Unit 4: What backpropagation is doing
About this unit
Backpropagation turns one example's output errors into nudges for every weight and bias, walking backwards a layer at a time; averaging those nudges over a mini-batch gives the step that stochastic gradient descent takes.
By the end, you can: Explain how one example's output errors become nudges for every weight and bias, layer by layer, and how averaging them over mini-batches trains the network.
Modules
5 modulesLearn
- One example's error reportArticle · An output's error, and the three levers that move it: bias, weights scaled by input activations, and the previous layer.Not started
- Backwards, then averagedArticle · Repeating the reasoning layer by layer, averaging nudges over examples, and mini-batches as stochastic gradient descent.Not started
- Watch the gradient flow backExperiment · Step a single neuron's computation graph forwards, then backwards to its weight and bias.Not started
- The whole methodReview · Error reports, levers, the backward walk, averaging and mini-batches in one place.Not started