A technique that accumulates and corrects for errors from previous steps to improve convergence in distributed training.