Software Firm Claims Breakthrough in Computer Vision Will Lead to Better AI

New brain-mimicking artificial intelligence software can read security images called CAPTCHAS as well as people do, its developer says–but scientists are skeptical

Luis von Ahn has heard it all before. As co-inventor of the CAPTCHA, those annoying images composed of wiggly letters and numbers that Web sites use to make sure you’re a human rather than a machine, von Ahn has received as many as 50 claims over the past decade of ways to beat his program.

Make that 51.

The start-up Vicarious, based in Union City, Calif., claims it has come up with artificial intelligence (AI) software that reads images nearly as well as humans and can crack a CAPTCHA 90 percent of the time. If the claims are true, they could signify a breakthrough in building AI that is indistinguishable from human cognition—at least when it comes to helping computers identify and understand images.

Vicarious calls the architecture its AI system is based on a “recursive cortical network,” meaning it is modeled along the line of the human neocortex—the brain’s gray matter that processes information. This approach allows AI software to learn new things from a few examples, much as a human child comes to understand the world by learning to recognize what he sees and figuring out how the images are connected.

Vicarious’s approach differs from AI methods such as “deep learning,” in which software trains an artificial neural network by providing thousands of training images for it to connect, according to the company. “The human brain is made up of a simple, replicated circuit—a single repeated element that happens over and over again in the neocortex,” Vicarious co-founder D. Scott Phoenix says, adding that his company’s software is likewise built from single, repeated elements.

Solving a CAPTCHA (for Completely Automated Public Turing test to tell Computers and Humans Apart) clears the bar that mathematician Alan Turing set in 1950 to determine whether a machine could be said to possess a humanlike intelligence, although in a limited way. Over the years other computer scientists and hackers have found ways to program computers to pass the CAPTCHA test, forcing Web publishers to employ increasingly more complicated CAPTCHAs that are difficult to decipher in their efforts to repel increasingly sophisticated spamming tactics.

Vicarious’s CAPTCHA-solving demonstration is an example of “narrow artificial intelligence,” a technology that can match or even exceed human performance on a narrowly defined task. IBM’s chess-playing Deep Blue is another such example. But Vicarious insists its computer perception software is the foundation of an AI that will learn the way humans do—by experiencing the world around it, principally via vision, and then identifying patterns. “If an algorithm solves vision in general, it is not narrow AI, it’s a general AI system,” says, Dileep George, also a Vicarious co-founder. “We are working on a general algorithm for solving [the] vision problem, and CAPTCHA is a stepping stone to that.”

One reason computers scientists are skeptical about Vicarious’s claim is that the company has kept its technology under wraps. It demonstrated the software on video, which showed its technology solving CAPTCHAs from major Web sites, rather than by publishing its findings in scientific journals. Nils Nilsson, emeritus professor at Stanford University and author of The Quest for Artificial Intelligence: A History of Ideas and Achievements, says Vicarious’s claim is significant but he has reservations. Vicarious, he says, uses “the CAPTCHA thing as just one test case to show how well their technology works. I’d say, okay, that’s probably a pretty good advance, but I would need to know more.”

Read more . . .