brainspikes 0.2.0

brainspike is the point-neuron spiking-network layer of the BrainX ecosystem: building, simulating, and training spiking neural networks of point neurons. A new kernel language unifies how models are built, how they are simulated, and how they are trained. It is built on JAX and brainstate.

brainspike provides:

• NEST-compatible models — JAX re-implementations of NEST simulator neurons, synapses, plasticity (STDP, STP), and devices that preserve NEST's parameter names and unit conventions, numerically validated against a live NEST install within documented tolerance bands, driven through a NEST-style Simulator (create / connect / simulate).
• A unified kernel (in development, still private) — one Network graph (populations, projections, devices, observers) over models with written contracts (declared ports, states, parameters, and spike rules), so the graph a simulation runs is the graph a learner trains (architecture spec, roadmap).

All parameters carry physical units via brainunit, catching unit errors at construction time.

Key features

• NEST parameter parity — port models from NEST with minimal friction; every model carries live-NEST parity evidence within documented tolerance bands.
• Physical units everywhere — parameters use brainunit quantities (mV, ms, nS, …), preventing unit errors at construction time.
• One graph for simulation and learning — the new kernel compiles the same Network for simulation and for gradient-based or online learning.
• Hardware-accelerated — JAX backend with JIT compilation for CPU, GPU, and TPU.

Scope

brainspike is for computational neuroscientists who want NEST models on GPU/TPU or want to train them, and for brain-inspired-computing researchers who need biologically constrained spiking networks.

• Model scope: point neurons only (no multi-compartment or population/mean-field models) and no single-unit rate neurons.
• Out of scope: detailed multi-compartment morphology, MPI cluster-scale simulation (single-host multi-device is in scope), MUSIC/SONATA co-simulation (SONATA import is in scope), and e-prop (see braintrace).

Kernel redesign in progress. The first kernel milestones have passed their go/no-go gates; the new kernel is still private. Meanwhile the current NEST-style Simulator API shown below is frozen (it keeps working; no new features or fixes) and will become a thin layer over the kernel's Network. CPU performance of the new kernel has been measured on one benchmark; GPU performance has not been measured yet.

See Positioning and scope.

Installation

brainspike requires Python ≥ 3.11 and runs on Linux, macOS, and Windows.

pip install brainspikes -U

The PyPI distribution is named brainspikes; the package is imported as import brainspike.

For hardware-specific JAX backends:

pip install brainspikes[cpu] -U     # CPU only
pip install brainspikes[cuda12] -U  # CUDA 12.x
pip install brainspikes[cuda13] -U  # CUDA 13.x
pip install brainspikes[tpu] -U     # TPU

Or install the full BrainX ecosystem:

pip install BrainX -U

Quickstart

A single iaf_psc_alpha driven by a constant current I_e and observed by a voltmeter (connected in the reversed direction, as in NEST):

import jax
import brainstate
import brainunit as u
import brainspike

jax.config.update("jax_enable_x64", True)   # NEST-parity runs use float64
brainstate.environ.set(precision=64)

sim = brainspike.Simulator(dt=0.1 * u.ms)
neuron = sim.create(brainspike.iaf_psc_alpha, 1, I_e=376.0 * u.pA)
vm = sim.create(brainspike.voltmeter)
sim.connect(vm, neuron)              # reversed: the voltmeter observes the neuron

res = sim.simulate(1000.0 * u.ms)
v_m = res.trace(vm, "V_m")           # membrane potential trace (with units)

See examples/nest_like/one_neuron.py for the full version with plotting, and the NEST-compatible hub for the tutorials, model directory, and validation evidence.

Documentation

Full documentation: https://brainx.chaobrain.com/brainspike/

• Get started: https://brainx.chaobrain.com/brainspike/get-started/index.html
• NEST-compatible hub (tutorials, models, validation): https://brainx.chaobrain.com/brainspike/nest-style/index.html
• API reference: https://brainx.chaobrain.com/brainspike/apis/index.html
• Examples: https://brainx.chaobrain.com/brainspike/examples/nest-gallery.html

Examples

Runnable NEST-compatible ports live in examples/nest_like/ and are catalogued in the NEST gallery. Highlights: Brunel random balanced networks, gap-junction networks, astrocyte networks, and STDP / STP plasticity demos.

Contributing

Contributions are welcome. Please read CONTRIBUTING.md for the development setup, coding conventions, testing guidance, and documentation workflow (including how to document new APIs and how to record live-NEST validation/parity evidence for NEST-compatible models).

Ecosystem

brainspike is one part of the BrainX ecosystem:

Citing

If you use brainspike, please consider citing the following:

@article {10.7554/eLife.86365,
    article_type = {journal},
    title = {BrainPy, a flexible, integrative, efficient, and extensible framework for general-purpose brain dynamics programming},
    author = {Wang, Chaoming and Zhang, Tianqiu and Chen, Xiaoyu and He, Sichao and Li, Shangyang and Wu, Si},
    editor = {Stimberg, Marcel},
    volume = 12,
    year = 2023,
    month = {dec},
    pub_date = {2023-12-22},
    pages = {e86365},
    citation = {eLife 2023;12:e86365},
    doi = {10.7554/eLife.86365},
    url = {https://doi.org/10.7554/eLife.86365},
    journal = {eLife},
    issn = {2050-084X},
    publisher = {eLife Sciences Publications, Ltd},
}

@inproceedings{wang2024a,
    title={A differentiable brain simulator bridging brain simulation and brain-inspired computing},
    author={Chaoming Wang and Tianqiu Zhang and Sichao He and Hongyaoxing Gu and Shangyang Li and Si Wu},
    booktitle={The Twelfth International Conference on Learning Representations},
    year={2024},
    url={https://openreview.net/forum?id=AU2gS9ut61}
}

License

brainspike is released under the Apache License 2.0. See LICENSE for the full text.