A new EM algorithm for flexibly tied GMMs with large number of components

作者：

Highlights：

• A flexible tying scheme is used to solve the memory and computational loads of Gaussian mixture models.

• To handle complexity of cost function, a fast Newton EM algorithm is proposed and is combined with a coordinate descent EM algorithm.

• Computation factorization technique is proposed to increase the speed and decrease the memory requirements for the case of large number of components.

摘要

•A flexible tying scheme is used to solve the memory and computational loads of Gaussian mixture models.•To handle complexity of cost function, a fast Newton EM algorithm is proposed and is combined with a coordinate descent EM algorithm.•Computation factorization technique is proposed to increase the speed and decrease the memory requirements for the case of large number of components.

论文关键词：Gaussian mixture model,Parameter sharing,Tied GMM,Computation factorization and reduction,Newton method,Fast minimal residual method,Clustering

论文评审过程：Received 28 January 2020, Revised 15 October 2020, Accepted 16 January 2021, Available online 22 January 2021, Version of Record 14 February 2021.

论文官网地址：https://doi.org/10.1016/j.patcog.2021.107836