Higher order conditional Boltzmann machines
Instead of modeling the density of image pairs, we could
model the conditional density   p(image(t+1) | image(t))
image transformation
image(t)
image(t+1)
 See the talk by Roland Memisevic