How do online preplay and offline replay optimize action selection in navigation?

One noise-driven mechanism yields three replay modes

The paper asks how the hippocampus produces such different neuronal sequences, and what each one does for navigation. Its answer is a closed-loop model of a two-dimensional open field built from leaky integrate-and-fire neurons, with a recurrent place-cell layer and an action-selection ring attractor that sets the agent's movement direction and velocity. Sequences in this network come from a common mechanism: background noise plus the balance of excitation and inhibition. What differs between modes is not a separate sequence generator but the synaptic modifications left by the task.

That distinction matters because the three modes are easy to conflate. Online forward replay is a sweep ahead of the animal's current location during goal-directed navigation. Offline forward replay occurs during rest, when sharp-wave-ripple sequences reactivate previously explored trajectories. Diffusive replay follows random foraging and resembles Brownian motion, with no correlation to behavior. The model reproduces all three within one architecture, so their differences can be attributed to learning history rather than to distinct circuits.

Learned paths become line attractors that bias replay forward

During exploration, synapses in the recurrent layer and in the place-to-action pathway are modified by dopamine-gated spike-timing-dependent plasticity with an eligibility trace, so only synapses active along the path to the goal are potentiated. The recurrent layer is arranged as a two-dimensional manifold in which neurons encoding nearby places are more strongly connected, with sparse long-range excitatory connections at a fixed 2% probability and strictly local inhibition.

This learning leaves a line attractor embedded in the two-dimensional network. When the agent pauses, both forward and backward replay can occur, but longer-lasting inhibitory currents in previously active populations suppress immediate reactivation and bias the network toward forward replay. The paper's simulations show forward propagation becoming dominant, prioritized, and more frequent when the agent is in the middle of the path. This is a model result about network dynamics, not a measurement of hippocampal activity.

Population bursts open a brief window for synaptic strengthening

The paper compares three settings for a six-neuron directional chain: a homogeneously connected network with asynchronous firing, a network with synchronous population bursts, and a network after exploration in which a line attractor has formed. In the asynchronous state, balanced inhibition decorrelates pre- and postsynaptic activity, so the covariance between firing rates is minimal and weight change is negligible.

During population bursts, inhibition follows excitation with a short delay, producing a small positive peak in the cross-covariance at short positive lags. That brief window favors potentiation along the chain. After exploration, synapses toward the highest-firing neurons are preferentially strengthened, and later bursts activate the chain more reliably and more frequently. Across a fixed 30-second simulation, the average normalized weight change was higher in the synchronous regime with an embedded line attractor than in the other scenarios for all tested initial weights, and activation rate over 60 seconds was measured at a median initial normalized weight of 0.4.

Offline replay shortens paths and raises trial success

During rest, spatially homogeneous Poisson noise replaces the environmental input, and replay modifies both recurrent connections and place-to-action synapses through spike-timing-dependent plasticity plus synaptic normalization. The reported behavioral effect is a shorter path on the subsequent exploration trial. Across twenty exploration phases, the agent gradually reached the reward more efficiently over successive trials.

The success rate is not perfect: in some fixed-duration trials the agent failed to reach the reward, which lowered the overall trial success rate. The authors attribute the improvement to stabilization of learned place and action sequences and to greater noise tolerance. This is a simulation-level claim about the model agent, and the paper does not report a corresponding behavioral experiment.

Diffusive replay generalizes actions to neighboring places

After random foraging, the network's spontaneous activity takes the form of wandering local bumps rather than ordered sequences. Synaptic normalization of excitatory-to-excitatory and excitatory-to-inhibitory weights, while maintaining excitation–inhibition balance, shifts the effective weight kernel and drives the network into this diffusive regime. The decoded bursts then resemble random walks.

The functional consequence is a more uniform action map. Diffusive replay modifies place-to-action connections so that the encoded action generalizes to neighboring locations visited during successful exploration, which supports more efficient goal-directed navigation from other starting positions. Across ten independent runs, the network was set to spontaneous firing for 5 seconds after a number of exploration trials, and success was measured from variable starting positions before and after replay. The paper reports a reduction in the average angular difference between the encoded action at each place and the actual goal-directed vector, alongside changes in the mean and variance of the fraction of successful trials.

A single architecture links replay type to navigation function

The paper's central claim is organizational: one network architecture can flexibly generate various replay types, and each type contributes differently to optimizing goal-directed behavior. Online forward replay supports shortcut discovery and flexible strategy adjustment by activating appropriate action cells for each place cell. Offline forward replay consolidates optimal sequences of actions. Diffusive replay generalizes the action and state map and homogenizes internal connectivity, which the authors say improves subsequent goal-directed navigation.

The mechanism behind this flexibility is the network's operating regime. The model sits near a phase transition, producing population bursts with power-law statistics and long spatial correlation lengths that let activity bridge distant neural ensembles. Homeostatic synaptic normalization keeps this sensitive operating point stable. The authors present this as an explanation of how weak synaptic inputs can trigger large-scale recruitment, not as a demonstrated property of CA3 recordings.

Model scope limits what the replay functions can claim

Several boundaries qualify the results. The work is a simulation of a spiking network, so the three replay modes and their navigation benefits are model outcomes rather than measured hippocampal dynamics. The authors note that CA3 burst statistics remain debated, with some studies reporting power-law scaling and others log-normal distributions, and they call for further experimental recordings. The critical exponents in the network do not coincide with standard universality classes such as directed percolation, which predicts a mean-field exponent of 1.5 for burst size and 2.0 for duration.

The model also omits mixed selectivity, the conjunctive coding of location, direction, context, and goal information observed in hippocampal neurons. The authors propose extending the model to conjunctive state–action coding or goal-modulated place representations. These limitations matter most for readers who want to treat the three replay functions as established biology rather than as a coherent computational account.

Task history decides which replay mode carries the benefit

The paper's most useful distinction is conditional rather than universal. In a goal-directed task analogous to the Morris water maze, online and offline forward replay activate the appropriate action cells for each place cell, enabling shortcut discovery and flexible strategy adjustment. After random foraging, diffusive replay instead produces local random-walk-like sequences that promote a more uniform spatial representation, strengthen recurrent connectivity, and generalize place-to-action mappings.

So the practical reading is that replay type should be matched to task history: forward replay is tied to learned goal-directed trajectories, while diffusive replay follows undirected exploration and pays off in later goal-directed navigation from new starting positions. The paper's own priority is this division of labor within one architecture, supported by simulations across twenty exploration phases and ten independent runs, rather than a claim that any single replay mode is universally optimal.

Sources used in this answer