acceptodds
Under review as a conference paper at ICLR 2027

Hi-Brain: Brain-Inspired Bidirectional Hierarchical Learning for Dynamic Brain Connectivity Analysis

Abstract

Modeling the brain’s dynamic information-processing mechanisms is a core challenge in neuroscience and artificial intelligence. Existing methods typically first construct dynamic functional brain connectivity (DFBC), and further use topological modeling or prior constraints to characterize such dynamic information transmission. However, they fail to comprehensively model the hierarchical organization of the brain, leading to learned features with limited biological plausibility. To address this issue, we propose a brain-inspired bidirectional hierarchical network (Hi-Brain) for DFBC analysis. Hi-Brain effectively simulates bottom-up functional aggregation and top-down functional regulation in the brain. Specifically, we first segment fMRI signals into multiple subsequences to model the brain's dynamic patterns. For each subsequence, we design a neuro-inspired node activation learner to capture the adaptive activation of neural nodes under different states. Building on this, we further introduce a high-order subnetwork learner to describe the functional differentiation and global collaboration of the sub-networks, enabling bottom-up high-order functional aggregation. Finally, the aggregated information is used to guide feature optimization in lower-level neural organizations, thereby simulating top-down functional regulation in the brain. Experiments on the ADNI and Parkinson’s disease datasets demonstrate that Hi-Brain not only achieves superior predictive performance, but also effectively explains the neural mechanisms underlying its predictions. More importantly, Hi-Brain provides a new perspective for developing brain-inspired machine learning models.

Primary area:
applications to neuroscience & cognitive science
Keywords:
brain-inspired, dynamic brain network, brain disease diagnosis, optimal transport

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ETwEOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a brain-inspired bidirectional hierarchical learning framework for dynamic functional brain connectivity analysis. The idea is to model brain information processing through both bottom-up functional aggregation and top-down functional regulation. The proposed framework consists of three main components: a neuro-inspired node activation learner, a high-order subnetwork organization learner, and a top-down functional regulation module. Multiple Hi-Brain layers are stacked to capture hierarchical interactions. The proposed Hi-Brain approach was evaluated on two fMRI datasets, ADNI and a Parkinson's disease dataset. The results generally look promising. It is argued by the authors that the proposed approach provides both improved diagnostic performance and a more biologically plausible framework for brain connectivity analysis.

Critical strengths

The main strength of this paper is, as I see, its attempt to integrate multiple levels of brain functional organization into a unified learning framework. Rather than focusing exclusively on node-level connectivity or predefined functional subnetworks, the proposed approach explicitly models node activation, subnetwork interactions, and global information feedback. The combination of spiking neural networks, MoE, and optimal transport is technically interesting, particularly the use of global representations to refine lower-level features.

Critical weaknesses
  1. Limited methodological novelty and insufficient justification of the brain-inspired design. Although the overall framework is interesting, its main components are largely established techniques, including e.g. SNNs, MoE, GCNs, optimal transport, and hierarchical feature refinement. The novelty of the proposed approach is not well articulated relative to existing approaches. A lot of related works have been published over the recent years. The literature review could be largely improved and in more depth (so that the novelty could be highlighted). The connection between the computational components and the claimed biological mechanisms is not sufficiently established. For example, the authors argue that SNNs are suitable for modeling neuronal activation, but fMRI BOLD signals represent indirect, temporally smoothed measurements of neural activity rather than neuronal spikes. The biological interpretation of applying an SNN directly to these signals therefore requires stronger justification. Similarly, the top-down regulation mechanism resembles conventional feature feedback or residual refinement. It is not clear what distinguishes this mechanism from existing hierarchical neural architectures beyond its biological interpretation.

  2. The empirical evidence does not fully substantiate the claims of biological plausibility and interpretability. Although the classification results are promising, the experiments primarily demonstrate predictive effectiveness rather than faithful modeling of brain information processing. The hierarchical temporal prediction experiments evaluate whether signals at one level can predict signals at another level, but successful prediction does not necessarily establish bottom-up information integration or top-down functional regulation. In particular, predicting a brain region from other regions demonstrates statistical dependence rather than directional regulatory influence. There is also insufficient analysis of whether the learned activation patterns and functional subnetworks are stable across subjects or cross-validation folds. Given that biological plausibility is a central motivation of the paper, more rigorous validation is needed to support this claim.

Questions for authors
  1. What is the specific methodological novelty (relative to the state of the art)?

  2. Could you justify the use of SNNs on fMRI BOLD signals? What evidence supports the claim that the SNN-based activation learner captures biologically meaningful neuronal activation?

  3. How does the optimal transport module explicitly capture high-order interactions among functional subnetworks? Why the resulting transport plan should be interpreted as high-order functional collaboration rather than a globally constrained pairwise alignment?

  4. Can you provide stronger evidence for the claimed bidirectional regulation mechanism? Could you clarify how the experiments distinguish genuine top-down regulation from statistical correlation or general representation learning?

  5. How were cross-validation and hyperparameter optimization conducted? Were the hyperparameters selected using nested cross-validation, and were the same data splits and tuning budgets applied to all baseline methods?

Reviewer_75KKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Hi-Brain combines SNN-based node experts, subnetwork GCN experts, optimal transport, and global feedback for fMRI-based disease classification. The authors interpret this architecture as modeling state-dependent activation and bidirectional hierarchical processing. Experiments on ADNI and a Parkinson’s disease cohort report improvements across eight classification tasks, supported by ablations and analyses of regional activity and cross-region signal prediction.

Critical strengths
  1. The node–subnetwork–whole-brain hierarchy is a reasonable inductive bias for this problem. Using global information to guide local processing is worth exploring, and the modular design makes the individual components amenable to ablation.

  2. The evaluation covers two cohorts, eight tasks, and nine relevant baselines, with reported accuracy gains of 2.3–6.3 percentage points over the strongest listed baseline. The module and loss ablations, together with the standard deviations in Appendix C, provide useful information about the model’s performance. Releasing code also helps assess the method.

Critical weaknesses
  1. The biological interpretation is stronger than the evidence supports. Different routing weights do not necessarily identify different brain states, and sparse expert selection does not establish functional specialization. Similarly, globally constrained OT is not sufficient evidence of high-order neural interactions. The signal-prediction experiment shows cross-region predictability, but does not distinguish the proposed regulation mechanism from ordinary statistical dependence.

  2. In the released code, the SNN recurs across ROIs rather than time. Subnetwork construction omits eight ROIs, including THA.L and CAU.L highlighted in Section 4.4, while feedback lacks ROI alignment. These issues complicate the regional interpretation. The OT cost differs from the Euclidean distance specified after Equation (8). Separately, Equations (9) and (10) give different feedback descriptions. The implementation underlying the reported results needs clarification.

Questions for authors
  1. What is the operational definition of “high-order” interaction in Equation (8), and how does the OT representation distinguish it from globally normalized pairwise affinity? Please specify the complete optimization objective, including nonnegativity and entropy regularization.

  2. Does “w/o TDFR” retain the same depth and comparable capacity? Please reconcile Equations (9) and (10) and clarify which modules receive the feedback signal.

  3. Does the released code correspond to the reported results, and how should the above discrepancies be resolved? In Figure 6(b), accuracy on NC vs. TPD ranges from 78.85% (N = 6) to 92.50% (N = 3), a difference of 13.65 percentage points. How were these and the other hyperparameters selected independently of the outer test folds?

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

Hi-Brain combines SNN-based node-level experts, template-defined subnetworks with GCN experts, and an optimal-transport representation of whole-brain interactions. A projected transport representation feeds information into the next hierarchical layer, and window-level outputs are combined for classification. Evaluation includes eight diagnostic tasks on ADNI and a hospital PD cohort, module and loss ablations, and cross-level BOLD prediction experiments.

Critical strengths
  1. The architecture links local temporal processing, subnetwork organization, and global feedback in an explicit hierarchy. The feedback pathway is a concrete design choice that can be ablated, rather than only a biological motivation for an otherwise conventional classifier.

  2. The comparisons cover nine brain-network methods across binary and multiclass tasks on two cohorts. The GRU replacement, attention replacement, feedback removal, and additional-loss ablations provide useful evidence that the proposed components contribute in this implementation.

Critical weaknesses
  1. The interpretation goes beyond what the computation establishes. Equation (8) optimizes pairwise transport with a pairwise Euclidean cost, while the cross-level prediction experiment measures predictive relationships between BOLD signals; neither alone establishes higher-order neural interaction or physiological top-down regulation. Because these claims are central to the paper’s contribution, they need a narrower formulation or controls that distinguish the proposed mechanism from ordinary global aggregation and feedback.

  2. The feedback computation and evaluation protocol remain insufficiently specified. Equation (9) changes the gate input using the original window, whereas Eq. (10) adds transport-derived feedback to learned node features before the next layer; the connection between these descriptions is unclear. In addition, the extensive hyperparameter search lacks a clear inner validation protocol, making it difficult to reproduce the model and assess the reported gains.

Questions for authors
  1. For Eq. (8), what are the precise source and target feature supports, and what entropy coefficient is used by Sinkhorn? How does the transport representation compare with row-normalized attention or the independent coupling formed from the same marginals?

  2. Please give the actual two-layer execution sequence and tensor shapes corresponding to Eqs. (9)–(10), distinguishing inputs to the gates from inputs to the experts. Does the advantage over the no-feedback variant remain against a parameter-matched global-residual feedback module?

  3. How are the window count, expert counts, and loss weights selected within each outer fold, and which subjects form the early-stopping validation set? All windows from a subject should remain together, and paired uncertainty is particularly relevant to small differences such as the MCI-versus-AD AUC of 77.52 versus 77.10.

  4. Does Section 4.5 predict future BOLD observations or reconstruct contemporaneous signals from other regions? Please specify the horizon and held-out protocol and evaluate the no-feedback variant on these same tasks; this would help determine what the experiment demonstrates about hierarchical feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.