Home

SynapticDistill.jl

Generated with Grok Build: Grok 4.5 · xAI Imagine (/imagine)

Modular online training for spiking neural networks in Julia — E-prop, OTTT, and more.

Version: 0.2.0 | Julia: 1.8+ (CI: 1.12) | License: MIT OR Apache-2.0

Framework-agnostic training utilities for pure SNNs or hybrid teacher–student setups. Callers inject a model step and loss; the package orchestrates a single train_step! and provides surrogate helpers and rule stubs.

Core philosophy

Principle	Meaning
Bring any model step	No global forward — inject (model, spikes::SpikeBatch) -> output
Bring any loss	Pure SNN MSE or distillation CE from a frozen teacher outside the package
Apply any rule	:eprop, :ottt (stubs today); extend with new rules
Update only the SNN	Hybrid teachers stay frozen in caller code

Public API

Symbol	Role
SpikeBatch	Spike trains + optional times/targets
TraceBatch	Eligibility-trace container
TrainingState	loss, metrics, traces, gradients
ModelStep	Abstract type for typed callables
train_step!	One online step (Zygote through model step + loss)
surrogate_heaviside / surrogate_sigmoid / surrogate_exponential	Surrogate ∂spike/∂v

Start Here

Page	Description
Overview	Wiki map + DeepWiki outline
Getting Started	Install and examples
Architecture	Modules and pipeline
Data Structures	SpikeBatch / TrainingState
Training Pipeline	train_step! contract
Learning Rules	e-prop / OTTT status
Surrogate Gradients	Surrogate helpers
Hybrid Distillation	Teacher–student pattern
Examples and Scripts	examples/ + scripts/
Benchmarks	eprop tick harness
Testing	Unit suite
CI and Quality	GHA
Ecosystem	Sibling packages
Glossary	Terms
Status and Roadmap	What is / is not implemented

Quick start

using SynapticDistill, Statistics

mutable struct MySNN
    weights::Matrix{Float32}
end

model = MySNN(rand(Float32, 10, 10))
batch = SpikeBatch(Float32.(rand(0:1, 10, 100)), nothing, nothing)

model_step(model, spikes::SpikeBatch) = (logits = model.weights * vec(mean(spikes.spikes; dims=2)),)
mse_loss(output) = sum(output.logits .^ 2)

model, state = train_step!(model, batch, mse_loss; forward_fn=model_step, rule=:eprop)

See examples/pure_snn_training.jl.

License

Dual-licensed under MIT or Apache-2.0 at your option.

---

Last updated: July 29, 2026 Updated by: Grok Build: Grok 4.5 Package tip reference: e986a66 (main, v0.2.0) Devin DeepWiki: commit e986a66 · SynapticDistill.jl