In this study, we proposed a multi-regional modular recurrent neural network to simulate the cognitive processes. The model is structured into three different modules: perception, information integration, and decision. Here a transfer learning approach is adopted to investigate generalizability across tasks. After training models on source tasks, we fixed the information integration layers, transferred the models to target tasks, and tested their performance. By comparing the performance of different source-target task pairs, we assessed the similarity between different cognitive tasks.
Youzhi Qu, Wenxin Che, Penghui Du, Xinyao Jian, Quanying Liu
