acceptodds
Under review as a conference paper at ICLR 2027

GeMoT: Encoding Class Knowledge in Geometry and Moments for Source-Free EEG Domain Adaptation

Abstract

Source-free domain adaptation has garnered increasing attention for cross-subject electroencephalography (EEG) decoding, as it eliminates the need for source-data sharing and labeled target calibration. However, strong source-domain performance does not guarantee transferability, and inter-subject variability can undermine source-model predictions. We propose a Geometry and Moment Transfer (GeMoT) framework that jointly encodes source class knowledge in classifier geometry and feature moments to improve source-model transferability and guide target-domain adaptation. For source-domain model training, we introduce a neural-collapse-inspired rule that replaces class-specific classifier weights with pooled source class prototypes, providing a structured initialization for transfer. Using these prototypes, we construct a memory bank of cross-channel means and variances to preserve complementary class statistics. Target features compensated by these feature moments are matched to corresponding classifier weights via cosine similarity to generate pseudo-labels. We design a neighbourhood attraction loss for local prediction consistency, a progressively weighted compensation-guided supervision loss to train the target-domain model with these pseudo-labels, and an inter-class dispersal loss to reduce prediction overlap across pseudo-classes. The adaptation process also incorporates a label-free early-stopping criterion based on the normalized entropy of the pseudo-label class distribution. Offline leave-one-subject-out evaluation covers three backbones and three EEG datasets spanning P300 detection, motor imagery, and vigilance monitoring. With one fixed backbone per dataset, mean accuracies reach 69.36%, 72.50%, and 82.45% on P300, BNCI2014002, and SADT, respectively. These results demonstrate that coupling classifier geometry with class-specific statistical memory provides a robust, generalizable paradigm for source-free cross-subject EEG decoding.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Source-Free Domain Adaptation, Cross-Subject EEG Decoding, Neural Collapse, Feature Moments

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

This paper investigates source-free domain adaptation for cross-subject EEG classification and proposes GeMoT, a framework combining classifier geometry with class-specific feature statistics. The method introduces neural-collapse-guided parameter replacement (NCPR) to align classifier weights with source class prototypes and utilizes stored feature moments for target pseudo-label generation. Target adaptation is performed through three complementary losses, together with entropy-based model selection. Experiments on three EEG datasets, P300, BNCI2014002, and SADT, demonstrate improvements in average classification accuracy over the compared baselines.

Critical strengths

The paper addresses source-free cross-subject EEG adaptation, a relevant and challenging problem arising from inter-subject distribution shifts and restricted access to source data. The proposed framework explores the integration of classifier geometry and class-specific feature statistics to facilitate knowledge transfer under source-free constraints.

Critical weaknesses
  1. The paper does not provide sufficient evidence to establish that the underutilization of class-specific feature statistics is a key limitation of existing SFDA methods. Furthermore, Equations (10) to (13) reduce class-specific feature moments to scalar coefficients, whose effects on cosine-based pseudo-label assignment may vanish under scale-invariant normalization. Although the authors acknowledge this limitation, the paper lacks rigorous theoretical analysis and sufficiently controlled empirical evidence to substantiate the effectiveness, necessity, and independent contribution of this mechanism to domain adaptation.
  2. The ablation studies are insufficient to establish the independent contributions and necessity of the proposed NCPR and moment compensation mechanisms. Moreover, the experimental comparisons are substantially outdated, as all nine baselines were published in or before 2023. The use of different backbone architectures and training configurations further complicates a fair assessment of the proposed method. These limitations weaken the empirical validation and make it difficult to assess the effectiveness and competitiveness of the proposed framework against recent advances.
Questions for authors
  1. Given the scale invariance of cosine similarity, how does the proposed moment compensation mechanism in Equations (10) to (13) influence pseudo-label assignment under different normalization strategies?
  2. Why are the experimental comparisons limited to baselines published in or before 2023? Can the authors provide comparisons with more recent SFDA methods or justify their exclusion?
  3. How were the dataset-specific and backbone-specific thresholds for entropy-based model selection determined? Were these thresholds selected independently of target-domain labels?
  4. Since the classifier weights continue to be updated after prototype replacement, how does the resulting change in classifier geometry affect its correspondence with the fixed source moment memory during target adaptation?
Reviewer_rwc2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GeMoT, a source-free domain adaptation framework for cross-subject EEG decoding. During source training, the classifier weights are replaced with pooled class prototypes whenever source-validation accuracy reaches a new maximum, and the cross-channel means and variances of these prototypes are stored as a moment memory. During target adaptation, pseudo-labels are generated by cosine similarity between moment-compensated target features and the classifier weights. The model is then trained with three losses, and adaptation is stopped using the normalized entropy of the pseudo-label distribution. Under leave-one-subject-out evaluation on three binary EEG datasets, GeMoT reports higher mean accuracy than nine baselines, along with component ablations and a hyperparameter sensitivity analysis.

Critical strengths
  • In SFDA, deciding when to stop adaptation and which checkpoint to use without target labels is a practical problem that is often overlooked. This paper explicitly proposes a label-free stopping criterion and acknowledges that it relies on a balanced-class assumption. This is a meaningful attempt to show that the reported performance does not result from post hoc selection using target labels.
Critical weaknesses
  • The paper motivates the method by removing the need for labeled target calibration. However, the target data in the experiments are cue-based trials for which labels already exist; the labels are simply hidden. Moreover, adaptation uses the complete set of target-subject trials. If this much target data is collected, labels can be obtained along with it, so the realistic alternative is supervised fine-tuning with a small number of labeled trials, and this comparison is missing. In addition, the P300 data were balanced from their original imbalanced ratio to 1:1 using target labels, which may favor the evaluation given that the stopping criterion prefers balanced pseudo-label distributions. Apart from SADT, there is no evaluation in settings where label-free adaptation is genuinely needed, so the paper does not convincingly establish why SFDA is necessary for EEG. If the practical rationale for the problem setting does not hold, the significance of the performance gains is itself weakened; I consider this the most consequential weakness.

  • The encoders are limited to lightweight models with embedding dimensions of 16–40, so it is unclear whether the proposed techniques hold for EEG foundation models or larger architectures. All datasets are binary classification tasks with only 8–14 subjects, so how the neural-collapse geometry and class-wise scalar compensation behave for C > 2 is not tested. Furthermore, the backbone used in the main results coincides with the best-performing backbone for each dataset; with a single fixed CNN backbone, GeMoT falls below DAN and CDAN on P300. As a result, the claim of a “generalizable paradigm” is insufficiently supported.

Questions for authors

Clarification

  • How does GeMoT compare with supervised fine-tuning of the source model using 5–20 labeled target trials per class? If a short calibration outperforms GeMoT, what is the practical need for the proposed method?
  • On what data were the per-dataset backbone, hyperparameters, and early-stopping threshold selected? Could you report results using a single backbone across all datasets?
  • The P300 data were balanced to 1:1 using target labels. How do h-norm-based selection and overall performance change when evaluated at the original imbalanced ratio?

Future work

  • Validation with larger encoders or EEG foundation models, and on multi-class tasks.
  • Evaluation in online or streaming adaptation settings that better reflect real-world use.
Reviewer_96NuOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GeMoT for source-free cross-subject EEG adaptation. During source training, the method replaces classifier weights with class prototypes and retains their mean and variance statistics. During target adaptation, these statistics are used to generate pseudo-labels, which guide model optimization through three complementary losses. Experiments on three EEG datasets show higher average accuracy than the selected baselines.

Critical strengths

I find the novelty of this work rather limited. The method mainly combines existing techniques, while the effectiveness of the proposed moment compensation mechanism remains unclear. I also do not think the current experiments provide sufficient evidence to support the claimed performance advantages.

Critical weaknesses

My main concern is the limited methodological novelty of this work. The proposed framework largely combines existing techniques, including class prototypes, mean and variance statistics, pseudo-labeling, and consistency regularization. The NCPR module is also closely related to existing prototype-based classifier alignment methods. In my view, the paper does not offer a clear methodological contribution beyond combining these standard techniques, and its connection to neural collapse is not sufficiently established. I am also not convinced by the effectiveness of the proposed moment compensation mechanism. In Eqs. (10)–(13), the stored statistics are reduced to a single scalar used to scale the target features. As the authors acknowledge in Appendix B.5, the effect of positive scalar scaling can be cancelled by subsequent normalization under certain conditions. I would like to see more direct experimental evidence showing that these coefficients meaningfully affect pseudo-label assignments. Moreover, the ablation results in Table 4 show that the NCPR configuration already achieves performance close to the full method, leaving the actual contribution of moment compensation unclear. I also find the experimental evaluation insufficient. The most recent baseline in Table 2 is LSFT (2023). Although the paper cites related studies from 2024 and 2026, it does not explain why these more recent methods were excluded from the comparisons. In addition, the improvements over the strongest baselines are not consistently statistically significant. I also think the paper would benefit from an overall framework diagram, as the current figures do not clearly illustrate how the different components interact.

Questions for authors

(1)What do the authors consider the main methodological contribution of this work beyond combining existing techniques? (2)To what extent does moment compensation actually change pseudo-label assignments? How would the results change if the compensation coefficients were replaced with constants? (3)Can the authors compare GeMoT with simpler prototype-based methods using the same backbone and training settings, and explain why more recent related methods were not included?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.