this is very timely now that imagenet is "solved" to a first approximation.

We think it’s time to think about how to do something deeper — something more at the level of human understanding of an image