acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Self-Supervised Learning for Intracranial EEG: Electrode-Origin Prediction

12 views

Abstract

Intracranial EEG (iEEG) foundation models aim to learn general neural representations that transfer across subjects, tasks, and institutions. While existing models pursue this goal using objectives adapted from other modalities, we design the pretraining target around iEEG's persistent implanted electrodes. Our model learns through electrode-origin prediction: it identifies which training electrode produced each channel's waveform across changing neural activity. Its shallow encoder, comprising one main Transformer block and a lightweight local-context module, achieves state-of-the-art overall iMindBench performance with frozen features and linear readouts, exceeding the prior foundation-model leader by approximately 1.7 AUROC percentage points. Separate models exceed that leader's published mean using ten minutes of unique neural recordings (0.41% of its pretraining duration) or 71k transferred encoder parameters (0.36% of its encoder size). Matched ablations show stronger transfer from electrode-origin prediction than from the tested reconstruction objectives under the same architecture. Scaling analyses show aggregate gains with more data and larger models, but the gains vary by task. These results suggest that objective design, not scale alone, matters for iEEG foundation models. Code: https://anonymous.4open.science/r/puppet-0926/README.md.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Neuroscience, Foundation Model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_zJBXOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PUPPET, an iEEG representation model pretrained to predict which implanted electrode produced each trial. Its main contribution is electrode-origin prediction as an alternative to reconstruction-based self-supervision. The model uses a shallow Transformer encoder, which is frozen for downstream evaluation with linear readouts. Experiments on iMindBench report improved overall AUROC over prior models, with additional comparisons against matched reconstruction objectives, architecture ablations, spectral probes, and analyses of model size, pretraining duration, and label efficiency.

Critical strengths
  1. The pre-training objective is intuitive and well-aligned with the characteristics of iEEG recordings. By leveraging the fact that the same implanted electrode continuously collects signals over time, the model uses the electrode source as a training target, thus eliminating the need for manual task labels or anatomical coordinates.
  2. The model demonstrates attractive efficiency in terms of both architecture and data requirements. Strong results can be achieved by freezing the encoder and employing a simple linear classifier.
Critical weaknesses
  1. Omission of directly relevant work. The novelty is not sufficiently established. The PPi [1] already employs self-supervised pre-training by determining whether two segments of SEEG signals originate from the same channel. The paper fails to cite or compare against this work, leaving the advantages of a multi-class electrode-based objective over existing channel-discrimination objectives unproven.
  2. The scope of comparisons for the pre-training objective is limited. Controlled experiments primarily focus on reconstruction-based objectives, neglecting more closely related alternatives such as cross-time contrastive learning for individual electrodes or pairwise channel discrimination. Additionally, the supervised baselines could be strengthened by including representative models like EEG Conformer [2] or DBConformer [3], adapted appropriately for this context.
  3. The underlying mechanism for effectiveness remains unclear. Electrode identity inherently encodes information regarding the patient, the session, and the recording conditions. Current evidence is insufficient to distinguish whether the model is learning neural functional features or stable characteristics related to signal amplitude, spectral content, or noise.
  4. The scope of generalizations is limited. Evaluations are confined to language, auditory, and visual decoding tasks during movie-watching scenarios, which is insufficient to support broad generalizations across the full spectrum of iEEG tasks.
  5. The primary innovation is concentrated on the pre-training objective. Since the architecture largely follows existing designs, the paper must provide more rigorous validation to demonstrate the independent contribution of its approach relative to similar existing methods. [1] Yuan, Zhizhang, et al. "PPi: Pretraining brain signal model for patient-independent seizure detection." Advances in Neural Information Processing Systems 36 (2023): 69586-69604. [2] Y. Song, Q. Zheng, B. Liu, and X. Gao, “EEG conformer: Convolutional transformer for EEG decoding and visualization,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 31, pp. 710-719, 2023. [3] Z. Wang, H. Wang, T. Jia, X. He, S. Li, and D. Wu, “DBConformer: Dual-branch convolutional Transformer for EEG decoding,” IEEE Journal of Biomedical and Health Informatics, vol. 30, no. 5, pp. 4134–4147, 2026.
Questions for authors
  1. Stronger supervised baselines are missing. Table 1 includes HTNet, CNN, and MLP but omits representative decoders such as EEG Conformer [1] and DBConformer [2]. The existing comparisons do not establish PUPPET’s advantage over strong supervised models.

  2. Closely related prior work is omitted. PPi [3] uses same-channel versus different-channel discrimination for SEEG self-supervised pretraining. It is neither cited nor compared, leaving the novelty of electrode-origin prediction insufficiently established. The author should claim that, beyond replacing pairwise discrimination with multiclass electrode identification, what distinguishes PUPPET’s objective from PPi? What evidence shows that this change improves transfer?

  3. Comparisons with related self-supervised objectives are insufficient. The matched controls mainly test reconstruction objectives, leaving other ways of exploiting channel identity unexplored.

  4. Recording-specific shortcuts have not been ruled out. Electrode IDs are tied to patients and recording conditions. Stable amplitude, background spectra, or noise could support electrode identification without requiring the functional features proposed by the authors. The synthetic spectral probes do not resolve this concern.

  5. The evaluated tasks cover multiple features, but share a narrow recording context. This limits broader claims about general iEEG representations. Can the authors provide more evidence from other iEEG tasks?

[1] Y. Song, Q. Zheng, B. Liu, and X. Gao, “EEG conformer: Convolutional transformer for EEG decoding and visualization,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 31, pp. 710-719, 2023. [2] Z. Wang, H. Wang, T. Jia, X. He, S. Li, and D. Wu, “DBConformer: Dual-branch convolutional Transformer for EEG decoding,” IEEE Journal of Biomedical and Health Informatics, vol. 30, no. 5, pp. 4134–4147, 2026. [3] Yuan, Zhizhang, et al. "PPi: Pretraining brain signal model for patient-independent seizure detection." Advances in Neural Information Processing Systems 36 (2023): 69586-69604.

6 comments
AuthorsOpenReview ↗

Response to Reviewer zJBX [1/3]

Thank you for the careful review. It led to new related work, three objective controls and three supervised baselines (new text in teal).

  1. W1, W2-a / Q2, Q3 — PPi and other channel-identity objectives

Thank you for pointing us to PPi, which we now cite. We ran matched experiments on the channel-discrimination losses you named (full PPi, its pairwise same-channel loss on its own, and per-electrode supervised contrastive learning across time), all with our encoder, data, training budget, frozen read-out and 3 seeds, and the learning rate chosen per objective (we re-implemented PPi, as its code is not public).

1-a. W1 / W2-a: comparison with PPi and other channel-discrimination losses

W1: "... The PPi [1] already employs self-supervised pre-training by determining whether two segments of SEEG signals originate from the same channel. The paper fails to cite or compare against this work ..."

W2-a: "... Controlled experiments primarily focus on reconstruction-based objectives, neglecting ... cross-time contrastive learning for individual electrodes or pairwise channel discrimination ..."

All three channel-discrimination losses transfer well, but electrode-origin prediction is best:

ObjectiveAUROCPUPPET's gain over it, AUROC points (95% CI)p (Holm)
Electrode-origin (ours)0.661——
Supervised contrastive, per electrode across time0.656+0.53 (+0.27 to +0.78)< .001
Full PPi0.651+1.01 (+0.52 to +1.53)< .001
PPi-style pairwise0.651+1.04 (+0.55 to +1.55).001
PopT-v2 (published, fine-tuned)0.645——

PPi is now cited in the introduction and Sec. 2, and both PPi variants are in Table 2.

1-b. Q2: what distinguishes electrode-origin prediction, and why it helps

Q2: "... beyond replacing pairwise discrimination with multiclass electrode identification, what distinguishes PUPPET's objective from PPi? What evidence shows that this change improves transfer?"

How it differs: PPi uses a binary same-channel loss on segment pairs, plus reconstruction and context-swap losses, for seizure detection. Ours gives each training electrode one fixed label and scores every channel against all 650 electrodes: no pairs, no reference segment, no auxiliary loss.

Part of why it improves transfer: electrode-origin prediction keeps one persistent target per electrode (a learned vector that summarizes the electrode over all of training and is present at every update), whereas pairwise and contrastive losses, including PPi's, compare each sample only with the current batch. We crossed the two factors with two variants from metric learning (three seeds each; App. Table 7):

LossPersistent targetsRebuilt from each batchΔ (targets)
Softmax cross-entropy66.13 (ours)65.71 (batch prototypes)+0.43 (p = 0.07)
Supervised contrastive65.84 (persistent proxies)65.61 (standard SupCon)+0.23 (p < 0.001)
Δ (loss form)+0.30 (p = 0.06)+0.10 (p = 0.45)

Main Overall AUROC (×100), three seeds per cell (App. Table 7). Of the 0.52-point gap to standard supervised contrastive learning, persistent targets account for about 0.33 and the softmax form for about 0.20, and only the persistent-target effect under supervised contrastive is significant (Movshovitz-Attias et al., 2017; Zhai & Wu, 2019).

1-c. Q3: other ways of using channel identity (coarser targets)

Q3: "... leaving other ways of exploiting channel identity unexplored."

To see how specific the identity target has to be, we also made it coarser. Anatomical classes within each patient, and patient identity, all transfer significantly worse (Holm p < .05; Fig. 4a):

Target within each patient (classes)AUROC
Electrode (ours, 650)0.661
Hemisphere × lobe (33)0.658*
Lobe (30)0.657*
Hemisphere × macro region (19)0.653*
Hemisphere × front/back half (14)0.652*
Macro region (16)0.648*
Front/back half (11)0.647*
Hemisphere (6)0.639*
Patient identity (5)0.598*

Main Overall AUROC, three seeds each. * Significantly below electrode-origin prediction (paired patient bootstrap, Holm p < .05).

We also tried groups of nearby contacts, to see whether anatomy matters or only how finely the target splits each patient's electrodes: any of an electrode's k nearest contacts in the same patient counts as correct (k = 1 is ours). Transfer stays flat up to about 8 contacts and drops sharply at k = 64 (Fig. 4a):

k (contacts per target)Median distance to the farthest (mm)AUROC
1 (ours)00.661
8130.660
16180.659
32240.653
64370.604
AuthorsOpenReview ↗

Response to Reviewer zJBX [2/3]

  1. W2-b / Q1 — Stronger supervised baselines (EEG Conformer, DBConformer)

W2-b: "... the supervised baselines could be strengthened by including ... EEG Conformer [2] or DBConformer [3] ..."

Q1: "... The existing comparisons do not establish PUPPET's advantage over strong supervised models."

Thank you for the suggestion. We first reported the supervised baselines that iMindBench already provides, which include HTNet, a strong decoder designed specifically for iEEG. As you suggested, we have now added EEG Conformer and DBConformer, and also ShallowFBCSPNet, each with a proper hyperparameter search: trained from scratch on every Main unit via the benchmark harness (learning rate on a x4 grid from 6.25e-5 to 1.6e-2, 6.4e-2 for ShallowFBCSPNet; early stopping on validation AUROC; App. D.2). Frozen PUPPET beats all three by 6 to 8 points, as shown in Table 1:

DecoderAUROCPUPPET's gain over it, AUROC points (95% CI)Patients PUPPET ahead
PUPPET (frozen, linear read-out)0.661reference
DBConformer0.598+6.32 (+4.85 to +7.94)26/26
EEG Conformer0.581+8.01 (+5.98 to +9.97)25/26
ShallowFBCSPNet0.581+8.06 (+6.89 to +9.45)26/26

All significant after Holm. Frozen PUPPET with a linear read-out beats each on every dataset, and no tested learning rate brings any within 6 points. Table 1 also marks every significant difference, and Sec. 4.5 describes the setup.

  1. W3 / Q4 — Neural features vs recording fingerprints

W3: "... Current evidence is insufficient to distinguish whether the model is learning neural functional features or stable characteristics related to signal amplitude, spectral content, or noise."

Q4: "Recording-specific shortcuts have not been ruled out. ... The synthetic spectral probes do not resolve this concern."

Thank you; we agree that electrode identity could in principle be solved from stable per-channel properties, and that the synthetic probes do not settle this. We therefore tested each kind of shortcut you name directly, always on PUPPET's encoder, objective and data.

Spectral content. If the features were a per-channel spectral fingerprint, a model given that fingerprint should transfer as well. It does not: an MLP on each channel's log spectrum and RMS identifies electrodes almost as well as PUPPET but scores 60.7 vs. 66.1, and PUPPET pretrained on spectral tokens identifies electrodes better, also on new patients, yet transfers significantly worse (Fig. 4b and App. Table 24):

Input tokensElectrode-origin pretrainingiMindBench Main AUROCElectrode ID, held-outNew patients, kNN
Raw waveformyes (PUPPET, first-session model)66.170.458.6
Raw waveformno (linear)57.70.76.3
Spectral fingerprintyes64.672.166.3
Spectral fingerprintno (linear)61.147.145.0
Multi-STFTyes64.776.470.1
Multi-STFTno (linear)63.446.056.2
Log spectrum + RMSyes (MLP, no PUPPET encoder)60.768.747.9

Spectral fingerprint: each channel's log power in 40 frequency bands plus log RMS; "no" = a linear read-out of the input itself. Spectral input makes identification easier but leaves less to learn; from the raw waveform the model must build its own features, and pretraining adds +8.5 points there, against +3.5 and +1.3 on spectral input.

Amplitude and noise. A read-out of five per-channel summaries (amplitude, 1/f exponent and offset, line noise) scores 54.9 against PUPPET's 62.5 on the same units, and removing them from PUPPET's features changes it by only −0.3 points (n.s.).

Sec. 5.3 and App. H give the details.

AuthorsOpenReview ↗

Response to Reviewer zJBX [3/3]

  1. W4 / Q5 — Movie-watching tasks only

W4: "... Evaluations are confined to language, auditory, and visual decoding tasks during movie-watching scenarios ..."

Q5: "... Can the authors provide more evidence from other iEEG tasks?"

Thank you for raising this. iMindBench is indeed limited to movie watching. We therefore added four public iEEG datasets with quite different tasks: working memory, picture naming, attention switching and free recall (new Appendix J: Table 35 describes each task and Table 36 compares every model on it). Frozen PUPPET is ahead of frozen PopT-v2 and DIVER-1 on all four, and none of the models we tested, including fine-tuned PopT-v2 and supervised decoders, does significantly better on any of them. On the pretraining side, PUPPET pretrained only on AJILE12 (non-movie ECoG from another institution) matches BrainTreebank pretraining (0.659 vs. 0.661, n.s.; Table 3).

Foundation models on the four datasets, AUROC (frozen with the same linear read-out, plus PopT-v2 and BaRISTA fine-tuned end to end; learning-rate grids and the score at every learning rate in Tables 37–38; all models, including supervised decoders, in Table 36):

ModelWorking memoryPicture namingAttentionFree recall
PUPPET (ours).679.641.977.586
PopT-v2 (frozen).610*.606*.972.574
PopT-v2 (fine-tuned).566*.589*.980.542*
DIVER-1.629*.582*.941*.552*
BrainBERT.585*.585*.975.558*
BaRISTA.557*.543*.899*.538*
BaRISTA (fine-tuned).544*.536*.831*.526*

*PUPPET significantly higher (paired patient bootstrap, Holm).

5. W5 — Independent contribution

"... the paper must provide more rigorous validation to demonstrate the independent contribution of its approach relative to similar existing methods."

Thank you; since the contribution rests on the objective, it has to hold up against similar existing methods. We believe the replies above now answer this:

  • PPi and other channel-discrimination losses: reply 1 (1-a, 1-b; Table 2 and App. Table 7)
  • Coarser identity targets (anatomical classes, patient identity, groups of nearby contacts): reply 1-c (Fig. 4a)
  • Reconstruction objectives, including DIVER-1's: Table 2
  • Stronger supervised baselines: reply 2 (Table 1)
  • Recording fingerprints and nuisance properties: reply 3
  • Other iEEG tasks and datasets: reply 4 (App. J)

Thank you again. We are happy to run any further control you think would settle a remaining point.

Reviewer_mNXsOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes electrode-origin prediction as a self-supervised pretraining objective for intracranial EEG, where the model predicts which of roughly 650 training electrodes produced each waveform segment. The one-block PUPPET encoder is pretrained on 24.7 hours of BrainTreebank from five subjects, with evaluation subjects fully excluded. With frozen features and linear readouts it reaches an Overall AUROC of 0.661 on iMindBench, above the published fine-tuned PopT-v2 result of 0.645, and variants pretrained on ten minutes of data or with 71k parameters reach 0.647. Under matched architecture, corpus, and budget, this objective outperforms waveform, Fourier, and joint reconstruction.

Critical strengths

The objective is simple and well motivated by a property specific to iEEG, namely that implanted contacts persist while neural activity changes. The claim that it transfers better than reconstruction is supported by genuinely matched controls that share the encoder, corpus, and budget in Table 2, by a random weight baseline in Fig. 9, and by a common-readout check in Table 4, with bootstrap significance over patients against all ten baselines in Table 1 as reported in Appendix D.3.

Critical weaknesses

The efficiency claims in the abstract overreach relative to the evidence. The gain of the ten minute variant over the PopT-v2 mean is 0.2 points, which is smaller than its seed standard deviation of plus or minus 0.25 in Appendix D.3, and Table 19 shows that the matched joint reconstruction control also exceeds the PopT-v2 mean at ten minutes with 64.81 versus 64.67 and even at 2.5 minutes. The ten minute result therefore does not specifically support electrode-origin prediction, and this caveat is disclosed only in the appendix.

Questions for authors
  1. What is the training and validation accuracy of the electrode classification task during pretraining? To what extent can the task be solved by non-neural electrode fingerprints such as impedance, artifacts, and referencing effects rather than by neural signal, and does electrode accuracy on held-out sessions differ from held-out windows?
  2. Can you add an unmasked reconstruction control, or a masked electrode-origin variant, with matched learning rate selection to decouple the objective from masking and optimization? Given Table 19, will you revise the ten minute claim in the abstract to note that it is not objective specific?
  3. All pretraining uses 24.7 hours from five BrainTreebank subjects. Does the advantage persist when pretraining on a second corpus such as AJILE12 or on a pooled corpus, particularly on the cross-institution tasks? This seems central to the foundation model framing.
  4. The gains concentrate on the Main subset of the benchmark, while the Challenge subset shows 0.537 versus 0.526 to 0.530 for baselines, which is near chance with no significance test in Table 12. How much of the Overall lead survives on unscreened units, and is the Main advantage driven by a few tasks or a few patients?
  5. For the 71k parameter claim, the per-unit linear readout has a feature dimension comparable to or exceeding the encoder. Can you report the total transferred system parameters for PUPPET and PopT-v2 under the same accounting?
2 comments
AuthorsOpenReview ↗

Response to Reviewer mNXs [1/2]

Thank you for the concrete questions; we ran each requested control. Revised text is in teal. Scores are Overall AUROC on iMindBench Main units unless stated.

  1. W1 / Q2b — 10-minute efficiency claim overreaches

W1: "The efficiency claims in the abstract overreach ... the matched joint reconstruction control also exceeds the PopT-v2 mean at ten minutes ... this caveat is disclosed only in the appendix."

Q2: "... will you revise the ten minute claim in the abstract to note that it is not objective specific?"

You are right that the ten-minute result is not specific to our objective. At ten minutes, matched reconstruction also reaches PopT-v2's mean. What separates the two objectives is how they scale: reconstruction does not improve with more pretraining data, while electrode-origin prediction keeps improving, as the data-scaling curves in Figure 5 already show. Electrode-origin prediction is significantly ahead of matched reconstruction from 1 hour of pretraining data and of PopT-v2 from 3.1 hours, and at every larger amount (Holm over the eight data rungs; 3 seeds each, Main Overall; App. Table 29, Fig. 5):

Pretraining dataOurs − matched reconstruction (95% CI)Ours − PopT-v2 (95% CI)
10 min−0.13 (−0.72 to +0.45), tied+0.41 (−0.97 to +1.64), n.s.
30 min+0.56 (+0.08 to +1.06), n.s. after Holm+1.12 (−0.04 to +2.23), n.s.
1 h+0.84 (+0.23 to +1.50), Holm p = 0.025+1.36 (+0.14 to +2.52), n.s. after Holm
3.1 h+1.23 (+0.66 to +1.93), Holm p < 0.001+1.85 (+0.79 to +3.04), Holm p = 0.002
24.7 h (full)+1.35 (+0.83 to +1.94), Holm p < 0.001+1.87 (+0.95 to +2.88), Holm p < 0.001

Against PopT-v2's released per-unit scores, the 10-minute and 71k variants lead by only +0.41 and +0.40 (n.s. after Holm).

We toned down the abstract, introduction and contributions (the variants now "match" PopT-v2's published mean) and added the comparison with matched reconstruction to Sec. 5.1.

  1. Q1 — Electrode-classification accuracy; held-out sessions vs windows; non-neural fingerprints

"What is the training and validation accuracy of the electrode classification task ...? To what extent can the task be solved by non-neural electrode fingerprints such as impedance, artifacts, and referencing effects ...? ... held-out sessions ... held-out windows?"

To answer directly: largely. A spectral-fingerprint MLP identifies 68.7% of held-out electrodes (PUPPET 70.4%), but its features transfer far worse (60.7 vs 66.1 Main AUROC). The electrode head is discarded after pretraining, so the question is whether the transferred features come from fingerprints. We tested this on unseen windows, sessions and patients, against models built on spectral fingerprints or nuisance summaries (tables below). Referencing effects are already suppressed: all recordings use a same-shaft Laplacian reference (App. A), which removes common-reference signal.

Identification accuracy. The head identifies sites, not memorised windows: training and held-out accuracy match, and identity carries over to later sessions and to new patients (App. H; App. Table 24).

Electrode identification (650 classes, chance 0.15%)Top-1Wrong guesses naming another patient
Training windows (main model, 3 seeds)70.0%—
Held-out windows, same sessions (main model, 3 seeds)70.2%0.8%
Later sessions, same patients (model pretrained on each patient's first session only, 6.6 h; 3 seeds)43.4 ± 0.7%11.3 ± 2.1%
New patients, nearest neighbours on frozen features (main model, 3 seeds; chance 1.4%)59.3% (untrained fingerprint read-out: 45.0%)—

Fingerprints and nuisance properties. If PUPPET relied on a per-channel fingerprint, a model given that fingerprint should transfer as well. It does not. Models trained on spectral input identify electrodes about as well or better, yet transfer worse (Fig. 4b; App. Table 24). A read-out of five per-channel nuisance summaries decodes poorly, and removing those summaries from PUPPET's features changes it by only −0.3 points (n.s.; App. H).

Model or control (same objective, data and budget)Electrode ID, held-outMain AUROC
PUPPET (raw waveform; first-session model, 3 seeds)70.4%66.1
MLP on each channel's log spectrum + RMS68.7%60.7
PUPPET with log-spectrum tokens72.1%64.6
PUPPET with multi-STFT tokens76.4%64.7
Read-out of 5 nuisance summaries per channel (amplitude, 1/f exponent and offset, line noise raw/referenced)*—54.9 (PUPPET 62.5)
PUPPET after regressing out those summaries*—−0.3 points (n.s.)

* On the same 97 Main units of 10 patients.

These analyses are in Sec. 5.3 and the new appendix "What electrode-origin prediction learns" (App. H).

AuthorsOpenReview ↗

Response to Reviewer mNXs [2/2]

  1. Q2a — Unmasked reconstruction / masked electrode-origin control

"Can you add an unmasked reconstruction control, or a masked electrode-origin variant, with matched learning rate selection ...?"

Yes. In both directions (3 seeds each), electrode-origin prediction stays ahead, and masking helps only reconstruction (Table 3):

ObjectiveUnmasked50% masked
Electrode-origin prediction0.6610.659
Joint reconstruction0.6240.648

Electrode-origin prediction scores higher than reconstruction in both settings, and with the same 50% masking it leads by 1.16 points (95% CI 0.70 to 1.59). Learning rates were selected the same way for all four runs (both objectives, with and without masking), by each run's own validation loss over a five-point grid (1.25e-3 to 0.32, in steps of ×4), and all four land at 2e-2. Details are in the appendices "Masking control" and "Learning-rate sweep" (App. B).

  1. Q3 — Second pretraining corpus

"... Does the advantage persist when pretraining on a second corpus such as AJILE12 or on a pooled corpus ...?"

Yes. PUPPET pretrained only on AJILE12 performs on par with the BrainTreebank-pretrained model (Table 3):

Pretraining corpusDataset size (h)NeuroprobeBYDPippiOverall AUROC
BrainTreebank (movie watching; main model)24.70.7420.6490.5930.661
AJILE12 only (naturalistic ECoG, no task or movie; other patients, another institution)27.40.7450.6450.5860.659

The difference, −0.27 points (95% CI −0.61 to +0.08), is not significant (Holm p = 0.17), and both are above every reconstruction objective (best 0.648) and PopT-v2 (0.645) (reconstruction was not retrained on AJILE12). Pretraining on a pooled multi-corpus dataset is a larger undertaking that we leave to future work.

  1. Q4 — Main vs Challenge; driven by a few tasks or patients?

"... the Challenge subset ... is near chance with no significance test ... How much of the Overall lead survives on unscreened units, and is the Main advantage driven by a few tasks or a few patients?"

App. D.3 now adds Main/All/Challenge, per-patient and drop-one consistency checks to its patient-stratified bootstrap tests (Holm over ten comparators). PUPPET is significantly ahead of PopT-v2 on Main and All, but not on Challenge:

SubsetPUPPET's gain over PopT-v2, AUROC points (95% CI)p (Holm)Patients ahead
Main+1.87 (+0.95 to +2.88).000521/26
All (unscreened)+1.53 (+0.87 to +2.28).000520/26
Challenge+1.09 (−0.11 to +2.54).23 (n.s.)17/26

Main is not driven by a few cases: ahead in 32/44 dataset-task pairs; dropping any one patient leaves +1.64 to +2.29, any one pair +1.47 to +2.24. On Challenge, PUPPET still has the highest score of all models (0.537 vs. 0.530 for the next best, the multi-STFT MLP), but the difference is not significant: every model is close to chance on these units, which leaves little signal to tell models apart. Challenge also has fewer units than Main (238 vs. 347). We see this as a limitation of the Challenge units themselves, and claim significance only for Main and All. They are summarized in Sec. 5.1, with markers in Table 1.

  1. Q5 — Total transferred parameters, same accounting as PopT-v2

"... Can you report the total transferred system parameters for PUPPET and PopT-v2 under the same accounting?"

Fair point. In iMindBench, a decoder is fit separately for each evaluation unit (one patient's recording session on one decoding task), and our linear read-out grows with that unit's channel count. Below, both models are counted the same way: everything trained for one unit with 109 channels, read-outs included and pretraining heads excluded. Even counted this way, PUPPET trains far fewer parameters per unit than PopT-v2:

ModelTrained, 109-channel unitTotal used (encoder + read-out), mean over the 347 units
PopT-v2whole encoder, 19.6M19.57M
PUPPET (main)read-out, 614k~4.8M
PUPPET (71k)read-out, 76.7k~130k (0.66% of PopT-v2)

As you noted, the 71k variant's linear read-out (58,849 parameters on average) is close in size to its encoder (71,222). Table 1 now has a "Trainable per unit" column, Sec. 5.1 gives 130k next to 71k, and App. B.1 "Transferred Parameter Counts" has the full accounting.

Thank you again; these questions made the paper more precise.

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

This paper tried to address the problem of choosing suitable self-supervised objectives for iEEG foundation models. The authors hypothesize that predicting electrode identity across different neural activities can help learn better transferable representations. They propose PUPPET, which trains an encoder to classify the origin of each electrode waveform among 650 training electrodes. The classifier is discarded after pretraining, and the frozen encoder is evaluated on iMindBench with linear readouts. The authors report improved performance over existing iEEG FMs and several reconstruction baselines. However, I find the methodological novelty limited and the experimental analysis does not convincingly explain why this objective works.

Critical strengths

None

Critical weaknesses
  1. I don't see much novelty in the proposed solution, because basically the method is just about predicting which electrode produced a waveform, using the loss in Eq. 1. There is previous work that has already investigated channel identification in time series representation learning [1], although with a different objective. Additionally, I don't think the paper sufficiently explains why identifying one of the electrodes from only 5 subjects leads to better transferable representations. I suggest the authors at least compare with simple objectives such as subject identification or electrode-group classification. I don't understand what is really learned from this objective.

  2. The experiment appears somewhat questionable. PUPPET uses the full unmasked iEEG, while the reconstruction baselines are trained with 50% masked signals. How can we know whether the improvement comes from the proposed objective or simply from the different input conditions?

  3. I don't feel convinced about the explanation for the improvement. Section 5.3 mainly relies on synthetic spectral features and correlations across different models, which have different architectures, pre-training data and preprocessing. How can these correlations explain the gain from electrode-origin prediction? For example, Table 15 shows that PopT-v2 recovers peak height and width better than PUPPET, despite having lower AUROC. The authors should investigate what the model actually learns from electrode identification, rather than just showing some spectral information can be recovered.

  4. I find the presentation extremely careless. Even Figure 1, which is supposed to explain the whole method, has poorly aligned arrows and overlapping annotations. Figure 2 is also very cluttered with different scales, making comparison unnecessarily difficult. I acknowledge that we can use AI tools to support with drawing, but at this level, you should spend a suitable amount of time to refine the AI-generated things. I am respectfully ask the author to be more responsible in the next submission.

References

[1] Channel Normalization for Time Series Channel Identification. ICML 2025.

Questions for authors

See Critical Weaknesses

6 comments
AuthorsOpenReview ↗

Response to Reviewer DAzE [1/3]

Thank you for the careful review. You are right that the first version did not explain why the objective works, and that Figures 1 and 2 were not made with enough care. We revised both. New and changed text in the revised PDF is shown in teal. Following your comments, we added the following (same encoder, data, budget and frozen read-out throughout):

  • Comparison with the simpler objectives you suggested: subject identification and electrode-group classification (1-b).
  • Tests of what the objective learns: PUPPET with spectral inputs, an identity-trained spectral fingerprint, and removal of nuisance properties (W3).
  • A masking control in both directions (W2).
  • Redrawn Figures 1 and 2 (W4).
  1. W1 — Limited novelty; channel ID already studied; compare to subject-ID / electrode-group

1-a. W1-a: novelty and prior work on channel identification

"I don't see much novelty in the proposed solution, because basically the method is just about predicting which electrode produced a waveform ... [W1-a] There is previous work that has already investigated channel identification in time series representation learning [1], although with a different objective."

Thank you for pointing us to Channel Normalization [1], which we also cite in Sec. 2. We would like to clarify the difference: in [1], telling channels apart is part of the task model, where each channel receives its own normalization parameters so that a forecasting model can distinguish them, whereas in our work classifying the channel serves only as a pretraining loss: we discard the electrode classifier after pretraining and evaluate the objective solely by how well the frozen encoder transfers to new patients and tasks.

On novelty (W1): the objective is simple by design, but it has not been used on its own as a pretraining target before; earlier work uses channel identity only as an auxiliary or pairwise term (PPi, PopT) or within the task model, as in [1]. In matched comparisons it transfers best, ahead not only of reconstruction but also of the existing channel-discrimination losses (supervised contrastive, PPi-style pairwise and full PPi; 1-c), so it is not equivalent to those objectives; and your suggestions, together with those of Reviewers zJBX and mNXs, helped us understand why: which target granularity it needs (1-b), which property of the loss drives the gain (1-c), and what it learns from the signal (W3).

1-b. W1-b: subject identification and electrode-group classification

"[W1-b] ... I suggest the authors at least compare with simple objectives such as subject identification or electrode-group classification."

Thank you for the suggestion. We ran both, plus a third family that relaxes the target to a group of nearby contacts, always with the same encoder, data, budget and frozen linear read-out (3 seeds; only the target changes; summary in Fig. 4a).

In short, no coarser target transfers significantly better than predicting the exact electrode, and only the exact electrode needs no channel metadata. Groups of nearby contacts and anatomical classes need electrode coordinates or anatomical labels for every pretraining electrode, which many iEEG recordings lack and which are hard to pool across sites; the channel a waveform came from is known by construction.

Anatomical classes within each patient, and subject identification. Every one scores significantly below electrode-origin prediction. Subject identification transfers 6.3 points worse (0.598); finer classes come closer, up to 0.658 for hemisphere × lobe.

Anatomical class within each patient (number of classes)AUROC
Electrode (ours, 650)0.661
Hemisphere × lobe (33)0.658*
Lobe (30)0.657*
Hemisphere × macro region (19)0.653*
Macro region (16)0.648*
Hemisphere × front/back half (14)0.652*
Front/back half (11)0.647*
Hemisphere (6)0.639*
Patient identity (5)0.598*

Main Overall AUROC, three seeds each. * Significantly below electrode-origin prediction (paired patient bootstrap, Holm p < .05). Class definitions and sizes: Table 26, App. H.

Groups of nearby contacts. To see whether it is the anatomy that matters or only how finely the target splits each patient's electrodes, we let any of an electrode's k nearest contacts in the same patient count as correct, from a single contact (k = 1, ours) to about half of a patient (k = 64). Transfer stays flat up to about 8 contacts and drops sharply once a group holds about half of a patient's contacts (k = 64).

k (contacts per target)Median distance to the farthest (mm)AUROC
1 (ours)00.661
230.661
470.662
8130.660
16180.659
32240.653
64370.604
AuthorsOpenReview ↗

Response to Reviewer DAzE [2/3]

1-c. Extra losses (channel discrimination) (requested by Reviewer zJBX)

Your novelty point concerns earlier uses of channel identity, and channel discrimination is the closest earlier form (e.g., PPi). As Reviewer zJBX also requested, we therefore added three channel-discrimination losses under the same setup; none transfers as well as electrode-origin prediction:

ObjectiveMain AUROCPUPPET's gain, AUROC points (95% CI)p (Holm)
Electrode-origin prediction (ours)0.661——
Supervised contrastive, per electrode across time0.656+0.53 (+0.27 to +0.78)< .001
Full PPi (our re-implementation)0.651+1.01 (+0.52 to +1.53)< .001
PPi-style pairwise (same channel or not)0.651+1.04 (+0.55 to +1.55).001

Same encoder, data, training budget and frozen read-out, 3 seeds, learning rate chosen per objective (Table 2). Reconstruction objectives are covered under W2.

We also tested part of why it transfers better. Electrode-origin prediction keeps one persistent target per electrode (a learned vector summarizing that electrode over all of training, present at every update), whereas contrastive losses compare each sample only with the current batch. We crossed the two factors with two variants from metric learning (three seeds each; new App. Table 7):

LossPersistent targetsRebuilt from each batchΔ (targets)
Softmax cross-entropy66.13 (ours)65.71 (batch prototypes)+0.43 (p = 0.07)
Supervised contrastive65.84 (persistent proxies)65.61 (standard SupCon)+0.23 (p < 0.001)
Δ (loss form)+0.30 (p = 0.06)+0.10 (p = 0.45)

Main Overall AUROC (×100), three seeds per cell (App. Table 7). Of the 0.52-point gap between standard supervised contrastive learning and ours, persistent targets account for about 0.33 and the softmax form for about 0.20 (averaged over both orders), and only the persistent-target effect under supervised contrastive is significant, as metric learning suggests for classification versus pair losses (Movshovitz-Attias et al., 2017; Zhai & Wu, 2019).

  1. W1 / W3 — What the objective learns; Spectral probes don't explain the gain (Table 15 in the submission, now App. Table 22)

W1: "... I don't think the paper sufficiently explains why identifying one of the electrodes from only 5 subjects leads to better transferable representations. ... I don't understand what is really learned from this objective."

W3: "... The authors should investigate what the model actually learns from electrode identification ..."

Thank you, this is a fair point. Cross-model correlations cannot isolate the objective, as your Table 15 example shows (PopT-v2 recovers spectral peaks better than PUPPET, yet transfers worse). We meant the spectral analysis as a property a good encoder should at least have, not as the explanation of the gain. To show what the objective itself contributes, we rebuilt Sec. 5.3 around same-encoder tests (the spectral analysis is now in App. F):

  • Identity-trained fingerprint: an MLP on each channel's log spectrum and RMS identifies electrodes almost as well as PUPPET (68.7% vs. 70.4%) but scores only 60.7 Main AUROC, like the untrained fingerprint (61.1).
  • New patients and wrong guesses: nearest neighbours on PUPPET's features (main model) identify 59.3% of new patients' electrodes (untrained fingerprint read-out 45.0%, chance 1.4%), and wrong guesses stay within the patient (0.8% name another patient).
  • Nuisance properties and inputs: removing amplitude, 1/f and line-noise summaries changes PUPPET by −0.3 points (n.s.); spectral input tokens identify electrodes better but lower transfer by 1.5 and 1.4 points.

The input comparisons in the bullets above, side by side (Fig. 4b; App. Table 24):

Input to PUPPETElectrode ID, held-outMain AUROC
Raw waveform (ours)70.4%66.1
Log-spectrum tokens72.1%64.6
Multi-STFT tokens (PopT-v2's input)76.4%64.7
MLP on log spectrum + RMS (no PUPPET encoder)68.7%60.7

In short, electrode identity helps only when it is learned from the raw signal.

The cross-model table stays in App. F. Sec. 5.3 is rewritten around these same-encoder tests.

On your W1 question of why identifying electrodes from only five pretraining subjects transfers: the label set is 650 electrodes, not 5 subjects, and the 26 test patients never appear in pretraining. Pretraining only on AJILE12 (other patients, another institution, no movie stimulus) scores 0.659 vs. 0.661 (n.s.; Table 3 and App. E.1).

AuthorsOpenReview ↗

Response to Reviewer DAzE [3/3]

  1. W2 — Unmasked PUPPET vs 50%-masked reconstruction baselines

"PUPPET uses the full unmasked iEEG, while the reconstruction baselines are trained with 50% masked signals. How can we know whether the improvement comes from the proposed objective or simply from the different input conditions?"

This is a valid concern, also raised by Reviewer mNXs, and we have trained both objectives under both input conditions (3 seeds each; Table 3, App. B "Masking control"). With the same 50% masking, electrode-origin prediction still leads by 1.16 points (95% CI 0.70 to 1.59), and it is ahead in every setting; masking helps only reconstruction:

ObjectiveUnmasked50% masked
Electrode-origin prediction0.6610.659
Joint reconstruction0.6240.648

Masking helps reconstruction (0.624 to 0.648) but not electrode-origin prediction, which scores 0.661 unmasked and 0.659 masked, a difference of only 0.19 points that is not significant (95% CI from -0.03 to +0.45). All runs were tuned the same way: same encoder, data and training budget, with each objective's learning rate chosen by its own validation loss (a sweep over all four settings, both objectives with and without masking, puts every one at 2e-2). At equal 50% masking, ours still leads by +1.16 points (95% CI +0.70 to +1.59). Therefore, this indicates that the gain comes from the objective, not from the input condition.

  1. W4 — Clarity of Figures 1 and 2

"... Figure 1 ... has poorly aligned arrows and overlapping annotations. Figure 2 is also very cluttered with different scales ..."

Thank you for pointing this out, and we are sorry. We understand how frustrating it is to review a paper whose figures look careless, and we take the point. We redrew both figures, checking every element by hand.

For Figure 1:

  • We fixed the misaligned arrows and overlapping annotations you noted.
  • Panel (b) now labels the pretraining data of PUPPET (24.7 h) and PopT-v2 (40.5 h).
  • The MLP that falls beyond the plotted range is marked with an arrow at the right edge.

For Figure 2:

  • The different scales were intended: AUROC ranges differ widely across tasks (from about 0.53 to 0.88), and one shared scale would hide the differences between models within each task.
  • We made this explicit by printing each task's AUROC range under its name.
  • We reduced clutter by greying the task-domain arcs and labelling the rank axis.

As a small additional fix, Fig. 6 now uses the same y range within each row, with labelled ticks. The redrawn figures are Figs. 1 and 2 and Fig. 6.

We hope these changes answer your concerns. We are glad to run further controls if any point remains unclear.

Comments on the paper

2 comments
Area_Chair_G6VCOpenReview ↗

Asking for the Reviewers' follow-up

Dear Reviewers, the Authors have repsonded to your questions. I would be inteterested to see if you find your questions addressed and if that changes your valuation of the paper.

Your AC.

Dear Reviewers and Area Chair,

Thank you all for your engagement throughout the discussion, which substantially improved the paper.

Current ratings are 1 (DAzE), 4 (mNXs) and 3 (zJBX). Reviewer DAzE has stated that our responses addressed most of the concerns and will raise the score in the final assessment; we will further improve the figures and tables for the camera-ready version, as suggested. Reviewer zJBX's follow-up on per-subject consistency led to a new appendix of per-patient results (App. K), after which Reviewer zJBX noted that the consistency of improvement across subjects is now much clearer. Reviewer mNXs recommended clear acceptance.

As the discussion closes, we summarise the main concerns, by theme, and what we added:

  1. Novelty and relation to channel discrimination (DAzE W1; zJBX W1/Q2/Q3). PPi is now cited and contrasted with Channel Normalization (Sec. 2). On the same encoder, data, budget and read-out, full PPi, PPi-style pairwise and supervised contrastive objectives all transfer worse (0.651, 0.651, 0.656 vs 0.661; Holm p ≤ .001 each; Table 2). A loss-form × persistent-target experiment shows why: persistent per-electrode targets explain part of the gap (App. Table 7). Coarser identity targets (patient, anatomical classes, groups of nearby contacts) also transfer worse (Fig. 4a).

  2. Confounds in the comparison (DAzE W2; mNXs Q2/Q3). We trained each objective with and without 50% masking (three seeds): masking helps reconstruction but not electrode-origin prediction, which still leads by +1.16 at equal masking (Table 3). Pretraining on a second corpus (AJILE12), from other patients and another institution, performs on par (0.659 vs 0.661, n.s.; Table 3).

  3. What the objective learns: fingerprint or shortcut? (DAzE W3; mNXs Q1; zJBX W3/Q4). Electrode identification holds on held-out windows, later sessions and new patients. Models given spectral or multi-STFT input identify electrodes as well or better but transfer 1.4–5.5 points worse, and removing amplitude, 1/f and line-noise summaries changes PUPPET by only −0.3 (n.s.). The gain comes from learning identity from the raw waveform (Sec. 5.3; new App. H).

  4. Generality beyond movie watching (zJBX W4/Q5). Four further public iEEG datasets (working memory, picture naming, attention switching, free recall): frozen PUPPET scores highest on three, and no model is significantly better on any (new App. J).

  5. Statistics, robustness and per-subject consistency (mNXs Q4/Q5; zJBX follow-up). Main/All/Challenge tests with per-patient and drop-one checks (App. D.3); stronger supervised baselines (Table 1); per-unit parameter accounting (App. B.1); and full per-patient results for every model, every ablation and all four additional datasets (new App. K, Tables 39–44).

  6. Presentation and claims (DAzE W4; mNXs). Figures 1 and 2 redrawn; small-model claims toned down to "match" PopT-v2's published mean. Further figure and table improvements will follow in the camera-ready version.

Best regards, The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.