96%.

On RNN, WBM, and Macaque, we kept the trained model weights, input window set, perturbation magnitude, and EC aggregation rule fixed and varied only three aspects of inference, the batch size (1, 8, 32, 64, 128, 256, 500, and all windows), the window processing order, and the window-to-batch partitioning (contiguous, randomized, boundary-shifted, and interleaved). We compared each inferred EC matrix with the all-window reference EC of the corresponding model.

Table 4 reports the minimum of each metric across all data instances, window orderings, and partitioning conditions at each batch size; Table 5 reports the corresponding means. To isolate ordering effects, we preserved batch membership and randomly permuted within-batch window order and batch processing order (Table 6); to isolate partitioning effects, we compared randomized membership, boundary-shifted, and interleaved partitions against the contiguous partition at the same batch size (Table 7). Changing the processing order alone is negligible (all minima round to 100.00%), and partitioning strategies differ only marginally on Macaque (minimum Top-20% Jaccard 97.40%, minimum dominant-direction agreement 98.15%).

Table 4. Stability of EC edge weights, strong-edge sets, and dominant directions across inference batch sizes. Each metric is the minimum across all data instances, window orderings, and partitioning conditions at the corresponding batch size. All values are in %.

DatasetBatch sizePearsonSpearmanTop-20% JaccardDominant-direction agreement
RNN1100.00100.00100.00100.00
RNN8100.00100.00100.00100.00
RNN32100.00100.00100.00100.00
RNN64100.00100.00100.00100.00
RNN128100.00100.00100.00100.00
RNN256100.00100.00100.00100.00
RNN500100.00100.00100.00100.00
RNNAll windows100.00100.00100.00100.00
WBM1100.00100.00100.00100.00
WBM8100.00100.00100.00100.00
WBM32100.00100.00100.00100.00
WBM64100.00100.00100.00100.00
WBM128100.00100.00100.00100.00
WBM256100.00100.00100.00100.00
WBM500100.00100.00100.00100.00
WBMAll windows100.00100.00100.00100.00
Macaque199.9799.9596.1396.82
Macaque899.9999.9997.4097.88
Macaque32100.00100.0098.6998.94
Macaque64100.00100.0098.6999.21
Macaque128100.00100.0098.6999.73
Macaque256100.00100.0098.6999.73
Macaque500100.00100.0098.6999.73
MacaqueAll windows100.00100.00100.00100.00

Table 5. EC inference stability across inference batch sizes, window orderings, and window partitions. Each metric is a mean. All values are in %.

DatasetMean PearsonMean SpearmanMean Top-20% JaccardMean dominant-direction agreement
RNN100.00%100.00%100.00%100.00%
WBM100.00%100.00%100.00%100.00%
Macaque100.00%100.00%99.74%99.77%

Table 6. EC inference stability under randomized within-batch window order and batch processing order, with batch membership preserved. Each metric is the minimum across all data instances and evaluated batch sizes, relative to the original contiguous partition at the same batch size. All values are in %.

DatasetPearsonSpearmanTop-20% JaccardDominant-direction agreement
RNN100.00100.00100.00100.00
WBM100.00100.00100.00100.00
Macaque100.00100.00100.00100.00
AuthorsOpenReview ↗

Response to Reviewer hEGC (3/6)

Table 7. EC inference stability across window-to-batch partitioning strategies at matched batch sizes. Each metric is the minimum across all data instances and evaluated batch sizes, relative to the contiguous partition at the corresponding batch size. All values are in %.

DatasetPartitioning strategyPearsonSpearmanTop-20% JaccardDominant-direction agreement
RNNContiguous (reference)100.00100.00100.00100.00
RNNRandomized membership100.00100.00100.00100.00
RNNBoundary-shifted100.00100.00100.00100.00
RNNInterleaved100.00100.00100.00100.00
WBMContiguous (reference)100.00100.00100.00100.00
WBMRandomized membership100.00100.00100.00100.00
WBMBoundary-shifted100.00100.00100.00100.00
WBMInterleaved100.00100.00100.00100.00
MacaqueContiguous (reference)100.00100.00100.00100.00
MacaqueRandomized membership100.0099.9997.4098.15
MacaqueBoundary-shifted100.00100.0098.6999.20
MacaqueInterleaved100.0099.9997.4098.67

These results quantify the practical effect of the batch dependence identified in Appendix A.1.4. They are also consistent with the estimator’s structure. Because EC responses are averaged over all inference windows (Eq. 11 and Eq. 26 in Appendix A.1.4), partition-specific effects are diluted rather than amplified.

We have added Section 3.5, “Stability of EC Inference” (page 9, line 463), to the main text and the detailed comparison in Appendix A.1.5 (page 18, line 918).

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

The paper learns a surrogate of regional BOLD dynamics using structural-connectivity-guided propagation and a latent hypergraph that modulates structural edges. Directed effective connectivity is estimated by perturbing a region in the trained surrogate and measuring the predicted response. Evaluation includes simulated recurrent systems, whole-brain simulations, macaque-derived synthetic networks, human MTL data, and exploratory clinical analyses.

Critical strengths
  1. The anatomical scaffold is incorporated into the dynamics predictor rather than applied only to its output. This is a useful design choice because it makes the role of structural information explicit and testable within perturbation-based connectivity estimation.
  2. The recovery experiments include several settings with reference connectivity rather than relying solely on downstream diagnostic accuracy. Averaging repeated training runs within each independently simulated system before the paired comparison is also an appropriate distinction between optimization variability and independent experimental units.
Critical weaknesses
  1. The inferred connectivity is batch-dependent in a way that affects its interpretation. Equation (7) shares an attention-averaged graph across windows, and Appendix A.1.4 confirms that perturbing one window can affect predictions for other windows through this graph. Without a stability analysis under alternative partitions of the same scan, it is unclear whether the reported matrix characterizes the subject or partly characterizes the chosen batching procedure.
  2. The experiments do not adequately isolate the contribution of the proposed hypergraph architecture from the structural prior and global prediction head. In Table 7, removing HGNN changes macaque recovery from 57.94 to 57.43, whereas removing the global predictor reduces it to 15.74. These results make matched structural-information and prediction-head controls important before attributing the improvement to anatomically grounded higher-order modeling.
Questions for authors
  1. For the same trained model and input windows, how stable are EC edge weights and directions across different batch sizes, window orderings, and partitions? A comparison with per-window attention or a baseline graph held fixed during perturbation would directly address the dependence identified in Appendix A.1.4.
  2. How is the structural scaffold obtained in each recovery benchmark, and does it encode support also present in the reference EC? Does NPI receive equivalent structural information, and how much of the improvement remains with a matched SC-informed predictor?
  3. Does a parameter-matched low-rank pairwise mixer recover the same benefit as the latent hypergraph when the global prediction head and training objective are unchanged? This control would distinguish the proposed higher-order representation from low-rank aggregation.
  4. At the reported perturbation magnitude of 1.0, how sensitive are the results to perturbation sign and amplitude? Please compare amplitude-normalized responses at smaller positive and negative perturbations, particularly the stability of connection signs and rankings.
11 comments
AuthorsOpenReview ↗

Response to Reviewer BFLG (1/7)

Dear Reviewer BFLG,

Q1: For the same trained model and input windows, how stable are EC edge weights and directions across different batch sizes, window orderings, and partitions? A comparison with per-window attention or a baseline graph held fixed during perturbation would directly address the dependence identified in Appendix A.1.4.

Response to Q1. Thank you for raising this important issue. We agree that the stability of EC inference across batch sizes, window orderings, and window partitions is essential for assessing inference reliability. In short, the inferred EC is highly stable under all tested conditions. For RNN and WBM, all stability metrics are numerically 100.00%, and for Macaque the worst-case Top-20% Jaccard similarity and dominant-direction agreement remain above 96%. We have added (1) a systematic stability analysis covering batch size, window ordering, and partitioning, and (2) the fixed-baseline-graph control you suggested, as detailed below.

(1) Stability across batch sizes, window orderings, and partitions.

On RNN, WBM, and Macaque, we kept the trained model weights, input window set, perturbation magnitude, and EC aggregation rule fixed and varied only three aspects of inference, the batch size (1, 8, 32, 64, 128, 256, 500, and all windows), the window processing order, and the window-to-batch partitioning (contiguous, randomized, boundary-shifted, and interleaved). We compared each inferred EC matrix with the all-window reference EC of the corresponding model.

Table 1 reports the minimum of each metric across all data instances, window orderings, and partitioning conditions at each batch size. For RNN and WBM, all four metrics are numerically 100.00%. For Macaque, the minimum Pearson and Spearman correlations are 99.97% and 99.95%, and the minimum Top-20% Jaccard similarity and dominant-direction agreement are 96.13% and 96.82%.

Table 2 summarizes the mean stability across all non-reference conditions. For RNN and WBM, all four metrics reach 100.00% within the reported precision. For Macaque, mean Pearson and Spearman correlations both round to 100.00%, while mean Top-20% Jaccard similarity and dominant-direction agreement are 99.74% and 99.77%.

To isolate ordering effects, we preserved batch membership and randomly permuted both the windows within each batch and the order in which the batches were processed. We then compared the resulting EC with the EC from the original contiguous partition at the same batch size. As shown in Table 3, the minimum values of all four metrics round to 100.00% for all three datasets. Changing the processing order alone therefore has a negligible effect.

We further compared randomized batch membership, boundary-shifted partitions, and interleaved partitions with the contiguous partition at the same batch size (Table 4). RNN and WBM remain perfectly stable, and for Macaque the minimum Top-20% Jaccard similarity and dominant-direction agreement are 97.40% and 98.15%.

Table 1. Stability of EC edge weights, strong-edge sets, and dominant directions across inference batch sizes. Each metric is the minimum across all data instances, window orderings, and partitioning conditions at the corresponding batch size. All values are in %.

DatasetBatch sizePearsonSpearmanTop-20% JaccardDominant-direction agreement
RNN1100.00100.00100.00100.00
RNN8100.00100.00100.00100.00
RNN32100.00100.00100.00100.00
RNN64100.00100.00100.00100.00
RNN128100.00100.00100.00100.00
RNN256100.00100.00100.00100.00
RNN500100.00100.00100.00100.00
RNNAll windows100.00100.00100.00100.00
WBM1100.00100.00100.00100.00
WBM8100.00100.00100.00100.00
WBM32100.00100.00100.00100.00
WBM64100.00100.00100.00100.00
WBM128100.00100.00100.00100.00
WBM256100.00100.00100.00100.00
WBM500100.00100.00100.00100.00
WBMAll windows100.00100.00100.00100.00
Macaque199.9799.9596.1396.82
Macaque899.9999.9997.4097.88
Macaque32100.00100.0098.6998.94
Macaque64100.00100.0098.6999.21
Macaque128100.00100.0098.6999.73
Macaque256100.00100.0098.6999.73
Macaque500100.00100.0098.6999.73
MacaqueAll windows100.00100.00100.00100.00
AuthorsOpenReview ↗

Response to Reviewer BFLG (2/7)

Table 2. EC inference stability across inference batch sizes, window orderings, and window partitions. Each metric is a mean. All values are in %.

DatasetMean PearsonMean SpearmanMean Top-20% JaccardMean dominant-direction agreement
RNN100.00%100.00%100.00%100.00%
WBM100.00%100.00%100.00%100.00%
Macaque100.00%100.00%99.74%99.77%

Table 3. EC inference stability under randomized within-batch window order and batch processing order, with batch membership preserved. Each metric is the minimum across all data instances and evaluated batch sizes, relative to the original contiguous partition at the same batch size. All values are in %.

DatasetPearsonSpearmanTop-20% JaccardDominant-direction agreement
RNN100.00100.00100.00100.00
WBM100.00100.00100.00100.00
Macaque100.00100.00100.00100.00

Table 4. EC inference stability across window-to-batch partitioning strategies at matched batch sizes. Each metric is the minimum across all data instances and evaluated batch sizes, relative to the contiguous partition at the corresponding batch size. All values are in %.

DatasetPartitioning strategyPearsonSpearmanTop-20% JaccardDominant-direction agreement
RNNContiguous (reference)100.00100.00100.00100.00
RNNRandomized membership100.00100.00100.00100.00
RNNBoundary-shifted100.00100.00100.00100.00
RNNInterleaved100.00100.00100.00100.00
WBMContiguous (reference)100.00100.00100.00100.00
WBMRandomized membership100.00100.00100.00100.00
WBMBoundary-shifted100.00100.00100.00100.00
WBMInterleaved100.00100.00100.00100.00
MacaqueContiguous (reference)100.00100.00100.00100.00
MacaqueRandomized membership100.0099.9997.4098.15
MacaqueBoundary-shifted100.00100.0098.6999.20
MacaqueInterleaved100.0099.9997.4098.67

These results are consistent with the structure of the estimator itself. Because EC is defined by responses averaged over all inference windows (Eq. 11 in the main text and Eq. 26 in Appendix A.1.4), partition-specific effects are diluted rather than amplified, and at small perturbation amplitudes the first-order term dominates the cross-window dependence identified in Appendix A.1.4. The experiments above quantify this intuition empirically.

To address this question, we have added Section 3.5, “Stability of EC Inference” (page 9, line 463), to the main text, and a detailed comparison across batch sizes, window orderings, and partitions in Appendix A.1.5 (page 18, line 918).

AuthorsOpenReview ↗

Response to Reviewer BFLG (3/7)

(2) Fixed-baseline-graph control.

Thank you for this suggestion. We agree that holding the baseline graph fixed directly probes the cross-window dependence mechanism, and we have therefore added this control.

Experimental setup. We conducted this comparison on Macaque, RNN, and WBM. Across the two inference modes, we kept the model parameters, input windows, preprocessing, prediction head, perturbation magnitude, and EC aggregation identical. The only difference was that, in the fixed-baseline mode, the baseline graph is held constant during perturbation instead of being recomputed.

Table 5. EC stability under dynamic graph recomputation and a fixed baseline graph. Each metric is a mean. All values are in %. ↑ indicates higher is better and ↓ indicates lower is better.

DatasetInference modePearson↑Spearman↑Top-20%Jaccard↑Dominant-direction agreement↑Relative L2 error↓
MacaqueOriginal dynamic graph100.00%100.00%99.66%99.68%0.29%
MacaqueFixed baseline graph100.00%100.00%99.80%99.77%0.21%
RNNOriginal dynamic graph100.00%100.00%100.00%100.00%0.00%
RNNFixed baseline graph100.00%100.00%100.00%100.00%0.00%
WBMOriginal dynamic graph100.00%100.00%100.00%100.00%0.00%
WBMFixed baseline graph100.00%100.00%100.00%100.00%0.00%

As shown in Table 5, the two inference modes agree closely. For Macaque, fixing the baseline graph changes mean Top-20% Jaccard similarity from 99.66% to 99.80% and dominant-direction agreement from 99.68% to 99.77% (relative L2 error from 0.29% to 0.21%), and RNN and WBM are identical in both modes. Two conclusions follow. First, the batch-mediated adaptive-graph effect on EC is small in practice, corroborating the stability analysis in part (1). Second, the small residual differences indicate that graph recomputation during perturbation is an active but bounded component of the response; holding the graph fixed removes it, at the cost of no longer modeling the system’s adaptive reconfiguration, the effect that the dynamic graph is designed to capture. Both modes are therefore highly stable, and the choice between them is a modeling decision rather than a source of instability.

We have added this analysis and discussion to Appendix A.1.6 (page 19, line 1023), which further examines the dependence identified in Appendix A.1.4.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.