The Generalized Gaussian Mixture Model Using ICA
Te-Won Lee, Michael S. Lewicki · 2000
An extension of the Gaussian mixture model is presented using Independent Component Analysis (ICA) and the generalized Gaussian density model. The mixture model assumes that the observed data can be categorized into mutually exclusive classes whose components are generated by a linear combination of independent sources. The source densities are modeled by generalized Gaussians (Box and Tiao, 1973) that provide a general method for modeling non-Gaussian statistical structure of univariate distributions that have the form p(x) / exp(\\Gammajxj q ). By inferring q, a wide class of statistical distributions can be characterized including uniform, Gaussian, Laplacian, and other suband super-Gaussian densities. The generalized Gaussian mixture model using ICA infers for each class the source parameters, the basis functions and bias vectors. The new method can improve classification accuracy compared with standard Gaussian mixture models and shows promise for accurately modeling structure in...