Towards robust machine learning methods for the analysis of brain data

In this short abstract I will discuss recent directions that machine learning and BCI efforts of the BBCI team and coworkers have taken. It is the nature of this short text that many pointers to research are given all of which show a high overlap to prior own contributions; this is not only unavoidable but intentional. When analysing Brain Data, it is challenging to combine data streams stemming from various modalities (see e.g. Biessmann et al., 2011, Sui et al., 2012, Fazli et al., 2015, Dähne et al., 2015). Hybrid BCIs are a successful example in this direction (Pfurtscheller et al., 2010, Müller-Putz et al. 2015, Dähne et al. 2015, Fazli et al. 2012, 2015). These techniques are firmly rooted in modern machine learning and signal processing that are now readily in use for analysing EEG, for decoding cognitive states etc. (Nikulin et al. 2007, and see Dornhege et al. 2004, Müller et al. 2008, Bünau et al. 2009, Tomioka and Müller, 2010, Blankertz et al., 2008, 2011, 2016, Lemm et al., 2011, for recent reviews and contributions to Machine Learning for BCI). Note that fusing information has also been a very common practice in the sciences and engineering (W altz and Llinas, 1990). The talk will discuss challenges for BCIs that are to be applied outside controlled lab spaces. Such complex and highly artifactual scenarios demand robust signal processing methods; see e.g. Samek et al. 2014, 2017b for recent reviews on robust methods for BCI. In addition I may expand on technical advances on the explanation framework for deep neural networks (Baehrens et al. 2010, Bach et al. 2015, Lapuschkin et al. 2016a and 2016b, Samek et al. 2017a, Montavon et al. 2017, 2018) to BCI data is given (Sturm et al. 2016). Furthermore, time permitting, I will revisit co-adaptive BCI systems (Vidaurre et al. 2011, Müller et al. 2017) and report on an upcoming study connecting fMRI and EEG data for co-adaptive training (Nierhaus et al. 2017). This abstract is based on joint work with Wojciech Samek, Benjamin Blankertz, Gabriel Curio, Michael Tangermann, Siamac Fazli, Vadim Nikulin, Gregoire Montavon, Sebastian Bach/Lapuschkin, Irene Sturm, Arno Villringer, Carmen Vidaurre, Till Nierhaus and many other members of the Berlin Brain Computer Interface team, the machine learning groups and many more esteemed collaborators. We greatly acknowledge funding by BMBF, EU, DFG and NRF.