DEFT: Disentanglement-Enhanced Fine-Tuning for EEG Foundation Models

Yuqiu Deng; Muyun Jiang; Zongpeng Zhang; Mingqing Xiao; Yansen Wang; Peiliang Gong; Xinliang Zhou; Yi Ding

cs.LGOct 10, 2026

DEFT: Disentanglement-Enhanced Fine-Tuning for EEG Foundation Models

Authors: Yuqiu Deng, Muyun Jiang, Zongpeng Zhang, Mingqing Xiao, Yansen Wang, Peiliang Gong, Xinliang Zhou, Yi Ding,

Organizations: Uploading Inc · College of Computing and Data Science, Nanyang Technological University · Department of Biostatistics, School of Public Health, Peking University · Microsoft Research Asia ·

Abstract

EEG foundation models (EFMs) have improved EEG representation learning by transferring knowledge from large-scale multi-subject pretraining to downstream tasks. However, their pretrained embeddings often retain substantial subject-related variation, placing an additional burden on the downstream classifier, especially in cross-subject settings. To address this issue, we propose Disentangled EFM Fine-Tuning (DEFT), a plug-and-play adapter for EFM fine-tuning. DEFT encourages branch disentanglement of pretrained tokens into task-oriented content and subject-consistent style, then performs downstream prediction from the content branch. It instantiates this factorize-before-classify design with recursive Gaussian refinement, temporal-variance regularization, subject-aware contrastive learning, and reconstruction. Experiments across multiple EEG datasets and four EFM backbones show that DEFT improves matched fine-tuning baselines. Further ablations and analyses support the intended content/style branch organization.

Figures & tables

Figure 1: Pretrained EFM embeddings amplify subject information. Frozen CBraMod embeddings improve linear subject probing over raw EEG and form subject-dependent clusters. Task probes use subject-disjoint splits; subject probes use held-out trials from known subjects.

Figure 2: Overview: motivation and empirical effect of DEFT . Left : standard EFM fine-tuning leaves the classifier to suppress subject-related variation inside mixed pretrained tokens, while DEFT models subject-consistent information before classification. Right : an illustrative Balanced Accuracy snapshot across representative datasets and foundation models, showing that DEFT improves over vanilla fine-tuning in most displayed comparisons.

Figure 3: Overall framework of DEFT . Mixed pretrained tokens are organized into task-content and subject-style branches, then optimized with temporal-variance regularization, classification, subject-aware contrastive learning, and reconstruction. The downstream classifier uses the content branch for prediction. Detailed processes of each stage are shown above the overall framework.

Model	Setting	BCIC-IV-2a			FACED			ISRUC-S1
		B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	Kappa	F1
CodeBrain	Vanilla	0.3678 $\pm$ 0.0220	0.1570 $\pm$ 0.0293	0.2980 $\pm$ 0.0369	0.5720 $\pm$ 0.0047	0.5156 $\pm$ 0.0055	0.5719 $\pm$ 0.0048	0.7803 $\pm$ 0.0071	0.7366 $\pm$ 0.0084	0.7943 $\pm$ 0.0059
	DEFT	0.4621 $\pm$ 0.0040	0.2828 $\pm$ 0.0054	0.4318 $\pm$ 0.0061	0.5870 $\pm$ 0.0041	0.5297 $\pm$ 0.0041	0.5799 $\pm$ 0.0031	0.7829 $\pm$ 0.0042	0.7391 $\pm$ 0.0069	0.7981 $\pm$ 0.0019
	Delta	+0.0943	+0.1258	+0.1339	+0.0150	+0.0141	+0.0080	+0.0026	+0.0025	+0.0039
CBraMod	Vanilla	0.4630 $\pm$ 0.0327	0.2840 $\pm$ 0.0437	0.4332 $\pm$ 0.0519	0.5387 $\pm$ 0.0108	0.4770 $\pm$ 0.0118	0.5387 $\pm$ 0.0100	0.7767 $\pm$ 0.0094	0.7514 $\pm$ 0.0109	0.8052 $\pm$ 0.0086
	DEFT	0.5579 $\pm$ 0.0310	0.4105 $\pm$ 0.0414	0.5522 $\pm$ 0.0341	0.6005 $\pm$ 0.0101	0.5467 $\pm$ 0.0122	0.6002 $\pm$ 0.0130	0.7818 $\pm$ 0.0080	0.7285 $\pm$ 0.0108	0.7906 $\pm$ 0.0110

Table 1: Main results on six representative downstream datasets. All metrics report multi-seed mean

\pm

std over available seeds. Bold marks the best value in each metric column, and Delta is the absolute improvement of DEFT over vanilla fine-tuning. Chance B-ACC is 0.250, 0.111, 0.200, 0.500, 0.333, and 0.500 for BCIC-IV-2a, FACED, ISRUC-S1, MentalArith., BCIC-Speech, and Mumtaz2016, respectively.

Dataset	Chance	Full	w/o StyleCL	w/o Invar.	w/o VarRatio	w/o Gaussian	w/o Rec.
BCIC-IV-2a	0.250	0.5579 $\pm$ 0.0310	0.4638 $\pm$ 0.0170	0.3973 $\pm$ 0.0664	0.3808 $\pm$ 0.0071	0.3782 $\pm$ 0.0198	0.4615 $\pm$ 0.0477
FACED	0.111	0.6005 $\pm$ 0.0101	0.5931 $\pm$ 0.0072	0.5992 $\pm$ 0.0011	0.5976 $\pm$ 0.0027	0.5998 $\pm$ 0.0087	0.6042 $\pm$ 0.0066
ISRUC-S1	0.200	0.7818 $\pm$ 0.0080	0.7897 $\pm$ 0.0012	0.7849 $\pm$ 0.0019	0.7838 $\pm$ 0.0121	0.7856 $\pm$ 0.0057	0.7782 $\pm$ 0.0090
MentalArith.	0.500	0.7292 $\pm$ 0.0059	0.5830 $\pm$ 0.0367	0.6549 $\pm$ 0.0478	0.5687 $\pm$ 0.0527	0.5958 $\pm$ 0.0321	0.6135 $\pm$ 0.0295
BCIC-Speech	0.333	0.6894 $\pm$ 0.0172	0.4908 $\pm$ 0.0034	0.4858 $\pm$ 0.0082	0.4870 $\pm$ 0.0109	0.4972 $\pm$ 0.0076	0.4776 $\pm$ 0.0120
Mumtaz2016	0.500	0.9473 $\pm$ 0.0142	0.8286 $\pm$ 0.0678	0.9049 $\pm$ 0.0475	0.7581 $\pm$ 0.1772	0.9126 $\pm$ 0.0114	0.8760 $\pm$ 0.0329

Table 2: Multi-seed CBraMod B-ACC ablations on the six representative datasets. Ablations use fixed seeds 0–2; Full is copied from Table 1 .

Figure 4: Representation geometry before and after disentanglement. Top: subject-controlled within-subject class separation after subject centering, where higher values indicate clearer class structure within each subject. Bottom: class t-SNE visualizations of the same before-module and content-branch representations.

Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

Scope	Training fits	GPU / fit	GPU-hours
Main Table 1 multi-seed reproduction	120–150	1 RTX 4090 GPU	90–160
Complete seed-0 suite	96	1 RTX 4090 GPU	55–130
Content/style probing diagnostic	16	1 RTX 4090 GPU	1–3
Parameter-matched adapter diagnostic	3	1 RTX 4090 GPU	2–5
GRL subject-confusion diagnostic	15–30	1 RTX 4090 GPU	3–8
Compute and memory overhead benchmark	2 one-epoch runs	1 RTX 4090 GPU	$<0.1$

Appendix

Table 3: Compute accounting for reproducing the reported experiments. All values are approximate.

Dataset	Source	License / usage terms
BCIC-IV-2a	BNCI Horizon 2020 dataset 001-2014, BCI Competition IV dataset 2a; https://bnci-horizon-2020.eu/database/data-sets	CC BY-ND 4.0.
SEED	SJTU BCMI Lab SEED collection; https://bcmi.sjtu.edu.cn/home/seed/	Application/signature required; academic research only; no commercial use.
SEED-IV	SJTU BCMI Lab SEED-IV collection; https://bcmi.sjtu.edu.cn/home/seed/	Application/signature required; academic research only; no commercial use.
SEED-V	SJTU BCMI Lab SEED-V collection; https://bcmi.sjtu.edu.cn/home/seed/	Application/signature required; academic research only; no commercial use.
FACED	FACED dataset, Synapse project syn50614194; https://www.nature.com/articles/s41597-023-02650-w	CC-BY metadata; the Scientific Data article is CC BY 4.0, with download terms governed by the Synapse page.
MI-KoreaU	Lee2019/OpenBMI, GigaDB DOI 10.5524/100542; https://moabb.neurotechx.com/docs/generated/moabb.datasets.Lee2019_MI.html	GPL 3.0 according to MOABB metadata.

Appendix

Table 4: Dataset sources and usage terms for the complete evaluation suite.

Dataset	Local shape	# Cls.	Task type	Statistic	Train	Val.	Test
BCIC-IV-2a	$(22,4,200)$	4	Motor imagery (multi.)	Subjects	5	2	2
				Samples	2784	1152	1152
SEED	$(60,4,200)$	2	Emotion (binary)	Subjects	10	2	3
				Samples	2700	540	810
SEED-IV	$(60,4,200)$	4	Emotion (multi.)	Subjects	10	2	3
				Samples	25050	5010	7515

Appendix

Table 5: Dataset shapes, class counts, task types, and split statistics used in the reported experiments. Each dataset is shown with subject counts and sample counts in separate rows.

Hyperparameter	Value / policy
Optimizer	AdamW
Random seed	0–2 for Table 1 . Complete results are reported under seed 0.
Weight decay	0.05
Dropout	0.1
Label smoothing	0.1
Task loss weight	1.0

Appendix

Table 6: General hyperparameter settings used for disentangled fine-tuning.

Backbone	Learning rate	Batch size	Epochs	$\boldsymbol{\lambda_{\mathrm{ord}}}$	$\boldsymbol{\lambda_{\mathrm{sty}}}$	$\boldsymbol{\lambda_{\mathrm{con}}}$	$\boldsymbol{\lambda_{\mathrm{rec}}}$
CodeBrain	$1\times 10^{-4}$	64	20	0.005	0.001	0.02	0.01
CBraMod	$1\times 10^{-4}$	64	20	0.1	0.01–0.02	0.2	0.02
LaBraM	$1\times 10^{-4}$	64	20	0.1	0.02	0.2	0.02
CSBrain	$1\times 10^{-4}$	64	20	0.1	0.01	0.01	0.02

Appendix

Table 7: Representative backbone-specific optimization and loss-weight settings. The four

\lambda

columns correspond directly to Eq. 14 .

Dataset	Original model	Added adapter	Overhead
BCIC-IV-2a	19.13M	2.07M	10.81%
FACED	133.29M	2.07M	1.55%
ISRUC-S1	26.47M	2.07M	7.81%
MentalArith.	25.09M	2.07M	8.24%
BCIC-Speech	26.61M	2.07M	7.77%
Mumtaz2016	24.09M	2.07M	8.58%

Appendix

Table 8: Parameter overhead of DEFT on the six representative CBraMod settings. The original model is the matched vanilla fine-tuning model, including the pretrained backbone and downstream prediction head.

Setting	Trainable params	Peak allocated memory	Peak reserved memory	One-step train FLOPs
Vanilla	19.13M	879.7 MiB	1216.0 MiB	85.54G
DEFT	21.44M	970.1 MiB	1300.0 MiB	93.86G
Overhead	+2.32M (+12.11%)	+90.4 MiB (+10.28%)	+84.0 MiB (+6.91%)	+8.33G (+9.74%)

Appendix

Table 9: Compute and memory overhead diagnostic for a representative CBraMod setting. FLOPs are measured for one training step.

Setting	BCIC-IV-2a			SEED			SEED-IV
	B-ACC	Kappa	F1	B-ACC	ROC-AUC	PR-AUC	B-ACC	Kappa	F1
Vanilla	0.3602	0.1470	0.2815	0.6827	0.7499	0.7559	0.3403	0.1151	0.2982
DEFT	0.4740	0.2986	0.4523	0.6914	0.7750	0.8068	0.3515	0.1277	0.3025
$\Delta$	+0.1138	+0.1516	+0.1708	+0.0087	+0.0251	+0.0508	+0.0112	+0.0126	+0.0043
Setting	SEED-V			FACED			MI-KoreaU
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	ROC-AUC	PR-AUC

Appendix

Table 10: Complete main results for CodeBrain across the evaluation suite and recorded metrics. Positive deltas are shaded orange, negative deltas are shaded blue, and the better value between Vanilla and DEFT is bold.

Setting	BCIC-IV-2a			SEED			SEED-IV
	B-ACC	Kappa	F1	B-ACC	ROC-AUC	PR-AUC	B-ACC	Kappa	F1
Vanilla	0.4870	0.3160	0.4650	0.7383	0.8346	0.8489	0.3773	0.1601	0.3301
DEFT	0.5981	0.4641	0.5954	0.7877	0.8589	0.8647	0.3989	0.1897	0.3444
$\Delta$	+0.1111	+0.1481	+0.1304	+0.0494	+0.0243	+0.0158	+0.0216	+0.0296	+0.0143
Setting	SEED-V			FACED			MI-KoreaU
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	ROC-AUC	PR-AUC

Appendix

Table 11: Complete main results for CBraMod across the evaluation suite and recorded metrics. Positive deltas are shaded orange, negative deltas are shaded blue, and the better value between Vanilla and DEFT is bold.

Setting	BCIC-IV-2a			SEED			SEED-IV
	B-ACC	Kappa	F1	B-ACC	ROC-AUC	PR-AUC	B-ACC	Kappa	F1
Vanilla	0.5312	0.3750	0.5226	0.7630	0.8261	0.8033	0.4078	0.2008	0.3540
DEFT	0.5703	0.4271	0.5554	0.7667	0.8288	0.8154	0.4132	0.2086	0.3584
$\Delta$	+0.0391	+0.0521	+0.0328	+0.0037	+0.0027	+0.0121	+0.0054	+0.0078	+0.0044
Setting	SEED-V			FACED			MI-KoreaU
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	ROC-AUC	PR-AUC

Appendix

Table 12: Complete main results for LaBraM across the evaluation suite and recorded metrics. Positive deltas are shaded orange, negative deltas are shaded blue, and the better value between Vanilla and DEFT is bold.

Setting	BCIC-IV-2a			SEED			SEED-IV
	B-ACC	Kappa	F1	B-ACC	ROC-AUC	PR-AUC	B-ACC	Kappa	F1
Vanilla	0.4792	0.3056	0.4545	0.7160	0.7920	0.8223	0.3341	0.1009	0.2590
DEFT	0.5000	0.3333	0.4785	0.7370	0.8106	0.8109	0.3357	0.1090	0.2840
$\Delta$	+0.0208	+0.0278	+0.0240	+0.0210	+0.0185	-0.0115	+0.0016	+0.0080	+0.0250
Setting	SEED-V			FACED			MI-KoreaU
	B-ACC	Kappa	F1	B-ACC	Kappa	F1	B-ACC	ROC-AUC	PR-AUC

Appendix

Table 13: Complete main results for CSBrain across the evaluation suite and recorded metrics. Positive deltas are shaded orange, negative deltas are shaded blue, and the better value between Vanilla and DEFT is bold.

Figure 5: Normalized training loss curves on CBraMod. Each curve is divided by its first-epoch loss. The auxiliary constraints in DEFT provide additional optimization signals and lead to faster normalized loss reduction on the shown datasets.

Backbone	Dataset	Task chance	Content task	Style task	Content subject	Style subject
CBraMod	FACED	0.1111	0.2615	0.2458	0.9206	0.9451
CBraMod	BCIC-IV-2a	0.2500	0.3409	0.3183	0.8222	0.8252
CBraMod	MentalArith.	0.5000	0.6705	0.6503	0.1734	0.1770
CBraMod	BCIC-Speech	0.3333	0.3958	0.3849	0.3364	0.3377
CodeBrain	FACED	0.1111	0.2684	0.2595	0.9705	0.9716
CBraMod	Mumtaz2016	0.5000	0.9759	0.9298	0.0633	0.1999

Appendix

Table 14: Frozen linear probes on content and style. All values are B-ACC. Task chance is determined by the dataset class count.

Backbone	Dataset	Generic Adapter	DEFT
CBraMod	BCIC-IV-2a	0.4806 $\pm$ 0.0403	0.5579 $\pm$ 0.0310
CBraMod	MentalArith.	0.6597 $\pm$ 0.0371	0.7292 $\pm$ 0.0059
LaBraM	BCIC-IV-2a	0.5483 $\pm$ 0.0067	0.5671 $\pm$ 0.0032
LaBraM	MentalArith.	0.6389 $\pm$ 0.0620	0.7236 $\pm$ 0.0149

Appendix

Table 15: Stable multi-seed capacity controls. Values are test B-ACC mean

\pm

std.

Backbone	GRL	FHVAE-token	DSAE-token	DEFT
CBraMod	0.5573 $\pm$ 0.0028	0.5883 $\pm$ 0.0052	0.5848 $\pm$ 0.0019	0.6005 $\pm$ 0.0101
LaBraM	0.5261 $\pm$ 0.0151	0.4709 $\pm$ 0.0070	0.4828 $\pm$ 0.0031	0.5409 $\pm$ 0.0086

Appendix

Table 16: Multi-seed subject-aware and sequential baselines on FACED. Values are test B-ACC mean

\pm

std.

Backbone	MLP	Attention	2 steps	4-step DEFT	6 steps
CBraMod	0.5794 $\pm$ 0.0045	0.5769 $\pm$ 0.0083	0.5957 $\pm$ 0.0008	0.6005 $\pm$ 0.0101	0.5932 $\pm$ 0.0003
LaBraM	0.5081 $\pm$ 0.0098	0.5068 $\pm$ 0.0085	0.5122 $\pm$ 0.0060	0.5409 $\pm$ 0.0086	0.5208 $\pm$ 0.0058

Appendix

Table 17: Same-loss architecture and recursion-depth controls on FACED. Values are test B-ACC mean

\pm

std.

Setting	Default	w/o Invar.	w/o VarRatio	Weight 0.5 $\times$	Weight 2 $\times$
Test B-ACC	0.7922 $\pm$ 0.0015	0.7849 $\pm$ 0.0019	0.7838 $\pm$ 0.0121	0.7809 $\pm$ 0.0133	0.7815 $\pm$ 0.0035

Appendix

Table 18: Temporal-prior sensitivity on ISRUC-S1 under a fixed multi-seed diagnostic protocol.

Dataset	Vanilla	Variance only	Variance + contrast	Variance + rec.	Full
FACED	0.5387	0.5961	0.6040	0.5805	0.6126
MentalArith.	0.6521	0.5740	–	–	0.7260

Appendix

Table 19: Objective-combination controls on CBraMod. Values are test B-ACC under the stated diagnostic protocol.

Dataset	Full	w/o Invar.	w/o Gaussian	w/o Rec.
FACED	0.5480	0.5425	0.4936	0.5157
MentalArith.	0.7417	0.6729	0.6365	0.6635

Appendix

Table 20: LaBraM component controls. Values are test B-ACC.

Dataset	REVE Vanilla	REVE + DEFT	$\boldsymbol{\Delta}$
MentalArith.	0.6531	0.7083	+0.0552
BCIC-Speech	0.6215	0.6393	+0.0178
BCIC-IV-2a	0.5712	0.5773	+0.0061

Appendix

Table 21: REVE compatibility checks. Values are test B-ACC.

Setting	Method	B-ACC	ROC-AUC	PR-AUC / added params
OpenNeuro REST	Vanilla	0.4968 $\pm$ 0.0028	0.4150 $\pm$ 0.0105	0.2712 $\pm$ 0.0035
OpenNeuro REST	DEFT	0.5439 $\pm$ 0.0278	0.6623 $\pm$ 0.0137	0.4276 $\pm$ 0.0227
MOMENT/UCI HAR	Vanilla	0.5665 $\pm$ 0.0078	–	0
MOMENT/UCI HAR	Generic MLP	0.7027 $\pm$ 0.0134	–	2,965,836
MOMENT/UCI HAR	DEFT	0.7742 $\pm$ 0.0249	–	2,966,051

Appendix

Table 22: Additional subject-disjoint checks. Values are mean

\pm

std.

Dataset	Raw EEG	CBraMod	CodeBrain	CSBrain	LaBraM
FACED	0.2297	0.6381	0.7580	0.5611	0.8598
BCIC-Speech	0.2434	0.3450	0.2666	0.3464	0.2874

Appendix

Table 23: Frozen subject-probe B-ACC across backbones.

Dataset	Random anchors	DEFT sampler	Positives/anchor
BCIC-IV-2a	99.85%	100%	6.9701
ISRUC-S1	7.78%	100%	1.0736
Mumtaz2016	53.17%	100%	1.7368
FACED	16.55%	100%	1.1593
MentalArith.	74.56%	100%	2.4592
BCIC-Speech	99.92%	100%	7.8036

Appendix

Table 24: Effective contrastive-anchor coverage and mean positives per anchor.

Dataset	Task chance	1-s task	Full task	Subject chance	1-s subject	Full subject
MentalArith.	0.5000	0.5585	0.6316	0.0280	0.2104	0.2570
BCIC-IV-2a	0.2500	0.3073	0.4630	0.1111	0.6406	0.6550

Appendix

Table 25: One-second versus full-sample task and subject probes. All values are B-ACC.

Setting	Metric	$\boldsymbol{\Delta}$	Welch 95% CI
CBraMod–ISRUC-S1	Kappa	-0.0229	$[-0.0475,0.0017]$
CBraMod–ISRUC-S1	F1	-0.0145	$[-0.0374,0.0084]$
CodeBrain–Mumtaz2016	ROC-AUC	-0.0074	$[-0.0368,0.0220]$
CodeBrain–Mumtaz2016	PR-AUC	-0.0045	$[-0.0236,0.0146]$
LaBraM–FACED	Kappa	+0.0093	$[-0.0114,0.0300]$
LaBraM–FACED	F1	+0.0112	$[-0.0088,0.0312]$

Appendix

Table 26: Confidence intervals for the reviewer-identified secondary-metric differences.

Dataset	Metric	Full	w/o StyleCL	$\Delta$	w/o Invar	$\Delta$	w/o VarRatio	$\Delta$
BCIC-IV-2a	B-ACC	0.5981	0.4844	-0.1137	0.3403	-0.2578	0.3715	-0.2266
	Kappa	0.4641	0.3125	-0.1516	0.1204	-0.3438	0.1620	-0.3021
	F1	0.5954	0.4673	-0.1281	0.2757	-0.3197	0.3043	-0.2910
BCIC-Speech	B-ACC	0.5116	0.4880	-0.0236	0.4911	-0.0204	0.4778	-0.0338
	Kappa	0.2673	0.2320	-0.0353	0.2367	-0.0307	0.2167	-0.0507
	F1	0.5124	0.4873	-0.0251	0.4910	-0.0214	0.4758	-0.0366

Appendix

Table 27: Complete seed-0 CBraMod ablations, Panel A: style contrast and variance constraints. Orange/blue cells mark positive/negative deltas; bold marks the best value.

Dataset	Metric	Full	w/o Gaussian	$\Delta$	w/o Rec.	$\Delta$
BCIC-IV-2a	B-ACC	0.5981	0.3967	-0.2014	0.5122	-0.0859
	Kappa	0.4641	0.1956	-0.2685	0.3495	-0.1146
	F1	0.5954	0.3560	-0.2394	0.5096	-0.0858
BCIC-Speech	B-ACC	0.5116	0.4920	-0.0196	0.4773	-0.0342
	Kappa	0.2673	0.2380	-0.0293	0.2160	-0.0513
	F1	0.5124	0.4917	-0.0207	0.4757	-0.0367

Appendix

Table 33

EEG Foundation Models

Electroencephalography (EEG) foundation models learn reusable representations of brain signals for transfer across people, tasks, and recording setups.

See all 96 papers

Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes
Apr 25, 2026Kuntal Kokate, Bruno Aristimunha, Dung Truong +1Fine-Tuning EEG Decoding
Swartz Center for Computational Neuroscience, Institute for Neural Computation, University of California San Diego, La Jolla, CA 92093, USA · Yneuro, Paris, France · Centre de Recherche Cerveau et Cognition, CNRS, Université Paul Sabatier, Toulouse, France
B[FM] $^2$ : Brain Foundation Model via Flow Matching with SplitUNet
Jun 18, 2026Jaedong Hwang, Kathleen Zhang, Wei Dai +5Flow Matching EEG Signal Generation
Massachusetts Institute of Technology · KU Leuven
CoCoT-EEG: Contrastive-Pretrained Multiscale Convolutional Transformer for EEG Decoding
Jul 10, 2026Gabriel Mahuas, Victoria Shevchenko, Ugo Tanielian +2Contrastive Learning Hybrid CNN-Transformer Architectures
Sigma Nova Paris, France · Goethe University Frankfurt Frankfurt, Germany

Apr 25, 2026J/K move · Enter open · S save

Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes

Kuntal Kokate, Bruno Aristimunha, Dung Truong, Arnaud Delorme

Swartz Center for Computational Neuroscience, Institute for Neural Computation, University of California San Diego, La Jolla, CA 92093, USA · Yneuro, Paris, France · Centre de Recherche Cerveau et Cognition, CNRS, Université Paul Sabatier, Toulouse, France

Fine-Tuning EEG Decoding Negative Transfer EEG Foundation Models

Open paper AskPDF arXiv

Scaling EEG foundation models requires pooling data across heterogeneous electrode montages, a prerequisite both for larger pretraining corpora and for downstream deployment. We present the first systematic comparison of four channel adaptation methods (Conv1d projection, spherical spline interpolation (SSI), source-space decomposition, and Riemannian re-centering) across five pretrained EEG foundation models (5M--157M parameters), five downstream tasks, and two training regimes with 10--15 random seeds each. We find that rigid-montage models (BENDR, Neuro-GPT) require external adaptation, while flexible models (EEGPT, CBraMod) match or exceed it natively when fine-tuned but benefit from external methods under frozen-encoder deployment. A probe-SFT asymmetry exists: external adaptation can cause severe negative transfer during fine-tuning of flexible models. The optimal method is architecture-dependent (Conv1d for BENDR, SSI/Riemannian for Neuro-GPT, source-space decomposition for depression detection), and 5M-parameter CBraMod outperforms models up to 31

\times

larger on 4/5 datasets, consistent with independent findings that compact EEG-specific architectures can match larger models.

Abstract

EEG Foundation Models

Explore similar work

Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes

B[FM]2^22: Brain Foundation Model via Flow Matching with SplitUNet

CoCoT-EEG: Contrastive-Pretrained Multiscale Convolutional Transformer for EEG Decoding

B[FM] $^2$ : Brain Foundation Model via Flow Matching with SplitUNet