We develop statistical machine learning methods for the analysis of multivariate and multi-modal time series, both at a mathematical and at an algorithmic level. Our focus in this is less on prediction and time series forecasting, but more on a mechanistic understanding of the underlying biological or physical system that generated the observed time series. For this, we take a dynamical systems perspective – for instance, we rely on recurrent neural networks to unravel the computational dynamics that may underlie a set of neurophysiological and behavioral recordings. The approaches we build are interpretable both in terms of their dynamical systems properties (e.g. different attractors) and in relation to the actual measurements. We are also actively involved in data analysis, heavily drawing on our own methods, mostly within the areas of (computational) neuroscience and psychiatry. For instance, we may train recurrent neural networks on simultaneously acquired multiple single unit recordings and behavioral data to gain insight into the computational principles that underlie the performance and neural implementation of cognitive tasks. Our long-term goals here are both to the understand the computational dynamics of the brain, as well as the neurodynamical and neurocomputational mechanisms that go astray in psychiatric conditions.