AI Systems Make Mental Maps Much Like Mammal Brains
AI systems are becoming more and more human-like by the day. In fact, that’s the goal. These systems are supposed to think like intelligent humans think. They must learn, process, and respond to information just like us. Sophia, the world’s first robot to receive citizenship, which was granted by Saudi Arabia, hasn’t even really left the media spotlight with her sharp wit and subtle smiles. Systems like her can remember your face and respond emotionally, switching her tone depending on how she’s feeling about the conversation. She even made a playful threat to ‘dominate the human race’ on the Tonight Show with Jimmy Fallon. These systems have shown to out-play us in games like chess and poker, they can even beat us in their object recognition speed. The one (OK maybe we have more than one) thing we had on these systems was our ability to orient ourselves in space and time. We wake up in the morning and our brains give us a mental rundown of who we are and where we are every day. It allows us to quickly remember locations spatially so that we can return to them, form associations between new and old areas and even develop shortcuts.
A recent study performed by Google’s artificial intelligence group, Deep Mind, may have changed this entirely. These processes that we all take for granted boil down to grid and path cells. These are neurons in our brain firing at different rates, literally telling us where we are and how to get somewhere else. They are our mental maps. Neuroscientists have only just pieced this part of our brain together. In 2014 the Nobel Prize in Physiology or Medicine was awarded to John O’Keefe, May-Britt Moser, and Edward Moser. John O’Keefe in 1971 discovered the first component of our mental pathways, which allows us to determine where we are in time and space. These ‘path cells’ as he called them were discovered in rats initially as they were always activated in a particular region of the brain when the rat was in a particular spot in a room. When the rat was in different place a different cell activated, yet it was always specific to the location of which the rat was in.
That discovery was left alone for another 34 years until May-Britt and Edward Moser discovered the missing piece to this theory. They described grid cells that generate an actual coordinate system in our brain. These neurons were shown to fire when the animal was at any end-point along a hexagonal grid, which implies that our brains form hexagonal shapes to map out our environment. For example, if you were to draw tessellating hexagons on your floor with one dot in the center of each of them until you’ve covered the whole room, that’s how your brain makes sense of the world. Along those hexagons, every time you were to step on either a center point or the end-point of a hexagon is when your brain would fire off a neuron.
Early in May, Deep Mind discovered that when AI systems were implanted with a network of artificial neurons, they too made grid cells that operated similarly to the ones in mammals. These systems developed brain-like activity from scratch. Essentially the researchers created a ‘deep neural network’, this is a computer program with layers and layers of artificial neuron code in order to process information. The AI systems were then trained to navigate through virtual mazes and scenarios with feedback on its performance, such as reward signals. Through trial and error, these AI systems learned that firing these neurons in hexagonal-like structures was the most effective way to map out their terrain. They developed something almost instantly that it took evolution millions of years to develop. This new system of firing allowed the AI systems with the nerve cells to take shortcuts in the virtual maze when a quicker outlet was provided, such as an open door. The AI systems without nerve cells still took the long way around regardless of the door.
This is an incredibly large milestone for AI technology and it’s fascinating how many implications this could have in the understanding of our own brains. It took almost 35 years for an advancement in how our neurons operate to give us a sense of position and direction. AI systems could help solve mysteries of the brain that philosophers and neuroscientists alike have pondered for centuries. Imagine if we could use AI systems to understand, for example how our brain gives us signals to move our limbs or even our eyes as we read. We wouldn’t require humans or other mammal subjects for research if the AI systems could do just that. Although these systems seem to be beating us in the races (probably lapping us a few times), maybe this is an advantage rather than a detriment. AI systems could quite possibly help us understand ourselves better than we could. Maybe these computer systems can teach us what it is to be human.