Improving Deep Learning Optimization through Constrained Parameter Regularization