AI researcher Patrick Koopmann has received funding from the Open Competition of the Dutch Research Council (NWO) for his project PICON – Practical Methods for Concept Interpolation in Realistic Ontologies, a study towards more explainable AI.
Koopmann works with so-called ontologies. These are collections of concepts and definitions that help computers understand information better. You can think of them as dictionaries for computer systems. Such systems are used in areas like medicine, biology, but also for artificial intelligence. Examples include the medical ontology SNOMED CT which is used by health providers all over the world to organize patient data, or the Gene Ontology, which helps researchers organize information about genes. Ontologies can also help robots understand and classify the objects in their surroundings.
The importance of ontologies
One major advantage of ontologies is that they enable computers to draw logical conclusions. If a computer has the right information, it can deduce new information—and explain why it has to be true. This allows for a special type of AI where all decisions are fully transparent and explainable.
The problem
However, there are also challenges. Creating and maintaining ontologies takes a lot of time and must be done manually, which can lead to mistakes. And even though the conclusions drawn by computers are logically correct, they can still be hard for humans to understand.
The solution: Concept interpolation
Koopmann aims to tackle these challenges using a technique called concept interpolation. Simply put: if you know that “x is a spider” and also that “x is not an insect,” a helpful intermediate statement might be: “x has eight legs.”. Such an in-between description is called a concept interpolant.
As the example illustrates, concept interpolants can be used by computers to explain their reasoning (e.g. explain why a spider is not an insect). But they can also help with other tasks that are useful for developing ontologies: they can be used to automatically generate new definitions, or for teaching computers how to come up with logical descriptions on their own. However, computing concept interpolants in practice is challenging, and so far, no automated method that works well in practice has been developed. The project aims to solve this issue, and will bring the first practical algorithms for concept interpolation for different ontology languages, and thus improve the potential of ontologies for many use cases in artificial intelligence and beyond.
About the grant
The NWO Open Competition ENW funds curiosity-driven research across all disciplines in the exact and natural sciences. The so-called M-grants are intended for innovative, fundamental research that is of high quality and/or of urgent scientific importance.