Technique to limit gradient magnitude during training to stabilize optimization.