Visual Perceptual to Conceptual First-Order Rule Learning Networks

TL;DR

This paper introduces {b3}ILP, a differentiable framework for rule learning from images that enhances explainability and reasoning in AI, working effectively on both symbolic and raw image data.

Contribution

The paper presents a novel differentiable rule learning framework that automatically induces rules from images without labels or predefined predicates.

Findings

{b3}ILP performs well on symbolic relational datasets.

{b3}ILP effectively learns from relational image data.

{b3}ILP achieves strong results on pure image datasets like 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 {\gamma}ILP, which provides a fully differentiable pipeline from image constant substitution to rule structure induction. Extensive experiments demonstrate that {\gamma}ILP achieves strong performance not only on classical symbolic relational datasets but also on relational image data and pure image datasets, such as Kandinsky patterns.