A technique to reduce gradient variance during training by maintaining reference statistics and reweighting samples.