Transfer-learning

[kaiolae]

Hi!

Thanks for this very exciting work and the great presentation of it on the wepage.
I was wondering, do you think this attention-agent may be able to do some sort of transfer learning, where the learned attention is reused on a new task? For instance, if one changed the task a bit (while staying in the same domain), could it be enough to re-train the controller, while retaining the same attention module?

Best wishes,
Kai