Department of Applied Statistics, Yonsei University
TITLE
Optimizing Embeddings using Contrastive Loss
ABSTRACT
Contrastive loss has emerged as a prevalent choice for optimizing embeddings across diverse contemporary applications. In this presentation, I will explore recent papers that offer theoretical analyses of optimal embeddings trained using contrastive losses, with a particular focus on InfoNCE loss and sigmoid loss. This examination will provide valuable insights into the underlying principles and effectiveness of these optimization techniques in the field of machine learning and representation learning.