acceptodds
Under review as a conference paper at ICLR 2027

WAVE: Optimal-Transport-based Domain-Incremental Learning and Generalization for Cross-Subject EEG Decoding

18 views

Abstract

Brain activity varies substantially across individuals, causing Electroencephalography (EEG) decoding models to degrade when applied to new subjects. Each individual shifts the signal underlying probability measure, and models trained on a fixed population generalize poorly to unseen subjects. We cast cross-subject adaptation as domain-incremental learning (DIL), treating each subject as a domain and adapting sequentially through a replay memory. We introduce WAVE, a probabilistic approach that represents each subject as a hierarchical probability measure over classes and a replay memory as a measure over subjects, which are natural objects in a Wasserstein over Wasserstein (WoW) space. As such, WAVE learns a replay buffer by minimizing an optimal-transport divergence in WoW space and augments data with virtual subjects generated between stored and incoming domains for smoother adaptation and improved generalization. Because the memory stores compressed, synthetic subjects rather than raw recordings, WAVE also avoids retaining privacy-sensitive EEG data. Applied to EEG foundation models, WAVE improves the stability-plasticity-generalization trade-off and matches strong replay baselines without storing real EEG.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Continual Learning, domain incremental learning, EEG decoding, optimal transport, replay memory, domain generalization

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_1JKeOpenReview ↗
Rating
1
Confidence
4
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

This paper points out the challenge of cross-subject EEG decoding and provides a general framework by domain-incremental learning, treating each new subject as a domain and adapting sequentially through replay memory. By comprehensively design the memory construction and the virtual subject generation, this paper gives clear framework on how do sequentially adapting EEG models for upcoming subjects. The evaluation metrics and datasets are also comprehensively defined and presented, and give clear result and analysis.

Critical strengths

Overall, the domain-incremental learning framework proposed in this work is elegant, well-defined, and general for cross-subject EEG decoding. The issue they want to solve is real and well motivated. The framework and the algorithm they propose are well-written and well-designed. The elaboration and the visualization are also well presented. The paper is well structured and friendly to follow. Everything before Section III looks strong to me.

Critical weaknesses

Major issues start to surface in Section III. First, the performance is just not good enough to make it play its role in the field. Second, the fundamental setup on the EEG FM they used is not clearly stated, which leaves the number in danger to represent the real behavior of how EEG FM would perform. The only major impact I can find after reading the section III is about the privacy de-identification potential that can be provided by the virtual subject, but this paper does not quantify and analyze the privacy aspect in the current manuscript.

Questions for authors
  1. In the current section III (Experiments), I don’t see how the author pre-train the EEG FM CBraMod. By the current description in line 347, readers would wonder: is CBraMod just pretrained on the “source subjects” they defined in table 1, without pretraining on any large corpus? If so, this is not a fully pretrained foundation model. This is an empty foundation model pre-trained with a super tiny pre-trained dataset (like N_s in ). The original CBroMod is pretrained on 9000 hours of EEG data. If the CBraMod is pre-trained using the official recipe, or just downloaded from the official release, please state it clearly. I assume this paper use CBraMod pretrained on the official recipe, and ask the following questions.
  2. Please calibrate the performance of Oracle (joint training) with the official performance of CBraMod on their fine-tuned datasets, if they are listed in the official paper, like PhysioNet-MI, SHU-MI, and FACED. If the number does not match, please specify and explain possible cause and difference in the implementation. The performance reported in this paper as Oracle (joint training) should match the number reported in the CBraMod's fine-tuned result.
  3. The current improvement of WAVE, compared to other baselines, like random replay, EWC, and LwF, are just too weak and marginal. Even on FACED, the observed outperfomed margin is not salient. Even though this paper has 4 seeds for each experiment, does each seed also randomly pick the subjects in source, incremental, and generalization domain? If not, the current numbers of subjects in each domain are just too small to make the result stable. In BCI-IV 2a, only 2 in source, 5 in incremental, and 2 in generalization domain. What if the 2 subjects in the generalization just happened to be supreme under WAVE? So, the subject pool should be shuffled to be reported, especially when the performance improvement is marginal.
  4. Only one FM is used. Since the performance margin of WAVE on CBraMod is too small compared to other baselines (random replay, EWC, and LwF), it is likely that such margin may disappear when applied on other EEG FM. To make this paper citable, this method should also be tested on other EEG FMs. This is not to ask the author to pre-train each FM from scratch, but to just download the official release of each EEG FM, and further fine-tune on the small source and incremental domain they define.
  5. How does this method perform on small and task-oriented model? For the SOTA task-oriented models on each of the dataset tested in this manuscript (BCI-IV-2a, PhysioNet-MI, SHU-MI, FACED, ISRUC), can each dataset's SOTA method also be benefited from the proposed domain-incremental learning framework?
  6. Does the privacy of each subject can actually be de-identified via the virtual subject augmentation? If so, this adds value to this paper, as the current performance analysis does not show satisfactory impact for this paper.
Reviewer_fdn8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed a new method based on continuous learning for domain generalisation. The experimental section is based on 5 different datasets where the model CBraMod is fine-tune with the proposed method and other strategy.

Critical strengths
  • The paper propose an interesting method based on optimal transport for incremental learning.
  • The method proposes nice visualisations to illustrate the method
Critical weaknesses
  • This paper has a lot on definition and sometimes it is hard to understand everything. For example, a metric name AAA is defined in the appendix and it is never say in the main paper that it is defined in the appendix. trying to reduce as maximum the number of abbreviation and definition and also defined everything needed to understand the table results would help the reading of the paper.

  • Of what I understood, at each incremental step all the source data are still available. So it is always possible to retrain from scratch everything, or continuing fine-tuning with all subjects. It will of course will be longer but the last step should reach the oracle performances. For all datasets, oracle is a lot higher, it motivates anyone to just retrain every time there are more subjects available rather than doing continual learning. Having this comparison vs the computational time could be interesting.

  • Only CBraMod is used for training. Having an experiment about backbones robustness could strengthen the paper.

Questions for authors
  • How are chosen the weights for subjects?

  • I do not know if CBraMod is used with random weight but in the original paper the performances of CBraMod are above the one in the paper. If only the architecture is used, it makes me wondering what the use if FM pre-trained models give better performance without CL. If it is CBraMod with pre-trained weights, do you think your CL is somehow detrimental for the performances ?

Reviewer_Y63QOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The proposed WAVE method treats sequentially arriving EEG subjects as domains. It maintains a fixed-size replay memory by optimizing synthetic samples with a class-additive, Sinkhorn-based discrepancy, then forms virtual subjects through transport-based interpolation between incoming subjects and stored memory.

Critical strengths

The tackled task is relevant, and combining synthetic replay with distribution-based augmentation is plausible. The experiments cover a broad range of datasets, with additional subject-order permutations. The empirical evaluation reports useful metrics including source retention, incremental retention, generalization, and runtime etc. The work also acknowledges that augmentation is not beneficial on every dataset which are solid observations.

Critical weaknesses

The geometry ablation changes several factors simultaneously and it is unclear of their individual influences. The flat Sinkhorn baseline differs from WAVE in more than subject hierarchy. It also uses optimized soft labels, permits cross-class transport, and reduces the diversity of subject-level augmentation etc. Therefore, this experiment does not demonstrate the benefit of hierarchical subject geometry.

Section 2.1 initializes M_1 by randomly drawing m samples from each source subject. The initial replay memory therefore contains real EEG recordings. The claim of storing only synthetic EEG conflicts with this memory initialization process.

The reported empirical evaluation results support competitiveness of the proposed method in some settings, however, for other settings such as ISRUC, enabling augmentation substantially worsens retention.

Questions for authors

Please see above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.