One of the main ways to protect the Internet from the invasion of bots is captcha - a computer test by which the system determines whether a user is a human or a computer. This is one example of a Turing test. The main idea of the test is to offer a problem that can be easily solved by a person, but impossible for a computer. And it seems that a group of scientists from the United States managed to create an artificial intelligence algorithm that is able to pass this test.
As reported by the editors of Science magazine, the American company Vicarious set a goal a few years ago: using machine learning methods to develop a program that can recognize captcha. The most common type of captcha today is ciphers of letters and numbers, superimposed on each other, translucent, crossed out, and so on. In general, any web user has come across them many times.
Vicarious has created an algorithm called the Recursive Cortical Network (RCN). It is a hierarchical structure that models an object as a combination of contours and surfaces. At the same time, the RCN algorithm is able not only to recognize captcha, but also to "see" handwritten characters. As the creator of the algorithm, Dilip George stated, “Using the experience of systems neurophysiology, we have created a new model of computer recognition that recognizes captcha better than deep neural networks, and at the same time it works 300 times more efficiently. Our program can independently learn to recognize numbers and letters of arbitrary shape drawn on captchas by studying only 260 examples."
Vladimir Kuznetsov