About DeepMind on Geektimes the lazy author did not write unless. This company really stands out for its achievements among other organizations that work in the field of AI. DeepMind's most famous project is
AlphaGo , an AI specializing in go. At the moment, this system is perhaps the most skillful go player in the world, both among people and among machines.
But the game of go - only a demonstration of the capabilities of AI, is not the only project of the company. Now she is engaged in teaching a weak form of AI navigation intricacies. The usual method of orientation in the surrounding space with the advancement to the desired goal includes a constant assessment of what surrounds the person (or animal) with the analysis of the information received. DeepMind specialists have developed a complex of neural networks that have trained to move around a square-shaped platform like rats.
We are talking about rats moving on the same platform. The computer system received data on the speed of movement of the rat, the main directions of movement, distance from the walls and all other parameters. On the basis of these data, the AI developed a practically similar method of movement, and not a spontaneous one, it is based on certain factors that are used in the navigation in space and the rat.
Interestingly, the computer system has developed a special layer for navigation, the principle of operation of which is similar to the principle of operation responsible for orientation in the mammalian brain space.
In humans, several groups of cells are responsible for this. These are neurons that are activated when a person passes by objects that he has previously met. These neurons are located in the hippocampus - a part of the brain that is responsible for memory formation and information storage. As far as can be judged, specialized cells are allocated for this purpose. The location of these cells is shown in the announcement picture.
Experts of DeepMind believe that the neural network during training has formed a similar structure, only, of course, digital, and not physical. And in the same company they say that only the neural networks that have formed such structures are able to navigate normally in a complex environment - not only on a square platform, but also in a building with several rooms. Moreover, neural networks can adapt the paths of advancement in such a complex environment in the event that something changes (for example, the doors are closed or furniture moves).
Based on the project results, several conclusions can be drawn. For example, one of them - the method of orientation in space, developed by animals, is optimal. Hippocampal neurons are a great way to memorize a path and pave the way among already known objects after a fairly long stay in a location.
It seems that not only biological, but also digital systems come to similar solutions in some cases.
DOI:
10.1038 / s41586-018-0102-6
