An accelerated learning algorithm of Gaussian mixture processes
Isamu Shioya, Takao Miura · 2012
This paper presents an accelerated algorithm of parametric learning, in Gaussian mixture processes, which employs Square-root Update method and erases the constraints of the log-likelihood function by utilizing auxiliary parameters embedding the constraints. The algorithm enables us to improve poor convergence, avoids us unstable implementation and removes unnecessary iterations in Gaussian mixture EM algorithm. Our algorithm also allows inexact searches for finding the parameters to maximize the log-likelihood function during the computation, and enables us to implement much efficiently.