Fast Sequences of Non-spatial State Representations in Humans

Fast internally generated sequences of neural representations are suggested to support learning and online planning. However, these sequences have only been studied in the context of spatial tasks and never in humans. Here, we recorded magnetoencephalography (MEG) while human subjects performed a novel non-spatial reasoning task. The task required selecting paths through a set of six visual objects. We trained pattern classifiers on the MEG activity elicited by direct presentation of the visual objects alone and tested these classifiers on activity recorded during periods when no object was presented. During these object-free periods, the brain spontaneously visited representations of approximately four objects in fast sequences lasting on the order of 120 ms. These sequences followed backward trajectories along the permissible paths in the task. Thus, spontaneous fast sequential representation of states can be measured non-invasively in humans, and these sequences may be a fundamental feature of neural computation across tasks. •Task requires navigating between objects whose neural representations can be decoded•After task has been learned, spontaneous MEG activity encodes reverse paths•These spontaneous sequences are up to four states long•Sequences might be cognate with rodent hippocampal “replay” The rat hippocampus spontaneously imagines past and future locations in rapid sequences (sometimes called “replay”), which link multiple locations into coherent neural representations. Kurth-Nelson et al. find similar sequences in humans, in a non-spatial task, suggesting they are a ubiquitous mechanism.