PhD Candidate: New Approximate Bayesian Inference Techniques at the Donders Centre for Cognition

Would you like to conduct cutting-edge research on approximate Bayesian inference and computational complexity theory? Then join the collaborative and supportive work environment of the AI section in the Donders Institute as a PhD candidate. You will be able to put your ideas to the test and push your boundaries. You will do this in a collaborative, multidisciplinary and supportive work environment, with a diverse international staff.
Bayesian inference (the computation of a posterior probability given a prior probability and new evidence) is one of the most crucial computational techniques in artificial intelligence. However, Bayesian inference is an intractable (NP-hard) problem even when only an approximate solution is sought, implying that well-known approximation techniques for Bayesian inference, including variational Bayes, Metropolis-Hasting sampling, and likelihood weighting, only work well on a subset of problem instances, and cannot give a guaranteed quality of approximation in general. This limits the applicability of Bayesian networks for real world applications.

In this PhD project we investigate a new approach towards approximate Bayesian inference, i.e. we translate an inference problem to a weighted satisfiability instance, apply approximation strategies to give an approximate count of the (weighted) number of models of the instance, and then translate the solution back to the inference problem. This approach may open up new avenues as it allows for a new class of approximation strategies based on hashing rather than sampling or model simplification to approximately count models. Currently, however, the state-of-the-art techniques are not yet well suited for the instances that arise from the translation from a Bayesian network to a weighted satisfiability problem. In this project we study how this translation can be adjusted such that hashing approaches work, and study both experimentally and by formal parameterised complexity analysis whether this allows for a novel sub-set of Bayesian inference problems that can be tractably approximated.

In addition to research in this domain, you will contribute to teaching in the BSc and MSc programmes in AI, attend courses offered by the Donders Graduate School and the national research schools IPA and SIKS, and collaborate with inspiring colleagues in the international PGM research community. The teaching contribution for PhD candidates is 10% of your work load, i.e., 0.1 fte in case of a full time contract.