Natural Selection

I’ve been playing with this all week. Go over to Mutating Pictures. You’ll see a computer-generated image. Rate how much (on a scale of 0-10) the image looks like a face. The system “learns” from the ratings people give it, and over time, the images improve.

I’ve just seen a faceWhat’s really going on? Philipp Lenssen started with 1000 random images. These images produce “offspring” with mutations. Initially, up to 15% of the image changed with each generation, though he has reduced that number as the site has gained popularity. As people rate the images, higher ratings extend the life of the image and allow it to produce more offspring. Lower ratings kill off the images. Over time, the images that are selected look more like faces, and the entire site “evolves.” In the discussion, Philipp explains more about the algorithm behind the experiment.

I told Darrell about this. “So we’re a subset of the infinite monkeys?” he asked.

“Not really. We’re the people reading what they’re typing. We then decide which monkeys should be allowed to reproduce, and which should be killed off.” While there’s no Shakespeare yet, we may be at the point of tabloids already, or even bad blogs. That’s pretty amazing.

Within the first day, all of the 1000 original images were killed off, and the average image was lasting 58 minutes. Yesterday, there were more than 71,000 hits on the site.

Recently, Philipp added two new sets of images to the pool. A second set of faces is currently evolving from the same originals, to see if they end up looking like the first set. At the same time, a new set of pictures is being used to produce an “animal.” He left the name ambiguous on purpose to see what develops. You can also see the progress by comparing a sample of the original pictures with the current generation.

No monkeys were harmed during the creation of this blog post.