arXiv paper questions cosine similarity's effectiveness in AI representations

A new arXiv paper titled "Semantics at an Angle: When Cosine Similarity Works Until It Doesn't" critically examines the widespread use of cosine similarity in machine learning representations. The paper, authored by Kisung You, argues that while cosine similarity is computationally convenient and effective for normalized embeddings, its limitations arise from its inability to capture radial variation and its potential mismatch with downstream decision-making objectives. The authors propose a review of alternative methods, including geometry-aware, hubness-aware, and norm-aware approaches, suggesting that the utility of cosine similarity is conditional on the specific representation, processing, scoring rule, and decision context. AI

IMPACT Highlights potential limitations of a widely used technique in AI representation learning, suggesting a need for alternative methods.

RANK_REASON Research paper published on arXiv discussing limitations of a common machine learning technique. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• computer vision
• cosine similarity
• Hugging Face
• information retrieval
• Kisung You
• multimodal learning
• natural language processing

• paper
• other

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