Reproduce “The Hyperspherical Geometry of CLIP Latent Space: A Semantic Mixture Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zijie Yu, Gaowen Liu, Ramana Rao Kompella, Philip S. Yu, Yue Song
MovMF-CLIP whitens CLIP image embeddings to reduce global anisotropy, then fits multiple directional distributions to normalized vectors. It yields density scores and soft component assignments. On MS-COCO versus OpenImages, FPR95 is 48.00% versus 67.76% for W-CLIP on full in-distribution data, and 33.48% versus 75.05% on tail-only data. COCO validation informs density estimation, so these are not held-out results.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.