core Estimated learning time: 6 h

6.2 Backpropagation by hand

You can derive gradients for a two-layer net on paper.

Before:05. Classical Machine LearningUnlocks:07. Natural Language Processing11. The ECE Bridge — Signals, Edge & Embedded AI

Backpropagation on paper for a two-layer network is the rite of passage: chain rule over the graph, local gradients, accumulation — and after it, framework errors about shapes and graphs become legible forever. It follows the perceptron topic because the mechanics need a network to run on. The safety habit taught alongside is gradient checking: a numerical finite-difference comparison catches sign and transpose errors that reread derivations miss.

Work through these

  • Chain rule over the computational graph

    Training is the chain rule applied over a graph of small operations, accumulating derivatives from the loss backwards. This is the entire algorithm, and it is worth deriving rather than accepting.

  • Gradients for common layers and losses

    The derivative expressions for the layers and losses you will actually use. Having these at hand is what makes hand-derivation feasible on a real network.

  • Implementing backprop in NumPy

    Writing the backward pass yourself in array code, without a framework doing it for you. Everyone who does this once reads framework code differently afterwards.

  • Gradient checking

    Comparing your analytic gradients against numerical estimates catches derivation errors that otherwise show up as a model that trains slowly for no visible reason. It is a short check with a large payoff.

Sign in to keep your progress.

Free resources

Links last checked 29 Aug 2026.

Stuck here?

Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.

Checking your session…

Topics shown in module order.