Google engineer's 2 a.m. code tweak reversed a vision neural net, which answered by growing extra eyes and snouts out of a stock photo of a kitten

At 2 a.m. on May 18, 2015, Google engineer Alexander Mordvintsev woke from a nightmare at his Zurich apartment, gave up on sleep, and wrote roughly thirty lines of code that ran a convolutional neural net backwards. Instead of identifying what was in an image, the network was told to amplify whatever it half-suspected it saw. He fed it a wallpaper-site photo of a beagle and a kitten perched on tree stumps. The network, trained on ImageNet's 1,000 categories including 118 dog breeds, built a dog out of the kitten: a second set of eyes on the forehead, another snout bulging from the haunches, a pair of eyes breaking through the lower jaw.

Mordvintsev worked on Safe Search and had no formal connection to Google Brain or DeepMind. He posted the images to Google's internal Plus at 2:32 a.m., writing, "Not sure if it was a good idea to try a DNN image enhancement at 2 a.m. How do I sleep now?" The first reply arrived seconds later from Mountain View: "MY EYES! MY EYES!" The post collected 162 +1s and more than 60 comments, and drew in Jeff Dean's intern Chris Olah and Seattle machine learning engineer Mike Tyka. The project became DeepDream.