Deep Learning meets Structured Prediction
Many holistic prediction challenges for real-world applications involve reasoning about several random variables which are statistically related. Markov random fields (MRFs), and energy minimization methods in general, are a great mathematical tool to encode such dependencies. Within this talk we'll show how to combine MRFs with deep learning algorithms to estimate more complex non-linear representations, while taking into account the dependencies between the output variables. Towards this goal, we propose a training algorithm that is able to learn structured models jointly with deep features that form MRF potentials. We then demonstrate its applicability and generality using a variety of 3D scene understanding tasks such as semantic image segmentation, 3D layout prediction and image tagging.
Alexander Schwing is currently a Fields Postdoctoral Fellow at University of Toronto working with Raquel Urtasun. He graduated with a diploma degree in electrical engineering and information technology from Technical University Munich (TUM) and completed his PhD in computer science at ETH Zurich, collaborating mainly with Raquel Urtasun (University of Toronto), Tamir Hazan (Haifa University) and Marc Pollefeys (ETH Zurich). His research focuses on optimization algorithms for inference and learning tasks and his work is motivated among others by applications arising from monocular 3D scene understanding topics. For more information, please browse to http://alexander-schwing.de.