Where are you from, and what is your background?
I grew up in the Palatinate region in southwest Germany, a quaint region mostly popular among tourists and wine enthusiasts. Afterwards, I studied a bachelor’s and master’s in Industrial Engineering and Management at the Karlsruhe Institute of Technology. During my bachelor’s degree I was mostly interested in optimisation, so for my final thesis I looked at how a company I worked with could best solve their machine scheduling problems. During my master’s my interests shifted more towards machine learning, so I ended up looking at Siamese Neural Networks for classification tasks as part of my master’s thesis.
What do you do in your spare time?
I really enjoy being outside and used to go mountain biking a lot, which is unfortunately quite difficult in London. Here, I’ve had to substitute my mountain bike for a nice folding bike which is quite convenient for commuting to uni and it also means I still get to cycle around a lot. Besides that, I enjoy playing tennis (and also watching some matches - being that close to Wimbledon does have its advantages!). After a few rainy weeks sailing the Baltic Sea I also obtained a sailing license in 2021 and quite enjoy spending some time on the water. When sailing around the Canary Islands last summer I definitely got to see some locations that wouldn’t have been accessible without a boat.
What influenced you to do a PhD?
I really enjoyed the research I did while completing my degrees! That means both of the theses I wrote, but I also worked as a student research assistant in the operations research area. I didn’t know much about AI Safety and verification before discovering the SAIL (formerly VAS) group, but when reading some papers I found the way in which research in this area combines optimisation concepts to formally verify the safety of AI models very interesting!
What are your research interests?
I am currently working on both neural network verification and robustification. In the verification domain I am focusing on extending verification tools to new specifications and model architectures. A recent paper that I worked on introduced methods for verifying the robustness of a model against convolutional perturbations such as motion blur. Recently, I have shifted to focus on robustness verification for object detection though. Since models trained with standard training methods are often vulnerable to attacks and difficult to verify, I am also working on methods for training robust networks. That work mostly focuses on training models that are robust to a wider range of perturbations beyond noise while retaining their standard accuracy.