NI Colloquium by Reinmar Kobler


On May 6th, 2024, at the invitation of the Neuroinformatics research group, Dipl.-lng. Reinmar Kobler will give a lecture titled "Geometric deep learning to advance EEG BCI generalization". The Faculty of Computer Science cordially invites all interested people to attend!


Reinmar Kobler is a Research Scientist at the Department of Dynamic Brain Imaging, ATR in Japan, where he is working on robust and interpretable machine learning techniques for functional neuroimaging. He received his Bachelor, Master and PhD (sub auspiciis praesidentis) degrees from Graz University of Technology in Austria. In his PhD thesis he pioneered functional neuroimaging tools to disentangle brain activity in magneto-/electroencephalography (M/EEG) from co-varying artifacts during goal-directed movements, enabling him to substantiate and extend the understanding of kinematics-related effects in M/EEG. In his recent works, he combined Riemannian geometry and deep learning approaches to advance robust and interpretable machine learning techniques for brain-computer interfacing (BCI) as well as multimodal neuroimaging data fusion.

Reinmar Kobler - Google Scholar Profile


Title and abstract of his presentation:

Geometric deep learning to advance EEG BCI generalization

Current brain-computer interfaces (BCIs) do not generalize well across domains (e.g., sessions and subjects) without expensive supervised re-calibration on small domain-specific data, which severely limits BCI utility and scalability. Fortunately, geometric deep learning offers a remedy via combining data-efficiency and invariances of Riemannian geometry aware methods with feature learning capabilities of neural nets. In this talk, I will present a geometric deep learning framework to perform unsupervised domain adaptation (UDA) on the symmetric, positive definite (SPD) manifold. The framework can be readily applied to multi-source/-target and online UDA scenarios. Using oscillatory EEG BCI datasets, we demonstrate that a simple, network architecture, which we denote TSMNet, can obtain state-of-the-art performance in inter-session and -subject transfer without compromising on neurophysiological interpretability. Additionally, the framework can be extended to extract and fuse latent brain dynamics encoded in EEG and fMRI via maximizing geodesic correlation among latent representations of simultaneous recordings.


Monday, May 6, 2024 at 3:00 p.m. // Währinger Straße 29, 1.UG, Seminarraum 4