A Comparative Study of Three Neural-Symbolic Approaches to Inductive Logic Programming

An interesting feature that traditional approaches to inductive logic programming are missing is the ability to treat noisy and non-logical data. Neural-symbolic approaches to inductive logic programming have been recently proposed to combine the advantages of inductive logic programming, in terms of interpretability and generalization capability, with the characteristic capacity of deep learning to treat noisy and non-logical data. This paper concisely surveys and briefly compares three promising neural-symbolic approaches to inductive logic programming that have been proposed in the last five years. The considered approaches use Datalog dialects to represent background knowledge, and they are capable of producing reusable logical rules from noisy and non-logical data. Therefore, they provide an effective means to combine logical reasoning with state-of-the-art machine learning.