How a Neural Network Learns

Give it a number, an array, or a matrix. Watch the weights become it — one step at a time.

June 22, 2026 · interactive

→

step size0.15 loss step 0loss 0

Surfaces

data

learning rate0.80 hidden activation

1Forward

2Loss

3Backward

4Update

Decision surface + data. Click to move the probe (◉).

epoch 0 loss — accuracy —

positive weight / class 1 negative weight / class 0

Transformers

encoder-only all tokens see each other → classify / embed decoder-only tokens see the past → predict the next token encoder–decoder read one sequence → write another

training datathe cat chased the mousethe dog chased the catthe mouse escaped the catthe dog escaped the mouse

learning rate0.06 attention focus1.00

example

updates 0loss — parameters —