Changelog

Changelog#

All notable changes to this project are documented in this file.

The format follows Keep a Changelog conventions, and this project adheres to Semantic Versioning.

0.1.0 – 2026-06-17#

Highlights#

Version 0.1.0 is the first release in which brainpy.state is the single public entry point for the package, and it consolidates both model families into a coherent, documented whole. It introduces two complementary ways to build networks — a high-level Network / Builder API and an explicit, NEST-vocabulary Simulator — together with first-class synaptic delays, new analog input generators and reduced neuron models, fully JIT-safe NEST-compatible models, and a live-NEST validation harness that reproduces published Brunel benchmarks to sub-percent accuracy. It also corrects several numerical bugs in the BrainPy-style models.

Added#

Network construction#

  • Network and Builder — a high-level API for assembling models into a runnable network. Network is a brainstate.nn.Module with a projection-first update traversal and a JIT-compiled simulate(duration, monitor=...); Builder produces the same module graph through an imperative add() / connect() interface. Includes rule-based projections (OneToOneProj, AllToAllProj, PairwiseBernoulliProj, SymmetricPairwiseBernoulliProj, FixedIndegreeProj, FixedOutdegreeProj, FixedTotalNumberProj, PairwisePoissonProj), a Recorder helper, and weight/parameter distributions (dist.Normal, dist.LogNormal, dist.Uniform).

  • Simulator — an explicit network builder that mirrors NEST’s vocabulary: create(model, size, params=...), connect(pre, post, *, rule, weight, delay, ...), and simulate(duration) returning a SimulationResult with rate(), n_events(), and spikes(). There is no global kernel; populations, devices, and projections form a flat module graph compiled into a single rollout.

    • Population algebra — NodeView supports concatenation (ne + ni) and slicing (ne[:n]), so one connect can target a combined or partial population.

    • Connection rules — all_to_all, one_to_one, fixed_indegree, pairwise_bernoulli, fixed_total_number, and more.

    • Sparse or dense connectivity — connect(..., comm='sparse') routes connectivity through a memory-light sparse event matmul (~1.9 GB at NEST’s native order=2500); comm='dense' remains the default for small networks. Both paths share one sampler and produce identical results.

    • Multi-receptor routing — connect(..., receptor_type=...) delivers input to a specific receptor port of a multi-receptor neuron.

Synaptic delays (BrainPy-style)#

  • AlignPostProj and CurrentProj accept an optional, unit-carrying delay= keyword: a scalar applies a homogeneous conduction delay; a (N_pre,) array applies a per-presynaptic-neuron axonal delay. Sub-dt delays are linearly interpolated. The default delay=None is unchanged and adds no overhead.

Analog current generators (BrainPy-style)#

  • SectionInput, ConstantInput, StepInput, RampInput, SinusoidalInput, WienerProcessInput, and OUProcessInput — stateful current sources that emit the instantaneous current at the network’s current time, staying synchronized with the simulation clock and composing under jit without materializing long offline arrays. Scalar parameters broadcast to in_size; the stochastic generators draw independent per-element samples.

Reduced neuron models (BrainPy-style)#

  • FitzHughNagumo (2-D excitable) and HindmarshRose (3-D bursting) reduced models, plus CubaLIF / CobaLIF convenience neurons that bundle an exponential synaptic term (current- and conductance-based, respectively).

Validation against live NEST#

  • A parity harness reproduces published benchmarks in both brainpy.state and a live NEST 3.x run, comparing population statistics (the random-number streams differ between simulators, so per-neuron matching is not meaningful). The tests skip automatically when NEST is not installed. Representative Brunel figures:

    Benchmark (vs live NEST)

    brainpy.state vs NEST

    rel. error

    Brunel alpha (order=200)

    56.9 vs 57.0 spks/s

    0.21 %

    Brunel alpha (order=2500)

    28.8 vs 28.5 spks/s

    0.91 %

    Brunel delta (order=200)

    58.5 vs 58.2 spks/s

    0.55 %

    Brunel exp-multisynapse (order=200, 4 seeds)

    25.8 vs 24.8 spks/s

    ≈ 4 %

    Brunel siegert (mean-field)

    32.03 vs 32.03 spks/s

    0.00 %

    Brunel + evolution strategies

    51.5 vs 51.5 spks/s

    0.08 %

Examples#

  • An expanded example library covering both families: BrainPy-style E/I networks and surrogate-gradient training, and faithful NEST ports including the Brunel variants (alpha, delta, exp-multisynapse, siegert, evolution strategies), single-neuron and plasticity demos, recording/device tours, spiking and rate-based networks, spatially-structured networks, and pedagogical models (AdEx figures, multi-compartment neurons, a reinforcement-learning Pong, and a Sudoku constraint solver). The Brunel flagship lives at examples/brainpy_like/brunel.py.

Changed#

  • Public API is brainpy.state only. The package is now used exclusively through the brainpy.state namespace; a direct import brainpy_state from user code raises ImportError pointing to the supported path (an escape hatch remains for tooling). brainpy.state.__version__ and brainpy.state.__version_info__ are now exposed, and all examples, docs, and docstrings use the public namespace.

  • NEST-compatible models are JIT-safe. Parameter validation is now tracer-aware (skipped during jit / vmap / grad tracing), and erfc_neuron / ginzburg_neuron no longer break under jax.jit. All 57 NEST neuron models are verified to trace under jit.

  • Supported Python is now ≥ 3.11, and the declared dependency floors were aligned across the packaging metadata (notably brainstate >= 0.5.0).

Fixed#

  • BrainPy-style model correctness:

    • Hodgkin-Huxley family (HH, MorrisLecar, WangBuzsakiHH). These models do not reset after a spike, so the previous per-step threshold test counted a single action potential as dozens of spikes (~55× over-count at dt = 0.01 ms). Spikes are now emitted once per upward threshold crossing, preserving a differentiable surrogate gradient.

    • Short-term plasticity (STP, STD). Transmitted output is computed from the resources available before depletion, so the first spike releases the correct amount.

    • WangBuzsakiHH. Corrected the K⁺ activation rate αₙ, which was 10× too large.

    • SpikeTime. Out-of-bounds time indices no longer read invalid rows; steps outside the defined window return zeros.

    • ExpIF / AdExIF (and refractory variants). The exponential spike-initiation term is clamped against overflow, preventing inf / NaN above the slope factor.

    • Poisson inputs (PoissonSpike, PoissonEncoder, poisson_input). Use the exact per-bin spike probability 1 − exp(−rate·dt), eliminating negative, fractional, or NaN counts at large rate·dt.

    • LeakyRateReadout. tau is sized to out_size, so a per-unit tau works when in_size != out_size.

    • Gap junctions (SymmetryGapJunction, AsymmetryGapJunction). Current inputs use a stable per-instance key, so repeated updates no longer leak input slots or accumulate stale currents.

    • Neuron base class. spk_reset is validated at construction (must be 'soft' or 'hard').

  • Network simulation. Projections and generators fanned out to several targets now derive a distinct, reproducible seed per target; previously a shared base seed gave every target identical connectivity and external drive.

0.0.4 – 2025-02-21#

Highlights#

Version 0.0.4 is a major feature release that introduces a comprehensive library of NEST-compatible neural models (initial version), reorganizes the public API into dedicated _brainpy and _nest submodules, and transitions the project to the Apache 2.0 license.

Added#

NEST-Compatible Model Library#

A complete port of the NEST simulator model catalogue, covering more than 250 models across all major categories:

Neuron Models

  • Integrate-and-Fire (IAF): iaf_psc_alpha, iaf_psc_alpha_multisynapse, iaf_psc_alpha_ps, iaf_psc_delta, iaf_psc_delta_ps, iaf_psc_exp, iaf_psc_exp_htum, iaf_psc_exp_multisynapse, iaf_psc_exp_ps, iaf_psc_exp_ps_lossless, iaf_cond_alpha, iaf_cond_alpha_mc, iaf_cond_beta, iaf_cond_exp, iaf_cond_exp_sfa_rr, iaf_bw_2001, iaf_bw_2001_exact, iaf_chs_2007, iaf_chxk_2008, iaf_tum_2000

  • Adaptive Exponential IF (AdEx / aeif): aeif_cond_alpha, aeif_cond_alpha_astro, aeif_cond_alpha_multisynapse, aeif_cond_beta_multisynapse, aeif_cond_exp, aeif_psc_alpha, aeif_psc_delta, aeif_psc_delta_clopath, aeif_psc_exp

  • Generalized IF (GIF): gif_cond_exp, gif_cond_exp_multisynapse, gif_pop_psc_exp, gif_psc_exp, gif_psc_exp_multisynapse

  • Multi-timescale Adaptive Threshold (MAT): amat2_psc_exp, mat2_psc_exp

  • Generalized LIF (GLIF): glif_cond, glif_psc, glif_psc_double_alpha

  • Hodgkin-Huxley family: hh_cond_beta_gap_traub, hh_cond_exp_traub, hh_psc_alpha, hh_psc_alpha_clopath, hh_psc_alpha_gap, ht_neuron

  • Izhikevich: izhikevich

  • Point-process neurons: pp_cond_exp_mc_urbanczik, pp_psc_delta

  • Binary neurons: erfc_neuron, ginzburg_neuron, mcculloch_pitts_neuron

  • Rate neurons: gauss_rate_ipn, lin_rate_ipn, lin_rate_opn, rate_neuron_ipn, rate_neuron_opn, rate_transformer_node, siegert_neuron, sigmoid_rate_ipn, sigmoid_rate_gg_1998_ipn, tanh_rate_ipn, tanh_rate_opn, threshold_lin_rate_ipn, threshold_lin_rate_opn

  • Miscellaneous: ignore_and_fire

Synapse Models

  • Static: static_synapse, static_synapse_hom_w, cont_delay_synapse, bernoulli_synapse

  • Short-term plasticity: tsodyks_synapse, tsodyks_synapse_hom, tsodyks2_synapse, quantal_stp_synapse

  • STDP: stdp_synapse, stdp_synapse_hom, stdp_dopamine_synapse, stdp_facetshw_synapse_hom, stdp_nn_pre_centered_synapse, stdp_nn_restr_synapse, stdp_nn_symm_synapse, stdp_pl_synapse_hom, stdp_triplet_synapse

  • Voltage-based / specialized: clopath_synapse, ht_synapse, jonke_synapse, urbanczik_synapse, vogels_sprekeler_synapse

  • Structural connections: diffusion_connection, gap_junction, rate_connection_delayed, rate_connection_instantaneous, sic_connection

Stimulation Devices

  • Current generators: ac_generator, dc_generator, noise_generator, step_current_generator, step_rate_generator

  • Spike generators: spike_generator, spike_train_injector, spike_dilutor

  • Poisson generators: poisson_generator, poisson_generator_ps, inhomogeneous_poisson_generator, sinusoidal_poisson_generator

  • Other generators: gamma_sup_generator, mip_generator, ppd_sup_generator, pulsepacket_generator, sinusoidal_gamma_generator

Recording Devices

  • correlation_detector, correlomatrix_detector, correlospinmatrix_detector, multimeter, spike_recorder, spin_detector, volume_transmitter, weight_recorder

Specialised Models

  • astrocyte_lr_1994: Leaky integrator astrocyte model

  • cm_default: Multi-compartment neuron model

NEST Base Infrastructure#

  • NESTNeuron, NESTSynapse, NESTDevice: abstract base classes for all NEST-compatible models, providing shared parameter management and state initialisation utilities (_nest/_base.py, _nest/_utils.py)

BrainPy-style Model Enhancements#

  • SpikeTime: added weight parameter and time-rounding option

  • AlignPostProj, DeltaProj, CurrentProj: new projection variants

  • align_pre_projection, align_post_projection: projection utility functions

  • SymmetryGapJunction, AsymmetryGapJunction: gap junction projection types

  • PoissonEncoder, PoissonInput, poisson_input: additional input generators

  • LeakyRateReadout, LeakySpikeReadout: renamed and expanded readout classes

Changed#

  • API layout: public models reorganised into brainpy_state._brainpy (BrainPy-style models) and brainpy_state._nest (NEST-compatible models) subpackages; all symbols remain importable from the top-level namespace

  • License: changed from GNU GPLv3 to Apache License 2.0

  • Data types: model state variables now use brainstate.environ.dftype() for consistent default floating-point precision across the ecosystem

  • Dependency: minimum brainpy requirement raised to >= 2.7.6

  • NEST models: all NEST models refactored onto shared base classes and utility helpers, eliminating duplicated boilerplate across model files

  • aeif_cond_alpha: streamlined initialisation of integration_step and I_stim fields

  • Documentation: mathematical equations and parameter descriptions expanded and standardised across all BrainPy-style and NEST-compatible model files

Fixed#

  • Documentation URLs updated in CONTRIBUTING.md, config.yml, and pyproject.toml to point to the correct hosted locations

  • Ecosystem cross-references in README.md and index.rst corrected

0.0.3 – 2025-01-01#

Highlights#

Version 0.0.3 consolidates the package rename from brainpy.state to brainpy_state, adds brainpy as a declared runtime dependency, and tightens internal state initialisation.

Added#

  • brainpy added as an explicit runtime dependency in requirements.txt and pyproject.toml

Changed#

  • Package renamed from brainpy.state to brainpy_state; all public import paths updated accordingly

  • HiddenState initialisation refactored for correctness and clarity

  • Function names updated and a simulation example added to the main script

  • Directory structure reorganised in preparation for the _brainpy / _nest split introduced in 0.0.4

  • Minimum brainpy requirement formalised

Fixed#

  • Import statements that still referenced the old brainpy.state namespace corrected throughout the codebase

0.0.1 – 2024-12-01#

Initial release of brainpy_state.

brainpy_state modernises the BrainPy spiking neural network simulator by adopting the state-based programming model introduced in brainstate.

Added#

Neuron Models#

  • Integrate-and-Fire (LIF) family

    • IF: basic integrate-and-fire neuron

    • LIF, LIFRef: leaky integrate-and-fire (with optional refractory period)

    • ExpIF, ExpIFRef: exponential integrate-and-fire

    • AdExIF, AdExIFRef: adaptive exponential integrate-and-fire

    • ALIF: adaptive leaky integrate-and-fire

    • QuaIF: quadratic integrate-and-fire

    • AdQuaIF, AdQuaIFRef: adaptive quadratic integrate-and-fire

    • Gif, GifRef: generalized integrate-and-fire

  • Hodgkin-Huxley family

    • HH: classic Hodgkin-Huxley conductance-based neuron

    • MorrisLecar: Morris-Lecar neuron

    • WangBuzsakiHH: Wang-Buzsaki modified Hodgkin-Huxley neuron

  • Izhikevich family

    • Izhikevich, IzhikevichRef: Izhikevich neuron (with optional refractory period)

Synapse Models#

  • Exponential synapses: Expon (single exponential decay), DualExpon (dual exponential rise-and-decay)

  • Receptor-based synapses: Alpha, AMPA, GABAa, BioNMDA

  • Short-term plasticity: STP (facilitation and depression), STD (pure depression)

Infrastructure#

  • Neuron, Synapse, Dynamics: abstract base classes for custom model development

  • Projection, AlignPostProj: network projection utilities

  • COBA, CUBA, MgBlock: synaptic output current handlers

  • SpikeTime, PoissonSpike: spike-train input generators

  • Readout, LeakyReadout, WeightedReadout: readout layer implementations

  • Runtime compatibility check: raises an informative error when an incompatible brainpy version (< 2.7.4) is detected

Dependencies#

Package

Minimum version

Python

3.10

jax

latest

brainstate

0.2.0

brainunit

latest

brainevent

0.0.4

braintools

0.0.9

numpy

1.15