Theoretical and applicational advances in variational autoencoders

Variational autoencoders (VAEs) is a strong family of deep generative models based on variational inference (VI). Despite its recent success in modelling probabilistic data distributions, the original VAE framework make a strong assumption on the overly-simplified prior and approximate posterior, leading to limited modelling capacity. In this thesis, we propose novel solutions to alleviate this problem by introducing a learnable Bayesian network as the latent space of VAEs. In addition to the theoretical contributions to VAE research community, we also propose to apply VAEs to various promising applications including controllable video generation, asynchronous event sequences modelling and scene layout generation.