Montag, 23. Oktober 2023, 16:45 - 17:45 iCal
ISOR Colloquium
"Early stopping in iterative learning methods"
Speaker: Markus Reiß (Humboldt-University Berlin, Germany)
HS 7 OMP1 (#1.303)
Oskar-Morgenstern-Platz 1, 1090 Wien
Vortrag
Many statistical and learning procedures rely on iterative solvers, e.g. gradient descent steps for minimizing an empirical risk criterion. It turns out that stopping the iterates before convergence ("early stopping") usually regularizes the problem and is thus recommendable from a statistical perspective while it also reduces computational costs considerably. The underlying statistical problem is to design early stopping rules that are completely data-driven and lead to optimal statistical estimation bounds. We clarify the basic ideas for spectral-cutoff projection methods with exact oracle inequalities and matching lower bounds under sequential information constraints. We then look at the nonlinear conjugate gradient or partial least squares algorithm for linear models and discuss how the basic ideas can be adapted to a modern greedy learning procedure.
Zur Webseite der Veranstaltung
Veranstalter
Institut für Statistik und Operations Research
Kontakt
Sabine Sobotka-Tompits, BA
Fakultät für Wirtschaftswissenschaften
Institut für Statistik und Operations Research
+43 1 4277 38631
sabine.sobotka-tompits@univie.ac.at
Erstellt am Montag, 17. Juli 2023, 12:52
Letzte Änderung am Montag, 09. Oktober 2023, 08:25