Abstract reasoning poses significant challenges to artificial intelligence algorithms, demanding a cognitive ability beyond that required for perceptual tasks. In this study, we introduce the Cross-Feature Network (CFN), a novel framework designed to separately extract concepts and features from images. This framework utilizes the responses of features to concepts as representations for reasoning, particularly in addressing the Bongard-Logo problem. By integrating an Expectation-Maximization process between the extracted concepts and features within the CFN, we have achieved notable results, albeit with certain limitations. To overcome these limitations, we propose the Triple-CFN, an efficient model that maximizes feature extraction from images and demonstrates effectiveness in both the Bongard-Logo and Raven's Progressive Matrices (RPM) problems. Furthermore, we introduce Meta Triple-CFN, an advanced version of Triple-CFN, which explicitly constructs a concept space tailored for RPM problems. This ensures high accuracy of reasoning and interpretability of the concepts involved. Overall, this work explores innovative network designs for abstract reasoning, thereby advancing the frontiers of machine intelligence.

翻译：抽象推理对人工智能算法构成了重大挑战，要求具备超越感知任务的认知能力。本研究提出了交叉特征网络（Cross-Feature Network, CFN），这是一个新颖的框架，旨在从图像中分别提取概念与特征。该框架利用特征对概念的响应作为推理表征，特别适用于解决Bongard-Logo问题。通过在CFN中整合概念与特征间的期望最大化（Expectation-Maximization）过程，我们取得了显著成果，但仍存在一定局限性。为克服这些局限，我们提出了三重概念特征网络（Triple-CFN）——一种高效模型，能够最大化从图像中提取特征，并在Bongard-Logo问题和Raven渐进矩阵（RPM）问题上均展现出有效性。此外，我们引入了Meta三重概念特征网络（Meta Triple-CFN），这是Triple-CFN的进阶版本，其显式构建了针对RPM问题的概念空间，确保了推理的高精度与所涉及概念的可解释性。整体而言，本研究探索了面向抽象推理的创新网络设计，从而推动了机器智能的前沿发展。