Akin to the behavior of a mini-batch in momentum SGD training, current block contributes to all the following global model updates, and noisy components of model-update resulting from current block will be attenuated while informative ones will be enlarged. We use the following formula