Investigating maximum likelihood based training of infinite mixtures for uncertainty quantification