How far away are truly hyperparameter-free learning algorithms?