An MLP learns a 2D classification boundary through backpropagation — the decision surface warps over epochs, weights visualize as connection thickness and color, loss descends as the network finds the pattern.