acceptodds
Withdrawn from ICLR 2027

TAHLP: A Topology-Aware Hierarchical Latent Prediction Foundation Model for Cross-Electrode Motor Imagery

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Motor imagery (MI) is a core electroencephalography (EEG) brain-computer interface paradigm, yet developing a single MI foundation model that transfers across heterogeneous electrode montages remains challenging. Generic EEG foundation models pre-trained on multi-paradigm mixtures may learn representations that do not align well with MI's frequency- and region-specific sensorimotor rhythms. Existing electrode-position encodings often rely on hand-defined channel orders or learn geometry-to-feature mappings that may not fully capture the electrode-coordinate topology. To overcome these challenges, we propose TAHLP, a Topology-Aware Hierarchical Latent Prediction foundation model for cross-electrode MI, pre-trained via hierarchical latent-prediction self-distillation, where a student predicts the multi-level cumulative residuals of an exponential moving average (EMA) teacher's continuous latent features using a weighted cosine loss. This objective is self-supervised and optimized in a single stage, requiring no codebook, negative pairs, or labels, which simplifies optimization and reduces reliance on large labeled datasets for pre-training. To explicitly model electrode-coordinate topology, we introduce interleaved 3D electrode rotary position embedding, a parameter-free rotary transform on adjacent query/key feature-dimension pairs, with angles from each electrode’s 3D scalp coordinates. By construction, attention scores depend on relative 3D electrode displacement, not absolute position or channel order/count. Extensive experiments across four public MI benchmarks demonstrate that TAHLP consistently achieves state-of-the-art performance while using less than 8.8% as much pre-training data as generic EEG foundation models do. Code and models will be publicly available.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Motor Imagery Foundation Model, Topology-aware Transformer, Hierarchical Latent Prediction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GhrSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors propose TAHLP, a self-supervised pretrained Transformer for cross-montage motor-imagery EEG. Two components are claimed as contributions. (1) Hierarchical latent-prediction self-distillation: an EMA teacher encodes the unmasked trial and provides, per token, the instance-normalized average of its last K block outputs as a continuous target; the student sees the trial under two complementary token masks and predicts the target at masked positions through n=3 cumulative residual levels with a weighted cosine loss. The objective is single-stage and uses no codebook, negative pairs, or labels. (2) I3D-ERoPE: a parameter-free rotary transform on query/key pairs whose angles come from each electrode’s 3D scalp coordinate, with frequency planes assigned round-robin to the x/y/z axes, so attention depends on relative electrode displacement rather than channel index. Heterogeneous montages are handled by a shared per-patch temporal conv stem, a learned electrode-name embedding, and dynamic padding with validity masks.

Validation: pretraining on 22 public MI datasets (559 subjects, 221 h, 6.2M-parameter encoder), then subject-independent fine-tuning on four held-out MI benchmarks (PhysioNet-MI, SHU-MI, BCICIV-2a, BNCI2015001) against seven specialist decoders and six generalist EEG foundation models. Supporting experiments: a same-corpus control re-pretraining all six generalists on the 221 h corpus (Appendix G), a component ablation (Table 2), a level-weight ablation, a pretraining-budget curve on BCICIV-2a, and an attention-geometry probe (Appendix C). The paper claims state-of-the-art on all four benchmarks with under 8.8% of the smallest baseline’s pretraining data.

Critical strengths
  1. I agree that the pretraining objective is the right kind of contribution for this problem, and the authors test it properly. Predicting continuous full-context latents instead of waveforms, codebook indices, or contrastive pairs addresses a real mismatch between existing EEG pretraining targets and the narrowband, nuisance-dominated structure of MI. The ablation attributes the largest single gain to this change (3.0 to 4.3 balanced-accuracy points over masked reconstruction), and the same-corpus control in Appendix G is the a very good experiment for separating objective from data budget.

  2. I very much like the proposed I3D-ERoPE, as it is a parameter-free way to make attention a function of relative 3D electrode displacement, which is the property cross-montage transfer actually needs. The ablation reports a strictly larger gain than a learned coordinate embedding across all benchmarks, and the pretraining-budget curve shows that the advantage is largest in the low-data regime. The mechanism is a direct adaptation of interleaved multi-axis RoPE, so the novelty lies in the application rather than the idea, but it is well executed and, by construction, generalizes to unseen electrodes (which is a crucial thing to have).

Critical weaknesses
  1. The component ablation (Table 2) is very strange and not credible as reported. The PhysioNet-MI and SHU-MI columns differ by exactly 1.27 points in all six rows (59.63/58.36, 62.63/61.36, 63.13/61.86, 64.33/63.06, 65.53/64.26, 66.30/65.03). Consequently every per-component delta is identical on the two benchmarks (3.00/3.00, 0.50/0.50, 1.70/1.70, 1.20/1.20, 0.77/0.77), and Section 3.5 reports these as independent findings. Five-seed means on two unrelated datasets with different class counts, subject pools, and montages do not produce a constant offset, or at least the chances of that are miniscule. Since the ablation is the primary evidence for both claimed contributions, this is consequential to the whole assessment. Per-seed raw results are required before the ablation can be evaluated.

  2. The state-of-the-art claim does not survive a baseline the authors themselves cite, but don't compare to. REVE (El Ouahidi et al., NeurIPS 2025) reports results on PhysioNet-MI and BCICIV-2a under the same CBraMod-derived subject splits used here (the CBraMod rows match in both papers), and its BCICIV-2a balanced accuracy is roughly 7 points above TAHLP’s 57.15. REVE is cited, its positional encoding is used as an ablation arm, yet it is absent from Table 1. This falsifies “state of the art on all four benchmarks” and weakens the paradigm-specialization thesis, since a generalist beats the MI specialist on an MI benchmark. The remaining margins are also small relative to the evaluation noise: BCICIV-2a and BNCI2015001 test on 2 and 3 subjects respectively, the 0.58-point lead over CSBrain on BCICIV-2a is within one reported std, and no significance test is given.

Questions for authors
  1. Please provide per-seed results for every row of Table 2 and explain the constant 1.27-point offset between the PhysioNet-MI and SHU-MI columns. I find it hard to believe that such runs would produce such a perfect offset.

  2. Why is REVE omitted from Table 1 given that it is cited, used in the ablation, and reports numbers under the same splits? Please add it, or state what in the protocol makes its published numbers incomparable. You could also compare to topology-agnostic baselines such as LUNA (also cited in the appendix).

  3. The paper never reports the 6.2M-parameter architecture trained from scratch on each downstream set (no pretraining). Table 2 starts from a reconstruction-pretrained baseline. What does pretraining contribute over no pretraining at all? This is a super important baseline to compare to and cleanly separates the effect of the pretraining from the model construction itself.

  4. The anti-collapse mechanism (EMA teacher plus BYOL-style predictor) is never ablated. Did you try a single-encoder variant with stop-gradient only, or a regularized JEPA formulation without a teacher? If latent prediction is the contribution, the reader needs to know which part of the machinery is load-bearing; also, thoughts from the authors on using a method that maybe doesn't require stop-gradients and teacher-student networks (such as LeJEPA) would be interesting to hear.

  5. Common average reference behaves very differently on a 13-channel montage than on 64 channels, so the reference is not actually common across the corpus. Was this considered, and how are datasets with bipolar or linked-ear references handled?

  6. The 50/60 Hz notch is applied together with a 0.5 to 40 Hz bandpass. What does the notch contribute in this setting, I would assume that doing bandpass alone would be sufficient?

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

This paper introduces TAHLP (Topology-Aware Hierarchical Latent Prediction), an EEG foundation model for cross-electrode motor imagery (MI) decoding across heterogeneous montages. Adapting MRoPE-Interleave (Huang et al., 2026) to continuous 3D scalp coordinates, the authors propose Interleaved 3D Electrode Rotary Position Embedding (I3D-ERoPE), so that the coordinate-dependent (rotary) factor of each attention score depends only on the relative electrode displacement (Section 2.2, Eq. 5–6). Grounded in continuous latent targets from data2vec (Baevski et al., 2022), the authors propose hierarchical latent-prediction self-distillation: an EMA teacher averages top K=6K=6 blocks under instance normalization, while a student predicts these targets across n=3n=3 cumulative residual levels under complementary masks (r=0.7r=0.7) using weighted cosine loss (w=[1.0,0.5,0.25]w = [1.0, 0.5, 0.25], Sections 2.1 and 2.3, Eq. 7–12).

Pre-trained on 221 hours of MI-EEG from 22 public datasets (N=559N=559 subjects, Table 4) and evaluated on four subject-independent benchmarks (PhysioNet-MI, SHU-MI, BCICIV-2a, BNCI2015001; Section 3.2, Tables 1 and 7), TAHLP achieves state-of-the-art Balanced Accuracy (+2.10, +0.86, +0.58, and +1.06 points over the strongest baseline on each benchmark) using under 8.8% of the pre-training data of the smallest-corpus generalist baseline (221 h vs. 2,534 h). In Section 3.3(ii) and Appendix G (Tables 7 & 8), re-pre-training all six generalist baselines on the same 221-hour MI corpus widens TAHLP's margin (+2.75, +1.42, +2.56, and +1.52 points), while the authors transparently note two structural caveats in Appendix G (keeping published epoch counts reduces baseline update steps by 11×–260×11\times\text{--}260\times and leaves large VQ codebooks and multi-thousand-hour schedules untuned).

Critical strengths
  1. Consistent Empirical Gains Supported by a Transparent Matched-Corpus Control (Section 3.3(ii), Appendix G Tables 7 & 8): To separate architecture/objective effects from corpus composition, the authors re-pre-train all six generalist baselines on the identical 221-hour MI corpus. In Section 3.2, they state: "To compare the pre-training objective and the electrode encoding under a matched data budget, we" re-run each baseline on the 221-hour MI corpus, and in Section 3.3(ii) they show that "Holding the corpus fixed preserves the ordering and widens the margin." across all four held-out benchmarks (+2.75 on PhysioNet-MI, +1.42 on SHU-MI, +2.56 on BCICIV-2a, and +1.52 on BNCI2015001). Commendably, the authors include a candid self-audit in Appendix G (Two biases that work against the baselines), noting that keeping published epoch counts reduces baseline optimizer update steps by 11×–260×11\times\text{--}260\times and leaves large codebooks (LaBraM 8,192 codes) and multi-thousand-hour schedules (BIOT 58,021h schedule) untuned.
  2. Effective Topology-Aware Modeling and Continuous Latent Distillation (Section 2.2, Section 3.5 Table 2): Interleaved 3D Electrode Rotary Position Embedding (I3D-ERoPE), detailed in Section 2.2 ("ERoPE(Interleaved 3D Electrode Rotary Position Embedding): every electrode is assigned a 3D"), adapts multi-axis rotary embeddings (specifically MRoPE-Interleave from multimodal models) to continuous 3D scalp coordinates, applying the rotation per token so that the coordinate-dependent factor of the query–key score depends only on the relative displacement ui′−uiu_{i'} - u_i (Equation 6). Concurrently, while the teacher target is rooted in data2vec (Baevski et al., 2022), TAHLP's 3-level residual refinement and complementary dual masking are a meaningful extension of latent-prediction self-distillation. In Section 3.5 and Table 2 ("the zero-parameter I3D-ERoPE adds 1.70 / 1.70 / 2.40"), replacing masked waveform reconstruction with latent self-distillation (+3.00 / +3.00 / +4.30) and adding I3D-ERoPE yield gains across PhysioNet-MI, SHU-MI, and BCICIV-2a, with I3D-ERoPE itself adding no learned parameters (pending confirmation that these values come from independent runs; see Question 6).
Critical weaknesses
  1. Prior Art Contextualization and Conceptual Disconnect Between MI Motivation and General Architecture (Abstract, Section 1, Appendix C): In Section 1, the authors scope their novelty claim to "the first MI foundation model built upon latent-prediction pretraining" and cite MIRepNet (Liu et al., 2026b; Knowledge-Based Systems) as prior MI-specific pre-trained work. The contribution should therefore be assessed as continuous latent-prediction pretraining plus I3D-ERoPE, not as MI-specific pre-training per se. Furthermore, in the Abstract and Introduction, the authors argue that generic models may learn representations that "do not align well with MI’s frequency- and region-specific" sensorimotor rhythms (μ/β\mu/\beta ERD/ERS). Yet neither I3D-ERoPE nor hierarchical latent self-distillation contains any MI-specific loss, μ/β\mu/\beta-band prior, or sensorimotor weighting; apart from a small attention-laterality probe in Appendix C (LI ≈+0.052\approx +0.052 vs. −0.049-0.049 for left- vs. right-hand trials), which the authors explicitly do not present as evidence of sensorimotor lateralization, the paper provides no band-limited (μ/β\mu/\beta, 8–30 Hz) or sensorimotor-channel (C3/Cz/C4) specificity analyses.
  2. Step-Budget Reduction (11×–260×11\times\text{--}260\times) and Codebook/Schedule Asymmetries in the Appendix G Control (Appendix G):
  • Matched-Budget Control Asymmetries: As the authors candidly acknowledge in Appendix G ("Two biases that work against the baselines"), re-pre-training the six generalist baselines on the 221-hour MI corpus while keeping published epoch counts reduces baseline optimizer update steps by 11×–260×11\times\text{--}260\times relative to their original 2,534–58,021-hour corpora, under-training two-stage discrete VQ tokenizers with large codebooks (LaBraM with 8,192 codes and BrainRVQ), and leaving learning-rate, warm-up, weight-decay, and masking schedules sized for multi-thousand-hour corpora (such as BIOT's 58,021-hour schedule) untuned for the 221-hour regime. As the authors note ("This asymmetry favours TAHLP, so we do not rest the headline claim of the paper on this control"), step-matched or hyperparameter-tuned controls for single-stage baselines (BIOT, CBraMod) are needed to isolate objective-level gains.

Score rationale

Presentation (3). Clearly written and logically structured, with high-quality architectural diagrams, mathematical exposition, and ablation curves.

Informativeness (2). The 22-dataset MI curation (Table 4) and component ablations (Table 2) are useful if they are confirmed to come from independent runs (Question 6), but the core components (I3D-ERoPE from MRoPE and continuous targets from data2vec) are domain-agnostic adaptations of established techniques, and the matched-budget control in Appendix G (Tables 7 & 8) suffers from an 11×–260×11\times\text{--}260\times update-step reduction against baselines.

Soundness (2). Empirical results are reported as five-seed means across four benchmarks. Soundness is rated 2 due to: (1) conceptual disconnect between MI-specific sensorimotor motivation (ERD/ERS) and domain-agnostic architecture (I3D-ERoPE and latent self-distillation), and (2) structural asymmetries in Appendix G (keeping published epochs reduces baseline update steps by 11×–260×11\times\text{--}260\times, disadvantages two-stage VQ tokenizers like LaBraM's 8,192 codes, and leaves schedules sized for BIOT's 58,021h untuned). My soundness assessment also depends on the answer to Question 6 about the regularity of the Table 2 and Figure 3 values.

Rating (2). While TAHLP achieves strong accuracy across four MI benchmarks and transparently documents a 221-hour matched-corpus control in Appendix G, two substantive issues preclude acceptance in its current form: (1) a disconnect between the MI-specific neurophysiological motivation (μ/β\mu/\beta ERD/ERS) and a domain-agnostic architecture assembled from established components (MRoPE-Interleave and data2vec targets, with MI-specific pre-training already explored in MIRepNet, Liu et al., 2026b, which the authors cite), with no μ/β\mu/\beta-band or sensorimotor-channel verification beyond a small attention-laterality probe (Appendix C); and (2) structural asymmetries in the Appendix G control, where keeping published epoch counts reduces baseline optimizer update steps by 11×–260×11\times\text{--}260\times and leaves large VQ codebooks (LaBraM 8,192 codes) and multi-thousand-hour schedules (BIOT 58,021h) untuned, leaving objective-level gains unisolated from training-budget and domain confounds.

Questions for authors
  1. [clarification] Reconciling MI Motivation with General Architecture & Collapse Metrics: Given that I3D-ERoPE and latent self-distillation are domain-agnostic (Sections 2.2–2.3), can you provide, beyond the attention-laterality probe in Appendix C, band-limited (μ/β\mu/\beta, 8–30 Hz) or channel (C3/Cz/C4) specificity analyses showing that TAHLP targets MI-specific rhythms better than generalist models? In Appendix B (Figure 4), how do collapse probe pwc and visibility gap compare between n=1n=1 and n=3n=3?

  2. [clarification] Step-Matched Baselines in Appendix G: In Appendix G, you note that keeping published epochs on 221h reduces baseline update steps by 11×–260×11\times\text{--}260\times and leaves schedules (BIOT 58,021h) and VQ codebooks (LaBraM 8,192 codes) untuned. If you match update steps and perform a modest sweep for single-stage continuous baselines (BIOT, CBraMod) on the 221-hour MI corpus, how much of the Appendix G margin (+1.42 to +2.75 points in Tables 7–8) remains?

  3. [clarification] Robustness to Coordinate Noise: How robust is I3D-ERoPE (Section 2.2) when 3D scalp coordinates are perturbed with Gaussian noise (e.g., σ∈[2,10]\sigma \in [2, 10] mm)?

  4. [future work] Transfer to Non-MI EEG Paradigms: Have you evaluated whether representations learned from your 221-hour MI corpus (Section 3.2) transfer to non-MI tasks (such as sleep staging on Sleep-EDF or event detection on TUEV)?

  5. [clarification] Computational Cost of Complementary Masking: In Section 2.1, complementary dual-mask passes (MM and Mˉ\bar{M}) require two student forward passes per step. What is the wall-clock training slowdown compared to a single-pass mask with an equivalent total prediction budget?

  6. [clarification] Regularity in reported ablation values: In Table 2, the SHU-MI column equals the PhysioNet-MI column minus 1.27 in every row, so all five per-step gains are identical on the two benchmarks (+3.00, +0.50 for the REVE-style embedding, +1.70 for I3D-ERoPE, +1.20, +0.77; also stated in Section 3.5). In Figure 3, the Uniform-to-Increasing difference is exactly 0.90 Balanced Accuracy points on all four benchmarks (65.35 to 64.45, 63.98 to 63.08, 56.15 to 55.25, 73.84 to 72.94). For independent five-seed runs on different datasets this regularity is unexpected. Could the authors confirm that these values come from separate runs on each benchmark and provide per-seed results?

Reviewer_Fn2LOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TAHLP, a motor imagery (MI) EEG foundation model aimed at handling heterogeneous electrode montages across datasets. The model pre-trains a 12-layer Transformer on approximately 221 hours (207,887 cue-locked trials) from 22 public datasets using a data2vec-style hierarchical latent prediction self-distillation objective. In this framework, an EMA teacher provides continuous latent targets and a student predicts them under complementary masks via cumulative residual refinement heads. To handle diverse channel layouts without spatial interpolation, the paper introduces I3D-ERoPE, a parameter-free 3D rotary position embedding based on native scalp coordinates, alongside learnable electrode-name embeddings. The proposed model is evaluated on four public MI benchmarks under a subject-independent setting, reporting improvements over existing generalist foundation models and specialist decoders.

Critical strengths
  • The motivation of handling heterogeneous electrode montages natively without spatial interpolation onto a standard template is sensible and practically important for EEG and BCI applications.

  • The idea of using a codebook-free, label-free continuous latent prediction objective is interesting, as it avoids multi-stage codebook fitting, negative pair sampling, and task-label reliance commonly seen in other self-supervised EEG models. Tracking training health diagnostics such as visibility gap and pairwise cosine similarity during pre-training is also a good practice.

  • The paper provides a detailed corpus inventory across 22 datasets with subject-level splits, and the step-by-step ablation study helps in understanding the contribution of different components.

Critical weaknesses
  • The core concepts — continuous EMA latent prediction and multidimensional RoPE — are already established in self-supervised learning (e.g., data2vec, I-JEPA, and M-RoPE). While combining them for MI may be new, the claim that this is the "first latent prediction for MI" overlooks related existing work. For example, EEGPT (NeurIPS 2024) predicts momentum-encoder representations from masked inputs and evaluates on MI. Other relevant foundation models like MIRepNet (cited in the introduction), REVE, LUNA, and EEG-X are missing from the comparisons.

  • The motivation behind the pre-training objective feels somewhat contradictory. The paper argues for self-supervised learning due to label scarcity, yet the pre-training corpus actually consists of ~208k cue-locked labeled trials where the class labels are simply discarded. Since trial windowing itself relies on event annotations, and no truly unlabeled continuous or resting EEG is used, I wonder why supervised or semi-supervised multi-dataset pre-training on this exact corpus was not explored as a natural baseline. Furthermore, the claim that waveform reconstruction mainly captures nuisance noise lacks direct empirical validation; the teacher still receives the same filtered recording, and there is no spectral or artifact-robustness analysis showing that latent targets preserve sensorimotor rhythms (μ\mu/β\beta) better.

  • The hierarchical cumulative residual refinement seems somewhat overkilled. All prediction heads regress to the exact same teacher target using a cosine loss, which functions more like deep supervision with a larger predictor rather than proving true multi-level refinement. The ablation in Table 2 shows that the initial level accounts for most of the gain. Missing here is a capacity-matched single-head or predictor-transformer control to justify the added complexity. Additionally, random 70% token masking on 1-second patches with common average referencing is likely too easy to reconstruct from temporally and spatially adjacent channels, as evidenced by the high masked alignment (cm≈0.89cm \approx 0.89). Using channel-block or time-block masking would be more convincing.

  • The claim that the model handles electrodes "by construction" is not fully substantiated. Even though the rotary transform depends on relative coordinates, each token still incorporates a learned channel-name embedding. If an unseen electrode is encountered, it must be assigned an untrained embedding vector, so computing a rotation alone does not provide a complete representation. Most importantly, there are no transfer experiments on unseen montages, channel subsets, or coordinate perturbations to verify these geometric claims on downstream tasks.

  • In my opinion, the experimental design for evaluation is not yet sufficient to support the claim of SOTA performance across all benchmarks. Several test sets are extremely small in terms of subject count, for example, BCICIV-2a has only 2 test subjects, BNCI2015001 has 3, and SHU-MI has 5. Reporting standard deviations over five random seeds only measures initialization variance, not inter-subject variability, which is well known to be large in EEG. The reported margins on BCICIV-2a (+0.58) and SHU-MI (+0.86) heavily overlap with the standard deviations, and no statistical significance tests or paired subject-level results are provided. PhysioNet-MI (20 test subjects) is the only adequately sized test set. I would defense for LOSO evaluation, which is probably better suit for academic settings and small datasets like these, as it maximizes training data and directly measures cross-subject generalization.

  • I am not convinced by the same-corpus comparison with the re-pre-trained generalist baselines in Tables 7 and 8. Retraining these models for their original epoch counts on a corpus that is 11x to 260x smaller drastically reduces the total optimizer steps. Their architectural hyperparameters (such as codebook sizes and learning rate schedules) were tuned for massive corpora, whereas TAHLP was tuned specifically on this 221h corpus. The authors acknowledge in Appendix G that this setup creates a bias favoring TAHLP; therefore, concluding in the main text that the performance gain comes purely from the objective rather than the data budget is not convincing.

  • Important BCI decoders and domain adaptation baselines are also missing, such as EEGNet or Conformer combined with Euclidean alignment (EA), as well as Riemannian / tangent space classifiers. The paper excludes Riemannian methods on the grounds that they require test labels, but unsupervised covariance alignment (such as EA) is routinely used in cross-subject MI. In addition, the exact fine-tuning and preprocessing details need clarification, particularly whether per-subject normalization or Euclidean alignment was applied to test subjects at inference time.

  • Finally, there are noticeable reporting anomalies that raise questions. In Table 2, the performance difference between PhysioNet-MI and SHU-MI is exactly 1.27 across all six rows, and in Figure 3, the drop between default and uniform weighting is exactly 0.90 across all four benchmarks. These identical numbers across distinct datasets seems weird and need clarification. There is also inconsistency regarding the training budget: while the paper mentiones 221 hours, the actual training split contains 178 hours (or ~199 hours including validation).

Questions for authors
  • Could you report LOSO cross-validation results or paired statistical significance tests across subjects to confirm whether the performance gains on the small-cohort benchmarks (BCICIV-2a, BNCI2015001, SHU-MI) are statistically significant?

  • How does TAHLP perform on downstream setups with completely unseen electrode names, where the learned channel-name embeddings are uninitialized? Have you tested transfer under channel dropout, low-density montages, or coordinate perturbations?

  • In the same-corpus control (Tables 7 and 8), the baselines received significantly fewer optimization steps due to the reduced dataset size. How would the comparison look if the generalist baselines were re-trained with a step-matched or compute-matched schedule?

  • Could you comment on the identical deltas observed in Table 2 (where the gap between PhysioNet-MI and SHU-MI is exactly 1.27 across all variants) and Figure 3 (where the drop from default to uniform weighting is exactly 0.90 across all four benchmarks)?

  • Could you clarify whether per-subject z-score normalization or Euclidean alignment was computed using the test subject's trials during downstream evaluation?

  • What is the performance of the backbone trained from scratch on the downstream benchmarks without pre-training, or when evaluated with a frozen backbone using a linear probe?

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

Related papers

Discussion (0)

Sign in to comment.