It's nothing specific, it's just that the FANN library, as of the latest version 2.2.0 at least, lets you create and manipulate training data structs directly, without any need to write any intermediate files to disk. All this string processing seems unnecessary when you already have the data in-memory, it's just that one needs to do a little fiddling around to hook it up to the neural network in the right way.

Also you can actually use int as fann_type, but I don't see a pressing need for it. The code might be cleaner I guess.