A0569
Title: On dimension-free uniform concentration of logistic regression
Authors: Shogo Nakakita - The University of Tokyo (Japan) [presenting]
Abstract: The purpose is to present a novel dimension-free uniform concentration inequality for the empirical risk function of logistic regression. This result establishes a uniform law of large numbers over an Euclidean ball under the natural sufficient condition that the effective rank of the problem divided by the sample size tends to zero. The derivation employs a PAC-Bayes bound in combination with a second-order expansion in place of McDiarmid's inequality and the Rademacher complexity argument.
