GloFND Automatically Identifies Global False Negatives in Contrastive Learning

GloFND is an optimization-based method that learns per-anchor thresholds to identify semantic false negatives during self-supervised training. Unlike mini-batch methods, it detects false negatives globally across the entire dataset without increasing per-iteration computation costs.