Welcome! This is the lab webpage of the theoretical neuroscience group at the Central Institute of Mental Health in Mannheim.

We follow four major lines of research:

**Statistical Machine Learning & Dynamical Systems**

We develop novel machine learning approaches for data analysis from a theory-driven statistical perspective. Our focus is on the analysis of high-dimensional and multimodal time series, such as brain signals or mobile data, with the major goal of identifying the underlying dynamical system that generated the observed time series.

Our major methodological framework for modeling and predicting such time series, and integrating them with other data, are deep generative recurrent neural networks (RNN; e.g. Koppe et al. 2019, PLoS Comp Biol).

Our method aims at reconstructing complex nonlinear dynamical systems with multiple time scales, using a range of training algorithms.

We have also developed novel machine learning tools and statistical tests for unsupervised spatio-temporal pattern detection in non-stationary environments (e.g. Russo & Durstewitz 2017, eLife), which we apply, for instance, to simultaneous electrophysiological recordings from tens to hundreds of neurons for decoding cognitive processes from neural activity.

**Computational Psychiatry & Biomedical Data Analysis**

We apply our innovative methodology for diagnostic and prognostic (predictive) purposes in psychiatry and neurology, but also to gain insight into underlying disease mechanisms. For instance, based on generative RNNs we can infer dynamical models of an individual’s brain from functional magnetic resonance imaging (fMRI), electroencephalographic (EEG), or magnetoencephalographic (MEG) measurements (e.g. Durstewitz et al. 2019, Molecular Psychiatry), or dynamical models of an individual’s behavior from mobile sensors and smartphone data (e.g. Koppe et al. 2018, Schiz. Bulletin), for which we develop algorithms for the integration of multiple data modalities (Kramer et al., 2022).

Such RNN-based dynamical systems models, trained on individual subjects, can then be used to forecast future behavioral or neuronal trajectories to enable early intervention. They can also be simulated to study the effect of potential pharmacological or behavioral treatments. In addition, we use formal reinforcement learning models of behavior to study learning processes in psychiatric conditions (e.g. Koppe et al. 2017 PLoS Biol).

**Mathematical Tractability**

We further aim at advancing our mathematical understanding of the reconstructed models, e.g. by investigating how ReLU-based recurrent neural networks can be transformed from discrete-time to continuous-time systems (Monfared&Durstewitz, 2020), how we can reduce the dimensionality of the reconstructed systems (Brenner et al. 2022), or how we can guarantee analytical access to important dynamical systems properties such as fixed points and limit cycles.

**Mathematical Models & Computational Dynamics of the Brain**

We also develop mathematical models of brain function at a more biophysical level (e.g., Hass et al. 2016, PLoS Comp Biol), as well as statistical approaches for inferring such models directly from experimental observations like multi-cell recordings or neuroimaging data (e.g. Hertäg et al. 2012, Front Comp Neurosci). These strongly data-driven models are then used to gain insight into the neuro-dynamical and neuro-computational processes underlying cognitive function, and in how these are altered in psychiatric conditions. Such models can be harvested to derive novel diagnostic and prognostic criteria for psychiatric and neurological conditions, and for devising more effective and individualized treatment options.