• Tech
  • Google

See the Fantastically Weird Images Google’s Self-Evolving Software Made

These pictures aren’t from the mind of a person on psychedelic drugs. Instead they represent the way Google’s AI software is able to reinterpret different images
Google’s software, called an artificial neural network, is able to learn in the same way a human brain does.
Google trains the network to see by feeding it millions of images that teach it how to interpret different objects.
The network is built of 10 to 30 layers of artificial neurons, each of which interpret different levels of complexity in an image.
These pictures were made by feeding the network an arbitrary image and allowing it to enhance whatever it deems most important
The images serve as a way to test how well the network has learned during training
The results are colorful, jarring and beautiful. Each layer of the network can come up with a different interpretationGoogle
Google calls this process of abstract re-interpretations of images “inceptionism”
Certain objects are often re-interpreted in similar ways. Horizon lines are often filled with towers and pagodas, rocks turn into buildings, and birds or insects appear in pictures of leaves
In addition to re-interpreting images, the network can create “dreams” from a random-noise image by continually building new impressions on top of its old ones Google
In the future neural networks could be used by artists as a new form of visual expression
Google will continue using neural networks for everything from voice recognition to identifying people in its Photos app

More Must-Reads From TIME

Contact us at letters@time.com