Picture Generation using Variational Auto-Encoders

In recent years probabilistic generative models have achieved an extensive use as an application in distribution modeling of high-dimensional data such as pictures. In the case of complex data structures or large datasets one of computationally tractable and fast models is Variational Auto-Encoders, the subject of this thesis. Variational Auto-Encoders rely on variational Bayesian inference methods, that are more robust and expedite learning by 10–30 times compared to traditional Monte Carlo methods. In this thesis we describe comprehensive theoretical background behind Variational Auto-Encoders, introduce our own practical implementation of Variational Auto-Encoder framework, describe its model hyper-parameters and demonstrate empirical behavior on two image datasets. During these experiments we notice that Variational Auto-Encoders can sometimes produce unrealistic, blurry images. In order to avoid this problem we implement new technique as diversifying regularization using side information to Variational Auto-Encoders, what constitutes the scientific novelty of our thesis. We show empirically that combination of different regularization terms can speed up learning and convergence process which would result in better clustering in the latent space and sharper reconstructed images.