ADIMEM

Automatische Daten-Getriebene Inferenz in Mechanistischen Modellen

01/08/2018 to 31/07/2021

Machine learning makes it possible to recognize patterns in complex data. Currently popular methods are based on generic, unstructured models whose inherent assumptions can have unexpected effects. An alternative are classical physical models, which represent the dynamics and structure of a system mechanically with the help of expert knowledge. Such models are more interpretable, but it is often difficult to adapt their parameters to empirical data. The aim of the project is to combine the efficiency of machine learning with the interpretability of classical modeling. Our methods will combine complex mechanistic models and simulations with high-dimensional data, and provide insights into the underlying processes. The planned approach is flexibly applicable to a variety of models in science, business and technology, and requires only limited prior knowledge from users. The project focuses on two exemplary questions of biophysical modelling of biological neural networks: Modelling of individual cells of the retina and simulations of the cerebral cortex. This specification enables direct evaluation of the methodology and combines technological development with scientific progress. It is intended to develop publicly available software that enables users to automatically infect their own models and simulations without implementation hurdles.