GitHub

V2 changelog

Changes of version 2, its known limitations, and the bugs fixed during version 1.

2.0.1: SpikingNeuralNetworks 2.0.1, SNNModels 2.0.2

  • Perturbation records are stored compactly (SNNModels 2.0.2): one PerturbationStore / PerturbationSpikeStore per (object, variable, condition) instead of one Dict per window, with perturbation_record output unchanged. Records written by earlier versions must be converted once with compact_perturbation_records! (then saved again); readers and new recordings refuse the old layout. New accessor perturbation_windows. Window numbers are never reused after a deletion.
  • Re-exported: perturbation_windows, compact_perturbation_records!, PerturbationStore, PerturbationSpikeStore. Compat SNNModels = "2.0.2".
  • Documentation: performance comparison page (v1, v2, Auryn, Brian2); monitor_rate written with units; COBA tutorial with the Brette et al. 2007 weights.

2.0.0: SNNModels 2.0.0, SpikingNeuralNetworks 2.0.0, SNNUtils 1.0.0, SNNPlots 1.0.0, CalciumSpike 1.0.0

All packages move to the next major version together. SNNModels 2 changes the types of synapses, stimuli and populations (concretely typed plasticity rules, spike lists, lazy synapse arrays); saved models of 1.9.1 and earlier are refused at load time. Full release summaries: CHANGELOG.md of SNNModels, SNNUtils and SNNPlots; every change with its migration: CHANGELOG_2.0.0.md of SNNModels.

  • Faster: E/I balanced network 8 to 12 times faster than 1.9.1, synapse memory 29 to 9 bytes per synapse, event-driven vSTDP, ring-buffer delays, one spike list per population, vectorised Tripod/BallAndStick, flush-to-zero of subnormals, no allocation per step.
  • New: Float64 by population type, with_plasticity/without_plasticity/change_plasticity, opt-in threading (neurons_per_thread).
  • Breaking: synapse/stimulus/population struct fields, iSTDPTime removed, delays rounded up to whole steps, random streams changed, vSTDPParameter is event-driven, Tripod/BallAndStick use param.physiology.
  • Removed code that was not loaded: Multipod, IF_CANAHP, AdExMultiTimescale, VariableInputParameter, stale SNNUtils parameter sets, SNNPlots Plots.jl backend and their examples.

Upgrading: see What changes from v1 to v2.

Known limitations

  • Float32 only: HetRec (refuses a Float64 step), CurrentStimulus, BalancedStimulus and Turnover (refuse a Float64 target), and the rate and FORCE connections.
  • The clock accumulates rounding at a step that is not exact in Float32 (e.g. dt = 0.1ms): get_time drifts, and a SpikeTimeStimulus spike can be delivered one step off.
  • Delays are counted in steps: after a change of dt mid-run, spikes in flight are dropped (if the delay in steps changes) or keep their old step count.
  • A record closed by a plasticity switch is not reopened by a later switch (clear_records! first).
  • Some components still do per-step dense work: iSTDPPotential, vSTDPDenseParameter, MarkramSTPParameterTimestep, ActivityDependentTurnover, AggregateScaling.
  • Tripod/BallAndStick reproduce 1.9 bit for bit only on x86-64 with FMA.
  • Files saved with version 1 cannot be loaded (load_model throws an ArgumentError).

Bugs fixed in version 1

SNNModels 1.9.0, SNNPlots 0.2.11, SNNUtils 0.2.10, SpikingNeuralNetworks 1.3.0

Bug fixes in SNNModels, SNNPlots, SNNUtils and SpikingNeuralNetworks found while documenting every public symbol. Several of them change simulation results.

Breaking changes

These are released as a minor version (1.9.0), but code that ran with 1.8.x can fail or give different results. Check this list before upgrading.

ChangeCode affectedMigration
AdExParameter/IFParameter: one membrane value alone (C, gl, R or τm) is an ArgumentErrorIFParameter(τm = 20ms), AdExParameter(C = 200pF)give a pair; to keep the 1.8 neuron: IFParameter(τm = x, R = 0.06), AdExParameter(C = x, gl = 40nS)
More than two membrane values must agree within 0.1 % (τm = C/gl, R = 1nS/gl)e.g. IFParameter(C = 281, gl = 40, τm = 20, R = 0.1)keep the pair the model integrates with: τm, R for IF/AdEx, C, gl for Tripod/BallAndStick
Changing one membrane value updates the others (@update!, p.τm = x, make_heterogeneous)@update! cfg.adex.τm = x on a Tripod or BallAndStick (no effect before, now C = τm gl); p.C = x on an AdExParameter (now also changes τm)intended in most cases; to change two values at once use with_membrane(p; τm = …, C = …)
A Tripod or BallAndStick built with AdExParameter(gl = …, τm = …) now uses C = τm gle.g. gl = 40nS, τm = 20ms: C 281 → 800 pFgive C, gl explicitly to keep the 1.8 neuron
Tsodyks-Markram STP update order (Mongillo 2008)every model with MarkramSTPParameter*none in code; parameters tuned with ≤ 1.8.4 are not equivalent (see below)
Tripod/BallAndStick somatic equation, Heun step and refractorinessevery Tripod/BallAndStick modelnone in code; results change (see below)
Exported names that were never defined are no longer exportedSNNModels: HUMAN, MOUSE, MultiRecetorSynapse, NMDA_CANAHP, PSParam, BSParam, SpikeTime, SpikingSynapseDelay, Synapse_CANAHP, autocorrelogram, filter_populations, gcamp6_kernel, get_path, get_synapse_symbols, isi_cv, no_PlasticityVariables, no_STDPParameter, record_plast!, synapsearray; SNNPlots: default_colors, nature_figure, plot_model, plot_stimulus, plot_connectionsthese names could not be used before either; remove them from using … : lists
SNNPlots stdp_test/stdp_kernel measure a single pre/post pairkernels computed with SNNPlots ≤ 0.2.10recompute; the old kernel included an extra causal pairing
SNNUtils quaresima2022_nar returns Tripod keyword arguments (it used the removed AdExSoma)code that called it (it threw before)pass the result to Tripod(; …)
Behaviour change: Tripod and BallAndStick
  • The somatic exponential term is now $g_L \Delta_T e^{(V-\theta)/\Delta_T}$, as in the AdEx model and in the published Tripod model (Quaresima et al. 2023, Eq. 1, and its code); it lacked $g_L$ (40 times smaller at gl = 40nS). Rheobase drops by about 29 % (Tripod with default parameters: 1134 -> 814 pA; paper parameters: 1097 -> 782 pA), the rate at 1500 pA rises from 24 to 44 Hz (19 to 32 Hz), the first ISI shortens from 18.7 to 10.9 ms, the adaptation current reaches about 1.5 times higher values; subthreshold dendritic EPSPs and NMDA plateaus are unchanged (< 0.1 mV). BallAndStick changes in the same way.
  • The Heun stage is now consistent (the adaptation stage lacked dt, the synaptic currents were not evaluated at the predicted state): spike times move by up to 7 ms over 500 ms and the scheme converges with dt.
  • The refractory counters keep at least one step per period: with up = τabs = 0.1ms at the default dt = 0.125ms the neuron used to fire at about 1 kHz.
  • In a 400-cell Tripod E/I network the E rate doubles (12.6 -> 24.8 Hz) and the interneuron rates double as well.
  • New: Vspike (spike detection threshold, default -10 mV), dendritic El used, Is/Id of BallAndStick applied.
Behaviour change: Tsodyks-Markram short-term plasticity

MarkramSTPParameter (and the Het and Timestep variants) now follow Mongillo, Barak & Tsodyks (2008): at a presynaptic spike $u$ jumps first and the same $u^+$ sets both the transmitted efficacy $u^+ x$ and the depletion of $x$. Before, the efficacy used $u$ before the jump while the depletion used $u$ after it, which depressed more than either published model. Efficacies are higher (first spike from rest $U(2-U)$ instead of $U$), and for some parameters the character of the synapse changes: with $U = 0.3$, $τ_F = 500$ ms, $τ_D = 70$ ms a 20 Hz train facilitated before (0.30 -> 0.44) and depresses now (0.51 -> 0.47).

Behaviour change: membrane parameters of AdExParameter and IFParameter (pair rule)

C, gl, R and τm are now always consistent ($τ_m = C/g_L$, $R = 1\,\mathrm{nS}/g_L$). AdEx and IF integrate with τm and R; Tripod and BallAndStick (through their adex::AdExParameter) with C and gl. Before, these were four independent fields after construction, so the same parameter object could describe two different membranes.

  • Construction takes a pair of independent values, any of (C, gl), (C, R), (C, τm), (gl, τm), (R, τm), and derives the rest, or none and uses the default pair (AdEx C = 281pF, gl = 40nS; IF τm = 15ms, R = 0.06). A single value is an ArgumentError (before, it was combined with the defaults: IFParameter(τm = 20ms) used R = 0.06, AdExParameter(τm = 20ms) kept C = 281pF for Tripod). More than two values must agree within 0.1 %.
  • Changing one value (@update!, p.τm = x on the mutable AdExParameter, with_membrane, make_heterogeneous) follows a fixed rule: τm keeps gl and changes C; C keeps gl and changes τm; gl or R keeps C and changes τm. Two values given together define a new pair (with_membrane(p; τm = 20ms, C = 281pF)).
  • Consequences: @update! cfg.adex.τm = x on a Tripod or BallAndStick used to have no effect at run time (only C and gl are read); it now sets $C = τ_m g_L$. A Tripod built with AdExParameter(gl = 40nS, τm = 20ms) used C = 281pF and now uses C = 800pF. A parameter set such as C = 281, gl = 40, τm = 20, R = 0.1 (two different membranes) is now rejected and must be fixed by choosing the intended pair.
  • New: with_membrane, resolve_membrane, membrane_update; @snn_kw structs can specialise snn_kw_finalize to derive dependent fields.
Other behaviour changes
  • HH and MorrisLecar flag one spike per action potential (they flagged about 60 per action potential at dt = 0.01ms).
  • Rate/WilsonCowan reset their input g every step (it accumulated all past inputs).
  • Confavreux2025Synapse: a spike of weight w increments the conductance by w (it was w dt).
  • AggregateScaling: time constants in ms, rate units, correct summed weight; HetRec baseline with dt; AdditiveNorm restores the summed weight.
  • vSTDP and iSTDPPotential voltage traces start at the membrane potential (no initial spurious depression); vSTDP weights are unchanged otherwise.
  • Recording: exact time axis of record/interpolated_record, sampling period rounded to the nearest step.
  • SNNPlots stdp_test/stdp_kernel measure a single pre/post pair (the kernel was shifted by an extra causal pairing).
  • BalancedStimulus, FORCE/PINning connections, train! with IZ/HH/MorrisLecar or without connections, connections onto MorrisLecar/ExtendedIF/WilsonCowan, several analysis, IO and SNNUtils functions now work; 35 exported-but-undefined names were removed.

Earlier releases

Two bugs of earlier releases change results:

  • iSTDPRate (SNNModels 1.5.0 to 1.8.1, SpikingNeuralNetworks from v1.0.0): potentiation at a postsynaptic spike was applied to the wrong synapses (a @turbo loop ignored a reassigned loop variable). Rerun simulations that used iSTDPRate with these versions.
  • STDPGerstner (before SNNModels 1.8.2): amplitudes were applied twice (effective $A^2$, sign lost). Old amplitude A corresponds to ±A^2 now.

Notes of SpikingNeuralNetworks 1.2.1, SNNModels 1.8.2 to 1.8.4 and SNNPlots 0.2.10: release notes of SpikingNeuralNetworks 1.3.0.