- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes EvoBrain, a continual learning framework for adapting pretrained EEG foundation models across heterogeneous BCI tasks. It introduces Neuro-Spectral Task Normalization (NSN) to handle task-dependent statistical and spectral shifts, and Response-Affinity Distillation (RAD) to preserve historical representations while encouraging transfer between spectrally similar tasks. The framework is evaluated on six EEG datasets and four pretrained backbones under three task orders, showing relatively low forgetting and improved performance over several continual learning baselines.
Critical strengths
- The paper addresses a relevant and relatively underexplored problem: continual adaptation of EEG foundation models across heterogeneous downstream tasks. The shared-backbone formulation offers an alternative to independently fine-tuning a separate model for each task.
- The method combines EEG-specific spectral adaptation with representation-level knowledge preservation. Evaluations across four backbones, together with ablation studies and additional representation analyses, provide useful evidence for the effectiveness of the proposed components.
Critical weaknesses
- The evaluation of task-order robustness is limited, as all three sequences begin with ISRUC, the largest dataset considered. This setting does not assess whether the method remains stable when substantially larger tasks arrive after several smaller ones. Including alternative initial tasks and larger, commonly used EEG benchmarks such as TUAB and TUEV would provide stronger evidence for robustness across heterogeneous task sequences.
- The proposed RGD preserves the response geometry of historical teacher checkpoints without considering whether the current student has already achieved better performance on previous tasks. This may constrain beneficial representation updates and potentially interfere with positive backward transfer.
- The assumption underlying SAA remains insufficiently validated. Similar spectral-gating patterns do not necessarily imply transferable task representations, and the experiments do not directly demonstrate that spectral affinity predicts positive transfer or that affinity weighting is preferable to unweighted alignment.
- The forward-transfer analysis in Figure 5 relies on target-task classifiers trained retrospectively and paired with earlier backbone checkpoints. Consequently, the reported improvements do not directly establish that prior task learning facilitates subsequent target-task adaptation.
- The claimed efficiency benefits are not fully supported by the reported comparisons. Although the final model shares a single backbone, the training procedure requires historical datasets and task-specific teacher checkpoints, while training-time computation and total storage costs are not quantified.
- The experimental conditions of competing methods are not sufficiently detailed, particularly regarding replay-memory access and resource budgets. This makes it difficult to assess whether the comparisons with existing continual learning approaches are conducted under equivalent conditions.
- Important methodological and experimental details remain insufficiently specified. The computation of global reference statistics from historical task prototypes is unclear, and the main results do not report the number of independent runs or measures of variability, making their reliability difficult to assess.
- The manuscript does not provide a source code repository or code availability statement, limiting independent reproducibility of the proposed framework and experimental comparisons.
Questions for authors
- Why is ISRUC fixed as the first task in all three sequences? Could the authors evaluate alternative task orders, particularly those in which a larger dataset is introduced after several smaller tasks?
- How does RGD account for cases where the current student has already surpassed the historical teacher on a previous task? Could the authors clarify whether preserving the earlier response geometry may constrain positive backward transfer?
- What evidence supports the assumption that spectral-gate similarity reflects cross-task transferability? Could the authors compare SAA with unweighted statistical alignment or examine the relationship between spectral affinity and actual transfer performance?
- Could the authors clarify the interpretation of forward transfer in Figure 5, given that the target-task classifiers are trained retrospectively? A comparison using independently initialized heads and identical target-task adaptation protocols would provide more direct evidence.
- Could the authors quantify the training-time computational and storage costs, including the overhead of historical datasets and teacher checkpoints, to substantiate the claimed scalability benefits?
- Were the competing continual learning methods evaluated under comparable replay-memory access, training budgets, and backbone-update settings? Additional implementation details would help establish the fairness of these comparisons.
- How are the global reference statistics computed from historical task prototypes? Also, how many independent runs were conducted for the main experiments, and could the authors report the corresponding variability?
8 comments
Response to Reviewer vmp3[2/N]
W2&Q2: How does RGD account for cases where the current student has already surpassed the historical teacher on a previous task? Could the authors clarify whether preserving the earlier response geometry may constrain positive backward transfer?
A2: Thank you for this insightful comment. Following your suggestion, we conducted a controlled experiment. Under CBraMod, Order 1, and seed 43, we observed positive backward transfer on PhysioNet-MI after FACED adaptation. Its test B-ACC increased from 0.5748 at M2, the original PhysioNet-MI teacher checkpoint, to 0.5836 at M3, the checkpoint saved after FACED adaptation.
Starting from the same M3 checkpoint, we continued adaptation to the fourth task, BCIC2020-3, and compared two strategies:
- Fixed historical teacher. PhysioNet-MI RGD continues to use its original post-adaptation checkpoint M2.
- Updated teacher. PhysioNet-MI RGD uses checkpoint M3, frozen throughout subsequent training.
We performed paired continuation runs using seeds 43, 432, and 4321. Within each pair, all settings were matched, changing only the teacher used for PhysioNet-MI RGD (M2 or M3). After 20 training epochs, we evaluated the M4 checkpoint selected by the highest BCIC2020-3 validation B-ACC.
| RGD strategy | PhysioNet B-ACC at M2 | PhysioNet B-ACC at M4 | BWT relative to M2 (pp) | Change from M3 (pp) | FACED B-ACC at M4 | BCIC2020-3 B-ACC at M4 | Four-task Average B-ACC |
|---|---|---|---|---|---|---|---|
| Fixed M2 teacher | 0.5748 | 0.5831±0.0010 | +0.8280±0.0963 | −0.0525±0.0963 | 0.4894±0.0049 | 0.5018±0.0239 | 0.5711±0.0060 |
| Updated M3 teacher | 0.5748 | 0.5796±0.0016 | +0.4774±0.1598 | −0.4031±0.1598 | 0.4901±0.0015 | 0.5084±0.0169 | 0.5695±0.0037 |
PhysioNet-MI performance with the fixed teacher remained close to the M3 starting point, decreasing by only 0.0525 percentage points on average while remaining 0.8280 points above M2. This is consistent with the design of RGD, which regularizes pairwise relation matrices of normalized features rather than directly matching classification outputs, allowing representation updates alongside historical-structure preservation. The updated teacher yielded slightly higher mean performance on FACED and BCIC2020-3, but lower four-task Average B-ACC. These findings demonstrate that, preserving the original response geometry does not prevent the student from retaining performance gains beyond its historical teacher. Moreover, the fixed-teacher strategy achieves a better balance between preserving positive backward transfer and maintaining overall performance.
We thank the reviewer again for this valuable suggestion. We hope this controlled comparison clarifies the practical role and limitations of historical-teacher preservation in RGD and addresses your concern. The controlled comparison is reported in Appendix E.12 (p. 34), Table 16.
Response to Reviewer vmp3[3/N]
W3&Q3: What evidence supports the assumption that spectral-gate similarity reflects cross-task transferability? Could the authors compare SAA with unweighted statistical alignment or examine the relationship between spectral affinity and actual transfer performance?
A3: Thank you for this insightful comment. During the rebuttal, we conducted additional experiments comparing SAA with three alternative alignment-weighting strategies using CBraMod, Order 1, and seed 432: (1) unweighted alignment, which assigns a uniform weight of 1; (2) intermediate-feature cosine, which computes weights based on the cosine similarity of intermediate representations; and (3) input-band-power cosine, which computes the nonnegative cosine similarity between normalized six-band input-power vectors. All other experimental settings were kept identical to ensure a fair comparison.
| Alignment-weighting rule | Average B-ACC | RFR (%) |
|---|---|---|
| Unweighted alignment | 0.5553 | 10.51 |
| Intermediate-feature cosine | 0.5973 | 3.35 |
| Input-band-power cosine | 0.5526 | 10.80 |
| Learned spectral affinity (SAA) | 0.6068 | 2.41 |
SAA achieved higher Average B-ACC and lower RFR than the alternative weighting rules in this comparison.
The spectral gates encode task-specific enhancement and suppression preferences learned during adaptation. Their similarity therefore provides a task-conditioned signal for regulating alignment strength. At the validation-selected FACED checkpoint, affinity was 0 for FACED–ISRUC and 0.5359 for FACED–PhysioNet-MI, substantially reducing alignment with ISRUC while retaining alignment with PhysioNet-MI. Consistent with the importance of selective weighting, final ISRUC B-ACC was 0.7352 with SAA, compared with 0.5744 under unweighted alignment.
We additionally tracked affinity during FACED adaptation through paired SAA-on/off experiments across three random seeds. With SAA enabled, the preference for PhysioNet-MI over ISRUC emerged early and persisted across all three runs, providing complementary evidence for the reproducibility of the learned affinity dynamics.
Together, these results support learned spectral affinity as an effective compatibility signal for regulating cross-task alignment and mitigating negative transfer. Notably, unweighted alignment performs substantially worse than removing SAA entirely (w/o SAA), indicating that indiscriminate cross-task alignment can introduce more severe negative transfer than omitting the alignment mechanism altogether. This further highlights the importance of selective, affinity-guided alignment across heterogeneous EEG tasks.
We thank the reviewer again for this valuable suggestion and hope these comparisons and analyses address your concern. The weighting comparisons and affinity dynamics are reported in Appendix E.6 (p. 24), Table 6 and Figure 8.
Response to Reviewer vmp3[1/N]
We sincerely thank the reviewer for recognizing the significance of continual EEG foundation model adaptation, the benefits of our shared-backbone framework, and the supporting experimental evidence. We also appreciate your constructive suggestions regarding task-order robustness, knowledge preservation and transferability, forward-transfer evaluation, computational efficiency, baseline fairness, methodological clarity, and reproducibility. We have carefully addressed these concerns through additional experiments, analyses, and clarifications, as detailed below.
Main Comments
A1: Thank you for this valuable suggestion. Following your suggestion, we conducted two additional sets of experiments to examine performance when large datasets arrive later in the task sequence and further evaluate EvoBrain’s robustness to different task orders.
1. Additional task orders
We evaluated CBraMod under three additional orders, beginning with BCIC-IV-2a or Mumtaz instead of ISRUC. Dataset columns below follow the training sequence, and each dataset cell reports final B-ACC (mean ± SD).
Additional Order 1
Additional Order 2
Additional Order 3
In Additional Orders 2 and 3, we kept the first four tasks in the same order and placed ISRUC sixth and fifth, respectively. Final Average B-ACC was 0.599 and 0.601, suggesting a modest overall difference between these two late-arrival positions. Across all three additional orders, final Average B-ACC ranged from 0.599 to 0.608. These results extend the evaluation beyond ISRUC-first sequences and support effective adaptation when ISRUC arrives after several smaller tasks.
2. Adding TUEV after the original six-task sequence
We additionally introduced TUEV as the seventh task after Order 1. Our processed TUEV split contains 112,491 five-second EEG event segments, including 67,436 training, 15,634 validation, and 29,421 test segments, providing an additional large clinical EEG task at the end of the stream.
Starting from the final checkpoint saved after BCIC-IV-2a adaptation, we continued training on TUEV and evaluated all seven tasks. Final B-ACC results are summarized below.
After TUEV adaptation, Average B-ACC over the original six tasks was 0.600, close to the original Order 1 reference of 0.602. This suggests that learning the additional large clinical task was compatible with retaining overall performance on the previously learned tasks. Together, these experiments provide further evidence of EvoBrain’s robustness to different task orders.
We thank the reviewer again for this valuable suggestion and hope these additional results address your concern. The additional task-order results and TUEV extension are reported in Appendix E.11 (pp. 33–34), Tables 12–15.