acceptodds
Under review as a conference paper at ICLR 2027

HarmTSFormer: Harmonic Encoding and Evidence-Preserving Fusion for Unseen-Frequency SSVEP Decoding

Abstract

Brain-computer interfaces based on steady-state visual evoked potentials (SSVEPs) identify attended flashing targets from electroencephalography (EEG) recordings. Decoders typically restrict their output to specific stimulus frequency categories, making it difficult to directly extend their performance to stimulus frequencies not encountered during training or not defined in the model. Although previous studies support unseen-frequency decoding, capturing within-candidate response structure and using information across candidates remain challenges for reliable cross-subject recognition. This paper proposes HarmTSFormer, a two-stage SSVEP decoding architecture. In the first stage, the Dual-Branch Harmonic Encoder (DBHE) constructs harmonic evidence from candidate frequencies and phases. Encoding parameters shared across candidates within each branch enable unified representations of seen and unseen stimuli, while low- and high-order Transformer branches model dependencies among channels, frequency bands, and harmonic responses. Aggregated candidate-set context further informs individual representations. In the second stage, the Evidence-Preserving Aggregation Module (EPAA) combines score normalization with learnable weighting to fuse learned scores and explicit reference-matching scores, integrating EEG response features with reference-matching information. HarmTSFormer uses only source participants' seen-frequency EEG for staged training, without target-user calibration. Evaluations on three public datasets demonstrate its effectiveness under cross-subject, unseen-frequency conditions. Ablation experiments, structural comparisons, and score decomposition analyses further examine the contributions of different information pathways to decoding performance.

Primary area:
applications to neuroscience & cognitive science
Keywords:
SSVEP, brain-computer interfaces, EEG decoding, Transformers, cross-subject generalization, unseen-frequency generalization

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_2SB9OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This study proposed HarmTSFormer for cross-subject SSVEP decoding at unseen stimulus frequencies. The model propose a Dual-Branch Harmonic Encoder (DBHE) to encode candidate harmonic evidence with shared parameters and an Evidence-Preserving Aggregation Module (EPAA) to fuse learned and reference-matching scores, and was evaluated on three public datasets without target-user calibration. Although extensive evaluations and ablations have been performed, the study lacks convincing evidence that the modest gains over simpler baselines justify the added architectural and training complexity.

Critical strengths

The authors combine explicit reference matching with learned representations for unseen-frequency decoding and evaluate the method across three datasets through a number of comparative experiments, ablations, and sensitivity analyses.

Critical weaknesses

The paper is difficult to follow, particularly regarding the motivation, technical details, scientific contribution, and justification for the necessity of its components.

Notably, the method relies heavily on handcrafted features and hyperparameters but lacks sufficient EEG signal-processing rationale and theoretical justification for their design and combination.

Questions for authors

Although the paper evaluates alternative frequency splits, broader testing is needed, including repeated random frequency splits with results averaged across repetitions.

More informative visualizations and interpretation analyses are needed to clarify what the model learns and substantiate the role of each component.

Compared with the candidate-conditioned Transformer and CNN baselines, the proposed architecture provides only modest accuracy improvements, whose statistical significance remains unclear, leaving insufficient evidence that these gains justify the added complexity.

The practical relevance remains unclear. Given the focus on cross-subject and cross-frequency generalization, evaluating transfer between datasets with different stimulus frequencies is needed to address the concern that the method may be overfitting to the dataset due to complexity.

The paper should provide more complete preprocessing details

5 comments
AuthorsOpenReview ↗

Response to Reviewer 2SB9 (Part 1/2)

[Q1] The motivation, technical details, scientific contribution, and necessity of the components are difficult to follow. What EEG signal-processing rationale supports the features and model design?

[A1]

  • Design and contribution. A candidate’s frequency and phase specify its harmonic reference. Separate encoders process H1–H3 and H4/H5. Each branch uses the mean representation of the candidates to update each one. EPAM combines the learned score with reference-matching strength and harmonic log power (Sections 2.2–2.3).

  • EEG features and parameter choices. Real and imaginary spectral responses retain phase relative to the reference, and log power summarizes amplitude. Single-harmonic matching measures both matching strength and spatial structure across channels. Subband features combine local responses with joint multiharmonic matching. Appendix A.2 explains the spectral offsets and subbands, Table 5 distinguishes fixed cue scales from learned weights and cutoffs, and Appendix D.3 describes source-only selection. With other cues fixed, increasing one reliability cue can only maintain or raise the H4/H5 trial gate (Equation 3). Removing only the F2 encoder tokens leaves the explicit F2 score and both residual paths in place. Accuracy at 0.6 s fell by 1.56 points on Benchmark, 0.92 on BETA, and 0.63 on Wearable (Table 22).

[Q2] Can the evaluation include repeated random frequency splits, with results averaged across repetitions?

[A2]

Yes. We ran five repetitions of randomized five-fold frequency holdouts across all three datasets and observation windows. Averaged over the five repetitions, HarmTSFormer outperformed independent fusion in all nine conditions and the designated comparator in eight. In each repetition we held out each of the 38 interior Benchmark/BETA frequencies or all 12 Wearable frequencies once. The Benchmark/BETA endpoints stayed in the fitting set. All methods in Table F.1 used the same splits and scored the full candidate set.

We fixed one prespecified training seed across repetitions. Each outer fold refit the trainable stages, scoring heads, source templates, and fitted statistics on the permitted source data. Within each repetition, we pooled the five holdouts per participant and averaged participants equally. Table F.1 reports means and sample SDs across repetitions, including the differences computed within each repetition.

DatasetWindowHarmTSFormer accuracy (%)Gain over independent fusion (pp)Gain over named comparator (pp)
Benchmark0.4 s27.11 ± 0.98+0.64 ± 0.59+0.50 ± 1.10
Benchmark0.6 s48.57 ± 1.35+1.08 ± 0.54+2.06 ± 0.38
Benchmark0.8 s67.24 ± 0.53+1.39 ± 0.89+2.13 ± 0.54
BETA0.4 s14.59 ± 0.55+6.13 ± 0.62+1.70 ± 0.27
BETA0.6 s20.89 ± 0.95+2.52 ± 0.61+1.07 ± 0.82
BETA0.8 s30.42 ± 1.78+1.45 ± 1.20+1.83 ± 0.76
Wearable0.4 s33.89 ± 0.60+2.11 ± 0.71+1.65 ± 0.60
Wearable0.6 s47.44 ± 0.38+0.30 ± 0.38−0.21 ± 0.38
Wearable0.8 s58.85 ± 0.91+0.66 ± 0.77+0.20 ± 0.66

The designated comparators are SRRNet for Benchmark 0.4 s and BETA 0.4/0.6 s, tlCCA-1 for Benchmark 0.6/0.8 s and BETA 0.8 s, FBCCA-H3 for Wearable 0.4/0.6 s, and LINEAR for Wearable 0.8 s.

The SDs quantify sensitivity to frequency partitioning under the fixed training seed. Tables F.2–F.4 give the individual repetitions, including the reversals on Wearable (Appendix F.6).

[Q3] Can you provide more informative visualizations and interpretation analyses to clarify what the model learns and the roles of its components?

[A3]

Figure 3(a) compares the low-order and combined DBHE representation readouts. The combined readout has higher accuracy in all nine conditions. Figure 3(b) scans perturbation frequency for targets at 9, 11, and 13 Hz. Its vertical axis gives the original minus perturbed correct-target score margin. Low-order margin drops concentrate near H1–H3, and the high-order readout also responds near H4/H5 (Section 3.5 and Appendix K).

Table O.1 reports a separate test of accuracy. We subtract one target harmonic at a time by fitting a sine–cosine pair to each EEG channel and subtracting the fitted component. We then recompute the scores with the model weights fixed. Removing H3 on Wearable at 0.6 s lowered full-model accuracy by 13.19 percentage points.

Figure 4 on page 10 gives accuracy for each readout. Table O.2 gives the proportions of predictions corrected or made incorrect by fusion. On Benchmark at 0.6 s, corrections account for 8.80% of evaluation weight and newly introduced errors for 7.71% relative to the learned readout, giving a net gain of 1.09 points.

AuthorsOpenReview ↗

Response to Reviewer 2SB9 (Part 2/2)

[Q4] Are the gains over the candidate-conditioned Transformer and CNN statistically significant, and how do they compare with the additional computational cost?

[A4]

Yes. HarmTSFormer gains 0.44–0.86 percentage points in mean accuracy over CCT and 0.39–1.96 over HCM. We used two-sided paired t-tests on participant accuracies and applied Holm correction across all 18 comparisons. All 18 remained significant (adjusted p ≤ 0.0063). Table 17(a) gives means and SDs across training runs. Table 17(b) gives paired differences and pointwise 95% confidence intervals. The following two comparisons are for Benchmark at 0.6 s.

ComparatorAccuracy gain95% confidence intervalHolm-adjusted p
CCT+0.86 pp[0.40, 1.32] pp0.0063
HCM+1.96 pp[0.92, 3.00] pp0.0063

Section 3.3 compares HarmTSFormer with CCT and HCM on accuracy and inference time (see Appendix M.3, Table 23, for the full comparison). We timed single-trial inference on the same host with matched inputs, 40 candidates and batch size one. The table gives median latency including preprocessing and scoring. Training time is reported per final source fit.

MethodBenchmark 0.6 s accuracyTraining time per final fitInference time per trial
HCM47.80%10.8 min23.6 ms
CCT48.90%19.2 min44.1 ms
HarmTSFormer49.76%38.4 min67.4 ms

On this workload, HarmTSFormer takes 1.53 times the inference time of CCT and 2.86 times that of HCM. The corresponding training-time ratios are 2.00 and 3.56 per final source fit (Appendices H.3 and M.3).

[Q5] Can you evaluate transfer between datasets with different stimulus frequencies to assess practical relevance and generalization?

[A5]

Yes. We fit on Benchmark participants S1–S16 and evaluated the frozen decoder on Wearable wet-electrode participants S17–S54 using eight common channels. Wearable’s 12 target frequencies are distinct from Benchmark’s 40 frequencies and lie within the source range. We used only source data for fitting and selection, then scored all 12 target candidates (Section 3.3, Table L.2, and Appendices L.4–L.5).

Method0.4 s accuracy (%)0.6 s accuracy (%)0.8 s accuracy (%)
FBCCA-H332.2447.6558.11
FBCCA-H515.4438.9056.07
FBCCA/tlCCA + MLP31.78 ± 0.3346.54 ± 0.4257.79 ± 0.33
HarmTSFormer33.64 ± 0.3847.93 ± 0.6559.08 ± 0.40
  • The FBCCA-H3 and FBCCA-H5 entries reproduce their mean accuracies on Wearable from Table 2. For the two fitted models, we report mean accuracy and SD across three runs.

  • HarmTSFormer has the highest mean accuracy at all three windows. After Holm correction across nine comparisons, the gains were significant at 0.4 and 0.6 s against independent fusion and at all three windows against FBCCA-H5. Against FBCCA-H3, mean accuracy was higher by 1.40, 0.28, and 0.96 percentage points at 0.4, 0.6, and 0.8 s, respectively. None of these three differences reached significance after Holm correction (Table E.1). The experiment demonstrates offline reuse of the frozen source model across participants, recording conditions, and stimulus grids without target-user calibration.

[Q6] Please provide more complete preprocessing details.

[A6]

Section 3.2 and Appendix D.4 specify the shared preprocessing. We select channels from the released 250-Hz arrays and crop each trial. Benchmark and BETA use Pz, PO5, PO3, POz, PO4, PO6, O1, Oz, and O2. Wearable and Benchmark-to-Wearable transfer use POz, PO3, PO4, PO5, PO6, Oz, O1, and O2, in that order. Table D.1 gives the starts relative to the beginning of the stored array.

DatasetStart within the stored trial arrayZero-based start index
Benchmark0.64 s160
BETA0.50 s125
Wearable wet0.64 s160

The 0.4-, 0.6-, and 0.8-s windows contain 100, 150, and 200 samples. Each channel receives a 6–90 Hz Butterworth bandpass with design-order parameter 4 using second-order sections and forward–backward filtering. Endpoint handling uses odd extension of the cropped window. We then subtract each channel’s window mean and divide by its population SD plus 10−610^{-6}. The first cropped sample defines the reference-time origin. Appendix A.2 describes the subsequent F2 subband processing.

Dear Reviewer,

We hope that our responses have thoroughly addressed your questions. As the discussion phase comes to a close, we eagerly await your feedback and would appreciate knowing if you have any remaining concerns we can address. We would be grateful if you find our revisions satisfactory and consider raising your evaluation of our paper.

Thank you once again for the time and effort you have dedicated to reviewing our manuscript.

Best regards,

The HarmTSFormer Authors

Reviewer_Hh5bOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SSVEP decoders typically restrict outputs to predefined frequency classes, making unseen-frequency recognition difficult. Prior work supports this setting, but within-candidate response structure and cross-candidate information remain challenging for reliable cross-subject decoding. This paper proposes HarmTSFormer, a two-stage architecture: DBHE builds harmonic evidence from candidate frequencies/phases with shared parameters and low-/high-order Transformers, using candidate-set context to update individual representations; EPAA then fuses learned scores with explicit reference-matching scores via normalization and learnable weights. Training uses only source participants’ seen-frequency EEG, without target-user calibration. Results on three public datasets under cross-subject, unseen-frequency conditions, plus ablations and score analyses, support the approach.

Critical strengths

Three datasets, nine window conditions, shared source-only splits, and systematic ablations / score-pathway analyses make the empirical case more trustworthy than a single-dataset SSVEP demo. This thoroughness is consequential: even when absolute margins are small, the paper shows repeated, same-direction gains under a well-specified protocol.

Critical weaknesses

HarmTSFormer combines dual Transformers, candidate-set gating, reliability modulation, teacher–student constraints, three training phases, residuals, and multi-view averaging, yet improvements over the strongest listed baseline are typically only about 1–2.5 percentage points, and the edge over a simple LINEAR scorer on Wearable is especially thin. This raises the question whether the paper establishes a reusable modeling insight, or mainly a heavily tuned SSVEP pipeline whose incremental accuracy may not justify the complexity. Component removals also affect multiple dependent paths at once, so gains are hard to attribute cleanly to any single design choice.

Questions for authors

Under the same source-only protocol, can you report a simpler strong baseline that already fuses FBCCA/tlCCA-style explicit scores with a lightweight learned head (no dual-branch Transformer / teacher–student stack)?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Hh5b

[Q1] Under the same source-only protocol, can you report a simpler strong baseline that fuses FBCCA/tlCCA-style explicit scores with a lightweight learned head, without the dual-branch Transformer or teacher–student stack?

[A1]

Yes. We added an independent decoder that combines three scores for each candidate, from FBCCA-H3, FBCCA-H5, and source-fitted tlCCA-1. We standardize each score across the current candidates. A shared 3→16→1 GELU head with 81 trainable parameters produces a logit for each candidate.

We train the head with cross-entropy under fixed AdamW settings and generate template-derived training scores by cross-fitting within the source pool. We use held-out source participants to select the hidden width (16 or 32) and training length, excluding their EEG from template estimation and head fitting during selection. With the selected width of 16 and training length fixed, we refit on all 16 source participants and retain the terminal checkpoint. Both decoders use the same splits, channels, windows, and full candidate lists (Appendix N).

DatasetWindowFBCCA/tlCCA + MLPHarmTSFormerDifference
Benchmark0.4 s27.30%28.10%+0.80 pp
Benchmark0.6 s48.20%49.76%+1.56 pp
Benchmark0.8 s66.30%68.23%+1.93 pp
BETA0.4 s8.50%14.43%+5.93 pp
BETA0.6 s18.50%21.24%+2.74 pp
BETA0.8 s30.10%31.72%+1.62 pp
Wearable0.4 s32.60%34.71%+2.11 pp
Wearable0.6 s47.50%48.32%+0.82 pp
Wearable0.8 s58.50%59.69%+1.19 pp

Independent fusion ranks first among the Table 2 comparators in four conditions. HarmTSFormer exceeds it by 0.80–5.93 percentage points across all nine.

[Q2] Do the improvements justify the additional architectural and training complexity, particularly given the small gap over LINEAR on Wearable?

[A2]

On Wearable at 0.8 s, HarmTSFormer reached 59.69% accuracy versus 58.97% for LINEAR. The mean gain was 0.72 percentage points, with higher accuracy for 23 of 38 participants (Holm-adjusted p = 0.0440, Table 11). At 0.4 and 0.6 s, the mean gains over LINEAR were 4.78 and 2.22 points, respectively (Table 2).

We use the same scoring rule for each candidate, so its harmonic reference lets us evaluate a frequency held out during training. We selected the model configuration using 12 source participants for fitting and four others for validation (Appendix D.3). To isolate the choice of encoder, we kept the candidate features, harmonic grouping, candidate context and scoring paths fixed. The Transformer exceeded the MLP encoder by 0.65–2.84 percentage points in mean accuracy across the nine conditions (Table 21).

Our matched timing comparison used Benchmark at 0.6 s with 40 candidates. Section 3.4 compares the two encoders, and Table 23 in Appendix M.3 also includes independent score fusion. The three rows below compare the full model with the two MLP baselines. Training time covers one final source fit. Median inference time includes preprocessing and scoring for one trial.

MethodAccuracyTraining time per final fitInference time per trial
FBCCA/tlCCA + MLP48.20%3.4 min24.6 ms
MLP encoder replacement46.92%24.6 min48.2 ms
HarmTSFormer49.76%38.4 min67.4 ms

On Benchmark at 0.6 s, the Transformer gained 2.84 percentage points over the MLP encoder and took an additional 19.2 ms per trial. Against independent score fusion, HarmTSFormer gained 1.56 points on average and achieved higher accuracy for 10 of the 12 participants (Table 11).

[Q3] Some component removals affect several dependent pathways. How can the gains be attributed more clearly to the design choices?

[A3]

  • Table 22 separates the F2 encoder input from the other paths that use F2. We remove the F2 tokens and retain the explicit F2 score and both residual paths. At 0.6 s, accuracy drops by 1.56, 0.92, and 0.63 percentage points on Benchmark, BETA, and Wearable, respectively. Table 3 gives the corresponding comparison for removal of the complete F2 pathway.

  • Table 21 reports two training comparisons. Removing the compatibility and preservation constraints together measures their combined contribution. We tested the additional source-group update separately. Each change lowered mean accuracy in eight of nine conditions. On Wearable at 0.8 s, removing the two constraints raised accuracy by 0.04 points and disabling the additional update raised it by 0.03 points.

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

The paper includes a new methodology to classify EEG in SSVEP-based brain-computer interfaces. A key claim is the ability to classify SSVEP frequencies not present in training. They contribute a new pipeline, which fuses concepts from classical SSVEP analyses (e.g., filter banks, CCA), with modern deep learning. Validation includes 3 datasets and cross-patient testing (which is important). Baseline selection seems reasonable, and the supplementary is extensive.

Critical strengths
  1. The paper trains and validates in different subjects. This is important, although already established in the field. Early SSVEP works trained and tested on EEG from the same patient, which is impractical in real-world applications.

  2. The paper uses 3 datasets, which is important to prove generalization.

  3. The supplementary is comprehensive.

  4. The paper includes a large selection of baselines, although many are old/classical baselines. The classical baselines are still relevant, given that the paper combines classical methods (e.g., filter banks and CCA) with deep learning.

  5. The paper includes ablations, which are important given the highly complex pipeline it proposes.

Critical weaknesses

1- The main weakness is a lack of clarity. The writing is very complex. From my experience, I would say this may be a sign of strongly LLM assisted writing: instead of explaining concepts in plain English and using only standard ML terms, the text creates a lot of new terms without a clear definitions of these terms. This makes understanding and reviewing this paper challenging.

2- Unclear novelty. The method seems a very complex, highly engineered pipeline. So, the main novelty becomes unclear. Novelty comparisons to previous works are not direct or concrete. This paper would benefit from a clear novelty list, with clear descriptions of how each main novelty differs from previous work.

3- Unclearly defined methods. Important methods are not clearly defined in the main paper. For example, a loss function is only defined as: ‘Ladj penalizes neighboring-frequency margin decreases against a stop-gradient original-view reference’. Is this a new loss function? If so, it should be clearly explained in the main paper (not in the appendix). If this is not novel, include a citation.

4- Inter-patient validation is a strength. Yet, it would be interesting to analyze inter-dataset validation (i.e., train in one dataset, test in another), since the method supports testing with frequencies and patients not seen in training.

5- The model surpasses the strongest baseline by 0.6 to 2.5 pp. Is this significant? It would be important to add significance tests in Table 2.

Questions for authors

Please see Critical Weaknesses above.

3 comments
AuthorsOpenReview ↗

Response to Reviewer DrFb (Part 1/2)

[Q1] The paper is difficult to follow because of complex terminology and insufficiently defined concepts.

[A1]

Sections 2.1–2.3 now show how HarmTSFormer scores each candidate frequency from an EEG window. For each candidate frequency, the model extracts harmonic features from the same EEG window. Within each branch, it averages the candidate representations and applies layer normalization. A learned gate adds this shared context to each candidate before scoring. The Evidence-Preserving Aggregation Module (EPAM) combines the learned score with single-harmonic reference-matching strength and harmonic log power across frequency bands. The model adds the spatial and matching-summary residuals, standardizes the resulting scores, and averages them over four reference-phase views.

Figure 2 uses EPAM throughout and labels its two direct inputs as F1 matching strength and F2 log power. Sections 2.2–2.3 give the corresponding scoring equations. Section 2.4 and Table 6 identify the parameters updated in each of the three training stages.

[Q2] The novelty is unclear. Please state the main contributions and explain concretely how they differ from previous work.

[A2]

We encode candidate responses in separate low- and high-order harmonic branches and combine the learned scores with direct measures of matching strength and harmonic power.

  • Harmonic encoding. GZSL-Lite uses separate branches for class-template and sine-template information. Our branches specialize by harmonic order, retaining spectral and spatial matching features for each candidate. Within each branch, a pooled representation of all candidates gives the scorer context for the current trial (Section 2.2).

  • Learned and reference scoring. GZSL scores latent correlations, and SRRNet matches predicted response templates to EEG. HarmTSFormer combines a learned readout with direct reference scores that bypass the Transformer. Standardization across candidates and source-trained weights bring the two paths together (Section 2.3). The MLP encoder replacement in Table 21 tests encoding, and the independent score-fusion decoder in Table 2 tests how well explicit scores perform with a lightweight head.

[Q3] Important methods are not sufficiently defined in the main paper. Is LadjL_{\mathrm{adj}} a new loss, and how is it computed?

[A3]

LadjL_{\mathrm{adj}} is the task-specific regularizer we introduce to preserve the scoring margin between a target and nearby frequencies across reference-phase views. Section 2.4 now defines it in Equation (13).

For trial mm and view vv, let zmvz_{mv} denote the standardized candidate scores. We exclude the target, find the other candidates nearest to it in physical frequency, and retain all ties. The target margin uses the highest score among those neighbors.

dmvadj=zmv,ym−max⁡i∈N(ym)zmv,i.d_{mv}^{\mathrm{adj}} = z_{mv,y_m} - \max_{i\in\mathcal N(y_m)}z_{mv,i}.

The zero-offset view supplies the reference margin. We stop its gradient and penalize decreases in the other three views using the squared positive part.

Ladj=13M∑m=1M∑v=13[sg(dm0adj)−dmvadj]+2.L_{\mathrm{adj}} = \frac{1}{3M} \sum_{m=1}^{M}\sum_{v=1}^{3} \left[ \mathrm{sg}(d_{m0}^{\mathrm{adj}}) - d_{mv}^{\mathrm{adj}} \right]_{+}^{2}.

Only the reference phases change across views. The EEG window is fixed. Equation (11) weights this term by 0.1, and Appendix B.1 specifies how we select neighbors. Section 2.4 also defines phase consistency in Equation (12), the teacher and student objectives in Equation (14), and the two student constraints in Equations (15)–(16).

AuthorsOpenReview ↗

Response to Reviewer DrFb (Part 2/2)

[Q4] Since the method supports unseen participants and frequencies, can it be trained on one dataset and tested on another?

[A4]

Yes. Section 3.3 and Appendix L report two transfer settings with frozen source models.

  • Shared frequency grid. We evaluate both Benchmark-to-BETA and BETA-to-Benchmark transfer using source participants S1–S16 and target participants S17–S28. In each of five frequency-holdout folds, source fitting excludes the target frequencies evaluated in that fold. Each prediction scores all 40 candidates. HarmTSFormer has the highest mean accuracy among the listed methods in all six conditions (Table L.1). Differences are in percentage points (pp).
Transfer directionWindowHarmTSFormer accuracyGain over the strongest listed baseline
Benchmark → BETA0.4 s12.11%+1.17 pp
Benchmark → BETA0.6 s18.94%+1.55 pp
Benchmark → BETA0.8 s27.66%+2.07 pp
BETA → Benchmark0.4 s22.81%+0.57 pp
BETA → Benchmark0.6 s41.59%+1.38 pp
BETA → Benchmark0.8 s57.55%+1.85 pp
  • Different frequency grid. We also fit on Benchmark and score all 12 Wearable wet-electrode targets using eight common channels. The target frequencies are distinct from the 40 source frequencies and lie within their range. Mean accuracy reaches 33.64%, 47.93%, and 59.08% at 0.4, 0.6, and 0.8 s. The gains over independent fusion reach the Holm-adjusted threshold at 0.4 and 0.6 s (Table L.2 and Table E.1).

[Q5] Are the gains over the strongest baseline statistically significant? Please add significance tests to Table 2.

[A5]

HarmTSFormer has higher mean accuracy than the strongest listed baseline in all nine Table 2 conditions. Six of the nine participant-paired gains are significant after Holm correction and carry a † in Table 2.

We pool disjoint frequency-holdout predictions within each participant and training run, then average repeated fits within participant. Two-sided sign-flip tests compare the paired accuracies. Holm correction includes all eleven baselines within each dataset and window. Table 11 gives the paired summary for the strongest baseline in each condition, including its confidence interval and the number of participants who improve.

DatasetWindowStrongest listed comparatorFull − comparatorHolm-adjusted p
Benchmark0.4 sSRRNet+0.63 pp0.4609
Benchmark0.6 sFBCCA/tlCCA + MLP+1.56 pp0.0215
Benchmark0.8 sFBCCA/tlCCA + MLP+1.93 pp0.0283
BETA0.4 sSRRNet+1.82 pp0.014
BETA0.6 sSRRNet+1.45 pp0.0312
BETA0.8 sFBCCA/tlCCA + MLP+1.62 pp0.0518
Wearable0.4 sFBCCA/tlCCA + MLP+2.11 pp0.0142
Wearable0.6 sFBCCA-H3+0.67 pp0.4951
Wearable0.8 sLINEAR+0.72 pp0.0440

Dear Reviewer,

We hope our response has addressed your questions. As the discussion phase is coming to a close, we are looking forward to your feedback and would like to know if you have any remaining concerns we can address. We are grateful if you find our revisions satisfactory and consider raising your score for our paper.

Thank you once again for the time and effort you have dedicated to reviewing our paper.

Best regards!

HarmTSFormer Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers

We thank all reviewers for their constructive feedback. We are encouraged that the reviewers recognized the value of evaluating HarmTSFormer across three public SSVEP datasets (DrFb, Hh5b, 2SB9), noted the importance of training and testing on different participants (DrFb), and highlighted our shared source-only protocol, systematic ablations, and score-pathway analyses as strengths of the study (Hh5b).

Below, we summarize our responses to the shared concerns and list the main revisions to the manuscript and supplementary material.

Shared concerns

  • Method clarity and contribution (DrFb, 2SB9). Section 1 now compares our low- and high-order harmonic branches with the class-template and sine-template branches of GZSL-Lite, and explains how our fusion of learned and direct reference scores differs from the scoring rules of GZSL and SRRNet. Sections 2.1–2.4 trace the decoding process from an EEG window to a prediction and define the training objectives, including the adjacent-frequency loss in Equation (13).

  • Simpler baselines and component contributions (Hh5b, 2SB9). Table 2 adds an independent FBCCA/tlCCA + MLP baseline, which HarmTSFormer exceeds by 0.80–5.93 percentage points in mean accuracy across the nine conditions. Table 21 compares the Transformer and MLP encoders with the same harmonic grouping, candidate context, and scoring paths, yielding gains of 0.65–2.84 points for the Transformer. Table 22 separately evaluates the F2 encoder inputs while retaining the explicit and residual scoring paths.

  • Statistical evidence and computational cost (DrFb, Hh5b, 2SB9). HarmTSFormer achieves the highest mean accuracy in all nine Table 2 conditions. Six gains over the strongest listed baseline are significant after Holm correction across eleven baselines within each condition. All 18 participant-paired accuracy comparisons with CCT and HCM are significant under a separate Holm correction (adjusted p ≤ 0.0063, Table 17(b)). Table 17(a) gives the standard deviations across training runs. On Benchmark at 0.6 s, HarmTSFormer took 38.4 minutes per final source fit and a median of 67.4 ms per trial including preprocessing (Appendix M.3, Table 23).

  • Frequency splits and cross-dataset transfer (DrFb, 2SB9). Appendix F.6 reports five repetitions of randomized five-fold frequency holdouts. Averaged over the five repetitions, HarmTSFormer exceeds independent fusion in all nine conditions and the designated comparator in eight. Section 3.3 and Tables L.1 and L.2 evaluate frozen source models in both directions between Benchmark and BETA on a shared frequency grid, and from Benchmark to Wearable on 12 new frequencies within the source range.

Main revisions in the manuscript and supplementary material

  • Section 1. Direct comparisons with GZSL-Lite, GZSL, and SRRNet on harmonic encoding and candidate scoring (DrFb, 2SB9).
  • Sections 2.1–2.4, Figure 2, and Appendices A, B, and K. Decoding steps, feature and score definitions, and training objectives (DrFb, 2SB9).
  • Section 3.2, Appendix D.4, and Table D.1. Channel order, window start indices, filtering, and temporal standardization (2SB9).
  • Sections 3.2–3.3, Table 2, and Appendix N. Independent FBCCA/tlCCA + MLP baseline and its source-only fitting protocol (Hh5b).
  • Tables 11, 17(a), 17(b), and 23. Tables 11 and 17(b) report participant-paired comparisons. Table 17(a) gives mean accuracy and training-run SD. Sections 3.3 and 3.4 compare accuracy and inference time. Table 23 in Appendix M.3 gives the full training and inference costs (DrFb, Hh5b, 2SB9).
  • Appendix F.6 and Tables F.1–F.4. Results averaged across randomized frequency splits and results for each repetition (2SB9).
  • Section 3.3, Tables L.1 and L.2, Appendix L, and Table E.1. Transfer protocols and results for shared and disjoint stimulus-frequency grids (DrFb, 2SB9).
  • Sections 3.4–3.5, Figures 3–4, and Appendices G, M, and O. Encoder and F2-token controls, harmonic sensitivity, and changes in individual predictions after score fusion (Hh5b, 2SB9).

Our detailed responses follow under the corresponding reviews.

Area_Chair_FbWcOpenReview ↗

Please Engage in the Author–Reviewer Discussion

Dear Reviewers,

As the author–reviewer discussion period is approaching its deadline (November 18, AoE), I would like to kindly remind you to review the authors' rebuttals and actively engage in the discussion as soon as possible.

Please address any outstanding questions or concerns and update your assessments and ratings if appropriate.

Thank you very much for your time and efforts!

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.