{"id":"0f3e2fb6-3901-4aa2-8f83-6092a4b06d54","arxiv_id":"2611.17031","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A frozen vision transformer (DINOv3) with a learned EEG-to-image display outperforms all seven tested EEG foundation models on eight benchmark datasets.","lead":"This paper shows that a vision foundation model (DINOv3) that has never seen EEG data can outperform specialized EEG foundation models on eight clinical and brain-computer interface tasks, once the EEG signals are rendered as images with a learned display.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Cross-fitting procedure may not fully prevent subject-specific shortcuts in within-subject benchmarks, and the paper provides no empirical test of this.","rationale":"The reader's weakest_assumption correctly identifies the cross-fitting procedure and the potential for subject-specific shortcuts. The reader notes that the paper acknowledges this but does not provide a formal guarantee or empirical test. My analysis agrees with this assessment. The reader's verdict is CONDITIONAL, and I find that this is appropriate. The paper's central claim is strong and well-supported by the ablation studies and image controls, but the lack of validation for the cross-fitting procedure is a notable weakness. The concrete test I propose would directly address this concern. I do not find any other load-bearing concern that would change the verdict. The paper is otherwise sound, and the method is novel and significant. The agreement_with_reader is 'agree' because the reader's weakest_assumption is the same as mine.","tokens_in":17814,"tokens_out":644,"duration_ms":8801,"concrete_test":"Run an ablation on a within-subject dataset (e.g., PhysioMI or SEED) where the cross-fitting procedure is replaced by a naive joint training (no cross-fitting). If the performance difference between the two is small (e.g., < 2% in balanced accuracy), then the cross-fitting is not critical and the concern is mitigated. If the difference is large, it indicates that the display is learning shortcuts that the cross-fitting suppresses, and the paper should report this and discuss the implications for generalization. Additionally, compute the correlation between the learned display's output and subject identity; if it is high, the display is exploiting subject-specific traits.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that a frozen ViT, with a learned display, outperforms all EEG foundation models. The weakest assumption is that the cross-fitting procedure (Eq. 7) prevents the display from exploiting subject-specific or trial-specific artifacts. The paper acknowledges this shortcut (b98-b99) and attempts to mitigate it by cross-fitting across subjects for between-subject tasks and across trials for within-subject tasks. However, the effectiveness of this mitigation is not empirically validated. For within-subject benchmarks (e.g., SEED, BCIC2020-3, PhysioMI), the folds split trials, not subjects. If the display learns trial-specific artifacts (e.g., noise patterns, eye blinks, or other non-neural signals that correlate with the label within a subject), the cross-fitting across trials may not prevent this, as the same subject appears in both training and validation folds. The paper does not provide a formal guarantee or an empirical test (e.g., comparing to a naive joint training, or showing that subject-specific traits are not learned). This is load-bearing because if the display exploits such shortcuts, the claim that the ViT is a better EEG encoder would be undermined, as the performance would not generalize to new subjects or new recording sessions. The paper's own limitation statement (b98-b99) confirms the concern but does not resolve it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper proposes SeeEEG, which turns a frozen vision transformer (DINOv3 ViT-B/16) into an EEG encoder by learning only the display: each channel is rendered as a dedicated row of 16-pixel patches in either a value (fixed-gain waveform) view or a scalogram view, with montage, display-filter, and contrast parameters initialized from clinical EEG-viewer settings and optimized through the frozen backbone at most 1,113 trainable parameters per view. SeeEEG is evaluated on eight datasets spanning clinical, sleep, affective, and brain-computer-interface tasks against seven frozen EEG foundation models using identical readout heads, hyperparameter grids, and validation-based selection. The paper reports that SeeEEG outperforms every EEG foundation model in balanced accuracy on all eight datasets (mean rank 1.3 in Cohen kappa/AUROC versus 2.9 for the strongest baseline REVE), that the learned displays transfer without re-optimization to V-JEPA 2.1 and a supervised ImageNet-21k ViT, and that replacing DINOv3 with a randomly initialized ViT-B costs 15.2 points on average. Ablations in Tables 3 and 4 attribute the gain to token-local rendering, spectral content within tokens, temporal order, and visual pretraining.","tokens_in":17942,"tokens_out":8760,"duration_ms":54167,"significance":"If the comparative result holds, this is a significant challenge to the premise that EEG-specific pretraining is necessary for frozen-encoder transfer: it shows that a vision backbone that never saw EEG, combined with a learned display, can match or beat seven EEG-specific foundation models under a controlled protocol. The experimental design is a genuine strength: 8 datasets, 7 baselines, all encoders frozen with identical heads, validation-based selection of display, head, and ensemble with the test set untouched, cross-fitting to prevent head-display shortcuts, transfer of the same learned displays to two other backbones, and a random-init control whose 15.2-point drop ties the gain to visual pretraining. The paper also argues fairly by noting that on TUEV, TUAB, and TUSL the EEG baselines saw the relevant TUH recordings during pretraining and still lose to SeeEEG.","major_comments":[{"comment":"The cross-fitting procedure prevents the display from exploiting subject-identity shortcuts on between-subject tasks, but for the three within-subject datasets (SEED, BCIC2020-3, PhysioMI) the folds split trials rather than subjects, so each head h−k still trains on other trials of the same subjects whose windows appear in fold Fk. The paper acknowledges the general shortcut (b98-b99) but provides no empirical measure of whether the display exploits session-, block-, or subject-level artifacts that correlate with labels within a subject (for example, recording-session drift, persistent eye-blink patterns, or impedance changes). This is load-bearing because two of the larger margins in Table 2 (BCIC2020-3 kappa +8.1, SEED kappa +6.3 over REVE) and the PhysioMI gain arise on within-subject tasks, so the 'all eight datasets' claim rests in part on this unresolved point. Because a trial-specific artifact would not generalize to held-out trials, the danger is not random trial noise but session- or block-level confounds shared across trials of the same subject. I request an empirical probe rather than a formal guarantee: for the within-subject datasets, either (i) split the cross-fitting folds by recording session or contiguous time blocks instead of random trials and report the resulting margins, or (ii) compare the cross-fitted display against a naively jointly trained display and against a display trained on permuted labels to bound the fraction of the gain that could come from trial- or session-level confounds.","section":"§3.3, Eq. (7)","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"Please report per-seed standard deviations or confidence intervals for the test scores: with three seeds, several margins over the best baseline are small (PhysioMI balanced accuracy +2.1, TUAB +0.5), and the reader cannot tell whether those particular comparisons are resolved beyond noise even though the direction is consistent across datasets.","section":"§5.1, Table 2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper repartitions validation and test subjects for HMC relative to the protocol of Xiong et al. [44]; please justify this deviation and state explicitly that all baselines were re-evaluated on the same repartitioned split so the comparison remains apples-to-apples.","section":"§4, Datasets","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The 'Best EEG FM' column should state explicitly that it is the per-dataset maximum over the seven baseline models rather than a single fixed model, since the gain of +4.6 points is computed against that per-dataset maximum.","section":"Table 3 caption","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The per-channel kappa improvements (median 15.8 to 25.8 overall, 16.0 to 36.0 for left- versus right-hand imagery) are reported without spread or sample counts; adding the distribution across channels would strengthen the claim that the learned display increases task-relevant information in the frozen tokens.","section":"§5.3, Figure 3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The per-window asymmetry index (0.45 for lateralized versus 0.07 for generalized discharges) is reported as a single number without variability; a per-window distribution or the number of recognized test windows would support the physiological localization claim.","section":"§5.5, Figure 4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The mean rank of 1.3 mixes Cohen kappa (multi-class tasks) with AUROC (binary tasks); please state in the main text which datasets contribute which metric to the ranking, as the abstract currently leaves this implicit.","section":"Abstract and §5.1","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Guoan,\n\nThis paper does something genuinely surprising: a frozen DINOv3 ViT-B/16, with no EEG pretraining, outperforms every EEG foundation model on all eight datasets. The method is clean—learn the display parameters (montage, filter, contrast) through the frozen backbone, keep the backbone frozen, and the display stays a valid page of EEG traces. The ablation in Table 3 is convincing: token-local strips already beat the best EEG foundation model on 6/8 datasets, and the learned display pushes that to 8/8. The image controls in Table 4 show that the performance depends on visual pretraining (random init drops 15.2 points) and on the token-local layout (half-row shift drops up to 29.7 points). The evidence maps in Figure 4 are a nice touch—they show the model localizes periodic discharges to the correct hemisphere, learned from window-level labels alone.\n\nThe cross-fitting concern the stress-test raises is real but the paper already acknowledges it (b98-b99) and implements a reasonable mitigation: cross-fitting across subjects for between-subject tasks, across trials for within-subject tasks. The paper does not provide an empirical test of whether this fully prevents subject-specific shortcuts, and that is a soft spot. But I think it's a minor one: the main result is robust across 8 datasets with different task structures, and the ablation in Table 3 shows that even the clinical montage (no learning) beats EEG foundation models on 6/8 datasets. The learned display adds 4.6 points on average, so the core claim does not rest entirely on the cross-fitting procedure. A more serious concern is the lack of error bars in Table 2—the paper reports means over three seeds but no standard deviations or confidence intervals. That is a standard expectation for a paper making comparative claims, and it should be addressed.\n\nWho is this for? Anyone working on EEG representation learning, foundation models, or cross-modal transfer. The paper is well-written, the experiments are thorough, and the central claim is supported. I would cite it. I would bring it to a reading group. I would send it to peer review—the authors should add error bars and ideally an empirical check of the cross-fitting assumption, but the paper is solid enough to deserve a serious referee.\n\nBest,\n[Your name]","headline":"Frozen ViT beats all EEG foundation models on 8 datasets—the evidence is strong and the method is clean. The cross-fitting concern is real but the paper acknowledges it and the main result is robust enough that I'd send this to review.","tokens_in":18541,"tokens_out":595,"duration_ms":5862},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"A vision model that has never seen EEG outperforms specialized EEG foundation models.","keywords":["EEG foundation models","vision transformers","cross-modal transfer","learned display","token-local rendering","DINOv3","SeeEEG","visual pretraining"],"falsifier":"If a carefully controlled experiment on a new EEG dataset (not seen during pretraining by any model) showed that a frozen EEG foundation model (e.g., REVE) with an optimized montage and readout head matched or exceeded the balanced accuracy of SeeEEG with DINOv3, the claim that vision pretraining is superior would be weakened. Alternatively, if SeeEEG's advantage disappeared when the task labels were perfectly confounded with subject identity despite cross-fitting, that would confirm the shortcut concern.","tokens_in":19085,"feed_emoji":"🧠","tokens_out":1558,"duration_ms":13642,"temperature":0.7,"pith_summary":"The paper argues that a frozen vision transformer (ViT), pretrained on natural images, can serve as a stronger encoder for EEG signals than any existing EEG foundation model—provided the EEG is rendered into an image that respects the ViT's patch structure. The authors introduce SeeEEG, a method that learns only how to display the EEG (montage, filtering, contrast) while keeping the vision backbone completely fixed. Across eight diverse EEG datasets, SeeEEG with DINOv3 ViT-B achieves a mean rank of 1.3 in Cohen kappa or AUROC, outperforming seven specialized EEG foundation models on every dataset. The key implication is that the visual priors from large-scale image pretraining already capture much of the structure needed for EEG analysis, reducing the need for massive EEG-specific pretraining.","feed_headline":"A vision model that has never seen EEG beats all EEG-specific models","feed_subtitle":"Frozen DINOv3 ViT with a learned display outperforms seven EEG foundation models on 8 datasets.","key_machinery":"Token-local rendering: each EEG channel occupies a dedicated row of 16×16 patches, so that every token holds one channel over 16 time samples (80 ms at 200 Hz), either as grayscale amplitude (value view) or as a scalogram with 16 logarithmically spaced frequencies from 1 to 64 Hz. This layout ensures that task-relevant patterns fall within individual tokens, leveraging the ViT's patch embedding while keeping spatial relationships across electrodes within the image. The display module (montage weights, filter gains, contrast parameters) is initialized from a clinical montage and optimized with the task loss through the frozen backbone, using at most 1,113 trainable parameters per view.","core_discovery":"A frozen DINOv3 ViT-B/16, when fed EEG signals rendered as token-local strips (one channel per row of patches, either as a waveform or scalogram), outperforms every EEG foundation model (BENDR, BIOT, LaBraM, EEGPT, CBraMod, CSBrain, REVE) in balanced accuracy on all eight datasets tested, reaching a mean rank of 1.3 in Cohen kappa or AUROC versus 2.9 for the strongest baseline REVE. The learned displays transfer to other frozen vision backbones (V-JEPA 2.1, supervised ViT) without re-optimization, and a randomly initialized ViT drops performance by 15.2 points on average, confirming that the ability comes from visual pretraining, not architecture.","pith_inferences":["The success of token-local strips suggests that the ViT's patch embedding and attention mechanism, tuned to natural image statistics, can be repurposed for time-frequency analysis when the signal is arranged to match its inductive biases—effectively treating the ViT as a universal feature extractor for locally structured signals.","The cross-fitting procedure (Section 3.3) is a practical but heuristic defense against shortcut learning; a more rigorous theoretical understanding of when this procedure guarantees out-of-distribution generalization would strengthen the claim.","The paper implicitly challenges the value of continued scaling of EEG-specific pretraining corpora, suggesting that the bottleneck may be in the input representation rather than the model capacity or data volume.","One could test the limits of this approach by applying SeeEEG to signals that lack a natural visual analogy (e.g., high-density EEG with many channels, or non-periodic signals) to see where the visual prior fails."],"forward_implications":["EEG foundation models may not need massive EEG-specific pretraining; a frozen vision backbone with a learned display can exceed their performance.","Future advances in vision foundation models (e.g., larger ViTs, better self-supervised methods) can directly benefit EEG analysis by simply replacing the backbone and refitting the lightweight display and head.","The same principle likely extends to other multichannel physiological signals that experts interpret visually, such as MEG and ECG.","The learned displays are interpretable: the evidence maps localize class-discriminative patterns to specific electrode pairs and time intervals, consistent with clinical knowledge.","Because the vision backbone stays frozen, the approach is computationally efficient for multi-task deployment—only the display and head (few thousand parameters) need to be stored per task."],"supporting_citations":[{"why":"Supplies the DINOv3 ViT-B/16 backbone that carries the visual pretraining.","marker":"[35]"},{"why":"BENDR baseline, a contrastive EEG foundation model pretrained on TUEG.","marker":"[17]"},{"why":"BIOT baseline, a biosignal transformer pretrained on SHHS+PREST.","marker":"[45]"},{"why":"LaBraM baseline, an EEG foundation model with masked patch prediction.","marker":"[15]"},{"why":"EEGPT baseline, a pretrained transformer for EEG with spatio-temporal alignment.","marker":"[40]"},{"why":"CBraMod baseline, a criss-cross attention EEG foundation model.","marker":"[41]"},{"why":"CSBrain baseline, a cross-scale spatiotemporal EEG foundation model.","marker":"[49]"},{"why":"REVE baseline, the strongest EEG foundation model in the comparison, pretrained on 60,000 hours.","marker":"[7]"},{"why":"Establishes the benchmark pipeline, preprocessing, and subject splits used in the evaluation.","marker":"[44]"},{"why":"The Platonic representation hypothesis that the paper invokes to explain why vision pretraining transfers to EEG.","marker":"[13]"}],"fun_headline_variants":[],"cache_read_input_tokens":0,"weakest_assumption_plain":"The cross-fitting procedure (Section 3.3) assumes that subject-specific traits are not correlated with the label in a way that survives splitting by subject (or by trial for within-subject tasks), and that the display cannot learn trial-specific artifacts that generalize to held-out subjects.","fun_headline_variants_meta":{"raw":{"variants":[]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000101,"raw_usage":{"cost":0.0001007552,"is_byok":false,"cost_details":{"upstream_inference_cost":0.0001007552,"upstream_inference_prompt_cost":8.32832e-05,"upstream_inference_completions_cost":1.7472e-05},"total_tokens":1216,"prompt_tokens":1120,"completion_tokens":96,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":0}},"tokens_in":1120,"tokens_out":96,"cost_source":"openrouter_usage","duration_ms":2292,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":0,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T07:14:48.665812+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"If a carefully controlled experiment on a new EEG dataset (not seen during pretraining by any model) showed that a frozen EEG foundation model (e.g., REVE) with an optimized montage and readout head matched or exceeded the balanced accuracy of SeeEEG with DINOv3, the claim that vision pretraining is superior would be weakened. Alternatively, if SeeEEG's advantage disappeared when the task labels were perfectly confounded with subject identity despite cross-fitting, that would confirm the shortcut concern.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DINOv3 ViT-B/16 backbone that carries the visual pretraining."},{"cited_title":"BENDR: Using transformers and a contrastive self- supervised learning task to learn from massive amounts of EEG data.Frontiers in Human Neuroscience, 15:653659,","cited_arxiv_id":null,"evidence_quote":"BENDR baseline, a contrastive EEG foundation model pretrained on TUEG."},{"cited_title":"Brandon Westover, and Jimeng Sun","cited_arxiv_id":null,"evidence_quote":"BIOT baseline, a biosignal transformer pretrained on SHHS+PREST."},{"cited_title":"Large brain model for learning generic representations with tremen- dous EEG data in BCI","cited_arxiv_id":null,"evidence_quote":"LaBraM baseline, an EEG foundation model with masked patch prediction."},{"cited_title":"EEGPT: Pretrained transformer for universal and reliable representation of EEG signals","cited_arxiv_id":null,"evidence_quote":"EEGPT baseline, a pretrained transformer for EEG with spatio-temporal alignment."},{"cited_title":"CBraMod: A criss-cross brain foundation model for EEG decoding","cited_arxiv_id":null,"evidence_quote":"CBraMod baseline, a criss-cross attention EEG foundation model."},{"cited_title":"CSBrain: A cross-scale spatiotempo- ral brain foundation model for EEG decoding","cited_arxiv_id":null,"evidence_quote":"CSBrain baseline, a cross-scale spatiotemporal EEG foundation model."},{"cited_title":"REVE: A foundation model for EEG – adapt- ing to any setup with large-scale pretraining on 25,000 sub- jects","cited_arxiv_id":null,"evidence_quote":"REVE baseline, the strongest EEG foundation model in the comparison, pretrained on 60,000 hours."},{"cited_title":"EEG-FM-Bench: A comprehensive benchmark for the systematic evaluation and diagnostic analyses of EEG foundation models","cited_arxiv_id":null,"evidence_quote":"Establishes the benchmark pipeline, preprocessing, and subject splits used in the evaluation."},{"cited_title":"Position: The Platonic representation hypothesis","cited_arxiv_id":null,"evidence_quote":"The Platonic representation hypothesis that the paper invokes to explain why vision pretraining transfers to EEG."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}