GPS Navigation System Resembles a Brain, How Does It Work?

Artificial intelligence that is trained to trace the route through the virtual labyrinth has developed a new architecture resembling a neural GPS system found in the brain. Where artificial intelligence can find its way around the maze with skills that have never existed before. Artificial intelligence that tracks the environment like mammals can help solve the mystery of our own internal GPS.

Equipped with a virtual version of special brain nerve cells called grid cells [2] , artificial intelligence can easily solve and plan new routes through a virtual maze. It shows grid cells in the animal's brain play an important role in road planning.

This discovery comes from DeepMind, a British company owned by Alphabet and dedicated to promoting general artificial intelligence. The study suggests how artificial neural networks, inspired by biology, can be used to explore aspects of the brain that remain mysterious. But this idea must be treated with caution, because there is much we don't know about how the brain works, and because the function of artificial neural networks is often difficult to explain. Researchers at DeepMind began to train artificial neural networks to mimic path integration, a method used by animals to calculate their movements through a space. The researchers trained neural networks with feedback to trace the labyrinthine route by feeding examples of routes taken by mice that crossed the real maze.

The researchers found that neural networks develop something similar to "grid cells" found in the biological brain. These cells, arranged in a triangular box, seem to provide a way for animals to position themselves in physical space. DeepMind researchers use trained networks to trace routes through unknown mazes by adding reinforcement learning to their approach. They found that newly trained networks can trace routes far more effectively than previous artificial intelligence systems, and that it explores its space more like a real animal.

Neuro scientists suspect these cells, which have also been found in humans, might help not only provide mammals with an internal coordination system, but also plan direct paths between points. To test the idea, Neuro Scientist Caswell Barry at University College London, along with his colleagues at Google DeepMind, created an AI containing virtual nerve cells, or neurons, whose activity is similar to real tissue cells. The researchers train this artificial intelligence to trace the route of the virtual labyrinth by giving a system signal when it reaches its destination.

Neural networks can be used to do many useful things, but until now they have not proven good in terms of navigation. "This research is an interesting demonstration that deep learning can be a value for tasks that not only depend on perceptual abilities but also on higher cognitive functions in this case, spatial navigation," saidFrancesco Savelli , a neurologist at Johns Hopkins University who study grid cells.

Research shows that tissue cells play a fundamental role in how animals including humans find their way around the world. This discovery might eventually have significant practical benefits, such as helping robots trace unusual building routes more easily.

"Our work is to build artificial general intelligence, and we think navigation is a fundamental part of that," said Andrea Banino, a member of the DeepMind team. His colleague, Dharshan Kumaran said the next step was to get artificial intelligence agents to learn more complex navigation skills. "We think of a more challenging environment," he said.

DeepMind has previously shown some remarkable advances in machine learning, including programs that are able to learn how to play video games, as well as board games like Go and chess, with superhuman skills. This achievement also relies on training in very large or deep artificial neural networks.

According to Demis Hassabis, co-founder and CEO of DeepMind, research into artificial intelligence can reveal new things about the brain. "The human brain is the only proof of our existence that the kind of general intelligence we are trying to build is possible," he said in a statement. "We believe that this inspiration must be a two-way street, with insight also flowing back from artificial intelligence research to explain open questions in neuroscience."

However, it is not clear how far the neural network, which is a very simple representation of biology, will lead us to explain the brain. Some neurologists note that the workings of deep neural networks cannot be interpreted much more than the function of the biological brain. But there are limitations to using artificial intelligence to study the brain [3] . Because this system is intended for self-study, researchers cannot say why the system made special decisions, said neurologist Francesco Savelli at Johns Hopkins University. Virtual tissue cells clearly help artificial intelligence trace more efficient routes, he said, but it is still unclear exactly how artificial intelligence uses these cells.

Reference

[1] MIT Technology Review. 2018. AI programs get really good at navigation by developing a brain-like GPS system. Accessed from: https://www.technologyreview.com/s/611105/robots-may-someday-explore-the-world-using-features-borrowed-from-your-brain/ on May 10, 2018

[2] A. Banino et al. Vector-based navigation using grid-like representations in artificial agents. Nature. Published online May 9, 2018. doi: 10.1038 / s41586-018-0102-6

[3] F. Savelli and JJ Knierim. AI Mimics brain codes for navigation. Nature. Published online May 9, 2018. doi: 10.1038 / d41586-018-04992-7

Subscribe to receive free email updates: