Architecture

Architecture

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

Module layout

SynapticDistill.jl/
├── src/
│   ├── SynapticDistill.jl   # module entry, exports
│   ├── types.jl             # SpikeBatch, TraceBatch, TrainingState
│   ├── training.jl          # ModelStep, train_step!, validate_model_step
│   ├── utils.jl
│   └── rules/
│       ├── surrogate.jl     # exported surrogates
│       ├── eprop.jl         # update_eprop! (placeholder)
│       ├── ottt.jl          # update_ottt! (placeholder)
│       └── predictive_eprop.jl  # placeholder (not exported)
├── examples/
├── scripts/                 # standalone tools (e.g. spikenaut_train)
├── benchmarks/
└── test/

Train-step data flow

SpikeBatch ──► model_step(model, batch) ──► output
                                              │
                                         loss_fn(output)  ──► scalar loss
                                              │
                                    Zygote.withgradient(model)
                                              │
                                    rule branch (:eprop / :ottt)
                                              │
                                    TrainingState(loss, gradients, …)

The package does not own neuron models, datasets, IPC, or teacher LLM execution — inject those at the boundary.

Related

• Training Pipeline
• Data Structures

---

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