Characterizing unknown events in MEG data with group factor analysis

Many current neuroscientific experiments can be seen as data analysis problems with two or more data sources: brain activity and stimulus features or, as in this paper, activity of two brains. These setups have been analyzed with Canonical Correlation Analysis or its multiple-source probabilistic extension Group Factor Analysis, which capture statistical dependencies between the data sources in correlating components. We relax the assumption of global correlations and search for correlating signals related to discrete events. The assumption is that the sources correlate only during events with known timings, inferred from a stimulus stream for instance, but the type or nature of each event is not known. The unsupervised modelling of the events can then be viewed as a generalization of conditional averaging. We apply the model on two-person MEG measurements, in a demonstration task of identifying which of the two persons utters a word.