Causal necessity of human hippocampus for structure-based inference in learning

Abstract

When meeting new individuals or encountering known individuals in new circumstances, we intuitively map out their relationships - not merely by direct experience, but by quickly inferring new connections based on prior relational knowledge. Using a novel task, we demonstrated that participants indeed employ knowledge of relational structures to facilitate learning of new relationships in a changing environment. Computational modelling revealed that participants leveraged relational knowledge to support inference, thus facilitating learning. Whole brain neuroimaging identified a uniquely robust representation of relational structure in the hippocampus. Neural networks trained on similar tasks demonstrated the emergence of relational structure representations, resembling those found in hippocampus. Lesioning network units sustaining such representations disrupted structure-based inference and predicted hippocampuss essential role. Transcranial ultrasound stimulation of human hippocampus, transiently modulating its activity without affecting overlying tissue, produced similar disruption effects, empirically confirming the causal necessity of hippocampal representations for structure-based inference in learning.