Dr. Jan Hendrik Metzen
Senior Expert in Robust Scalable Perception
I am a Senior Expert in Robust Scalable Perception at Bosch Research. My research is motivated by the desire to advance the understanding, robustness, and applicability of deep learning-based perception. In particular, I focus on evaluating and increasing the robustness of neural networks. While neural networks excel in classifying data coming from the distribution they have been trained on, they often perform poorly on data with slightly different properties. Addressing this shortcoming is crucial for applying neural network-based perception in safety-critical domains such as automated driving, where they will inevitably be faced with situations not encountered or foreseen during training.
Please tell us what fascinates you most about research.
Research enables to obtain a more in-depth understanding of complex systems. For me, intelligence and the ability to learn are among the most exciting phenomena and our research in the area of machine learning enables me to address outstanding topics in this thematic field in a very in-depth manner. If applied responsibly, research results in this area (especially in the industrial environment) can also make a key contribution toward improving the quality of life and safety of many people.
What makes research done at Bosch so special?
Research at Bosch is distinguished by the high degree of freedoms associated with academic research. Regular publication of research results, intensive exchanges with external scientists, and mutual support of dissertations and doctoral theses enable us to further develop the current state of scientific knowledge together with our partners in the academic community. At the same time, close collaboration with Bosch divisions gives us the possibility to swiftly transfer methods developed by us into applications such as driver assistance systems.
What research topics are you currently working on at Bosch?
My research involves understanding and improving the robustness of neural networks for perception tasks (for instance pedestrian detection in the emergency braking assistant). An exact understanding and quantification of robustness are key conditions for the usefulness of a perception system in safety-critical applications such as highly automated driving. Based on this understanding of current limitations, we are researching methods for improving the robustness of neural networks.
What are the biggest scientific challenges in your field of research?
As yet, we still do not know how to make neural networks as robust as human perception. Malicious and practically imperceptible modifications (“adversarial attacks”), and harmless distortions such as background noise or blurred images can result in entirely misleading predictions by neural networks. Such incomprehensible errors are very problematic when it comes to the acceptance and approval of perception systems even if the system has, on average, better detection rates than human users. Our research contributes toward increasing the robustness of neural networks.
How do the results of your research become part of solutions "Invented for life"?
We can only conscionably designate a system as providing technology “Invented for life” where we are certain that it works reliably and safely even under unforeseen conditions. This presents us with major challenges, especially for systems based on machine learning. Our research results contribute toward improving the robustness and understanding of systems based on machine learning. Accordingly, they are key contributions toward reliable and safe systems capable of protecting people from accidents and improving their quality of life.
Curriculum vitae
Since 2019
Senior Expert in Robust Scalable Perception, focusing on increasing the robustness of perception systems based on deep learning and developing automated deep learning techniques, Bosch Research
2016-2019
Senior Expert in Robust Scalable Perception, Bosch Center for Artificial Intelligence
2013-2015
Team leader of the "Sustained Learning" team in the Robotics Innovation Center, focusing on robot learning, Bayesian optimization and supervised learning, German Research Center for Artificial Intelligence (DFKI)
2014
Ph.D. in Hierarchical Reinforcement Learning on "Learning the Structure of Continuous Markov Decision Processes", University of Bremen
Selected publications
Elsken et al. (2023)
- Rohit Mohan, Thomas Elsken, Arber Zela, Jan Hendrik Metzen, Benedikt Staffler, Thomas Brox, Abhinav Valada and Frank Hutter
- International Journal of Computer Vision, 2023
Lovisotto et al. (2022)
- Giulio Lovisotto, Nicole Finnie, Mauricio Munoz, Chaithanya Kumar Mummadi and Jan Hendrik Metzen
- In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern RecognitionRepresentations
Yatsura et al. (2021)
- Maksym Yatsura, Jan Hendrik Metzen, Matthias Hein
- In Advances in Neural Information Processing Systems
Elsken et al. (2019)
- Thomas Elsken, Jan Hendrik Metzen, Frank Hutter
- 7th International Conference on Learning Representations
Get in touch with me
Dr. Jan Hendrik Metzen
Senior Expert in Robust Scalable Perception
