A step which was taken a long time ago and does not seem to have played much of a role in recent developments; for the most part, people don't bother with extensive hyperparameter tuning. Better initialization, better algorithms like dropout or residual learning, better architectures, but not hyperparameters.