Spike-Timing-Dependent Plasticity¶

A synaptic learning rule where the sign of a weight change depends on the millisecond order of pre- and postsynaptic spikes — pre-before-post potentiates, post-before-pre depresses — turning a coincidence detector into a causality detector that grows directed connectivity.

Core Idea¶

Spike-timing-dependent plasticity (STDP) is the synaptic learning rule in which the direction and magnitude of long-lasting change in synaptic efficacy are determined by the relative timing — on a millisecond scale — of presynaptic and postsynaptic action potentials: when the presynaptic spike precedes the postsynaptic spike within a window of roughly 10–40 milliseconds, the synapse is potentiated (LTP); when the order is reversed — postsynaptic before presynaptic — the synapse is depressed (LTD); and the magnitude of the change decays as the absolute timing interval grows, producing a characteristic asymmetric biphasic STDP curve. The rule was systematically demonstrated in cortical and hippocampal preparations by Guo-Qiang Bi and Mu-Ming Poo (1998) and in developing cortex by Henry Markram and colleagues (1997). The biophysical mechanism is NMDA-receptor coincidence detection: the NMDA receptor requires both presynaptic glutamate binding and postsynaptic membrane depolarisation to admit calcium; the pre-then-post order reliably satisfies both conditions within the millisecond window, producing the calcium influx that recruits CaMKII and triggers LTP-related downstream cascades; the post-then-pre order produces a weaker, differently-timed calcium signal that activates phosphatases and drives LTD. The consequence at the network level is that STDP is a causality detector: it selectively strengthens synapses from neurons that could plausibly have caused the postsynaptic firing (presynaptic spike precedes postsynaptic spike) and weakens those from neurons that could not (postsynaptic spike precedes the presynaptic one), so a network with STDP develops directed connectivity that tracks the causal-temporal structure of its input statistics rather than the mere co-occurrence structure that symmetric Hebbian learning would capture.

Structural Signature¶

Sig role-phrases:

the synapse-connected pair — a presynaptic and a postsynaptic neuron joined by a synapse with an existing weight
the millisecond-resolved spike times — pre- and postsynaptic action potentials timed at fine resolution, yielding a signed interval Δt
the NMDA coincidence detector — the receptor admitting calcium only when presynaptic glutamate binding and postsynaptic depolarisation coincide, implementing the timing window biophysically
the locality constraint — only the two endpoint spikes and the current weight enter the update; no global controller
the sign-asymmetric update — pre-before-post (within ~10–40 ms) potentiates via strong calcium → CaMKII (LTP); post-before-pre depresses via weaker calcium → phosphatases (LTD)
the biphasic STDP curve — weight change as one function of Δt, two opposite-sign lobes decaying as |Δt| grows
the causality-detector outcome — at network scale, crediting plausible causal predecessors and debiting non-predecessors, growing directed connectivity that tracks causal-temporal (not mere co-occurrence) structure

What It Is Not¶

Not plain "fire together, wire together." STDP refines Hebbian learning by making the sign of the weight change depend on the millisecond order of pre- and postsynaptic spikes — pre-before-post potentiates, post-before-pre depresses. Symmetric Hebbian rules treat co-activation without regard to order and cannot tell whether the presynaptic neuron led or followed; STDP can, which is the whole point.
Not a coincidence detector. What it reads is temporal order, not mere overlap. Two cells firing "together" is not enough: the asymmetric biphasic curve credits inputs that preceded the output and debits those that followed it. Treating STDP as detecting simultaneity collapses the order-asymmetry that turns it into a causality detector.
Not a bare phenomenology. The sign rule is grounded in a specific mechanism — NMDA-receptor coincidence detection, where the pre-then-post order admits the strong calcium influx that recruits CaMKII for LTP, and the reverse order gives a weaker, differently-timed calcium signal that drives LTD. The same machinery yields opposite outcomes by order; the curve is a prediction about biophysics, not a curve fit.
Not performing causal inference. STDP exploits a causality-like asymmetry but does not reason about causes; it merely makes synaptic weights track the causal-temporal statistics of its inputs. The network ends up wired to reflect who plausibly drove whom, but no inference is computed — the arrow of time is written into weights by a local update, not inferred by an agent.
Not a globally coordinated rule. The update is strictly local: only the two endpoint spikes and the existing weight enter it, with no global controller surveying the network. The directed connectivity a population grows is an emergent consequence of applying one curve per synapse over input statistics, not the product of any centralised allocation of credit.

Scope of Application¶

STDP lives across the synaptic-plasticity subfields of neuroscience and extends intact to the engineered spiking and neuromorphic substrates built to host it; its reach is bounded by the presence of spikes whose order carries information and a coincidence-detecting mechanism to read it — one abstraction layer up, the temporal-credit-assignment pattern it instantiates travels under that parent, but the calcium-and-spike rule itself does not.

Hippocampal LTP/LTD — the classical Bi & Poo (1998) demonstrations of order-sensitive potentiation and depression at Schaffer-collateral synapses.
Cortical synaptic plasticity — the rule measured across species and developmental stages, including the developing-cortex findings of Markram and colleagues (1997).
Cerebellar plasticity — climbing-fibre-driven LTD sharing the temporal-order sensitivity that defines the rule.
Computational neuroscience — STDP as core machinery in spiking-neural-network models of learning, perception, and motor adaptation.
Neuromorphic hardware — the canonical local learning rule on chips such as Loihi, TrueNorth, and BrainScaleS, prized for its locality, low memory footprint, and unsupervised character; here it is genuinely the same rule on a different physical substrate, the spike, window, and sign-asymmetric update all literally present.

Clarity¶

STDP's clarifying force is to replace the slogan "neurons that fire together wire together" with a precise, order-sensitive rule — and in doing so to convert a coincidence detector into a causality detector. Plain Hebbian learning treats co-activation symmetrically: it cannot tell whether the presynaptic neuron led or followed. By making the sign of the weight change depend on the millisecond order of pre- and postsynaptic spikes, STDP makes legible that timing, not mere overlap, carries the information — pre-before-post strengthens (the input could have helped cause the output), post-before-pre weakens (it could not). The sharper question a synaptic physiologist can now ask is not "did these two cells fire together?" but "in what order, within what window, and therefore is this connection being credited or debited?"

The deeper distinction it sharpens is between co-occurrence structure and causal-temporal structure in a network's wiring. Because the rule is order-asymmetric, a population governed by STDP develops directed connectivity that tracks which neurons plausibly drive which others, whereas symmetric Hebbian learning could only carve out undirected co-activation clusters. That reframing — a learning rule reading the arrow of time off spike order and writing it into the connectome — is what links the millisecond biophysics of NMDA-receptor coincidence detection to a network-level claim about how causal chains get reinforced over mere correlations, and it is what keeps STDP analytically separate from the broad family of timing-insensitive plasticity rules it refines.

Manages Complexity¶

Two problems that look sprawling collapse under STDP onto one compact rule. The first is the heterogeneity of synaptic change itself — the seemingly separate phenomena of potentiation here and depression there, induced under varied stimulation protocols across cortical and hippocampal preparations. The second is the network-scale mystery of how a population's connectivity comes to reflect anything about its inputs. STDP tames both by reducing the weight change at any synapse to a single function of a single variable: the signed timing interval Δt between presynaptic and postsynaptic spikes, read off the asymmetric biphasic curve. Potentiation and depression stop being two phenomena and become the two lobes of one curve; the whole induction story reduces to "where on the Δt axis did this spike pair land." Because the rule is local — only the two endpoint spikes and the existing weight matter — the entire emergent behaviour of an arbitrarily large network follows from applying that one curve per synapse over the input statistics, with no global controller to track.

What the analyst then reads off operates at two levels from the same rule. At the synapse, the sign of Δt is a credit/debit verdict: pre-before-post (within roughly 10–40 ms) credits the connection — the input could have helped cause the output — and post-before-pre debits it, so "is this connection strengthening or weakening, and by how much" is read directly from spike order and interval rather than re-derived from the biochemistry. At the network, the rule's order-asymmetry licenses a sharp qualitative prediction that symmetric Hebbian learning cannot make: a population under STDP develops directed connectivity tracking the causal-temporal structure of its inputs (which neurons plausibly drive which), whereas a timing-insensitive rule could only carve undirected co-occurrence clusters. So the analyst predicts the kind of wiring a network will grow — causal chains selectively reinforced over mere correlations — by tracking one property of the rule, its order-asymmetry, instead of simulating the molecular detail of every connection. The move is from a scatter of plasticity protocols plus an open question about emergent wiring to a single biphasic curve over Δt whose sign and magnitude the analyst reads to settle per-synapse credit and, at scale, to read the arrow of time off spike order and predict directed connectivity — collapsing a high-dimensional cellular-and-network problem onto one local rule with one input variable.

Abstract Reasoning¶

STDP's signature inference reads causality off temporal order, converting a coincidence detector into a causality detector. At a single synapse the analyst reasons from the sign of the timing interval Δt to a credit-or-debit verdict on the connection: a presynaptic spike preceding the postsynaptic one (within roughly 10–40 ms) means the input could have helped cause the output, so the synapse is credited and potentiated; the reverse order means the input could not have caused the output, so it is debited and depressed. The magnitude follows the asymmetric biphasic curve, decaying as |Δt| grows. The move is to treat spike order as evidence about who drove whom, and to write that evidence into the weight — so the question shifts from "did these two cells fire together?" to "in what order, within what window, and therefore is this connection being credited or debited?" This is the inference that distinguishes STDP from timing-insensitive Hebbian rules, which treat co-activation symmetrically and cannot tell whether the presynaptic neuron led or followed.

The mechanistic inference grounds the sign rule in coincidence detection and lets the analyst reason from biophysics to outcome. The NMDA receptor admits calcium only when presynaptic glutamate binding and postsynaptic depolarisation coincide, so the analyst infers that the pre-then-post order reliably satisfies both conditions within the millisecond window, producing the strong calcium influx that recruits CaMKII and triggers LTP, while the post-then-pre order produces a weaker, differently-timed calcium signal that activates phosphatases and drives LTD. The reasoning runs from a molecular coincidence requirement to the direction of the weight change — same machinery, opposite outcomes depending on order — which is what makes the sign rule a prediction about a specific mechanism rather than a bare phenomenology.

The most distinctive move operates at the network level: predict the kind of connectivity a population will grow by tracking one property of the rule, its order-asymmetry. Because the local sign rule credits causal predecessors and debits non-predecessors, the analyst infers that a population under STDP develops directed connectivity tracking the causal-temporal structure of its inputs — which neurons plausibly drive which — and selectively reinforces causal chains over mere correlations. This is a qualitative prediction a symmetric Hebbian rule cannot make: timing-insensitive learning could only carve undirected co-occurrence clusters. So the analyst reasons forward from "this network uses STDP" to "it will read the arrow of time off spike order and write it into the connectome," and predicts, for instance, that an upstream neuron that consistently leads a downstream one will become a progressively more reliable trigger of it, while a connection whose order is reversed by other inputs will fade. Crucially, because the rule is local — only the two endpoint spikes and the existing weight matter — this network-scale behaviour follows from applying one curve per synapse over the input statistics, with no global controller to track.

The boundary on all these inferences is the spike-timing substrate they presuppose. The credit/debit verdict, the calcium-mechanism reasoning, and the directed-connectivity prediction depend on discrete pre- and postsynaptic action potentials with millisecond-resolved timing, a narrow asymmetric window, sign-asymmetric update, and the NMDA-receptor coincidence biophysics that implements the window. The analyst draws a sharp distinction the rule is built to honour — causal-temporal structure versus mere co-occurrence structure — and the moves have force precisely where there are spikes whose order carries information and a calcium-dependent mechanism to read it; it is that fine-timing, locally-computed, order-asymmetric machinery, not a general notion of order-mattering, that makes the causality-detection and directed-wiring inferences load-bearing.

Knowledge Transfer¶

Within neuroscience STDP transfers as mechanism, and the transfer is exact because the rule is local and biophysically specified. The same signed-Δt biphasic curve, the same NMDA-receptor coincidence detection, and the same causality-detector reading carry across hippocampal LTP/LTD (the classical Bi & Poo 1998 findings), cortical synaptic plasticity (measured across species and developmental stages), and cerebellar plasticity (climbing-fibre-driven LTD sharing the temporal-order sensitivity). What carries is the working content: the per-synapse credit/debit verdict (read the sign of Δt — pre-before-post credits, post-before-pre debits), the mechanistic inference (same calcium machinery, opposite outcomes by order), and the network-scale prediction (a population under STDP grows directed connectivity tracking causal-temporal structure, where symmetric Hebbian learning could only carve undirected co-occurrence clusters). Notably, the transfer extends intact to an engineered substrate that was built to host it: STDP is the canonical local learning rule in spiking neural networks and on neuromorphic hardware (Loihi, TrueNorth, BrainScaleS), prized for its locality, low memory footprint, and unsupervised character — here it is genuinely the same rule, a different physical implementation, not an analogy, because the spike, the timing window, and the sign-asymmetric update are all literally present.

Beyond those spike-bearing substrates the honest reading is mixed, and worth stating precisely. There is a real shared abstract mechanism (case B): the substrate-general residue STDP instantiates is causal-order-dependent credit assignment — order matters, the earlier event credits the later one, and the sign of the update respects the arrow of time — and that pattern genuinely recurs as co-instances across distinct learning substrates: temporal-difference learning (credit assigned backward from rewards to predecessor states), eligibility traces in reinforcement learning (a temporal window over which a state stays eligible for credit), and backward induction in decision theory. These are not metaphors; they are independent solutions to the same temporal-credit-assignment problem, which is why STDP and TD learning are best related as siblings under that parent rather than one borrowing the other's name. But the cross-domain lesson should carry that parent — temporal-order-sensitive credit assignment (a composition of temporal_dynamics, credit_assignment, and causality) — not "STDP," because the STDP-specific cargo stays home: the millisecond window, the sign-asymmetric biphasic curve, the calcium influx, and the NMDA coincidence biophysics have no counterpart outside neural-or-neuromorphic substrates. And where the reach goes further still — "early collaborators got more credit for the project than late joiners," touchpoints preceding a marketing conversion credited while later ones are not — it lapses into analogy (case A): the temporal-order shape is borrowed while the narrow window, the biphasic sign rule, and the local-only-information constraint are all absent, so only the silhouette of the rule travels. The disciplined statement is therefore: STDP refines hebbian_learning by adding order-asymmetry and ports exactly to spiking and neuromorphic substrates; one abstraction layer up, the temporal-credit-assignment pattern it instantiates travels literally as a co-instance to other learning systems; and the named neural rule itself, with its calcium-and-spike machinery, does not.

Examples¶

Canonical¶

Guo-Qiang Bi and Mu-Ming Poo's 1998 experiment on cultured hippocampal neurons is the defining demonstration. They paired presynaptic and postsynaptic firing at precisely controlled relative times and measured the resulting lasting change in synaptic strength. The result was the asymmetric biphasic curve: when the presynaptic neuron fired before the postsynaptic one, within a window of roughly a few tens of milliseconds, the synapse potentiated (strengthened, LTP), and the effect was largest at the shortest positive intervals; when the order was reversed — postsynaptic before presynaptic — the synapse depressed (weakened, LTD); and the magnitude of both effects fell off as the absolute timing gap widened, vanishing beyond the window. Crucially, the sign of the change flipped with the order of the two spikes, even when the pair fired only a millisecond or two apart across the crossover point.

Mapped back: The precisely-timed pre/post pairs are the millisecond-resolved spike times yielding a signed Δt. Potentiation for pre-before-post and depression for post-before-pre is the sign-asymmetric update, and the whole measured shape — two opposite lobes decaying with |Δt| — is the biphasic STDP curve. The order-flip at the crossover is what makes the synapse a causality detector rather than a mere coincidence detector.

Applied / In Practice¶

STDP is deployed as an actual learning algorithm on neuromorphic hardware. On chips such as Intel's Loihi, spiking neural networks learn unsupervised feature detectors using STDP as the local weight-update rule: when an input pattern repeatedly drives a neuron to fire, the synapses whose presynaptic spikes reliably preceded that firing are potentiated and those that lagged are depressed, so the neuron sharpens into a detector of the recurring, causally-leading input pattern. Because the rule needs only each synapse's own two spike times and current weight — no global error signal, no backpropagation — it maps efficiently onto hardware with local memory at each synapse, which is precisely why neuromorphic designers favor it for low-power, on-chip learning.

Mapped back: Here STDP is genuinely the same rule on an engineered substrate: the spike, the timing window, and the sign-asymmetric update are all literally present. That each update uses only the synapse's own endpoint spikes and weight is the locality constraint, the property that makes on-chip implementation efficient. The neuron sharpening into a detector of its causally-leading input is the causality-detector outcome — directed connectivity tracking causal-temporal structure — realized in silicon.

Structural Tensions¶

T1: Temporal order versus genuine causality (the "causality detector" detects order, a fallible proxy). STDP's headline reframing is that order-asymmetry converts a coincidence detector into a causality detector — pre-before-post credits an input that could have caused the output. But what the rule actually reads is temporal precedence, which correlates with causation without being it: a common upstream driver that makes neuron A lead neuron B will get A→B credited even when A does not drive B, and the entry itself insists STDP "does not reason about causes" — it merely makes weights track causal-temporal statistics. The tension is that the powerful "causality" label overstates what a local order-sign rule can do, so the network's directed connectivity reflects lead-lag structure that is a proxy for, and sometimes a spurious stand-in for, real causal influence. Diagnostic: Is the strengthened connection here tracking genuine causal drive, or merely a temporal-precedence correlation (common cause, spurious lead-lag) that STDP credits identically because it reads order, not causation?

T2: Locality versus distal credit assignment (the millisecond window cannot bridge to delayed consequences). The rule's locality — only the two endpoint spikes and the current weight enter the update — is its great virtue: biologically plausible, hardware-efficient, no global controller. But that same locality confines credit to the narrow ~10–40 ms window, so STDP cannot assign credit across the seconds-to-minutes gaps that separate an action from its reward, the distal temporal-credit-assignment problem that eligibility traces and backpropagation-through-time are built to solve. The tension is that the property making STDP cheap and local is exactly what prevents it from bridging long delays, so the rule that elegantly captures millisecond causal order is blind to the longer causal chains that matter for goal-directed learning. Diagnostic: Is the credit assignment needed here within STDP's millisecond window, or does it span delays the local rule cannot bridge without an added eligibility-trace or global signal?

T3: The canonical biphasic curve versus real-synapse variability (the clean rule is an idealization synapses deviate from). The asymmetric biphasic curve over Δt — pre-before-post potentiates, post-before-pre depresses, magnitude decaying with |Δt| — is what gives STDP its precise, mechanistic content. But the canonical curve is not universal: real STDP shape varies by synapse type, dendritic location, developmental stage, firing rate (frequency-dependent and triplet effects), and neuromodulatory state, so some synapses show reversed or symmetric windows and the rate can override the timing. The tension is that the crisp textbook curve that makes STDP a sharp, predictive rule is a partial idealization of a family of context-dependent behaviors, so applying "the STDP curve" uncritically imports a specificity that many real synapses do not honor. Diagnostic: Does the synapse in question actually follow the canonical pre/post biphasic window, or is its plasticity rate-dominated, neuromodulator-gated, or otherwise deviating from the idealized curve?

T4: Self-reinforcing potentiation versus stability (the causal-credit rule runs away without homeostasis). STDP's crediting of causal predecessors is inherently a positive-feedback process: potentiating a synapse whose presynaptic spike led the postsynaptic firing makes that input more likely to lead again, earning further potentiation, so causally-leading connections can grow without bound while others vanish. The order-sensitivity that makes the rule a causality detector is thus intrinsically destabilizing, and functional networks require added mechanisms — weight bounds, synaptic scaling, homeostatic plasticity — to keep it from saturating. The tension is that the very asymmetry giving STDP its discriminating power also gives it a runaway tendency, so the elegant local rule is not self-stabilizing and must be paired with global or homeostatic machinery it does not itself provide. Diagnostic: Is the STDP here bounded by a stabilizing mechanism (weight caps, synaptic scaling), or is the positive feedback of order-credit allowed to run toward saturation and silence?

T5: Unsupervised self-organization versus task-directedness (structure tracked is not structure that serves a goal). STDP grows directed connectivity that tracks the causal-temporal statistics of its inputs with no error signal and no objective — a strength for biological plausibility and low-power on-chip learning. But tracking input statistics is not optimizing a task: the structure STDP wires up reflects whatever lead-lag regularities the input contains, which need not be the regularities useful for a downstream goal, so an STDP network can faithfully learn causal structure that is behaviorally irrelevant. The tension is that the same objective-free locality that makes STDP cheap and unsupervised also means it optimizes nothing in particular, so its self-organized wiring must be shaped by input curation or reward-modulated variants to become task-useful. Diagnostic: Is the connectivity STDP is growing here aligned with a useful task, or is it faithfully tracking input lead-lag structure that carries no benefit for the network's actual function?

T6: Autonomy versus reduction (a named neural rule, a refinement of Hebbian learning, or an instance of temporal-credit-assignment). STDP is a specific, biophysically-specified rule with proprietary cargo — the millisecond window, the sign-asymmetric biphasic curve, NMDA-receptor coincidence detection, the calcium-CaMKII/phosphatase cascades — and it transfers literally (not by analogy) to spiking and neuromorphic substrates built to host spikes, timing windows, and sign-asymmetric updates. Below, it refines the prime hebbian_learning by adding order-asymmetry. Above, it instantiates a substrate-general pattern — causal-order-dependent credit assignment (a composition of temporal_dynamics, credit_assignment, and causality) — that recurs as genuine co-instances in TD learning, eligibility traces, and backward induction. The calcium-and-spike machinery stays home; the temporal-credit-assignment pattern travels. Diagnostic: Resolve toward the temporal-credit-assignment pattern (and hebbian_learning) when carrying the earlier-event-credits-later lesson to non-spiking learning systems; toward STDP when spikes, a timing window, and sign-asymmetric updates are literally present.

Structural–Framed Character¶

Spike-timing-dependent plasticity sits toward the structural end of the spectrum but stops short of the pole — best read as mixed-structural, in the family of isostasy and spatial updating: a genuine, evaluatively-neutral natural mechanism wearing heavy domain vocabulary. Its evaluative_weight is nil: a synapse potentiating or depressing by spike order is neither good nor bad, and "LTP"/"LTD" name directions of change, not verdicts. It is not human_practice_bound: synapses run this rule in living brains automatically, below any observer — remove every neuroscientist and the Schaffer-collateral synapse still credits its causal predecessors. Its institutional_origin is none: the rule is a fact of how NMDA-receptor biophysics reads spike timing, discovered rather than stipulated (Bi and Poo, Markram documented a mechanism nature already ran). And cross-substrate reuse is, unusually strongly, recognition rather than import: STDP ports literally — not by analogy — to engineered spiking and neuromorphic substrates (Loihi, TrueNorth, BrainScaleS), where the spike, the timing window, and the sign-asymmetric update are all physically present, and one abstraction layer up its causal-order credit-assignment recurs as genuine co-instance in TD learning, eligibility traces, and backward induction.

What keeps it off the structural pole is vocab_travels: STDP's distinctive cargo — the millisecond window, the sign-asymmetric biphasic curve, the calcium influx, the NMDA coincidence detection, the CaMKII/phosphatase cascades — is irreducibly neuro-biophysical and has no counterpart outside spike-bearing substrates; where the reach goes further ("early collaborators got more credit than late joiners") it lapses into mere analogy. The portable skeleton is causal-order-dependent credit assignment — the earlier event credits the later one, the sign of the update respecting the arrow of time — a composition of the parent primes temporal_dynamics, credit_assignment, and causality, and below it a refinement of hebbian_learning by added order-asymmetry. That skeleton is what STDP instantiates from its parents and what travels as co-instance to other learning systems; the cross-domain reach belongs to them, while the calcium-and-spike machinery stays home. Its character: structural in skeleton — a real, evaluatively-neutral, recognized-in-nature (and literally silicon-portable) causal-order credit-assignment rule — but stated in neurobiological vocabulary that pins it to spike-bearing substrates, leaving it mixed-structural rather than a free-floating prime.

Structural Core vs. Domain Accent¶

This section decides why spike-timing-dependent plasticity is a domain-specific abstraction and not a prime, and carries the case for its domain-specificity — so it is worth being exact about what could lift and what cannot.

What is skeletal (could lift toward a cross-domain prime). Strip the biophysics and a thin relational structure survives: the sign of an update to a directed link between two events depends on their temporal order, with the earlier event credited for the later one and the later-then-earlier order debited, the magnitude decaying as the interval grows. The portable pieces are abstract — a pair of time-stamped events, a signed interval between them, an asymmetric rule that credits precedence and debits succession, and an update that respects the arrow of time. That is causal-order-dependent credit assignment: order matters, the earlier event earns credit toward the later one, and the sign of the change tracks temporal direction. The skeleton is genuinely substrate-portable — which is why it recurs as genuine co-instances in temporal-difference learning, eligibility traces, and backward induction, and is best named as a composition of temporal_dynamics, credit_assignment, and causality, refining hebbian_learning by the added order-asymmetry. But it is the core the entry shares, not what makes STDP distinctive.

What is domain-bound. Almost everything that makes the rule STDP in particular is neurobiological furniture, and none of it survives extraction into a non-spiking substrate. The synapse-connected pre/postsynaptic pair; the millisecond-resolved action potentials that yield the signed Δt; the roughly 10–40 ms window; the sign-asymmetric biphasic curve with its two decaying lobes; the NMDA-receptor coincidence detection that admits calcium only when presynaptic glutamate binding and postsynaptic depolarisation coincide; the strong-calcium→CaMKII cascade for LTP and the weaker-calcium→phosphatase cascade for LTD — these are the worked mechanism, the instruments, and the empirical cases, all specific to spike-bearing tissue. The decisive test: remove the spike and the calcium-dependent coincidence detector and there is no window to land on and no biphasic curve to read; what is left is a bare "earlier-credits-later" shape with none of STDP's machinery. Notably, that machinery is not removed by moving to neuromorphic hardware — on Loihi, TrueNorth, and BrainScaleS the spike, the timing window, and the sign-asymmetric update are all literally present, which is why that crossing is genuine mechanism-recognition, not analogy; the substrate stays spike-bearing.

Why this does not clear the prime bar. A prime is a relational structure whose vocabulary travels and whose transfer is recognition of the same mechanism, not analogy. STDP's transfer is bimodal in a sharper-than-usual way, because its literal reach extends one substrate wider than its home. Within neuroscience and its engineered spiking substrates the rule travels intact — hippocampal, cortical, and cerebellar plasticity and neuromorphic silicon all supply spikes whose order carries information and a coincidence detector to read it, so the signed-Δt curve, the calcium mechanism, and the directed-connectivity outcome recur as the same rule, not an analogy. Beyond spike-bearing substrates it travels only by analogy: "early collaborators got more credit than late joiners," a marketing touchpoint preceding a conversion credited over later ones — these borrow the temporal-order silhouette while the narrow window, the biphasic sign rule, and the local-only-information constraint are all absent. Crucially, when the bare structural lesson is needed across non-spiking learning systems, it is already carried, in more general form, by the pattern STDP instantiates: hebbian_learning supplies the correlational base, and the composition of temporal_dynamics, credit_assignment, and causality supplies the order-sensitive credit assignment that recurs as a literal co-instance in TD learning, eligibility traces, and backward induction. The cross-domain reach belongs to those parents; the calcium-and-spike machinery is domain baggage that should stay home.

Relationships to Other Abstractions¶

Current abstraction Spike-Timing-Dependent Plasticity Domain-specific

Parents (1) — more general patterns this builds on

Spike-Timing-Dependent Plasticity is a kind of Synaptic Plasticity Domain-specific

STDP is the synaptic-plasticity subtype whose direction and magnitude axis is fixed by the signed millisecond interval between endpoint spikes.

Not to Be Confused With¶

Hebbian learning ("fire together, wire together"). The super-type STDP refines. Plain Hebbian rules strengthen a synapse when pre- and postsynaptic cells co-activate, but treat that co-activation symmetrically — they cannot tell whether the presynaptic neuron led or followed. STDP is the species that adds the millisecond sign-asymmetry: pre-before-post potentiates, post-before-pre depresses. Part-whole: every STDP synapse is Hebbian, but not every Hebbian rule is order-sensitive. Tell: does the rule flip the sign of the weight change when you swap which spike came first (STDP), or respond only to the fact of co-activation regardless of order (symmetric Hebbian)?
LTP / LTD (long-term potentiation and depression). These name the two directions of lasting synaptic change — the outcomes STDP selects between, the two lobes of the biphasic curve — not a rule for choosing between them. LTP and LTD can be induced by many protocols (high- or low-frequency tetanus, pharmacological manipulation) that have nothing to do with spike order. STDP is one specific induction rule that picks the direction by the signed interval Δt. Tell: is the potentiation or depression here selected by the millisecond order of pre/post spikes (STDP), or driven by stimulation frequency or chemistry with no spike-order dependence (LTP/LTD induced by other means)?
Coincidence detection. What the NMDA receptor performs at the molecular level — admitting calcium only when presynaptic glutamate binding and postsynaptic depolarisation overlap. But coincidence is symmetric in time, whereas STDP reads temporal order: the asymmetric curve credits precedence and debits succession, flipping sign across the crossover even for pairs a millisecond or two apart. The coincidence detector is the biophysical substrate; the order-asymmetry is what turns it into a causality detector. Tell: does the mechanism reverse its verdict when you swap which spike leads (STDP), or fire identically whenever the two events overlap regardless of which came first (pure coincidence detection)?
Rate-based plasticity (BCM / frequency-dependent LTP). A rival plasticity family in which the sign and size of the weight change depend on postsynaptic firing rate — often against a sliding threshold — rather than on millisecond spike order. Real synapses show both, and at high rates the rate rule can override timing (the entry's T3), which is exactly why "the STDP curve" cannot be applied uncritically. Tell: does the weight change track the signed pre/post interval Δt (STDP), or the average postsynaptic activity level crossing a threshold, with timing irrelevant (rate-based/BCM)?
Homeostatic synaptic scaling / metaplasticity. The slow, cell-wide, order-insensitive machinery that renormalizes total synaptic drive to keep a neuron near a firing setpoint — the stabilizer STDP requires but does not itself supply (the entry's T4: the order-credit rule runs away toward saturation without it). STDP is the fast, local, order-sensitive rule; scaling is the global counterweight. Tell: is the change a fast per-synapse credit/debit read off spike order (STDP), or a slow multiplicative rescaling that preserves relative weights to hold activity at a setpoint (synaptic scaling)?
Backpropagation. The dominant artificial-network learning rule, which assigns credit by a global error signal propagated backward through every layer from an output objective. That is the categorical opposite of STDP's strict locality — only the two endpoint spikes and the current weight enter the update, with no global controller (the entry's locality constraint, and the source of its T2 limitation on distal credit). Tell: does the update use only each synapse's own two spike times and weight (STDP), or a network-wide gradient computed from an output error and routed backward (backpropagation)?
The temporal-credit-assignment parent it instances (hebbian_learning plus temporal_dynamics × credit_assignment × causality). The substrate-general pattern — the earlier event credits the later one, the sign of the update respecting the arrow of time — that STDP instantiates and that recurs as genuine co-instances in temporal-difference learning, eligibility traces, and backward induction. Those siblings are not borrowings of STDP; they are independent solutions to the same problem, lacking the millisecond window, the biphasic sign curve, and the calcium biophysics. Tell: strip away the spike, the timing window, and the NMDA coincidence machinery — if what remains is bare "order-sensitive credit assignment" in a non-spiking learner, you are using the parent pattern (or a sibling like TD), not STDP. (Treated fully in Knowledge Transfer and Structural Core vs. Domain Accent.)

Neighborhood in Abstraction Space¶

Spike-Timing-Dependent Plasticity sits in a sparse region of the domain-specific corpus (60^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Neural Circuitry & Synaptic Plasticity (9 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-07-12