Research paper questions shared temperature assumption in contrastive learning

A new research paper explores the relationship between shared temperature and angular scale in probabilistic contrastive learning, particularly within high-dimensional settings. The study, using the von Mises-Fisher (vMF) probabilistic score, demonstrates that a shared temperature does not necessarily imply a shared similarity scale for class representations. The research accurately predicts changes in decision boundaries and feature gradients on real-world datasets like CIFAR-LT and ImageNet-LT, suggesting that vMF concentration plays a crucial role in both decision-making and learning scales in this domain. AI

IMPACT This research could refine how contrastive learning models are trained and evaluated, potentially leading to more robust and accurate representations in high-dimensional data.

RANK_REASON The cluster contains an academic paper detailing novel research findings in machine learning. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• CIFAR-10
• CIFAR-100
• CIFAR-LT
• ImageNet-LT
• iNaturalist
• von Mises-Fisher distribution

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →