Backpropagation

Gradient descent needs a gradient for every weight; backpropagation is how the network computes them. Send the cross-entropy error backward through the same tiny MLP and watch the chain rule hand each weight its share of the blame.

1 exhibit Computed teaching network Offline, single file

StudioStart hereComputed
Backprop Studio

Pick the target token, then step the error backward: p − y at the logits, gradients for the output weights, the ReLU gate deciding which hidden neurons learn, and a final gradient — plus an update preview — for every weight.
Single screen · SVG + MathML→

Backpropagation

Backprop Studio