A network learns by being wrong. It guesses, measures the miss, then works backward to ask how much each connection contributed. Backpropagation is the bookkeeping for that question: each weight's share is a product of small local numbers, multiplied along the path from the output back to that weight. Then every weight moves a little, and the network guesses again. Tap any edge to see its product. Tap a point to make it the step example.