Californian computer scientists have developed a new form of AI to crack the popular Captcha security system.
Captcha stop bots from automatically signing up to their sites by asking users to identify distorted numbers and letters.
It’s a task that has traditionally been difficult for computers to perform, but the system devised by the researchers appears to have picked it up exceptionally quickly.
Rather than deploying neural networks, which have to be trained on thousands of images, the team built a system inspired by the brain’s visual cortex.
“The model outperforms deep neural networks on a challenging scene text recognition benchmark while being 300-fold more data efficient,” the researchers write in the journal Science.
The team, which works for Vicarious – an AI startup backed by Amazon’s Jeff Bezos and Facebook’s Mark Zuckerberg – say the software can beat Google’s reCaptcha test 66.6 per cent of the time. The human pass rate is 87 per cent.
The researchers say it may have a wider impact on computer science too: “Our model emphasizes aspects like data efficiency and compositionality that may be important in the path toward general artificial intelligence.”