According to research conducted by the University of Sheffield, connecting artificial intelligence (AI) systems to the physical world through robots and designing them using principles from evolution is the most promising approach for achieving human-like cognition in AI. In their paper published in Science Robotics, Professor Tony Prescott and Dr. Stuart Wilson explain that even as AI systems grow in size and complexity, they are unlikely to resemble the processing of real brains if they remain disembodied.
Current AI systems, such as ChatGPT, utilize large neural networks to tackle complex problems, such as generating coherent written text. These networks are trained to process data in a manner inspired by the human brain and learn from their mistakes to enhance accuracy.
While these models share some similarities with the human brain, the researchers from Sheffield emphasize that there are significant differences preventing AI from attaining biological-like intelligence. Firstly, human brains are embodied in physical systems—the human body—allowing for direct sensory perception and interaction with the world. This embodiment gives meaning to brain processes in a way that disembodied AI systems cannot achieve. Although these AIs can recognize and generate complex patterns in data, they lack a direct connection to the physical world and, therefore, lack understanding or awareness of their surroundings.
Secondly, human brains consist of multiple subsystems organized in a specific configuration, known as architecture, which is similar across vertebrate animals. However, this architecture is not replicated in AI systems.
The Sheffield study suggests that biological intelligence, as seen in the human brain, has developed due to its specific architecture and its interaction with the real world, enabling it to overcome challenges, learn, and evolve throughout the ages. The researchers highlight that the design of AI systems rarely incorporates this interaction between evolution and development.
Professor Tony Prescott, Director of Sheffield Robotics and Professor of Cognitive Robotics at the University of Sheffield, states that while models like ChatGPT are impressive in solving complex tasks such as understanding human language, they are unlikely to achieve full human-like cognition if they continue to be designed using current methods. He proposes that AI systems are more likely to develop human-like cognition if they are constructed with architectures that learn and improve in a similar fashion to the human brain, utilizing connections to the physical world facilitated by robotics, such as sensors and actuators. By sensing and interacting with the world, AI systems can learn and develop akin to the human brain.
The researchers note that progress has been made in developing AI systems for controlling robots, with the use of recurrent neural network models showing promise in making robots more adaptable. However, these robot AIs still have a long way to go in terms of replicating the comprehensive cognitive architecture of real brains and how different subsystems interact.
Dr. Stuart Wilson, Senior Lecturer in Computational Neuroscience at the University of Sheffield, suggests that significant advancements in AI will arise from closely mimicking how real brains develop and evolve, building artificial brains for robots to understand how they control physical bodies.
Overall, the research from the University of Sheffield indicates that connecting AI systems to the real world through robotics and incorporating principles of evolution in their design is a more promising avenue for achieving human-like cognition compared to purely disembodied AI systems.
Source: University of Sheffield