Where exactly does the gradient occur in the gradient boosting algorithm if the base algorithm is a simple decision tree?