Help the AI with a fragile snowflake
What the study of snowflakes and neurons in the brain can give
Scientists at Imperial College London have put forward and confirmed the hypothesis that the use of different - and not identical, as is done now - artificial neurons in the creation of artificial intelligence (AI) increases the efficiency of artificial neural networks. The idea that artificial neurons should be different was prompted by the neurons of the human brain and snowflakes.
As you know, in nature there are no two identical snowflakes. It's the same with the neurons in the mammalian brain - each one is unlike any other. At the same time, artificial neural networks created by humans, although they imitate the work of the brain, however, the neurons used in them are absolutely the same. According to scientists at Imperial College London, this is why the human brain is still superior to AI in many ways, for example, it learns faster, adapts to changing conditions, and switches from one task to another. Their research is published in the journal Nature Communications.
Computational neuroscientist Daniel Goodman, who participated in the study, in an interview with NEO.LIFE, explained the difference between the ability to learn and adapt to changing circumstances in the human brain and AI. Artificial intelligence can be trained to play, for example, the arcade video game Pong, famous in the 80s. “There are two rackets moving along the edges of the field and alternately hitting the ball. A trained AI will play this game perfectly. Better than a human, - says the scientist. - 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. " A person will not have such a problem, says Mr. Goodman. The reason for this, the scientist believes, is that all neurons in the human brain are different.
The Imperial College Intelligent Systems and Networks Lab tried to slightly alter every cell in an artificial neural network modeled after the brain. As a result, the efficiency and accuracy of the artificial neural network has increased by 20%.
As part of the study, scientists at the Imperial College studied the principle of communication between neurons in the mammalian brain. It does this through electrical impulses that are precisely timed, and these impulses are fundamentally different from digital and analog computational operations. By modifying the neurons in AI systems so that they differ from each other, like neurons in the brain, and making them as closely as possible reproduce the impulse work of the brain's networks, scientists have been able to improve the efficiency of these systems. In particular, the performance of AI has improved in speech recognition, receiving and interpreting voice commands. And the change in the activation time of artificial neurons made it possible to increase the efficiency of performing tasks with a time component, such as recognizing numbers pronounced in a row.
Meanwhile, Parta Mitra, a neuroscientist at Cold Spring Harbor, New York, USA, is convinced that it’s not so much the variety of neurons in AI systems, but how they are arranged.
For example, he says, in electronic circuits, all the components are the same, but depending on how they are lined up, this circuit will be used for a specific type of appliance, be it a radio or a washer-dryer.
His British colleague notes that this can also contribute to increasing the efficiency of AI systems, for example, in solving problems related not to temporal, but to spatial components. And both scientists agree that in the near future there will be AI systems built from artificial neurons (made of silicone) - neuromorphic systems. Such systems will mimic another important feature of natural neural networks - plasticity. It is the ability of the brain to change its wiring by growing and reorganizing physical connections with other cells.
It is plasticity that can become the most important property for creating autonomous AI systems that can interact with the real world, as the human brain does, adapting to new circumstances, allowing a person to recognize danger and avoid it, regulate motor activity by analyzing sensory information, etc. And this plasticity could allow AI to learn, for example, to play Pong with variable parameters without error, scientists say.
Post a Comment