Kurtosis-Switching Based Blind Source Separation for Linear Mixed Signals

For blind source separation of linear mixed signals, when determining source probability density functions and corresponding activation functions proves difficult - especially when source signals contain both super-Gaussian and sub-Gaussian components - we employ kurtosis as a statistical measure of signal PDFs to adaptively configure activation functions. This kurtosis-switching based blind separation algorithm effectively handles complex signal mixing scenarios through implementation features including: kurtosis calculation for signal classification, adaptive switching logic between hyperbolic tangent (tanh) and polynomial functions based on kurtosis thresholds, and iterative optimization using natural gradient descent. The core algorithm structure involves computing kurtosis values for each estimated source, selecting appropriate nonlinearities for separation, and updating the demixing matrix through stability-ensured learning rules.