Computing analogies between images. Clever idea: one wants to complete analogies of the form A:A'::B:B', where those are all pictures. The solution is to find a filter, such that both A' and B' are filtered versions of A and B. The even neater application is to be able to fill in the blanks: learn the filter from A and A', then find what the analog for B is. (For instance, landscapes in the style of Lucian Freud!) Especially look at the textures-by-numbers examples.
There are vast possibilities for fraud here; also for jokes. Presumably the same approach would work for movies. In fact, you could build an analogical predictor, not just a filter, in this way: train it on pairs of successive frames from a movie, and then apply it to the last frame, to see how the movie continues! (Memo to self: see if you can't work in a reference to this in the new paper on predicting spatiotemporal systems.)