We approach the problems of blind source separation and blind deconvolution from the point of view of a single neuron. Two simple non-linear rules for a neuron are presented. When used for blind source separation, the first rule learns to separate one (arbitrary) source which has a negative kurtosis (i.e. is sub-Gaussian), and the second rule separates a source with positive kurtosis (i.e. a super-Gaussian source). Formulating the problem of blind deconvolution as a special case of blind soruce separation, we can also apply our learning rules for that problem. Then, the single unit learns to deconvolve the signal.
展开▼
