Measurement-Gated Provenance Attenuation for Frozen EEG Representations

Anuar Aimoldin; Yankai Chen; Ayana Mussabayeva; Nurdaulet Akhanov; Xue Liu

cs.LGSep 26, 2026

Measurement-Gated Provenance Attenuation for Frozen EEG Representations

Authors: Anuar Aimoldin, Yankai Chen, Ayana Mussabayeva, Nurdaulet Akhanov, Xue Liu

Organizations: Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)

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 keeps the same information as any 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, 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 2.7x 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 (lowest over device-label shifts) by .057 and .019 over LEACE, at a cost to those heads while the training association holds. A reusable correction shows its value in how an existing predictor behaves once acquisition cues stop being reliable, not only in what a probe can read.

Figures & tables

Figure 1: Two routes to source-predictive frozen EEG representations. Source $U$ reaches the representation $H$ of observed EEG $X$ through the acquisition pipeline ( $A$ ) or through sampling association with biological or population factors ( $B$ ; route labels local to this diagram). MGPA edits only the gate coordinates $Z$ of $H=Q+BZ$ ( $B$ : gate basis, Section 3 ) and preserves $Q$ .

Figure 2: Three choices in measurement-gated correction. (a) Embedding space $H=Q+BZ$ . Dashed segments join the two recordings of one trial; their differences (grey fan) span the gate, and their mean $\widehat{\delta}$ is the closed-form direction; a new input $h$ moves to $h^{\prime}$ along $\widehat{\delta}$ . (b) Slice at fixed $Q=q$ ; rows are population groups. Selective correction moves points only along the device direction $\delta$ ; whole-block replacement collapses them. (c) Same slice; arrows show the movement to the zero target of erasure and to $\mu(q)=\mathbb{E}[s\mid Q=q]$ , the mean source log-odds score given $q$ . Gaussian illustration (source means $(\pm 1,0)$ , covariance $I_{2}$ , $\Pr(U=1\mid q)=.8$ ), not EEG data.

Figure 3: Effects of the gate and target in Iterative MGPA. (a–b) Source AUROC and paired frozen-task change from identity (pp: percentage points) at a common

.05

validation movement budget; dashed: identity, dotted: chance; random averages five orientations. (c) Conditional versus empirical unconditional mean-score anchoring at native 48-stage outputs. Three fits, 81 participants; whiskers and movement brackets give 95% participant-bootstrap intervals.

Interface	Method	Source $\downarrow$	Frozen $\uparrow$	Refitted $\uparrow$	Movement
Source-blind correction	Identity	0.683	0.753	0.753	0.000
	LEACE	0.660	0.656	0.665	0.120
	MGPA-CF	0.510	0.753	0.753	0.028
Source-aware alignment	CORAL $\to$ ref. 0	0.502	0.753	0.753	0.012
	CORAL $\to$ ref. 1	0.502	0.753	0.753	0.012
	FEATMAP $\to$ ref. 0	0.502	0.753	0.753	0.012

Table 1: Selective correction on controlled SSVEP. Native outputs; MGPA-CF: Closed-form MGPA; ref. 0/1: target view (scaled common-average/Oz). One reader realization over four rotations, 81 evaluation participants; paired contrasts in Appendix C.1 .

		A. Temporal N170				B. Recorded SSVEP
Interface	Method	$S\downarrow$	$T_{f}\uparrow$	$T_{r}\uparrow$	$M$	$S\downarrow$	$T_{f}\uparrow$	$T_{r}\uparrow$	$M$
Source-blind correction	Identity	0.989	0.728	0.728	0.000	0.916	0.688	0.688	0.000
	LEACE	0.980	0.728	0.727	0.234	0.918	0.689	0.689	0.102
	MGPA-CF	0.982	0.727	0.728	0.237	0.895	0.689	0.690	0.056
	MGPA-Iter	0.946	0.735	0.746	0.347	0.872	0.688	0.689	0.149
	Linear-only ablation	0.987	0.725	0.729	0.345	0.872	0.689	0.689	0.069

Table 2: Source attenuation and task utility on recorded EEG. Native outputs (movements differ) for N170 (A, 10 participants) and wet/dry SSVEP (B, 21).

S

: maximum source AUROC over the eleven-reader bank (A) or the 41-reader full-input bank (B);

T_{f}/T_{r}

: frozen/refitted task AUROC;

M

: participant-mean relative movement; MGPA-Iter: Iterative MGPA. Ref. 0/1: Neuroscan/Flex (A), dry/wet (B). N/A: FEATMAP needs paired trials, unavailable in B.

		Worst-association AUROC $\uparrow$			Control-head AUROC $\uparrow$
Interface	Method	P300	N170	MMN	P300	N170	MMN
Source-blind correction	Identity	0.355	0.122	0.446	0.627	0.622	0.640
	LEACE	0.601	0.384	0.614	0.622	0.621	0.639
	MGPA-CF	0.599	0.420	0.642	0.627	0.621	0.642
	MGPA-Iter	0.599	0.441	0.632	0.627	0.621	0.641
Source-aware alignment	CORAL	0.550	0.277	0.624	0.627	0.622	0.641

Table 3: One P300-selected correction, three frozen task heads. Eight common participants. Worst: minimum participant-mean AUROC of the shortcut-trained head over device–label association strength; Control: separately trained head with device assignment independent of the label. Paired intervals and per-fit results: Appendix C.4 .

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Study	Coordinates, calibration and gate	Participants and operating point
Controlled SSVEP	Pooled EEGPT512; full same-trial CAR/Oz contrasts; 90% directional energy, ranks $3,3,3,2$	Four role rotations; 81 distinct EVAL participants. Native correction/target; $.05$ VAL budget for permission.
Temporal N170	240 temporal features; simultaneous Neuroscan/Flex pairs; 90% energy, rank 56	$4/2/4/10$ ; native maps and matched-budget IGBP comparison.
Recorded SSVEP	Mean512 correction, $15\times 512$ readouts; six virtual FIT-local contrasts; 99% energy, rank 328	$41/20/20/21$ ; native maps.
P300-selected reuse	EEGPT/PCA256, P300 FIT reference coordinates; 400 simultaneous CAL pairs, rank 8	$4/2/4/8$ ; P300 DEV-selected map, unchanged on N170/MMN.

Appendix

Table 4: Protocol map. Participant counts are FIT/VAL/HEAD/EVAL (DEV replaces VAL in reuse); CAL denotes the calibration pairs. CAR is common-average referencing, Oz is the reference electrode, and PCA256 retains 256 principal components. Gates are fixed during correction.

Component	Setting
Iterative fits / horizon	Seeds 17, 29, 43; at most 48 stages
Source critics	Balanced logistic ( $C=1$ ); GELU MLP, hidden width 128
Conditional-mean anchor	Ridge with intercept and penalty 10 (both constructions)
Neural-critic optimizer	AdamW, batch 256, weight decay $.001$
Gradient stabilizer / damping	$.1$ / $.001$
Controlled and real-representation fits	12 critic epochs, LR $.001$ , radius $.05\sqrt{d}$

Appendix

Table 5: Shared MGPA settings. Component ablations change only the specified critic or target. LR denotes learning rate,

C

is the logistic regularization parameter, and

d

is the corrected representation’s dimension. Split, reader and assignment seeds are separate.

Input	Readers and training
Common representation	Ordinary and covariance-whitened logistic ( $C=1$ , whitening shrinkage $.01$ ); GELU MLP with width 128, AdamW, LR $.001$ , batch 256, weight decay $.001$ .
Direct full tokens	Raw logistic ( $C=1$ ), HEAD-standardized logistic ( $C=.01$ ); flat $7680\to 128\to 2$ GELU with common training; token network with LayerNorm, shared $512\to 64$ GELU, concatenation and binary output, LR $.0005$ , batch 128, weight decay $.001$ , gradient clipping 5.
Centered token residuals	Standardized logistic and the same token network with its three restarts/checkpoints.

Appendix

Table 6: Independent source-reader bank. Controlled/N170 use the 11 common readers. Recorded SSVEP uses all

11+20+10=41

readers; its common bank alone defines

S_{\rm mean}

, the maximum source AUROC readable from the pooled token mean. LR denotes learning rate and

C

the logistic regularization parameter. Arrows between network widths indicate successive layer dimensions.

Target	$S\downarrow$	$T_{f}\uparrow$	$T_{r}\uparrow$	$M$
Conditional mean	0.5105	0.7496	0.7527	0.1500
Unconditional mean (FIT)	0.5214	0.7488	0.7574	0.4063
Zero score	0.5199	0.7479	0.7562	0.4172

Appendix

Table 7: Conditional versus constant targets. Native controlled outputs for the same 81 participants and fits as Figure 3 c.

S

: maximum source AUROC over the reader bank;

T_{f}/T_{r}

: task AUROC with frozen/refitted heads;

M

: participant-mean relative movement. FIT is the adapter-fitting set. Arrows indicate the preferred direction.

A. Native source accessibility
Correction	$L\downarrow$	$S\downarrow$
Identity	0.763	0.989
MGPA-Iter	0.794	0.946
Linear-only	0.663	0.987

Appendix

Table 8: Nonlinear critics on N170. A:

L

and

S

are maximum source AUROCs over linear readers and the full reader bank, respectively, for the ten participants in Table 2 A. B:

\Delta

denotes Iterative minus linear-only, in AUROC percentage points;

T_{f}/T_{r}

denote frozen/refitted task AUROC. Brackets give paired 95% confidence intervals (CIs), and arrows indicate the preferred direction. Iterative estimates average fits; linear-only is deterministic.

VAL budget	Method	$S\downarrow$	$T_{f}\uparrow$	$T_{r}\uparrow$	$M$
.10	MGPA-Iter	0.9850	0.7274	0.7300	0.0974
.10	IGBP	0.9893	0.7292	0.7284	0.0898
.25	MGPA-Iter	0.9643	0.7292	0.7394	0.2422
.25	IGBP	0.9877	0.7210	0.7237	0.2239

Appendix

Table 9: N170: MGPA versus IGBP at matched VAL budgets. VAL/EVAL denote validation/evaluation participants.

S

is the maximum source-reader AUROC;

T_{f}/T_{r}

are frozen/refitted task AUROC, and

M

is participant-mean relative movement. Arrows indicate the preferred direction. First-crossing endpoints follow Appendix B.2 ; attained EVAL movement

M

may differ from the VAL budget. Same ten participants and shared fits.

Figure 4: Cross-task robustness of a P300-selected correction. Absolute worst-association AUROC of frozen N170/MMN shortcut-trained heads; whiskers show marginal 95% participant-bootstrap intervals. The horizontal axis reports AUROC, with larger values indicating better worst-association performance. Dashed: LEACE; dotted: chance; hollow: source-aware methods.

\dagger

: paired 95% interval for the gain over LEACE entirely above zero. Intervals condition on fitted maps and are unadjusted for multiplicity; panel ranges differ.

	P300			N170			MMN
Method	A	I	R	A	I	R	A	I	R
Identity	0.770	0.568	0.355	0.914	0.547	0.122	0.753	0.603	0.446
LEACE	0.604	0.602	0.601	0.782	0.590	0.384	0.614	0.633	0.652
MGPA-CF	0.643	0.621	0.599	0.735	0.579	0.420	0.642	0.644	0.646
MGPA-Iter	0.636	0.617	0.599	0.744	0.594	0.441	0.632	0.639	0.646
CORAL	0.662	0.607	0.550	0.818	0.564	0.277	0.639	0.632	0.624

Appendix

Table 10: Absolute reuse outcomes. A/I/R: aligned/independent/reversed AUROC of the shortcut-sensitive head. These regimes respectively preserve, remove and reverse the device–label association used to train that head; worst/control results are in Table 3 . Iterative entries average fits; other maps are evaluated once. CORAL/FEATMAP require device identity.

Task	Method	Reference	$\Delta$ worst [95% CI]	$\Delta$ control head [95% CI]
P300	MGPA-CF	Identity	$+0.244\ [+0.220,+0.268]$	$+0.0005\ [-0.0010,+0.0022]$
P300	MGPA-CF	LEACE	$-0.002\ [-0.022,+0.027]$	$+0.0050\ [+0.0036,+0.0066]$
P300	MGPA-Iter	Identity	$+0.244\ [+0.224,+0.262]$	$+0.0003\ [-0.0016,+0.0025]$
P300	MGPA-Iter	LEACE	$-0.002\ [-0.016,+0.021]$	$+0.0047\ [+0.0034,+0.0062]$
N170	MGPA-CF	Identity	$+0.298\ [+0.244,+0.340]$	$-0.0014\ [-0.0044,+0.0019]$
N170	MGPA-CF	LEACE	$+0.036\ [+0.005,+0.067]$	$+0.0002\ [-0.0023,+0.0037]$

Appendix

Table 11: Paired reuse gains.

\Delta

is the named method’s AUROC minus that of the reference method, on the original AUROC scale; positive values indicate improvement. “Worst” refers to worst-association AUROC of the shortcut-sensitive head; “control head” uses independent device assignment. Brackets give paired 95% confidence intervals (CIs), with worst AUROC recomputed in each bootstrap draw (Appendix B.4 ).

Invariant Representation Learning

Invariant representation learning seeks features stable to irrelevant changes while preserving information needed for prediction and transfer.

See all 227 papers

Jul 31, 2026

EEG-JEPA: Structured Latent Prediction for EEG Foundation Models

Electroencephalography (EEG) foundation models aim to learn reusable representations from large-scale unlabeled recordings. A common pretraining strategy is masked waveform reconstruction, but applying supervision directly to noisy EEG may encourage models to recover predictable background activity, acquisition effects, and artifacts rather than neural structure that transfers across tasks. This raises a central question: what should an EEG foundation model predict to learn transferable representations? We introduce EEG-JEPA a structured latent-prediction framework for EEG foundation modeling. Rather than reconstructing masked voltage samples, a masked context encoder and predictor infer contextual latent states produced by an exponential-moving-average target encoder that observes the complete input. EEG-JEPA organizes target design along three complementary dimensions: target content specifies what representation is predicted, target support specifies where prediction occurs over structured electrode--time regions through Neurotopology-Aware Multi-scale Electrode-Temporal Masking (N-MET), and target depth specifies at which encoder layers supervision is applied. Together, these designs shift EEG pretraining from recovering missing measurements to inferring latent states from structured electrode--time context. We evaluate EEG-JEPA through controlled objective comparisons, frozen multitask transfer, and full fine-tuning. Under the same backbone, pretraining corpus, and training duration, EEG-JEPA improves the 14-task frozen macro balanced accuracy from 40.49% to 50.42% over CBraMod-style masked waveform reconstruction. Multi-source continuation further raises this result to 52.94%, the highest average among the EEG foundation models evaluated on EEG-FM-Bench. Under protocol-matched full fine-tuning, EEG-JEPA also improves the nine-task average balanced accuracy from 68.98% to 70.65%.

Jinhao Li, Zhiyuan Ma, Xueqiao Han +8

Tsinghua Laboratory of Brain and Intelligence, Tsinghua University · School of Basic Medical Sciences, Tsinghua Medicine, Tsinghua University · School of Biomedical Engineering, Tsinghua Medicine, Tsinghua University +2

Self-Supervised Time Series Representation Learning Joint-Embedding Predictive Architecture Predictive Representation Learning EEG Foundation Models

Ask PDF

May 8, 2026

Same Brain, Different Prediction: How Preprocessing Choices Undermine EEG Decoding Reliability

Electroencephalography (EEG) is a cornerstone of brain-computer interfaces and clinical neuroscience, yet deep learning models are typically trained and evaluated under a single, unreported preprocessing pipeline. We formalize preprocessing choices as a counterfactual intervention space and show that EEG predictions are surprisingly unstable under this space: across six datasets spanning four paradigms, up to 42% of trial-level predictions flip when only the preprocessing changes, a variability that standard uncertainty methods do not explicitly quantify because they condition on a fixed preprocessing pipeline. We provide three tools to make this instability measurable, decomposable, and reducible. First, a Walsh-Hadamard decomposition of the 2^7 pipeline space reveals that sensitivity is near-additive in practice under the binary intervention design, enabling efficient step-by-step optimization. Second, we introduce Preprocessing Uncertainty (PU), a per-trial diagnostic that captures a dimension of instability complementary to model-based confidence. Third, we study Normalized Adaptive PGI (NA-PGI), a graph-structured regularizer that exploits the compositional structure of preprocessing interventions as one mitigation strategy with clear scope conditions.

Dengzhe Hou, Zihao Wu, Lingyu Jiang +3

Tohoku University · University of Georgia · Texas A&M University +1

Uncertainty Quantification EEG Decoding Walsh-Hadamard Transform

Ask PDF

Jun 4, 2026

The Identity Trap in EEG Foundation Models: A Diagnostic Audit

Objective. EEG foundation models (FMs) report strong accuracy on clinical resting-state EEG. However, high accuracy under subject-disjoint cross-validation remains ambiguous: it can reflect a genuine clinical biomarker, or subject-identity features that correlate with the label. We name this the Identity Trap and ask whether it can be diagnosed at the representation level before fine-tuning. Approach. We propose FMScope, a frozen-representation protocol packaging five diagnostics: variance decomposition, subject-axis erasure, aperiodic 1/f ablation, layer-wise label probing, and within-subject direction consistency. We apply it to three pretrained FMs (LaBraM, CBraMod, REVE) across four datasets in a 2x2 layout: subject relation of label x presence of a consensus cross-subject EEG marker. Main results. (i) The Identity Trap is universal: frozen subject-variance is 13-89x a random null in 12/12 pairs, rising in all 12 under fine-tuning (+10 to +63 pp). This dominance is a removable linear axis: erasing it improves label decoding where the label varies within subject (+6 to +12 pp in primary cells; +4 to +27 pp across external cohorts). (ii) Aperiodic 1/f is one subject carrier: removing it drops the subject probe by 9-19 pp on LaBraM and CBraMod. REVE saturates subject identity without measurable aperiodic dependence. (iii) Fine-tuning amplifies label-variance only in cells with a literature-established cross-subject marker. Significance. The Identity Trap is a physically-grounded instance of shortcut learning: the preferred cue has a measurable physiological component, and subject-disjoint splitting alone cannot rule it out. FMScope separates gains reflecting a biological marker from those reflecting subject identity.

Jun-You Lin, Ying Choon Wu, Tzyy-Ping Jung

School of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan · Swartz Center for Computational Neuroscience, University of California, San Diego, La Jolla, CA 92037, USA

Cross-Subject EEG Decoding EEG Decoding Shortcut Learning EEG Foundation Models

Ask PDF

Abstract

Invariant Representation Learning

Explore similar work

EEG-JEPA: Structured Latent Prediction for EEG Foundation Models

Same Brain, Different Prediction: How Preprocessing Choices Undermine EEG Decoding Reliability

The Identity Trap in EEG Foundation Models: A Diagnostic Audit