Séminaire du 10 décembre 2024 à 10h30 dans la salle 204, bâtiment 11 sur le campus de La Gaillarde
Recent Advances in Scalable Concept Learning par Axel-Cyrille NGONGA NGOMO l'Université de Paderborn (http://dice-research.org)
Résumé : Concept learning (CL) exploits background knowledge to compute ante-hoc explainable machine learning models in the form of description logic concepts. While CL is well suited for data on the Semantic Web in theory, it was only led to scale up to the requirements of real applications in the recent past. In this talk, we give insights into some of those results through which CL can now be deployed to large-scale knowledge graphs. We focus on recent results in the areas of querying, embeddings, and neural reasoning, which decrease the runtime of CL by orders of magnitude. In particular, we present an update to the OWL to SPARQL bridge that facilitates significantly lower runtimes for instance retrieval. Our unified theory of multiplicative embeddings allows for vector space selection is the second pillar of the talk. We also present how this theory can be exploited for neural reasoning, the third pillar of the talk. We conclude with potential next steps towards robust CL.