| Life learned to fly at least four times. First the insects, then pterosaurs (flying reptiles of the Jurassic and Cretaceous), then avian dinosaurs (the ancestors of modern birds) and finally bats. Yet, intriguingly, every group of animals that has succeeded in taking to the skies does so slightly differently. Birds' feathers work like airfoils, providing lift. The long, webbed wings of bats can curve and stretch to best exploit aerodynamic forces. Insects flap their wings in a flattened figure-eight pattern. And while it would take an extraordinary act of time travel to study the exact physics of pterosaur wings, it's likely that those would surprise us too. The many, varied solutions to the problem of flight is an example of life's creativity, biologists say. Though evolution happens without foresight, the pressures imposed by natural selection have produced a world of unexpected and delightful results. But creativity is not the exclusive purview of life. Thanks to neural network expert Janelle Shane, who pointed it out on Twitter, I recently stumbled across this fascinating collection of anecdotes from three dozen computer scientists. They argue that machine learning algorithms — computer programs that figure out how to solve problems through trial and error — can be creative too. Among their examples: - A robot that was supposed to evolve to move as fast as possible didn't develop legs or wheels; it learned to somersault. "Evolution discovers that it is simpler to design tall creatures that fall strategically than it is to uncover active locomotion strategies," the researchers wrote.
- A computer program designed to automatically fix a buggy sorting algorithm found a simple — if unhelpful — solution to the problem: it short-circuited the buggy program, deleting the contents of the lists it was supposed to sort. If there are no items on a list, it can't be out of order!
- And algorithm tasked with figuring out how to use the minimum amount of force needed to land a plane reached a solution unexpectedly fast. When researchers took a closer look, they realized that the algorithm was instead applying forces too large for the system to handle, causing it to register as a "zero." As Shane wrote, "The pilot would die but, hey, perfect score."
"Artificial organisms evolving in computational environments have ... elicited surprise and wonder from the researchers studying them," the researchers wrote. But too often, they say, those bursts of AI brilliance are treated as "obstacles to be overcome, rather than results that warrant study in their own right." And as a result, "many natural scientists do not realize how interesting and lifelike digital organisms are and how natural their evolution can be." I found this paper captivating, intriguing, and maybe just a little bit disquieting. It's a testament to how easily the tools humans create can turn into something we never imagined. Something interesting to mull over on this Wednesday afternoon. — Sarah Kaplan |
Twitter
Facebook
