Blind separation of mixed-kurtosis signals using soft switch activation function

A soft switch nonlinear activation function for the blind separation of signals with mixed kurtosis signs is presented. The parameter of the nonlinear activation function is adaptively determined during the blind separation process, which does not need to estimate the kurtosis of the recovered signals. Numerical simulations demonstrate that the proposed algorithm is suitable for recovering sub-Gaussian, Gaussian and super-Gaussian signals from their mixtures.