Contrastive Learning with Uncertainty

1. A SIMPLE FRAMEWORK FOR UNCERTAINTY IN CONTRASTIVE LEARNING (2020, 9회 인용)

• 기존 대조학습 encoder는 uncertainty or confidence에 대한 언급 없이 contrastive model을 구현함
• 본 논문은 uncertainty for pretrained contrastive representations을 구하기 위해 “contrasting distributions”을 기반으로 함
• 제안한 uncertainty을 EMBEDDING INTERPRETABILITY와 ANOMALY DETECTION와 OUT-OF-DISTRIBUTION IMAGE DETECTION에 적용함
  
  
  
  

2. UNCERTAINTY IN CONTRASTIVE LEARNING: ON THE PREDICTABILITY OF DOWNSTREAM PERFORMANCE (Netflix 2022, 2회 인용)

• In this work, we study whether the uncertainty of such a representation can be quantified for a single datapoint in a meaningful way
• In other words, we explore if the downstream performance on a given datapoint is predictable, directly from its pre-trained embedding.
  

3. Probabilistic Contrastive Learning Recovers the Correct Aleatoric Uncertainty of Ambiguous Inputs (ICML 2023)

• 현실에서는 이미지가 흐릿하거나 2d나 3d만 보여준다던지 본질적인 ambiguities가 존재함.