Surrogate Gradients
Generated with Grok Build: Grok 4.5 · xAI Imagine (/imagine)
Module: src/rules/surrogate.jl — exported and covered by tests.
| Function | Form | Default param |
|---|---|---|
surrogate_heaviside(x, γ=10) |
γ * max(0, 1 − |x|) |
γ = 10 |
surrogate_sigmoid(x, β=1) |
sigmoid derivative-like | β = 1 |
surrogate_exponential(x, α=1) |
α * exp(−α|x|) |
α = 1 |
Broadcast over membrane arrays with dotted calls, e.g. surrogate_heaviside.(v).
Use inside custom model steps when differentiating through spike thresholds with Zygote-friendly surrogates.
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