Improving robustness of deep neural networks to amplitude pre-processing: Introduction of a “reweighting” layer.