In 1968, science fiction author Philip K. Dick asked the question in the eponymous novel, "Do Androids Dream of Electric Sheep?" Cut to nearly 50 years later and we've finally stumbled onto the answer... yes they do, sort of.
To process Google's images, as well as the most complicated of the world's data, Google runs off stacked artificial neuron networks (ANNs). To understand this as the "brain" of Google isn't oversimplifying as much as you might imagine: the ANNs, like a real brain, need to be "taught" through finding patterns in a series of repetitions. The neural networks are fed a series of images by their programmers and, over a long enough time period, will pick up a pattern.
For example, this neural network is learning what a dumbbell is. Because arms were often attached to the dumbbells, the AI questioned if the arm was an extension of the dumbbell itself that maybe some dumbbells have arms.
Here's where it gets super sci-fi: as the computer begins to amplify patterns, different neural networks will focus on different elements. From these impressions, the AI will begin to generate its own images from free-association. It's the computer's version of finding shapes in the clouds.
Actually, Google's AI is a champion of that game.
When a neural network still in the learning stage it is tasked with finding something specific within an image, and it can produce troubling hallucinations like this animal-automobile hybrid.
As it grows, it learns to differentiate separate elements within an image, creating a painterly, edge-finding effect on the image.
As the artificial intelligence develops, the computer can take full reign over its neural processing to create its own pictures from a mechanical mind, like this psychedelic landscape.
The computer starts with a white screen and is allowed to do what its programmers call "dreaming."
No electric sheep were harmed in the making of this artificially intelligent dreamscape.
Via: Google Research