Evaluating Relaxations of Logic for Neural Networks: A Comprehensive Study