Domain-Specific Language for Transfer Learning / Uncertainty Quantification

Recent breakthroughs in biomedical image analysis have been based on the application of advanced machine learning for large imaging data sets. For instance, convolutional neural network-based classifiers have recently been shown to reach or even exceed human-level performance in a range of disease- and abnormality detection tasks, based on tens of thousands of images. However, such machine learning approaches often fail when confronted with changes to the input data, e.g. images from a different MRI machine, or patients of a different age group or ethnicity. Uncertainty quantification and transfer learning are used to identify and alleviate such problems.

In this project, we are looking for a student to design a domain specific language (DSL) for uncertainty quantification and transfer learning on biomedical images based on our existing code base in python. Specifically, the task is to identify commonalities between neuroimaging and microscopy data analysis, and design and build a framework to evaluate state-of-the-art methods for uncertainty quantification and transfer learning in both analysis settings.

The thesis will be co-supervised by Prof. Dr. Kerstin Ritter, Assistant Professor for Computational Neuroscience at Charité - Universitätsmedizin Berlin.