How we can use mathematics to get a better understanding of neural networks training dynamic?