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.
