Plasticity, Networks and Consolidation

Shine blue light on cells tagged during learning in a mouse’s dentate gyrus and the mouse freezes, as it did when it learned to fear a box it is no longer in. Headlines said the memory had been found. To decide what was found, you need three kinds of evidence: cells whose firing tracks an experience, signs that the experience is replayed after it ends, and manipulations that make a memory appear or disappear. This lesson takes them in that order, then builds a small network that recalls a whole pattern from part of it.

Cells that know where the animal is

In 1971 John O’Keefe and Jonathan Dostrovsky reported, in a short preliminary paper in Brain Research, single units in the hippocampus of freely moving rats that fired when the animal was in a particular part of a platform, often facing a particular way. These became place cells: hippocampal neurons that fire when the animal is in one region of an environment, the cell’s place field. The paper already framed the findings as support for a hippocampal cognitive, or spatial, map; O’Keefe and Lynn Nadel developed that into a full theory in their 1978 book.

You read a place cell from a rate map. Track the rat with a camera, divide the floor into bins, and for each bin divide the spikes by the seconds spent there. One hot spot on a cool floor is a place field.

In 2005 Torkel Hafting, Marianne Fyhn, Sturla Molden, May-Britt Moser and Edvard Moser recorded upstream, in the dorsocaudal medial entorhinal cortex of the rat. These grid cells fired at many places, and the places formed a lattice of roughly equilateral triangles. To test the regularity they used a spatial autocorrelogram: correlate the rate map with copies of itself shifted in every direction. A triangular lattice gives six peaks around the centre, at angles that are multiples of 60 degrees. Neighbouring grid cells shared an orientation but differed in phase, where the vertices fell, and grid spacing increased from dorsal to ventral. In the first large-enclosure sample, spacing ran from about 39 to 73 cm across cells. The grids persisted without external landmarks, which the authors took as consistent with path integration, keeping track of position from one’s own movement.

Two cautions. The lattice is in the rate map, not in the tissue: grid cells are not laid out in a honeycomb. And a cell that fires in one place is a correlation between firing and position. It does not show that the cell encodes only position, or that the animal uses it to navigate. O’Keefe and both Mosers shared the 2014 Nobel Prize in Physiology or Medicine.

Making a memory last

Consolidation is the transformation of a temporary, labile memory into a more stable form. It comes at two scales. Synaptic consolidation includes the processes that stabilize synaptic plasticity within hours of learning: the late, protein-synthesis-dependent phase of LTP you met in synaptic plasticity, with the gene expression behind it in signaling to the nucleus. Systems consolidation is the gradual reorganization of the brain systems that support a memory, over days to years.

What that reorganization is remains debated. The standard model holds that the hippocampus is needed for a time and the cortex eventually supports the memory alone, which fits patients whose hippocampal damage spares older memories more than recent ones (memory systems). Multiple-trace and transformation accounts hold that detailed episodic memories keep needing the hippocampus, while a more schematic version comes to rest in cortex. A memory does not simply move from one box to another. (New dentate granule cells, and the open question of whether adult humans make them, belong to adult neurogenesis.)

Replay and ripples: correlation, then perturbation

If the hippocampus teaches the cortex, it has to repeat experience after it happens. Matthew Wilson and Bruce McNaughton recorded many hippocampal cells at once in three rats, in slow-wave sleep before a spatial task, during the task, and in sleep afterwards. Cells that co-fired during the task were more likely to co-fire in the sleep after it than in the sleep before. Cells silent during the task showed no such increase. This was pairwise co-firing, not a decoded route.

Later work decoded order. In rats running on a track, Kamran Diba and György Buzsáki in 2007 found place-cell sequences replayed in reverse during pauses at the end of runs, and forward sequences before runs; Foster and Wilson had reported reverse replay in awake rats a year earlier. Much replay rides on sharp-wave ripples, brief bursts of fast oscillation in the CA1 local field potential. Replay in both directions was reported in awake rats (Diba and Buzsaki 2007).

All of this is correlation. The causal step came from cutting ripples short. Gabrielle Girardeau and colleagues in 2009 detected ripples online in rats during sleep after training on a hippocampus-dependent spatial task and triggered a stimulus that ended each one; rats given the same stimuli at a delay served as controls. Selective elimination of the ripples impaired performance. In 2019 Antonio Fernández-Ruiz and colleagues went the other way in awake rats: closed-loop optogenetic stimulation for 100 ms after detected ripples prolonged them, and maze working-memory performance improved. The comparison was control light delivered at random delays of 400 to 1000 ms after ripples, and the authors distinguish prolonging spontaneous ripples from inducing them at random. Note what disruption removes: every ripple-linked event, not one sequence.

Engrams: sufficiency and necessity

An engram is the physical change a learning experience leaves. Kitamura and colleagues define engram cells as cells activated by learning that undergo lasting change and can be reactivated by part of the original stimulus.

Liu and colleagues (2012) in Tonegawa’s laboratory used the promoter of c-fos, an immediate early gene turned on in recently active neurons, to put channelrhodopsin-2 into dentate-gyrus cells active while mice learned fear conditioning, a tone paired with shock in a chamber; the authors judged the engram likely contextual. Mice were off doxycycline for one or two days before training and back on it afterwards, which restricted tagging (the tools are in hacking the cell). Light on those cells evoked freezing in a different context. Two controls made it specific: cells tagged in a neutral context did not evoke fear freezing, and conditioned mice expressing EYFP without the opsin did not freeze to light. The authors themselves noted that sufficiency is not necessity.

Necessity needs silencing during natural recall. Kitamura and colleagues (2017) tagged prefrontal cells during learning, so the cortical engram formed at once, not after transfer. Inhibiting them disrupted natural-cue recall on day 12 but not day 2, and their dendritic-spine density was higher on day 12. Dentate engram cells, meanwhile, became less reactivated by natural cues by day 15 yet could still drive recall by light. Roy and colleagues (2022) mapped learning-activated cells across 247 mouse brain regions and flagged 117 as likely engram candidates; reactivating several ensembles together gave more recall than one. All of this is mouse contextual fear over days, not human memories over years.

Completing a pattern

How could a fragment recall a whole? The CA3 region of the hippocampus has dense recurrent connections between its pyramidal cells, so it is often modelled as an attractor network, in which a stored pattern is a stable state the activity falls into. John Hopfield’s 1982 network makes this concrete. Units are +1 or −1, weights are set by Hebb’s idea that co-active units strengthen their link, and units update one at a time.

Worked problem. These values are teaching choices, not hippocampal measurements. Store x = (1, 1, −1, −1) in four units:

w_ij = x_i × x_j / 4 for i ≠ j; w_ii = 0; s_i = sign(sum over j of w_ij × s_j)

Thresholds are zero, and a unit keeps its old value if its field is exactly zero.

Step 1: the weights. Unit 1’s row is (0, +1/4, −1/4, −1/4): units 1 and 2 agree, units 1 and 3, 1 and 4 disagree.

Step 2: a corrupted cue. Start at s = (−1, 1, −1, −1), unit 1 wrong. Its field is

h_1 = (1/4)(1) + (−1/4)(−1) + (−1/4)(−1) = 3/4

Positive, so s_1 becomes +1 and the state is (1, 1, −1, −1), the stored pattern.

Step 3: is it stable? At the stored pattern each field is x_i × 3/4, so the fields are (3/4, 3/4, −3/4, −3/4). Every sign matches its unit, so nothing changes.

Step 4: the energy. With E = −(1/2) × sum of w_ij × s_i × s_j, this network gives

E = −[(overlap)² − 4] / 8, where overlap = sum of x_i × s_i

The cue has overlap 2, so E = −(4 − 4)/8 = 0. The pattern has overlap 4, so E = −(16 − 4)/8 = −1.5. The update moved the state downhill.

Cue: toggle a unit to flip it (or click it in the picture)

Update next unit Run to stable Reset to the lesson’s cue

Overlap with x = Σ x_i s_i

—

Energy E = −[(overlap)² − 4] / 8

—

Updates so far | energy rises

—

—

In this sketch: four units of +1 or −1, one stored pattern x = (1, 1, −1, −1), w_ij = x_i x_j / 4 with w_ii = 0, zero thresholds, units updated one at a time in order 1 to 4, s_i = sign of the field, and a unit keeps its value if its field is exactly zero. These are the lesson’s teaching values, not hippocampal measurements.

Rule to keep: each unit takes the sign of the weighted sum of the others, so the state slides downhill in energy to a stored pattern or its inverse; CA3’s recurrent connections are often modelled this way, the model completes a pattern from part of it, capacity is limited, and more stored patterns give spurious mixtures.

The inverted pattern (−1, −1, 1, 1) is also stable, and as more patterns are stored, recall degrades into spurious mixtures. Capacity is limited.

Nakazawa and colleagues (2002) tested CA3 in mice lacking NMDA receptors only in CA3 pyramidal cells (the receptor’s coincidence detection is in ion channels as receptors). The mice learned a water-maze platform and recalled it normally with all the room cues, but were impaired when only a fraction of the cues remained. That is the deficit an attractor model predicts. How to record such ensembles across the brain is the subject of large-scale neuroscience.

Design problem

Light activation of dentate-gyrus cells tagged during fear conditioning makes mice freeze in a new box. A headline says scientists found where a memory is stored. Design the experiments that would test that headline, and say what each control rules out.

One way to do it

Start from what is already shown: sufficiency, with specificity. Liu’s neutral-context tags and EYFP-only mice did not freeze, so light, heating and tagging alone are excluded.

Test necessity next. Tag the same cells with an inhibitory opsin and silence them while the mouse is returned to the conditioning box, the natural cue. Include a fluorophore-only group given the same light. If freezing falls only in the opsin group, the cells are needed for natural recall at that time.

Vary the time and the cue. Test at a recent and a remote delay; Kitamura’s prefrontal engram was needed on day 12 and not day 2. Test partial cues, which engram cells are defined to respond to. Silence other regions too, since Roy’s map points to a distributed complex: “where the memory is” may have more than one answer.

Finally, “stored” needs a measured change in the tagged cells, such as Kitamura’s higher spine density or stronger synapses onto them, and a demonstration that blocking that change blocks the memory. Every result stays bounded to mouse contextual fear over days. Designing a circuit experiment takes these controls further.