Google DeepMind has developed an algorithm that orientates itself in space using an artificial analogue of neurons in a lattice.
DeepMind, the AI research arm of Google, has created a program that is capable of constructing optimal routes using an analogue of the neurons of the grid. These cells are part of the brain network that provides navigation in all mammals, including humans. In the future, the new development will allow us to study our orienteering abilities without testing on animals. The technology article was published in the journal Nature.
Another program, created by DeepMind, has repeatedly defeated the world's strongest Go masters, a game that has long been considered immune to artificial intelligence.
The authors of the new algorithm have created an artificial analogue of the lattice neurons. These brain cells are activated when the mammal crosses the boundary of an imaginary grid "superimposed" on the space in which the animal is located. In humans, the destruction of these neurons becomes one of the symptoms of Alzheimer's disease, and people lose the ability to navigate. Scientists suggest that lattice neurons help find the shortest paths in familiar environments.
In a new study, the developers modeled two artificial recurrent neural networks. In such networks, communications between elements form a directional sequence: the program uses its previous steps to plan the next action.
One algorithm used artificial lattice neurons, the second did without them. The programs were trained to look for a path in virtual mazes, where the shortest path to the goal was blocked by a locked "door". Then the algorithms moved on to larger mazes of a similar configuration: the program using the neurons of the lattice looked for a path more efficiently. When the doors were opened, the algorithm was able to take this fact into account and found the shortest route. The program, which worked without special neurons, ignored the opened passage and looked for a path in the maze longer.
The results of the experiment confirmed the hypothesis of neuroscientists: lattice neurons are indeed involved in the search for the fastest path. Artificial intelligence modeling could replace some types of animal experiments over time, experts say.
Natalia Pelezneva
Promotional video: