Théo Beuzeville

Toulouse Institute for Computer Science Research (IRIT),

University of Toulouse,

Toulouse, France

I am a postdoctoral researcher specializing in numerical error analysis for algorithms in artificial intelligence. I received an Engineering Degree in Applied Mathematics and Computer Science from the French Grande École ENSEEIHT and a Master of Science in Computer Science from Université de Toulouse in 2020. I completed my PhD thesis in 2024 at Université de Toulouse, under the supervision of Alfredo Buttari, Théo Mary, and Serge Gratton.

The aim of my research is to provide tools to better understand, explain, and predict the sensitivity of artificial neural networks to various types of perturbations. To this end, I extend to artificial neural networks some well-known concepts from numerical linear algebra, such as condition number and backward error, which led to contributions in adversarial attacks and rounding error analysis, as well as practical applications.

Selected publications

Journal article

Deterministic and probabilistic rounding error analysis of neural networks in floating-point arithmetic

Théo Beuzeville, Alfredo Buttari, Serge Gratton, and Theo Mary

IMA Journal of Numerical Analysis, 2025

In press

Bib PDF

@article{beuzeville:hal-04663142,
  title = {{Deterministic and probabilistic rounding error analysis of neural networks in floating-point arithmetic}},
  author = {Beuzeville, Th{\'e}o and Buttari, Alfredo and Gratton, Serge and Mary, Theo},
  url = {https://hal.science/hal-04663142},
  journal = {{IMA Journal of Numerical Analysis}},
  publisher = {{Oxford University Press (OUP)}},
  year = {2025},
  keywords = {rounding errors ; backward error ; probabilistic error analysis ; artificial neural networks ; error analysis ; floating-point arithmetic},
  hal_id = {hal-04663142},
  note = {In press},
  hal_version = {v2}
}

Thesis

Backward error analysis of artificial neural networks with applications to floating-point computations and adversarial attacks

Théo Beuzeville

Jun 2024

Bib PDF

@phdthesis{beuzeville:tel-04622129,
  title = {{Backward error analysis of artificial neural networks with applications to floating-point computations and adversarial attacks}},
  author = {Beuzeville, Th{\'e}o},
  url = {https://theses.hal.science/tel-04622129},
  number = {2024TLSEP054},
  school = {{Universit{\'e} de Toulouse}},
  year = {2024},
  month = jun,
  keywords = {Backward error ; Rounding errors ; Adversarial attacks ; Error analysis ; Floating-point ; Artificial neural networks ; Erreur inverse ; Erreurs d'arrondi ; Attaques adverses ; Analyse d'erreur ; Virgule flottante ; R{\'e}seaux de neurones artificiels},
  type = {Theses},
  hal_id = {tel-04622129},
  hal_version = {v1}
}

Technical report

Adversarial attacks via backward error analysis

Théo Beuzeville, Pierre Boudier, Alfredo Buttari, Serge Gratton, Théo Mary, and Stéphane Pralet

Dec 2021

Bib PDF

@unpublished{beuzeville:hal-03296180,
  title = {{Adversarial attacks via backward error analysis}},
  author = {Beuzeville, Th{\'e}o and Boudier, Pierre and Buttari, Alfredo and Gratton, Serge and Mary, Th{\'e}o and Pralet, St{\'e}phane},
  url = {https://ut3-toulouseinp.hal.science/hal-03296180},
  year = {2021},
  month = dec,
  hal_id = {hal-03296180},
  hal_version = {v3}
}