Low-Regret Active learning