Dreamscapes of MIT

Here are visualisations of the 205 scene categories in the MIT Places Database using a variant of the same technique used to generate the thousand faces of Caffenet.

These are rendered on random grayscale noise rather than a photograph, so they're purely a manifestation of what the neural net has "learned" from the training image set. I don't find the architectural fantasies here anywhere near as disturbing as the classic deepdreams images. Some of them are remarkably "readable" (like the butcher's shop in 34, or the volcanos in 197) and they've all got a kind of trippy beauty.