Google’s DeepMind research team isn’t just creating game algorithms like the well-known AlphaGo that can beat the champions in Go. Artificial intelligence specialists are also trying to understand how the human brain works in order to apply this knowledge to create AI.
The foundation of any training is positive reinforcement. When a person receives new information or solves a difficult problem, he receives a surge of dopamine. This is reminiscent of animal training: for the right trick, the trainer throws a banana or piece of meat into the cage.
When training a neural network, the same method is sometimes used: artificial intelligence discovers that for the necessary actions it is awarded points and tries to maximize profits.
However, it is much more interesting how people manage to motivate themselves for long-term results. Learning a foreign language or continuing education is not always fun. Intellectually, we understand that this will lead to good grades, well-paid jobs and the ability to travel more. But this will not happen immediately.
Researchers at DeepMind have suggested that the brain is able to predict rewards. Moreover, they claim that the brain does not just predict, but also uses a complex mathematical system. This theory is covered in a Nature article written by Google.
This is how developed artificial intelligence works: it does not calculate the average value, but calculates the probability. This approach helps him achieve better results.
The proof of the theory is a study on mice by Harvard scientists. The animals were placed in an environment where for various actions they could receive both a very large and a very small reward. Naturally, the mice struggled to figure out how to hit the big jackpot. Scientists measured the neural activity of the experimental subjects and found that neurons actually create something like a mathematical system.
It turns out that while we are busy with our own affairs, our brain is constantly looking for what actions will lead us to the most beneficial results.