The effect of mislabeled samples on the performance of the linear learning machine