Not every network learns by a global backward pass. A spiking network can also learn locally — each synapse adjusting itself from the relative timing of the spikes it sees, with no error signal shipped back across the whole graph. That is what biology appears to do, and it is what makes online, on-chip learning possible. SC-NeuroCore carries a family of these rules; the most famous of them draws the timing window you can shape below.
The core rule is disarmingly simple. If a presynaptic spike arrives just before the postsynaptic neuron fires, it plausibly helped cause that firing, so the synapse is strengthened. If it arrives just after, it could not have contributed, so the synapse is weakened. The size of the change decays exponentially with the time gap $\Delta t = t_{\text{post}} - t_{\text{pre}}$ — a causal spike a millisecond early counts for far more than one twenty milliseconds early. Shape the window and watch the balance shift.
The curve is the weight change a synapse receives for each timing gap. The library realises it with a pair of exponentially-decaying traces — one left by each presynaptic spike, one by each postsynaptic spike — so no explicit $\Delta t$ is ever stored; the trace values at the moment of a spike reproduce this window on average. The net balance $\tfrac{A_{+}\tau_{+}}{A_{-}\tau_{-}}$ decides the rule's long-run bias: above $1$ it tends to strengthen, below $1$ to prune. Its default here uses the library's $\tau = 20\,\text{ms}$.
Pairwise STDP is the starting point, not the whole story. Local learning needs more than timing — it needs a way to carry credit across a delay, to be gated by reward, and to keep the network stable while it learns. SC-NeuroCore implements the standard rules for each, transcribed from their primary sources.
Timing alone cannot bridge a reward that arrives seconds after the spikes that earned it. An eligibility trace is a fading memory of each synapse's recent pre×post coincidence; when a later error or reward signal arrives, the weight change is that trace times the third factor. It is the local, online counterpart to backpropagation-through-time.
The same eligibility trace, gated by a global reward signal that stands in for a neuromodulator such as dopamine. Synapses that were recently active are the ones a reward reinforces — a biologically-grounded route to reinforcement learning without a differentiable objective.
Left alone, Hebbian rules run away — strong synapses get stronger without bound. Homeostasis is the brake: it multiplicatively rescales a neuron's incoming weights to hold its firing rate near a set-point, preserving relative strengths while keeping the whole network in a usable regime. It is what lets a network be deployed and left to adapt without supervision.
A faster, transient effect that does not change stored weights at all: each spike temporarily depletes a synapse's resources (depression, $\tau_d$) and raises its release probability (facilitation, $\tau_f$). The effective weight becomes history-dependent on a timescale of tens to hundreds of milliseconds — a built-in dynamic filter on incoming spike trains.
Also in the toolkit: MAML-style meta-learning (learning a good initialisation), structural plasticity (growing and pruning connections), and BPTT / truncated-BPTT for when a full backward pass is wanted — the surrogate-gradient route on its own page.
Local plasticity is the other half of how a spiking network learns. Surrogate gradients train a network offline, through a smooth stand-in derivative; these rules let it keep learning online, on-device, from its own activity — no stored dataset, no global backward pass, and an update that is cheap enough to run on the same hardware that runs inference. The two are complementary, and the toolkit carries both.
Evidence boundary: these plasticity modules are tier research (experimental). The demo plots the exact STDP window equation in your browser; the rules run over full networks through the toolkit's NumPy / PyTorch / Rust learning backends, not on this page. Rule forms and constants are transcribed from the cited primary sources; no learning-speed or accuracy figures are quoted here.