Mixture conditional density estimation with the EM algorithm

It is well-known that training a neural networkwith least squares corresponds to estimatinga parametrized form of the conditionalaverage of targets given inputs. In orderto approximate multi-valued mappings,e.g., those occurring in inverse problems,a mixture of conditional densities must beused. In this paper we apply the EM algorithmto fit a mixture of Gaussian conditionaldensities when the parameters of themixture, i.e., priors, means, and variancesare all functions of the inputs....