Data-dependent online learning regret bounds

Abstract

This work is still very preliminary. We hope to relate the notion of ‘nearly independent’ data to online prediction.