I remember trying for hours to understand Adaboost before it all clicked (it's an ingenious algorithm, by the way). The paper could certainly do with some more explanation, rather than just "this is the weight updating function, this is the evaluation step, done".
I think AI and machine learning are worse than math in that respect for some reason. Part of it might be the greater focus on 6-page conference papers, which tends to require everything to get squished.
That's exactly the opposite of my impression. Most papers are full of text, and symbols are not the main feature. For instance:
http://ttic.uchicago.edu/~yury/papers/kuratowski.pdf Graph theory, the symbolism is next to nonexistent.
http://math.berkeley.edu/~aboocher/math/tietze.pdf Topology, still symbolism does not take much space.
http://www.jstor.org/stable/1989708 Classical and very highly technical, yet the ratio of text to symbolism is still in favour.