Reasoning the hidden relational structure from sequences of events is a crucial ability humans possess, which help them to predict the future and make inferences. Besides simple statistical properties, humans also excel in learning more complex relational networks. Several brain regions are engaged in the process, yet the time-resolved neural implementation of relational structure learning and its behavioral relevance remains unknown. Here human subjects performed a probabilistic sequential prediction task on image sequences generated from a transition graph-like network, with their brain activities recorded using electroencephalography (EEG). We demonstrate the emergence of two key aspects of relational knowledge – lower-order transition probability and higher-order community structure, which arise around 840 msec after image onset and well predict behavioral performance. Furthermore, computational modeling suggests that the formed higher-order community structure, i.e., compressed clusters in the network, could be well characterized by a successor representation operation. Overall, human brains are constantly computing the temporal statistical relationship among discrete inputs, based on which new abstract knowledge could be inferred.

No abstract