We use machine learning, high-dimensional statistics, physics driven modeling and data visualization to develop algorithms to analyze large and complex datasets in neuroscience and ophthalmology. In the retina, we aim to understand the link between the computations performed by individual cell types, the circuits they are part of and the underlying genetics and anatomy. Also, we are interested in developing interpretable, uncertainty-aware algorithms for improving clinical decision making. We collaborate closely with our experimental and clinical partners to answer fundamental questions about how visual information is processed and how degenerative diseases of the eye can be detected early.

(Group leader: Philipp Berens)

We develop new machine learning algorithms for analyzing neurons and neural circuits based on large scale functional two-photon imaging data, high-resolution anatomical reconstructions, and single-cell transcriptomes. We are interested in developing a new class of tractable models of an intermediate complexity, between traditional biophysically realistic models and purely statistical ones. For learning the model parameters from data, we resort to the latest developments in machine learning such as likelihood-free inference. We believe that such hybrid models will help us reach an integrative understanding of retinal circuits across scales, and help integrating complex datasets from disparate sources.

(Group leader: Lisa Koch)

We work on machine learning in ophthalmic imaging for safe medical diagnostics. We are particularly interested in critical aspects for clinical acceptance of machine learning, such as interpretability, uncertainty quantification, robustness, and generalisation to real world settings. Methodologically, our research focuses on Bayesian deep learning and generative modelling. Our research is mostly applied to ophthalmic image modalities such as fundus photography and optical coherence tomography. We collaborate closely with the university hospital, and ultimately aim to improve clinical decision making in the context of major causes of vision impairments such as age-related macular degeneration and diabetic retinopathy.

(Group leader: Dmitry Kobak)

We are interested in dimensionality reduction and data visualization of transcriptomic and multi-omic data. We work on developing neighbor embedding methods that yield a biologically meaningful two-dimensional representation and preserve not only local but also global structure of the data. We study implicit trade-offs between existing neighbor embedding algorithms such as t-SNE and its variants. We also develop methods for dimensionality reduction of multi-omic datasets, such as e.g. Patch-seq data combining transcriptomic and electrophysiological recordings.

We believe that the best way to achieve reproducible scientific results is to promote open science, including sharing data and making software for scientific research freely available. Therefore, the code for our publications as well as associated data sets are openly available, where possible (see our github.com/berenslab/). Also, we are interested in the ethical ramifications of our work and collaborate with philosophers in this regard.