Blind Signal Separation with Kernel Probability Density Estimation Based on MMI Criterion Optimized by Conjugate Gradient

This paper presents a novel blind signal separation (BSS) approach based on the theory of independent component analysis. In the proposed BSS approach, the learning rule is derived by the conjugate gradient optimization algorithm rather than the ordinary gradient and natural gradient algorithm based on the minimum mutual information (MMI) criterion. The score function is a key point in solving the BSS problem. Instead of choosing nonlinear activity functions empirically, a kernel probability density function estimation method is used in order to estimate the probability density functions and their derivatives of the separated signals. Thus the score function is then estimated directly. The proposed BSS approach is applied to separate the mixtures of sub-Gaussian and super-Gaussian source signals simultaneously. Computer simulations are provided to demonstrate the superior learning performance of the proposed BSS approach. 