A key difference to standard LTH work is that You et al.
prune complete convolution channels primarily based on their batch normalization scaling issue.