a visual course

How a Network
Actually Learns

(Backpropagation)

How a network turns 'that's wrong' into a fix for millions of weights — the backprop algorithm, visualized.

A free, visual course on backpropagation — how neural networks actually learn. From loss and the gradient to the chain rule, the backward pass, and the gradient-descent update that trains every modern model.

in the real world

Backprop is the one algorithm that trains almost everything — image models, LLMs, diffusion, speech and game-playing AI. Learn it and you understand how AI learns at all.

Start with a wrong guess →

Part 1 — The setup

1

A guess, and how wrong it is

The network predicts, a loss measures the error — and now millions of weights need tuning. But which way?

2

Downhill on the loss

Wiggle one weight and watch the loss change — that slope is the gradient, and it points the way.

Part 2 — The trick

3

The chain rule

A deep weight affects the loss through a chain of steps — so multiply the local slopes along the way.

4

The backward pass

Run forward once, then send the error backward layer by layer — one pass computes every weight's gradient.

Part 3 — Learning

5

Take a step, repeat

Nudge every weight against its gradient, scaled by a learning rate — then do it a million times.

Part 4 — Out in the world

6

Where backprop shows up

The same algorithm trains image models, LLMs, diffusion, speech and game-playing AI — essentially everything.