Spike Encoding

Spike Encoding#

Convert numbers and time series into spike trains for spiking networks. The tutorial covers all ten encoders in braintools.spike: rate codes (RateEncoder, PoissonEncoder, BernoulliEncoder, PopulationEncoder), timing codes (LatencyEncoder, RankOrderEncoder, SpikeCountEncoder), signal codes (DeltaEncoder, StepCurrentEncoder), and symbol codes (TemporalEncoder). Each is checked against the rates, timings, or counts it should produce, and the tutorial ends with a guide to choosing one.

The encoders moved from the top level to braintools.spike in 0.4.0. The old names (braintools.RateEncoder and so on) still work but emit a BraintoolsDeprecationWarning, and will be removed in 1.0.0.

From rasters to spike times: SpikeTrains#

Encoders return dense rasters with shape (n_time, n_neurons). Most of those entries are zero. braintools.spike.SpikeTrains stores only the spike times, one padded row per neuron. Row n holds the counts[n] spike times of neuron n in ascending order, followed by +inf padding up to a fixed capacity.

Because the shapes are fixed, a SpikeTrains is a JAX pytree. You can build and use it inside jax.jit and jax.vmap. When dt carries a unit, the spike times carry it too:

import brainstate
import brainunit as u
import jax.numpy as jnp
import braintools
from braintools.spike import SpikeTrains

brainstate.environ.set(dt=0.1 * u.ms)
brainstate.random.seed(0)
raster = braintools.spike.PoissonEncoder()(jnp.array([20., 50., 100.]), n_time=10_000)  # 1 s

st = SpikeTrains.from_dense(raster)       # dt comes from brainstate.environ
st.counts                                 # spikes per neuron
st.bin(100 * u.ms)                        # (10, 3) spike counts in 100 ms bins
st.window(200 * u.ms, 500 * u.ms)         # spikes in [200 ms, 500 ms), same capacity
times, ids = st.to_events()               # flat, time-sorted events (host-only)
st.to_dense()                             # back to the boolean raster
braintools.visualize.spike_raster(st)     # the spike plots accept SpikeTrains directly

Two other constructors build a SpikeTrains from other formats:

  • SpikeTrains.from_events(times, neuron_ids, n_neurons) takes flat event arrays.

  • SpikeTrains.from_list([...]) takes one array per neuron, and the arrays may differ in length.

capacity is the padded row length. It is part of the array shapes, so it must be fixed when a function is traced:

  • With concrete data, it defaults to the largest spike count.

  • Inside jit or vmap, pass it explicitly, for example SpikeTrains.from_dense(raster, capacity=200).

  • If a traced neuron fires more often than capacity, its extra spikes are dropped and overflow is set. Call st.check() outside the transform to turn that flag into an error.