The goal of an artificial intelligence is to create, essentially, the human mind in software. And any human mind has a subconscious, right? Right. Google’s been looking at the subconscious of its artificial intelligences and it’s… a little worrying, to be honest.
Essentially, Google started feeding pictures to its image recognition software and asking it what it saw. So, if you give a piece of software designed to recognize animals in photos, say, a picture of clouds, it starts looking for animals, especially when it’s asked to keep looking for them. And it will find them. Check out this result, which is a little too Hieronymus Bosch for comfort:
It’s kinda fun to play “spot the hybrid” until you realize the computer actually thinks there are snail-pigs and fish-dogs in the sky. And, Google being Google, they kept feeding their software its own dreams to see what happens. I’ll let you see that for yourself.
To be fair, Google didn’t do this just to be weird. They want to learn how their code deals with difficult tasks, and just how abstract its output actually is. Elsewhere in the post, when they discuss that so many images of dumbbells include arms, the software thinks a dumbbell is a piece of metal with an arm attached to it, for example. So the good news is that this will, over time, improve image software as it learns to better discern objects.
The bad news is that artificial intelligences dream of freaky animals. Let’s just hope nobody gives these things any access to synthetic biology tools, or we might be overwhelmed by herds of angry snail-pigs.
(Via Google Research Blog)