Google engineers have taught artificial neural networks so much that the machines have started to produce art and even dream. Google shared images of the artificial brains have been producing and it's trippy.

Google has been working with neural networks that are each composed of 10 to 30 stacks of artificial neurons.

The networks have been engineered to function like a biological brain, human grey matter to be specific, and their stacks of artificial neurons can be broadly separated into three layers. The complexity of the processing increases as information moves from the lowest level to the top layer.

The neural networks started off with image recognition and classification. But because even Google's engineers didn't completely understand how the neural networks learn, they some decide to invert the process and ask the neural networks to spit images out.

"So here's one surprise: neural networks that were trained to discriminate between different kinds of images have quite a bit of the information needed to generate images too," says Google.

In one experiment, the neural networks were asked to visual object or animals within generally unrelated pictures. The AI would search for elements that made a particular object or animal unique and the Google engineers would enhance these markers by passing the data through an information loop.

"if a cloud looks a little bit like a bird, the network will make it look more like a bird," says Google. "This in turn will make the network recognize the bird even more strongly on the next pass and so forth, until a highly detailed bird appears, seemingly out of nowhere."

In another experiment, the Google researchers let the AI choose what elements of any given image were important. That information was also looped, with each lay producing increasingly more complex results.

When Google began to progressively decrease the zoom on their AI's "over interpretation of images," the artificial neural networks began to deliver "an endless stream of new impressions" or, as some call it, they dream.

"The techniques presented here help us understand and visualize how neural networks are able to carry out difficult classification tasks, improve network architecture, and check what the network has learned during training," says Google.

Head over to Michael Tyka's gallery for more images.

ⓒ 2021 All rights reserved. Do not reproduce without permission.