AI has a habit of doing things you never expected, but this isn’t creativity – it’s generally lack of imagination or analysis on our part. It isn’t that AI s creative, it is that that we are not as clever as we think we are.
Way back when I first started in AI, I was assigned a project to classify photos of cracks in materials – all the program had to do was say there was a crack or there wasn’t a crack. I used a very simple analysis procedure that today you probably wouldn’t even consider AI – discriminant analysis. I was amazed I got 100% correct classification. Then I did some image processing and normalized the brightness so that all of the images had the same average luminance. I got 0% correct classification. It turned out that the cracks were bright and the analysis had used average brightness to classify the images. As it happened this was a practical success and the complex “AI” was replaced by a light meter and a threshold gate.
The point of this story is that I had convinced myself that something clever was going on, and it wasn’t. I didn’t really understand the techniques I was using. The same is true, only much more so, today. Researchers are often unable to predict what their complex optimization approaches will lead to. It is just too tempting to jump back, hold your hands in the air, and say “its creative!”. Well no; it is just that you missed seeing a whole part of the parameter space that the optimizing algorithm didn’t miss.
This said there are some lovely stories that will keep you amused for ages. So much so that a group of researchers have got together to tell their tales, but not over a camp fire in the dead of night. Instead they have collaborated on a paper: The Surprising Creativity of Digital Evolution: A Collection of Anecdotes from the Evolutionary Computation and Artificial Life Research Communities. The list of authors and establishments is far too long to give here, but as they describe it:
“This paper is the crowd-sourced product of researchers in the fields of artificial life and evolutionary computation who have provided first-hand accounts of such cases. It thus serves as a written, fact-checked collection of scientifically important and even entertaining stories. In doing so we also present here substantial evidence that the existence and importance of evolutionary surprises extends beyond the natural world, and may indeed be a universal property of all complex evolving systems.”
You can tell that I’m not really in agreement with the idea that any of this proves that random algorithms are creative in any sense other than random surprise. I am more sympathetic to the idea that this is a form of blind creativity that occurs within evolution, but without us to notice it and call it creative it is just another random adaptation.
The paper presents 27 tales of the unexpected. To save you from having to read it, Two Minute Papers has made a You Tube video showing four of the stories:
The rest of the stories are just as interesting, but the point is made in the paper that:
“Exacerbating the issue, it is often functionally simpler
for evolution to exploit loopholes in the quantitative measure than it is to achieve the actual desired outcome. Just as well-intentioned metrics in human society can become corrupted by direct pressure to optimize them (known as Campbell’s law or Goodhart’s law), digital evolution often acts to fulfill the letter of the law (i.e. the fitness function) while ignoring its spirit. We often ascribe creativity to lawyers who find subtle legal loopholes, and digital evolution is often frustratingly adept at similar trickery.”
Recalling the meme that there is always an obligatory xkcd cartoon – here it is:
Click for larger image
More cartoon fun at xkcd a webcomic of romance,sarcasm, math, and language
Many of the stories are about failures of the algorithm or simulation. For example, an AI agent learned to exploit bugs in a game to win by committing joint suicide with its enemy – the game didn’t count the suicide as death of the agent so the game continued.
One class of story that is particularly rewarding is where the algorithm exceeds its expectations. For example, artificial organisms were trained to follow a trail of nutrients. While organisms were given the ability to sense whether there was nutrient underneath them, and if it was necessary to turn left or right to stay on the nutrient trail, their sensors couldn’t detect arriving at the end of the trail.
“Organisms were rewarded for reaching more of the trail, and were penalized for stepping away from the trail. Because it was impossible to directly sense where the trail ended, the best expected solution was to correctly follow the trail one step past where it ended, which would incur a slight unavoidable fitness penalty. However, in one run of evolution, the system achieved a perfect fitness score – an analysis of the organism revealed that it had invented a step-counter, allowing it to stop precisely after a fixed number of steps, exactly at the trail’s end!”
There are plenty more where that came from.
If you are working in AI you should read the paper just to learn to avoid some of the mistakes it outlines.
To be informed about new articles on I Programmer, sign up for our weekly newsletter, subscribe to the RSS feed and follow us on, Twitter, Facebook, Google+ or Linkedin.
Make a Comment or View Existing Comments Using Disqusor email your comment to: firstname.lastname@example.org