Knowledge-based AI approaches are characterized by the fact that knowledge (e.g. expert knowledge) can be stored, generated, used and queried explicitly and in a comprehensible way. Machine learning approaches, on the other hand, implicitly derive relationships "on the fly" from the data presented to them. They usually require significant computer capacities and massive amounts of high-quality and sufficiently task-specific data. Hybrid AI approaches that combine knowledge-based and data-driven AI contribute to the improved applicability of machine learning processes and, last but not least, to their interpretability.
On this basis, we develop IT solutions that make it possible to provide human operators in networked systems with relevant information and knowledge at the right time, at the right level and according to their specific task. Essential for this are not only suitable procedures for information acquisition and exploitation, but also adequately designed human-machine interfaces, in particular for the realization of interactive assistance systems.
In the EU projects MAGNETO and PREVISION, an ontology is being developed for knowledge representation, in which all relevant elements of a police investigation are integrated and represented. In the knowledge base that is based on this ontology, the results of text and data mining are stored. By means of semantic information processing, information fusion and reasoning, they are linked together, implicit relationships between the individual pieces of information are uncovered and presented to the investigator in a clear form.
At the Fraunhofer Research Center for Machine Learning, intensive research is being conducted into the topic of "Informed Machine Learning", which aims at systematically integrating expert knowledge into statistical learning processes. It is precisely this integration that enables users to understand decisions made by models that otherwise function as black boxes and to intervene in the models if necessary. In the Fraunhofer lighthouse project Machine Learning for Production - ML4P a process model for the integration of machine learning in production processes is being developed. One focus in this project is the integration of the domain knowledge of the machine operators into the learning processes.