Training Fully Connected Neural Networks is ∃R-Complete

Dettagli

Il Collegio del DOTTORATO in INGEGNERIA INDUSTRIALE e DELL'INFORMAZIONE ha organizzato per il giorno Giovedì 19 Maggio ore 14:30 il seminario dal titolo Training Fully Connected Neural Networks is ∃R-Complete, dedicato agli studenti del dottorato.

Speaker:  Paul Jungeblut, Karlsruhe Institute of Technology (KIT)

Abstract: We consider the algorithmic complexity of training fully connected neural networks. In particular, we see that this problem is ER-complete. This means that it is (up to polynomial transformations) equally difficult as finding the solutions of a system of polynomial equations and inequalities in several unknowns.
In the talk I will introduce the neural network training problem and give some background on ER-completeness. Then we combine both to get an idea how ER-completeness is proven.
The talk is based on our recent paper available at https://arxiv.org/abs/2204.01368

L'incontro si svolgerà sulla piattaforma Teams al seguente link: https://bit.ly/3EAuTnI.

Notizie ed eventi - Elenco