Visual Perceptual to Conceptual First-Order Rule Learning Networks

2026-04-09

AI Summary

This work addresses the challenge that existing inductive logic programming (ILP) methods struggle to learn first-order logical rules directly from unlabeled images. To overcome this limitation, the paper introduces γILP, a novel framework that, for the first time, enables end-to-end learning of first-order logical rules from purely visual data without any human-provided annotations. γILP integrates a neural perception module with differentiable ILP to form a fully differentiable rule-learning pipeline, which automatically invents visual predicates and jointly models the entire process—from extracting constants from images to inducing rule structures. Experimental results demonstrate that γILP achieves strong performance across diverse benchmarks, including symbolic relational datasets, relational image tasks, and purely visual Kandinsky patterns.

Abstract

Learning rules plays a crucial role in deep learning, particularly in explainable artificial intelligence and enhancing the reasoning capabilities of large language models. While existing rule learning methods are primarily designed for symbolic data, learning rules from image data without supporting image labels and automatically inventing predicates remains a challenge. In this paper, we tackle these inductive rule learning problems from images with a framework called γILP, which provides a fully differentiable pipeline from image constant substitution to rule structure induction. Extensive experiments demonstrate that γILP achieves strong performance not only on classical symbolic relational datasets but also on relational image data and pure image datasets, such as Kandinsky patterns.

Problem

• rule learning
• image data
• predicate invention
• inductive logic programming
• explainable AI

Innovation

• differentiable rule learning
• inductive logic programming
• visual reasoning
• predicate invention
• Kandinsky patterns