Q9 — type the backprop steps you would write on the exam.

Answer

1. forward pass: O = f(sum w*x), layer by layer 2. output delta: d_i = (target-O)*O*(1-O) 3. hidden delta: d_h = O*(1-O)*sum(d_j*w_hj) 4. update: w += c * delta * input (gradient descent)

Forward to get outputs, compute output error, propagate blame backward to hidden weights via the chain rule, then a gradient-descent step on every weight. Repeat for many epochs.