Experiments

The Bayesian ICA algorithm was tested on MEG data, which is identical to the data used in [10]. The data has 122 channels of measurements over two minutes digitized at 148 Hz. The measurements contain signals resulting in the electrical activity of the brain but also signals which can be considered artifacts. These include signals caused by muscular activity, eye movements, cardiac rhythm and even a signal caused by a digital watch that the test subject was wearing. Since it can be assumed that most of the artifacts are independent of the brain activity, it is hoped that ICA can find the artifacts.

The Bayesian ICA algorithm was used to separate 30 sources from the 122 measures channels. The results obtained were comparable to those reported in [10]. In figure 1, five measurements and five non-Gaussian sources found by the algorithm are illustrated.

 

Figure 1: Above: five MEG measurements. Below: five separated sources found by the Bayesian ICA algorithm.

The modification of the estimate of a by the estimate a_Gtypically reduces the convergence time by a factor of ten; the iteration typically converged in 30 iterations.