# Back\-Propagation Line Animation

One bundle of glowing lines becomes a loss landscape, a neural network, a rolling optimizer and a dense network to show how back\-propagation teaches a network

By greenlig

Topic: Computing, AI & Electronics

[Open the interactive Jig](https://ask.jig.com/j/KyERg3yppVU5)

## Guide

### 1. One bundle of light

Every picture in this Jig is drawn by the same 640 glowing threads\. Watch them twist and reshape to show how a neural network learns from its mistakes\. The trick is called back\-propagation\.

Drag empty space to turn the drawing\. Scroll or pinch to zoom\.

### 2. A landscape of mistakes

Imagine every setting of a network's knobs, called weights, spread out like a map\. The height above each spot is the loss: how wrong the network is with those settings\. Learning means finding the lowest valley\.

Real networks have millions of weights, so their landscape has millions of directions\. This is a two\-knob cartoon of it\.

### 3. A network makes a guess

A neural network is layers of simple units joined by connections of adjustable strength\. A guess flows forward, layer by layer, from left to right\. At the end, the loss neuron measures how far the guess was from the right answer\.

### 4. Blame flows backward

Back\-propagation sends the error backward from the loss through every layer\. A maths rule called the chain rule lets each connection work out how much it added to the mistake\. Orange means turn this weight down; blue means turn it up\.

One backward pass costs about as much as one forward pass, however many connections there are\.

### 5. Rolling downhill

The gradient is the slope under the ball, and it points uphill\. The optimizer, the part that changes the weights, steps the other way, like a ball rolling into a valley\. Momentum, the speed it has built up, carries it over small bumps until it settles in a minimum\. Drag on the landscape to drop the ball somewhere new\.

Start it in the small hollow on the left and it can get stuck there: a local minimum\.

### 6. Bigger networks, same trick

Real networks are far denser, but they learn the same way: guess forward, send the blame backward, nudge every weight a little, and repeat\. A 1986 paper by Rumelhart, Hinton and Williams made this method famous\.

A small network that reads handwritten digits has 13,002 weights and biases\. Large modern ones have billions\.