MS.c. Seminar: Dylan Loker

TITLE:

Compilation of the Non-Impeding Noisy-AND model
Dylan Loker

ABSTRACT:

Bayesian Networks are a useful tool for representing uncertain knowledge about the world. However, they suffer from intractable inference when the family size of variables is large. Non-Impeding Noisy-AND Trees (NAT) are capable of representing reinforcing and undermining relationships within a variable family in a linear fashion, which improves the speed of inference. We hope to improve this speed further by developing a compilation method from Bayesian Networks which utilize NAT, into Sum-Product Networks (SPN) which encode the inference process structurally and allow for substantial computational savings when compiled from networks with local structure (such as NAT).

Advisor: Dr. Yang Xiang
Advisory Committee: Dr. Luiza Antonie