Abstract
In this paper we propose two novel bounds for the log-likelihood based onKullback-Leibler and the R\'{e}nyi divergences, which can be used forvariational inference and in particular for the training of VariationalAutoEncoders. Our proposal is motivated by the difficulties encountered intraining VAEs on continuous datasets with high contrast images, such as thosewith handwritten digits and characters, where numerical issues often appearunless noise is added, either to the dataset during training or to thegenerative model given by the decoder. The new bounds we propose, which areobtained from the maximization of the likelihood of an interval for theobservations, allow numerically stable training procedures without thenecessity of adding any extra source of noise to the data.