Recording
https://www.bilibili.com/video/BV1Sj411C77e/?vd_source=e9626f9767e6e22ece9d765f34ba01c5
Speaker
Zhaoyu Zuo
Bio
Zhaoyu Zuo obtained a master’s degree in pattern recognition and intelligent systems from the University of Science and Technology of China and is now a research assistant at Shanghai Jiao Tong University. He is concerned with the nature of human intelligence and the mechanisms that lead to behavior and mental disorders. He uses brain-inspired models to study the cognitive computational mechanisms behind complex decisions, particularly the cooperative relationship between memory systems and reinforcement learning.
Abstract
Model-based (goal-directed) decision-making requires prudent evaluation of the ultimate consequences of multi-stage choices. Previous studies have suggested that such evaluation relies on the reward experience accumulated by the reinforcement learning process in the brain. However, the core component of model-based decision-making–working memory (WM), also retains reward information, and it is still unclear whether WM contributes to the evaluation. The current study analyzes four two-stage decision experiments, which separately manipulate two WM-related variables (delay and load). We found that time delay interfered with evaluating the ultimate consequences, while increased task load reduced cognitive effort in the feedback process and the probability of selecting the optimal option. Notably, our proposed models that corporated the reward-retained mechanism of WM could replicate the behavioral effects of delay and load, whereas the classical hybrid reinforcement learning model could not. Furthermore, individual-level analysis revealed a close correlation between model parameters and WM scores. Together, these results provide a deeper understanding of how reinforcement learning and WM co-work in complex decision-making and facilitate analysis of impaired model-based decision-making in clinical populations.