If dropout limits trainable depth, does critical initialisation still matter? A large-scale statistical analysis on ReLU networks

Re-assign community