Abductive subconcept learning

Zhongyi Han

doi:10.1007/s11432-020-3569-0

Abductive subconcept learning

2023-01-12

0

PRE

AI

DOI：10.1007/s11432-020-3569-0

Original

Original request for help

Save

Abstract

En 中文

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.

Keywords:

abductive learning

subconcept learning

logical reasoning

subconcept set selection

meta-interpretive learning

Abductive subconcept learning

Abstract

Journal

Organization