acceptodds
Under review as a conference paper at ICLR 2027

BrainDuo: Dual-Domain Tokenization and Pretraining for EEG Foundation Models

19 views

Abstract

Electroencephalography (EEG) foundation models use large-scale pretraining to learn a general-purpose EEG encoder that can be adapted to diverse brain-decoding tasks. Although current EEG foundation models benefit from large-scale self-supervised pretraining, learning transferable representations from high-dimensional, non-stationary EEG signals remains challenging. Existing approaches can struggle to jointly preserve time series and spectral information and capture dependencies at multiple spatiotemporal scales, due to inherent limitations in tailoring tokenizer design and pretraining strategies to the characteristics of EEG signals. To address these challenges, we introduce BrainDuo, an EEG foundation model that combines a dual-domain spectrotemporal codebook with discrete and continuous supervision to learn local and long-range dependencies among EEG tokens. Our BrainDuo features three novel components, including Duo Codebook, Duo Latent and DuoView Pretraining to learn effective, robust EEG representation. Specifically, Duo Codebook learns a shared spectrotemporal vocabulary, Duo Latent transfers both discrete codes and continuous embeddings into pretraining, and DuoView Pretraining applies cyclic and clustered attention to encode the temporal context along each channel and the topographic context among neighboring electrodes. We further incorporate brain-computer interface (BCI)-native skills into downstream tasks to help EEG foundation models exploit task-specific prior knowledge for various BCI decoding. Extensive experiments across 10 representative EEG tasks and 14 datasets demonstrate that BrainDuo outperforms strong EEG foundation baselines across general EEG benchmarks and moves closer to the models performance specialized in BCI tasks. These results highlight the potential of leveraging a hierarchical, flexible design to learn general-purpose EEG representations, demonstrating BrainDuo’s versatility as a backbone for diverse EEG decoding and BCI applications.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, Foundation Model, BCI, EFM

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_bELwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BrainDuo, an EEG foundation model designed to improve EEG representation learning through dual-domain tokenization and structured spatiotemporal pretraining. The framework combines a shared spectrotemporal codebook, discrete and continuous supervision, and multi-scale attention mechanisms. Extensive experiments across 10 EEG tasks and 14 datasets demonstrate strong performance on general EEG benchmarks and competitive results under task-specific BCI evaluation protocols.

Critical strengths
  1. The paper presents extensive experiments with well-designed evaluation protocols, covering both general EEG foundation model benchmarks and task-specific BCI settings. BrainDuo achieves consistent improvements across the general EEG benchmarks and competitive performance in several BCI-specific evaluations.

  2. The paper provides detailed ablation studies to investigate the contributions of individual components. Additional analyses of pretraining versus training from scratch, pretraining data scale, and architectural configurations further support the effectiveness of the proposed design choices.

Critical weaknesses
  1. In my view, the main concern with this work is what new insight or core technical contribution it offers, given the extensive prior research on EEG foundation models. BrainDuo introduces several incremental components, but I do not see a substantial conceptual difference between Duo Codebook, Duo Latent, DuoView Pretraining, and existing approaches. I appreciate the authors' extensive experiments and the strong empirical results. However, it remains unclear whether the improvements come from a genuinely new modeling insight or simply from combining and refining multiple existing techniques. Although the paper provides a number of ablation studies, the contributions of some finer-grained design choices are not sufficiently isolated. Overall, the work feels more like an extensive engineering refinement of existing methods than a significant methodological advance, and as a reader, I find the broader technical takeaway rather limited. Furthermore, I could not find publicly available code or a repository link in the submission, which raises additional concerns about the reproducibility of such a relatively complex framework with many interacting components.

  2. The figures are rather difficult to follow, with dense text and multiple components compressed into limited space, particularly in Figure 1. This makes it unnecessarily difficult to understand the overall framework and its key contributions. I also find the term "BCI training skills" somewhat misleading. It initially gave me the impression that the paper involved some form of agentic or skill-learning framework, whereas these "skills" are essentially conventional BCI preprocessing and task-specific adaptation techniques. A more precise term would help avoid this confusion.

  3. The paper lacks linear probing experiments, making it difficult to directly assess the representation quality of the pretrained model beyond full fine-tuning performance.

Questions for authors
  1. Could the authors provide an anonymized code repository? This would help reviewers reproduce and verify the experimental results.
4 comments
AuthorsOpenReview ↗

Response to Reviewer bELw

We thank the reviewer for recognizing the general-benchmark results and existing ablations, and for the critical questions about novelty, component attribution, presentation, and pretrained representation quality. We have expanded the methodological comparisons and component experiments and added a six-dataset frozen-feature evaluation. The additions are marked in blue.

[W1] Core technical contribution and evidence for individual components

BrainDuo contributes a coupled target-and-context design for EEG pretraining, rather than claiming that waveform reconstruction, STFT, vector quantization, or structured attention is new in isolation. Duo Codebook learns a shared waveform–STFT tokenizer. From that frozen tokenizer, Duo Latent obtains both patch-level discrete codes and electrode-level continuous targets by pooling pre-quantization features over time. DuoView then supplies factorized cyclic context and joint CTopo–time neighborhoods for predicting those targets. This differs from the target sources and context organizations of LaBraM, CodeBrain, PATCHCODE, and CSBrain, as clarified in Appendix J.1, Table 31.

To examine what each choice contributes, we extended the ablations to two additional datasets (Tables 39–41) and added an unmasked-input control (Table 38). In balanced accuracy:

ComparisonFACED-9ISRUC-S1
Full Duo Codebook / shared-MSE tokenizer0.6172 / 0.59710.8098 / 0.7930
CTopo / random masking0.6172 / 0.60340.8098 / 0.7802
Joint / discrete-only supervision0.6172 / 0.63460.8098 / 0.7817

The complete tokenizer has the highest reported means among the tested reconstruction variants, and CTopo outperforms random and montage-distance masking on both added datasets. Joint supervision is not uniformly better: discrete-only wins on FACED-9, whereas joint supervision wins on ISRUC-S1. Separately, masked pretraining improves TUEV balanced accuracy from 0.6118 to 0.7242 over the clean-input control (Appendix L.5, Table 38). These controls attribute benefits to specific choices without claiming that every component helps on every task or that separate ablations establish a higher-order synergy. Appendix J.4 now discusses that scope explicitly.

[W2] Figure readability and the term “BCI training skills”

We appreciate the concern about the dense overview. The original Figures 1 and 2 are retained, but Appendix J.3 adds a panel-level reading guide distinguishing tokenizer learning, the two attention-context patterns, complementary masked passes, and downstream adaptation. We also clarify the terminology in the prose: “BCI training skills” refers to task-specific preprocessing and adaptation, such as filter banks, channel selection, denoising, and trial averaging. These are established BCI procedures integrated into the evaluation pipeline, not a learned skill library or an agentic mechanism. We do not claim that the original figure density has been eliminated.

[W3] Frozen-backbone linear probing

Following the reviewer's suggestion, we added linear probing on six datasets (Appendix L.3, Table 36). The pretrained EEG encoder is frozen and only a linear classifier is trained on its representations. Balanced accuracy against CSBrain is:

DatasetBrainDuoCSBrain
TUEV0.56330.5062
FACED-90.36550.2999
ISRUC0.66280.6703
PhysioNet-MI0.54690.5169
Mumtaz0.94310.8900
BCIC-IV-2a0.45660.4219

BrainDuo has higher balanced accuracy on five of six datasets, with ISRUC favoring CSBrain. These results directly assess pretrained features without full backbone fine-tuning. We also clarify that the existing t-SNE figures visualize fine-tuned rather than frozen representations (Appendix J.5).

[Q1] Anonymized code repository and reproducibility

We understand the importance of executable code for reproducing the results. An anonymized repository is not available during review; we plan to release the implementation and configurations after acceptance. To make the current manuscript more useful for reproduction, the appendices clarify the target definitions, training objectives, data partitions, and evaluation protocols (Appendices A, J, and L). We recognize that these descriptions do not substitute for a runnable repository.

We appreciate the reviewer's suggestions, particularly the request for frozen-feature evaluation, which provides a more direct assessment of the learned representations. We welcome any further questions.

Thank you for the detailed response and additional experiments.

However, my primary concern regarding the limited methodological novelty of BrainDuo remains unresolved. Although the authors have provided further comparisons and explanations in the appendices, I still find that the three core components—Duo Codebook, Duo Latent, and DuoView Pretraining—are largely incremental extensions or combinations of existing techniques, rather than substantial methodological advances. As Reviewer aFYP also pointed out, similar ideas involving waveform and frequency-domain reconstruction have already been explored in CodeBrain, PATCHCODE, and TFM-Tokenizer. While BrainDuo combines these techniques differently, I remain unconvinced that such integration offers a sufficiently distinct technical insight beyond existing EEG foundation models. The additional ablations provide further evidence of empirical effectiveness, but do not fundamentally resolve this concern.

Furthermore, the integration of multiple interacting components introduces additional implementation complexity, making independent reproduction and verification particularly important. Unfortunately, the authors have not provided an anonymized code repository during the review process, and the proposed post-acceptance release does not address the current reproducibility concern.

I therefore maintain my initial rating of 2 (weak rejection).

AuthorsOpenReview ↗

Response to Reviewer bELw's Follow-up

We sincerely thank the reviewer for the continued discussion and for acknowledging our additional experimental evidence. We appreciate the remaining concerns regarding methodological novelty and reproducibility. We would like to further clarify how BrainDuo differs from existing EEG foundation models and why we consider its pretraining design a methodological contribution.

1. Methodological novelty beyond existing EEG foundation models

BrainDuo's contribution is not any single isolated component, but a pretraining framework that connects dual-domain tokenization, supervision at different granularities, and complementary spatiotemporal context modeling. The central idea is to use a common learned tokenizer to construct both local discrete prediction targets and broader continuous alignment targets, while learning EEG representations from distinct spatial and temporal contexts.

First, Duo Codebook and Duo Latent establish a direct connection between tokenization and backbone supervision. Existing approaches construct their learning targets differently. LaBraM uses discrete-code prediction from a separately pretrained tokenizer; CodeBrain learns separate temporal and frequency codebooks; and TFM-Tokenizer learns discrete time–frequency motif tokens. PATCHCODE additionally combines discrete-code supervision with continuous patch-level prediction from a momentum teacher.

BrainDuo instead uses a shared waveform–STFT tokenizer to provide two forms of supervision at different granularities: discrete patch prediction and continuous channel-level alignment. This allows the backbone to learn both local EEG patterns and broader temporal representations from a common target source. The contribution lies in how these targets are constructed and jointly used for pretraining, beyond simply combining existing learning objectives.

Second, DuoView introduces a distinct spatiotemporal context and masking formulation. CodeBrain combines global convolution with sliding-window attention, while CSBrain employs cross-scale temporal and regional modeling. PATCHCODE and REVE also investigate structured masking. DuoView differs by combining periodic factorized temporal/spatial attention with joint CTopo-based electrode–time neighborhoods. The same topology guides structured cluster masking, while complementary masked passes ensure that every patch is predicted from a context in which it is hidden.

The methodological contribution is therefore in how these mechanisms are organized for pretraining: Duo Codebook defines discrete and continuous targets from a common representation, Duo Latent applies them at different granularities, and DuoView provides the spatiotemporal context for learning from masked inputs. We do not claim novelty for STFT, vector quantization, attention, or structured masking individually. The contribution lies in this specific target-construction and contextual prediction formulation.

The expanded ablations support the effectiveness of these design choices across multiple EEG tasks, while also revealing task-dependent behavior. We have clarified the methodological distinctions and their supporting evidence in Appendix J.1–J.4, Table 31, and Tables 39–41 of the revised manuscript.

2. Reproducibility

We understand the reviewer's concern that the complexity of the framework makes independent verification important. We have expanded the manuscript's descriptions of the architecture, training objectives, target construction, data partitions, and evaluation protocols. We recognize that these details do not replace executable code. An anonymized repository is not available during review, and we plan to release the implementation and configurations after acceptance.

We sincerely appreciate the reviewer's careful consideration and constructive feedback. We hope these clarifications better communicate the methodological contribution and its distinction from existing EEG pretraining approaches.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer bELw,

We sincerely appreciate your time and constructive feedback. We hope our responses have helped clarify the concerns raised in your review and response. We would welcome your further thoughts and would be glad to address any remaining questions.

Best regards, The Authors

Reviewer_RiU4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a EEG foundation model pretraining framework by integrating Duo Codebook, Duo Latent and DuoView Pretraining. The Duo Latent module incorporates both discrete codes and continuous embeddings into the pretraining process. DuoView Pretraining utilizes clustered attention by incorporating the spatial topology across neighboring electrodes.

Critical strengths
  1. The paper is well-written and easy to follow. The proposed framework is clearly illustrated with well-designed figures.
  2. The authors conduct extensive comparison and ablation experiments to show the effectiveness of the method.
Critical weaknesses
  1. The motivation behind neighborhood-based pre-training is not sufficiently justified. Furthermore, the multiple strategies proposed in the DuoView pretraining module seem fragmented and incremental. It remains unclear what the two "views" in DuoView explicitly represent and how they specifically map to the introduced architectural modules. The authors should clarify these connections.
  2. The motivation of multi-rate block and clustering masking is not clear. The idea of modeling local and long-range spatiotemporal dependencies is similar to CSBrain. Please specify differences.
  3. The framework of codebook training and code-prediction-based pretraining is very similar to LaBraM, except for additional supervision loss, which makes the current contribution seem somewhat incremental. Please specify the differences.
  4. In the experiment section, what is the fundamental difference between the general EEG foundation-model benchmark and the specialist-facing BCI evaluation? Most EEG foundation models compare on the former, while it seems that BrainDUO performs much better on specialist-facing BCI evaluation compared to other methods. The two tracks also have overlapping datasets. Please clarify the necessity of using this specific evaluation track.
  5. The ablation study lacks the results of Duo Pretraining without any masking strategy.
  6. In the main results (Table 2 &3), could authors report average performance. It facilitates a more direct comparison among different methods.
Questions for authors

Please see weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer RiU4 [1/2]

We thank the reviewer for the positive assessment of the presentation and experimental comparisons, and for the questions about DuoView, its relationship to earlier methods, and the evaluation design. We have clarified the architectural choices, added controls for masking and supervision, and reported aggregate benchmark results. The revisions are marked in blue in the manuscript.

[W1] Motivation for the two views in DuoView

The two views refer to complementary ways of organizing EEG context, not to the two masked forward passes. The Multi-rate (M) block factorizes spatiotemporal interactions: temporal heads connect patches of the same electrode in cyclic groups, while spatial heads connect electrodes at the same time. The Long-range (L) block instead attends jointly to nearby electrode–time pairs, using CTopo distance and temporal separation. Interleaving the blocks exposes the backbone to both factorized and joint context. Separately, two complementary masks ensure that every patch is predicted from a pass in which it is hidden.

The original Table 5 also compares the M-only and L-only variants. On TUEV, balanced accuracy is 0.6853 (M-only), 0.7194 (L-only), and 0.7242 (full); on TUAB, it is 0.8160, 0.8278, and 0.8307, respectively. These results favor using both block types, although the improvement over L-only is small. We have made the distinction between context organization and masked prediction explicit in Appendix J.2.

[W2] Multi-rate grouping, CTopo masking, and distinction from CSBrain

The periods 2 and 3 define cyclic groups of patch indices, not EEG frequency bands. M-block temporal attention and same-time spatial attention are factorized; L-block attention uses joint channel–time neighborhoods. Unlike the matching-position groups and regional descriptors in CSBrain, BrainDuo uses multiple temporal residue classes and CTopo-based electrode proximity. CTopo is calculated on the canonical electrode graph before selecting the channels present in a recording, so a missing electrode does not redefine the distance between retained electrodes. Attention and masking also have different supports: the L block uses temporal radius 7 and CTopo radius 2, while masking uses temporal radius 1 and CTopo radius 2.

Cluster masking is intended to make the model infer hidden patches from context outside their local neighborhood. We added masking comparisons on FACED-9 and ISRUC-S1 (Appendix L.8, Table 41):

Dataset (balanced accuracy)CTopoRandomMontage-distance
FACED-90.61720.60340.5961
ISRUC-S10.80980.78020.7703

Together with the original Table 5, these results support CTopo-based masking across four datasets. We do not interpret this comparison as proof of a specific neural-connectivity mechanism. The structural differences are summarized in Appendix J.1–J.2 and Table 31.

[W3] Relationship to LaBraM and the role of dual supervision

Both BrainDuo and LaBraM use tokenizer-based masked prediction, but their pretraining targets differ. LaBraM predicts discrete codes produced by a spectrum-trained tokenizer. BrainDuo obtains two targets from the same frozen waveform–STFT tokenizer: categorical code identities for individual patches and continuous pre-quantization features averaged over time for each electrode. The backbone predicts the codes and aligns its channel-level summaries with the continuous targets. The distinction is in the source and granularity of the targets, not simply the presence of an additional loss. We also distinguish related continuous-target approaches, including PATCHCODE, in Appendix J.1, Table 31.

The ablations show a task-dependent benefit. Joint versus discrete-only supervision gives balanced accuracy 0.7242 vs. 0.7074 on TUEV, 0.8307 vs. 0.8202 on TUAB, and 0.8098 vs. 0.7817 on ISRUC-S1. On FACED-9, discrete-only is stronger (0.6346 vs. 0.6172). We now state this exception explicitly (original Table 4; Appendix L.7, Table 40).

AuthorsOpenReview ↗

Response to Reviewer RiU4 [2/2]

[W4] Why evaluate both general EEG benchmarks and specialist-facing BCI tasks?

The two tracks answer different questions. The general benchmark tests representation transfer under the common EEG foundation-model evaluation protocols. The specialist-facing track tests the complete decoding pipeline with task-native input construction and adaptation, allowing comparison with task-specific BCI systems. Filter banks, channel selection, denoising, and trial averaging are established operations; we do not present them as individually novel. The added value is to examine where a general pretrained model remains competitive after task-specific adaptation and where specialist methods still lead.

The distinction matters because some dataset names occur under different protocols. For example, SEED-V and FACED-9 have different windowing and subject splits across the two tracks, so their scores are not interchangeable. PhysioNet-MI and BCIC-IV-2a appear in both tracks with the same evaluations to place the foundation models alongside specialist references; they are not counted as independent experiments. Comparisons against specialist pipelines measure entire systems rather than isolating the pretrained backbone. These points and the task-level results are clarified in Appendix J.3 and summarized separately in Appendix K, Table 33.

[W5] Ablation without input masking

We added the requested no-masking control on four datasets (Appendix L.5, Table 38). In both settings, the frozen tokenizer supplies clean-signal targets; the sampled set and its complement determine which code predictions are scored in the two passes, and the alignment loss is retained. The only change is that the no-masking model can see all input patches in both passes.

Dataset (balanced accuracy)Masked pretrainingNo input masking
TUAB0.83070.7887
TUEV0.72420.6118
FACED-90.61720.5141
ISRUC-S10.80980.7170

The masked configuration achieves higher means on all 12 reported metrics. This control supports hiding target patches during pretraining, rather than merely supervising the same positions in an uncorrupted input.

[W6] Average performance for the two evaluation tracks

We added Appendix K, Tables 32–33, rather than changing the original main-text Tables 2–3. For the general benchmark, the equal-dataset mean balanced accuracy is 0.7125 for BrainDuo versus 0.6937 for CSBrain across ten datasets. On the seven datasets shared with CodeBrain, the means are 0.7032 (BrainDuo), 0.6824 (CSBrain), and 0.6775 (CodeBrain) (Table 32).

For the specialist-facing track, Table 33 reports separate means for SSVEP (accuracy 0.7670), emotion recognition (accuracy 0.2819), visual retrieval (Top-1 0.4433), and motor imagery (balanced accuracy 0.6116). We keep these task families separate because their metrics and protocols are not directly interchangeable. All means are descriptive averages of the reported settings, not pooled predictions.

We thank the reviewer again for the constructive suggestions. They helped us clarify the pretraining mechanism and make the two evaluation tracks easier to interpret. We welcome further questions.

Reviewer_aFYPOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes BrainDuo to learn both temporal and spectral information for EEG foundation models. The authors combine a shared time-frequency codebook, discrete code prediction, continuous feature alignment, and spatial-temporal attention. They also use several BCI-specific techniques for downstream tasks. Experiments on 14 datasets show improvements over existing EEG FMs.

Critical strengths

The authors evaluate the model on various EEG tasks.

Critical weaknesses
  1. The novelty appears somewhat limited as far as I know about the literature. Similar ideas of combining waveform and STFT reconstruction have already been investigated in CodeBrain [1], PATCHCODE [2] and TFM-Tokenizer [3]. I understand the authors combine them differently, but what is actually new here? I think simply putting several existing techniques together is not a sufficient contribution.

  2. The tokenizer design should be presented clearer. For example, App. B.1 showed that the authors cropped to 200 coefficients. With the specified FFT and hop size, the original representation contains much more information, why do the authors simply discard most coefficients, which parts are retained?

  3. The significance is questionable here. In Table 4, Duo Codebook only improves BA by 0.0078 on TUEV and 0.0027 on TUAB compared with shared STFT-waveform reconstruction using MSE. Also, Tables 4–5 only evaluate the components on these 2 datasets. I want to see more tests on other datasets, given that the paper claims general representation learning? The paper needs much investigation into where the improvement actually comes from.

  4. I also have more concerns about the generalization claim. For example, Tab. 2 shows only a small improvement on BCIC-IV-2a compared with CSBrain (0.5715 vs 0.5657), with only two subjects in the test set. CHB-MIT also uses only two held-out cases (Appendix C.6).

  5. (minor) Fig.1 is overloaded with tiny labels and difficult to read, broken sentence "Dataset- specific settembedding designs for downstream EEG decoding.ings are detailed in Appendix C.1.", App. C.5 even has unresolved citations written as "(Luck, 2014; ?; ?)". I don't understand why these basic issues were not checked before submission.

Questions for authors

See Weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer aFYP [1/3]

We sincerely thank the reviewer for the careful assessment and for recognizing the breadth of our experimental evaluation. The comments have helped us clarify the methodological contributions and strengthen the validation of BrainDuo. We have uploaded an updated PDF, with all revisions highlighted in blue. We address each concern below.

[W1] Novelty beyond existing EEG pretraining methods

We thank the reviewer for the opportunity to clarify our methodological contributions. BrainDuo addresses how discrete local EEG patterns and continuous representations can jointly guide masked pretraining. Our contributions are dual-domain tokenization with supervision at different granularities and complementary spatiotemporal context modeling with topology-guided masking.

1. Dual-domain tokenization and cross-granularity supervision.

Duo Codebook learns a shared waveform–STFT tokenizer that produces both continuous representations and their quantized codes. Duo Latent uses these outputs at two granularities: discrete codes supervise masked patch prediction, while temporally pooled pre-quantization features provide channel-level alignment targets. Both targets originate from the same frozen tokenizer and jointly supervise the backbone.

This differs from CodeBrain's separate temporal/frequency codebooks, LaBraM's discrete-code prediction, and PATCHCODE's momentum-teacher continuous prediction at the patch level. The methodological distinction lies in how categorical and continuous targets are derived from a common tokenizer and applied at different granularities.

2. Complementary spatiotemporal context modeling and topology-guided masking.

DuoView introduces two complementary attention structures. The Multi-rate (M) block separately models periodic temporal dependencies within each channel and spatial interactions across channels. The Long-range (L) block captures joint spatial–temporal relationships through neighborhoods defined by CTopo distance and temporal separation.

DuoView further employs topology-guided cluster masking, jointly hiding neighboring electrodes and adjacent temporal patches to encourage prediction beyond local correlations. A mask and its complement are processed through two shared-weight passes, ensuring that every patch is predicted while hidden.

This design differs from CodeBrain's global convolution and sliding-window attention, CSBrain's cross-scale regional modeling, and the region- or block-masking approaches of PATCHCODE and REVE. The methodological distinction is the combination of periodic factorized attention, joint topology-aware attention, and complementary topology-guided masked prediction within one backbone.

3. Why this constitutes a methodological contribution.

BrainDuo introduces a pretraining design in which a single waveform–STFT tokenizer provides two forms of supervision: discrete codes for masked patch prediction and temporally averaged continuous features for channel-level alignment. Discrete prediction encourages learning local EEG patterns, while continuous alignment provides a training signal at a broader temporal granularity without requiring patch-level feature reconstruction. DuoView further supports this learning process through distinct spatiotemporal attention structures and topology-guided masking. The contribution lies in how these training targets are constructed and applied at different granularities, together with the proposed context modeling and masking strategy.

4. Quantitative evidence for the proposed components.

To validate these contributions, we extended the component ablations to four datasets. Each entry reports the control's balanced accuracy, followed by the improvement from the complete BrainDuo configuration in percentage points.

Component comparisonTUEVTUABFACED-9ISRUC-S1
Duo Codebook vs. separate codebooks0.6918 (+3.24)0.8297 (+0.10)0.5873 (+2.99)0.7768 (+3.30)
Duo Latent vs. discrete-only0.7074 (+1.68)0.8202 (+1.05)0.6346 (−1.74)0.7817 (+2.81)
CTopo vs. random masking0.7091 (+1.51)0.8201 (+1.06)0.6034 (+1.38)0.7802 (+2.96)
DuoView M+L vs. L-only0.7194 (+0.48)0.8278 (+0.29)——
Complete BrainDuo0.72420.83070.61720.8098

The complete Duo Codebook outperforms separate codebooks across all four datasets, while CTopo-guided masking consistently outperforms random masking. Joint discrete–continuous supervision improves three of four datasets, with discrete-only performing better on FACED-9.

These results support the proposed design choices while identifying task-dependent effects. We have clarified the methodological distinctions from prior work in Appendix J.1–J.2 and Table 31, with additional ablations in Tables 39–41 and the original Table 5.

AuthorsOpenReview ↗

Response to Reviewer aFYP [2/3]

[W2] Selection of 200 STFT coefficients

The choice of 200 coefficients follows the tokenizer decoder width, rather than an empirically optimized EEG frequency threshold. For each one-second patch, the STFT produces 33×26=85833 \times 26 = 858 frequency–time coefficients. Frequency-major flattening and truncation retain the lowest frequency bins, extending to approximately 22 Hz. Although higher frequencies are excluded from direct spectral reconstruction, the waveform reconstruction objective still provides full-band supervision.

To assess this choice, we added a full-STFT tokenizer comparison (Appendix L.4, Table 37):

DatasetOriginal tokenizer (BA)Full-STFT (BA)
TUAB0.83070.8244
TUEV0.72420.7048
FACED-90.61720.6490
ISRUC-S10.80980.8045

Full-STFT provides better higher-frequency spectral reconstruction, whereas the original configuration achieves higher waveform reconstruction correlation (0.692 vs. 0.633). Downstream results are mixed: the original tokenizer performs better on three datasets, while full-STFT performs better on FACED-9.

These results indicate that retaining all STFT coefficients does not consistently improve downstream decoding. We have clarified the rationale and limitations of this design choice in Appendix L.4.

[W3] Modest reconstruction gains and limited component ablations

We appreciate the reviewer's concern about the small reconstruction-loss improvements and limited evaluation scope. The additional gains from the full reconstruction objective over shared STFT–waveform MSE are indeed modest: 0.78 percentage points on TUEV and 0.27 on TUAB (original Table 4). We have clarified that these differences should not be interpreted as evidence of statistical significance or as the main contribution of Duo Codebook.

More importantly, we extended the Duo Codebook, Duo Latent, and masking ablations to FACED-9 and ISRUC-S1. The new results are summarized below in balanced accuracy.

ComponentVariantFACED-9ISRUC-S1
Duo CodebookFFT only0.55800.7580
Waveform only0.54800.7523
STFT only0.57440.7613
Shared STFT + waveform MSE0.59710.7930
Separate codebooks0.58730.7768
Duo Codebook0.61720.8098
Duo LatentDiscrete-only0.63460.7817
Continuous-only0.56340.7554
Joint supervision0.61720.8098
MaskingRandom0.60340.7802
Montage-distance0.59610.7703
CTopo-guided0.61720.8098

The complete Duo Codebook achieves the highest reported means among the six tokenizer variants on both additional datasets. CTopo-guided masking also outperforms the two alternative strategies. For Duo Latent, joint supervision improves ISRUC-S1, while discrete-only supervision remains stronger on FACED-9.

Together with the original TUEV and TUAB experiments, these additions establish a broader component-level evaluation across four datasets, while also identifying where the gains depend on the downstream task. The new experiments are included in Appendix L.6–L.8, Tables 39–41, with the discussion updated in Appendix J.4.

AuthorsOpenReview ↗

Response to Reviewer aFYP [3/3]

[W4] Generalization beyond the original test subjects and cases

Following the reviewer's suggestion, we substantially expanded the evaluation coverage. We now report nine subject-held-out folds on BCIC-IV-2a and five patient-grouped folds covering 22 distinct patients on CHB-MIT. BrainDuo and CSBrain use matched folds and validation-based checkpoint selection.

BCIC-IV-2a (Appendix L.2, Table 35). Balanced accuracy across all nine held-out subjects:

SubjectBrainDuoCSBrain
10.58510.5833
20.41490.4670
30.67010.7118
40.46530.4306
50.50000.4844
60.49830.4392
70.60760.6094
80.60070.5434
90.60420.5660
Mean0.54960.5372

CHB-MIT (Appendix L.1, Table 34). Results across five patient-grouped folds:

FoldBrainDuo BACSBrain BABrainDuo AUC-PRCSBrain AUC-PR
00.8310.7640.2690.217
10.7170.6840.1080.146
20.8230.7340.2180.135
30.8820.8600.4480.440
40.6500.7050.1490.164
Mean0.78060.74940.23840.2204

For CHB-MIT, chb01 and chb21 are grouped as the same patient, preventing patient overlap between training and testing.

Across the expanded evaluations, BrainDuo achieves higher mean balanced accuracy than CSBrain on both BCIC-IV-2a (+1.24 percentage points) and CHB-MIT (+3.12 percentage points). Individual subject and fold results are included to show the variation in performance, including settings where CSBrain performs better.

These experiments provide broader evidence of subject- and patient-level generalization beyond the original fixed partitions.

[W5] Presentation and editorial issues

We have corrected the incomplete sentence in Section 3.5, unresolved citations, and additional language inconsistencies. We also expanded the explanation of the overview figures in Appendix J.3 to clarify the pretraining targets, context modeling, and evaluation setup. The original figures have been retained.

We sincerely appreciate the reviewer's constructive suggestions. They have helped us clarify the methodological contributions and strengthen the experimental validation of BrainDuo. We hope the revised manuscript and these additional results address the concerns, and we welcome further discussion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their careful reading, constructive feedback, and detailed assessment of our work. We appreciate the positive recognition of BrainDuo's broad evaluation across EEG tasks, competitive benchmark performance, and extensive experimental comparisons.

The reviews raised important questions regarding methodological novelty, the contributions of individual components, and the generalizability of the learned representations. We have revised the manuscript and conducted additional experiments to address these concerns. All revisions are highlighted in blue. The main updates are summarized below.

  1. Methodological contributions. We clarify how Duo Codebook, Duo Latent, and DuoView form a unified pretraining framework, and distinguish their target construction and context modeling from CodeBrain, LaBraM, CSBrain, and related methods (Appendix J.1–J.3, Table 31).

  2. Expanded component ablations. We extend tokenizer, dual-supervision, and masking comparisons from TUEV/TUAB to FACED-9 and ISRUC-S1 (Tables 39–41), and add a no-masking control across four datasets (Table 38) and a full-STFT comparison (Table 37). These experiments clarify individual design choices and identify task-dependent effects, especially for joint supervision.

  3. Broader subject-level generalization. We evaluate nine subject-held-out folds on BCIC-IV-2a and five patient-grouped folds covering 22 patients on CHB-MIT. BrainDuo achieves higher mean balanced accuracy than CSBrain, 0.5496 vs. 0.5372 and 0.7806 vs. 0.7494, respectively (Tables 34–35). We also report individual folds to show variability across subjects.

  4. Pretrained representation quality. We add frozen-backbone linear probing on six datasets (Table 36), where BrainDuo achieves higher balanced accuracy than CSBrain on five of six datasets, independently of full fine-tuning.

  5. Additional clarification and reporting. We explain the general EEG and specialist-facing BCI evaluation tracks, add aggregate benchmark summaries (Tables 32–33), clarify the STFT coefficient selection, and correct presentation and citation issues.

We believe these revisions clarify BrainDuo's contribution and strengthen the experimental evidence. We are grateful for the reviewers' suggestions and welcome further discussion.

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.