Video: What masking geometry works best for EEG foundation models?

In this 2-minute video, Pierre Guetschel explains the main findings of the new paper on pre-training EEG foundation models. The study compares different masking geometries, i.e. which parts of the EEG signal are hidden during pre-training, across two self-supervised approaches: masked autoencoders (MAE) and joint-embedding predictive architectures (JEPA).

The full paper is available on arXiv, and more details can be found on the project page. The code and the pre-trained models are publicly available.

This work was done together with Bruno Aristimunha, Yassine El Ouahidi, Arnaud Delorme, Thomas Moreau, and Michael Tangermann.