Discovering Global False Negatives On the Fly for Self-supervised Contrastive Learning

Name: Discovering Global False Negatives On the Fly for Self-supervised Contrastive Learning
Uploaded: 2026-10-11T07:34:29.420469+00:00
Duration: 293 s
Description: GLOFND automatically learns dynamic thresholds for each anchor data point during training to identify false negatives in self-supervised contrastive learning. This addresses the problem where similar negative pairs are incorrectly separated, which harms representation quality. GLOFND improves existing contrastive learning methods with minimal computational overhead.

GLOFND automatically learns dynamic thresholds for each anchor data point during training to identify false negatives in self-supervised contrastive learning. This addresses the problem where similar negative pairs are incorrectly separated, which harms representation quality. GLOFND improves existing contrastive learning methods with minimal computational overhead.

Open the full episode