Semantic divergences

Semantic divergences#

When you port a NEST model to brainpy.state, the parameter names match — but a few semantics differ in ways the source code does not surface at a glance. This catalog states each divergence exactly. Every entry is backed by a live-NEST parity test, so the catalog consolidates what those tests proved (see Validation status for the harness and tolerance categories).

There are five kinds of divergence a port must account for.

1. Where state lives#

NEST sometimes keeps learning or trace state on the neuron; brainpy.state moves it onto the synapse (or a broadcast node) so a rule runs standalone — without requiring an archiving-capable postsynaptic neuron. The canonical case is the STDP post-synaptic trace time constant tau_minus: a neuron parameter in NEST, a synapse-spec attribute here.

Setting a parameter on the wrong object relative to NEST is the single most common porting mistake. The full treatment — with the porting rule and the parameter table — is on the STDP page:

2. Parameter-location maps#

Wherever NEST precomputes a weight change on an archiving neuron, brainpy.state relocates the governing parameters onto the synapse spec (or, for the dopamine modulator, onto the volume_transmitter). Parity always sets identical values on both sides. The exhaustive table — including which Clopath parameters move to the synapse and which voltage filters stay on the neuron — is in Parameter-location map for the STDP family.

3. Documented numerical bands#

A few rules do not reproduce NEST bit-for-bit. In each case the direction and ordering of the effect are exact; only a small magnitude band remains, and each band is asserted by a parity test. Pull the numbers from the STDP page, not from a fresh measurement:

Divergence

Band

Where

Clopath online LTP vs delayed ring buffer

≤ 5 %

Clopath delay_u_bars and online LTP (≤ 5 % band)

Dopamine online-vs-deferred integration

~0.2 %

Dopamine online-vs-deferred integration (~0.2 % band)

Nearest-neighbour “phantom pre at 0”

not reproduced (documented)

Nearest-neighbour “phantom pre at 0” (stdp_nn_symm / stdp_nn_restr only)

The online-vs-deferred origin of these bands — brainpy.state integrates every step while NEST defers the weight integral to the next pre spike — is explained in the STDP overview.

4. Device and recording conventions#

The Simulator device model differs from NEST’s in a few load-bearing, but easy-to-remember, ways (full detail in Devices):

  • Reversed analog tap. A voltmeter / multimeter is connected in NEST’s reversed direction — connect(voltmeter, neuron) — because the recorder observes the neuron rather than receiving events from it.

  • Current vs spike sources. Current devices (dc_/noise_/step_/ ac_generator) inject through the neuron’s current ring buffer (a NEST-faithful one-step delay); spike sources (poisson_generator, spike_generator) deliver delayed delta events.

  • In-memory recording. There is no file/ascii backend yet, so NEST’s record_to backend axis collapses to the in-memory equivalent; data is read back after the run with res.trace / res.spikes / res.rate / res.n_events / res.weight_trace.

  • Plastic edges need a spike source. In NEST a plastic synapse cannot be driven by a device, so a parrot_neuron relays the train; on the Simulator API a spike_generator can drive the plastic edge directly (with a one-step holder lag the parity tests align to).

5. PRNG-divergent drives → distributional parity#

NEST and JAX draw from independent PRNG streams, so any Poisson drive, random connectivity, or stochastic neuron diverges sample-by-sample even when the model is correct. These are validated distributionally — the seed-aggregated statistic (the mean over ≥ 5 seeds) is compared, never a per-sample trace. This is comparison category D; see Validation status.

See also#