Class disjointness constraints as specific objective functions in neural network classifiers

摘要

Increasing performance of deep learning techniques on computer vision tasks like object detection has led to systems able to detect a large number of classes of objects. Most deep learning models use simple unstructured labels and assume that any domain knowledge will be learned from the data. However when the domain is complex and the data limited, it may be useful to use domain knowledge encoded in an ontology to guide the learning process. In this paper, we conduct experiments to introduce constraints into the training process of a neural network. We show that explicitly modeling a disjointness axiom between a set of classes as a specific objective function leads to reducing violations for this constraint, while also reducing the overall classification error. This opens a way to import domain knowledge modeled in an ontology into a deep learning process.