Late interaction methods compute representations for the query and corpus graphs separately, and compare these representations using simple similarity functions at the last stage, leading to highly scalable systems. Early interaction methods combine information from both graphs right from the input stages, are usually considerablymoreaccurate,butslower.

More advanced methods dealwith missing data byautoregressivelyreplacing missing observations with predicted ones, eventually using bidirectional architectures [5,6]toexploit both forwardandbackwardtemporal dependencies.

Object, attribute or stylistic changes can be learned from visually static datasets. On the other hand, high-quality data for action and reasoning-centric edits is scarce and has to come from entirely different sources that cover e.g.

This work focuses on the active learning in video captioning.

Neural Information Processing Systems http://nips.cc/