Research has found that artificial intelligence (AI) systems are unlikely to acquire human-like perception unless they are connected to the real world through robots and designed using evolutionary principles.
Cognition is the mental process of acquiring knowledge and understanding through thoughts, experiences, and sensations.
The study, published in Science Robotics, found that if an AI system remained disembodied, it would be able to adapt to real brain processing, no matter how large the neural networks and datasets used to train them. turned out to be dissimilar.
Researchers at the University of Sheffield in the UK noted that current AI systems such as ChatGPT use large-scale neural networks to solve difficult problems, such as generating easy-to-understand sentences.
These networks teach AI to process data in ways inspired by the human brain, and can also learn from mistakes to improve and become more accurate.
Although these models have similarities to the human brain, they also have important differences that hinder the acquisition of biological intelligence, the researchers said.
First, they said, the actual brain is embedded in the physical system that directly senses and acts on the world: the human body.
Being embodied makes brain processes meaningful in ways that disembodied AI cannot. AI can learn to recognize and generate complex patterns in data, but lacks a direct connection to the physical world, researchers said.
Such an AI would therefore be unable to understand or perceive the world around it, they said.
Second, the human brain is made up of multiple subsystems, organized in specific configurations known as architectures. This configuration is similar in all vertebrates, from fish to humans, but it is different in AI.
This research shows how biological intelligence, like the human brain, was developed by this particular structure and how throughout evolution it has used its connection to the real world to overcome challenges, learn and improve. suggests.
Researchers say this interplay between evolution and development is rarely factored into AI designs.
“ChatGPT and other large-scale neural network models are exciting developments in AI that show that they can solve very difficult problems, such as learning the structure of human language,” said Professor Tony Prescott of the University of Sheffield. Stated.
“But these kinds of AI systems, even if you keep designing them in the same way, are unlikely to progress to the point where they think completely like the human brain,” Prescott said.
“It is unlikely that AI systems will develop human-like cognitive capabilities if they are built with architectures that use their connections to the real world to learn and improve in ways similar to the human brain. It will be very expensive,” he added.
