This Neural Network can Replicate our Brain’s Navigation System, with Human Level Accuracy!
- Researchers, primarily from DeepMind, have developed an artificial neural network that can grow brain-like cells
- Navigational artificial intelligence will now be able to identify shortcuts in a maze-like environment
- The grid units in neural nets were remarkably similar to grid cells in our human brains
Finding a shortcut to anything has always fascinated human beings. From saving time to cutting down on budgets, finding the shortest path to the end goal is quite often the ultimate aim. But what if machines could do that for us, with the same level of accuracy and precision that we manage?
It might be a very real possibility now.
According to a report presented in Nature last week by researchers at DeepMind, a navigational artificial intelligence system can explore complex simulated environments (like a maze) and find the shortest route to an end destination, like humans and animals! It works like a GPS system, and the AI manages to carve a path around the environment (or maze) with remarkable skill and accuracy.
The AI consists of an artificial neural network that uses the concept of grid cells. A grid cell is a type of neuron in the brains of many species that allows them to understand their position in space. While learning how to navigate, the neural net spontaneously develops the equivalent of grid cells.
For investigating the role of grid cells in navigational functions, the researchers attempted to use deep-learning neural networks. They have also released the approach they followed, which we have summarized below:
- The researchers used the ANN to navigate through previously unseen environments by adding reinforcement learning to their model
- A neural net was set up to learn to perform path integration for a simulated agent moving through a small space
- The grid units that emerged in the network were remarkably similar to what’s seen in animals’ brains, right down to the hexagonal grid. (see the image below for this)
- The research team then joined the neural network’s abilities to systems that helped simulated agents find their way through maze-like virtual environments, towards the set target, or goal
- The system with grid units was far superior to systems without. For example, the networks with grid cells could tell if a previously closed door offered a shortcut to the goal, and it would take that route preferentially, while other systems continued to take any available routes.
- The grid units in the neural net were performing vector-based navigation because they were identifying a shorter, more direct route based on knowledge of the target goal’s position
Our take on this
The researchers have admitted they came upon this algorithm unexpectedly but it’s ended up being a truly valuable research for the community. AI researchers can use this for improving existing automated navigation systems (maybe self-driving cars, or helping robots). This approach is not only restricted to navigation though. It can also be used for testing theories related to brain functioning, a field which has seen tons of coverage but little breakthrough.
Subscribe to AVBytes here to get regular data science, machine learning and AI updates in your inbox!
Leave a Reply Your email address will not be published. Required fields are marked *