Some medical imaging procedures are inherently noisy processes. When your target quantities are defined through this data, areas of high noise should produce higher model uncertainty. In this work, we show how we can communicate the model uncertainty to the underlying numerical algorithm, thus yielding more realistic results off the achieved accuracy.
Publication: Soren Hauberg, Michael Schober, Matthew Liptrot, Philipp Hennig, Aasa Feragen: A Random Riemannian Metric for Probabilistic Shortest-Path Tractography . In: Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015.
We show that some Gaussian process ordinary differential equation solvers have the exact same mean as other well-known ODE solvers. This recovers all theoretical guarantees known for the traditional solvers while putting them in a probabilistic context.
Publication: Michael Schober, David Duvenaud, Philipp Hennig: Probabilistic ODE Solvers with Runge-Kutta Means. In: Advances of Neural Information Processing Systems (NIPS), 2014. Selected for oral presentation.
We use Gaussian process ODE solvers to quantify uncertainty in the approximation quality of ODE solvers. Thus, we can visualize the approximation quality and incorporate the solution quality in downstream computations.
Publication: Michael Schober, Niklas Kasenburg, Aasa Feragen, Philipp Hennig, Soren Hauberg: Probabilistic shortest path tractography in DTI using Gaussian Process ODE solvers. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2014.
There is lots of noise in astronomical images due to long exposure times and low-light settings. I test whether image denoising can be improved by only considering this special case against state-of-the-art algorithms for the general case of any image with i.i.d. Gaussian noise.
Publication: Michael Schober: Camera-specific Image Denoising. Diploma thesis, 2014.
Using training labeled data, I use machine learning algorithms to learn pairwise similarities between instances from a user-specified metric over their labels. Thus, new instances can be compared to labeled instances and predictions of this algorithm are more interprative.
Publication: Michael Schober: Using label metrics for Distance Metric Learning. Research thesis, 2009.
A short collection of tips and tricks to design your first conference poster.
An article I have written for a student magazine outlining basic ideas of causal inference. Unfortunately, only available in German.