UCLA

The world is woven by our interactions with other people, animals, and nowadays, artificial intelligence. Cooperation is a fundamental aspect of human behavior throughout evolution, yet its underlying neural mechanisms remain poorly understood. This dissertation employs a cross-modal approach to investigate the neural mechanisms of cooperation in both the biological brain and multi-agent artificial intelligence (AI). Chapter 1 contextualizes this research by summarizing the behavioral origins of cooperation across animal species alongside the emergence of cooperative AI.In Chapters 2 and 3, pairs of mice were trained to perform a cooperative nose-poke coordination task while recording neural activity from the anterior cingulate cortex (ACC). We found that ACC neurons encode successful cooperative actions, the movement trajectories of both self and partner, and strategic interaction dynamics such as waiting and coordinated engagement. These neural representations were especially pronounced in mice that exhibited stronger cooperative performance. Recent advances in multi-agent artificial intelligence (AI) provide an opportunity to further investigate the principles of cooperation in controlled artificial environments. In Chapter 4, I trained recurrent neural network (RNN)-based artificial agents to perform the same nose-poke coordination task. The agents developed waiting strategies similar to those observed in mice. Selective silencing of neurons associated with waiting versus initiating movement produced distinct behavioral phenotypes, suggesting that functionally specialized neural populations contribute to cooperative behavior. These results also demonstrate the potential of multi-agent AI systems as experimental testbeds for studying brain function. In Chapter 5, I applied neural principles of cooperation to design multi-agent AI systems that are more robust in novel social environments. In the brain, neural manifolds emerge that separately process self-related and partner-related information, while partner location is represented in egocentric coordinates. Inspired by these findings, we developed a two-head attention network that processes visual information egocentrically. We found that the agents spontaneously specialized their attention, with one head focusing on partner-related information and the other on environmental variables. Compared with one-head attention models and convolutional neural network (CNN) baselines, the two-head attention agents generalized significantly better to novel social scenarios and environments. Together, these findings demonstrate that neuroscience-inspired principles can inform the design of more adaptive and socially robust AI architectures.