I study the mathematical structure of neural networks, with the goal of contributing to AI alignment. My research connects discrete geometry, loss landscapes, and interpretability: how networks learn, how they compute, and when their explanations can be trusted.
I am a master's student in Computer Science at Technical University Berlin, a researcher in the Discrete Geometry Group at Free University Berlin, and a working student at Fraunhofer HHI in Berlin.
I will be an ILIAD Fellow from December 2026 to March 2027.
With professor Georg Loho, I work on the mathematical foundations of Deep Learning. We investigate the geometry of the Newton polytopes of neural networks and develop new interpretability methods from the Difference-of-Convex decomposition of networks. Using agentic proof search in Lean 4, we study loss landscapes of shallow ReLU networks — where skip connections remove spurious minima and how teacher parameters determine the critical points.
With professor Wojciech Samek I am working on new backpropagation based interpretability methods for the Transformer Architecture.
I have worked on runtime monitoring of image recognition models and representation learning. I developed a monitoring approach that is knowledge-guided and works on internal representations of image recognition models and embedding models such as DINO.
I want to help build AI systems whose behaviour we can understand and reliably evaluate. My approach starts with the mathematical structure of neural networks: how their geometry shapes learning, what their internal computations reveal, and where explanations fail to capture what a model actually does.
I work on geometric approaches to interpretability, attribution methods grounded in network computations, and monitoring methods for detecting model failures.
Human understanding is my north star: I want explanations that help people understand how a model reaches its conclusions. The same principle guides my use of agentic tools in mathematics: beyond obtaining a correct proof, I want to understand the human intuitions behind it.
Network Newton polytopes. A ReLU network \(x\mapsto\sum_i s_i\,\mathrm{ReLU}(\langle w_i,x\rangle)\) with weights \(s_i\ge 0\) is the support function of the zonotope \(P=\sum_i s_i[0,w_i]\), its Newton polytope: the scaled first-layer weights \(s_iw_i\) are its generators. The learned linear skip only translates it: drag \(P\); it snaps to the centered \(P_c=\sum_i s_i[-\tfrac12 w_i,\tfrac12 w_i]\) at the origin and back to its Steiner point \(\tau=\tfrac12\sum_i s_iw_i\).
Support functions. The support function of a convex body \(K\) is \(f_K(u)=\max_{v\in K}\langle v,u\rangle\). Translating by \(\tau\) adds exactly the odd term: \(f_P=f_{P_c}+\langle\tau,u\rangle\). Show whose maximum it is, and of the maximising vertex.
I completed my bachelor's degrees in Mathematics and Computer Science at KIT (Karlsruhe Institute of Technology) in Karlsruhe.
My bachelor thesis “Induced Turán problems” was supervised by Maria Axenovich and Thorsten Ueckerdt. The thesis studies extremal problems for induced and biinduced subgraphs, connections to Vapnik–Chervonenkis dimension, and includes a simplified proof of the Erdős–Hajnal conjecture for graphs with bounded VC dimension.