There may be a paper on it, but it’s not a common view.

In particular, this paper neglected to do the obvious thing: ensemble networks trained with dropout. It improves performance over dropout alone.