Philipp Czerner

We received the teaching award for the best mandatory course in informatics at TU Munich in the summer term of 2021. The course was “Introduction to Theory of Computer Science”, held by Prof. Javier Esparza and organised by Martin Helfrich and me. Being a mandatory course it was large (~1000 students registered), and it needed to be fully remote. We invested a lot of effort into having a high-quality stream that was as interactive as possible. I am especially fond of the homework sheets, for which I created entirely new exercises set in a fictional world where the students have to help Theo and Dora overcome various problems, and thwart the supervillain Dr. Evilsparza.

Our paper Fast and Succinct Population Protocols for Presburger Arithmetic was recognised by the program committee. In this paper, we manage to construct population protocols that are both succinct (they use only few states relative to the predicate they are deciding) and fast (they stabilise to an output quicly). Our construction is in some sense optimal along both axes, matching known lower bounds. We introduce a new model in which population protocols can be constructed, use the idea of generalised output functions to simplify protocol design, and apply a technically involved potential function analysis. Other innovations include a fast multi-way implementation, a slightly faster output broadcast and an input distribution (population splitting) scheme that cannot fail.

Since then, Nils Harmsen, whom I advised during his BSc Thesis, has implemented our construction. Despite the 50-page paper, the protocols are smaller as soon as the constants in the predicate exceed ~1000, or if the predicate is a boolean combination of 2-3 predicates (as compared to the simplest construction).

The “Förderpreis” of the “Verein der Freunde der TU Clausthal” is given to a small number of outstanding PhD, MSc and BSc Theses each year. I received the award for my BSc thesis, where I built a neural network to classify certain graphs resulting from software repositories. We wanted to capture how realistic such a graph looked. To this end, I used metadata from publicly available Git repositories. To make the scraping more efficient, I re-implemented the relevant parts of the Git protocol, so that I could e.g. only save the metadata and avoid downloading the actual contents. I then converted the metadata into change-coupling graphs, from which I sampled training data (there were some interesting optimisations here). On top of this, I then implemented the neural network. While not going full bare-metal, I did stay reasonably close to the ground, e.g. by using the C++ API of TensorFlow and implementing many important convenience features (snapshotting, training visualisation) myself.

The thesis was advised by Niklas Fiekas and supervised by Prof. Jürgen Dix. I would also like to thank Noura Al-Moubayed for helpful discussions.