Title:Triple-CFN: Restructuring Conceptual Spaces for Enhancing Abstract Reasoning process

Abstract:Abstract reasoning poses significant challenges to artificial intelligence algorithms, demanding a higher level of cognitive ability than that required for perceptual tasks. In this study, we introduce the Triple-CFN method to tackle the Bongard Logo problem, achieving remarkable reasoning accuracy by implicitly reorganizing the conflicting concept spaces of instances. Furthermore, with necessary modifications, the Triple-CFN paradigm has also proven effective on the RPM (Raven's Progressive Matrices) problem, yielding competitive results. To further enhance Triple-CFN's performance on the RPM problem, we have upgraded it to the Meta Triple-CFN network, which explicitly constructs the concept space of RPM problems, ensuring high reasoning accuracy while achieving conceptual interpretability. The success of Meta Triple-CFN can be attributed to its paradigm of modeling the concept space, which is tantamount to normalizing reasoning information. Based on this idea, we have introduced the Re-space layer, boosting the performance of both Meta Triple-CFN and Triple-CFN. This paper aims to contribute to the advancement of machine intelligence and pave the way for further breakthroughs in this field by exploring innovative network designs for solving abstract reasoning problems.