Question 29
UnclassifiedBackpropagation computes gradients in a neural network by:
Correct answer: B
Explanation
Backpropagation uses the chain rule to compute how the loss changes with each parameter by moving backward through the computation graph. This is why it "propagate[s] gradients backward through the computation graph," allowing each layer to receive its gradient from later layers.
Why each option is right or wrong
A. Estimating gradients with finite differences by perturbing each parameter slightly
B. Applying the chain rule to propagate gradients backward through the computation graph
Backpropagation is the standard reverse-mode differentiation procedure for a computational graph: it evaluates partial derivatives at each node and combines them using the chain rule so the gradient of the loss with respect to earlier parameters can be obtained from later ones. In a feedforward network, this means the error signal is passed from the output layer back through each preceding layer, with each layer’s contribution computed from the derivatives of the operations that produced it.
C. Sampling random directions and averaging their effects on the loss
D. Solving a system of linear equations for each layer's parameters