MSc Seminar: Benjamin Baird

Title: Towards more efficient CPT compression and lazy inference with NAT-Modeled Bayesian Networks

Abstract:

A Bayesian network (BN) is used to represent conditional dependencies among a set of random variables. However, the acquisition of conditional probabilities tables (CPT) for the variables can be quite expensive as the family size of variables increases. This problem is solved by compressing a BN's CPT into a Non-Impeding Noisy-AND Tree (NAT), which can model reinforcement and undermining relationships in linear time. This compression step currently utilizes gradient descent to compress the CPT. By modifying the gradient descent to become swarm-like we are able to improve the efficiency of the compression step. When the compression step is completed the NAT is used in place of the BN for probabilistic inference. In this step, the NAT can be converted into a junction tree and lazy propagation can be exploited. It has been shown that for some BNs lazy propagation shows a substantial improvement in inference speed. Further exploration will be done to investigate how the efficiency gain of lazy inference is affected by various properties of the BN. 

Advisor: Dr. Yang Xiang
Advisory Committee Member:  Dr. Andrew Hamilton-Wright