"I consider the approach of reversing a sentence a “hack”. It makes things work better in practice, but it’s not a principled solution." I had the same feeling

"I consider the approach of reversing a sentence a “hack”. It makes things work better in practice, but it’s not a principled solution." I had the same feeling about boundary box recommendations/guesses that were used to speed up object recognition with Deep Learning fairly recently. Just as with a sliding box approach it is intuitive and works, but it also seems quite inelegant and like a better approach should be possible. Visual attention seems like it should work much better in the long term, so it is exciting the field has come to a point where it has been developed.