ILP at 30 - An introduction and recent advances

A AAAI 23 tutorial

Tutorial outline. Major limitations of standard ML approaches include poor generalisation, a lack of interpretability, and a need for large numbers of training examples. However, unbeknown to many researchers, recent work in Inductive Logic Programming (ILP) has shown promise at addressing these issues. ILP is a form of ML based on computational logic. The goal is to induce a logic program (a set of logic rules) that generalises training examples. In this tutorial, we will provide an introduction to ILP. We will cover major recent breakthroughs, notably in recursion and predicate invention. We will also emphasise the connections between ILP and the wider AI community, principally the logic programming and constraint-solving communities. By highlighting these connections, we hope to bridge the gap between these communities and foster future research.

Goals of the tutorial. The two main goals of this tutorial are to (G1) introduce ILP to a broad AI audience, and (G2) help bridge the gap between the ML and constraint-solving communities. For G1, recent breakthroughs in ILP have shown promise at addressing the major limitations of standard ML approaches, chiefly through learning recursive theories and predicate invention (the automatic discovery of novel high-level concepts). Our goal is to disseminate these breakthroughs to a wider ML audience. For G2, ILP is uniquely placed to attract a broad AI audience and help bridge the gap between ML and other areas. Because it uses logic programming as a uniform representation for data, the tutorial will interest the logic programming and knowledge representation and reasoning communities. For instance, many recent approaches use answer set programming. Similarly, many recent advances in ILP come from using state-of-the-art techniques from constraint programming. We want to bridge the gap between ILP and these communities. By doing so, ILP can greatly benefit from the ideas and expertise of researchers in these fields. Similarly, other communities can benefit from the challenges faced by ILP. By the end of the tutorial, we hope to have increased the likelihood of collaboration and interchange of ideas between the communities.

Slides: PDF

Tutorial text: ILP at 30:a new introduction