Investigating Foundation Models, Disentanglement and Latent Alignment for Subject-Independent EEG Learning

Abstract

Brain–Computer Interfaces (BCIs) require models that generalize across subjects, yet EEG signals exhibit strong inter-subject variability and non-stationarity, leading to performance degradation on unseen users. This limitation is particularly critical in interactive and multimodal systems, where reliable, calibration-free decoding is essential for real-time interaction, and even minor errors can directly affect user experience. To address this challenge, several paradigms have been explored: foundation models leverage large-scale pretraining, disentangled representation learning aims to separate subject, task, and noise factors, and alignment methods focus on reducing distribution shifts in the feature space. Despite their promise, a systematic comparison of these approaches is still lacking. In this work, we propose a modular framework for the unified evaluation of these paradigms in subject-independent learning. We conduct extensive experiments on Motor Imagery (MI) and Event-Related Potential (ERP) datasets, representative of interactive BCI scenarios spanning both oscillatory (MI) and event-related (ERP) paradigms, to analyze their relative effectiveness and practical trade-offs. The code is publicly available at: https://github.com/phuselab/eeg_subj_indep_learning.