A spiking model built on György Buzsáki's Rhythms of the Brain and The Brain from Inside Out. One network runs a track, rests and sleeps: what it learns while running, it replays at rest and hands to the cortex in sleep. Two further circuits show how movement decides what gets perceived, and how theta links hippocampus and prefrontal cortex when a choice has to be made.
The animal runs a track. Theta frames place-cell firing into compressed sequences, and spike-timing plasticity writes that order into the recurrent synapses of CA3. Run a few laps here before visiting the other scenes.
CA1, unprompted
180 pyramidal cells (triangles) and 45 basket interneurons (dots), driven by noise and a septal-like pacemaker. Ringed cells are place cells. Click to stimulate the cells under the pointer. The dial shows theta phase, with its excitable half shaded.
Field potential and spikes, on one clock
Violet columns mark the excitable half of each theta cycle; with nesting on, gamma bursts gather there. In the raster, alternate theta cycles are shaded and teal hairlines mark interneuron volleys, the gamma windows that bind assemblies. CA3 place cells sit at the top.
Phase precession
Every CA1 place cell's spikes, aligned to its own field. Firing starts late in the theta cycle and slides earlier as the animal crosses. This is what orders cells within each cycle, and what plasticity learns from.
Sequence stored in CA3
Synapse strength from each place cell (rows) to each other (columns), ordered along the track. A band just right of the diagonal means each cell now excites the cells that come after it.
Power spectrum
Drawn as power × frequency to flatten the 1/f slope: one slow hump and one near gamma.
Gamma amplitude by phase
Flat means independent rhythms; a peak near the trough means the slow wave decides when gamma may happen.
The animal stops. Septal pacing and acetylcholine fall away, recurrent synapses in CA3 come back to full strength, and the circuit starts to burst on its own. When a burst lands on a stored sequence, it plays back as a sharp-wave ripple, much faster than it was lived.
CA3 drives CA1
Click inside CA3 to start a burst at the place cells nearest the pointer. During a replay the wave of activity moves through the ringed place cells in track order.
Sharp waves and ripples
The CA3 burst depolarises CA1 dendrites, the slow sharp wave. Basket cells in CA1 answer with a 150–220 Hz ripple. In the raster, place cells are sorted by where their fields sit on the track, so a replay reads as a diagonal.
Latest replay
First spike of each CA3 place cell during the most recent ripple, against its position on the track. ρ is the rank correlation between the two.
Sequence stored in CA3
The same synapses the replay travels along.
What sets the replay speed
In the run, neighbouring place fields were a few hundred milliseconds apart. In the replay, each cell recruits the next through a recurrent synapse with a 7 ms axonal delay, so the whole track plays back in a fraction of a second. Only cells linked in the stored order recruit each other; weak synapses stay silent, which is why an untrained network bursts without replaying anything.
In slow-wave sleep the cortex alternates between active up-states and silent down-states, and thalamic spindles ride on the up-states. Hippocampal replays keep arriving. The ones that land inside a spindle are allowed to change cortical synapses, so the sequence is copied, slowly, into cortex.
Hippocampus talks to cortex
CA3 bursts, CA1 relays, and 40 cortical memory cells (ringed) receive the replay. The cortex glows during up-states. Click CA3 to start a burst.
Slow oscillation, spindles, ripples
Grey columns are down-states; amber columns are spindles; red ticks mark ripples. With coupling on, ripples cluster inside spindles. In the raster, replays appear as diagonals in CA1 and, when cortex is receptive, in the memory cells below.
Consolidation over time
How strongly each network's synapses encode the forward order of the track, from 0 (none) to 1. Changes of setting are marked, so you can compare how fast cortex learns with and without coupling.
Sequence in cortex
The cortical copy of the CA3 matrix, built only from replays that arrived during spindles.
When ripples happen
Ripples over the last two minutes of sleep, split by what the cortex was doing, against the time it spent in each state.
A whisker sweeps back and forth at theta frequency. Every sweep stimulates the whisker's own sensors, and a copy of each motor command, the corollary discharge, learns to predict that input. Only the difference between what arrives and what was predicted reaches perception.
Whisking
Drag the pole to move it along the whisker's path. The label above the head appears when barrel cortex reports something it did not predict.
Command, input, prediction
The prediction (dashed) is built from the motor copy alone; it knows nothing about the pole or taps except what repeats with every sweep. Barrel cortex has two output groups: one fires when input exceeds the prediction, the other when an expected touch fails to arrive.
What each sweep should feel like
Average input across one whisk cycle (solid) and the corollary discharge's learned prediction (dashed). When they overlap, the whisker's own movement has become invisible.
Recent percepts
Nothing perceived yet.
Try switching the corollary discharge off: the circuit keeps reporting touches caused by its own whisking, so it can no longer tell self from world. It is the reason you can't tickle yourself.
At the fork of a T-maze, hippocampal theta sequences sweep ahead into one arm, then the other, cycle by cycle. Prefrontal cells lock to hippocampal theta, so their excitable phase lines up with the look-ahead as it arrives. Prefrontal cortex accumulates what the sweeps show, weighted by what each arm has paid out, until it commits.
The maze
Dots along the maze are place fields, lit when their cell fires. Near the fork, the violet or amber streak is the look-ahead sweep of the current theta cycle. The meters show the two prefrontal pools and the decision variable against its threshold.
Two rhythms, one decision
Shaded columns mark time at the choice point. Watch prefrontal theta slide into a fixed phase relationship with hippocampal theta as the animal approaches the fork, and the arm cells in the raster fire in alternating cycles before the choice (red lines).
Theta cycles at the fork
Each bar is one theta cycle at the choice point: left-arm cell spikes up, right-arm spikes down. Successive cycles tend to alternate, as Kay and colleagues found in rats (2020).
Trial by trial
Height is time spent deciding. Filled dots were rewarded, open red ones were not; rings mark guesses at the deadline. Grey background marks trials run with weak coupling.
Prefrontal locking by place
How consistently prefrontal spikes fall at the same hippocampal theta phase, since the coupling was last changed.