An approach that enables large-batch contrastive learning under memory constraints, similar to gradient accumulation.