New AI knows how to take shortcuts

Struggling with a labyrinth, deep learning , a particular approach of artificial intelligence that has as its model the networks of neurons, has proven to be able to find creative solutions, such as taking shortcuts, just like a mammal would. The analysis then showed that for this task the deep learning networks spontaneously reproduce the activation patterns of the grid neurons, the GPS system of the brain.

Credits

---

Deep learning , is an approach to artificial intelligence that, by drawing inspiration from the brain's neural networks, is helping to develop different technologies, from automatic video analysis to language interpretation. But it can also reveal - with a curious feedback - important information on how the natural neuronal networks work, helping in particular to understand how the regular geometric representations of space can facilitate flexible strategies for the orientation of an animal.

This is demonstrated by a new study published in "Nature" by Andrea Banino of DeepMind, perhaps the most famous private company that deals with artificial intelligence, and colleagues of the University College of London.

Deep learning networks have the characteristic of being able to learn how to process the inputs to obtain a particular result. They are defined as "deep" because they are made up of a hierarchy of computational units repeated in several layers. Each unit receives input from similar units in the upstream level and sends its outputs to those downstream.

Deep learning demonstrates its almost prodigious potential in the perceptive field: for example, learn to choose a particular face by analyzing, following certain instructions, many photos of different people.

Few studies have been dedicated instead to trying to reproduce complex tasks of the animal mind, such as orientation in space, with deep learning.

Credits

---

In the human brain, navigation skills are dedicated to specialized neurons also known as GPS neurons, whose discovery is worth the Nobel to John O'Keefe, May-Britt Moser and Edvard Moser in 2014. Among the GPS neurons, present within hippocampal formation - a region that, in humans, is involved in space learning, in autobiographical memory and in knowledge of generic information - grid neurons are particularly important.

The latter are activated according to a regular spatial scheme to help humans and other animals keep track of their position. Their working principle is therefore known, but their specific computational functions have so far been difficult to understand.

Based on the principles of deep learning networks, Banino and colleagues have trained an artificial agent - basically a virtual mouse, able to imitate the basic behavior of a real mouse - to navigate in simulated unfamiliar environments. After a certain period of learning in finding the way from point A to point B in a maze, the agent has become so competent as to start taking shortcuts, just like a mammal would do, overcoming a busy expert human being in efficiency in the same task.

Credits

The researchers found that these amazing abilities were based on an activation scheme of deep learning networks very similar to that of grid neurons.

This structure has not been intentionally imposed a priori by the researchers, but has emerged spontaneously, reinforcing the remarkable capabilities of the virtual agent. This shows that the role of the grid neurons goes some way beyond processing a GPS-like position signal and gets to plan the most direct route between two points, a bit like navigation programs like Google do Maps, also based on GPS coordinates.

The study helps to explain how grid neurons are able to encode spatial information. The emergence of similar units in a simulation is an impressive example of a specific potential of deep learning: to invent an original internal representation, often not foreseen, which helps to solve a task.

References for Further Reading

Phys

Gadgets