Résumé : We consider online learning algorithms that guarantee worst-case
regret rates in adversarial environments (so they can be deployed safely and will perform robustly), yet adapt optimally to favorable environments (so they will perform well in a variety of settings of practical importance). In this talk we introduce the MetaGrad algorithm for online convex optimization, its luckiness bound, and consequent unified adaptivity to two common scenarios :
* curvature : we show that MetaGrad exploits exp-concavity
* stochastic case : we show that MetaGrad adapts to the Bernstein exponent (generalization of Tsybakov margin condition). MetaGrad’s computational efficiency is comparable to AdaGrad’s. We expect its new adaptivity to be especially useful in practice.
Lieu : bâtiment 1R3, salle MIP (1er étage)