An artificial neural network reproduces the brain's navigation system

byPenn State

To the animal's brain,Lattice cellThere is a nerve cell that recognizes the sense of direction called relative positional relationship. Through the work of lattice cells, humans intuitively do shortcuts from a certain point to a certain point, or they can draw a map in the brain that is somehow walking. AI researchers succeeded in reproducing the function of such lattice cells using artificial neural networks.

Artificial Neural Nets Grow Brainlike Navigation Cells | Quanta Magazine
https://www.quantamagazine.org/artificial-neural-nets-grow-brainlike-navigation-cells-20180509/

The intuitive navigation ability of animals including humans includes Norwegian neurologistsEdward MoselMr. and his wifeMay Brit MoselLattice cells discovered in 2005 by Mr. are greatly involved. Mosels and Mr. and Mrs. Received the achievement of the discovery of the lattice cell, it won the Nobel Prize in physiology in 2014 in 9 years after discovery.

British AI companyDeep MindofAndrea BagninoMr. and University College LondonCaswell BerryDr. has questioned lattice cells also called "brain GPS" and started research to reproduce the function of lattice cells using artificial neural network. "Lattice cells update information on where you are based, based on how you have moved," Dr. Berry said.

Neuroscientists say lattice cells have a navigation system without external clues called "pathway integration". For example, the lattice cell recognizes navigation such as "walk 30 steps and turn 90 degrees to the right", not a route based on a landmark "turn around the convenience store next to the second signal" .

Also, lattice cells not only recognize paths but also have the function of measuring the distance taken from a certain point to a certain point,It has been shown in the research. Dr. Berry says, "Lattice cells can perform vector-based navigation based on objects and current position."

byAbhijit Bhaduri

AI's research team decided to investigate the role of lattice cells using deep learning and neural networks. Dr. Berry et al. Set up a neural network which learns route integration in a short distance and tried to determine whether it could construct an appropriate artificial neural network. Then, the artificial "lattice neuron" that appeared in the artificial neural network appeared to have a grid-like neuron firing model very similar to that seen in the animal's brain.

Dr. Berry who was surprised that "this is going to work" subsequently made a search for the maze of the virtual environment using systems with and without artificial lattice neurons. Then, it seems that the system with artificial grid neuron has far superior maze searching ability compared to the system without grid neuron. According to Mr. Banino, the lattice neuron has chosen a shorter, more linear route "vector-based" based on the knowledge of the target position.

"I think that we could prove the theory that lattice cells are used for path shortcuts in this study," Bunino said, even in experiments where it is difficult to use living animals, using AI to create a virtual environment By doing experiments with, we can obtain certain results.

byAnders Sandberg

Dr. Berry says, "The method of experimenting on the neural circuits of animals using this AI can be applied to other neuroscience issues." For example, it is possible to deepen insight into living animals by training artificial neural networks, such as limb control of animals, and performing experiments using robot arms and the like.

Although it is pointed out that "The result obtained by calculation is not necessarily used for actual animals as it is", it is pointed out that at least experiments are carried out with an artificial neural network to advance research on real organisms About the method, it is greatly appreciated.

"Some Deep Mind researchers may think that the navigation system using the artificial neural network obtained this time will be used to raise the performance of expedition robots and autonomous drone," but Mr. Banino According to the report, this experiment is the first step of a more ambitious plan. "Navigation systems are only one aspect of our intellect, creating my largest general-purpose algorithm called" brain "is my goal, I do not think about anything else," Bunino said Said.

byBryan Jones