Reproduce “Boosting Contrastive Self-Supervised Learning with False Negative Cancellation” on OpenResearch

About this paper

Tri Huynh, Simon Kornblith, Matthew R. Walter, Michael Maire, Maryam Khademi

False Negative Cancellation (FNC) improves contrastive self-supervised visual learning by identifying likely same-class images that are incorrectly used as negatives. Using additional support views and similarity-based screening, the method either removes these samples from the negative set or promotes them to positives; the full system reaches 74.4% ImageNet top-1 accuracy, improves low-label and transfer performance, and can achieve comparable accuracy to long-trained SimCLR with substantially less pretraining.

Key insights

• Support views are particularly important for attraction because a candidate may resemble only one crop or viewpoint of the anchor; maximum aggregation preserves this signal better than averaging.
• Even removing roughly two suspected false negatives from thousands of candidates can improve representation quality, while attraction generally provides larger gains but is more sensitive to detection errors.
• False-negative cancellation complements multi-crop augmentation and momentum encoders, adding semantic links between different images rather than only enforcing invariance across augmentations of the same image.

Summary by alphaXiv.