Parallel Inference and Learning with Deep Structured Distributions

Many problems in real-world applications involve predicting several random variables which are statistically related. A structured model, like a Markov random field, 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. I will present a model-parallel inference algorithm and illustrate its effectiveness in 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 to estimate complex representations which take into account the dependencies between the random variables. To model those deep structured distributions I will present a sample-parallel training algorithm and show its applicability, among others, by using a 3D scene understanding task.

Date:
Speakers:
Alexander G. Schwing
Affiliation:
University of Toronto