Systematic Abductive Reasoning via Diverse Relation Representations in Vector-Symbolic Architecture.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Computer Vision

Background:

• Monolithic deep learning models lack interpretability and generalization for abstract visual reasoning.
• Existing neuro-symbolic methods struggle with diverse attribute and relation representations.

Purpose of the Study:

• To propose a systematic abductive reasoning model (Rel-SAR) for Raven's progressive matrices (RPM).
• To enhance attribute representation and systematic reasoning capabilities in AI models.

Main Methods:

• Developed a vector-symbolic architecture (VSA) incorporating diverse relation representations.
• Introduced high-dimensional (HD) encodings and structured HD representation (SHDR) for attributes.
• Proposed novel numerical and logical relation functions for rule abduction and execution.

Main Results:

• Rel-SAR demonstrated significant performance improvements on RPM tasks.
• The model effectively combines HD attribute representations with symbolic reasoning.
• Achieved systematic abductive reasoning with interpretable and computable semantics.

Conclusions:

• Rel-SAR offers a promising approach to overcome limitations in current AI reasoning models.
• The synergy between HD representations and symbolic reasoning enhances AI's ability for complex tasks.
• This work advances the field of abstract visual reasoning and neuro-symbolic AI.