The importance of grid cells in mammalian navigation
Deep learning can defeat any Go player in the world without human help, and can recognize human faces in seconds, but until now, deep learning algorithms have failed to do something that seems simple: spatial perception.
Most animals, including humans, can navigate flexibly in a living environment, such as exploring new areas, quickly returning to previous locations, or taking shortcuts. Bypassing obstacles, remembering a store, turning at an intersection... These capabilities are so natural to humans that we simply do not realize how complicated it is to complete the process. On the contrary, space navigation is an important challenge faced by agents, and in this respect they are far from animals.
When the brain thinks about where it is, it uses many cells. For example, place cells stimulate the cell when it reaches a specific location. Another example is the orientation of the animal's head that stimulates "head-directioin cells."
In addition, another important neuron in space navigation is a grid cell (also referred to as a "grid unit"), which is the area in the brain responsible for spatial learning, spatial memory, and common sense. In 2005, the researchers found that unlike other neurons, the shape of the grid neuron is a perfect hexagon. When the animal walks around in the environment, the neuron is excited, tracking and recording the trajectory of the object's movement. Points in a grid of hexagons are considered to support spatial navigation similar to the latitude and longitude lines on the map. In addition, these neurons are constantly updated. When an animal enters a new environment, it reactivates the same grid and adapts to the new environment. In addition to the internal coordinate systems of animals, recent researchers have assumed that their neurons (also known as grid cells) can also support vector-based navigation. In other words, let the brain calculate the direction and distance to the destination, just like "raven flying," which allows the animal to find its own route in a strange environment.
In 2014, the team that first discovered the grid cells won the Nobel Prize for Physiology or Medicine in recognition of their contribution to spatial cognition. But although it has been more than a decade since the discovery of grid cells, the computational power of grid cells and how they support vector-based navigation remain unclear.
Now, the brain behind the AI ​​that beats the world's best go players is becoming less mysterious, and DeepMind's discovery helps us think of the answer.
DeepMind's research expert Andrea Banino said: “What we think, remember, and feel is coded in the brain in some way. To understand this, we must learn how to study neurons, how to measure their activities, and to These activities are linked to our behavior. However, this is difficult to achieve in the real brain."
DeepMind did it, but instead of experimenting with real brains, it used neural networks and algorithms inspired by brain neurons to explain this problem. In DeepMind's recent paper published in Nature, they developed an agent to test the theory that grid cells support vector-based navigation. "This is consistent with our important idea of ​​using algorithms and brains for AI. The elements are very similar," the researchers said.
In the first step, DeepMind trained a Recurrent Network (RNN) to position itself in a virtual environment using the main velocity signals associated with the action. This ability usually occurs when the mammal walks into an unfamiliar environment or when it is not easy to find obvious landmarks (such as navigating in the dark).
They found that a grid-like representation (hereinafter referred to as "grid unit" representation) would spontaneously occur in the network, which is very similar to the pattern of neural activity observed in mammals that are foraging, and conforms to grid cells as The view that space provides valid code.
In their experiment, the agent generated a grid-like representation: the grid unit, which is very similar to the mammalian foraging bio-grid cells.
Next, the researchers tested whether the grid unit can support the theory of vector-based navigation by creating an agent. They see the agent as a virtual mouse. By combining the initial "mesh network" with a larger network architecture, an agent can be built with deep reinforcement learning, and its goal is to navigate in a virtual reality gaming environment. The researchers found that when it began to look for its own position, the grid-like shape began to form in the network. Some nodes are used more often than others, which is very similar to the shape of the mesh cells that the real animal is generating in the direction of the search.
The performance of this agent surpasses humans, is even better than professional game players, and demonstrates the flexible navigation behavior that animals have. It can open up new paths when necessary, and even “send near the roadâ€.
Through a series of experiments, we have proved that the grid-like representation is very important for vector-based navigation. For example, when the grid cells in the network disappear, the function of the agent navigation is weakened, and the key indicators (distance and direction to the target) become inaccurate.
Demonstration of vector navigation with grid cells. The bottom dot represents three different sizes of grid cells. When the agent moves, the grid unit will be activated to represent the current grid code, which will change, and the reaction agent will enter different fields. The grid unit is used to calculate the shortest distance to the target
According to Demis Hassabis, CEO of DeepMind, “The human brain is the only proof that we can create strong artificial intelligence, so it is entirely feasible to think of neuroscience as a new inspiration for algorithms. But we think this inspiration should be two-way. , it can both be inspired by artificial intelligence and can also help unsolved problems in the neuroscience community. This work is a good proof that by creating an agent that can navigate in a complex environment, we have proved that The importance of cells in mammalian navigation has deepened their understanding of it."
The research team believes that this study is very important for understanding the basic calculation purposes of the grid cells in the brain, and also shows that their help to the agents is also great. The results of the experiment provide support for the theory that the grid unit can provide a Euclidean space framework, which supports vector-based navigation.
More broadly, this study reconfirmed that the algorithm used by the brain can provide a reference for machine learning architecture. Previous research on grid cells by neuroscience has made the interpretability of agents easier. This is in itself a major topic in artificial intelligence research. When they try to understand its internal representation, previous research can give a reference. . This work also demonstrates the potential of using agents in virtual complex environments to verify how the brain works.
In the long run, similar methods can also be used to detect certain regions of the brain, such as where sounds are captured or where limbs are controlled. In the future, such networks may provide scientists with different experimental methods or propose new theories and even complete experiments that can only be done on small animals.
TV Antenna ,HDTV Antenna ,TV Antenna Best ,TV Antenna Walmart ,TV Antenna digital
Yetnorson Antenna Co., Ltd. , https://www.yetnorson.com