Effective Regularization Through Loss-Function Metalearning