City Research Online

Abstract

This paper introduces and evaluates a neural-symbolic cycle for Convolutional Neural Networks (CNNs). Knowledge in the form of logic programming rules is extracted from a trained (teacher) CNN model. Domain experts can interact with the rules to assign concepts, intervene and make changes to the model. Distillation is then used to re-train a simplified CNN, closing the neural-symbolic cycle. The approach is evaluated in the classification of medical images (chest x-rays). Experiments indicate that the approach can generate symbolic rules for pleural effusion detection with high accuracy (94.5%) and fidelity (98.2%) in comparison with the original CNN with 96.2% accuracy. Expert intervention produces symbolic rules with clinically relevant concepts while preserving predictive accuracy (94.8%). The approach also enables effective transfer of learning from clinically-relevant rules onto a much simplified (student) CNN that is almost 90% more compact while maintaining accuracy of 93.8%. The goal of this work is to offer an auditable record of network training, elaboration and deployment in the medical domain.

Publication Type:	Conference or Workshop Item (Paper)
Additional Information:	© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0)
Publisher Keywords:	Neural-Symbolic System, Knowledge Extraction, Symbolic Reasoning, Human-in-the-loop, Knowledge Distillation
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science R Medicine > RC Internal medicine > RC0321 Neuroscience. Biological psychiatry. Neuropsychiatry
Departments:	School of Science & Technology > Computer Science

Export

Downloads