Many problems in real-world applications involve predicting several random variables which are statistically related. A structured model, like a Markov random field (MRF), is a great mathematical tool to encode those dependencies.
Within the first part of this talk I will discuss the difficulties in finding the most likely configuration described by a structured distribution. Subsequently I will present a model-parallel inference algorithm and demonstrate its effectiveness by jointly estimating the disparity of more than 12 million variables.
In the second part, I will show how to combine structured distributions with deep learning algorithms to estimate complex representations while taking into account the dependencies between the output variables. To estimate those deep structured distributions I will discuss a sample-parallel training algorithm that is able to learn structured models jointly with deep features. I will then illustrate its applicability by using, among others, a 3D scene understanding task.