acceptodds
Under review as a conference paper at ICLR 2027

Collaborating Shared and Routed Experts for Cross-Subject EEG Generalization

Abstract

Cross-subject EEG generalization is challenging due to substantial heterogeneity across subjects. Existing methods typically learn either a shared subject-invariant model or multiple subject-specialized experts, but these two paradigms fail in complementary ways: the former may over-reduce subject-specific discriminative signals, while the latter may under-reduce transferable structure. We show that their suitability depends on the reducibility cost of branch-specific functions to branch-invariant ones, and we further provide a theory-to-method mapping that instantiates alignment principles in cross-subject EEG learning. Based on this insight, we propose Shared-Routed Expert Alignment (SREA), a collaborative framework that couples a shared expert for reducible invariant functions with routed experts for irreducible subject-specific functions. SREA trains the shared branch with joint embedding over augmented temporal neighbors, the routed branch with prototype-based sparse routing and expert specialization, and both branches with numerically stable mutual-guided reweighting based on cross-branch learnability gaps. Experiments on seven public EEG benchmarks across different tasks show that SREA consistently outperforms state-of-the-art methods and EEG foundation models.

Primary area:
applications to neuroscience & cognitive science
Keywords:
cross-subject decoding, test-time adaptation, mixture of adapters

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

The paper proposes a framework combining mixture of experts and shared experts, along with a theoretical analysis to try to prove this architecture has collaborative mechanism that allows the two to correct each other. Experiments are extensive, on 7 datasets and multiple baselines.

Critical strengths
  1. The experiment amount is big, on multiple baselines and datasets across different tasks.
  2. The theoretical analysis makes effort to demonstrate the generalization ability of combining shared experts and MoE, though there are some flaws in the analysis.
  3. Figures are clear to understand.
Critical weaknesses
  1. The theoretical analysis is right but has great limitations, section 2.1 is the same as [1], proving that ERM has the ability to learn generaliziable neural networks if the Theorem 1 (a) branches holds. However, in the following analysis, you do not discuss when the model aligns with invariant correlations. Noting that in [1], only one of Theorem 1 (a) and (b) holds, not both, but in Assumption 4, you didn't care about which correlations the S or R aligns. Furthuremore, Theorem 3 shows the alignment errors of the sum(ϵS\epsilon_S+ϵR\epsilon_R) in epoch (t) is smaller than epoch 0, but this does not support the architecture has the ability to avoid over-reduction and under-reduction. For example, S or R aligns with spurious correlations, or whether the model relies heavily on shared experts—a situation that does indeed occur. Thus, the conclusion of line 145-149 is too strong and can not directly been drawn from the analysis.

  2. I appreciate the effort of Table 4, but it can sitll not answer the flaws in the theoretical analysis. I only say one: since you use LJELL_{JEL} in the shared branch, which can be regarded as a kind of regularization loss, the Gap of shared branch in Table 4 is samller than the gap of routed branch can only prove the effectiveness of the LJELL_{JEL} regularization.

  3. The qualitative experiments are not enough. And some validation experiments are not conducted like in [1], for example, in [1], the author do some validation experiments on toy datasets to show that sometimes neural networks align with spurious correlations, sometimes with invariant correlations.

[1] Li B, Shen Y, Yang J, et al. Sparse mixture-of-experts are domain generalizable learners, ICLR 2023.

Questions for authors

See above.

Reviewer_eH8EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Shared-Routed Expert Alignment (SREA) for cross-subject EEG generalization. The framework combines a shared expert for learning subject-invariant representations with routed experts for modeling subject-dependent patterns. The two branches are jointly optimized using a mutual-guided reweighting strategy based on their relative sample-wise losses. The authors also provide a theoretical analysis using algorithmic alignment to motivate the design.Experiments are conducted on seven EEG datasets covering motor imagery, sleep stage detection, and auditory attention decoding. The reported results show consistent improvements over several task-specific baselines and EEG foundation models.

Critical strengths
  1. The paper evaluates the framework on seven datasets from three EEG decoding tasks. I appreciate this relatively broad evaluation, as cross-subject generalization is challenging and may behave differently across tasks. The reported improvements suggest that the proposed framework could be useful in practice.
  2. The motivation for preserving both subject-invariant and subject-dependent information is reasonable. The shared-routed architecture provides an intuitive way to address these two aspects, and the ablation results indicate that combining the branches is beneficial.
Critical weaknesses
  1. My main concern is whether the paper provides a sufficiently substantial contribution beyond existing shared-expert and mixture-of-experts approaches.The use of shared and specialized experts is not new, and closely related ideas have been explored in EEGMoE. Previous works such as GMoE have also investigated expert-based architectures from an algorithmic alignment perspective. Although SREA applies these ideas to cross-subject EEG generalization, the differences from the existing literature are not sufficiently established.The mutual-guided reweighting mechanism appears to be the main methodological contribution. However, its benefit is relatively modest in Table 3. Removing mutual guidance reduces accuracy from 81.7% to 80.8% on KUL and from 58.3% to 57.8% on DTU. This makes it difficult to determine whether the main improvements arise from the proposed collaboration mechanism or simply from combining two complementary branches. I also have concerns about the explanation of how the routed experts generalize to unseen subjects. The subject-level specialization loss explicitly encourages consistent expert assignments within each training subject, but the evaluation concerns subjects that are never observed during training. The paper does not provide sufficient evidence explaining how the learned specialization transfers to these unseen subjects.
  2. The theoretical analysis is an important part of the paper, but I have concerns about the proof of Theorem 3. In Appendix A.7, Lemma 2 interprets the weighted cross-entropy objective as a weighted knowledge distillation objective. However, the weighting factor depends on the predictions of the branch being optimized. Consequently, the entropy-related term treated as constant in Eq. (29) is generally not independent of the optimization variables. The subsequent gradient updates in Eqs. (30)–(31) therefore do not appear to follow directly from the proposed objective. Regarding the experiments, the baseline evaluation procedure is not sufficiently transparent. It is unclear whether all baseline results were independently reproduced or taken from previous publications. Moreover, the statistical significance claims require clarification. For example, Table 2 reports a statistically significant improvement on AVED despite a difference of only 0.1 percentage points between SREA and CBraMod.
Questions for authors

Q1. How does SREA differ fundamentally from existing shared/specialized expert frameworks such as EEGMoE and expert-based domain generalization methods such as GMoE? In particular, given the relatively small improvements from mutual-guided learning in Table 3, could the authors compare the proposed exponential reweighting with simpler alternatives, such as fixed weighting or other loss-based sample-weighting strategies?

Q2. Eq. (7) encourages samples from the same training subject to be consistently assigned to particular experts. However, the evaluation follows a leave-one-subject-out setting. What prevents these experts from primarily specializing in the observed training subjects? Could the authors analyze the expert assignments for source and unseen target subjects, and provide evidence that the routed branch captures transferable subject-dependent patterns rather than training-subject-specific correlations?

Q3. In Eq. (29), the weighting factor depends on the shared expert's predictions, yet the entropy-related term is treated as constant.How is this dependence handled in the derivation? Do Eqs. (30)–(31) represent the actual gradients of the training objectives in Eqs. (10)–(12)? If not, under what additional assumptions does the non-increasing alignment error guarantee in Theorem 3 hold?

Q4. In Table 2(c), the improvement on AVED is only 0.1 percentage points (53.8% to 53.9%), but it is marked as statistically significant.Could the authors clarify whether the paired t-tests were performed across independent subjects or EEG windows? Please provide the subject-level paired results, standard deviations, and exact p-values for this comparison.

Q5. Were the baseline results in Tables 1 and 2 obtained by independently running the methods, or were some taken from previously published papers? For independently reproduced baselines, were the same preprocessing procedures, subject-wise splits, and model selection protocols used? Additionally, under the leave-one-subject-out evaluation, how were the validation sets constructed for early stopping?

Reviewer_q19eOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses the challenge of cross-subject generalization. The authors propose SREA that uses a shared expert with routed experts, making use of the benefits of both.

Critical strengths

The paper is well-written. The proposed SREA framework is well-motivated to tackle the issue of inter-subject variability. The mutual-guiding reweighting mechanism is also very interesting idea to correct the two branches (expert and general) by using one another.

Critical weaknesses

Although a nice framework to tackle inter-subject variability, I find the novelty extremely limited. There are several works that use the idea of routing experts in EEG (for example doi:10.1109/TNSRE.2023.3337802). In addition, there is no comparison in the paper to prove that this method is better than all the previously introduced ones.

Although I do like the theoretical aspects of the paper with the various theorems, these do not constitute contributions, but rather mathematically framed observations already known in the EEG community.

High CKA in the comparison section between the two branches show high similarity between generalists and experts branch which weakens the claim of the method.

Questions for authors

How did the authors choose M=5? Any ablation on this choice? How sensitive is performance of this parameter? It also seems that the larger the M, the greater separability.

Can the authors report routing confidence? What happens if the unseen subject falls outside any known expert?

How does a temporal augmentation can approximate subject distribution shifts? Have the authors tried other methods as well?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.