Dreaming of Electric Sheep

When you see animals in the clouds or faces in the patterns of a stucco ceiling, that’s pareidolia: the tendency to interpret random noise as meaningful images.

This was an exclusively human phenomenon until researchers discovered that neural networks trained to identify images also have sufficient information to generate those images. When the neural net is fed a random image and instructed to enhance whatever it finds there, it begins to manufacture specific images out of random noise.

The results are beautiful and surreal.

The following images were produced by the MIT Computer Science and AI Laboratory, generated from pure noise as interpreted by a network trained to recognize places.

Read more about the process in the Google Research Blog, and view a gallery of specimens in Michael Tyka’s Inceptionism: Going Deeper into Neural Networks.