I understand the zig-zag nature of the cost function when applying gradient descent, but what bothers me is that the cost started out at a low 300 only to increase to 1600 in the end.