# Why to use moving statistics instead of population statistics in batch normalization implementation at inference?

Many (tensorflow) implementations of batch normalization seem to use moving statistics at its inference phase (R2RT, tf.contrib.layers.batch_norm).

But in the algorithm from the original paper, they use population statistics in the batch normalization function.

The thing is, pardon me, I don't understand why most of the implementations do not precisely follow the original algorithm.

The paper itself very briefly mentions about 'moving average' but does not give specific instructions for algorithm implementations.