Subsequent projects | Special projects

Start of funding 01.01.2019

Scalable probabilistic inference for mechanistic models

Prof. Dr. Jakob Macke
Technische Universität München
Professorship for Computational Neuroengineering

Ph.D. Scott Lindermann
Stanford University
Department of Statistics


Many systems in engineering, natural and social sciences and economics are best described by high- dimensional and nonlinear dynamical system models that are derived from knowledge about the mechanisms underlying the data-generating process. Inferring the parameters and states of such mechanistic dynamical models from incomplete, noisy observations can be challenging. For example, in neuroprosthetics, one needs to describe the relationship between neural activity in the brain and the motor outputs it produces; this relationship is highly dynamic and must be inferred from noisy electrophysiological measurements. Bayesian inference provides a general framework for inferring the parameters and latent states of dynamical models. We will tackle two central challenges: We will scale up variational inference techniques for nonlinear dynamics to high-dimensional observations, and will make them applicable to models which are implicitly defined through simulators. While our techniques will be applicable more generally, we will specifically explore their utility in the context of neuroprosthetics.