The Spatio-Temporal Dynamics of Awake Hippocampal Sharp-Wave Ripples and Their Relation to Planning

Abstract

Planning through mental simulation, or the anticipation of future action-outcome sequences, is a powerful mechanism for improving action selection. Strikingly, neural activity reminiscent of mental simulation has been observed in rodents when performing spatial navigation tasks. Within sharp-wave ripples (SWRs) during brief pauses in movement, place cells in the hippocampus—which usually represent the animal’s current location—shift to generating “simulated” trajectories through the environment. This phenomenon, termed awake hippocampal replay, is thought to support planning. Here, we first examine a potential computational mechanism underlying hippocampal replay by casting the question of selecting what to mentally simulate into a reinforcement learning problem. We ask how simulations should be prioritized to maximize utility for action selection, probe how the simulation objective impacts the optimal order of prioritization, and propose experiments for identifying these objectives. We then characterize the spatio-temporal dynamics of replay trajectories, as this forms an essential foundation for theories of their computational function. Notably, currently employed methods for detection and classification of trajectories in open-field environments rely on heuristics, which discard the majority of SWRs as unrelated to trajectory simulation and, thus, risk biasing their computational interpretation. Using Bayesian model comparison of state-space models, we identify replay events by the presence/absence of structural features, such as trajectory-continuity, spatial diffusion, or momentum-dependence. Applied to hippocampal data collected during a foraging task, we find that almost all SWRs—including many discarded in previous studies—encode spatial trajectories through the environment. Further, we find that these trajectories feature momentum, that is, inertia in their velocities, mirroring the animals’ natural movement. Comparing the identified replay trajectories to the animal’s actual movement through the maze, we find that the statistics of neural activity during replay closely resemble neural activity during behavior, and that the simulated trajectories are modulated by the animal’s putative intentions. Specifically, replay trajectories preceding random foraging tend to be shorter, slower, and less predictive of future behavior than replay events preceding directed movement. Overall, this theoretical study provides a systematic and comprehensive characterization of the spatio-temporal dynamics of SWRs and lays a principled foundation for future work examining their computational role.