Safety Properties of Inductive Logic Programming
Gavin Leech, Nandi Schoots, Joar Skalse · Bristol Research (University of Bristol) · 2021
This paper investigates the safety properties of inductive logic programming (ILP), particularly as compared to deep learning systems. We consider the following properties: ease of model specification; robustness to input change; control over inductive bias; verification of specifications; post-hoc model editing; transfer learning; and interpretability. We find that ILP satisfies many of these properties in some domains. Lastly, we propose a hybrid system using ILP as a preprocessor to generate specifications for other ML systems.