Unless linked to the actual world via robotics, AI is unlikely to develop human-like cognition

According to research from the University of Sheffield, connecting artificial intelligence systems to the physical world through robots and building them using evolutionary principles are the most likely ways for AI to develop human-like cognition. No matter how large their neural networks or the datasets used to train them might become if AI systems remain disembodied, they are unlikely to resemble accurate brain processing, according to Professor Tony Prescott and Dr. Stuart Wilson from the University's Department of Computer Science. According to the Sheffield study, biological intelligence, such as that seen in the human brain, has evolved due to this design and how it has leveraged connections to the outside world to overcome obstacles, pick up new skills, and improve over time. According to the study, the relationship between evolution and development should be considered while designing AI. Modern AI systems like ChatGPT use large neural networks to tackle challenging issues like producing understandable written language. These networks train AI to process data in a manner that is modeled after the human brain while also learning from errors to become better and more accurate. Although these models resemble the human brain in specific ways, the Sheffield researchers claim significant discrepancies prevent them from developing biologically equivalent intelligence.

First, brains are physically present in the human body, directly perceiving and responding to the outside world. For disembodied AIs, which can learn to recognize and develop sophisticated patterns in data but lack a direct link to the real world, being embodied gives brain functions significance in a manner that is not conceivable. As a result, such AIs are unaware of and unable to comprehend their surroundings. Second, human brains are made up of various subsystems arranged in a certain way, or architecture, shared by all vertebrate creatures, including fish and humans, but not AI.

"ChatGPT, and other large neural network models, are exciting developments in AI that show that tough challenges like learning the structure of human language can be solved," said Professor Tony Prescott, Professor of Cognitive Robotics at the University of Sheffield and Director of Sheffield Robotics.

However, if they continue to be created using the same techniques, these AI systems are unlikely to get to the point where they can thoroughly think like a human brain. Suppose AI systems are created with designs that learn and grow similarly to how the human brain works, leveraging its links to the actual world. In that case, the likelihood that they will eventually attain human-like cognition increases significantly. These links can be given to AI systems through robotics, such as actuators like wheels and grippers and sensors like cameras and microphones. Then, AI systems could detect their environment and learn similarly to the human brain.

The Sheffield academics claim that recent developments in the creation of AIs for robot control have made some progress. Recurrent neural network models are one effective method comprising several feedback loops and are taught to produce better predictions about what can happen next.  These models represent significant advancements in the field of robot adaptability. The study contends that robot AIs still need to accurately simulate how various brain subsystems interact as a part of a more comprehensive cognitive architecture.

"Efforts to understand how real brains control bodies by building artificial brains for robots have led to exciting developments in robotics and neuroscience in recent decades," stated Dr. Stuart Wilson, Senior Lecturer in Computational Neuroscience at the University of Sheffield. After looking through some of these initiatives, which have mainly concentrated on how artificial brains can learn, we believe the subsequent advances in AI will result from more closely imitating how actual brains develop and evolve.