How a network computes, why gradients vanish, and what makes depth trainable
calc2.24 already did most of this lesson's work. Given a dense layer z = Wx + b and the upstream gradient δ = ∂L/∂z, it derived all three gradients the layer needs: ∂L/∂x = Wᵀδ, ∂L/∂W = δxᵀ, ∂L/∂b = δ. Those are not being re-derived here.
🔒 This is a Pro lesson — the interactive figure, worked examples, quiz and practice open with Pro access.
▶ Backprop Through a Layer Block: the Activation's Half