Blind Nonlinear Source Separation Using EKENS Learning and MLP Network
Wai Yie Leong, John P. Homer · 2006
We propose an equivariant kernel nonlinear separation (EKENS) learning algorithm to extract independent sources from their nonlinear mixtures. Generally, unmixing signals from the nonlinear model in an unsupervised manner is very complicated, because both the nonlinear mapping and the sources distribution are not-known apriori, and should be learned from the observations. The observations are modelled based on nonlinear generative multilayer perceptrons analysis. The theory of the EKENS learning algorithm is discussed. In simulations with artificial data, the EKENS algorithm is able to find the underlying sources from the observation only, even though the data generating mapping is strongly nonlinear and flexible