AI System Self-Organizes Like the Human Brain

AI System Self-Organizes Like the Human Brain

Researchers at the University of Cambridge have recently unveiled an artificial intelligence system that, by adopting physical constraints analogous to those of the human brain, has developed a self-organization capacity similar to that of the human brain. This innovation holds promise for advancements that will foster the development of more efficient AI and maybe be a step closer to the coveted goal of artificial general intelligence (AGI), a pursuit actively undertaken by leaders in the field, including OpenAI.

As neural systems organize and establish connections, they must balance competing demands. Energy and resources are required to develop and maintain the network in physical space while optimizing it for information processing. This compromise shapes all brains within and across species, possibly explaining why many brains converge toward similar organizational solutions.

Artificial Intelligence (AI) may have reached a new milestone, approaching closer to the complexity and efficiency of the human brain, thanks to the work of researchers at the University of Cambridge. They have recently developed an AI system that, under physical constraints similar to those of the brain, demonstrates remarkable self-organization capabilities. This advancement, detailed in the journal Nature Machine Intelligence, raises fundamental questions about the links between brain structure and function while opening new avenues for designing more efficient AI systems.

Self-organization in AI: a mirror of the human brain

seRNNs have a structural topology similar to that of a brain. Image: J. Achterberg.

The researchers explain in a statement that they have developed an AI system based on spatially integrated recurrent neural networks (seRNNs). These networks are designed to mimic the structure and functioning of the human brain, incorporating elements that replicate the physical and biological constraints the brain faces.

In the human brain, neurons are constrained by physical and biological factors such as the distance between neurons and the amount of energy available to establish and maintain connections, as mentioned earlier.

To test their AI system, the researchers chose a maze navigation task similar to those used in behavioral studies on animals like rats and macaques. The AI system had to determine the shortest path to reach a final point, combining several pieces of information: the starting point, the endpoint, and the intermediate steps.

This task required the system to retain and process these elements to succeed. Once the task was mastered, it was possible to analyze which nodes in the network were active at different times, identifying those encoding for specific aspects, such as arrival locations or possible routes.

Initially, the system made errors in task execution, but it progressively improved through feedback, adjusting the strength of connections between its nodes, akin to brain cells modifying their connections during learning. However, a physical constraint imposed on the system made creating connections between distant nodes more challenging, simulating the energy costs of long-distance connections in the human brain.

In response to this constraint, the system developed hubs, highly connected nodes facilitating information transmission. Moreover, these nodes adopted a flexible coding system where the same node could represent different properties of the maze at different times, a feature also observed in the brains of complex organisms.

Implications for neuroscience and mental health

This study goes beyond merely enhancing artificial intelligence systems. By examining the impact of physical constraints, such as distance between neurons and energy limits, on the development and functioning of the brain, researchers gain valuable insights into cognitive processes, even in humans. This research could reveal how the brain overcomes these constraints to process information, learn, and adapt, shedding light on individual variations in cognitive abilities and the potential causes of certain cognitive or mental disorders.

On the other hand, applying these physical constraints to AI systems allows the study of differences between human brains and simulates the effects of structural variations on cognitive functioning. This approach could provide clues about the neuronal manifestations of conditions such as autism, dyslexia, or schizophrenia. Additionally, by analyzing how AI adapts under these constraints, researchers could propose more targeted and effective treatments.

Towards more energy-efficient AI systems?

Drawing inspiration from the structure and constraints of the human brain, the AI system developed by Cambridge researchers proves to be more efficient, requiring fewer resources for its operation. This innovation is particularly relevant for devices and systems where energy efficiency is essential.

It also paves the way for brain-inspired AI models that are potentially more flexible and adaptive than current systems, especially for complex tasks. In other words, these systems are more sustainable and suited to real-world challenges. Ultimately, this type of system may bring us a step closer to true general artificial intelligence (GAI).