Square root update acceleration of the EM algorithm in Gaussian mixture processes
Isamu Shioya, Takao Miura · 2011
This paper presents a new expectation maximization (EM) algorithm, which employees Square-root Update method combined by conventional Gaussian mixture EM algorithm, to accelerate the parameter learning of Gaussian mixture models. The algorithm enables us to improve poor convergence, avoids us unstable implementation and removes unnecessary iterations by employing inexact searches during the maximization processes. The convergence is faster compared to conventional EM algorithm. Furthermore, our proposal algorithm can be applied to autoregressive Gaussian mixture stationary processes.