Title:Replay and compositional computation

Abstract:It is still a mystery how the human brain radically generalizes from previous experiences to solve new problems with very little data, an ability that far surpasses deep learning. Here, we propose a speculative hypothesis: that 'replay' in the brain implements a form of compositional computation where entities are assembled into meaningful structures. We review recent advances in the neuroscience of replay, highlighting how the hippocampus flexibly binds objects to generalizable roles and how replay strings these role-bound objects into compound statements. We suggest new experiments to test this hypothesis. We end by noting the bidirectional interactions between replay and other brain systems and their relationship to AI methods that hybridize neural networks with compositional computation.