Things that are completely obvious to humans for artificial intelligence can become an insoluble task, so DeepMind decided to teach a neural network to think abstractly and connect to other neural networks in order to help and jointly solve complex problems.
Teaching AI to recognize pictures is as easy as shelling pears, but determining which of the objects is larger than the other, if they are at different distances, is already more difficult. To do this, DeepMind developed a neural network that specializes in such tasks, and then connected AI to it to improve its capabilities.
During the AI training, he was shown images of various three-dimensional objects, and then they began to ask the new program difficult questions, forcing him to think rationally. When asked what is the size of the cylinder located to the left of the brown metal piece, which is to the left of the large ball, the artificial intelligence connected to the neural network answered fairly quickly and correctly. Developers prepared quite a lot of such tasks for AI, but in 95 percent of cases the program answered correctly, showing a small gap from human rivals.
Google's neural network began teaching other AIs back in January this year, but now it can teach them much more interesting lessons. Practical application of the development has not yet been invented, but the very fact that artificial intelligence has become smarter and smarter is definitely pleasing.