Improving Contrastive Learning on Imbalanced Seed Data via Open-World Sampling

note