Spike Encoding: Rate, Temporal & Population Coding (2D)

Input Signal

Input frequency 0.40 Hz

Neuron population size 8

Playback speed 1.0×

Live Readouts

True input v(t)

0.00

Sim time

0.0 s

Rate coding — v̂ 0.00 Temporal coding — v̂ 0.00 Population coding — v̂ 0.00

Rate error

Rate spikes/s

Temporal error

Temporal spikes/s

Population error

Population spikes/s

How it works

A neuromorphic system never passes a raw number between neurons — it passes discrete spikes. This simulator drives three classic encoding schemes from the same continuous signal v(t) ∈ [0,1] and renders each as a live 2D spike raster (time flows right→left, one row per neuron), so you can see the tradeoffs directly. Drag the raster to pan and scroll to zoom in on individual spikes.

Rate coding (population of N independent neurons):
  P(spike in Δt) = v · r_max · Δt      (Poisson-like)
  decode: v̂ = spikes(window) / (N · r_max · window)

Temporal / latency coding (one code window per cycle):
  latency_j = (1 − v₀) · W · 0.85 + jitter_j
  decode: v̂ = 1 − mean(latency) / (W · 0.85)

Population coding (tuning curves, preferred values p_j):
  r_j(v) = r_max · exp(−(v − p_j)² / (2σ²))
  decode: v̂ = Σ p_j·count_j / Σ count_j   (population vector)

Rate coding is simple and robust to noise but needs many spikes averaged over a window to be accurate — the slowest and least energy-efficient of the three.
Temporal coding needs as little as one spike per neuron per cycle: the exact timing carries the value, giving very low latency and very few spikes — the scheme real neuromorphic chips (Loihi, TrueNorth) exploit for energy efficiency.
Population coding spreads the value across a bank of broadly-tuned neurons; no single neuron needs to be precise, only the population's activity bump has to peak near the true value — this is how the brain encodes continuous variables like orientation or joint angle.

Switch the input waveform, frequency, and population size to see how each scheme's decode error and spike rate respond — temporal coding should stay accurate with far fewer spikes/s than rate coding.

drag to pan · scroll to zoom