DeduDeep: An Extensible Framework for Combining Deep Learning and ASP-Based Models

Pierangela Bruno, Francesco Calimeri, Cinzia Marte

Abstract: In the last decades, Deep Learning (DL)-based approaches have been fruitfully employed in many tasks, such as providing valuable support to computer-aided diagnosis and medicine. However, DL-based approaches are known to suffer from some limitations; for instance, they lack of proper means for providing clear explanations and interpretations of the results, or explicitly including available knowledge to drive decisions. In this work, we present DeduDeep, the prototypical implementation of a framework explicitly conceived with the aim of tackling such limitations by making use of deductive declarative formalisms. In particular, the framework aims at enabling the declarative encoding of explicit knowledge, and, by relying on the use of Answer Set Programming (ASP), taking advantage of it for driving decisions taken by neural networks and refining the output. The framework has been tested using different artificial neural networks tailored to semantic segmentation tasks over Laryngeal Endoscopic Images and Freiburg Sitting People Images.