On a groundbreaking studyScientists from Cambridge have taken a new approach to artificial intelligence, demonstrating how physical constraints can profoundly affect the development of an AI system.
This research, reminiscent of the developmental and operational limitations of the human brain, provides new insights into the evolution of complex neural systems. By integrating these limitations, AI not only mirrors aspects of human intelligence, but also unravels the complicated balance between resource expenditure and information processing efficiency.
The concept of physical limitations in AI
The human brain, an example of natural neural networks, evolves and functions within a wide range of physical and biological constraints. These limitations are not barriers, but play an important role in shaping its structure and function. I
In the words of Jascha Achterberg, a Gates Scholar from the Medical Research Council Cognition and Brain Sciences Unit (MRC CBSU) at the University of Cambridge: “The brain is not only great at solving complex problems, it also does this while use very little energy. . In our new work, we show that considering the brain’s problem-solving capabilities, in addition to the goal of spending as few resources as possible, can help us understand why brains look the way they do.”
The experiment and its meaning
The Cambridge team embarked on an ambitious project to create an artificial system that models a highly simplified version of the brain. This system was distinguished by the application of ‘physical’ limitations, comparable to those in the human brain.
Each computing node within the system was assigned a specific location in a virtual space, simulating the spatial organization of neurons. The greater the distance between two nodes, the more challenging their communication, which reflects the neuronal organization in human brains.
This virtual brain was then tasked with navigating a maze, a simplified version of the maze navigation tasks often given to animals in brain studies. The importance of this task lies in the requirement that the system integrate multiple pieces of information, such as the starting and ending locations and the intermediate steps, to find the shortest route. This task not only tests the problem-solving capabilities of the system, but also allows to observe how different nodes and clusters become critical at different stages of the task.
Learning and adaptation in the AI system
The artificial system’s journey from novice to expert in maze navigation is a testament to AI’s adaptability. Initially, much like a human learning a new skill, the system struggled with the task and made numerous errors. However, through a process of trial and error and subsequent feedback, the system gradually refined its approach.
Crucially, this learning occurred through changes in the strength of connections between computing nodes, which reflect the synaptic plasticity observed in human brains. What is especially fascinating is how the physical limitations affected this learning process. The difficulty in establishing connections between remote nodes meant that the system had to find more efficient, localized solutions, mimicking the energy and resource efficiency seen in biological brains.
Emergent features in the artificial system
As the system evolved, it began to exhibit characteristics surprisingly similar to those of the human brain. One such development was the formation of hubs – highly connected nodes that acted as information channels across the network, similar to neural hubs in the human brain.
Even more intriguing, however, was the shift in the way individual nodes processed information. Instead of a rigid coding where each node was responsible for a specific aspect of the maze, the nodes adopted a flexible coding scheme. This meant that a single node could represent multiple aspects of the maze at different times, a feature reminiscent of the adaptive nature of neurons in complex organisms.
Professor Duncan Astle from Cambridge’s Department of Psychiatry highlighted this aspect, saying: “This simple limitation – it is more difficult to wire nodes that are far apart – forces artificial systems to produce some quite complicated features. Interestingly, they are features shared by biological systems such as the human brain.”
The implications of this research extend far beyond the domain of artificial intelligence and into the understanding of human cognition itself. By replicating the limitations of the human brain in an AI system, researchers can gain valuable insights into how these limitations shape brain organization and contribute to individual cognitive differences.
This approach provides unique insight into the complexity of the brain, especially in understanding conditions that affect cognitive and mental health. Professor John Duncan from MRC CBSU added: “These artificial brains give us a way to understand the rich and mind-boggling data we see when the activity of real neurons in real brains is recorded.”
Future of AI design
This groundbreaking research has significant implications for the future design of AI systems. The study vividly illustrates how integrating biological principles, especially those related to physical limitations, can lead to more efficient and adaptive artificial neural networks.
Dr. Danyal Akarca of the MRC CBSU underlines this by stating: “AI researchers are constantly trying to figure out how to create complex, neural systems that can code and perform in a flexible and efficient way. To achieve this, we think that neurobiology will give us a lot of inspiration.”
Jascha Achterberg elaborates on the potential of these findings for building AI systems that closely mimic human problem-solving skills. He suggests that AI systems that tackle challenges similar to those faced by humans are likely to develop structures similar to the human brain, especially when operating within physical constraints such as energy constraints. “The brains of robots deployed in the real physical world,” Achterberg explains, “will likely become more similar to our brains, as they may face the same challenges as we do.”
The Cambridge team’s research marks an important step in understanding the parallels between human neural systems and artificial intelligence. By imposing physical constraints on an AI system, they have not only replicated key features of the human brain, but also opened new avenues for designing more efficient and adaptable AI.