Year of Graduation
Tensoral Methods in Generative Models Consisting Discrete Stochastic Nodes
Statistical Learning Theory
Some real world data, such as images from several distinct classes, can be modeled with discrete latent variables. However, training discrete latent variable models remains a challenge and all methods suffer from either high bias or high variance of the gradient estimate compared to training continuous models. In this Master's thesis, we propose a family of unbiased methods for training Variational Autoencoders with discrete latent variables that are based on a structured decomposition of the log-likelihood tensor. The proposed family generalizes a known algorithm —— MuProp —— to more expressive decompositions of the loss tensor and shows state of the art results in terms of Evidence Lower Bound convergence on a set of vision tasks.