[gfodor]

The point isn't that you shouldn't learn latent dirichlet allocation, but that the best way to learn it is to come upon a problem that warrants it and then have a mentor point this fact out to you, motivating you to learn it so you can expand your capabilities.

There are plenty of engineers who love to pour over academic research of the bleeding edge who have no problems to solve with it, yet when you ask them to hack together a prototype of an idea they don't know where to start. I think there is a clear personality type that prefers to continue reading before doing, since there is an ever-present fear that just around the corner there is some kernel of knowledge that will remove the need for a massive body of work. In reality, this is almost never the case, and elbow grease and experimentation turns out to be the solution to many problems, not digging through libraries, frameworks, algorithms, and academic research.