Spike-Timing-Dependent Plasticity

Presynaptic Input Groups

Group A — 6 correlated (co-fire together)

Group B — 6 independent background

Group A volley rate 12 Hz

Group B background rate 8 Hz

STDP learning rate A₊ 0.060

Live Readouts

Postsynaptic rate

0.0 Hz

Membrane V / V_th

Avg weight — Group A

0.220

Avg weight — Group B

0.220

Simulated time

0.0 s

How it works

Each of the 12 rim neurons is a synapse onto one leaky integrate-and-fire (LIF) postsynaptic neuron. Every synapse keeps an exponentially-decaying eligibility trace of its own recent spikes, and the postsynaptic neuron keeps one trace of its own recent spikes — the same trace-based rule real neuromorphic chips (Intel Loihi's on-chip learning engine, SpiNNaker) implement in silicon:

dx_pre/dt  = -x_pre/τ₊     (jumps +1 on a presynaptic spike)
dx_post/dt = -x_post/τ₋    (jumps +1 on a postsynaptic spike)

on presynaptic spike i:   w_i ← w_i - A₋·x_post(t)     (depression)
on postsynaptic spike:    w_i ← w_i + A₊·x_pre,i(t)  ∀i (potentiation)

membrane:  dV/dt = -V/τ_m + Σ_i w_i·δ(t-t_pre,i),  fire & reset when V ≥ V_th

Group A fires as one synchronized volley, so its 6 synapses reliably arrive just before the postsynaptic spike they help trigger — Δt = t_post − t_pre > 0 on almost every event, so A₊ dominates and their weights climb toward the ceiling.
Group B fires independently of the postsynaptic neuron, so its spikes land at random phase relative to a spike — roughly half the time just after one, catching the elevated x_post trace and getting depressed by A₋. Net effect: weights drift down.
This is the classic STDP result — a neuron with no supervision learns to respond selectively to whichever inputs are temporally correlated with its own output, which is exactly the on-chip, unsupervised learning mechanism neuromorphic hardware exploits instead of backpropagation.

Group A (correlated) Group B (background) Postsynaptic neuron

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.