References: What backpropagation is really doing

Source material

Source curriculum (structural mirror, cited as further study):
• 3Blue1Brown, Neural Networks, Chapter 4: "What is backpropagation really doing?"
  Creator: Grant Sanderson (text adaptation by Josh Pullen)
  Lesson page: https://www.3blue1brown.com/lessons/backpropagation
  Series index: https://www.3blue1brown.com/?topic=neural-networks
  License: copyright Grant Sanderson; videos published on his site and YouTube
This lesson mirrors the intuition chapter on backpropagation (the companion
calculus chapter is mirrored in the next lesson). Note: live 3B1B Chapter 3
("Analyzing our neural network", at /lessons/neural-network-analysis) sits
between the gradient-descent chapter (Ch2, mirrored in T11 lesson 7) and this
backpropagation chapter (Ch4). T11 deliberately does not mirror Ch3 as a
standalone lesson; its central insight, that a trained network's hidden layers
do not cleanly detect "edges then loops" as the hopeful story suggests, is
folded into lesson 2 as the "hold the edges-to-loops story loosely" caveat.
Live Ch3-5 numbering verified 2026-05-25. Clawdemy's lessons are original
prose that follows the pedagogical arc of this series. We do not reproduce or
transcribe the videos; we cite them as the recommended companion. All rights
to the original videos remain with the creator.

Watch this next

What is backpropagation really doing? (3Blue1Brown) by Grant Sanderson. The chapter this lesson mirrors. It animates exactly the move described here: an output neuron’s desire to change, the three ways to grant it, and the wishes fanning backward and summing into the previous layer’s desires. Seeing the backward flow in motion is the fastest way to make “propagate the desires” feel concrete.

Going deeper

A short, durable list. Each link is a specific next step, not a generic pile.

Neural Networks and Deep Learning, Chapter 2 (“How the backpropagation algorithm works”) by Michael Nielsen. The same algorithm explained from the ground up, building toward the equations. The natural deeper read once the intuition here has landed and you want the mechanism the next lesson begins.
TensorFlow Playground. Every time you press play, backpropagation is running invisibly under the hood: each step computes the gradient by a backward sweep, then takes a gradient-descent step. You have now seen what that hidden computation is actually doing.

Adjacent topics

Where this leads inside this track.

Gradient descent (lesson 7). The previous lesson assumed the gradient was available and used it to take steps. This lesson supplied the missing half: where that gradient comes from. The two together are the complete training loop.
Backpropagation and the chain rule (lesson 9). This lesson stayed at the level of “desires” and “wishes.” The next one opens the hood: those desires are derivatives, and the backward sweep is the chain rule from calculus applied layer by layer. If you took Track 8, the chain rule will be familiar machinery.