Abductive subconcept learning  

作　　者：Zhongyi HAN Le-Wen CAI Wang-Zhou DAI Yu-Xuan HUANG Benzheng WEI Wei WANG Yilong YIN 

机构地区：[1]School of Software,Shandong University,Jinan 250101,China [2]National Key Laboratory for Novel Software Technology,Nanjing University,Nanjing 210023,China [3]Department of Computing,Imperial College London,London SW72AZ,UK [4]Center for Medical Artificial Intelligence,Shandong University of Traditional Chinese Medicine,Qingdao 266112,China

出　　处：《Science China(Information Sciences)》2023年第2期110-122,共13页中国科学（信息科学）（英文版）

基　　金：supported in part by National Natural Science Foundation of China (Grant Nos. 62176139, 61872225, 61876098);Major Basic Research Project of Natural Science Foundation of Shandong Province (Grant No. ZR2021ZD15)。

摘　　要：Bridging neural network learning and symbolic reasoning is crucial for strong AI. Few pioneering studies have made some progress on logical reasoning tasks that require partitioned inputs of instances(e.g.,sequential data), from which a final concept is formed based on the complex(perhaps logical) relationships between them. However, they cannot apply to low-level cognitive tasks that require unpartitioned inputs(e.g., raw images), such as object recognition and text classification. In this paper, we propose abductive subconcept learning(ASL) to bridge neural network learning and symbolic reasoning on unsegmented image classification tasks. ASL uses deep learning and abductive logical reasoning to jointly learn subconcept perception and secondary reasoning. Specifically, it first employs meta-interpretive learning(MIL) to induce first-order logical hypotheses capturing the relationships between the high-level subconcepts that account for the target concept. Then, it uses the groundings of the logical hypotheses as labels to train a deep learning model for identifying the subconcepts from unpartitioned data. ASL jointly trains the deep learning model and learns the MIL theory by minimizing the inconsistency between their grounded outputs. Experimental results show that ASL successfully integrates machine learning and logical reasoning with accurate and interpretable results in several object recognition tasks.

关 键 词：abductive learning subconcept learning logical reasoning subconcept set selection metainterpretive learning 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]