MicrogradEx.Value.Edge (MicrogradEx v0.1.0)

A single local derivative from one operation output back to one parent value.

In the Python version each Value stores a _backward closure that mutates its parents. This struct is the functional replacement for those closures. An edge says:

d(output) / d(parent) = local_gradient

During backpropagation we multiply that local gradient by the gradient that has already reached the output node. That is the chain rule in one line.

Summary

Types

t()

Types

t()

@type t() :: %MicrogradEx.Value.Edge{
  local_gradient: float(),
  parent_id: pos_integer()
}