Each example’s weight changes with its predicted probability, decreasing for confident examples as training progresses.