acceptodds
Under review as a conference paper at ICLR 2027

EvoBrain: Continual Learning of EEG Foundation Models Across Heterogeneous BCI Tasks

Abstract

Electroencephalography (EEG) is the predominant modality for non-invasive Brain-Computer Interfaces (BCIs) due to its high temporal resolution. However, conventional decoding approaches rely on fragmented, task-specific architectures, limiting scalability across heterogeneous BCI tasks. EEG foundation models pre-trained on large-scale corpora offer a promising solution for universal brain decoding, yet current post-training strategies remain dominated by task-isolated fine-tuning. This static paradigm restricts knowledge transfer across tasks and incurs computational and storage costs that grow linearly with the number of downstream tasks.In this paper, we formulate downstream adaptation of EEG foundation models as a cross-task continual learning problem and propose EvoBrain, a dynamic task-aware post-training framework that enables a unified foundation model to continually generalize across diverse EEG tasks. EvoBrain addresses the plasticity-stability trade-off through two complementary components. Neuro-Spectral Task Normalization (NSN) facilitates adaptation to distributional and neuro-spectral shifts by aligning task statistics and recalibrating spectral responses. Response-Affinity Distillation (RAD), together with time-dependent replay, preserves historical task knowledge while promoting selective transfer between spectrally compatible tasks, thereby mitigating catastrophic forgetting. Extensive evaluations on six distinct BCI tasks demonstrate that our approach consistently surpasses state-of-the-art methods across diverse foundation backbones, effectively balancing the plasticity-stability trade-off.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG Decoding, Foundation Model, Continual Learning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_vmp3OpenReview ↗
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
AuthorsOpenReview ↗

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

W1&Q1: 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.

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

BackboneBCIC-IV-2aPhysioNet-MIBCIC2020-3FACEDMumtazISRUCAverage B-ACC(Mi/MNM_i/M_N)RFR (%)
CBraMod0.489±0.0390.545±0.0190.464±0.0050.516±0.0380.895±0.0220.739±0.0130.636±0.005/0.608±0.0064.3%

Additional Order 2

BackboneMumtazPhysioNet-MIFACEDBCIC2020-3BCIC-IV-2aISRUCAverage B-ACC(Mi/MNM_i/M_N)RFR (%)
CBraMod0.846±0.0130.546±0.0050.513±0.0160.476±0.0300.486±0.0490.726±0.0170.622±0.011/0.599±0.0053.7%

Additional Order 3

BackboneMumtazPhysioNet-MIFACEDBCIC2020-3ISRUCBCIC-IV-2aAverage B-ACC(Mi/MNM_i/M_N)RFR (%)
CBraMod0.859±0.0240.540±0.0080.511±0.0190.497±0.0260.705±0.0230.497±0.0420.624±0.009/0.601±0.0053.7%

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.

BackboneISRUCPhysioNet-MIFACEDBCIC2020-3MumtazBCIC-IV-2aTUEVOriginal-six Average B-ACCSeven-task Average B-ACC
CBraMod0.708±0.0300.571±0.0230.489±0.0120.465±0.0190.874±0.0140.494±0.0080.555±0.0110.6000.594±0.004

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.

AuthorsOpenReview ↗

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:

  1. Fixed historical teacher. PhysioNet-MI RGD continues to use its original post-adaptation checkpoint M2.
  2. 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 strategyPhysioNet B-ACC at M2PhysioNet B-ACC at M4BWT relative to M2 (pp)Change from M3 (pp)FACED B-ACC at M4BCIC2020-3 B-ACC at M4Four-task Average B-ACC
Fixed M2 teacher0.57480.5831±0.0010+0.8280±0.0963−0.0525±0.09630.4894±0.00490.5018±0.02390.5711±0.0060
Updated M3 teacher0.57480.5796±0.0016+0.4774±0.1598−0.4031±0.15980.4901±0.00150.5084±0.01690.5695±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.

AuthorsOpenReview ↗

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 ruleAverage B-ACCRFR (%)
Unweighted alignment0.555310.51
Intermediate-feature cosine0.59733.35
Input-band-power cosine0.552610.80
Learned spectral affinity (SAA)0.60682.41

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.

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

This paper proposes EvoBrain, a continual learning framework for adapting pretrained EEG foundation models across heterogeneous brain-computer interface (BCI) tasks. Unlike conventional task-specific fine-tuning, EvoBrain aims to incrementally learn new EEG decoding tasks while preserving previously acquired knowledge within a shared foundation backbone. The proposed framework consists of two main components. Neuro-Spectral Task Normalization (NSN) combines prototype-guided temporal normalization and task-conditioned spectral residual modulation to accommodate task-specific distribution shifts. Response-Affinity Distillation (RAD) incorporates response-geometry distillation and spectral-affinity alignment to mitigate catastrophic forgetting and encourage knowledge transfer between related tasks. A time-dependent replay mechanism further supports the preservation of historical task knowledge.

Critical strengths
  • Most EEG foundation model studies focus on pretraining or isolated downstream adaptation. This paper instead considers sequential adaptation across distinct BCI paradigms, including sleep staging, motor imagery, emotion recognition, imagined speech, and mental-disorder classification. This setting is practically meaningful and introduces substantial distributional and label-space heterogeneity.

  • The framework is evaluated with four pretrained EEG foundation models and three task orders. The inclusion of component-wise ablations and comparisons with alternative adaptation strategies strengthens the empirical investigation.

Critical weaknesses
  • Although the authors emphasize that only one backbone is retained for inference, the training procedure appears to require access to previous-task data and historical model checkpoints. Consequently, the claimed scalability and memory efficiency are not fully supported without a quantitative accounting of training-time storage, computational costs, and replay requirements.

  • The current ablations demonstrate the usefulness of several components, but do not fully establish whether the proposed spectral-affinity mechanism provides a meaningful advantage over simpler similarity measures, conventional feature alignment, or standard replay-based distillation. In particular, the spectral-affinity score is computed from learned band-gating parameters. It is unclear whether this learned similarity reliably reflects actual cross-task transferability.

Questions for authors
  • How does the proposed learned spectral affinity compare with cosine similarity of intermediate representations, unweighted statistical alignment, or other simple task-similarity measures? Can the authors demonstrate that higher spectral affinity is empirically associated with more beneficial cross-task transfer?

  • Are all competing continual learning methods evaluated with the same historical data access, replay budget, backbone update policy, and training resources? How does EvoBrain perform when the replay budget is substantially reduced?

  • How many historical teacher checkpoints are retained during continual adaptation? Please report the total training-time storage requirements and computational overhead as the number of tasks increases.

5 comments
AuthorsOpenReview ↗

Response to Reviewer puPL[1/N]

We sincerely thank the reviewer for the careful reading and for recognizing the practical importance of sequential adaptation across heterogeneous BCI paradigms, as well as the empirical evaluation across four pretrained EEG foundation models and three task orders. We also appreciate your constructive suggestions regarding training-time storage and computational costs, the effectiveness of learned spectral affinity, fair baseline comparisons, and performance under reduced replay budgets. We have addressed these points as detailed below.

Main Comments

Q1: Although the authors emphasize that only one backbone is retained for inference, the training procedure appears to require access to previous-task data and historical model checkpoints. Consequently, the claimed scalability and memory efficiency are not fully supported without a quantitative accounting of training-time storage, computational costs, and replay requirements.

A1: Thank you for highlighting the need to quantify the training-time resource requirements of continual adaptation. EvoBrain enables a single shared EEG backbone to acquire heterogeneous tasks sequentially while retaining previously learned capabilities, with only one backbone retained for inference. However, continual adaptation requires access to historical samples and teacher checkpoints during training.

Following your suggestion, we conducted reduced-replay experiments and quantified training time under full replay and cumulative additional storage requirements using CBraMod on Order 1. Additional continual-learning (CL) storage includes historical teacher checkpoints and retained historical data, excluding the current backbone, task-specific classifiers, and current-task data. Historical-data storage is calculated according to the replay fraction, while teacher-checkpoint storage remains unchanged across replay settings.

TaskTraining time (100%, s/epoch)B-ACC (100%)B-ACC (10%)B-ACC (1%)Additional CL storage (100%, GiB)Additional CL storage (10%, GiB)Additional CL storage (1%, GiB)
ISRUC44.810.73520.73060.73610.000.000.00
PhysioNet-MI46.600.54690.54410.505523.962.410.26
FACED54.430.48000.48390.480727.772.810.31
BCIC2020-334.410.47600.46530.432032.753.320.38
Mumtaz30.120.88270.85940.854334.513.520.42
BCIC-IV-2a15.020.52000.50780.519135.573.640.45
Average225.38 (total)0.60680.59850.587925.762.620.30

With 10% and 1% replay, the final mean B-ACC is 0.83 and 1.88 percentage points below full replay, respectively. At the sixth task, additional CL storage decreases from 35.57 GiB to 3.64 GiB and 0.45 GiB, corresponding to reductions of 89.8% and 98.7%. Each of these storage totals includes five historical teacher checkpoints occupying 93.65 MiB in total. These results quantify the performance–storage trade-off and show that, in the evaluated setting, reduced replay maintains performance close to full replay while substantially lowering training-time storage requirements.

We thank the reviewer again for this valuable suggestion. The manuscript summarizes the replay-budget trade-off and resource accounting in Appendix F (pp. 34–35), Tables 17–18 (p. 35).

AuthorsOpenReview ↗

Response to Reviewer puPL[2/N]

Q2: The current ablations demonstrate the usefulness of several components, but do not fully establish whether the proposed spectral-affinity mechanism provides a meaningful advantage over simpler similarity measures, conventional feature alignment, or standard replay-based distillation. In particular, the spectral-affinity score is computed from learned band-gating parameters. It is unclear whether this learned similarity reliably reflects actual cross-task transferability.

A2: Thank you for your insightful comments and valuable suggestions on our spectral-affinity alignment module. Following your recommendation, we added two controlled comparisons using CBraMod, Order 1, and seed 432:

  1. Unweighted alignment: conventional statistical feature alignment with a constant cross-task alignment weight of 1.
  2. Intermediate-feature cosine: we compute the cosine similarity between the new-task and historical-teacher feature vectors at each layer and average the resulting values across layers to obtain the cross-task alignment weight.

All three conditions use the same alignment loss, NSN, RGD, replay budget, and training schedule, changing only the alignment weight. Each dataset is evaluated immediately after its initial adaptation MiM_i and after all six tasks MNM_N. Average B-ACC denotes the six-task mean at MNM_N; RFR measures the relative decrease from the corresponding mean at MiM_i.

ConditionISRUC (Mi/MNM_i/M_N)PhysioNet-MI (Mi/MNM_i/M_N)FACED (Mi/MNM_i/M_N)BCIC2020-3 (Mi/MNM_i/M_N)Mumtaz (Mi/MNM_i/M_N)BCIC-IV-2a (Mi/MNM_i/M_N)Average B-ACCRFR (%)
Unweighted alignment0.7505/0.57440.5951/0.57400.4899/0.46240.5173/0.49330.8635/0.72090.5069/0.50690.555310.51
Intermediate-feature cosine0.7505/0.73810.5737/0.53860.4930/0.45480.4893/0.47730.8883/0.86180.5130/0.51300.59733.35
SAA0.7505/0.73520.5725/0.54690.4972/0.48000.4880/0.47600.9024/0.88270.5200/0.52000.60682.41

SAA achieves the highest final mean B-ACC and the lowest RFR, followed by intermediate-feature cosine and then unweighted alignment. SAA improves final mean B-ACC by 5.15 and 0.95 percentage points over unweighted alignment and feature cosine, respectively. Mean post-adaptation B-ACC is similar across the three conditions, indicating that the main difference lies in retaining knowledge through subsequent adaptations.
The w/o SAA ablation in Table 2 (p. 8) further isolates the contribution of SAA beyond replay-based RGD by removing the entire SAA loss while retaining NSN and RGD. For CBraMod, adding SAA increases final mean B-ACC from 0.592 to 0.607 and reduces RFR from 4.2% to 2.4%.

Interestingly, unweighted alignment produces a particularly large decline on ISRUC. We therefore examine its retention trajectory across the six-task sequence.

ConditionISRUCAfter PhysioNet-MIAfter FACEDAfter BCIC2020-3After MumtazAfter BCIC-IV-2a
Unweighted alignment0.75050.73060.62700.52480.53900.5744
Intermediate-feature cosine0.75050.73660.74440.73960.73690.7381
SAA0.75050.73780.74380.73840.72810.7352

Unweighted alignment causes its largest ISRUC declines after learning FACED and BCIC2020-3, by 10.36 and 10.23 percentage points, respectively. Subsequent training fails to fully reverse this performance loss, leaving the final ISRUC B-ACC 17.62 percentage points below its initial value. With SAA, this decrease is limited to 1.54 percentage points.

Figure 7 (p. 22) provides a complementary interpretation. The learned spectral modulation for ISRUC enhances low-frequency components while suppressing the β band. In contrast, FACED, which involves video-elicited emotions, and BCIC2020-3, which involves imagined speech, exhibit the opposite modulation pattern: β-band enhancement and low-frequency suppression. This contrast may reflect different task-specific spectral requirements, given the prominence of slow-wave activity in non-REM sleep and the task-dependent modulation of neural rhythms during speech imagery and visually elicited emotional processing.

Together with the retention trajectories, these opposing modulation patterns suggest that imposing equally strong statistical alignment on spectrally dissimilar tasks can introduce conflicting constraints and increase the risk of negative transfer. SAA regulates alignment strength according to spectral compatibility to reduce such conflicts and preserve representations that support positive cross-task transfer.

We sincerely thank the reviewer for this insightful suggestion. It prompted a more direct validation of our core design and exposed a consequential failure mode of unweighted alignment, substantially strengthening the technical evidence and completeness of our work. The aggregate alignment-weighting comparison is reported in Appendix E.6, Table 6 (p. 24).

AuthorsOpenReview ↗

Response to Reviewer puPL[3/N]

Q3: How does the proposed learned spectral affinity compare with cosine similarity of intermediate representations, unweighted statistical alignment, or other simple task-similarity measures? Can the authors demonstrate that higher spectral affinity is empirically associated with more beneficial cross-task transfer?

A3: Thank you for highlighting this valuable point. Our full-sequence experiments support learned spectral affinity as a useful compatibility signal for selective alignment.

The comparisons with alternative similarity measures and the ISRUC retention trajectories are provided in our response to Q2.

To make the affinity calculation explicit, we provide the learned six-band gating coefficients of the first three tasks at the validation-selected FACED checkpoint. Positive and negative coefficients indicate spectral enhancement and suppression, respectively.

Taskδθασβγ
ISRUC0.0265460.0563220.0890910.013785−0.0858510.169273
PhysioNet-MI−0.001763−0.000754−0.005324−0.0028780.352117−0.236840
FACED−0.001448−0.003818−0.000168−0.0971790.916060−0.005784

The negative β gate for ISRUC and positive β gates for PhysioNet-MI and FACED are consistent with their β-band suppression and enhancement patterns in Figure 7 (p. 22).

Following Eqs. (18)–(19), spectral affinity combines the cosine similarities of the enhancement and suppression components:

\frac{\cos(\mathbf{g}_t^+,\mathbf{g}_i^+)+\cos(\mathbf{g}_t^-,\mathbf{g}_i^-)}{2} \right), \qquad \mathbf{g}^+=\operatorname{ReLU}(\mathbf{g}),\quad \mathbf{g}^-=\operatorname{ReLU}(-\mathbf{g}).$$ The resulting affinities are: $$\alpha_s(\mathrm{FACED},\mathrm{ISRUC}) =\frac{0+0}{2} =\mathbf{0},$$$$\alpha_s(\mathrm{FACED},\mathrm{Physionet}) =\frac{1+0.07173}{2} \approx\mathbf{0.5359}.$$ At the validation-selected FACED checkpoint, the mean affinity weights for ISRUC and PhysioNet-MI are 0 and 0.5359, respectively. The learned weighting therefore nearly suppresses ISRUC alignment while retaining substantial alignment with PhysioNet-MI, consistent with their different spectral-modulation patterns in Figure 7 (p. 22). With this weighting, final mean B-ACC reaches 0.6068, compared with 0.5973 for feature cosine and 0.5553 for unweighted alignment, while RFR decreases to 2.41%. Final ISRUC B-ACC is also preserved at 0.7352, compared with 0.5744 under unweighted alignment. **These results demonstrate the practical value of affinity-guided alignment for improving retention and reducing harmful cross-task interference.** They support using affinity as an alignment-compatibility signal. The affinity formulation appears in Section 2.5, Eqs. (18)–(19) (pp. 6–7); the checkpoint affinity values and aggregate comparison are summarized in Appendix E.6 (p. 24). --- > #### **Q4: Are all competing continual learning methods evaluated with the same historical data access, replay budget, backbone update policy, and training resources? How does EvoBrain perform when the replay budget is substantially reduced?** **A4: Thank you for emphasizing fair comparison and performance under limited replay.** In our implementation, all replay-based methods share the same historical training pool and replay sampler, while parameter updates follow each method’s design. Following your suggestion, we evaluated EvoBrain with replay budgets of **1%, 10%, 50%, and 100%**. Here, the replay budget denotes the proportion of each historical task’s training data randomly retained for replay. The retained subsets remain fixed and are reused throughout subsequent continual-learning stages. **All other settings remain unchanged.** All experiments use CBraMod, Order 1, and seed 432. | Replay budget | Retained replay samples | Final Average B-ACC | RFR (%) | B-ACC change from 100% (pp) | | --- | --- | --- | --- | --- | |100%|28,754 | 0.6068 | 2.41 | 0 | |50%|14,376 | 0.6072 | 2.41 | +0.04 | |10%|2,874 | 0.5985 | 2.88 | −0.83 | |1%|285|0.5879|3.64| −1.88 | **Retaining only 50% of historical training data yields nearly identical final mean B-ACC to full replay (0.6072 versus 0.6068).** Reducing the replay budget to 10% and 1% lowers final mean B-ACC by 0.83 and 1.88 percentage points, respectively. Even with only 1% retained data, EvoBrain achieves a final mean B-ACC of 0.5879 and an RFR of 3.64%. **These results show that EvoBrain can substantially reduce its historical-data storage requirements while limiting performance degradation in the evaluated setting.** **We thank the reviewer again for this valuable suggestion.** We hope these additional experiments clarify the impact of the replay budget on EvoBrain’s performance and address your concerns. The comparison settings are clarified in Section 3.4 (pp. 9–10), and the replay-budget results are summarized in Appendix F, Table 17 (p. 35).

Comments on the paper

9 comments
AuthorsOpenReview ↗

Response to Reviewer wAu7[1/N]

We thank the reviewer for recognizing the value of our problem formulation, broad empirical evaluation, and EEG-specific spectral-affinity mechanism. We also appreciate the constructive suggestions on storage accounting, quantitative baseline comparisons, run-to-run variability, affinity stability, and task-order robustness. We address each point below.

Main Comments

W1: Although the authors emphasize that only one backbone is retained for inference, the training procedure appears to require access to previous-task data and historical model checkpoints. Consequently, the claimed scalability and memory efficiency are not fully supported without a quantitative accounting of training-time storage, computational costs, and replay requirements.

A1: Thank you for your careful reading and valuable suggestions. EvoBrain consolidates the backbone used for inference, while continual adaptation requires retained historical samples and teacher checkpoints. We address your concerns through the following three points.

1. Motivation and inference-time storage

“A core motivational claim — avoiding linear storage growth — is overstated and not consistent with the method as described.”

Thank you for this valuable suggestion. We narrow the motivation to backbone storage at deployment: task-isolated fine-tuning requires a separate adapted backbone for each task, whereas EvoBrain supports previously learned tasks through one shared backbone. The corresponding replacements are as follows.

(1) Abstract (p. 1)

This static paradigm restricts knowledge transfer across tasks and requires a separately fine-tuned backbone to be stored for each downstream task at deployment.

(2) Introduction (p. 2)

Maintaining a separate adapted backbone for each supported task makes deployment backbone storage grow linearly with task count.

2. Continual-learning storage and computational costs

“The storage/overhead argument should be restated precisely (training-time vs inference-time; rehearsal assumption)”

(1) Additional replay-budget experiments

Following Reviewer puPL’s suggestion, we conducted additional replay-budget experiments using CBraMod, Order 1, and seed 432. A replay budget of 1% or 10% restricts replay to a fixed, randomly selected subset containing that fraction of each historical task’s training samples. Current-task training data, training schedules, replay batch sizes, and replay frequency remain unchanged.

(2) Resource requirements and performance

Learning task ii uses i−1i-1 historical teachers. Additional CL storage includes historical teacher checkpoints and retained historical data; total training storage additionally includes the current backbone, task classifiers, and current-task data. Historical-data storage is calculated proportionally to the replay budget.

Values in grouped columns correspond to replay budgets of 100%, 10%, and 1%, respectively. Average B-ACC denotes final performance after all six tasks.

TaskHistorical teachers, count (MiB)Average B-ACC, 100% / 10% / 1%Additional CL storage (GiB), 100% / 10% / 1%Total training storage (GiB), 100% / 10% / 1%Time at 100% (s/epoch)FLOPs at 100% (TFLOPs/epoch)
ISRUC0 (0.00)0.7352 / 0.7306 / 0.73610.00 / 0.00 / 0.0024.04 / 24.04 / 24.0444.81390.822
PhysioNet-MI1 (18.73)0.5469 / 0.5441 / 0.505523.96 / 2.41 / 0.2628.00 / 6.46 / 4.3046.6095.139
FACED2 (37.46)0.4800 / 0.4839 / 0.480727.77 / 2.81 / 0.3133.46 / 8.51 / 6.0154.43126.738
BCIC2020-33 (56.19)0.4760 / 0.4653 / 0.432032.75 / 3.32 / 0.3835.31 / 5.88 / 2.9434.4170.739
Mumtaz4 (74.92)0.8827 / 0.8594 / 0.854334.51 / 3.52 / 0.4236.44 / 5.45 / 2.3530.1259.245
BCIC-IV-2a5 (93.65)0.5200 / 0.5078 / 0.519135.57 / 3.64 / 0.4537.19 / 5.26 / 2.0715.0231.190
Average—0.6068 / 0.5985 / 0.5879————

With 10% and 1% replay, final mean B-ACC decreases by only 0.83 and 1.88 percentage points, respectively. At the sixth task, 1% replay reduces additional CL storage from 35.57 GiB to 0.45 GiB, a 98.7% reduction, and total training storage from 37.19 GiB to 2.07 GiB. These results support effective continual adaptation with a small historical-data budget.

Each replay-enabled update uses one historical teacher forward pass shared by RGD and SAA, so the teacher-forward count per update remains constant as task history grows. In the 1% replay experiment, test-set evaluation takes 0.20–3.46 seconds per task, using the shared backbone and task head without historical teachers or replay. The manuscript’s replay-budget and training-resource summaries are presented in Appendix F (pp. 34–35), Tables 17–18 (p. 35).

AuthorsOpenReview ↗

Response to Reviewer wAu7[2/N]

3. Fair comparison with rehearsal-free and rehearsal-based methods

“the comparison to rehearsal-free CL baselines made fair”

We use the same backbone initialization, data splits, task order, training epochs, and checkpoint-selection criterion across methods, while preserving their native learning protocols. EWC and LwF use no historical samples during adaptation, whereas rehearsal-based baselines share EvoBrain’s historical-data pool and replay-sampling settings. We explicitly account for EvoBrain’s retained samples and teacher checkpoints in the training-time storage and computational costs reported above. This allows performance differences between rehearsal-free and rehearsal-based methods to be assessed alongside their respective resource requirements. These comparison settings are clarified in Section 3.4 (pp. 9–10).

We sincerely thank the reviewer for the suggestions on deployment storage, training requirements, and fair baseline comparisons. We hope these clarifications and additional results address your concerns and make the scope and practical value of EvoBrain clearer.


W2: The key baseline comparison lacks numbers and variance ... Fig. 3 [contains] “no numeric table and no significance testing” ... “the main Tables 1–2 report no seeds and no standard deviations” ... “Without numeric comparisons and run-to-run variance, I cannot tell whether EvoBrain's advantage over established CL methods, the RFR improvements, and the ablation deltas exceed noise.”

A2: Thank you for this valuable suggestion. We agree. During the rebuttal, we expanded the main, ablation, and baseline-comparison studies with repeated runs using random seeds 4321, 43, and 4, and summarized these alongside the original experiments. We report mean ± standard deviation in the following three tables.

We have updated the manuscript to include the complete experimental results and their variability across runs. The numerical results are provided in Appendix E.8 (Tables 7–9, pp. 26–28), Appendix E.9 (Table 10, pp. 29–30), and Appendix E.10 (Table 11, pp. 31–32).

1. Overview performance

“the main Tables 1–2 report no seeds and no standard deviations”

We repeated the main study across four EEG foundation backbones and three task orders using random seeds 4321, 43, and 4. For clearer presentation, we report B-ACC (mean ± SD) in the tables below. Complete repeated-run results are detailed in Appendix E.8 (pp. 26–28), Tables 7–9.

Task orderBackboneISRUCPhysioNetFACEDBCIC2020-3MumtazBCIC-IV-2aAverage B-ACC(Mi/MNM_i/M_N)RFR (%)
Ord.1CBraMod0.718±0.0130.562±0.0130.486±0.0060.467±0.0260.876±0.0330.503±0.0150.620±0.003/0.602±0.0082.9%
Ord.1EEGMamba0.740±0.0060.543±0.0050.457±0.0040.479±0.0150.860±0.0090.483±0.0010.592±0.002/0.594±0.003-0.3%
Ord.1CSBrain0.746±0.0150.535±0.0110.471±0.0140.500±0.0070.911±0.0060.507±0.0090.624±0.008/0.612±0.0062.0%
Ord.1DeeperBrain0.721±0.0080.576±0.0030.526±0.0080.416±0.0180.910±0.0080.570±0.0140.630±0.004/0.620±0.0021.6%
Ord.2CBraMod0.726±0.0220.584±0.0020.489±0.0130.501±0.0140.866±0.0250.419±0.0110.612±0.001/0.598±0.0042.3%
Ord.2EEGMamba0.745±0.0120.532±0.0090.437±0.0070.476±0.0060.826±0.0180.446±0.0060.589±0.001/0.577±0.0012.0%
Ord.2CSBrain0.737±0.0160.526±0.0220.484±0.0210.482±0.0200.893±0.0100.427±0.0250.613±0.009/0.591±0.0093.5%
Ord.2DeeperBrain0.690±0.0200.574±0.0130.522±0.0110.429±0.0150.908±0.0110.551±0.0160.628±0.007/0.612±0.0052.5%
Ord.3CBraMod0.716±0.0190.534±0.0380.466±0.0020.478±0.0350.895±0.0110.495±0.0160.617±0.005/0.597±0.0083.2%
Ord.3EEGMamba0.734±0.0070.541±0.0130.435±0.0110.446±0.0060.850±0.0100.453±0.0180.583±0.002/0.577±0.0041.0%
Ord.3CSBrain0.737±0.0120.534±0.0140.459±0.0100.481±0.0080.906±0.0070.473±0.0060.613±0.007/0.598±0.0052.3%
Ord.3DeeperBrain0.708±0.0180.573±0.0080.482±0.0130.394±0.0200.919±0.0100.573±0.0170.623±0.006/0.608±0.0112.3%
AuthorsOpenReview ↗

Response to Reviewer wAu7[3/N]

2. Ablation study

“the method is often distinguished from ablations by <1 point of average B-ACC”

Following your suggestion, we repeated the ablation study under Order 1 across all four backbones using random seeds 4321, 43, and 4. For clearer presentation, we report B-ACC (mean ± SD) in the tables below. Complete repeated-run ablation results are detailed in Appendix E.9 (pp. 29–30), Table 10.

BackboneConditionISRUCPhysioNetFACEDBCIC2020-3MumtazBCIC-IV-2aAverage B-ACC(Mi/MNM_i/M_N)RFR (%)
CBraModEvoBrain0.718±0.0130.562±0.0130.486±0.0060.467±0.0260.876±0.0330.503±0.0150.620±0.003/0.602±0.0082.9%
CBraModw/o NSN0.730±0.0030.512±0.0030.391±0.0230.355±0.0070.884±0.0110.435±0.0040.584±0.006/0.551±0.0075.6%
CBraModw/o PTN0.735±0.0080.529±0.0120.419±0.0260.342±0.0110.895±0.0050.443±0.0100.584±0.000/0.561±0.0064.0%
CBraModw/o SRM0.668±0.0350.572±0.0060.456±0.0060.464±0.0040.854±0.0310.502±0.0130.619±0.006/0.586±0.0035.3%
CBraModw/o RAD0.653±0.0050.553±0.0090.486±0.0120.450±0.0080.759±0.0050.492±0.0130.612±0.006/0.566±0.0067.6%
CBraModw/o RGD0.686±0.0260.554±0.0180.474±0.0280.434±0.0230.809±0.0210.514±0.0110.618±0.007/0.579±0.0076.4%
CBraModw/o SAA0.650±0.0020.551±0.0270.478±0.0130.438±0.0250.777±0.0590.495±0.0170.613±0.004/0.565±0.0077.9%
CBraModw/o NSN+RAD0.736±0.0080.442±0.0520.360±0.0070.346±0.0120.874±0.0110.435±0.0160.588±0.004/0.532±0.0099.5%
EEGMambaEvoBrain0.740±0.0060.543±0.0050.457±0.0040.479±0.0150.860±0.0090.483±0.0010.592±0.002/0.594±0.003-0.3%
EEGMambaw/o NSN0.748±0.0060.536±0.0160.447±0.0040.475±0.0250.860±0.0020.484±0.0050.592±0.005/0.592±0.0020.1%
EEGMambaw/o PTN0.746±0.0120.531±0.0090.434±0.0090.480±0.0110.862±0.0040.482±0.0030.588±0.001/0.589±0.001-0.2%
EEGMambaw/o SRM0.741±0.0120.531±0.0100.438±0.0100.478±0.0130.862±0.0090.484±0.0070.589±0.006/0.589±0.0040.0%
EEGMambaw/o RAD0.696±0.0120.523±0.0120.443±0.0120.452±0.0130.858±0.0070.490±0.0020.581±0.001/0.577±0.0050.7%
EEGMambaw/o RGD0.702±0.0300.531±0.0040.420±0.0060.468±0.0230.849±0.0110.486±0.0040.585±0.004/0.576±0.0031.5%
EEGMambaw/o SAA0.750±0.0050.533±0.0020.445±0.0060.443±0.0180.855±0.0040.485±0.0040.585±0.005/0.585±0.0030.0%
EEGMambaw/o NSN+RAD0.683±0.0120.530±0.0040.452±0.0070.483±0.0070.848±0.0040.484±0.0030.588±0.004/0.580±0.0031.4%
CSBrainEvoBrain0.746±0.0150.535±0.0110.471±0.0140.500±0.0070.911±0.0060.507±0.0090.624±0.008/0.612±0.0062.0%
CSBrainw/o NSN0.749±0.0130.561±0.0090.444±0.0110.396±0.0060.890±0.0070.516±0.0060.609±0.004/0.593±0.0042.7%
CSBrainw/o PTN0.727±0.0100.560±0.0170.464±0.0090.403±0.0380.895±0.0120.510±0.0090.609±0.004/0.593±0.0082.6%
CSBrainw/o SRM0.743±0.0130.538±0.0130.468±0.0150.480±0.0130.893±0.0170.502±0.0160.617±0.008/0.604±0.0052.1%
CSBrainw/o RAD0.687±0.0420.501±0.0090.438±0.0460.451±0.0120.899±0.0170.489±0.0160.620±0.007/0.578±0.0156.9%
CSBrainw/o RGD0.648±0.0570.515±0.0080.454±0.0230.469±0.0080.895±0.0120.480±0.0090.617±0.006/0.577±0.0126.5%
CSBrainw/o SAA0.735±0.0040.530±0.0110.473±0.0120.459±0.0220.891±0.0100.501±0.0160.618±0.005/0.598±0.0053.2%
CSBrainw/o NSN+RAD0.565±0.1500.530±0.0260.388±0.0560.377±0.0820.899±0.0100.488±0.0240.607±0.013/0.541±0.03310.8%
DeeperBrainEvoBrain0.721±0.0080.576±0.0030.526±0.0080.416±0.0180.910±0.0080.570±0.0140.630±0.004/0.620±0.0021.6%
DeeperBrainw/o NSN0.642±0.0570.557±0.0190.514±0.0100.426±0.0390.879±0.0090.558±0.0090.624±0.006/0.596±0.0054.5%
DeeperBrainw/o PTN0.669±0.0330.570±0.0250.519±0.0110.409±0.0130.897±0.0130.554±0.0140.622±0.001/0.603±0.0073.1%
DeeperBrainw/o SRM0.692±0.0010.567±0.0140.519±0.0080.395±0.0270.896±0.0180.556±0.0100.620±0.004/0.604±0.0042.6%
DeeperBrainw/o RAD0.553±0.0660.565±0.0200.510±0.0040.427±0.0150.911±0.0140.554±0.0330.618±0.009/0.587±0.0105.1%
DeeperBrainw/o RGD0.600±0.0400.557±0.0270.544±0.0080.413±0.0210.884±0.0300.553±0.0180.627±0.005/0.592±0.0035.6%
DeeperBrainw/o SAA0.668±0.0440.562±0.0130.530±0.0060.419±0.0100.894±0.0140.554±0.0100.627±0.005/0.605±0.0083.6%
DeeperBrainw/o NSN+RAD0.578±0.0130.565±0.0110.512±0.0100.428±0.0150.878±0.0130.550±0.0070.630±0.002/0.585±0.0157.1%

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.