Homepage Inria website

Section: New Results

EEG Signal Processing

Participant : Fabien Lotte.

To make BCI practical and useful, we need to make them reliable, i.e., able to recognize the users' mental commands, despite noise and non-stationarities [42] . We also need to reduce their calibration time, as current systems needs many examples from each user to calibrate the system for this specific user. This year we addressed these two issues with two different studies.

In order to reduce BCI calibration times, we first surveyed existing approaches, these approaches being notably based on regularization, user-to-user transfer, semi-supervised learning and a-priori physiological information. We then proposed new tools to reduce BCI calibration time. In particular, we proposed to generate artificial EEG trials from the few EEG trials initially available, in order to augment the training set size. These artificial EEG trials are obtained by relevant combinations and distortions of the original trials available. We proposed 3 different methods to do so. We also proposed a new, fast and simple approach to perform user-to-user transfer for BCI. Finally, we studied and compared offline different approaches, both old and new ones, on the data of 50 users from 3 different BCI data sets. This enabled us to identify guidelines about how to reduce or suppress calibration time for BCI [16] .

In order to increased BCI robustness, we performed an empirical comparison of covariance matrix averaging methods for EEG signal classification. Indeed, averaging EEG signal covariance matrices is a key step in designing brain-computer interfaces (BCI) based on the popular common spatial pattern (CSP) algorithm. BCI paradigms are typically structured into trials and we argue that this structure should be taken into account. Moreover, the non-Euclidean structure of covariance matrices should be taken into consideration as well. We reviewed several approaches from the literature for averaging covariance matrices in CSP and compared them empirically on three publicly available data sets. Our results showed that using Riemannian geometry for averaging covariance matrices improves performances for small dimensional problems, but also the limits of this approach when the dimensionality increases [36] .