• I would guess that maybe it's simply not necessary in common cases? If groups of inputs are often related to each other, such that some features being non-zero automatically implies a bunch of other features being "real" (even if they equal $0$), then a NN should be able to automatically learn that the $0$s are only padded if they are in a bunch together, all $0$? Really just a guess though. Intuitively, I do think your idea could be beneficial in some cases. – Dennis Soemers Feb 4 '19 at 16:58