Improving Probabilistic Inference
- Master Thesis
- Announcement date
- 10 Mar 2021
- Research Areas
Probabilistic graphical models are one of the most important concepts for representing uncertainties in large and possibly distributed systems. For probabilistic reasoning, it is necessary to rely on efficient message-passing algorithms as belief propagation (BP). Over the years, BP has been successfully used in a wide range of applications; these include artificial intelligence, graph neural networks, social network analysis, computer vision, signal detection wireless networks, error-correcting codes, and many more.
Despite this success story, BP is only an approximate method that sometimes fails to converge or to provide accurate results. Our research focuses on (i) theoretical work to understand the behavior of BP and to give guarantees under which BP performs well and on (ii) algorithmic improvements of BP by pruning the computation graph or intelligent scheduling methods.
- an existing toolbox with implementations of BP-variants
- various implemented application scenarios
- possibility to do fundamental research on a highly relevant topic
- implement enhanced BP method(s)
- validate the proposed algorithm(s) empirically
- (optional) theoretically analyze the proposed algorithm(s)
- motivation and interest in the topic
- experience in python programming
- background in machine learning is preferable
This thesis project is planned for a duration of 6 months starting immediately.
- Christian Knoll (firstname.lastname@example.org)
- Franz Pernkopf (email@example.com)