AI Startup Says It Has Defeated Captchas
Brain-mimicking software can reliably solve a test meant to separate humans from machines.
Image recognition is a key challenge in artificial intelligence.
Captchas, those hard-to-read jumbles of letters and numbers that many websites use to foil spammers and automated bots, aren’t necessarily impossible for computers to handle. An artificial-intelligence company called Vicarious says its technology can solve numerous types of Captchas more than 90 percent of the time.
It’s not the first time that computer scientists have managed to fool this method of separating man from machine. But Vicarious says its technique is more reliable and more useful than others because it doesn’t require mountains of training data for it to recognize letters and numbers consistently. Nor does it take a lot of computing power. Vicarious does it with a visual perception system that can mimic the brain’s ability to process visual information and recognize objects.
The purposes go well beyond Captchas: Vicarious hopes to eventually sell systems that can easily extract text and numbers from images (such as in Google’s Street View maps), diagnose diseases by checking out medical images, or let you know how many calories you’re about to eat by looking at your lunch. “Anything people do with their eyes right now is something we aim to be able to automate,” says cofounder D. Scott Phoenix.
Vicarious expands on an old idea of using an artificial neural network that is modeled on the brain and builds connections between artificial neurons (see “10 Breakthrough Technologies: Deep Learning”). One big difference in Vicarious’s approach, says cofounder Dileep George, is that its system can be trained with moving images rather than only static ones.
Vicarious set its cognition algorithms to work on solving Captchas as a way of testing its approach. After training its system to recognize numbers and letters, it could solve Captchas from PayPal, Yahoo, Google, and other online services. The company says its average accuracy rate ranges from 90 to 99 percent, depending on the type of Captcha (for example, some feature characters arranged within a grid of rectangles, while others might have characters in front of a wavy background). The system performed best with Captchas composed of letters that look like they’re made out of fingerprints.
“Captcha” stands for “completely automated public Turing test to tell computers and humans apart.” They were created in 2000 by researchers at Carnegie Mellon University and are solved by millions of Web users daily.
That’s not about to change: Vicarious isn’t going to release its system publicly. And besides, as Luis von Ahn, one of the creators of the Captcha, points out, many people have shown evidence of computerized Captcha-solving over the years. Von Ahn even helpfully passed along a link to a list of such instances.
Become an MIT Technology Review Insider for in-depth analysis and unparalleled perspective.Subscribe today