What’s the advantage of learning this positional embedding vs hardcoding one that has the desired properties of location similarity?