Plasticity and how circuits learn

In BCI-2.1 you saw a neuron as a threshold unit summing thousands of inputs, each input weighted by how strong its synapse is. It is tempting to picture those weights as constants, numbers baked in once and read forever. They are not. The brain rewrites its own weights while it runs. That single fact is what learning is, what memory is, and, as you will see by the end, the reason a brain-computer interface decoder that works beautifully today can drift out of tune by tomorrow. This lesson derives how a synapse changes its strength, from the molecule up, and then cashes that out for the BCI.

The weights are not frozen, and they cannot be

Start from the job. A circuit that could never change its wiring could never learn anything. It would react to the world with whatever weights it was born with and never update from experience. So if brains learn at all, something must be able to change synaptic strength based on activity. That something is called plasticity: the capacity of a synapse (and by extension a whole circuit) to change how strongly it transmits, driven by the pattern of activity passing through it.

Plasticity is not one thing. It runs in both directions. Some activity patterns make a synapse respond more strongly to the same input, a durable strengthening called long-term potentiation (LTP). Other patterns do the reverse, a durable weakening called long-term depression (LTD). Depression here means turned down, not sad. Together LTP and LTD give a circuit a volume knob on every connection that experience can turn up or down and, crucially, leave turned.

Here is the programmer's hook, and its limit in the same breath. A synaptic weight really is like a trainable parameter in a neural network, and LTP and LTD really are like nudging that parameter up or down during training. That is a good intuition. Now the failure edge, because an analogy without its limit is a bug. In an artificial network, every weight is updated by a global optimizer that computes a single loss over the whole network and pushes each parameter along its gradient. The brain has no global loss and no backward pass. Each synapse changes from signals it can sense locally, its own presynaptic and postsynaptic activity, and nothing tells it whether the network as a whole did well. Biological learning is local and unsupervised at the synapse. Keep the weight-update picture for intuition, and drop the idea of a global gradient reaching down to set each weight.

Fire together, wire together (and the honest other half)

If the update is local, what local signal drives it? The oldest and most durable answer is Hebb's rule, usually compressed to four words: neurons that fire together wire together. Stated carefully: if a presynaptic neuron repeatedly helps make its postsynaptic partner fire, so the two are active in a correlated way, the synapse between them strengthens.

The half people forget is the honest corollary: out of sync, lose the link. If the presynaptic neuron keeps firing at times unrelated to whether the postsynaptic neuron fires, that connection weakens. Strengthening correlated inputs is only half a learning rule. Weakening the uncorrelated ones is what stops every synapse from creeping to maximum and going useless. LTP is the fire-together case. LTD is (roughly) the out-of-sync case. A circuit tuned this way slowly amplifies the inputs that reliably predict its own firing and fades the ones that do not.

Notice what Hebb's rule demands of the hardware. To know whether two neurons fired together, some physical part of the synapse has to detect coincidence, the presynaptic side being active and the postsynaptic side being active at the same time. A weight that updated on presynaptic activity alone could not tell a correlated input from a random one. So the biology needs a coincidence detector. Remarkably, one molecule is built to be exactly that.

The coincidence detector is a receptor with two locks

Recall from BCI-2.1 that at an excitatory synapse the presynaptic neuron releases glutamate, the brain's main excitatory neurotransmitter, and the postsynaptic membrane carries receptors that respond to it. There are two kinds of glutamate receptor at the synapses that learn, and the difference between them is the whole trick.

The first is the AMPA receptor. It is the everyday workhorse: glutamate binds, the channel opens, positive ions flow in, and the postsynaptic cell depolarizes a little. That is ordinary fast transmission, the input doing its normal job. AMPA receptors carry the signal.

The second is the NMDA receptor, and it is the coincidence detector. It is a channel that opens only when two conditions hold at once, which makes it a molecular AND gate. Condition one: glutamate must be bound, which means the presynaptic neuron just fired. Condition two: the postsynaptic membrane must already be depolarized. The reason for condition two is beautifully physical. At the resting voltage (recall the roughly -70 millivolts from S9.1 and BCI-1.1), a magnesium ion, written Mg2+, sits lodged inside the NMDA channel like a cork, and it plugs the pore even when glutamate is bound. Only when the postsynaptic cell is already depolarized by other activity does the voltage change shove that Mg2+ cork out of the way. So the channel passes ions only when glutamate is present (presynaptic fired) AND the postsynaptic cell is depolarized, typically because it is firing. Presynaptic alone: glutamate binds but the Mg2+ cork stays and nothing passes. Postsynaptic alone: no glutamate, nothing to open it. Both together: the cork pops and the channel conducts.

And what it lets through is the signal that matters: calcium, written Ca2+. The NMDA receptor, when both conditions are met, admits a pulse of Ca2+ into the postsynaptic spine (the small dendritic protrusion that holds the synapse). That Ca2+ pulse is the physical readout of "these two neurons just fired together." Hebb's abstract rule is implemented by a single ion passing a single gated pore that only opens on coincidence.

From a calcium pulse to a lasting change: signaling, then genes

A pulse of Ca2+ inside the spine is not itself a stronger synapse. It is a trigger. What happens next is signaling, exactly the kind of receptor-to-cascade machinery you met in S9.3: the calcium binds calcium-sensing proteins, which switch on enzymes, which act on targets, an amplifying relay that turns one brief ion pulse into a coordinated cellular response. Importantly, the size and timing of the Ca2+ pulse decide which way the synapse moves. A large, sharp Ca2+ influx (strong coincidence) drives the strengthening cascade toward LTP. A smaller, more drawn-out Ca2+ trickle (weak or mistimed activity) drives a different cascade toward LTD. Same ion, different amount, opposite outcome. That is how one detector produces both up and down.

Strengthening then happens on two timescales, and the split is the key to why memory is more than a chemical flicker. The fast, early phase does not wait for anything new to be built. The signaling cascade grabs AMPA receptors the cell already has and inserts more of them into the synapse. More AMPA receptors means the very same amount of glutamate now produces a bigger response. The weight went up, within minutes, using parts on hand.

But receptors inserted this way drift back out over hours. For a change that lasts days or a lifetime, the synapse has to do something that early LTP does not: it has to reach into the nucleus and turn on genes. The calcium cascade activates transcription factors, which switch on gene expression: they flip genes on (the gene regulation from S8), running the transcription and translation machinery from S5 to build new proteins that are shipped out to physically rebuild and enlarge the synapse. This late phase is why blocking transcription or protein synthesis leaves the early strengthening intact but erases the durable one. It is a genuine, load-bearing fact about biology: a fleeting memory and a permanent one are separated by a trip to the genome. Durable learning is, quite literally, a change in which genes a neuron is expressing.

The BCI payoff: you are decoding a moving target

Now the part that matters most for this track, and it falls straight out of everything above. If synaptic weights change with activity, then the pattern of neural activity a decoder learns to read is not a fixed function of intent. It is a function that the brain is continuously editing.

Here is the concrete failure. A decoder, which you will meet properly in BCI-3.2, is trained by recording a user's neurons while they attempt a movement and fitting a model that maps firing patterns to intended output. Train it today and it works. But overnight the brain does what brains do: plasticity keeps adjusting weights across the recorded circuits, partly from ordinary learning and partly in response to the implant itself. The neurons that coded "move right" with one firing pattern on Monday may code it with a shifted pattern by Tuesday. Nothing broke. No electrode has to fail for this. A large part of the drift is biological, since plasticity keeps adjusting the weights, and recording instability, the electrodes shifting relative to the neurons, adds more on top. A model fit to Monday's patterns degrades on Tuesday's. This is drift, and it comes from both the biology and the recording, not from a single broken part.

There is a second layer, and it is the honest and hopeful one. The brain adapts to the device too. Give a user a decoder and consistent feedback, and their motor circuits will, over sessions, reshape their own activity to drive that decoder better, the same LTP and LTD machinery tuning the very neurons the electrodes are reading. This is co-adaptation: the decoder is learning the brain while the brain is learning the decoder, two adaptive systems chasing each other. It is why users genuinely get better at operating a BCI with practice, sometimes dramatically. It is also why the mapping never fully settles.

The programmer's frame, with its limit. This is the machine-learning problem of a nonstationary signal, where the data distribution changes over time, so the honest engineering answers are the familiar ones: recalibrate periodically, or build a decoder that keeps learning online instead of freezing after training. But mark the failure edge against ordinary concept drift. In a normal deployment you can, in principle, freeze the data source and study a stationary snapshot. Here you cannot freeze the brain, and worse, your own model is one of the things making it move, because the feedback the decoder gives is what the brain adapts to. It is not one drifting distribution. It is two coupled learners, and the coupling is the point. The correct mental model of a BCI decoder is not a function you fit once. It is a control loop you keep closed with a partner that is also learning.

Key terms

plasticity

The capacity of a synapse or circuit to change how strongly it transmits, driven by the pattern of activity passing through it. It is the physical basis of learning and memory.

long-term potentiation (LTP)

A durable strengthening of a synapse, so the same presynaptic input produces a larger postsynaptic response, triggered by strong correlated activity.

long-term depression (LTD)

A durable weakening of a synapse, triggered by weak or uncorrelated activity. The counterweight that keeps synapses from all saturating high.

Hebbian rule

The principle that neurons which fire together wire together, and its corollary that out-of-sync connections weaken. A local, unsupervised rule for changing synaptic weights.

NMDA receptor

A glutamate receptor that acts as a coincidence detector: it passes calcium only when glutamate is bound AND the cell is already depolarized, which relieves its magnesium block.

AMPA receptor

The workhorse glutamate receptor that carries ordinary fast transmission. Adding or removing AMPA receptors is how a synapse's early strength is turned up or down.

co-adaptation

The two-way learning between a BCI user and a decoder: the brain reshapes its activity to drive the device while the decoder is retrained on the brain, each chasing the other.

drift

The gradual change over hours and days in the neural patterns a decoder was trained on, caused partly by ongoing plasticity and partly by recording instability, which degrades a fixed decoder and forces recalibration or online learning.

Timing, not just togetherness: spike-timing-dependent plasticity

Hebb's rule says fire together, but neurons take real experiments seriously, and the sharper story is about order and timing on the scale of milliseconds. In spike-timing-dependent plasticity (STDP), if the presynaptic spike arrives just before the postsynaptic neuron fires, so the input plausibly helped cause the output, the synapse strengthens (LTP). If the presynaptic spike arrives just after, so it could not have contributed, the same synapse weakens (LTD). The switch between strengthening and weakening can hinge on a difference of only a few milliseconds. This is a lovely fit to the calcium story: the exact timing of presynaptic glutamate relative to postsynaptic depolarization sets how much Ca2+ the NMDA receptor admits and in what shape, and the amount of calcium is what tips the cascade toward LTP or LTD. STDP is Hebb's rule made causal: not merely correlated, but in the right order to look like cause and effect. It is also a hint at why the brain is such a moving target, since a rule this sensitive to millisecond timing is constantly nudging weights as ordinary activity flows through.