Martin Carrasco

I am ELLIS Doctoral Student @ AIDOS supervised by Prof. Bastian Rieck and co-supervised by Prof. Søren Hauberg.

My research focuses on geometrical and topological methods in machine learning. Coming from CS and having ventured a bit on competitive programming, I’ve taken to really like graphs on their incredible usefulnes as a modeling tool and to study other spaces.

Past work

On efficient TDL

I started my journey in applications of topological deep learning when working with Dr. Telyatnikov and Dr. Bernardez from REAL AI. We have a preprint out on efficient ways to perform higher-order message passing. Check it out here .

On graph metrics

For my M.Sc. thesis I was supervised by Prof. Erik Bekkers from AMLab and Prof. Bastian Rieck. I explored an alternative metric that fully characterizes attributed graphs (in expectation), it’s relationship to homomorphism counts and how it can be used as an inductive bias in message passing neural netowrks (MPNN). We put out a preprint with some of our findings here.

I also got to work with professionals on well-studied metrics on graphs in here. It was a joy to work with Katharina and Nadja for the first time :).

On topological deep learning

On my first official project in the lab Johannes and I look into higher-order message passing under the hood and find strong evidence it’s more brittle than we thought!

Naturally we were lead by a great supervision team, Nello, Bastian along with a new collaboration with Guy.

We find that graph-based models on more parsimonious representations of simplicial complexes (one skeleton, dual graph or hasse diagram) are good alternative to HOMP. Suprisingly, we learn all of these models are terrible at topological generalization. You can try your hand at our benchmark using the MANTRA dataset and/or read our preprint.