The research department of Neuro-mechanistic Modeling focuses on the question of how to combine established knowledge with data-driven learning AI methods. Since such knowledge is often in the form of so-called mechanistic models, the research department pursues hybrid approaches in which mechanistic and neural-based models complement each other. The associated research projects will develop solutions to problems that require both the high model complexity of neural networks and the integration of mechanistic descriptions, or that particularly benefit from the combined advantages. In contrast to purely neural models, neuro-mechanistic models not only allow the integration of domain knowledge but can also achieve excellent results when only moderate amounts of data are available – which is often the case in life sciences, for example. In addition, they are also easier to interpret and generalize to unknown inputs.
"In the new research department, we would like to extend the potential of AI-based methods to areas where the amount of data is insufficient for existing AI techniques. Often the collection of data is very complex and expensive or even impossible. There, we can provide new computational solutions by combining AI and traditional mathematical modeling. Integrating existing domain knowledge delivers very good results even with a small amount of data and is easier to interpret," explains Prof. Dr. Verena Wolf.
Prof. Dr. Antonio Krüger, CEO and head of the research department Cognitive Assistants, adds: "DFKI is increasingly focusing on hybrid AI systems that combine methods based on explicit knowledge with statistical methods such as classical machine learning and deep learning. The neuro-mechanistic approaches are a very important building block for this and strengthen DFKI's expertise in the more general area of neuro-explicit AI."
Models of data-driven learning, especially artificial neural networks, can recognize very complex data patterns automatically and in real-time and can be used to solve regression or classification problems. However, the training of neural models requires vast amounts of data in which world knowledge is only implicitly contained, so that the user can explicitly comprehend the situational adequacy of the results.
On the other hand, mathematical models for computer-aided simulation play a fundamental role in research and the technological progress based on them because simulations make it possible to discover mechanistic relationships and regularities. Still, the applicability of the findings from such simulations is limited because one has to make many strongly simplifying assumptions in the models.
A hybrid approach combines the advantages and delivers good results not only for problems in the life sciences but also for complex artificial systems such as dynamic and highly flexible Industry 4.0 manufacturing processes.