Model-Driven AI

The limitations of AI methods are well known: they are extremely greedy in terms of data and computation time for training and they lead to solutions that are sometimes not very robust, that present erratic behaviours, in situations, moreover, that are difficult to characterise. One solution to these two limitations is to integrate into these data-driven techniques any a priori knowledge available on the process we are trying to model.

This knowledge can be expressed in a variety of forms, ranging from simple constraints on expected predictions to complex mathematical models to expert intuition on how to break the problem down into subtasks. Conversely, in contexts where models or simulators already exist, machine learning also provides elegant solutions for analysing and exploiting the data generated by these models/simulators. In this context, this work package aims at developing innovative strategies combining AI and models or knowledge, in order to obtain more efficient and robust solutions.

The work package is coordinated by Prof. Pierre Geurts and post-doc Vân Anh Huynh-Thu at ULiège. On this occasion, we will ask researchers to present their research projects related to the “Model-IA Integration” axis. In addition, as most of the major challenges have been activated, this kick-off will also be an opportunity to take stock of the research activities carried out in relation to these challenges since the progress already made at the Paris workshop (see paragraph 5 below). After the kick-off, regular meetings will be planned. The two-month frequency proposed by the other universities will probably also be adopted for this axis.

Currently, 22 researchers affiliated to TRAIL are working on WP3. Among these 22 researchers, 9 PhD students and two post-docs have been engaged in the ARIAC project.

In addition to the coordination of WP3, ULiège has also been active in launching the TRAIL doctoral school (see Section 2).