Researchers have shown that imposing hard constraints on the direction of input-output effects during neural network training ...