Zeroes out extremely large values for robustness to data corruption on
clients, clips to moderately high norm for robustness to outliers. After
weighting in mean, the weighted values are uniformly quantized to reduce the
size of the model update communicated from clients to the server. For details,
see Suresh et al. (2017)
http://proceedings.mlr.press/v70/suresh17a/suresh17a.pdf The default
configuration is chosen such that compression does not have adverse effect on
trained model quality in typical tasks.