acceptodds
Under review as a conference paper at ICLR 2027

MGT: Multi-Granularity Transfer for Cross-Subject EEG Adaptation

Abstract

Building EEG models that can adapt reliably to unseen subjects across diverse paradigms remains challenging because EEG signals vary substantially across subjects and tasks. Although existing approaches increasingly emphasize transferable EEG representations, knowledge acquired from different sources does not contribute equally when adapting to unseen subjects, and indiscriminate transfer inevitably induces negative transfer and degrades cross-subject generalization. We therefore propose Multi-Granularity Transfer (MGT), a general transfer framework for cross-subject EEG adaptation. MGT jointly models transfer relationships in heterogeneous EEG knowledge at multiple granularities, enabling the model to preserve universally transferable patterns while suppressing domain-mismatched knowledge. We further introduce a progressive transfer mechanism to stabilize knowledge adaptation as learning proceeds, thereby improving the robustness of cross-subject adaptation. We systematically evaluate MGT on seven EEG datasets spanning five paradigms, including motor imagery, emotion recognition, and imagined speech. Under cross-subject evaluation protocols, MGT consistently achieves robust decoding performance across these diverse paradigms. The results demonstrate the effectiveness of modeling transferability at multiple granularities and highlight the potential of MGT for cross-subject EEG decoding.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Electroencephalogram (EEG), General-Purpose EEG Models, Multi-Granularity Transfer, Cross-Subject EEG Decoding, Brain-Computer Interface (BCI)

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_p33BOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EEG decoders generalize poorly to unseen subjects. MGT weights source data by estimated transferability at two levels. A reference model is first trained on source and target-calibration data and then frozen. In its feature space, a subject-level weight comes from entropy-regularized optimal transport between each source subject's class-wise features and the target's (Eq. 2-3), and a sample-level weight from distances to target class prototypes and a cluster-to-target OT cost (Eq. 4-8). The two are multiplied, normalized per class and clipped (Eq. 9-10), kept fixed, and ramped in over the first 100 epochs (Eq. 11) on the source terms of the GAT objective (Eq. 12). Inference needs neither source data nor target calibration labels (Sec. 2.3).

Evaluation covers seven datasets and five paradigms (motor imagery, emotion, imagined speech, RSVP, ERN), with subject-averaged balanced accuracy, its subject-wise standard deviation, Kappa and AUROC (Sec. 3.1). MGT reports the best primary metric on all six datasets in the main text, with margins over re-run GAT (official updated) from +0.20 on ERN to +2.35 on RSVP (Tables 2 and 7).

Critical strengths

S1. The method is simple to add. Transfer weights are computed once in a frozen reference space and stay fixed during final training (Sec. 2.1, 2.3), and they only replace the source average in the loss (Eq. 12). In principle the approach could be added to backbones other than GAT, although only GAT is tested.

S2. For six of the seven datasets, preprocessing is described carefully enough to rule out the leakage that is easy to introduce in cross-subject EEG work. App. B.1 states that normalization statistics come only from the training or calibration partition (on BCI-IV-2a, "Per-channel z-score statistics are estimated exclusively from the training partition"). The SEED paragraph gives no normalization procedure.

Critical weaknesses

W1. The gaps to the literature baselines in Table 2 cannot be attributed to MGT. MGT and GAT (official updated) use a substantial labeled target-calibration set (80% of the target subject on RSVP, Sessions 1-4 with 240 trials on ERN, and the target's full official training partition on BCI-IV-2a; App. B.1), while App. C.1 says the other results are "literature-reported" and never states their calibration budget. The 18.0-point gap over BioCodec on ERN (78.89 vs 60.90) and the 9.8-point gap over CST-TVA-DRTL on RSVP (83.70 vs 73.95) are already present for GAT (official updated) before MGT is applied. Either rerun the baselines under this protocol or restrict quantitative claims to the GAT comparison.

W2. Hyperparameters appear to have been selected on the evaluation data, and the per-subject evidence for the coarse-grained factor is weak. App. B.1 defines only a calibration and an evaluation partition for each dataset, and the word validation does not appear. Figure 6(a) reports balanced accuracy on BCI-IV-2a and peaks at lambda_proto = 0.5, the default used for all reported results; with no validation split, the only held-out data are the target evaluation partition. Using the per-subject values in App. D.2-D.4, a two-sided paired sign test does not find the removal of Coarse-Grained Transfer significant on any dataset: the full model wins on 11 of 15 subjects on SEED (p = 0.12), 6 of 8 decided on BCI-IV-2a (p = 0.29) and 3 of 5 decided on NM000252 (p = 1.00). A Wilcoxon signed-rank test reaches p = 0.03 on SEED but not on BCI-IV-2a (p = 0.08) or NM000252 (p = 1.00), and with five decided pairs the sign test cannot reach 0.05 on NM000252 at all. By the sign test, removing Progressive Transfer on SEED (p = 0.002) and Fine-Grained Transfer on BCI-IV-2a (p = 0.039) is significant. The conclusion's statement that the ablations support all three components should be qualified for the coarse-grained one.

W3. The related work is too narrow. MEKT (Zhang & Wu, 2020) already estimates subject-level transferability to choose source subjects for a target. It appears in the paper only as the MEKT-R baseline on RSVP and is not discussed with the selective-transfer work, so what the coarse-grained weight adds over it is not addressed.

W1 and W2 can be addressed in revision: W1 by rerunning the baselines under the same calibration protocol or narrowing the claims to the GAT comparison, and W2 by stating how the hyperparameters were chosen and qualifying the ablation conclusion.

Questions for authors

Q1. How many runs and seeds are behind Tables 2, 3 and 7, and which data were used to select lambda_proto, E_ramp, m, beta_min, tau_t and the clipping bounds, and to stop training early in the Figure 6 runs? The Std columns are defined across subjects (Sec. 3.1), and four of seven datasets show gains under one point over GAT (official updated) against subject-wise standard deviations of 6.00 to 13.22.

Q2. Table 3 has no row with both granularities removed. Setting all w_i = 1 reduces Eq. (12) to the ordinary source average, so GAT (official updated) should be that reference, but the paper does not say whether it shares the seed and schedule of the ablation runs. Could you add that row, with paired per-subject tests for each ablation?

Q3. How does MGT relate to existing work on weighting sources? Besides MEKT, source instances have been reweighted using labeled target data, for example in TrAdaBoost (Dai et al., 2007), and multi-source adaptation has learned per-domain weights, for example Domain Aggregation Networks (Wen et al., ICML 2020). In its discussion of selective transfer, the introduction cites Wang et al. (2024b) and Cui et al. (2025). Is the claim that the two levels had not been combined before, or that these particular estimators are new?

Q4. In Table 2 the Std column is exactly 0.00 for xDAWN-SVM on RSVP (11 subjects) and for CBraMod and STELAR-base on ERN (16 subjects), and the other ERN literature baselines lie between 0.20 and 0.89, whereas GAT (official updated) and MGT show 10.80 and 10.27 on the same data. These values do not look like the subject-wise standard deviation defined in Sec. 3.1. What does the Std column measure for each literature baseline, and under what calibration budget were these values obtained?

Q5. For every new target subject, MGT trains an 80-epoch reference model on the source pool together with that subject's calibration data, estimates the transfer weights, and then trains the final model from a new initialization (Sec. 2.1; 80 epochs in Sec. 3.1). The conclusion notes this added computation, but no training time is reported. How long does this take per target compared with GAT alone? Together with the labeled calibration set it requires (App. B.1), this decides how practical the method is for a new user.

6 comments
AuthorsOpenReview ↗

Response to Reviewer p33B(Part 1)

We sincerely thank the reviewer for the thoughtful and constructive comments. We appreciate the positive assessment of the simplicity of MGT and its evaluation across diverse EEG paradigms. We have carefully considered the concerns regarding comparison fairness, model selection, statistical evidence, related work, and computational efficiency. In the following, we address each point in detail and describe the corresponding revisions.

W1. Attribution of Improvements over Literature Baselines

Response: Thank you for raising this important concern. We agree that performance differences between methods using different target-calibration protocols cannot be attributed solely to MGT. The original GAT study (Song et al., 2023) provides relevant evidence for the contribution of source knowledge and domain adaptation. On BCI-IV-2a, the target-only model achieved 62.96% accuracy, compared with 64.54% for source–target training without adaptation and 76.58% for full GAT. On BCI-IV-2b, full GAT achieved 84.44% accuracy, compared with 79.95% for the target-only model. These results are summarized in Appendix C.2, Table 8 of our revised manuscript. They demonstrate the importance of source–target adaptation in the original GAT framework and help explain why differences in training protocols can substantially affect cross-method comparisons.

To ensure a clearer comparison, we distinguish literature-reported results from controlled experiments. On BCI-IV-2a, seven comparison methods were re-implemented using the same target calibration/test partition as MGT. For the remaining datasets, literature-reported results are retained as references rather than treated as strictly controlled comparisons. Our primary method comparison focuses on GAT (official updated), which uses the same v0.72 backbone and evaluation protocol as MGT. These distinctions are clarified in Appendix C.1–C.2.

Under the shared evaluation settings, MGT improves BAcc over GAT (official updated) by 2.02, 0.45, 1.60, 0.84, 2.41, and 0.25 percentage points on BCI-IV-2a, OpenBMI, SEED, NM000252, RSVP, and ERN, respectively. An additional improvement of 0.55 percentage points is observed on BCI-IV-2b (Appendix D.1).

We have also added subject-wise statistical analyses on BCI-IV-2a to further assess the significance of these improvements. The Friedman test identifies significant differences among the nine compared methods (χ²(8) = 48.80, p < 0.001). Two-sided paired Wilcoxon signed-rank tests with Holm correction show significant improvements over all eight comparison methods (all adjusted p ≤ 0.0391). In particular, MGT outperforms GAT (official updated) on seven of nine target subjects, with a mean paired improvement of 2.02 percentage points (adjusted p = 0.0391). These results are reported in Section 3.2, Table 3, with subject-wise data and five-seed results provided in Appendix D.6–D.7.

We have therefore clarified the scope of our quantitative claims. The literature-reported results provide broader context, while the controlled comparisons with GAT (official updated) and the supplementary statistical analyses provide more direct evidence of MGT's effectiveness.

AuthorsOpenReview ↗

Response to Reviewer p33B(Part 2)

W2. Hyperparameter Selection and Statistical Evidence for Coarse-Grained Transfer

Response: Thank you for raising these important questions. We clarify the target-data partitioning and model selection protocol, as well as the contribution of coarse-grained transfer. Target data partitioning and model selection. The Target Calibration Set and Target Test Set are two separate, non-overlapping partitions of the target subject's data. The labeled calibration set is used for adaptation, transferability estimation, and model selection, while the test set is reserved for final evaluation. Specifically, BCI-IV-2a uses the official training and test partitions, RSVP adopts a stratified 80%/20% split, and ERN uses Sessions 1–4 for calibration and Session 5 for testing. These protocols are described in Section 3.1 and Appendix B.1.

As stated in Appendix B.2, “The key hyperparameters were selected using a simple grid search based on performance on the target calibration set.” The selected configuration is summarized in Table 6. The nominal cluster size (m) is fixed at 64 across datasets. Checkpoints are selected according to calibration performance, as described in Section 3.1, with the separate test set used for final evaluation.

For the sensitivity experiments in Figure 6, each target subject has a maximum training budget of 200 epochs, with early stopping patience set to 30 epochs. Training terminates when the monitored calibration performance fails to improve for 30 consecutive epochs. Early stopping is applied independently to each target subject.

Contribution of coarse-grained transfer. As shown in Table 4, removing coarse-grained transfer reduces the average BAcc across the five reported datasets, demonstrating its contribution to the overall framework. The subject-wise results in Appendix D.2–D.4 further reveal that its contribution varies across target subjects.

Regarding the reviewer's statistical analysis, the Wilcoxon signed-rank test yields (p=0.03) on SEED, while the corresponding results are (p=0.08) on BCI-IV-2a and (p=1.00) on NM000252. The sign-test results are (p=0.29), (p=0.12), and (p=1.00), respectively. These results reflect differences in how the two tests evaluate paired subject-level improvements, with the Wilcoxon test additionally accounting for the magnitude and ranking of paired differences.

We further distinguish these component-level analyses from the evaluation of the complete MGT framework. As reported in Section 3.2, Table 3, MGT improves BAcc over GAT (official updated) by 2.02 percentage points across the nine target subjects, with improvements on seven subjects and a Holm-adjusted (p=0.0391). This provides additional statistical evidence for the effectiveness of the complete MGT framework.

Taken together, the average ablation results support the contribution of coarse-grained transfer, while the subject-wise analyses characterize its varying effects across target subjects. The statistical comparison with GAT further demonstrates the effectiveness of the integrated MGT framework.

AuthorsOpenReview ↗

Response to Reviewer p33B(Part 3)

W3. Related Work on Selective Transfer Methods

Response: Thank you for suggesting these relevant studies. We have expanded Appendix A (Related Work) to discuss existing selective transfer methods and clarify the methodological contributions of MGT. The revised manuscript introduces three complementary directions:

1.MEKT (Zhang & Wu, 2020): “aligns EEG covariance representations on a Riemannian manifold and performs domain adaptation in the tangent space for unsupervised cross-subject classification.”

2.TrAdaBoost (Dai et al., 2007): “iteratively adjusts source-instance contributions within a boosting framework to exploit labeled target data.”

3.Domain Aggregation Networks (Wen et al., 2020): “dynamically balance contributions from multiple source domains to reduce negative transfer.”

These discussions are included in Appendix A, together with the corresponding references.

We further describe the specific design of MGT in the same section: “MGT jointly estimates subject-level transferability through class-conditional optimal transport and trial-level transferability through prototype and local-cluster costs in a frozen reference space.”

“The resulting transfer factors are fused and progressively incorporated into source-dependent training objectives, capturing complementary variations across source subjects and individual trials.”

The methodological formulation is detailed in Sections 2.1–2.3, Eqs. (1)–(12). Specifically, the coarse-grained factor characterizes source-subject compatibility, while the fine-grained factor evaluates individual trial compatibility using target-class prototypes and local cluster structures. Their multiplicative fusion and progressive integration provide a unified mechanism for regulating source contributions during cross-subject EEG adaptation.

Q1. Experimental Repetitions, Hyperparameter Selection, and Stopping Criteria

Response: Thank you for requesting these clarifications.Regarding experimental repetitions, MGT and GAT (official updated) on BCI-IV-2a were evaluated using five random seeds. MGT was also evaluated using five seeds (2022–2026) on RSVP and ERN. The corresponding results are reported in Appendix D.6 and D.9–D.10. The other MGT benchmark results use single runs, while literature baselines retain the results reported in their original studies.

Five-seed evaluation summary across three EEG datasets.

DatasetMethodBAcc (Mean ± SD) (%)Additional Metric
BCI-IV-2aMGT83.11 ± 0.08--
BCI-IV-2aGAT (official updated)81.09 ± 0.28--
RSVPMGT86.11 ± 1.43AUROC: 91.67 ± 0.25
ERNMGT79.14 ± 0.20AUROC: 83.95 ± 0.03

For the statistical comparisons in Table 3, each target subject is treated as one paired observation. The five random seeds are averaged at the subject level rather than treated as independent observations, as clarified in Appendix D.7. The BCI-IV-2b comparison, originally presented in Table 7, is now included in Appendix D.1.

Regarding hyperparameter selection, the Target Calibration Set and Target Test Set are separate data partitions. Hyperparameters are selected through grid search based on target calibration performance, as described in Appendix B.2. The final configurations are summarized in Table 6. The nominal cluster size m is fixed at 64 across datasets. Figure 6 further examines the sensitivity to the prototype–cluster balance and progressive ramp duration on BCI-IV-2a.

For the Figure 6 experiments, each target subject has a maximum training budget of 200 epochs. Early stopping is applied independently to each subject, with a patience of 30 epochs. Training terminates when no performance improvement is observed for 30 consecutive epochs.

Checkpoint selection is based on target calibration performance, while the separate Target Test Set is reserved for final evaluation, as described in Section 3.1.

Reviewer_UKkbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

MGT addresses supervised cross-subject EEG adaptation by assigning source-level and sample-level transfer weights. These weights are estimated using optimal-transport discrepancies, target-class prototypes, and source clusters in a frozen reference space. They are progressively introduced during training of an independently initialized GAT model. Experiments cover seven datasets across five EEG paradigms.

Critical strengths
  1. The evaluation covers seven datasets and explicitly documents the labeled target-calibration protocols. The distinction between the original GAT and the updated implementation is also clearly acknowledged.

  2. MGT improves subject-averaged balanced accuracy over the updated GAT on all seven datasets. The subject-wise ablations provide useful detail about variation across targets rather than relying exclusively on aggregate scores.

Critical weaknesses
  1. MGT uses substantial labeled target calibration, including 288 trials on BCI-IV-2a and two sessions on SEED. A target-only model under the same architecture and augmentation would help establish how much source knowledge contributes and whether negative transfer is actually avoided.

  2. Balanced-accuracy gains over updated GAT are 0.45 points on OpenBMI, 0.55 on BCI-IV-2b, and 0.20 on ERN. Subject-wise standard deviations do not establish uncertainty in these paired differences, and repeated-training variability is not reported.

  3. Figures 2 and 5 and Table 4 show weight heterogeneity and proximity to target calibration samples. However, distance-based weights and representations trained using those weights make these patterns insufficient evidence that high-weight samples actually improve held-out target performance more than low-weight samples.

  4. The study lacks matched comparisons with simpler distance-based weighting and separate prototype-only or cluster-only variants. The sensitivity sweep uses mixtures from 0.1 to 0.9, so it does not establish that both fine-grained terms are necessary.

  5. All experiments instantiate MGT with the updated GAT backbone, while component ablations cover only three datasets. Results with another backbone and component analyses on RSVP or ERN would help establish whether the mechanism generalizes beyond this particular adaptation system.

  6. Although Table 5 specifies many hyperparameters, the manuscript does not clearly state the main final-training duration, checkpoint-selection procedure, or training-seed protocol. It also provides no measured cost for the additional reference-model training and transport estimation, limiting assessment of the gain relative to its computational expense.

Questions for authors
  1. Which literature baselines use the same labeled target-calibration budget and evaluation partitions? How does MGT compare with target-only learning under the same GAT implementation and augmentation?

  2. Can paired subject-level confidence intervals and results across independent training seeds be provided for MGT versus updated GAT?

  3. What are the post-clipping mean weights for each class, and what final-training, checkpoint-selection, and computational-budget settings are used?

2 comments
AuthorsOpenReview ↗

Response to Reviewer UKkb(Part 1)

We thank the reviewer for the constructive feedback and for recognizing the extensive evaluation and subject-wise analysis. We have conducted additional experiments and analyses to address the main concerns.

1.Target Calibration and Source Knowledge Contribution

Response: MGT is designed for supervised target-calibrated cross-subject adaptation rather than zero-shot transfer. The labeled calibration budgets are explicitly reported in Table 1, including 288 trials for BCI-IV-2a and two sessions for SEED.

Regarding the target-only comparison, GAT is a cross-subject domain adaptation method that employs adversarial training on source and target data to learn subject-invariant representations. Removing the source domains would eliminate the source–target adaptation component and result in a different target-only learning setting. When only target-domain data are considered, GAT reduces to a hybrid deep learning model based on a CNN–Transformer architecture. Therefore, we did not separately implement a target-only GAT variant. Instead, we cite the target-only results reported in the original GAT study, which correspond to approximately 62.96% and 79.95% classification accuracy on BCI-IV-2a and BCI-IV-2b, respectively, as derived from the reported performance differences (seen in Appendix C.2, Table 8). These results are based on the original GAT architecture rather than the updated v0.72 implementation. We retain all comparison methods in Table 2 while using GAT (official updated) as the primary reference for protocol-matched comparisons across datasets, without claiming superiority over target-only learning.

  1. Statistical Significance and Random-Seed Variability

Response: We conducted five-seed experiments (2022–2026) on BCI-IV-2a for both MGT and GAT (official updated), using identical evaluation partitions. MGT achieves 83.11 ± 0.08% BAcc, compared with 81.09 ± 0.28% for GAT, yielding a 2.02 percentage-point improvement (seen in Appendix D.6, Table 17).

The paired analysis across nine subjects shows that MGT performs better on seven subjects, with a 95% paired confidence interval of [0.15, 3.89] percentage points and a Holm-adjusted exact Wilcoxon p-value of 0.0391. Importantly, the confidence interval is calculated using subjects as paired units, rather than treating different random seeds as independent subjects (seen in Table 3; Appendix D.7, Table 18).

We also conducted five-seed experiments on RSVP and ERN to assess random-seed variability. The five-seed mean results are now reported in Table 2, with detailed results for individual seeds provided in Appendix D.9–D.10, Tables 20–21.

3.Practical Utility of Transfer Factors

Response: Beyond the visualization of transfer-factor distributions, we have added a direct decoding comparison between high- and low-factor source samples on BCI-IV-2a. High-factor sources achieve 72.46% average BAcc, compared with 67.75% for low-factor sources, with better results on seven of nine target subjects.

Target decoding performance using high- and low-factor source samples on BCI-IV-2a.

SubjectHigh-factor BAcc (%)Low-factor BAcc (%)Difference (pp)
S0183.3376.74+6.59
S0260.4247.92+12.50
S0389.4887.50+1.98
S0470.8371.18−0.35
S0542.3634.03+8.33
S0659.3357.64+1.69
S0781.6069.79+11.81
S0882.1980.56+1.63
S0982.6484.38−1.74
Mean72.4667.75+4.72

Note: Difference is calculated as High-factor BAcc minus Low-factor BAcc.

This result provides quantitative evidence that the estimated transfer factors distinguish source samples with different downstream adaptation utility, rather than merely reflecting their geometric proximity to target samples.(seen in Section 3.4; Appendix D.5, Table 16).

  1. Distance-based Baselines and Fine-grained Components

Response: Our existing sensitivity analysis varies the relative contributions of prototype and cluster costs, with the best BAcc obtained when both are equally weighted (seen in Section 3.4, Fig. 6(a)).

We acknowledge that additional controlled comparisons would help further isolate the contributions of these components. Due to the limited rebuttal period, we have not yet completed the distance-only weighting, prototype-only transfer, and cluster-only transfer experiments. We will conduct these additional experiments after the rebuttal period to provide further validation under the same evaluation protocol as MGT.

AuthorsOpenReview ↗

Response to Reviewer UKkb(Part 2)

  1. Backbone Generality and Component Ablations

Response: We have extended component ablations to five datasets, including the newly added RSVP and ERN experiments. On both RSVP and ERN, removing coarse-grained, fine-grained, or progressive transfer consistently reduces BAcc and AUROC compared with Full MGT. The detailed results are reported in Table 4 (seen in Section 3.3).

All current experiments use the updated GAT backbone. Due to the limited rebuttal period, we have not yet completed an evaluation using a second EEG backbone. We will further investigate the applicability of the proposed transfer mechanism to other EEG architectures after the rebuttal period.

  1. Training Configuration and Computational Cost

Response: We clarify that the reference model is pretrained for 80 epochs, followed by final training with a maximum budget of 1,000 epochs. The progressive transfer period is 100 epochs, with a batch size of 64 and an Adam learning rate of 2 × 10⁻⁴ (seen in Section 3.1; Appendix B.2, Table 6). These settings refer to BCI-IV-2a. Apart from the additional reference-model pretraining stage, the standard training configuration is kept consistent with GAT (official updated), while progressive transfer is specific to MGT.

Checkpoint selection is based on the highest balanced accuracy on the target calibration set, retaining the earliest checkpoint in case of ties. The target test set is reserved for final evaluation (seen in Section 3.1). Due to the limited rebuttal period, we have not yet completed the analysis of post-clipping mean transfer factors for each class or the stage-wise computational cost evaluation. We will provide these additional analyses after the rebuttal period, including a runtime comparison covering reference-model pretraining, transfer-factor estimation, and final-model training.

Answers to the Questions for Authors

Q1.Baseline calibration budgets and target-only comparison

Response: For BCI-IV-2a, seven comparison methods were reimplemented using the same target calibration and test partitions as MGT. For other datasets, results are primarily taken from the original publications, whose calibration protocols are not necessarily identical. The updated GAT provides our main matched comparison. The distinction between target-only learning and cross-subject adaptation is explained in Section 1 above.

Q2. Paired confidence intervals and independent training seeds

Response: The BCI-IV-2a five-seed comparison provides a 2.02 percentage-point improvement, a 95% paired confidence interval of [0.15, 3.89], and an adjusted p-value of 0.0391. Additional five-seed results for MGT on RSVP and ERN are reported in Table 2, with detailed results provided in Appendix D.9–D.10, Tables 20–21.

Q3. Post-clipping factors and training settings

Response: Fused factors are normalized to a class-wise mean of one before clipping. The final-training budget and checkpoint-selection criterion are clarified in Section 6. Due to the limited rebuttal period, we have not yet completed the analysis of post-clipping mean transfer factors for each class or the stage-wise computational cost evaluation. We will provide these additional analyses after the rebuttal period.

We thank the reviewer again for the valuable suggestions. We believe that the additional statistical evidence, transferability analysis, and expanded ablation experiments provide stronger support for the effectiveness of MGT across different EEG paradigms.

Code Availability. BCI-IV-2a implementation: https://anonymous.4open.science/r/MGT-2C13/

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to Reviewer fd1j(Part 1)

We thank the reviewer for the careful evaluation and constructive feedback. We have conducted additional experiments and analyses to address the two major concerns regarding the statistical reliability of MGT's performance gains and the comparability of cross-dataset baselines. Specifically, we provide multi-seed evaluations and paired statistical analyses, clarify the contribution of the updated GAT backbone, and strengthen the evidence for the practical utility of the proposed transfer factors. We also address the reviewer's six questions individually below.

  1. Statistical Significance and Random-Seed Variability

Response: We have conducted five-seed experiments (2022–2026) comparing MGT with GAT (official updated) on BCI-IV-2a under identical evaluation protocols. MGT achieves 83.11 ± 0.08% BAcc, compared with 81.09 ± 0.28% for GAT, yielding a 2.02 percentage-point improvement. MGT outperforms GAT on seven of nine subjects, with a 95% paired confidence interval of [0.15, 3.89] percentage points and a Holm-adjusted Wilcoxon p-value of 0.0391.

Subject-wise comparison between MGT and GAT (official updated) on BCI-IV-2a.

SubjectMGT (Mean ± SD) (%)GAT (Mean ± SD) (%)Difference (pp)
S0190.50 ± 0.6990.33 ± 0.69+0.17
S0268.33 ± 0.8362.26 ± 1.86+6.07
S0391.28 ± 0.6191.45 ± 0.29−0.17
S0484.14 ± 1.2382.25 ± 0.87+1.89
S0573.27 ± 1.1768.67 ± 0.81+4.60
S0672.22 ± 1.6167.56 ± 0.39+4.66
S0792.16 ± 1.1491.34 ± 0.57+0.82
S0888.31 ± 0.7587.88 ± 0.79+0.43
S0987.81 ± 0.8688.07 ± 1.07−0.26
Overall BAcc83.11 ± 0.0881.09 ± 0.28+2.02

Note: Mean ± SD represents the mean and sample standard deviation across five random seeds. The overall BAcc standard deviation is calculated across the five run-level averages. Difference is calculated using the five-seed subject means.

We have also conducted five-seed MGT experiments on RSVP and ERN. To provide a more comprehensive assessment, we will extend the matched multi-seed comparison to GAT and MGT across all seven datasets and report per-dataset paired significance tests and confidence intervals.

2.Separating Backbone Improvements from the MGT Contribution

Response: We agree that GAT (official updated) is the appropriate reference for evaluating the contribution of MGT, since both methods use the same v0.72 backbone. As described in Appendix B.3, this backbone consists of complementary temporal–spatial and spatial–temporal convolutional branches, followed by token projection and a Transformer with self-attention and cross-attention operations. MGT retains this architecture and introduces multi-granularity transfer factors to regulate source-dependent training objectives.

On BCI-IV-2a, GAT (official updated) achieves 81.09% BAcc, while MGT achieves 83.11%, demonstrating a 2.02 percentage-point improvement under the same backbone and evaluation setting (seen in Table 2).

Performance comparison between GAT (official updated) and MGT across seven datasets.

DatasetGAT BAcc (%)MGT BAcc (%)Improvement (pp)
BCI-IV-2a81.0983.11+2.02
BCI-IV-2b84.7885.33+0.55
OpenBMI82.2082.65+0.45
SEED79.8481.44+1.60
NM00025261.9462.78+0.84
RSVP83.7086.11+2.41
ERN78.8979.14+0.25

Regarding other comparison methods, seven baselines on BCI-IV-2a were reimplemented using the same target calibration and test partitions as MGT. For the remaining datasets, comparison results are primarily taken from the corresponding original publications. We therefore use GAT (official updated) as the primary reference for matched comparisons across datasets, while retaining the literature-reported results as additional context (seen in Appendix C.1–C.2).

AuthorsOpenReview ↗

Response to Reviewer fd1j(Part 2)

  1. Independent Validation of Transferability

Response: We have added a direct downstream decoding comparison between high- and low-factor source samples on BCI-IV-2a. High-factor sources achieve 72.46% average BAcc, compared with 67.75% for low-factor sources, corresponding to a 4.72 percentage-point improvement. High-factor sources perform better on seven of nine target subjects. The paired t-test and Wilcoxon signed-rank test yield p-values of 0.0274 and 0.0391, respectively.

Target decoding performance using high- and low-factor source samples on BCI-IV-2a.

SubjectHigh-factor BAcc (%)Low-factor BAcc (%)Difference (pp)
S0183.3376.74+6.59
S0260.4247.92+12.50
S0389.4887.50+1.98
S0470.8371.18−0.35
S0542.3634.03+8.33
S0659.3357.64+1.69
S0781.6069.79+11.81
S0882.1980.56+1.63
S0982.6484.38−1.74
Mean72.4667.75+4.72

Note: Difference is calculated as High-factor BAcc minus Low-factor BAcc.

Unlike the feature-space proximity analysis, this experiment evaluates actual target decoding performance and therefore provides independent evidence of the practical utility of the estimated transfer factors (seen in Section 3.4; Appendix D.5, Table 16).

  1. Effect of Target Calibration Budget

Response: We acknowledge that evaluating MGT under different target calibration budgets would provide further insight into its applicability with limited target supervision. After the rebuttal period, we will conduct a calibration-budget sensitivity experiment on BCI-IV-2a, using progressively smaller, class-stratified subsets of the 288 labeled target calibration trials. MGT and GAT will use identical calibration subsets and training configurations. This experiment will directly examine whether the advantage of multi-granularity transfer is maintained as target supervision becomes increasingly limited.

  1. Reference-to-Final Representation Consistency

Response: The reference encoder provides a fixed representation space for estimating source–target transferability. The estimated factors characterize the relative compatibility of source subjects and samples with the target distribution. They are then transferred to an independently initialized final model with the same GAT architecture. Keeping the reference encoder frozen avoids continuously changing the transfer-factor estimation space during final optimization.

To further examine this design choice, we will conduct a controlled comparison between fixed reference-space factors and periodically re-estimated factors from the evolving final encoder after the rebuttal period.

  1. ERN AUROC Performance

Response: On ERN, MGT improves BAcc from 78.89% to 79.14%, while AUROC changes slightly from 83.99% to 83.95%. These differences are small and reflect the distinct characteristics of the two metrics: BAcc evaluates balanced classification performance at a decision threshold, whereas AUROC measures ranking performance across thresholds. Importantly, the ERN component ablations show that removing any of the three proposed transfer components reduces both BAcc and AUROC relative to Full MGT. Thus, the results support the contribution of the proposed components, although we do not claim an AUROC improvement over GAT.

We thank the reviewer again for the constructive feedback. The additional statistical comparisons, controlled baseline evaluations, component ablations, and targeted sensitivity analyses included in the revised manuscript provide a more comprehensive assessment of the contribution and applicability of MGT.

Code Availability. BCI-IV-2a implementation: https://anonymous.4open.science/r/MGT-2C13/

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.