Hebb

Pure-Rust spiking-neural-network substrate — embed-anywhere, with Python bindings.

hebb is the simulator substrate underneath the broader Hebb project: a fast, no-I/O-by-default Rust crate that implements the spiking-network primitives — neuron models, synapse models, plasticity, deterministic seed generators, and an optional on-disk format. The desktop app and visualizer consume this crate; you can also embed it directly in your own Rust or Python work.

What is this?

The hebb crate is two things at once:

1. A spiking-neural-network simulator you can drop into existing code. A SimEngine state machine you tick forward in time, with built-in neuron models (LIF, Izhikevich, AdEx, Hodgkin–Huxley) and plastic synapses (STDP, dopamine-gated R-STDP). No filesystem, no async runtime, no database — it's pure compute. If you do computational neuroscience, neuromorphic, or SNN research, this is a scriptable spiking substrate you can use today.

2. The portable core of the larger Hebb cortex-mechanism project. The same crate is the seed of an event-driven, continually-learning AI substrate where neural networks are modules, not the core — see the umbrella project README.

Don't know what a "cortex" means in this context? Read it as a brain-inspired memory + compute substrate. The simplest useful form is a graph of neurons that learn from a stream of events — which is exactly what SimEngine gives you.

Who is this for?

You are…	Use hebb as…	Start here
A computational-neuroscience / SNN / neuromorphic researcher	A fast, scriptable spiking-network simulator (Rust crate or import hebb_py from Python)	Use as a library
A Rust developer integrating spiking models into a larger system	A pure-Rust, no-I/O crate that drops cleanly into anything (wasm, FFI, embedded sim, server)	Use as a library
A PyTorch / SNN-ML researcher	A fast event-driven runtime for inference / online learning, complementing surrogate-gradient training in snnTorch / Norse / BindsNET / SpikingJelly	Vision
A Hebb desktop / visualizer contributor	The substrate the app depends on. New neuron / synapse / format work lands here.	Repository layout

Features

• Neuron models — LIF, Izhikevich, AdEx, Hodgkin–Huxley (with internal substepping for production dt).
• Plastic synapses — STDP and dopamine-gated R-STDP, with parameters streamable through the on-disk format.
• Deterministic seed generators — random, ring, small-world, layered. Same seed → same network.
• Embed-anywhere — pure-Rust, no I/O, no async runtime, no unsafe. WASM-ready. Filesystem support is opt-in behind the disk feature.
• Python bindings — import hebb_py; same engine, same domain types, same on-disk format.

Repository layout

• src/ — the Rust library (hebb crate, published on crates.io).
• python/ — PyO3 bindings, built with maturin and published as hebb-py on PyPI. Python module name is hebb.
• tests/ — integration tests against the public Rust API.
• SCHEMA.md — on-disk format spec for .cortex/ folders (gated behind the disk feature).

Use as a library

Install:

# Rust
cargo add hebb

# Python
pip install hebb-py

Both the PyPI distribution and the Python module are hebb-py / hebb_py. The bare hebb name on PyPI is taken by an unrelated astronomy package, and import hebb_py avoids colliding with it.

Drive the substrate from Python:

import hebb_py

sim = hebb_py.Sim()
a = sim.add_neuron()
b = sim.add_neuron()
sim.add_edge(a, b, weight=0.9)
sim.stimulate(a, current=50.0, duration_ms=30.0)

spikes = sim.run(dt_ms=1.0, n_steps=200)   # -> [(neuron_id, t_ms), ...]

…or from Rust:

use hebb::SimEngine;
use uuid::Uuid;

let mut sim = SimEngine::new();
let a = Uuid::new_v4();
let b = Uuid::new_v4();
sim.add_neuron(a);
sim.add_neuron(b);
sim.add_edge(Uuid::new_v4(), a, b, 0.9);
sim.inject(a, 50.0, 30.0);

for _ in 0..200 {
    let frame = sim.tick(1.0);
    for event in frame.events {
        println!("{} spiked at t={}ms", event.node_id, event.t_ms);
    }
}

For the on-disk .cortex/ folder format (open the same network you build here in the Hebb desktop app's visualizer), enable the disk feature:

hebb = { version = "0.1", features = ["disk"] }

Vision

The computational-neuroscience and SNN ecosystem is already rich: PyTorch-based ML frameworks (snnTorch, BindsNET, Norse, SpikingJelly, Rockpool) train spiking networks with surrogate gradients; biophysical simulators (NEST, Brian2, NEURON, GeNN, Arbor) model networks down to morphology; neuromorphic platforms (Intel Loihi / Lava, SpiNNaker, BrainChip Akida, SynSense) run trained networks on event-driven hardware. Each of these is excellent in its niche, but they don't talk to each other, and very few of them ship as an embeddable, no-runtime-deps crate you can drop into a larger system.

hebb aims to be a lightweight, interchange-friendly runtime that complements those tools rather than competes with them. The roadmap (loose ordering, contributions welcome):

• PyTorch interop. Round-trip with snnTorch / Norse / SpikingJelly: train weights with surrogate gradients in PyTorch, export to hebb for sparse event-driven inference and online plasticity. Inverse path too — initialize a hebb network from a .pth checkpoint.
• Model-description format support. Importers/exporters for NeuroML, SONATA (Allen Brain / BMTK), and (eventually) PyNN models. The goal is that a network defined for NEST or NEURON can run a sparse simulation under hebb without rewriting.
• Experimental-data ingestion. Stream NWB files into a stimulator so you can replay recorded spike trains against a learning network.
• Event-camera datasets. First-class support for DVS / N-MNIST / N-Caltech101 / DVS-Gesture so event-driven workloads are easy to wire up.
• Neuromorphic targets. Long-term: emit a hebb network to Intel Loihi (via Lava IR), SpiNNaker, or Akida. Short-term: keep the data layout and event-loop semantics close enough to those platforms that the mapping is mechanical.
• More neuron and plasticity models. Multi-compartment neurons; e-prop / equilibrium-propagation rules; calcium-based plasticity; reward-modulated variants.
• Performance. SIMD-friendly batched ticking; GPU backend for dense-region simulation (likely via wgpu); deterministic parallel ticking.
• WASM. Run hebb in the browser so the visualizer (and arbitrary educational toys) can simulate without a server.

If you maintain one of the ecosystem tools above and want to talk about interop, please open an issue.

Local development

# Rust
cargo test --features disk

# Python bindings (requires maturin)
pip install maturin
maturin develop --features pyo3/extension-module
python -c "import hebb_py; print(hebb_py.__version__)"

Contributing

• CONTRIBUTING.md — workflow basics (coming soon).
• PRs are welcome. Keep them small and single-purpose. New neuron models go in src/domain/, new plasticity rules in src/domain/synapse_plastic.rs (or a new sibling file).

License

hebb has an Apache 2.0 license, as found in the LICENSE file.