Why don't we simply use running statistics for batch normalization during training? - View it on GitHub