Equi-convergence Algorithm Based on Asymmetric Generalized Gaussian System for Blind Source Separation

A new equi-convergence learning algorithm for blind source separation (BSS) is proposed in this paper. Taking into account the asymmetry of the distributions, the asymmetric generalized Gaussian (AGG) model is employed to model source distributions. To avoid directly estimating the source distributions, we update the activation functions adaptively. And also we use the posterior distributions of source signals to estimate the minor property parameter. The learning rule is compatible with minimization of mutual information for training demixing model. Combining the AGG model with this adaptation approach, we propose our method eICA. Finally the simulation examples are given to demonstrate the reliable performance and validity of the proposed method