ISSN 1671-3710
CN 11-4766/R

心理科学进展 ›› 2021, Vol. 29 ›› Issue (4): 677-696.doi: 10.3724/SP.J.1042.2021.00677

• ·研究前沿· • 上一篇    下一篇


黎穗卿, 陈新玲, 翟瑜竹, 张怡洁, 章植鑫, 封春亮()   

  1. 教育部脑认知与教育科学重点实验室(华南师范大学); 华南师范大学心理学院; 华南师范大学心理应用研究中心; 华南师范大学广东省心理健康与认知科学重点实验室, 广州 510631
  • 收稿日期:2020-08-10 出版日期:2021-04-15 发布日期:2021-02-22
  • 基金资助:

The computational and neural substrates underlying social learning

LI Suiqing, CHEN Xinling, ZHAI Yuzhu, ZHANG Yijie, ZHANG Zhixing, FENG Chunliang()   

  1. Key Laboratory of Brain, Cognition and Education Sciences (South China Normal University), Ministry of Education; School of Psychology, South China Normal University; Center for Studies of Psychological Application, South China Normal University; Guangdong Key Laboratory of Mental Health and Cognitive Science, South China Normal University, Guangzhou 510631, China
  • Received:2020-08-10 Online:2021-04-15 Published:2021-02-22


人类在社会互动中通过他人的行为对他人特质、意图及特定情境下的社会规范进行学习, 是优化决策、维护积极社会互动的重要条件。近年来, 越来越多的研究通过结合计算模型与神经影像技术对社会学习的认知计算机制及其神经基础进行了深入考察。已有研究发现, 人类的社会学习过程能够较好地被强化学习模型与贝叶斯模型刻画, 主要涉及的认知计算过程包括主观期望、预期误差和不确定性的表征以及信息整合的过程。大脑对这些计算过程的执行主要涉及奖惩加工相关脑区(如腹侧纹状体与腹内侧前额叶)、社会认知加工相关脑区(如背内侧前额叶和颞顶联合区)及认知控制相关脑区(如背外侧前额叶)。需要指出的是, 计算过程与大脑区域之间并不是一一映射的关系, 提示未来研究可借助多变量分析与脑网络分析等技术从系统神经科学的角度来考察大尺度脑网络如何执行不同计算过程。此外, 将来研究应注重生态效度, 利用超扫描技术考察真实互动下的社会学习过程, 并更多地关注内隐社会学习的计算与神经机制。

关键词: 社会学习, 计算模型, 神经影像, 强化学习模型, 贝叶斯模型


Social learning refers to the belief updates of others’ personal attributes and intentions as well as social norms under various circumstances during social interactions, which helps to optimize social decision-making and maintain positive social interactions. Due to its critical role in human decisions and social interactions, the past years have witnessed a growing body of studies that examine computational and neural basis of social learning combining computational models and human brain imaging techniques. The current literature indicates that human social learning can be well captured by reinforcement learning model and Bayesian model. 
In the framework of reinforcement learning models, an active agent adaptively adjusts his behaviors according to the feedback in social interactions to achieve a certain goal, with  positive feedback will increasing the possibility of the previous behavior and negative feedback weakening it. Accordingly, social learning mainly engages the computation of subjective expectation and prediction error. Consistent with the findings in nonsocial learning, these computations involve brain regions associated with reward and punishment processing (e.g., the ventral striatum and ventromedial prefrontal cortex). Notably, in social situations, brain regions associated with social cognition (e.g., the dorsomedial prefrontal cortex and the temporal-parietal junctions) are also involved due to the inference of the traits and intentions of others.
Although reinforcement learning models provide powerful explanations for social learning processes, they did not account for the representation of social uncertainty. Instead, the Bayesian models assume that the social learning process follows the Bayesian information updating, and the perceived uncertainty is represented in the posterior distribution of psychological variables. Therefore, the Bayesian models can depict the representation of uncertainty. People represent their prior beliefs about others and calculate the deviation between actual feedback and prior beliefs, which is similar to the representation of subjective expectations and expected errors respectively in reinforcement learning style. In addition, representation of uncertainty and information integration are involved, engaging brain regions associated with reward and punishment processing, social cognition, and cognitive control (e.g., dorsolateral prefrontal cortex).
However, it should be noted that there is no one-to-one mapping between computational processes and brain regions, rather, it is in a many-to-many-pattern, that is, a single cognitive process involves multiple brain regions, and a specific brain region can be involved in multiple calculations. Therefore, multivoxel pattern analysis and brain network analysis should be utilized in future studies to reveal how different computational processes are implemented in large-scale networks according to systems neuroscience. Moreover, future studies should try to increase the ecological validity by creating real social interactions between people and by leveraging novel neuroimaging approaches (e.g. hyperscanning techniques). Finally, more efforts are needed to unravel the neural and computational signatures of implicit social learning.

Key words: social learning, computational modeling, neuroimaging, reinforcement learning model, Bayesian model