There's a huge difference between being able to apply a set of known techniques and being able to create novel techniques. I suspect most of the autodidacts you are referring to are able to read papers and implement them, and are well-versed in the practical aspects of how to tune models, but would likely struggle with doing something novel enough that is actually worth publishing.

There are exceptions, of course, but for certain types of work that extensive background in math/stats/cs is absolutely necessary.