This is chapter three of 3Blue1Brown’s deep learning series and it explains backpropagation, the algorithm that computes the gradients gradient descent depends on. Grant Sanderson builds the intuition from the output backward: each training example wants certain weights to go up and others to go down, and backpropagation is the bookkeeping that adds up all of those competing demands across the network.
The video frames the algorithm as distributing responsibility for an error back through the layers, showing how a change in any one weight ripples forward to affect the final output. It keeps the heavy chain-rule mechanics for a companion video and focuses here on the picture that makes the math feel inevitable rather than arbitrary.
Backpropagation is the single idea that made training many-layered networks practical, and it is notoriously hard to convey clearly. By grounding it in a concrete, visual story of how blame flows backward, this explainer gives newcomers a genuine understanding of how learning happens inside a neural network. It is a natural follow-on to the channel’s chapter on gradient descent.