Learning Structured Representations and their Applications in Computer Vision

This thesis investigates the representations learned by Deep Neural Networks. We explore how structure can be induced by using prior information we have about the task or data at hand. The research presented includes an exploration of the Generative models domain where a statistical prior is used to better estimate the distribution of given data. The applications of such techniques to real world problems in Computer Vision, such as video prediction and camera pose estimation, is also explored.