ANDRE: An Attention-based Neuro-symbolic Differentiable Rule Extractor for Inductive Logic Programming
Date:
I presented our paper ANDRE: An Attention-based Neuro-symbolic Differentiable Rule Extractor for Inductive Logic Programming online at NeSy 2026, the International Conference on Neurosymbolic Learning and Reasoning, held in Lisbon, Portugal, on September 4, 2026.
The talk walks through why differentiable inductive logic programming struggles in practice (hand-written rule templates, product-based logic operators with vanishing gradients, and brittleness under label noise) and how ANDRE addresses each one:
- A continuous, template-free rule space. Every candidate body predicate can be included, negated, or excluded, and these choices are learned as probabilities by gradient descent.
- Attention as logic. Softmin and softmax attention give a close, differentiable approximation of min and max, so conjunction and disjunction stay faithful while gradients stay alive.
- Results. Correct rules on all ten classical ILP tasks, an average accuracy gain of 21.6% over DFORL on probabilistic datasets, and exact recovery of 9 of 14 ground-truth rules under noisy supervision.
