Mixtures of several hundred signals can be blindly separated by Independent Components Analysis (ICA), an new unsupervised neural network learning algorithm that generalizes Principal Component Analysis to nongaussian signals that are nonorthogonal. This new technique can be applied to data at many different spatial and temporal scales and has many areas of application in signal processing and datamining. When applied to patches from natural images, ICA finds components that resemble localized and oriented Gabor filters, similar to responses of neurons in the primary visual cortex of primates. This suggests that the visual cortex preprocesses visual information into channels that are maximally independent. When applied to functional magnetic resonance imaging data (fMRI), which allows cognitive brain activity to be measured in humans noninvasively, the resulting ICA components consist of spatially-fixed 3-D maps of distributed activity and associated time courses of activation which identify spatially independent brain processing systems.
A list of all publications for this speaker can be found at http://www.cnl.salk.edu/cgi-bin/pub-search.