Computational neuroscience

How do circuits link related memories?

A bilateral CA3–CA1 spiking network to study memory formation and generalisation.

Peking University · Guozhang Chen’s group2026–present

The question

When a fear response transfers from one context to another, have the two memories become associated, or has the network lost the information that distinguishes them? I use a bilateral hippocampal model to ask how circuit structure and plasticity might produce these different outcomes.

Building a model when the connectome is incomplete

I built a CA3–CA1 spiking network with nearly 60,000 neurons and roughly 30 million synapses. Both hemispheres include excitatory, PV, and SST populations. The model combines constraints on cell-type firing properties, neuron numbers and positions, and axonal projections.

My main methodological contribution is a Bayesian connectivity framework using conditional relative-entropy reconstruction. It combines complementary anatomical sources to estimate connection probabilities conditioned on neuronal type and spatial position, while respecting available marginal constraints.

Spatial organisation of the bilateral CA3–CA1 model, with CA3 in blue and CA1 in teal. The network is reconstructed from anatomical constraints.

My contributions

Connectivity structure of the reconstructed network, organised by hippocampal region and hemisphere. Colour indicates connection probability.

Population dynamics

I examine network activity at both population and single-cell scales. Spike rasters show how activity is distributed across CA1 and CA3, alongside excitatory, PV, and SST populations.

CA1 (top) and CA3 (bottom) spike rasters over 2–5 s, after warm-up. Each region contains 300 sampled cells; the E/PV/SST composition is given in the legends. These are example dynamics from the model snapshot shown in my research presentation.

Single-neuron dynamics

Inspecting individual neurons links the population activity to the interaction of excitatory and inhibitory currents with membrane potential. The examples below show this relationship across excitatory, PV, and SST cells.

Twelve example neurons, with excitatory, inhibitory, and net current traces above their corresponding membrane-potential traces. The plotted interval is 0.5–2.5 s. These selected examples illustrate cellular dynamics rather than population-average statistics.

Pattern completion: partial-cue retrieval

I test retrieval in three learned ensembles, A, B, and C, by comparing no-cue activity with the response to a partial cue. Separating directly cued cells from the rest of the target ensemble makes it possible to examine recruitment beyond the stimulated subset.

Rows target ensembles A, B, and C. Left: no-cue activity. Centre: the first 120 ms around the partial cue, separating 138 directly cued cells (blue) from 137 uncued cells (salmon); the shaded cue window is 0–20 ms. Right: an overview through 1 s, with ensembles A/B/C coloured cyan/red/green. Activity in the uncued subset is used to assess recruitment beyond the direct input.
Ongoing work. Partial-cue recruitment is an intermediate result toward the question of memory generalisation. The context–shock association mechanism, offline-stage hypotheses, and freezing-related readout remain under development.

What comes next

I am developing a four-stage protocol linking context learning, context–shock association, offline activity, and retrieval. Candidate mechanisms include BTSP-mediated hetero-association, EC3/SST modulation of plateau gating, and offline reactivation.

The key comparison will measure transfer of a shock-related representation alongside retained context identity, to distinguish memory association from a loss of context-specific information.

Get in touch

Let’s talk about brains & ideas.

Interested in memory, neural circuits, and biologically grounded computation.

Email me
Open full-resolution image ↗