A predictive learning model for cognitive maps that generate replay

The hippocampus maintains an online representation of an animal’s situation in the environment which supports navigation, and generates offline simulations of plausible trajectories in that environment which are used for recall, imagination, planning, and the consolidation of memories. A leading theory is that these capacities rely on an attractive neuronal manifold, or cognitive map, in CA3, with topological correspondence to space. Supporting this idea, continuous attractor neural networks (CANNs) show place cell-like tuning for spatial locations, and can be used to simulate plausible trajectories in the environment in the absence of input. However, these networks require specific wiring between units with pre-assigned spatial locations or learning from inputs or outputs with preexisting spatial tuning, and it’s unclear how such a network can be learned from sensory information alone. Recently, it’s been found that self-supervised predictive learning - learning to predict subsequent or held-out sensory observations - can produce networks with spatially tuned cells, suggesting that predictive learning may be a plausible mechanism to develop a cognitive map in hippocampal circuits.

Here we show that predictive learning can form a cognitive map which can autonomously generate offline replay. However, we find that the presence of spatially-tuned cells is insufficient to indicate the presence of a continuous attractor. While recurrent neural nets trained to predict next time step observations reliably develop spatially-tuned units, they don’t produce a neural manifold with topological correspondence to space, and as a result, do not produce coherent replay in the absence of sensory input. These fundamental properties of a cognitive map emerge only when the network is trained using a learning algorithm in which recurrent connections are used to project multiple timesteps ahead of the agent. Once learned, this cognitive map can autonomously generate behaviorally plausible “replay” trajectories offline, including those not taken by the agent during wakefulness.