We use Machine Learning methods to uncover patterns in high-dimensional amplicon- and metagenome-based microbiome data. This helps us to understand microbial communities and their ecosystem.

We try to develop deep neural network models that outperform tree-based models for tabular data. Specifically, we think about how categorical features can be handled best in neural networks.

We are developing a domain adaptation algorithm for Human Activity Recognition, where a deep learning model trained on one dataset (source) is applied to a different dataset (target) with distinct characteristics.

These models make use of prior domain knowledge and thus need less training data than purely data-driven models. Also, we can guarantee that certain constraints hold.

These probabilistic models come with guarantees on the computational complexity of inference operations. This is really useful if you want to answer many queries about the same distribution.

Probabilistic activity models of dynamic systems often have symmetries. Lifted Marginal Filtering makes use of them to make inference in dynamic systems much more efficient.