acceptodds
Under review as a conference paper at ICLR 2027

NeuroWeave: Frequency-Factorized Multi-Slot Discrete Targets for Masked EEG Pre-Training

Abstract

Discrete masked EEG pre-training depends on both target structure and target capacity. A single broadband code must compress heterogeneous rhythms into one discrete decision; increasing its vocabulary does not create independently representable signal components. We introduce \model, which factorizes each one-second channel patch into delta, theta, and alpha bands and represents each band with multiple discrete slots. Independent tokenizers generate frozen targets, and a shared spatiotemporal Transformer predicts them from masked broadband input. We evaluate target design in stages, separating tokenizer capacity, masked reconstruction, target-band selection, and downstream transfer. Increasing slot count improves tokenizer fidelity more than enlarging a single-slot codebook, while masked reconstruction reveals finite, frequency-dependent useful capacity. Under matched controls, frequency-factorized targets improve low-frequency reconstruction over an equally sized broadband VQ target. Downstream ablations select delta, theta, and alpha, whereas beta and gamma provide no consistent benefit. Across six datasets, the resulting three-band model surpasses single-slot, broadband VQ, and waveform-regression controls on every reported metric and achieves state-of-the-art performance on the reported benchmarks. Code will be released.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG foundation models, self-supervised learning, vector quantization, discrete representation learning

Reviews

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

NeuroWeave is a masked EEG pre-training method. Its targets are discrete codes from separate band-specific VQ tokenizers (delta, theta, alpha). Each tokenizer gives SbS_b code slots per one-second channel patch (Sections 3.2-3.4). A shared Transformer takes broadband masked patches and predicts all codes with a band- and slot-normalised cross-entropy (Eqs. 8-9). The tokenizers are frozen and dropped after pre-training. Section 4 is a staged study. It covers tokenizer reconstruction when codebook size KK or slot count SS changes (Fig. 2), masked-patch reconstruction against broadband VQ and waveform-regression controls (Fig. 3, Table 1), band selection on TUAB and TUEV (Table 2), and fine-tuning on six datasets against random-init models, BIOT, LaBraM, CBraMod, BrainOmni and three internal controls (Tables 3, 4, 8). The paper reports that the three-band S=8S=8 model beats the controls on every metric, and it claims state-of-the-art results.

Critical strengths
  1. The design is controlled and checkable. The paper separates which bands are supervised from how many slots each gets, and tests each with a named control: S1, RawVQ-S32 and RawRegression downstream (Tables 3, 4, 8), and RawVQ-5 and RawVQ-80 for reconstruction (Table 1). I checked the statement "lower NMAE than their matched broadband VQ controls in every dataset-band comparison" (L362-364) against Table 1, and it holds for all ten rows. This makes the central design claim testable, and it is the main reason I take the paper seriously.

  2. Variability and limits are reported with care. Downstream results are five-seed mean ± std (L230-231), and Appendix E gives subject-level partitions and sample counts for TUAB and TUEV. The paper also says the slot sweep is not parameter- or compute-matched and uses one seed (L324-327), and that the external comparisons are descriptive (L428-431). A reader can judge how far the results go.

Critical weaknesses
  1. Band selection uses the reported data, and the state-of-the-art claim is too broad. The delta-theta-alpha set is picked from Table 2 on TUAB and TUEV (L406-407, L420-421). The selected row of Table 2 (82.85 / 91.27 / 91.63 and 69.68 / 0.7039 / 85.28) is the same as the NeuroWeave (S8) row of Table 3. So the numbers that pick the configuration are the numbers reported for it. The paper does say TUAB and TUEV are selection tasks and the other four datasets are not, so the issue is limited to these two. The margins are mixed. On TUAB, the AUROC and PR-AUC gaps to the four- and five-band models are 0.6 to 1.1 points, bigger than the stds. For TUAB balanced accuracy (82.85±0.06 vs 82.36±0.51 and 82.48±0.59), they are about one std. On TUEV, the gaps to the four-band model are within one std, and the four-band model has the higher TUEV balanced accuracy (69.70±2.42 vs 69.68±1.70). Against the five-band model, the TUEV gap is large. The paper doesn't say which split Table 2 uses. Separately, the abstract (L026-028) and conclusion (L484-485) claim state-of-the-art, but L428-431 says the external comparisons are descriptive because corpora, architectures and adaptation pipelines differ. Several margins are inside one std: SHU-MI against BrainOmni (66.71±2.08 vs 65.98±1.17, Table 4), PhysioNet-MI against CBraMod (59.80±0.70 vs 59.32±1.38), and S8 against S1 on SHU-MI (66.71±2.08 vs 66.32±2.00). I couldn't find parameter counts, pre-training hours or compute for NeuroWeave or the baselines (Appendix D gives hours only for Sleep-EDF). If Table 2 is a test split, the authors could select on validation, report test only for the chosen set, and add a significance test for the band comparison. That would fix the first part. Narrowing the claim or adding matched-data comparisons would fix the second.

  2. Neither the pre-training benefit nor the case for multi-slot targets is separated out. Tables 3, 4 and 8 have no randomly initialised NeuroWeave backbone. The random-init rows are other architectures. This matters, because on SSVEP the random-init ST-Transformer is close to the pre-trained models: 48.32±1.20 vs 49.09±1.37 on BETA-SSVEP and 75.90±1.88 vs 78.96±1.85 on Benchmark-SSVEP (Table 8). For slot count, the paper compares the KK sweep (S=1S=1, up to 12 bits per band) with the SS sweep (up to 128 slots of 10 bits). To call slot count "the dominant tokenizer-capacity axis" (L068-069), it needs equal-budget pairs, for example S=2S=2 with K=32K=32 against S=1S=1 with K=1024K=1024. At S=128S=128, NMAE is still high for a band limited to 3 Hz (delta about 0.35 on TUAB, Fig. 2). Codebook usage isn't reported, so I can't tell whether the codebooks are used. Downstream, only S=1S=1 and S=8S=8 are tested, so the case for multi-slot targets rests on two points.

A few smaller points. Beta and gamma are dropped based on two clinical datasets, but the abstract says they "provide no consistent benefit" without that scope. No four-band variant is run on the motor-imagery or SSVEP sets. The "negative transfer" explanation (L405-406) is untested, since adding a band also adds 8 targets and changes the per-band weight 1/∣B∣1/|B| in Eq. 9. Every band gets S=8S=8, but Fig. 3 shows a different best SS per band and Eq. 9 allows different SbS_b. Table 7 gives one set of fine-tuning settings, and I couldn't find whether the baselines used it or were tuned (BrainOmni gets 15.00±7.66 on BETA-SSVEP, below ContraWR and ST-Transformer, Table 8). As far as I can tell, the five seeds are downstream only. Appendix A describes BandVQ (Sukhbaatar et al., 2026) as the closest work, since it also uses band-specific VQ tokenizers. From the paper's description, the stated difference is that factorization and per-band slot count are treated as separate choices, with several slots per band (L148-150, L624-626). No experiment compares with BandVQ, FAME, EEG-X or MANAS-2, and I couldn't check these 2026 preprints. Minor: Eq. 5 doesn't give the forms of LcodeL_{code} and LcommitL_{commit} or the value of λcom\lambda_{com}; the beta and gamma filters are not in Section 3.2 or Table 5; the per-patch NMAE used in Fig. 3 and Table 1 is not written out; y^\hat y in Eq. 8 is not defined in the text; Appendix B (L646-647) mentions entries marked "–" that I did not find in Tables 5-7; Fig. 1 has "Mutli"; there is no limitations section.

Questions for authors
  1. Which split is Table 2 computed on, and what do the five band sets give on validation?
  2. What does the same backbone give on all six datasets without pre-training?
  3. In the SS sweep, what happens if KK is reduced so that total bits match the K=4096K=4096, S=1S=1 case? And what is codebook utilisation at S=64S=64 and 128128?
  4. Do the TUAR, TUEP, TUSZ and TUSL pre-training recordings share patients or sessions with the TUAB and TUEV test partitions? I couldn't find a statement that they don't.
  5. How do parameter count and pre-training hours compare with LaBraM, CBraMod and BrainOmni? Were the baselines fine-tuned with the Table 7 settings?
4 comments
AuthorsOpenReview ↗

Response to Reviewer Twyt

Thank you for your precise questions and for identifying limitations in the original band-selection and capacity claims.

Q1. Table 2 split and validation results for five band sets

A1. The original Table 2 used test results to motivate band selection; this was inappropriate. We repeated the full comparison on TUAB and TUEV validation splits. Revised Section 4.4/Table 2 reports the five-seed means below (standard deviations are in the table); Table 3 is held-out test evaluation.

BandsTUAB BAcc / AUROC / PR-AUC (%)TUEV BAcc (%) / κ / W. F1 (%)
δ+θ+α+β+γ81.83±0.22 / 89.70±0.23 / 90.77±0.3164.87±6.97 / 0.7203±0.0077 / 87.72±0.40
δ+θ+α+β82.01±0.29 / 90.10±0.29 / 91.33±0.4171.10±0.80 / 0.7671±0.0077 / 89.79±0.31
δ+θ+α82.90±0.16 / 91.10±0.19 / 92.55±0.1571.35±2.69 / 0.7760±0.0081 / 90.03±0.36
δ+θ82.18±0.46 / 90.96±0.22 / 92.42±0.1971.00±1.03 / 0.7721±0.0080 / 89.68±0.39
δ81.96±0.72 / 90.67±0.12 / 92.12±0.1170.73±1.64 / 0.7585±0.0106 / 89.36±0.41

Validation again selects δ+θ+α on both datasets, but this does not excuse the original test-based choice. We make no significance claim for differences between band sets.

Q2. Same backbone without pre-training

A2. Appendix C (Table 8) compares the same backbone trained from random initialization with NeuroWeave (S=8S=8) across all six datasets. Both settings use five downstream seeds and validation-selected checkpoints. Difference is the pre-trained mean minus the from-scratch mean, in each metric's displayed units.

DatasetMetricWithout pre-trainingNeuroWeave (S8)Difference
TUABBAcc (%)75.83 ± 0.6882.85 ± 0.06+7.02
AUROC (%)84.36 ± 0.3091.27 ± 0.11+6.91
PR-AUC (%)83.58 ± 0.3991.63 ± 0.11+8.05
TUEVBAcc (%)33.39 ± 2.4669.68 ± 1.70+36.29
Cohen's κ0.3675 ± 0.04810.7039 ± 0.0114+0.3364
Weighted F1 (%)67.92 ± 2.7985.28 ± 0.54+17.36
SHU-MIBAcc (%)58.00 ± 0.7766.71 ± 2.08+8.71
AUROC (%)60.64 ± 1.1473.59 ± 2.43+12.95
PR-AUC (%)59.03 ± 1.6974.11 ± 1.83+15.08
PhysioNet-MIBAcc (%)28.47 ± 1.2559.80 ± 0.70+31.33
Cohen's κ0.0463 ± 0.01670.4640 ± 0.0093+0.4177
Weighted F1 (%)27.93 ± 1.3059.95 ± 0.81+32.02
Benchmark-SSVEPBAcc (%)68.29 ± 2.0778.96 ± 1.85+10.67
Cohen's κ0.6748 ± 0.02120.7842 ± 0.0189+0.1094
Weighted F1 (%)68.42 ± 2.1979.14 ± 1.87+10.72
BETA-SSVEPBAcc (%)9.39 ± 0.9649.09 ± 1.37+39.70
Cohen's κ0.0707 ± 0.00980.4778 ± 0.0141+0.4071
Weighted F1 (%)8.96 ± 0.8449.33 ± 1.48+40.37

The pre-trained model has a higher mean on every reported metric under this same-backbone control.

AuthorsOpenReview ↗

Response to Reviewer Twyt

Q3. Equal-bit slot comparison and codebook use at large SS

A3. Appendix E.1 (Tables 10–11) compares TUAB validation global NMAE at B=Slog⁡2KB=S\log_2K bits. At 12 bits, all tested settings are:

K,SK,Sδ / θ / α / β / γ NMAE
4096; 10.51709 / 0.68806 / 0.71840 / 0.86350 / 0.85320
64; 20.57265 / 0.69782 / 0.71984 / 0.88286 / 0.90061
16; 30.62249 / 0.74608 / 0.74790 / 0.88975 / 0.92955
8; 40.64894 / 0.73719 / 0.75372 / 0.89241 / 0.94107
4; 60.67361 / 0.74607 / 0.78669 / 0.89807 / 0.93734
2; 120.72450 / 0.82032 / 0.84468 / 0.92108 / 0.95787

At 12 bits, one slot has the lowest error in every band. At 20 bits, the larger-codebook comparison is:

K,SK,Sδθαβγ
1,048,576; 10.473090.598960.634440.814400.84213
1024; 20.483920.547950.599790.827640.86869

The 20-bit result is mixed. Section 4.2 now limits the slot-count claim to its tested sweeps; Section 5 notes that downstream SS was evaluated only at 1 and 8. Appendix E.2 (Tables 12–13) reports active-code fraction across S=1S=1–128. For the requested large settings:

Validation setSSδ / θ / α / β / γ active codes (%)
TUAB6497.37 / 99.96 / 100 / 100 / 100
TUAB12897.36 / 99.98 / 100 / 100 / 100
TUEV6485.35 / 96.29 / 98.74 / 99.74 / 99.81
TUEV12888.20 / 94.72 / 97.77 / 99.68 / 99.59

This measure counts entries seen at least once per slot, averaged across slots. It shows no severe complete codebook collapse under this measure, while coverage varies by dataset, band, and SS.

AuthorsOpenReview ↗

Response to Reviewer Twyt

Q4. Patient/session overlap with TUAB and TUEV test partitions

A4. We exclude TUH-derived recordings matching the patient or recording session of a TUAB or TUEV evaluation example before building the pre-training corpus. The retained TUAR, TUEP, TUSZ, and TUSL recordings share neither patients nor sessions with the evaluation partitions. This is now explicit in Section 4.1 and Appendices G and H.

Q5. Model scale, pre-training duration, and baseline fine-tuning

A5. Appendix I (Table 16) reports the parameter counts of the evaluated implementations and their pre-training corpus durations:

ModelParametersCorpus duration
NeuroWeave5.820M3,415.48 h EEG
LaBraM-Base5.820Mapproximately 2,500 h EEG
CBraMod69.286Mover 9,000 h EEG
BrainOmni14.524M1,997 h EEG + 656 h MEG
EEG-X0.040Mapproximately 4 h EEG (STEW)
MANAS-292.8Mapproximately 60,000 h EEG

NeuroWeave and LaBraM-Base use the same backbone and have the same parameter count; NeuroWeave is larger than EEG-X and smaller than BrainOmni, CBraMod, and MANAS-2. Corpus durations describe scale rather than a controlled data-efficiency comparison. Appendix B states that NeuroWeave and all baseline models use the downstream settings in Table 7 for each dataset.

Addressed weaknesses: performance claims and scope. Section 4.5 treats external comparisons descriptively; the Abstract qualifies the result as the highest reported mean among evaluated methods under our protocol. The Abstract and Section 4.4 restrict the β/γ finding to TUAB/TUEV validation. Section 5 states that other downstream slot counts and nonuniform per-band slot allocations were not tested, and clarifies that standard deviations cover five fine-tuning seeds from one pre-trained checkpoint.

Addressed weaknesses: omitted baselines. Appendix D (Table 9) adds BandVQ-style, MANAS-2, and EEG-X comparisons. Because BandVQ's code was unavailable, we independently implemented its paper-described target design within the NeuroWeave backbone and used the same downstream protocols; this is not an official-code result. MANAS-2 and EEG-X are separate model baselines. Balanced accuracy (%) is:

DatasetMANAS-2EEG-XNeuroWeave (S8)
TUAB80.99±0.6975.24±0.2282.85±0.06
TUEV65.57±2.6638.91±5.6269.68±1.70
SHU-MI62.58±0.6153.87±0.9166.71±2.08
PhysioNet-MI56.87±1.5338.01±2.7059.80±0.70
BETA-SSVEP48.09±1.709.68±0.7749.09±1.37
Benchmark-SSVEP75.71±2.2417.15±1.3578.96±1.85

Appendix D explains that FAME pre-trained weights were unavailable at evaluation time, so we do not report a fine-tuning result.

Addressed weaknesses: method notation and metric. Section 3.2 defines the tokenizer loss terms and commitment weight in Eq. (5); Section 3.4 defines the prediction logits in Eq. (8). Sections 4.2–4.3 now use the same global NMAE of Eq. (10) for Figure 3 and Table 1. The earlier “median patch-level NMAE” wording was a writing error; no patch-wise median was used and the results are unchanged. Appendix B specifies the β and γ filter cutoffs. Section 5 supplies the requested limitations.

Thank you for the thorough response. Your response addresses my main soundness concerns, and I appreciate your candor about the original test-based selection. The equal-bit results are informative: at 12 bits, a single slot gives the lowest NMAE in every band, and at 20 bits, the picture is mixed. This suggests the benefit of multi-slot targets may come largely from total target capacity rather than slot count per se. Given that downstream K is tested only at 1 and 8, I would encourage the final version to frame the multi-slot contribution accordingly. I have raised my soundness score to 3. I keep my overall rating at weak acceptance, as the core design choices are now more honestly scoped, but their advantage over equal-budget alternatives is limited.

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

The authors propose NeuroWeave, a masked EEG pre-training scheme with frequency-factorized discrete targets. To this end, each one-second channel patch is filtered into delta, theta, and alpha bands, each band is represented by several discrete codes from its own tokenizer, and a Transformer predicts these codes for masked patches from the broadband input. The method is evaluated in stages, from tokenizer capacity to transfer on six datasets, against internal controls and published EEG foundation models.

Critical strengths

Firstly, the paper poses a very clear design question: which frequency bands should be supervised, and how much discrete capacity should each receive? The internal controls share the backbone, masking, and downstream protocol, and they do in fact favour the selected configuration across all six datasets, four of which were not used to select the bands. I therefore think the gain over the internal controls is credible.

Secondly, the comparison with a broadband target of the same total size shows that factorising by band leads to better reconstruction of theta and alpha. This is a useful result in the context of the current debate on spectral bias in EEG foundation models.

Critical weaknesses

First, it seems that the band sets were chosen using the test set. The band selection results in Table 2 are the same as the test results in Table 3 for both TUAB and TUEV. This matters because on TUAB the only band set that beats the one-slot control on all metrics is the one that was chosen. The authors mention in Section 4.5 that TUAB and TUEV are used to choose the band sets, but they still include them in the state-of-the-art claim in the abstract. In addition, all datasets use one split, the seeds appear to be used only for fine-tuning, and many margins on the motor imagery and SSVEP datasets are within one standard deviation.

Second, the capacity study does not seem to support its claim. Figure 3 looks at reconstruction for each band, but the paper never tests an unbalanced distribution of slots across bands, since all bands are given eight slots. The sweep is also not balanced in information: adding slots increases the bits per patch linearly, whereas increasing the codebook size does so only logarithmically (1280 vs 12 bits). Lower reconstruction error with more slots is therefore expected, and it does not show that slot count is the dominant capacity axis. Finally, Section 4.5 states that the external comparisons are descriptive, yet the state-of-the-art claim is based on re-runs of these models, some of which are well below their published results (e.g., CBraMod gets 59.32 balanced accuracy on PhysioNet-MI here vs 64.17 published, a value NeuroWeave does not reach either, although the splits differ). Finaly, EEG-X is discussed but not compared.

Questions for authors

Does Table 2 report validation or test results? If the selection was made on test results, it should be repeated on the validation splits. It would be useful to see a tokenizer comparison at an equal bit budget, ideally including residual quantization, as well as a BandVQ-style target under the same backbone. Regarding the pre-training data, it is unclear whether patients in the TUAB and TUEV test splits were excluded from the TUH corpora used for pre-training, and results from more than one pre-training seed would help. Finally, the SHU-MI result should be qualified, since the trial-based split in Appendix E is not directly comparable to the cross-subject splits used in published results.

3 comments
AuthorsOpenReview ↗

Response to Reviewer sLtr

Thank you for your detailed methodological comments. We have corrected the band-selection procedure and added controlled tokenizer comparisons.

Q1. Test-set band selection and Table 2

A1. We agree that using test results to choose the target bands in the original submission was inappropriate. We reran all five configurations on the TUAB and TUEV validation splits. Table 2 in Section 4.4 now contains these values; the test sets are reserved for final evaluation in Table 3.

BandsTUAB BAcc / AUROC / PR-AUC (%)TUEV BAcc (%) / κ / W. F1 (%)
δ+θ+α+β+γ81.83±0.22 / 89.70±0.23 / 90.77±0.3164.87±6.97 / 0.7203±0.0077 / 87.72±0.40
δ+θ+α+β82.01±0.29 / 90.10±0.29 / 91.33±0.4171.10±0.80 / 0.7671±0.0077 / 89.79±0.31
δ+θ+α82.90±0.16 / 91.10±0.19 / 92.55±0.1571.35±2.69 / 0.7760±0.0081 / 90.03±0.36
δ+θ82.18±0.46 / 90.96±0.22 / 92.42±0.1971.00±1.03 / 0.7721±0.0080 / 89.68±0.39
δ81.96±0.72 / 90.67±0.12 / 92.12±0.1170.73±1.64 / 0.7585±0.0106 / 89.36±0.41

Values are five-seed mean ± standard deviation. Validation selects δ+θ+α for both datasets. Its agreement with the originally selected set does not validate the original test-based procedure.

Q2. Equal-bit comparison and residual quantization

A2. The original capacity sweeps increased KK with S=1S=1 or increased SS with K=1024K=1024. They show that additional slots alleviate the single-slot reconstruction bottleneck within those settings, but do not establish which capacity axis is preferable at a fixed bit budget. We have refined the conclusion in Section 4.2 to reflect this scope. To examine the budget question directly, Appendix E.1 compares B=Slog⁡2KB=S\log_2K on the complete TUAB validation set using global NMAE (lower is better):

BitsKKSSδθαβγ
12409610.517090.688060.718400.863500.85320
126420.572650.697820.719840.882860.90061
121630.622490.746080.747900.889750.92955
12840.648940.737190.753720.892410.94107
12460.673610.746070.786690.898070.93734
122120.724500.820320.844680.921080.95787
201,048,57610.473090.598960.634440.814400.84213
20102420.483920.547950.599790.827640.86869

At 12 bits, the single-slot setting has the lowest error in every band. At 20 bits, two slots improve θ and α, while one slot improves δ, β, and γ. The 20-bit comparison targets a larger-codebook regime; matching two 1024-entry slots with one slot requires a 1,048,576-entry codebook. Neither budget establishes a universal slot advantage.

Appendix E.3 (Table 14) adds RVQ under matched K=1024K=1024 per stage/slot, nominal bit budget, training data, optimization steps, and optimizer. Here SS means sequential residual stages for RVQ and parallel slots for NeuroWeave. The TUAB validation NMAE values are:

SSRVQ δ/θ/α/β/γNeuroWeave δ/θ/α/β/γ
10.558480 / 0.743944 / 0.788387 / 0.905440 / 0.8722500.547505 / 0.732223 / 0.745642 / 0.881264 / 0.865470
80.463363 / 0.513915 / 0.516163 / 0.751109 / 0.8989790.399396 / 0.490465 / 0.436568 / 0.578077 / 0.836892
320.398493 / 0.440170 / 0.416580 / 0.756437 / 0.9413380.372308 / 0.419636 / 0.359199 / 0.450268 / 0.745243

NeuroWeave has lower NMAE in all 15 matched comparisons. Together, these analyses define the scope of the capacity finding while reinforcing the paper's central result: frequency-factorized multi-slot targets achieve lower NMAE than matched broadband VQ controls in every comparison in Section 4.3 (Table 1), and outperform the internal target-design controls across all six downstream datasets in Section 4.5 (Tables 3–4) and Appendix F (Table 15).

AuthorsOpenReview ↗

Response to Reviewer sLtr

Q3. BandVQ-style target under the same backbone

A3. BandVQ's implementation code was not publicly available at the time of our experiments, so we could not run the authors' implementation. We instead independently implemented a BandVQ-style pre-training target following the method described in their paper, used it with the same backbone and downstream protocols as NeuroWeave, and compared both targets at S=8S=8 on all six datasets. Thus, this is a paper-based implementation of the target design, not a result from BandVQ's official code. Appendix D (Table 9) documents the setup and results:

DatasetMetricBandVQ-styleNeuroWeave (S8)
TUABBAcc (%)81.70 ± 0.2482.85 ± 0.06
AUROC (%)90.35 ± 0.3091.27 ± 0.11
PR-AUC (%)90.19 ± 0.3791.63 ± 0.11
TUEVBAcc (%)64.68 ± 0.7869.68 ± 1.70
Cohen's κ0.6453 ± 0.01230.7039 ± 0.0114
Weighted F1 (%)82.46 ± 0.6885.28 ± 0.54
SHU-MIBAcc (%)65.82 ± 0.6966.71 ± 2.08
AUROC (%)72.64 ± 0.6773.59 ± 2.43
PR-AUC (%)73.16 ± 0.5474.11 ± 1.83
PhysioNet-MIBAcc (%)59.27 ± 0.7059.80 ± 0.70
Cohen's κ0.4569 ± 0.00930.4640 ± 0.0093
Weighted F1 (%)59.02 ± 0.7059.95 ± 0.81
BETA-SSVEPBAcc (%)48.37 ± 2.1149.09 ± 1.37
Cohen's κ0.4705 ± 0.02170.4778 ± 0.0141
Weighted F1 (%)48.89 ± 2.3149.33 ± 1.48
Benchmark-SSVEPBAcc (%)77.23 ± 1.9678.96 ± 1.85
Cohen's κ0.7665 ± 0.02010.7842 ± 0.0189
Weighted F1 (%)77.59 ± 1.9379.14 ± 1.87

NeuroWeave has the higher reported mean on every metric in this comparison.

Q4. Patient/session overlap with the TUH pre-training corpus

A4. Before corpus construction, we excluded TUH-derived recordings matching a patient or recording session in the TUAB or TUEV evaluation partitions. The retained TUAR, TUEP, TUSZ, and TUSL pre-training recordings share neither patients nor sessions with those partitions. This rule is now stated in Section 4.1 and Appendices G and H.

Q5. Multiple pre-training seeds

A5. We used one pre-training seed and five fine-tuning seeds per downstream dataset, as in common EEG foundation-model evaluation protocols. Thus, the reported standard deviations measure fine-tuning variability from one checkpoint; they do not measure variation across independent pre-training runs. We now state this limitation in Section 5.

Q6. SHU-MI trial split versus published cross-subject results

A6. We agree that our trial-based SHU-MI split is not directly comparable with published cross-subject results. Every method in our Table 4 was retrained and evaluated on the same stratified trial split; no published SHU-MI score was reused. Appendix H now states this distinction explicitly.

Addressed weaknesses. Section 4.5 treats external-model comparisons as descriptive because corpora and adaptation pipelines differ; the Abstract says “highest mean performance among the evaluated methods under our protocol” rather than an unrestricted state-of-the-art claim. Section 4.4 reports the validation results without attributing the four-/five-band differences to negative transfer. Section 5 states that nonuniform per-band slot allocations and additional downstream slot counts remain untested. Appendix D adds EEG-X and MANAS-2 as separate model baselines. EEG-X versus NeuroWeave BAcc (%) is 75.24 versus 82.85 (TUAB), 38.91 versus 69.68 (TUEV), 53.87 versus 66.71 (SHU-MI), 38.01 versus 59.80 (PhysioNet-MI), 9.68 versus 49.09 (BETA-SSVEP), and 17.15 versus 78.96 (Benchmark-SSVEP). FAME could not be fine-tuned under our protocol because its pre-trained weights were unavailable at evaluation time.

Thank you for the detailed response and the additional experiments. My main concerns have been addressed, and I will raise my score.

Reviewer_7hA4OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposed a method to factorize raw EEG signal into multiple discrete slots, with several independent tokenizers and a shared spatiotemporal Transformer.

Critical strengths

The idea is interesting and might create large inspiration to EEG large models: for model construction, training and potential explanations.

Critical weaknesses

The experimantal results lack comparation with other EEG transformers, so the perfromance is not perfectly convincing.

Questions for authors

Why comparation with state of art EEG transformers/large models?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 7hA4

Thank you for highlighting the need for comparisons with EEG Transformers and foundation models.

Q1. EEG Transformer and foundation-model comparisons

A1. These comparisons were already included in the original submission: ST-Transformer, BIOT, LaBraM, CBraMod, and BrainOmni were evaluated alongside task-specific baselines on all six datasets. They appeared in the submitted Tables 3, 4, and 8; in the revised paper, they appear in Tables 3 and 4 (Section 4.5) and Table 15 (Appendix F). The balanced-accuracy results (%) below are five-seed mean ± standard deviation.

DatasetNeuroWeave (S8)Best external comparator in the corresponding table
TUAB82.85 ± 0.06LaBraM 81.43 ± 0.30
TUEV69.68 ± 1.70LaBraM 66.10 ± 0.98
SHU-MI66.71 ± 2.08BrainOmni 65.98 ± 1.17
PhysioNet-MI59.80 ± 0.70CBraMod 59.32 ± 1.38
BETA-SSVEP49.09 ± 1.37ST-Transformer 48.32 ± 1.20
Benchmark-SSVEP78.96 ± 1.85ST-Transformer 75.90 ± 1.88

NeuroWeave has the highest reported mean balanced accuracy in these six comparisons. Some margins are small relative to the reported standard deviations; we do not claim statistical significance. Section 4.5 clarifies that comparisons across different pre-training corpora and adaptation pipelines are descriptive.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.