Quick reference

Two-layer backpropagation

For h = w₁x, o = w₂h, and L = (o − y)²:

dL/do = 2(o − y) dL/dw₂ = (dL/do) × h dL/dh = (dL/do) × w₂ dL/dw₁ = (dL/dh) × x wᵢ ← wᵢ − η(dL/dwᵢ)

The output weight receives the error signal directly. The first-layer weight receives it through the output layer, which is why its gradient includes w₂. This is “credit assignment”: how much each earlier component contributed to the final error.

Primary source: Rumelhart, Hinton & Williams (1986).