|
For the purpose of obtaining an improved quantitative understanding of the dynamics of human brain, contemporary research in the neurosciences employs various observational approaches, such as electroencephalography (EEG), magnetoencephalography (MEG), functional magnetic resonance imaging (FMRI), near-infrared spectroscopy (NIRS) and others; the resulting data consists mainly of multivariate time series. Decomposition of multivariate time series into source components can be accomplished by methods such as Principal Component Analysis (PCA) or Independent Component Analysis (ICA). ICA aims at estimating independent source components for the data; for this purpose, residual dependencies between source components are minimised. However, it has to be expected that part of these dependencies would be coincidental, resulting from finite time series length. In this talk an alternative to typical ICA algorithms will be presented, based on state space modelling. Through simulations it will be demonstrated that state space modelling produces source components with approximately the correct amount of residual dependency. |
|