Séminaire APR
Reachability of quantified problems to guarantee control and neural network properties
Mardi 31 mars 2026Nuwan Herath (LIX)
In this talk, I will give a brief overview of my previous works and present the current ones. The common thread is providing guarantees for different applications, by characterizing safe domains of numerical problems. I will first briefly mention how I produced guaranteed visualization and how I improved a set-based inner- and outer-approximating projection operator, using interval methods. I will then present how quantified reachability can be applied in control theory and for neural network robustness. I rely on and extend techniques by Goubault and Putot with sound approximations to properties that can be expressed as quantified problems, with any alternation of existential and universal quantifiers. I show that this set-based inversion approach is suited for different applications and present some first experimental results.
Antoine.Mine (at)
nulllip6.fr