Researcher at Laboratory of Data Engineering

I am Sven Heydenreich, an AI interpretability researcher at LADE at AREA science park. I recently joined Matteo Biagetti’s team, aiming to leverage tools from topological data analysis to make the inner workings of AI models more transparent, understandable and safe.
Before joining LADE I was a researcher in observational cosmology at the University of California Santa Cruz, chairing a working group within DESI that combines data from different survey types to achieve precise and robust constraints on the cosmological standard model.

Research Interests

I believe that modern machine learning models are far too complex to truly understand them through brute-force approaches within mechanistic interpretability. Our best chance is arguably posed by trying to find (or develop) the right types of mathematical abstractions that allow us to study the behavior and emergence of capabilities in an abstract and general way.
I am generally interested in all such efforts, right now my particular interest is for methods from algebraic topology (most notably persistent homology) and algebraic geometry (singular learning theory).

Experience & Education

• B.Sc. & M.Sc. in Mathematics (2014 & 2016) at University of Münster, Germany (focus: Algebraic Topology)
• M.Sc. in Astrophysics (2019) at University of Bonn, Germany (focus: systematic biases in gravitational lensing observations)
• Ph.D. in Astrophysics (2022) at University of Bonn, Germany (focus: higher-order statstics for gravitational lensing observations)
• Post-doc in Cosmology (2022-2026) at UC Santa Cruz, USA (focus: synergies between spectroscopic and imaging surveys within the DESI collaboration)
• Post-doc in AI interpretability (2026-present) at LADE, Area Science Park, Trieste, Italy (focus: topological data analysis for AI interpretability)