I think they trained on a bunch of random patterns, and can recognize basically anything of approximately the same resolution.