Differential geometry of ML: a geometric interpretation of gradient descent