acceptodds
Withdrawn from ICLR 2027

Unraveling Brain Signals through Expert Collaboration: Mechanism-Guided Routing for Heterogeneous EEG Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

EEG decoding across heterogeneous paradigms requires models to capture distinct yet complementary signal characteristics. Motor imagery, error-related potentials, and steady-state visual evoked potentials place different demands on temporal, spectral, and spatial representations, making it challenging for a single representation to capture the cues relevant to each paradigm. We propose MGNet, a mechanism-guided expert-routing framework that enables adaptive collaboration among specialized experts while preserving their complementary representations. MGNet combines a gated causal temporal model with three experts: a temporal module, a spectral dynamic graph module, and a Riemannian covariance module. A lightweight router and a Manhattan-distance-based fusion mechanism adaptively integrate expert representations to accommodate paradigm-dependent EEG characteristics. Experiments on three public benchmarks—BCI Competition IV-2a, BCI Challenge, and MAMEM—show that MGNet consistently outperforms the evaluated state-of-the-art baselines. Ablation studies further reveal paradigm-dependent expert contributions consistent with the signal characteristics motivating their design. These results support mechanism-guided specialization and adaptive expert collaboration as an effective approach to multi-paradigm EEG decoding.

Primary area:
learning on time series and dynamical systems
Keywords:
EEG decoding, time series classification, adaptive expert routing, Riemannian learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_T2PBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes MGNet, an EEG classification framework combining three specialized experts for temporal, spectral-spatial, and Riemannian covariance representations. A lightweight router dynamically weights these experts, while Manhattan-distance-based attention fuses their outputs. Evaluated on three BCI datasets (BCI IV-2a, BCI Challenge, and MAMEM), MGNet outperforms the evaluated baselines. The authors also investigate computational efficiency, robustness, and expert contributions through ablation studies.

Critical strengths

1.Matching expert types to known EEG mechanisms is a sensible, interpretable idea. 2. The comparison includes classical paradigm-specific pipelines (FBCSP, xDAWN, CCA/TRCA) with paired tests, which is good practice. 3. The paper reports results that don't favor it, including the router collapse in Table 6 and no_router beating the full model, and it is upfront about the limits of the router analysis.

Critical weaknesses
  1. The main motivation is never tested. The paper argues that one shared encoder suffers interference across paradigms, but every main result trains a separate model per subject on a single dataset, so the model never sees more than one paradigm. The only pooled multi-paradigm experiment (Table 6) shows the router collapsing onto one expert (max weight 0.997).

  2. The ablations don't support the full design.

a) On MAMEM, several reduced variants beat the full model (59.18): no_router (62.64), no_A (61.45), and no_C (64.27). Expert B alone gets 65.18

b) On BCI Challenge, no_router (81.33), no_fusion (81.04), and no_B (80.37) also beat the full model (79.77).

c) The full model is best only on BCI2a.

With these results what exactly is the motivation behind the design. These results do not support the design choices authors made which make the choices seem unncessary?

  1. The router doesn't really route. Average weights are about 1/3 per expert on BCI2a and MAMEM, which amounts to doing nothing. On BCI Challenge the weight goes 0.999 to Expert A, which is collapse, and it chooses A even though B alone scores better there (76.26 vs 74.16).

  2. The baseline comparison is questionable

a) MAtt (Pan et al., NeurIPS 2022 is miscited in the paper as Congedo 2017 / Huang & Van Gool 2017) evaluated on these same three datasets, and the paper says it follows MAtt's preprocessing. MAtt reported about 74.7% on BCI2a and 65.5% on MAMEM, higher than MGNet's numbers. Here MAtt is listed at 62.96% and 54.54%. Please explain this gap or report MAtt's published numbers.

b) DeepConvNet is at 35.6% on BCI2a which is near chance. Is it possible the moel was undertrained? Or is it just a lack of the model's capabilities?

  1. The BCI Challenge dataset is cited as Chavarriaga & Millán (2010). It is actually the Kaggle BCI Challenge @ NER 2015 P300-speller ErrP data.

  2. Fig. 2 places tangent-space mapping after fusion, while Section 3.4 places it before. Please clarify.

  3. Novelty over existing multi-branch and MoE EEG models is incremental.

Questions for authors

See weakness

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

The paper proposes MGNet, an EEG decoding framework with temporal, spectral dynamic-graph, and Riemannian covariance experts. A shared gated causal convolutional module extracts initial features, a lightweight router weights the experts, and Manhattan-distance fusion combines their representations. The evaluation covers motor imagery, error-related potentials, and steady-state visual evoked potentials. Table 1 reports 69.87% accuracy on BCI2a, 79.77% AUC on BCIcha, and 59.18% accuracy on MAMEM.

Critical strengths
  1. The expert design has a sensible motivation, temporal patterns, spectral and spatial relationships, and covariance geometry capture different aspects of EEG signals. Preserving these cues in dedicated branches is an intuitive design.
  2. The evaluation is broad and transparent. The paper covers three EEG paradigms and includes neural and classical baselines, component ablations, and noise and missing-channel tests.
  3. The authors also report unfavorable findings, including better results without the router and routing collapse during joint training. These disclosures help clarify the method’s limitations.
Critical weaknesses
  1. The incremental contribution over earlier EEG expert models needs clearer justification. EEG-based MoE modeling has precedents, including NeurIPT [1] and CogMoE [2]. The paper discusses EvoMoE, but its relation to these additional works remains unclear. A comparison with a generic expert model of similar capacity would help isolate this contribution.
  2. The benefit of the central router is inconsistent. In Table 4, removing the router improves MAMEM accuracy from 59.18% to 62.64% and BCIcha AUC from 79.77% to 81.33%, while reducing BCI2a accuracy from 69.87% to 66.74%. The authors acknowledge this result, but it remains unclear when dynamic routing is preferable to simpler fusion and whether its added complexity is justified.
  3. The main evaluation does not directly establish reduced interference during heterogeneous joint learning. Appendix B.3 states that the main experiments train, validate, and test separately for each subject. This supports using the same architecture across paradigms, but does not directly demonstrate reduced cross-paradigm interference within a jointly trained model. Table 6 compares the original router with balance regularization in a separate joint-learning setting, where the original router collapses. Matched shared-encoder and fixed-fusion controls in that setting are still needed to assess this motivation. [1] NeurIPT: Foundation Model for Neural Interfaces. NeurIPS 2025. [2] CogMoE: Signal-Quality-Guided Multimodal MoE for Cognitive Load Prediction. ICLR 2026.
Questions for authors
  1. What is the main contribution beyond NeurIPT, CogMoE, and the EvoMoE work already discussed? Which results distinguish the benefits of mechanism-specific experts from those of additional branches or model capacity?
  2. Why does removing the router improve results on two datasets? Could you compare against fixed expert weights and report multiple independent training seeds for the full and no-router variants?
  3. Under the joint-training setup in Table 6, does MGNet outperform a shared encoder or fixed fusion? Please clarify data mixing, handling of different channel configurations.
Reviewer_mRzQOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes MGNet, an EEG decoder combining a gated causal temporal front end with temporal, spectral-dynamic graph, and Riemannian covariance experts. A sample-dependent router weights the expert outputs, which are projected into a common space and fused using Manhattan-distance attention. Evaluation covers motor imagery on BCI2a, error-related potential detection on BCIcha, and SSVEP classification on MAMEM. The paper reports improvements over the evaluated neural baselines, comparisons with classical pipelines, inference measurements, corruption tests, and expert/routing analyses.

Critical strengths
  1. The architecture uses relevant EEG priors, and evaluating the same architectural design on motor imagery, ErrP, and SSVEP is useful. Including classical pipelines alongside neural baselines makes the empirical comparison more informative.

  2. The reported main results improve over the evaluated neural baselines: 69.87% accuracy on BCI2a, 79.77% AUC on BCIcha, and 59.18% accuracy on MAMEM. Table 2 also provides paired subject-level tests for the classical comparisons and acknowledges the non-significant comparison with CCA.

Critical weaknesses
  1. The paper motivates a shared model for heterogeneous paradigms, but Appendix B.3 trains models independently for each subject and dataset. These experiments do not test the claimed reduction of cross-paradigm interference.
  2. Table 6 compares routing with and without balance regularization, without a matched shared-encoder or fixed-ensemble baseline. The near-collapsed router and missing per-task results leave the main heterogeneous-learning claim unresolved.
  3. Removing the router improves MAMEM accuracy from 59.18% to 62.64% and BCIcha AUC from 79.77% to 81.33% (Table 4). The benefit of a central proposed component is therefore not established.
  4. Removing Expert C raises MAMEM accuracy from 59.18% to 64.27%. This calls for a clearer demonstration that the full architecture offers an advantage over simpler task-appropriate configurations.
  5. The architecture combines established temporal, graph, covariance, and gating components. Without capacity- and budget-matched multi-branch controls, the results do not establish what the proposed mechanism-guided organization contributes beyond those ingredients.
  6. Figure 4 tests corruption only on the full model; Table 4 compares fusion rules on clean data. Neither establishes that Manhattan fusion is more robust than alternative fusion rules under identical perturbations.
  7. Tables 1 and 4 omit uncertainty for neural-baseline and component results. The paired tests provided for classical methods do not establish the reliability of the routing and fusion improvements.
  8. Forward-pass timings on prepared segments do not establish acquisition-to-decision latency or continuous-stream behavior. Causality of the temporal block alone is insufficient evidence for the broader online-system framing.
Questions for authors
  1. What parameters are actually shared in Table 6, and how are the differing channel counts, subjects, and Riemannian references handled? Please provide per-task results and a matched comparison with a shared encoder and a fixed expert ensemble under that same joint-learning protocol.

  2. How does the fusion handle an isolated corrupted expert given its self-attention term and the averaging in Eq. 7? Under the existing corruption protocol, how do Manhattan, dot-product, Euclidean-distance, and fixed fusion compare?

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

Related papers

Discussion (0)

Sign in to comment.