Deep Generative Models for Visual Understanding

Generative models have always been a long-standing topic for computer vision. From a machine learning perspective, a thorough understanding of the data requires one to generate it. From a neuroscience perspective, a majority of synapses in human visual system comprise of feedback generative connections. Recent advances of deep learning bring new tools and opportunities for generative modeling. This thesis discusses the role of generative models for visual understanding through deep learning. In particular, we build deep generative models for detailed low-level generation and then we advance it for high-level understanding and robotics vision.