New Theory Cracks Open the Black Box of Deep Neural Networks