AAAI 2026
Exploiting Space Folding by Neural Networks
In Proceedings of the AAAI Conference on Artificial Intelligence, 2026
I am currently a Senior Researcher (AI) at Software Competence Center Hagenberg (SCCH). My background is in mathematics, physics, and statistics, and my doctoral work, carried out within the S3AI project, focused on invariant geometric structures, notably convexity interpreted as space folding. I work across machine learning, neural-network geometry, mechanistic interpretability, and surrogate models for physical simulation, with additional experience in computer vision, natural language processing, and time-series modeling.
Current focus
- Space folding, symmetries, and geometry of neural networks
- LLMs, post-training and catastrophic forgetting
- Machine-learning surrogates for rotary-kiln calcination processes
Selected Publications
Journal and Conference Papers
TMLR 2025
IJCAI 2022
Tessellation-Filtering ReLU Neural Networks
In International Joint Conference on Artificial Intelligence, 2022
BCP 2017
Geometric features of Vessiot-Guldberg Lie algebras of conformal and Killing vector fields on \(\mathbb{R}^2\)
In Banach Center Publications, 2017
Workshop Papers (workshops are cool)
ICLR 2026
Workshop
ICLR 2025
Workshop
NeurIPS 2024
Workshop
NeurIPS 2024
Workshop
The Turing Game
In NeurIPS Workshop on System-2 Reasoning at Scale, 2024
Selected Applied Work
Surrogate
Models
Machine-learning surrogates for rotary-kiln calcination
Details
At SCCH I work on machine-learning surrogates for calcination processes of limestone and magnesium carbonate in rotary kilns. The challenge is to approximate expensive physical simulations (CFD-based) with neural operator-based approaches. This simulation work is carried out within the PRIM-ROCK project.
Mechanistic
Interpretability
Mechanistic interpretability of activation regions
Details
I work on questions around the semanticity of activation regions in LLMs and the geometry of CoT. The common theme is whether the information contained in the activation space can be used to mechanistically interpret model behavior.
Education
PhD
PhD in Artificial Intelligence
On Tessellations of ReLU Neural Networks
Study of tessellations and geometric structures induced by ReLU neural networks, with emphasis on how input-space regions map through activation patterns.
MSc
MSc in Applied Mathematics
Estimating Extreme Quantiles with the Fisher Approximation of the Tail
Extreme-value modeling for tail quantile estimation, comparing GPD and Fisher-tail approaches, parametric and non-parametric estimators, and Bayesian inference with weakly informative Jeffreys priors.
BSc
BSc in Mathematical Physics
Geometric features of Vessiot-Guldberg Lie algebras of conformal and Killing vector fields on \(\mathbb{R}^2\)
Local classification and geometric study of finite-dimensional Lie algebras of conformal and Killing vector fields on \(\mathbb{R}^2\). A related paper appeared in Banach Center Publications in 2017.
Further study
Stats
Post-graduate MSc in Statistics
Bocconi University · Sep 2018 - Jul 2019
Erasmus
Erasmus study period
Université Grenoble Alpes · Sep 2017 - Jul 2018