Researchers at the University of Cambridge implemented a straightforward physical constraint on an artificial intelligence (AI) system, leading it to exhibit characteristics reminiscent of the human brain. By applying constraints similar to those experienced by neurons in the brain, the AI system, using computational nodes instead of real neurons, developed features observed in complex organisms’ brains to solve tasks. The researchers believe that insights gained from this study could contribute to the creation of more efficient AI models with simpler internal structures, offering potential benefits such as improved performance on computer chips and better distribution of large AI models across multiple chips in extensive compute clusters. The study, conducted by Jascha Achterberg and Danyal Akarca from the Medical Research Council Cognition and Brain Sciences Unit (MRC CBSU) at the University of Cambridge, was published in the journal Nature Machine Intelligence. The approach involved simulating a simplified version of the brain, introducing physical constraints on computational nodes to emulate the challenges faced by neurons in terms of communication distance. The AI system was then tasked with solving a maze navigation problem, gradually improving its performance through feedback despite the communication constraints. Interestingly, the system exhibited similar problem-solving strategies to those observed in real human brains, such as the development of highly connected hubs to pass information efficiently. Additionally, the individual nodes in the system displayed a “flexible coding scheme,” firing for a mix of maze properties at different moments, akin to the behavior observed in complex animal brains. The researchers believe that the study’s findings can inform the development of more efficient AI models with simplified internal structures, potentially enhancing their performance on computer chips and facilitating better distribution across large-scale compute clusters.
