Go to ICLR 2026 Conference homepagestLaBraM: Interpretable EEG Foundation Models with Factorized Spatial-Temporal Patch Embedding for Self-Supervised Learning
Keywords: EEG foundation models, self-supervised learning, masked modeling, spatial-temporal filters, interpretability, EEG classification
TL;DR: stLaBraM is an interpretable EEG foundation model with factorized spatial-temporal patch embedding, improving self-supervised reconstruction and classification while providing insights into brain dynamics and relevant cortical regions.
Abstract: Electroencephalography (EEG) foundation models aim to learn robust and transferable representations from large-scale EEG datasets, which are essential for enabling clinical and cognitive applications, such as rapid neurological screening, seizure detection, and brain state decoding. Current architectures struggle to combine interpretability and high performance for self-supervised masked modeling of EEG signals. In medical contexts, interpretability is especially important, because transparent models foster trust and facilitate clinical adoption. In this work, we introduce novel, interpretable spatial and temporal filters in the patch embedding module, advancing EEG foundation models and outperforming the previous state-of-the-art LaBraM architecture. We demonstrate that our approach significantly reduces reconstruction loss during self-supervised pre-training, enhancing the performance of the masked language model (MLM). Our new model outperforms the original LaBraM in standard EEG classification benchmarks and offers unique insights into the second-order dynamical properties and cortical locations of neuronal sources pivotal for self-supervised masked modeling. These results position stLaBraM as a compelling foundation model for EEG, advancing both performance and interpretability in self-supervised neurophysiological representation learning.
Primary Area: applications to neuroscience & cognitive science
Submission Number: 13594
Loading