Author mentions style transfer as a product of NN research.

Yet this paper [1] from 2014 works extremely well, no neural networks. Here's one from 2008 [2] that worked extremely well too, better than most early NN techniques.

Although, deep learning is far from the intuitive approach that existed 10 years ago. It's clearer now how to reason about models, layers, activation functions etc. As much as there's no mathematical foundation, I do not believe it helps that much in the case of SVMs or whatever else. There still has to be experimentation, approximation and proper testing.

[1]: https://dspace.mit.edu/handle/1721.1/100018 [2]: http://link.springer.com/chapter/10.1007%2F978-3-540-88690-7...