A Brain-inspired Theory of Mind Spiking Neural Network Improves Multi-agent Cooperation and Competition
Jul 10, 2023Author:
PrintText Size A A

Social cognitive abilities such as theory of mind (ToM) play a crucial role in the emergence of social intelligence. ToM refers to the ability to distinguish between oneself and others and to infer the mental states of others, including beliefs, intentions, desires and so on. In recent years, this cognitive function has been extensively studied in psychology and cognitive neuroscience, gradually uncovering the neural mechanisms underlying ToM. These neural mechanisms provide important insights and innovative resources for studying and exploring social interactions among multi-agent systems and human-computer interactions based on ToM.


Inspired by the mechanisms of ToM, a research team led by Prof. ZENG Yi from the Institute of Automation of the Chinese Academy of Sciences has proposed a multi-agent theory of mind spiking neural network (MAToM-SNN) which improves multi-agent cooperation and competition. The associated study was published in Patterns (https://www.cell.com/patterns/fulltext/S2666-3899 (23)00126-5) on June 23.


“When AI obtain the functionality of Theory of Mind/Cognitive Empathy, it has to be inspired by the brain so that we know the Theory of Mind model in AI is comparable to the human version so that it is more trustworthy for human and society. More specifically, inspired by the neural mechanisms in the ventral medial prefrontal cortex (vmPFC) and dorsal medial prefrontal cortex (dmPFC) for mentalizing and storing information related to self and others, as well as the neural mechanisms in the dorsolateral prefrontal cortex (dlPFC) for simulating others' decisions, the proposed model consists of two modules: Self-MAToM module for inferring others based on self-experience and Other-MAToM module for inferring others based on historical observations of others. The predicted behaviors of others by MAToM-SNN provide rich state representations for the decision-making model, enabling the decision network to adjust its policies adaptively.” said Prof. ZENG Yi, corresponding author of the study.


Agents with MAToM-SNN can use their own experiences or observations of others to infer their behaviors and adjust their policies to interact better with others. Besides, MAToM-SNN enhances the performance of multi-agent systems in cooperative and competitive tasks. MAToM-SNN demonstrates high levels of generalization on multi-agent reinforcement learning tasks based on spiking neural networks and recurrent neural networks , according to ZHAO Zhuoya, the first author of the study.


“The ablation study experiments validate that MAToM-SNN helps teams with larger numbers improve performance on competitive tasks. Besides, we find that Self-MAToM helps Other-MAToM learn quickly. The self is a prerequisite for inferring others. Thus it is essential to infer about others based on self-experience when information about them is incomplete.” " said Associate Professor ZHAO Feifei, one of the authors.


This study is part of the Brain-inspired Cognitive Intelligence Engine (BrainCog) project initiated by Prof. ZENG Yi's team, an on-going scientific exploration of the infrastructure of brain-inspired Artificial Intelligence.


ZHANG Xiaohan, PIO, Institute of Automation, Chinese Academy of Sciences

Email: xiaohan.zhang@ia.ac.cn