Information geometry meets BCI spatial filtering using divergences

Algorithms using concepts from information geometry have recently become very popular in machine learning and signal processing. These methods not only have a solid mathematical foundation but they also allow to interpret the optimization process and the solution from a geometric perspective. In this paper we apply information geometry to Brain-Computer Interfacing (BCI). More precisely, we show that the spatial filter computation in BCI can be cast into an information geometric framework based on divergence maximization. This formulation not only allows to integrate many of the recently proposed CSP algorithms in a principled manner, but also enables us to easily develop novel CSP variants with different properties. We evaluate the potentials of our information geometric framework on a data set containing recordings from 80 subjects.