CFE 2019: Start Registration
View Submission - CMStatistics
Title: Improving on discrepancy-based early stopping with Tikhonov smoothing Authors:  Alain Celisse - Lille University (France) [presenting]
Martin Wahl - Humboldt University of Berlin (Germany)
Abstract: The purpose is to describe the early-stopping challenge and illustrate some deficiencies of classical discrepancy-based stopping rules. This motivates considering smoothing-based strategies such as the one inspired from Tikhonov regularization. For this rule, we prove several theoretical (non-)asymptotic guarantees, and also illustrate its promising practical behavior on simulation experiments carried out by means of spectral filter algorithms.