acceptodds
Under review as a conference paper at ICLR 2027

Measurement-Gated Provenance Attenuation for Frozen EEG Representations

Abstract

Frozen EEG representations retain acquisition signatures as well as neural activity. Source predictability alone does not identify what should be removed: it can reflect measurement effects or genuine biological and population differences, which should not be erased. We propose Measurement-Gated Provenance Attenuation (MGPA), built on one principle: measurement evidence determines where correction may act, and preserved information determines what it should aim for. Paired measurement contrasts define a gate outside which nothing changes; inside it, the source score is moved to the value the preserved coordinates already predict: for a fixed affine score, this retains the same information as any other target set by those coordinates and needs the least expected squared movement. Closed-form and critic-guided iterative constructions apply it without source identity or encoder retraining. Three studies test the principle at increasing distance from its assumptions. Under controlled reference changes, where the source–task association is known exactly, MGPA brings source to near chance with task performance unchanged, whereas erasing what predicts source (LEACE) lowers frozen-task AUROC from .753 to .656 while barely touching source; ablations attribute the attenuation to the gate's directions and less movement to the conditional target. Across recordings from different devices and electrodes, iterative correction lowers source accessibility while preserving or improving task performance. Finally, one iterative map selected on one task and reused unchanged on existing heads for two others raises their worst-association AUROC (the lowest over device–label association shifts) by .057 and .019 over LEACE, at a cost to those heads while the training association holds. A reusable correction thus shows its value in how an existing predictor behaves once acquisition cues stop being reliable, not only in what a probe can read.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
EEG foundation models, representation learning, measurement shift, provenance leakage, post-hoc adaptation, concept erasure, conditional invariance

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tu4QOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper is motivated by the idea that source-label association can be a learning shortcut in EEG adaptation, especially when the association changes. They conjecture that removing this association entirely will also remove relevant biological features useful for prediction. Hence, they propose limited artifact (device, setup, reference) removal to preserve device-population factors, which should be meaningful predictive features. They provide limited evaluation to simulate device changes, showing that MGPA detaches irrelevant device-label information and makes the classifier more robust to changes in device-label association without degrading performance.

Critical strengths
  1. Interesting idea of separating setup-relevant factors and biological factors from source-label association.

  2. Well-motivated approach to create an editable space BB to update representation HH without affecting fixed representation component QQ.

  3. The theoretical analysis part (Sec.5) to explain anchoring source-label association to the conditional mean seems rigorous, although I mark that I cannot verify this analysis as it is beyond my scope.

Critical weaknesses
  1. Concerns about the measurement gate: MGPA rests on the assumption that calibration contrasts identify directions for editing acquisition effects. However, I do not see clear evidence that content contribution is additive, so we can separate the preserved component from the editable component by current contrast calibration. The authors should provide further evidence for this, especially for complex, unseen measurement shifts.

  2. Limited external validity: MGPA addresses a real problem of source-label association; however, the current experiments remain unconvincing. Cross-task reuse (Sec. 6.4) is implemented on only same device pair (Neuroscan-Flex), same setups. This does not truly reflect adaptation across diverse source-label associations, as the paper states. The authors should evaluate the method on the same downstream task under various unseen acquisition conditions, as this is the main claim.

  3. The paper suffers from poor writing and a lack of clarity in its presentation. The paper introduces several key concepts and notations before defining them, making the arguments hard to follow (e.g., BB is introduced in line 151 and explained in line 165, making Sec. 3 hard to follow). Even the abstract and contribution sections are confusing and require effort to understand. The authors should improve organization and ensure that all definitions and notations are clearly introduced, and that all terms and jargon are well explained before use.

Minor concern: MGPA is proposed as a general framework for correcting frozen EEG representations, but its experiments rely solely on EEGPT (using a summary token). Extra evaluation with other architectures using full-token representation (such as CBraMod[1] or CSBrain[2]) would be necessary and useful.

[1] Jiquan Wang, Sha Zhao, Zhiling Luo, Yangxuan Zhou, Haiteng Jiang, Shijian Li, Tao Li, and Gang Pan. Cbramod: A criss-cross brain foundation model for EEG decoding. In International conference on learning representations, volume 2025, pp. 75310–75346, 2025.

[2] Juchen Zhou, Jiamin Wu, Zichen Ren, Zhouheng Yao, Weiheng Lu, Kunyu Peng, Qihao Zheng, Chunfeng Song, Wanli Ouyang, and Chao Gou. Csbrain: A cross-scale spatiotemporal brain foundation model for EEG decoding. Advances in Neural Information Processing Systems, 38:87150–87195, 2026.

Questions for authors

Weaknesses above.

  1. In which scenario does the additive context assumption hold?

  2. What does BB in Fig.1 mean, does it differ from BB in line 151?

  3. Editable rank rr seems sensitive across tasks and setups: 328 for recorded SSVEP, 8 for P300/N170/MMN reuse. Could the authors explain this choice in more detail? Is this based on a mechanism, or is it treated as a hyperparameter? If so, does it affect the robustness of reuse?

  4. In Eq.2, chain rule mutual information is proposed, and second-term I(U;R(H)∣Q)I(U;R(H)|Q) measure source information beyond Q that remains in the refined representation. This term seems to be minimized to exclude irrelevant device-related information. However, MGPA does not directly minimize it; instead, it updates the representation using the conditional mean. This lacks a clear link; more explanation would be useful

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

This paper introduces Measurement-Gated Provenance Attenuation (MGPA), a method designed to remove hardware and acquisition artifacts from frozen EEG representations. The primary aim is not accidentally removing biological or population differences. First, MGPA uses paired calibration measurements to restrict edits to a subspace. It then shifts source scores toward the conditional mean of the preserved coordinates. This avoids removing feature that predicts the source blindly.

Critical strengths

Unlike standard alignment baselines that require knowing which site a sample came from, MGPA can run without it at deployment. That makes it much more practical.

Critical weaknesses
  • Paper is almost impossible to digest. Somehow I find the writing very hard to follow and understand, even the abstract. The written presentation of the paper requires substantial revision.

  • Real hardware experiments rely on just 8 to 10 evaluation participants, which is too small to draw strong conclusions.

Questions for authors

NA

Reviewer_887GOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how frozen EEG representations retain acquisition signatures like device and reference configurations and shows why indiscriminately erasing source-predictive features is problematic when those features also encode task-related differences. The paper proposes MGPA which selectively edits only directions supported by paired measurement contrasts and adjusts source scores toward predictions from preserved coordinates using a reusable closed form or iterative map without requiring encoder retraining. In the controlled reference study, the authors reduce source AUROC to 0.510 while maintaining frozen-task AUROC at 0.753 whereas LEACE only slightly reduces source AUROC to 0.660 and causes task performance to drop to 0.656.

Critical strengths
  1. Controlled SSVEP results show closed-form MGPA reduces source accessibility to 0.510 while maintaining 0.753 task AUROC unlike LEACE showing the advantage of gated correction over blunt erasure.
  2. At matched movement the measurement gate beats PCA, random, and ungated edits on source attenuation, and conditional-mean targeting needs about 2.7 times less movement than an unconditional target.
  3. The reuse study shows that applying a P300-selected map for N170 and MMN tasks lifts worst-association AUROC above LEACE by 0.057 and 0.019 with control-head scores staying near identity.
Critical weaknesses
  1. On recorded N170 (Table 2), Iterative MGPA only moves max source AUROC from 0.989 to 0.946, and Closed form barely changes it (0.982). Although task scores slightly improved the practical claim of meaningful acquisition cue attenuation is weaker here than in the controlled reference study (Table 1).
  2. Several key results rely on small evaluation cohorts particularly the reuse study with only eight participants. After Iterative MGPA the N170 worst-association AUROC remains at 0.441 which is below chance indicating the map mitigates a severe failure mode without fully restoring a reliable head.
Questions for authors
  1. On recorded N170 in Table 2, why does source AUROC stay so high after MGPA (0.946 for Iterative, 0.982 for Closed form), and can the authors explain this?
  2. The reuse study involves only eight participants and the N170 worst-association AUROC remains at 0.441 (Table 3) after Iterative MGPA. Can the authors comment on how sensitive this result is to cohort size, and what they would expect with a larger evaluation set?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.