I did skim the article when I posted my comment and I don't mean to throw shade on folks.

That said, selecting good data and then intrinsic or derived features is the key for successful ML applications. Reducing input data without losing relevant features is also a common transformation/abstraction technique, e.g., how do you test for the negative impact of a change to an algorithm without testing it against a TB/PB of year's worth of data? So, since this paper is doing these for DL, I am lost on its contribution; specifically, along the dimension of novelty and being interesting.