LAtte: Hyperbolic Lorentz Attention for Cross-Subject EEG Classification

Ahmad Bdeir; Johannes Burchert; Tom Hanika; Lars Schmidt-Thieme; Niels Landwehr

cs.LGMar 11, 2026

LAtte: Hyperbolic Lorentz Attention for Cross-Subject EEG Classification

Authors: Ahmad Bdeir, Johannes Burchert, Tom Hanika, Lars Schmidt-Thieme, Niels Landwehr

Organizations: Data Science Group, Universität Hildesheim · ISMLL, Universität Hildesheim

Abstract

Electroencephalogram (EEG) classification plays a key role in medical diagnosis and brain-computer interfaces, but remains challenging due to low signal-to-noise ratios and high inter-subject variability. As a result, many existing approaches rely on subject-specific models, which fail to exploit shared structure in neural signals and do not generalize to unseen subjects. To address these limitations, we propose LAtte, a framework that combines Lorentz attention with a hyperbolic InceptionTime-based encoder to improve cross-subject generalization in EEG classification. The model explicitly decomposes EEG signals into a learned baseline component and task-relevant deviations, enabling more structured representation learning. To further improve robustness and adaptability, we incorporate subject-specific low-rank adaptation (LoRA) modules at both encoder and decoder levels, augmented with a Lorentz boost-based LoRA mechanism and hyperbolic projection layers to reduce overfitting in geometric representations. We evaluate LAtte with and without finetuning in three settings: subject-specific, subject-conditional, and leave-one-subject-out (LOSO) on five established EEG datasets, achieving a consistent improvement in performance over current state-of-the-art methods for smaller datasets and maintaining performance for larger datasets.

Figures & tables

	Models	BCIC-IV 2a	MAMEM II	BCI Challenge
Subject-Specific	ShallowConvNet	61.84 $\pm$ 6.39	56.93 $\pm$ 6.97	71.86 $\pm$ 2.64
	EEGNet	57.43 $\pm$ 6.25	53.72 $\pm$ 7.23	74.28 $\pm$ 2.47
	SCCNet	71.95 $\pm$ 5.05	62.11 $\pm$ 7.70	70.93 $\pm$ 2.31
	EEG-TCNet	67.09 $\pm$ 4.66	55.45 $\pm$ 7.66	77.05 $\pm$ 2.46
	TCNet-Fusion	56.52 $\pm$ 3.07	45.00 $\pm$ 6.45	70.46 $\pm$ 2.94
	FBCNet	71.45 $\pm$ 4.45	53.09 $\pm$ 5.67	60.47 $\pm$ 3.06

Table 1: Performance comparison on the EEG datasets BCIC-IV 2a, MAMEM II, and BCI Challenge. We repeated the experiments 10 times on random seeds and report the average accuracy for BCIC-IV 2a, and MAMEM II, and the AUC for BCI Challenge. We present t-test results as ** for extreme significance and * for significance.

Models	BCICIV-2b	HGD
EEGNet	83.65 $\pm$ 0.46	85.59 $\pm$ 0.45
ShallowConvNet	81.45 $\pm$ 0.50	89.75 $\pm$ 0.54
BaseNet	86.11 $\pm$ 0.40	93.64 $\pm$ 0.86
EEG-TCNet	86.74 $\pm$ 0.12	91.83 $\pm$ 1.8
TS-SEFFNet	84.18 $\pm$ 0.42	92.45 $\pm$ 0.85
CTNet	86.91 $\pm$ 0.29	94.21 $\pm$ 0.52

Table 2: Performance comparison on the EEG datasets BCICIV-2b, and HGD for the subject-specific setting. FT represents finetuning. Experiments are repeated 3 times for HGD and 5 times for BCICIV-2b to match the reporting from Altaheri et al. (2025) . We present t-test results as ** for extreme significance and * for significance.

Models	BCICIV-2a	BCICIV-2b	HGD
EEGNet	52.03 $\pm$ 0.88	77.89 $\pm$ 0.73	57.95 $\pm$ 0.51
ShallowConvNet	47.31 $\pm$ 1.06	75.58 $\pm$ 1.01	72.47 $\pm$ 0.94
BaseNet	56.89 $\pm$ 0.88	78.61 $\pm$ 0.57	68.55 $\pm$ 1.74
EEG-TCNet	55.99 $\pm$ 0.84	80.56 $\pm$ 0.20	60.59 $\pm$ 1.98
TS-SEFFNet	56.74 $\pm$ 0.83	77.82 $\pm$ 0.78	69.99 $\pm$ 0.57
CTNet	60.09 $\pm$ 0.94	80.29 $\pm$ 0.42	64.6 $\pm$ 0.9

Table 3: LOSO performance comparison on BCICIV-2a, BCICIV-2b, and HGD. Experiments are repeated 3 times for HGD and 5 times for the others to match the reporting from Altaheri et al. (2025) . We present t-test results as ** for extreme significance and * for significance.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 1 — Figure 1: LAtte architecture overview

Dataset	Task	Classes	Trials used
MAMEM	SSVEP	5	3,300
BCICha	Error monitoring	2	2,880
BCIC IV 2a	Motor imagery	4	2,592
BCIC IV 2b	Motor imagery	2	3,680
HGD	Motor imagery	4	5,000

Appendix

Table 4: EEG datasets used in the

\delta

-hyperbolicity analysis.

Dataset	$D_{\mathrm{enc}}$	Baseline	$\delta_{\mathrm{rel}}^{q95}$ (Real)	$\delta_{\mathrm{rel}}^{q95}$ (Baseline)	$\Delta_{q95}$	$\\|\Delta\\|/\sigma$
MAMEM	1488	Shuffle	0.0038	0.0116	$-$ 0.0078	11.1
		Gaussian	0.0042	0.0222	$-$ 0.0180	352.5
BCICha	96	Shuffle	0.0238	0.0388	$-$ 0.0150	26.5
		Gaussian	0.0239	0.0565	$-$ 0.0326	102.5
BCIC IV 2a	816	Shuffle	0.0167	0.0193	$-$ 0.0026	8.9
		Gaussian	0.0165	0.0281	$-$ 0.0116	79.5

Appendix

Table 5: Gromov

\delta

-hyperbolicity of untrained TCFormer encoder features.

\delta_{\mathrm{rel}}^{q95}

is the diameter-normalised 95th-percentile

\delta

(lower = more tree-like).

\Delta_{q95}

is the difference vs. baseline; negative values indicate stronger hyperbolicity than expected by chance.

\|\Delta\|/\sigma

quantifies statistical separation.

Figure 2: A graphical representation of the normalized hyperbolicity values and deviation from baseline for RAW and Encoded data.

Dataset	$D_{\mathrm{raw}}$	Raw $\Delta_{q95}$	$D_{\mathrm{enc}}$	Encoder $\Delta_{q95}$	Geometric change
MAMEM	1,000	$-$ 0.0097	1,488	$-$ 0.0078	Remains hyperbolic
BCICha	8,960	$+$ 0.0112	96	$-$ 0.0150	Euclidean $\rightarrow$ Hyperbolic
BCIC IV 2a	22,000	$+$ 0.0078	816	$-$ 0.0026	Euclidean $\rightarrow$ Hyperbolic
BCIC IV 2b	3,375	$-$ 0.0031	960	$-$ 0.0065	Strengthened
HGD	44,000	$-$ 0.0003	816	$-$ 0.0093	Strengthened

Appendix

Table 6: Encoder vs. no-encoder comparison (shuffle baseline). Negative

\Delta_{q95}

= more hyperbolic than baseline.

Models	MAMEM II	BCI Challenge
EEGFormer	29.55 $\pm$ 5.40	73.30 $\pm$ 3.55
ATCNet	50.31 $\pm$ 3.82	73.58 $\pm$ 4.32
CBraMod	43.58 $\pm$ 2.63	66.70 $\pm$ 299
TCFormer	42.18 $\pm$ 3.91	74.28 $\pm$ 4.10
LAtte	57.86 $\pm$ 3.66	75.31 $\pm$ 3.17
Relative $\Delta$	15.00%	1.38%

Appendix

Table 7: Performance comparison for LOSO protocol. The experiments are repeated 5 times on random seeds, and we report accuracy for BCI-IV 2a and MAMEM II, and AUC for BCI Challenge.

Subject	InceptionTime	MAtt	HyperMAtt	LAtte
1	$80.40\pm 2.06$	$81.60\pm 2.87$	$\underline{89.94}\pm 1.29$	$\textbf{90.00}\pm 2.18$
2	$86.60\pm 1.62$	$89.40\pm 1.36$	$\underline{90.31}\pm 0.96$	$\textbf{94.50}\pm 4.62$
3	$61.60\pm 3.07$	$58.20\pm 5.64$	$\underline{68.05}\pm 3.06$	$\textbf{74.20}\pm 0.16$
4	$25.00\pm 4.00$	$20.60\pm 3.88$	$\underline{30.00}\pm 0.58$	$\textbf{42.67}\pm 1.89$
5	$25.00\pm 6.72$	$26.40\pm 4.80$	$\underline{30.96}\pm 1.53$	$\textbf{53.20}\pm 2.34$
6	$79.20\pm 1.72$	$79.00\pm 2.68$	$\underline{80.90}\pm 1.55$	$\textbf{86.40}\pm 0.70$

Appendix

Table 8: Performance Comparison on the MAMEM II dataset for the subject-specific case. We report the average accuracy over 10 runs, respectively.

Models		BCIC IV 2a	MAMEM II	BCI Challenge
CBraMod SS		45.22 $\pm$ 1.26	40.45 $\pm$ 2.12	68.49 $\pm$ 1.17
CBraMod SA		39.14 $\pm$ 1.55	49.91 $\pm$ 0.00	72.90 $\pm$ 0.46
CBraMod +	IncShallow	40.93 $\pm$ 2.41	45.47 $\pm$ 1.63	70.77 $\pm$ 0.14
CBraMod❄ +	IncShallow	30.01 $\pm$ 0.06	24.00 $\pm$ 0.00	53.52 $\pm$ 0.01
CBraMod +	IncDeep	41.71 $\pm$ 2.23	48.09 $\pm$ 0.42	70.03 $\pm$ 0.10
CBraMod❄ +	IncDeep	32.83 $\pm$ 0.00	26.87 $\pm$ 1.57	57.39 $\pm$ 9.32

Appendix

Table 9: Performance comparison of foundation model configurations on the EEG datasets BCI-IV 2a, MAMEM II, and BCI Challenge. SS and SA denote subject-specific and subject-agnostic training of the native CBraMod model, respectively. ❄ indicates that the CBraMod backbone weights are frozen during training. We report accuracy for BCI-IV 2a and SSVEP and AUC for ERN, averaged over 5 runs. The last row (shaded) shows the standard LAtte model without any foundation model component for reference. The best result per dataset is highlighted in bold.

Model Variant	Accuracy	$\Delta$
LAtte Full	71.53 $\pm$ 1.00	-
LAtte w/o Lorentz	57.40 $\pm$ 6.53	-19.75%
LAtte w/o LoRA	62.87 $\pm$ 0.51	-12.11%
LAtte w/o Proj	69.34 $\pm$ 1.37	-3.06%
LAtte w/o Cut/Fill	69.55 $\pm$ 0.71	-2.77%

Appendix

Table 10: Component Analysis of LAtte. We compare the performance of the full LAtte model to versions without the main modules on MAMEM II. We report the accuracy and std over 5 runs.

Model Variant	BCIC-IV 2a	$\Delta$	BCIC-IV 2b	$\Delta$
LAtte Full	78.85 $\pm$ 2.55	-	88.87 $\pm$ 0.17	-
LAtte w/o Baseline Block	72.90 $\pm$ 1.67	-7.54%	84.50 $\pm$ 0.47	-4.9%
LAtte w/o Hyperbolic MaxPool	77.01 $\pm$ 1.34	-2.33%	87.39 $\pm$ 0.48	-1.66%
LAtte w/o LB-LoRA	77.21 $\pm$ 2.13	-2.1%	86.29 $\pm$ 0.30	-2.9%

Appendix

Table 11: Additional Component Analysis of LAtte. We compare the performance of the full LAtte model to versions without the main modules on BCIC-IV 2a and BCIC-IV 2b.

Model Variant	Accuracy
LAtte + Identity LoRA	44.22 $\pm$ 4.56
LAtte + Embedding Distance	57.32 $\pm$ 1.51
LAtte + SPD Covariance Metric	56.69 $\pm$ 2.22
LAtte + Slot-0 generalization	65.32 $\pm$ 0.71

Appendix

Table 12: Alternative LOSO approaches for the BCIC-2A dataset

Cross-Subject EEG Decoding

Cross-subject EEG decoding develops methods that classify or predict brain activity for people not seen during training.

See all 77 papers

Stacked LoRA for Subject-Adaptive EEG Foundation Models in Motor Imagery Decoding
Jul 3, 2026Aymen Sarhane, Fouad Lbakali, Mouad Souissi +2Cross-Subject EEG Decoding Brain-Computer Interfaces
IMT Atlantique, Lab-STICC, Brest, France
Are EEG Sequences Time Series? EEG Classification with Time Series Models and Joint Subject Training
Apr 10, 2024Johannes Burchert, Thorben Werner, Vijaya Krishna Yalavarthi +3Time Series Classification Cross-Subject EEG Decoding
Institute for Computer Science, University of Hildesheim, Hildesheim, German
Zero-Shot Neural Priors for Generalizable Cross-Subject and Cross-Task EEG Decoding
Jun 12, 2026Baimam Boukar Jean Jacques, Brandone Fonya, Nchofon Tagha Ghogomu +2Cross-Subject EEG Decoding Zero-Shot Learning
Carnegie Mellon University

Jul 3, 2026J/K move · Enter open · S save

Stacked LoRA for Subject-Adaptive EEG Foundation Models in Motor Imagery Decoding

Aymen Sarhane, Fouad Lbakali, Mouad Souissi, Jonathan Lys+1

IMT Atlantique, Lab-STICC, Brest, France

Cross-Subject EEG Decoding Brain-Computer Interfaces Neural Decoding EEG Decoding Low-Rank Adaptation EEG Foundation Models

Open paper AskPDF arXiv

Electroencephalography (EEG) decoding for brain-computer interfaces (BCIs) faces a major challenge: substantial inter-subject variability limits effective cross-subject generalization. Consequently, practical systems still rely largely on subject-specific models trained from scratch and requiring individual recalibration. EEG foundation models have recently emerged as a promising alternative; however, even large pretrained models cannot simply be used as fixed feature extractors and still require additional adaptation before they can be reliably applied to downstream tasks. In this work, we address this challenge through targeted adaptation strategies. Building on recent EEG foundation models such as REVE, LaBraM, and LUNA, we examine the impact of different low-rank adaptation strategies on motor imagery classification. We propose a framework that structurally decouples subject-invariant knowledge from subject-specific neural signatures: the low-rank update at each adapted layer is split into a Global adapter, trained jointly across all subjects, and Subject-Specific adapters, each absorbing individual variability. To assess the contribution of each path, we compare three adaptation strategies: (i) subject-specific LoRA (ii) global LoRA and (iii) stacked LoRA, combining both Global and Subject Specific adapters. Experiments on BCI Competition IV-2a, PhysioNet Motor Imagery, and the clinical Zuo2025 benchmark show that Stacked LoRA effectively mitigates inter-subject variability, achieving the best accuracy in the large majority of backbone and dataset combinations. Our analysis further reveals that the optimal balance between the global and subject-specific paths depends on the target population: a shared adapter is sufficient for large, diverse cohorts, whereas subject-specific adaptation is decisive under the high inter-session variability of clinical recordings.

Abstract

Cross-Subject EEG Decoding

Explore similar work

Stacked LoRA for Subject-Adaptive EEG Foundation Models in Motor Imagery Decoding

Are EEG Sequences Time Series? EEG Classification with Time Series Models and Joint Subject Training

Zero-Shot Neural Priors for Generalizable Cross-Subject and Cross-Task EEG Decoding