It might be some time before it's abstracted enough. For example, there's software packages for Finite Element Analysis, but if you don't know the underlying math, you can't do anything more complicated than the basics. Plus, if there's a problem without that knowledge you can't really debug it. I'm guessing ML will be like that for a while. If you want to do something simple using a package is fine, but as soon as something goes wrong you'll need that knowledge to figure out what's happening.
For example, for a while there's been work on doing sentiment analysis using machine learning and they typically train them on a data set of movie reviews. It turns out that as soon as you apply that trained system to anything other than movie reviews the actual results are quite poor, but you might not catch it.
For example, for a while there's been work on doing sentiment analysis using machine learning and they typically train them on a data set of movie reviews. It turns out that as soon as you apply that trained system to anything other than movie reviews the actual results are quite poor, but you might not catch it.