报告题目：Theoretical Insights into Self-Supervised Contrastive Learning

报告时间：2024-05-15  10:00-11:00

报 告 人：黄维然 副教授(上海交通大学)

腾讯会议ID：784 515 754

Abstract：Self-supervised learning has recently attracted significant attention because it requires only unlabeled data for model training. Contrastive learning, a popular method within self-supervised learning, has demonstrated promising empirical performance. However, the theoretical understanding of its generalization ability remains limited. In this talk, I will analyze self-supervised contrastive learning from a theoretical perspective, demonstrating that its generalization ability is linked to three key factors: the alignment of positive samples, the divergence of class centers, and the concentration of augmented data.  Moreover, I will show that self-supervised contrastive learning fails to learn domain-invariant features, thereby limiting its transferability. To address this issue, I will introduce Augmentation-robust Contrastive Learning (ArCL) and show how it significantly improves the transferability of self-supervised contrastive learning.