Model-regularized Learning of Complex Dynamical Behavior
This project is planned to couple machine learning approaches, especially from the field of Deep Learning, with (reduced) ODE models in the sense that the model becomes an integral part of the learning iteration. In this way, the training of the deep network can build on the available - but possibly small and incomplete - data, but is additionally regularised by the relevant physics. For many scenarios, these reduced (or coarsened) models are available.
Although they are significantly less complex and often based on only a few basic structural properties, they still contain the basic physics or structure of the problem. The developed methods will then be used to analyze real-world data, e.g., from opinion formation in social networks, the dynamics of molecules, or single-cell analysis.
Research from the very top
BIFOLD scientist Prof. Dr. Begüm Demir explains the benefits of using Explainable Artificial Intelligence (XAI) to enhance earth observation satellite data analysis in the Tagesspiegel TU Berlin supplement.
Fellow
Fellow
Postdoctoral Researcher