Robust parameter estimation of finite mixture models with self-paced learning

Abstract

Self-paced learning (SPL) is a training strategy that mitigates the impact of non-typical observations. SPL introduces observations in a meaningful order by considering the likelihood for each observation. The proposed algorithm considers a finite mixture model that includes a distributional structure for non-typical observations in the SPL weight calculation. Two new self-paced learning (SPL) algorithms is proposed for finite mixture models (FMM). This includes self-paced component learning FMMs and a self-paced learning algorithm that includes a distributional structure for non-typical observations. The properties of these algorithms are presented through a simulation study along with an application on real data. A comparison is made with the properties of well known models. The algorithms shows a reduction in parameter estimation bias which indicates an improvement in the estimation accuracy of the parameters.