Google’s most advanced artificial intelligence, Artificial Neural Networks, has quite possibly demonstrated the beginnings of an imagination. What does an artificial neural network imagine when it looks at pictures?
Curious, terrifying and beautiful things.
The images were created by presenting a picture to the neural network, asking it to identify its features and then requesting it to modify the image so that the features it identifies are emphasized.
Engineers can task the network with identifying and emphasizing almost any object, like animals for example. In a way, the network uses its imagination in the same way humans do when we see familiar shapes in clouds.
A feedback loop is created when engineers ask the computer to identify and emphasize features on the already-modified picture again and again, until the picture is nothing like the original.
When engineers ask the neural network to identify and only modify an image slightly, the effect makes the modified image look like a painting.
When the network is asked to identify and emphasize animals in an image, the results are almost psychedelic.
In the image below, the network was tasked with finding building features in a featureless image with nothing but its imagination.
This is probably one of the coolest examples of machine learning, but chances are you’ve already used this technology before. Google’s new photo service, Google Photos, features the option to search images with text. According to The Guardian, if you were to search the term “dog,” Google would pull up every image that has a dog in it (but photos of other four-legged and furry animals may also show up).
Is this cool or what?