acceptodds
Under review as a conference paper at ICLR 2027

Learning EEG Structure in Its Native Coordinates: Operator-Preconditioned Multi-View Pretraining

Abstract

EEG foundation models are trained to learn general-purpose latent representations, but different physiological structures are not equally easy to discover in the same feature space of the signal. We study feature-space mismatch: under finite data, model capacity, optimization, and readout budgets, a structure may be present in EEG while remaining difficult for a learned representation to expose and use. A deterministic change of feature space does not add information, but can make such structure more accessible to a finite learner. Motivated by this observation, we propose operator preconditioning: explicitly compute structure-aware EEG measurements, then use self-supervision to learn their organization in the resulting feature space. A learning-free operator already reads a known physiological change more accurately than the frontends, and even the pretrained encoders, of the EEG foundation models we evaluate, highlighting the importance of the feature space before representation learning itself. We instantiate this principle as OpView, a multi-view EEG foundation model with independently pretrained waveform, spectral, cross-spectral, spatial, and dynamical views. Operators can be replaced or extended according to the structures and invariances relevant to different task families. Using operator-only, matched random-encoder, and pretrained-encoder controls, we separate the contributions of measurement, architecture, and learning. Under the EEG-FM-Compass evaluation protocol, OpView achieves strong performance across multiple EEG paradigms. These results support operator-preconditioned representation learning as an extensible approach for directing foundation-model capacity toward structures that downstream EEG decoding can effectively use.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
EEG foundation models, self-supervised learning, operator preconditioning, multi-view representation learning, representational accessibility, transfer learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_i42GOpenReview ↗
Rating
4(+2)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper analyzes the problem from the perspective of feature-space mismatch: while a physiological structure may be present in the raw EEG data, it is not necessarily easily decodable from pre-trained representations—especially given constraints such as limited model capacity, scarce annotations, and the specific readout mechanisms employed. The authors propose operator preconditioning, which involves using deterministic physiological signal operators to construct an input space tailored to the target structures, followed by pre-training an encoder within this space. The resulting framework, dubbed OpView, incorporates five independently trained "views"—Waveform, Frequency, Relation, Projective, and Dynamics—each with its own specific objective. Theoretically, the paper employs the conditional expectation of the squared loss to characterize the "reward" that a prediction target provides regarding the underlying physiological structures. The approach is evaluated across nine datasets and various cross-subject and intra-subject scenarios, with comparisons drawn against several baselines, including fixed operators, randomly initialized encoders, standard pre-trained encoders, and 12 EEG foundation models featured in the EEG-FM-Compass benchmark.

Critical strengths
  1. Clearly disentangles "information inherent to the operator," "inductive biases of the architecture," and "learning gains derived from pre-training." By fixing the coefficient of determination (R2R^2) for the power operator at approximately 0.99—compared to ~0.63/0.61 for patch embeddings and ~0.93/0.90 for pre-trained encoders—the study provides empirical grounding for its arguments regarding feature space mismatches.

  2. Demonstrates an awareness of structural adaptation across physiological tasks, while transparently reporting instances of failure. The framework employs distinct views to capture temporal transients, spectral power, inter-channel relationships, and polarity-invariant topology and dynamics. Notably, for the P300 task, a view that discards sign information performs no better than random chance, illustrating that no single architectural view is a "universal solution" (pp. 6–8). The appendix further clarifies key procedural details, such as the presence of evaluation data within the unsupervised pre-training set and the use of transductive alignment protocols.

Critical weaknesses
  1. The reported performance, which claims to be "better than the best foundation model," relies on post-hoc selection of individual views; it does not yet constitute a label-free selection strategy suitable for direct deployment. In Table 2, the "View" column reflects the selection of the optimal pre-trained view for each specific dataset/setting—an optimal view that varies across different tasks. While the "Fusion" approach avoids the use of test labels, it combines features from operators, random encoders, and pre-trained models, and employs a ridge regression or ensemble readout mechanism that differs from those used with the published foundation models.

  2. The benchmarking comparisons suffer from significant incompleteness and issues regarding resource comparability. Appendix Table 14 (page 28) acknowledges that some entries are based on only a single classifier seed, some cover only the first seven held-out subjects, and the window counts for the CHB-MIT dataset are inconsistent with the established benchmarks. Furthermore, the unsupervised pre-training phase incorporates data from a subset of subjects across the seven evaluation datasets; the total amount of data exposed to each model during pre-training—and the overall fairness of these comparisons—requires standardized reporting. Additionally, the cumulative computational cost of the five independent encoders and the fused readout mechanism should be compared against that of a single foundation model.

Questions for authors
  1. If a fixed set of view selection and readout rules is determined using only the validation set—prohibiting the selection of the "best" view based on the test set—what would the overall win/tie/loss statistics be across the 9 datasets when compared against the 12 publicly available baseline models? Please provide a comparison following the same protocol as the main table.

  2. Please address the missing subject-fold entries and classifier random seeds noted in Table 14, as well as the window alignment for the CHB-MIT dataset; then, report how these revisions affect the core conclusions.

  3. What are the differences in performance and computational cost for the Fusion approach when using only pretrained views, pretrained views plus operators, and the full configuration (including random encoders)? This will help determine whether the performance gains stem from representation learning or from the engineering of the readout mechanism.

9 comments
AuthorsOpenReview ↗

Response to Reviewer i42G (1/2)

Response to Reviewer i42G (1/2)

TL;DR: Across Parts 1 and 2, we replace post-hoc per-view reporting with a single deployment-ready model combining only the five pretrained views via one validation-calibrated rule (R-cal, fitted strictly on training subjects without test labels) under a shared ridge probe across all models (W1, Q1), isolate parameter-free operators and random encoders as standalone control models (Q3), and complete all 131 held-out subject folds across 3 seeds with corrected CHB-MIT seizure parsing (29,841 windows), a unified zero-exposure pretraining corpus (61 sources, 2,321 h), and full computational cost comparisons (W2, Q2).

We sincerely thank the reviewer for the detailed and constructive review, which has helped us make our evaluation fully unified and deployment-ready. Below we address W1, Q1, and Q3, followed by W2 and Q2 in Part 2.

W1 and Q1. Deployment-ready validation-calibrated fusion (R-cal) and win/tie/loss statistics (Section 5.5, Tables 2, 26–28)

While the initial View column in Table 2 was included alongside Fusion to illustrate which physiological view aligns with each paradigm, we agree that main benchmark comparisons should evaluate a single, label-free deployment model under an identical readout protocol across all baselines:

  • Single fixed model (no post-hoc view selection): We removed the single-view column from Table 2. On every dataset, OpView combines the five pretrained views using a single rule (R-cal) whose weights are fitted strictly on the training subjects of each leave-one-subject-out fold, using no test labels.
  • Unified protocol across all models: OpView and 12 published encoders (10 from EEG-FM-Compass, with BrainOmni in two sizes, plus ST-EEGFormer; EEGMamba is added in our Nov. 15 update, and MIRepNet is a motor-imagery-only model) are read by the exact same ridge probe and random projection cap (32,768 dimensions), tested via paired Wilcoxon signed-rank tests over held-out subjects (p<0.05p < 0.05).

Under this unified protocol, OpView is statistically higher than almost all published encoders (Table 28; ‡ indicates baseline pretraining exposure to evaluation datasets):

Published encoderOpView higher / tied / lower (p<0.05p < 0.05)
CBraMod6 / 3 / 0
EEGPT‡ (8 datasets)5 / 3 / 0
LaBraM‡6 / 3 / 0
ST-EEGFormer‡ (8 datasets)2 / 5 / 1
LUNA7 / 2 / 0
BENDR, BIOT‡, Neuro-GPT, TFM-Tokenizer‡, BrainOmni (2 sizes), SingLEM7 to 9 higher each, 0 lower

Against 11 of the 12 encoders, OpView is higher on 5 to 9 datasets and lower on none; ST-EEGFormer is the one encoder at our level. Even against the post-hoc best baseline on each individual dataset (a choice made on test scores), our fixed model is statistically indistinguishable on 7 datasets, higher on CHB-MIT, and lower on SEED (Table 2; with per-subject alignment on both sides, OpView is higher on BCI IV-2b and indistinguishable on the other 8). For reference only, Table 26 reads OpView with the benchmark's unregularized linear head on the raw concatenation of all five views (leading on EEGMat, 77.4 vs. 68.6, and Nakanishi2015, 86.2 vs. 85.5, but trailing where the informative view holds only a few percent of total variance, such as Dynamics at 2–5% on motor imagery—showing why per-view normalization and calibration in R-cal are essential).

Q3. Isolating pretrained views from operators and random encoders (Appendix F, Table 18)

To determine whether OpView's gains stem from representation learning or from mixing operators and random networks in the readout, our main model contains only the five pretrained views, while operators and randomly initialized encoders form standalone control models built by the same rule:

9 datasets pooled (mean over 131 held-out subjects)Fitted view weights (R-cal)Equal weights
Pretrained views (OpView) − operators only+2.6 (p=4×10−5p = 4 \times 10^{-5})+1.9 (p=0.017p = 0.017)
Pretrained views (OpView) − random encoders only+1.0 (p=0.018p = 0.018)+0.3 (p=0.67p = 0.67)
Adding operators to pretrained views (10 blocks)+0.2 (p=0.31p = 0.31)+0.2 (p=0.46p = 0.46)
Adding operators + random encoders (15 blocks)+0.2 (p=0.22p = 0.22)+0.6 (p=0.10p = 0.10)
  • The five pretrained views are sufficient: Adding operators or random encoders changes the pooled score by only +0.2 points (not statistically significant; higher on at most 1–2 single datasets by 0.4–0.7 points and lower on none) while requiring 2× and 3× as many per-view probes (and 2× encoder compute for random encoders).
  • Component attribution (rows 1–2): Most of the baseline performance comes from the operator spaces and architecture, and self-supervised pretraining adds a significant further gain (+2.6 over operators, +1.0 unaligned and +1.3 aligned over random initialization, driven by Dynamics +5.0 and Projective +2.9; Table 1).
AuthorsOpenReview ↗

Response to Reviewer i42G (2/2)

Response to Reviewer i42G (2/2)

W2 and Q2. Complete 3-seed subject-fold coverage and CHB-MIT window alignment (Appendix E, Appendix G)

  • Motivation: To eliminate any partial-fold or seed-sampling variance and resolve the window-count discrepancy on CHB-MIT, we completed all 131 held-out subject folds across all 9 datasets with 3 classifier seeds and audited the benchmark's public CHB-MIT data generator.
  • Result:
    1. Complete folds and seeds: Every frozen evaluation table in the revised paper covers 100% of subjects across all datasets with 3 random seeds.
    2. Root cause of the CHB-MIT window discrepancy: In 20 lines of the upstream PhysioNet CHB-MIT summary files, the colon is followed by two or three spaces instead of one. The benchmark's public regex parser skips those 20 lines and mis-pairs subsequent start/end timestamps, dropping 23 true seizures and mislabeling 831 interictal windows as seizure. Fixing only this whitespace regex yields 29,841 windows (matching the benchmark's intended 29,840 count), and we evaluate OpView and every published baseline on the exact same corrected dataset. (The 75.2% score reported for CBraMod in the benchmark paper comes from a tuned head on the mis-parsed labels; under the shared ridge probe on the corrected labels CBraMod scores 67.5%, with the difference driven primarily by decision thresholding under class imbalance; Appendix G.)
  • Significance for our core conclusions: Completing all folds, seeds, and CHB-MIT windows confirms and sharpens all three conclusions: (1) operator-defined feature spaces expose task-relevant structure and distinct views are complementary; (2) under a single validation-calibrated rule (R-cal) and shared probe without test-label selection, OpView matches the post-hoc best baseline on 7 of 9 datasets, beats it on 1 (CHB-MIT), and trails on 1 (SEED); and (3) frozen pretraining adds +1.0 point unaligned and +1.3 aligned over random initialization across the 9 datasets (and +8.3 to +13.8 points on large-scale TUAB and Sleep-EDFx; see our Nov. 15 update), whereas end-to-end supervised fine-tuning adapts random and pretrained initializations comparably within the operator space (Table 23).

W2. Standardized zero-exposure pretraining corpus and cross-model exposure audit (Appendix D.1, Appendix J, Table 34)

  • Motivation: To ensure strictly leak-free evaluation and transparent comparison against published models whose pretraining corpora overlap with downstream benchmarks, we standardized all OpView checkpoints on a zero-exposure corpus and documented corpus exposure across all baselines.
  • Result & Significance: Every OpView result in the revised paper uses standardized 30,000-step checkpoints trained on 61 public sources (2,321 hours) that exclude all 13 benchmark datasets entirely (Appendix D.1, PROVENANCE.md). In addition, Table 34 documents pretraining data, compute budgets, and benchmark exposure for all compared models: notably, SEED is included in the pretraining corpora of EEGPT and LaBraM, and BCI IV-2a and 2b are included in ST-EEGFormer. Even with zero exposure to those datasets, OpView is statistically higher than EEGPT on SEED (+7.3 points, p=0.005p = 0.005) and ST-EEGFormer on BCI IV-2b (+10.8 points, p=0.008p = 0.008), and statistically indistinguishable from LaBraM on SEED (+1.9) and ST-EEGFormer on BCI IV-2a (+1.1).

W2. Total five-branch computational cost vs. single foundation models (Table 33)

  • Motivation: Because OpView comprises five independent branches that share no parameters, its deployment and pretraining cost must be evaluated as the sum of all five branches against single-backbone foundation models.
  • Result & Significance: Summing across all five branches, OpView totals 19.3M parameters, 33.5 GFLOPs, and 16.6 ms latency per 22-channel, 4 s window (on one RTX 5060 Ti, batch size 32), plus ~2 ms on CPU for the five ridge probes and R-cal fusion. Total pretraining across all five branches requires 24–27 A100-equivalent GPU-hours (about 101810^{18} total FLOPs)—4 to 8 times fewer GPU-hours than reported for single foundation models such as CodeBrain, CBraMod, and LUNA (Table 33).

We thank the reviewer again and would be glad to clarify any remaining points.

AuthorsOpenReview ↗

Update for Reviewer i42G: completed comparisons

Update for Reviewer i42G: completed comparisons

TL;DR: We report three completed sets of experiments addressing Q1, Q2/W2, and Q3: (1) direct evaluation of our five parameter-free operators alone against all 12 published encoders under the shared ridge probe, showing that the operators alone beat 10 of 12 published encoders when pooled (Q3, Table 28); (2) side-by-side unaligned and per-subject aligned pretraining gains (Q3, Table 1: +1.3 over random initialization and +2.2 over operators under alignment, with Dynamics, Waveform, and Projective all gaining significantly) plus initial equal-weight results on three extension datasets (Q1, Appendix K); and (3) complete 3-seed fine-tuning across all 9 datasets (Q2/W2, Table 23) and anonymous weight release.

Thank you again for your review. We have uploaded a revised manuscript and supplementary package with improved readability (Section 5), completed experimental runs, and anonymous pretrained weights on Hugging Face (https://huggingface.co/Anony-0x3f/opview-v2-30k). All table, section, and appendix numbers match our first reply. Below we summarize the new results most relevant to your questions.

1. Q3: Disentangling representation learning from feature-space priors and readout (Table 28, new columns; Table 1)

  • Motivation: To test directly how much of OpView's advantage over published foundation models comes from the operator-defined feature spaces versus pretraining—under the exact same ridge readout for every model—we added a head-to-head comparison of the five parameter-free operators alone against each published encoder (Table 28) and placed unaligned and aligned pretraining decompositions side by side (Table 1).
  • Result & Significance:
    • Operators alone vs. published encoders (Table 28): Combined by the same rule and read by the same ridge probe without any pretraining, the five parameter-free operators are statistically higher than 10 of the 12 published encoders when pooled across the 9 datasets, and not statistically different from EEGPT and ST-EEGFormer. Most of the gap over existing encoders is therefore already established by the operator-defined coordinate spaces under a unified linear probe.
    • Pretraining contribution without and with alignment (Table 1): Alongside the unaligned gains (+1.0 over random initialization, +2.6 over operators), per-subject alignment yields +1.3 points over random initialization (p=1×10−4p = 1 \times 10^{-4}) and +2.2 points over the operators alone (p=7×10−6p = 7 \times 10^{-6}), with three views gaining statistically (Dynamics +7.8, Waveform +4.5, and Projective +4.4).

2. Q1: Extending evaluation beyond the 9 BCI datasets (Appendix K, Tables 36–37)

  • Motivation & Result: To test generality on larger clinical, vigilance, and visual-retrieval benchmarks, we applied the same five pretrained views and ridge probe to three additional EEG-FM-Compass datasets. Under an initial uncalibrated equal-weight control (with validation-calibrated R-cal and Sleep-EDFx completed in our Nov. 15 update):
    • TUAB (clinical abnormality): OpView reaches 81.0%, statistically higher than 11 of the 12 published encoders and matching EEGPT (80.5%), even though 7 of the 12 baselines include TUAB in their pretraining corpus while OpView has zero exposure.
    • SEED-VIG (vigilance regression): Prediction error is statistically lower than 8 published encoders and not different from the other 4.
    • Things-EEG2 (visual retrieval): Under equal weights—where two non-visual views (Frequency, Relation) are at chance and weighted equally with the others—OpView beats 5 encoders, ties 1, and trails 6 (54.5% vs. 60.1% for ST-EEGFormer; as shown in our Nov. 15 update, applying our standard validation-calibrated R-cal rule raises Things-EEG2 by +3.6 points to 58.1%/58.3%, beating 12 of 13 encoders).

3. Q2 and W2: Complete 9-dataset 3-seed fine-tuning and cost discussion (Tables 2 and 23, Section 6)

  • Motivation, Result & Significance: Completing 3-seed end-to-end fine-tuning of the Dynamics encoder across all 9 datasets (adding BNCI2014008, SEED, and CHB-MIT; Table 23) shows that pretrained and randomly initialized encoders do not differ statistically on any dataset without alignment (−1.8 to +0.9 points), while with alignment pretrained initialization is higher on BNCI2015001 (+1.7, p=0.036p = 0.036) and lower on BNCI2014009 (−1.1 AUROC, p=0.002p = 0.002) and CHB-MIT (−4.3, p=7×10−5p = 7 \times 10^{-5}). This confirms that full supervised optimization adapts weights effectively inside the operator space from scratch, whereas self-supervised pretraining specifically benefits frozen transfer to unseen subjects (Section 5.3). Section 6 also explicitly notes the five-branch inference cost alongside Table 33.

We would be glad to clarify any remaining points.

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

This paper investigates feature-space mismatch in EEG foundation models, arguing that physiologically meaningful information may be present in raw EEG signals but remain difficult to recover from learned representations under finite data, model capacity, optimization, and readout constraints. The authors demonstrate this phenomenon through diagnostic experiments on existing EEG foundation models, showing that simple deterministic operators can sometimes expose relevant physiological structures more effectively than pretrained neural representations. Experiments follow the EEG-FM-Compass benchmark across nine datasets, six EEG paradigms, and four evaluation settings. The authors compare operator-only features, randomly initialized encoders, pretrained encoders, and multi-view combinations. The results demonstrate meaningful improvements from pretraining in several settings, particularly for the Frequency and Dynamics views, while also revealing substantial task dependence and limitations in cross-subject generalization.

Critical strengths

The most compelling contribution is the systematic examination of whether existing EEG foundation models make physiologically relevant information accessible to downstream learners. Rather than assuming that a powerful pretrained encoder necessarily produces more useful representations, the authors directly compare fixed physiological operators, patch embeddings, and complete pretrained encoders.

Critical weaknesses

W1. The empirical evidence does not convincingly establish a consistent advantage of the proposed pretraining framework over its simpler components. Although the paper demonstrates that operator preconditioning can make certain physiological structures easier to learn, the downstream results raise questions about how much additional benefit is attributable to the =trained representations. Although this is a legitimate design choice, it makes the overall system performance insufficient evidence for the effectiveness of operator-preconditioned pretraining itself.

W2. The generality and practical benefits of the multi-view framework remain insufficiently established. The proposed approach relies on five independently pretrained encoders, each with a dedicated operator, architecture, and objective. While this design is modular, it introduces additional training and inference costs and requires selecting or combining views according to downstream tasks.

Questions for authors

Q1. How much of the improvement of the complete OpView framework can be attributed specifically to self-supervised pretraining?

Q2. What is the benefit of independently pretrained views compared with simpler alternatives under matched total computational budgets?

Q3. How robust are the results to the choice of readout protocol and representation dimensionality?

6 comments
AuthorsOpenReview ↗

Response to Reviewer rMDs (1/2)

Response to Reviewer rMDs (1/2)

TL;DR: Across Parts 1 and 2, we isolate the contribution of self-supervised pretraining from parameter-free operators and random initialization under a single frozen probe (W1, Q1), explain what physiological structures each pretext objective learns and why label-free subject alignment unlocks cross-subject transfer (W1, Q1), show that five independent operator views beat compute-matched single-view and ensemble alternatives (Q2), verify robustness across 12 readout protocols and widths (Q3), and show that one validation-calibrated rule (R-cal) combines all five views without per-task selection at 4–8× lower training cost than recent EEG foundation models (W2).

We sincerely thank the reviewer for the insightful review. Your questions go directly to the core of our work. Below we address W1 and Q1, followed by Q2, Q3, and W2 in Part 2.

W1 and Q1. How much of the improvement comes from self-supervised pretraining?

  1. Separating operator spaces, architecture, and pretraining (Table 1). Because an overall downstream score conflates input representation with learning, we evaluate three controlled variants per view—the parameter-free operator alone, the identical architecture at random initialization, and the pretrained encoder—under the exact same ridge probe across all 131 held-out subjects of the 9 datasets (paired Wilcoxon test):
Variant comparison (9 datasets, 131 subjects)Pretrained − random init.Pretrained − operator
Model of five views (calibrated weights, R-cal)+1.0 (p=0.018p = 0.018)+2.6 (p=4×10−5p = 4 \times 10^{-5})
Dynamics view alone+5.0 (p=6×10−7p = 6 \times 10^{-7})+10.2 (p=3×10−11p = 3 \times 10^{-11})
Projective view alone+2.9 (p=1×10−5p = 1 \times 10^{-5})+0.6 (p=0.14p = 0.14)

Consistent with our thesis and findings on structured random networks (Saxe et al., 2011), the operator-defined feature spaces provide the primary foundation on compact BCI tasks, while self-supervised pretraining adds statistically significant gains where tasks depend on temporal forecasting (Dynamics) or spatial topography (Projective) rather than static first-order statistics. Under uncalibrated equal weights without alignment, the three views where the operator already captures the linear structure (Waveform, Frequency, Relation; Table 19) dilute the two improved views to +0.3, motivating validation-calibrated fusion (R-cal) and subject alignment.

  1. Unlocking cross-subject transfer via label-free subject standardization (Table 22, Appendix H). Unsupervised pretraining naturally captures subject-specific traits alongside task structure, which can obscure pretraining gains when a frozen probe is applied directly across unseen subjects. Variance partitioning confirms that pretraining increases the between-subject share of feature variance in 10 of 12 view–dataset pairs (Appendix H). Standardizing each test subject's features using only their unlabeled trials makes the pretrained model statistically higher than random initialization on 4 of 9 datasets and lower on none, raising the pooled five-view gain to +1.3 points with calibrated weights (p=1×10−4p = 1 \times 10^{-4}) and +1.4 points with equal weights (both statistically significant; Table 22). Removing inter-subject baseline shifts without labels thus lets a frozen probe access the learned representations consistently.

  2. How pretext objectives shape learned physiological structure (Section 5.3, Appendix H, Tables 5 and 15). To test what pretraining learns and why gains vary across views, we probed physiological targets on held-out BCI IV-2a subjects. A linear probe on pretrained Waveform features tracks trial-to-trial mu-band power at r=0.77r = 0.77 versus 0.370.37 at random initialization; however, because raw-patch MSE weights amplitude, Waveform trades slow 0–2 Hz trial drift (r=0.89→0.70r = 0.89 \to 0.70) for rhythm, lowering single-view BCI IV-2a accuracy by 6.1 points. By contrast, Dynamics—trained to forecast temporal change—retains slow drift (0.81→0.870.81 \to 0.87) while gaining mu rhythm (0.65→0.920.65 \to 0.92), boosting classification by +8.5 points. A parallel trade-off appears in published models (Table 5): raw-patch ST-EEGFormer gains +7.2 points on BNCI2015001 and +15.0 on Nakanishi2015 from pretraining, whereas models with stronger built-in priors can lose task cues under a single objective (e.g., LUNA drops from 47.8% to 30.4% on Nakanishi2015). For dense operator views (Frequency, Relation, Projective), full-output autoencoding and finer 25% masks avoid the cue starvation of coarse patch masks (0.32–0.97 vs. 0.77–1.05 relative error, Table 15), and per-dataset input scaling avoids corpus-wide range compression on EEGMat (Appendix H). Pretraining thus amplifies the structure rewarded by its objective, showing why pairing complementary operators with tailored objectives is essential.

AuthorsOpenReview ↗

Response to Reviewer rMDs (2/2)

Response to Reviewer rMDs (2/2)

Q2. Independent views vs. simpler alternatives at matched compute (Section 5.4, Table 25)

To test whether OpView's gains arise from combining complementary operator spaces rather than from more training steps or ensembling multiple networks, we ran three compute-matched controls:

  1. Matched total training steps (5 × 5k vs. 1 × 25k): Five views trained for 5k steps each are statistically higher than one Waveform encoder trained for 25k steps on 4 of 6 datasets and lower on none. Against a 25k-step Dynamics encoder (our strongest single view), they are higher on 1 dataset, lower on 1, and tied on 4.
  2. Extended single-view training (75k vs. 30k): Training a Waveform encoder for 75k steps instead of 30k yields no statistically significant gain on any of the 9 datasets (all differences within 1.5 points), whereas the five views at 30k steps each are statistically higher than the 75k Waveform encoder on 6 of 9 datasets and lower on none.
  3. Distinct operator views vs. homogeneous ensemble: At random initialization, five distinct operator views are statistically higher than five copies of one view on 5 of 9 datasets and worse on none, while five copies stay within 1.8 points of one copy.

These controls show that scaling training steps or ensembling identical encoders within one space quickly saturates; the improvement comes from exposing distinct physiological structures across complementary operator spaces. Independent pretraining is also structurally required: while a time–frequency model like CBraMod can train branches jointly toward a shared raw-patch target, several of our operators map EEG segments to summary statistics (spectral power, sensor coupling, spherical-harmonic energy) whose inversion back to raw waveforms is ill-posed. Without a shared reconstruction target, joint training could only sum five disparate losses, letting the easiest objective dominate and suppress harder modality-specific features.

Q3. Robustness to the linear probe and representation dimensionality (Appendix G, Table 26)

To verify that pretraining gains are an intrinsic property of the representations rather than a probe artifact, we evaluated 12 readout settings across three linear classifiers (the benchmark's head, ridge regression, and logistic regression) and feature widths from 512 to 32,768. The sign of the pretrained-minus-random difference agrees in at least 10 of 12 settings for 20 of 29 view–dataset pairs, and in 7 of the remaining 9 pairs the difference is below 1 point. Because our ridge probe also improves published encoders (CBraMod, EEGPT, LaBraM) by +3.9 points on average over the benchmark's head in 10 of 15 pairs, all main comparisons use the exact same ridge probe (with benchmark-head reference scores in Table 26). Whether pretraining helps is thus governed by the view and task structure rather than readout tuning.

W2. Deployment-ready task fusion, generality, and computational cost (Appendices F and J, Table 33)

To ensure practical deployment without per-task engineering, one fixed rule (R-cal) combines all five views on every dataset, fitting weights strictly on the training subjects of each fold without test labels. Across 9 datasets in 6 paradigms (and all 13 datasets in our follow-up updates), the complete five-branch model totals 19.3M parameters, 33.5 GFLOPs, and 16.6 ms per window, and pretrains in 24–27 A100-equivalent GPU-hours (about 101810^{18} FLOPs)—4 to 8 times fewer GPU-hours than reported by CodeBrain, CBraMod, and LUNA (Table 33).

W1 and Q1 (continued). Frozen transfer vs. full supervised fine-tuning (Table 23)

To test whether pretraining adds gains when all weights are updated with downstream labels across 3 seeds, we fine-tuned the Dynamics encoder end-to-end. In the well-conditioned operator space, random initialization already reaches strong supervised accuracy (e.g., 55.2% on BCI IV-2a) and performs comparably to pretrained initialization across the 9 datasets (−1.8 to +0.9 points without alignment; +1.7 on BNCI2015001, p=0.036p = 0.036, and −1.1 AUROC on BNCI2014009, p=0.002p = 0.002, −4.3 on CHB-MIT, p=7×10−5p = 7 \times 10^{-5} with alignment). Full-network supervision adapts the architecture directly within the operator space; self-supervised pretraining specifically provides its value in frozen transfer to unseen subjects.

We thank the reviewer again and would be glad to clarify any remaining points.

AuthorsOpenReview ↗

Update for Reviewer rMDs: the share of pretraining, and why released encoders lose structure

Update for Reviewer rMDs: the share of pretraining, and why released encoders lose structure

TL;DR: This update provides direct mechanistic evidence for why single-space EEG foundation models lose structure (Section 3.3, Appendix B.8: probe gradients for band power and cross-channel correlation oppose each other in CBraMod with cosine −0.50), shows that our five parameter-free operators alone beat 10 of 12 published encoders under a shared probe (Table 28), extends evaluation to three further benchmarks where pretraining adds +1.7 points on clinical TUAB (p=0.002p = 0.002, beating 11 of 12 baselines; Appendix K), and completes 9-dataset fine-tuning and readout-width analyses (Section 5.3, Table 23).

Thank you again for your review. We have uploaded a revised manuscript with improved figures, completed controls, and anonymous weights (https://huggingface.co/Anony-0x3f/opview-v2-30k).

1. Why a released encoder loses structure: gradients at its checkpoint (Section 3.3, Appendix B.8)

You highlighted our examination of whether existing EEG foundation models make physiological structure accessible as the paper's most compelling contribution. To explain why four of five released encoders decode steady-state responses worse than their own patch embeddings, Section 3.3 extends Appendix B.8. At CBraMod's released checkpoint on BNCI2015001, we fitted two linear probes on eight subjects—one for band power and one for cross-channel correlation beyond it—and compared their improvement directions with the descent direction of masked reconstruction on two held-out subjects:

  • Opposing physiological gradients: The two probe directions oppose each other (cosine −0.50), so an update that makes one structure easier to read makes the other harder.
  • Pretext objective bias: The descent direction agrees with the band-power probe (cosine +0.028) and opposes the cross-channel probe (−0.129; 0.00005 for a matched random direction). A step that lowers pretraining loss improves the first probe and degrades the second.

Together with our finding that cross-electrode attention in CBraMod lowers reconstruction error while reducing stimulus separability, this illustrates a concrete mechanism: under one objective on one shared representation, structures compete and the objective decides which is favored. It motivates pretraining our views independently, each with an objective in its own observation space.

2. W1/Q1 & W2: Operator-space foundation and pretraining gains across frozen transfer and larger benchmarks (Tables 1, 28, Appendix K)

  • Operators alone vs. published encoders (Table 28): Combined by the same rule and read by the same ridge probe without any pretraining, our five parameter-free operators are statistically higher than 10 of the 12 published encoders when pooled across the 9 datasets, and not different from EEGPT and ST-EEGFormer. A pretrained EEG encoder should therefore be reported beside its parameter-free operator and a matched random encoder to isolate what was learned (Section 6).
  • Frozen transfer with alignment and on clinical EEG (Table 1, Appendix K): Alongside the unaligned gains (+1.0 over random initialization, +2.6 over operators), Table 1 shows per-subject alignment side by side: alignment raises the five-view gain over random initialization to +1.3 points (p=1×10−4p = 1 \times 10^{-4}), and Waveform (+4.5) joins Dynamics (+7.8) and Projective (+4.4) as views that gain statistically. On clinical TUAB (2,329 patients), pretraining adds +1.7 points over random initialization (p=0.002p = 0.002) under equal weights (and +10.0 to +11.2 points in Dynamics and Waveform; see our Nov. 15 update).
  • Three further benchmarks (Tables 36–37): Under an initial equal-weight control, OpView reaches 81.0% on TUAB (higher than 11 of 12 encoders and matching EEGPT at 80.5%), achieves lower error than 8 of 12 encoders on SEED-VIG (matching the other 4), and beats 5 of 12 on Things-EEG2 (54.5% under equal weights, where two views are at chance; see our Nov. 15 update for R-cal results across all 13 datasets).

3. Readout width analysis and full 9-dataset fine-tuning (Section 5.3, Table 23, Section 6)

Besides corpus-wide input scaling, an 8,192-column probe compresses the Frequency encoder's token states (at least C×75×192C \times 75 \times 192 values) far more than the operator output (Section 5.3). Across all 9 datasets under 3-seed fine-tuning (Table 23), pretrained and random Dynamics encoders do not differ statistically without alignment (−1.8 to +0.9 points; with alignment, pretrained is higher on BNCI2015001 by +1.7, p=0.036p = 0.036, and lower on BNCI2014009 by −1.1 AUROC, p=0.002p = 0.002, and CHB-MIT by −4.3, p=7×10−5p = 7 \times 10^{-5}), confirming that pretraining helps frozen transfer to unseen subjects. Section 6 notes the five-branch inference cost and per-view objective design. We are glad to clarify any remaining points.

Reviewer_SCg6OpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies feature-space mismatch in EEG foundation models: task-relevant structure may be present in the signal yet remain difficult to extract from a learned representation under limited learning and readout budgets. This paper proposes OpView, an EEG pretraining framework motivated by feature-space mismatch. Five independently pretrained views pair fixed signal operators with objectives tailored to waveform, spectral, relational, spatial, and temporal structure. Evaluation across nine datasets examines downstream performance, with operator-only and matched random-encoder controls used to distinguish feature construction from pretraining benefits.

Critical strengths
  1. The paper provides an solid framework for understanding representation quality. The distinction between information being present and being accessible under a fixed readout budget is well motivated.
  2. The overall design is coherent and practically extensible. Each view pairs an observation space with explicit invariances, a prediction target, and a dedicated encoder. This accommodates complementary EEG structures while allowing individual views to be added or replaced without retraining the others.
  3. The experiments identify both benefits and limitations of pretraining.
Critical weaknesses
  1. The incremental contribution needs clearer positioning against closely related approaches. Prior work already incorporates temporal–spectral separation, physiological-statistic prediction, complementary multi-view learning, and independently pretrained experts. OpView offers a coherent integration through view-specific observation spaces, invariances, objectives, and encoders. However, the current controls primarily establish benefits over fixed operators and random initialization; they provide less direct evidence of what this integration adds beyond the closest existing approaches. The paper should clarify and substantiate this additional methodological or empirical value.
  2. The manuscript and supplementary package give inconsistent descriptions of the pretraining corpus. Please clarify which configurations produced the reported results and whether the 40k–160k comparisons held the training data fixed.
Questions for authors
  1. Could you provide a checkpoint-to-result mapping covering training sources, subject partitions, sampling weights, schedule endpoints, and evaluated steps? Did the reported 40k–160k comparisons use identical data mixtures
  2. Could you clarify the preprocessing and dataset versions underlying the benchmark results?
7 comments
AuthorsOpenReview ↗

Response to Reviewer SCg6 (1/2)

Response to Reviewer SCg6 (1/2)

TL;DR: Across Parts 1 and 2, we standardize all benchmark tables on a single unified set of zero-exposure 30,000-step checkpoints trained on 61 public sources with complete provenance and preprocessing manifests (W2, Q1, Q2; Part 1), and clarify OpView's methodological and empirical contribution beyond prior multi-view and statistic-prediction approaches (W1; Part 2): preconditioning the input observation space of independently pretrained encoders exposes higher-order and non-stationary EEG structures that single-space and joint-objective models lose (confirmed by our operator-as-target control in our Nov. 12 update).

We sincerely thank the reviewer for the encouraging and constructive review. Below we address checkpoint standardization, provenance, and preprocessing (W2, Q1, Q2), followed by positioning (W1) in Part 2.

W2 and Q1. Standardized zero-exposure pretraining corpus, checkpoint mapping, and duration control (Appendix D.1, Tables 14 and 25, PROVENANCE.md)

To ensure that every benchmark comparison and training-duration ablation reflects an identical, fully documented, zero-exposure data mixture, we synchronized Appendix D.1 with our run manifests (PROVENANCE.md) and standardized all results on a single unified set of checkpoints:

  • Unified zero-exposure corpus (Appendix D.1, Table 14): All results in the revised manuscript use a single standardized set of checkpoints trained on 61 public sources (2,321 hours), excluding all 13 benchmark sources entirely (Relation uses 60 sources and Projective 59 due to electrode-coordinate and reference requirements).
  • Subject splits and sampling: Within each source, subjects are split 70/15/15 into pretraining, validation, and test partitions via a fixed seeded permutation, using only the 70% pretraining partition. Each source contributes a fixed share of sampled signal time (pretraining.json), with 32 windows per update.
  • Schedule and objectives: Each view is trained for 30,000 AdamW updates (peak learning rate 4×10−44 \times 10^{-4}, 1,500-step warm-up, cosine decay, weight decay 0.05) and evaluated at the final checkpoint. Waveform reconstructs masked patches and Dynamics forecasts signal change; Frequency, Relation, and Projective reconstruct the full operator output, with 25%-masked variants reported alongside (Section 5.2).
  • Controlled single-corpus duration check (30k vs. 75k steps, Section 5.4, Table 25): Holding the 61-source data mixture strictly fixed, extending training from 30,000 to 75,000 steps yields no statistically significant change on any of the 9 datasets (all differences within 1.5 points).
  • Clarification of initial configurations and complete mapping: Thank you for noting that Appendix D of the initial submission summarized the corpus at a higher level than TRAINING.md. In the initial submission, the 40k checkpoints were trained on compact paradigm-matched subsets (Lee2019 MI and ERP CORE for Waveform, Dynamics, Relation, Projective; Lee2019 SSVEP and six pretraining-split subjects of Nakanishi2015 for Frequency), whereas the 160k Waveform and Dynamics checkpoints used the 70-source collection (which, as marked by ‡, included unlabeled pretraining-split subjects from seven benchmark datasets). Because the 40k and 160k runs differed in both update count and source mixture, our revised 30k vs. 75k comparison on the unified 61-source corpus provides a strictly controlled test of training duration. PROVENANCE.md maps every initial and standardized checkpoint to its sources, partitions, weights, schedules, and steps.

Q2. Standardized preprocessing and dataset versions (Appendix E, Tables 16 and 17)

  • Raw versions: BNCI datasets and Nakanishi2015 from MOABB 1.5.0; EEGMat and CHB-MIT from PhysioNet 1.0.0; SEED from the publisher's Preprocessed_EEG release (session 1, as in the benchmark).
  • Generators: The benchmark's released scripts for seven of the nine datasets; our own MOABB scripts for Nakanishi2015 (no benchmark generator is released) and BCI IV-2b (matching the paper's filter and window). Under MOABB 1.5.0, upstream session names changed on three datasets, so we alias the first session to the expected name; on CHB-MIT we correct a whitespace-parsing bug in the seizure annotations (see our reply to Reviewer i42G).
  • Filters and shared reference rule: Pretraining recordings are band-passed and notch-filtered as described in the paper; evaluation data follow the benchmark's generators (no notch filter; 0.5–45 Hz band-pass in our two scripts). CMS/DRL and monopolar motor-imagery recordings are re-referenced to the common average and scaled by one label-free constant per dataset (applied to OpView and all published models alike), while Relation and Projective read every recording as acquired. Tables 16–17 give per-dataset details and all scripts are in the supplementary package.
AuthorsOpenReview ↗

Response to Reviewer SCg6 (2/2)

Response to Reviewer SCg6 (2/2)

W1. Methodological positioning and empirical value beyond closest prior approaches (Section 2, Table 32)

OpView builds on a rich body of prior work, and we appreciate the request to state precisely what our framework adds methodologically and empirically. The distinction lies in two coupled design choices:

  1. Self-supervised learning inside operator-defined observation spaces. Models that use spectra or physiological statistics (BIOT, LUNA, LaBraM, DeeperBrain) add them as an input or prediction target to a single representation that still has to hold every kind of structure. In OpView, the operator is neither an appended feature nor a target for a raw-waveform encoder: it defines the coordinate space in which an encoder both observes the signal and poses its self-supervised objective, with one space per structure. Frequency, Relation, and Projective remove nuisance factors (gain, reference, polarity)—encoding reference- and gain-invariant cross-spectral coupling shapes (Relation) and polarity-invariant spherical-harmonic scalp states (Projective)—while Waveform and Dynamics keep the time-domain signal and differ in their objectives.
  2. Independent pretraining for complementary structure. Multi-view methods such as CMC and TF-C train views to agree and so keep only what the views share; our views are chosen for what the other views hide, sharing no parameters or objectives and meeting only at the linear probe. Whereas CBraMod can train time and frequency branches jointly toward a shared raw-patch target, several of our operators map EEG segments to statistics whose inversion back to the raw signal is ill-posed; without a shared target, summing five separate losses would let the easiest one dominate, consistent with findings that joint training under-learns modality-specific features (Wang et al., 2020; Du et al., 2023). Unlike data-routed mixtures of experts (SpecMoE) or multi-task raw-window encoders (Kazatzidis & Mehrkanoon), OpView trains every encoder on the same data in distinct observation spaces.

Empirical substantiation of what this integration adds

  • Complementary to time–frequency (CBraMod) and statistic-target (LaBraM) models (Section 5.5, Table 27): To test whether OpView captures structure missed by closely related models, we combined our views with published encoders under the same linear probe. Adding one of our views to a published encoder improves accuracy (p<0.05p < 0.05) in 11 of 35 prespecified combinations and lowers it in none, and adding all five views to five published encoders raises performance on 8 of 9 datasets (p<0.05p < 0.05 on CHB-MIT; Table 27). On Nakanishi2015, our Frequency view lifts CBraMod, LaBraM, and LUNA from 32–73% to 95% (with the Frequency operator alone reaching 94.1%).
  • Higher-order, non-stationary coupling (Section 3.3, Table 31): On a test signal whose two-channel coupling changes within a window while power spectra stay constant, published encoders reach only 19–65% of the score of our Relation operator, whereas for a first-order band-power shift CBraMod and EEGPT reach 91–94% of a fixed operator.
  • Avoiding background-signal shortcuts (Appendix B.7): In CBraMod, cross-electrode attention in the last six blocks lowers masked-reconstruction error by 23.8% (96% of it outside the stimulation frequencies and harmonics) while making the SSVEP stimulus less readable; switching it off raises the benchmark's probe from 44.3% to 63.3% across all nine subjects on Nakanishi2015. Reconstruction loss can fall through background activity rather than task structure; input preconditioning prevents this shortcut.
  • Distinct feature spaces and learning within each space (Tables 1 and 25): With randomly initialized encoders, five distinct views are statistically better than five copies of one view on 5 of 9 datasets and worse on none (Table 25). Within these spaces, pretraining adds significant frozen-transfer gains over both operators (+2.6 points overall; +10.2 in Dynamics) and random initialization (+5.0 in Dynamics, +2.9 in Projective unaligned; +7.8, +4.4, and +4.5 in Dynamics, Projective, and Waveform under alignment in Table 1, and +8.3 to +13.8 points on TUAB and Sleep-EDFx in our Nov. 15 update).
  • Direct operator-as-target control (Section 5.4, Table 35): The most direct test of your question is a raw-waveform encoder trained to predict our operator outputs from masked input, as statistic-target models do, with the same corpus, architecture, and compute; as reported in our Nov. 12 update, it behaves as a second Waveform view and trails our Frequency view by 6.1 points on SSVEP (p=0.016p = 0.016).

We thank the reviewer again and would be glad to clarify any remaining points.

AuthorsOpenReview ↗

Update for Reviewer SCg6: the direct control, and positioning

Update for Reviewer SCg6: the direct control, and positioning

TL;DR: We report the completed direct control testing our core design choice (W1, Section 5.4, Table 35): a raw-waveform encoder trained to predict the output of our Frequency operator behaves as a second Waveform view and scores 6.1 points below our input-preconditioned Frequency view on steady-state responses (88.9% vs. 95.0%, p=0.016p = 0.016). We also highlight expanded positioning in Section 2, show that our five parameter-free operators alone beat 10 of 12 published encoders (Table 28), and release updated provenance records (PROVENANCE.md) and anonymous weights.

Thank you again for your review. We have uploaded a revised manuscript and supplementary package with improved exposition and redrawn figures (Section 5; Figures 2, 6, and 7), completed experimental controls, and anonymous pretrained weights on Hugging Face (https://huggingface.co/Anony-0x3f/opview-v2-30k). All table, section, and appendix numbers match our first reply. Below we summarize the findings most directly connected to your review.

1. W1: Direct control — placing the operator at the input vs. using the operator as a prediction target (Section 5.4; Appendix K, Table 35)

  • Motivation: Existing foundation models that incorporate spectral or physiological statistics as prediction targets (e.g., LaBraM, DeeperBrain) still feed raw waveforms into the encoder. To test whether predicting an operator's output from raw EEG can achieve the same representation as preconditioning the encoder's input space, we trained a controlled raw-waveform encoder (using the exact input, patching, masking, and backbone of our Waveform view) to predict the output of our Frequency operator under the same target normalization, 61-source corpus, and 30,000-step schedule as the Frequency view.
  • Result:
    1. It remains a waveform view: Across all benchmark datasets, the operator-target control stays within 4.2 points of the Waveform view. It excels where the Waveform view excels—scoring statistically higher than the Frequency view on the three motor-imagery and two P300 datasets (+3.2 points across all subjects, p=3×10−4p = 3 \times 10^{-4}).
    2. It fails to expose what the input-preconditioned Frequency view provides: On steady-state visual evoked responses (Nakanishi2015)—the spectral task for which the Frequency view is designed—the operator-target control scores 6.1 points below our Frequency view (88.9% vs. 95.0%, p=0.016p = 0.016) and even falls below its own random initialization (91.8%).
    3. Effect on the multi-view model: Replacing our Frequency view with the operator-target control reduces SSVEP decoding by 3.5 points across the nine subjects (p=0.16p = 0.16) and changes the pooled 9-dataset score by −0.3 points (p=0.25p = 0.25).
  • Significance: Training a raw-waveform encoder to predict an operator's output simply yields another time-domain waveform representation that OpView already possesses; placing the operator at the input is essential to construct a genuinely complementary representation space.

2. Clarified positioning and baseline strength of the five operators (Section 2, Table 28, Table 1)

  • Motivation & Result: We expanded Section 2 to contrast OpView with prior multi-view time-series methods: whereas conventional multi-view self-supervision maximizes cross-view agreement between time and frequency domains, our operators target higher-order EEG structure (sensor coupling, scalp topography, temporal dynamics) and intentionally retain what other views remove. To quantify the value of these observation spaces directly, Table 28 adds a head-to-head comparison showing that our five parameter-free operators alone, without any pretraining, are statistically higher than 10 of the 12 published encoders when pooled across the 9 datasets (and match EEGPT and ST-EEGFormer).
  • Side-by-side alignment gains in Table 1: Table 1 now displays unaligned and per-subject aligned results side by side. Alongside the unaligned gains quoted in our first reply (Dynamics +10.2 over its operator and +5.0 over random initialization; Projective +2.9), per-subject alignment yields +9.2, +7.8, and +4.4, and Waveform (+4.5) also gains significantly from pretraining.

3. Q1 and Q2: Expanded provenance and reproducibility package (PROVENANCE.md, rebuttal_scripts/)

The checkpoint-to-result mapping and dataset versions for the nine main datasets are unchanged. In the updated supplementary package, PROVENANCE.md additionally documents the operator-target control checkpoint, the 9-dataset 3-seed fine-tuning runs, and the extension benchmark splits (with all 13 datasets completed in our Nov. 15 update), while rebuttal_scripts/ provides the exact evaluation scripts for every new result.

We would be glad to clarify any remaining points.

Comments on the paper

6 comments
AuthorsOpenReview ↗

Summary of our response to all reviewers (1/2)

Summary of our response to all reviewers (1/2)

TL;DR: Across Parts 1 and 2, we evaluate a single fixed five-view OpView model (combined by validation-calibrated task fusion, R-cal, using no test labels) against published EEG foundation models under a shared ridge probe across all 131 held-out subjects of 9 benchmark datasets (extended to all 13 EEG-FM-Compass datasets and 13 encoders in our Nov. 15 update). Part 1 shows that operator-defined feature spaces provide the primary foundation, self-supervised pretraining adds significant frozen-transfer gains at 4–8× lower training compute than recent models, and OpView outperforms or matches the strongest published encoders. Part 2 presents new diagnostic evidence on why single-space encoders lose physiological structure (Section 3.3) and summarizes our unified zero-exposure evaluation protocol and released artifacts.

We sincerely thank all three reviewers for their careful and constructive feedback, which has substantially improved the paper. Reviewer rMDs highlighted our systematic comparison of how much physiological structure a linear probe can recover from fixed operators, patch embeddings, and pretrained encoders. Reviewer SCg6 highlighted the distinction between information being present in a signal and being accessible under a fixed readout budget. Reviewer i42G noted the value of separating the contributions of the operator, the architecture, and pretraining.

To carry that separation through the main benchmark, we evaluate a fixed five-view model with fusion weights fitted strictly on the training subjects of each fold (R-cal, Appendix F) and read OpView and all published encoders with the same ridge probe (leave-one-subject-out across 131 held-out subjects on 9 datasets, 3 seeds, paired Wilcoxon signed-rank tests at p<0.05p < 0.05). For each view, we compare the pretrained encoder, the same architecture at random initialization, and the parameter-free operator alone:

  1. Operator-defined feature spaces account for most of OpView's performance (Tables 1, 25, 31). To test whether exposing distinct physiological coordinates matters more than ensembling multiple networks in raw signal space, we compared distinct operator views against single-view copies and tested higher-order coupling recovery. At random initialization, five distinct views outperform five copies of one view on 5 of 9 datasets and underperform on none (Table 25). Adding one of our views to a published encoder improves accuracy in 11 of 35 prespecified combinations and lowers it in none (Section 5.5). On a test signal whose two-channel coupling varies within a window while channel power spectra stay constant, published encoders reach only 19–65% of the score of our parameter-free Relation operator (Table 31), showing that operator preconditioning directly resolves feature-space mismatch.
  2. Self-supervised pretraining adds measurable gains in frozen transfer (Tables 1, 21–23, Appendix H). Without alignment, the pretrained five-view model gains an average of +2.6 percentage points over the operators alone (p=4×10−5p = 4 \times 10^{-5}) and +1.0 point over matched random initialization (p=0.018p = 0.018), driven on the 9 BCI datasets by Dynamics (+5.0 points) and Projective (+2.9 points) (and reaching +8.3 to +13.8 points in Waveform and Dynamics on large-scale TUAB and Sleep-EDFx; see our Nov. 15 update). With label-free per-subject alignment, the five-view gain over random initialization rises to +1.3 points with calibrated weights (p=1×10−4p = 1 \times 10^{-4}) and is also significant with equal weights (Tables 1, 21, 22). A linear probe on pretrained Waveform features tracks trial-to-trial mu-band power at r=0.77r = 0.77 versus r=0.37r = 0.37 at random initialization (Appendix H), while full fine-tuning adapts both initializations comparably within the operator space (Table 23).
  3. Head-to-head comparison with published encoders under the same probe (Tables 2 and 28). Without alignment, our single fixed OpView model is statistically higher on 5 to 9 datasets against each of 11 published encoders, with no statistically lower result in those comparisons (and against ST-EEGFormer, higher on 2, indistinguishable on 5, lower on SEED). Even against the post-hoc highest-scoring baseline on each dataset, OpView is higher on 1 dataset, statistically indistinguishable on 7, and lower on 1 (and with alignment on both sides, higher on 1 and indistinguishable on 8).
  4. Computational efficiency (Table 33). Summing across all five branches, OpView has 19.3M parameters, requires 33.5 GFLOPs and 16.6 ms per window, and pretrains in 24–27 A100-equivalent GPU-hours—roughly one-eighth to one-quarter the reported pretraining costs of CodeBrain, CBraMod, and LUNA.

Each reviewer has a dedicated reply. Part 2 summarizes our new diagnostic evidence in Section 3.3 and our unified evaluation protocol.

AuthorsOpenReview ↗

Summary of our response to all reviewers (2/2)

Summary of our response to all reviewers (2/2)

1. New diagnostic evidence on why single-space pretraining loses structure (Section 3.3, Appendices B.2, B.7, B.8, G, H)

  • Motivation: To understand why existing EEG foundation models often fail to expose physiological structures that are present in the raw recording, we added three diagnostic tests examining how the input space and pretext objective shape linear accessibility.
  • Result & Significance:
    1. Higher-order non-stationary coupling vs. band power (Appendix H, Table 31): While published encoders recover 91–94% of a fixed operator's score on first-order band-power shifts, they recover only 19–65% of our Relation operator's score when two-channel phase coupling changes over time at constant spectral power.
    2. Encoder layers degrading steady-state responses (Appendices B.2 and G): Under the benchmark's own linear probe, four of five complete pretrained encoders decode steady-state visual evoked potentials (SSVEP) less accurately than their own initial patch embeddings; adding our Frequency view restores decoding to 95%.
    3. Objective competition inside a shared representation (Appendices B.7 and B.8): In CBraMod, cross-electrode attention in the last six blocks reduces masked-reconstruction error by 23.8% (96% of which comes from background activity outside the stimulation frequencies) while making the SSVEP stimulus harder to decode (Appendix B.7); furthermore, parameter gradients that improve band-power and cross-channel correlation probes directly oppose each other (cosine −0.50; see our Nov. 12 update and Appendix B.8). Together, these findings show that a single reconstruction objective on a shared representation trades off competing physiological structures, directly motivating OpView's independent operator-preconditioned views.

2. Guide to the revised manuscript, follow-up controls, and anonymous weights

Sections 2, 3.3, and 5 have been revised for clarity, and the supplementary package includes a marked diff PDF, evaluation scripts, and complete provenance manifests (PROVENANCE.md). Our November 12 and November 15 updates report the completed 3-seed fine-tuning comparisons across all 9 datasets (Table 23), the full 13-dataset evaluation across all 13 released encoders under validation-calibrated fusion (Appendix K, Tables 36–37), the gradient conflict analysis in Appendix B.8, and the direct control training a raw-waveform encoder to predict our Frequency operator output (Section 5.4, Table 35). Anonymous pretrained weights are available at OpView-30k and OpView-mask25.

3. Unified evaluation protocol and zero-exposure corpus standardization (Appendices D.1, F, J)

  • Motivation: Following the reviewers' constructive suggestions on deployment-ready view combination and corpus documentation, we unified all benchmark tables under a single standardized pretraining and evaluation protocol so that every comparison isolates representation quality cleanly:
Protocol dimensionStandardized design across the revised manuscript
Deployment-ready task fusionEvaluates a single fixed five-view model (R-cal, with weights fitted strictly on training subjects) against each published encoder under the same ridge probe (Tables 2, 28, 36–37), removing post-hoc single-view selection.
Strict component separationThe main OpView model combines only the five pretrained encoders; randomly initialized encoders and parameter-free operators are evaluated as separate standalone controls built by the same rule (Tables 1 and 18).
Objective design for dense operatorsUses full-output reconstruction for dense operator views (Frequency, Relation, Projective) alongside 25%-masked variants as a controlled comparison (Section 5.2, Tables 1 and 15).
Complete 3-seed & alignment coverageReports all 131 held-out subjects (3 seeds) with side-by-side unaligned and per-subject aligned results for both frozen transfer (Table 1; Dynamics, Projective, and Waveform gaining significantly) and 9-dataset fine-tuning (Table 23).
  • Zero-exposure pretraining corpus and duration control (Appendix D.1, Tables 14, 25, 34): We synchronized Appendix D.1 with PROVENANCE.md and standardized all checkpoints on a unified 30,000-step run trained on 61 public sources (2,321 hours) with zero exposure to any of the 13 benchmark datasets. To evaluate training duration on a strictly fixed data mixture, Table 25 compares 30k vs. 75k updates on identical corpus splits, and Table 34 documents benchmark dataset exposure across all compared published foundation models.

We would be very happy to discuss these findings further and welcome any follow-up questions or thoughts from the reviewers.

AuthorsOpenReview ↗

Update to all reviewers: revised writing and figures, and further results

Update to all reviewers: revised writing and figures, and further results

TL;DR: We have uploaded a revised manuscript with clearer exposition and redrawn figures (Figures 2, 6, 7), released anonymous pretrained weights on Hugging Face, and added four new results (Sections 3.3 and 5.4, Appendix K, Tables 23, 28, 35–37): (1) a direct control showing that a raw-waveform encoder trained to predict our Frequency operator output trails the input-preconditioned Frequency view by 6.1 points on SSVEP (p=0.016p = 0.016); (2) gradient analysis showing that two physiological probes directly oppose each other in CBraMod (cosine −0.50); (3) a comparison showing that our five parameter-free operators alone beat 10 of 12 published encoders; and (4) extended benchmark evaluations and 9-dataset 3-seed fine-tuning.

1. Direct control: operator at the input vs. operator as a prediction target (Section 5.4, Table 35)

Models that use spectral or physiological statistics as prediction targets still feed raw waveforms into a single encoder. To test whether predicting an operator's output can substitute for placing the operator at the input, we trained a raw-waveform encoder (using our Waveform patching, masking, and backbone) to predict the output of our Frequency operator under the same corpus and 30,000-step schedule. It behaves as a second Waveform view (within 4.2 points of Waveform across datasets) and fails to capture what the input-preconditioned Frequency view provides: on steady-state responses (Nakanishi2015), it is 6.1 points below the Frequency view (88.9% vs. 95.0%, p=0.016p = 0.016) and below its own random initialization (91.8%). Placing the operator at the input is thus essential to resolve feature-space mismatch. Details are in our update to Reviewer SCg6.

2. Mechanistic analysis: opposing physiological gradients in a released encoder (Section 3.3, Appendix B.8)

To explain why single-space pretrained encoders lose decodable physiological structure present in their own patch embeddings, Section 3.3 extends the gradient analysis of Appendix B.8 on CBraMod: at its released checkpoint, the parameter directions that improve a band-power linear probe and a cross-channel correlation probe directly oppose each other (cosine −0.50), and the reconstruction objective aligns with band power (+0.028) while opposing cross-channel coupling (−0.129). A single pretext objective on a shared representation forces competing structures into a trade-off, directly motivating independent operator-preconditioned views. Details are in our update to Reviewer rMDs.

3. Parameter-free operators alone vs. published encoders & extension benchmarks (Table 28; Appendix K, Tables 36–37)

  • Operators alone (Table 28): Combined by the same rule and read by the same ridge probe without any pretraining, our five parameter-free operators are statistically higher than 10 of 12 published encoders when pooled across the 9 datasets, and not different from EEGPT and ST-EEGFormer, confirming that most of the gap over existing models is established by the feature spaces.
  • Extension benchmarks (Tables 36–37): Under an initial equal-weight control on three further datasets, OpView reaches 81.0% on TUAB (statistically higher than 11 of 12 published encoders and matching EEGPT at 80.5%), achieves lower error than 8 of 12 encoders on SEED-VIG (matching the other 4), and beats 5 of 12 on Things-EEG2 (54.5% vs. 60.1% for ST-EEGFormer under equal weights, where two views are at chance; see our Nov. 15 update for calibrated R-cal results across all 13 datasets including Sleep-EDFx and EEGMamba).

4. Complete 9-dataset fine-tuning, readout width analysis, and limitations (Sections 5.3 and 6, Tables 2 and 23)

  • Fine-tuning across all 9 datasets (Table 23): With BNCI2014008, SEED, and CHB-MIT completed (3 seeds each), the Dynamics encoder fine-tuned from pretraining and from random initialization does not differ statistically without alignment (−1.8 to +0.9 points; with alignment, pretrained is higher on BNCI2015001 by +1.7, p=0.036p = 0.036, and lower on BNCI2014009 by −1.1 AUROC, p=0.002p = 0.002, and CHB-MIT by −4.3, p=7×10−5p = 7 \times 10^{-5}). Supervised end-to-end training adapts weights directly in the operator space, so pretraining specifically benefits frozen transfer to unseen subjects (Section 5.3).
  • Readout width and limitations (Sections 5.3 and 6): An 8,192-column probe compresses the Frequency encoder's token states (at least C×75×192C \times 75 \times 192 values) far more than the compact operator output (Section 5.3). Section 6 notes four limitations: per-view objective design, five-branch inference cost, role of subject alignment, and readout-level fusion.

Pretrained weights (anonymous): https://huggingface.co/Anony-0x3f/opview-v2-30k and https://huggingface.co/Anony-0x3f/opview-v2-mask25. We would be glad to clarify any remaining points.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.