Toward the Improvement of Probabilistic Classifiers Using Ontologies.

Incorporating ontological knowledge into supervised machine learning pipelines is a promising avenue of research aimed at enhancing classification performance. Specifically, when probabilistic ontologies are accessible, it is plausible to combine the corresponding probabilities with the outputs of a probabilistic classifier. This integration can enhance the classification process by incorporating the probabilistic information from the ontologies alongside the classifier outputs. This work investigates the gain given by using ontologies in an image classification task. We start from the hypothesis that context information of the elements present in an image can lead the classifier toward the correct class. An implementation of this integration is proposed and its effectiveness is demonstrated, especially in situations where the classifier accuracy is relatively low. The initial experimental results support the efficacy of this approach. To assess the impact of the ontology, performance was first measured on simulated classifiers on a subset of MIT-Indoor. Then, as a real case, we ran experiments using a Neural Network pre-trained on the ImageNet dataset.