HDS-LEE: Helmholtz School for Data Science

HDS-LEE Homepage
Contact: Ira Assent

The Helmholtz School for Data Science in Life, Earth and Energy (HDS-LEE) provides an interdisciplinary environment for educating the next generation of data scientists in close contact to domain-specific knowledge and research.

IAS-8 is involved via a PhD project on transfer learning methods for computational medicine. Transfer learning has recently attracted research interests in the machine learning community as a technique that benefits data-hungry analysis tasks with related but differing data sources in imaging, natural language processing and more. Examples include computational medicine where we see huge potential in scaling data analysis, but where manual annotation by medical experts is often prohibitive.

In this project we contribute a characterization of data and model properties that benefit learning tasks in computational medicine, and support the identification and integration of data from publicly available experimental databases. We propose transfer learning models that optimize target predictions while reducing the manual effort in curating training data.

We devise efficient learning algorithms that allow incremental addition of data sources, and that make use of available hardware resources.

We create explanations of transfer learning models that provide the domain experts with an understanding of the models and their role in transfer learning, allowing them to scrutinize the model and ensure that conclusions based on the model are reliable.

Last Modified: 25.03.2022