New geometrical approach for blind separation of sources mapped to a neural network

A new method is proposed for the blind separation of mixed digital or analog sources, based on geometrical considerations concerning the observation space. For p mixed sources, where p is greater than or equal to two, the new approach considers the p-dimensional hyperparallelepiped formed in the observation space, and by means of a neural network with w _if weights, computes the coordinates of p vectors corresponding to the image of orthogonal inputs in the source space. These coordinates provide the columns of the unknown mixture matrix A, with a_if elements, and the neural network recursively separates the unknown sources, S₀. This geometrical procedure does not need the computation of any order of statistics, using instead primitives that may easily be implemented by hardware and it has a polynomial complexity (p³) which depends on the number of sources (p)