Backward Pass Overview
The backward pass asks how the loss changes with earlier values.
If:
L = (a - y)^2
then:
dL/da = 2(a - y)
This gradient moves backward to the operation that produced a.
If:
a = relu(z)
then:
dL/dz = dL/da * da/dz
That is the pattern of backpropagation:
upstream gradient x local derivative = downstream gradient
Here "downstream" means closer to the inputs and parameters.
Exercise: Loss derivative
Let L = (a - y)^2, with a = 5 and y = 2. What is dL/da?
Compute it first, then check your number.
HintUse the derivative
dL/da = 2(a - y).
SolutionWork it out
dL/da = 2(5 - 2) = 6.
Exercise: Multiply by local derivative
If dL/da = 6 and da/dz = 1, what is dL/dz?
Compute it first, then check your number.
HintChain rule
Multiply the upstream gradient by the local derivative.
SolutionWork it out
dL/dz = dL/da * da/dz = 6 x 1 = 6.