NPR reports that the new company Vicarious said that they have used AI to develop an algorithmic model that can be used to "completely differentiate the computer and human brain from the public Turing test", which is now very common, known as CAPTCHA verification mechanism The
CAPTCHA is mainly in the picture shows a set of verification code, by the confusion of letters, corrugations and some background noise combination, similar to the test in the past for computer interpretation is very difficult, and therefore CAPTCHA's ability to be some AI researchers As a key benchmark for testing.
This is not the first time someone has successfully cracked CAPTCHA's verification, about 10 years ago, a technology company had been trafficking tickets Ticketmaster prosecution, because the CAPTCHA system around the purchase of concert tickets.
But in the past, the use of CAPTCHA in the specific weaknesses to prevent minor changes in the system to bypass the system, but Vicarious co-founder Dileep George pointed out that this new study is "fundamentally" down CAPTCHA.
In the past, the use of deep learning (Deep learning) to train AI to respond to specific things, through such training, AI will be able to learn to identify the different letters in the picture, but once the letters overlap each other, AI will become difficult Identify.
George explained that in-depth learning is only part of the way to copy the human brain, as people will learn from the same experience, but the human child does not need to read a large number of different forms of letters to distinguish, even if the letter is relatively large Italics, children can still easily identify out.
George said that they found the brain in the visual, there will be some assumptions, but the depth of learning in the past did not have the same situation.To make AI also have this feature, Vicarious team to use recursive cortical network (RCN) technology, So that AI even in less training, but also be able to better reason to see the content.
In the training phase, AI will see the letter after it will construct the concept of the letter model, understand and guess the outline of the letter, internal, background, etc., after the new picture appears, AI will try to use the past to establish To understand the picture, when the letter appears partially overlapping, it can understand some of the missing because the part of the letter hidden behind another letter.
Vicarious team mentioned, CAPTCHA has a very many types, but no matter what the algorithm to achieve more than 1% accuracy, can be considered as a crack. This AI algorithm model in the reCAPTCHAs test, the accuracy rate of 66.6 %, BotDetect reached 64.4%, Yahoo and PayPal are up to 57%.
Although Vicarious's calculus model does achieve good results through the CAPTCHA verification mechanism, the main goal of the study is to make the robot as visually as human as rational thinking and to have little to do with CAPTCHA.
George says the team's long-term goal is to build AIs like the human brain, CAPTCHA is a natural test, because it is able to detect whether the system can work like a human brain.
George emphasizes that robots need to know the world around and understand and manipulate objects. In areas where people intend to apply, AI has to deal with things with less inaccuracies and less flexibility in less training.
"This is the inevitable direction of technology, people must accept this thing: the computer will work like a human brain."
It is unclear how much impact the study will have on information security, and George points out that Google has begun to move from text-based CAPTCHA to more advanced testing. As artificial intelligence gets smarter, these validation mechanisms must A new way to ensure that the user's human identity.