TORONTO – Have you ever wondered if computers dream?
Google was curious about its own artificial intelligence — the sort it trains to identify images, for example. It does this by showing it millions of images and “slowly adjusting the network parameters (called backpropagation) until it classifies it correctly using about 10 to 30 layers of artificial neurons. Each image is fed into an input layer which then communicates with the next one and so on. Finally, the output layer — the network’s equivalent of an answer — is created.
But instead of feeding it an image and asking it to identify it, engineers turned the tables: it asked it to generate images.
Asking it to give an image of dumbbells, the network came up with the images below.
Kind of creepy since it has arms attached to it. So, the thinking is that maybe the network had never been shown dumbbells without someone lifting them or holding them. So it’s an easy mistake for the network to make.
But then engineers then asked it to enhance what it thinks it sees in a photo of clouds. The results? Well… more creepy:
Then, instead of just something abstract, engineers gave it several photographs of real objects and asked it to enhance what it saw.
Freaky Google network result
Now, giving it random “noise” images, engineers asked it to make out something, it’s “dreams.”
Why bother doing this at all? It helps engineers determine if the network is learning, which clearly it is.
For more images, visit Google’s gallery. And prepare for some crazy dreams tonight.
All hail our robot overlords. (We might as well get started now.)