Matthias Hein

Timeline

1. 2002–2007

   A PhD in Schölkopf's group

   After studying physics in Tübingen, he works in Bernhard Schölkopf's group at the Max Planck Institute for Biological Cybernetics from 2002 to 2007. At TU Darmstadt, he earns a Dr. rer. nat. in 2005 with the dissertation "Geometrical Aspects of Statistical Learning Theory", advised by Thomas Hofmann and Bernhard Schölkopf.

2. 2007–2011

   Professor at Saarland University

   He becomes a professor of computer science at Saarland University in 2007, where he heads the machine learning research group and serves as a principal investigator of the Cluster of Excellence on Multimodal Computing and Interaction. In 2011, the German Association for Pattern Recognition (DAGM) gives him the German Pattern Recognition Award for his work on graph-based learning methods, in particular manifold learning and nonlinear eigenproblems.

3. July 2012

   An ERC Starting Grant

   In July 2012, the European Research Council awards him an ERC Starting Grant worth 1.27 million euros over five years. The grant supports his group's use of nonlinear eigenproblems for the analysis of complex data, an approach he says still works when data have been manipulated or quantified incorrectly.

4. June 2018

   A Bosch-endowed chair

   In June 2018, Bosch announces that Hein, aged 42, joins the University of Tübingen in a Bosch-endowed professorship, which the company supports with 5.5 million euros over ten years as part of Cyber Valley. His research covers statistical learning for image processing and genetics, with a focus on reliable and comprehensible learning processes.

5. 2018–2019

   Overconfident ReLU networks

   With Maksym Andriushchenko and Julian Bitterwolf, he posts a paper in December 2018 showing that ReLU networks almost always produce high-confidence predictions far away from the training data. The paper proposes a robust optimization technique, similar to adversarial training, that enforces low confidence there, and is accepted with an oral presentation at CVPR 2019.

6. 2020

   Attack ensemble and RobustBench

   In an ICML 2020 paper with Francesco Croce, he combines two new extensions of the PGD attack with two existing attacks into a parameter-free ensemble for testing adversarial robustness. Applied to over 50 published models, the ensemble finds lower robust accuracy than reported in all but one case. In 2020, he also coauthors RobustBench, a standardized adversarial robustness benchmark that evaluates models with AutoAttack.

7. 2026

   Co-founding Exponential Security Labs

   In April 2026, he becomes involved in a start-up for AI security, especially of agent systems. In June 2026, the ELLIS Institute Tübingen announces Exponential Security Labs, which he co-founds with Maksym Andriushchenko, Naman Deep Singh and Anselm Paulus to secure agentic systems through self-improving red-teaming and guardrail agents. Commenting for the Science Media Center Germany in June 2026, he says every large language model is vulnerable to jailbreaks.

8. As of October 2026

   Professor at Tübingen

   He is Bosch endowed Professor of Machine Learning at the University of Tübingen, coordinates its international master program in machine learning, and belongs to the Tübingen AI Center and the Excellence Cluster "Machine Learning: New Perspectives for Science". An ELLIS Fellow in the ELLIS Unit Tübingen, he works on making machine learning systems robust, safe and explainable. A project by him and Dr. Geiping in Schmidt Sciences' Science of Trustworthy AI program investigates how and why AI systems become less safe when they operate for extended periods of time.