Supervisor: Johannes Fürnkranz

Motivation:

One of the key factors for the success of deep learning is the ability to learn intermediate concepts in the hidden layers of the network. Inductive rule learning has not yet found a good solution for this problem, also known as constructive induction or predicate invention. Current methods operate on a few data with strong templates in first-order logic rules.

Objective:

The task of this thesis is to investigate the potential of optimizing a fixed rule structure on large datasets, drawing upon ideas from ensemble learning, sum-product networks, and deep learning.