Learning Decision Trees With Gradient Descent