How an American scientists created AI, able to bypass any captcha
American company Vicarious has taught artificial intelligence to solve CAPTCHA with record accuracy. The project is fundamentally different from past developments: instead of analyzing thousands of examples of ready-made captcha, the system sees the letters as a person. Researchers called the technology a recursive cortical network, told about the principle of its work and efficiency.
In human vision, different groups of neurons recognize the surfaces and outlines of objects. Instead of perceiving objects and views as a collection of different details, neurons communicate with each other and determine which parts belong to the same thing. After reconstruction and recognition of the object, the general view is built hierarchically, based on entire objects, and not on individual details.As a result, a person is able to recognize objects, even if they are mutated or distorted. For example, we can read inverted and blurry text.
The recursive cortical network imitates these principles, recognizing the contours, features and internal structure of objects. Other network structures are responsible for recognizing the surface characteristics specified by these loops. Then the received information is combined into arrays, which are linked together and base groups of adjacent parameters of the object. They are built into a hierarchy, on top of which are hypotheses about the final subject. The network evaluates each assumption, determines the most probable and recounts them for location in a general two-dimensional space. After several such checks, the technology detects the object even in spite of minor changes in shape and position.
Vicarious AI tested the efficiency of a neural network using CAPTCHA as an example: with distorted symbols from the font Georgia, the system coped with an accuracy of 94%. It is interesting that a person solves such problems from the first time only in 87% of cases. With the BotDetect system from Yahoo and PayPal, the development handled somewhat worse – with 57% accuracy. Typical artificial intelligence would need 50,000 pre-decided “captcha”, while a recursive cortical network was taught only on 260 pictures with individual symbols.
The authors of the technology called the project a big step for artificial intelligence. In the future, the network can be trained and other tasks: for example, the recognition of plain text for instant translation.