Robust Contrastive Learning Against Noisy Views

By Ching-Yao Chuang et al
Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
2. Related Work
3. Prelim: From Noisy Labels to Noisy Views
4. Robust InfoNCE Loss
5. Intuition behind RINCE

Summary

Contrastive learning relies on an assumption that positive pairs contain related views, but what if this assumption is violated? In this work, a new contrastive loss function, RINCE, is proposed to be robust against noisy views. The proposed loss function is modality-agnostic and provides consistent improvements over the state-of-the-art on various benchmarks. The paper discusses the importance of designing the right contrasting views and addresses the issue of deteriorating effects of noisy views in contrastive learning. It introduces a theoretical analysis and demonstrates the robustness of RINCE through empirical evidence on different modalities and noise types.
×
This is where the content will go.