Machine Learning and Signal Processing for Brain-Computer Interfaces
Duration of project
Swiss National Science Foundation
Brain-computer interfaces (BCIs) are systems that enable communication with other persons or control of devices, only through cerebral activity, without using muscles. In this project we have concentrated mainly on the development of signal processing and machine learning methods for the classification of brain-signals. Almost all BCI systems rely on such methods in order to learn classification rules from training data and to translate measurements of cerebral activity into commands for a computer. The main outcome of this project are machine learning methods that make brain-computer interfaces (BCIs) more practical. In particular the proposed methods are a) regularized and hence relatively robust with respect to outliers, b) fully automatic, c) allow to automatically reduce the number of sensors for acquisition of brain signals, and d) provide probabilistic outputs. All of these properties are important for building practical BCI systems. Some of the above mentioned algorithms and a database of electroencephalogram recordings that has been used for evaluation have been made available on the internet. This was done in order to allow other researchers to reproduce the experiments and to further analyze the data we recorded.
K. Diserens, U. Hoffmann, E. Girardet, R. Leroy , A. Massy, N. Gremaud, T. Ebrahimi, P.A. Despland, J. Bogousslavsky, “Brain-Computer Interface: Communication by cognition for neurological patients with severe cerebellar and brainstem lesions”, Abstract in proceedings of the 16th meeting of the European Neurological Society, 2006.