This work introduces a fast and efficient method for epistemic uncertainty in ensemble models for regression. Our method significantly improves baseline active learning methods on high-dimensional tasks.