CIBSS-D

Data Efficient Deep Learning

Prof. Dr. Thomas Brox (CIBSS-PI), Department of Computer Science (Faculty of Engineering), University of Freiburg

CIBSS Profile page

 

Computational quantification in imaging data has gained importance for biological studies in recent years, as data sets get larger and manual annotation becomes more and more tedious. Especially the availability of recent deep learning techniques promise to be applicable in a wide variety of scenarios. A major challenge to make deep learning practically applicable is the amount of annotated data to teach the computer the quantification task. Only if this amount is much lower than the data to be processed in total, the use of computational methods is beneficial.

The project will advance learning-based techniques to make the best use of available data and, thus, to reduce the required amount of data to set up a new experiment. We will explore semi- and self-supervised learning strategies to better leverage unlabelled data than with the techniques available today. The project will also work on the problem of domain transfer, i.e., the efficient transfer of models learned on data from one data domain to another data domain. Techniques for automated parameter search will allow the automated adaptation of the learning technology and will further increase the usability for scientists without a machine learning background.

Deep learning is also well-suited to integrate images or image sequences with additional data from other modalities, e.g.  meta-data like text or sequencing data. Especially in conjunction with generative learning approaches, this will lay the ground for techniques that can consume multiple types of available data and generate various visualizations of this data.