train complex image deep networks even with little supervised im- age data. This is contrary to the popular strategy of trading the representation power of the deep networks for a better generaliza- tion performance in the literature. In this paper, we will present how to use transferred cross...