acceptodds
Under review as a conference paper at ICLR 2027

Interpretable Multi-View Dynamic Graph Learning for Intention Prediction Via Information Bottleneck

Abstract

Predicting intention before behavioral onset provides valuable lead time for responsive brain-computer interfaces. However, pre-behavior neural activity exhibits nonstationary spectral dynamics, evolving functional connectivity, and redundant graph structures, posing challenges to both predictive performance and interpretability. We propose an interpretable dynamic graph learning framework for anticipatory intention prediction using ECoG recordings from non-human primates. The framework constructs a multi-view spatio-temporal-frequency representation, models the evolving coupling between neural features and functional topology through dynamic graphs, and learns dynamic predictive subgraphs with an information bottleneck motivated objective. On the vocalization prediction task, our framework achieves accuracy and compares favorably with 20 representative baselines. Further analyses reveal distinct predictive neural structures between rest and vocal preparation and show that these structures reorganize throughout the pre-vocalization interval, with window-specific spectral dynamics providing informative predictive evidence, supporting interpretable advance prediction for responsive brain-computer interfaces.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Anticipatory Neural Decoding, Dynamic Graph Neural Networks, Interpretable Graph Learning, Graph Information Bottleneck

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_FmHjOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies prediction of neural activity preceding behavior on the marmoset ECoG data. The authors propose to align multi views of the data including spectral features, window-specific functional connectivity and information bottleneck subgraph learning module. Improved classification accuracy is reported, as well as interpretation of learned masks as evolving predictive neural activity. However, I find the paper not very well-written. The contribution is mostly a combination of existing components with no technical novelty. Methodology lacks insights, with repeated emphasis on of "aligning spectral representations and functional topology". However, I do not see challenges in combining these methods and more importantly why the combination would lead to better scores.

Critical strengths

I believe predicting behavior before its onset is an important problem for responsive brain–computer interfaces. The authors have done a good job in introducing the background in the introduction. The authors attempt to facilitate understanding of which neural signals support these predictions by the use of ECoG data.

Critical weaknesses

My main concern is that the paper does not bring forward much technical novelty and there seems to lack a convincing motivation for the combination of existing modules. The proposed cross-view coupling, which the authors repeatedly emphasize throughout the paper, amounts to applying graph attention to node features and adjacency matrices, which are standard steps. It is unclear to me what challenges or failures of the existing methods that this proposed "alignment" resolves.

The experiments also does not seem to sufficiently support the proposal: destorying the feature–graph correspondence reduces accuracy by only 1.181.18\\%, and removing temporal attention slightly reduces accuracy while improving ROC-AUC. The interpretability claims are also not very well supported. Table 9 indicates that random retention outperforms learned retention. While improved scores are appreciated, what is more important is the necessity of the proposed mechanisms, the insights that the combination would work, and empirical evidence that tells what enables the improved scores.

Questions for authors

Please refer to weaknesses for questions.

Reviewer_kh26OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper focuses on the interesting problem of improving the responsiveness of brain-computer interfaces, specifically via more effective early (anticipatory) prediction of intent from electrocorticography (ECoG) signals. To achieve this, the paper proposes a multi-view representation (covering spatial, temporal, and frequency representations) which are aggregated into a dynamic graph structure, and learn predictive subgraphs using a novel objective employing an information bottleneck style design. This system is applied to one specific problem, that of vocalization prediction in a marmoset species. The system and experimental results are the main contributions.

Critical strengths

I have two main comments.

First, the paper is very readable, even to someone who is familiar with the technology applied but not the specific application area (though with a small negative note below). The system architecture is very detailed and comprehensive, with clear notation throughout. The system diagram makes it very easy to understand the overall architecture and how the components relate to each other. This matters a lot in my opinion, as it means that the methods can be applicable to a broader category of readers than just researchers deeply enmeshed in this problem space.

Second, I greatly appreciated that the paper uses a broad variety of baselines, and that it does extensive analysis of the results beyond the main set of numbers shown in tables 1 and 2.

Critical weaknesses

My main concern is related to the strength of the findings and scope, with a smaller concern about presentation.

In terms of strength of findings, I cannot help but note that, while the proposed system tends to perform at or near the top in most evaluations as compared to a broad set of baselines, the variance in the results is high enough as to likely not provide statistically significant improvements over the strongest baselines (often, though not always, BrainIB). Consequently, I would expect a discussion of what value the proposed system offers, if the results are, indeed, not statistically significant. Some options might involve cost (in terms of training and/or inference speed), amount of data required for training, robustness to mismatched training-evaluation conditions, etc. Essentially, it would be useful to have some intuition as to how important these results are. Furthermore, looking at the ablation study (table 3), it appears that the only meaningfully different result is obtained when removing the frequency encoder; all the other results appear to be within the margin of error of the configurations listed. It would be interesting to determine whether a much simpler system can be built which nevertheless performs near the top of the baselines.

In terms of scope, while admittedly an outsider to the field, I will flag that using only two specimens in the study feels fairly limiting, especially as the generalization analysis (table 2) shows significant degradation compared to the aggregate results shown in table 1. So, while it is true that the proposed system still performs at or near the top in many metrics, the drop in performance suggests that the impact of the method is relatively small in terms of generalization power, which is not how the results are presented.

Finally, a small note, again from the perspective of an outsider. Some terms, such as ECoG, are never introduced formally. I would recommend a pass over the paper to improve readability even further.

Questions for authors

A suggestion for improvement as much as a request for future work: while I greatly appreciated the in-depth analysis discussion, I think that it would benefit from having a set of research questions listed out explicitly at the onset (beyond just the three modeling contributions included in section 1).

Reviewer_x3swOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an interpretable dynamic graph learning framework for anticipatory vocalization prediction using ECoG recordings from two marmosets. The framework divides 128-channel ECoG signals into overlapping temporal windows, constructs window-specific functional connectivity graphs using Pearson correlations, and integrates global and local spectral features. A shared GAT encodes each graph state, followed by temporal attention for classification.

The authors further introduce a node-and-edge mask-based subgraph selector, optimized using classification loss, normalized HSIC regularization, and structural constraints. The proposed method achieves 95.55% accuracy under 10-fold cross-validation, with additional evaluation through baseline comparisons, ablation studies, cross-animal generalization, and subgraph analyses.

Critical strengths

1. The research problem is practically important.

Predicting behavioral intentions before their onset is important for reducing response latency and improving the responsiveness of brain-computer interfaces. This paper investigates anticipatory vocalization prediction using pre-behavioral neural activity, addressing a meaningful problem with potential practical applications.

2. The experimental evaluation is relatively comprehensive.

The paper provides comparisons with multiple baselines, component ablations, cross-animal generalization experiments, leave-one-session-out evaluation, temporal perturbation analyses, and subgraph retention and deletion experiments. These analyses provide additional insights into the contributions of different model components and the reliability of the learned representations beyond standard classification metrics.

Critical weaknesses

1. Limited methodological novelty, as the framework primarily combines existing components.

The proposed framework integrates existing techniques, including frequency-domain encoding, Pearson correlation-based graph construction, GAT, temporal attention, and subgraph selection. In particular, the proposed cross-view coupling mainly involves applying a standard GAT to spectral node features using their corresponding adjacency matrices. The technical novelty of this mechanism is therefore not sufficiently established. The authors should better clarify the unique contributions of the proposed framework and justify why such a complex combination of existing components is necessary.

2. Insufficient theoretical justification for the information bottleneck objective.

The authors use normalized HSIC to measure the statistical dependence between full-graph and masked-graph representations. However, HSIC is not equivalent to mutual information, and minimizing it does not directly guarantee the information compression required by the information bottleneck principle. Furthermore, the objective lacks an explicit constraint on subgraph sparsity. Therefore, the current formulation does not sufficiently justify the claims of reducing redundant information and learning compact predictive subgraphs.

3. Limited performance improvements and unclear statistical significance.

The proposed method achieves an accuracy of 95.55%, outperforming BrainIB (94.37%) by only 1.18 percentage points. However, no paired statistical significance tests are provided to establish whether this improvement is statistically reliable. Moreover, the ablation results suggest that the frequency encoder contributes substantially more than several other components. Removing temporal attention only slightly reduces accuracy while improving ROC-AUC. These results raise questions about the necessity of the overall complex architecture and whether comparable performance could be achieved using a simpler model.

Questions for authors
  • Please see the weaknesses.
  • Alos, given the motivation of real-time brain-computer interfaces, could the authors report the end-to-end computational cost and inference latency, including frequency-domain encoding, functional connectivity graph construction, and dynamic graph processing?
Reviewer_SHDCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for predicting anticipatory intention, aiming to improve decoding accuracy through experiments on a pre vocalisation task using ECoG recordings collected from the primary auditory cortex and prefrontal cortex of two awake adult common marmosets. The proposed method learns 1) physically aligned multi-scale representations 2) the evolution of fused node representation coupling with its corresponding functional topology 3) temporally resolved predictive subgraphs complementarily in a hierarchical order, and is compared across 20 baselines.

Critical strengths
  1. Relatively well placed research question, and combining temporal, spatial, and spectral views of neural activity is well motivated. Modelling the relationships between the complementary views is an important direction in (dynamic) representation learning for brain systems.

  2. The experimental pipeline is relatively complete, with comparisons with multiple baselines, ablations, and cross animal evaluations.

Critical weaknesses
  1. Frequency encoder shows substantial contribution in ablation, but the spectral representation depends on a few fixed parameters, e.g, the 128hz cutoff, frequency quantiles (which frequency components receive greater weights), and the rationale and sensitivity are not sufficiently discussed. Likewise, the temporal representation’s dependence on input interval, 200ms window length and the stride.
  2. The interpretability and performance claims need some caution. The biological interpretation of the changing subgraphs is unclear; in table 1, accuracy gain over BrainIB is relatively modest, and the uncertainty for the difference is not reported, adding this would strengthen the robustness of the improvement.
  3. Presentation point – figures (particularly figs 2, 4) would benefit from more informative descriptions/captions and better connection to corresponding discussion.
Questions for authors
  1. The current method combines different components of several existing methods, could authors clarify the contribution of this integration? Does the combination provide further insights into pre vocalisation neural dynamics?
  2. Could the authors also explain the leave-one-session-out result in the supplementary, which shows 80.13% ± 13.81%?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.