acceptodds
Under review as a conference paper at ICLR 2027

TopoMontage: Topology-Preserved Cross-Montage Alignment for Universal EEG Representation Learning

21 views

Abstract

Electroencephalography (EEG) pretraining across datasets remains difficult because recordings are often acquired under heterogeneous montages, with substantial variation in cap layouts, channel availability, channel ordering, and reference schemes. Such inconsistencies weaken spatial correspondence at the sensor level and limit the transferability of learned representations. We present TopoMontage, a self-supervised framework for EEG representation learning that explicitly models cross-montage correspondence through topology-preserving alignment. Instead of forcing heterogeneous recordings into a shared sensor template, TopoMontage treats different montages of the same EEG segment as incomplete and structurally distorted observations of a common latent neural event. Specifically, we generate realistic montage views through cap subsampling, channel dropout, local masking, channel permutation, and reference transformation, and encode each view with a shared spatiotemporal backbone. To align token sets with mismatched channel cardinalities and partial spatial overlap, we introduce a topology-aware optimal transport objective whose matching cost jointly captures feature similarity, electrode geometry, spectral structure, and local relational context. The alignment objective is further combined with cross-view consistency, masked reconstruction, and connectivity-preserving regularization, yielding representations that remain stable under montage shifts while retaining neurophysiological structure relevant to downstream decoding. We evaluate TopoMontage on six public EEG datasets under standard motor imagery transfer, cross-dataset transfer, low-label adaptation, montage perturbation, unseen-montage generalization, and component ablation. Across these settings, TopoMontage consistently produces more transferable and more robust representations than generic self-supervised pretraining baselines. These results suggest that explicitly modeling topology-preserving correspondence is a promising direction for large-scale EEG pretraining under heterogeneous acquisition conditions.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Electroencephalography (EEG), self-supervised learning, cross-montage transfer, topology-aware alignment, optimal transport

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_xzx7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose TopoMontage, a self-supervised pre-training method (like SimCLR/BYOL for EEG) designed to make an EEG encoder invariant to different headset layouts ("montages"). During pre-training, they take a single 4-second EEG clip and create two "corrupted views" of it by dropping different subsets of sensors, shuffling the channel order, masking time patches, and changing the voltage reference. Because View A might have 30 channels and View B might have 45 different channels, you can't just compare Channel 1 to Channel 1. So, they use Optimal Transport (Sinkhorn matching) to softly match each sensor in View A to a weighted average of similar/nearby sensors in View B, and train the model using a combination of 5 different loss functions so the two views agree both locally and globally.

The central proposed contribution is an explicit alignment mechanism between channel-token sets using entropy-regularized optimal transport. Its matching cost combines learned feature similarity, electrode geometry, spectral descriptors, and local-neighborhood information. The resulting soft correspondence supports alignment between views with different channel counts. The framework also introduces transported token consistency, global representation consistency, masked signal reconstruction, and a regularizer matching latent similarities to within-montage sensor correlations. Together, these objectives are intended to encourage montage robustness while retaining information useful for downstream decoding.

The described validation uses three pretraining datasets and three downstream motor-imagery benchmarks, with evaluations covering binary left-versus-right classification, linear probing and fine-tuning, channel reduction, permutation and reference shifts, unseen montage templates, and component ablations. Comparisons include training from scratch and generic masked and contrastive self-supervised baselines. However, all seven result tables explicitly identify their values as draft or simulated results that must be replaced by actual experimental runs. Thus, the manuscript currently provides a proposed method and evaluation design, rather than empirical validation of its claimed improvements in transferability, robustness, or preservation of neural structure.

Critical strengths
  • One key strength is the paper’s explicit treatment of correspondence across heterogeneous EEG montages. Rather than merely designing an encoder that accepts different channel counts, TopoMontage proposes aligning channel representations using feature similarity, electrode geometry, spectral information, and local spatial context. This directly addresses a meaningful challenge: accepting different sensor configurations does not necessarily produce comparable representations. The formulation also exposes individual components that can be tested through controlled ablations.
Critical weaknesses
  • Every single number in every table in this paper is fake/placeholder data! I am basing this on the fact that every table states that the numbers are placeholders. For this reason, it's difficult to trust what the paper offers, since it's unclear whether the numbers are fabricated. The authors generated synthetic numbers (or had an LLM draft the tables and results section with placeholder numbers) and forgot to delete the prompt/drafting notes or actually run the experiments before submitting. Similarly, there are issues with the references (on page 10, reference 2 is listed as anonymous, but a quick Google Scholar search indicates that this paper is a rejected ICLR 2026 submission and could be cited accordingly).
Questions for authors
  • Do any real experimental results exist? The captions for all seven tables state that the numbers are "draft" or "simulated drafting results" and should be replaced before submission, yet the abstract and conclusion claim that the method demonstrates consistent improvements. Have any of these experiments actually been run? If so, please provide the real results and clearly separate them from placeholders.

  • How does the matching handle non-overlapping brain regions? Equation 8 uses balanced optimal transport, which ensures that 100% of each sensor's weight in View A is matched to sensors in View B. However, if View A only covers the front of the head and View B mostly covers the back, sensors with no true counterpart are still forced to match and have their features pulled together (Eq. 10). Why use balanced transport instead of partial/unbalanced optimal transport, which allows unmatched sensors to be ignored?

  • Why does shuffling channel order change accuracy in Table 4? According to Section 2, the "Permuted" test shuffles the channel row order while keeping each channel paired with its true 3D coordinate. However, your encoder processes each channel independently over time (Eqs. 3-4), builds neighborhoods using 3D coordinates rather than row indices (Eqs. 5-6), and averages across channels (Eq. 11). Mathematically, shuffling the row order should produce the exact same pooled output. Why, then, does Table 4 show TopoMontage dropping from 90.5% to 87.0% on B2a when channels are permuted? What downstream classifier head is being used that breaks permutation invariance?

  • Missing definitions in Section 2: Could you specify the exact equations and hyperparameters for the missing components in Section 2, specifically the full segment loss in Eq. 16 (which omits LrecL_{\text{rec}} and LconnL_{\text{conn}}), the architecture of EtE_t, Γ(pi)\Gamma(p_i), and Ψ(Xi,:)\Psi(X_{i,:}), and the loss weights (λf,λc,λs,λr,β1,β2,β3\lambda_f, \lambda_c, \lambda_s, \lambda_r, \beta_1, \beta_2, \beta_3)?

Reviewer_hB54OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes TopoMontage, a self-supervised framework addressing EEG cross-montage heterogeneity. Rather than interpolating onto rigid sensor grids, TopoMontage models diverse electrode montages as incomplete observations of a continuous neural field s:Ω×{1,…,T}→Rs: \Omega \times \{1,\dots,T\} \to \mathbb{R}.

Multi-view perturbations use cap subsampling, channel dropout, permutation, and reference shifts (Section 2.1, Eq. 2). A shared backbone applies temporal patch embeddings with distance-biased local kk-NN spatial attention (K=6K=6) (Section 2.2, Eq. 3–6). For cross-view token alignment under unequal channels, entropic optimal transport (OT) defines cost DijabD_{ij}^{ab} combining latent distances, electrode geometry (coordinate embeddings), spectra, and local topology (Section 2.3, Eq. 7–8), paired with barycentric projection Llocab\mathcal{L}_{loc}^{ab} (Eq. 9–10), semantic consistency Lglobab\mathcal{L}_{glob}^{ab} (Eq. 11), masked reconstruction Lrec(v)\mathcal{L}_{rec}^{(v)} (Eq. 12), and connectivity preservation Lconn(v)\mathcal{L}_{conn}^{(v)} (Eq. 13–14). Pretrained on TUH and OpenNeuro (ds004448, ds006921), evaluation spans BNCI2014 001, PhysioNet MI, and Lee2019 MI.

Critical strengths
  1. Principled Conceptual Framing of Cross-Montage Shift (Section 1, lines 077–083; Section 2.1, lines 132–135): Modeling heterogeneous EEG montages as incomplete observations of an underlying continuous neural field is conceptually sound, avoiding artificial spatial interpolation artifacts while providing structured geometric priors for spatial attention beyond unguided positional embeddings.
  2. Multi-Modal Optimal Transport Cost Formulation (Section 2.3, Eq. 7–8, lines 186–198): Formulating soft channel correspondence across unequal sensor counts via entropic optimal transport is mathematically principled. The 4-component cost matrix DijabD_{ij}^{ab} (balancing latent features, electrode geometry, spectral profiles, and local topology) is designed to establish geometry-aware correspondences without requiring fixed channel indices.
Critical weaknesses
  1. Unreplaced Simulated/Draft Results Across All Tables (Section 3, Tables 1–7; caption lines 325–326, 336–338, 379–380, 388–389, 433–434, 444–445, 478–479): Every quantitative table in the submission carries a caption statement that the reported numbers are draft or simulated drafting results to be replaced before submission:
  • Table 1 (lines 325–326): "Numbers are draft results for manuscript preparation and should be replaced by actual experimental runs before submission."
  • Table 2 (lines 336–338): "Numbers are simulated drafting results for manuscript preparation and should be replaced by actual experimental results before submission."
  • Tables 3, 4, 5, 6, and 7: Identical notices. The headline claim that "TopoMontage achieves the best performance on all three benchmarks" (line 301) therefore rests on numbers that the authors themselves describe as draft or simulated. As submitted, the tables do not provide verifiable empirical evidence for the paper's claims; I would revisit my assessment if the authors confirm and document actual runs.
  1. Unreported OT Cost Scaling, Missing Baseline Benchmarking, and Task Simplification (Sections 2.3, 2.5, 3.1, 3.2):
  • Unreported OT Cost Scaling: In Section 2.3 Eq. (7), the pairwise cost Dijab=λf∥zi(a)−zj(b)∥22+λc∥ci(a)−cj(b)∥22+λs∥si(a)−sj(b)∥1+λr∥ri(a)−rj(b)∥22D_{ij}^{ab} = \lambda_f \| z_i^{(a)} - z_j^{(b)} \|_2^2 + \lambda_c \| c_i^{(a)} - c_j^{(b)} \|_2^2 + \lambda_s \| s_i^{(a)} - s_j^{(b)} \|_1 + \lambda_r \| r_i^{(a)} - r_j^{(b)} \|_2^2 sums four terms that can live on very different scales: squared distances between latent tokens, squared distances between coordinate embeddings ci=Γ(pi)c_i = \Gamma(p_i), an ℓ1\ell_1 distance between spectral descriptors si=Ψ(Xi,:)s_i = \Psi(X_{i,:}), and squared distances between neighborhood aggregates rir_i. The manuscript does not define Γ\Gamma or Ψ\Psi, does not state whether tokens, coordinates, or spectra are normalized, and does not give coordinate units. If, for example, Γ\Gamma stays close to raw coordinates in millimetres without normalization, the geometric term could dominate DijabD_{ij}^{ab}; under entropic OT (Eq. 8) the Gibbs kernel exp⁡(−Dijab/ε)\exp(-D_{ij}^{ab}/\varepsilon) would then be driven mainly by fixed sensor geometry (or spectra) rather than by the learned features. The reader cannot check this because Section 3.1 (lines 287–292) does not report λf,λc,λs,λr\lambda_f, \lambda_c, \lambda_s, \lambda_r or ε\varepsilon. Furthermore, Eq. (16) introduces Lseg(X)\mathcal{L}_{seg}(X) without an algebraic definition.
  • Baseline Deficit & Missing OT Citations: While Section 1 cites BIOT (Yang et al., 2023) and REVE (Ouahidi et al., 2025) and discusses EEGMirror, E-MAP, and ITSA, none are benchmarked in Tables 1–5; comparisons are limited to same-backbone baselines (Scratch, Masked-SSL, Contrastive-SSL). In addition, despite the "topology-preserving" framing, Eqs. (7)–(8) define a standard linear entropic OT problem whose cost includes a node-wise neighborhood descriptor rir_i (an attention-weighted average of neighboring tokens); they do not compare intra-view relational structure as quadratic (fused) Gromov-Wasserstein matching does (Vayer et al. 2019; Chen et al. 2020). The manuscript also does not cite prior OT-based EEG transfer work (e.g., Chambon et al. 2018) or the OT domain-adaptation work that uses the same barycentric mapping as Llocab\mathcal{L}_{loc}^{ab} (Courty et al. 2017).
  • Task Simplification: The authors evaluate BNCI2014 001, which they themselves describe as "a four-class cue-based MI setting" (line 281), under a binary left-vs-right protocol (lines 296–298; motivated as keeping "the label space consistent across datasets") rather than the standard 4-class setting, which limits comparability with published 4-class results; results are averaged over three random seeds (lines 289–290) and no statistical hypothesis tests are reported.

Score rationale

Presentation (2). The conceptual motivation in Sections 1–2 is clear, but presentation suffers from critical omissions:

  1. Mathematical Incompleteness in Objective Specification: Composite loss Lseg(X)\mathcal{L}_{seg}(X) in Eq. (16) is undefined without explaining how view-wise reconstruction Lrec(v)\mathcal{L}_{rec}^{(v)} (Eq. 12) and connectivity preservation Lconn(v)\mathcal{L}_{conn}^{(v)} (Eq. 14) combine with pairwise Lpair\mathcal{L}_{pair} (Eq. 15). Spectral operator Ψ\Psi in Eq. (7) lacks definition, and Section 3.1 (lines 287–292) does not report the loss weights or OT/temperature hyperparameters (β1,β2,β3,ε,λf,λc,λs,λr,τc\beta_1, \beta_2, \beta_3, \varepsilon, \lambda_f, \lambda_c, \lambda_s, \lambda_r, \tau_c).
  2. Formatting Hygiene: All empirical tables in Section 3 retain visible caption disclaimers stating that the numbers are draft or simulated drafting results.

Informativeness (1). While formulating cross-montage alignment via soft optimal transport over continuous neural fields is conceptually appealing, the paper provides little reliable empirical knowledge because all quantitative tables (Tables 1–7) are described in their captions as draft (Table 1) or simulated drafting (Tables 2–7) results that should be replaced by actual experimental results before submission.

Soundness (1). The submission has major soundness problems; the first cannot be resolved without documented experimental runs:

  1. Unreplaced Simulated Draft Data: Experimental tables (Tables 1–7) contain caption notices stating that the reported numbers are draft or simulated drafting results to be replaced before submission. As submitted, the manuscript therefore does not provide verifiable empirical findings.
  2. Missing State-of-the-Art Baselines: In Tables 1–5, the proposed method is compared only against same-backbone baselines (Scratch, Masked-SSL, Contrastive-SSL). Published foundation models and montage-invariant methods cited in Section 1 (BIOT, EEGMirror, E-MAP, ITSA, REVE) are not included.
  3. Unreported OT Cost Scaling: Eq. (7) sums latent, coordinate-embedding, spectral, and neighborhood terms without stated normalization, and the balance weights λf,λc,λs,λr\lambda_f, \lambda_c, \lambda_s, \lambda_r and ε\varepsilon are not reported, so it is unclear whether the learned features or fixed sensor geometry/spectra drive the transport plan.
  4. Task Simplification and Statistical Deficits: Evaluating BNCI2014 001 under a binary left-vs-right protocol limits comparability with published results on the standard 4-class task, and results rely on three random seeds without statistical hypothesis testing.

Rating (1). Rubric Rationale: Every quantitative experimental table in the manuscript (Tables 1–7; caption lines 325–326, 336–338, 379–380, 388–389, 433–434, 444–445, 478–479) carries an author statement that the reported numbers are draft (Table 1) or simulated drafting (Tables 2–7) results that should be replaced by actual experimental results before submission. As submitted, the tables do not provide verifiable empirical evidence that TopoMontage achieves superior transfer. Combined with the unreported weights and scaling of the OT cost, missing citations to OT prior art, and the incompletely specified training objective (Eqs. 7, 16), I cannot recommend acceptance in its present form. I would revisit my assessment if the authors confirm and document actual runs.

Questions for authors
  1. [clarification] Regarding Table 1–7 Captions: Can the authors explicitly clarify the status of the quantitative data in Tables 1–7? Were any of the reported results produced by actual pretraining and fine-tuning runs on the named EEG datasets, or do all tables reflect simulated drafting projections?
  2. [clarification] Regarding Eq. (7), Eq. (15), and Eq. (16): How are the four terms of the cost DijabD_{ij}^{ab} in Eq. (7) scaled relative to each other: how are the coordinate embedding Γ\Gamma and the spectral operator Ψ\Psi defined, are tokens, coordinates, and spectra normalized, and in what units are electrode coordinates given? If coordinates are in millimetres and unnormalized, could the geometric term dominate the latent-feature term in the transport plan? What is the exact algebraic definition of Lseg(X)\mathcal{L}_{seg}(X), and what values were used for λf,λc,λs,λr,ε,β1,β2,β3\lambda_f, \lambda_c, \lambda_s, \lambda_r, \varepsilon, \beta_1, \beta_2, \beta_3?
  3. [clarification] Sinkhorn vs. Gromov-Wasserstein & Prior Art: In what sense is the alignment in Section 2.3 "topology-preserving", given that Eqs. (7)–(8) define a linear entropic OT problem with a node-wise neighborhood descriptor rir_i rather than a quadratic (fused) Gromov-Wasserstein objective that compares intra-view structure (Vayer et al. 2019; Chen et al. 2020)? How does the method differ from prior OT-based EEG transfer (e.g., Chambon et al. 2018), and how are gradients backpropagated through Eq. (8)?
  4. [future work] Baseline Benchmarking: Can the authors benchmark TopoMontage against public checkpoints of established EEG foundation models (e.g., BIOT, REVE) on the standard 4-class BNCI2014 001 benchmark with paired statistical tests across subjects?
Reviewer_Pn1JOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes an EEG self-supervised learning framework designed to address one specific challenge: How can we learn transferable EEG representations when datasets use different electrode configurations, channel counts, spatial layouts, and reference schemes? The key idea is to treat different EEG montages as different, incomplete observations of the same underlying neural activity, and explicitly learn correspondences between those observations.

Critical strengths

None

Critical weaknesses

The manuscript must be completely AI-generated without verification. All experimental results are placeholders with unrealistic numbers, based on the table captions: “Numbers are simulated drafting results for manuscript preparation and should be replaced by actual experimental results before submission.”

Questions for authors

None

Comments on the paper

1 comment

Captions of Tables 1, 2, 3, 4, 5, 6, 7

“Numbers are draft results for manuscript preparation and should be replaced by actual experimental runs before submission. Best results are in bold.” and similar statements

#6 of https://researchonresearch.blog/2026/10/20/self-confessed-fabrications-and-rise-of-tbds/

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.