Robust Learning from Discriminative Feature Feedback