Lourens Touwen

PhD candidate at TU Delft in the Scalable Graph Learning group with Dr. Kubilay Atasu. Working on graph neural networks in a federated setting.

Contact and socials

Plan a meeting with me on Cal.com.

X — Google Scholar — GitHub — LinkedIn — touwenlourens[at]gmail.com

Research

Paper: Learning the mechanisms of network growth, L. Touwen, D. Bucur, R. van der Hofstad, A. Garavaglia, N. Litvak, Scientific Reports, 2024 — Nature.com

MSc thesis: Triangle Count reveals Communities and improves Graph Neural Networks — Thesis

• We show that triangle count can be used to reveal communities in graphs, and that this information can be used to improve the power of GNNs on whole-graph tasks.
• I visited Toronto Metropolitan University for a research stay with Prof. Pawel Pralat.

Writing

Substack - Re:LT

Last (and only) post: Observations in China, on tech diffusion and more. Subscribe to get posts right in your inbox.

Excerpts

Some projects

• Scanned PDF To EPUB with GPT — Repository
  • Script that converts a scanned PDF to EPUB, using GPT-4o for the character recognition.
  • This is expensive (~1$/100 pages), but works significantly better than "dumb" OCR. The key idea is to also provide metadata in the prompt, deduced from earlier pages (title, author, chapter title, last paragraph). This way GPT-4o can infer the correct formatting of the given image.
• Data-mined political compass of the Dutch congressional chamber — Repository
• From Scribble To Readable — Webshop
  • A notebook containing the guide to writing mathematical symbols like you're in kindergarten. This was my COVID summer project with Tristan Trouwen. Fun article about the whole thing.
• GPT as Markov Chain — Repository
  • Can we model GPT as a Markov Chain, and obtain probability bounds on reaching "naughty" states?
  • The answer is yes, but the dimension of the state space is too large to be of practical use.
• Team HART — Website
  • University and industry funded team of ~18 students. We developed a sleeve that can give haptic feedback representing any type of information. The user learns to interpret these vibrations, effectively learning a new sense. Our first application allowed the user to understand a foreign language, in real time. I joined HART as the first generation in early 2021. Some media highlights:
  • AD: Studenten in Eindhoven bedenken trilmouw
  • BNR: Mouw vertaalt elke taal naar trilpatronen op de arm
  • Brainport Eindhoven: meet our pioneer: Lourens Touwen