Section: New Results

A new notion of stability for learning algorithms

Participants : Alain Celisse, Benjamin Guedj.

We introduced a new notion of stability for learning algorithms, which bridges the gap between the earlier uniform and hypothesis stability notions. It allows us to derive new PAC exponential concentration inequalities that apply to the Ridge regression algorithm as a first step. The first version of this work is presented in the preprint [41] and is now an active line of research.