acceptodds
Under review as a conference paper at ICLR 2027

NeuroAccord: Time–Space–Frequency Guided Continual Learning for Large EEG Models

Abstract

EEG foundation models learn transferable representations for diverse decoding tasks. As EEG datasets and paradigms expand, continual learning is needed to acquire new capabilities while retaining performance on earlier tasks. Existing approaches mainly rely on two perspectives as protecting important model parameters and replaying representative historical samples. We observe that protected parameters gradually lose importance for their original tasks as the shared model evolves. Meanwhile, fixed replay buffers cover a diminishing share of historical EEG knowledge amid heterogeneity across tasks and subjects. These observations call for a reference that remains meaningful as parameter roles and sample coverage change. We introduce NeuroAccord, a Time–Space–Frequency (TSF)-guided framework linking physiologically motivated EEG descriptors to foundation-model features. It constructs a TSF basis from current-task descriptors and frozen previous-model features, transfers the teacher's relations among examples, and preserves TSF coupling during adaptation. Across stages, accumulated TSF-related directions constrain feature drift without replaying earlier recordings. In task-incremental experiments spanning nine datasets, six paradigms, and four EEG foundation models, NeuroAccord achieves the highest average anytime accuracy and backward transfer and the lowest forgetting among 18 baselines. These results establish state-of-the-art performance among the evaluated methods and a favorable plasticity–stability balance. Ablation, robustness, and continued-learning analyses further support this balance. More broadly, the findings show the value of grounding evolving model representations in EEG signal structure to accumulate new knowledge while retaining earlier decoding capabilities.

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

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_L9eROpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EEG foundation models are pretrained once and then need to learn new decoding tasks over time. Figure 1 makes two observations: parameters protected by importance-based methods lose their importance as the shared model changes, and fixed replay buffers cover less of the history as tasks and subjects accumulate. In both cases the reference for what to preserve lives inside the model or the buffer, so it moves.

NeuroAccord moves that reference outside the model, to physiologically motivated time-space-frequency (TSF) descriptors of the EEG, and adds three objectives (Sec. 3.3-3.4): relational distillation on the TSF basis (TSFRD), alignment of the orientation of the feature-TSF coupling (COA), and historical knowledge anchoring (HKA), which accumulates TSF directions across stages (Q_t = orth([Q_{t-1}, E_t])) and penalizes feature drift inside that span. Evaluation is task-incremental with task identity available at test time (App. C.1), on nine datasets, six paradigms, four EEG foundation models and 18 baselines. NeuroAccord reports the best average anytime accuracy, backward transfer and forgetting on all four backbones (Table 2).

Critical strengths

S1. The paper shows that TSFRD and COA are not redundant instead of assuming it. App. B.2 gives explicit counterexamples: O = -I_r gives L_TSFRD = 0 with L_COA = 2, and a common shift H(t) = H(t-1) + 1a^T gives L_COA = 0 while changing the cosine relations. Table 3 on CBraMod is consistent with HKA doing most of the retention. Removing it raises forgetting from 4.39% to 8.65% and lowers AAA from 61.95 to 59.61, the largest drop among the three objectives.

Critical weaknesses

W1. The setup is not described in enough detail to reproduce the results or to judge the comparison as fair. No replay-buffer size or memory budget is given for any of the six sample-side baselines (App. C.3.2), while NeuroAccord itself keeps a frozen copy of the stage-start backbone (Sec. 3.2) plus the basis Q_t; the abstract presents "without replaying earlier recordings" as an advantage without accounting for either side's storage. App. C.4 says the "data splits" are fixed but never defines them, and the words subject-independent and cross-subject do not appear. This matters because the motivation is subject heterogeneity (Figure 1 IV), and a subject-overlapping split would make retention much easier. The optimizer, learning rate, epochs and batch size are not given, and the rank r of the TSF basis (Eq. 4) has no value. App. C.4 refers to "default values" of lambda_R, lambda_C and lambda_H without stating them. Figure 5(a-c) marks a "Best" setting for CBraMod and EEGMamba, and the CBraMod point matches Table 2, which suggests the weights were picked on the reported metric; no values are given for CSBrain or CodeBrain.

W2. Several comparisons cannot be separated from noise or from weak baselines. Table 2 has one number per cell with no variance, while Sec. 4.5 reports NeuroAccord's own AAA standard deviation as 0.93-1.32% across task orders. In Figure 4(b), the experiment that isolates the main idea, TSF beats the best alternative reference by 0.39 AAA on CBraMod (58.39 vs 58.00 for Weights), within the 0.38-0.68% seed standard deviation the paper reports for its own runs; the EEGMamba margin is larger (58.40 vs 56.97). Several forgetting-reduction baselines also forget more than doing nothing. On CodeBrain, LwF (13.75) and SI (13.94) are above Naive CL (13.66), and on EEGMamba, ER (12.22) is nearly double Naive CL (6.86). No tuning protocol is given for any baseline.

W3. The relation between HKA and gradient-projection methods is not discussed, so it is hard to tell what HKA adds. GPM, TRGP and SGP are among the baselines (Table 2), and App. C.3.1 describes GPM as extracting principal directions from activations by SVD and accumulating them in a memory basis. HKA's rule Q_t = orth([Q_{t-1}, E_t]) (Eq. 10) accumulates a basis in the same way. Figure 4(b) varies the source of the basis, and its Latent PCA reference resembles a GPM-style activation basis (55.61 vs 58.39 AAA on CBraMod), but App. C.4 does not say whether the swap applies to U_t, to E_t, or to both. The other difference, penalizing feature drift on current inputs instead of projecting gradients away from the stored span, is not tested. GPM is weak in this setting (on CBraMod, FOR 16.33 against 20.95 for Naive CL), and with no tuning protocol it is unclear whether that reflects the method or its threshold. A variant that projects gradients away from Q_{t-1} in place of the HKA penalty would separate the two.

Questions for authors

Q1. What replay-buffer size was used for ER, DER, DER++, A-GEM, OCAR and CSReL, and how much storage does NeuroAccord need at stage t (frozen teacher plus Q_t)? A matched-storage comparison would show whether the gain comes from the TSF reference or from a larger retention budget.

Q2. For each of the nine datasets, are the train and evaluation partitions subject-disjoint, or session-disjoint within subject? Given the motivation in Figure 1 (IV), this changes how the retention results should be read.

Q3. On CodeBrain, LwF and SI forget more than Naive CL, and EWC's AAA (53.65) is below Naive CL's (54.76). How were the baselines' regularization strengths chosen, and on which split?

Q4. HKA adds up to r directions per stage (Eq. 10), so q_t grows toward min(d, T*r). Once q_t approaches d, the penalty covers all feature drift on current-task inputs, which may limit plasticity. Could you report r, the actual growth of q_t over the nine stages, and AAA/FOR as a function of stage?

6 comments
AuthorsOpenReview ↗

Response to Reviewer L9eR (1/3)

Thank you for your time and detailed comments on the experimental setup. Below, we provide the requested settings and controlled comparisons to clarify reproducibility, storage costs, and HKA's contribution.

W1. Reproducibility, storage fairness, and hyperparameter selection.

Response.{\color{green}\mathbf{Response.}} We detail the storage budget, training settings, and selection protocol below.

Storage. Each replay baseline stores at most 256 examples in total across tasks. For CBraMod at stage 9, the accounting includes NeuroAccord's frozen teacher and historical basis:

MethodStored EEG examplesAdditional state countedMiB
NeuroAccord0Frozen teacher + Q8Q_818.6485
ER256EEG, labels, routing40.7085
DER256EEG, logits, routing40.7109
DER++256EEG, labels, logits, routing40.7129
A-GEM256EEG, labels, routing40.7085
OCAR256Replay buffer, curvature factors and inverses165.9095
CSReL256Replay buffer, selection model and candidate state226.9289

NeuroAccord's teacher and historical basis occupy 18.6485 MiB. Current-stage bases, optimizer state, and activations are outside this subtotal. CSReL's figure includes its selection-stage auxiliary state. This accounting makes the cost of retaining the teacher explicit, alongside the storage saved by keeping no historical EEG. Q1 reports results for three replay capacities.

Training configuration. All four backbones use AdamW, 50 epochs per task, and TSF rank r=3r=3. Initial learning rates, batch sizes, and final coefficients are:

BackboneBackbone LRHead LRBatch(λR,λC,λH)(\lambda_R,\lambda_C,\lambda_H)
CBraMod10−410^{-4}5×10−45\times10^{-4}64(0.5, 0.25, 0.5)
CSBrain5×10−45\times10^{-4}5×10−45\times10^{-4}64(0.5, 0.375, 0.5)
EEGMamba10−410^{-4}5×10−45\times10^{-4}32(0.75, 0.25, 0.5)
CodeBrain10−410^{-4}10−410^{-4}64(0.25, 0.125, 0.75)

The coefficient candidates are

undefined
AuthorsOpenReview ↗

Response to Reviewer L9eR (2/3)

W3. What does HKA contribute beyond accumulated bases and gradient projection?

Response.{\color{green}\mathbf{Response.}} You correctly identify the shared use of an accumulated basis. NeuroAccord uses EEG signal structure to select historical directions and directly penalizes feature drift along them. GPM [1] derives layerwise activation bases and projects gradients. SGP [2] scales updates by subspace importance, and TRGP [3] uses task relationships to guide transfer. HKA compares the current model's features with those of a frozen teacher along the TSF-associated directions.

The replacement experiment keeps the TSF historical basis QQ, TSFRD, COA, and all other training settings fixed. It removes HKA or replaces it with a GPM-, SGP-, or TRGP-style update rule:

ConfigurationCBraMod AAA ↑ / FOR ↓EEGMamba AAA ↑ / FOR ↓
Full NeuroAccord61.95 / 4.3959.01 / 2.43
Remove HKA59.61 / 8.6557.26 / 4.28
Replace HKA with GPM57.42 / 10.7956.56 / 6.43
Replace HKA with SGP57.34 / 11.5756.79 / 6.07
Replace HKA with TRGP56.84 / 10.8755.70 / 5.47

Compared with the best replacement on each metric, HKA improves AAA by 4.53 pp on CBraMod and 2.22 pp on EEGMamba, and reduces FOR by 6.40 and 3.04 pp, respectively. Removing HKA also worsens both metrics. With the basis and other objectives held fixed, these results support the benefit of the feature-drift penalty.

Equation (11) also clarifies the mechanism. Let P=QQ⊤P=QQ^{\top}, with qq historical directions, mm feature tokens, and teacher variance normalizer ν\nu. Then

=\frac{2}{mq\nu}(F-F^{\mathrm{teacher}})P.$$ HKA responds to the feature mismatch produced by the current model and is optimized jointly with the task loss. Gradient projection and scaling act directly on the update direction. These operations impose different constraints during learning. In Figure 4(b), only the source of the historical basis changes. TSFRD and COA are disabled, so $U_t$ is unused. Latent PCA and TSF share the same feature-drift penalty, isolating the effect of basis construction: | Reference in Figure 4(b) | CBraMod AAA ↑ / FOR ↓ | EEGMamba AAA ↑ / FOR ↓ | | --- | --- | --- | | Weights | 58.00 / 10.38 | 56.75 / 12.57 | | Gradients | 55.96 / 12.17 | 56.97 / 9.28 | | Latent PCA | 55.61 / 10.05 | 56.05 / 11.32 | | TSF | 58.39 / 8.71 | 58.40 / 8.57 | Figure 4(b) supports the TSF reference, and the replacement experiment supports HKA's use of that reference to preserve features. [1] Gradient Projection Memory for Continual Learning. [2] Continual Learning with Scaled Gradient Projection. [3] TRGP: Trust Region Gradient Projection for Continual Learning. --- > #### **Q1. Replay budget and teacher-plus-basis storage at stage t.** ${\color{green}\mathbf{Response.}}$ All six replay methods store at most 256 preprocessed EEG examples across tasks. During task $t$, NeuroAccord keeps the frozen previous-stage teacher and $Q_{t-1}$, then updates the basis to $Q_t$ after training. Their FP32 storage cost is $$S_t^{\mathrm{teacher+history}}=S_{\mathrm{teacher}}+4d q_{t-1}\quad\text{bytes}.$$ At stage 9, the CBraMod teacher occupies 18.6302 MiB and the $200\times24$ historical basis occupies 0.0183 MiB, for a total of 18.6485 MiB: | Method | Stored EEG examples | Additional state counted | MiB | | --- | --- | --- | --- | | NeuroAccord | 0 | Frozen teacher + $Q_8$ | 18.6485 | | ER | 256 | EEG, labels, routing | 40.7085 | | DER | 256 | EEG, logits, routing | 40.7109 | | DER++ | 256 | EEG, labels, logits, routing | 40.7129 | | A-GEM | 256 | EEG, labels, routing | 40.7085 | | OCAR | 256 | Replay buffer, curvature factors and inverses | 165.9095 | | CSReL | 256 | Replay buffer, selection model and candidate state | 226.9289 | The table counts the listed retained and auxiliary states. It excludes the shared training model, optimizer, and activations. The current $200\times3$ TSF basis adds approximately 0.0023 MiB. CSReL's figure includes its selection-stage auxiliary state. The CBraMod results across replay capacities are: | Replay examples | ER AAA / FOR | DER++ AAA / FOR | A-GEM AAA / FOR | CSReL AAA / FOR | | --- | --- | --- | --- | --- | | 128 | 53.67 / 15.66 | 54.49 / 15.76 | 54.91 / 12.64 | 57.71 / 13.43 | | 256 | 54.69 / 16.70 | 55.50 / 15.54 | 55.88 / 8.75 | 58.54 / 12.86 | | 512 | 55.04 / 14.64 | 56.30 / 12.34 | 56.64 / 9.07 | 60.33 / 8.95 | With no replay, NeuroAccord achieves **61.95 AAA / 4.39 FOR**. The best evaluated 512-example replay configuration, CSReL, reaches 60.33/8.95. NeuroAccord therefore gains 1.62 pp AAA and lowers FOR by 4.56 pp. Its advantage holds across the tested replay capacities, with a smaller teacher-plus-history storage cost than the listed replay states. ---
AuthorsOpenReview ↗

Response to Reviewer L9eR (3/3)

Q2. Which datasets use subject-disjoint evaluation?

Response.{\color{green}\mathbf{Response.}} Eight datasets use subject-disjoint partitions, as listed below:

DatasetPartition unitTrainValidationTest
BCIC-IV-2aSubject1–56–78–9
PhysioNet-MISubject1–7071–8990–109
Mumtaz2016Subject1–4344–5253–62
Mental ArithmeticSubject1–2829–3233–36
BCIC2020-3Non-subject indices†1–6061–7071–80
BCIC-UpperLimbSubject1–1011–1213–15
eldBETASubject1–7576–8081–100
SHU-MISubject1–1516–2021–25
FACEDSubject1–8081–100101–123

† BCIC2020-3 has 15 participants. Its 60/10/10 allocation uses non-subject indices. The other eight datasets use subject-disjoint partitions.

Each subject's sessions, trials, and derived windows stay together in these eight datasets, allowing retention to be evaluated on unseen subjects. The subject-mean results for five datasets are:

DatasetWithin-subject TSF distanceBetween-subject TSF distanceRatioSubject-mean BAcc gain over Naïve CL (pp)
BCIC-IV-2a0.1880.6103.242+7.90
PhysioNet-MI0.4410.9922.252+5.92
BCIC-UpperLimb0.2952.2237.532+1.11
SHU-MI0.6031.0041.665+2.26
eldBETA0.7971.2101.518+1.59

Subject-mean BAcc improves on all five datasets, showing that the retention gains extend to held-out subjects.


Q3. Baseline regularization strengths and the poor CodeBrain results.

Response.{\color{green}\mathbf{Response.}} Configurations are selected by validation performance. The additional LwF, SI, and ER searches use the following candidates and selected values:

Backbone / methodSearch candidatesSelected settingAAA / FOR
CodeBrain / LwFλKD∈{0.1,1,10}\lambda_{\mathrm{KD}}\in\{0.1,1,10\}, T∈{1,2,4}T\in\{1,2,4\}λKD=1, T=1\lambda_{\mathrm{KD}}=1,\ T=157.11 / 12.78
CodeBrain / SIc∈{0.01,0.1,1}c\in\{0.01,0.1,1\}, ϵ∈{0.01,0.1,1}\epsilon\in\{0.01,0.1,1\}c=0.1, ϵ=0.1c=0.1,\ \epsilon=0.156.73 / 12.17
EEGMamba / ERBatch ∈{16,32,64}\in\{16,32,64\}, α∈{0.5,1,2}\alpha\in\{0.5,1,2\}Batch =32, α=0.5=32,\ \alpha=0.555.22 / 11.27

On CodeBrain, tuned LwF reaches 57.11/12.78 and SI reaches 56.73/12.17, improving both metrics over Naïve CL's 54.76/13.66. NeuroAccord achieves 60.86/4.69. These additional configurations provide a stronger comparison alongside Table 2.

EWC obtains 53.65/13.47 in Table 2. Its lower FOR than Naïve CL comes with lower AAA, illustrating the need to assess accuracy and forgetting together. NeuroAccord improves both metrics over EWC and the tuned LwF/SI configurations.


Q4. Actual rank growth and the stability–plasticity tradeoff.

Response.{\color{green}\mathbf{Response.}} Both analyzed backbones use d=200d=200 and a default rank of r=3r=3. The measured historical dimensions during training are:

Task being learned123456789
Historical dimension qt−1q_{t-1}03691215182124

By task 9, the historical basis covers 12% of the feature space. Varying the per-stage rank gives q8=8rq_8=8r in the measured runs:

Per-stage rank rrActual q8q_8 (both backbones)CBraMod AAA ↑ / FOR ↓EEGMamba AAA ↑ / FOR ↓
21659.68 / 7.2857.64 / 4.22
32461.95 / 4.3959.01 / 2.43
43260.94 / 5.1358.47 / 2.66
64860.82 / 4.0657.95 / 2.81
86460.50 / 4.3558.37 / 2.25

The default r=3r=3 gives the highest AAA on both backbones. On CBraMod, r=6r=6 reduces FOR by 0.33 pp at a cost of 1.13 pp AAA. On EEGMamba, r=8r=8 reduces FOR by 0.18 pp at a cost of 0.64 pp AAA. Increasing rank therefore trades accuracy for slightly lower forgetting in these configurations.

The capacity experiment at fixed r=3r=3 also measures later-task learning:

qmax⁡q_{\max}CBraMod AAA / FORCBraMod late-task BAccEEGMamba AAA / FOREEGMamba late-task BAcc
959.44 / 11.2065.7957.49 / 3.3162.95
1260.11 / 8.1965.9157.45 / 2.9663.17
1561.98 / 7.2966.3558.54 / 2.5363.22
1860.54 / 6.4465.8158.79 / 3.2163.27
2161.04 / 5.9766.0257.45 / 2.7062.57
2461.95 / 4.3966.0659.01 / 2.4363.54

Late-task BAcc averages the initial accuracy on the final three tasks. At the default 24-dimensional capacity, it reaches 66.06% on CBraMod and 63.54% on EEGMamba. These results show that the cumulative basis preserves old-task performance while supporting learning of later tasks in the nine-task sequence.

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

This paper suggests a way to build knowledgebase based on temporal-spatial-frequency basis per task, while this knowledge base would be transferred across task to provide robustness to data and parameters knowledge drifts. NeuroAccord is modularly applicable to various EFMs, and show good performances across datasets compared to prior continual learning frameworks.

Critical strengths

Thanks to several novelties, the paper's method section is quite intensive to read, but still renders the idea in a clean manner. It was easy to follow, with clear motivation.

Explicit control of the TSF Basis would also provide interpretability.

The results are also strong, with general applicability. Ablations are well done and described.

Interpretation across each dimension of TSF is well described in the appendix.

Critical weaknesses

Subject-level heterogeneity is a motivation (Fig. 1(VI)), but I do not see clearly where in the paper justifies / addresses this concern.

Questions for authors

How would TSF encode subject-level variability?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Geye

Thank you for your time and helpful comments on subject heterogeneity. Below, we explain how TSF captures individual signal differences and present the subject-level analyses.

W1. How does the method address the subject heterogeneity motivating Figure 1?

Response.{\color{green}\mathbf{Response.}} TSF captures subject differences through each sample's waveform dynamics, channel dependence, and spectral composition. The descriptor definitions are shared across subjects, while their values reflect individual signal differences. Their association with model features guides knowledge preservation. Training uses these signal properties without subject labels or subject-specific models.

The following analysis measures subject differences and accuracy gains on held-out subjects:

DatasetWithin-subject TSF distanceBetween-subject TSF distanceRatioSubject-mean BAcc gain over Naïve CL (pp)
BCIC-IV-2a0.1880.6103.242+7.90
PhysioNet-MI0.4410.9922.252+5.92
BCIC-UpperLimb0.2952.2237.532+1.11
SHU-MI0.6031.0041.665+2.26
eldBETA0.7971.2101.518+1.59

Between-subject distances exceed within-subject distances in all five datasets, and subject-mean BAcc improves by 1.11–7.90 pp over Naïve CL. Thus, TSF captures individual differences, and NeuroAccord's retention benefit extends to unseen subjects.


Q1. How does TSF encode subject-level variability?

Response.{\color{green}\mathbf{Response.}} We examine this directly on the 25 SHU-MI subjects. Descriptors are standardized using training-partition statistics and averaged within each subject, session, and class. We then compare within-subject and between-subject distances across sessions, with class and session combinations matched. We use dimension-normalized Euclidean distance and 2,000 subject-level bootstrap resamples to estimate confidence intervals.

ViewWithin-subject distanceBetween-subject distanceRatio95% bootstrap CI
Temporal0.4260.7001.641[1.393, 1.944]
Spatial0.4500.7911.756[1.431, 2.160]
Frequency0.6491.1621.789[1.418, 2.168]
Joint TSF0.5580.9561.713[1.472, 1.977]

All 95% intervals lie above one, showing that subject-related structure persists after matching class and session conditions. The joint ratio is 1.713 in this matched analysis and 1.665 in the pooled comparison in W1. Subject identities are used only for this offline diagnostic.


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

This paper proposes NeuroAccord, a Time-Space-Frequency (TSF)-guided continual learning framework for EEG foundation models. It identifies TSF-associated representation subspaces and combines relational distillation, coupling alignment, and historical knowledge anchoring to mitigate catastrophic forgetting without replaying historical EEG recordings. Experiments across nine datasets, four EEG foundation models, and 18 continual learning baselines demonstrate improvements in average anytime accuracy and forgetting.

Critical strengths

S1. Important research problem and reasonable methodological motivation.

The paper addresses knowledge retention during continual adaptation of EEG foundation models across heterogeneous tasks. This is an important problem, and the proposed framework is reasonably motivated.

S2. Comprehensive experimental evaluation.

The experiments cover nine EEG datasets, four foundation models, and 18 continual learning baselines. The ablation, sensitivity, and robustness analyses further support the empirical evaluation.

Critical weaknesses

W1. Limited discussion of relevant prior work.

The paper would benefit from discussing EvoBrain, Subspace Distillation for Continual Learning, and Prototype-Sample Relation Distillation, particularly their connections to EEG continual learning, subspace preservation, and relational distillation. What Do EEG Foundation Models Capture from Human Brain Signals? is also relevant to distinguishing encoded signal properties from task-discriminative information.

W2. Insufficient justification for the discriminative importance of TSF subspaces.

TSF subspaces are selected based on linear associations between model representations and 16 EEG descriptors. However, directions strongly associated with spectral or waveform statistics are not necessarily important for previous tasks' decision boundaries. Preserving these directions therefore does not directly establish preservation of historical discriminative knowledge.

W3. Overstated stability of TSF descriptors as a knowledge reference.

The 16 TSF descriptors provide a compact and broadly applicable summary of EEG signal characteristics. However, treating them as a physiologically stable knowledge reference across heterogeneous tasks requires stronger justification.

W4. Unclear stability of TSF subspaces across learning stages.

The SVD-derived TSF subspaces depend on the data distribution, previous-stage model representations, and cross-covariance statistics. Their consistency across learning stages remains unclear, as does whether the selected directions consistently capture historical knowledge worth preserving.

W5. Potential reduction in plasticity as historical subspaces accumulate.

Historical knowledge anchoring progressively preserves feature directions from earlier tasks. As these directions accumulate, the remaining representation space available for learning new tasks may become increasingly restricted. An analysis of historical subspace dimensionality and long-term plasticity would help clarify this concern.

W6. Numerical discrepancies between Figure 4(b) and Table 2.

The TSF results in Figure 4(b) differ noticeably from the corresponding NeuroAccord results in Table 2. For CBraMod, AAA/FOR changes from 58.39/8.71 to 61.95/4.39; for EEGMamba, from 58.40/8.57 to 59.01/2.43. Please clarify whether these results use different experimental configurations or evaluation protocols.

W7. Insufficient details on dataset partitioning.

The paper does not clearly specify the training, validation, and test partitioning protocols for all datasets, particularly regarding subject-, session-, trial-, and window-level separation. These details are important for interpreting EEG generalization performance and ensuring reproducibility.

Questions for authors

Q1. Could the authors clarify the relationship between TSF-associated directions and the discriminative information required by previous tasks? Is there evidence that these directions remain stable and relevant to historical predictions across learning stages?

Q2. Appendix B.4 provides a conditional forgetting bound involving cross-task TSF distribution shift, projected-drift approximation error, and off-subspace logit changes. Could the authors examine whether these conditions hold in the evaluated task sequences? This would help establish when HKA on current-task inputs can effectively preserve historical performance.

Q3. Since Eq. (10) progressively accumulates historical TSF directions, could the authors analyze how the retained subspace dimension and overlap evolve across learning stages?

I would be open to increasing my rating if the authors can address my concerns during the discussion period.

4 comments
AuthorsOpenReview ↗

Response to Reviewer sbkX (1/4)

Thank you for your time, careful reading, and thoughtful questions about the TSF reference. We address them below with clarifications and additional experiments on its contribution to old-task retention.

W1. Connections to EvoBrain, subspace distillation, and relational distillation.

Response.{\color{green}\mathbf{Response.}} These works focus on samples or parameters retention mechanisms, whereas NeuroAccord investigates the question of how EEG signal structure guides what to preserve. EvoBrain [1] combines spectral adaptation, relational distillation, and replay. SD [2] preserves latent subspaces, and PRD [3] preserves prototype–sample relations. Our central insight is that EEG signal structure can guide the choice of representation directions to preserve across heterogeneous tasks. Cross-covariance links observable signal properties to these directions. They then provide a shared reference for preserving sample relationships and historical features without storing earlier EEG recordings.

The additional comparisons use the same backbones and nine-task protocol:

MethodCBraMod AAA ↑ / FOR ↓EEGMamba AAA ↑ / FOR ↓
Naïve CL53.17 / 20.9553.69 / 6.86
EvoBrain60.44 / 7.5957.98 / 3.85
SD56.18 / 6.8955.65 / 3.15
PRD57.06 / 6.9256.93 / 4.89
NeuroAccord61.95 / 4.3959.01 / 2.43

Compared with EvoBrain, NeuroAccord gains 1.51 pp AAA on CBraMod and 1.03 pp on EEGMamba, while reducing FOR by 3.20 and 1.42 pp, respectively. The distinction between encoded signal properties and their predictive use in [4] is also relevant here. Our descriptor ablations in W2 test their contribution to old-task accuracy.

[1] EvoBrain: Continual Learning of EEG Foundation Models Across Heterogeneous BCI Tasks.
[2] Subspace Distillation for Continual Learning.
[3] Prototype-Sample Relation Distillation: Towards Replay-Free Continual Learning.
[4] What Do EEG Foundation Models Capture from Human Brain Signals?


W2. Do TSF-associated directions preserve discriminative information?

Response.{\color{green}\mathbf{Response.}} You are right to distinguish signal association from decision relevance. Equations (3)–(4) select directions associated with EEG properties, and the ablations test how this signal information contributes to old-task accuracy. Training jointly optimizes the supervised task loss and the preservation objective.

Table 4 and the additional partial ablations test each descriptor view:

Descriptors retainedCBraMod AAA ↑ / FOR ↓EEGMamba AAA ↑ / FOR ↓
All 1661.95 / 4.3959.01 / 2.43
Remove 3 of 6 temporal descriptors60.88 / 6.7558.25 / 3.39
Remove all temporal descriptors60.70 / 7.2357.99 / 3.72
Remove 2 of 5 spatial descriptors61.26 / 6.0558.27 / 3.96
Remove all spatial descriptors61.10 / 6.1458.06 / 4.28
Remove 2 of 5 frequency descriptors61.04 / 7.2758.12 / 3.41
Remove all frequency descriptors60.84 / 7.3257.92 / 3.77

Partial removal of any view lowers AAA and increases FOR on both backbones, with larger changes after complete removal. For CBraMod, removing the temporal or frequency view raises FOR from 4.39% to 7.23% or 7.32%, respectively. These results support the value of the selected signal information for retaining old-task accuracy.


W3. Why is TSF a stable reference across heterogeneous tasks?

Response.{\color{green}\mathbf{Response.}} “Stable” refers to the descriptor definitions, which remain fixed as the network learns. Their values and distributions can vary across tasks. Appendix A defines six temporal descriptors, five channel-dependence descriptors, and five relative band-power descriptors. Under consistent preprocessing, these give heterogeneous EEG data a common description. Figure 4(a) shows the cross-dataset variation, and Appendix B.4 accounts for this distribution shift in the retention analysis. TSF therefore provides a consistent reference while allowing task-specific signal differences.


AuthorsOpenReview ↗

Response to Reviewer sbkX (2/4)

W4. Stability of TSF subspaces across learning stages.

Response.{\color{green}\mathbf{Response.}} Newly estimated TSF directions can change as the data and representations evolve. NeuroAccord retains the directions from earlier stages in its cumulative historical basis. Equation (10) gives:

undefined
AuthorsOpenReview ↗

Response to Reviewer sbkX (3/4)

W7. Dataset partitions and subject/session/trial/window separation.

Response.{\color{green}\mathbf{Response.}} The dataset partitions are listed below:

DatasetPartition unitTrainValidationTest
BCIC-IV-2aSubject1–56–78–9
PhysioNet-MISubject1–7071–8990–109
Mumtaz2016Subject1–4344–5253–62
Mental ArithmeticSubject1–2829–3233–36
BCIC2020-3Non-subject indices†1–6061–7071–80
BCIC-UpperLimbSubject1–1011–1213–15
eldBETASubject1–7576–8081–100
SHU-MISubject1–1516–2021–25
FACEDSubject1–8081–100101–123

† BCIC2020-3 has 15 participants. Its 60/10/10 allocation uses non-subject indices. The other eight datasets use subject-disjoint partitions.

For the eight subject-partitioned datasets, each subject's sessions, trials, and derived windows stay in the same partition. All methods use identical partitions and task order. Retention on these datasets is therefore evaluated on unseen subjects from previously learned tasks.


Q1. Evidence linking TSF directions to historical predictions.

Response.{\color{green}\mathbf{Response.}} We address your question through descriptor ablations and old-task accuracy trajectories. The ablations test the contribution of each signal view:

Descriptors retainedCBraMod AAA ↑ / FOR ↓EEGMamba AAA ↑ / FOR ↓
All 1661.95 / 4.3959.01 / 2.43
Remove 3 of 6 temporal descriptors60.88 / 6.7558.25 / 3.39
Remove all temporal descriptors60.70 / 7.2357.99 / 3.72
Remove 2 of 5 spatial descriptors61.26 / 6.0558.27 / 3.96
Remove all spatial descriptors61.10 / 6.1458.06 / 4.28
Remove 2 of 5 frequency descriptors61.04 / 7.2758.12 / 3.41
Remove all frequency descriptors60.84 / 7.3257.92 / 3.77

We also track first-task BAcc as subsequent tasks are learned:

StageTask 1 BAcc: NeuroAccordTask 1 BAcc: Naïve CL
156.3456.08
355.9951.48
554.1744.79
754.3429.34
953.9125.43

Each TSF view contributes to retention, and first-task accuracy remains substantially higher than Naïve CL after eight further tasks. These findings support the usefulness of the signal reference for preserving old-task predictions.


Q2. Empirical applicability of the conditional forgetting bound.

Response.{\color{green}\mathbf{Response.}} Equation (43) explains when controlling drift on current inputs also controls drift on historical inputs:

\mathbb{E}_{\mu_j} D_t \leq \sqrt{E_t} + K_t W_1(\phi_{\\#}\mu_j, \phi_{\\#}\mu_t) + \delta_{j,t}.

Here EtE_t is the scalar drift quantity in Appendix B.4. HKA penalizes projected drift on current inputs. The distribution-shift and approximation-error terms govern its transfer to historical inputs. Equations (40)–(42) also account for classification-head sensitivity, off-subspace logit changes, and old-task margins.

The diagnostics below are averaged over the nine-task sequence. “RHS” denotes the empirical sum of the three terms on the right of Eq. (43).

Backbone / settingEt\sqrt{E_t}KtW1K_tW_1δj,t\delta_{j,t}Historical drift / RHSOff-subspace logit changeAAA / FOR
CB / full0.1860.4100.2690.438 / 0.8651.14261.95 / 4.39
CB / no HKA1.1721.0550.3650.921 / 2.5921.80959.61 / 8.65
EEG / full0.0660.1960.0440.082 / 0.3060.39459.01 / 2.43
EEG / no HKA0.3770.4470.1460.338 / 0.9710.61357.26 / 4.28

HKA reduces current and historical drift, the empirical RHS, and off-subspace logit changes. FOR also falls from 8.65% to 4.39% on CBraMod and from 4.28% to 2.43% on EEGMamba. These sequence-level results support the drift-transfer mechanism. The bound is most informative when head-scaled drift and off-subspace logit changes are small relative to old-task margins.


AuthorsOpenReview ↗

Response to Reviewer sbkX (4/4)

Q3. Growth and overlap of the retained historical subspace.

Response.{\color{green}\mathbf{Response.}} For Pt−1=Qt−1Qt−1⊤P_{t-1}=Q_{t-1}Q_{t-1}^{\top}, exact orthogonalization gives

undefined

Comments on the paper

1 comment

Dear Program Chairs, Senior Area Chair, Area Chair, and Reviewers,

We thank Reviewers sbkX, Geye, and L9eR for their careful reading and thoughtful follow-up. As an EEG foundation model learns new tasks, the importance of protected parameters can change, and a fixed replay buffer covers a shrinking share of past recordings. This motivated us to use the physiological structure of EEG to guide which learned features to preserve. NeuroAccord describes waveform dynamics, relationships across electrodes, and spectral rhythms consistently across tasks. Cross-covariance between these time–space–frequency (TSF) descriptors and model features identifies representation directions to protect. The reviewers’ suggestions helped us test whether these directions still matter for decoding. We are grateful for their input and summarize below the experiments and analyses addressing their concerns.

  1. Discriminative relevance of TSF (sbkX). Removing descriptors from any temporal, spatial, or frequency view reduced accuracy and increased forgetting on both backbones. We then froze the model and original task head and perturbed early TSF directions or random directions matched in rank and magnitude. Across two old tasks, two backbones, and stages 3, 6, and 9, perturbing TSF directions caused 2.57–3.84 percentage points more loss in balanced accuracy in all 12 comparisons.

  2. Practical value of the forgetting bound (sbkX). For 36 old-task–update pairs per configuration, we connected the historical drift in Eq. (43) to old-task decision margins, classifier sensitivity, and logit changes outside the protected subspace. Historical knowledge anchoring (HKA) tightened the mean unclipped empirical bound in Eq. (42) by 35.4% on CBraMod and 51.1% on EEGMamba. At the thresholds used, the near-boundary sample term was the largest remaining contribution on both backbones.

  3. Subject variability and evaluation splits (Geye, sbkX, L9eR). On SHU-MI, after matching classes and sessions, between-subject TSF distance was 1.71 times within-subject distance. Across five datasets, subject-averaged balanced accuracy on held-out participants improved by 1.11–7.90 percentage points over naïve continual learning. We documented all nine splits. Eight are subject-disjoint, and BCIC2020-3 uses non-subject indices.

  4. Choice of directions and how they are protected (sbkX, L9eR). Figure 4(b) varies the reference used to select directions under a common projection penalty. Table 2 evaluates the full method. In a separate control, we fixed the historical TSF basis and other objectives and varied the update rule. HKA achieved higher average anytime accuracy (AAA) and lower forgetting (FOR) than GPM-, SGP-, and TRGP-style replacements on both backbones.

  5. Subspace growth and late-task learning (sbkX, L9eR). At the default rank of three directions per stage, the historical basis occupied 24 of 200 dimensions at the start of task nine. After eight further tasks, CBraMod retained 53.91% balanced accuracy on BCIC-IV-2a, compared with 56.34% immediately after learning it. Initial balanced accuracy on the final three tasks averaged 66.06% on CBraMod and 63.54% on EEGMamba.

  6. Baseline comparisons and variability (sbkX, L9eR). We added direct comparisons with EvoBrain, Subspace Distillation, and Prototype–Sample Relation Distillation, along with tuning results for LwF, SI, and ER. In a separate three-seed evaluation against nine representative baselines, NeuroAccord achieved the highest mean AAA and lowest mean FOR on both CBraMod and EEGMamba.

  7. Storage and reproducibility (L9eR). We documented training settings, validation-based coefficient selection, and replay budgets of 128, 256, and 512 examples. At stage nine on CBraMod, the frozen teacher and historical basis occupied 18.65 MiB, compared with 40.71 MiB for ER with 256 examples. At 512 examples, the strongest evaluated replay method, CSReL, achieved 60.33 AAA and 8.95 FOR, compared with 61.95 and 4.39 for NeuroAccord.

We are especially grateful to the reviewers for encouraging us to test what is retained at the level of individual decisions. The margin analysis explains when limiting their drift can preserve an old prediction despite changes elsewhere in the representation. Together, these findings offer a way to choose what to preserve using observable EEG properties and evaluate that choice through its effect on decoding.

We sincerely thank all three reviewers for the time and care they devoted to our paper, from their initial reviews through the follow-up discussions. Their thoughtful feedback helped us strengthen both the study and its presentation.

Best regards,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.