Scientists at Imperial College London have confirmed their hypothesis that the use of different types of artificial neurons (computing units) in the creation of artificial intelligence (AI) systems increases the efficiency of neural networks. This idea was pushed to scientists by the structure of the human brain and snowflakes.
In nature, there are no two identical snowflakes, as well as identical neurons in the mammalian brain. According to British scientists, precisely because artificial networks have exactly the same neurons, the human brain still surpasses AI in many ways – it learns faster, adapts to changing conditions, switches from one task to another.
Computational neuroscientist Daniel Goodman, who participated in the study published in the journal Nature Communications, explained to NEO.LIFE the difference between learning and adapting to changing circumstances in the human brain and AI.
AI can be trained, for example, in the famous arcade video game Pong. “Two rackets move along the edges of the field, which alternately hit the ball. A trained AI will play this game perfectly. Better than humanGoodman says. – However, it is worth moving the rackets at least a pixel closer to each other, and the AI will not be able to play it, since it is trained only for specific parameters of the game and cannot cope with any, even the most insignificant, changes in it.“. Such a problem does not exist in humans, and the reason, according to the scientist, lies in the fact that all neurons in the human brain are different.
The Imperial College Intelligent Systems and Networks Lab has slightly altered every component in the neural network, modeled after the brain, and this has increased the efficiency and accuracy of its work by 20%. In addition, scientists tried to reproduce the impulse work of the brain networks as accurately as possible, which also increased the efficiency of the neural network: the performance of AI in speech recognition, receiving and interpreting voice commands improved. In addition, changing the activation time of artificial neurons made it possible to increase the efficiency of performing tasks with a time component, such as recognizing numbers spoken in a row.
In turn, Partha Mitra, a neuroscientist at the Cold Spring Harbor laboratory in New York, believes that it is more about the way neurons are arranged. And depending on how the neurons are lined up (even if they are the same), they can be used to solve different problems.
Mitra and his British colleague Goodman believe that in the near future there will be plastic AI systems built on chips with various artificial neurons – neuromorphic systems. And thanks to plasticity, an important feature of natural neural networks, AI will be able to learn, for example, to accurately play Pong with changing parameters, scientists say.