How does gradient descent iteratively discover local minimas? How exactly was this algorithm invented from the ground up? What are the limitations of traditional gradient descent?