Info about me

Hi there!
I’m Niklas, currently a Postdoc @Meta AI (FAIR), where I work on memory in language models, reasoning with structured data, cryptography and different types of interpretability.
Previously I was an AI and Fundamental Physics Research Associate @MIT, where I worked on interpretability and robustness, both researching in the AI space aswell as using gained insights for more trustable data collection in collider physics.
In 2021 I completed a PhD in Physics from CERN, where I developed trigger & core software for the upgrade of the LHCb experiment.
Python and C++ are the languages I do best.

See my CV for more details.

When not at work, you might find me playing video games, in the gym, outside, playing lasertag, skiing, biking or so.

Papers

Last update: Mar 25, 2024

While at CERN and MIT, I was part of a large experimental physics collaboration, LHCb, which jointly publishes everything. I did not have significant contributions to the papers individually, but I was part of the team that developed the software and algorithms to be able to collect the data in the first place. This makes my Scholar page a little wonky, therefore I keep track here, where I actually did a significant chunk of the things presented.

• Memory Mosaics (Preprint)
• Transforming the Bootstrap: Using Transformers to Compute Scattering Amplitudes in Planar N = 4 Super Yang-Mills Theory (Preprint)
• The cool and the cruel: separating hard parts of LWE secrets (AfricaCrypt)
• From Neurons to Neutrons: A Case Study in Interpretability (ICML)
• Salsa Fresca: Angular Embeddings and Pre-Training for ML Attacks on Learning With Errors (Preprint)
• DiSK: A Diffusion Model for Structured Knowledge (Preprint)
• NuCLR: Nuclear Co-Learned Representations (IMCL SynS&ML)
• Development of the Topological Trigger for LHCb Run 3 (ACAT)
• Finding NEEMo: Geometric Fitting using Neural Estimation of the Energy Movers Distance (ML4PS NeurIPS)
• Towards Understanding Grokking: An Effective Theory of Representation Learning (NeurIPS)
• Robust and Provably Monotonic Networks (ML: Sci. Technol.) & Expressive Monotonic Neural Networks (ICLR) (once fledged out a little more for physics, once for ML audience)
• A Comparison of CPU and GPU Implementations for the LHCb Experiment Run 3 Trigger (Comp. Soft. Big Sci.)
• Thesis: A Selection Framework for LHCb’s Upgrade Trigger
• Evolution of the energy efficiency of LHCb’s real-time processing (CHEP)
• Configuration and scheduling of the LHCb trigger application (CHEP)
• A new scheduling algorithm for the LHCb upgrade trigger application (ACAT)
• New approaches for track reconstruction in LHCb’s Vertex Locator (CHEP)