complexity of the supervised learning of nonlinearity compensation algorithms. We
demonstrate both numerically and experimentally that the augmentation allows reducing the
training dataset size up to 6 times while keeping the same post-compensation bit-error rate.