Exploring Abstract Reasoning Methods Working with Disentangled Visual Attributes

Abstract

The capacity for abstract reasoning is a cornerstone of human intelligence, and replicating this in machine learning models presents a complex challenge. This study focuses on the RAVEN dataset, a standard for assessing abstract reasoning abilities. Building upon the foundational work of Hersche et al., which employed Vector Symbolic Architectures (VSA) for reasoning, we delve deeper into making the reasoning mechanism learnable. Central to our exploration is the introduction of the AxR-Learners Framework, which incorporates the Learnable Formula. This innovative model is designed to autonomously adapt, allowing for a more flexible and tailored approach to abstract reasoning tasks. Our findings highlight the efficacy of the Learnable Formula in enhancing machine reasoning, suggesting a promising avenue for the evolution of more sophisticated artificial intelligence systems.