(1) The paper presents a method called LASO for few-shot multi-label learning that combines pairs of examples in feature space to synthesize new examples corresponding to label set operations like intersection and union. (2) LASO trains a network to predict these synthesized label sets using losses that encourage the network to learn label set relationships. (3) The method is evaluated on multi-label datasets like COCO and CelebA, outperforming baselines on classification and retrieval metrics for few-shot learning settings.