Advances in Differentiable Program Learning

Abstract: Inductive Logic Programming (ILP) is a subfield of Artificial Intelligence that learns Logic Programs for a concept from positive and negative examples of the concept. Learning Logic Programs allow for interpretability, can benefit from background knowledge, and require small training set. However, traditional ILP techniques are not noise-tolerant, and do not scale well to large/high-dimensional domains. In recent years, there have been several attempts to use differentiable representations of logic programs and learn them using gradient descent based techniques. This talk introduces these attempts, and our efforts at extending them to learn logic programs with negations and higher-order logic programs.

In both cases, considerable care is needed from a theoretical standpoint. Negation should be restricted to avoid paradoxical scenarios. We learned logic programs with stratified negation (in the style of Datalog). Anti-unification (i.e., generalization) of arbitrary higher-order terms is not unique. We learned second order logic programs that are generalizations of first order programs.

computation and languagemachine learninglogic in computer scienceprogramming languageslogic

Audience: researchers in the topic

Online logic seminar

Series comments: Description: Seminar on all areas of logic