Backpropagation
The forward pass makes a prediction; backprop decides how much to blame each weight. Follow one blame path from loss to weight.
Fixed 2 → 2 → 1 sigmoid net, one example, loss L = (yhat − y)² / 2. Edge width carries |weight|; the bright path is the one being blamed.
Try this
- Step to Backward and read one full chain left to right.
- Crank the learning rate to 2 and watch the update overshoot.
- Switch to example B and see which gradients nearly vanish.
How backpropagation works here
Backpropagation is just the chain rule, applied layer by layer: the loss's sensitivity to each weight factors into the prediction error times the local slopes along its path. This lab shows one complete chain at a time on a fixed tiny net, with exact numbers.
Computing the gradient is only half the story — the update step then moves every weight against its gradient, scaled by the learning rate. Watch the new loss: a good step shrinks it, and too big a step can grow it.
Learn the ideas behind the update
ManySpace turns AI mechanics like this into guided lessons.