Predicate Logic Network: Vision Concept Formation

Bang Chen, Maonian Wu, Bo Zheng, Shaojun Zhu, Wei Peng

Abstract: Although deep learning has shown good performance in many fields, it still lacks the most basic human intelligence, which we often called the ability to draw inferences about other cases from one instance. Therefore, how to empower model with logical reasoning ability has received much attention. Thus, we propose neural predicate networks, a model that combines deep learning methods with first-order logic. It converts visual tasks into first-order logic problems by deconstructing them into objects, concepts and relations. Then, achieve first-order logic differentiable by learning logical predicates as neural networks. Finally, the differentiable model can be trained by back propagation to simulate the formation of concepts in the human brain and solve the problem. Experimental results on two image concept classification datasets demonstrate the effectiveness and advantages of our approach.