Cell types
Principal cells and interneurons of the hippocampal formation. Each card shows where the soma sits, where the dendrites (colour) collect input and where the axon (red) delivers output, in the style of Klausberger & Somogyi's laminar diagrams.
Firing lab
Inject current into a model of each cell type and watch its characteristic discharge: adaptation, bursting, fast spiking, Ih sag and rebound, late spiking. Models are Izhikevich neurons with an added h-current, tuned to each type's published behaviour; voltages are scaled to typical resting potentials.
Rhythms
Each cell type fires at its own moment of the network oscillation. Switch between running (theta with nested gamma) and quiet rest (sharp-wave ripples) to see which cells lead, follow or fall silent.
Top: CA1 pyramidal-layer LFP. Below: simulated spike rasters drawn from published firing rates and phase preferences (rat CA1, Klausberger & Somogyi 2008).
Function
What the circuit computes. Run the demos: spatial codes in an open arena, phase precession on a track, pattern separation in the dentate gyrus and pattern completion in CA3.
Spatially tuned cells
Theta phase precession
As the animal crosses a place field, the cell fires at progressively earlier theta phases (O'Keefe & Recce 1993). Each theta cycle therefore compresses a sequence of upcoming places into ~125 ms, a time scale suited to spike-timing-dependent plasticity.
Pattern separation (dentate gyrus)
Two entorhinal patterns with a given overlap are projected onto 2,000 model granule cells with random connections; only the most strongly driven fraction fires (lateral inhibition). Sparser codes push similar inputs further apart: points below the diagonal mean separation.
Pattern completion (CA3)
A 144-unit recurrent network stores three patterns in its recurrent weights, as CA3 recurrent collaterals are thought to. A partial or noisy cue settles into the nearest stored memory (an attractor). Push degradation past ~40% and recall fails or lands on the wrong memory.
The long axis: dorsal versus ventral
In rodents the hippocampus curves from dorsal (septal) to ventral (temporal). In humans these correspond to the posterior body/tail and anterior head. Function, connections and gene expression change along this axis (Fanselow & Dong 2010; Strange et al. 2014).
Space & time
Hippocampal neurons respond to places, moments, routes, laps, objects, other animals and even ideas. This view collects those functional cell types, asks what single principle could produce all of them, and shows how the same machinery may let us travel mentally into the past and the future.
Space is a latent sequence
The paper argues that the brain's map of space is not built from coordinates. It emerges when an animal learns the higher-order sequences of what it senses and does. Animals have no GPS. They only receive egocentric sensations, and those are aliased: the same view can occur in many places. To predict what comes next, a learner must split one sensation into several latent states according to the sequence it appears in. Once those states are learned, they behave like places.
The model is a clone-structured causal graph (CSCG), an action-conditioned hidden Markov model in which each observation owns a fixed set of hidden "clones". Expectation-maximisation learns which clone follows which under each action; Viterbi training then prunes unused clones. A clone stands for a sensation in a particular sequential context, so when the experimenter plots a clone's activity on a floor map, it looks like a place cell.
- Maps from aliased egocentric input. CSCGs learned the topology of 2D rooms and 3D environments from random walks, with no location input and no Euclidean assumptions. Clones also came to represent heading, which is never sensed directly.
- Stitching and separating. Overlapping episodes collected at different times were stitched into one global map. Experience from disjoint mazes was kept as separate maps, and the right map is recalled by inferring the hidden state.
- Planning as replay. To reach a remembered goal, messages are passed forward from the current clone and backward from the goal clone, which the authors liken to hippocampal replay. Blocking a route means editing one link of the graph.
- Schemas. Keeping the learned transition graph fixed and relearning only the emissions let the agent find a shortcut through unvisited rooms in a new environment that shared the old one's structure.
- Place fields as a visualisation. In the paper's words, many "diverse fascinating hippocampal phenomena might be artifacts of Euclidean place field mapping." The agent itself never computes place fields.
Phenomena the model reproduces
Each row is a classic experiment the authors re-ran in simulation with one model and one learning rule; only the environment changed.
Testable predictions
- Uniqueness, not visual change, controls field size. A floor with a repeating checkerboard should give larger central place fields than a floor with a random pattern, because the checkerboard makes many sequences look the same (Fig. 4F).
- Local landmarks decide which fields remap. When a room with landmarks on one side is elongated, fields whose local sequence is anchored by two landmarks should stay put. Fields anchored to a landmark and the moved wall should stretch (Fig. 4G).
- Connectivity changes show up in planning, not in fields. Locking doors should alter replay and route choice while place fields stay put.
- Grid cells are helpful but optional. They act as extra sensations that speed learning in featureless areas, consistent with place fields surviving grid disruption.
Limits the authors note
- Animals form place fields with far less experience than the model needs; reusing schemas may close that gap.
- Remapping to new objects, active exploration, reward learning and automatic graph editing after surprise are not yet modelled.
- The work is computational. Clones might be neuronal assemblies and EM resembles spike-timing-dependent plasticity, but no circuit implementation is claimed.
- Unlike the Kantian view that space and time come first, the results support a Leibnizian view: only sequential order needs to be built in.
How cloning creates splitter and lap cells
Replay: mental time travel in a rat
The rat shuttles between two reward wells while 16 place cells (sorted by field position) fire in order. At each well it stops, and during sharp-wave ripples the same sequence is replayed about 20 times faster: backward on arrival, as if reviewing the path that led to reward, and forward before leaving, as if planning (Foster & Wilson 2006; Diba & Buzsáki 2007). In a sequence model, recall and planning are both just running the learned graph.
Mental time travel
Functional cell types
Cells named for what they respond to. Where Raju et al. (2024) or related sequence theories offer an explanation, it is given under "Sequence view".
How might the hippocampus encode space and time?
Eight influential accounts, roughly in historical order. They are not mutually exclusive; most current work combines several.
Plasticity
Each hippocampal synapse has its own rules for getting stronger or weaker. The rules differ in where they are induced (pre- or postsynaptically), what triggers them, how long they last and what they are for. Run a virtual slice experiment, explore the learning rules, then browse the mechanisms by region.
Virtual slice experiment
Timing rules
Frequency, calcium and the sliding threshold
Low-frequency stimulation causes a modest, prolonged Ca²⁺ rise that activates phosphatases (calcineurin, PP1) and depresses the synapse. High frequency gives a large, fast Ca²⁺ rise that activates kinases (CaMKII) and potentiates it. The crossover point slides with the neuron’s history (the BCM rule; Bienenstock, Cooper & Munro 1982). After high activity, LTP becomes harder; after deprivation, easier. This is metaplasticity (Abraham & Bear 1996).
Which mechanism operates where
Select a synapse name to see its pathway on the circuit map. A dash means not described or not prominent.
Mechanisms
From synapse to memory
Theory & math
The equations behind the atlas: how single neurons, learning rules and networks are modelled, how function (memory, maps, rhythms) falls out of the maths, and how changing one parameter in a model reproduces a disease. Three live models let you break the circuit yourself.
Three levels of explanation
David Marr argued that a brain system is understood only when it is explained at three levels. His 1971 “Simple memory: a theory for archicortex” was the first mathematical theory of the hippocampus: sparse codes in a recurrently connected CA3 store events from a single exposure and recall them from a fragment. Nearly every model on this page descends from it.
- Computational
- What problem is solved, and why? Store episodes after one exposure; keep similar ones apart; recall from partial cues; map space and time; replay for consolidation.
- Algorithmic
- Which representations and procedures? Sparse expansion (DG), attractor dynamics (CA3), comparison of prediction and input (CA1), Bayesian decoding, successor maps, latent sequences.
- Implementation
- Which neurons, synapses and rhythms? Mossy-fibre detonators, recurrent collaterals with NMDA-receptor plasticity, PV-cell gamma, theta from the septum, plateau-driven BTSP.
Attractor memory: capacity and synapse loss
A 200-unit Hopfield network stores P random patterns with a Hebbian rule, then recalls each from a noisy cue. Recall stays near perfect until the load α = P/N approaches ≈0.14, then collapses (Amit, Gutfreund & Sompolinsky 1985). Removing synapses, as in Alzheimer’s disease, lowers the collapse point: memory fails abruptly, not gradually. The grey curve is the intact network.
Excitation–inhibition dynamics: rhythm and seizure
Reading out place: Bayesian decoding
Each cell fires Poisson spikes at a rate set by its Gaussian field on a 1 m track. The posterior over position (top, accent) is computed from the spike counts alone (Zhang et al. 1998). Error (bottom) falls roughly as 1/√N; broader fields and shorter windows blur the code. Unstable or broadened fields in ageing and Alzheimer models act like turning these dials down.
From model parameter to disease
Equation library
The core equations, what each predicts and where it is used in this atlas.
Some numbers worth knowing
Connectivity
Major projections into, within and out of the hippocampal formation. Rows send, columns receive. Shade gives relative strength (1 sparse, 2 moderate, 3 dense); colour gives transmitter class. Select a cell for details.