All News

“Robust Machine Learning for Critical Applications” funded by Hasler Stiftung

We are very pleased that our project “Managing the trade-off between distributional and adversarial robustness – Reliable machine learning in diabetology and beyond” was recently granted funding by Hasler Stiftung.

The research team led by Prof. Felix Wortmann and Prof. Elgar Fleisch from the ETH Zurich and the University of St. Gallen, together with researchers Ass.-Prof Krikamol Muandet from the CISPA Helmholtz Center for Information Security and PD Dr. med. Thomas Züger from the Kantonsspital Olten successfully obtained a Hasler Stiftung grant to strengthen the foundation of robust machine learning. The overall objective of this project is to provide the theoretical fundament to build machine learning models that can be reliably deployed, especially in critical applications, such as in healthcare.

Research and industry seek promise from machine learning models to enable more applications than ever, tackling some of the grand challenges of humankind. Unfortunately, in the last few years, numerous research papers and practical applications have provided staggering evidence that machine learning is often subject to unexpected drops in performance in real-world deployments, and they are susceptible to strategic manipulation. Two fundamental reasons for this failure are a lack of robustness against distributional shifts between the training and testing data and adversarial attacks.

In the field of deep learning, our novel neural network architectures called Gated Domain Units enable robust deployment on unseen domains, such as during unseen hospitals. For this, we are the first to decode information on a lower-level distributional level to learn the right invariance that enables the generalization to unseen domains. Our project funded by the Hasler Stiftung aims to strengthen further the adversarial robustness of the Gated Domain Units to provide researchers and practitioners with novel methods that increase the robustness of their models’ deployment in critical applications.

2023-02-01T21:29:13+01:00January 4th, 2023|