- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The authors audit frozen representations of five released EEG-FMs across five canonical BCI benchmarks using linear probing, variance decomposition, t-SNE, and representational similarity analysis. Finds that pretraining strengthens subject identity information without comparable gains in task-related information, revealing subject identity shortcuts as a shared failure mode of prevailing EEG selfsupervised paradigms.
To further explore underlying mechanisms, the authors pretrain both MAE-style and MoCo-style EEG-FMs and revealed several phenomena: 1 In masked reconstruction models, temporal masking is more challenging because predicting EEG signals along the temporal dimension is inherently difficult, whereas spatial masking is relatively easier since EEG channels exhibit strong inter-channel dependencies. 2 In contrastive pretraining, subject identity encoding arises from positive-negative pair construction. Because positive pair are from same subject, negative pair are very likely from difference subjects, therefore furnishing an efficient shortcut for the contrastive objective.
3 Interventions on the above two shortcuts do not promote task decoding. Use MAE with Euclidean Alignment (EA) reconstruction targets does not strengthens task-related information; Using negative samples only from the same subject does not lead to significant improvements in task-related decoding accuracy.
Critical strengths
1 The authors conducted systematic experiments to analyse representations obtained by pretrained EEG-FMs, and found that the latent representations contain more subject identity information than task-related information.
2 The authors further investigate some factors that might influence MAE-style and MoCo-style EEG-FMs. Both pretraining strategies favor subject identity encoding over task-related encoding. For MAE-style model, temporal masking is much harder than spatial masking; for MoCo-style model, introducing more negative samples from same subject can mitigate subject identity encoding, but does not help improve task-related encoding.
The above findings are obtained through comprehensive experiments and analysis, makes the conclusion robust.
Critical weaknesses
-
The current amount of work is insufficient to support solid work. For example, based on the above findings, the authors could further investigate how to improve task-related decoding accuracy, or, given that pretrained models are more sensitive to subject identity information, it would also be interesting to explore whether these models can be applied to tasks such as EEG biometric identification. Both directions are not explored in the current work.
-
The novelty of this work is moderate; some of the contributions listed in the paper such as “Interventions on shortcuts do not promote task encoding,” are more likely negative findings rather than real contributions that can help or inspire future work in this field.
Questions for authors
-
The authors attribute the subject-identity bias in contrastive pretraining primarily to positive-negative pair construction. Have the authors tested alternative strategies, such as treating same task trials from different subjects as positive pairs? Such experiments would directly test the proposed shortcut mechanism and determine whether suppressing subject identity encoding can improve task-related representations.
-
Given that pretraining consistently improves subject-identity decodability, have the authors evaluated whether these representations support robust EEG biometric identification across recording sessions? Such an experiment would help determine whether the learned subject information reflects stable physiological identity, thereby clarifying the practical significance of the main finding.
3 comments
Thank you for the detailed response and clarification. I appreciate the authors’ approach to analyzing and exploring the problem, as well as the clear presentation. However, I remain concerned that the overall novelty is limited. Although the paper explores the potential of pretrained EEG-FMs for subject identification, this investigation is not sufficiently systematic to constitute a substantial contribution. Therefore, I will maintain my rating of 2 (Weak Reject).
Response to Reviewer Rmww
Thank you for your further assessment.
We acknowledge that the biometric identification experiments are not a systematic benchmark. Their purpose is to test whether subject-discriminative information persists across sessions and to illustrate a potential application of the observed identity encoding. The main contribution remains the investigation of why self-supervised EEG pretraining may favor subject identity under MAE- and MoCo-style objectives. We hope these findings offer a useful perspective for designing and evaluating EEG pretraining objectives.
More broadly, this paper does not claim novelty through a new method or a new state-of-the-art result. In our experience, despite substantial progress in EEG decoding, performance gains over compact baselines such as EEGNet are not always consistent under comparable evaluations. Similarly, many recent EEG-FMs report strong performance after full downstream fine-tuning, but such results alone do not establish whether self-supervised pretraining has learned transferable neural representations. In our evaluations, pretraining did not yield reliable task-decoding gains over matched random initialization, even when it substantially enhanced subject-identity encoding.
We believe that, at this stage of the field, understanding why existing pretraining approaches may fall short of their intended goal is a valuable contribution in its own right. We hope our analysis will help the community develop and evaluate objectives that more directly promote transferable task-related representations.
Regardless of your rating or the final decision on this paper, we sincerely appreciate the time and thought you have devoted to reviewing our work. Your comments, particularly your suggestion to explore biometric identification, have helped us clarify our contributions and strengthen the revised manuscript. Thank you again.
Response to Reviewer Rmww
Dear Reviewer,
Thank you for your careful review and constructive comments. We respond point by point below.
1. Can the work be strengthened by improving task decoding or evaluating EEG biometric identification?
In fact, our initial goal was to improve task decoding, but our preliminary attempts did not yield consistent gains. Observing similar limitations across EEG-FMs led us to investigate identity shortcuts. However, identity suppression through adversarial learning or disentanglement, as well as alternative objectives such as Jigsaw Puzzle Solving for spatiotemporal position prediction, did not improve task decoding. These observations led us to question whether task-related EEG representations are too subtle or abstract to be readily learned through current pretraining objectives.
We thank the reviewer for suggesting biometric identification as an application. We have added cross-session biometric identification experiments to strengthen the work (Section 3.3). Specifically, three of the five downstream datasets—BCI-IV-2a, SEED-V, and Tsinghua-RSVP—contain recordings from the same subjects across sessions. In the revised manuscript, we repartitioned these datasets (Appendix B) so that test data come from an independent session, and updated the task and subject decoding results accordingly. We compared frozen pretrained EEG-FMs with linear probes against three fully supervised baselines trained from scratch: EEGNet, SPaRCNet, and EEG Conformer. The results (Section 4.4) show that selected frozen EEG-FMs can match or outperform these baselines with fewer trainable parameters. Together with the matched random-initialization comparisons, these results highlight the tendency toward subject-identity encoding under cross-session evaluation and suggest that suitably pretrained EEG-FMs can serve as parameter-efficient brainprint identifiers.
2. How do the negative intervention results contribute to the field?
We agree that the interventions yield negative results in terms of task improvement. However, they test a practical hypothesis: does suppressing identity information promote task-related representations? Our results suggest that it does not do so automatically.
We decompose the total variation in pretrained representations as: SStotal=SSsubj+SStask+SSinter+SSres and calculate the relative contribution of each component as Pc=SSc/SStotal, where c∈{subj,task,inter,res}. As shown in Appendix K and discussed in Section 4.5, identity suppression and task encoding do not necessarily trade off, reductions in Psubj are largely accompanied by increases in Pres rather than Ptask, potentially reflecting trial-specific, view-consistent variation such as waveform details and artifacts. This finding provides insight into the design of future EEG pretraining objectives: preventing identity shortcuts does not guarantee more useful task representations, so objectives may also need to explicitly promote transferable task-related information. We believe this provides useful guidance for the design and evaluation of EEG-FMs.
3. Would using same-task trials from different subjects as positive pairs suppress identity encoding and improve task representations?
We agree that the reviewer’s proposed strategy could help test the mechanism. However, it requires task labels and therefore moves beyond the label-free self-supervised pretraining studied here.
Furthermore, large EEG corpora available for pretraining often lack complete task labels. Smaller labeled datasets may also have labels that do not align with those of downstream datasets. For example, Stieger21, included in our pretraining corpus, contains left-hand, right-hand, both-hands, and rest classes, whereas BCI-IV-2a contains left-hand, right-hand, feet, and tongue imagery. In fact, we previously attempted supervised contrastive pretraining using only Stieger21. Explicit supervision reduced subject-identity encoding, but linear-probe performance on BCI-IV-2a remained at chance level. However, this setting is a smaller cross-dataset transfer problem, distinct from the large-scale foundation model pretraining considered here, making its findings difficult to compare directly.
We consider the reviewer’s proposed strategy a valuable direction for future work, but have not tested it in this study.
4. Do pretrained representations support robust cross-session EEG biometric identification?
As noted in Response 1, we added cross-session biometric identification experiments. The results suggest that suitably pretrained EEG-FMs have potential for this application.
We hope these clarifications and additional results address your concerns. Thank you again for your time and thoughtful feedback.