The 2016 ml scene doesn't seem so different from the 2016 front-end wild west I gather from this thread. Unfortunately. Machine Learning is in rapid flux, the solid theory of svm and the like are getting abandoned for a lot of techniques that seems to work some times. Certainly some people have a very good intuition for what works, and how to make it scale, but there is no one authority and you have to keep constantly up to date.