Paper ID: 2403.03190

Triple-CFN: Restructuring Concept and Feature Spaces for Enhancing Abstract Reasoning Process

Ruizhuo Song, Beiming Yuan

Visual abstract reasoning poses challenges to AI algorithms, requiring cognitive abilities beyond perception. For methodology, this study emphasizes the need to separately extract concepts and features from visual abstract reasoning problems, employing the responses of features to concepts as elements in the reasoning process. It also advocates for clear concept and feature spaces to tackle visual abstract reasoning tasks effectively. For technology, we introduce the Cross-Feature Network (CFN), a framework that separately extracts concepts and features from reasoning problems, utilizing their responses as reasoning representations. The CFN integrates a dual Expectation-Maximization process to actively seek an ideal concept space for problem-solving, yielding notable results despite limitations in generalization tasks. To overcome these limitations, we propose the Triple-CFN, maximizing feature extraction and demonstrating effectiveness in Bongard-Logo and Raven's Progressive Matrices (RPM) problems. Additionally, we present Meta Triple-CFN, which constructs a promising concept space for RPM, ensuring high reasoning accuracy and concept interpretability. Furthermore, we design the Re-space layer, defining a clear feature space for (Meta) Triple-CFN, with its unique warm-start process aiding generalization. Overall, this work advances machine intelligence through innovative network designs for abstract reasoning.

Submitted: Mar 5, 2024