The mystery of over-parametrization in neural networks