In machine learning, we take gradient descent for granted. We rarely question why it works. What’s usually told is the mountain-climbing analogue : to find the valley, step towards the steepest descent. But why does th...