I have seen many links about MA for batch normalization but nothing answered my question.

In Batch normalization, you get means and variance for each mini-batches in the training process. And the default process, you calculate through the moving average.

Those mini-batches are chosen randomly so there is no higher priority for the recent mini-batch. But why should we calculate the mean and variance for the evaluation process with MA and not just the normal average?


1 Answer 1


The mean and variance of each mini-batch varies and each one may not represent the original input distribution fully. Also in the presence of outliers, the statistics of mini-batch containing them might be biased towards the outliers. So in order to avoid equal weightage assignment to each mini-batch moving average is used.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.