acceptodds
Under review as a conference paper at ICLR 2027

SemanticBN: From Individual Brain Networks To Evidence-Grounded Semantic Units

Abstract

Brain disorder identification from resting-state functional MRI requires traceable connections between individual connectivity measurements, model decisions, and language descriptions. We propose SemanticBN, which links attribute-level prediction to claim-level language generation through traceable brain-state records. Within shared functional units, SemanticBN learns subject-specific within- and between-unit connectivity representations through joint disease identification, attribute preservation, and bidirectional brain–text alignment. A decomposable predictor quantifies each measured attribute’s contribution to the decision score relative to an explicit baseline. Each contribution is linked to its underlying measurement, training-derived reference, and data source, enabling decision-related statements to be checked against the predictor’s computations. These linked records also govern the language interface: claim-level permissions specify supported statements, and a frozen language model uses authorized evidence to answer questions and describe individual brain states. An output verifier checks numerical consistency, evidence citations, and interpretive scope. By coupling attribute-level prediction with claim-level evidence control, SemanticBN provides a unified basis for disease identification and auditable language description, connecting what is measured, how the model uses it, and which statements the evidence supports.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Brain Network Representation Learning, Self-Supervised Community Discovery, Brain Semantic Units, Brain–Language Alignment, Evidence-Grounded Descriptions, Resting-State fMRI

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_txZaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors proposed SemanticBN, with an attempt to develop a novel neural network that bridges brain units and langauge. The framework starts with establishing correlations between signals from different brain regions, and has three stages: self-supervised learning forms individual soft communities and static unit vectors linked to spatial relations and available temporal records; neuroscientific reference checks determine which interpretations the evidence supports and govern controlled description generation; and brain–language alignment pairs unit vectors.

Critical strengths

This paper raises an interesting and profound concept brain semantic units and tries to link this with language. Or ideally, this semantic unit as semantic representation of language.

Critical weaknesses

Although intersting, this work neglects the research work in psychological linguistics, spatial linguistic description, etc., that have profound results in meaning representation in mind. For short, geographical concepts are represented as regions in mind. A geographcial taxonomy will be represented as partial hierarchical tree in mind. This also supported by the barrier effects in experiments. From this perspective, the semantic unit shall be regions or mathematical manifolds.

Questions for authors

Have you searched related literature in cognitive science, qualitative spatial representation, and psychological linguistics on the topics of semantics of linguistic spatial description?

Why do you think it is sufficient to use vector embeddings to encode spatial relations, for example, the part-whole relation, or the connection relation?

1 comment
AuthorsOpenReview ↗

Cognitive Foundations and Relational Representation

We thank the reviewer for raising an important question about the cognitive foundations of brain semantic units. We agree that spatial meaning cannot be reduced to generic vector similarity, and clarify the scope of our representation below.

  1. Connection to cognitive science and spatial semantics. We recognize the extensive literature on spatial language and cognitive representation, including Landau and Jackendoff (1993), Hayward and Tarr (1995), Gärdenfors (2000), Zlatev (2012), and Landau (2017). These studies show that spatial meaning involves structured regions, reference frames, geometric and topological relations, and functional relations. In particular, conceptual-space theory represents concepts as regions in structured domains, while spatial-language research distinguishes geometric relations from functional and force-dynamic relations. These perspectives are highly relevant to the long-term goal of grounding brain-network units in cognitive meaning.

However, SemanticBN does not claim to provide a complete cognitive theory of spatial or linguistic semantics. Its present objective is more specific: to construct measurable brain-network candidates, associate them with explicit connectivity and spatial-composition records, and constrain language descriptions according to available evidence. We will clarify this distinction and better position the work in relation to cognitive science literature.

  1. Why vector embeddings are not sufficient by themselves. We do not assume that a vector embedding alone fully represents part–whole, spatial, or connectivity relations. In SemanticBN, the brain vector serves as a computational representation for alignment and retrieval; it is not the sole representation of the unit's meaning. Each candidate unit is also associated with structured records, including ROI composition and pooling weights, within-unit connectivity, between-unit connectivity, and available temporal measurements. These records preserve explicit measurements and relations that cannot be recovered reliably from an unconstrained embedding alone.

For example, a part–whole relation is represented through the unit's ROI membership and weighted composition, whereas connectivity relations are represented through measured within- and between-unit FC profiles. The vector provides a compact representation for learning and matching, while the structured evidence record retains the underlying relational information. The two serve complementary purposes.

  1. Regions, manifolds, and the meaning of brain semantic units. We agree that region-based or manifold-based representations are important theoretical possibilities. Our soft communities can be viewed as distributed, weighted groupings of ROI nodes in the brain-network space, rather than isolated points. Nevertheless, we do not claim that these communities have been demonstrated to form cognitive-semantic regions or mathematically validated manifolds. Establishing such a correspondence would require additional experiments linking neural organization to independently measured spatial and conceptual structures.

Thus, our use of “candidate brain semantic units” is intentional: it identifies measurable neural-network objects that may support semantic interpretation, not already established cognitive concepts. The proposed framework provides a way to connect their measured properties to controlled descriptions, while leaving the stronger claim of cognitive-semantic correspondence open for future investigation.

In summary, we view cognitive spatial semantics as an important theoretical foundation and future direction, not as something that can be replaced by generic vector embeddings. SemanticBN combines compact vector representations with explicit relational evidence records; it does not claim that vector similarity alone captures the full structure of meaning.

Reviewer_k8p5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SemanticBN proposes a six-stage pipeline that maps individual functional-connectivity patterns to soft brain-network communities, measured properties, conditional text descriptions, brain-text alignment, and record-based question answering. The study evaluates ADNI and MDD cohorts and reports community stability, property prediction, retrieval, and QA results. Notably, the manuscript includes several negative or inconclusive findings.

Critical strengths
  1. The unusually transparent reporting of weak and negative results is valuable. Comparisons against raw functional connectivity, fixed-reference pooling, and alternative retrieval settings make the empirical limitations visible rather than hiding them.
  2. The attempt to separate measured properties from generated descriptions, with explicit permission checks in the QA pipeline, is a sensible design choice for an application where unsupported anatomical statements would be problematic.
Critical weaknesses
  1. The proposed anatomical interpretation remains conditional on unresolved provenance. Section 4.1 explicitly reports unverified ROI-column-to-atlas mapping in both cohorts and unconfirmed data-access/secondary-use clearance. This does not establish wrongdoing, but it limits anatomical conclusions and requires resolution before reliable interpretation. In addition, the paper acknowledges that Eq. (3) penalizes an unavoidable positive diagonal in QQ^T against a zero-diagonal target; the downstream impact of correcting that objective has not been tested.
  2. The experiments do not yet demonstrate a clear advantage for the proposed semantic representation over strong controls. Table 2 reports MDD joint two-property accuracy of 69.03% for individualized pooling versus 92.58% for raw-FC ridge. In the separate information-matched MDD paraphrase task, Table 3 reports 7.91% brain-to-text and 4.90% text-to-brain accuracy for the individualized MiniLM condition. Section 4.6 reports 92.22% QA correctness with full context versus 64.44% with learned top-three retrieval. The authors candidly report these negative results, but they weaken the case for the proposed pipeline as an effective alternative.
Questions for authors
  1. Can the ROI order/atlas binding and relevant data-use permissions be verified for both cohorts, and how would this change the anatomical analyses?
  2. Would an off-diagonal reconstruction objective, evaluated through the full downstream pipeline, improve stability, property prediction, and brain-text correspondence?
  3. After rerunning the revised permission policy end-to-end, what measurable advantage remains over raw-FC, fixed-reference, and full-context controls?
Reviewer_BvEtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SemanticBN, a framework that learns individualized brain network communities from functional connectivity and links them to evidence-grounded language descriptions.

Critical strengths

The paper addresses an interesting problem of connecting individualized brain networks with interpretable language descriptions. The motivation for evidence-grounded brain representations is relevant to trustworthy neuroimaging analysis.

Critical weaknesses
  1. The motivation is interesting, but the technical novelty is very limited.

  2. More importantly, the paper does not establish why this combination is necessary or what fundamental limitation of existing methods it overcomes. The contributions are primarily framed around combining existing components into a pipeline.

  3. The paper repeatedly presents the learned communities as meaningful brain semantic units, yet provides little direct evidence that these units possess stable, interpretable, or independently validated semantic properties.

  4. The authors acknowledge that the ROI-to-atlas correspondence remains unverified in both ADNI and MDD, which seriously undermines the neuroscientific validity of the proposed semantic units.

  5. There is insufficient evaluation of individual-level reproducibility, cross-subject semantic consistency, or independent neuroscientific validation.

  6. Although the authors include BrainLM and BrainGFM baselines, these are simplified, locally trained adaptations rather than the original pretrained models. Therefore, the paper does not establish competitiveness with existing brain foundation models.

Questions for authors

see weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.