Deep Dreaming — GoogLeNet on the MIT Places database
zenith parsec
Jul 20, 2015
A two-minute DeepDream zoom from the first month, run on a GoogLeNet trained on MIT's Places database instead of ImageNet — so the network dreams in architecture and landscapes rather than dogs.
DeepDream
Alexander Mordvintsev / Google
Google published Inceptionism in June 2015 and the internet met DeepDream's psychedelic, dog-eyed images. Alexander Mordvintsev made the first ones by running Google's Inception classifier backwards, amplifying whatever the network half-saw in a photo. Trained on ImageNet's 120 dog breeds, everything it dreamed came out canine.
The link to video was immediate. DeepDream is iterative, and iteration is animation: run it on a frame, zoom in, feed the result back. Johan Nordberg's July 2015 "Inside an Artificial Brain" is the direct ancestor of every AI zoom video since. It was the first time a neural network's imagination had been made to move.