Including Uncertainty when Learning from Human Corrections