Unit 5: Backpropagation calculus
About this unit
Each derivative backpropagation needs is a product of three small local derivatives, and the derivative with respect to a previous activation hands the same calculation to the layer before.
By the end, you can: Use the chain rule to compute a network's derivatives with respect to its weights, biases and earlier activations, one layer at a time.
Modules
5 modulesLearn
- The chain behind one weightArticle · z, a and C for a one-neuron chain, and the chain-rule products for ∂C/∂w, ∂C/∂b and ∂C/∂a_prev.Not started
- One weight's derivativeWalkthrough · Compute C, ∂C/∂w and ∂C/∂b for the last neuron from concrete values.Not started
- One layer further backWalkthrough · Turn ∂C/∂a_prev into the derivative of an earlier weight.Not started
- The formulas in one placeReview · The three local formulas, the recursion, and the sum over paths with many neurons.Not started