In the Ilya Sutskever Talking Machines podcast[1] he describes exactly how it is "magic". He talks about how there is no theoretical basis to think that a deep neural network should converge, and prior to around 2006 the accepted wisdom was that networks deep enough to outperform other methods of machine learning were useless because they couldn't be trained.
And then they discovered how to initialize the random starting values in such a way that they do converge - for reasons no one really understands.
I've heard a lot of people claim there is no magic, but I tend to think Sutskever[2] knows a little bit more than most people about deep neural networks.
[1] http://www.thetalkingmachines.com/blog/2015/1/15/machine-lea...