Research

I work on the mathematical foundations of deep learning. I am particularly interested in understanding:

  • Generalization: How and when can we predict the performance of trained neural networks on previously unseen data?

  • Stability & Signal Propagation: How can we ensure stable information flow in deep neural networks?

  • Feature Learning: How do neural networks develop useful internal representations during training?

My research is funded by the DFG Emmy Noether project RMT4DL (Random Matrix Theory for Deep Learning), in which I develop probabilistic and random matrix theory methods to study these questions.

Publications & Preprints

How Long Does Infinite Width Last? Signal Propagation in Long-Range Linear Recurrences (link)
M. Seleznova.
arXiv preprint, 2026.

GradPCA: Leveraging NTK Alignment for Reliable Out-of-Distribution Detection (link)
M. Seleznova, H. H. Chou, C. M. Verdun, G. Kutyniok.
International Conference on Learning Representations (ICLR), 2026.

Conditional Deep Generative Modeling of Blood-Based Infrared Spectra Enables Controlled In-Silico Phenotyping Studies (link)
N. Leopold-Kerschbaumer, T. Halenke, S. Süzeroğlu, M. Jung, M. Seleznova, N. Thorisch, J. Lorenz, T. Bocklitz, B. Eskofier, G. Kutyniok, F. Krausz, K. Kepesidis.
Research Square preprint, 2026.

Revisiting Glorot Initialization for Long-Range Linear Recurrences (link)
N. Bar*, M. Seleznova*, Y. Alexander, G. Kutyniok, R. Giryes.
Advances in Neural Information Processing Systems (NeurIPS), 2025. (*Equal contribution.)

Analyzing Training Dynamics of Deep Neural Networks: Insights and Limitations of the Neural Tangent Kernel Regime (link)
M. Seleznova.
PhD thesis, Ludwig Maximilian University of Munich, 2024.

Neural (Tangent Kernel) Collapse (link)
M. Seleznova, D. Weitzner, R. Giryes, G. Kutyniok, H. H. Chou.
Advances in Neural Information Processing Systems (NeurIPS), 2023.

Neural Tangent Kernel Beyond the Infinite-Width Limit: Effects of Depth and Initialization (link)
M. Seleznova, G. Kutyniok.
International Conference on Machine Learning (ICML), Spotlight, 2022.

Analyzing Finite Neural Networks: Can We Trust Neural Tangent Kernel Theory? (link)
M. Seleznova, G. Kutyniok.
Mathematical and Scientific Machine Learning (MSML), 2022.

Guided Exploration of User Groups (link)
M. Seleznova, B. Omidvar-Tehrani, S. Amer-Yahia, E. Simon.
Proceedings of the VLDB Endowment (PVLDB), 2020.