2ba61cc3a8f44143e1f2f13b2b729ab3-AuthorFeedback.pdf

A1: We progress the line of works [3, A, B] by introducing novel probabilistic insights that also bring significant11 empiricalimprovements. A3: Weretrained all compared methods since prior works chose different baselines and cannot be fairly compared50 with. Thus, end-to-end training with it is slower.