Conceptual
Browse Instruct Curate

Generalization Analysis of Difficult-to-Learn Examples in Contrastive Learning

A theoretical framework for spectral contrastive learning that models pairwise sample similarity to explain why difficult-to-learn examples (points near the decision boundary) degrade the generalization of unsupervised contrastive learning. Students learn how removing such examples, along with margin tuning and temperature scaling, tightens generalization bounds and improves downstream classification, and how to select difficult-to-learn examples in practice.