Loss function based second-order Jensen inequality and its application to particle variational inference

Bayesian model averaging, obtained as the expectation of a likelihood function by a posterior distribution, has been widely used for prediction, evaluation of uncertainty, and model selection.