Learning Rules
Generated with Grok Build: Grok 4.5 · xAI Imagine (/imagine)
Public training entry: train_step!(…; rule=:eprop | :ottt).
Internal helpers (included, not all re-exported):
| File | Function | Status |
|---|---|---|
rules/eprop.jl |
update_eprop! |
Placeholder (dummy gradients / prints) |
rules/ottt.jl |
update_ottt! |
Placeholder |
rules/predictive_eprop.jl |
— | Empty placeholder; not exported |
rules/surrogate.jl |
surrogates | Implemented — see Surrogate Gradients |
- Maintain per-synapse eligibility traces with decay
trace_lambda(default0.95f0in stub kwargs). - Learning signal from loss / error.
- Gradient ∝ learning_signal × eligibility.
- Presynaptic / postsynaptic traces.
- Updates correlated with a global learning signal (loss).
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