← Reference Library

REFERENCE

Synaptic PlasticityLTP / STDP

The junction between two neurons changes its own strength according to use — strengthening when used one way, weakening when used another. That's how learning and memory get stored in a lump of tissue; it's also the most direct correspondence between biological brains and artificial networks.

The one-line version: fire together, wire together

In 1949 Hebb proposed a conjecture that later got compressed into a catchy line: fire together, wire together — if neuron A consistently fires just before B, so that A looks like it helped cause B to fire, then the A→B junction should get stronger.

What makes the conjecture powerful is that it is purely local: a synapse only needs to know what happened at its own two ends. No instruction from above, no knowledge of what the rest of the network is doing, and experience gets written into structure. Two decades later it was actually measured in the hippocampus: give a bundle of axons a burst of high-frequency stimulation and the response to the same stimulation afterwards is durably larger, lasting hours to months — that's long-term potentiation (LTP). Conversely, prolonged low-frequency stimulation weakens it: long-term depression (LTD).

How it's done at the molecular level

At the centre is a protein called the NMDA receptor, which is by construction a coincidence detector — it only opens when two conditions hold at once:

Condition one: the presynaptic neuron actually fired

It releases glutamate (the brain's main excitatory transmitter), which binds the receptor.

Condition two: the postsynaptic neuron is excited right now

The NMDA receptor's pore is normally plugged by a magnesium ion; only depolarisation of the postsynaptic neuron itself pops that plug out.

Both satisfied → calcium floods in

Calcium is the decisive signal. A large, fast influx takes the strengthening route; a slow trickle takes the weakening one. Same signal — amount and speed set the direction.

The result: what changes is receiving capacity

Strengthening mostly means inserting more AMPA receptors (the ones that actually do the fast signal reception) into the postsynaptic membrane; weakening removes them. That's how the synapse's volume knob is turned.

Order sets the direction: tens of milliseconds decide

Hebb's line hides a directionality — "A helped cause B," not "A and B at the same time." Experiments in the 1990s quantified exactly that: whether a synapse strengthens or weakens depends on the order in which the two neurons fired, with a tolerance of only tens of milliseconds. This is spike-timing-dependent plasticity (STDP).

The curve is doing something rather clever: it detects causation, not correlation. Whatever came first is the more likely cause, so the synapse rewards it; whatever arrived late didn't help, so it gets cut. A junction that watches nothing but the time difference across its own two ends manages a small verdict on "what caused what."

Why "long-term" needs new protein

Potentiation comes in two phases of quite different character:

Early LTP (tens of minutes)

Just shuffling and modifying receptors that already exist; no new materials needed. Fast to arrive, and it decays on its own.

Late LTP (hours and beyond)

Requires gene expression and new protein synthesis — actually growing new structure. Neurotrophic factors like BDNF act at precisely this step: they're among the key agents converting "temporarily stronger" into "durably stronger." This is why things apparently unrelated to studying — exercise, sleep — can affect memory: they change the conditions of this step.

Synaptic tagging and capture

An elegant solution to a real problem: the new proteins are made by the neuron as a whole and carry no address, so how do they know which synapse to reinforce? Recently activated synapses hang out a temporary tag, and newly made proteins drifting past get captured by those tags. Global resources meet local bookkeeping.

It can't only strengthen: the homeostatic half

A pure Hebbian rule has a fatal flaw — it runs away: a strong synapse makes the postsynaptic neuron fire more easily, more firing strengthens it further, and the positive feedback rolls until the network is either fully saturated or completely silent.

So the brain also runs homeostatic plasticity on a slower timescale as a backstop. The classic case is synaptic scaling: when a neuron notices its overall activity has drifted from its normal set point, it turns all of its synapses up or down by the same factor. The beauty is in "by the same factor" — the relative strengths between synapses (i.e. the stored memory) are preserved; only the overall volume changes. The fast system learns; the slow one keeps it from learning itself to death.