- 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
Response to Reviewer L9eR (2/3)
W3. What does HKA contribute beyond accumulated bases and gradient projection?
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 , TSFRD, COA, and all other training settings fixed. It removes HKA or replaces it with a GPM-, SGP-, or TRGP-style update rule:
| Configuration | CBraMod AAA ↑ / FOR ↓ | EEGMamba AAA ↑ / FOR ↓ |
|---|---|---|
| Full NeuroAccord | 61.95 / 4.39 | 59.01 / 2.43 |
| Remove HKA | 59.61 / 8.65 | 57.26 / 4.28 |
| Replace HKA with GPM | 57.42 / 10.79 | 56.56 / 6.43 |
| Replace HKA with SGP | 57.34 / 11.57 | 56.79 / 6.07 |
| Replace HKA with TRGP | 56.84 / 10.87 | 55.70 / 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 , with historical directions, feature tokens, and teacher variance normalizer . 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. ---Response to Reviewer L9eR (3/3)
Q2. Which datasets use subject-disjoint evaluation?
Eight datasets use subject-disjoint partitions, as listed below:
| Dataset | Partition unit | Train | Validation | Test |
|---|---|---|---|---|
| BCIC-IV-2a | Subject | 1–5 | 6–7 | 8–9 |
| PhysioNet-MI | Subject | 1–70 | 71–89 | 90–109 |
| Mumtaz2016 | Subject | 1–43 | 44–52 | 53–62 |
| Mental Arithmetic | Subject | 1–28 | 29–32 | 33–36 |
| BCIC2020-3 | Non-subject indices† | 1–60 | 61–70 | 71–80 |
| BCIC-UpperLimb | Subject | 1–10 | 11–12 | 13–15 |
| eldBETA | Subject | 1–75 | 76–80 | 81–100 |
| SHU-MI | Subject | 1–15 | 16–20 | 21–25 |
| FACED | Subject | 1–80 | 81–100 | 101–123 |
† 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:
| Dataset | Within-subject TSF distance | Between-subject TSF distance | Ratio | Subject-mean BAcc gain over Naïve CL (pp) |
|---|---|---|---|---|
| BCIC-IV-2a | 0.188 | 0.610 | 3.242 | +7.90 |
| PhysioNet-MI | 0.441 | 0.992 | 2.252 | +5.92 |
| BCIC-UpperLimb | 0.295 | 2.223 | 7.532 | +1.11 |
| SHU-MI | 0.603 | 1.004 | 1.665 | +2.26 |
| eldBETA | 0.797 | 1.210 | 1.518 | +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.
Configurations are selected by validation performance. The additional LwF, SI, and ER searches use the following candidates and selected values:
| Backbone / method | Search candidates | Selected setting | AAA / FOR |
|---|---|---|---|
| CodeBrain / LwF | , | 57.11 / 12.78 | |
| CodeBrain / SI | , | 56.73 / 12.17 | |
| EEGMamba / ER | Batch , | Batch | 55.22 / 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.
Both analyzed backbones use and a default rank of . The measured historical dimensions during training are:
| Task being learned | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 |
|---|---|---|---|---|---|---|---|---|---|
| Historical dimension | 0 | 3 | 6 | 9 | 12 | 15 | 18 | 21 | 24 |
By task 9, the historical basis covers 12% of the feature space. Varying the per-stage rank gives in the measured runs:
| Per-stage rank | Actual (both backbones) | CBraMod AAA ↑ / FOR ↓ | EEGMamba AAA ↑ / FOR ↓ |
|---|---|---|---|
| 2 | 16 | 59.68 / 7.28 | 57.64 / 4.22 |
| 3 | 24 | 61.95 / 4.39 | 59.01 / 2.43 |
| 4 | 32 | 60.94 / 5.13 | 58.47 / 2.66 |
| 6 | 48 | 60.82 / 4.06 | 57.95 / 2.81 |
| 8 | 64 | 60.50 / 4.35 | 58.37 / 2.25 |
The default gives the highest AAA on both backbones. On CBraMod, reduces FOR by 0.33 pp at a cost of 1.13 pp AAA. On EEGMamba, 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 also measures later-task learning:
| CBraMod AAA / FOR | CBraMod late-task BAcc | EEGMamba AAA / FOR | EEGMamba late-task BAcc | |
|---|---|---|---|---|
| 9 | 59.44 / 11.20 | 65.79 | 57.49 / 3.31 | 62.95 |
| 12 | 60.11 / 8.19 | 65.91 | 57.45 / 2.96 | 63.17 |
| 15 | 61.98 / 7.29 | 66.35 | 58.54 / 2.53 | 63.22 |
| 18 | 60.54 / 6.44 | 65.81 | 58.79 / 3.21 | 63.27 |
| 21 | 61.04 / 5.97 | 66.02 | 57.45 / 2.70 | 62.57 |
| 24 | 61.95 / 4.39 | 66.06 | 59.01 / 2.43 | 63.54 |
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.
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.
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:
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=3. Initial learning rates, batch sizes, and final coefficients are:
The coefficient candidates are
undefined