Post-Hoc Feature Selection Layer for Neural Networks Interpretability