Scientists from the U.S. created a learning AI, capable of breaking any captcha

American engineers have created a new system of artificial intelligence capable of independent learning and adapted it for breaking text captchas used by webmasters to protect sites from bots. Instructions for creating this AI were published in the journal Science.

“The ability to learn using only a small set of examples, and ability to find similarities in different situations are the hallmarks of people who still remain inaccessible to machines. Using the experience of system neurophysiology, we have created a new model for computer vision that recognizes the captcha’s doing it better than deep neural networks, and is 300 times more effective,” said Dileep George (Dileep George), one of the founders of IT-AI startup Vicarious.

The main drawback of all existing neural networks and artificial intelligence systems is that they are unable to learn new skills. Engineers literally have to teach them what they should do, using a huge data archives that are manually processed by the person, or with the direct participation of the people, showing the machine how to solve the problem.

This property severely limits the applicability of AI systems in real life, such as the program in principle not able to learn by watching the actions of people or other machines, and to pick up new skills and knowledge «on the fly» as a man. Moreover, the capabilities of such AI systems, in fact, are limited to how well they have taught the people that, in principle, does not allow them to enter the «superhuman» level.

George, a neurophysiologist by training, and an IT entrepreneur Scott Phoenix (Phoenix Scott) six years ago founded the project Vicarious aimed at eliminating this disadvantage the AI. In 2013 they presented their first design – the so-called «recursive cortical network»(RCN), a special subspecies of the neural network, imitating the work of the visual cortex.

AI system built on the basis of the RCN, as they said George and Phoenix, were able to crack approximately 90% of text captchas used at that time to protect the sites. Almost all of the experts in this field do not believe the founders of Vicarious, saying that they could manipulate the data or to build and train a neural network specifically for the selection of the answers to these puzzles.

This time Phoenix and George presented the results of their work in the prestigious scientific journal Science, creating a new version of the RCN, which can learn to recognize letters and digits of arbitrary shape drawn on the captchas, self-examining only 260 examples of such puzzles.

A key feature of this system, explains George, is that the perceived divides her picture on two types of objects, surfaces and contours. For their detection is responsible for two separate neural networks, organized roughly the same way as the layers of neurons in the visual cortex of the brain, and the results of their work integrates another network consisting of two layers.

This approach, as they note, allows you to split letters, numbers and other objects into sets of individual elements, some of which is common to many characters that accelerates the process of learning the network and increase the recognition accuracy for even the most stylized and ambiguous labels.

As shown by the first experiments with this AI system, only three hundred rounds of practice tests on simple captchas enough to RCN learned to crack the security of Paypal, Yahoo, reCAPTCHA and many other popular systems «bot-check» with 66% probability. Interestingly, the chances of success did not depend on the distance between the letters and their shapes, the two main «enemies» of conventional neural networks.

This indicator, as they note, exceeds the accuracy of all available machine learning systems designed for similar purposes, and it can be considerably increased if you allow the AI train longer, and use the fonts that are used by creators of these puzzles.

What is important, breaking captchas is not the only purpose of existence RCN – this same system can be successfully applied to very different problems, for example, sorting images, or determine which items are present in the pictures and the pictures that she makes with extremely high accuracy. As scientists hope that their algorithm will find application in solving scientific and practical problems.