We humans beings and most of the animals can easily navigate in our surroundings through complex processing systems in the brain that grants the ability to explore new areas, memory retention of previously visited places and taking shortcuts.
While such abilities come naturally to us without ever realizing how complex the underlying processes are, the same cannot be said for artificial agents — facing difficulty in spatial navigation.
But new research conducted by DeepMind shows that an artificial neural network was able to spontaneously develop a system equivalent to the mammalian brains which helped the system to navigate successfully in space.
The mammalian brain has an internal coordinate system that works through neurons known as “grid cells.” These cells are hypothesized to support vector-based navigation which enables the brain to calculate the distance and direction between source and destination.
Scientists at DeepMind first trained the artificial neural network to navigate in a virtual environment by using movement-related velocity signals — commonly used by mammals to traverse through unfamiliar territory.
The neural net kept on amassing information such as speed, direction, distance from walls and other details. Later, they found that the network had successfully learned to navigate this virtual space through a self-developed layer of grid-like representations.
This was quite surprising as the spontaneously emerged grid units had a striking similarity to the neural activity patterns observed in foraging mammals and it used the exact same system that mammalian brains use to navigate.
The researchers also found through a series of experimental manipulations that silencing the grid-like units resulted in impairment of the artificial agent’s ability to navigate accurately, thereby confirming that grid-cells are crucial for vector-based navigation.
This finding now serves as another stepping stone towards our understanding of the inner workings of the human brain. It has also opened up new potential aspects where artificial beings could be used to display more complex behaviors within realistic environments.