acceptodds
Under review as a conference paper at ICLR 2027

BrainNEST: Negotiated Share with Routed Memory in Federated Continual Learning of Large EEG Models

Abstract

Large EEG Models (LEMs) learn transferable representations from diverse datasets, yet their knowledge must keep pace with new tasks, recording protocols, and subject populations. These datasets arrive over time and remain distributed across institutions, calling for a model that can learn collaboratively while retaining earlier capabilities. We present the first systematic study of federated continual adaptation of pretrained LEMs across evolving EEG datasets and task families. Our framework, BrainNEST, organizes adaptation around evidence-gated knowledge consolidation. Clients encode multiscale channel and frequency-band structure into compact sketches. The server establishes partial cross-client correspondences and uses this evidence to control both the blockwise contribution of local updates and their allocation to shared or routed state. Routed parameter memory retains updates for subsequent reuse, retrieving offsets by structural and acquisition compatibility to initialize future clients. Experiments on 14 datasets spanning seven BCI task families and two pretrained backbones demonstrate improved adaptation, retention, and transfer to unseen datasets. BrainNEST provides a practical approach to maintaining LEMs as distributed EEG resources continue to grow.

Primary area:
applications to neuroscience & cognitive science
Keywords:
federated continual learning, large EEG models, electroencephalography, knowledge consolidation, brain-computer interfaces

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_tbNwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how pretrained EEG models can learn from new datasets over time in a federated setting, without revisiting earlier training data. BrainNEST compares summaries of multiscale channel and frequency-band structure across clients to decide which updates to share. Updates with broad support are merged into the shared backbone, while others are stored as parameter offsets for later reuse. The authors evaluate BrainNEST on 14 datasets from seven BCI task families over four stages, using CBraMod and CodeBrain. Experiments include federated and continual learning baselines, ablations, and transfer by fine-tuning the final backbone on two held-out datasets.

Critical strengths
  1. The paper addresses a practical challenge: adapting pretrained EEG models across heterogeneous datasets without revisiting earlier training data. Local prediction heads accommodate different label spaces, while the federated continual setting combines collaboration across datasets with retention across stages.

  2. BrainNEST consistently improves adaptation and reduces forgetting compared with ReFed under matched communication and training budgets. These gains hold across both backbones, and the ablations support the contribution of the main components.

Critical weaknesses
  1. The analysis bounds aggregation weights, but larger clients take more optimizer steps under the one-epoch schedule with a fixed batch size. The resulting differences in local updates are neither covered by the bound nor isolated experimentally, leaving their effect on consolidation unclear.

  2. The removals in Table 4 also change regularization, update filtering, or parameter reuse. The performance drops show that these modules help, but do not distinguish the benefit of structural matching from these other effects.

Questions for authors

1.How was Re-Fed’s replay mechanism adapted after clients left and were no longer retrained? Which historical samples, if any, remained available for replay?

2.When computing Rt,jR_{t,j}, was each historical dataset’s head paired with the current shared backbone, a personalized model retrieved from memory, or a saved checkpoint? Which memory state was used, and how was this handled across methods?

3.Were the comparisons and ablations in Tables 3–4 repeated with independent seeds beyond Seed 42? If so, how much did the AAA and forgetting gaps between BrainNEST and the strongest baselines vary?

4 comments
AuthorsOpenReview ↗

Response to Reviewer tbNw (1/3)

Thank you for your time and detailed suggestions on the analysis and experimental protocol. Below, we address your concerns about data-size imbalance, mechanism attribution, and evaluation with controlled results and precise definitions.

W1. Aggregation-weight bounds do not control differences in local optimizer steps.

Response.{\color{green}\mathbf{Response.}} Your question separates server-side weighting from local optimization. Proposition C.3 bounds αiG\alpha_i^G, while the full contribution also contains the exposed local update ϱi→G,ℓΔi(ℓ)\varrho_{i\to G,\ell}\Delta_i^{(\ell)} (main-text Eq. (8)). Local data and optimization determine that update’s magnitude and direction.

Tables R11–R12 directly control the number of local updates. Every client receives 45 steps per round in Stage 1 or 47 in Stage 3. Evidence-balanced weighting achieves higher macro BA than both alternatives: 0.5560/0.4609, compared with 0.5418/0.4452 for uniform weighting and 0.5310/0.4334 for sample weighting. The advantage therefore persists with equal local update counts. These controls complement the coefficient bound by testing the optimization effect you identify.

Table R11. Stage 1 aggregation controls.

WeightingLocal budget per roundBCIC-IV-2aPhysioNet-MIDEAP-VASEEDMacro BA ↑
Sample-weighted1 epoch0.55580.60130.32490.68460.5417
Uniform-client1 epoch0.56320.61150.33410.69980.5522
Evidence-balanced1 epoch0.57640.62630.34870.72470.5690
Sample-weighted45 steps0.54610.58170.33280.66350.5310
Uniform-client45 steps0.55490.59340.34060.67820.5418
Evidence-balanced45 steps0.56530.60790.34950.70140.5560

Every client receives the same 45-step budget in the fixed-step rows. One step denotes one optimizer update. This control fixes update counts per client. All dataset entries are BA.

Table R12. Stage 3 aggregation controls.

WeightingLocal budget per roundSEED-VBCIC-UpperLimbBCIC2020-3BETAMacro BA ↑
Sample-weighted1 epoch0.37850.43890.34510.63560.4495
Uniform-client1 epoch0.38720.44960.35480.64830.4600
Evidence-balanced1 epoch0.40200.46110.36670.66340.4733
Sample-weighted47 steps0.35260.42970.33840.61290.4334
Uniform-client47 steps0.36410.44150.34760.62740.4452
Evidence-balanced47 steps0.38150.45580.35930.64710.4609

Every client receives the same 47-step budget in the fixed-step rows. One step denotes one optimizer update. This control fixes update counts per client. All dataset entries are BA.


AuthorsOpenReview ↗

Response to Reviewer tbNw (2/3)

W2. Removal ablations confound structural matching with other mechanisms.

Response.{\color{green}\mathbf{Response.}} The last two rows of Table R3 address your concern by retaining equal-capacity parameter memory, personalization, and the Eq. (14) objective form. Evidence-guided consolidation and routing yield +0.57 BA points, +1.87 AAA points, and −0.73 Fgt points.

This comparison measures the joint contribution of consolidation and routing, including how the policy shapes memory contents and alignment references. Table R13 quantifies the role of each component. Removing Family reduces mean AAA by 5.42/5.43 points, while removing Routed reduces it by 3.19/3.21 points across the two backbones. The matched controls and component ablations together show that evidence-guided knowledge organization contributes gains beyond memory capacity, personalization, and objective form.

Table R3. Memory, personalization, and local-objective controls.

MethodMatched memoryPersonalizationEq. (14)BA ↑AAA ↑Fgt ↓
FedAvgNoYesNo0.57370.51690.0594
FedAvg without personalizationNoNoNo0.55180.50140.0697
FedAvg + matched memory, without personalizationYesNoNo0.58460.52790.0523
FedAvg + matched memoryYesYesNo0.59310.54120.0468
FedAvg + matched memory + local objectiveYesYesYes0.59720.55240.0426
BrainNESTYesYesYes0.60290.57110.0353

BA is the equal-dataset macro-average across the four independent stages. AAA/Fgt use CBraMod. The last two rows retain equal-capacity parameter memory, stage-end personalization, and the Eq. (14) objective form. Their consolidation and routing policies differ.

Table R13. Repeated component ablations.

MethodCBraMod AAA ↑CBraMod Fgt ↓CodeBrain AAA ↑CodeBrain Fgt ↓
BrainNEST (Full)0.5738 ± 0.00680.0329 ± 0.00390.5867 ± 0.00740.0351 ± 0.0042
w/o Family0.5196 ± 0.01240.0457 ± 0.00760.5324 ± 0.01310.0483 ± 0.0081
w/o Sparse0.5372 ± 0.01030.0412 ± 0.00650.5498 ± 0.01120.0436 ± 0.0069
w/o Gate0.5485 ± 0.00910.0478 ± 0.00520.5613 ± 0.00970.0374 ± 0.0056
w/o Routed0.5419 ± 0.01170.0435 ± 0.00710.5546 ± 0.01230.0459 ± 0.0075
w/o Feedback0.5593 ± 0.00960.0394 ± 0.00590.5721 ± 0.01040.0418 ± 0.0063

Mean ± sample SD over training seeds 42, 52, 2025, 2222, and 99. Full reuses the same runs as Table R6. These are component-removal experiments.


Q1. Historical-sample access in the ReFed comparison.

Response.{\color{green}\mathbf{Response.}} For the client-exit case you ask about, the ReFed* reference retains selected historical training samples for replay after their source clients leave. Only active clients optimize, with access to these retained samples. Re-Fed [1] uses sample caching and replay, so this comparison includes additional historical training-data access.

BrainNEST follows strict client exit and retains knowledge through parameter offsets and retrieval information. Historical test partitions are used exclusively for evaluation. ReFed* therefore serves as a replay-enabled reference. FedWeIT and the matched-memory control provide comparisons under the historical-training-data restriction.

[1] Towards Efficient Replay in Federated Incremental Learning.


AuthorsOpenReview ↗

Response to Reviewer tbNw (3/3)

Q2. Backbone, memory, and prediction-head state for historical evaluation.

Response.{\color{green}\mathbf{Response.}} For the historical evaluation you ask about, Rt,jR_{t,j} uses the current stage-end shared backbone and current routed memory, paired with dataset jj’s frozen prediction head:

\sum_{k\in\mathcal A_{t,j}}\alpha_{t,jk}\Xi_k^t, \qquad R_{t,j}=\operatorname{BA}\!\left(\mathcal D_j^{\mathrm{test}}, h_j^{\mathrm{frz}}\circ f_{\Theta_{t,j}}\right).$$ Evaluation reads the state available at stage $t$, using the frozen training-time query and acquisition descriptor. Predictions are forward-only, and the current bank determines which memory entries are available. Consensus memory serves training-time alignment. All methods use the same stage-end timing, historical test splits, and task identity. Shared-model baselines use their current backbone with frozen historical heads. FedWeIT uses its retained task-specific parameters, and BrainNEST uses its current routed readout. This evaluates retention in each complete system. The unseen fine-tuning in Table 5 starts from BrainNEST’s final shared backbone alone and evaluates the quality of that initialization. --- > #### **Q3. Variation across independent training runs and ablations.** ${\color{green}\mathbf{Response.}}$ Yes. Tables R6 and R13 report **five independent training seeds**, including seed 42 and four additional seeds. BrainNEST’s mean AAA gains over FedWeIT are **3.89/4.93 points** on CBraMod/CodeBrain, and its mean Fgt reductions are **1.67/1.52 points**. Against ReFed*, the corresponding gains are **2.51/2.55 AAA points** and **0.96/1.10 Fgt points**. For BrainNEST, AAA SDs are **0.0068/0.0074**, compared with **0.0132/0.0141** for FedWeIT and **0.0108/0.0107** for ReFed*. The full ablation statistics are repeated below so you can compare each component’s effect with the observed variation. All ± values denote per-method sample SDs. Reported gains are differences between method means. **Table R6. Repeated continual evaluation.** | Method | CBraMod AAA ↑ | CBraMod Fgt ↓ | CodeBrain AAA ↑ | CodeBrain Fgt ↓ | | --- | --- | --- | --- | --- | | FedWeIT | 0.5349 ± 0.0132 | 0.0496 ± 0.0071 | 0.5374 ± 0.0141 | 0.0503 ± 0.0079 | | ReFed* | 0.5487 ± 0.0108 | 0.0425 ± 0.0052 | 0.5612 ± 0.0107 | 0.0461 ± 0.0064 | | BrainNEST (Ours) | 0.5738 ± 0.0068 | 0.0329 ± 0.0039 | 0.5867 ± 0.0074 | 0.0351 ± 0.0042 | Mean ± sample SD over training seeds 42, 52, 2025, 2222, and 99. ReFed* permits historical training-sample replay and is a separate reference. **Table R13. Repeated component ablations.** | Method | CBraMod AAA ↑ | CBraMod Fgt ↓ | CodeBrain AAA ↑ | CodeBrain Fgt ↓ | | --- | --- | --- | --- | --- | | BrainNEST (Full) | 0.5738 ± 0.0068 | 0.0329 ± 0.0039 | 0.5867 ± 0.0074 | 0.0351 ± 0.0042 | | w/o Family | 0.5196 ± 0.0124 | 0.0457 ± 0.0076 | 0.5324 ± 0.0131 | 0.0483 ± 0.0081 | | w/o Sparse | 0.5372 ± 0.0103 | 0.0412 ± 0.0065 | 0.5498 ± 0.0112 | 0.0436 ± 0.0069 | | w/o Gate | 0.5485 ± 0.0091 | 0.0478 ± 0.0052 | 0.5613 ± 0.0097 | 0.0374 ± 0.0056 | | w/o Routed | 0.5419 ± 0.0117 | 0.0435 ± 0.0071 | 0.5546 ± 0.0123 | 0.0459 ± 0.0075 | | w/o Feedback | 0.5593 ± 0.0096 | 0.0394 ± 0.0059 | 0.5721 ± 0.0104 | 0.0418 ± 0.0063 | Mean ± sample SD over training seeds 42, 52, 2025, 2222, and 99. Full reuses the same runs as Table R6. These are component-removal experiments. ---

Dear Reviewer tbNw,

Thank you for reading our paper so carefully and for your thoughtful questions. We sincerely apologize for the delay in posting our response. The additional experiments on client dataset sizes, the contribution of structural matching, and variability across training runs took longer than we anticipated. We also clarify which historical samples were available for replay and how we evaluated earlier tasks after their clients left. We appreciate your patience and would welcome any further questions or discussion.

Best regards,
The authors

Reviewer_EaReOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The proposed BrainNEST framework studies continual adaptation of EEG foundation models when different clients participate over successive stages. It constructs structural summaries, matches them across clients, and uses a gate to allocate updates to either the shared backbone or a memory of reusable parameters. The authors performed detailed empirical evaluation with multiple sets of ablation settings.

Critical strengths

The problem setting is relevant. Separating broadly transferable updates from specialized updates is a valid and reasonable approach. The evaluation covers a large number of datasets, across multiple task families, with different backbones. Significant performance improvements is demonstrated with CBraMod and CodeBrain, alongside lower forgetting. The appendix also provides useful information on operators, splits, hyperparameters, and computational costs.

Critical weaknesses

The default sharing gate may not open under certain scenarios. Per the definition of \omega_G, whenever there is exactly one validated cross-client group with positive dispersion, its normalized dispersion equals one. Consequently, a_G is a very small number less than sharing threshold. The group therefore cannot update the shared backbone or consensus memory even if it includes every active client.

The method used t to represent stages, while experiments use 120 communication rounds per stage. It is unclear which operations occur every round and which occur only at stage boundaries.

The family-to-block attribution weights whole minibatch gradient norms by family activity. It is unclear whether an update helps other clients or not. Two clients can exhibit similar structural summaries while requiring conflicting discriminative updates.

Questions for authors

Please see above

3 comments
AuthorsOpenReview ↗

Response to Reviewer EaRe (1/2)

Thank you for your time, careful reading, and helpful comments on the method. Below, we clarify the sharing gate, execution schedule, and attribution mechanism.

W1. A single positive-dispersion group may never pass the sharing gate.

Response.{\color{green}\mathbf{Response.}} Your calculation applies to the 0.08 threshold. From Eq. (9) and Appendix B.2, let cG=∣G∣/∣Ct∣c_G=|G|/|\mathcal C_t|. For one validated group with positive dispersion, σ~G=1\widetilde\sigma_G=1, hence

\bar m_G\leq\rho_{\mathrm{pt}}=0.20, \qquad a_G\leq0.20/e\approx0.07358.$$ **Stage-wise federated and continual runs use $\tau_{\mathrm{sh}}=0.03$.** Simultaneous 14-client federation uses 0.08. At 0.03, a single group passes exactly when $\bar m_Gc_G\geq0.03e\approx0.08155$. As an algebraic illustration, full coverage with $\bar m_G=0.15$ gives $a_G\approx0.05518>0.03$. At 0.08, the candidate in your example enters routed memory while the shared backbone and consensus memory stay unchanged. Its parameter update remains available for selective reuse. The inequality above gives the single-group sharing condition. Tables R9–R10 show the performance trade-off as the threshold changes. **Table R9. Sharing-threshold sensitivity in stage-wise experiments.** | $\tau_{\mathrm{sh}}$ | CBraMod S1 BA | CBraMod S2 BA | CBraMod AAA ↑ | CBraMod Fgt ↓ | CodeBrain AAA ↑ | CodeBrain Fgt ↓ | | --- | --- | --- | --- | --- | --- | --- | | 0.0100 | 0.5568 | 0.7753 | 0.5627 | 0.0442 | 0.5716 | 0.0465 | | 0.0200 | 0.5637 | 0.7798 | 0.5674 | 0.0391 | 0.5783 | 0.0412 | | 0.0300 | 0.5690 | 0.7825 | 0.5711 | 0.0353 | 0.5832 | 0.0380 | | 0.0400 | 0.5609 | 0.7761 | 0.5638 | 0.0354 | 0.5747 | 0.0379 | | 0.0500 | 0.5584 | 0.7739 | 0.5606 | 0.0345 | 0.5721 | 0.0368 | Single-run results. Stage-wise federated and continual configurations use $\tau_{\mathrm{sh}}=0.03$. The default row agrees with paper Tables 3, 13, and 14. **Table R10. Threshold sensitivity with all 14 clients active together.** | $\tau_{\mathrm{sh}}$ | Macro BA ↑ | | --- | --- | | 0.0200 | 0.5816 | | 0.0400 | 0.5994 | | 0.0600 | 0.6023 | | 0.0800 | 0.6071 | | 0.1000 | 0.5927 | This separate simultaneous-federation setting uses $\tau_{\mathrm{sh}}=0.08$. --- > #### **W2. Distinguish communication-round operations from stage transitions.** ${\color{green}\mathbf{Response.}}$ You can distinguish the two timescales using stage $t$ and round $r=1,\ldots,120$. At each round, clients initialize from the preceding shared backbone and routed memory. They perform one local epoch and upload updates and summaries. The server then performs matching, group validation, block exposure, gating, and memory writes to form the next round’s state. After round 120, active clients receive the common **10-epoch personalization** budget and stage-end evaluation. Historical datasets receive forward-only evaluation. The next stage changes the active-client set and inherits the shared backbone and memories. Evidence processing and memory writes therefore occur every round. Client arrival/departure and stage-level reporting occur at stage boundaries. ---
AuthorsOpenReview ↗

Response to Reviewer EaRe (2/2)

W3. Structural similarity and gradient-norm attribution need not imply beneficial transfer.

Response.{\color{green}\mathbf{Response.}} Your question concerns block exposure and transfer direction. Attribution in Appendix B.1, Eq. (4), combines family activity with minibatch block-gradient norms. It measures the association between a family and optimization activity in each block. The recipient’s loss change also depends on gradient alignment.

For a fixed recipient head and an external update δ\delta, an LL-smooth recipient loss satisfies

\leq\eta\langle\nabla\mathcal L_k(\theta),\delta\rangle +\frac{L\eta^2}{2}\|\delta\|_2^2.$$ A sufficiently negative inner product yields a decrease in the recipient’s loss. BrainNEST uses structural evidence to select update contributions and route knowledge. Tables R3 and R13 test the resulting learning benefit. With memory, personalization, and objective form retained, Table R3 shows **+1.87 AAA points and −0.73 Fgt points**. Table R13 quantifies component effects across five seeds. Together, these results support the contribution of evidence-guided consolidation and reuse. **Table R3 (excerpt). Consolidation and routing with matched supporting capabilities.** | Method | BA ↑ | AAA ↑ | Fgt ↓ | | --- | --- | --- | --- | | FedAvg + matched memory + local objective | 0.5972 | 0.5524 | 0.0426 | | BrainNEST | 0.6029 | 0.5711 | 0.0353 | BA is the equal-dataset macro-average across the four independent stages. AAA/Fgt use CBraMod. The last two rows retain equal-capacity parameter memory, stage-end personalization, and the Eq. (14) objective form. Their consolidation and routing policies differ. **Table R13. Repeated component ablations.** | Method | CBraMod AAA ↑ | CBraMod Fgt ↓ | CodeBrain AAA ↑ | CodeBrain Fgt ↓ | | --- | --- | --- | --- | --- | | BrainNEST (Full) | 0.5738 ± 0.0068 | 0.0329 ± 0.0039 | 0.5867 ± 0.0074 | 0.0351 ± 0.0042 | | w/o Family | 0.5196 ± 0.0124 | 0.0457 ± 0.0076 | 0.5324 ± 0.0131 | 0.0483 ± 0.0081 | | w/o Sparse | 0.5372 ± 0.0103 | 0.0412 ± 0.0065 | 0.5498 ± 0.0112 | 0.0436 ± 0.0069 | | w/o Gate | 0.5485 ± 0.0091 | 0.0478 ± 0.0052 | 0.5613 ± 0.0097 | 0.0374 ± 0.0056 | | w/o Routed | 0.5419 ± 0.0117 | 0.0435 ± 0.0071 | 0.5546 ± 0.0123 | 0.0459 ± 0.0075 | | w/o Feedback | 0.5593 ± 0.0096 | 0.0394 ± 0.0059 | 0.5721 ± 0.0104 | 0.0418 ± 0.0063 | Mean ± sample SD over training seeds 42, 52, 2025, 2222, and 99. Full reuses the same runs as Table R6. These are component-removal experiments. ---

Dear Reviewer EaRe,

Thank you for reading our paper so carefully and for your thoughtful questions. We sincerely apologize for the delay in posting our response. The additional experiments on the sharing threshold took longer than we anticipated. In our response, we explain when a single validated group can pass the gate, which operations run each round or at stage boundaries, and how clients with similar EEG structure can still require conflicting updates. We appreciate your patience and would welcome any further questions or discussion.

Best regards,
The authors

Reviewer_5KE6OpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper studies federated continual adaptation of pretrained EEG models across datasets that arrive over time. BrainNEST summarizes local channel, frequency, and temporal structure, matches compatible summaries across clients, and uses this evidence to control parameter sharing. Broadly supported updates enter the shared backbone, while specialized updates remain in routed memory for later reuse.

Experiments cover 14 datasets, seven task families, and two pretrained backbones. The reported results show better adaptation and retention than the evaluated federated continual baselines, with additional gains when fine-tuning on two unseen datasets.

Critical strengths

A well-motivated combination with a meaningful framework-level contribution. Prototype matching, selective aggregation, shared/private parameters, and memory reuse have precedents. The contribution is their coordination through a common structural-evidence interface: correspondence controls both blockwise contributions and the destination of updates. This is more specific than simply applying federated continual learning to EEG.

Broad and encouraging empirical results. Table 3 shows improvements in all four continual-learning metrics on both backbones. Compared with ReFed, anytime average accuracy increases from 0.5478 to 0.5711 for CBraMod and from 0.5583 to 0.5832 for CodeBrain, while forgetting decreases. Table 4 also shows meaningful degradation when sketching, gating, or routed memory is removed. Together with the dataset diversity, this provides useful evidence for the complete design.

Critical weaknesses

The experiments do not fully isolate the distinctive mechanism from the benefits of added structure and memory. FedWeIT already separates shared and specialized parameters, and EvoBrain already addresses continual adaptation across heterogeneous EEG tasks. BrainNEST’s novelty therefore depends on the evidence-controlled consolidation and retrieval mechanism.

The removal ablations are helpful, but do not fully establish how much improvement comes from this particular mechanism rather than additional memory capacity, personalization, or auxiliary representation learning. A simpler capacity-matched sharing-and-memory control would make that distinction clearer. Similarly, Table 5 compares the adapted backbone mainly against other pretrained models; it does not isolate BrainNEST’s contribution from additional exposure to the 14 training datasets.

The breadth of evaluation exceeds the evidence for robustness. Appendix F.1 specifies seed 42, and the continual evaluation follows one four-stage chronological grouping. No repeated-run uncertainty or alternative-order results are reported. Several single-stage and unseen-transfer margins are small, making their reliability difficult to assess.

Some wording also needs qualification. The claim of best results across all datasets and metrics has exceptions, including SEED-V κ and SHU-MI F1 in the appendix. Unseen transfer uses 100 epochs of full fine-tuning, so these results demonstrate improved adaptation initialization rather than immediate transfer without training.

Questions for authors

Novelty clarification: What most clearly distinguishes evidence-gated consolidation from a simpler shared/private memory system with similarity-based retrieval? Are any existing capacity-matched controls available?

Robustness clarification: Are repeated-seed or alternative-stage-order results available for the main continual comparison? How sensitive are sharing decisions to the fixed gate threshold?

Transfer clarification: How do the final backbones from ReFed or another matched continual baseline perform under the same unseen-dataset fine-tuning protocol?

8 comments
AuthorsOpenReview ↗

Response to Reviewer 5KE6 (1/5)

Thank you for your time and thoughtful feedback on the novelty and evaluation. Below, we address your concerns with matched controls, robustness results, and transfer comparisons.

W1. Distinctive contribution beyond existing shared/private and continual EEG methods.

Response.{\color{green}\mathbf{Response.}} BrainNEST asks which parts of locally learned EEG knowledge are ready to be shared, and which should remain available for selective reuse. Different EEG datasets can share local patterns despite differences in their tasks and overall distributions. BrainNEST uses these partial correspondences to decide which parameter blocks contribute to a joint update and where that update is stored (Figure 2 and Eqs. (6)–(13)). This links collaboration among current clients to knowledge reuse across later stages.

For the methods you mention, FedWeIT [1] transfers weighted task-specific parameters. EvoBrain [2] handles sequential EEG adaptation through spectral normalization, distillation, and replay. FedGLoRA [3] adapts frozen EEG backbones with global/private LoRA and subspace aggregation. BrainNEST uses partial EEG correspondence to coordinate update contributions and storage decisions. Tables R1–R2 compare these approaches. Table R3 tests the benefit of this policy with memory capacity and training objectives matched.

Table R1. All 14 clients trained simultaneously with CBraMod.

MethodMacro BA ↑
Central0.5672
FedAvg0.5768
FedPer0.5649
FedGLoRA0.5986
FedWeIT0.5803
BrainNEST0.6071

Results from a separate experiment with all 14 clients participating simultaneously.

Table R2. Continual adaptation across related approaches.

MethodCBraMod AAA ↑CBraMod Fgt ↓CodeBrain AAA ↑CodeBrain Fgt ↓
FedAvg0.51690.05940.52350.0600
FedGLoRA0.54700.04920.55600.0469
FedWeIT0.53900.04540.54250.0459
EvoBrain*0.53400.05680.53800.0534
BrainNEST0.57110.03530.58320.0380

Single-run results. EvoBrain* is a non-native-federated replay reference. FedGLoRA retains its global/private adaptation mechanism across stages. BrainNEST and FedWeIT values agree with paper Table 3.

Table R3 (excerpt). Consolidation and routing with matched supporting capabilities.

MethodBA ↑AAA ↑Fgt ↓
FedAvg + matched memory + local objective0.59720.55240.0426
BrainNEST0.60290.57110.0353

BA is the equal-dataset macro-average across the four independent stages. AAA/Fgt use CBraMod. The last two rows retain equal-capacity parameter memory, stage-end personalization, and the Eq. (14) objective form. Their consolidation and routing policies differ.

[1] Federated Continual Learning with Weighted Inter-client Transfer.

[2] EvoBrain: Continual Learning of EEG Foundation Models Across Heterogeneous BCI Tasks.

[3] FedGLoRA: Grassmann-Manifold Federated Learning via Dual LoRA for Large EEG Models.


AuthorsOpenReview ↗

Response to Reviewer 5KE6 (2/5)

W2. Separate the mechanism from memory, personalization, auxiliary losses, and additional data exposure.

Response.{\color{green}\mathbf{Response.}} The capacity-matched comparison you request is in Table R3. Its last two rows retain equal-capacity parameter memory, personalization, and the Eq. (14) objective form. BrainNEST improves BA by 0.57 points, AAA by 1.87 points, and reduces Fgt by 0.73 points (17.1% relative). These gains quantify the contribution of evidence-guided consolidation and routing under matched capacity and training objectives.

Table R4 addresses data exposure: the adapted backbones have encountered the same 14 datasets and receive the same subsequent 100-epoch fine-tuning budget. Relative to FedWeIT, BrainNEST improves eldBETA/TUAB BA by 1.03/1.51 points with CBraMod and 1.27/1.89 points with CodeBrain. The benefit therefore persists when the source datasets for prior adaptation are matched.

Table R3. Memory, personalization, and local-objective controls.

MethodMatched memoryPersonalizationEq. (14)BA ↑AAA ↑Fgt ↓
FedAvgNoYesNo0.57370.51690.0594
FedAvg without personalizationNoNoNo0.55180.50140.0697
FedAvg + matched memory, without personalizationYesNoNo0.58460.52790.0523
FedAvg + matched memoryYesYesNo0.59310.54120.0468
FedAvg + matched memory + local objectiveYesYesYes0.59720.55240.0426
BrainNESTYesYesYes0.60290.57110.0353

BA is the equal-dataset macro-average across the four independent stages. AAA/Fgt use CBraMod. The last two rows retain equal-capacity parameter memory, stage-end personalization, and the Eq. (14) objective form. Their consolidation and routing policies differ.

Table R4. Unseen-dataset fine-tuning from final backbones.

InitializationCBraMod eldBETACBraMod TUABCodeBrain eldBETACodeBrain TUAB
Original pretrained0.58970.78860.59920.7919
FedWeIT0.61350.78650.61270.7842
FOT0.59050.76080.59840.7742
ReFed*0.61430.78310.60950.7906
BrainNEST0.62380.80160.62540.8031

BA after 100 epochs of full fine-tuning under the same unseen-data protocol. Adapted initializations have encountered the same 14 training datasets. BrainNEST uses its final shared backbone alone. ReFed* additionally permits historical replay during continual adaptation. Pretrained and BrainNEST rows reproduce paper Table 5. Baseline-transfer rows are supplementary.


AuthorsOpenReview ↗

Response to Reviewer 5KE6 (3/5)

W3. Reliability across seeds, stage orders, and unseen-dataset fine-tuning.

Response.{\color{green}\mathbf{Response.}} You can assess these sources of variability separately in Tables R5–R8. Five training seeds give BrainNEST higher mean BA in all four independent stages. In continual learning, its mean AAA exceeds FedWeIT by 3.89/4.93 points on CBraMod/CodeBrain, with Fgt lower by 1.67/1.52 points. Against the replay-enabled ReFed reference, the corresponding AAA gains are 2.51/2.55 points.

BrainNEST also retains higher AAA and lower Fgt in every tested stage order. Five downstream fine-tuning seeds show higher mean BA than each original backbone initialization. Table R8 reports mean performance and run-to-run variation for all initializations, including ECHO, so you can assess the smaller margins in context.

Table R5. Repeated single-stage federated evaluation with CBraMod.

MethodStage 1 BA ↑Stage 2 BA ↑Stage 3 BA ↑Stage 4 BA ↑
FedAvg0.5412 ± 0.01240.7498 ± 0.01170.4426 ± 0.01480.6153 ± 0.0136
FedROD0.5487 ± 0.01030.7635 ± 0.00950.4271 ± 0.01290.6089 ± 0.0118
FedALA0.5354 ± 0.01160.7712 ± 0.00820.4593 ± 0.01010.6347 ± 0.0094
BrainNEST (Ours)0.5714 ± 0.00810.7846 ± 0.00670.4762 ± 0.00950.6493 ± 0.0084

Mean ± sample SD over training seeds 42, 52, 2025, 2222, and 99. Each stage is trained independently. BA is its dataset macro-average.

Table R6. Repeated continual evaluation.

MethodCBraMod AAA ↑CBraMod Fgt ↓CodeBrain AAA ↑CodeBrain Fgt ↓
FedWeIT0.5349 ± 0.01320.0496 ± 0.00710.5374 ± 0.01410.0503 ± 0.0079
ReFed*0.5487 ± 0.01080.0425 ± 0.00520.5612 ± 0.01070.0461 ± 0.0064
BrainNEST (Ours)0.5738 ± 0.00680.0329 ± 0.00390.5867 ± 0.00740.0351 ± 0.0042

Mean ± sample SD over training seeds 42, 52, 2025, 2222, and 99. ReFed* permits historical training-sample replay and is a separate reference.

Table R7. Alternative stage orders.

BackboneOrderFedAvgFedWeITReFed*BrainNEST
CBraModS1→S2→S3→S40.5169 / 0.05940.5390 / 0.04540.5478 / 0.04120.5711 / 0.0353
CBraModS4→S3→S2→S10.5074 / 0.06770.5258 / 0.05310.5314 / 0.04860.5576 / 0.0419
CBraModS2→S4→S1→S30.5048 / 0.06930.5213 / 0.05520.5289 / 0.05070.5528 / 0.0435
CodeBrainS1→S2→S3→S40.5235 / 0.06000.5425 / 0.04590.5583 / 0.04410.5832 / 0.0380
CodeBrainS4→S3→S2→S10.5138 / 0.07340.5307 / 0.05480.5452 / 0.05130.5703 / 0.0447
CodeBrainS2→S4→S1→S30.5094 / 0.06910.5276 / 0.05390.5418 / 0.04960.5671 / 0.0431

Each entry is AAA ↑ / Fgt ↓ for a single run. Methods are compared within the same order. Reordering also changes each dataset’s contribution to the cumulative AAA metric. ReFed* permits historical replay.

Table R8. Repeated unseen-dataset fine-tuning.

InitializationeldBETA BA ↑TUAB BA ↑
CBraMod0.5864 ± 0.00970.7852 ± 0.0103
CodeBrain0.5971 ± 0.01050.7938 ± 0.0098
ECHO0.6168 ± 0.00860.8014 ± 0.0089
BrainNEST (CBraMod)0.6215 ± 0.00780.8067 ± 0.0082
BrainNEST (CodeBrain)0.6261 ± 0.00720.8103 ± 0.0075

Mean ± sample SD over fine-tuning seeds 42, 52, 2025, 2222, and 99, using 100 epochs of full fine-tuning. These statistics describe downstream fine-tuning variability. Table R6 reports variability across continual-training runs.


W4. Scope of the performance and unseen-transfer claims.

Response.{\color{green}\mathbf{Response.}} Your examples help clarify the level of comparison. BrainNEST improves overall adaptation and retention on both backbones, leading on all four continual metrics in Table 3. Individual dataset metrics have exceptions. Table 15 reports SEED-V κ=0.2541\kappa=0.2541 for BrainNEST versus 0.2580 for FedAvg. Table 16 reports SHU-MI F1 = 0.6672 versus 0.6758 for FedALA.

Table 5 evaluates initialization quality under 100-epoch full fine-tuning. eldBETA and TUAB are held out from the 14-dataset adaptation stream. The results show that BrainNEST’s shared initialization supports better adaptation to new datasets under the same fine-tuning budget.


Comments on the paper

1 comment

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

We apologize for posting our detailed responses late in the discussion period. Repeating the continual-learning experiments and ablations across different seeds, comparing additional baselines on a 14-dataset benchmark spanning seven BCI task families, and checking the stage-order and transfer results took longer than we anticipated. We appreciate your patience and thank Reviewers 5KE6, EaRe, and tbNw for the care they took in identifying the comparisons that would best test our claims.

BrainNEST addresses a practical problem. EEG datasets arrive at different institutions over time. Two clients may share useful patterns across channels or frequency bands even when their tasks differ, yet sharing an entire model update can interfere with what either client has learned. BrainNEST uses partial EEG correspondences to decide which parameter blocks to share now and which updates to retain for selective reuse by future clients. The additional experiments test whether this decision improves learning under matched conditions.

  1. Contribution of evidence-guided sharing (Tables R1–R3 and R13). We added comparisons with FedGLoRA, FedWeIT, and EvoBrain. With all 14 clients active simultaneously, CBraMod reached 60.71% macro balanced accuracy with BrainNEST and 59.86% with FedGLoRA. We then gave FedAvg the same parameter-memory capacity, personalization budget, and local objective as BrainNEST. On CBraMod, BrainNEST improved anytime average accuracy (AAA) by 1.87 percentage points and reduced forgetting by 0.73 points over this matched control. Across five seeds, removing the multiscale family sketches reduced mean AAA by more than five points on both backbones.

  2. Reliability across seeds and stage orders (Tables R5–R7). Across five training seeds, BrainNEST exceeded FedWeIT in mean AAA by 3.89 points on CBraMod and 4.93 on CodeBrain, while reducing mean forgetting by 1.67 and 1.52 points, respectively. In a separate comparison of three stage orders, BrainNEST achieved higher AAA and lower forgetting than FedAvg, FedWeIT, and the replay-enabled ReFed reference in every tested order on both backbones.

  3. When a local update is shared (Tables R9–R10). A single validated group with sufficient support can pass the stage-wise gate. Updates below the threshold remain available in routed memory. The server evaluates support every round, and the active clients change at stage boundaries. In the stage-wise experiments, a threshold of 0.03 gave the highest AAA on both backbones, while 0.05 gave the lowest forgetting. With all 14 clients active simultaneously, 0.08 gave the highest macro balanced accuracy among the thresholds tested. Structural agreement guides sharing, while the effect on a recipient also depends on the alignment of its loss gradient with the update.

  4. Fairness across client sizes (Tables R11–R12). To separate aggregation weights from the number of local optimizer steps, we gave every client 45 steps per round in Stage 1 and 47 in Stage 3. Evidence-balanced aggregation exceeded uniform weighting by 1.42 and 1.57 points in macro balanced accuracy, respectively, and also exceeded sample-count weighting in both stages.

  5. Retention after clients leave. BrainNEST retains parameter offsets and retrieval information when a client leaves. Earlier tasks are evaluated by forward inference with the current shared backbone, current routed memory, and the frozen task head. We explicitly identified ReFed as a replay-enabled reference because it can retain historical training samples after their source clients leave.

  6. Transfer to unseen datasets (Tables R4 and R8). We compared final backbones previously adapted on the same 14 datasets, giving each the same 100 epochs of downstream fine-tuning on eldBETA and TUAB. Using its shared backbone alone, BrainNEST exceeded FedWeIT on both unseen datasets with both backbones by 1.03–1.89 points in balanced accuracy. A separate five-seed fine-tuning study also found higher mean balanced accuracy than the original pretrained initialization in all four backbone–dataset combinations. These results measure the quality of the initialization for subsequent adaptation.

We are grateful that the reviewers encouraged us to distinguish the benefit of the sharing policy from the effects of memory capacity, local computation, and prior data exposure. We would welcome the opportunity to discuss any remaining questions in greater depth.

If these results and clarifications address your concerns, we would greatly appreciate your consideration of an updated assessment or rating. Thank you again for your time and thoughtful feedback.

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.