EKENS: A Learning on Nonlinear Blindly Mixed Signals
Wai Yie Leong, John P. Homer · 2006
We present experimental results of the blind separation of independent sources from their nonlinear mixtures. The proposed EKENS (equivariant kernel nonlinear separation) algorithm is a generalization of a natural gradient algorithm and the Gram-Charlier series, which is extended in two ways: (1) to deal with nonlinear mapping; (2) to be able to adapt to the actual statistical distributions of the sources by estimating the kernel density distribution at the output signals. The observations are modelled based on nonlinear generative multilayer perceptron analysis. The theory of the EKENS learning algorithm is discussed. Simulations show that the EKENS algorithm is able to find the underlying sources from the observation, even though the data generating mapping is nonlinear and unknown.